Advances in Psychology Volume 77 Understanding Word & Sentences

UNDERSTANDING WORD AND SENTENCE

ADVANCES IN PSYCHOLOGY

77 Etlirol-.s:

G. E. STELMACH

P. A. VROON

N OKT tl- H O L I .AN L) AMSTERDAM * NEW YOI1K * OXFORD * I'OKYO

UNDERSTANDING WORD AND SENTENCE

Greg B. SIMPSON Utii\>crsityc~fNcht~u.sku at Oniuhu Oniuhu Nchruska. U.S.A, ~

NORTH-HOLLAND AMSTERDAM * N E W Y O K K OXFORD *TOKYO

NORTH-HOLLAND ELSEVIER SCIENCE PUBLISHERS B.V. Sara Burgerhartstraat 25 P.O. Box 21 I , 1000 AE Amsterdam, The Netherlands

Distributors for the United States and Canada: ELSEVIER SCIENCE PUBLISHING COMPANY, INC. 655 Avenue of the Americas New York, N.Y. 10010, U.S.A.

ISBN: 0 444 XX487 4 ELSEVIER SCIENCE PUBLISHERS B.V., 1991 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying. recording or otherwise, without the prior written permission of the publisher, Elsevier Science Publishers B.V.1 Physical Sciences and Engineering Division, P.O. Box 103. 1000 AC Am\sterdam. The Netherlands. Special regulations for readers in the U.S.A. - This publication has been registered with the Copyright Clearance Center Inc. (CCC). Salem, Massachusetts. Information can be obtained from the CCC about conditions under which photocopies of parts of this publication may be made in the U.S.A. All other copyright questions, including photocopying outside of the U.S.A.. should be referred to the copyright owner, Elsevier Science Publishers B.V., unless otherwise specified. No responsibility is assumed by the Publisher for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions or ideas contained in the material herein. Pp. 97-128: copyright not transferred. Printed in The Netherlands

To

Mary Margaret and May

This Page Intentionally Left Blank

vii

Preface Research concerning structure and processing in the mental lexicon has achieved central prominence within cognitive psychology and psycholinguistics. The importance of lexical processing for consideration of higher levels of language comprehension is taken for granted. Historically, however, much of the research on the lexicon originated not with an eye to understanding language processing, but rather as a way of studying semantic memory. That is, words are an obvious and convenient medium with which to examine human semantic processing. Of course, the relevance of semantic memory for language comprehension was also assumed, but only rarely addressed directly. Consequently, lexical research was for many years dominated by “priming” studies, which focus on the effects of processing one meaningful stimulus (most often a single word) on the subsequent recognition of another. Through this research, we have gained some understanding of how lexical information may be organized, but considerably less about the ways in which that information is used in understanding natural language. The picture has changed dramatically in the past several years, with proportionally less work devoted to word recognition per se, and more to exploring the role of the lexicon, its processes and output, in other aspects of comprehension. This volume represents an attempt to gather together the work of some of those researchers who are responsible for this shift of emphasis. The modern descendants of the earlier word recognition research are well represented, as is that research which emphasizes the place of lexical information in syntactic and pragmatic processing. The first several chapters extend the priming literature to explore more fully the effects of sentence context on word recognition. Tabossi, Schwanenflugel, and Kellas and colleagues all consider the roles of sentence constraint and the activation of featural information in word recognition. Tabossi and Kellas et al. focus also on the problem of lexical ambiguity (the processing of multiplemeaning words) a topic that has figured especially prominently in the debate over whether the various component stages of language processing are autonomous or interactive. O’Seaghdha examines more completely the problem mentioned above, namely, that sentence context research occupies a position between psycholinguistic and more general cognitive concerns. O’Seaghdha discusses his own research on the relation between syntactic and lexical information in word recognition, and also givcs consideration to some of the methodological issues often raised in this research. Gernsbacher and Faust discuss two general cognitive processes, the enhancement and suppression of information in memory, and the role that the latter plays in Comprehension. They argue that the efficiency of suppression processes is a key factor underlying individual differences in comprehension skill. Van Petten and Kutas dcscribe their research on event-related brain potentials (ERP) as indicators of lexical processing in sentence context. Their discus-

...

Vlll

Preface

sion entails examination of scvcral variables important to word recognition (e.g., word frequency, ambiguity, and sentence constraint), as well as an assessment of several prominent word recognition models in light of their behavioral and ERP data. Rayner and Morris also expand on the ambiguity research, considering not only word-level ambiguity, but also word sense, syntactic, and syntactic-category ambiguity. Their eye-movement studies indicate a discontinuity among these types of ambiguity, and provide insight into the different processing operations that occur at these levels of comprehension. Whitney and Waring complete the discussion of the effects of context on the activation of semantic information. The contexts they consider, however, range from the single word to the prose passage, and the activated information similarly ranges from the single word to the elaborative inference. Cacciari and Glucksberg provide a thorough discussion of the role that lexical information plays in understanding figurative language, specifically, idiomatic expressions. They present a taxonomy of idiom typcs, and consider the different contributions made by lexical information to each type. Wisniewski and Gentner review how the meanings of words are combined to yield new concepts. Thcir work makes clear that there is no unitary set of processes by which concepts are always combined, and that the mapping of word meaning to higher-level comprehension will be very complex. Oden and colleagues use their FuzzyProp framework to describe how listeners and readers identify a linguistic message despite the noise typically present in the signal. This is considered at both the lexical and sentence Icvcls, as contextual information combines with sensory information to yield the best match to the input message. Ferreira and Henderson, and Boland and Tanenhaus examine the role of lexical information in syntactic processing. Ferreira and Henderson consider how verb information is used in parsing sentences, and provide data showing that this information is used not in the initial syntactic analysis of a sentence, but rather in the reanalysis following a parsing error. Boland and Tanenhaus also provide a very thorough treatment of the kinds of information carried in lexical entries, and how these types of information are used in sentence parsing and Sentence interpretation. Finally, the paper by Swinney touches on a number of the issues raised in the preceding chapters. Specifically, the issue of indeterminacy (ambiguity) is addressed at two levels (co-reference assignment and lexical ambiguity), and data bearing on the nature and timing of context effects are reported. It is quite clear that the authors of the papers contained here are not in complete agreement on every issue. Had the authors gathered to present these papers, a series of lively debates would undoubtedly have ensued. This must always be the case in any collection that represents the new directions taken by a field. Indeed, it is most desirable. It is hoped that the papers herein will inspire continued debate, both among the authors, and their readers.

ix

Contents Dedication ................................................................................................................ v .. Preface ................................................................................................................... VII Acknowledgments ..................................................................................................xi ... Contributors .......................................................................................................... XIII 1. Understanding Words in Context Patrizia Tabossi

.....................................................................

1

2. Contextual Constraint and Lexical Processing .............................................. Paula J . Schwanenflugel

23

3. Contextual Feature Activation and Meaning Access ..................................... George Kellas, Stephen T. Paul, Michael Martin, and Greg B . Simpson

47

4. A Perspective on Sentence Context Research ................................................ Padraig G . O'Seaghdha

73

5. The Role of Suppression in Sentence Comprehension .................................. Morton Ann Gernsbacher and Mark Faust

97

6. Electrophysiological Evidence for the Flexibility of Lexical Processing ........................................................................................ Cyma Van Petten and Marta Kutas

129

7. Comprehension Processes in Reading Ambiguous Sentences: Reflections from Eye Movements ............................................. Keith Rayner and Robin K . Morris

175

8. The Role of Knowledge in Comprehension: A Cognitive Control Perspective .................................................................. Paul Whitney and Douglas A . Waring

199

9. Understanding Idiomatic Expressions: The Contribution of Word Meanings ........................................................... Cristina Cacciari and Sam Glucksberg

217

10. On the Combinatorial Semantics of Noun Pairs: Minor and Major Adjustments to Meaning ................................................. Edward J . Wisniewski and Dedre Gentner

24 1

Contents

X

1 1 . Making Sentences Make Sense, or Words to That Effect ....................................................................................................

285

Gregg C . Oden, Jay G . Rueckl, and Thomas Sanocki 12. How is Verb Information Used During Syntactic Parsing? ........................................................................................ 305 Fernanda Ferreira and John M . Henderson 13. The Role of Lexical Representations in Sentence Processing ..................................................................................... Julie E. Boland and Michael K . Tanenhaus

33 1

14. The Resolution of Indeterminacy During Language Comprehension:Perspectives on Modularity in Lexical, Structural and Pragmatic Processing ........................................................... David A. Swinney

367

Author Index ........................................................................................................ Subject Index .......................................................................................................

387 397

xi

Acknowledgments I would like to express my thanks to a number of people who, directly or indirectly, assisted in the completion of this volume. I am grateful, first, to Dr. George Stelmach, for encouragement to undertake the project, and to Dr. Kees Michielsen for his continued encouragement, advice, and patience. Mr.John Butterfield, Elsevier’s technical editor, was most helpful during every stage of the book’s preparation. Closer to home, I would like to express my deepest appreciation to Ms. Jane Johnson for her patience and her desktop-publishing expertise, which contributed immeasurably to the quality of the volume. Thanks go to several of my students, Robin Beyer, Paula Felchner, Rose Kroeker, Tim Riley, and Hyewon Suh, who kept the lab running while I was occupied with this project. Finally, I thank my wife and daughter, the former for her editorial advice, and both for their understanding and patience during my long preoccupation with this book.

Omaha, Nebraska July, 1990

This Page Intentionally Left Blank

xiii

Contributors Sam Glucksberg

Julie E. Boland Department of Psychology University of Rochester Rochester, New York 14627 U.S.A.

Department of Psychology Princeton University Princeton, New Jersey 08544 U.S.A.

Cristina Cacciari Dipartimento di Psicologia Viale Berti-Pichat, 5 40 127 Bologna Italia

John M. Henderson Department of Psychology University of Alberta Edmonton, Alberta T6G 2E9 Canada

Mark Faust Department of Psychology University of Oregon Eugene, Oregon 97403 U.S.A.

Fernanda Ferreira Department of Psychology University of Alberta Edmonton, Alberta T6G 2E9 Canada Dedre Gentner Department of Psychology University of Illinois at Urbana-Champaign Champaign, Illinois 61820 U.S.A. Morton Ann Gernsbacher Department of Psychology University of Oregon Eugene, Oregon 97403 U.S.A.

George Kellas Department of Psychology University of Kansas Lawrence, Kansas 66045 U.S.A. Marta Kutas Department of Neurosciences University of California, San Diego La Jolla, California 92093 U.S.A. Michael Martin Department of Psychology University of Kansas Lawrence, Kansas 66045 U.S.A. Gregg C. Oden Department of Psychology University of Iowa Iowa City, Iowa 52242 U.S.A.

xiv

Contributors

Padraig G. O’Seaghdha Beckman Institute University of Illinois at Urbana-Champaign Urbana, Illinois 61801 U.S.A.

David A. Swinney Linguistics Department Graduate Center City University of New York Ncw York, New York 10036 U.S.A.

Stephen T. Paul Department of Psychology University of Kansas Lawrence, Kansas 66045 U.S.A.

Patrizia Tabossi Dipartimento de Psicologia Wale Berti-Pichat, 5 40126 Bologna Italia

Keith Rayner Department of Psychology University of Massachusetts Amherst, Massachusetts 01003 U.S.A.

Michael K. Tanenhaus Departmcnt of Psychology University of Rochester Rochester, New York 14627 U.S.A.

Jay G. Rueckl Department of Psychology Harvard University Cambridge, Massachusetts 02138 U.S.A.

Douglas A. Waring Department of Psychology Washington State University Pullman, Washington 99164 U.S.A.

Thomas Sanocki Department of Psychology University of South Florida Tampa, Florida 33620 U.S.A.

Paul Whitney Department of Psychology Washington State University Pullman, Washington 99164 U.S.A.

Paula J. Schwanenflugel Department of Educational Psychology University of Georgia Athens, Georgia 30602 U.S.A.

Edward J. Wisniewski Department of Psychology University of Michigan Ann Arbor, Michigan 48104 U.S.A.

Greg B. Simpson Department of Psychology University of Nebraska at Omaha Omaha, Nebraska 68182 U.S.A.

Cyma Van Petten Department of Neurosciences University of California, San Diego La Jolla, California 92093 U.S.A.

Understanding Word and Sentence G.B. Simpson (Editor) 0 Elsevier Science Publishers B.V. (North-Holland), 1991

Chapter 1 Understanding Words in Context

Palrizia Tabossi Universith di Bologna Bologna, Italy

This paper deals with how sentential context affects the comprehension of words. In understanding a sentence, people use the information provided by lexical items to construct an internal representation of what is said in the sentence. The semantic information about words must be recovered from the mental lexicon and combined according to the syntax of the language before more complex elaborative processes can take place. Thus, the individual words constitute the building blocks of comprehension. But not only do words contribute to make up the meaning of the sentence in which they occur: In many occasions the reverse may also be true, and the internal representation constructed from a sentential context may help the various processes connected with the comprehension of a lexical item in the sentence. Consider, for instance, the following sentence: The cook put the sugar on the cake and lert it in the fridge. Here it may refer to either sugar or cake and in order to resolve the referential indeterminacy and interprct it correctly as referring to cake, one has to take into account the ovcrall meaning of the sentence and the general knowledge it elicits, in particular the fact that sugar is not kept in fridges, whereas cakes often are. The example illustrates the contribution that context can give to the comprehension of anaphoric expressions. The phenomenon, however, is not restricted to these words: It is to the more general case of sentential context effects on the comprehension of content nouns that the present article is devoted. But what is it to understand a word? A word can be considered fully comprehended when it has been adequately interpreted in its context of occurrence: It is understood not when it has been recognized as the neutral

2

P. Tabossi

pronoun, but when it has been interpreted as referring to cake, and before this result can be achieved several processes must take place. There is considerable disagreement among researchers as to how these processes should be characterized, and the terminological heterogeneity of the current literature does not contribute to the clarification of the matter. In any case, it will be sufficient for present purposes to consider those processes that are the object of interest of three of the major areas in lexical processing research: lexical interpretation, word recognition or identification, and lexical access. Lexical interpretation is illustrated in the above example where ii receives its interpretation according to context. Word recognition is used here to refer to the processes by which the visual or sound pattern corresponding to a word makes contact with the various kinds of information-semantic, morphological, syntactic, phonological, etc.-available to the readerflistener about that word. Finally, lexical access refers to retrieval of the semantic information related to a word, when the word is recognized. For instance, what information about the meaning of dog becomes available to a listener/reader when s h e recognizes the word? Does one recover all the available information about dogs or only that which is contextually relevant? How and under what conditions sentential contexts can affect these processes is still an open question, and one whose answer has implications for models of lexical processing. These implications will be discussed in the concluding section, after considering the available evidence on the issue, starting from the least controversial of the effects of context: the interpretation of words.

WORDINTERPRETATION Perhaps the most obvious case of sentential context effects on the interpretation of a lexical item is ambiguity. Although we hardly notice it, ambiguity is an extremely common phenomenon in language, and it is handled so efficiently by the language system that pcople find it easier to deal with a relatively small number of ambiguous words than with larger numbers of unambiguous lexical items. In fact, apart from function words, the more frequently a word is used, the more likely it is to be ambiguous (Miller, 1951). Context plays a central role in the comprehension of ambiguous words, and indeed it seems easier to understand an ambiguous item in context than to think of its meanings in isolation, as clearly illustrated in the following example by Phil Johnson-Laird (1983). Consider first the various meanings of plane. Has the word called to mind all the meanings involved in the following sentences? The plane landed on the runway. Imagine a sphere divided equally by a plane. The carpenter smoothed the surface of h e wood with a plane. All the trees have been cut down except the tall plane at the end.

Understanding Words in Context

3

Although ambiguity is the most striking example of sentential context effects, the phenomenon also applies to unambiguous words, which can be flexibly interpreted. In a cued recall study, Barclay, Bransford, Franks, McCarre11 and Nitsch (1974) presented their subjects with sentences such as: 1. The man lifted the piano. 2. The man tuned the piano. They found that “something heavy” was a better memory cue for Sentence 1 than for Sentence 2, whereas the reverse was true when “something with a nice sound” was the cue. Likewise, Halff, Ortony. and Anderson (1976) gave the subjects a list of paired sentences, each containing the word red (e.g., “The red fire engine raced down the street,” “The skin was red due to sunburn”), and asked them to judge, for each pair, whether the red in one sentence (e.g., the fire engine red) was definitely redder than, definitely less red than, or possibly equally as red as the red in the other sentence (e.g., sunburn red). The results indicated that the interpretation of red consistently varied according to its contexts. Following the seminal studies in the 70’s. subsequent work has further specified the nature of semantic flexibility showing, for instance, that not all the aspects of the meaning of a word are equally prone to context effects. Rather, some aspects - the ‘core’ meaning of the word - tend always to be present, whereas more peripheral aspects may become more or less salient depending on the contexts of Occurrence of the word (Barsalou, 1982; Greenspan, 1986). Related to semantic flexibility is the instantiation of general terms. Anderson and Ortony (1975) presented their subjects with sentences like the following:

3. The container held the apples. 4. The container held the cola.

They found that the subjects were better at remembering Sentence 3 when the cue was basket than when it was bottle, whereas the reverse was true for Sentence 4. Anderson, Pichert, Goetz, Schallen, Stevens, and Trollip (1976) extended these results, showing that basket, which did not occur in Sentence 3, was a better memory cue for that sentence than container, which did occur in it. Container, however, was better than basket for Sentence 5:

5 . The container stood near the apples showing that basket is not in general a better memory cue than container.

4

P. Tabossi

These data suggest that when people encounter general terms, like container, they tend to interpret them as more specific ones, according to context. The container was likely to be a basket in the situation described in Sentence 3 and a bottle in the situation described in Sentence 4: This is why the two words were better cues for Sentences 3 and 4, respectively. In contrast, Sentence 5 did not provide enough information to instantiate the general term that was therefore held unspecified. Disamabiguation, semantic flexibility and instantiation show unquestionably that sentential context can affect the interpretation of a word. In addition to being well established, these phenomena have also been given a convincing theoretical account within the framework of the mental model theory (JohnsonLaird, 1983). According to this theory, understanding a sentence involves building a mental model of the state of affairs described by the sentence. A mental model is not a description of a sentence: It is the mental construction of the situation described by it, and its structure is analogous to the corresponding events in the world. Within this theory, words are cues to the construction of the model, and their meanings are functions which contribute to determine the referents of the words in the model. A central assumption of the theory is that understanding requires linguistic capability to interact with knowledge of the world. Thus information provided by individual words in the sentence is understood in relation to sentential context and general knowledge, reducing the indeterminacy, vagueness and ambiguity of lexical items. One of the advantages of this theory is that it provides a homogeneous account for ambiguity, lexical flexibility and instantiation: In all cases, linguistc and nonlinguistic context sharpens their interpretation. But interpretation is a late process in the comprehension of a word, which in order to be interpreted must previously be identified. Indeed, as soon as the temporal dimension of lexical processing is taken into account, ambiguity, semantic flexibility and instantiaton, which look akin at the interpretative level, begin to show their differences. In the case of general terms, the semantic information relevant to their instantiation is not part of their meaning: The meaning of container does not include BASKET or BOlTLE a n y more than the meaning of it includes CAKE (Garnham, 1979; Johnson-Laird, 1983; for an alternative view see Halff et al., 1976). Hence, in order for instantiation to occur, the meaning of a general term must be made available by early lexical processes, and only subsequently may context and general knowledge allow inferences to make the general term more specific. Ambiguity and semantic flexibility are different. The information to be integrated in the context is part of the meaning of the word: HARBOUR is one of the meanings of porf, and MUSICAL or HEAVY are part of the semantic information about piano. It does therefore become relevant to establish when during the processes involved in the comprehension of these words context operates: at the time of their recognition and/or while accessing their meaning,

Understanding Words in Context

5

or only at the later interpretive stage.

WORDRECOGNITION A prior word (e.g.. NURSE) semantically associated to a subsequent one (e.g.. DOCTOR) facilitates the recognition of the latter (Meyer & Schvaneveldt, 1971). This phenomenon, where the prime provides a single-word context to the target, is well attested and various theories have been proposed to explain it (Becker, 1980; Collins & Loftus, 1975; Morton, 1970). Single-word context effects are lexical. That is, they originate entirely in the lexicon and can be accounted for on the basis of how information-semantic or otherwise-is organized within it. In order to establish whether sentential context operates on word recognition, it is important that single-word and sentential context effects are not confounded, making sure that effects that appear to be produced by a sentential context are not, in fact, lexical (Forster, 1981). Stanovich and West (1983). for instance, invoked lexical priming to explain the facilitation effects they observed in an early study, where their subjects were faster both at deciding that the target (e.g., SNOW) was a word and at pronouncing it after reading a predictive context (“The skier was buried in the”) than after reading a neutral one (e.g., ‘‘ They said it was the”) (West and Stanovich, 1982). There, it was skier that lexically primed SNOW. But even granting that this explanation can account for some of the available data (see Conclusions), several studies have reported sentential context effects in the absence of semantic association or relation. In particular, the recognition of a word appears to be facilitated by a prior predictive context (Balota, Pollatsek, & Rayrier, 1985; Fischler & Bloom, 1979, 1980; Morton & Long, 1976; Stanovich & West, 1983; West & Stanovich, 1982). Fischler and Bloom (1979), for example, had their subjects perform a lexical decision task on a word presented to them two seconds after the offset of a sentential context. The target word was a predictable completion of the sentence, a possible but unpredictable completion, or an anomalous completion as in the following example:

She cleaned the dirt from her SHOES HANDS TERMS. The results showed that SHOES was responded to faster than both HANDS and TERMS. But although in Fischler and Bloom (1979) sentences did not contain words semantically associated to the targets, their findings might not be conclusive. It is often argued, in fact, that the lexical decison task may not be adequate to investigate word recognition (Seidenberg, Waters,

6

P.Tabossi

Sanders, & Langer, 1984; Stanovich &West, 1983). Methodological problems are rather serious in the study of lexical processing, and are not restricted to one experimental paradigm only; hence it is very important that comparable results can be obtained with different techniques. Indeed, this is the case with predictive sentential contexts whose effects were observed, for example, by Stanovich and West (1983) using the same materials as Fischler and Bloom (1979), but with the naming rather than the lexical decision task. Probably the most serious problem with the above evidence is that it may well reflect genuine facilitative processes, but these processes are very unusual in natural language. Language, in fact, is rarely predictable, and people’s ability to guess correctly a word in different types of contexts is often not over 20% (Gough, Alford, & Holley-Wilcox. 1981). Hence, the above results may have very little generality. There are studies, however, showing facilitation effects produced by congruent rather than predictive contexts both with visual and auditory presentation of the materials (Schuberth & Eimas, 1977; Stanovich & West, 1983; Tyler & Wessels, 1983). Needless to say, the interpretation of these findings is not straightforward: In Schuberth and Eimas (1977) the subjects were fastest at recognizing a word in a congruous context, next fastest at recognizing the word in isolation, and slowest of all at recognizing it in an uncongruous context. Recognition time, however, was measured with a lexical decision task. Moreover, there are reasons to believe that recognition time of a word in isolation may not be an adequate baseline against which to measure the effects of a sentential context on the identification of that word (Stanovkh & West, 1983). Comparable difficulties are faced by Tyler and Wessels (1983), whose subjects were faster at recognizing a spoken word when it occurred in a semantically and syntactically congruous context than when it occurred in a syntactically correct, but semantically anomalous context. Here, the task used to assess word recognition was gating (Grosjean,l980). The subjects hear increasingly longer fragments of a target word until the whole word is presented. After each fragment they write down what word they think it is and how confident they are of their decision. These judgements provide a measure of the amount of perceptual information that is needed in order for the subject in to recognize a word and be sure of the identification. Performance differences for a word in different contcxts are interpreted as due to context effects on the recognition of the word. This task relies on the assumption that gating reflects the on-line processes involved in word recognition. However, the decision component that is so harmful in the lexical decision task is even more evident in gating, where the subjects are directly required to make a guess. Moreover, the adequacy of anomalous sentential frames as neutral contexts is problematic. More persuasive is the study by Stanovich and West (1983). who found that the prior presentation of a sentential frame such as “The puppy chewed on

Understanding Words in Context

7

the...” facilitates recognition not only of the predictable target BONE, but also of the unpredictable congruous word SHOE. In this study, facilitation was assessed by comparing recognition in the above context with recognition in three neutral contexts - “They said it was the,” “The next word will be,” and “They were thinking about the” - all of which produced similar results. In addition, recognition of the targets was established using both lexical decision and naming. Thus, none of the standard objections concerning the adequacy either of the task or of the baseline, or else the Occurrence of words in the sentence semantically related to the target seems to apply to this study. This is therefore one of the most convincing pieces of evidence in favour of genuine effects produced by sentential context, not only on the intepretation of a word, but also on the earlier processes involved in word recognition.

LEXICAL ACCESS Whether a prior sentential frame can affect word recognition has been and still is a matter of debate. By contrast, that context cannot affect access to the different meanings of an ambiguity has been probably the most widely accepted result in current psycholinguistics, and the strongest piece of evidence in favour of the autonomy of the lexical processing system. Although early research had suggested that lexical access might be context insensitive (Cairns & Kamerman, 1975; Dooling, 1972; Foss. 1970; Holmes, Arwas, & Garrett, 1977), strong and convincing evidence to this effect was provided by Swinney and his colleagues (Onifer & Swinney. 1981; Swinney, 1979). Onifer and Swinney (1981) had their subjects listen to a sentence that biased either the dominant or the subordinate meaning of an ambiguity occurring in it, as in the following example:

6. The housewife’s face literally lit up as the plumber extracted her lost wedding ring from the sink. 7. The office walls were so thin that they could hear the ring of their neighbour’s phone whenever a call came in. At the offset of the ambiguity, the subjects were presented with a visual word, on which they had to perform a lexical decison. The word was related to the dominant meaning of the ambiguity or to its subordinate meaning, or else was an unrelated control (e.g., FINGER-TALENT; BELL-WHIP). The results showed that both words related to the ambiguity were responded to faster than their controls, regardless of the biasing context, thus supporting the view that lexical access is an autonomous process which occurs exhaustively and is affected neither by extralexical contextual information, nor by such lexical factors as the relative frequency of the meanings of the ambiguity. The latter claim was particularly surprising, since it is well-known that

8

P. Tabossi

frequency is one of the most robust and pervasive phenomena in lexical processing. In fact, further work has modified that position and now most researchers would agree that frequency has an effect (Carpenter & Daneman, 1981; Duffy, Moms, & Rayner, 1988; Rayner & Frazier, in press; Simpson & Burgess. 1985; Tabossi, Colombo, & Job, 1987). But dominance is an intralexical phenomenon: It is produced by the way in which semantic information is organized in the lexicon. Hence, it can be easily accommodated within a context-insensitive model of lexical access. According to the ‘ordered search model,’ originally proposed by Hogaboam and Perfetti (1975). for instance, the meanings of an ambiguous item are accessed serially from the most frequent. The retrieved meaning is then matched for congruence with context. If this process succeeds, the search terminates; otherwise the next most frequent meaning is accessed and matched with context. Unlike the claim that lexical access is exhaustive, the insensitivity of this process to contextual information has gone almost unquestioned for several years. On the one hand, new results strengthened that view (Oden & Spira, 1983; Tanenhaus, Leiman, & Seidenberg, 1979); on the other hand methodological challenges to the soundness of Swinney’s results turned out not to be very cogent (Burgess, Tanenhaus. & Seidenberg, 1989; Glucksberg, Kreuz, & Rho, 1986). In addition, evidence contradicting those results was hardly conclusive. Simpson’s study (1981), for instance, showed effects both of dominance and context using a cross-modal paradigm. The interval between the offset of the ambigous item and the presentation of the visual target, however, was 120 msec long and this interval might have been sufficent to allow post-access interpretative processes to occur. The context-insensitive model has been so predominant that even potentially challenging results have been interpreted within its framework. Using a phoneme-monitoring paradigm, Cairns and Hsu (1980) found that an ambiguous word took longer to be processed than an unambiguous one in a neutral context. In a biasing context, however, this difference disappeared. Although the results indicated context effects and the methodology was supposed to reflect access rather than post-access phenomena, the authors interpreted their findings as post-access effects, supporting the autonomous view. No complete list of the materials was available in that study, but the examples suggest that the biasing contexts contained lexical items sematically related to one meaning of the ambiguous words. If so, then intralexical priming could perhaps be invoked to account for Cairns and Hsu’s results (Fischler, 1977; Seidenberg et al., 1984). Indeed, this explanation has been adopted in a series of studies by Seidenberg, Tanenhaus, Leiman, and Bienkowski (1982). They found, among other things, that subjects were faster at naming a target word related to one meaning of an ambiguity (e.g., HAY) after listening to ambiguous items in a context biasing that meaning (e.g., “Although the farmer bought the straw”) than after an unrelated control context that replaced the

Understanding Words in Context

9

ambiguous word with a word related to the other meaning (e.g., “Although the farmer bought the soda”). The intralexical explanation is plausible; however, as O’Seaghdha (1989) has pointed out, not onlyfarmer is related to straw. but the overall interpretation of the phrase prior to the ambiguous word favoured an agricultural meaning. Moreover, neither have semantic priming effects between related but unassociated words been conclusively established (Burani, Tabossi, Silveri, & Monteleone, in press; Lupker, 1984). nor can their putative existence be extented to words in sentential context (see Conclusions). In any event, the bulk of the findings support the context-insensitive hypothesis for ambiguous as well as unambiguous words. Whitney, McKay, Kellas, and Emerson (1985) had their subjects to name the colour of the ink with which a target word was written. This word named a high- or low-saliency property (e.g., BRANCHES, LUMBER) of a noun (e.g.. oak) which occurred in a sentence auditorily presented to the subjects prior to the target word. The sentence made salient either of the two properties of the noun (e.g., “The man trimmed the oak,” or “The man used the oak”). Whitney et al. (1985) found that when the target appeared exactly at the end of the unambiguous word, both targets interfered with colour naming, regardless of the preceding context. When the target occurred later in the sentence, limited context effects were observed: Whereas the high-saliency properties continued to interfere regardless of context, the low-saliency properties produced interference only when they occurred in conjunction with the congruent context. How conclusive is the above evidence on lexical access? Recently, findings have been presented that seem to be at odds with the data supporting context insensitive access to the lexicon. Tabossi (1988a) used feature priming contexts to investigate access to unambiguous words. In Experiment 1 the subjects listened to a sentence that contained a noun (e.g., butter) and either made salient an aspect of its meaning or did not make particularly salient any of its aspects, as in the following example:

8. To follow her diet, the woman eliminated the use of butter (butter is fat) 9. Before paying, the man checked the price of butter (no specific aspect of butter) 10. To soften it, the woman heated the butter (butter is meltable). At the offset of butter the subjects performed a lexical decision task to a word referring to the property made salient by one of the sentences (e.g., FAT). The results showed that lexical decision was fastest after Sentence 8, next fastest after Sentence 9, and slowest of all after Sentence 10. Experiment 2 indicated that those results were not attributable to words in the prior context directly priming the target: when Sentence 8 was substituted with another sentence containing a different unambigous noun (e.g., “To

P. Tabossi

10

follow her diet, the woman eliminated the use of wine”),the resulting sentence failed to facilitate the target word. Effects of context, as well as of dominance, were also observed when feature priming contexts were employed in order to bias the interpretation of ambiguous words (Tabossi et al., 1987). After listening to a sentence that made salient a central aspect of the dominant meaning of an ambiguity (e.g., “The violent hurricane did not damage the ships which were in the port, one of the best equipped along the coast”: a port is safe), lexical decisions were faster for targets related to the dominant meaning of the ambiguity (e.g., SAFE) than for words related to the subordinate meaning (e.g., RED) or control words (e.g., SHORT). In contrast, when the sentence made salient the subordinate meaning of the word (e.g., “Deceived by the identical colour, the host took a bottle of barolo, instead of one of port’“: port is red), targets related to either meaning of the ambiguity were responded to faster than the control target. A subsequent cross-modal study (Tabossi. 1988b) compared the effects of feature-priming contexts biasing the dominant meaning of an ambiguity with contexts that biased the same meaning, but without priming any aspect of it, as in the following example: 11. The violent hurricane did not damage the ships which were in the port, one of the best equipped along the coast.

12. The man had to be at five o’clock at the port, for a very important meeting. The results showed that feature priming contexts produced selective effects, thus replicating Tabossi et al. (1987). In the nonpriming contexts, however, targets related to both meanings of the ambiguity were responded to faster than the controls, replicating Swinney (1979; Onifer & Swinney, 1981). Hence, whether or not a sentence can guide selective access to the dominant meaning of an ambiguity may depend on the nature of the sentential bias. Consider, for instance, Sentence 12. In this sentence, there is no doubt of the interpretation of port ; yet, the context prior to the ambiguity does not provide much information about the upcoming word. The situation is similar to the example in Whitney et al. (1985). In both studies, after the word has been received its interpretation becomes clear, but there is nothing before it to give clues to the listener as to what information should be in the subsequent word. By contrast, the featurepriming contexts do not not render an ambiguous item predictable, but place constraints on the semantic information that will be included in it, and presumably, when the bias is toward the dominant meaning, this information can be used to direct access to the ambiguous word. The hypothesis that access may not always be insensitive to effects of context has recently been strengthened by new findings that, taken together, are rather challenging for the autonomous view of lexical access (Rayner & Frazier,

Understanding Words in Context

11

in press; Simpson & Krueger, in preparation; Van Petten & Kutas. 1987). But before discussing the implications of these findings on the issue of word processing, there is a preliminary point that needs to be considered: One must make sure that findings that appear to indicate effects of context on lexical access do not reflect instead effects on word recognition.

LEXICALACCESSOR WORDRECOGNITION? Suppose that prior sentential context can influence the various processes connected with the recognition of a lexical item, but has no influence on the recovery of the semantic information related to that item. If this is the case, experimental results apparently suggesting context effects on access may be observed without lexical access being affected at all. Context, in fact, might speed up the processes involved in the recognition of a word, leaving enough time for context-insensitiveaccess and subsequent interpretation to take place. Although early recognition is more likely to occur with long than with short words, and in general word processing studies use very short (mono- or bisyllabic) lexical items, the above hypothesis is a plausible explanation of the selective results observed in Tabossi’s studies. Because there are no monosyllabic content nouns in Italian, in the studies by Tabossi (1988b; Tabossi et al., 1987) the ambiguous items were two syllables long and lasted, on average, 471 msec. Assuming that a spoken word can be identified soon after the first syllable, roughly corresponding to about 200/250 msec from its onset (MarslenWilson & Tyler, 1980), the subjects in those studies might have recognized the ambiguous words and still have had more than 200 msec available before being tested for access at the offset of the words. Unfortunately, this time is long enough for post-access selection to take place (Seidenberg et al., 1982). Two lines of argument suggest that this explanation, though plausible, is not the most likely. First, the perceptual information received by a listener after 200/250 msec from a word onset may be sufficient to initiate lexical processing, but there are no grounds to believe that that amount of time is also sufficient to complete the recognition. This is even more true with words like the ambiguities used in Tabossi (1988b; Tabossi et al., 1987) whose initial fragments - with the exception of gemma - have alternative completions until very late: port- can be continued as portone, portalettere, porta; pol- can continue as polo, polio, polipo, etc. Second, if early recognition and subsequent selection is what actually happened in those studies, why was the dominant meaning still active in the context that biased the subordinate meaning? While speculations favour the lexical access interpretation of those results, Tabossi and Zardon (in preparation) have recently explored the issue empirically in a cross-modal lexical decision experiment. The materials sentences and visual targets - where the same as those used Tabossi et al. (1987) to bias the dominant meaning of the ambiguities. Here, however, three

P. Tabossi

12

different groups of subjects were tested: One group performed the lexical decision task 100 msec before the physical ending of the ambiguous prime, another group performed the task 60 msec before the end, and the third group saw the target at the offset of the ambiguity as in the prior work. The main result was an interaction: no reliable difference among any of the targets was observed at 100 and 60 msec. At 0 rnsec, however, previous findings were replicated: The target related to the dominant, contextually congruent meaning of the ambiguity was responded to faster than both the subordinate and the control targets, which did not differ from each other. Thus, either the ambiguous word has not been recognized, as indicated by its ineffectiveness as a prime for both the related targets at 100 and 60 msec from its ending, or else as soon as it is recognized and becomes effective (0 msec), it selectively facilitates a lexical decision on thc target related to its dominant, contextually pertinent meaning. Hence, the effects observed in Tabossi (1988b; Tabossi et al., 1987) are not due (at least not entirely) to the context speeding up the identification of the ambiguous items, and contextual information seems able to directly affect the access to the meaning of an ambiguous word. But how can these and similar results be reconciled with the earlier evidence on lexical access, and what is the picture that emerges from these studies and those on word recognition? The final section of this paper is devoted to consideration of these questions.

CONCLUSIONS The aim of this paper was to establish whether sentential context can influence the comprehension of a word. Of the three processes we have considered, interpretation is certainly the least problematic. The phenomenon itself is robust and widespread: It applies to anaphoric expressions, to ambiguous words, to general terms (nouns as well as verbs), and to unambiguous nouns. Moreover, theoretical accounts have been provided to explain it, and various factors - lexical and contextual - which are likely to constrain it have recently been explored. But interpretation is an integrative, post-lexical phenomenon that takes place after a word has been recognized and its semantic information recovered from the mental lexicon. Therefore, it does not clarify whether sentential context can influence earlier, lexical processes. More relevant to the issue are the results of studies on word recognition, suggesting the existence of effects that, to the best of our present knowledge, cannot be explained only on the basis of intralexical phenomena or methodological artifacts. Furthermore, those effects do not seem to be restricted to the identification of only the few words that in ordinary language are entirely predictable. Hence, it might be reasonably justified to conclude that word recognition is not a completely autonomous subprocess of the language comprehension system - a view that, interestingly enough, has been recently taken up

Understanding Words in Context

13

by West and Stanovich (1988). earlier proponents of the autonomy of lexical processing. But there are several difficulties, one of which is the reliability of the experimental paradigms. Researchers are well aware of the problem, and this is probably why they have devoted a considerable amount of their work trying to develope new methodologies, or to understand better and possibly improve old ones (Conrad, 1974; Cutler & Noms, 1979; Dyer, 1973; Forster, 1981; Foss & Gernsbacher, 1983; Francolini & Egeth, 1980; Grosjean, 1980; Kahneman & Henik, 1981; Van Petten & Kutas, 1988). In particular, as has already been mentioned, the characteristics of the two most popular tasks employed in the study of word recognition - lexical decision and naming - have been widely investigated, leading to the conclusion that lexical decision is open to strategic factors and is therefore not adequate to the study of early, automatic processes. Naming, on the other hand, which is faster and does not require decision making, is better suited to the study of such processes. Evidence to this effect relies mostly on comparisons between the two tasks,typically showing facilitation effects with both lexical decision and naming, but inhibitory, strategy-driven effects only with lexical decision (Balota & Chumbley, 1984; de Groot, 1985; den Heyer, Briand. & Dannenbring, 1983; Lorch, Balota. & Stamm, 1986; Seidenberg et al., 1984; Stanovich & West, 1983). Unfortunately, evidence once more is not conclusive. Exactly the opposite pattern was observed, for instance, in Forster (1981). where inappropriate sentential contexts gave rise to strong inhibitory effects when the subjects’ task was to name the targets, but not when it was to perform lexical decision on them. Inhibitory effects on the naming of contextually incongruous targets were also obtained by Simpson, Peterson, Casteel and Burgess (1989). and indeed various effects typically attributed to strategic tasks have been observed with naming (Balota & Chumbley, 1985). Probably the biggest problem is that one is never sure of whether response time is genuinly sensitive to word recognition or else is sensitive to criteria adopted by the subjects to perform the task (or both). To make an extreme example of the influence of criteria, imagine a task in which subjects are asked to perform a lexical decision on a number of ‘experimental’ words occurring either in a list made up of words and nonwords or else a list including words only. Clearly, the ‘experimental’ words will be responded to much faster in the latter than in the former case. In the word-only condition, in fact, all that is needed for the subjects to make a correct positive decision is to perceive that a stimulus is being presented. In the word-nonword condition, however, in order to perform the task correctly h e subjects must adopt a different criterion and decide to respond only after the word has actually been recognized. A solution to this difficulty would be a method such as the signal detection analysis which makes it possible to separate the relative contribution of

14

P. Tabossi

recognition from criteria used in the performance of an experimental task. Masson (1988) used this procedure in order to investigate the effects of sentential context on the perceptual analysis and on the integration processes involved in word recognition. His subjects were visually presented with a word, and their task was to search for that word in a sentence that followed immediately. The sentence, the words of which were presented visually one at a time, either contained the target word or did not contain it. There were two sets of materials: one set in which, whether present or absent, the target words were relevant to the meaning of the subsequent sentence, and one set where the targets where irrelevant. The results showed that relevant targets were recognized more often than irrelevant ones. However. when hit and false alarm rates were considered, it turned out that with relevant targets, the hit rate was balanced by a high percentage of false alarms. Thus, provided that ‘pure’ word recognition can be identified with the perceptual analysis of the stimulus word, and granted that the requirements of the paradigm-searching for a word-do not trigger a special strategy, Masson’s findings suggest that integration rather than recognition is affected by prior context. However, in an earlier signal detection study, Samuel (1981) found that a prior sentential frame, in addition to affecting the post-perceptual stages of word recognition, also increased the discriminability of a word (but see his interpretation). Taken together, these two studies illustrate the difficulty of obtaining conclusive evidence even with very promising analyses. Lexical access literature does not offer a more homogeneous picture. The same methodological problems discussed for word recognition also apply to lexical access, and attempts to account for the contrasting results on methodological grounds are unlikely to be very successful. In fact, even though several studies may have used experimental paradigms and/or materials which render them difficult to interpret, results still remain which, to the best of our present knowlege, cannot be reconciled on those bases. Thus, the available massive evidence is not capable of establishing whether sentential context can affect word recognition and/or lexical access, and fails to provide convincing support for or against the modular view of the language system (Fodor, 1983; Forster, 1979). According to this hypothesis, there are two different types of linguistic processes: those that take place within a module, like the lexicon, and those that operate in an integrative fashion outside modules. The former processes are unaffected by information not available in the module, are fast and automatic, do not require attention, do not produce interference, are out of any voluntary control, and are mandatory. In contrast, post-lexical integrative processes are context sensitive and slow, produce interference, require conscious attention, and are under voluntary or strategic control. Clearly, if establishing whether or not word recognition and lexical access are context sensitive - and hence whether or not the lexicon is a module - is

Understanding Words in Context

15

the aim of the investigation, its results are disarmingly inconclusive and contradictory. It is possible, however, that research does provide useful evidence, though on issues different from the modularity/non-modularityof the lexicon. Consider, for example, those studies that compare single-word and sentential context effects on word recognition. A consistent result in that line of work is that a word that facilitates recognition on a semantically associated target when presented in isolation, may fail to do so when occurring in a sentence (Auble & Franks, 1983). This is so even when the prime and target words are adjacent, unless the associated prime is also relevant to the topic of the sentence (Foss & Ross, 1983). Moreover, facilitation effects produced by a prior context (e.g., “The author of this”) containing a word associated to the target (e.g., BOOK) fail to be observed when the same associated prime occurs in a scrambled context (e.g.. “The author the from”) (Foss, 1982; O’Seaghdha, 1989; Simpson et al., 1989; Stanovich, Nathan, West, & Vala-Rossi, 1985). The relevance of these findings is twofold. First, they weaken considerably the strength of the intralexical interpretation of some of the results supporting the context sensitive hypothesis. Second, and more to the point, they suggest that although associative priming is one of the most typical instances of automatic intralexical effects, it may not be mandatory, and hence violates a crucial characteristic of modular processes (Hoffman & MacMillan, 1985). Similar violations are also suggested by those studies which challenge the automaticity of processes like the semantic analysis of unattended messages or the semantic priming produced by subliminally presented words on the ground of their sensitivity to attentional conditions (Johnston & Dark, 1982; Kahneman & Treisman, 1984). This situation is not unlike what one finds in current research on perception, where even very basic, typically automatic processes, such as visual feature extraction and integration or the perception of subjective contours, may nevertheless be influenced by focal attention (Prinzmetal, Presti, & Posner, 1986; Pritchard & Warm, 1983). Even the perception of ambiguous figures or the well-known Mueller-Lyer illusion may be modified through attention (Coren & Porac, 1983; Goryo, Robinson & Wilson, 1984; Reisberg, 1983; Tsal, 1984). Not only modular processes, but also post-lexical ones often fail to fit their characterization: Sentence reading, for instance, is typically considered an integrative process which illustrates the difference between automatic and controlled operations (Forster, 198 1; LaBerge & Samuels, 1974). Yet, Hirst, Spelke, Reaves, Cahavack, and Neisser (1980) found that after a long training, the performance of two subjects on copying dictated sentences and reading was comparable when the two tasks were executed together or separately. Even for undoubtedly complex processes their characterization as sequential, resource-demanding, strategic, and interfering may not be completely adequate.

16

P. Tabossi

What the above discussion amounts to is that a sharp distinction between automatic and controlled processes is probably untenable: Neither are automatic processes immune from strategic factors, nor are controlled processes bound to capacity limitations. Whether these two types of processes are best charcterized as lying along a continuum or else as qualitatively different is still under debate (Hirst el al., 1980; Posner & Snyder, 1975). but even those who prefer the twoprocess approach agree that “it must be realized that automatism and controlled processing are theoretical states and that performance in almost all tasks will be carried out with a contribution from both types of processes....[It is] difficult to set forth necessary and sufficient distinguishing characteristics” (Shiffrin, Dumais, & Schneider, 1981, p. 224). If indeed the distinction between modular and interactive processes is not a simple dichotomy, asking whether or not early lexical processing is modular might not be very fruitful, and a more productive approach would be trying to specify what types of information - whether intra- or extralexical - are likely to affect lexical processing and under what attentional, temporal, contextual conditions. Considered in this perspective, evidence of context effects on word recognition and lexical access may no longer be contradictory and inconsistent, but reflect genuine differences in the way in which these processes are accomplished in different situations. Existing results give already a few indications in this direction. Syntactic violations, like those in scrambled contexts, for instance, are so disruptive that they interfere even with such automatic effects as associative priming. Nevertheless, syntactic information by itself does not appear to be able to influence lexical processing (Tanenhaus & Lucas, 1987). Predictability, instead, is likely to be effective. Moreover, although recognition and access are very fast and strongly automatized processes which do not often rely on context information, still they appear to make use of sentential information that places strong constraints on the to-be-processed word. Not only studies on lexical access, but also work on word recogniton seems to justify this conclusion (Schwanenflugel & Shoben, 1985). In sum, whether or not lexical processing is modular and context insensitive cannot be established on the basis of the available evidence. Perhaps this state of affairs is analogous to the impasse on the modularity/penetrability of perceptive processes. There, according to Umilth (1988), research has failed so far to translate the ideological problem into an empirical one. Moreover, scholars on both camps face the almost insoluble problem of demonstrating two negative claims: 1) Perceptive processes are not affected by other cognitive processes, and 2) seemingly cognitive effects on perception do not take place within the perceptual system. A less pessimistic view, however, is that granted the methodological difficulties and provided that evidence is not forced into resolving an impossible dichotomy, the available research does give useful indications strongly speaking for a more flexible conceptualization of the mental

Understanding Words in Context

17

processes involved in the comprehension of words. Acknowledgments

This chapter was supported by grant. no. 8900064 Fondi 60%, and grant no. 8900135 Fondi 40%. I am grateful to Corrado Cavallero and Carlo Umild who both helped me durning various discussions to clarify some of the points made in the paper. Of course, the interpretation of these points is entirely my responsibility. References

Anderson, R. C. & Ortony, A. (1975) On putting apples into bottles. Cognitive Psychology, 7 , 167-180. Anderson, R. C., Pichert, J. W., Goetz, E. T., Schallert, D.L., Stevens, K. V., & Trollip, S. R. (1976) Instantiation of general terms. Journal of Verbal Learning and Verbal Behavior, 15,667-679. Auble, P. & Franks, J. J. (1983) Sentence comprehension processes. Journal of Verbal Learning and Verbal Behavior, 22,395-405. Balota, D. A. & Chumbley, J. I. (1984) Are lexical decisions a good measure of lexical access? A role of frequency effect in the neglected decision stage. Journal of Experimental Psychology: Human Perception and Performance, 10, 340-357. Balota, D. A. & Chumbley, J. I. (1985) The locus of word-frequency effects in the pronunciation tak: laxical access and/or production? Journal of Memory and Language, 24,89-106. Balota, D. A., Pollatsek, A., & Rayner K. (1985) The inetraction of contextual constraints and parafoveal visual information in reading. Cognirive PSyCholOgy, 17, 364-390. Barclay, J R., Bransford, J. D., Franks, J. J . , McCarrell, N. S . , & Nitsch, K. (1974) Comprehension and semantic flexibility. Journal of Verbal Learning and Verbal Behavior, 13,471-481. Barsalou, L. W. (1982) Context-independent and context-dependent information in concepts. Memory and Cognition, l , 82-93. Becker, C. A. (1980) Semantic context effecls in visual word recognition: An analysis of semantic strategies. Memory and Cognition, 8 , 493-512. Burani, C., Tabossi, P., Silveri, C., & Monteleone, D. (in press) Relazioni semantiche associative e non associative: effetti di priming. Giornale Italianoo di Psicologia Burgess, C., Tanenhaus, M. K., & Seidenberg, M. S. (1989) Context and lexical access: Implications of nonword interference for lexical ambiguity resolution. Journal of Experimental Psychology: Learning, Memory, and Cognilion, 15, 620-632.

18

P. Tabossi

Cairns, H. S. & Hsu. J. R. (1980) Effects of prior context upon lexical access during sentence comprehension: A replication and reinterpretation. Journal of Psycholinguistic Research, 9, 3 19-326. Cairns, H. S. & Kamerman, J. (1975) Lexical information processing, during sentence comprehension. Journal of Verbal Learning and Verbal Behavior, 14, 170-179. Carpenter, P. A. & Daneman, M. (198 1) Lexical retrieval and erroe recovery in reading: A model based on eye fixations. Journal of Verbal Learning and Verbal Behavior, 20, 137-160. Collins, A. M. & Loftus, E. F. (1975) A spreading activation theory of semantic processing. Psychological Review, 82,407-428. Conrad, C. (1974) Context effects in sentence comprehension: A study on the subjective lexicon. Memory and Cognition, 2 , 130-138. Coren, S. & Porac, C. (1983) The creation and reversal of the Mueller-Lyer illusion through attentional manipulation. Perception, 12,49-54. Cutler, A. & Norris, D. G. (1979) Monitoring sentence comprehension. W. E. Cooper & E. C. T. Walker (Eds.) Sentence processing: Psycholinguistic studies presented to Merrill Garrett. Hillsdale, N.J.: Erlbaum. de Groot, A. M. B. (1975) Word-context effects in word naming and lexical decision. Quarterly Journal of Experimenatl Psychology, 37A, 28 1-297. den Hyer, K., Briand, K, & Dannenbring, G. L. (1983) Strategic factors in a lexical decision task: Evidence for automatic and attention-driven processes. Memory and Cognition. 1 1 , 374-381. Dooling, D. J. (1972) Some context effects in the speeded Comprehension of sentence. Journal of Experimental Psychology, 93, 56-62. Duffy, S. A., Morris, R. K., & Rayner, K. (1988) Lexical ambiguity and fixation times in reading. Journal of Memory and Language, 27,429-446. Dyer, F. N. (1973) The Stroop phenomenon and its use in teh study of perceptual, cognitive, and response processes. Memory and Cognition, 1 , 106120. Fischler, I. (1977) Semantic facilitation without association in a lexical decision task. Memory and Cognition, 5 , 335-339. Fischler, I. & Bloom, P. A. (1979) Automatic and attentional processes in the effects of sentence contexts on word recognition. Journal of Verbal Learning and Verbal Behavior, 18, 1-20. Fischler, I. & Bloom, P. A. (1980) Rapid processing of the meaning of sentences. Memory and Cognition, 8,216-225. Fodor, J. A. (1983) The modulariv of mind. Cambridge, MA: MIT Press. Forster, K. I. (1979) Levels of processing and the structure of the language processor. W. E. Cooper & E. C. T. Walker (Eds.) Sentence processing: Psycholinguistic studies presented to Merrill Garrett. Hillsdale, N.J.: Erlbaum. Forster, K. I. (1981) Priming and the effects of sentence and lexical contexts on

Understanding Words in Context

19

naming time: Evidence for autonomous lexical processing. Quarterly Journal of Experimental Psychology, 33A. 465-495. Foss, D. J. (1970) Some effects of ambiguity upon sentence comprehension. Journal of Verbal Learning and Verbal Behavior, 8,457-462. Foss, D. J. & Gernsbacher, M. A. (1983) Cracking the dual code: Toward an unitary model of phoneme identification. Journal of Verbal Learning and Verbal Behavior, 22,609-632. Foss, D. J. & Ross, J. R. (1983) Great expectations: Context effects in sentence processing. G. B. Flores d’Arcais & R. J. Jervella (Eds.) The process of language understanding. New York: Wiley. Francolini, C. M. & Egeth, H. A. (1980) On the nonautomaticity of ‘automatic’ activation: Evidence of selective seeing. Perception & Psychophysics, 27, 331-342. Garnham, A. (1979) Istantaition of verbs. Quarterly Journal of Experimental Psychology, 31,207-214. Glucksberg, S . , Kreuz, R. J., & Rho, S. H. (1986) Context can constrain lexical access: Implications for models of language comprehension. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 323335. Goryo, K. , Robinson, J. O., & Wilson, J. A.(1984) Selective looking and the Mueller-Lyer illusion: The effect of changes in the focus of attention on the Mueller-Lyer illusion. Perception, 13.647-654. Gough. P., Alford, J. A,, & Holley-Wilcox, P. (1981) Words and contexts. 0. J. L. Tzeng & H. Singer (Eds.) Perception of print: Reading research in experimental psychology. Hillsdale, N.J.: Erlbaum. Greenspan, S. L. (1986) Semantic flexibility and referential specificity in concrete nouns. Journal of Memory and Language, 25,539-557. Grosjean, F. (1980) Spoken word recognition processes and the gating paradigm. Perception and Psychophysics, 28,267-283. Halff, H. M., Ortony, A., & Anderson, R. C. (1976) A context-sensitive representation of word meanings. Memory and Cognition, 4,378-383. Hirst, W., Spelke, E. S., Raves, C. C., Cahavack, G., & Neisser, U. (1980) Dividing attention without alternation or automaticity. Journal of Experimental Psychology: General, 109, 98-117. Hoffman, J.E. & MacMillan, F. W. (1985) Is semantic priming automatic? R. Parasuraman & D. R. Davies (Eds.) Varieties of attention. New York: Academic Press. Hogaboam, T. W. & Perfetti, C. A. (1975) Lexical ambiguity and sentence comprehension. Journal of Verbal Learning and Verbal Behavior, 14, 265-274. Holmes, V. M., Arwas, R., & Garrett, M. F. (1977) Prior context and the perception of lexically ambiguous sentences. Memory and Cognition, 5 , 103110.

20

P. Tabossi

Johnson-Laud, P. N. (1983) Mental Models. Cambridge, MA: Harvard University Press. Johnston, W. A. & Dark, V. J. (1982) In defence of intraperceptual theories of attention. Journal of Experimental Psychology: Human Perception and Performance, 8,407-421. Kahneman, D. & Henik, A. (1979) Perceptual organization and attention. M. Kubovy & J.R. Pomeranz (Eds.) Perceptual organization. Hillsdale, N.J.: Erlbaum. Kahneman, D. & Treisman. A. (1984) Changing views of attention and automaticity. R. Parasuraman & D. R. Davies (Eds.) Varieties of attention. New York: Academic Press. LaBerge, D. & Samuels. S. J. (1974) Toward a theory of automatic information processing in reading. Cognitive Psychology, 6,293-323. Lorch, R. F., Jr., Balota, D. A., & Stamm, E. G. (1986) Locus of inhibition effects in the priming of lexical decisions: pre- or postlexical access? Memory and Cognition, 14,95-103. Lupker, S. J. (1984) Semantic priming without association: A second look. Journal of Verbal Learning and Verbal Behavior, 23, 709-733. Marslen-Wilson, W. & Tyler, L.K. (1980) The temporal structure of spoken language understanding. Cognition, 8, 1-71. Masson, M. E. J. (1988) The interaction of sentence context and perceptual analysis in word identification. Memory and Cognition. 16,489-496. Meyer, D. E. & Schvaneveldt, R. W. (1971) Facilitation in recognizing pairs of words: Evidence of a dependence between retrieval operations. Journal of Experimental Psychology, 90,227-234. Miller, G. A. (1951) Language and Communication. New York: McGraw-Hill. Morton, J. (1970) A functional model for memory. D. A. Norman (Ed.) Models of human memory. New York: Academic Press. Morton, J. & Long, J. (1976) Effect of word transitional probability on phonemic identification. Journal of Verbal Learning and Verbal Behavior, 1.5, 43-51. Oden, G . C. & Spira, J. L. (1983) Influence of context on the activation and selection of ambiguous word senses. Quarterly Journal of Experimental Psychology, 35A, 51-64. Onifer, W. & Swinney, D. A. (1981) Accessing lexical ambiguity during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory and Cognition, 9,225-236. O’Seadhdha, P. G. (1989) The dependence of lexical relatedness effects on syntactic cennectedness. Journal of Experimental Psychology: Language, Memory, and Cognition, 15, 73-87. Posner, M. I. & Snyder, C. R. R. (1975) Attention and cognitive conlrol. R.L. Solso (Ed.) Information processing and cognition: The Layola symposium. Hillsdale. N.J.: Erlbaum.

Understanding Words in Conrext

21

Prinzmetal, W., Presti, D. E., & Posner, M. I. (1986) Does attention affect visual feature intergration? Journal of Experimental Psychology: Human Perception and Performance, 12,361-369. Pritchard, W. S. & Warm, J. S. (1983) Attentional processes and the subjective contour illusion. Journal of Experimental Psychology: General, 112, 145175. Rayner, K. & Frazier, L. (in press) Selection mechanisms in reading lexically ambiguous words. Journal of Experimental Psychology: Learning, Memory and Cognition. Reisberg, D. (1983) General mental resources and perceptual judgement. Journal of Experimental Psychology: Human Perception and Performance, 9, 966-979. Samuel, A. G. (1981) Phonemic restoration: Insights from a new methodology. Journal of Experimental Psychology: General, I1 0,474-494. Schuberth, R. E. & Eimas, P. D. (1977) Effects of context on the classification of words and nonwords. Journal of Experimental Psychology: Human Perception and Performance, 3.27-36. Schwanenflugel, P. & Shoben, E. (1985) The influence of sentence constraint on the scope of facilitation for upcoming words. Journal of Memory and Language, 24,232-252. Seidenberg, M. S., Tanenhaus, M. K., Leiman, J. M., & Bienkowski, M. (1982) Automatic access of the meanings of ambiguous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14, 489-537. Seidenberg, M. S., Waters, G. S., Sanders, M.. & Langer, P. (1984) Pre- and post-lexical loci of contextual effecs on word recognition. Memory and Cognition, 12, 315-328. Shiffrin, R. M., Dumais, S. T., & Schneider, W. (1981) Characteristics of automatism. J. Long & A. Baddeley (Eds.) Attention and Performance I X . Hillsdale, N.J.: Erlbaum. Simpson, G. B. (1981) Meaning dominance and semantic context in the pracessing of lexical ambiguity. Journal of Verbal Learning and Verbal Behavior, 20, 120-136. Simpson, G. B. & Burgess, C. (1985) Activation and selection processes in the recognition of ambiguous words. Journal of Experimental Psychology: Human Perception and Performance, 11.28-39. Simpson, G. B. & Krueger, M. A. (in preparation) Sentence context effects in reading ambiguous words. Simpson, G. B., Peterson, R. R., Casteel, M. A., & Burgess, C. (1989) Lexical and sentence context effects in word recognition. Journal of Experimental Psychology: Learning, Memory and Cognition. 15,88-97. Stanovich, K. E., Nathan, R. G.. West, R. F.. & Vala-Rossi, M. (1985) Children’s word recognition in context: Sreading activation, expectancy,

22

P. Tabossi

and modularity. Child Development, 56, 14 18-1428. Stanovich, K. E. & West, R. F. (1983) On priming by a sentence context. Journal of Experimental Psychology: General, 112, 1-36. Swinney, D. A. (1979) Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 15,681-689. Tabossi, P. (1988a) Effects of context on the immediate interpretation of unambiguous nouns. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 153-162. Tabossi, P. (1988b) Accessing lexical ambiguity in different types of sentential context. Journal of Memory and Language, 27.324-340. Tabossi, P., Colombo, L., & Job, R. (1987) Accessing Lexiacla ambiguity: Effects of context and dominance. Psychological Resaerch, 49, 161-167. Tabossi, P. & Zardon, F. (in preparation) Processing ambiguous words in context. Tanenhaus, M. K., Leiman, J. M., & Seidenberg, M. S. (1979) Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18,427-440. Tanenhaus, M. K. & Lucas, M. M. (1987) Context effects in lexical processing. Cognition, 25, 213-234. Tsal, Y. (1985) A Mueller-Lyer illusion induced by selective attention. Quarterly Journal of Experimental Psychology, 37A, 25-37. Tyler, L. K. & Wessels, J. (1983). Quantifying contextual contributions to wordrecognition processes. Perception and Psychophysics. 34,409-420. Umilta, C. (1988) Attenzione e peneuabilita dei processi cognitivi. G. Kanizsa & N. Caramelli (Eds.) L'ereditd della psicologia della Gestalt. Bologna: IL Mulino. Van Petten, C. & Kutas, M. (1987) Ambiguous words in context: An eventrelated potential analysis of the time course of meaning activation. Journal of Memory and Language, 26, 188-208. Van Petten, C. & Kutas, M. (1988) Tracking the time course of meaning activation. S. L. Small, G.W. Cotuell, & M. K. Tanenhaus (Eds.) Lexical ambiguity resolution. San Mateo: Morgan Kaufman. West, R. F. & Stanovich, K. E. (1982) Source of inhibition in experiments on the effect of sentence context on word recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 8, 385-399. West, R. F. & Stanovich, K. E. (1988) How much of sentence priming is word priming. Bulletin of the Psychonomic Society, 26, 1-4. Whitney, P., McKay, T., Kellas, G.,& Emerson, W.A., Jr. (1985) Semantic activation of noun concepts in context. Journal of Experimental Psychology: Learning, Memory, and Cognition, 11, 126-135.

Understanding Word and Scntence

G.B.Simpson (Editor) 0 Elsevicr Science Publishers R.V. (North-Holland). 1991

Chapter 2 Contextual Constraint and Lexical Processing Paula J . Schwanenflugel University of Georgia Athens, Georgia U.S.A.

Developing an understanding of the influence of contextual constraint on lexical processing is important for several reasons. One is that contexts vary considerably in terms of the degree to which they lend themselves to the anticipation of particular words. For example, a simple word context such as blackleads one to anticipate the upcoming occurrence of the word white, although other words such as brown or ink are also possible. To a similar degree, the sentence John kept his gym clothes in the leads us to strongly expect that locker will follow, although other completions such as closet or hamper are also reasonable. However, a sentence such as In the valley, there were three small does not lead us to develop strong expectations for any particular upcoming word. This tendency for contexts to be differentially predictable has been shown through numerous word association (cf. Jenkins, 1970; Keppel & Strand, 1970) and sentence constraint (also called cloze frequency) norms (Bloom & Fischler, 1980; Schwanenflugel, 1986). as well as the large literature on cloze production (see McKenna & Robinson, 1980, for a review). It seems likely that the strength of a person's expectations should have a profound influence on their processing of words in reading, listening, and production. Contextual constraint has been isolated as a consistently important factor in several applied domains. For example, contextual constraint (or cloze frequency) is sometimes included as an important (although atheoretical) component of readability formulas (eg., Bormuth, 1968; Taylor, 1953). Activities involving contextual constraint have also proven useful as part of reading assessment (Shanahan, Kamil, & Tobin, 1982) and instructional devices (Jongsma, 1980). Most importantly, understanding contextual constraint will enable us to understand in more detail how contextual processes operate in language processing. By doing so, we should gain a better understanding of the processes involved as persons interrelate and combine information from their internal knowledge base with information from the external stimulus world as they process upcoming words.

24

P.J. Schwanenfugel

In this chapter, I will discuss the varied literatures describing influences from word-level constraint and sentence constraint. Throughout, contextual constraint will be defined simply as the degree to which contexts elicit an individual word when persons are asked to generate completions or continuations. It should be recognized that there may be other kinds of contextual constraints operating in lexical processing beyond the simple predictability of particular words such as verb selection restrictions and thematic roles (Tanenhaus & Carlson, in press), or general syntactic information (Seidenberg, Waters, Sanders, & Langer, 1984; West & Stanovich, 1986; Wright & Garrett, 1984). However, the scope of this chapter will be limited to describing how changes in the overt predictability of words from contexts (as defined by lexical generation tasks) influence the processing of upcoming words. I will also concentrate on research describing on-line processing of words, rather than memory for or explicit decisions about those words. I will then discuss how various theories of lexical processing might (or not) account for these influences.

WORD-LEVEL CONTEXTUAL CONSTRAINTS Existing research on word-level constraints on lexical processing generally involves one of two types of relationships: simple word-association relatedness and categorical relatedness. Constraints based on word association are said to reflect the co-occurrence frequency in the real world (Anderson, 1976). Thus, the word pair docror-nurse would be highly associated because doctors and nurses tend to co-occur in similar contexts. In word-association tasks, subjects are usually given a list of words and asked to list either the first word or several words that come to mind for each word within a certain time frame (Jenkins, 1970). In studies of word-association relatedness on lexical processing, word pairs are usually selected from one of the many existing word association norms. Category term constraints are said to come from one of several sources. One view is that such constraints are also simple reflections of co-occurrence frequency (Glass & Holyoak, 1975) such that a word pair such as bird-robin is more associated than bird-chicken because the word robin co-occurs more often than chicken with the category term bird. Another view is that category constraints are derived from either the attribute overlap of the representation of the category term with their exemplars (McCloskey & Glucksberg, 1979; Smith, Shoben, & Rips, 1974) or the similarity of the category prototype with individual exemplars (Mervis & Rosch, 1981). In most of the studies relevant to the influence of category relatedness on lexical processing, category-exemplar pairs have been selected from one of several category production norms, most typically the Battig and Montague (1969) category norms. In studies of word-level constraints on lexical processing, a typical trial would entail presenting subjects with a single stimulus word eilher visually or

Contextual Constraint and Lexical Processing

25

auditorally for varying amounts of time (termed stimulus onset asynchrony or SOA), followed by a response item. Usually this response item varies considerably in the degree to which the word is either categorically or associatively related to the stimulus word. The subject’s task is either to name the response word aloud or make a lexical decision for it. The costs or benefits associated with the processing of these one word contexts are assessed by comparing context trials to some kind of control trial in which the response item is preceded by a control stimulus such as the word blank, ready, or X X X , or an unrelated word. Intuition would tell us that highly related words should benefit more than less related words from prior presentation of a context word compared to a control. In most cases, this intuition has been confirmed (Becker. 1980, Experiment 3; Balota & Duchek, 1988; deGroot, Thomassen, & Hudson, 1982; Fischler & Goodman, 1978, at 40 ms SOA; Howard, 1983, Experiment 2; Lorch, 1982; Lorch, Balota, & Stamm, 1986; Neely, in press; Massaro, Jones, Lipscomb, & Scholz, 1978, for perceptually degraded trials; Schwanenflugel & Rey, 1986, Experiment 1) but in a startling number of cases it has not (Becker, 1980, Experiments 2 and 4; den Heyer, Briand, & Smith, 1985; Fischler, 1977; Howard, 1983, Experiment 1; Neely, 1977; Massaro, et al., 1978, for intact trials; Schwanenflugel & Rey, 1986, Experiment 2; Warren, 1977). Given the disparity of results on the degree to which word level constraints influence the processing of upcoming words, it seems reasonable to probe deeper to see whether some regularities can be discerned. Associative and Category Constraints One factor that might potentially distinguish cases for which word contexts differentially influence the processing of upcoming words from those that do not is whcther the prime is associatively or merely categorically related to the target word. That is, beyond differentially predicting exemplar words, category terms may also be associated to lots of other words not included in category norms. For example, the category termfruit might actually be more highly associated with the word vegetable than it is with any of its exemplars, thereby producing smaller benefits in processing for any exemplar terms. An examination of the literature on word-level constraint and lexical processing suggests that this factor will not distinguish studies showing differential contextual benefits from those that do not. In studies using category word primes, some find differential benefits as a function of categorical relatedness (Becker, 1980, Experiment 5; Lorch et al., 1986; Lorch, 1982, Experiment 2; Howard, 1983, Experiment 2; Massaro et al., 1978, for degraded words; Neely, in press; Schwanenflugel & Rey, 1986, Experiment 1). whereas others do not (Becker, 1980, Experiments 2 and 4; den Heyer et al., 1985; Howard, 1983, Experiment 1; Massaro et al.. 1978, for intact words; Neely, 1977; Schwanenflugel & Rey,

26

P.J. Schwanenjlugel

1986, Experiment 2). Similarly, in experiments employing associative word primes, some find differential contextual benefits for highly associated compared to less associated words (deGroot, et al., 1982; Fischler & Goodman, 1978, at 40 ms SOA), whereas others do not (Fischler, 1977; Fischler & Goodman, 1978, at 550 ms SOA; Warren, 1977). Moreover, studies explicitly comparing the categorically and associatively related primes have not reported a difference in the way that the prime types have acted to produce contextual benefits (Lorch, 1982, Experiment 3; Balota & Duchek, 1988). In fact, both of these studies reported that highly related words benefitted more from appearing in context than less related words. Therefore, it can be concluded that the category primes and associative primes act similarly to produce contextual benefits.

Task Differences Task differences might also act as potential determinants of differential context effects for high and low related words. Virtually all the studies discussed thus far have employed either lexical decision or naming to investigate contextual effects on lexical processing. It is now widely accepted tliat lexical decision involves postlexical decision processes occurring after the word has been recognized (Balota & Chumbley, 1984; Balota & Lorch. 1986; Lorch et al.. 1986; Lupker, 1984; Seidenberg et al., 1984; Stanovich & West, 1983; West & Stanovich, 1982, 1986). It is conceivable that this task difference might somehow alter the contextual benefits that related words receive as a function of task. Using naming as the processing task, Larch (1982) found greater contextual benefits for highly related words than less related words in two experiments. Keefe and Neely (Neely, in press) showed differential benefit for highly related words in naming that grew larger as the proportion of related trials in the experimental context increased. Warren (1977). however, did not find such greater benefits for highly related words, but this null effect (as well as that of Fischler and Goodman, 1978, for lexical decision) might have been partly due to his contrasting only high and moderately related words. For lexical decision, where there has been more research, the evidence has been quite mixed. Several studies found a greater priming effect for highly related words (Becker, 1980, Experiment 5 ; deGroot et al., 1982; Howard, 1983, Experiment 2; Neely, Keefe, & Ross, in press; Schwanenflugel & Rey, 1986, Experiment l), whereas others did not (Becker, 1980, Experiments 2 and 4; den Heyer et al., 1985; Fischler, 1977; Howard, 1983, Experiment 1; Neely, 1977). Moreover, three studies by Lorch, et al. (1986) Massaro et al. (1978), and Neely (in press) explicitly comparing the two tasks detected no particular differences as a function of task. Despite this apparent confusion in the literature, there seems to be an

Contextual Constraint and Lexical Processing

27

overall tendency for there to be less of a relatedness effect in lexical decision than in naming. However, in at least some of these lexical decision studies, this lack of difference seems at least partially attributable to an overall minimal beneficial effect of context being demonstrated for even the most highly related items (Becker, 1980, Experiments 2 and 4; den Heyer et al., 1985; Schwanenflugel & Rey, 1986). Such a pattern would leave little room for relatedness effects to be displayed. Given the way that the differences between lexical decision and naming are typically explained, it would seem that, if anything, highly related words should enjoy particularly larger benefits in lexical decision. That is, lexical decisions are usually described as having a semantic integration or plausibility decision associated with them (Balota & Lorch, 1986; deGroot et al., 1982). A priori, it would seem that such a semantic integration decision would be easier for highly related words, yielding an extra degree of contextual benefit in lexical decision for such words (deGroot, 1985). There is no hint in the literature that such a pattern is the case. Lexical decision benefits for highly related words are not discernably different than they are in naming tasks. It is unclear just what kind of post-lexical checking is going on here. On the other hand, despite the fairly large number of studies examining relatedness effects on lexical processing, there have been very few direct comparisons between lexical decision and naming. Before any firm conclusions can be drawn regarding relatedness effects in lexical decision and naming, more research explicitly comparing this factor needs to be carried out. However, given the current research, task seems to make only a minimal difference at best in determining the degree of contextual benefit derived from high and low related words.

Benefits for Low Related Items Another issue to be considered is whether low related items typically receive any benefits from appearing in a word context at all. Whether lexical processing in a particular situation can be shown to display a contextual benefit or not depends very much on the choice of baseline from which those benefits are measured (Antos, 1979; deGroot et al., 1982; Jonides & Mack, 1984; Stanovich & West, 1983). Typically, the baseline of choice is a neutral word or string of Xs rather than an unrelated word that might be inhibited by the context. Thus, in thinking about contextual benefits, only the studies with at least somewhat appropriate baselines will be considered. Of the experiments reporting facilitation scores for low related word pairs, the majority of these display either extremely small or no benefits of context at all (Balota & Duchek, 1982, c 10 ms; Becker, 1980, 13 ms; Lorch et al., 1986, 6 ms; deGroot et al., 1982, 10 ms; Neely, Keefe et al., in press, 20 ms; Schwanenflugel & Rey, 1986,21 ms; except for Lorch. 1982, 32 ms) with the vast

P.J. Schwanenflugel

28

majority of experiments reporting null effects of context for low related words. This pattern of small to null effects of context for low related items can be discerned at virtually all SOAs. When averaged over those studies reporting facilitation scores for low related items, the benefits such items derive from context in studies employing SOAs of less than 300 ms is approximately < 10 ms, whereas a benefit of < 15 ms is displayed for all SOAs greater than 300 ms. Therefore, it seems that the benefits that words derive from appearing in a word context to which they are only minimally related range from small to nonexistent.

Summary There are several conclusions that we can draw from this analysis of word-level effects on lexical processing. First, there is great disparity in the literature regarding whether highly related items receive greater benefits in lexical processing than low related words. However, when there are significant differential benefits, it is always in the direction of highly related items accruing greater benefits than low related. If context effects truly did not differ between high and low related words, one would expect that in at least some of these cases benefits would be significantly larger for low related items. In fact, there is no report of such a finding. Therefore, this regularity should probably not be ignored (see Simpson, 1984, for a similar analysis in a somewhat different domain). It also seems to make little difference whether items are associatively related or categorically related to the target word. For both kinds of relationships, when a difference between high and low related words is shown, high related words accrue greater benefits. Task also seems to make surprisingly little difference in producing differential context effects as a function of relatedness. There may be a tendency for lexical decision studies to report slightly fewer differential benefits of context, but this appears to be at least partly attributable to several studies finding small overall benefits of context in general. Finally, very low related words with only a minimal relation between the context and the target word appear to benefit little from appearing in a single word context. This seems true regardless of task timing.

SENTENCE CONSTRAINT Sentence constraint is usually defined as the probability of eliciting a particular final word given the sentence context. Typically, sentence constraint is derived by presenting persons with a list of sentence frames with their final words deleted and asking them to generate either a single completion (called the cloze procedure, Bloom & Fischler, 1980) or several potential completions (called the multiple production procedure, Schwanenflugel, 1986) for each sen-

Contextual Constraint and Lexical Processing

29

tence. The frequency or probability of the most generated word is used to define each sentence’s constraint. Beyond simply having a highly predictable completion, high constraint sentences also tend to have relatively fewer potential completions (Bloom & Fischler, 1980; Schwanenflugel, 1986). Therefore, high constraint sentences tend to constrain both the predictability of individual words and the range of words that can complete individual sentences. Assessed sentence constraint appears to be a fairly reliable construct. The relative predictability of particular words doesn’t seem especially influenced by the characteristics of those who perform the task (see Cohen & Faulkner, 1983, for a comparison of young and old adult subjects; Fischler, 1983. for comparison of deaf and hearing subjects; Nebes, Boller, & Holland, 1986, for a comparison of young adults, and old adults with and without Alzheimer’s). Relative sentence constraint also appears to be comparable regardless of the method used to derive it. The multiple production method tends to shift the absolute estimates of sentence constraint upward compared to the cloze method (Schwanenflugel, 1986), but doesn’t appear to alter the relative standing of various sentences in terms of their constraint. The multiple production measure is likely to be better able to distinguish truly unexpected words. Therefore, the relative estimates of sentence constraint appear to be fairly robust across a variety of subject populations and methods used to obtain them. One point to consider is the degree to which the words persons list in such generation tasks are an outcome of word-level constraints such as those described in the first section. That is, can the probability of expected completions from sentence contexts be derived from simple relations between single words in the sentence and the generated word? An unpublished pilot study discussed in Masson (1986) suggested that this could be so. He reported that individual content words in high constraint sentences tended to be more associated with their most expected completion than content words in low constraint sentences. On the other hand, findings from a senior research project by Alice Tam in my laboratory suggested that the individual content words from high constraint sentences used by Schwanenflugel and LaCount (1988) were no more semantically related to their most expected completions than those from low constraint sentences. Clearly, more research needs to address the question of whether sentence constraint is an emergent property of sentences or whether it is attributable to a collection of word-level constraints. This is an issue that I will take up further later. Sentence constraint appears to have a large influence on the processing of upcoming words. Throughout this section, high constraint will be used to describe those sentences having an average predictability of greater than .70 for the multiple production method (or .60 using the cloze method, to account for its generally lower assessment of predictability), whereas low constraint will be used to describe those having a predictability below that. Overall, studies show that words appearing in high constraint sentences are easier to process than

30

P.J. Schwanenflugel

those appearing in low constraint sentences (Cohen & Faulkner, 1983; Ehrlich & Rayner, 1981; Fischler, 1985; Fischler & Bloom, 1979; Kutas & Hillyard, 1984; Nebes et al., 1986; McClelland & O’Regan. 1981; Schuberth, Spoehr, & Lane, 1981; Schwanenflugel & Shoben, 1985; Schwanenflugel & LaCount, 1988; but see Masson, 1986). However, in thinking about sentence constraint effects on lexical processing, it is important to distinguish words that are the expected responses from the context (sometimes called primary, dominant, or likely words) from those that are unexpected (sometimes called unlikely or low probability words). As we shall see, the influence of a sentence context is likely to be very different depending on whether the word that follows is an expected or an unexpected word. Therefore, I will discuss the form and scope of this influence in more detail below.

Sentence Constraint and the Processing of Expected Completions When studies directly comparing high and low constraint sentence contexts are considered, it is clear that sentence constraint effects on expected completions occur across a wide variety of processing tasks. In all of these tasks, we find that processing of expected completions following high constraint sentences is faster than the processing of words in low constraint sentences. Production of potential completions appears to be faster for high constraint sentence frames than low constraint contexts. In such tasks, subjects are presented with the sentence frame and asked to generate a sentence completion. This differential benefit in word production for high constraint sentence contexts has been displayed by young and old adults (Cohen & Faulkner. 1983). as well as old adults with Alzheimer’s disease (Nebes et al., 1986). although the constraint effect is much larger for this latter group. Therefore, it appears that high constraint sentences better enable persons to retrieve appropriate completions from long term memory than low constraint sentences. In studies using lexical decisions, processing of expected words have been significantly faster following high constraint than low constraint sentences in virtually all cases (Cohen & Faulkner, 1983; Fischler, 1985; Fischler & Bloom, 1979; Schuberth et al., 1981; Schwanenflugel & Shoben, 1985; Schwanenflugel & LaCount, 1988). Compared to normal young adults, sentence constraint effects are larger for deaf subjects (Fischler, 1985) and old subjects (Cohen & Faulkner, 1983). suggesting a greater reliance on context for these groups. Whether the larger sentence constraint effects for these latter groups come about because of overall slowed processing of verbal information is unclear, however, because an unpublished study (Schwanenflugel, 1985) in my laboratory comparing good and poor college readers suggests that sentence constraint effeck for those two groups are actually quite similar despite the longer reading times for the poor readers. Nevertheless, these studies on individual differences in context use concur with the view that words following

Contexrual Constraint and Lexical Processing

31

high constraint sentences are faster to make lexical decisions for than those following low constraint sentences. There has been much less research comparing naming times from high and low constraint sentences. McClelland and O’Regan (1981) found naming times for words appearing in moderately constraining contexts to be approximately 55 ms faster than those appearing in low constraint contexts. On the other hand, Masson (1986) found naming times to be only a marginally significant 12 ms faster in high than low constraint contexts. Clearly, further research will be necessary to determine whether sentence constraint effects are typically found in naming. Sentence constraint effects have also been demonstrated in tasks that monitor normal silent reading by assessing event-related brain potentials and eye movements. Event-related brain potentials for expected completions following high constraint sentences tend to produce late broad positivity whereas very low constraint sentences produce the opposite, suggesting an overall mismatch of the word with subjects’ expectations for the latter (Kutas & Hillyard, 1984; Kutas, Lindamood, and Hillyard, 1984). Eye fixations on expected completions tend to be shorter for high constraint than low constraint sentences (Ehrlich & Rayner, 1981). Therefore, sentence constraint effects are evident even when task demands are minimal. In sum, expected words appearing in high constraint sentences tend to be processed faster than those appearing in low constraint contexts. This is true for tasks with highly contrasting processing requirements. On the other hand, it is important to distinguish contextual influences that benefit processing from those that do not. As noted earlier, a neutral baseline control needs to be included to make this distinction. When contextual costs versus benefits are examined across studies, it is clear that the benefits in processing expected completions appearing in high constraint sentence contexts are greater and more consistent than those following low constraint sentences. Studies employing high constraint sentences almost universally find expected completions to benefit from appearing in a sentence context compared to a neutral control. Lexical decisions are faster for expected words when they appear in a high constraint context compared to a neutral context (Cohen & Faulkner, 1983; Fischler & Bloom, 1979, 1985; Forster, 1981; Schwanenflugel & Shoben, 1985; Schwanenflugel & LaCount, 1988; Stanovich & West, 1983). Naming times for expected words are also generally faster, although significantly so only in some studies (Schwantes, 1985; Stanovich & West, 1983) and not others (Forster, 1981; Masson, 1986). However, these latter studies with nonsignificant findings both used an RSVP technique to present the sentences, which may have been marginally disruptive for reading for meaning. Furthermore, the null finding by Masson was largely dependent on which of the several included baselines one used to calculate context effects. For expected completions following low constraint sentences, the situ-

32

P.J. Schwanenflugel

ation seems to be much more mixed regardless of the task used. Some lexical decision studies have found significant benefits from low constraint contexts while others have not. For example, Cohen & Faulkner (1983) noted significant facilitation from such contexts for old adults, but not young adults. Sanocki & Oden (1984) found facilitation in high cue validity experimental contexts, but not low validity contexts. In most of my own research, I (Schwanenflugel & LaCount, 1988; Schwanenflugel & Shoben, 1985) have usually found expected words to benefit from appearing in low constraint sentences, but not for concrete words where nonwords derived from meaningful sentence completions were also used (Schwanenflugel & Shoben. 1983; Schwanenflugel. Harnishfeger, & Stowe, 1988). Fischler & Bloom (1979; Fischler, 1985) detected no benefits for words in low constraint contexts, but this result seems partly due to blocking context trials and neutral trials enabling context specific strategies to develop. For naming, the picture is quite similar. Stanovich & West (1981, 1983) found expected words appearing in low constraint contexts to benefit significantly in most, but not all experiments. Facilitation was demonstrated by young and old adults, and adults with Alzheimer’s disease in a study by Nebes et al. (1986). McClelland and O’Regan (1981) noted benefits for words appearing in moderately low. but not very low, constraint contexts. Again, Forster (1981) and Masson (1986) did not find facilitation for low constraint contexts using an RSVP technique for presenting the contexts. In sum, the overall pattern of results from these studies regarding the processing of expected completions is that words appearing in high constraint sentences derive greater facilitation than words in low constraint contexts. Whether words in low constraint context can be said to derive benefits from context seems to depend partly on choice of neutral baseline as well as the availability of optimal conditions for persons to derive maximum understanding of the context. Again, even when significant facilitation for expected words is not shown, it is almost invariably in the direction showing benefits and not costs. Therefore, it seems, on average, that the influence of a low constraint context on the processing of upcoming words is small, but positive. The Scope of Facilitation from Sentence Contexts Not all words that persons will process during reading normal text are ones that they expect to see. For example, when people receive a sentence ,” they are very context such as “The combustible chemicals caused the likely to anticipate that the final word of the sentence will be explosion, although it might just as likely be another reasonable completion such as blast or fire. In fact, by one estimate, only about 25 percent of the words people receive in normal text are ones they could have guessed on the first attempt given the context (Gough, Alford, & Holley-Wilcox, 1981). The issue to be addressed

Contextual Constraint and Lexical Processing

33

here is whether sentence contexts are helpful in the processing of words that persons do not expect, but that are highly acceptable sentence completions. In other words, how general or broad is the scope of facilitation from high and low constraint sentences for upcoming words in sentences? Across studies, there appears to be considerable empirical disagreement regarding whether unexpected words such as blast above benefit from appearing in a meaningful sentence context. Several investigators find that sentence context effects are general enough to facilitate the processing of unexpected words (Schwantes, 1985; Stanovich & West, 1981, 1983; Tyler & Wessels, 1983) whereas others fail to find such facilitation (Balota, Pollatsek, & Rayner, 1985; Fischler & Bloom, 1979, 1985; Forster, 1981; Kleiman, 1980; Kutas & Hillyard, 1984; Norris, 1987). A closer look at these studies, however, suggests that reported sentence constraint may go a long way toward delineating when sentence contexts will facilitate the processing of unexpected words and when they will not. In a large proportion of cases where no benefits of context for unexpected words have been shown, relatively high-constraint sentences have been used. For example, Kleiman (1980), using sentences with an average cloze frequency of 78%. found no facilitation for unexpected words in lexical decision. Fischler and Bloom (1985), using sentences with cloze frequencies of 85%, also found no contextual benefits for lexical decisions for unexpected words following high constraint sentences. Schwanenflugel and Shoben (1 9 8 9 , and Schwanenflugel and LaCount (1988) using high constraint sentences with an average multiple production frequency of 87% and 88%, respectively, also report no facilitation for unexpected words in lexical decision. Forster (1981) detected no facilitation for unexpected words in high constraint sentences with a cloze frequency of 76% for either naming or lexical decision. Using moderately high constraint sentences with a cloze frequency of 61%. Stanovich & West (1983, Experiments 10 and 11) found no facilitation for naming or lexical decision for unexpected words following sentence frames ending with an article in either naming or lexical decision. However, when the unexpected word was immediately preceded by a non-article, facilitation was found for naming, but not lexical decision, suggesting that local priming effects might at least briefly ovemde those based on sentence contexts. The only true exception to this general finding is one by Schwantes (1985), whose adult subjects demonstrated a small, but significant, 28 ms benefit from high constraint sentences with a cloze frequency of 96%. Overall, then. it seems that high constraint sentences produce a very narrow scope of facilitation for upcoming words which generally does not include congruous, but unexpected words. Moreover, it does not seem to matter whether those unexpected words are semantically related to the expectations engendered from the context (Schwanenflugel & LaCount, 1988). A different picture emerges for low-constraint sentences. For these sen-

34

P.J. Schwanenflugel

tences. some studies have found facilitation for unexpected completions (Forster, 1981, Experiment 5 ; Schwanenflugel & Shoben, 1985; Stanovich & West, 1981, 1983). whereas others have not (Fischler & Bloom, 1979, 1985; Forster, 1981, Experiment 4). Schwanenflugel and LaCount (1988) reasoned that at least one source of the empirical disagreements between researchers is the degree of semantic relatedness of the unexpected words to the expectations subjects generate from the context, which is best indicated by the word that subjects expect given the context. When they explicitly varied this factor in a series of experiments, they found that unexpected words related to the expected completion for low constraint sentences tended to be facilitated whereas semantically unrelated, unexpected words did not. In conclusion, the pattern that emerges from the various studies on the scope of facilitation for upcoming words in sentences is that sentence context effects are not as general as might be assumed. High constraint sentences seem to engender a very narrow scope of facilitation including only expected, but not unexpected, words. Low constraint sentences yield a somewhat broader, although weaker, scope of facilitation than high constraint sentences. However, this scope of facilitation for low constraint sentences does not appear to extend to words semantically unrelated to subjects’ expectations. In fact, words that are unrelated to subjects’ expectation do not benefit from appearing in a sentence context, despite their being congruous and highly acceptable sentence completions.

WORDAND SENTENCE CONSTRAINT COMPARED Having summarized the available literature on word-level and sentencelevel contextual constraints, there are several issues regarding contextual constraint in general that need to be considered at this point. First, are sentence constraint effects merely a reflection of word-level constraints? Second, does contextual contraint operate similarly in word and sentence contexts? I shall deal with each of these in turn. Several researchers have proposed that sentence context effects on word processing, when they occur, are attributable to simple word-level relationships among words in the sentcnce (Forster, 1979; Masson, 1986). and not to emergent properties of the sentence as a whole. The view is that the higher level language processes that are engaged by sentences and larger amounts of text cannot affect the recognition of upcoming words because lexical mechanisms operate autonomously from those external sources of information. Thus, sentence context effects are said to be the result of accumulated simple word-level spreading activation effects. There are several reasons to question that account. First, there are now several studies suggesting that there are emergent properties of the sentence beyond simple word-level effects. Comparing normal intact sentences to their

Conlextual Constraint and Lexical Processing

35

scrambled counterparts, Simpson, Peterson, Casteel, and Burgess (1989), Foss (1982). and Masson (1986, Experiment 1) all found that scrambled sentences did not facilitate their final words as much as intact sentences did. Simpson et al. (1989) provided a particularly strong test of this word-level view by constructing sentences explicitly containing highly associated words and noted, at best, minimal effects of such associations outside of intact sentence contexts. Thus, although scrambled sentences preserve the word-level relationships in the sentences, they cannot account for the totality of sentence-level effects. Second, if lexical-level effects are primarily responsible for sentence context effects, then it is hard to see how the differential scope of facilitation for high and low constraint sentence context could be explained by such a simple word-level factor. That is, in the Schwanenflugel and Shoben (1985) and Schwanenflugel and LaCount (1988) studies, the scope of facilitation for high constraint sentences included expected completions, but not unexpected completions. However, for low constraint sentences, the scope of facilitation was wider and included unexpected words related to the expected completion. If sentence context effects are word-level effects, why should those word-level effects operate differentially in low constraint sentences to include related, although unexpected congruous words. It would seem that lexical-level models would have to make uniform predictions regarding the fate of all related, unexpected words. Last, we have conducted a preliminary study in my laboratory to assess the degree to which semantic relatedness among words in sentences can predict facilitation from sentence contexts. Subjects were asked to rate the semantic relatedness of content words and expected target completions taken from the sentences used by Schwanenflugel and LaCount (1988). By using semantic relatedness ratings as predictors of facilitation scores from that study, we hoped to be able to determine the degree to which the semantic relatedness of individual content words could account for variations in contextual benefits for expected sentence completions. One model tested employed the summed semantic relatedness of all content words (SUM) to assess the view that sentence context effects were merely an aggregate of all simple relations in the sentence. Another model tested used semantic relatedness of the closest content word (CLOSEST) to test the view that semantic relatedness effects are brief and easily disrupted by intervening information, but add temporarily to the benefits in processing that words gain from sentence context (cf. Gough et al., 1980; Stanovich & West, 1983). These simple word-level constraint models were contrasted with a sentence-level model using sentence constraint as a predictor of contextual facilitation scores. This latter model tested the view that benefits from a sentence context are attributable to an emergent characteristic of the context taken as a whole which is best indicated by sentence constraint. Lastly, word frequency was also included as a factor in all models to contxol for characteristics of the words themselves and to address the general finding that contextual benefits are gen-

36

P.J. Schwanenflugel

erally greater for high frequency than low frequency words (Schuberth et al., 1981). The conclusion of these analyses was that neither of the lexical-level models (SUM or CLOSEST) accounted for significant proportions of the variance in facilitation scores beyond that accounted for by word frequency alone. On the other hand, the sentence constraint did add significantly to a model including word frequency alone as well as to both of the lexical-level models. Therefore, we concluded that the simplest model to account for contextual benefits in word processing is one in which word-level effects, at best, comprise a minor part and sentence-level context effects are central. Given the conclusion that sentence-level context effects are not merely a conglomeration of word-level effects, we can consider the degree to which contextual constraint as a construct operates similarly in word and sentence contexts. However, this comparison is not a straightforward one. That is, the relationship that characterizes word-level constraint in word prime studies actually resembles the relationship between expected and unexpected words discussed in sentence context studies more than it does sentence constraint. For the most part, word context studies on this topic have tended to employ targets that are either very highly associated or good category member words and have contrasted them with some other, less associated or poor category member words. This is much more similar to the practice of selecting the most expected completion from sentence contexts and contrasting them wilh some other unlikely completion to examine expectancy effects in sentence context studies. What is striking about this latter contrast between word and sentence context studies is how similar these contextual influences are. Word contexts appear to have a beneficial influence on the processing of expected completions but only minimal benefits for unexpected completions. Similarly, sentence contexts tend to benefit the processing of expected completions, but do so for unexpected completions under some highly circumscribed conditions only. Therefore, it seems that the scope of facilitation for upcoming words from both word and sentence contexts are fairly narrow indeed. Despite this gross similarity in the way that contextual constraint seems to operate in word and sentence contexts, we cannot be certain that, in fact, similar processes are acting uniformly in both types of contexts. Clearly further research needs to be directed at distinguishing the similarities and differences in the way that contextual constraint behaves in word and sentence contexts.

THEORIIB OF WORD RECOGNITION A N D CONTEXTUAL CONSTRAINT In previous work, I have proposed a mechanism that may enable us to characterize the basic patterns of contextual benefits that occur as a function of contextual constraint. Contextual constraint, in this view, might be viewed as influencing the number and kind of featural restrictions generated as a result of

Contextual Constrain t and LexicaI Processing

37

processing the context. Specifically, readers are said to generate a greater number of more specific features for high constraint than low constraint contexts. So, for example, readers might generate many elaborate featural restrictions for the high constraint sentence “The tired mother gave her dirty child a ,” such as [cleans], [given by mothers], [taken by humans], [common to children], etc. The view is that as the number of featural restrictions generated from the context increases, then the number of potential completions in the lexicon whose meanings match that description decreases. According to this description, in order for a word to benefit from appearing in a context, it must match the entire featural restriction set generated by the reader. So, for example, in the above sentence, if enough features are generated from the context it may happen that nearly the entire featural description for a word such as bath might be generated, excluding even related, plausible words such as shower, and other less related, plausible words like scolding. For low constraint contexts, fewer featural restrictions are said to be generated from the context such that a greater number of possible completions will match the description and show facilitation. Thus, for the low constraint sentence “Hank reached into his pocket to get the ,” readers might generate only the features [usually found in pockets] and [small], which would match the expected completion money, and the unexpected completion coin, but not an unexpected, unrelated word such as watch. So, we would expect a broader scope of facilitation for low constraint than high constraint contexts, but not so broad so as to include words minimally related to the featural restrictions generated from the context. This featural restriction mechanism for describing the influence of contextual constraint accounts well for the sentence constraint effects shown by Schwanenflugel and Shoben (1985) and Schwanenflugel and LaCount (1988). This featural restriction mechanism might also be able to describe various other kinds of findings related to contextual constraint. For instance, it was previously noted that low constraint sentences don’t appear to facilitate even their expected completions as consistently as high constraint sentences. If it is assumed that the sampling of potential featural restrictions varies probabalistically across subjects for low constraint sentences because of their wider array of potential completions, sometimes featural restrictions might be sampled that eliminate even plausible, most expected completions. This would create more tenuous facilitation patterns for expected completions in low than high constraint contexts. Moreover, it is likely that the particular features that indivual low constraint sentences activate will also determine the degree to which different alternative completions are benefitted by such contexts (Tabossi, 1988a). This mechanism would be able to account for the overall finding of minimal facilitation for low dominant words following associate or category primes in a similar manner. That is, in most of those word context studies, it seems likely that the low dominant words used were relatively unrelated to the domi-

38

P.J. Schwanenflugel

nant target words. Given that the dominant target words best represent the featural expectations subjects generate from the context, the less dominant target words are likely to mismatch on some features. As a result, these words would not benefit from prior presentation of the context, particularly when the dominant word is highly expected. A similar kind of mechanism may also be responsible for determining the number of meanings that are accessed for ambiguous words. Qpically, both meanings of an ambiguous word are accessed initially regardless of context (Swinney, 1979; Tanenhaus. Leiman, & Seidenberg, 1979; see Simpson, 1981. for a review). Several researchers have shown that, when context places a lot of featural constraint on the meaning of ambiguous words, only meanings related to context are accessed (Simpson, 1981; Tabossi, 1988b; Tabossi, Colombo, & Job, 1987). This is true even when lexical-level priming explanations can be eliminated (Tabossi, 1988b). This featural restriction mechanism can be incorporated easily into some existing models of word recognition during reading. Models of word recognition can generally be divided into three types (Simpson et al.. 1989): 1) those that place the influence of context at the earliest phases of lexical access, 2) those that assign the influence of context to the resolution of word identity among several candidates, and 3) those that assign the influence of context (particularly sentence contexts) to decisional processes occurring after the word has already been fully identified. Below I will describe how some current models of word recognition might incorporate contextual constraint. Among the models placing the influence of context at the earliest phases of lexical access are the Two-Process Model (Neely, 1977; Neely, in press), the Verification Model (Becker, 1980), and the Interactive Activation Model (McClelland & Rumelhart, 1981). There are two points at which contextual constraint can potentially operate in the Two-Process Model. The first is during the quick acting, spreading activation process that is said to activate all words related to the context indiscriminately. It is doubtful that locus of the complex pattern of constraint effects occur here, because it would be unable to explain the lack of facilitation for unexpccted words related to high constraint sentences (Schwanenflugel & LaCount, 1988; Schwanenflugel & Shoben, 1985). On the other hand, there is relatively limited research monitoring such facilitation patterns at short SOAs, so this possibility of locating some effects here cannot be entirely ruled out. The second process is a slower acting, expectancy mechanism which both inhibits recognition of inappropriate words and further facilitates the recognition of appropriate ones. This process seems ideal for characterizing contextual constraint effects in that, if subjects are attending to the features generated by the context, words mismatching on some features would not be facilitated in their recognition. Such a process could yield greater facilitation for high constraint contexts than low constraint contexts because the activation level of words meeting a greater number features is higher than when

Contextual Constraint and Lexical Processing

39

fewer features are generated. Related, unexpected words would be facilitated only in low constraint contexts because they are more likely to meet the few featural restrictions generated from the context. The Verification Model (Becker, 1980) would place the influence of contextual constraint at the point of semantic set formation. According the model, when words are presented in semantic context, persons form a semantic set containing words that are consistent with the context. When target words are presented, the stimuli are matched against this set and facilitated if found there. The featural restrictions generated from the context might be able to determine the number and kind of words that enter the semantic set. Only words matching the entire featural restriction set would remain in the semantic set, so that for high constraint sentences only expected completions would be included whereas for low constraint sentences a greater number of words would be included in the set. The Interactive Activation Model (McClelland & Rumelhart, 1981) would place the locus of contextual constraint at the earliest stages of activation and inhibition of words in the lexicon. According to this model, immediately when context is presented, recognition of consistent words is benefitted through the activation of positive faciliatory connections and inconsistent words are interfered with through lateral inhibitory connections. The current featural restriction formulation could be incorporated by assuming that the influence of contextual constraint would take place at the semantic feature level. Thus, when only a few semantic features are activated, words consistent with those features would become activated and would immediately start to inhibit incompatible words. The particular strength of this model is that the greater number of words activated through compatability with the featural restrictions, the less any particular word would become activated because of lateral inhibition for words within the same level. Thus, this model would particularly be able to predict the reduced amount of facilitation shown for even expected words appearing in low constraint sentences. In other models, contextual constraint would serve to assist in resolving the identity of target words from a set of candidates defined by the bottom-up visual information. For example, the checking model described by Norris (1986) views context as serving to lower the recognition threshold for words consistent with context. Therefore, while the context does not itself activate word candidates, it can produce contextual facilitation through changes in response bias. The featural restriction mechanism can be implemented in this model by stating that only words that exhaustively match the featural description set up by the context will have their thresholds lowered and, only then, in proportion to the number of features they match. While this approach could account for the influence of contextual constraint when bottom-up information is available, it is hard to see how contextual constraint effects in word production could be accounted for.

P.J. Schwanenfugel

40

Lastly, other models, such as the lexical autonomy view of Forster (1979) and the integration view espoused by Masson (1986). would have to view the contextual constraint effects emanating from sentences as being attributable either to intralexical priming or post-lexical decisional processes and not to priming from sentence-level information. However, as noted earlier, the similarity of the word-level and sentence-level constraint effects lead me to believe that there is no need to account for these contextual constraint effects separately. Consequently, if all contextual constraint effects are then attributed to post-lexical decisional processes, then it must assume that the these processes monitor the feature match between expectations derived from the context and features of the target word as the word is being integrated into the context. However, this view would have greater difficulty accounting for contextual constraint effects in production as well as in naming (which has been shown to be less susceptible to post-lexical processes in other research).

SUMMARY One thing that has become very clear from all the related work on contextual constraint is that the influence of context is not as all-encompassing as was once thought to be the case. Specifically, high constraint contexts seem to engender a much narrower scope of facilitation than low constraint contexts. Moreover, words that are minimally related to the context appear rarely to benefit from appearance in a meaningful context. For the most part, there appears to be a great deal of commonality in the way that contextual constraint operates from words and sentences. On the other hand, some of the critical comparisons between the two context types have yet to be examined experimentally. Without further evidence, it can be at least tentatively concluded that contextual constraint operates uniformly across contexts. Finally, studies of contextual constraint have the potential to lend us information relevant for determining the locus and manner in which context influences the processing of words. As a result, such studies should assist us in deciding among alternative theories of lexical processing. Regardless the view of lexical processing that is eventually decided upon, contextual constraint has emerged as an important factor in the processing of words and will need to be included in any theory of lexical processing. Acknowledgments

This work was supported in part by a National Science Foundation Grant BNS-8808453 and an Elva Knight Grant from the International Reading Association. I thank R. White for his helpful comments on an earlier version of the chapter.

Contextual Constraint and Lexical Processing

41

References Anderson, J. R. (1976). Language, memory, and thought. New York: Erlbaum. Antos, S. J. (1979). Processing facilitation in a lexical decision task. Journal of Experimental Psychology: Human Perception and Performance 5 , 527545.

Balota, D. A., & Duchek, J. M. (1988). Age-related differences in lexical access, spreading activation, and simple pronunciation. Psychology and Aging 3, 84-93. Balota, D. A., & Chumbley, J. I. (1984). Are lexical decisions a good measure of lexical access? The role of word frequency in the neglected decision stage. Journal of Experimental Psychology: Human Perception and Performance, 10, 340-357. Balota, D. A., & Lorch, R. F. (1986). Depth of automatic spreading activation: Mediated priming effects in pronunciation but not in lexical decision. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 336-345. Balota, D. A., Pollatsek, A., & Rayner, K. (1985). The interaction of contextual constraints and parafoveal visual information in reading. Cognitive PSyCholOgy, 17, 364-390. Battig, W. F., & Montague, W. E. (1969). Category norms for verbal items in 56 categories: A replication and extension of the Connecticut category norms. Journal of Experimental Psychology Monograph, 80 (3, Pt. 2). Becker, C. A. (1980). Semantic context effects in visual word recognition: An analysis of semantic strategies. Memory & Cognition, 8,483-512. Bloom, P.A., & Fischler, I. (1980). Completion norms for 329 sentence contexts. Memory & Cognition, 8,63 1-642. Bormuth, J.R. (1968). The cloze readability procedure. In J.R. Bormuth (Ed.), Readability in 1968. Champaign, IL: National Council of Teachers of English. Cohen, G., & Faulkner, D. (1983). Word recognition: Age differences in contextual facilitation effects. British Journal of Psychology, 74, 239-251. deGroot, A. M. B., Thomassen, A. J., & Hudson, P. T. (1982). Associative facilitation of word recognition as measured from a neutral prime. Memory & Cognition, 10, 358-370. den Heyer, K., Briand, K., & Smith, L. (1985). Automatic and strategic effects in semantic priming: An examination of Becker’s Verification Model. Memory & Cognition, 13,228-232. Ehrlich, S . F., & Rayner, K. (1981). Contextual effects on word perception and eye movements during reading. Journal of Verbal Learning and Verbal Behavior, 20, 641-655. Fischler, I. (1977). Semantic facilitation without association in a lexical decision task. Memory & Cognition, 5 , 335-339.

42

P.J. Schwanenfugel

Fischler, I. (1983). Contextual constraint and the comprehension of written sentences by deaf college students. American Annals of the Deaf, 228, 4 18-424. Fischler, I. (1985). Word recognition, use of context, and reading skill among deaf college students. Reading Research Quarterly, 20,203-218. Fischler, I., & Bloom, P. A. (1979). Automatic and attentional processes in the effects of sentence contexts on word recognition. Journal of Verbal Learning and Verbal Behavior, 18, 1-20. Fischler, I., & Bloom, P. A. (1985). Effects of constraint and cue validity of sentence contexts on lexical decisions. Memory & Cognition, 13, 128139. Fischler, I., & Goodman, G. 0. (1978). Latency of associative activation in memory. Journal of Experimental Psychology: Human Perception and Performance, 4,455-470. Forster, K. I. (1979). Levels of processing and the structure of the language processor. In W. E. Cooper & E. C. T. Walker (Eds.). Sentence processing: Psycholinguistic studies presented to Merrill Garrett. Cambridge, MA: MIT Press. Forster, K. I. (1981). Priming and the effects of sentence and lexical contexts on naming time: Evidence for autonomous lexical processing. Quarterly Journal of Experimental Psychology, 33A, 465-495. Foss, D. J. (1982). A discourse on semantic priming. Cognitive Psychology, 14, 590-607. Glass, A. L., & Holyoak, K. J. (1975). Alternative conceptions of semantic memory. Cognition, 3 , 313-339. Howard, D. V. (1983). The effects of aging and degree of association on the semantic priming of lexical decisions. Experimenfal Aging Research, 9, 145-151 . Jenkins, J.J. (1970). The 1952 Minnesota word association norms. In L. Postman and G. Keppel (Eds.), Norms of word association. New York: Academic Press. Jongsma, E.A. (1980). Cloze instruction research: A second look. Newark, DE: International Reading Association. Jonides, J.. & Mack, R. (1984). On the cost and benefit of cost and benefit. Psychological Bulletin, 96, 2944. Keppel, G., & Strand, B.Z.(1970). Free-association responses to the primary responses and other responses selected from the Palermo-Jenkins norms. In L. Postman and G. Keppel (Eds.), Norms of word association. New York: Academic Press. Kleiman, G. M. (1980). Sentence frame contexts and lexical decisions: Sentence acceplability and word relatedness effects. Memory & Cognition, 8, 336-344. Kutas, M., & Hillyard, S. A. (1984). Brain potentials during reading reflect

Contextual Constraint and Lexical Processing

43

word expectancy and semantic association. Nature, 307, 161-163. Kutas, M., Lindamood, T. E., & Hillyard, S. A. (1984). Word expectancy and event-related brain potentials during sentence processing. In S. Kornblum and J. Requin (Ed.), Preparatory states and processes (pp. 217-238). Hillsdale, NJ: Erlbaum. Lorch, R. F. (1982). Priming and search processes in semantic memory: A test of three models of spreading activation. Journal of Verbal Learning and Verbal Behavior, 21,468-492. Lorch, R. F., Balota, D. A., & Stamm, E. G. (1986). Locus of inhibition effects in the priming of lexical decisions: Re- or postlexical access. Memory & Cognition, 14, 95-103. Lupker, S. J. (1984). Semantic priming without association: A second look. Journal of Verbal Learning and Verbal Behavior, 23,709-733. Massaro, D. W., & Jones, R. D., Lipscomb, C., & Scholz, R. (1978). Role of prior knowledge on naming and lexical decisions with good and poor stimulus information. Journal of Experimental Psychology: Human Learning and Memory, 4,498-512. Masson, M. E. J. (1986). Comprehension of rapidly presented sentences: The mind is quicker than the eye. Journal of Memory and Language, 25,588604. McClelland, J. L., & O’Regan, J. K. (1981). Expectations increase the benefit derived from parafoveal information in reading words. Journal of Experimental Psychology: Human Perception and Performance, 7,634-644. McClelland. J. L., & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part 1. An account of basic findings. Psychological Review, 88, 375-407. McCloskey, M. & Glucksberg, S. (1979). Decision processes in verifying category membership statements: Implications for models of semantic memory. Cognitive Psychology, 11. 1-37. McKenna, M.C., & Robinson, R.D. (1980). An introduction to the Cloze procedure. Newark, DE: International Reading Association. Mervis, C.B., & Rosch, E. H. (1981). Categorization of natural objects. Annual Review of Psychology, 32, 89-115. Nebes, R. D., Roller, F., & Holland, A. (1986). Use of semantic context by patients with Alzheimer’s disease. Psychology and Aging, I, 261-269. Neely, J. H. (1977). Semantic priming and retrieval from lexical memory: Evidence for faciliatory and inhibitory processes. Journal of Experimental Psychology: General, 106, 226-254. Neely, J. H. (in press). Semantic priming effects in visual word recognition A selective review of current findings and theories. In D. Besner & G. Humphreys (Eds.), Basic processes in reading: Visual word recognition. Hillsdale, NJ: Erlbaum. Neely, J. H., Keefe, D. E., & Ross, K. L. (1989). Semantic priming in the

44

P.J. Schwanenflugel

lexical decision task: Roles of prospective prime-generated expectancies and retrospective semantic matching. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 1003-1019. Norris, D. (1986). Word recognition: Context effects without priming, Cognition, 22, 93-136. Norris, D. (1987). Strategic control of sentence context effects in a naming task. Quarterly Journal of Experimental Psychology, 39A, 253-275. Schuberth, R. E., Spoehr, K. T., & Lane, D. M. (1981). Effects of stimulus and contextual information on lexical decision processes. Memory & Cognition, 9, 68-77. Schwanenflugel, P. J. (1985). Reading skill and the recognition of unexpected words. Paper presented at the Georgia Educational Research Association meeting, Atlanta, GA. Schwanenflugel, P.J. (1986). Completion norms for sentences using a multiple production measure. Behavior Research Methods, Instrumentation, & Computers, 18, 363-371. Schwanenflugel, P. J., Harnishfeger. K. K., & Stowe, R. W. (1988). Context availability and lexical decisions for abstract and concrete words. Journal of Memory and Language, 27,499-520. Schwanenflugel, P. J., & LaCount, K. (1988). Semantic relatedness and the scope of facilitation for upcoming words in sentences. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 344-354. Schwanenflugel, P. J., & Rey, M. (1986). Interlingual semantic facilitation: Evidence for a common representational system in the bilingual lexicon. Journal of Memory and Language, 25,605-618. Schwanenflugel, P. J., & Shoben, E. J. (1985). The influence of sentence constraint on the scope of facilitation for upcoming words. Journal of Memory and Language, 24,232-252. Schwantes, F. M. (1985). Expectancy, integration and interactional processes: Age differences in the nature of words affected by sentence context. Journal of Experimental Child Psychology, 39,212-229. Seidenberg, M., Waters, G, Sanders, M., & Langer, P. (1984). Pre- and post lexical loci of contextual effects on word recognition. Memory & Cognition, 1 2 , 315-328. Shanahan, T., Kamil, M.L., & Tobin, A.W. (1982). Cloze as a measure of intersentential comprehension. Reading Research Quarterly, 17, 229-255. Simpson, G. B. (1981). Meaning dominance and semantic context in the processing of lexical ambiguity. Journal of Verbal Learning and Verbal Behavior, 20, 120-136. Simpson, G. B. (1984). Lexical ambiguity and its role in models of word recognition. Psychological Bulletin, 96, 3 16-340. Simpson, G. B., Peterson, R. R., Casteel, M. A., & Burgess, C. (1989). Lexical

Contextual Constraint and Lexical Processing

45

and sentence context effects in word recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15.88-97. Smith, E. E., Shoben, E. J., & Rips, L. J. (1974). Structure and process in semantic memory: A featural model for semantic decisions. Psychological Review, 81,214-241. Stanovich, K. E., & West, R. F. (1983). On priming by a sentence context. Journal of Experimental Psychology: General, 112, 1-36. Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18,645-660. Tabossi, P. (1988a). Effects of context on the immediate interpretation of unambiguous nouns. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 153-162. Tabossi, P. (1988b). Accessing lexical ambiguity in different types of sentential contexts. Journal of Memory and Language, 27,324-340. Tabossi, P., Colombo, L., & Job, R. (1987). Accessing lexical ambiguity: Effects of context and dominance. Psychological Research, 49, 161-167. Tanenhaus, M.K., & Carlson, G.N. (in press). Lexical structure and language comprehension. In W.D. Marslen-Wilson (Ed.), Lexical representation and process. Cambridge, MA: MIT Press. Taylor, W. (1953). Cloze procedure: A new tool for measuring readability. Journalism Quarterly, 30,415-433. Tyler, L. K., & Wessels, J. (1983). Quantifying contextual contributions to word-recognition processes. Perception & Psychophysics, 34,409-420. Warren, R. E. (1977). Time and the spread of activation in memory. Journal of Experimental Psychology: Human Learning and Memory, 3,458-466. West, R. F., & Stanovich, K. E. (1982). Source of inhibition in experiments on the effect of sentence context on word recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 8, 395-399. West, R.F., & Stanovich, K.E. (1986). Robust effects of syntactic structure on visual word processing. Memory & Cognition, 14, 104-112. Wright, B., & Garrett, M. (1984). Lexical decision in sentences: Effects of syntactic structure. Memory & Cognition, 12, 31-45.

This Page Intentionally Left Blank

Understanding Word and Sentence G.B.Simpson (Editor) (B Elsevier Science Publishers B.V. @’orth-HoUand), 1991

Chapter 3 Contextual Feature Activation and Meaning Access George Kellas, Stephen T. Paul, Michael Martin University of Kansas Lawrence, Kansas U.S.A. and Greg B. Simpson University of Nebraska at Omaha Omaha, Nebraska U.S.A.

A word is a string of letters that symbolizes an abstract or concrete, internal or external, environmental or situational referent. As youngsters learning the alphabet we were taught that “A” is for “apple,” and later that A-P-P-LE spells “apple.” During this early period of learning to read we became aware of the seeming one-to-one correspondence between words and objects apart from the specific contexts in which these associations occurred. If one knows apples only through a dictionary definition, the word “apple” is likely to refer to an edible red fruit and nothing more. However, the son of an apple-picker may see apples as representing a way of life and use the word “apple” accordingly. Ultimately, our personal expcricnce with referents in context gives words their meaning. Hence meaning (i.e., reference) can be viewed as a product of experience that mediates the relationship bctween referents and their word symbols (cf. Olson, 1970). Consider the reference of piano for musicians versus that for movers (cf. Barclay, Bransford, Franks, McCarrell, & Nitsch, 1974). Although musicians know pianos are heavy and movers know they are musical instruments, experience makes different features of pianos more or less salient (important to the meaning of the referent). Thus a musician and mover can refer to the same piano yet the two references may be quite different. However, suppose the mover is taking piano lessons or the musician is moving. Changes in context will differentially emphasize aspccts of a word’s meaning (cf. Schoen, 1988; Tabossi & Johnson-Laird, 1980). Changes in context, and the corresponding shift in meaning within and across individuals. reflect the dynamic nature of language comprehension (Simpson & Kellas, 1989). For these reasons, a speaker or writer must construct a context that emphasizes the features of the

48

G.Kellas, S.T. Paul, M . Martin and G.B. Simpson

intended meaning. The listener or reader can then use the context provided as a common frame of reference from which to retrieve a specific meaning (Grice, 1975). Effective communication makes use of the fact that the listener or reader will be able to infer much of the information necessary for comprehension, based on shared world knowledge. Breakdowns in communication will occur when the intended reference is not adequately specified, explicitly or implicitly. In such a circumstance, further clarification on the speaker or writer’s part is necessary until the message is sufficiently restricted to what was intended (Osgood, 1968). Communication, then, occurs when the range of perceived alternatives to an intended reference has been contextually restricted to the appropriate meaning (cf. Olson, 1970). When the pattern of features activated does not reduce the potential number of references to a specific sense, ambiguity results. Therefore it becomes important to determine the role context plays in specifying the salient features of word meanings and, consequently, the intended meaning of a sentence. Barsalou (1982) has argued that the meaning of a word can be partitioned into context-independent and context-dependent features or properties. Those properties activated by the presence of a word, regardless of context, are described as context-independent. Context-dependent properties are those which become activated as a consequence of a word in a specific context. Word meanings may be thought of as constellations of features that are activated in response to perceptual and linguistic inputs (Ashcraft, 19%; Barclay, et al., 1974; Barsalou, 1982; Just & Carpenter, 1987). The particular pattern of features activated by a specific context delineates meaning. In addition to the features associated with a single meaning of a word, multiple meanings can be associated with certain individual words, each with their own set of features. For example, the word “duck” may be uttered in the context of ordering a meal, hunting, or stowing luggage in the overhead compartment of an airplane. An ambiguous word, by definition, symbolizes more than one referent; however, ideally an ambiguous word in well-formed discourse should function to symbolize the meaning of a single referent. In fact, ambiguity is rarely perceived during normal discourse because words may be organized so as to make sensible only a single meaning, such as: The ship began to sink. However, such clarity is often not possible by syntax alone. Oftentimes, it is necessary to construct sentences by a judicious selection of words which activate, and thereby emphasize, specific aspects of a word’s meaning. If we are successful, the pattern of activated features will constrain meaning to a single reference.’ Ambiguous words are useful for examining the processes leading toward comprehension. More specifically, examination of the ways in which context specifies the meaning of an ambiguous word lends itself to the study of meaning access. At least two opposing views of the effects of context on meaning

Contextual Feature Activation and Meaning Access

49

retrieval have been proposed. The debate centers on whether or not prior context and/or expectation can determine which meaning of an ambiguous word is accessed. The issue is generally expressed in terms of the locus of context effects: pre- versus post-access. It may be the case that only contextually appropriate meanings are retrieved (selective or direct access) as found by Tabossi (1988a. 1988b) and others (e.g., Simpson, 1981; Tabossi, Colombo, & Job, 1987) under certain conditions. In these cases meaning access may be described as context-dependent because the meaning retrieved depends primarily on the context in which the word occurs (i.e., pre-access effects of context). Although, intuitively, it seems that only one meaning of a word has been retrieved, those who hold a multiple-access view of contextual effects on meaning access (e.g., Fodor, 1983; Onifer & Swinney, 1981; Seidenberg, Tanenhaus. Leiman, & Bienkowski, 1982; West & Stanovich, 1982) argue that all meanings are initially accessed, independent of context. The appropriate meaning is subsequently selected based on its semantic congruency with the context and inappropriate meanings are suppressed. Facilitation from prior context in this case is often assumed to derive almost entirely as a result of post-access phenomena (e.g., Fodor, 1983; Seidenberg, Waters, Sanders, & Langer, 1984; Tanenhaus & Lucas, 1987; West & Stanovich, 1982). If true, context may not facilitate meaning access per se but meaning selection, after all meanings have been accessed. In the following text we have three major goals: (1) to propose that context is critical in determining what aspects of word meaning become activated during reading comprehension, (2) to critically evaluate a task (cross-modal priming) currently accepted as least problematic for the study of meaning access, followed by a discussion of an alternate methodology, and (3) to describe research which makes use of normatively derived stimuli that represent the features activated by specific sentences. Finally, we will conclude that current efforts to specify the locus of context effects on meaning retrieval may be premature considering the greater issue of what is activated by context in general.

Context and Constraint There appears to be an implicit assumption in much of the literature that words elicit the same information from memory regardless of changes in surrounding context.* This assumption may be due, in part, to the early semantic priming studies in which context was defined by the semantic relationship between a single word prime and the response stimulus (target). The semantic relationship between the word (the context) and a target was ascertained by collecting normative data on the word in isolation (i.e.. associative norms). The targets selected from these procedures accurately reflected the information activated by the context (in this case, the word), as demonstrated by experimental

50

G. Kellas, S.T. Paul, M.Martin and G.B. Simpson

outcomes. When a sentence is used as a prime, however, it is not clear whether the information activated by a critical word in the sentence context is the same as that activated by the critical word in isolation. In principle, words may be considered the smallest independent carriers of meaning available for use within a language. As such, words may sometimes be considered the functional units of meaning. Clearly, though, context can be used to color a word’s meaning. Once a word is placed in a sentence, it may no longer serve as an isolable unit of meaning (Barclay et al., 1974). The meaning of a sentence, then, may not simply be the sum of the meanings of the individual words that constitute it. For instance, different contexts may change the semantic relationship between a critical “prime” word embedded in a sentence and a target obtained from norms collected on the word in isolation. Consequently, targets obtained from such procedures may be of little use when exploring meaning access during sentence comprehension. It follows that the selection of stimuli to assess meaning access in sentence contexts should be based on a consideration of the modifying influence context has on word meanings. Researchers would be well advised to use targets whose semantic relationship with the sentence is known, rather than extrapolated from normative procedures not designed to reflect the functional meaning of a word as it is used in context. Schwanenflugel and Shoben (1985) used normatively defined sentence constraint to explore the scope of facilitation for upcoming unambiguous words. The strength of sentence constraint was determined by a modified Cloze procedure in which subjects generated completions for sentence frames. A high constraint sentence frame reliably generated a particular word as its completion, whereas the completions for low constraint frames were less consistent. The sentence frames were used as primes for lexical decisions to sentence-final words. Targets selected from among subjects’ most frequent responses were considered expected completions. Targets used that were not generated frequently by subjects were considered unexpected completions. The context-specific, expected and unexpected targets were semantically related to, and congruent completions of, their respective sentence frames. In highly constraining contexts, facilitation was obtained for expected completions but not for semantically related, unexpected completions. Low constraint contexts, on the other hand, facilitated both expected and unexpected sentence completions. Schwanenflugel and Shoben (1985) interpreted their results in terms of the scope of facilitation provided by the context. Highly constraining contexts led to a more narrow scope of facilitation than low-constraining contexts. They suggested that the mechanism underlying the scope of facilitation for upcoming words involved the featural restrictions imposed by context. If the meaning of an upcoming word does not converge on the meaning constrained by prior context, processing will be impaired. High-constraint contexts activate very specific meanings. The meaning activated by a high-constraint sentence is so specific that not even words semantically related to the most expected word are

Contextual Feature Activation and Meaning Access

51

facilitated. In contrast, low constraint contexts activate meanings compatible with a variety of semantically related words. These findings were extended (Schwanenflugel & LaCount, 1988) to include an examination of semantically unrelated, but acceptable, sentence completions. It was found that unrelated completions were not included within the scope of facilitation for either high- or low-constraining contexts, even though acceptable as completions. This strongly suggests that featural restrictions from context determines the scope of facilitation. Acceptable but unrelated targets would have little, if any, featural overlap with the features activated during sentence processing. In a similar vein, Tabossi (1988a) used sentences which emphasized specific aspects of the meaning of sentence-final, unambiguous nouns. All targets were related to the sentence-final noun. Target identification was facilitated when a sentence prime was followed by a target related to the contextually emphasized aspect. Processing was inhibited, though, whenever a target not emphasized by the context was encountered. Tabossi’s (1988a) results converge on the research of Schwanenflugel and colleagues (Schwanenflugel & LaCount, 1988; Schwanenflugel & Shoben, 1985) to support the notion that context delineates a word’s meaning via feature priming. According to Olson (1970), the purpose of context is to discriminate among potential alternatives to an intended reference. Schwanenflugel and LaCount (1988). Schwanenflugel and Shoben (1985), and Tabossi (1988a) have provided evidence in support of this for unambiguous words. Simpson (1981). Tabossi (1988b). and Tabossi et al., (1987) have provided evidence in support of Olson’s principle for ambiguous words as well. Context which succeeds in specifying a reference may be considered well-formed discourse. It should be noted that sentences containing ambiguous words which do not adhere to the principle of well-formed discourse may well result in initial access of multiple meanings followed by post-access selection. For example, if the sentence, “They went to the bar.” is encountered in an impoverished experimental setting, it may refer to a drinking establishment, a playground apparatus, or perhaps a sand bar somewhere along the eastern shore. Evidence of multiple access in this instance would merely reflect insufficient context. Simpson (1981) demonstrated that the nature of meaning access is dependent on the strength to which a sentence is biased toward a single meaning. When a context strongly biased a specific meaning of an ambiguous word, only that meaning was accessed. However, when the context weakly biased a meaning, access was selective for the most frequently used meaning (dominant) and exhaustive for the subordinate sense. Tabossi et al., (1987) also have obtained support for selective context effects on meaning access. However, these priming contexts were not designed so as to bias certain meanings, but to activate salient features of ambiguous word senses. The results were consistent with a selective access interpretation when contexts made features of the dominant

52

G . Kellas, S.T. Paul, M . Marlin and G.B. Simpson

sense salient. Yet multiple access obtained when features of the subordinate sense were made salient. Tabossi (1988b) conducted three experiments using dominant biasing sentences to explore further the nature of meaning access. However, in her first two experiments, she used sentences that did not emphasize a specific aspect of meaning. Rather, the sentences used were determined by a panel of judges to bias only the dominant sense of the embedded homonym without making a specific feature salient. These sentences can be considered low-constraint because they did not activate specific features associated with the context (cf. Schwanenflugel & LaCount, 1988; Schwanenflugel & Shoben, 1985). In Experiment 1, associates to the homonym in isolation were used as targets. Evidence of multiple access was obtained, converging on Schwanenflugel and Shoben’s finding that low-constraint contexts facilitate both expected and unexpected targets. In Experiment 2, the same sentences were used as primes. Targets, however, were specific features related to the dominant or subordinate sense of the homonym. Note that the features used as targets were not specifically primed by the low-constraint contexts as determined by the panel of judges. Once again, multiple access was obtained. In terms of Schwanenflugel and colleagues’ (Schwanenflugel & LaCount. 1988; Schwanenflugel & Shoben, 1985) findings, low-constraint Sentences may bias the interpretation of an ambiguous word, but have a wider scope of feature activation than high-constraint sentences. Therefore, features of both the dominant and subordinate sense may become activated. In Experiment 3, sentences were used which emphasized a specific aspect of meaning (high constraint) followed by a target associatively related to the ambiguity. Based on Schwanenflugel and Shoben’s (1985) research, these high-constraint sentences should produce a narrow scope of facilitation and increase the likelihood of related targets being selectively accessed. In fact, Experiment 3 yielded evidence in support of selective access for associatively related targets. However, we claim that for this outcome to occur, the associative targets must have represented features activated during sentence processing. Conjointly, the research of Tabossi (1988b) and Schwanenflugel and Shoben (1985) serves to emphasize the importance of an underlying feature priming mechanism, as well as to highlight the function of context in restricting the features activated. The strong bias condition in Simpson’s (1981) research also yielded evidence of selective access using targets from association norms to the homonym in isolation. Both Tabossi (1988b) and Simpson (1981). then, have demonstrated that evidence for selective access can be obtained using associatively related targets of the homonym in isolation, if the sentence context highly constrains meaning. There remains, however, some discrepancy bctween predictions from Schwanenflugel and Shobcn (1985) and Tabossi’s (1988b) Experiment 3 that are not immediately obvious. At first glance, the outcomes of Tabossi (1988b) and Simpson (1981)

Contextual Feature Activarion and Meaning Access

53

seem to suggest that using associative norms to obtain targets for exploring meaning access of words in sentences is acceptable, as long as the context is highly constraining. However, the research of Schwanenflugel and Shoben (1985) must be taken as a warning. Schwanenflugel and Shoben’s results suggest that words semantically related to a meaning of an ambiguous word in isolation might not be included in the scope of facilitation from a sentence. Therefore, a sentence which highly constrains the meaning of an ambiguous word may not yield evidence of selective access if the probe words are taken from norms based on words in isolation. The research of Schwanenflugel (Schwanenflugel & LaCount, 1988; Schwanenflugel & Shoben, 1985), Simpson (1981), and Tabossi (Tabossi, 1988a; 1988b; Tabossi, et al., 1987) converge on the notion that increasing contextual constraint serves to increase the specificity of meaning accessed relative to low constraint or more impoverished sentences. These studies and related findings (cf., Whitney, McKay, Kellas, & Emerson, 1985) suggest a strong relationship between contextual feature activation and the processing of subsequent words. The features primed by a sentence will determine the scope of facilitation. One might, therefore, expect sentence constraint to be a function of the degree of overlap between features activated by context and the features represented by upcoming words. The mechanism of feature priming provides the instantiation of meaning for both unambiguous and ambiguous words. Highconstraint sentences activate features which serve to instantiate a specific meaning of either an unambiguous or ambiguous word. In the case of unambiguous words, high-constraint sentences prime a reduced range of features. In a similar fashion, contextually activated features of ambiguous words will facilitate processing of words representing only those features related to the appropriate sense. Low-constraint sentences, on the other hand, increase the range of features activated for both unambiguous and ambiguous words. In the case of ambiguous words, this is evidenced by the access of multiple meanings. The influence of context on sentence integration and comprehension processes can be evaluated by examining the features of a particular referent activated in different sentences (e.g., Tabossi, et al., 1987). The role of context in the access of an intended meaning of an ambiguous word may be similarly evaluated. Indeed, contextual feature priming may be the primary mechanism involved in selective or multiple access outcomes. If it is the case that sentence constraint is critical in determining the extent to which prior context facilitates access of word meanings, then the issue is not simply whether context affects meaning access. Rather the research emphasis placed on the time-course of meaning access should include a consideration of the dynamic nature of context and word meaning. It has been proposed above that meaning is represented by features activated by context which need not be specified by word-symbols alone. The

54

G. Kellas, S.T. Paul, M. Martin and G.B. Simpson

intent or purpose of a speaker or writer determines which features of a referent are to be made most salient for communication (i.e., word choice). Therefore, the functional context must include the features activated by inference processes, world knowledge, and the listener’s representation of the speaker’s intent. It is our view that context provides the information necessary to discriminate among alternate word senses by differentially emphasizing a referent’s composite features. It is hypothesized that the concept of lexical ambiguity derives solely as a result of unspecified knowledge. When a reader or listener perceives ambiguity, it is a consequence of inadequate contextual specification. Sentence context is not used to resolve ambiguity, rather, ambiguity exists only as a result of insufficient context. It becomes important, then, to determine how the interaction of individual words results in eventual comprehension. In order to do this, the experimental tasks chosen for research must precisely reflect when and what features are activated by sentences.

METHODOLOGICAL ISSUES Cross-Modal Priming Access to word meaning seems to occur over time (e.g., McClelland, 1987). Consequently, researchers must employ tasks which carefully control for the amount of processing an ambiguous word receives before measures of meaning access are taken. To this end, many researchers use the cross-modal priming procedure offered by Swinney (1979). This procedure rcquires that subjects listen to auditorialy presented material and at some point respond (lexical decision or naming) to a visual target shown at a critical moment during or after the acoustic stimulus. By implication, the cross-modal task has been offered as a panacea for examining contextual effects on meaning access (Tanenhaus, Leiman, & Seidenberg, 1979; Seidenberg et al., 1982; Swinney, 1979). One reason that crossmodal tasks enjoy such favor is that visual targets can be presented to subjects with experimental precision relative to an auditorialy presentcd stimulus that may not be possible with visually presented sentences. However, given the flexibility of a language processing system, it may be the case that subtle variations in procedure will qualitatively alter the evidence regarding meaning access. Three procedural variants are currently available for use within the cross-modal paradigm. Differences among procedures essentially arise from variations in the structure and presentation of the stimuli. The specific form of the cross-modal procedure depends on the location of an ambiguous word, and therefore, presentation of a target during processing. In some studies, the auditorialy presented ambiguous word concludes the sentence, at which point a visually presented target appears (e.g.. Seidenberg et al., 1982). Presumably the target is presented after the sentence has been fully encoded and is unaffected by an acoustic trace extending over time. This will

Contextual Feature Activation and Meaning Access

55

be referred to as the sentence-final procedure. Another task variation involves embedding the ambiguous word within the sentence. Auditory presentation may terminate following the ambiguous word, essentially leaving subjects with only a sentence fragment to process (e.g., Seidenberg et al., 1982). This sentencemedial procedure, which superficially resembles the sentence-final procedure, results in presentation of the target at the offset of the critical stimulus word. An obvious criticism of sentence-medial procedures is that priming ambiguous words with sentence fragments will likely provide weaker contexts than methods which present post-sentential or sentence-find targets. Rather than terminating the sentence at the ambiguous word, the visual target may be presented while subjects continue to listen to the remainder of the sentence (e.g., Onifer & Swinney, 1981; Tabossi, 1988a, 1988b). With this sentence-interrupt variation, it is assumed that presentation of a visual target during auditory processing does not interfere with comprehension (Swinney, 1979). Although it has not been empirically verified, the sentence-interrupt version closely resembles a dual-task procedure in which subjects must continue processing the auditory sentence while performing a task related to a visual target. If this is true, it might be expected that the attentional resources required for sentence processing would reduce the resources available for visual target recognition and influence the experimental results. The problem would be exacerbated with increasing sentence complexity. Responses to targets presented in the sentence-interrupt method are the most likely to be affected by the dual-task nature of the cross-modal procedure. The subject’s allocation of resources to sentence processing and word recognition will be sensitive to the nature of the sentence task (comprehension, recognition, verbatim recall, etc.). The more demanding that task, the fewer resources will be available for word recognition, and this will particularly be the case when the sentence information is not yet complete. In the case of a relatively demanding sentence task, the subject may delay before switching to the word recognition task, thus functionally lengthening the processing time for the homograph. Consequently, it is difficult to assess the relationship between sentence context and lexical access even at an “immediate” homograph-target interval. In other words, the sentence-interrupt version may not be able to take full advantage of the temporal precision that is promised by the cross-modal procedure. As we have no taxonomy of the demands of the sentence tasks that may be used, we do not know the exent of this problem in existing research. Finally, meaning may be accessed prior to the termination of the auditorialy presented stimulus (cf. Marslen-Wilson, 1984). In this way, the functional interstimulus interval (ISI) may be underestimated with the cross-modal procedure, (however, see Tabossi, 1989). The time-locked nature of auditory presentation along with an emphasis on the time-course of meaning access requires that the interval between homograph onset and target onset be examined more critically.

56

G . Kellas, S.T. Paul, M . Martin and G.B. Simpson

Visual Unfolding

Tasks more clearly related to reading processs must utilize visually presented texts. However, stimulus presentation for such tasks makes it difficult to control the amount of processing that occcurs prior to target presentation. Since subjects generally are reading silently, it is impossible to determine when they begin to process specific words within the sentence (however, see Carpenter & Just, 1981). For example, assuming that a sentence ends in a homograph, and a target must be presented within 200 msec of its presentation, the researcher can only estimate when the subject will have read up to the last word. It is difficult and certainly less convenient to achieve control over visually presented texts in the same manner as is possible with auditorily presented stimuli. Essentially this is because the speeds at which listeners encode sentences for comprehension do not vary across individuals to the same degree as the speeds at which individuals read for comprehension. Auditorily presented sentences can be presented to all subjects at the same (normal speaking) rate. Visually presented stimuli, on the other hand, must be presented either at some average display speed or based on the individual’s reading speed. A problem with using an average display speed is that the researcher can never be certain of when, during sentence processing, a target is functionally being presented. Therefore, some means must be established by which the processing duration of visually presented material can be experimentally controlled. Attempts to resolve this problem include rapid serial visual presentation (RSVP) tasks which present sentences one word at a time (e.g., Till, Mross, & Kintsch, 1988). Each word is displayed, for some fixed duration, at a single focal point in front of the subject. Sentences presented in this manner may alter normal reading processes in ways essential to comprehension (e.g., peripheral processing). For example, text presented in this way would tend to place unusual emphasis on function words (Just & Carpenter, 1987, pp. 39-40). Alternately, sentences may be displayed all at once for a limited duration. The advantage of this mode is that it allows for examinations of gaze duration, eye movements, and eye fixations, using more normally displayed text. However, it is likely that presentation of a naming or lexical decision target during text processing will interfere with normal comprehension. Hence, even in this case a target must be presented as the sentence-final word or following the sentence, in order to avoid the same kind of dual-task problem associated with crossmodal priming. The RSVP procedure allows for precise control over stimulus presentation However, reading may be disruptcd by the unusual method of presentation. Ideally, visual text should be displayed in as normal a manner as possible (i.e., left-to-right). In a sense, then, what is needed resembles the sliding window technique described by Just and Carpenter (1987). or a sliding-RSVP, described by Simpson, Peterson, Casteel, and Burgess (1989) as the unfolding procedure

Contextual Feature Activation and Meaning Access

57

(see also West & Stanovich, 1986). The unfolding procedure is similar to RSVP because each word of the sentence can be displayed for a duration determined by each subject’s estimated reading speed, or to any processing speed of interest to the researcher. What results is a visual “unfolding” of the sentence across a computer screen. The first word appears, followed in time by the second word immediately to the right of the first, and so on to the end of the sentence. It is up to the experimenter whether each word of the sentence remains visible following its initial exposure duration. The unfolding procedure allows researchers to examine reading comprehension processes within certain limitations. Obviously, in order to present the material at a comparable rate for each subject, some initial time must be spent calibrating the presentation rate of the stimuli, In addition, interrupting a visually presented text (in any manner) with a demand to respond to some target will likely result in an interruption in normal reading processes as well. Therefore, in order to avoid such interruptions, the target stimuli must be presented prior to (e.g., as a memory load) or following the unfolded sentence. If one is concerned with the relationship of a target to a lexical item in a sentence, the researcher would have the greatest control over the timecourse associated with the sentence-final position and the interval between the homograph and subsequent target. Subjects who read faster than calibrated would not be able to see the sentence-final word any sooner than the presentation rate would allow. In effect, this procedure limits the functional stimulus onset asynchrony (SOA) between prime and target to no more than the subject’s calibrated reading speed (plus the ISI). In the case where subjects read more slowly than calibrated, the critical interval between prime and target, if experimentally short (e.g., zero msec ISI), would not be open to criticisms involving functionally extended processing. If support were obtained for selective access, it could not be argued that subjects were speed-reading to the sentence-final word, thereby functionally increasing processing time prior to presentation of the target. Alternately, support for initial access of all word meanings would leave little doubt as to the role of context during reading comprehension. While subjects may not be able to read ahead any faster than the presentation rate would allow, one might argue that in certain contexts the sentencefinal word is predictable (e.g., “The farmer went to the barn to milk the...”). Predictability may result either through pragmatic implication or spreading activation from certain words in the sentence. Such arguments specify functionally increased intervals between a nominal prime and the presented target. To avoid this difficulty, steps should be taken to ensure stimuli are not used that make such on-line predictions possible.

G . Kellas, S.T.Paul, M . Martin and G.B. Simpson Stimulus Norms The stimulus materials generally used to examine the effects of context on meaning access are those whose properties or associates in isolation are fairly well known. The experimental properties of the ambiguous words have been determined through some operational procedure (e.g., normative responses). However, as suggested by Olson (1970). the precise meaning of one sense of an ambiguity may change with variations in the features made salient by surrounding text. Therefore, in order to make use of these stimuli for examining the effects of context, one is forced to assume that meaning is invariant, regardless of the contextual environment. Most homograph norms currently available for research represent single responses to homographs in isolation (e.g., Nelson, McEvoy, Walling, & Wheeler, 1980; Perfetti, Lindsey, & Garson, 1971). Although such procedures are important for determining meaning frequency (i.e., meaning dominance), it remains unclear whether the information activated by the word in isolation accurately reflects that which is activated when the homograph appears in a sentence context. Schwanenflugel and Shoben’s (1985) research suggests that words semantically related to a meaning of an ambiguous word in isolation may not be included in the scope of facilitation provided by a sentence containing that ambiguous word. Responses generated to ambiguous words in isolation most likely represent the features of the referent as contextually determined in the history of its use. The normatively derived responses, then, reflect features of the ambiguity that will be made salient for some messages and not for others (Schoen, 1988). If a given experiment employs target words without regard to the prior features activated by context, evidence may be obtained for either multiple access, selective access, or some mixture of the two. The specific outcome would be determined by the overlap of contextually activated features with the features represented by subsequent targets. When the target represents a feature of an ambiguity which has not been made contextually salient, one would expect evidence for multiple access early in processing, even though meaning selection may eventually obtain from post-access semantic matching strategies. A solution to the above problems is to employ normative data from ambiguous words in constraining contexts. As suggested by Barclay et al. over a decade ago, norms derived from words in sentence contexts are required in order to generalize about the nature of extended discourse comprehension and processing. To describe more precisely the features activated by sentences, we have gathered normative data on 150 ambiguous words in context. In order to collect features related to both dominant and subordinate senses of the homographs, 300 sentences were constructed. Sentences were simple, active-declarative sentences with a range in length of three to seven words. Each sentence

Contextual Feature Activation and Meaning Access

59

ended in a homograph. Sentences were accepted or rejected for inclusion by a panel of five judges (majority vote). Three rules were followed during sentence construction: 1) Each sentence context must render the sentence-final homograph unambiguous. 2) The homograph itself must not be predictable from the context. That is, the sentence-final word need not be restricted to the ambiguous word in order for the sentence to make sense. For example, “He bought a ...“ may be followed reasonably by many different words, anything that may be purchased. However, when completed with CHEST, the resulting sentence does not make the body-part meaning of “chest” sensible. 3) No words may be included in the sentence context that would likely be generated as features of the sense of the ambiguous word constrained, or deemed as semantically related to the homograph. We would not have included, for instance, the following sentence: He plugged in the drill. It is reasonable that “plug” might be generated as a feature of the sense of drill intended. It was a concern that inclusion of words representing potential features in the sentence might influence responses such that the resulting features may not accurately reflect the meaning accessed through the context. Subjects might avoid recording the presented feature with their own, original, responses. The 300 sentences were randomly assigned to six sets of 50 sentences each. During data collection, each sense of an ambiguous word was responded to by 50 subjects. Each subject responded to an equal number of dominant (25) and subordinate (25) constrained sentences. No subject saw the same homograph in more than one context. Three hundred subjects were randomly assigned to sets, and sampled in groups of between 20 and 30. All subjects were instructed to generate as many fearures of the constrained ambiguity as possible within the time allotted for each stimulus, in the order they thought of them. Subjects were given one minute to generate responses to each stimulus and were not allowed to return to previous sentences, nor were they allowed to preview upcoming sentences. When the full minute had elapsed, the expenmen ter informed them to proceed. Responses were compiled and organized over a total interval of fivehundred hours. The frequency of specific responses to each word meaning (dominant, subordinate) was calculated and organized to reflect the production frequency of each response. It was assumed that the most frequently generated response to a single word sense would represent the most salient feature of the particular sense constrained by the sentence. The fact that a word is frequently generated in a given context suggests that the feature it represents is highly activated during sentence comprehension. Examples of the stimuli constructed, as well as high- and low-salient responses generated by subjects to the sentence contexts, can be found in Table 1. It may be helpful to emphasize, at this point, the major importance of these data. The stimuli now at our disposal represent the features of single senses of ambiguous words as activated within a specific sentence context. The

60

G . Kellas, S.T. Paul, M. Martin and G.B. Simpson

advantage of this becomes obvious when one considers that until recently very little of the research has controlled for the possibility that the target representing the sense of an ambiguous word was not activated sufficiently, or at all, by the context. Table 1

Sample Stimuli Sense of Sentence-Final Homograph

Target Salience

Dominant The structure was made of marble. They like to have company. It was time for them to park. I wish that it was spring. The girl's bag was made of straw.

High HARD FRIENDS CAR WARM HAY

Low GRAY HOME SHOP SCHOOL FLIMSY

Subordinate She had found her favorite marble. The woman bought the company. There were many things in the park. She sat on a big spring. He quickly finished it with a straw.

High ROUND BUSINESS GRASS BOUNCE PLASTIC

Low COLLECT RICH RANGER COMFORT STIR

A SENTENCE CONTEXT STUDY Employing the unfolding procedure discussed earlier, the data generated from the norms were used to examine how context might affect meaning access under an ideal set of conditions. The most obvious predictions to make derive from context-dependent and context-independent models of meaning access. For instance, assuming that we have constrained our sentence contexts to only a single meaning of the ambiguous word, one might reasonably conclude that evidence supporting selective access would obtain for all features appropriate to the sense of the homograph activated by the con~ext.~ Our norms represent those features made salient by the sentence context. As such, targets representing features primed by context should have been selectively activated during sentence comprehension. This outcome, however, is completely reliant on the fact that the sentence context constrains the meaning sense of the homograph immediately upon its presentation. Recall that the sentences were constructed to constrain the homograph to a single sense without making it predictable from the context. If all meanings are accessed initially, despite the constraint im-

Contextual Feature Activation and Meaning Access

61

posed upon a single interpretation of the homograph, one could reasonably predict facilitation for targets related to either the dominant or subordinate sense. The need to determine more clearly the factors that contribute to sentence constraint resulted in an additional manipulation and hypothesis. Although the sentences were constructed to constrain single senses of ambiguous words, they did not explicitly contain features of the intended meaning. It is tempting, then, to define them as low-constraining. We hypothesized that by explicitly embedding words representing high-salient features within the sentence, additional benefit might accrue as a result of lexical priming. Since access may be closely linked to contextual variations (cf. Seidenberg, et al., 1982; Simpson, 1984; Tabossi, 1988b). it was decided that inserting words representing high-salient features in the prime sentences might produce stronger context effects than sentences which merely constrain meaning. Sentences explicitly containing highly salient features might be more constraining (in terms of the sense of a homograph initially accessed) than sentences containing no such features. However, since the sentences were designed to constrain one sense of the homograph, it may not follow that they should be less constraining than a similar context which, in addition, contains a feature of the word meaning.4 If the normative context activated these features, then explicit presentation would render them redundant and no benefit should accrue relative to the absence of such features. The inclusion of feature-present sentence types allowed a comparison with feature-absent sentences to determine the extent that explicit high-salient features contribute to activation of appropriate homograph meanings. Although the time-course of contextual effects on meaning access is important to evaluate, we decided it was more critical to restrict our examination to early processing. The controversy as to whether meaning is selectively accessed or not has focused on the time-course of processing. There is little debate that, eventually, only contextually appropriate meanings remain accessed. Thus, most of the debate has focused on initial access. To examine whether initial access is selective or multiple, we concentrated on the earliest point during processing. That is, targets followed primes immediately (0-msec ISI). Contextual effects were evaluated for both dominant and subordinate meanings of sentence-final ambiguous words. Sentences constrained either the dominant meaning or the subordinate meaning of each homograph. In order that facilitation and inhibition might be estimated, a neutral prime was used. The neutral primes were constructed so as to provide a baseline measure of response time to targets. The neutral sentences ended in a homograph but contained no other words semantically related to the target nor did they constrain meaning (e.g., “The last word you will read is bank). Using this homograph-final control, we were able to examine meaning access in the biasing sentences inde-

62

G.Kellas, S.T. Paul, M . Martin and G.B. Simpson

pendently of lexical priming from the ambiguity itself. Finally, based on pilot research, it seemed likely that subjects could become aware of the neutrality of a sentence prime prior to its completion, thus affecting the subjects’ processing of these sentences. Therefore, 48 different neutral sentences were constructed to reduce the likelihood that subjects might recognize repeated exposures and generate some response strategy incompatible with responses to the other prime-types (i.e., dominant and subordinate biased sentences). Each neutral sentence was presented to subjects only once. Targets were selected to represent features related either to the dominant or subordinate sense of the sentence homograph. Biased sentences were followed by either dominant or subordinate related features. If access is indeed selective, facilitation should obtain only for the appropriate prime-target pairs, In order for multiple access to be indicated, facilitation must be obtained for both appropriate and inappropriate prime-target pairs. Targets were selected from the norms to represent either high salient (generated by approximately 70% of the subjects) or low salient (generated by only 4% of the subjects) features of the sense of the homograph constrained by the sentence context. This allowed for an examination of the scope of feature activation for each meaning sense of the homograph. Prior to the experimental trials, subjects’ reading speeds were estimated by presenting a set of 20 sentences, one word at a time, at an initial rate of 250 msec per word on a computer monitor. As soon as a word reached its exposure duration it was cleared from the screen. Immediately after, the next word in the sentence was displayed directly to the right of the previously displayed word. This “moving window” procedure ensured that subjects would attempt to follow the flow of presentation as closely as possible with their normal reading pace. After a sentence was displayed, subjects were asked to recall the sentence and describe whether the display rate was comfortable. The following sentence was then adjusted or not adjusted accordingly. The display rate could be slowed if subjects were unable to recall the sentence or if they indicated that it was currently too fast. Alternatively, the display rate was made faster if subjects indicated that they would prefer a faster display. After the 20 reading-speed calibration trials, subjects were given a brief rest before beginning the actual experiment. During this rest, subjects were instructed as to the nature and demands of the lexical decision task. They were told that they would be shown a number of sentences, similar to those presented during the calibration trials. Following each sentence would be either a real word or a pronounceable nonword to which they were to respond by pressing the appropriate key. To allow for more normal reading, each word of the sentence remained visible through the duration of the sentence-final word, at which point the entire sentence was removed from the screen and the target displayed. To avoid the possibility that subjects might develop the strategy of not reading the sentence prime, in order to focus on the upcoming target, subjects were

Contextual Feature Activation and Meaning Access

63

asked to recall the sentence on a random 20% of the trials.

Results Our preliminary results are described in Figure 1. The figure displays the facilitation (benefit) and inhibition (cost) obtained for context sentences compared with neutral homograph sentences for both target-saliency conditions. Difference scores were calculated to represent costs and benefits associated with sentence contexts. Cost reflects responses that took longer to initiate than the corresponding neutral condition, benefit represents responses that were initiated more quickly than the corresponding neutral condition. Including a highly salient feature in the priming sentence (feature-present) did not differ from leaving it out of the context (feature-absent) altogether. Initially, it had seemed quite plausible that meaning access (or some subsequent decision) would be enhanced by the presence of highly salient features in the context relative to a context which contained no features. However, Tabossi et al. (1987) have shown that the context as a whole accounts for feature priming, rather than simply associates of the target presented out of context. Thus, this aspect of our research corroborates their findings. High Salient

::-

35

30

25

20

10 15 I0

25 50

35

Figure 1. Mean difference scores for correct response times to high- and low-salient. dominant ( h m ) and subordinate (Sub)word targets. Sentence primes were constrained toward dominant and subordinate interpretations of the homograph. The appropriate condition is reflected by a congruent semantic relationship between the context and prime (h., dominant prime, domiqant target; rubordinate prime, subordinate target). The inappropriate condition reflects the opposite relationship (ix., dominant prime, subordinate subordinate target; subordinate prime, dominant target).

Interpreting the outcome for the high-salient condition is relatively straightforward. In line with a selective access interpretation. it was found that

64

G . Kellas, S.T. Paul, M.Martin and G.B. Simpson

responses to appropriately primed targets (e.g., dominant prime followed by a dominant target or a subordinate prime followed by a subordinate target) were facilitated, while inappropriately primed targets were inhibited, relative to the neutral condition. The results for dominant constraining sentences are similar to those reported by Tabossi et al. (1987). However, contrary to their results we have obtained evidence for selective access following subordinate constraining sentences as well, with no evidence of a meaning frequency involvement. On the basis of our data we must conclude that with sufficiently constrained context in which features of ambiguous words have been made salient, features associated with these referents will, in fact, be selectively activated. The source of the discrepancy between the present data and the Tabossi research is not immediately obvious. It may be due to task differences, representativeness of the ambiguous stimuli, procedures (cross-modal vs. visual), the sentence-final position of the homograph, or even reliability of the normative data. The pattern of results obtained for the low-salient condition presents an interesting extension of meaning frequency effects. There was no facilitation associated with appropriately primed dominant targets. However, for subordinate targets facilitation was obtained when appropriately primed. A comparison between saliency conditions indicates that when a context primes the dominant meaning of an ambiguous word, only the high-salient features appear to be activated, There seems to be little, if any, activation of low-salient features of the dominant sense. Responses associated with subordinate low- salient features, on the other hand, reflect a pattern of activation similar to that obtained in the high-salient condition. Appropriately primed responses were facilitated, while inappropriately primed responses were inhibited.

Discussion Overall, we argue that our data support a model of selective access when the context constrains a single sense of the ambiguous word. This outcome is especially clear when the targets are highly salient to the intended meaning of the homograph. When targets reflect low-salient properties of the constrained meaning, the outcome becomes more qualified. While sentence constraint is sufficient to eliminate meaning dominance effects for high salient targets, the overall pattern of results describe a more complicated view of meaning dorninance in terms of the scope of features activated by the context. Low-salient features appear to be selectively activated in the context of subordinate priming sentences. However, the data do not describe a pattern of selective activation for low salient features of dominant sentences. This may seem surprising if one expected that all features of a dominant meaning of a homograph should be highly salient, or more salient than low salient features of subordinate meanings. However, the fact that a word meaning is dominant reflects that the meaning is more frequently used. A word meaning that is used

Contextual Feature Activation and Meaning Access

65

relatively frequently and with consistent application may become much more defined in terms of its associated features. It may be the case that dominant meanings are employed in more consistent contexts than subordinate meanings. The consistency of application would be likely to lead to a rather homogeneous encoding of specific high-salient properties. As a result, low-salient properties of dominant senses may not become activated at all. Subordinate senses, on the other hand, may experience a greater variety of contextual use, thereby accruing a greater range of encodings. This acquired semantic flexibility for subordinate senses would ensure that a wider range of features are activated. What results is selective activation for highly salient features of a dominant meaning, and virtually no activation of the less salient features. Activation of a low salient feature will, as a function of its history of usage, be more likely to occur within subordinate meaning contexts than dominant meaning contexts. Consequently, selective activation under the low-salient condition occurs only for the subordinate sense.

SUMMARY AND CONCLUSIONS In sum, the present research, along with that of Barsalou (1982). Barclay et al. (1974), Tabossi (1988a. 1988b). Schwanenflugel and LaCount (1988), and Schwanenflugel and Shoben (1985), is compatible with a distributed model of meaning representation, Each of these investigations has demonstrated the differential importance of various properties of word meanings depending on context. As such, the meaning of a word can not be represented as a single conceptual node in a meaning network. Rather, meaning must be represented as a dynamic constellation of distributed features which reflect the meaning of the word in context. Some features may be more closely related to the lexical item than others as a result of frequency of use. Some features may be related to the lexical item only as a consequence of a given context. Certainly we would not presume to argue that our thinking is novel to this area of semantic research. The computationally intuitive appeal of a fixedfeature approach (e.g., Smith. Shoben, and Rips, 1974). indeed, constituted an important empirical step in the right direction. The extent to which lexical processing can be affected by context, in the manner we have applied here to reconcile the literature, suggests that arguments examining the loci @re- and post-lexical) of context effects may be premature. The confines of fixed-feature approaches to context effects on meaning access may have provoked the notion of encoding specificity as applied to text processing. Lines of reasoning similar to ours can be traced back to Gumenik (1979), who challenged the Anderson, Pichert, Goetz, Schallen, Stevens, and Trollip (1976) conclusion that general terms are encoded as specific instantiations. Gumenik preferred to interpret the Anderson et al. findings as evidence

66

G. Kellas, S.T. Paul, M. Martin and G.B. Simpson

for featural overlap. These ideas were extended by Greenspan (1986), who described core (highly associated to a word meaning) and peripheral (less frequently used aspects of word meaning) features. Barsalou (1982) also made a distinction between property types, describing features as either context-dependent or context-independent.The former are features of words that are activated only as a consequence of what has been specified in a preceding context, either directly or through inference. Contextindependent features are those activated by the presence of the stimulus itself, that is, they are activated regardless of context. We would suggest a qualification to this distinction, however. We would argue that even the most commonly associated features of a concept are sensitive to the context in which that concept appears. Although our present resr!!ts do not allow us to make the stronger claims that all features require the proper context for their activation, we do hold that the relative salience of any feature is sensitive to contextual constraints. That is, context structures the salience of all the features of the concepts contained therein. Other researchers, such as Schreuder, Flores d' Arcais, and Glazenborg (1985) have examined specific aspects of features. They made a distinction between perceptual and conceptual features of a word's meaning. Perceptual features, they concluded, are activated early during processing with conceptual features later dominating. Support for conceptual feature activation was obtained with a lexical decision task while the naming task resulted only in facilitation for perceptual features. It was concluded that the naming task tapped into an earlier stage of processing, when perceptual features are activated, than the lexical decision task. In conclusion, we would argue that the features activated by a context must be specified prior to an analysis of meaning access of individual words within a sentence (Olson, 1970). To ignore the effect of context on meaning representation, as our results indicate, may yield support for either selective or multiple access as a consequence of happenstance. A lack of understanding of the subtle, or sometimes dramatic, shifts in the meaning of a word (ambiguous or otherwise) due to changes in contextual usage will serve to misrepresent the processes underlying language comprehension. In order to understand the mechanisms involved in word comprehension in context, researchers are attempting to determine the nature of meaning access resulting from exposure to lexical items. While extended texts may be reduced to the study of individual words, it does not follow that studying meaning access of words in isolation will necessarily tell us much about normal reading comprehension. Nor does it follow that single-word contexts, preceding target words, are at all representative of normal reading conditions. And sentence or short paragraph primes, while certainly more representative of normal reading, merely resemble some reading encounters. It is unlikely that many people normally read lists of unrelated sentences or short, unrelated, paragraphs in a

Contextual Feature Activation and Meaning Access

67

single sitting. In all such attempts we must realize how contextually impoverished our language samples are. However, it seems reasonable to expect that those sentences (or contexts) which do not restrict an ambiguous word’s meaning will result in evidence favoring multiple access. There would not be sufficient context with which to constrain appropriate senses since “appropriateness,” in such cases, is arbitrary. Selective access would be expected only when the reader must commit to a single interpretation of the sentence. Therefore, we conclude that rather than endeavor to find support for either selective or multiple access, researchers should first focus on determining what actually is activated by sentence contexts. Acknowledgments This research was supported by General Research Fund Grant number 3213-XX-0038 from the University of Kansas (awarded to the first author). We would like to thank Cheryl Anagnopoulos, Kelly Hopkins, Mark Ludorf, Kelly Lyons, Cynthia Paul and Suzanne Wood for their selfless offer (which we quickly accepted) to help with the collection and organization of the normative data. Please address all correspondence to George Kellas, Department of Psychology, 426 Fraser Hall, University of Kansas, Lawrence, Kansas 66045. Electronic mail may be sent to QKELLASG@UKANVM. Notes ‘Consider the previous sample sentence, “The ship began to sink.” Within the context of the discussion in which this sentence first occurred and because of the emphasis placed on the word “sink,” you may likely have been aware of more than one meaning. However, note that “sink” is not the only ambiguous word in the sentence. 2This is essentially a modularity position (e.g., Fodor, 1983; Forster, 1979, 1981; Seidenberg, 1985) where lexical processing occurs in discrete stages. Information from any stage is made available to other stages only after having completed its process, regardless of whether processes are occurring in serial or parallel. This view may be contrasted with an interactive activation account (e.g.. McClelland, 1987) which does not restrict processing to discrete stages. However, the debate between modular and interactive views of language processing has been explored vigorously elsewhere (cf. Tanenhaus, Dell, & Carlson, 1987) and is beyond the scope of this chapter. 3Responsesto our normative stimuli back this assumption up. In less than 1% of the responses to the sentences used in the experiment were features generated that could be classified with an unintended reading of the sentence. This figure includes both the alternate sense of the homograph and any misreading of the sentence or context attributable to printing imperfections.

68

G . Kellas, S.T. Paul, M. Marlin and G.B. Simpson

‘Perhaps contexts which bias no sense of a homograph should be considered low constraining. Sentences such as. “He bought the straw.” or “They walked to the bank,” which render more than a single sense of the homograph sensible, would likely support a multiple access view.

References Anderson, R. C., Pichert, J. W.. Goetz, E. T., Schallert, D. J., Stevens, K. V., & Trollip, S. R. (1976). Instantiation of general terms. Journal of Verbal Learning and Verbal Behavior, 15,661-679. Ashcraft, M . H. (1976). Priming and property dominance effects in semantic memory. Memory & Cognition, 4,490-500. Barclay, J. R., Bransford, J. D., Franks, I. J., McCarrell, N. S., and Nitsch, K. (1974). Comprehension and semantic flexibility. Journal of Verbal Learning and Verbal Behavior, 13,47 1-481. Barsalou, L. W. (1982). Context-independent and context- dependent inforrnation in concepts. Memory & Cognirion, 10,82-93. Carpenter, P. A., & Just, M. A. (1981). Cognitive processes in reading: Models based on readers’ eye fixations. In A. M. Lesgold & C. A. Perfetti (Eds.), Interactive processes in reading. Hillsdale, NJ: Lawrence Erlbaum Associates. Fodor, J. A. (1983). Modularity of mind. Cambridge, MA: MIT Press. Forster, K. I. (1979). Levels of processing and the structure of the language processor. In W. E. Cooper, and E. C. T. Walker (Eds.), Sentence processing: Psycholinguistic srudies presented to Merrill Garrett. Cambridge, MA: MIT Press. Forster, K. I. (1981). Priming and the effects of sentence and lexical contexts on naming time: Evidence for autonomous lexical processing. Quarterly Journal of Experimental Psychology, 33A, 465-495. Greenspan, S . L. (1986). Semantic flexibility and referential specificity of concrete nouns. Journal of Memory and Language, 25, 539-557. Grice, H. P. (1975). Logic and conversation. In P. Cole, & J. L. Morgan (Eds.), Syntax and semantics, Vol. 3: Speech acts (pp. 41-58). New York: Seminar Press. Gumenik, W. E. (1979) The advantages of specific terms over general terms as cues for sentence recall: Instantiations or retrieval? Memory & Cognition, 7,240-244. Just, M. A., & Carpenter, P. A. (1987). The psychology of reading and language comprehension. Boston: Allyn & Bacon. Marslen-Wilson, W. (1984). Function and process in spoken word recognition: A tutorial review. In H. Bouma & D. G. Bouwhuis (Eds.), Aftention and performance X . Control of language processes. Hillsdale NJ: Lawrence Erlbaum Associates, 125- 150.

Contextual Feature Activation and Meaning Access

69

McClelland, J. L. (1987). The case for interactionism in language processing. In M. Coltheart (Ed.), Attention and performance X I I . Hillsdale, NJ: Lawrence Erlbaum Associates Inc. Nelson, D. L., McEvoy, C. L., Walling, J. R., & Wheeler, J. W. (1980). The University of South Florida homograph norms. Behavior Research Methods & Instrumentation, 12, 16-37. Olson, D. R. (1970). Language and thought: Aspects of a cognitive theory of semantics. Psychological Review, 77,257-273. Onifer, W., & Swinney, D. A. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory & Cognition, 9,225-236. Osgood, C. E. (1968). Toward a wedding of insufficiencies. In T. R. Dixon & D. L. Horton (Eds.), Verbal behavior and general behavior theory. Englewood Cliffs, NJ: Prentice Hall. Perfetti, C. A., Lindsey, R., & Garson, B. (1971). Association and uncertainty: Norms of association to ambiguous words. Pittsburgh, Pennsylvania: University of Pittsburgh, Learning Research and Development Center. Schoen, L. M. (1988). Semantic flexibility and core meaning. Journal of Psycholinguistic Research, 17, 113-123. Schreuder, R., Flores d’Arcais, G. B., & Glazenborg, G. (1985). Semantic decomposition and word recognition. In G. A. J. Hoppenbrouwers, P. A. M. Seuren, and C. Weijters (Eds.), Meaning and the lexicon. Dordrecht-Holland/Cinnaminson. USA: Foris Publications. Schwanenflugel, P. J., & LaCount, K. L. (1988). Semantic relatedness and the scope of facilitation for upcoming words in sentences. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 344-354. Schwanenflugel, P. J., & Shoben, E. J. (1985). The influence of sentence constraint on the scope of facilitation for upcoming words. Journal of Memory and Language, 24,232-252. Seidenberg, M. S. (1985). The time course of information activation and utilization in visual word recognition. In D. Besner, T. Waller, & G.E. MacKinnon (Eds.), Reading research: Advances in theory and practice. (Vol. 5 ) . N Y Academic Press. Seidenberg, M. S . , Tanenhaus, M. K., Leiman, J. M., & Bienkowski. M. (1982). Automatic access of the meanings of ambiguous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14, 489-537. Seidenberg, M. S., Waters, G. S., Sanders, M., & Langer, P. (1984). Pre- and postlexical loci of contextual effects on word recognition. Memory & Cognition, 12, 315-328. Simpson, G. B. (1981). Meaning dominance and semantic context in the processing of lexical ambiguity. Journal of Verbal Learning and Verbal Behavior, 20, 120-136.

70

G . Kellas, S.T.Paul, M. Martin and G.B. Simpson

Simpson, G. B. (1984). Lexical ambiguity and its role in models of word recognition. Psychological Bulletin, 96, 3 16-340. Simpson, G. B., & Burgess, C. (1985). Activation and selection processes in the recognition of ambiguous words. Journal of Experimental Psychology: Human Perception and Performance. 11,28-39. Simpson, G . B., & Kellas, G. (1989). Dynamic contextual processes and lexical access. In D. S. Gorfein (Ed.) Resolving semantic ambiguity. New York: Springer-Verlag. Simpson, G. B., Peterson, R. R., Casteel, M. A., & Burgess, C. (1989). Lexical and sentence contexts effects on word recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 88-97. Smith, E. E., Shoben, E. J., & Rips, L. J. (1974). Structure and process in semantic memory: A featural modcl for semantic decisions. Psychological Review, 81,214-241. Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)Consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18,645-659. Tabossi, P. (1988a). Effects of context on the immediate interpretation of unambiguous nouns. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 153-162. Tabossi, P. (1988b). Accessing lexical ambiguity in different types of sentential contexts. Journal of Memory and Language, 27,324-340. Tabossi, P. (1989, November). Processing ambiguous words in context. Paper presented at the 30th annual meeting of the Psychonomic Society, Atlanta, GA. Tabossi. P., Colombo, L., & Job, R. (1987). Accessing lexical ambiguity: Effects of context and dominance. Psychological Research, 49, 161-167. Tabossi, P., & Johnson-Laird, P. N. (1980). Linguistic context and the priming of semantic information. Quarterly Journal of Experimental Psychology, 32, 595-603. Tanenhaus, M. K., Dell, G. S.. & Carlson. G. (1987). Context effects in lexical processing: A connectionist approach to modularity. In J. Garfield (Ed.) Modularity in knowledge representation and natural language understanding. Cambridge: MIT Press. Tanenhaus, M. K., Leiman, J. L., & Seidenberg, M. S. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18,427-441. Tanenhaus, M. K., & Lucas, M. M. (1987). Context effects in lexical processing. Cognition, 25, 213-234. Till, R. E., Mross, E. F., & Kintsch, W. (1988). Time course of priming for associate and inference words in a discourse context. Memory & Cognition, 16, 283-298. West, R. F., & Stanovich, K. E. (1982). Source of inhibition in experiments on

Contextual Feature Activation and Meaning Access

71

the effect of sentence context on word recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 8,385-399. West, R. F., & Stanovich, K. E. (1986). Robust effects of syntactic structure on visual word processing. Memory & Cognirion, 14, 104-112. Whitney, P., McKay, T., Kellas, G., & Emerson, W. A. (1985). Semantic activation of noun concepts in context. Journal of Experimental Psychology: Learning Memory and Cognition, 11, 126-135.

This Page Intentionally Left Blank

Understanding Word and Sentence G.B. Simpsm (Editor) Q Elscvier Science Publishers B.V. (North-Holland), 1991

Chapter 4 A Perspective on Sentence Context Research

Padraig C. O’Seaghdha University of Illinois Champaign, Illinois U.S.A.

The essential feature of sentence context research, as considered here, is that it involves studying the response to a single word as a function of its relation to a preceding extended linguistic context. The sentence-context paradigm is formally very similar to a single-word priming procedure (e.g., Meyer & Schvaneveldt, 1971; Fischler, 1977a; Neely, 1977). In single-word priming, the effect of a context word (e.g., cow) is measured in terms of the response to a subsequently presented target word (e.g., milk or an unrelated control word). The response measure is usually lexical decision or naming latency. The cited studies, and many others, show that latencies are faster to related than to unrelated targets. As we shall see, a dominant theme in the sentence context literature concerns the role of such lexical-lexicalpriming effects in natural language understanding. This issue aside for the moment, sentence context was treated very much like single-word context, at least in early studies. One consequence was that sentence context was for the most part treated as undifferentiated information. This is seen clearly in early models of word recognition, particularly the logogen model (Morton, 1969), where all information, except perceptual information about the target, was lumped together in a “context system.” It is also reflected in the practice of using sentence completion norms, where context is operationally reduced to a simple measure of probability, to develop experimental materials. Just as single word primes were selected on the basis of normative associative strength (Meyer & Schvaneveldt, 1971). or rated semantic relatedness (Fischler, 1977b), sentence targets could be selected using the Clozc technique, a norming procedure in which subjects generate best completions of sentences (Bloom & Fischler, 1980; and see Fischler & Bloom, 1979, 1980; Morton, 1964; Schubcrth & Eimas, 1977; Stanovich & West, 1981). Thus, the focus of these early sentence context studies was on the response to the target words. The context, for the most part, was considered only indirectly, in terms of its effect on the target response measure. In what follows, I will call this the information processing approach to sentence context.

74

P.G. O’Seaghdha

The context may warrant consideration as a subject of study in its own right, however. From a psycholinguistic perspective, a sentence context is not undifferentiated information,but comprises several more or less articulate, more or less transient, structural representations. Interpreting a sentence entails parsing it, and this involves constructing a syntactic representation of the sentence, or at least of a local portion of it (O’Seaghdha, 1989; Wright & Garrett, 1984). The end-product of interpretation, which could be characterised in relatively formal (e.g., propositional) or informal (e.g., mental model) terms, also requires representation (see Kintsch & van Dijk, 1978; Foss, 1982; Johnson-Laird, 1983). This sets the stage for the major debate in the sentence context literature, which concerns the roles of lexical context, and of the non-lexical syntactic and meaning-level representations that have just been introduced. One view is that nonlexical contextual representations do not impinge directly on word recognition (e.g., Forster. 1976, 1979; Fodor, 1983; Seidenberg, Tanenhaus, Leiman, & Bienkowski. 1982). These authors account for context effects in terms of priming from lexical relatives of the targets in the contexts. For example, in the sentence She mailed the letter wifhout a sfamp, mailed and leffer are lexical relatives of stamp and may produce a priming effect independent of concurrent syntactic and semantic processes. Thus, priming is attributed to elements of the context, but in the very impoverished form of individual words acting independently. Other context effects are effectively lumped together in a postlexical category, and designated as not impacting directly on lexical access, though certainly influencing the later disposition of the accessed lexical information. Two criticisms can be levelled against this position, however. First, empirical evidence argues against the assumption that words embedded in sentences necessarilly activate the same associations as they do in isolation (e.g., Auble & Franks, 1983; Foss, 1982; Foss & Ross, 1983; Williams, 1988). Second, by lumping post-lexical effects together, these modular theorists, just like the information processors, ignore a large domain of research. Other researchers who are less concerned with the question of whether effects are strictly prelexical or post-lexical (e.g.. Foss, 1982; Johnson-Laird, 1983; Sharkey & Mitchell, 1985; Williams, 1988), have focussed on this neglected domain of text-level representation. From the text-oriented perspective, even if sentence context does not affect the earliest stages of word recognition, the process of assimilating new lexical elements to sentence level representations warrants detailed study. At the risk of oversimplification then, I have identified three factions engaged in sentence context research. First, there are the “information processors,” concerned to apply general information processing principles to this domain. Second, there are the “modularists,” concerned with defining the precise limits of lexical access, and particularly with the question of whether lexical access is insulated from the influence of higher level processes. And third, there are “text processors” who are more interested in the way in which text-level

Sentence Context Research

75

representations are established and maintained than with the exact nature of lexical access. Yet another distinguishable group of researchers should also be mentioned. This group can be labelled “sentence processors” and may be seen as the avatars of an earlier phase of linguistically-driven psycholinguistic research (see e.g., Clifton, 1989). This group is primarily concerned with aspects of syntactic processing, rather than with either word recognition or text-level representation (see e.g., Clifton, 1989; Frazier, 1987; Boland & Tanenhaus, this volume). However, it seems obvious that the structural representations studied under the rubric of sentence processing are vital to a full understanding of the nature of sentence context effects. As analyses of the nature of the contextual representation become more sophisticated, we should see an erosion of the artificial boundary between information processing studies of sentence context and linguistically guided studies of sentence processing. This is already evident in the interest of sentence processing researchers in the syntactic effects of sentence contexts, and in some recent investigations of the syntactic regulation of semantic priming (e.g.. Carroll & Slowiaczek, 1986; Duffy, Henderson, & Morris, 1989). In this chapter, I will establish a perspective on sentence-contextresearch primarily in terms of a contrast between a target-centered information-processing approach and a context-centered linguistically-based approach. In principle, and perhaps in the long historical perspective, all the factions I have identified may be seen as contributing to the same scientific effort, but in practice, I will argue, the sentence context literature occupies an uncomfortable position between the information processing approach which attempts to deal with context effects in terms of general cognitive processes, and the psycholinguistic approach that focusses on the nature of the representation of the context more than on the processes involved in responding to targets. Sentence context research was initially dominated by the information processing approach. This was partly because the development of new techniques (lexical decision, naming) made it natural for experimentalists to apply these more articulate procedures to the already defined problems of how perceptual and contextual information are combined in word recognition, and partly because the psycholinguistic approach was out of favor at the time that sentence context effects were most actively investigated. The failure to demonstrate the psychological reality of Chomskian representations in the previous decade led many psychologists to conclude that relating performance data to the constructs of linguistic theory was a futile exercise (see e.g., Clifton, 1989; Garnham, 1985, for discussion of this background). I will review the implications of the tension between information processing and psycholinguistic approaches, and argue that it explains some of the theoretical problems of the area. I will anchor the discussion to a review of a program of research which attempts to clarify the theoretical issues by consid-

76

P.G. O’Seaghdha

ering extremely simple cases of sentence-like context (O’Seaghdha, 1988, 1989, 1990). The research addresses three main issues in the sentence context literature: a) the nature of semantic context effects, and particularly the question of the role of inualexical activation in on-line comprehension; b) the nature and role of syntactic context effects; and c) the subsidiary issue of the relation between various on-line tasks (particularly naming and lexical decision) and the representations they index. Many other issues, including effects of strategy, contextual cnsuaint, and item characterisics, have been studied in the sentence context literature, but I will concentrate on the three topics mentioned as being most relevant to general processes of language comprehension and to the relation of the study of language to general cognitive psychology. I will assess the achievments of sentence context research after more than a decade of work, and evaluate its current status in light of the recent resurgence of more linguistically oriented sentence processing work.

HISTORY Research on sentence context effects as defined here began as long ago as the early 1960s (e.g., Morton, 1964; Tulving & Gold, 1963), but was most intensively pursued in the late 1970s and early 1980s (Fischler & Bloom, 1979, 1980; Forster, 1981; Kleiman, 1980; Schuberth & Eimas, 1977; Stanovich, 1981; Stanovich & West, 1979, 1981, 1983a. 1983b; West & Stanovich, 1982, 1986). Whereas the dependent measure in the early work of Morton, Tulving and Gold, and others, was tachistoscopic recognition of degraded target words, the bulk of the recent work with which we are concerned uses naming or lexical decision latency as the dependent measure. For example, Fischler and Bloom (1979) presented contexts such as She cleaned the dirt from her..., followed by related (shoes) or unrelated (terms) lexical decision targets. Relative to a control condition, they found facilitation of the related targets, and inhibition of the inconguous, unrelated targets. Stanovich and West (1979) obtained similar results in experiments in which the response task was word-naming, except that the facilitation effect was relatively large, and inhibition relatively small. Automatic and Attentional Effects Both Fischler and Bloom and Stanovich and West interpreted their findings in relation to the influential two-process attentional theory of Posner and Snyder (1979, following Neely’s (1977) application of this framework to single-word context priming. The two processes of the theory are automatic, associative activation and a slower-acting, non-automatic, attentional expectancy. According to the two-process theory, automatic activation is indicated by a pattern of bencfits without costs, whereas attentional expectancies yield benefits when they are correct, but incur costs when they are not. As described

Sentence Context Research

77

above, Fischler and Bloom observed facilitation of likely sentence completions, and inhibition of improbable, incongruent completions. In terms of two-process theory, the facilitation effect could be automatic andlor attentional, but the inhibition effect indicated the presence of an attentional process. However, Fischler and Bloom did not find inhibition of appropriate but normatively unlikely sentence completions (e.g., She cieaned the dirt from her hands). According to the two-process theory, inhibition should be observed in this condition if the subjects were expecting the likely completions of the sentences (in the example, shoes). Stanovich and West also used the two-process framework in their investigations. In their initial experiments with the naming task, Stanovich and West (1979) found facilitation of appropriate sentence completions when the contexttarget delay was short, and both facilitation of appropriate completions and inhibition of inappropriate completions when the delay was longer. Thus, in keeping with two-process theory, there appeared to be an early automatic process producing only facilitation, and a later, slower acting, attentional process that produced inhibition. More specifically, Stanovich and West postulated that the inhibition was a cost incurred when lexical expectancies were violated. However, in light of Fischler and Bloom’s evidence of benefits to probable sentence completions without cost to unlikely completions, West and Stanovich (1982) later revised the interpretation of inhibition. They suggested that inhibition simply reflects the detection of incongruity between a recognised target word and the sentence context. Although it severely limits the theoretical interest of inhibitory effects, there is widespread agreement that this revised interpretation is correct. In any case, following the early work I have described, most studies concentrated on the nature of the facilitatory effect for congruent words. However, the reason for this shift of focus had as much to do with the emergence of the new theoretical issue of lexical modularity as with the limitations of two-process theory.

Lexical Modularity Whereas the early work I have outlined was simply an exploration of the applicability of two-process theory in a new domain, sentence context research soon became a testing ground for modular conceptions of the language processing architecture. For modular theorists such as Forster (1976, 1979, 1981; see also Fodor, 1983) language comprehension is accomplished by hierarchically organized lexical, syntactic, and message-level processors, in such a way that the higher level processes do not impinge on the operation of the lower. In particular, the lexical level is insulated from syntactic and text or message-level processes. However, to accommodate the evidence of single-word context facilitation, associative processes were incorporated in the lexical module. Facilitation could therefore occur within the lexical module without regard to concur-

78

P.G. O’Seaghdha

rent syntactic and message-level operations. Thus, returning to the example I used earlier, facilitation of stamp in the sentence She mailed the letter without a stamp could be attributed to intralexical priming from mailed and letter. To account for facilitation in contexts such as She cleaned the dirt from her shoes, where cleaned and dirt are not strong associates of shoes, Forster (1981) proposed a response bias based on predictability. In effect, this means that whatever processes are responsible for the relatively high Cloze probability of shoes also make it relatively predictable in an on-line task. Although this is a plausible argument, evidence provided by Fischler and Bloom (1980). indicates that facilitation is not a function of predictability. Whatever the exact status of nonlexical facilitation effects, however, the issue of intralexical priming has become a dominant one in the sentence context literature (see Duffy et al. 1989; Forster, 1981; Seidenberg et al., 1982; Stanovich & West, 1983; Ratcliff, 1987; Zwitserlood. 1989). In spite of the influence of this idea, however, little or no direct evidence supports it. On the other hand, an increasing body of evidence suggests that lexical elements do not prime their associates autonomously when they are incorporated in natural language structures. Fischler and Bloom (1980) provided indirect evidence against inualexical priming in two of their findings. First, inhibition of unrelated targets without facilitation of related targets was observed when the contexts were presented serially at rapid rates, indicating that the contexts were interpreted but that automatic activation did not accrue from words in the contexts. Second, for the slower presentation rates where facilitation was observed, a post hoc comparison of contexts containing associates of the targets with those not containing such associates showed no effect of lexical association. Several other studies showing that contextual effects are better understood in terms of conceptual activation in a text-level representation (Auble & Franks, 1983; Foss, 1982; Foss & Ross, 1983) are reviewed in O’Seaghdha (1989). In addition, particularly clear evidence of a divergence between the activation produced by words presented in isolation and by the same words when they are incorporated in sentence or textual representations was provided by Williams (1988). Williams used a cross-modal priming procedure in which visual targets (e.g., nail) were presented at the offset of auditorily presented primes (e.g., hummer). He found the standard semantic facilitation effect when the primes were tested in isolation, but not if they were embedded in contexts as slight as In fact a hammer... Finally, direct assessment of the contribution of lexical-level priming in natural language comprehension is a main focus of the research summarised below (O’Seaghdha, 1989, 1990). Syntactic Priming Though most sentence context research has been concerned with semantic relations, syntactic context effects have been explored in a smaller set of stud-

Sentence Context Research

19

ies. Syntactic priming may be defined in terms of the relative response latencies to targets that respect and violate the syntactic category requirements of contexts. For example, a verb (enter) is more appropriate than a noun (units) in the context The essential beauty and order of the universe might... (see Wright & Garrett, 1984). In the Wright and Garrett study, lexical decision latencies were longer to targets that violated syntactic assignment. This effect, like the inhibition effect for semantically anomalous targets, has generally been interpreted as a postlexical one, triggered by the subjects’ noticing the incongruity. If so, one might expect the naming task to be less sensitive to syntactic anomaly than lexical decision. And indeed some early studies indicated that this might be the case. Initial demonstrations of syntactic priming effects in single word contexts (e.g., whose planet vs it planet) showed rather weak lexical decision effects (Goodman, McClelland, & Gibbs, 1981; see also Lukatela, Kostic, Feldman & Turvey, 1983). Seidenberg, Waters, Sanders and Langer (1984) replicated the Goodman et al. lexical decision effect but found no significant syntactic priming effect in a naming task. Therefore, they argued that syntactic priming effects should be grouped with a range of other postlexical effects which are manifest in lexical decision but not in naming. However, the syntactic priming effect observed by Wright and Garrett (1984) with fuller sentence contexts was much larger than the Goodman et al. effect. When West and Stanovich (1986) tested the hypothesis that syntactic priming does not affect the naming task, they found, somewhat to their surprise, that it did. This finding is corroborated by other evidence reported by Tanenhaus, Dell and Carlson (1987). Tanenhaus et al. presented semantically neutral contexts such as The next word i s ... followed by appropriately or inappropriately inflected words or nonwords. The most interesting result concerns the nonword conditions where the targets were inflected or uninflected nonword strings such as condetelcondetes. A robust syntactic priming effect was observed for these materials in the naming task. That is, subjects were slower to say condetes than condete. This confirms West and Stanovich’s evidence that the effect is not limited to lexical decision and, because it was obtained with nonword targets for which lexical access is precluded, also provides rather clear evidcnce that syntactic priming is not mediated by lexical access. Other evidence of a distinct status of syntactic processing (O’Seaghdha, 1990) is presented below. Naming and Lexical Decision Evidently then, the choice of response task in studies of sentence context has important consequences, and in recent years it has become a common precaution to compare results in the lexical decision and naming tasks. This is a substantive and not merely a methodological concern, because, at least in some circumstances, the tasks appear to indcx different representations. The naming

P.G.O’Seaghdha

80

task is thought to be primarily sensitive to current lexical-level activation (Stanovich & West, 1983). while the decision component of the lexical decision task makes it sensitive to other later or higher-level influences, and particularly to the subject’s evaluation of the fit between the target and the interpretation of the context.

575

Related Neutral

n

v)

E

0

525

Unrelated

W

>r

0

c

CI

a

475

4

425

1

Lexical Decision

Naming

Figure I. Relatedness effects in representative lexical decision and naming experiments.

In general, the following observations summarise the status of the tasks in the sentence context literature. Typically, lexical decision shows moderate facilitation of semantically appropriate targets and substantial inhibition of unrelated targets (Fischler & Bloom, 1979, 1980; and see Figure 1). Naming, on the other hand, shows a “facilitation-dominant pattern” with little inhibition (Stanovich & West, 1983). The difference between lexical decision and naming with regard to inhibition is explainable in terms of the lexical decision task’s tendency to reflect processes of postlexical evaluation. Since the inhibition effect is not the focus of much theoretical interest this difference could be seen as unimportant. However, two factors complicate the picture considerably. First, there is the evidence that both naming and lexical decision are sensitive to syntactic priming effects. This raises the interesting possibility, explored below, that the nature of the syntactic and semantic inhibition effects is different. The second complication concerns the evaluative aspect of lexical decision. If the lexical decision task is influenced by the outcome of an evaluation, this means that in particular circumstances a facilitation effect could be augmented, cancelled, or even reversed. In the sentcnce context literature, facilitation ef-

Sentence Context Research

81

fects are usually stronger with the naming task (Stanovich & West, 1983), so it seems that for this paradigm relatedness effects are not augmented in lexical decision. At least two recent studies, however, argue that there are conditions where lexical decision overrides a real facilitation effect. In the first of these studies, Balota and Lorch (1986) reexamined the issue of mediated priming in single-word contexts in light of the concerns about the lexical decision task. Mediated priming is priming between items that share an associate but are not themselves directly associated. Thus the words bull and milk are not directly associated, but both are associated to the potential mediator cow. In spreading activation terms, if bull is presented as a prime, activation should spread to cow and thence indirectly to milk. Previous research (e.g., de Groot, 1983) with the lexical decision task found no evidence of mediated priming. Balota and Lorch, however, though confirming the absence of an effect in lexical decision did observe mediated priming in the naming task. To explain this discrepancy, they proposed that lexical decision involves an evaluative phase which checks for a relation between the prime and target. If a relation is discovered, lexical decision is fast and therefore capitalises on any concurrent automatic lexical facilitation. However, if a relation is not discovered, decision is delayed until a response deadline expires. In this circumstance, no effect of automatic lexical facilitation would be evident. In the naming task on the other hand, there is no evaluative component, so the facilitation effect should be apparent. A similar analysis was recently proposed by Colombo and Williams (1990) to account for task differences in cross-modal sentence context effects. In this study, subjects heard sentences and responded to visual probes presented at the offset of a critical word. The probes were synonyms or antonyms of the critical words in the related conditions. For example, the context might be In order to be kept dry... and the target might be a synonym (arid), an antonym (wet), or an unrelated word. The outcomes were facilitation of synonyms but not antonyms in lexical decision, but facilitation of both kinds of relation in naming. In addition, and in keeping with the findings of Williams (1988) described above, both kinds of relation showed facilitation in lexical decision in a control study where the contexts were the synonyms and antonyms presented in isolation. To explain the absence of facilitation for embedded antonyms in the lexical decision task, Colombo and Williams argued that the contradictory nature of the antonym probes inhibited lexical decision, thus cancelling the lexical facilitation effect. Like Balota and Lorch, they suggested that evaluation is absent in naming. Both these studies, as well as a body of other work (e.g., Lupker, 1984; Seidenberg et al., 1984) appear to point to limitations of the lexical decision task. Given lexical decision’s susceptibility to bias, it is difficult to be confident in interpreting any given set of data. Therefore it has become almost customary to compare effects in the lexical decision and naming tasks in the recent litera-

82

P.G. O'Seaghdha

t u x In the next section, I will follow this convention in exploring the nature of semantic and syntactic priming in very simple sentence-like contexts.

A MINIMALIST APPROACH In my own research on context effects, I have adopted the approach of reducing the context to the greatest extent possible, that is, to the extent that it consists of only one content word (a noun) and some function words. The idea is that the nature of the context is poorly defined in much of the literature, and that by simplifying it to this rudimentary level specific hypotheses could be tested in a straightforward fashion. In addition, it appears that little is gained in the sentence context literature by using more complex contexts since, as I show below, the essential findings of the sentence-frame literature are captured by the reduced contexts. The contexts I used are in fact not sentences but sentence fragments or phrases. Table la shows the kinds of materials used to mimic typical sentence context conditions. Each phrase consisted of a pair of nouns embedded in a phrase-frame of the form function word - noun - function word function word - noun. The first noun could be related, neutral, or unrelated with respect to the second. The first four words of each phrase constituted the context and the final noun was the target. The targets were fully balanced over subjects and conditions. In all the experiments reported here, the context words, which appeared in lower case, were presented at a constant rate to a fixed central location on a CRT. The rate was 500 ms/word except where otherwise specified. The target was presented at the end of the sequence, and remained on the screen until the appropriate response was made. In some experiments the task was lexical decision, and in some it was target naming. The critical trials were the same in comparable conditions of the naming and lexical decision experiments, but nonword filler trials were included in the lexical decision experriments. Figure 1 shows representativc benchmark data for each of the tasks. The two main features of the sentence context literature are shown in these data. First, a substantial facilitation effect is observed for related targets in both lexical decision and naming.' Second, there is inhibition of unrelated targets in lexical decision, but not in naming. In the next sections, I will show how some of the main issues in the sentence context literature can be addressed using variations on these simplified contexts. The Intralexical Priming Issue The first issue to be addressed is the nature of the facilitatory effect in related contexts, and specifically the role of intralexical priming in accounting for facilitation. O'Seaghdha (1989. 1990) examined this question in two extended series of experiments. The basic manipulation employed in both series

83

Sentence Context Research

involved the grammatical coherence of the context-target sequences. O’Seaghdha (1989) employed a scrambling manipulation (see also Foss. 1982; Simpson, Peterson, Casteel, & Burgess, 1989; Masson, 1986) in a series of lexical decision experiments. The scrambling manipulation involved substituting inappropriate function words between the critical context word and the target (see Table 1b). The rationale of this manipulation was that, by comparing lexical decision latencies to related and unrelated targets in normal and scrambled conditions, the contribution of intralexical priming to the sentencelike context effect could be isolated. Note that, because of the serial mode of context presentation, subjects did not know until after the critical context word was presented whether the trial was a normal or scrambled one. Therefore, in contrast to other studies that have employed a scrambling manipulation, they could not make any special adjustments at the beginning of a trial. The prediction was that if intralexical priming is responsible for the facilitation effect in normal contexts, it should also show in the scrambled conditions.

Related

6oo

Unrelated

. Normal

Scrambled

Figure 2. Effect of relatedness with normal and scrambled contexts in the lexical decision task.

The general outcome obtained in a series of experiments is shown representatively in Figure 2.2 There was a substantial relatedness effect in normal contexts, but only a marginal, nonsignificant effect in the scrambled contexts. Thus, although an intralexical process cannot be entirely ruled out, it is clearly not strong enough to account for substantial variance in on-line comprehension. The role of intralexical activation is therefore at best moot. Note that this outcome also discounts h e importance of unmediated conceptual priming in

P.G. O'Seaghdha

84

natural language comprehesion. That is, if a relatedness effect had emerged in the scrambled condition, it might be attributed to intralexical priming (Forster, 1981) or to conceptual-lexicalpriming (see e.g.. Collins & Loftus. 1975). However, since the relatedness effect was substantial only in the natural contexts, it appears that neither lexical-lexical nor conceptual-lexical priming plays a substantial role independent of the computation of contextual representations.

Table 1 Examples of Materials Used in (he Experiments a. Grammatical nount target conditions (see all figures) Related Neutral Unrelated

The message of that letter The color of that letter The nose and the letter

b. Scrambled noun target conditions (see Figure 2) Related Unrelated

The message this near letter The nose that of letter

c. Ungrammatical verb target conditions (see Figures 3 and 4) Related Neutral Unrelated

The message of that send The color of that send The nose and the send

d. Grammatical verb target conditions (see Figure 4) Related Neutral Unrelated

The message that was sent The color that was sent The nose that he sent

e. Ungrammatical noun target conditions (see Figure 4) Related Neutral Unrelated

The message that was letter The color that was letter The nose that he letter

The second series of experiments employed an even stronger manipulation to address the intralexical issue. One explanation of the relatedness effecl

85

Sentence Context Research

observed in the normal condition is what I call the lexical transmission hypothesis, which states that lexical activation is maintained in coherent sequences. Duffy et al. (1989) recently proposed a version of this hypothesis, and argued that it is compatible with lexical modularity. However, this claim appears to be invalid since a process that is “a byproduct of the higher level syntactic and integrative processes” (Duffy et al., p. 797) is by definition not autonomous. Nevertheless, because the processes of reading scrambled text are not well understood, the lexical transmission hypothesis should be evaluated. The second series of experiments therefore employed a manipulation of the syntactic category of the target instead of scrambling the context. For example, instead of contrasting The message of that letter with The message this near letter, the critical conditions were The message of that letter and The message of that send (see Table 1). where send is a verb semantically related to message. The related targets were therefore either grammatically appropriate nouns or inappropriate verbs. In effect the semantic relatedness and syntactic priming paradigms were conflated, so that both contextual relatedness and the syntactic status of the sequences were defined by the targets. Thus, in contrast to the scrambling manipulation, subjects had no indication of the experimental condition before the target was presented. This meets the concern of Duffy et al. that when contexts are scrambled “the individual lexical items are unlikely

1

Related

475 I-

1

Noun Verb Target Category Figure 3. Effect of relatedness for appropriate noun and inappropriate verb targets in the lexical decision task.

86

P.G.O’Seaghdha

to be available as primes by the time the target word is reached” (p. 797). Figure 3 shows that this stronger test of intralexical priming yielded exactly the same results as before as far as lexical relatedness is concerned. There was a strong relatedness effect for noun targets, comprising both facilitation of related targets and inhibition of unrelated targets. However, there was virtually no effect for the verb targets.’ Thus, even when the target item itself determines whether the sequence is syntactic, the relatedness effect depends on syntactic coherence. This provides very strong evidence that lexical priming per se can play little role in on-line comprehension. Taking Naming to Task Before concluding that relatedness effects depend on syntactic coherence and that they are non-lexical, it may be prudent to confirm these results with the naming task. We have already seen that the expected “facilitation dominant” pattern in the naming task is obtained with syntactically coherent phrasal materials (see Figure 1). The presence of facilitation for related targets together with the lack of inhibition for incongruous, unrelated targets has been taken to indicate that the naming task is sensitive to automatic activation but relatively insensitive to slower acting attentional or sentence integration processes (Stanovich & West, 1983). However, it has also been argued that facilitation effects are sometimes obscured in the lexical decision task (Balota & Lorch, 1986; Colombo & Williams, 1990). If so, following our previous discussion of naming and lexical decision, a relatedness effect could be observed for both grammatically appropriate and grammatically inappropriate targets in the naming task. This might occur if the detection of syntactic incongruity caused the kind of delay in lexical decision responding suggested by Balota and Lorch, masking an otherwise observable relatedness effect. A s discussed above, Colombo and Williams (1990) also observed a facilitation effect in naming but not in lexical decision. However, in this case the null result appeared to be a result of cancelling lexical-conceptual and sentence evaluation effects. This kind of contradiction does not arise in the present case because relatedness and syntactic appropriateness are fully crossed. An alternative and much simpler outcome might also occur: The same outcomes could be observed in naming and in lexical decision except that, as before, inhibition of incongruent grammatical targets should not occur in naming, In this case, the interpretation would be that the relatedness effect in both naming and lexical decision is a nonlexical, text-level effect. In the event, neither the task equivalence nor the particular task difference prediction outlined above was supported. Figure 4a shows the resulls of a large experiment in which both contexts favoring noun targets (see Table 1, a and c) and contexts favoring verb targets (Table 1, d and e) were presented, followed by grammatically appropriate or inappropriate targets. The striking result is that

87

Sentence Context Research

the relatedness effect in all conditions was virtually eliminated in this experiment. There was neither a main effect of relatedness as the lexical priming hypothesis would predict, nor the interaction of grammaticality and relatedness that was found in the lexical decision experiment (Figure 3). Most surprisingly, there was virtually no indication of a relatedness effect in the noun-context, noun-target conditions which yielded robust facilitation in the naming task experiment shown in Figure 1. Because it shows that the effect is not even obligatory under standard conditions, this result provides new strong evidence that automatic intralexical priming does not play an important role in on-line comprehension. The outcome does not depend on the particular combination of conditions tested in this experiment since it was replicated in a separate experiment in which only the noun context conditions (those shown in Figure 3 for lexical decision) were tested (see O’Seaghdha, 1990, for details). Rather, it appears that the inclusion of syntactically inappropriate targets influenced the mode of responding in the naming task in an unanticipated way. In other words, to understand the lack of a semantic relatedness effect in Figure 4a, we need to consider the effect of syntactic priming in the naming task.

550 n

1

Related

I

ea Verb Context

Noun Context

0

I

Neutral Unrelated

I

500 W

>r

0 C

Q

450

A

400

Noun

Noun Verb Target Category

Verb

Figure 4a. Effect of relatedness in the naming task for appropriate and inappropriate targets in noun- and verb-disposing contexts.

Is Syntactic Assignment Obligatory? Although no reliable semantic priming was observed in Figure 4a, another interesting effect was observed in that experiment. By presenting all combina-

P.G. O'Seaghdha

88

tions of context (noun-disposing or verb-disposing), relatedness (related, neutral, unrelated) and target (noun or verb) to the same subjects, it was possible to derive precise estimates of syntactic as well semantic priming. A syntactic priming effect is simply the difference in naming latencies to the same words in syntactically appropriate versus syntactically inappropritate contexts. Thus, the syntactic priming effect for nouns is the difference between latencies to nouns in noun contexts and latencies to nouns in verb contexts, disregarding relatedness. Figure 4b shows that the effect of syntactic appropriateness was substantial, especially for nouns. This is a very interesting result for several reasons. First, in agreement with West and Stanovich (1986) and Tanenhaus et al. (1987), it confirms that syntactic priming effects are obtained in the naming task. Indeed, the fact that the effect is obtained in the absence of semantic priming suggests that syntactic assignment may be obligatory in a sense that the effect of semantic relatedness is not. Carello, Lukatela, and Turvey (1988) argued that previous reports of syntactic priming in English (West & Stanovich, 1986; Wright & Garrett, 1984) confounded syntactic and message-level effects, but this demonstration of syntactic inhibition independent of semantic relatedness clearly addresses that concern.

.

A

550

I

Appropriate Inappropriate

r1

Verb Noun Target Category F i g w e 4 b . The effect of syntactic appropriateness for noun and verb targets.

The existence of a syntactic priming effect is also important because it shows that the absence of a semantic relatedness effect cannot be attributed to a

Sentence Context Research

89

failure to parse the sequences. If subjects responded to the intermixture of syntactic and nonsyntactic sequences within the experiment by largely ignoring the contexts, neither syntactic nor semantic priming should be observed. Third, the presence of syntactic inhibition in the absence of semantic inhibition means that subjects were sensitive to syntactic category violations but not to violations of sense. Syntactic and semantic violations therefore are not registered by a general anomaly detector. Rather, they appear to be different in kind. Corroborative evidence of this difference may be found in the evoked potential work of Kutas and Hillyard (1983) who showed that semantic anomaly engendered one pattern of potentials (N400)whereas grammatical violations created a quite distinct pattern. The distinction observed here in the naming task offers the prospect of being able to study such qualitative differences by the judicious use of simple latency measures. The most surprising result, and the most difficult to reconcile with previous evidence, is of course the absence of facilitation in the congruent, syntactic conditions in Figure 4. The facilitation effect has been observed in many studies, and has usually been interpreted as the product of an automatic, associative process (e.g., Stanovich & West, 1983). Yet it was absent in an experiment where a syntactic inhibition effect that is assumed to occur much later in processing was observed. This apparent paradox is discussed in detail in O’Seaghdha (1990), but the essentials of the account that seems to be required are as follows. First, the evidence points to the conclusion that lexical priming, in the intralexical sense, plays little or no role in the sentence context paradigm and in the experiments described here. Rather, target processing is determined by the nature of computed contextual representations and by the relation of the target to these representations. In addition, however, the nature of the response task must be taken into account. In lexical decision, all sources of information feed into a binary choice. When a syntactically congruous related word target is presented, the coherence of the context and target constitutes an additional source of evidence that the target is a word, leading to a facilitation effect relative to the neutral condition. Inhibition in the unrelated condition is explained in the standard way as a side-effect of detecting the incongruity. Finally, syntactic anomaly disrupts responding and prevents integration of the target with the context, thus precluding semantic relatedness effects. Note that syntactic inhibition is observed in the lexical decision task (O’Seaghdha, 1990; West & Stanovich, 1986; Wright & Garrett, 1984). but the inclusion of syntactic anomalies in the experiment does not influence the semantic relatedness effect in the syntactically appropriate conditions. That is, the same semantic facilitation and inhibition effects are evident in the lexical decision data of Figure 1 and in the equivalent conditions of Figure 3. The naming task differs from lexical decision in that information does not feed into a decision process. Rather, naming involves the computation of a

90

P.G.O'Seaghdha

pronunciation followed by utterance of the word. In the data reported here, naming responses were sensitive to syntactic anomaly (Figure 4b) but, unlike lexical decision responses, not to semantic incongruity (Figures 1 and 4a). In addition, in the experiment where syntactic violations occurred in some conditions, semantic facilitation was virtually eliminated in all conditions. These outcomes suggest that syntactic anomalies a) have a separate status than semantic violations, b) are obligatorily registered, and c) impact on naming in a different way than on lexical decision. One explanation of this pattern of results is that the syntactic priming effect in naming has to do with the productive aspect of the task: Speakers are apparently reluctant to utter words that violate syntactic category assignment. Thus the syntactic inhibition effects observed in lexical decision and naming may be different in kind: In lexical decision, syntactic inhibition is simply an effect of noticing an anomaly, but in naming it may be a genuine hesitancy in speaking. Finally, how do we account for the absence of a semantic relatedness effect in coherent sequences when ungrammatical conditions are included in a naming experiment? In the standard sentence context experiment in which there are no grammatical violations, rapid integration of the target with the context in a related condition, by allowing naming to take place sooner, could explain the facilitation effect. However, when syntactic violations occur in the experiment, the mode of responding in the naming task is altered. Instead of assuming that targets will be grammatically acceptable, subjects check the syntactic status of the targets before responding. A checking or monitoring process by itself, however, would merely add a constant to response times, leaving the relatedness effect intact. Therefore, to account for the virtual elimination of facilitation, it is necessary to assume that the checking process entails a delay in sentence integration. Finally, I assume, on common sense grounds, that sentence integration is completed, probably concurrently with naming the target, in the case of syntactically congruent phrases. Therefore, the absence of a relatedness effect in Figure 4a need not mean that the phrases are not fully interpreted in these conditions.

CONCLUSIONS: INFORMATION PROCESSJNG AND SENTENCE PROCESSING Sentence context research is not so much a theoretically motivated research topic as a nexus where the interests of rather disparate groups of researchers temporarily converged. At the beginning of this chapter I identified four such groups, whom I called information processors, modularists, text processors, and sentence processors. For the most part I have concentrated on the concerns of the first two groups, and especially on the intralexical priming hypothesis, but my findings point to the importance of the syntactic and message-level representations studied by the other two groups. For reasons I outlined earlier, sentence context research was initially

Sentence Context Research

91

dominated by the information processing approach. However, in the work of Fischler and Bloom (1979, 1980) and Stanovich and West (1979, 1981, 1983). analyses of context effects in terms of the dichotomy of automatic and attentional processes did not fare very well. I have suggested elsewhere that the reason the Posner-Snyder framework is inapplicable to the sentence context problem is that the effects are determined by the nature of constructed representations (O’Seaghdha, 1989, p. 84) rather than by elementary processes of activation and expectancy. Thus, the failure of the Posner-Snyder framework to account for sentence context effects may be contrasted with Neely’s (1977) quite successful application in the case of categorical information where the contexts were simple category labels. Although Forster’s (1976, 1979, 1981) linguistic modularity theory brought a concern with the overall architecture of the language comprehension apparatus to the field, in practice the influence of modular theorising was to narrow the scope of inquiry rather than to broaden it. Language processing was conveniently partitioned into autonomous subcomponents, and psychologists were provided with a subdomain, the lexicon, which they could explore without regard to the complications of syntactic and message-level representations. Though the modularity debate has had a large impact on research into language processing in the last decade, the particular idea that had most impact on sentence context research, intralexical priming, is not well supported either in the literature or by the results presented above. Though it may be impossible to discount the notion of intralexical priming completely, it now seems clear that such a process does not have a significant impact in on-line comprehension. Rather, the operative level is the computed representation of context (see Foss, 1982; Williams, 1988). Dismissal of the intralexical hypothesis, however, need not entail a return to a strongly interactionist conception of language processing. The data presented above are quite compatible with an emphasis on bottom-up processing in early word recognition, though they do not rule out the possibility of priming from higher level representations in some circumstances. Such priming, however, may occur in a selection rather than in an initial activation stage of word recognition (see Zwitserlood, 1989, for a recent discussion of this issue). In addition, the syntactic effect depicted in Figure 4b can be interpreted as indicative of an autonomous or isolable (Bock & Kroch, 1989) component of syntactic processing. In recent years, thcre has been increasing concern with the relation of lexical and syntactic processing (Carello et al., 1988; West & Stanovich, 1986; Wright & Garrett, 1984). and with the nature and representational status of contextual representations (Foss, 1982; Sharkey & Mitchell, 1985; Williams, 1988). In addition, the research I have summarised above, togeiner with some other studies (e.g., Carroll & Slowiaczek, 1986) has begun to examine syntactic regulation of semantic relatcdness, thus bringing the sentence context literature

92

P.G. O‘Seaghdha

into closer contact with the structural concerns of the sentence processors. It may be that sentence context research will gradually be subsumed by a revitalised linguistically oriented psycholinguistics. Unfortunately, but perhaps inevitably, the strengthening of psycholinguistics as a subdiscipine may be accompanied by an increasing separation from the concerns of general cognitive psychology. In any case, the lasting contribution of the phase of sentence context research I have described may be that by sharpening the debate on a rather narrow set of issues, it provided methodological and analytical sophistication that can continue to be used in future more linguistically oriented work. Acknowledgment

This research was supported by NIH Grant NS25502 and by NSF Grant BNS-8910546. Notes

‘The lexical decision data are from O’Seaghdha (1989) Experiment 5 , and are averaged over context presentation rates of 200,400, and 800 ms/word. The naming data are from O’Seaghdha (1990). *These data are averaged over Experiments 1 and 2 of O’Seaghdha (1989) which showed equivalent effects in these conditions. The presentation rate was 400 ms/word in Experiment 1 and 800 ms/word in Experiment 2. ’Note that this conclusion also holds in a symmetrical experiment reported in O’Seaghdha (1990) where the context-target relations were reversed. In this experiment, the contexts required verb targets (see Table Id and l e for examples). The appropriate verb targets were only weakly facilitated, suggesting that they may be less strongly primed by the contexts than nouns, but the inappropriate noun targets also showed only a marginal relatedness effect. That is, whether one compares the relatedness effect for nouns in noun contexts to that of verbs in noun contexts. or to that of nouns in verb contexts, the conclusion is the same: The relatedness effect is strong in the normal syntactic conditions and at best marginal in the ungrammatical conditions. (See also discussion of Figure 4). References

Auble, P., & Franks, J. J. (1983). Sentence comprehension processes. Journal of Verbal Learning and Verbal Behavior, 2 2 , 395-405. Balota, D. A. & Lorch, R. F. (1986). Depth of automatic spreading activation: Mediated priming effects in pronunciation but not in lexical decision. Journal of Experimental Psychology: Learning, Memory & Cognition, 1 2 , 336-345.

Sentence Context Research

93

Bloom, P. A., & Fischler, I. (1980). Completion norms for 329 sentence contexts. Memory & Cognition, 8, 631-642. Bock, J. K. & Kroch, A. S. (1989). The isolability of syntactic processing. In G. N. Carlson & M. K. Tanenhaus (Eds.), Linguistic structure in language processing (pp. 157-196). Kluwer. Boland, J. E. & Tanenhaus, M. K. The role of lexical representations in sentence processing. This volume. Carello, C., Lukatela, G., & Turvey, M. T. (1988). Rapid naming is affected by association but not by syntax. Memory & Cognition, 16, 187-195. Carroll, P. & Slowiaczek, M. L. (1986). Constraints on semantic priming in reading: A fixation time analysis. Memory & Cognition, 14, 509-522. Clifton, C. Jr.. (1989). Syntactic modularity in sentence comprehension. In R. R. Hoffman & D. S. Palermo (Eds.), Cognitive psychology: The state of the art. Hillsdale, NJ: Erlbaum. Collins, A. M., & Loftus, E. F. (1975). A spreading activation theory of semantic processing. Psychological Review, 82,407-428. Colombo, L., & Williams, J. (1990). Effects of word- and sentence-level contexts upon word recognition. Memory & Cognition, 19, 153-163. Duffy, S. A., Henderson, J. M. & Morris, R. K. (1989). Semantic facilitation of lexical access during sentence processing. Journal of ExperimentalPsychology: Learning, Memory, & Cognition, 15, 791-801. Fischler, I. (1977a). Associative facilitation without expectancy in a lexical decision task. Journal of Experimental Psychology: Human Perception and Performance, 3, 18-26. Fischler, I. (1977b). Semantic facilitation without association in a lexical decision task. Memory & Cognition, 5 , 335-339. Fischler, I., & Bloom, P. A. (1979). Automatic and attentional processes in the effects of sentence contexts on word recognition. Journal of Verbal Learning and Verbal Behavior, 18, 1-20. Fischler, I., & Bloom, P. A. (1980). Rapid processing of the meaning of sentences. Memory & Cognition, 8, 216-225. Fodor, J. A. (1983). The modularity of mind. Cambridge, MA: MIT. Forster, K. I. (1976). Accessing the mental lexicon. In R. J. Wales & E. Walker (Eds.), New approaches to language mechanisms (pp. 257-287). Amsterdam: North-Holland. Forster, K. I. (1979). Levels of processing and the structure of the language processor. In W. E. Cooper & E. C. T. Walker (Eds.), Sentence processing: Psycholinguistic studies presented to Merrill Garrett (pp. 27-85). Hillsdale, New Jersey: Erlbaum. Forster, K. I. (1981). Priming and the effects of sentence and lexical contexts on naming time: Evidence for autonomous lexical processing. Quarterly Journal of Experimenlal Psychology, 33A, 465-495. Foss, D. J. (1982). A discourse on priming. Cognitive Psychology, 14, 590-607.

94

P.G. O‘Seaghdha

Foss, D. J., & Ross, J. R. (1983). Great expectations: Context effects in sentence processing. In G. B. Flores d’kcais & R. J. Jarvella (Eds.), The process of language understanding (pp. 169-191). New York: Wiley. Frazier, L. (1987). Sentence processing: A tutorial review. In M. Coltheart (Ed.), Attention and Performance X U : The psychology of reading (pp. 559-586). Hillsdale, NJ: Erlbaum. Gamham, A. (1985). Psycholinguistics: Central topics. London: Methuen. Goodman. G.0.. McClelland, J. L. & Gibbs. R. W.(1981). The role of syntactic context in word recognition. Memory & Cognition, 9, 580-586. Johnson-Laird, P. N. (1983). Mental models. Cambridge, MA: b a r d University Press. Kintsch, W. & van Dijk, T.A. (1978). Towards a model of text comprehension and reproduction. Psychological Review, 8 5 , 363-394. Kleiman, G. M. (1980). Sentence frame contexts and lexical decisions: Sentence acceptability and word-relatedness effects. Memory & Cognition, 8 , 336-344. Kutas, M. & Hillyard, S . A. (1983). Event-related brain potentials to grammatical errors and semantic anomalies. Memory & Cognition, 1 1 , 539-550. Lukatela, G.,Kostic, A., Feldman.. L.B., & Turvey, M.T. (1983). Grammatical priming of inflected nouns. Memory 8i Cognition, 1 1 , 59-63. Lupker, S . J. (1984). Semantic priming without association: A second look. Journal of Verbal Learning and Verbal Behavior, 23, 709-733. Masson, M. E. J. (1986). Comprehension of rapidly presented sentences: The mind is quicker than the eye. Journal of Memory and Language, 25, 588604. Meyer. D. E.. & Schvaneveldt, R. W. (1971). Facilitation in recognizing pairs of words: Evidence of a dependence between retrieval operations. Journal of Experimental Psychology, 90,227-234. Morton, J. (1969). The interaction of information in word recognition. Psychological Review, 76, 165-178. Morton, J. (1964). The effect of context on the visual duration threshold for words. British Journal of Psychology, 55, 165-180. Neely, J. H. (1977). Semantic priming and retrieval from lexical memory: Roles of inhibitionless spreading activation and limited capacity attention. Journal of Experimental Psychology: Gencral, 106, 226-254. O’Seaghdha, P. G. (1988). Conjoint syntactic and semantic context effects: Tasks and representations. Proceedings of the 10th Annual Conference of the Cognitive Science Society (pp. 673-679). Hillsdale. NJ: Erlbaum. O’Seaghdha, P. G. (1989). The dependence of lexical relatedness effects on syntactic connectedness. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 73-87. O’Seaghdha, P. G. (1990). Conjoint and dissociable effects of syntactic and semantic context. Manuscript submitted for publication.

Sentence Context Research

95

Posner, M. I., & Snyder, C. R. R. (1975). Facilitation and inhibition in the processing of signals. In P. M. A. Rabbitt & S. Dornic (Eds.), Attention and performance V . New York: Academic Press. Ratcliff, J. E. (1987). The plausibility effect: Lexical priming or sentential processing. Memory & Cognition, 15,482-496. Schuberth, R. E., & Eimas, P. D. (1977). Effects of context on the classification of words and nonwords. Journal of Experimental Psychology: Human Perception and Performance, 3,27-36. Seidenberg, M. S., Tanenhaus, M. K., Leiman, J. M., & Bienkowski, M. (1982). Automatic access of the meanings of ambiguous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14, 489-537. Seidenberg, M. S., Waters, G. S., Sanders, M., & Langer, P. (1984). Pre- and post-perceptual loci of contextual effects on word recognition. Memory & Cognition, 12, 315-328. Sharkey, N. E., & Mitchell, D. C. (1985). Word recognition in a functional context: The use of scripts in reading. Journal of Memory and Language, 24, 253-270. Simpson, G. B., Peterson, R. R., Casteel, M. A., & Burgess, C. (1989). Lexical and sentence context effects in word recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 88-97. Stanovich, K. E.. & West, R. F. (1979). Mechanisms of sentence context effects in reading: Automatic activation and conscious attention. Memory & Cognition, 7, 77-85. Stanovich, K. E., & West, R. F. (1981). The effect of sentence context on ongoing word recognition: Tests of a two-process theory. Journal of Experimental Psychology: Human Perception and Performance, 7,658-672. Stanovich, K. E., & West, R. F. (1983). On priming by a sentence context. Journal of Experimental Psychology: General, 112, 1-36. Tanenhaus, M. K., Dell, G. S., & Carlson, G. (1987). Context effects and lexical processing: A connectionist approach to modularity. In J. L. Garfield (Ed.), Modularity in knowledge representation and natural language understanding (pp. 83-108). Cambridge, MA: MIT. Tulving, E. & Gold, C. (1963). Stimulus information and contextual information as determinants of tachistoscopic recognition of words. Journal of Experimental Psychology, 66, 3 19-327. West, R. F., & Stanovich, K. E. (1982). Source of inhibition in experiments on the effect of sentence context on word recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 8, 385-399. West, R. F., & Stanovich, K. E. (1986). Robust effects of syntactic structure on visual word processing. Memory & Cognition, 14, 104-112. Williams, I. N. (1988). Constraints upon semantic activation during sentence comprehension. Language and Cognitive Processes, 3, 165-206.

96

P.G. O'Seaghdha

Wright, B. & Garrett, M. F. (1984). Lexical decision in sentences: Effects of syntactic structure. Memory & Cognition, 12, 31-45. Zwitserlood. P. (1989). The locus of the effects of sentential-semantic context in spoken word-recognition. Cognition, 32, 25-64.

Understanding Word and Scntcnce

G.B.Simpson (Editor) Elscvicr Scicncc Publishers I1.V. (hbrth-Holland), 1991

Chapter 5 The Role of Suppression in Sentence Comprehension Morton Ann Gernsbacher and Mark Fausl

University of Oregon Eugene, Oregon U.S.A.

Nearly twenty years ago, Gough (1971) wrote: “The problem of when and how a sentence is understood is, in my view, the central problem of experimental psycholinguistics. Its solution in the form of a machine which could understand sentences would, at the least, earn its inventor an invaluable patent. But while a machine which could understand sentences would be something to marvel at, a person who could do only that would not even make good company.” (p. 64). Two decades later, we continue to share Gough’s appreciation, amazement, and curiosity. How do people comprehend sentences? We have approached this question by tracing the cognitive processes and mechanisms that underlie sentence comprehension (and more generally, language comprehension). We have identified a few of those cognitive processes and mechanisms in a framework we call the Structure Building Framework (Gernsbacher, in pressa; Gernsbacher, in press-b). According to the Structure Building Framework, the goal of comprehension is to build a coherent mental representation or structure. These structures represent sentences, paragraphs, passages, and any other meaningful unit. For instance, comprehending a sentence requires building a mental structure to represent that sentence. The building blocks of mental structures are memory cells. Memory cells represent previously stored memory traces. Their representation might be in either the traditional sense of an individual cell representing an individual trace, or the distributed sense of a group of cells representing an individual trace. Memory cells are automatically activated by incoming stimuli. Once activated, the information they represent can be used by cognitive processes. Furthermore, according to the Structure Building Framework, once activated, memory cells transmit processing signals. These processing signals either suppress or enhance the activation of other memory cells. So, once memory cells are activated, two mechanisms modulzte their level of activation: They are suppression and enhancement. Suppression decreases or dampens the activation of memory cells when

98

M.A. Gernsbacher and M.Faust

the information they represent is no longer as necessary for the structure being built. Enhancement increases or boosts the activation of memory cells when the information they represent is relevant to the structure being built. By modulating the activation of memory cells, suppression and enhancement contribute to comprehension. The notion that incoming stimuli activate memory representations is familiar. What is novel about the Structure Building Framework is its proposal that activated memory cells transmit processing signals. This additional proposal more fully captures the analogy of neural activity - an analogy that inspires many models of cognition. The familiar notion that incoming stimuli activate memory representations captures only one aspect of the analogy, the electrical transmission of information (along axons). But the novel proposal that activated memory cells transmit processing signals (for suppression and enhancement) continues the analogy by paralleling the chemical transmission of information (across synapses, via neurotransmitters). We propose that the mechanisms of suppression and enhancement are general cognitive mechanisms. They are not dedicated to language; they play vital roles in many nonlinguistic phenomena, too. Yet, they are crucial to language comprehension. In this chapter, we focus on the mechanism of suppression and the vital role that suppression plays in sentence comprehension. In the first half of the chapter, we illustrate the vital role that suppression plays in sentence comprehension by demonstrating how suppression fine tunes the meanings of words. In the second half of the chapter, we illustrate the vital role that suppression plays in sentence comprehension by documenting that less-skilled comprehenders suffer from less-efficient suppression mechanisms.

THEROLEOF SUPPRESSION IN FINETUNING THE MEANINGS OF WORDS According to many models of word understanding, when comprehenders first hear or read a word, information provided by that word activates various potential meanings. Then, constraints provided by lexical, semantic, syntactic, and other sources of information alter those meanings’ levels of activation. Eventually, one meaning becomes most strongly activated. That meaning is what comprehenders access and incorporate into their developing mental structures (these ideas are culled from the models of (Becker, 1976; Kintsch, 1988; Marslen-Wilson, Tyler, & Seidenberg, 1978; McClelland & Kawamoto, 1986; Norris, 1986). What the Structure Building Framework adds to these ideas is the proposal that the mechanisms of suppression and enhancement modulate the different meanings’ levels of activation. In particular, according to the Structure Building Framework, the mechanism of suppression helps fine tune the mean-

The Role of Suppression in Sentence Comprehension

99

ings of words by suppressing the less likely meanings. An excellent arena for demonstrating the vital role that suppression plays in fine tuning word meaning is provided by ambiguous words - for instance, words like bug that have at least two diverse meanings. Ambiguous words have clearly distinct meanings, and in sentence contexts one meaning is usually more appropriate. But contrary to intuition, immediately after comprehenders hear or read an ambiguous word in context, multiple meanings are often activated, even when only one meaning is suggested by the context. For example, immediately after comprehenders hear the word bug, both the “insect” meaning and the “covert microphone” meaning are activated (Swinney, 1979). Both meanings are activated even when the context is biased toward the “insect” meaning, as in (1) The man was not surprised when he found several spiders, roaches, and other bugs .... This immediate activation of multiple meanings, regardless of context, is demonstrated by the following experimental task: Subjects listen to a series of sentences. At a critical point during each sentence, the subjects see a test word. The subjects must decide rapidly whether that test word is an English word. For example, if sentence (1) was presented in such an experiment, then immediately after subjects heard the word bug, they might see the test word ANT. That test word is related to the contextually appropriate meaning of bug (the meaning implied by the context). In another condition of the same experiment, subjects might see the test word SPY immediately after they hear bug. The test word SPY is related to a contextually inappropriate meaning of bug (a meaning not implied by the context). In a third condition, the subjects might see the test word SEW. That test word is unrelated to any meaning of bug and serves as a control. If subjects are tested immediately after they hear the word bug, they respond just as rapidly to SPY as they respond to ANT. And they respond to both SPY and ANT more rapidly than they respond to the unrelated test word SEW. In other words, subjects respond to test words that are related to the contextually inappropriate meanings just as rapidly as they respond to test words that are related to the contextually appropriate meanings. This result suggests that immediately after comprehenders hear ambiguous words, both appropriate and inappropriate meanings are activated - and both meanings are more activated than unrelated concepts. But this is only what happens when activation is measured immediately after comprehenders hear ambiguous words. Comprehenders do not keep multiple meanings activated forever. If they did, they would never unambiguously understand any utterance or passage. Instead, multiple meanings are activated

M.A. Gernsbacher and M. Faust

100

only momentarily. For instance, when subjects continue listening to sentence (1) and are tested only four syllables after hearing the word bug, they still respond rapidly to ANT. But they respond no more rapidly to SPY than they respond to SEW. In other words, tliey respond no more rapidly to test words that are related to contextually inappropriate meanings than they respond to test words that are unrelated to any meaning. This finding suggests that after one and a half syllables, the inappropriate meanings have decreased in activation. Other experiments have demonstrated that inappropriate meanings decrease in activation even more quickly, often within only 200 ms. That is probably why comprehenders are typically aware of only one meaning - the contextually appropriate one. This phenomenon, immediate activation of multiple meanings but continued activation of only appropriate meanings, is also demonstrated with other laboratory tasks. It is demonstrated when subjects read sentences one word at a time, and occasionally, instead of seeing the next word of a sentence, they see the test words. They decide rapidly whether each test word is an English word (Kintsch & Mross. 1985; Till, Mross, & Kintsch, 1988). The phenomenon is also demonstrated when subjects listen to sentences and are visually presented with test words, But instead of rapidly deciding whether each test word is an English word, they simply pronounce each test word as rapidly as possible. Or they simply name the color of ink in which each test word is printed (Conrad, 1974; Lucas, 1987; Seidenberg, Tanenhaus, Leiman, & Bienkowski, 1982; Tanenhaus, Leiman. & Seidenberg, 1979). Each of these laboratory tasks demonstrates that multiple meanings of ambiguous words are often immediately activated - regardless of semantic context. But only contextually appropriate meanings remain activated a short while later, This phenomenon occurs even when one meaning is a noun and the other is a verb (Seidenberg et al., 1982). For example, watch refers to both an object, a timepiece, and an action, looking. Sentence (2) implies the noun meaning of watch, while sentence ( 3 ) implies the verb meaning. (2) I like the watch.

(3) I like to watch. Why multiple meanings are immediately activated without regard to context intrigues researchers, perhaps because the phenomenon challenges introspection. Many laboratory investigations have searched for its boundary conditions (Blumer & Sommer, 1988; Burgess. Tanenhaus, & Seidenberg, 1989; Glucksberg, Kreuz, & Rho, 1986; Tabossi, 1988; Tabossi, Colombo, & Job, 1987; Van Petten & Kutas, 1987; Williams, 1988). But equally intriguing are the following questions: What happens to the inappropriate meanings? How do they become less activated? Unfortunately,

The Role of Suppression in Sentence Comprehension

101

scant empirical attention has been directed toward answering these questions. According to the Structure Building Framework, inappropriate meanings become less activated via the mechanism of suppression. The memory cells representing the semantic or syntactic context transmit processing signals. These processing signals suppress the contextually inappropriate meanings. Dampening the activation of inappropriate meanings could be one of the most important roles that the mechanism of suppression plays in sentence comprehension. But other theories assume that inappropriate meanings become less activated via other mechanisms. For instance, according to some theories, inappropriate meanings are inhibited by appropriate meanings, and according to other theories, inappropriate meanings simply decay. Unfortunately neither assumption has been tested empirically. That was the purpose of the experiments we shall describe next. Are Inappropriate Meanings Mutually Inhibited? Some theories propose that inappropriate meanings become less activated through a mechanism we shall call compensatory inhibition (McClelland & Kawamoto, 1986; Waltz & Pollack, 1985). These theories assume that all concepts compete for a fixed amount of activation. So when multiple meanings of ambiguous words are immediately activated, they are sharing this fixed sum. Later, inappropriate meanings must decrease in activation presumably because appropriate meanings have increased. Like a seesaw, when one meaning becomes more activated, the other must become less activated. But if reaction times reflect activation, which is what many reaction time researchers assume (Posner. 1978), the behavioral data do not demonstrate this compensatory pattern. Simply put: The appropriate meanings do not increase in activation when the inappropriate meanings decrease. For instance, in neither Swinney’s (1979) nor Seidenberg et al.3 (1982) data did the appropriate meanings increase in activation from the immediate to the delayed test point. But in both sets of data, the inappropriate meanings decreased. This is the pattern typically observed in these experiments. Perhaps appropriate meanings do not observably increase in activation because during the delay they are competing with other concepts for the fixed sum of activation. By definition, Swinney’s (1 979) four-syllable delay introduced new syllables (four, to be precise). Perhaps during these four syllables, new concepts were introduced. For example, sentence (1) continued, (4) The man was not surprised when he found several spiders, roaches, and other bugs in the corner of the room.

We need some way to introduce a delay without introducing new concepts. In the following experiment we did just that. We selected 48 ambiguous

M.A. Gernsbacher and M.Faust

102

words that were just as likely to be thought of as verbs as nouns, according to ambiguity norms (Cramer, 1970; Kausler & Kollasch, 1970; Nelson, McEvoy, Walling, & Wheeler, 1980). For each ambiguous word, we constructed two experimental sentences. The two sentences were identical until after the ambiguous word occurred, with the following exception: In one sentence, the ambiguous word was preceded with the infinitive marker to, whereas in the other sentence, the ambiguous word was preceded with the definite article the. For example,

(5) Jack tried to punch .... (6) Jack tried the punch .... For each ambiguous word, we selected two test words: One test word was related to the verb meaning, and the other was related to the noun meaning. The two test words for sentences ( 5 ) and (6) are illustrated in Table 1.

TEST WORDS

SENTENCES HIT

DRINK

Related to APPROPRIATE Meaning

Related to INAPPROPRIATE Meaning

Related to INAPPROPRIATE Meaning

Related to APPROPRIATE Meaning

...

Unrelated to Any Meaning

Unrelated to Any Meaning

Jack tried the rolls ...

Unrelated to Any Meaning

Unrelated to Any Meaning

Jack tried to punch ...

Jack tried the punch

Jack tried to bluff

...

For each ambiguous word, we also constructed two control sentences, which were identical to the two experimental sentences up to the point where the ambiguous words occurred. In the control sentences, the experimental ambiguous words were replaced with other ambiguous words (which they matched

The Role of Suppression in Sentence Comprehension

103

in length and familiarity). For example,

....

(7) Jack tried to bluff (8) Jack tried the rolls ....

The control words (e.g., bluff or rolls) were unrelated to the test words (e.g.,

HIT or DRINK). This relationship is also illustrated in Table 1. Finally, we constructed 48 “lure” sentences that resembled the expenmental and control sentences. The test words for the lure sentences were pronounceable strings of letters that did not form English words (e.g., HUP, DRACK). All of the sentences were presented visually, word-by-word in the center of a computer screen. Immediately after the ambiguous word disappeared (e.g., punch), or the control word disappeared (e.g., bluffl, a test word appeared. The test words appeared at the top of the screen in capital letters. Subjects decided rapidly whether each test word was an English word. After the ambiguous or control words occurred, their sentences continued in meaningful but different ways. For example,

(9) Jack tried the punch but he didn’t think it tasted very good. However, remember that the test words always appeared immediately after the ambiguous or control words; so, activation was always measured before the sentences diverged. We measured activation at two test intervals. These test intervals were produced by manipulating the rate at which the words in the sentences appeared. There were two presentation rates: At the faster rate, each word ap-

Faster Rate EfIl Word presan t;, tion

punch

Jack tried to

0 150 ms interval

IT

Slower Rate tried

Jack

.

punch

to

1

I

I

I

I

I

I

I

I

I

I

I

I

I

r

0

300

.

I

HIT I

1

I

I

I

600 900 1200 15001800 2100240027003000 3300

Figure 1. Rates at which the words of the sentences were presented.

M.A. Gernsbacher and M.Faust

104

peared for 16.667 ms per character, plus a constant 150 ms. At the slower rate, each word appeared for 50 ms per character, plus a constant 450 ms. Figure 1 illustrates these presentation rates. In both the fast and slow presentation rate, a constant 150 ms intervened between the appearance of each word in a sentence. And in both the fast and slow presentation rate, a constant 150 ms intervened between the ambiguous (or control) word and its test word. The difference in these rates created the difference between the two test points. With the faster rate, a five-letter word (like punch) appeared for 233 ms; with the slower rate, the same five-letter word

I

Faster Rate &'.................. ... .,..B@$E.

(punchi"iJ~i~ i....

.......x.:.:.:.:.:.:.:&+ ........................

I

Slower Rate

HIT

I ;

rio

2b0 3:o

460

560

600 760

*o;

goo

Id00

Figure 2. Rates at which the words of the sentences were presented.

appeared for 700 ms. So, the difference between the two test points for fiveletter words was 467 ms. Figure 2 illustrates this difference. For continuity with the other experiments we have discussed, we shall call the test point produced by the faster rate Immediare, and the test point produced by the slower rate Delayed. Figure 3 displays our 80 subjects' data. We estimated activation by subtracting subjects' latencies to respond to test words that were related to the appropriate or inappropriate meanings of the ambiguous words from their latencies to respond to test words that were unrelated to any meaning of the ambiguous words. For instance, we estimated the activation of the contextually appropriate meaning of to punch by subtracting subjects' latencies to respond to HIT after reading Jack tried to punch from their latencies to respond to HIT after reading Jack tried 10 blufl. Similarly, we estimated the activation of the contextually inappropriate meaning of 10 punch by subtracting subjects' latencies to respond to DRINK after reading Jack tried 10 punch from their latencies to respond to DRINK after reading Jack rried lo bluff. First examine what happened at the immediate test point. As Figure 3

The Role of Suppression in Sentence Comprehension

105

illustrates, at the immediate test point (caused by the faster presentation rate), both the appropriate and the inappropriate meanings were reliably more activated than unrelated concepts, minF' (1.83) = 6.495, p c .01 for appropriate

6o T

50t

Estimated Activation

40

munrelated 30

- RTrelated

20 10

0 Immediate Appropriate Meanino

Delayed

0 Inappropriate Meaninq

Figure 3. Subjects' average activation scores.

meanings, and minF' (1.82) = 10.26, p < .005 for inappropriate meanings. Indeed, at this immediate point, the appropriate and inappropriate meanings were activated at the same level (i.e., their activation levels did not differ, both Fs c

3.

Now, examine what happed at the delayed test point. As Figure 3 illustrates, after the delay (caused by the slower presentation rate), only the appropriate meanings were reliably activated, minF' (1,82) = 4.562, p < .05. In contrast, the inappropriate meanings were considerably less activated than the appropriate meanings, minF' (1,82) = 3.919, p c .05. Indeed, the inappropriate meanings were no more activated than unrelated concepts, both Fs c 1.0. These data replicate those of Swinney (1979) and Seidenberg et al. (1982). They also demonstrate that when inappropriate meanings decrease in activation, appropriate meanings do not increase; in other words, there is no compensation. If reaction times reflect activation levels, then there is no evidence that inappropriate meanings lose activation because appropriate meanings take a larger

M.A. Gernsbacher and M. Faust

106

share (of a fixed sum). In other words, there is no evidence to support the compensatory inhibition explanation for why inappropriate meanings lose activation, especially when no new concepts are introduced during the test delay.

Do Inappropriate Meanings Simply Decay? Another explanation for why inappropriate meanings become less activated is that they decay. In many models of cognition, mental representations automatically decay when they are not continuously stimulated (Anderson, 1983). Inappropriate meanings might therefore decay because they do not continuously receive stimulation from a biasing semantic or syntactic context. We empirically tested this decay explanation in the following experiment. We selected 48 ambiguous words that were just as likely to be thought of as one noun as another (according to ambiguity norms). For example, the word quack is just as likely to be interpreted as “an incompetent doctor” as “the sound a duck makes.” For each of the 48 ambiguous words, we constructed three experimental sentences. One experimental sentence was biased toward one meaning of the ambiguous word, for example, (10) Pam was diagnosed by a quack

....

A second experimental sentence was biased toward another meaning of the ambiguous word, for example, (11) Pam heard a sound like a quack

....

But the third experimental sentence was neutral: Neither its semantic nor its syntactic context was biased toward either meaning of the ambiguous word, for example, (12) Pam was annoyed by the quack

....

To ensure that our sentences were effectively biased or neutral, we had 50 subjects read the beginnings of the sentcnces (e.g.. Pam wus annoyed by the quack ....). These subjects decided which meaning was intended. We used biased sentences only if 95% of these subjects agreed with the meaning we intended, and we used neutral sentences only if these subjects were roughly split over which of the two meanings we intended. For each of the 48 ambiguous words, we selected two test words. One was related to one of the biased meanings (e.g.. DOCTOR), and the other was related to the other biased meaning (e.g., DUCK).The test words and experimental sentences are illustrated in Table 2.

The Role of Suppression in Sentence Comprehension

107

For each of the 48 ambiguous words, we also constructed a control sentence. The control sentences were identical to the neutral experimental sentences to the point where the ambiguous words occurred. In the control sentences, the experimental ambiguous words were replaced with unrelated ambiguous words (which matched the experimental words in length and familiarity). For example, (13) Pam was annoyed by the pupil

....

The ambiguous words in the control sentences were unrelated to the test words. This relationship is also illustrated in Table 2.

TABLE 2 SENTENCES

TEST WORDS

DOCTOR

DUCK

Pam was diagnosed by a quack ...

Related to APPROPRIATE Meaning

Related to INAPPROPRIATE Meaning

Pam heard a sound like a quack ...

Related to INAPPROPRIATE Meaning

Related to APPROPRIATE Meaning

Pam was annoyed by the quack ...

Neutral

Neutral

Pam was annoyed by the pupil ...

Unrelated to Any Meaning

Unrelated to Any Meaning

We also constructed 48 lure sentences that resembled the experimental and control sentences, but the test words for the lure sentences were pronounceable strings of letters that did not form English words. All the sentences were presented visually, as in the experiment we described before. And as in the experiment we described before, the sentences continued in meaningful but different ways after the ambiguous or control words. For example, (14) Pam heard a sound like a quack but couldn’t imagine where it was

coming from.

108

M.A. Gernsbacher and M. Faust

However, it was before the sentences diverged that we measured activation. We again manipulated the presentation rate (as illustrated in Figures 1 and 2). so that we could measure activation at two test points without introducing new concepts. To summarize, there were three experimental sentences. One was biased toward one meaning of the ambiguous words; one was biased toward another meaning; and the third was neutral - there was no semantic or syntactic bias. While subjects read these experimental sentences, we measured how activated the multiple meanings were. And we made this measurement at two test points. The decay explanation and the suppression explanation make identical predictions about the biased sentences; these sentences should replicate earlier experiments: At the immediate test point, both appropriate and inappropriate meanings should be activated, but at the delayed test point, the inappropriate meanings should be less activated (in relation to the unrelated control condition). Where the decay and the suppression explanations differ is their predictions about the neutral sentences. According to the decay explanation, inappropriate meanings become less activated because they automatically decay. And they decay because they lack stimulation from a semantic or syntactic context. Because neutral sentences also lack stimulation from a semantic or syntactic context, multiple meanings of ambiguous words should also decay. In other words, the decay explanation predicts that with neutral sentences, both meanings should be less activated after the delay than they are immediately. This is because neither meaning receives stimulation from a semantic or syntactic context. In contrast, according to the suppression explanation, inappropriate meanings become less activated because the memory cells representing semantic or syntactic contexts transmit processing signals; these processing signals suppress the inappropriate meanings’ activation. So, the suppression explanation predicts that only the inappropriate meanings of the biased sentences should become less activated after the delay; the multiple meanings of the neutral sentences should be just as activated after the delay as they are immediately. This is because there are no bases from which suppression signals can be transmitted. So, the decay explanation predicts that with the neutral sentences, both meanings should be less activated after the delay than they are immediately. But the suppression explanation predicts that both meanings should be just as activated after the delay as they are immediately. Figure 4 displays our 80 subjects’ data. We estimated activation by subtracting subjects’ latencies to respond to test words that were related (to the appropriate, inappropriate, or both meanings of the neutral sentences) from their latencies to respond to test words that were unrelated to any meaning. First, examine what happened at the immediate test point. As Figure 4

The Role of Suppression in Sentence Comprehension

109

illustrates, with the biased sentences, both the appropriate and inappropriate meanings were reliably activated as were both meanings with the neutral sentences, minF' (1.73) = 4.503, p e .05 for appropriate meanings, minF' (1,83) = 4.773, p < .05 for inappropriate meanings, and minF' (1.83) = 3.711, p e .05 for both meanings with the neutral sentences.

Estimated Activation 40

t

Immediate Appropriate Meaning

Neutral

Delayed

0 lnawropriate Meaning

Figure 4. Subjects' average activation scores

Now, examine what happened after the delay. As Figure 4 illustrates, after the delay, the inappropriate meanings of the biased sentences were less activated; indeed, they were (statistically) no more activated than unrelated concepts, minF' e 1.0. In contrast, with the neutral sentences, both meanings were still reliably more activated than unrelated concepts, minF' (1,83) = 3.846, p e .05. The same was true of the appropriate meanings (with the biased sentences), minF' (1,83) = 4.702, p e .05. Indeed, as Figure 4 illustrates, with the neutral sentences, the ambiguous words' multiple meanings were just as activated after the delay as they were immediately. These results confirm the prediction made by the suppression explanation, not the decay explanation. The suppression explanation, drawn from the Structure Building Framework, predicts that inappropriate meanings become less activated because the memory cells representing semantic or syntactic contexts transmit processing signals; these processing signals suppress the

M.A. Gernsbacher and M . Fausr

110

inappropriate meanings' activation. With a neutral context, multiple meanings remain activated because there are no bases from which suppression signals can be transmitted. These two experiments demonstrate that the inappropriate meanings of ambiguous words do not decrease in activation because they are mutually inhibited; neither do they decrease in activation because they decay. Rather, we suggest that they are suppressed. In this way, the mechanism of suppression plays a vital role in sentence comprehension: It fine tunes the meanings of words.

THEROLE OF SUPPRESSION IN COMPREHENSION SKILL There are many situations in which irrelevant or inappropriate information is automatically activated, unconsciously retrieved, or naturally perceived. But for successful comprehension, this irrelevant or inappropriate information must not affect ongoing processes. According to the Structure Building Framework, irrelevant or inappropriate information is suppressed.But what if a comprehender's suppression mechanism was faulty? Irrelevant or inappropriate information would remain activated. Surely that would affect the comprehender's success. Perhaps that is one of the reasons why some comprehenders are less successful: They have less-efficient suppression mechanisms. In several experiments, we have investigated whether less-skilled comprehenders are indeed characterized by less-efficient suppression mechanisms. Are Less-skilled Comprehenders Less Efficient at Suppressing the Inappropriate Meanings of Ambiguous Words? We have suggested that successful comprehension requires suppressing the contextually inappropriate meanings of ambiguous words. For example, successfully comprehending sentence (15) requires suppressing the playing card meaning of the word spade. (15) He dug with the spade.

If less-skilled comprehenders are plagued by less-efficient suppression mechanisms, then they should be less able to suppress these contextually inappropriate meanings. We tested this hypothesis in Gernsbacher, Varner. and Faust (1990). We selected two samples of more- versus less-skilled comprehenders from a disuibution of 270 University of Oregon students. All 270 students had previously been tested on our Multi-Media Comprehension Battery (Gernsbacher & Varner, 1988). The more-skilled comprehenders' scores were from the upper third of the distribution of Comprehension Battery scores; the less-skilled comprehen-

The Role of Suppression in Sentence Comprehension

111

ders were from the lower third of the distribution. When these more- and lessskilled comprehenders returned to the lab, they performed the following task: They read short sentences, and after each sentence, they saw a test word. Their task was to verify whether the test word fit the meaning of the sentence they just read. On 80 trials, the test word did indeed fit the sentence, but we were more interested in the 80 trials in which the test word did not fit the sentence. On half of those trials, the last word of the sentence was an ambiguous word, for example, (15) He dug with the spade. The test word on these trials was a meaning of the ambiguous word that was inappropriate to the context, for example, ACE We measured how long subjects took to reject a test word like ACE after reading a sentence like (15). And we compared that latency with how long subjects took to reject ACE after reading the same sentence but with the last word replaced by an unambiguous word, for example, (16) He dug with the shovel.

This comparison showed us the activation level of the inappropriate meanings; the more time subjects took to reject ACE after the spade- versus the shovelsentence, the more activated the inappropriate meaning must have been. We presented the test words at two points: immediately (100 ms) after subjects finished reading each sentence, and after an 850 ms delay. We predicted that at the immediate test point, both the more- and less-skilled comprehenders would take longer to reject test words after reading the ambiguous words as opposed to the unambiguous words. For example, both groups would take longer to reject ACE after reading the spade sentence than after reading the shovel sentence. This prediction was based on the studies we described earlier which demonstrate that immediately after ambiguous words are read, contextually inappropriate meanings are often activated. We particularly expected the inappropriate meanings to be activated because our task required comprehenders to focus their attention on a subsequent word and try to integrate that word into the previous context (Glucksberg et al., 1986; Van Petten & Kutas, 1987). Our novel predictions concerned what would happen after the 850 ms delay. We predicted that by that point the more-skilled comprehenders would not take longer to reject test words following ambiguous words. This is because more-skilled comprehenders should be able to successfully suppress the inappropriate meanings. But we made a different prediction for our less-skilled comprehenders. If less-skilled comprehenders are plagued by less-efficient suppression mechanisms, then even after the delay, the inappropriate meanings should still be activated.

M.A. Gernsbacher and M . Faust

112

Figure 5 displays our 64 subjects' data. We estimated activation by subtracting subjects' latencies to reject test words like ACE after reading ambiguous words like spade from their latencies to reject test words like ACE after reading unambiguous words like shovel. The more-skilled comprehenders are represented by hashed lines, and the less-skilled comprehenders are represented by unfilled bars.

100 ESTIMATED ACTIVATION 80 RTambig -

RTunambig

60 40

20 0 Immediate More-Skilled Comprehenders

Delayed

0 Less-Skilled Comprehenders

Figure 5. Subjects' average activation s c o m (from Gernsbacher et al., 1990)

First, examine what happened at the immediate test point. As Figure 5 illustrates, immediately after both the more- and less-skilled comprehenders read the ambiguous words, the inappropriate meanings were highly activated. Now, examine what happened after the delay. As Figure 5 illustrates, 850 ms after the more-skilled comprehenders read the ambiguous words, the inappropriate meanings were no longer reliably activated; by this time, the more-skilled comprehenders had successfully suppressed them. But the less-skilled comprehenders were less fortunate: As Figure 5 illustrates, even after the 850 ms delay, the inappropriate meanings were still highly activated. In fact, they were as highly activated after the delay as they were immediately. So, almost a second after the less-skilled comprehenders read the ambiguous words, they were unable to suppress the inappropriate meanings. These results support the hypothesis that less-skilled comprehcnders are plagued by less rapid (and therefore less-efficient) suppression mechanisms.

The Role of Suppression in Sentence Comprehension

113

Are Less-skilled Comprehenders Less Efficient at Suppressing the Incorrect Forms of Homophones? Reading a string of letters activates an array of information. Virtually always reading a letter string activates orthographic information - information about the individual letters in the string and their relative position to one another. Often, reading a letter string activates semantic information, lexical information, and phonological information. In fact, semantic, lexical, and phonological information is often activated even when the string does not compose an English word (Coltheart, Davelaar, Jonasson, & Besner, 1977; Rosson, 1985). Automatic activation of phonological information was the focus of OUT next experiment. By automatic activation of phonological information we meant the phenomenon in which reading the letter string rows activates the phonological sequence lrozl. In fact, reading rows can activate lrozl, which can activate rose. In other words, reading a homophone (rows) can activate a phonological sequence (/rod),which can then activate another form of the homophone (rose). How do we know that a letter string often activates phonological information, which in turn activates other forms of homophones? Consider the following finding: Comprehenders have difficulty quickly rejecting the word rows as not being an exemplar of the category FLOWER (van Orden, 1987; van Orden, Johnston, & Hale, 1988). But to successfully comprehend a written passage, these incorrect forms cannot remain activated. According to the Structure Building Framework, sentence comprehension involves the mechanism of suppression. The same cognitive mechanism that suppresses the inappropriate meanings of ambiguous words, could also suppress the incorrect forms of homophones. If this is the same mechanism, and if this general suppression mechanism is less efficient in less-skilled comprehenders, then less-skilled comprehenders should also less efficiently suppress the incorrect forms of homophones. Related evidence already supports this prediction. Consider the sentence: (17) She blue up the balloon.

Six-year olds are more likely to accept that sentence than are 10-year olds even when they clearly know the difference between blue and blew (Coltheart, Laxon, Rickard, & Elton, 1988; Doctor & Coltheart, 1980). If we assume that 6-year olds are less skilled than 10-year olds at comprehension, this finding suggests that less-skilled comprehenders are less able to suppress the incorrect forms of homophones that are often automatically activated. In Gernsbacher and Faust (in press), we tested this hypothesis more directly, with adult subjects whom we knew differed in comprehension skill. Our subjects were US Air Force recruits who were drawn from a sample of 455 subjects whom we had previously tested with the Multi-Media Comprehension

M.A. Gernsbacher and M. Faust

114

Battery.’ We drew 48 subjects from the top third of the distribution (those who scored the highest) and 48 subjects from the bottom third of the distribution (those who scored the lowest). When these more- versus less-skilled comprehenders returned to the lab, they performed a laboratory task similar to the task we used in Gemsbacher et al. (1990). The subjects read short sentences, and following each sentence, they saw a test word. The subjects’ task was to verify whether the test word fit the meaning of the sentence they just read. On 80 trials, the test word did indeed fit the sentence’s meaning, but on 80 trials it did not. We were interested in those trials in which the test word did nor fit the meaning. On half of those trials, the last word of the sentence was one form of a homophone, for example, (18) He had lots of patients.

On these trials. the test word was related to the homophone’s other form; for example, the test word CALM is related to patience. We compared how long subjects took to reject CALM after reading sentence (18) with how long they took to reject CALM after reading the same sentence with the last word replaced by a nonhomophone, for example, (19) He had lots of students.

This comparison showed us the activation levels of the incorrect forms; the more time subjects took to reject CALM after the parienls- versus studentssentence, the more activated the patients form of the homophone must have beenS2 We presented the test words at two test points: immediately (100 ms) after subjects finished reading each sentence, and after a one-second delay. We predicted that at the immediate test point, both the more- and less-skilled comprehenders would take longer to reject test words following homophones than nonhomophones. For example, both groups would take longer to reject CALM after reading the patients sentence than after reading the students sentence. This result would corroborate the results of van Orden (1987; van Orden et al., 1988). This result would also demonstrate that comprehenders of both skill levels often activate phonological information during reading. Our novel predictions concerned what would happen after the one-second delay. We predicted that after the one-second delay, the more-skilled comprehenders would not take longer to reject test words following homophones versus nonhomophones; this is because more-skilled comprehenders should be able to successfully suppress incorrect forms. But we made a different prediction for our less-skilled comprehenders. If less-skilled comprehenders are characterized by less-efficient suppression mechanisms, then even after the one-

The Role of Suppression in Senfence Comprehension

115

second delay, the incorrect forms of the homophones should still be highly activated. Figure 6 illustrates our 96 subjects’ data. We estimated activation by sub-

120

ESTIMATED ACTIVATION 100

80 RThphone - RTnonhphone 60

40 20

0

-I

Immediate More-Skilled Comprehenders

Delayed

Less-Skilled Comprehenders

Figwe 6. Subjects’ average acdvation scores (from Gemsbacher & Faust. in press)

tracting subjects’ latencies to reject test words like CALM after reading nonhomophones like students from their latencies to reject test words like CALM after reading homophones like patients. First examine what happened at the immediate test test point. As Figure 6 illustrates, immediately after both the more- and less-skilled comprehenders read the homophones, the inappropriate forms were highly activated; in fact, they were almost equally activated for the more- versus less-skilled comprehenders. So, 100 ms after comprehenders of both skill levels read homophones, other forms are often activated. Now, examine what happened after the one-second delay. As Figure 6 illustrates, one second after the more-skilled comprehenders read the homophones, the incorrect forms were no longer reliably activated; the more-skilled comprehenders had successfully suppressed them. But as Figure 6 also illustrates, the less-skilled comprehenders were less fortunate: Even after the onesecond dclay, the incorrect forms were still highly activated; in fact, they were as activated after one second as they were immediately. So, a second after the less-skilled comprehenders read the homophones, they were unable to suppress the incorrect forms. These data support the hypothesis that less-skilled compre-

M.A. Gernsbacher and M . Faust

116

henders are plagued by less-efficient suppression mechanisms. Are Less-skilled Comprehenders Less Efficient at Suppressing Information Across Modalities? Comprehension often requires making sense of stimuli that originate from various modalities. We would be severely handicapped if we were skilled at only reading written words, or only listening to spoken words, or only comprehending graphic displays. Information originates from different modalities, often simultaneously. We read while listening to music, and we drive while carrying on a conversation. Comprehenders often experience interference across modalities. For in-

PICTURE TRIAL Context Display

Test Display

I

I

WORDTRIAL Context Display

rest Display

&rREAM

Figure 7.(From Gernsbacher & Faust. in press)

stance, it is harder to name a pictured object such as an ashtray if a letter string such as INCN is written across the picture, as illustrated in the uppcr lcft panel

The Role of Suppression in Sentence Comprehension

117

of Figure 7. The opposite is also true: It is harder to read a word such as RIVER if it is superimposed on a picture, as illustrated in the bottom left panel of Figure 7 (Smith & McGee, 1980). Successful comprehension often requires suppressing information across modalities. The same mechanism that suppresses information within modality, could suppress information across modalities. If this is the same mechanism, and if this general suppression mechanism is less efficient in less-skilled comprehenders, then less-skilled comprehenders should also be less efficient in suppressing information across modalities. We tested this hypothesis in the following way. We modified Tipper and Driver’s (1988) experimental task. In our modification, subjects first viewed a context display. Each context display contained a line-drawn picture of a common object and a familiar word. For example, the top panel in Figure 7 illustrates a picture of an ashtray with the word INCH written across it. The bottom panel of Figure 7 illustrates the word RIVER superimposed on a picture of a baseball player. All context displays contained both a picture and a word. After subjects viewed each context display, they were shown a test display. Each test display contained either another picture or another word. Half the time, the test display contained another picture, and we referred to those trials as Picture trials: half the time, the test display contained another word, and we referred to those trials as Word trials. Subjects were told before each trial whether that trial would be a Picture trial or a Word trial. The top panel of Figure 7 illustrates a Picture trial. On Picture trials, subjects were told to focus on the picture in the context display and ignore the word. For example, for the Picture trial shown in Figure 7, subjects should have focused on the ashtray and ignored the word INCH. Following each context display, subjects were shown a test display. On the Picture trials, the test display contained another picture. The subjects’ task (on Picture trials) was to verify whether the picture shown in the test display was related to the picture shown in the context display. For the Picture trial shown in Figure 7, subjects should have responded “yes,” because the picture shown in the test display, the pipe, was related to the picture shown in the context display, the ashtray. The bottom panel of Figure 7 illustrates a Word trial. On Word trials, subjects were supposed to focus on the word in the context display and ignore the picture. For example, for the Word trial shown in Figure 7, subjects should have focused on the word RIVER and ignored the baseball player. The test display on Word trials contained another word. The subjects’ task was to verify whether the word writtcn in the test display was related to the word written in the context display. For the Word trial shown in Figure 7, subjects should have responded “yes,” because the word written in the test display, STREAM, was related to the word written in the context display, RIVER. On 40 Picture trials and 40 Word trials, the test display was related to what the subjects were to focus on in the context display, just as they are in

M.A. Gernsbacher and M . Faust

118

Figure 7. However, we were more interested in the 80 trials in which the test display was unrelated to what the subjects were supposed to focus on in the context display. On half of those trials, the test display was unrelated to what the subjects were to focus on in the context display, but it was related to what they were supposed to ignore. For example, the top panel in Figure 8 illustrates an experimental Picture

PICTURE TRIAL Context Display

Test Display I

WORDTRIAL

Context Display

Test Display I

Figure 8 . (From Gernsbacher & Faust. in press)

The Role of Suppression in Sentence Comprehension

119

trial. The context display contains a picture of a hand with the superimposed word RAIN. Because this is a Picture trial, subjects should have focused on the picture (the hand) and ignored the word. The test display is a picture of an umbrella. So the test display, the umbrella, is unrelated to what the subjects were supposed to focus on in the context display, the hand; therefore, the subjects should have responded “no.” But the test display is related to what the subjects were supposed to ignore, the word RAIN. We measured how long subjects took to reject the test display, the picture of the umbrella, after viewing the context display, the picture of the hand with the superimposed word RAIN. And we compared that to how long subjects took to reject the same test display, the picture of the umbrella, after viewing the same context display, the picture of the hand, but with another word superimposed, SOUP. This comparison showed us how quickly comprehenders could suppress information across modalities. Experimental Word trials worked similarly, as illustrated by the third panel of Figure 8. When reading this context display, subjects should have focused on the word MONTH and ignored the surrounding picture of a broom. Then, they should have rejected the test display, the word SWEEP, because it is unrelated to the word MONTH. We compared how long subjects took to reject the word SWEEP after reading the word MONTH surrounded by the broom. And we compared that to how long subjects took to reject SWEEP after viewing the same context display with the picture of a broom replaced by a picture of a sandwich (as illustrated by the bottom panel of Figure 8). This comparison showed us how quickly comprehenders could suppress information across modalities. As in our other experiments, we presented the test displays at two test points: Immediately (50 ms) after the context-setting display, and after a onesecond delay. We predicted that at the immediate test point, both the more- and less-skilled comprehenders would take longer to reject a test display when it was related to the ignored picture or word in the context display. This result would corroborate Tipper and Driver (1988). This result would also demonstrate that for both more-and less-skilled comprehenders, ignored pictures or words are often activated. Our novel predictions concerned what would happen after the delay. We predicted that after the one-second delay, the more-skilled comprehenders would not take longer to reject test displays when they were related to the ignored pictures or words. After one second, more-skilled comprehenders should be able to successfully suppress information across modalities. We made a different prediction for our less-skilled comprehenders. If less-skilled comprehenders are characterized by less-efficient suppression mechanisms, then even after the one-second delay, the ignored pictures and words should still be highly activated. Figure 9 displays our 160 subjects’ data. We estimated activation by sub-

M.A. Gernsbacher and M. Fausr

120

loo

T

ESTIMATED ACTIVATION

Immediate More-Skiled Comprehenders

Delayed

0 Less-Skilled Comprehenden

F i g w c 9. Subjects’ average activation scores (from Gemsbacher & Faust, in press)

tracting subjects’ latencies to reject test displays that were unrelated to ignored pictures/words from their latencies to reject test displays that were related to ignored pictures/words.’ First examine what happened at the immediate test test point. As Figure 9 illustrates, immediately after both the more- and less-skilled comprehenders saw the context displays, the ignored pictures/words were highly activated; in fact, they were almost equally activated for the more- versus less-skilled comprehenders. So, 50 ms after viewing pictures with superimposed words or reading words surrounded by pictures, comprehenders of both skill levels have difficulty suppressing related pictures or words, even when they are told explicitly to ignore them. Now examine what happened after the one-second delay. As Figure 9 illustrates, one second after the more-skilled comprehenders saw the context displays. the ignored pictures/words were no longer reliably activated; the more-skilled comprehenders had successfully suppressed them. But as Figure 9 also illustrates, the less-skilled comprehenders were less fortunate: Even after the one-second delay, the ignored pictures/words were still highly activated; in fact, they were as activated after the delay as they were immediately. So, a second after less-skilled comprehenders view pictures with superimposed words or read words surrounded by pictures, they still have difficulty suppressing the ignored pictures or words. These data support the hypothesis that less-skilled comprehenders are plagued by less-efficient suppression mechanisms.

The Role of Suppression in Sentence Comprehension

121

Are Less-skilled Comprehenders Simply Less Appreciative of Sentence Con text?

The three experiments we just described demonstrate that less-skilled comprehenders are less able to reject inappropriate meanings of ambiguous words, incorrect forms of homophones, and ignored pictures and words.We suggest that this inability arises because less-skilled comprehendersare plagued by less-efficient suppression mechanisms. Another explanation is that less-skilled comprehenders have difficulty rejecting inappropriate information because they less fully appreciate what is contextually appropriate. For instance, by this logic, less-skilled comprehenders have difficulty rejecting ACE after reading H e dug with the spade simply because they less fully appreciate that the context of digging with (I spade implies a garden tool, not a playing card. This explanation seems unlikely given the repeated finding that lessskilled comprehenders are not less appreciative of predictable sentence contexts -just the opposite: Less-skilled comprehenders often benefit from predictable contexts more than more-skilled comprehenders do. For example, the word dump is very predictable in the following context: (20) The garbage men had loaded as much as they could onto the truck. They would have to drop off a load at the garbage dump.

In contrast, dump is less predictable in the following context: (21) Albert didn’t have the money he needed to buy the part to fix his car. Luckily, he found the part he wanted at the dump.

All comprehenders pronounce the word dump more rapidly when it occurs in the very predictable context than when it occurs in the less predictable context; in other words, all comprehenders benefit from the predictable contexts. But less-skilled comprehenders benefit even more than more-skilled comprehenders (Perfetti & Roth, 1981). We evaluated this counter-explanation with adult comprehenders and a task similar to those we used in our previous experiments. Subjects read short sentences, and following each sentence, they saw a test word. As in our other experiments, the subjects’ task was to verify whether the test word fit the meaning of the sentence they just read. However, in this experiment we were most interested in the 80 trials in which the test word did indeed match the meaning of the sentence (and, therefore, the subjects should have responded “yes”). On half of those trials, the last word of the sentence was an ambiguous word, for example, spade, and the verb in the sentence biased one meaning of

M.A. Gernsbacher and M . Fausr

122

the ambiguous word, for example, (22) He dug with the spade.

The test word was related to the meaning of the ambiguous word that was biased by the verb, for example, GARDEN. In a comparison condition we presented the same sentence, but the biasing verb was replaced with a neutral verb, for example, (23) He picked up the spade.

The spade in sentence (23) could be either a garden tool or a playing card. We measured how rapidly subjects accepted test words after reading sentences with biasing verbs versus neutral verbs.’ This comparison showed us how fully comprehenders could appreciate the biasing contexts: The faster subjects were to accept GARDEN after reading the sentence with the biasing verb phrase dug with versus the neutral verb phrase picked up, the more fully they appreciated the biasing context. We presented the test words at two test points: Immediately (100 ms) after subjects finished reading each sentence, and after a one-second delay. We predicted that both the more- and less-skilled comprehenders would benefit from the biasing contexts; that is, both groups of comprehenders would accept test words more rapidly when the sentences contained biasing as opposed to neutral verbs. However, we were especially interested in whether the less-skilled comprehenders would benefit less than the more-skilled comprehenders. If less-skilled comprehenders are less able to reject contextually inappropriate information (as we found in our previous experiments) because they are less appreciative of context, then the less-skilled comprehenders should have benefitted less from the biasing contexts. In contrast, if less-skilled comprehenders are less able to reject inappropriate information because they have less efficient suppression mechanisms, then the less-skilled comprehenders should have benefitted just as much from the biasing contexts as the more-skilled comprehenders did. Based on previous experiments, we predicted that the lessskilled comprehenders would benefit even more from the biasing contexts than the more-skilled comprehenders did. Figure 10 displays our 120 subjects’ data. We estimated activation by subtracting subjects’ latencies to accept test words like GARDEN after reading sentences with biasing verbs like dug with from their latencies to accept GARDEN after reading sentences with unbiased verbs like picked up. As Figure 10 illustrates, at both the immediate and the delayed test test points, the biased verbs led to greater activation, and this occurred for both more- and less-skilled comprehenders. Indeed, as Figure 10 also illustrates, at both test test points, the less-skilled comprehenders benefitted from the biasing verbs more than the

The Role of Suppression in Sentence Comprehension

123

more-skilled comprehenders benefitted. These data do not support the hypothesis that less-skilled comprehenders are are less able to reject contextually inappropriate information because they are less appreciative of context.

180

T

ESTIMATED ACTIVATION

Immediate More-Skilled Comprehenders

Delayed

0 Less-Skilled Comprehenders

Figure 10. Subjects' average activation scores (from Gemsbacher & Faust. in press)

CONCLUSIONS The experiments we have described here demonstrate the vital role that the mechanism of suppression plays in comprehension. Suppression helps fine tune the meanings of ambiguous words by decreasing the activation of the contextually inappropriate meanings. Indeed, less-skilled comprehenders are less able to suppress contextually inappropriate meanings. Less-skilled comprehenders are also less able to suppress the incorrect forms of homophones, and they are less able to suppress words while viewing pictures or suppress pictures while reading words. The mechanism of suppression and the mechanism of enhancement play other important roles in sentence comprehension. For instance, the mechanisms of suppression and enhancement are vital to anaphoric reference (Gernsbacher, 1989). Anaphoric reference is the process by which speakers and writers use an anaphor, such as a repeated noun phrase or a pronoun, to refer to a previously

124

M.A. Gernsbacher and M . Faust

mentioned concept (called an antecedent). Many anaphors improve their antecedents’ accessibility by enhancing the activation of their antecedents (the concepts they refer to). Many anaphors also improve their antecedents’ accessibility by suppression; they suppress the activation of other concepts (the concepts not referred to by the anaphors). When other concepts are suppressed, a rementioned concept can rise to the top of the queue of potential referents. Anaphors differ in how much suppression and enhancement they trigger: The more explicit the anaphor, the more suppression and enhancement it triggers. Some anaphors are very explicit; for instance, repeated noun phrases match their antecedents exactly, and are, therefore, very explicit (e.g., The man went to the store. The man bought a quart of milk). Repeated noun phrase anaphors trigger a lot of suppression and enhancement, and they do so immediately. Other anaphors are less explicit; for instance, zero anaphors provide no information about their antecedents (e.g., The man went to the store. The manbought a quart of milk). Zero anaphors trigger very little suppression and virtually no enhancement (e.g., The man went to the store and gi bought a quart of milk). Information from other sources that identifies antecedents, for instance, information from the semantic, syntactic, and pragmatic context, also triggers suppression (although not enhancement). The mechanisms of suppression and enhancement are also crucial to a process we have called cataphoric access (Gernsbacher & Jescheniak, 1990; Gernsbacher & Shroyer, 1989). Just as there are anaphoric devices which enable access to previously mentioned concepts, we have demonstrated that there are cataphoric devices which improve access to subsequently mentioned concepts. For instance, we have demonstrated that spoken stress operates as a cataphoric device. The unstressed indefinite this also operates as a cataphoric device, as in So this man walks into a bar and says ... One way that cataphoric devices improve their concepts’ accessibility is by enhancing the activation of the concepts they mark. Another way that cataphoric devices improve their concepts’ accessibility is by triggering the suppression of other concepts. And a third way that cataphoric devices improve their concepts’ accessibility is by making the concepts they mark more resistant to being suppressed by other concepts. Thus, the general cognitive mechanisms of suppression and enhancement play a vital role in language comprehension. These mechanisms enable us to build mental structures that represent sentences; in other words, these mechanisms enable us to understand sentences, which as Gough (1971) noted, is a paradoxically common but complex behavior. Acknowledgements This research was supported by National Science Foundation Grant BNS 85-10096, National Institutes of Health Research Career Development Award

The Role of Suppression in Sentence Comprehension

125

KO4 NS-01376, and Air Force Office of Sponsored Research Grants 89-0258 and 89-0111. We are indebted to Rachel R.W. Robertson for her invaluable contribution to our research program.

Notes ‘Air Force recruits are high school graduates, and typically 20% have completed some college courses. Our subjects’ ages ranged from 17 to 23, and approximately 18% were female. *To ensure that the homophones would be familiar to our subjects, 25 students from the University of Oregon judged - without time pressure whether the test words fit the meanings of our experimental and filler sentences. We only used experimental sentences and test words if 95% of our students agreed that the test words did not fit their sentences’ meanings, and we only used filler sentences and test words if 95% of our students agreed that the test words did fit their sentences’ meanings. 3Although both more- and less-skilled comprehenders responded more rapidly on Picture trials than Word trials, there were no interactions with modality (Picture vs Word). So, we have collapsed across this variable in our figures. ‘TO ensure that the biased verbs were biased and the neutral verbs were neutral, 25 students at University of Oregon read all of the experimental and comparison sentences and made unspeeded judgments about the meanings of the ambiguous words. We only used biased verbs if 95% of our students selected the meaning of the ambiguous word that we intended, and we only used neutral verbs if our students were roughly split over which meaning we intended (e.g., when given the sentence He picked up the spade, approximately 50% chose GARDEN TOOL and approximately 50% chose PLAYING CARD). References Anderson, J. R. (1983). The architecture of cognition. Cambridge, MA: Harvard University Press. Becker, C. A. (1976). Semantic context and word frequency effects in visual word recognition. Journal of Experimental Psychology: Human Perception and Performance, 2,556-566. Blutner, R., & Sommer, R. (1988). Sentence processing and lexical access: The influence of the focus-identifying task. Journal of Memory and Language, 27, 359-367. Burgess, C., Tanenhaus, M. K., & Seidenberg, M. S . (1989). Context and lexical access: Implications of nonword interference for lexical ambiguity resolution. Journal of Experimental Psychology: Learning, Memory, and Cognition. 15, 620-632. Colthean, M., Davelaar, E., Jonasson, J. T., & Besner, D. (1977). Access to the

126

M.A. Gernsbacher and M . Faust

internal lexicon. In S. Dornic (Ed.), Attention and Performance VI (pp. 535-555). New York: Academic Press. Coltheart, V., Laxon, V., Rickard, M., & Elton, C. (1988). Phonological recoding in reading for meaning by adults and children. Journal of Experimental Psychology: Learning, Memory, & Cognition, 14,387-397. Conrad, C. (1974). Context effects in sentence comprehension: A study of the subjective lexicon. Memory & Cognition, 2 , 130-138. Cramer, P. (1970). A study of homographs. New York: Academic Press. Doctor, E. A., & Coltheart, M. (1980). Children’s use of phonological encoding when reading for meaning. Memory & Cognition, 8,195-209. Gernsbacher, M. A. (1989). Mechanisms that improve referential access. Cognition, 32, 99-156. Gernsbacher, M. A. (in press-a). Cognitive processes and mechanisms in language comprehension: The structure building framework. In G. H. Bower (Ed.), The psychclogy of learning and motivation. New York: Academic Press. Gernsbacher, M. A. (in press-b). Language comprehension as structure building. Hillsdale, NJ: Erlbaum. Gernsbacher, M. A., & Faust, M. (in press). The mechanism of suppression: A component of several comprehension skills. Journal of Experimental Psychology: Learning, Memory, and Cognition. Gernsbacher, M. A., & Jeschcniak, J. D. (1990). Cataphoric devices in spoken discourse. Manuscript submitted for publication. Gernsbacher, M. A., & Shroyer, S. (1989). The cataphoric use of the indefinite this in spoken narratives. Memory & Cognition, 17, 536-540. Gernsbacher, M. A., & Varner, K. R. (1988). The mulri-media Comprehension batrery (No. 88-04). Institute of Cognitive and Decision Sciences, University of Oregon, Eugene, OR. Gernsbacher, M. A., Varner, K. R., & Faust, M. (1990). Investigating differences in general comprehension skill. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16,430-445. Glucksberg, S., Kreuz, R. J., & Rho, S. H. (1986). Context can constrain lexical access: Implications for models of language comprehension. Journal of Experimenral Psychology: Learning, Memory, and Cognition, 12, 323335. Gough, P. B. (1971). (Almost a decade of) Experimental psycholinguistics. In W. 0. Dingwall (Ed.), A survey of linguistic science . College Park, MA: Kausler, D. H., & Kollasch, S. F. (1970). Word associations to homographs. Journal of Verbal Learning and Verbal Behavior, 9,444-449. Kintsch, W. (1988). The role of knowledge in discourse comprehension: A construction-integration model. Psychological Review, 95, 163- 182. Kintsch, W., & Mross, E. F. (1985). Context effects in word identification. Journal of Memory and Language, 24, 336-349.

The Role of Suppression in Sentence Comprehension

127

Lucas, M. (1987). Frequency effects on the processing of ambiguous words in sentence context. Language and Speech, 30,2546. Marslen-Wilson, W., l)der, L. K., & Seidenberg, M. (1978). Sentence processing and the clause boundary. In W. J. M.Levelt, & G. B. Flores d’Arcais (Ed.), Studies in the perception of language (pp. 219-246). London: Wiley. McClelland, J. L.. & Kawamoto, A. H. (1986). Mechanisms of sentence processing: Assigning roles to constituents of sentences. In J. L. McClelland, & D. E. Rumelhart (Ed.), Parallel distributed processing: Explorations in the microstructure of cognition . Cambridge, MA: MIT Press. Nelson, D., McEvoy, C. L., Walling, J. R., & Wheeler, J. W. (1980). The University of South Florida homograph norms. Behavior Research Methods & Instrumentation, 12, 16-37. Norris, D. (1986). Word recognition: Context effects without priming. Cognition, 22, 93-136. Perfetti, C. A., & Roth, S. (1981). Some of the interactive processes in reading and their role in reading skill. In A. M. Lesgold, & C. A. Perfetti (Ed.), Interactive processes in reading (pp. 269-297). Hillsdale, NJ: Lawrence Erlbaum Associates. Posner, M. I. ( 1 978). Chronometric explorations of mind. Hillsdale. NJ: Erlbaum. Rosson, M. B. (1985). The interaction of pronunciation rules and lexical representations in reading aloud. Memory & Cognition, 13,90-99. Seidenberg, M. S., Tanenhaus. M. K., Leiman, J. M., & Bienkowski, M. (1982). Automatic access of the meanings of ambiguous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14, 489-537. Smith, M. C., & McGee, L. E. (1980). Tracing the time course of picture-word processing. Journal of Experimental Psychology: General, 109, 373-392. Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18,645-659. Tabossi, P. (1988). Accessing lexical ambiguity in different types of sentential contexts. Journal of Memory and Language, 27,324-340. Tabossi, P., Colombo, L., & Job, R. (1987). Accessing lexical ambiguity: Effects of context and dominance. Psychological Research, 49, 161-167. Tanenhaus, M. K., Leiman, J. M., & Seidenberg. M. S. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18,427-440. Till, R . E., Mross, E. F., & Kintsch, W. (1988). Time course of priming for associate and inference words in a discourse context. Memory & Cognition, 16, 283-299. Tipper, S. P., & Driver, J. (1988). Negative priming between pictures and words in a selective attention task: Evidence for semantic processing of ignored

128

M . A . Gernsbacher and M . Faust

stimuli. Memory & Cognition, 16.64-70. van Orden, G. C. (1987). A rows is a rose: Spelling, sound, and reading. Memory & Cognition, IS, 181-198. van Orden, G. C., Johnston, J. C., & Hale, B. L. (1988). Word identification in reading proceeds from spelling to sound to meaning. Journal of Experimental Psychology: Learning, Memory, & Cognilion, 14, 371-386. van Petten, C., & Kutas, M. (1987). Ambiguous words in context: An eventrelated potential analysis of the time course of meaning activation. Journal of Memory and Language, 26, 188-208. Waltz, D. L., & Pollack, J. B. (1985). Massively parallel parsing: A strongly interactive model of natural language interpretation. Cognirive Science, 9, 51-74. Williams, J. N. (1988). Constraints upon semantic activation during sentence comprehension. Language and Cognitive Processes, 3, 165-206.

Understanding Word and Scnlcnce G.B. Simpson (Editor) 0 Elsevier Science Publishers 13.V. (R’orlh-Holland), 1991

Chapter 6 Electrophysiological Evidence for the Flexibility of Lexical Processing Cyma Van Pelten and Marta Kutas University of California, San Diego La Jolla, California U.S.A. The frequency of a word’s occurrence in common usage and the relationship of a word to a prior word have proven to be two of the most powerful determinants of performance in experimental studies of word recognition. In reaction time tasks such as pronunciation or lexical decision, subjects typically respond more slowly to rare than to common words; with tachistoscopic presentation, subjects require longer exposure durations to report rare words accurately (e.g., Rubenstein, Garfield, & Millikan, 1970; Solomon & Howes, 1951). Similarly, the results of a number of tasks, including pronunciation, lexical decision, verbal report of briefly presented or visually degraded words, naming the color of ink in which a word is printed, and monitoring for a target word, have led to the conclusion that processing of a single word is facilitated by the prior occurrence of a related word (Becker & Killion, 1977; Massaro, Jones, Lipscomb, & Scholz, 1979; Meyer & Schvaneveldt, 1971; Rouse & Verinis. 1962; Warren, 1974, 1977). Both of these findings can be explained by postulating a fairly simple self-contained lexicon wherein 1) commonly accessed enuies are either closer to some threshold activation level (Morton, 1969) or have priority in a search procedure (Bradley & Forster, 1987) and 2) some entries are strongly linked such that they can influence each other via some type of “spreading activation.” The stimulus pairs used in these receptive language task have typically been judged as related via production norms (i.e., people tend to produce one when presented with the other). The pairs are often members of the same category of objects, or antonyms. It is unlikely that such pairs will often co-occur in the same sentence. When they do, they will comprise only a small subset of the words in the sentence. A priming effect between words that are semantically similar but not associatively related (e.g., “prince - boy”) appears to be less robust; it has been observed to be either very small or dependent on the use of a lexical decision task (Fischler, 1977a; Huttenlocher & Kubicek, 1983; Lupker,

130

C. Van Petten and M. Kutas

1984; Seidenberg,Waters, Langer, & Sanders, 1984; Warren, 1977). The extension from single word to sentence contexts is thus not obvious. Moreover, it can be argued that single word and sentence contexts must affect subsequent words in a fundamentally different way because lexical context effects can take place within a lexicon, whereas sentence-level processes, by definition, involve novel combinations of words that cannot be pre-stored in a lexicon. There has been some doubt as to whether sentence-level semantic context effects occur in situations which approximate fluent reading, namely those wherein skilled readers quickly process words which are not highly predictable, not strongly associa!ed to previous words, and not visually degraded (see Fischler & Bloom, 1979; Forster, 1981; Henderson, 1982, pp. 351-353; Mitchell & Green, 1978; Perfetti, Goldman, & Hogaboam, 1979). Of course, this position is not universally accepted and a number of recent studies clearly demonstrate the influence of sentence or sentence-like context (O’Seaghdha, 1989; Sanocki, Goldman, Waltz, Cook, Epstein, & Oden, 1985; Ratcliff, 1987; Simpson, Peterson, Casteel, & Burgess, 1989). Results from our lab have consistently shown a pervasive influence of sentence-level context as well; these will be presented in a later section of the present chapter. Because sentence-level context effects are seen in the same experimental tasks that yield associative context and frequency effects, the question arises as to which, if any, of these effects originates in the lexicon. This leads in turn to the more general question concerning the constituents of a “lexical entry”: abstract orthographic and/or phonemic information only (i.e. only the physical form of a word, perhaps with a function that allows some mapping between modalities), the syntactic category of the word, some basic semantic information in addition, or all of the above plus a complete encyclopedic listing? If we were to grant that the lexicon contained everything a person knows about words and their referents, the original conception of the lexicon as only one among several componcnts in the language processing system would cease to be thcoretically functional. Rather than postulate some division between operations “in” the lexicon and “post-lexical” processes, perhaps we should retreat to the most general issue of all, namely the sequence and organization of mental operations that yield the intended meaning of a sentence. In the research to be described, we have focused primarily on only one aspect of this issue - how and when the specific connotation of a word in a sentence is singled out or constructed by the reader. Even a brief intuitive analysis will suggest that the potential meaning of a single word is broader and more diverse than that used in any one sentence, so that perceiving the author’s intended meaning indeed poses a major task for the reader. The semantic attributes of a word may be placed on a continuum ranging from the most basic, those which are frequently used and held in common by most users of the language, to those which are more encylopedic in that they are less often pertinent and perhaps belong only to the vocabulary of the spe-

Electrophysiology and Lexical Processing

131

cialist. For the word “bird,” for example, “typically feathered” and “typically flies” are attributes which are often relevant to the interpretation of a sentence containing the word, and most English speakers will have little trouble accessing these attributes as needed. The facts that birds typically have hollow bones and high metabolic rates will less often be required to understand a sentence using the word “bird” and these attributes may be known by only a subset of English speakers. The basic and encyclopedic features of “bird” are not in conflict with one another; indeed there are logical entailments between them because having a high metabolic rate and hollow bones contribute to the ability to fly. The encylopedic features of the specialist’s vocabulary simply reflect more extensive knowledge of the referent (see Langacker, 1987). In contrast, other words may possess semantic attributes which have no necessary relationship, despite the fact that they may often be conflated in usage. To borrow Lakoff’s (1987) example, “mother” may often imply a genetic, a birth, and a nurturance relationship all at once, but a single relationship can be picked out by the appropriate context as in “surrogate mother” (implying only the birth relationship) or “adoptive mother” (implying only the nurturance relationship). Unlike the basic-encyclopedic continuum, these different attributes of “mother” exist at the same level of detail, but may be more or less compatible with the intended meaning of a particular sentence. A third manner in which the potential senses of a word may be related seems to be distinct from the previous two in that the different senses are always incompatible with one another. One may use either the spatial or temporal sense of “over.” but not at the same time: a sentence context clearly selects one or the other (see Brugman & Lakoff, 1988 for an analysis of how even the spatial sense of ‘‘over” contains several distinct meanings or Lindner, 1981 for an analysis of the polysemy of “up” and “out”). These two senses of “over,” however, do seem to be closely related to one another. In the most extreme case of polysemy even this family resemblance is lacking. For unsystematic homographs such as “spoke” (past tense of “speak”) and “spoke” (“part of a wheel”) it is merely an accident of linguistic change that the same physical form has come to be associated with completely different meanings. Like the lexical entry for “bird,” both meanings of “spoke” may contain core and encyclopedic attributes, but there are two non-overlapping sets of core attributes, only one of which will be appropriate in any given sentence. In the last ten years or so, one of the most influential views of language comprehension has been that the initial recognition of a word’s meaning is context-invariant: while isolated word meanings may be inherently ambiguous, polysemous,or vague, the responsibility of selecting among the multiplicity of retrieved meanings for those that contribute to the ongoing discourse model is left to subsequent processes. A hierarchical processing sequence has prevailed as a means of explaining how the reader arrives at the final interpretation of a word. In broad outline, the work of a number of authors has suggested some-

C.Van Petten and M. Kutas

132 thing like the following sequence of processes:

1) an orthographic (or perhaps phonemic) access code is derived from the visual input; 2) the code is used to tap into the mental lexicon and find a matching word;

3) the lexical entry for the word is accessed such that all of its basic core meanings are activated; 4) the appropriate core meaning is selected based on prior context; 5 ) inferential processes combined with knowledge of the world and the speaker’s intentions are used to elaborate the relevant encyclopedic attributes of the word’s meaning and integrate these with previous words to establish the meaning of the sentence.

In this scheme, the first three stages take place within the lexicon, while subsequent steps are carried out by distinct higher-level integrative mechanisms. Evidence for a sequential processing arrangement here, as in other domains of cognitive psychology, arises from two major sources. First, the presence or absence of interactions among different experimental effects: An effect which is constant despite the addition of other effects must arise at a stage of processing prior to that which yields the other effects. It has been argued that the influence of word frequency must arise prior to the completion of Stage 3 because it is impervious to lexical or sentential context manipulations (Bradley & Forster, 1987; Forster, 1981a. 1981b; Schuberth & Eimas, 1977; Schuberth, Spoehr, & Lane, 1981). However, this result has been inconsistent in that an attenuation of frequency effects by context effects has been observed in some experiments (Becker 1979; Grosjean & Itzler, 1984; Stanovich & West, 1983). A second source of experimental support for placing contextual factors late in the processing sequence could come from measurements of the temporal onset and duration of different experimental effects. But because most dependent measures are temporally punctate, consisting of a single discrete motor response such as a button press, it has been possible to estimate these temporal factors only indirectly. It has been argued that manipulations of the stimulusonset-asynchrony (SOA) between a stimulus word and a target can be used to limit the amount of time devoted to the stimulus word. Thus with a short SOA, responses to the target should be influenced only by the products of the initial, as-yet-incomplete analysis of the stimulus word. This type of experimental paradigm has been used primarily to contrast different varieties of semantic analyses by varying the nature of the semantic relation between the target word, the immediately preceding stimulus word, and the rest of the prior text. Results have suggested Stages 1 through 3 above can be separated from the “higher

Electrophysiology and Lexical Processing

133

level” processes of Stages 4 and 5. So for instance. most of the plausible inferences supported by a text seem to be drawn relatively late. This follows from the finding that inference words do not show priming with a short SOA which does yield priming for semantic associates of the stimulus word (McKoon & Ratcliff, 1989; Swinney & Osterhout, 1990; Till, Mross, & Kintsch, 1988). To illustrate, following the sentence “Thinking of the amount of garlic in his dinner, the guest asked for a mint.”, a word such as “candy” which is semantically associated to “mint” in isolation would show priming with a short SOA, while the inference probe “breath” would show priming with a longer SOA. This dissociation thus indicates that access to the core meaning of the stimulus word (and consequently priming of its associates) must precede inferential processes. Manipulations of SOA have also led to the more remarkable conclusion that even the initial activation of the stimulus word’s core semantic attributes is indiscriminate and context-invariant. When the prime word possesses mutually incompatible core meanings, as in the case of a homograph, semantic associates of both meanings show priming effects when the interval between the stimulus word and its associates is brief. In the example above, “money” would also show a priming effect if presented a short time after “mint.” With a longer interval, only the associate of the sententially relevant meaning of the homographic word would continue to show a priming effect. This result has been taken as the final bit of evidence for the five-stage sequence outlined above in that it subdivides even core meaning processes into two discrete steps and appears to place the first of these beyond the reach of sentence context (Kintsch & Mross, 1985; Onifer & Swinney, 1985; Seidenberg, Tanenhaus, Leiman, & Bienkowski, 1982; Swinney, 1979;Till et al., 1988; see Tabossi, 1988a. 1988b for different results and conclusions). However, this conclusion is debatable; we will critique the experimental logic underlying it and present some contradictory data shortly. Several experiments conducted in our laboratory have led us to question the hierarchical nature of the processing sequence above, particularly the postulate of a self-contained lexicon which is insensitive to sentence meaning but produces both word frequency and associative priming effects. In this chapter, we review several data sets which suggest that: 1) the process which yields frequency effects for words presented in isolation is neither mandatory nor immune to sentence-level context; 2) the influence of sentence-level context can be as powerful and act as early as that of a single lexically associated word; and 3) sentence context can be used to pick out the appropriate core meaning of an ambiguous word without first passing through an early stage of

C. Van Perren and M.Kuras

134

indiscriminate semantic activation. As all of the empirical evidence to follow consists of event-related brain potential data, some description of the technique and a review of its applicability to psycholinguistic research is in order.

EVENT-RELATED POTENTIALS Electrodes placed on the scalp can be used to record voltage fluctuations known as the electroencephalogram (EEG). It is generally believed that the electrical activity recorded at the scalp is a summation of graded post-synaptic potentials (PSPs) generated by the depolarization and hyperpolarization of brain cells (see Wood & Allison, 1981 for a review of the neurophysiological basis of the EEG or Nunez, 1981 for a treatise on the physics of EEG). At any given moment the observed EEG is likely to reflect the activity of a number of functionally distinct neuronal populations. With the advent of computer averaging some two decades ago, it became possible to obtain an estimate of activity which is time-locked to an arbitrary point, such as the onset of a stimulus. Averaging many epochs of EEG following each of a set of similar stimuli tends to cancel the random background EEG, leaving a record of the evoked or eventrelated potentials (EPs or ERPs) which were synchronized to the time of stimulus presentation. Which stimuli are defined as “similar” depends on the goals of the experiment and is established a priori by the experimenter. The resulting waveform of voltage plotted against post-stimulus time typically includes a series of positive and negative peaks. Much ERP research has focused on the decomposition of these voltage fluctuations into experimentally dissociable “components” which can be linked to a specific physiological and/or cognitive process. Attempts to identify a functionally distinct component may include manipulations of the physical (e.g., size, luminance, pitch, etc.) or psychological (e.g.. task-relevance, meaningfulness, predictability, etc.) atuibutes of the stimuli, or the physiological state of the subject (e.g., drug administration, selecting a population with a particular type of brain damage, etc.). Other factors called upon for component identification are voltage polarity, peak or onset latency and duration, distribution across the scalp, and general waveshape. Knowledge of the neural generator can offer yet another criterion for the identity of a scalp-recorded component, one which is of interest in and of itself. A component which has been closely linked to language processing is the N400 (so called because it is negative-going in polarity and has a typical peak latency of 400 msec post-stimulus onset). To date, evidence from work with comrnissurotimized individuals, depth recordings from epileptic patients with temporarily-implanted electrodes, correlations between scalp recordings and regions of high glucose metabolism, and scalp recordings of evoked magnetic fields have all converged to suggest a left temporal lobe substrate for the N400

Electrophysiology and Lexical Processing

135

(Halgren, in press; Heit, Smith, & Halgren, 1988; Kutas, Hillyard, & Gazzaniga, 1988; McCarihy & Wood, 1984; Schmidt, Arthur, Kutas, George, & Flynn, 1989; Smith, Stapleton, & Halgren, 1986). However, it would be premature to conclude at this point that all N400 effects are produced by the same population of cells. Sensitivity of the N400 to Psycholinguistic Manipulations The N400 was first described in experiments which compared semantically predictable to semantically incongruent sentence completions. Subjects in this experiment, as in most of those here, read (silently) sentences as they were presented one word at a time on a CRT. Incongruous final words elicited a negative wave which was largest over posterior scalp locations and somewhat larger over the right than left hemisphere, whereas congruous words elicited a positive-going wave instead. The first separation between the congruous and incongruous waveforms occurred at about 200 msec after the onset of the visual word; the difference peaked at about 400 msec poststimulus (Kutas & Hillyard, 198Oa, 1980b, 198Oc, 1982; Kulas, Van Pettcn, & Besson, 1988). These first incongruity experiments were conducted in 1978. Since then it has become clear that the positivity for a wholly predictable word is the exceptional case: most words elicit an N400. Its amplitude and latency vary with experimental manipulation. When lists of letter strings have been presented to subjects the following pattern of results has emerged: Unrepeated words which are semantically unrelated to previous words elicit the largest N400; orthographically legal, pronounceable nonwords (pseudowords) also elicit large N400s; and unpronounceable nonwords elicit little or no N400 activity (Bentin, 1987; Bentin, McCarthy & Wood, 1985; Holcomb, 1988; Rugg & Nagy, 1987; Smith & Halgren, 1987).’ If the component were produced only after the meaning of a word had been accessed, there should be no N400 for pseudowords. On the other hand, if the N400 reflected simply the “wrongness” of a letter string, there should be a sizeable N400 for illegal nonwords. The results from the two classes of nonwords thus suggest that the N400 reflects some of the earlier processes in visual word recognition, wherein illegal nonwords can be quickly rejected but pseudowords require some additional processing to determine that they are not, in fact, words. The frequency of a word’s usage also modulates the amplitude of the N400 elicited by words presented in lists. Whether the task is lexical decision or detection of repeated words, rare words elicit larger N400s than do common words (Rugg, in press; Smith & Halgrcn, 1987). Finally, the amplitude of the N400 elicited by words in lists is reduced if the same word, or a semantically associated word occurred earlier in the list (Bentin et at., 1985; Holcomb, 1988; Kutas, 1985; Rugg, Furda, & Lorist, 1988). A similar reduction of N400 ampli-

136

C. Van Petten and M.Kurus

tude is apparent for words occurring in repeated sets of sentences, (Besson, Kutas, & Van Petten, in press) and for words which are reused in a text although they occur in different sentences (Van Petten, Kutas, Kluender, & McIsaac, unpublished observations). Negative results are also important in establishing that the N400 component is specific enough to be useful as a measure of language processing. There has been some controversy among ERP researchers as to the relationship between the N400 and another ERP component, the N2, and whether or not the N400 should be regarded as “language-specific”(see Kutas & Van Petten, 1988; Ritter, Ford, Gaillard, Harter, Kutas, Naatanen, Polich, Renault, & Rohrbaugh, 1984; Stuss, Picton, & Cerri, 1986; Stuss. Sarazin, Leech, & Picton, 1983 for some discussion of this issue). However, it is clear that N400s are not universally elicited by any stimulus which fails to fit into a previously established context. So, for instance, ending a sentence with a word which is semantically congruous but of a larger typeface than the preceding sentence fragment, or presenting a slide of a complex abstract drawing rather than a word, does not yield an N400 component but rather a late positive component of the P300 family (Kutas & Hillyard, 1980a, 1984b). Similarly, altering a note in a wellknown melody, or interrupting a progression of geometric figures which have been ordered by increasing size, do not produce N400s (Besson & Macar, 1987). In experiments using sentences, the amplitude of the N400 elicited by the final word is sensitive not only to whether the word is, roughly, congruous or incongruous with the preceding fragment, but how congruous the terminal word is. An a priori metric of the amount of semantic constraint imposed on a terminal word by the preceding fragment can be obtained via the off-line technique of cloze probability, e.g., what proportion of subjects will fill in a particular word as being the most likely completion of a sentence fragment (Taylor, 1953). Cloze probability proportions and N400 amplitude have been shown to be inversely correlated at a level above 90%. It is important to note, however, the subtle distinction between the cloze probability of a terminal word and the contextual constraint of the sentence fragment per se. For example, the sentence fragment “The bill was due at the end of the ...” is of high contextual constraint in that most people will fill in “month” while “He was soothed by the gentle ...” is of low contextual constraint because there are a number of acceptable endings, no one of which is clearly preferred over the others (Bloom & Fischler, 1980). But both fragments can be completed by words of equal (low) cloze probability as in “The bill was due at the end of the hour.” and “He was, soothed by the gentle wind.” The results of experiments which crossed several levels of contextual constraint with several levels of cloze probability showed that the N400 was correlated with the cloze probability of the final word but generally independent of the contextual constraint of the preceding sentence fragment (Kutas & Hillyard, 1984a; Kutas, Lindamood, & Hillyard, 1984). This

Electrophysiology and Lexical Processing

137

result was critical in establishing that N400 amplitude does not index the violation of previously established expectancies for a particular word which was not presented, but rather is sensitive to the degree to which the sentence fragment prepared the way for the word which actually followed. Note however that in the absence of an explicit attempt to dissociate cloze probability and contextual constraint, the two factors are generally correlated. In what follows we will use the term “contextual constraint” in reference to this more typical situation?

Sentence Context and Word Frequency We have conducted a number of experiments wherein subjects’ primary task was to read a number of unrelated sentences. Across experiments, we have assigned a variety of secondary tasks in order to ensure subjects’ continued alertness. These have included detecting an occasional repeated sentence, indicating whether or not a subsequent probe word had occurred in the preceding sentence, and indicating whether or not a particular letter of the alphabet occurred in a single word presented subsequent to each sentence. These tasks all were constructed in such a way that subjects were neither required to make overt responses, while reading nor, with the exception of the repeated sentence task (wherein the repeated sentences were excluded from analysis), to recognize task-related stimuli while reading (in the recognition probe and letter detection tasks, subjects did not know which word or which letter would be their target until some time after the completion of the sentence). We have seen little variability in the pattern of ERP results dependent on which secondary task was used, and thus believe that the data reflect general mechanisms of word recognition and sentence comprehension rather than task-specific factor^.^ The cloze probability results reviewed above suggest that the amplitude of the N400 is a sensitive index of the semantic constraints imposed by a sentence fragment on the processing of its final word. More recently, we have evaluated whether or not this holds for the intermediate words of sentences as well. With a set of unrelated stimulus sentences, the subjectheader must begin each sentence with no information concerning the topic, but as it progresses he or she should begin building a mental model of the concept expressed by the sentence and have more information available concerning what sorts of words may occur next. Accordingly, we sorted intermediate open class words according to sentence position and observed a decrement in N400 amplitude, with increasing position (Kutas et al., 1988). The linear decrease in N400 amplitude has been observed in an experiment in which all of the stimulus materials were normal congruent sentences, as shown in Figure 1. Moreover, no behavioral task which could have fostered differential levels of preparedness was assigncd to the subjects (Van Petten & Kutas, in press). The sequential position of a word within a sentence, in and of itself, is not likely to be a critical variable in language comprehension. However, word posi-

C.Van Petten and M. Kuta:

138

tion is a simple index of the semantic and structural links which differentiate a sentence from a string of unconnected words. Thus we have interpreted the decrement in N400 amplitude across the course of a sentence as a consequence of the build-up of contextual constraints. A similar word position effect has been reported for word-monitoring times with speech materials (MarslenWilson & Tyler, 1980). The N400 word position effect during silent reading supports the the existence of sentence-level context effects in the visual modality. ix.P.6. i t i i

. I

I

T

2.0 pvl

+

4-++-+ o

800 ms

400

Word ..................... ....... Word - - - - - - - - Word Word -.-.-.-.-.-.-.-.-

*

3 5 7 9

Figure 1 . Grand average ERPs elicited by intermediate open-class words in four different sentence positions. Recorded at a central midline scalp site (Cz). Note that while the late negative components elicited at each word position all reach peak amplitude around 400 msec. the differentiation between word positions is apparent as early as 200 msec post stimulus onset (data from Van Petten & Kutas, in press).

Electrophysiology and Lexical Processing

139

In our initial experiment examining word position effects in detail, we also subdivided open class words according to their frequency of usage (Van Petten & Kutas, in press). Six categories of frequency were used, ranging from less than ten to over 450 per million in the Francis and KuCera count (1982). The relatively fine-grained frequency analysis mandated a coarser breakdown of word position in order to have a sufficient number of trials in each frequency-by-position category to form an ERP average with an adequate signalto-noise ratio.‘ Thus we examined the ERPs elicited by the first open class word of each sentence (typically the second or third ordinal position), all of the intermediate open class words, and the final word at each of the six levels of frequency. For the category of “first open class words” we observed a frequency effect consisting of a larger N400 for rare words as shown in Figure 2. Although the data evidenced a gradient of N400 amplitude across the six frequency categories, the only statistically significant difference was between words of less-than versus greater-than 30/million. We have adopted this as our cutoff point for defining high versus low frequency words in subsequent experiments. The more important result of this experiment was that there was no hint of a frequency effect for intermediate or final open class words. We have observed the same interaction between word frequency and the position of a word in its sentence in two subsequent experiments. The subsequent experiments answered some additional questions about the interaction. The first of these was actually a reservation as to whether the absence of a frequency effect for intermediate words should be considered a true interaction. When the ERPs for all intermediate words are averaged together as they were in the first experiment, the resulting N400 is fairly small because the incremental effect of sentence context drives amplitude down with each new word. It may thus have been possible that the absence of an observed frequency effect for intermediate words reflected a “floor effect” in N400 amplitude; that is, the influence of sentence context was so great that it swamped a persisting frequency factor. By examining a greater number of word position categories we were able to rule out this possibility. Figure 3 shows that the influence of word frequency is eliminated quite early in a sentence, although N400 amplitude continues to decline with increasing word position (Van Petten, 1989). We can thus regard the absence of a frequency effect late in a sentence despite its presence earlier as a true interaction between word frequency and sentence context. A second question concerned whether the semantic or syntactic aspects of sentential context, or both, were responsible for attenuating the word frequency effect. In addition to offering some semantic clues about upcoming words, a sentence fragment can also place form class and agreement constraints on words which are legal continuations. We might have suspected from the first experiment that these syntactic constraints are not adequate to suppress the word frequency effect since the first open class words in those stimulus sentences

C.Van Petten and M. Kutas

140

-3.0

-2.0

-1

.o

0.0

M

-2-o

i C

r 0 V 0

I

-1.0

1I

Middle

OC

words

0.0

t S

i n 1 .o

Final O C

-

2.0 3.0 -

-

I

4 .n -

B

&

A

words 7

-

!.

Figure 2 . Mean voltage during the peak latency range of the N400 (350-500 msec post-stimulus onset) relative to a 100 msec pre-stimulus baseline, collapsed across recording site. Word frequency has been broken down into six categories for each of three sentence positions: the first open-class words, other intermediate open-class words, and sentence-final open-class words. The error bars represent the standard error across subjects (data from Van Peuen & Kutas, in press).

141

Electrophysiology and Lexical Processing

Congruent sentences High Frequency

Low Frequency

-1.5

I

-1.0

\ \ \ \ \

-0.5

\ \

T

\

0.0

0.5

1.o \ \ \

‘i

1.5

2.0 1

PV

3

5

7

9

11

13

1

Sentence position

Figure 3. Mean voltage in the N400 peak latency range, plotted against word position for intermediate open-class words in congruent sentences from a different experiment from that shown in Figures 1 and 2 (data from Van Petten. 1989).

142

C. Van Petten and M.Kutas

were preceded by one or two closed class words. In most cases the first open class words were nouns preceded by articles, or main verbs preceded by auxiliaries so that the syntactic constraints were fairly strong. However, we conducted a second experiment to explicitly test whether or not a legal syntactic structure alone would produce either a word position effect or a position by frequency interaction for open class words (Van Petten, 1989; Van Petten & Kutas, submitted). The materials were modelled on those used by MarslenWilson and Tyler in their (1980) word monitoring experiments, consisting of normal congruent sentences, “syntactic” strings which were constructed by replacing each of the open class words in a normal sentence with one of equivalent form class and frequency, and random strings wherein the open class words were again replaced and the closed class words re-ordered within each sentence.5 Thus, the Same closed-class items were present in the Random as in the Syntactic condition, but they were arranged in a non-meaningful way. Examples of each sentence type are shown in Table 1.

Table 1 Sample sentences

Congruent Sentences

The tenants were evicted when they did not pay the last two months rent. It is supposed to bring seven years bad luck to break a mirror. Most new drugs are tested on white lab rats. He was so wrapped up in the past that he never thought about the present. Everything she owned was in a brown paper bag. Syntactic Sentences

He ran the half white car even though he couldn’t name the raise. The necklace pulled the certain cat and borrowed the spoon. He went out of right food and had to go to the black bed. In the wet levels fathers were smoking by congress. They married their uranium in store and cigarettes. Random Sentences

To prided the bury she room she of peanut the had china. Into thumb cable male the effort his into group rowboat. She which had jazz anchor a she to straight couldn’t gun. Was reason and ash the angry with technician. Every opened the gripped they stepping kind steel pine.

Electrophysiology and Lexical Processing

143

The results of this experiment were clear in indicating that semantic processes alone are responsible for both the word position effect and the suppression of the frequency effect. As seen in Figure 4,a frequency effect was apparent for all of the open class words in the Syntactic and Random strings, but only for early words in the Congruent sentences. The usual decrement of N400 amplitude was present across the course of meaningful sentences but absent in both classes of meaningless strings. Of course, there are two alternative interpretations for this pattern of results. On the one hand, the experiment may have demonstrated that a syntactic structure provides very little information concerning upcoming words. On the other hand, we might suppose that such information is present and used by the reader, but that semantic and syntactic information are handled by different brain processors and the N400 measure is insensitive to the use of one class of information. We were able to rule out the second alternative by an examination of the ERPs elicited by the closed class words in the three sentence types. Although much smaller than the open class N400, the amplitude of the N400 to closed class words showed a three-way difference across the conditions. It was smallest when the closed class words were embedded in congruent sentences, of intermediate amplitude in the Syntactic strings, and largest in the Random strings. We thus have an indication that the N400 can serve as an index of the utilization of both sources of information. Several studies have demonstrated that, when confronted with a “fill-inthe-blank,” or cloze, procedure, subjects are much more accurate in predicting function than content words (Aborn, Rubenstein, & Sterling, 1959; Gough. 1983; Smith-Burke & Gingrich, 1979). The present results suggest that this predictive power derives from both semantic and syntactic constraints. In contrast to closed class words, the similarity of the open class ERPs between the Syntactic and Random conditions indicates that syntactic structure alone places few constraints on these words. However, in meaningful sentences, N400 amplitude begins to decline after only a few words, suggesting that even the most incomplete semantic structure is quite powerful (see Tyler & Wessels, 1983; Tyler & Marslen-Wilson, 1986 for similar conclusions based on a different dependent measure). The foregoing experiments established that the word position effect for open class words was a consequence of semantic analyses. Thus the absence of a word frequency effect late in a sentence must likewise be attributed to semantic processes. We can now consider what these results might tell us about the respective loci and timing of frequency and meaning-based effects on lexical processing. Word frequency effects have received such a great deal of research attention over the last thirty years that a number of sources have been proposed for their existence, including each of the stages in the five-stage sequence listed above. We first will describe those proposals which we can eliminate from consideration as explanations of the N400 frequency effect, and then discuss

C. Van Petten and M.Kutas

144

how suppression of this frequency effect by context might be interpreted in various models of word recognition.

Early

------.

Low frequency 2.0

Late

pv +

High frequency

400

800

msec

Figure 4. ERPs elicited by openclass words in thnx sentence types. Early and late refer here to approximately the first and second halves of the sentences, excluding the initial and final words (data from Van Petten. 1989; Van Petten & Kutas. submitted).

Electrophysiology and Lexical Processing

145

Visual encoding

After demonstrating that the letter bigrams and phonemes comprising low frequency words are themselves rare, Landauer and Streeter (1973) put forth the following proposal: “Suppose that some phonemes or graphemes are easier to produce or perceive than others. Then those words composed of good units should have an advantage over those composed of poorer units’’ @. 121). In other words, frequency might have an impact at a perceptual stage prior to any expressly lexical processes, as in the first stage of the five-stage sequence outlined above. In support of this idea, Landauer and Streeter demonstrated that words containing low frequency phonemes were less intelligible in noise than words composed of high frequency phonemes even when the two sets of words were matched for both printed and spoken frequency. However, Gernsbacher (1984) found that bigram frequency did not contribute to word frequency effects for printed words: when factorially combined with experiental familiarity. bigram frequency did not influence lexical decision times whereas familiarity did. Sentence integration A proponent of the sequential processing scheme outlined earlier might argue that some late phase of sentence comprehension, as in Stage 5 (e.g., beyond the lexicon), is the sole source of the the larger N400 to low frequency words. By this view, there may be an “earlier” frequency-sensitive phase of lexical access, but this is simply not indexed by N400 amplitude. Accordingly, the observed interaction between frequency and sentence context would not reveal fundamental processes in word recognition but rather some increased difficulty during the integration of low frequency words with an established context. We find this argument untenable for two reasons. First, other laboratories have reported that low frequency words elicit larger N400s than high when the words are seen in lists for lexical decision, clearly a situation where there is little call for integrative processes (Rugg, in press; Smith & Halgren, 1987). Furthermore, within a sentence context, we have repeatedly observed word frequency effects for the first open class of a sentence and find the argument that it would be more difficult to integrate a low frequency noun than a high frequency one with a preceding article (e.g.. “the squirrel” versus “the rock”) a bit strained.

Task specific stages

Frequency-sensitive mechanisms have also been proposed for the opposite end of the perceptual-motor continuum, near the output stages of the most common tasks employed in word recognition research. The decision stage of the lexical

C. Van Petten and M.Kuras

146

decision task may be prolonged, or in naming tasks it may take longer to assemble the articulatory code for low frequency words (Balota & Chumbley, 1984, 1985; McCann, Besner, & Davelaar, 1988; Theios & Muise, 1977). Attributing the word frequency effect to these task-specific factors in its entirety has been a controversial idea (see Monsell, Doyle, & Haggard, 1989) but one which we need not consider here. The N400 frequency effects we have observed occurred several seconds or minutes before a motor response was required of our subjects.

MODELSOF WORDRECOGNITION WHICH PREDICT ADDITIVITY OF FREQUENCY AND CONTEXT Having ruled out very early and very late operations as accounts of the N400 frequency effect, we can conclude that it arises sometime during the intermediate stages of lexical processing, recognition and meaning access, loosely speaking. In fact, this has also been the standard assumption of most models of word recognition of how frequency influences behavioral responses (for summaries, see Norris, 1986; or Van Petten & Kutas, in press). However, among these models are some which postulate that word frequency and context effects should be additive, either because frequency of usage imposes a stable difference in the recognition criterion for a word (Morton, 1969) or because higher frequency entries are examined first in a serial search process (Bradley & Forster, 1987; Forster, 1981a. 1981b). These models thus cannot accommodate an interaction between frequency and context because frequency is presumed to have had its impact sometime prior to the start of meaning access. We should next consider some models which allow semantic analyses to overlap in time with frequency-sensitive processes. We will see that while this stipulation allows for the possibility of an interaction between frequency and context, it is more difficult to describe the form of the interaction we have observed within extant models, namely, a complete elimination of a frequency effect by sentence context. The Checking Model

In Norris’s Checking model (1986), slower responses to low frequency words are accounted for by a higher recognition criterion much as in the Logogen model. The distinction between the two models is that the recognition criterion can be dynamically altered by reccurring contextual plausibility “checks.” The initial high criterion for low frequency words allows them to accrue more “checks” and consequently more criterion reductions than high frequency words. The model thus predicts a disproportionate influence of context on low frequency words. However, because both high and low frequency

Electrophysiology and Lexical Processing

147

words are subject to such criterion reductions on the basis of context, the criterion for low frequency words can approach but never reach that of high frequency words. This form of interaction is what has been observed in the lexical decision and naming latency data which Norris attempts to explain (Becker, 1979; and Stanovich & West, 1983), but not what we have observed. It is possible that the concept of “criterion” is too bound up with the notion of passing a threshold for the elicitation of an all-or-none behavioral response to be relevant to situations where no motor response is called for. The Cohort Model In the most recent formulation of the Cohort model it is supposed that perceptual recognition processes (e.g., Stage 2 above) proceed more quickly for high frequency words, so that they can have access to semantic processes earlier than low frequency words. As it is a model of spoken-word recognition, it is possible to collect data more directly related to this claim than in the visual modality. In her dissertation work, Zwitserlood (1989, see also Marslen-Wilson, 1987) presented visual probe words for lexical decision at different time points during an auditory word embedded in a sentence. High and low frequency words were interrupted at the same temporal positions, thus yielding data which is more difficult to obtain in the visual modality. At the earlier probe points, the acoustic signal had not yet unambiguously specified the identity of the eventual word. For instance, when her subjects heard “The men mourned the loss of their cap...”, both “captain” (the actual word) and “captive” were equally consistent with the acoustic input. At this time point, visual probe words related to both possible words displayed a priming effect in lexical decision times, but the larger priming effect was for probes related to the higher frequency member of the competing pair. This result does therefore suggest that high frequency words might begin to participate in semantic processes earlier than low frequency words during auditory word recognition. This frequency advantage proved to be transient, at later probe points lexical decision times for associates of high and low frequency words became more equivalent. We have not observed a similar early difference between high and low frequency printed words to suggest that high frequency words gain access to meaning more quickly. Sentence context effects begin at the same time relative to the onset of the word in the ERP. Figure 5 shows a word position effect for high and low frequency words (data from Van Petten, 1989). The inco:sistency between the two patterns of results may reflect a difference between the auditory and visual modalities of language comprehension. Or it may reflect a methodological difference: we have based our conclusions on the measurement of responses to words in sentences, Zwitserlood’s were based on responses to probe words presented subsequent to sentence words. We will return to this methodological point later.

148

C. Van Petten and M.Kutas

Frequency Effects in Connectionist Models Word frequency effects arise naturally in connectionist models because the weights between units are modified with each presentation of a word; more frequently presented words have a greater impact on the strength of the connections between units in different layers (see Seidenberg & McClelland, 1989). In various models, semantic context effects have arisen in two ways: via the abstraction of patterns of word co-occurrences in sentences (Harris, 1989; McClelland, St. John, & Taraban, 1989). or via the explicit encoding of a “word” as a collection of semantic features which may be shared with other words (Kawamoto. 1988). The single (to our knowledge) connectionist model incorporating both word frequency and semantics used the latter scheme for the semantic representation (Sharkey, 1989). In this model, the strength of the connections between graphemic input features and semantic features increased with the frequency of the word, thereby allowing more rapid activation of corresponding semantic features for high frequency inputs. As prior semantic context brought the state of the network closer to that for the next word less adjustment was needed to reach the new target state when the next word was actually encountered. The model thus yielded an interaction between frequency and context because the rate of movement toward the target state (i.e. the effect of frequency) became less important as the distance to be travelled decreased. Nonetheless, this model likewise cannot explain an absolute elimination of the frequency effect by context: unless the network is already in the desired target state when a new word arrives, speed of movement will always be somewhat important.

The Verification Model The sole proposal which we find able to accommodate the form of the frequency by context interaction observed in the ERP data is Becker’s Verification model (1976, 1979, 1980). In this model, the initial sensory analysis of a word yields a set of candidate words which are ordered by frequency. However, even before the word is actually encountered, the reader begins to generate a “semantic set” of candidates based on the prior context: this set is ordered by strength of semantic relationship rather than frequency of usage. Because readers are presumed to search the “semantic set” first, frequency effects obtain only when the stimulus word is not included in this set (i.e., not related to the previous context) and the reader is forced to check the “sensory set.” Thus, given appropriate context, no frequency effects would be observed. While we appreciate the priority of semantics offered by this model, the serial nature of the search process leads to some additional predictions which we find implausible. If the “semantic set” proves to be larger than the sensory set because prior context

Electrophysiology and Lexical Processing

149

Sentence position effect

Fkontal

T

T R &

Temporal

Occipital

----

Low frequency

High frequency

1.0 p v ( , 0

I

400

,

I

800 ms

Figure 5 . ERP difference waves formed by subtracting the responses to open-class words occurring late in congruent sentences from those occumng early. Data recorded from left hemisphere electrode sites is shown in the left-hand column, midline sites in the middle column, right hemisphere sites in the right-hand column. Note that the early (100 to 200 msec post-stimulus) negative wave seen at temporal and occipital sites is not part of the N400 effect. Rather, the difference wave reflects a refractory effect in this visual sensory component; amplitude of the "N180" falls off rapidly during a serial train of stimuli (see Hillyard, Munte. & Neville, 1985). Data from Van Petten (1989).

150

C . Van Petten and M.Kutas

Table 2

Sample Sentences ~~~

Congruent Associated When the MOON is full it is hard to see many STARS or the Milky Way. There were advantages to living in a CITY but Martha moved to a small town for the peace and quiet. After studying the map she realized they should have turned LEFT instead of RIGHT at the light. She was glad she had brought a BOOK since there was nothing to READ in the waiting room.

Congruent Unassociated When the INSURANCE investigators found out that he’d been drinking they REFUSED to pay the claim. The biologist went to the desert every WEEK to collcct a particular SPECIES of lizard that he hopcd to study. The union officials were womcd about the long term health HAZARDS of breathing CHEMICAL fumes every day. She picked up a wallet on the STREET and was honest enough to TRY to locate the owner.

Anomalous Associated When the MOON is rusted it is available to buy Many STARS or the Santa Ana. There was jewelry to drumming in a CITY but Martha turned to a grey TOWN for the lizard and scones. After fixing the movie she found they should have killed LEFT instead of RIGHT at the pot. She was glad she had waved a BOOK since thcre was everyone to READ in the security child.

Anomalous Unassociated When the INSURANCE supplics cxplaincd that he’d been complaining they REFUSED to speak the keys. The shirt went to thc gun every WEEK to keep a good SPECIES of fumes that it hired to see. She scrambled up an official black on the STREET and was deep enough to TRY to ring the glue. The star hair was worried about the bared hard drinking HAZARDS of signing CHEMICAL boxes evcry town. Note: The critical pairs of words are shown capitalized, although the subjects saw them in normal

typeface.

Electrophysiology and Lexical Processing

151

only loosely constrains the current word, preferential search through this set would result in slower word recognition than if there had been no context whatsoever. As sentence fragments are rarely predictive of a single word or even a handful of words, it seems as if the Verification model would rarely stipulate a beneficial role for sentence context. After reviewing a number of existing models of word recognition, we conclude that none provide an accurate account of the form of the context by frequency interaction we have observed in several experiments. In large part, this is due to the fact that each model presupposes that frequency and context act via different mechanisms. Even in the case where a formally identical mechanism is proposed to account for both (i.e., Norris’s recognition criterion), its application would entail a different time course for the two: frequency is presumed to have pre-set the criterion long before the currently processed word is encountered, whereas context can alter the criterion only after the word’s presentation. Although we do not have a new model of word recognition to contrast to those above, we would like to call attention to one possible locus for the word frequency effect that has been relatively overlooked. Perhaps semantic analysis is the source of frequency effects. Where a semantically-based frequency effect has been considered, it has been restricted to quantitative dimensions such as concreteness or number of meanings, neither of which seems to contribute to frequency effects (Gernsbacher, 1984; see, however, Schreuder & Flores d’Arcais, 1989 for a discussion of the ease of accessing different types of semantic attributes). However, many low frequency words are little used because they possess semantic attributes that are not often called for. If the rarity of a word’s semantic features rather than its physical form were responsible for frequency effects, we would expect the frequency effect to disappear with contexts that increased the predictability of these uncommon features. This speculation could be tested by comparing responses to low frequency words which possess high frequency synonyms to those which do not. If rarity of meaning is the basis of word frequency effects, then semantically isolated words should result in slower response times and larger N400s than words whose meanings overlap with more commonly used words.

LEXICAL VERSUS SENTENTIAL CONTEXTS Lexical priming has often been assigned a different status from sentential context effects. A common theoretical stance views lexical-associative priming as a reflcction of pre-stored connections within an autonomous lexicon and sentence-level context effects as a consequence of subsequent comprehension processes that act on the output of the lexicon (Fodor, 1983; Forster, 1981b; Kintsch, 1988; Seidenberg et al., 1982; Swinney, 1979). Recently, we have completed an experiment which was designed to compare these two types of

152

C. Van Petten and M. Kutas

context effects, taking advantage of the temporally continuous dependent measure offered by the ERP as a way of evaluating their relative time courses (Van Petten, 1989). The materials for this experiment consisted of four types of sentences, each containing a critical pair of words. The same associated pairs of words (such as SALT-PEPPER) were embedded in both congruent and syntactically legal but semantically anomalous sentences (“Congruent Associated” and “Anomalous Associated” respectively). Likewise. pairs of words which were not particularly related to one another outside of a sentence context also were embedded in both congruent and anomalous sentences (“Congruent Unassociated” and “Anomalous Unassociated”). The overall lengths of the sentences, the positions of the two members of a critical pair in their sentences and the number and lexical class (open or closed) of the words intervening between the two members of the pair were matched across the four sentence types. Examples of the four sentence types are shown in Table 2. Outside of the critical associated pairs, an attempt was made to avoid including sentence words which seemed clearly associated to one anotherS6With this design, the second word of a pair could thus benefit from 1) both sentential and lexical context in the Congruent Associated condition, 2) lexical context alone in the Anomalous Associated condition, 3) sentential context alone in the Congruent Associated condition, or 4)neither in the Anomalous Unassociated condition. The ERPs elicited by the critical pairs are shown in Figure 6. As expected, a reduction in N400 amplitude from the first to the second word of a pair was observed in all but one of the conditions, the Anomalous Unassociated. The overall word position effect from this experiment is shown in Figure 3. We have taken this more general word position effect as a demonstration of the influence of sentence context on N400 amplitude, and view the decrement in amplitude from the first to the second members of the Congruent Unassociated pairs as a subset of the general trend across a sentence. The associative priming effects were not surprising either, given that a number of laboratories have demonstrated modulations of N400 amplitude in word pair studies (see citations earlier). However, the present experiment allows a comparison of the two types of context effects when obtained in similar circumstances from the same group of subjects, Our analysis will thus focus on the timecourse of the two context effects, taking the difference between the first and second words of the Anomalous Associated pairs as a purely lexical context effect, and the difference between the first and second words of the Congruent Unassociated to be a purely sentential effect.’ The time course of a priming or context effect can logically be divided into two parameters: rise time and decay rate. In behavioral experiments, both are usually inferred from the effects of varying the interval between a context and a target word. Rise time is described by the shortest context-target interval that affords priming, and decay rate by the longest interval which sustains

153

Electrophysiology and Lexical Processing

significant priming effects. However, both rise time and decay rate can be further subdivided. For rise time, one of these is the length of time it takes a subject to process previous words to the point where they become a potential source of context. It is this component which can be estimated by manipulating the SOA between a current word and prior context. A second component of rise time reflects the exact moment at which this processed information is actually

Congruent Associated. The HOT water tank sprang a leak so they had to wash everything in COLD water.

Congruent Unassociated. The VETERANS were suing the United States government because they had been exposed to TOXIC chemicals.

Anomalous Associated. The HOT visitor plastic reached a leak so they had t o intake everything in COLD hours.

Anomalous Unassociated. The VETERANS were bolting the Nancy Jane expense because they had been started with TOXIC owners.

- - - - - - - First word

2.0

Second word F i g w c 6 . ERPs to the critical pairs of words (shown capitalized) in each of the four sentence types

used to contrast sentential and lexical contexts.

154

C. Van Petten and M.Kutas

applied to the analysis of the current word. Speech researchers have been scnsitive to the distinction between these different components of rise time because the sequential nature of the acoustic signal allows measurements to be taken at different time points during the input signal (see Zwitserlood, 1989 or Grosjean, 1980). Because printed words are not spread out across time, these two rise time parameters have often been conflated in studies of visual word recognition. There is little question that it takes longer to extract meaning from a sentence fragment than a single word. But this fact has little bearing on the issue of whether or not word recognition draws on all potentially available scinantic constraints, or if some information has privileged access. The decay rate parameter can similarly be subdivided. On the one hand, we can estimate the temporal extent of a context effect during the processing of a single word. On the other hand, we can wonder how many words downstream from the relevant context are still susceptible to its influence. In the present experiment we were able to evaluate three of these four temporal factors.

Rise Time A comparison of the onset latencies of the lexical and sentential context effects offers a window onto the second component of “rise time,” namely when a context effect is evident during the processing of a given word. As these onset times were indistinguishable, we found no evidence for the view that there is a distinct stage in the processing of a word which can be influenced by associative links, but not by sentence-level context.s The first component of “rise time.” the time it takes for a sentence context to become available (as opposed to the time it takes to be used, once available) can only be evaluated by varying the rate at which sentence words are presented. Throughout this experiment, words were presented at a single rate of one word every 600 msec. However, we are currently conducting an experiment wherein the same stimulus materials are presented at a rate of one word every 300 msec; comparison between the two experiments will allow for an evaluation of this temporal factor. Given that this faster rate approximates that of fluent reading, similar results will provide strong support not only for the theoretical position that associative and sentential context can have the same timecourse, but also for the practical position that readers are ablc to avail themselves of sentence context during the processing of individual words in more natural situations.

Decay Rate For sentential context effects, we have already seen that there is essentially no limit to the number of words downstream which may be influenced by prior context. The general word position effect indicates that sentential context is incremental, each new word benefitting from the entire previous fragment. Lexi-

Electrophysiology and Lexical Processing

155

cal context effects have been thought to be more prone to disruption by unrelated words or long delays between the two members of the associated pair. This seems to be true when words are presented in lists (Foss, 1982; Gough, Alford, & Holley-Wilcox, 1981; Neely, 1977; Warren, 1972). but not when they occur in congruent sentences (Carrol & Slowiaczek, 1986; Foss, 1982; Simpson et al., 1989). The anomalous sentences used here fall somewhere between the two extremes in being more structured than a random word list, but less so than a congruent sentence. We divided the associated pairs occurring in anomalous sentences into two equal-sized classes of “near” and “far” pairs. The members of “near” pairs occurred either immediately adjacent to one another, or had one word intervening. “Far” pairs had more than one word intervening, with an average of 4.8 words. No difference in the magnitude of the lexical priming effect was observed as a function of number of intervening words. Together with the general finding that associative priming effects do fall off with distance in random lists, the present null result is consistent with the suggestion of O’Seaghdha (1989) that a syntactic structure licenses the more effective use of lexical associations. It is also consistent with our previous suggestion that subjects are probably more actively engaged or attentive when reading sentencelike materials than when reading random word strings (Van Petten, 1989; Van Petten & Kutas, submitted). Evaluation of the other component of decay time yielded interesting results in that we observed a difference between lexical and sentential context effects. While both effects had the same onset latency at about 300 msec relative to stimulus presentation. the sentential priming effect proved to have a longer duration as is apparent in Figure 6. The sentence context effect persisted late in the recording epoch (500 to 700 msec post-stimulus) while there was no lexical effect in this latency window. A broad duration sentential effect was observed for the congruent sentences both with and without lexical associates (see Van Petten, 1989 for a lengthier discussion of this result). We have found across several experiments that all of the N400 effects which are due to sentence context (i.e. comparisons of responses to words occurring early versus late in congruent sentences, words in congruent versus anomalous sentences, congruent versus incongruent words in otherwise normal sentences) have a broader temporal duration than N400 effects arising from lexical-associative priming or differences in word frequency (see Figures 1, 4 and 6). This durational difference may indicate that while sentence context word processing in the same time frame as these other factors, it also demands more extended processing of single words. Within the present comparison of associative and sentence contexts, it seems likely that comprehension of a sentence encourages a more detailed semantic analysis of each word than does lexical context alone. In isolation, or in an anomalous sentence, one might analyze the rclationship between “salt” and “pepper” rather quickly and superficially as “those black and white substances in the twin shakers.” Compare this

156

C. Van Petten and M.Kutas

to the senses of “salt” and “pepper” evoked by one of the congruent sentences used here: “He was trying to cut down on his SALT intake so he used a lot of PEPPER and other spices.” This Sentence suggests not only the white substance in the shaker, but also the fact that it may lead to high blood pressure, and perhaps the fact that salt contains sodium. On reaching “pepper,” the reader is invited to infer that it has not been associated with high blood pressure, can serve as an alternative flavoring agent, etc. Accordingly, the longer duration of sentence context effects in the ERP may reflect prolonged processing of more encyclopedic semantic attributes of words as they are generally needed for sentence comprehension. Such a speculation could be tested by comparing ERPs to the same words embedded in contexts which refer to their core or peripheral senses.

SENTENCE CONTEXT AND LEXICAL AMBIGUITY Above, we speculated that the later portion of the N400 effects we observed in the context of meaningful sentences might reflect the encyclopedic elaboration of single word meanings and/or inferential processes. However, the results also indicated that some aspects of sentence context operate as quickly as lexical-associative context; perhaps these aspects of sentence context aid in focusing on the relevant core semantic attributes of the current word. We began this chapter by outlining a five-stage sequence of events occurring during sentence comprehension, one which we take to be a widely-accepted view, but one which much of our results have been inconsistent with. In this view, sentence context cannot select the relevant semantic attributes of a word until all of its core attributes are generated by a prior lexical processor. This early phase of generating meanings has been held to be mandatory, indiscriminate, and exhaustive (Fodor, 1983). One of the key findings in supporting this view has arisen from the literature on lexical ambiguity. The typical experimental paradigm in this literature consists of embedding a homographic word in a sentence which clearly disambiguates it, and recording responses to target words which follow the homograph. It is usually found that, with a brief SOA between homograph and target, targets related to either meaning of the homograph show a priming effect. With a longer SOA, priming is observed only for the associates of the sententially appropriate sense of the homograph (Kintsch & Mross, 1985; Onifer & Swinney, 1985; Seidenberg et al., 1982; Swinney, 1979; Till et al., 1988). Priming for both appropriate and inappropriate associates is observed with SOAs of less than 200 msec or so, whereas longer intervals between the two words yields selective priming. This finding has been the primary evidence for a contrast between an early stage of word processing wherein all core (at least) meanings are activated (and lexical-associative priming occurs) and a later phase wherein sentence context is used to select the relevant meaning from among the candi-

Electrophysiology and Lexical Processing

157

dates. The result has been so persuasive that even Kintsch, who has otherwise described his most recent model of language comprehension as being one where “word identification is ... deeply embedded into the process of discourse understanding” (1988, pp 173). has included a stage involving random sampling of associated words prior to any sentence integration. When a single experimental finding supports so much theoretical infrastructure, it is reasonable to scrutinize both the data and the interpretation carefully. We believe there are both empirical and logical flaws in the conclusion. On the empirical side, the postulate that the initial access to a word’s meaning is context-invariant should predict not only that priming be found for inappropriate associates of ambiguous words, but that it be equivalent in all respects to the priming for appropriate associates. Thus, greater priming for the appropriate probe at a short SOA would violate the modularity of the activation process and suggest an early influence of the preceding sentence context. This is indeed what has been reported in the majority of ambiguity studies (see Blutner & Sommer, 1988; Oden & Spira, 1983; Onifer & Swinney, 1981; Seidenberg et al., 1982; Simpson, 1984; Till et al.. 1988; Swinney, 1979; Van Petten & Kutas, 1987, Exp. 1). In some cases this difference has been little noted because it did not reach statistical significance. Recently, however, St. John (1988) conducted a meta-analysis in which the results of these different experiments were analyzed jointly and found that the difference between appropriate and inappropriate associates was statistically significant. There are thus good empirical grounds to question the conclusion of context-invariant meaning activation. Likewise, we can question the conceptual underpinnings of the experimental procedure. The logic of the homograph/probe-word paradigm is that, due to spreading activation, a response to the probe can reveal which meanings of the preceding word were activated and in what sequence. It is assumed that the processing of the probe itself is of no consequence. Although they are presented very close together in time, their is an assumption of strict seriality in the processing of the two words? There has been an additional assumption that setting the SOA between the homograph and the probe to a particular interval is equivalent to taking a “snapshot” of the processing state of the homograph at that time point. For instance, based on manipulations of SOA, Till and colleagues have stated that: 1) “associates of a priming word are equally facilitated ... whether or not they are appropriate to the discourse context - at least for the first 300 msec of processing,” 2) “Sense activation is achieved within 400 msec” and 3) “inference words are not activated by the discourse context ... until the prime word has been processed for more than 500 msec” (Till et al., 1988). These time estimates for the processing of prime words were based on reaction times to subsequent probe words, reaction times which exceeded 600 msec. The danger of equating SOA to processing time is most clear if we pursue

C. Van Pelten and M.Kutas

158

the logic to its extreme. Associative priming effects have been observed with an SOA of 16 msec (Simpson gL Burgess, 1985). but we surely don't want to conclude that the meaning of the first word has been accessed within 16 msec, placing the processes of interest somewhere in the retina or the lateral geniculate. With interstimulus intervals as short as this, some processing of the prime clearly occurs after the presentation of the target. In this situation, the priming effect must be due to coincident processing of the two words rather than to a preactivation of the second word by the first.

PRIME-TARGET STIMULUS ONSET ASYNCHRONY 200 SOA

700 SOA A

Filler

1m

4

1 0

1 u L . 500 1000

5 PV J + 1500

5P

+-

1-500

0

500

v

msec

Contextually appropriate targets . . . ....... Contextually inappropriate targets - - - - - Unrelated targets Figure 7. ERPs elicited by sentence-terminal words and subsequent probe words at a central midline scalp site (Cz). "Filler" sentences ended with unambiguous words. In all four plots, the sentenceterminal words were presented at the time indicated by the arrow and probe words at time 0.

Electrophysiology and Lexical Processing

159

The standard interpretation of the ambiguity literature is that readers or listeners simultaneously access all meanings of homographic words upon seeing or hearing them, and this yields the subsequent priming effect for associates of all meanings. However, if there is temporal overlap in the processing of prime and target words in the short SOA conditions of lexical ambiguity experiments. then the usual interpretation would be invalidated. The facilitated response to the inappropriate probe could reflect mutual priming between the homograph and the target; this situation would only arise in the laboratory setting. The reasonableness of this interpretation is suggested by the fact that, independent of the priming condition, reaction times are often slower following short SOAs (by some 160 msec in Exp 2 of Till et al.). This pattern of results would suggest that processing of the prime word is not truncated by presentation of the target, rather the target word is added to the subject’s processing load, allowing for interactive processing of the two current words. We are then presented with an empirical question that must be resolved before the ambiguity results can be interpreted unequivocally: what is the longest separation between two words which will result in temporal overlap? There has been a paucity of data that is directly applicable to this question. We will describe that which exists shortly, but first digress to describe the results of an ERP ambiguity study which underscores our belief that this is a critical question. While the probe word technique itself has some built-in interpretative complications, the most severe of these arise in conjunction with the use of a discrete dependent measure. Some years ago, we adopted the lexical ambiguity probe word technique from the behavioral literature but used the ERP to provide a more detailed picture of the timecourse of context/probe processing (Van Petten & Kutas. 1987). We presented sentences which biased one reading of an ambiguous word, followed by a target word that was either 1) related to the sententially appropriate sense, 2) related to the inappropriate sense, or 3) unrelated. With a 700 msec interval (onset to onset) between ambiguity and target, the appropriate target elicited a smaller N400 than did the unrelated. As seen in Figure 7, this N400 difference was apparent beginning about 300 msec after the onset of the probe word. In contrast, the inappropriate target elicited an N400 which was equivalent to the unrelated target. Given this long interval, no priming was apparent for the inappropriate target. However, with a short SOA between ambiguity and target (200 msec onset-to-onset), we observed a priming effect even for inappropriate targets. The ERP data thus demonstrated the same sensitivity to the temporal interval between ambiguity and target as reaction time data. However, the priming effect observed for inappropriate probes in the short SOA condition was substantially delayed relative to that for the appropriate probe: it was not apparent until some 500 msec after target onset.lo These data therefore did not support the view that both meanings of the ambiguous words were simultaneously activated. In line with the principle of parsimony, we attributed the different time courses of the two priming effects seen in the

160

C . Van Petten and M . Kutas

short SOA condition to the different times at which contextual information was delivered. In one case, the preceding sentence served as context to interpret the ambiguous word thereby producing an early priming effect for the subsequent probe related to this meaning. In the other case, the inappropriate probe itself took on a double role not only as probe but as context for the alternative interpretation of the ambiguous word. The ERP data are conclusive as an indication that the sententially appropriate and inappropriate senses of ambiguous words are not equally weighted. The proposed mechanism for how the inappropriate meaning of the ambiguous word is activated in studies of this type is more speculative. It hinges largely on the extent to which mutual priming between (slightly) temporally separated words is a genuine phenomena. The first report that lexical decision times could be influenced by a prime word presented afier its target was that of Kiger and Glass (1983) who reported speeded RTs at SOAs of less than 130 msec. When a prime has been visually masked, “backward” priming effects on accuracy of reporting the prime (Dark, 1988) or lexical decision times (Briand, den Heyer, & Dannenbring, 1988) have been obtained with SOAs of up to 1 second. However, these three studies used word pair stimuli because they were not designed to investigate the issue of lexical ambiguity. As pointed out by Peterson and Simpson (1989), sentence materials would bear more directly on the results in the ambiguity literature. These investigators found that both lexical decision and naming latency times to unambiguous words were speeded by the presentation of a related word 200 msec later, but only if the words were presented in pairs and not when the prime words occurred at the ends of sentences. Because the ambiguity results have originated from sentence paradigms, Peterson and Simpson thus concluded that “backward” priming could not have accounted for the context-invariant priming observed in those studies. We find some of the issues raised by Peterson and Simpson to be intriguing. For instance, they suggested that backward priming was observed in word pairs but not sentences because the preceding sentence fragment allowed faster processing of the prime thereby allowing less temporal overlap with the subsequent target. Additionally, these authors noted that their data “do not rule out the possibility that backward priming might occur for target words that have, in fact, received some forward priming” (1989, pp. 1028). They do not believe this idea to be inconsistent with the postulate of context-insensitive meaning retrieval since by “forward” priming they mean forward priming of the target word’s meaning. However, the possibility that “backward” priming could be supported by “forward” priming could be at odds with the context-insensitive view if we consider other varieties of forward priming. This idea will become more clear as we describe some of our experiments in progress. Like Peterson and Simpson, we think that the backward priming paradigm needs to be brought closer to the standard ambiguity paradigm if it is to have any relevance. In our attempt to do so without using ambiguous words, we have

Eiectrophysiology and Lexical Processing

161

focused on the defining attribute of homographic words, namely that two distinct meanings are represented by the same orthographic form. When a homograph is used to terminate a sentence, one of its meanings will be congruent with the sentence context and one incongruent although the shared visual form will match the sentence context in both cases. We have tried to incorporate these aspects of lexical ambiguity in an experiment by starting with pairs of words which have quite distinct meanings but are orthographically similar, such as SOUP and SOAP, or BREAD and BROAD. Sentence contexts which suggested one member of the pair were generated, as in “He ate a bowl of hot chicken ...” Subjects were presented with such sentences terminated by the incongruent but visually similar member of the pair (e.g., SOAP) mixed in with normal congruent sentences (e.g., “Water and sunshine help plants grow”), and sentences with incongruent final words (e.g., “He ordered french fries with his cable”). As in the ambiguity paradigm, each sentence was followed by a word which was either related or unrelated to the meaning of the sentence terminal word, with an SOA of 200 msec. The condition including a “visually similar” sentence completion followed by a related word approximates the condition in an ambiguity experiment of a homograph followed by a related word in that the subsequent word is related not to the semantically congruent sentence completion, but instead to another word which shares some visual characteristics of the congruent completion. We recorded naming latencies to the final words of the sentences in this experiment in order to assess the influence of the subsequent word. No backward priming was observed for the final words of either the congruent or the incongruent sentences. However, in the “visually similar” condition, we found that both error rate and reaction time were sensitive to the relationship between the sentence terminal word (the target) and the subsequent word. When the subsequent word was unrelated, subjects often committed the error of pronouncing the word suggested by the sentence fragment (e.g., SOUP) rather than the word actually presented (e.g., SOAP). This type of error was significantly reduced when the sentence was followed 200 msec later by a related word. In addition mean reaction time for correct pronunciations was reduced by the presence of a subsequent related word. Although the backward priming effect on RTs was only 10 msec, it was statistically significant. This compares favorably to the magnitude of priming effects reported for contextually inappropriate probes of ambiguous words in experiments using a naming latency measure and SOAs close to 200 msec. This backward priming effect was about half the magnitude of the forward effect observed when we asked a separate group of subjects to pronounce the subsequent words rather than the final words of the sentences. We are currently pursuing the impact of including subsequent words that are related to the congruent completions. increasing the similarity between the congruent and “visually similar” words, and other manipulations. However, even as it stands, OUT data indicate that it is possible to obtain a backward

162

C. Van Petten and M.Kutas

priming effect in the sentencehaming latency paradigm that has been a mainstay of the ambiguity literature.

THEFIVE-STAGE SEQUENCE RECONSIDERED We began this chapter by outlining a common view of word recognition in which the interpretation of words in sentences is inevitably delayed by passage through a lexicon which is both ordered by the frequency of usage and unresponsive to the weight of sentence context. We set out to test several of the predictions of this view, primarily with the ERP experiments summarized above. The results of these have persuaded us that much of the interpretation of words in sentences is fairly immediate. In retrospect, the concept of the lexicon as a passive store of information which is acted upon by subsequent processes should have been suspect for a variety of logical considerations. Firstly, the distinction between memory and process is clearly a metaphor drawn from computer science which should not be applied seriously to biological systems, where information processing transactions are conducted by the same neural elements which are modified by experience. Secondly, while it was suggest.ed that focusing on processes within the lexicon would make the research questions more tractable (see Fodor, 1983). this has had the result of making thosc “post-lexical” processes required for sentence comprehension a somewhat more vague and intractable issue. Gerrig (1986) has noted that any adequate theory of comprehension should “specify both the information that is derived from the lexicon and the processes that operate on that information. Since trade-offs of process and products lead to the same end result for comprehension, the validity of a theory that considers either aspect alone is indeterminate” (pp 188-189). Autonomous models of lexical processing have focused only on the first half of this equation. The non-trivial nature of the work that remains has been argued by Clark (1983) in his accounts of the ubiquity of “contextual expressions,” or phrases that will not yield their intended meaning via a simple concatenation of their lexical components. Such expressions include a number of common constructions such as 1) compound nouns (“finger cup,” “tea garden”) where the relationship between the two nouns is not specified by the words themselves, 2) indirect or shorthand noun phrases (“one water” taken to mean “one glass of water,” “a Beethoven” taken to mean “symphony written by Beethoven,” or “a talented composer”), 3) possessive nouns (in the appropriate context “John’s dog” could refer to “the dog that attacked John yesterday” or “the dog John is standing is front o f ) , 4) pro-act verbs (“Alice did the lawn” might indicate that she planted it or mowed it or fertilized it, etc). 5 ) eponymous adjectives (“Churchillian” might mean “with a face like Churchill,” “smoking a cigar like Churchill” or “with a speaking style like Churchill”). Clark’s criteria for a contextual expression are that its possible senses are not denumerable, and that

Electrophysiology and Lexical Processing

163

the intended sense depends on coordination between speaker and addressee: “who uttered the expression, where, what he was pointing at, who had just been mentioned in the conversation, what his addressee knew and didn’t know” (1986, pp 301). When contextual expressions are used innovatively, the intended sense of their lexical components will not be discovered in static elements of long term memory. Thus a first-pass activation process like that proposed in Stage 3 of the five-stage sequence would not yield the correct candidate for subsequent selection by context-sensitive processes. A context-invariant activation process would similarly fail to produce the correct sense for a novel metaphor which does not possess a pre-stored lexical representation. Of course, as pointed out by both Gemg (1986) and Clark (1983) both contextual expressions and metaphors do draw on some pre-stored information for their interpretation. In Gemg’s example “The chimney belched forth soiled wisps of cotton,” there is something about the perceptual quality of “cotton” which lends itself to a metaphorical identification with “smoke” much better than, say, “wool.” So while we should not expect a lexical retrieval system to be able to deliver a fully specified, contextually appropriate definition of every word i t encounters, we would like it to yield some relevant bits of information which can then be amplified and elaborated by continued processing. This sort of flexibility and susceptibility to subsequent processing is what Stage 3 of the original sequence was supposed to provide. We believe that the failure of that original model is due primarily to the rigidity of the initial activation phase; as it has been characterized, failure of the activation phase to find an existing, apt definition would be catastrophic. A more efficient and less failure-prone system would allow prior context to highlight the relevant aspects of a word’s meaning without requiring the retrieval of a fully-specified definition. Acknowledgments

We are grateful to Cathy Harris and Rob Kluender for helpful discussions and comments on a previous version of this paper. During some of the time this research was conducted, Cyma Van Petten was supported by an NSF graduate fellowship, and Marta Kutas by a Research Scientist Development Award (MH00322). The research was supported by a grant from NIH (HD22614). Notes ‘In various studies using different materials, the pseudoword N400 has been either somewhat larger or somewhat smaller than that elicited by real words. It is not yet clear how the amplitude of the pseudoword N400 will compare to the largest possible N400 elicited by real words, namely that to a set of words which are unrepeated, unrelated to previous words, and low in fre-

164

C. Van Petren and M.Kuras

quency of usage. *All of the experiments cited above were conducted in the visual modality. The sentence incongruity results have been replicated in both speech and American Sign Language (Kutas, Neville, & Holcomb, 1987; Karniski, Vanderploeg, Diehl, Lease, 1988). Associative priming in word lists and repetition priming effects have also been obtained in the auditory modality as well as in an ideographic writing system (Holcomb, 1985; Katayama, Teraji, & Yagi, 1989; McCallum, Farmer, & Pocock, 1984; Feldstein, Smith, & Halgren, 1987). ’Another reason for explicitly designing tasks which prevent the subject from making task-related decisions while reading is that a large positive ERP component (the P300) generally appears in any task requiring a binary decision, as in go/no-go tasks or choice RT tasks. The P300 occurs in the same latency range as the N400 which makes it difficult to determine which of the two overlapping components is affected by an experimental manipulation (see Donchin, 1981; Johnson, 1988; Kutas & Hillyard, 1989; Pritchard, 1981). ‘The “noise” in an ERP average is ongoing EEG activity which is unsynchronized to stimulus presentation. Residual EEG activity in an ERP declines with the square root of the number of trials in the average (see Regan, 1989). In practice, the noise level can be evaluated by examining the difference between two averages during the prestimulus baseline. We have found 20 to be a minimally adequate number of trials for most N400 experiments. However, there is an inverse relationship between the size of the experimental effect which one hopes to obtain (i.e.. one hopes to rise above the noise level) and the number of trials required. The difference in N400 amplitude between adjacent word positions may be as small as 1 UV and demands a somewhat larger number of trials. 5Care was taken to maintain verb argument structure in the replacement process. In addition, only “ly” adverbs were replaced; quantifiers such as “some” and “many” were not (see Cowart, 1982). There were two cases where this was difficult to avoid: sentence terminal words, and two-word proper names such as “United States.” However, the critical word pairs never occurred as either the first or last word of a sentence, nor were they proper names, so these factors will not influence the analysis of responses to the critical pairs. ’Note that our analysis could focus on only the second words of each pair, regarding the Anomalous Unassociated as the baseline condition to define both the lexical and sentential effects. This strategy yields similar conclusions concerning the timecourse of the two effects as that pursued here. *It should be noted that because neither the cellular events underlying the generation of the N400 nor those underlying word recognition are known in any detail, we do not make the claim that the ERP provides a real-time record of those events. However, the goal of the present experiment was to compare the latencies of two ERP context effects. Given that the waveshape and scalp distributions of these were the same, we found it parsimonious to assume that they

Electrophysiology and Lexical Processing

165

bore the same temporal relationship to the underlying neurophysiology. 91nteractiveprocessing of words that are close together in time or space has fallen out of several recent network models of word recognition which did not set out to explicitly incorporate this factor. Mozer found that a network trained to identify words presented at varying spatial locations suffered “crosstalk” or letter migration errors when multiple words were presented close together, as do human subjects (McClelland & Mozer, 1986; Mozer, 1987). Masson (1989) has modelled mutual priming between words which share some semantic features. ’OBoth the early and the late priming effects began prior to the average reaction times we obtained in a behavioral version of an experiment using the same stimuli. References Aborn. M., Rubenstein, H.. & Sterling, T.D. (1959). Sources of contextual constraint upon words in sentences. Journal of Experimental Psychology, 57, 171-180. Balota, D.A. & Chumbley, J.I. (1984). Are lexical decisions a good measure of lexical access? The role of word frequency in the neglected decision stage. Journal of Experimental Psychology: Human Perception and Performance, 10, 340-357. Balota, D.A. & Chumbley, J.I. (1985). The locus of word-frequency effects in the pronunciation task: Lexical and/or production frequency? Journal of Memory and Language, 24,89-106. Becker, C.A. (1976). Allocation of attention during visual word recognition. Journal of Experimental Psychology: Human Perception and Performance, 2, 556-566. Becker, C.A. (1979). Semantic context and word frequency effects in visual word recognition. Journal of Experimental Psychology: Human Perceplion and Performance, 5,252-259. Becker, C.A. (1980). Semantic context effects in visual word recognition: An analysis of semantic strategies. Memory & Cognition, 8,493-512. Becker, C.A., & Killion, T.H. (1977). Interaction of visual and cognitive effects in word recognition. Journal of Experimental Psychology: Human Perception and Performance, 3 , 389-401. Bentin, S. (1987). Event-related potentials. semantic processes, and expectancy factors in word recognition. Brain and Language, 31, 308-327. Bentin. S . , McCarthy, G.& Wood, C.C. (1985). Event-related potentials associated with semantic priming. Electroencephalography and Clinical Neurophysiology, 60, 343-355. Besson, M., Kutas, M., and Van Petten, C. (in press). ERP signs of semantic congruity and word repetition in sentences. In C.H.M.Brunia, A.W.K.

C . Van Petten and M . Kutas

166

Gaillard, A. Kok, G. Mulder, M.N. Verbaten @ids.), Proceedings of the ninth international conference on event-related potentials of the brain. Tilburg, Germany: Tilburg University Press. Besson, M., & Macar, F. (1987). An event-related potential analysis of incongruity in music and other non-linguistic contexts. Psychophysiology, 24, 14-25. Bloom, P.A. & Fischler, I. (1980). Completion norms for 329 sentence contexts. Memory & Cognition, 8, 631-642. Blutner, R., & Sommer, R. (1988). Sentence processing and lexical access: The influence of the focus-identifying task. Journal of Memory and Language, 27, 359-367. Bradley, D.C., & Forster, K.I.(1987). A reader’s view of listening. Cognition, 25.

Brugman, C., & Lakoff, G . (1988). Cognitive topology and lexical networks. In S.L. Small, G. Cottrell. and M.K. Tanenhaus (Eds.), Lexical ambiguity resolution (pp 477-508). San Mateo, CA: Morgan Kaufmann. Carroll, P., & Slowiaczek, M.L. (1986). Constraints on semantic priming in reading: A fixation time analysis. Memory & Cognition, 14, 509-522. Clark, H.H. (1983). Making sense of nonce sense. In G.B. Flores d’Arcais and R.J. Jarvella (Eds.), The process of language understanding (pp 297-332). New York: John Wiley and Sons. Cowart, W. (1982). Autonomy and interaction in the language processing system: A reply to Marslen-Wilson and Tyler. Cognition, 12, 109-117. Dark, V.J. (1988). Semantic priming, prime reportability. and retroactive priming are interdependent. Memory & Cognition, 16,299-308. den Heyer, K.,Briand, K.,& Dannenbring, G.L. (1988). Retroactive semantic priming in a lexical decision task. Quarterly Journal of Experimental Psychology, 40A, 341-259. Feldstein, P., Smith, M.E.,& Halgren, E. (1987, June). Cross-modal repetition effects on the N4. Paper presented at the Fourth International Conference on Cognitive Neuroscience, Dourdan, France. Fischler, I. (1977). Associative facilitation without expectancy in a lexical decision task. Journal of Experimental Psychology: Human Perception and Performance, 3 , 18-26. Fischler, I.S., & Bloom, P.A. (1979). Automatic and attentional processes in the effects of sentence contexts on word recognition. Journal of Verbal Learning and Verbal Behavior, 5 , 1-20. Fodor, J.A. (1983). The modularity of mind. Cambridge, MA: MIT Press. Forster, K.I. (1981a). Lexical access and lexical decision: Mechanisms of frequency sensitivity. Journal of Verbal Learning and Verbal Behavior, 22, 24-44.

Forster, K.I.(1981b). Priming and the effects of sentence and lexical contexts on naming time: Evidence for autonomous lexical processing. Quarterly

Electrophysiology and Lexical Processing

167

Journal of Experimental Psychology, 33A, 465-495. Foss, D.J. (1982). A discourse on semantic priming. Cognitive Psychology, 14, 590-607. Francis, W.N., & Kucera, H. (1982). Frequency analysis of English usage: Lexicon and grammar. Boston: Houghton Mifflin Company. Gernsbacher, M.A. (1984). Resolving 20 years of inconsistent interactions between lexical familiarity and orthography, concreteness, and polysemy. Journal of Experimental Psychology: General, 113, 256-28 1. Gerrig, R. J. (1986). Process and products of lexical access. Language and Cognitive Process, 1 , 187-195. Gough, P.B., Alford, J.A., Jr., & Holley-Wilcox, P. (1981). Words and contexts. In O.J.L. Tzeng and H. Singer (Eds.), Perception of prinf: reading research in experimental psychology. Hillsdale, NJ: Erlbaum. Grosjean, F. & Itzler, J. (1984). Can semantic constraint reduce the role of word frequency during spoken-word recognition? Bulletin of the Psychonomic Society, 22, 180-182. Halgren, E. (in press). Insights from evoked potentials into the neuropsychological mechanisms of reading. In A. Scheibel and A. Weschler (Eds.), Neurobiological substrates of higher cognitive function. New York: Guilford. Harris, C.L. (in press). Connectionism and cognitive linguistics. Connection Science. Heit, G., Smith, M.E., & Halgren, E. (1988). Neural encoding of individual words and faces by hippocampus and amygdala. Nature, 333, 773-775. Henderson, L. (1982). Orthography and word recognition in reading. London: Academic Press. Hillyard, S.A. & Kutas, M. (1983). Electrophysiology of cognitive processing. Annual Review of Psychology, 34, 33-61. Hillyard, S.A., Munte, T.F., & Neville, H.J. (1985). Visual-spatial attention, orienting, and brain physiology. In M.I. Posner and O.S. Marin (Eds.), Attention and Performance X I : Mechanism of Attention (pp. 63-84). Hillsdale, NJ: Erlbaum. Holcomb, P.J. (1985). Unimodal and multimodal models of lexical memory: An ERP analysis. Psychophysiology, 22, 576. [Abstract]. Holcomb, P.J. (1988). Automatic and attentional processing: An event-related brain potential analysis of semantic processing. Brain and Language, 35, 66-85. Huttenlocher, J., & Kubicek, L.F. (1983). The source of relatedness effects on naming latency. Journal of Experimental Psychology: Learning, Memory, and Cognition, 9,486-495. Johnson, R., Jr. (1988). The amplitude of the P300 component of the eventrelated potential: Review and synthesis. In P.K. Ackles, J.R. Jennings, and M.G.H. Coles (Eds.), Advances in Psychophysiology, Vol. 3 (pp. 69-138).

168

C . Van Petten and M . Kutas

Greenwich, CT.JAI Press. Karniski, W., Vanderploeg. R.D., Diehl, S.. & Lease, L. (1988). Analysis of the N400 event-related potential component in a simple auditory sentence processing task. Psychophysiology, 25, 460-461. [Abstract]. Katayama, J., Teraji, M.. & Yagi, A. (1989). N400 in orthographic decision tasks. Psychophysiology, 26, S38. [Abstract]. Kawamoto, A.H. (1988). Distributed representations of ambiguous words and their resolution in a connectionist network. In S.L. Small, G.W. Cottrell, and M.K. Tanenhaus (Eds.), Lexical ambiguity resolution: Perspectives from psycholinguistics, neuropsychology, and artificial intelligence. San Mateo, CA: Morgan Kaufmann. Kiger, J.I., & Glass, A.L. (1983). The facilitation of lexical decisions by a prime occurring after the target. Memory & Cognition, 11, 356-365. Kintsch, W. (1988). The role of knowledge in discourse comprehension: A construction-integrationmodel. Psychological Review, 9.5, 163-182. Kintsch, W., & Mross, E.F. (1985). Context effects in word identification. Journal of Memory and Language, 24,336-349. Kutas, M. (1985). ERP comparisons of the effects of single word and sentence contexts on word processing. P sychophysiology, 22, 575-576. [Abstract]. Kutas, M. (1988). Review of event-related potential studies of memoryh M.S. Gazzaniga (Ed.), Perspectives ir. memory research (pp 181-217). Cambridge, MA: MIT Press. Kutas, M., & Hillyard, S.A. (1980a). Event-related brain potentials to semantically inappropriate and surprisingly large words. Biological Psychology, 11,99- 116. Kutas, M., & Hillyard, S.A. (1980b). Reading senseless sentences: brain potentials reflect semantic incongruity. Science, 207. 203-205. Kutas, M., & Hillyard, S.A. (1980~).Reading between the lines: Event related brain potentials during natural sentence processing. Brain and Language, 11, 354-313. Kutas, M., & Hillyard, S.A. (1982). The lateral distribution of event-related potentials during sentence processing. Neuropsychologia, 20, 579-590. Kutas, M., & Hillyard, S.A. (1983). Event-related brain potentials to grammatical errors and semantic anomalies. Memory & Cognition, 1 1 , 539-550. Kutas, M., & Hillyard, S.A. (1984a). Brain potentials during reading reflect word expectancy and semantic association. Nature, 307, 161-163. Kutas, M. & Hillyard S.A. (1984b). Event related brain potentials (ERPs) elicited by novel stimuli during sentence processing. In R. Karrer, J. Cohen, & P. Tueting (Eds.), Brain and informalion: Event-related potentials, Annals of the New York Academy of Sciences, (Vol. 425, pp. 236-241). New York: New York Academy of Sciences. Kutas, M., & Hillyard, S.A. (1989). An electrophysiological probe of incidental semantic association. Journal of Cognitive Neuroscience, 1 , 38-49.

Electrophysiology and Lexical Processing

169

Kutas, M., Hillyard, S.A., & Gazzaniga. M.S.(1988). Processing of semantic anomaly by right and left hemispheres of commissurotomy patients: Evidence from event-related brain potentials, Brain, 111,553-576. Kutas, M., Lindamood, T., & Hillyard, S.A. (1984). Word expectancy and eventrelated brain potentials during sentence processing. In s. Kornblum & J.Requin (Eds.), Preparatory states and processes. (pp 217-238). Hillsdale, NJ: Erlbaum. Kutas, M.,Neville, H.J., & Holcomb. P.J. (1987). A preliminary comparison of the N400 response to semantic anomalies during reading, listening, and signing. Electroencephalography and Clinical Neurophysiology Supplement 39, 325-330. Kutas, M., & Van Petten, C. (1988). Event-related brain potential studies of language. In P.K. Ackles, J.R. Jennings, and M.G.H. Coles (Eds.), Advances in Psychophysiology, Vol. 3 (pp. 139-187). Greenwich, C T JAI Press. Kutas, M., Van Petten, C., & Besson, M. (1988). Event-related potential asymmetries during the reading of sentences. Electroencephalography and Clinical Neurophysiology, 69, 218-233. Lakoff, G.(1987). Women, fire, and dangerous things: What categories reveal about the mind. Chicago: University of Chicago Press. Landauer, T.K. & Streetcr, L.A. (1973). Structural differences between common and rare words: Failure of equivalence assumptions for theories of word recognition. Journal of Verbal Learning and Verbal Behavior, 12, 119131. Langacker, R.W. (1987). Foundations of cognitive grammar, Vol.1: Theoretical prerequisites. Stanford, CA: Stanford University Press. Lindner, S. (1981). A Lexico-semantic analysis of verb-particle constructions with UP and OUT. Doctoral dissertation. University of California at San Diego. Lupker, S.J. (1984). Semantic priming without association: A second look. Journal of Verbal Learning and Verbal Behavior, 23, 709-733. Marslen-Wilson, W.D. (1987). Functional parallelism in spoken wordrecognition. Cognition, 25, 71-102. Marslen-Wilson, W., & Tyler, L.K. (1980). The temporal structure of spoken language understanding. Cognition, 8, 1-71. Massaro, D.W., Jones, R.D., Lipscomb, D., & Scholz, R. (1978). Role of prior knowledge on lexical decisions with good and poor stimulus information. Journal of Experimental Psychology: Human Learning and Memory, 4 , 498-512. Masson, M.E.J. (1989). Lexical ambiguity resolution in a constraint satisfaction network. Program of the eleventh annual conference of the Cognitive Science Society, 757-764. Hillsdalc, New Jersey: Erlbaum. McCallum, W.C., Farmer, S.F. & Pocock, P.K. (1984). The effects of physical

170

C. Van Petten and M.Kutas

and semantic incongruities on auditory event-related potentials. Electroencephalography and Clinical Neurophysiology, 59,477-488. McCann, R.S., Besner, D., & Davelaar, E. (1988). Word recognition and identification: Do word frequency effects reflect lexical access? Journal of Experimental Psychology: Human Perception and Performance, 14, 693706. McCarthy, G., & Wood, C.C. (1984). Intracranially recorded event-related potentials during sentence processing. Society for Neuroscience Abstracts, 10, 847. McClelland, J.L., St. John, M., & Taraban, R. (1989). Sentence comprehension: A parallel distributed processing approach. Language and Cognitive Processes, 4, 287-336. McClelland, J.L.. & Mozer, M.C. (1986). Perceptual interactions in two-word displays: Familiarity and similarity effects. Journal of Experimental Psychology: Human Perception and Performance, 12, 18-35. McKoon. G., & Ratcliff, R. (1989). Semantic associations and elaborative inference. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15. 326-338. Meyer, D.E.. & Schvaneveldt, R.W. (1971). Facilitation in recognizing pairs of words: Evidence of a dependence between retrieval operations. Journal of Experimental Psychology, 90,227-234. Mitchell, D.C., & Green, D.W. (1978). The effects of context and content on immediate processing in reading. Quarterly Journal of Experimental Psychology, 30, 609-636. Monsell, S., Doyle, M.C., & Haggard, P.N. (1989). Effects of frequency on visual word recognition tasks: Where are they? Journal of Experimental Psychology: General, 118,43-71. Morton, J. (1964). The effects of context on the visual duration threshold for words. British Journal of Psychology, 5 5 , 165-180. Morton, J. (1969). The interaction of information in word recognition. Psychological Review, 76, 165-178. Mozer, M.C. (1987). Early parallel processing in reading: A connectionist approach. In M. Collheart (Ed.), Attention and Performance X I I , The Psychology of Reading (pp 83-104). Hillsdale, New Jersey: Erlbaum. Neely, J.H. (1977). Semantic priming and retrieval from lexical memory: Roles of inhibitionless spreading activation and limited-capacity attention. Journal of Experimental Psychology: General, 106,226-254. Nenov, V.I., Halgren, E., Smith, M.E., Badier, J., Ropchan, J.R., Blahd, W.H., & Mandelkern, M. (submitted). Metabolic localization of brain potentials to words. Norris, D. (1986). Word recognition: Context effects without priming. Cognition, 22, 93-136. Nunez, P.L. (1981). Electricfields of the brain: The neurophysics of EEG. New

Electrophysiology and Lexical Processing

171

York: Oxford University Press. O’Seaghdha, P. (1989). The dependence of lexical relatedness effects on syntactic connectedness. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 73-87. Oden, G.L., & Spira, J.L. (1983). Influence of context on the activation and selection of ambiguous word senses. Quarterly Journal of Experimental Psychology, 35A, 51-64. Onifer, W., & Swinney, D.A. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory & Cognition, 9, 225-236. Perfetti, C.A.. Goldman. S.R. & Hogaboam. T.W.(1979). Reading skill and the identification of words in discourse context. Memory & Cognition, 7 , 273-282. Peterson, R.R., & Simpson, G.B. (1989). Effects of backward priming on word recognition in single word and sentence contexts. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 1020-1032. Ratcliff, J.E. (1987). The plausibility effect: Lexical priming or sentential processing. Memory & Cognition, 15, 482-496. Regan, D. (1989). Human brain electrophysiology: Evoked potentials and evoked magnetic fields in science and medicine. New York: Elsevier. Ritter, W., Ford, J.M., Gaillard, A.W.K.. Harter, M.R., Kutas,M., Naatanen, R., Polich, J., Renault, B. & Rohrbaugh, J. (1984). Cognition and eventrelated potentials: I. The relation of negative potentials and cognitive processes. In R. Karrer, J. Cohen, & P.Tueting (Eds.), Brain and information: Event related potentials, (pp 24-38). New York: New York Academy of Sciences. Rouse, R.O., & Verinis, J.S. (1962). The effect of associative connections on the recognition of flashed words. Journal of Verbal Learning ad Verbal Behavior, 88, 454-462. Rubenstein, H. Garfield, L., & Millikan, J.A. (1970). Homographic entries in the internal lexicon. Journal of Verbal Learning and Verbal Behavior, 9 , 487-494, Rugg, M.D. (in press). Event-related potentials dissociate repetition effects of high and low frequency words. Memory & Cognition. Rugg, M.D., Furda. J., & Lorist, M. (1988). The effects of task on the modulation of event-related potentials by word repetition. Psychophysiology, 25, 55-63. Rugg, M.D., & Nagy, M.E. (1987). Lexical contribution to nonword-repetition effects: evidence from event-related potentials. Memory & Cognition, 15, 473-481. St. John, M. (1988). Hitting the right pitch: A meta-analysis of the processing of ambiguous words in context. Unpublished manuscript, Cognitive Science Department, University of California. San Diego.

172

C . Van Petten and M . Kutas

Sanocki, T., Goldman, K., Waltz, J., Cook, C., Epstein, W., & Oden. G. C. (1985). Interaction of stimulus and contextual information during reading: Identifying words within sentences. Memory & Cognition, 13, 145-157. Schmidt, A.L., Arthur, D.L., Kutas, M., George, J., & Flynn, E. (1989). Neuromagnetic responses evoked during reading meaningful and meaningless sentences. Psychophysiology, 26, S6 [Abstract]. Schuberth, R.E., & Eimas, P.D.(1977). Effects of context on the classification of words and nonwords. Journal of Experimental Psychology: Human Perception and Performance, 3,27-36. Schuberth, R.E., Spoehr, K.T., & Lane, D.M. (1981). Effects of stimulus and contextual information on the lexical decision process. Memory & Cognition, 9 , 68-77. Schreuder, R., & Flores d’Arcais, G.B. (1989). Psycholinguistic issues in the lexical representation of meaning. In W. Marslen-Wilson (Ed.), Lexical representation and process (pp. 409-436). Cambridge: MIT Press. Seidenberg, M.S., & McClelland, J.L. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96, 523568. Seidenberg, M.S., Tanenhaus, M.K., Lieman. J.M.,& Bienkowski, M. (1982). Automatic access of the meanings of ambiguous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14, 489-537. Sharkey, N.O. (1989). The lexical distance model and word priming. Program of the Eleventh Annual Conference of the Cognitive Science Society, 860867. Hillsdale, NJ: Erlbaum. Simpson, G.B. (1984). Lexical ambiguity and its role in models of word recognition. Psychological Bulletin, 96, 3 16-340. Simpson, G.B, & Burgess, C. (1985). Activation and selection processes in the recognition of ambiguous words. Journal of Experimental Psychology: Human Perception and Performance, 1 1 , 28-39. Simpson, G.B., Peterson, R.R., Casteel, M.A., & Burgess, C. (1989). Lexical and sentence context effects in word recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 88-97. Smith-Burke, M., & Gingrich, P.S. (1979). The differential role of function words and lexical items in narrative and expository text. In: M. Kamil and A.J. Moe (Eds.), Reading research: Studies and applications. Clemson, SC: National Reading Conference. Smith, M.E., & Halgren, E. (1987). Event-related potentials during lexical decision: Effects of repetition, word frequency, pronounceability, and concreteness. Electroencephalography and Clinical Neurophysiology Supplement 40,417-421. Smith, M.E., Stapleton, J.M., & Halgren, E. (1986). Human medial temporal

Electrophysiology and Lexicul Processing

173

lobe potentials evoked in memory and language tasks. Electroencephalography and Clinical Neurophysiology, 63, 145- 149. Stanovich, K.E., & West, R.F. (1983). On priming by a sentence context. Journal of Experimental Psychology: General, 112, 1-36. Stuss, D.T., Picton, T.W. & Cerri, A.M. (1986). Searching for the names of pictures: An event-related potential study. Psychophysiology, 23, 21 5-223. Stuss, D.T.. Sarazin. F.F., Leech, E.E. & Picton, T.W. (1983). Event-related potentials during naming and mental rotation. Electroencephalography and Clinical Neurophysiology, 56, 133- 146. Swinney, D.A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18, 645-659. Swinney, D.A., & Osterhout, L. (1990). Inference generation during auditory language comprehension. In A. Graesser and G.H. Bower (Eds.), The psychology of learning and motivation. New York: Academic Press. Tabossi, P. (1988a). Effects of context on the immediate interpretation of unambiguous nouns. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 153-162. Tabossi, P. (1988b). Accessing lexical ambiguity in different types of sentential contexts. Journal of Memory and Language, 27, 324-340. Taylor, W.L. (1953). “Cloze” procedure: A new tool for measuring readability. Journalism Quarterly, 30, 4 15-417. Theios, J., & Muise, J.G. (1977). The word identification process in reading. In N.J. Castellan, Jr., D.B. Pisoni, and G.R. Potts (Eds.), Cognitive theory, Vol. 2. Hillsdale, NJ: Erlbaum. Till, R.E., Mross, E.F., & Kintsch, W. (1988). Time course of priming for associate and inference words in a discourse context. Memory & Cognition, 16, 283-290. Tulving, E. & Gold, C. (1963). Stimulus information and contextual information as determinants of tachistoscopic recognition of words. Journal of Experimental Psychology, 6 6 , 319-327. Tyler, L.K., & Marslen-Wilson, W. (1986). The effects of context on the recognition of polymorphemic words. Journal of Memory and Language, 25, 74 1-752. Tyler, L.K., & Wessels, J. (1983). Quantifying contextual contributions to wordrecognition processes. Perception and Psychophysics, 34,409-420. Van Petten, C. (1989). Context effects in word recognition: Studies using eventrelated brain potentials. Doctoral dissertation, University of California at San Diego. Van Petten, C. and Kutas, M. (1987). Ambiguous words in context: An eventrelated potential analysis of the time course of meaning activation. Journal of Memory and Language, 26, 188-208.

174

C . Van Petten and M . Kutas

Van Petten, C., & Kutas, M.(submittcd). Influences of semantic and syntactic context on open and closed class words. Van Petten, C. & Kutas, M.(in press). Interactions between sentence context and word frequency in event-related brain potentials. Memory & Cognition. Warren, R.E. (1972). Stimulus encoding and memory. Journal of Experimental Psychology, 94, 90-100. Warren, R.E. (1974). Association, directionality, and stimulus encoding. Journal of Experimental Psychology, 102, 151-158. Warren, R.E. (1977). Time and the spread of activation in memory. Journal of Experimental Psychology: Human Learning and Memory, 3,458-466. Wood, C.C., & Allison, T.(1981). Interpretation of evoked potentials: A neurophysiological perspective. Canadian Journal of Psychology, 35, 113- 135. Zwitserlood, C.M.E.(1989). Words and sentences: the effects of sententialsemantic context on spoken-word processing, Doctoral dissertation, MaxPlanck Institute for Psycholinguistics, Nijmegen, the Netherlands.

Understanding Word and Sentence C.B. Simps; (Editor) 0 Elsevier Science Publishers B.V. (North-Holland), 1991

Chapter 7 Comprehension Processes in Reading Ambiguous Sentences: Reflections from Eye Movements Keith Rayner and Robin K . Morris University of Massachusetts Amherst. Massachusetts U.S.A.

Reading is a complex skill that is extremely important in our society. In the past twenty years, a considerable amount of research by cognitive psychologists has dealt with how skilled readers process text. A great deal has been learned (see Just & Carpenter, 1987; Rayner & Pollatsek, 1989), but much remains to be investigated. One area where there has been considerable interest of late relates to how readers comprehend text on a moment-to-moment basis. Most of the research on comprehension has focused on the product of reading. However, there is a Considerable amount of current interest in the process of reading comprehension. One way in which on-line comprehension processes can successfully be addressed is by having subjects read sentences that contain some type of ambiguity, where more than one meaning or interpretation is possible, and to determine how they process the ambiguous material. Do readers take longer to process ambiguous material on their first pass through the sentence? If not, do they need to reread it more frequently than unambiguous material? Or, does it take no longer to read ambiguous material than unambiguous material? Using the appropriate material, questions such as these can be framed to determine how on-line processing of normal text takes place. Early psycholinguistic research (conducted in the late 1960s and early 1970s) focused on the processing of ambiguous sentences. Unfortunately, it was the case that many of the tasks that were used were not on-line. For example, studies in which subjects are asked to make grammaticality judgments (a frequently used task) do not provide information about on-line processing activities. Another problem with much of the early research was that different types of ambiguity were not systematically considered or investigated. For example, consider the following three sentences: 1. John knew the boxer was angry when he started barking at him.

176

K . Rayner and R . K . Morris 2. While Mary was mending the sock fell off her lap. 3. Jeff heard that the dessert trains were hot and dusty.

Each of these sentences contain ambiguity, but the ways in which they are ambiguous differ. Sentence 1 contains an example of lexical ambiguity in that there is a word (boxer) that has more than one meaning. Sentence 2 has a syntactic ambiguity in that most readers want to parse “the sock” as the object of “was mending.” only to find when they encounter “fell” that it is really the beginning of a new constituent. Thus, in this type of sentence, readers will often be “led down the garden path” because they expect one thing and something different was intended. Sentence 3 contains a combination of the two types of ambiguity, which we will refer to as syntactic category ambiguity. The word “trains” is lexically ambiguous (it can refer either to locomotives or to the process of teaching) and its syntactic category is also ambiguous (it can be either a noun or a verb). In early psycholinguistic research, it was not uncommon to find all of these different types of ambiguity included in an experiment on sentence ambiguity without any differentiation among them. In this chapter, we will review research we have done dealing with these different types of ambiguity. One of our primary findings is that the different types of ambiguity are handled in different ways by the language processing system. Our research relies very heavily upon eye movement recording during reading to make inferences about how ambiguous sentences are processed. We do this for two reasons. First, we are convinced that eye movements provide a very good on-line record of moment-to-moment comprehension processes. Second, eye movement recording provides a relatively natural reading situation so that we can get a good sense of how people normally process ambiguous sentences. For example, much of the most influential research on lexical ambiguity has relied upon the cross-modal priming task (see Swinney, 1979; Seidenberg, Tanenhaus, Leiman, & Bienkowski, 1982) in which subjects listen to a sentence containing a lexically ambiguous word and at some point close to hearing the ambiguous word (for example, boxer in Sentence 1) they must respond to a word presented visually which is either related to one of the meanings of the ambiguous word or is unrelated to that word. While we find the results from such studies quite plausible and convincing, the cross-modal priming situation rarely (if ever) occurs during more normal real-world situations (i.e., outside of the experimental psychology laboratory). Thus, we believe that the results of our studies using eye movement data provide a good estimate of what readers normally do. Since we rely on eye movement data in studying how readers understand ambiguous sentences, we will begin by presenting an overview of the characteristics of eye movements during reading. From there, we will move to a series of studies that we have carried out on lexical ambiguity. From there, we will turn to research that has been done in our lab on other types of ambiguity to provide

Ambiguity and Eye Movements

177

a contrast with our findings concerning lexical ambiguity. Since our research suggests that different types of ambiguity are treated in different ways, we will conclude the chapter by attempting to account for why this is the case.

Eye Movements in Reading When we read, we make a series of fixations and saccades (rapid movements of the eyes from one position to another). The fixations last about 200250 ms and we typically move our eyes about 8 character spaces. Approximately 15% of the time, we make regressions in which we move our eyes backwards in the text M look at material that we have already read. While the values that we have just presented represent the basic characteristics of eye movements during reading, it is very important to realize that they are only averages and that there is considerable variability between and within subjects. Reading ability greatly influences the mean values; poor readers and beginning readers make longer fixations, shorter saccades, and more regressions than skilled readers. Text difficulty also has a profound influence on the mean values so that as text difficulty increases, readers make more and longer fixations, shorter saccades, and more regressions. While the differences in fixation time, saccade length, and number of regressions due to reading ability and text difficulty are interesting in their own right, for our purposes the more interesting fact is that there is considerable variability within a given reader reading a single passage of text. Thus, though the average fixation duration for a single reader may be 225 rns, the range of fixations may well be from 50 ms to over 500 ms. Likewise, the range of saccade lengths may be from 1 character space to over 15 spaces (though such long saccades only typically occur following a regression, when readers return to the place in the text from which they initiated the regression). While many fixations cluster between 175 and 350 ms, there is obviously variability in how long readers fixate on a word. What causes this variability? The argument that we wish to make (see also Rayner, 1978, 1984; Rayner & Pollatsek, 1987, 1989; Rayner, Sereno, Morris, Schmauder, & Clifton, 1990) is that much of the variability associated with both fixation time and saccade length is related to cognitive processes activated during language comprehension. However, we also hasten to point out that it is clearly the case that some of the variability is also due to purely motoric components of the saccadic eye movement system (Kowler & Anton, 1987; Rayner, Slowiaczek, Bertera, & Clifton, 1983). In addition, low level visual and perceptual factors can influence fixation times (O’Regan & LevySchoen, 1987). Each of these factors distorts the signal that is being taken to reflect immediate language processes. However, useful information about language processes can be gleaned from eye movement data despite this noise in the signal.

K . Rayner and R.K. Morris

178

CRITICAL ISSUESIN USINGEYEMOVEMENT DATA In using eye movement data to investigate on-line language processes there are a number of issues inherent in the methodology that need to be considered prior to turning to our results dealing with lexical ambiguity. We will discuss three relevant issues: (1) measures of processing time associated with fixation times, (2) the perceptual span in reading, and (3) the eye-mind span in reading.

Measures of Processing Time Currently, there is some controversy associated with choosing the appropriate measure of processing time to be used with eye movement data (Blanchard, 1985; Rayner & Pollatsek, 1987; Rayner et d., 1990). If readers made only one fixation on each word, then there would be little problem. Unfortunately, things are not that simple since words are sometimes skipped altogether and some words are fixated more than once before the reader moves on to another word. The occurence of multiple fixations on a word has been treated in several different ways. Some researchers have argued for using the first fixation duration on a word as a measure of lexical access (Inhoff, 1984). The assumption seems to be that what occurs after the first fixation reflects higher-order processing or is noise. However, the opposite assumption has also been made: O'Regan and Levy-Schoen (1987) have argued that the second fixation on a word results when the reader initially lands in a "bad" place and then moves to a more informative position in the word. According to this account, the second fixaton on a word is more diagnostic of lexical access than the first. Our view is that while multiple fixations on a word may be due in part to factors like landing in a non-optimal position in a word, many such cases are also due to the processing associated with the word. Another solution to the problem of multiple fixations is to use the gaze duralion on a word. Gaze duration represents the total amount of time Lhat a reader looks at a word before moving to another word; it is the sum of the fixations on a word excluding any that result from regressions to that word. While some researchers favor it as the best measure of processing time (Just & Carpenter, 1980), others view it with skepticism (Hogaboam & McConkie, 1981; Kliegl, Olson, & Davidson, 1982). Our argument is that first fixation duration and gaze duration reflect related processes. When readers have difficulty processing a word either they maintain their initial fixation on the word or they program a second fixation on the word. When readers have extreme difficulty with a word (particularly, a long word) they may make three or four fixations on the word. Most of the data that we are familiar with has indicated that first fixation and gaze duration yield similar results. For example, one of the robust effects in the literature is that low frequency words are fixated longer than high frequency words (Inhoff &

Ambiguity and Eye Movements

179

Rayner, 1986; Just & Carpenter, 1980; Rayner, 1977; Rayner & Duffy, 1986; Rayner et al., 1990). In one such study (Rayner & Duffy, 1986). first fixation duration was 37 ms longer on low frequency words than high frequency words and gaze duration was 87 ms longer. Both effects were highly significant and support the same basic conclusion: low frequency words are more difficult to process than high frequency words. Finally, it is the case that some words are not fixated directly; readers skip about 20% of the words in the text. Three things are important to note with respect to word skipping. First, word skipping is highly correlated with word length so that short words are skipped much more frequently than long words (Rayner & McConkie, 1976); words that are 7-8 letters or longer are fixated most of the time. Second. words that are highly constrained by the context are skipped more frequently than words that are not constrained (Balota, Pollatsek, & Rayner, 1985; Ehrlich & Rayner, 1981; O’Regan, 1979). Third, when a word is skipped, it is most likely identified on the fixation prior to skipping since that fixation is inflated (Hogaboam, 1983; Pollatsek, Rayner, & Balota, 1986). Our general conclusion is that as much information as possible should be examined in inferring cognitive activities associated with word processing. Examining first fixation duration, gaze duration, and the probability of fixating a word provides researchers with a great deal of information to use to construct a coherent explanation of how words are processed. The Perceptual Span The perceptual span is the region around a reader’s fixation point from which useful information can be obtained. Knowing the size of this region and the type of information obtained different distances from fixation is important. If the perceptual span is large enough to encompass several words, it would not be possible to determine which word was being processed at any point in time, and therefore, eye movement data would not provide a very good measure of the moment-to-moment processing of text. The ideal situation would be that the reader identifies only the fixated word and then moves to the next word on the ensuing saccade. Fortunately, the data suggest that it is frequently the case that readers only identify the word they are fixating. As we mentioned above, readers typically move about 8 character spaces per saccade (which is approximately 1.5 words). Practically, what this means is that some words are skipped (that is, they do not receive a fixation even though they are perceived and processed). However, as we mentioned above, there are systematic differences in what kinds of words get skipped. Short words (three letters or less) are much more likely to be skipped than words that are six letters or more; words that are eight letters or more are rarely skipped; and words that are six letters long are fixated most of the time (Rayner & McConkie, 1976). Content words are fixated over 80% of

180

K. Rayner and R.K.Morris

the time, while function words are fixated 20-40% of the time (Carpenter & Just, 1983; Rayner & Duffy, 1988). The most definitive research on the size of the perceptual span has utilized the eye-contingent display change technique introduced by McConkie and Rayner (1975) and Rayner (1975). In this chapter, we will simply assert the major findings from that research endeavor. More thorough reviews of the research can be found in Rayner and Pollatsek (1987, 1989). The primary finding from the research is that the perceptual span is quite limited during an individual fixation. Readers extract information from a region extending from the beginning of the currently fixated word, but no more than three or four letters to the left of fixation, to about fifteen character spaces to the right of fixation. However, information used to idenfify a word during the current fixation is generally confined to a region extending no more than about 5-7 character spaces to the right of fixation. Thus, the word identification span is smaller than the total perceptual span. Functionally, this often means that the reader is able to identify only the currently fixated word and must then make an eye movement to the next word. However, if two short words (or even three very short words) fall within the word identification span, they may all be identified during a single fixation. Likewise, when a reader fixates on a content word and a short function word follows the content word, the function word can be identified without a direct fixation. The region within which words can be identified is thus quite small, though it clearly isn’t the case that only one word is identified on each and every fixation. A second point that we need to stress is that when the word to the right of fixation is not identified, some processing of that word is done before the reader fixates on it. However, the processing that is done is not at a semantic level (Rayrler, Balota, & Pollatsek, 1986). Indeed, a fair amount of research (see Rayner & Pollatsek, 1987) has documented a preview eflecf in which readers process tt,e first few letters of the word to the right of fixation. The existence of the preview effect and the fact that more than one word sometimes c m be identified on a fixation are problems that must be dealt with in interpreting the eye movement record. But neither problem is insurmountable as we will document later in this chapter. The Eye-Mind Span

The eye-mind span issue has to do with the extent to which the eyes and the mind are fairly tightly linked in reading. If the eyes lead the mind by an appreciable amount, then eye movement data would not be a very sensitive measure of the processing time associated with a given word because the experimenter would never be sure which word was being processed during a particular eye fixation. On the other hand, if the eye-mind span is negligible, then eye movement data can provide a good reflection of moment-to-moment

Ambiguity and Eye Movements

181

word processing. Just and Carpenter (1980) argued against the existence of an eye-mind span and in favor of what they termed the immediacy hypothesis. They argued that the reader completes all processing that can be completed for a given word prior to moving on to the next word. While there is now clear evidence that the ease or difficulty of processing a word influences how long the reader looks at the word, the relationship is not perfect. There are two problems in claiming that fixation times on a word are a pure reflection of the processes associated with comprehending that word. The first is the parafoveal preview effect which we discussed in the prior section: some of the processing associated with a word is initiated on the prior fixation. The second problem has to do with spillover effects. Spillover effects can be illustrated in the two following sentences: 4. The slow music (waltz) captured her attention.

5. The shiny new vehicle (gondola) moved slowly. In each of these sentences a critical target word can be replaced by another word (such as those in parentheses). The difference between the two words is that one (“music”) is a high frequency word, while the other is a low frequency word (“waltz”). A number of studies (Inhoff & Rayner, 1986; Rayner & Duffy, 1986; Rayner et al.. 1990) that have varied word frequency in identical sentence frames such as these have found that fixations on low frequency words are 3090 ms longer on low frequency words than on high frequency words. This is the word frequency effect that we discussed earlier. The interesting point with respect to spillover effects is that when fixation time on the next word (“captured” in Sentence 4) is examined, it is also found to increase by 30-40ms. It seems that the processing associated with the low frequency word spilled over onto the processing of the next word in the text. On the one hand then, studies of the word frequency effect demonstrate clear evidence that the difficulty associated with a specific target word influences the amount of time that the word is looked at. On the other hand, such studies also provide clear evidence for spillover effects. When the spillover problem is added to the fact that some processing of a word is done prior to fixating it (the preview effect), it is reasonable to ask questions about the extent to which the processing time for a word can be determined from the fixation time on that word alone. Ultimately, we will need a theory of just what processing is completed during a fixation on a word, and what processing is done only after the eyes have moved on. In the meantime, however, we would want to argue that by carefully controlling the properties of words that are of interest and by controlling the context in which these words occur, we can make safe conclusions when we find differences in eye movement behavior on target words.

182

K . Rayner and R.K. Morris

One final issue we will address in this section concerns the mechanism that triggers the movements of the eyes from one word to the next in text. In particular, the issue is: to what extent does eye fixation time reflect lexical access processes? Morrison (1984) proposed a model of eye movement control in which lexical access is the trigger for an eye movement program. In Morrison’s model, attentional shifts precede the actual movement of the eyes. Thus, the sequence is (1) lexical access, (2) an attention shift to the next location to be fixated, and (3) movement of the eye. We will not discuss this model in detail here, but it does account for many of the aspects of eye movements in reading. Recently, a number of modifications have been made in the model to account for other aspects of the data (see Rayner & Pollatsek, 1989; Pollatsek & Rayner, 1990). One major question concerning the issue of what determines when the eyes move from one word to another, and one that is critical for the present chapter, concerns the extent to which fixation time on a word reflects primarily lexical access processes. That is, do fixation times only refiect lexical access? Or, are other processes also reflected in the amount of time that a reader looks at a word? Elsewhere, we (Rayner et al., 1989) have discussed the extent to which the amount of time that the eyes remain fixated on a word reflects lexical access, text integration processes, or both. The issue is complicated and factors related to the points we made above about the appropriate measure of processing time on a word are involved. For instance, Inhoff (1984) argued (and presented data to support his contention) that first fixation on a word reflected lexical access and subsequent fixations reflected integration processes. However, subsequent data have suggested that such a clear distinction is probably not justified. Perhaps the most direct test of the issue was provided in a series of experiments reported by Schustack, Ehrlich, and Rayner (1987). Schustack et al. found that fixation time on a target noun varied as a function of how far it was from a previous referent and as a function of how constraining the verb was that immediately preceded it. In order to discriminate between effects due to access and those due to text integration they sought converging evidence using a naming task. Thus, subjects read the same passages of text as in the eye movement study and named the target word when it appeared on the screen. The logic was that since naming time is assumed to be a “pure” measure of lexical access, those effects which reflect lexical access processes should replicate when subjects name a target word, while those effects due to text integration would not be reflected in the naming time measure. The results of the naming time experiment were that there was still an effect of verb constraint, but not an effect of distance. Thus, the results of this study suggest that fixation times reflect both lexical access and text integration processes. While text integration processes do influence fixation time, our suspicion is that most of the

Ambiguity and Eye Movements

183

variability in fixation time is due to lexical access processes. When readers have considerable difficulty at the text integration stage, the evidence indicates that the processes spill over onto the next fixation and further eye fixations are made in the region of processing difficulty. As we stated above, eventually it will be necessary to have a theory of just what processes are completed when in relation to when the eyes move. In the meantime, we believe that it is possible to infer the source of difficulty associated with different levels of processing from the eye movement data. Experiments conducted in our lab exploring issues associated with the processing of lexically ambiguous words further illustrate the utility of eye movement data in investigating critical processes associated with on-line comprehension processes.

LEXICAL AMBIGUITY Our initial work dealing with lexical ambiguity can be traced to a study in which we were interested in the effect of lexical complexity on fixation times in reading (Rayner & Duffy, 1986). One of the primary goals in the research was to provide evidence concerning the extent to which fixation times on words reflect the processing associated with that word. We reasoned that if words are lexically complex, and if fixation times are sensitive to such characteristics of the word, then they should increase for lexically complex words (in comparison to words that were not as complex). Thus, we varied word frequency (holding factors such as word length, number of syllables, bigram frequency, and predictability in the sentence context constant) and verb complexity, in addition to lexical ambiguity. Briefly, there were major effects of word frequency, but no effects on fixation times of verb complexity. In the study dealing with lexical ambiguity, Rayner and Duffy (1986) compared a set of ambiguous words with unambiguous control words presented in the same sentence contexts. The control words were matched on word frequency, word length, and number of syllables. In addition, the context preceding the target word was constant in both the experimental and control conditions and it was neutral with respect to which meaning of the ambiguous word should be instantiated. Again, the reasoning was that a polysemous word should be more difficult to process than a word with only one meaning; the assumption was that there would be competition between alternative meanings if all meanings were activated. The results of the initial study were somewhat disappointing because no differences were found between the ambiguous target words and their matched control (unambiguous) words. However, a post-hoc analysis of the data revealed that there may well have been an effect on nonbiased ambiguous words (words for which the two interpretations of the word are relatively equal). In a subsequent experiment, Rayner and Duffy (1986) used half biased

K.Rayner and R . K . Morris

184

ambiguous words (words with one strongly dominant interpretation) and half nonbiased words. Again, the disambiguating information followed the target word and fixation times on the ambiguous words were compared to the control words. In addition, the amount of time readers spent in the disambiguating region of the sentence was examined. The disambiguating information was always consistent with the subordinate interpretation. The basic results of the study were quite clear (see Table 1). While subjects looked no longer at h e biased words than the control words, they looked significantly longer at the nonbiased words than the control words. However, it took them significantly longer to read the disambiguating region with the biased words than with the nonbiased words.

Table 1 Mean Gaze Durations (in milliseconds) on the Ambiguous Target Words and the Control Words in Rayner and Duffy (1986).

Ambiguous

Control

Biased

260

263

Nonbiased

275

258

As a result of this initial study, another study (Duffy, Morris, & Rayner, 1988) was designed in which the disambiguating information either preceded or followed the target word. This was accomplished by using two clause sentences wherein the order of the two clauses could be reversed as in: 6. When she finally served it to her guests, the port was a great success. 7. Last night the port was a great success when she finally served it to her guests.

Table 2 Mean Gaze Durations (in milliseconds) on the Ambiguous Target Words and the Control Words in Duffy et al. (1 988).

Location of Disambiguating Clause Before

After

Ambiguous

Control

Ambiguous

Control

Biased

276

255

26 1

259

Nonbiased

264

264

219

26 1

185

Ambiguity and Eye Movements

In this experiment, the pattern of results (see Table 2) reported by Rayner and Duffy was replicated when the disambiguating information followed the target word. However, when the disambiguating information preceded the target, the pattern of results was somewhat different. That is, subjects looked at the nonbiased ambiguous target word no longer than it’s matched control. On the other hand, they looked at the biased word (in all cases the subordinate meaning had been instantiated by the preceding context) longer than it’s control. Furthermore, the post-target region (“was a great success” in the example) took a considerable amount of time to process when the disambiguating information preceded the target suggesting that not only lexical access processes were affected, but also integration processes. Before discussing the implications of these findings, we will discuss one further experiment by Rayner and Frazier (1989). Like Duffy et al., Rayner and Frazier varied whether or not the disambiguating information preceded or followed the ambiguous target word. However, unlike Duffy et al., the disambiguating information was consistent with either the dominant or subordinate meaning of both biased and equibiased ambiguous words. In addition, Rayner and Frazier did not compare the ambiguous words to matched control words, but rather compared the dominant and subordinate sense of each word. Table 3

Mean Gaze Durations (in milliseconds) on the Ambiguous Target Words in Rayner and Frazier (1989). Location of Disambiguating Clause

Before

Dominant

After

Subordinate

Dominant

Subordinate

Biased

224

286

220

22 1

Nonbiased

223

224

250

250

Note: Dominant vs. subordinate refers to which meaning of the ambiguous word was instantiated by the context. The results of the Rayner and Frazier experiment (see Table 3) are quite consistent with the Duffy et al. results. Once again, when the disambiguating

186

K . Rayner and R.K. Morris

information followed the target word, subjects looked longer at nonbiased words than biased words. When the disambiguating information preceded the target word, subjects looked at the biased word for a long time only if the subordinate meaning was instantiated by the context. In this experiment, like the others, when the disambiguating information followed the target word, readers had considerable difficulty in the region of the disambiguation when the subordinate meaning was instantiated; when the disambiguating information preceded the target word and the subordinate meaning was instantiated they had considerable difficulty processing the post-target region.

Theoretical Mechanisms What can be made of these results? Much of the research on lexical ambiguity has been undertaken to differentiate between selective versus exhaustive access models of word processing. According to selective access models, only the contextually relevant meaning of an ambiguous word is activated while according to exhaustive access models more than one (possibily all when there are more than two meanings) meaning is automatically activated. However, a fair amount of research using cross-modal priming techniques and other techniques (see Simpson, 1984) has made it clear that there are a number of critical factors involved including the nature of the context and the nature of the ambiguous word (for example, biased versus nonbiased alternative meanings). These findings, along with the findings reported in our lab, suggest that pure exhaustive access and pure selective access models cannot accurately capture the processing activities associated with the processing of ambiguous words. Thus, as we shall discuss below, it probably makes most sense to talk about different models of lexical ambiguity in the context of where the disambiguating information appears. Recent research has converged on a two-stage model to account for the processing of ambiguous words which are ambiguous when encountered (that is, when the preceding context does not disambiguate the intended meaning). In the lexical access stage, all meanings of an ambiguous word are initially accessed. In the subsequent selection stage, one meaning is selected. The timing of access seems to depend on the relative frequency of the various meanings. For nonbiased ambiguous words, the two meanings are accessed simultaneously (Seidenberg et al., 1982; Swinney, 1979). For biased ambiguous words, although both meanings are accessed (Onifer & Swinney, 1981) the dominant meaning becomes available earlier than the subordinate meaning (Simpson & Burgess, 1985). Thus, under this model, the meanings of an ambiguous word are accessed exhaustively in order of frequency. This exhaustive access process is followed by the selection of one meaning, which is accomplished quite quickly. While the evidence seems to converge on an exhaustive access model for

Ambiguity and Eye Movements

187

ambiguous words preceded by a neutral context, the picture is less clear for the processing of ambiguous words preceded by disambiguating context. There are two kinds of models of how prior disambiguating information could affect the processing of an ambiguous word. These models differ in their assumptions about the effect of context on the lexical access process for ambiguous words. Under the assumptions of selective access models, prior disambiguating information provides enough information to allow access of only the appropriate meaning of the word. Under exhaustive access models, all meanings are accessed as long as prior disambiguating information does not contain a word which strongly primes one meaning of the ambiguous word. Let’s now consider two versions of the exhaustive access model. Under the uutonomous access model, prior disambiguating context has no effect at all on the access process (unless the context contains words which strongly prime one meaning). The access process proceeds exactly as it would when preceded by neutral context. Under the reordered access model, prior context affects the access process by speeding the availability of the context appropriate meaning without influencing the alternative meaning. Thus, the appropriate meaning for nonbiased words would be expected to become available earlier than the inappropriate meaning (in essence, the nonbiased word becomes biased). For the case in which the appropriate meaning is the subordinate meaning of a biased word, the model predicts that this less frequent meaning would become available earlier than usual, possibly simultaneously with the more frequent meaning. Such a model has been advocated by Hogaboam and Perfetti (1975), Carpenter and Daneman (1981), and Simpson (1984). This is also the type of model that Duffy et al. (1988) argued could best account for their eye movemen t data. Recall again that the primary results in the Duffy et al. study were that readers fixated longer on nonbiased words than control words and no longer on biased than control words when the disambiguating information followed the word. However, when the disambiguating information preceded the target word, readers looked at the biased word longer than the control word (since the subordinate meaning had been instantiated) and no longer at the nonbiased than the control word. The reordered access model predicts just these results. Under this model, both frequency of meaning and context can influence the order in which the meanings of an ambiguous word are accessed. When the disambiguating information follows the target word, the two meanings of the nonbiased word become available simultaneously and therefore compete with one another at the integration stage, resulting in longer fixation times on the word. In the case of the biased words, the dominant meaning of the word becomes available so much earlier than the subordinate meaning that there is no competition at the integration stage; integration of the dominant meaning would begin before the subordinate meaning was fully accessed. When the disambiguating information precedes the target word, the availability of one meaning of the nonbiased word

188

K . Rayner and R.K. Morris

is speeded, thus making it available to the integration mechanism first, and hence, processing proceeds relatively smoothly. On the other hand, when the preceding context speeds access of the subordinate meaning of a biased ambiguous word, the two meanings of the word may be accessed simultaneously and therefore compete with one another at the integration stage. In this case, the biased word has become much like the nonbiased word when the disambiguating information follows the target. In contrast to the reordered access model, both the selective access and the autonomous access models have difficulty accounting for the pattern of data. The selective access model, for example. can not account for the increased fixation times on biased words when they are preceded by context that disambiguates to the subordinate meaning of the word, nor can it account for the inflated processing time in the post-target region in this case. The autonomous access model, on the other hand, does not give a well-motivated account of the shift in the pattern of fixation times on the target words when the disambiguating information is moved from after the target word to before it. This model predicts that the same pattern of results should be obtained independently of where the disambiguating information occurs. Hence, like the selective access model, it is unable to account for the full pattern of results obtained in the Rayner and Duffy (1986), Duffy et al. (1988), and Rayner and Frazier (1989) studies. While the reordered access model can nicely account for the data, an alternative account of the data pattern in these studies has been offered by Rayner and Frazier (1989). They basically offered an alternative way of conceptualizing the processing associated with ambiguous words in what they referred to as the integration model. According to the integration model, the language processing mechanism automatically attempts to access any meanings stored for the currently fixated word; only the integration process can be influenced by prior context. Access procedures operating on the current word are terminated only after one or more meanings have been successfully integrated with prior context. This type of model can thus account for the finding that gaze durations on nonbiased words are longer than on biased words when disambiguating information follows the target word: Integration of the dominant meaning with the context occurs before access of the subordinate meaning only in the case of meanings with large frequency disparities (i.e., biased ambiguous words). When disambiguating information precedes the target word, this information can be used to speed integration of one of the meanings (thus terminating the search for the subordinate meaning). For nonbiased words, the meaning which is consistent with the context would thus be quickly integrated so that the process would be quite selective. For biased words, a prior context favoring the subordinate meaning would prevent acceptance of the dominant meaning, thereby preventing termination of the access process operating on the current word (and, hence, increasing the gaze duration on the word).

Ambiguity and Eye Movements

189

To summarize, the reordered access model relies on access mechanisms to account for the effects of lexical ambiguity on gaze durations while the integration model localizes the effects at a stage subsequent to lexical access. In both models, more than one meaning of an ambiguous word is automatically activated. In the reordered access model, the speed of activation of alternative meanings is influenced by the type of word (biased versus nonbiased) and by the location of disambiguating information. In the integration model, more than one meaning is automatically activated, but selection mechanisms operating on the output of the access stage are critical; as in the reordered access model, the type of ambiguous word and the location of the disambiguating information influence the speed of the integration process. At the moment, we would argue that on the basis of the data that have been collected it is difficult to discriminate between the two models. However, we would like to note that we view the studies that have been done in our laboratory as the beginning of a series of studies that we will undertake to address the processing of lexical ambiguity. We are currently designing experiments to discriminate between the two models. In addition, we are running studies to investigate the extent to which a number of prior encounters with the subordinate sense of a biased ambiguous word will “prime” that meaning when it is encountered in a target location in text. Perhaps such encounters come closer to representing the more normal conditions under which readers deal with biased ambiguous words in the reading of connected discourse.

Sense Ambiguity Before turning to other types of ambiguity, we will now briefly describe some research from our laboratory dealing with what we will call sense ambiguity. The type of ambiguous words that we have been discussing are interesting because they have more than one distinct meaning. In this section, we will consider words that have multiple senses. For example, the word book can refer either to a concrete object (a particular book) or to an abstract concept (the book that you’re supposed to begin writing next week). Do readers treat such words in the same way as they treat lexically ambiguous words? Frazier and Rayner (1990) explicitly compared the reading of lexically ambiguous words with the reading of words with multiple senses. Subjects read sentences containing target words that were either lexially ambiguous or were words with multiple senses. In both cases, the disambiguating information followed the target word. As a baseline, sentences with target words that were unambiguous were also included in the study. The results of the study were quite clear. Frazier and Rayner found that readers tended to treat the lexically ambiguous words differently than words with multiple senses or the control words (as indexed by fixation times on the target word). On the other hand, there were no major differences between fixation times on the words with

190

K . Rayner and R.K. Morris

multiple senses and the control words. We believe that this finding has interesting implications for how lexical access proceeds for lexically ambiguous words and words with mu1tiple senses. Frazier and Rayner’s findings suggest different implications for the processing associated with the two different types of words. Since the multiple senses of a word are not incompatible with one another, immediate selection of one sense over the other may not be necessary in order for processing to proceed. However, in the case of words with multiple meanings, these meanings are, by definition, mutually exclusive. Therefore, processing can not proceed until one meaning is selected. and this process of selecting the appropriate meaning takes time.

SYNTACTIC AMBIGUITY So far we have considered the influence of word level ambiguities on reading. Eye movement data have been particularly useful in inferring the processing associated with such ambiguous words. Eye movement data have also been particularly useful for investigating how readers parse sentences during comprehension processing. Much of the research that has been undertaken in our laboratory (Ferreira & Clifton, 1986; Ferreira & Henderson, 1990; Frazier & Rayner, 1982; Rayner, Carlson, & Frazier, 1983; Rayner & Frazier, 1987) has been carried out to test a general garden-path theory of sentence processing. In particular, two general parsing principles called lute closure and minimal aitachment, which we have argued account for how readers parse sentences, have motivated most of h e research. The late closure principle is that if grammatically permissable, the reader attaches new items to the phrase or clause currently being processed. Late closure favors attachment to preceding items over attachment to subsequent items, allowing sentences like 8 to be parsed correctly, but causing sentence 9 to be parsed incorrectly. That is, readers typically parse “the sock” as the object of “mending,” only to find when they encounter the word “fell” (the disambiguating word) that it is the beginning of a new constituent. 8. While Mary was mending the sock it fell off her lap. 9. While Mary was mending the sock fell off her lap. According to the mimimal attachment principle, readers attach incoming material into the phrase marker being constructed using the fewest nodes consistent with the well-formed rules of the language. For example, minimal attachment predicts in sentences like 10 and 11 the phrase “the answer” will initially be taken as the direct object of the verb “knew.” even though this leads to a revision of this analysis in Sentence 11 where “the answer” is in fact the object of the new clause. The reason, according to the application of the mini-

Ambiguity and Eye Movements

191

ma1 attachment principle, is that the analysis of Sentence 11 includes an additional node to the one in Sentence 10. In Sentence 12, the prepositional phrase “with the binoculars” is ambiguous because it could be either John or the cop that has the binoculars. Minimal attachment predicts that readers will attach the prepositional phrase to the verb resulting in misanalysis. 10. The girl knew the answer by heart. 11. The girl knew the answer was wrong. 12. John saw the cop with the binoculars but not the other one.

Each of these strategies can be viewed as a consequence of a general tendency of the syntactic processor to adopt the first available syntactic analysis of an input (Frazier, 1985. 1987). In this chapter, we will not dwell on these strategies. Rather, we will outline in broad strokes the general pattern of results that we have obtained so as to contrast them with the results from the lexical ambiguity studies we have done. In these studies, subjects have been asked to read sentences that contain temporary structural ambiguities such as 9, 11, and 12 (they also read alternative versions like 8 and 10). Sentences like these are syntactically ambiguous (or have temporary structural ambiguities) because “the sock,” “with the binoculars,” and “the answer” all have two potential ways in which they can be parsed. Do readers compute both structures or are they committed to a single interpretation which forces them to reparse the sentence when that analysis turns out to be incorrect? Frazier and Rayner (1982) found strong evidence in favor of the latter position. They recorded readers eye movements and found that there was considerable disruption in the disambiguating region of the sentence in comparison to control versions of the sentences. When readers encountered the disambiguating information they tended to exhibit three different patterns of eye movement behavior: (1) they immediately made a regression to the ambiguous phrase and then read the sentence quite smoothly, (2) they fixated for a long time on the disambiguating word and reanalyzed the sentence without going back in the sentence, or (3) they showed considerable confusion and disruption (as evidenced by long fixations and short saccades) and read to the end of the sentence and then went back to the beginning of the sentence and started over. Unlike the general finding with lexical ambiguity then, the pattern of results obtained in a number of studies (Ferreira & Clifton, 1986; Ferreira & Henderson, 1990; Frazier & Rayner. 1982; Rayner et al., 1983; Rayner & Frazier, 1987) suggests that readers only assign a single representation for sentences containing temporary structural ambiguities. Thus, in comparison to lexical ambiguity, where multiple meanings are considered by the language processing system, with syntactic ambiguity, only a single representation is assigned. This representation, it has been argued (Frazier & Rayner, 1982) is

K.Rayner and R.K.Morris

192

based on strategies related to the underlying structure of sentence processing and the late closure and minimal attachment principles. We will discuss at a later point why the language processing system adopts different strategies depending upon the type of ambiguity. In addition to the general conclusion that readers only assign a single meaning to structurally ambiguous sentences, the research from our laboratory has revealed a number of somewhat surprising findings. It has been found that (1) pragmatic or world knowledge information does not influence the parser in making it’s initial assignment (Rayner et al., 1983), (2) contextual information does not prevent initial garden-pathing (Ferreira & Clifton, 1986). and (3) the characteristics of the verb, such as whether or not it prefers a direct object, does not strongly affect the initial parsing assignment (Ferreira & Henderson, 1990). All of these findings are quite controversial and the results of our studies have been countered by others (for example, see Altmann. 1988; Altmann & Steedman, 1988; Taraban & McClelland, 1988). We will not discuss the relative merits of the various claims here. However, we will point out that, for the most part, the studies that are inconsistent with our claims have used self-paced reading tasks. As we have argued elsehwere (Rayner et al., 1990), eye movement data provide much more resolution in dealing with immediate on-line language processing activities than do self-paced reading tasks which slow reading down to about half its normal rate.

SYNTACTIC CATEGORY AMBIGUITY In this section, we will discuss one final type of ambiguity. By syntactic category ambiguity, we mean that a given word is ambiguous as to the possible grammatical role it plays in a sentence. Consider sentences like (13 and 14): 13. The desert trains are hot and dusty. 14. The desert trains boys to be men.

The words “desert” and “trains” are ambiguous when encountered here because they can be an adjective and a noun (as in sentence 13) or a noun and a verb (as in sentence 14). What should the reader do when situations such as this are encountered? One possibility is to access both meanings of the word, but this means that a decision will need to be made about which of the grammatical categories are relevant or else to compute both possible syntactic readings and hold onto them. This latter alternative seems unlikely, given the limitations in short-term memory and given our findings discussed in the prior section. A second, perhaps related, possibility would be to make an immediate decision about the syntactic role of the word (“desert” in the example) as soon as it is encountered. Such a decision could be based on the relative frequency of usage of the word in each of the two possible syntactic categories. A third possibility

Ambiguity and Eye Movements

193

is that the decision is delayed until more information becomes available to clarify the syntactic role of the word in question. Frazier and Rayner (1987) reported the results of three experiments in which subjects read sentences containing syntactic category ambiguities. Half of the sentences were ambiguous (as in 13 and 14) and in the other half of the sentences the syntactic role of the target words was disambiguated by preceding context. The major finding across the three studies was that subjects looked at the target words (“desert trains” in the example) for less time when they were ambiguous than when preceding context disambiguated their syntactic categories. In addition, they took longer to read the remainder of the sentence in the ambiguous condition than in the unambiguous condition. Thus, the results suggest that readers delay in assigning a syntactic category when words are ambiguous and their syntactic category is not obvious in hopes of obtaining later relevant information that will aid them in making the appropriate decision. One interesting question unanswered by the Frazier and Rayner studies relates to how long readers might delay in making an assignment. In the studies that Frazier and Rayner reported, the disambiguating information was only a few words away from the target words. We suspect that if the disambiguating information was further away, short-term memory limitations would force readers to make a decision.

AMBIGUITY IN READING: SUMMARY In this chapter, we have reviewed a number of different types of studies that have been undertaken in our laboratory dealing with various types of ambiguity. As we pointed out in the Introduction, in some past research different types of ambiguity have not been systematically investigated. In our lab, we have conducted studies dealing with lexical ambiguity, syntactic ambiguity, and syntactic category ambiguity. Our results suggest that readers deal with these different types of ambiguity in differing ways. Our results with lexically ambiguous words are generally consistent with the common finding that more than one meaning of an ambiguous word is automatically activated. However, we also found clear evidence that (1) the characteristics of the ambiguous word and (2) the prior context are both important. With respect to the type of ambiguous word, we found that biased words are treated differently than nonbiased words. With respect to the prior context, we found that biasing and neutral contexts have differing effects on biased and nonbiased ambiguous words. The two models that we have proposed to account for our results (the reordered access model and the integration model) are variants of exhaustive access models that take into account the characteristics of the ambiguous word and the location of disambiguating information. While we cannot at this point discriminate between the two models, the important point is that our work is consistent with other work which suggests that more than one

194

K.Rayner and R . K . Morris

meaning of a lexically ambiguous word is activated. In contrast, our results concerning syntactic ambiguity reveal a very different general conclusion. That is, work in our lab suggests that readers only entertain one alternative interpretation for a syntactically ambiguous string of words. The interpretation they choose is based on structural principles, but when they are wrong about the analysis they have chosen (i.e.. they get led “down the garden path”), they must revise their analysis and reparse the sentence. Research from our lab also indicates that in the case of syntactic category ambiguities, readers delay in making an interpretation and move forward in the text hoping to find information that will aid them in making the appropriate syntactic assignment. Why should it be the case that these different types of ambiguity are dealt with in dramatically different ways by the language processing system? Our basic assumption is that the differences reflect an important characteristic of the processing system. That is, we suspect that the system acts differently when looking up a stored representation (as in the case of lexical ambiguity) than when required to make computations on-line (as in the case of sentence parsing). Capacity limitations may force the system to only compute one alternative (syntactic ambiguity) or to delay making an assignment (syntactic category ambiguity) when language processing decisions are made on-the-fly. On the other hand, when looking up the meaning of stored representations in the lexicon, more than one meaning may automatically be activated. Finally, it is important to note that these differences in how readers process different types of ambiguity are transparent in the eye movement record. The types of effects that we have documented in this chapter show up in fixation times on critical words and in the pattern of eye movements (such as when readers make a regression). Our goal is to continue research on the types of issues addressed in this chapter in the hope of obtaining more information about moment-to-moment comprehension processes during reading. Acknowledgments

Preparation of this chapter was supported by Grant HD17246 from the National Institute of Child Health and Human Development. Appreciation is expressed to our colleagues Chuck Clifton, Susan Duffy, and Lyn Frazier for their collaboration on many of the studies on ambiguity described in this chapter; we would also like to acknowledge their contributions to the ideas presented. References

Altmann, G. (1988). Ambiguity, parsing strategies, and computational models. Language and Cognitive Processes, 3.73-97.

Ambiguity and Eye Movements

195

Alunann, G., & Stecdman, M. (1988). Interaction with context during human sentence processing. Cognition, 30, 191-238. Balota, D.A., Pollatsek, A., & Rayner, K. (1985). The interaction of contextual constraints and parafoveal visual information. Cognitive Psychology, 17, 364-390. Blanchard, H.E. (1985). A comparison of some processing time measures based on eye movements. Acta Psychologia, 58, 1-15. Carpenter, P.A., & Daneman, M. (1981). Lexical retrieval and error recovery in reading: A model based on eye fixations. Journal of Verbal Learning and Verbal Behavior, 20, 137-160. Carpenter, P.A., & Just, M.A. (1983). What your eyes do while your mind is reading. In K. Rayncr (Ed.), Eye movements in reading: Perceptual and language processes. New York: Academic Press. Duffy, S.A., Morris, R.K., & Rayner, K. (1988). Lexical ambiguity and fixation times in reading. Journal of Memory and Language, 27,429-446. Ehrlich, S.F., & Rayncr, K. (1981). Contextual effects on word perception and eye movements during reading. Journal of Verbal Learning and Verbal Behavior, 20, 641-655. Ferreira, F., & Clifton, C. (1986). The independence of syntactic processing. Journal of Memory & Language, 25,348-368. Ferreira, F., & Hcndcrson, J.M. (1990). The use of verb information in syntactic parsing: A comparison of evidence from eye movements and word-byword self-paced rcading. Journal of Experimental Psychology: Learning, Memory, and Cognilion, 16, 555-568. Frazier, L. (1985). Syntactic complexity. In D. Dowty, L. Karttunen, & H. Zwicky (Eds.), Nalural language parsing. Cambridge: Cambridge University Press. Frazier, L. (1987). Sentence processing: A tutorial review. In M. Coltheart (Ed.), Attenlion and performance X I I . London: Erlbaum. Frazier, L., & Rayncr, K. (1982). Making and correcting errors during sentence comprchcnsion: Eye movements in the analysis of structurally ambiguous sentences. Cognilive Psychology, 14, 178-210. Frazier, L., & Rayncr, K. (1987). Resolution of syntactic category ambiguities: Eye movemcnts in parsing lexically ambiguous sentences. Journal of Memory and Language, 26, 505-526. Frazicr, L., & Rayncr, K. (1990). Taking on semantic commitments: Processing multiple mcanings vs. multiple senses. Journal of Memory and Language, 29, 181-200. Hogaboam, T.W. (1983). Reading patterns in eye-movement data. In K. Rayner (Ed.), Eye movements in reading: Perceptual and language processes. New York: Academic Press. Hogaboam, T.W., & Pcrfctti, C.A. (1975). Lexical ambiguity and sentence comprchcnsion. Journal of Verbal Learning and Verbal Behavior, 14,265-

196

K . Rayner and R.K. Morris

274. Hogaboam, T.W., & McConkie, G.W. (1981). The rocky road from eye fixations to comprehension. Technical Report 207, Center for the Study of Reading, University of Illinois. Inhoff, A.W. (1984). Two stages of word processing during eye fixations in the reading of prose. Journal of Verbal Learning and Verbal Behavior, 23, 612-624. Inhoff, A.W., & Rayner, K. (1986). Parafoveal word processing during eye fixations in reading: Effects of word frequency. Perception & Psychophysics, 40.43 1-439. Just, M.A., & Carpenter, P.A. (1980). A theory of reading: From eye fixations to comprehension. Psychological Review, 87, 329-354. Just, M.A., & Carpenter, P.A. (1987). The psychology of reading and language comprehension. Newton, MA: Allyn and Bacon. Kleigl, R., Olson, R.K., & Davidson, B.J. (1982). Regression analysis as a tool for studying reading processes: Comment on Just and Carpenter’s cye fixation theory. Memory & Cognition, 13, 107- 1 1 1 . Kowler, E., & Anton, S. (1987). Reading twisted text: Implications for the role of saccades. Vision Research, 27,45-60. McConkie. G.W., & Rayner, K. (1975). The span of the effective stimulus during an eye fixation in reading. Perception & Psychophysics, 17, 578586. Morrison, R.E. (1984). Manipulation of stimulus onset delay in reading: Evidence for parallel programming of saccades. Journal of Experimental Psychology: Human Perception and Performance, 10,667-682. Onifer, W., & Swinney, D.A. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory & Cognition, 9, 225-236. O’Regan, J.K. (1979). Eye guidance in reading: Evidence for the linguistic control hypothesis. Perception & Psychophysics, 25, 501 -509. O’Regan, J.K., & Levy-Schoen, A. (1987). Eye-movement strategy and tactics in word recognition and reading. In M. Coltheart (Ed.), Attention and performance Xff. London: Erlbaum. Pollatsek, A., & Rayner, K. (1990). Eye movements and lexical access in reading. In D.A. Balota, G.B. Flores d’Arcais (Eds.), Comprehension processes in reading. Hillsdale, NJ: Erlbaum. Pollatsek, A., Rayner, K., & Balota, D.A. (1986). Inferences about eye movement control from the perceptual span in reading. Perceplion & Psychophysics, 40, 123-130. Rayner, K. (1975). The perceptual span and peripheral cues in reading. Cognitive Psychology, 7, 65-81. Rayner, K. (1977). Visual attention in reading: Eye movements reflect cognitive processes. Memory di Cognilion, 4,443-448.

Ambiguity and Eye Movements

197

Rayner, K. (1978). Eye movements in reading and information processing, Psychological Bulletin, 85,618-660. Rayner, K., Balota, D.A., & Pollatsek. A. (1986). Against parafoveal semantic preprocessing during eye fixations in reading. Canadian Journal of Psychology, 40,473-483. Rayner, K., Carlson, M., & Frazier, L. (1983). The interaction of syntax and semantics during sentence processing: Eye movements in the analysis of semantically biased sentences. Journal of Verbal Learning and Verbal Behavior, 22, 358-374. Rayner, K., & Duffy, S.A. (1986). Lexical complexity and fixation times in reading: Effects of word frequency, verb complexity. and lexical ambiguity. Memory & Cognition, 14, 191-201. Rayner, K., & Duffy, S.A. (1988). On-line comprehension processes and eye movements during reading. In M. Daneman, G.E. MacKinnon, & T.G. Waller (Eds.), Reading research: Advances in theory and practice, Volume 6. New York: Academic Press. Rayner, K., & Frazier, L. (1987). Parsing temporarily ambiguous complements. Quarterly Journal of Experimental Psychology, 39A, 657-673. Rayner, K., & Frazier, L. (1989). Selection mechanisms in reading lexically ambiguous words. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15,779-790. Rayner, K., & McConkie, G.W. (1986). What guides a reader’s eye movements. Vision Research, 16, 829-837. Rayner, K., & Pollatsek, A. (1987). Eye movements in reading: A tutorial review. In M. Coltheart (Ed.), Attention and performance XII. London: Erlbaum. Rayner, K., & Pollatsek, A. (1989). The psychology of reading. Englewood Cliffs, NJ: Prentice Hall. Rayner, K., Sereno, S.C., Morris, R.K., Schmauder, A.R., & Clifton, C. (1989). Eye movements on on-line comprehension processes. Language and Cognitive Processes, in press. Rayner, K., Slowiaczek, M.L., Clifton, C., & Bertera, J.H. (1983). Latency of sequential eye movements: Implications for reading. Journal of Experimental Psychology: Human Perception and Performance, 9 , 9 12-922. Schustack, M., Ehrlich, S.F., & Rayner, K. (1987). The complexity of contextual facilitation in reading: Local and global influences. Journal of Memory and Language, 26,322-340. Seidenberg, M.S., Tanenhaus, M.K., Leiman, J.M., & Bienkowski, M. (1982). Automatic access of the meanings of ambiguous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14, 489-537. Simpson, G.B. (1984). Lexical ambiguity and its role in models of word recognition. Psychological Bulletin, 96, 3 16-340.

198

K . Rayner and R.K. Morris

Simpson, G.B., & Burgess, C. (1985). Activation and selection processes in the recognition of ambiguous words. Journal of Experimental Psychology: Human Perception and Performance, 11,28-39. Swinney, D.A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, I S , 645-659. Taraban, R., & McClelland, J.L. (1988). Constituent attachment and thematic role assignment in sentence processing: Influences of content-based expectations. Journal of Memory and Language, 27,597-632.

Understanding Word and Scntence G.B. Simpson (Editor) Q Elsevier Science Publishers B.V. (North-Holland), 1991

Chapter 8 The Role of Knowledge in Comprehension: A Cognitive Control Perspective Paul Whitney and Douglas A. Waring Washington State University Pullman, Washington U.S.A. Analysis proceeds in a top-down predictive manner. Understanding is expectation based. It is only when the expectations are useless or wrong that bottom-up processing begins. Schank (1978) What readers say they expect at a certain place in a text has no effect on sense activation, or in other words, there are no top-down effects of thematic context in discourse comprehension on the sense activation phase of word identification. Kintsch and Mross (1985) To say that context does or does not influence lexical processing ignores a potentially large number of dimensions along which context may differ, and which may have profound influences on subsequent lexical processing. Simpson and Kellas (1989) The quotations above serve as clear indications of how much disagreement exists over the role of top-down processes in comprehension. If you imagine a continuum that represents comprehension processes as mainly perceptual at one end and mainly problem solving at the other end. it would be easy to find theoretical positions falling at almost any point along the continuum, including the extremes (cf., Collins, Brown, & Larkin, 1980; Gough, 1972). Our view is that there is no single correct position one could select on such a continuum. The degree to which comprehension processes are guided by pre-existing knowledge (i.e.. top-down processing) varies significantly with the reading context. The purpose of this chapter is to show that conflicting views over the importance of top-down processes emerge from a failure to fully appreciate the context-sensitivity of the processes that guide comprehension. Much of the recent research on how prior knowledge and context influence comprehension

P. Whitney and D.A. Waring

200

has centered on issues of lexical access (cf., Tanenhaus, Dell, & Carlson, 1987). We will take a broader view here and attempt to provide a unifying perspective from which it makes sense to talk about contextual effects that influence the use of prior knowledge at the lexical, sentence, and passage levels. Our thesis is that understanding the role of top-down processes in comprehension requires a cognitive control perspective. That is, our focus is on the way the readers adjust their processing of text to fit the current context. Because readers differ in their information processing abilities, they do not all make the same adjustments to context (cf., Whitney & Clark, 1989). Accordingly, our analysis will include a consideration of individual differences in comprehension. The point of examining individual differences is not just to understand individual differences for their own sake, but also to redefine what is meant by “normal” comprehension processes through understanding variations. The remainder of this chapter will be organized into Lhree parts. In order to to clarify what is meant by a cognitive control perspective, .we will first specify what is entailed in our use of the term “context.” Second, we will review different conceptions of the knowledge base and how knowledge influences comprehension. This will begin with a discussion of schema theory and then move to some recently proposed alternatives. Third, we will review research on the role of top down processing from the lexical to the passage level, and evaluate the different theories of knowledge organization in light of recent research.

CONTEXT AND

THE

COGNITIVE CONTROL PERSPECTIVE

As Simpson and Kellas (1989) noted (see above), to say that some process is or is not contextually sensitive begs some important questions about the nature of the context. A useful scheme for considering the effects of different kinds of contexts can be adapted from Jenkins’ (1979) work on the nature of contextual influences on memory. Jenkins proposed a “theorist’s tetrahedron” for organizing research on context effects. Each vertex of the tetrahedron represented different classes of contextual variables known to affect memory: subject variables, orienting tasks, criterion tasks, and materials. A similar scheme can be proposed for research on comprehension processes. Although we do not often give subjects different “orienting tasks” to perform on a text, the other three classes clearly apply as much to comprehension research as to memory research. Several kinds of subject variables could affect comprehension, including such factors as well-entrenched background knowledge and working memory capacity. Another aspect of processing context that deserves attention is the criterion used to measure comprehension. This factor could affect a reader’s goal in comprehending a text. At least some readers alter their comprehension

The Role of Knowledge in Comprehension

20 1

processes to fit the type of comprehension task used (e.g., Black, 1981; Lorch, Lorch, & Mogan, 1987). The greatest attention, however, has been given to the category of materials as context. Most of the research in this area has dealt with the effect of sentence constraints on lexical access (e.g., Kintsch & Mross, 1985; Simpson, 1984; Tanenhaus, Leiman, & Seidenberg, 1979). Our point is that it is helpful to think of the context of any experiment as a position in a space bounded by different dimensions. Therefore, when we consider the effects of “context” on some process, such as the activation of prior knowledge, we should be cognizant of the fact that several elements of context may interact to produce the effects obtained. Of course, different variables within a dimension of context may lead to different effects. As discussed below, there is evidence that different kinds of constraints within sentence materials have different effects on lexical access (Simpson & Kellas, 1989). What considering Jenkins’ tetrahedron makes salient is that there may be interactions among variables from different classes. For example, sentence constraints may affect some types of subjects differently from others. To undcrstand “context effects” we must understand something of the interactions among variables that affect a given comprehension process. Such interactions can be viewed as evidence that the information processes involved in comprehension are flexibly adapted to context in a very broad sense. (This is not to say that all the adaptations will necessarily be facilitative or optimal.) The purpose behind adopting a cognitive control perspective is to develop theories of comprehension that can capture the flexibility that readers display in adapting to different conditions. Researchers have only begun to explore many of the possible interactions. However, even in limiting the discussion to context effects on top-down processing, it is possible to show that understanding adaptations to context is crucial to understanding comprehension.

SCHEMA THEORY AND TOP-DOWN PROCESSING The Traditional View Several studies in the early 1970s seemed to indicate that readers are very constructive in their processing of text. That is, the representation formed during reading is elaborated extensively based on previous knowledge. Evidence for this assertion came from studies in which subjects falsely recognized sentences containing elaborative inferences and from studies in which possible inferences served as effective retrieval cues for previously read material (e.g, Anderson, Pichert, Goetz, Shallert, Stevens, & Trollip, 1976; Johnson, Bransford, & Soloman, 1973; Paris & Lindauer, 1976). These data were typically interpreted in schema-theoretic terms (e.g., Anderson, 1978). According to this view, the reader retrieves a generic knowl-

202

P. Whitney and D.A. Waring

edge structure, or schema, to organize and interpret a text. The schema contains slots for values that are expected to be present. Slots not filled by the text are filled with default values (an elaborative inference is made). Problems with the Traditional View

Traditional schema theoretic views of comprehension were proposed, in part, to explain the extensive use of elaborative inferences. However, the data indicating that readers are extremely constructive in reading has been qualified by recent work using priming paradigms rather than memory paradigms. It is now clear that the memory paradigms used could reflect elaborations generated during encoding or elaborations taking place at the time of retrieval when the memory test is given (Corbett & Dosher, 1978; Gumenik. 1978). In some cases priming studies have contradicted earlier research that used memory paradigms. A good example of a priming study in this domain is the work reported by Dosher & Corbett (1982) in which instrumental inferences were examined using a modified Stroop task. The experimental stimuli were priming sentences in which the instruments for actions were implied. The probable instruments served as target words for the Stroop color naming task. For example, the sentence The man dug a hole would be followed by the target shovel. Across four experiments, there were no differences in color naming after priming sentences in comparison to color naming after unrelated control sentences. In the fifth experiment, subjects were instructed to make the inferences and Stroop effects were obtained. In this case, color naming was faster after priming sentences because the target word was explicitly generated before it appeared as the Stroop target. S u m p data have also be used to show that there are limitations to readers’ use of instantiation (Whitney & Kellas, 1984; Whitney 1986). In these studies, category terms appeared in sentences in which the context biased interpretation toward a typical or an atypical member of the category (e.g., The fish attacked the swimmer- shark). When typical exemplars followed typical-biasing sentences, the Stroop effects obtained in the Whitney (1986) study were like those obtained by Dosher and Corbett (1982). That is, color naming times were facilitated in comparison to control. The standard Stroop interference effects were obtained when atypical targets followed the appropriately biasing primes. The difference in the direction of the effects seems to be related to whether the target is consciously generated (leading to facilitation) versus passively activated (leading to interference; see Whitney, 1986. Experiment 2). In contrast to the effects obtained in Whitney (1986). the earlier study (Whimey & Kellas, 1984) found no evidence for instantiation. The difference in the two studies was in the role that the category terms played in the priming sentences. Obtaining evidence for instantiation depended on foregrounding of the category terms. Foregrounding was accomplished by pronominal reference

The Role of Knowledge in Comprehension

203

to the category and, in another experiment, by having the category terms serve as the sentence topics. These studies as well as other priming work by McKoon and Ratcliff (1981) suggest that readers are more selective in their use of elaborative inferences than traditional schema theories would suggest (see Whitney, 1987, for a more complete review). Alternatives to Traditional Schema Theory Schema Assembly Theory Several researchers have proposed that the theory can be modified to be made more contextually sensitive (e.g., Abbott, Black, & Smith, 1985; Schank, 1982; Sharkey, 1986; Whitney, 1987; Yekovich & Walker, 1987). All of these various proposals can be classified under the general heading of schema assembly theory. The core idea of a schema assembly view is that schemata exist only loosely in memory but they are assembled ad hoc to meet current processing demands. Such a system can be implemented within the commonly used localist semantic network or within a parallel distributed processing network (see Rumelhart, Smolensky, McClelland, & Hinton, 1986, for a PDP version of schema assembly theory). For simplicity, we will use the localist semantic network as the architecture for our examples. If we view the knowledge base as an associative network consisting of semantic links and concept nodes, then schemata can be thought of as groups of nodes that are strongly linked such that activation of one element of the "structure" tends to activate other elements through a spreading activation mechanism (cf., Anderson, 1983). However, even within the schema, some nodes are more strongly linked than others, depending on their frequency of coactivation. Thus, there is a gradation of the tendency for the activation of a given node to activate other nodes in the knowledge base. As a result, groups of associated nodes vary in how schematic they are. That is, the memory structures known as schemata are loosely organized in memory and show variation in their degree of systematicity. Only for very stereotypic situations will the nodes be so strongly linked that the assembled structure would direct processing in the way described by a traditional schema theory. In most other processing situations, the specific nodes that actually become part of the schema that is assembled will be more context sensitive. In such a system, contextual sensitivity comes from the fact that there are two sources for activation of a node-the activation it receives based on the current stimulation and its baseline level of activation within the surrounding network structures. Some nodes will reach a critical threshold for envy to working memory very easily (hence affect ongoing processing) while others, with lower baseline strength, will require activation by several related nodes or

204

P. Whitney and D.A. Waring

direct activation from information in the stimulus environment to reach a critical threshold. Thus, the probability that some information from the knowledge base will become part of the memory representation for a stimulus depends jointly on the stimulus environment and the strength of pre-existing connections in memory. To put it another way, the schema used to understand some stimulus is assembled from the pattern of activation caused by the stimulus and directed by strengths of associations already in memory. Therefore, the schema used in a given context is a function of both past and present stimulation. The system is constantly being updated by altering the strengths of baseline connections to reflect new learning (Anderson, 1983; Rumelhart et al., 1986). From this viewpoint, schemata display short-term malleability (activation of nodes is somewhat context sensitive) and long-term evolution (they are altered by our learning experiences). How would such a system account for the data on elaborative inferences that were problematic for traditional schema theory? Consider, for example, the instantiation data reported in Whitney (1986). Recall that subjects received sentences biasing interpretation of a category term either toward a typical or atypical member. Despite equating the strength of the biasing contexts, typical exemplars were more highly activated after the appropriate context than were atypical exemplars. This apparently led to a Stroop facilitation effect for typical exemplar targets. Typical and atypical exemplars can be viewed as nodes connected with the category term, but the two types of nodes have different resting levels of activation. It thus takes a more strongly biasing context to activate atypical exemplars strongly enough to become integrated with the memory representation being formed in working memory (see Whitney, 1987). That is, the degree of instantiation depends jointly upon sentence context and pre-existing strengths of activation. The same processes may operate for more inclusive portions of the semantic network. Among those connections that together constitute the restaurant "schema," some subscenes (ie., clusters of concepts) may have lower baseline levels of activation than others and require more direct reference in the text to reach a critical threshold (cf. Schank, 1982). Only when the text follows a very common script will the connections be strong enough to activate extensive associations that strongly direct processing in a top-down manner. In most other processing contexts, the degree of top-down direction will vary from moderate to none.

Construction-Integration Theory Kintsch (1988) proposed that schema theories could be scrapped altogether in favor of the idea that meaning construction is completely situation specific. However, Kintsch argues, the process begins with a completely context-independent activation process. Similar to the schema assembly view discussed above, the construction-integrationmodel represents the knowledge base

The Role of Knowledge in Comprehension

205

as an associative net. Unlike the schema assembly view, Kintsch posits that the initial activation of the network is never knowledge or context driven. Kintsch separates the context-dependent and context-independent aspects of meaning construction into different stages. The entire process of building a mental representation of text is depicted as a series of four stages. First, a propositional representation is constructed solely on activation of individual word meanings, with no regard for context. Second, each proposition activates associated information in the knowledge base. This produces a largely incoherent and overly elaborated representation of the text segment. Third, additional inferences may be generated as needed as part of a problem solving strategy to understand the text. The resulting distributed network settles into a stable activation pattern as in other connectionist models (cf., Rumelhart & McClelland, 1986). The associative value of interconnections is a result of two factors: 1) associative values derived from proximity of propositions in the text base, and 2) the pre-existing strength of connections to concepts in the knowledge base. This fourth stage is the integration process of Kintsch’s model. This is essentially a computational process in which the strengths of positive and negative connections to the knowledge base are used to produce an integrated, coherent representation. Thus, Kintsch believes knowledge construction begins with context-independent activation and ends with a context-dependent representation computed from strengths and directions of association among concepts in the textbase and information is already in memory. The idea of initial context-independent activation fits with much of the data on access of meanings of ambiguous words in context. For example, Swinney (1979) found that lexical decisions to words associated with both meanings of an ambiguous word were speeded in comparison to control words if the lexical decision came immediately after the ambiguous word. This was true even if the intended meaning of the ambiguity was clear from the context (e.g., “The man fished from the bank” clearly does not refer to the financial institution). A few hundred milliseconds after the ambiguous word, only the target word associated with the meaning of the ambiguity intended in the context was facilitated. Similar results have been obtained by several investigators including Kintsch and Moss (1985) and Till, Mross, and Kintsch (1988). The data that indicate that lexical access is context-independent is certainly consistent with the construction integration view. However, obtaining context-independent access in some contexts is also consistent with the schema assembly theory. Thus, the issue that best distinguishes the two views, schema assembly and construction-integration, is whether top-down effects on lexical access can be obtained in some natural reading conditions. From the schema assembly view, some contexts will lead to top-down effects on lexical access. According to Kintsch’s model, no such effects are possible. In the next section, we will review current evidence on this issue and also consider how to characterize top-down effects beyond the lexical level.

P. Whitney and D.A. Waring

206

CONTEXT AND TOP-DOWN PROCESSING Studies of Lexical Access The principal evidence Kintsch cites in favor of separate construction and integration phases in lexical access is the data regarding access of homograph meanings in sentence contexts. Certainly, if the data were uniform in their support of context-independent access, there would be little to recommend even modified versions of schema theory. However, whether one finds contextdependent or context-independent lexical access depends very much on the nature of the sentence contexts used. Several recent studies have shown that when contexts are not only high in bias (the degree to which a particular meaning is suggested) but also are high in constraint (predictability of the sentence ending) the sentence contexts can lead to context-dependent lexical access (e.g., Neill, Hilliard, t Cooper, 1988; Tabossi, 1988). Another dimension of context that can be vaned is the repetition of homographs with either the same or a different meaning of the homograph used on the second occurrence (Simpson & Kellas, 1989). On the second occurrence of the homograph, meanings other than the one used earlier are quite suppressed. This is an important finding because it suggests that in natural discourse words that are repeated may show context-dependent lexical access. Thus, consistent with the schema assembly view outlined above, we cannot say that lexical access is always context-independent and free from top-down effects. Several other lines of evidence converge on the conclusion that the degree of contextual effects on lexical access varies by context. For example, Schwanenflugel and LaCount (1988) found that high constraint sentence contexts differ from low constraint contexts in the number of featural restrictions that are generated for possible sentence completions. A similar conclusion can be drawn from a series of studies by Sharkey and Mitchell (1981; 1985). They noted that many of the studies reporting minimal effects of context on lexical access used very brief and relatively weak contexts. They examined larger units of text and found that texts that were script-like influenced lexical access by making script relevant words significantly more available than script irrelevant words. We can now sum up the evidence on how context (in terms of materials) influences lexical access: The activation processes involved in lexical access may proceed exactly as Kintsch (1988) describes when contexts are relatively weak, but when contexts contain information that is strongly connected with other information already stored in memory the activation pattern is apparently schematic enough to direct activation in a top-down fashion. These results are thus completely consistent with a schema assembly view. It is worth noting here that it is the mature reader who can be character-

The Role of Knowledge in Comprehension

207

ized as flexibly and automatically adjusting the degree of top-down processing to fit the type of material being read. It is well established that younger and poorer readers are less flexible in this regard (see Stanovich, 1986). In fact, some immature readers attempt to compensate for poor lexical access skills by an extreme reliance on context to guess at upcoming words (Perfetti, 1985; Stanovich, 1980; 1984). In terms of the dimensions of context introduced above, top-down effects on lexical access vary as a function of at least two classes of variables: subjects and materials. In fact, readers of different ability levels may show different patterns of adaptation to the specifics of the materials being read (Stanovich, 1986). We will elaborate on this point in the next section. For present purposes, the point is simply that it is not accurate to conclude that the are no top-down effects on lexical access. Beyond Words: Top-Down Influences on Larger Units Much of the interest in exploring the role of knowledge in comprehension comes from the paradox that comprehension presents: Individual words are used to arrive at a sentence meaning and yet overall sentence meaning determines the meaning of individual words (cf., Lachman, Lachman, & Butterfield, 1979). The research discussed above shows that world knowledge and sentence or passage context can, under some conditions, influence lexical access processes. Of course, the influence of overall text meaning on individual words is only half of the “paradox of comprehension.” We now turn to the question of how extensively expectations and background knowledge guide the process of building sentence level and passage level representations. One important line of research relevant to this question has been mentioned already: the research on elaborative inferences. Just as in the area of topdown influences on lexical access, there are very divergent views concerning the extent to which readers incorporate information into sentence representations that goes beyond the textbase itself. Recall that the original formulations of schema theory grew out of a desire to explain why readers are very elaborative and constructive in their processing (e.g., Anderson, 1978). Other theorists have stressed the close correspondence between the textbase and the propositional representation formed for that textbase (e.g., Kintsch & van Dijk, 1978). Interestingly, much of the more recent work in this area shows a strong parallel with the research on lexical access. That is, the extent to which processing is knowledge-driven depends on how strongly the context is associated with background knowledge (McKoon & Ratcliff, 1989; O’Brien, Shank, Myers, & Rayner, 1988; Whitney, 1986; Whitney & Williams-Whitney, in press). For example, McKoon & Ratcliff (1989) gave subjects a series of texts to read and followed each with a recognition task in which they had to decide if a given target word was present in the text. Based on their previous work (e.g., McKoon & Ratcliff, 1986). they expected that responses to target words that represented

P. Whitney and D.A. Waring

208

inferences made during processing would take longer and be less accurate. This result was obtained only when the texts contained strong semantic associative information that supported the inference. McKoon and Ratcliff (1989) concluded that unless the context contains semantic associates that support a particular inference, elaborative inferences are only minimally encoded. This is reminiscent of the finding that it takes a strongly biasing context to show top-down effects on lexical access. However, McKoon & Ratcliff’s “mimimalist view” of elaborative inferences may underestimate how often background knowledge is incorporated into text representations. Just as we suggested with regard to studies of lexical access, it is important to note that there are other dimensions of context that can be manipulated which give a somewhat different view of the limitations of haowledge-driven processing. A Study of Sentence-Level Elaboration

One of these “other dimensions” that we have investigated is the effect of the readers’ goals (as set by the type of comprehension questions they are given). We recently obtained evidence that the type of comprehension question readers are given exerts some influence over the degree to which sentence representations are elaborated, although readers were surprisingly elaborative even when such inferences were completely optional (Whitney & Waring, 1989).

Table 1 Example Stimulus set from Whitney & Waring (1989).

Condition

Sentence

Typical Biasing

The fish from the mountain stream ended up in the pan.

Typical Control

The nugget from the mountain stream ended up in the pan.

Atypical Biasing

The large and dangerous fish oved slowly in the water.

Atypical Control

The large and dangerous explosive moved slowly in the water. ~~

Note: For the sentences above, the typical target was “trout,” and the atypical target was “shark.”

The Role of Knowledge in Cornprehension

209

To test for use of elaborative inferences, we inserted an implicit memory test for the inferred concept between each priming sentence and the explicit

memory test for the prime. The goals of the readers were manipulated by varying the type of explicit test they were given. Half the subjects were asked to recall the primes verbatim and the other subjects answered comprehension questions that required simple inferences. The latter were designed to encourage elaborative processing in general without specifically asking for instantiation. An example stimulus set is shown in Table 1. The implicit test used to detect inferences was a constrained word stem (CWS) completion test we have developed. In previous work (Whitney & Williams-Whitney, in press), we found that this task was a very sensitive index of inference use. Using the task, we replicated the instantiation results in Whitney (1986) and also the Dosher & Corbett (1982) study of instrumental inference.

Table 2

Proportion of targets completed by task, prime, and typicality. Verbatim Test Prime Type Target Type

Control

Typical

Atypical

Typical

0.11017

0.19917

0.155 17

Atypical

0.066 17

0.06617

0.12167

Inferential Test Prime Type Target s p e

Control

spical

Atypical

Typical

0.15483

0.29400 '

0.16055

A typical

0.088 17

0.08267

0.21050

Note: * significantly different from control (p c .05).

210

P. Whitney and D.A. Waring

The advantage of the CWS task in the present study is that subjects view the task as a distractor activity before their memory for the sentence is tested. This focused more attention on the explicit memory test and allowed us to better manipulate the readers’ goals. To further disguise the nature of the task, the experimental trials were interspersed among filler trials designed to have no relation between the sentence and probable target completions. The data from the experiment are shown in Table 2. The data were clear cut. When elaboration was encouraged by giving questions that required simple inferences, both typical and atypical exemplars were instantiated in appropriately biasing contexts. Notably, even when elaboration was discouraged by testing for verbatim memory some inferences were drawn. In particular, typical exemplars were inferred; priming of atypical targets did not reach significance. These data indicate that, to a limited extent, readers adjust their inference strategies to fit the goals of comprehension. However, the data also support the idea that some knowledge-driven processing in the form of elaborative inferences is fairly automatic. Certainly, there was no reason for subjects in the verbatim condition to elaborate as an optional strategy. If anything, such a strategy would be expected to hurt performance on the explicit memory test. A Study of Passage Level Elaboration

As a final example of the context-sensitive nature of top-down processing, we turn to an investigation of text level comprehension. Whitney and Clark (1988; 1989) examined the use of knowledge to form text representations as a function of working memory span (WMS). Virtually every current model of text processing (e.g., Kieras, 1981; Just & Carpenter, 1980) proposes that WM has the dual role of maintaining sentence-to-sentence connections (local coherence) and an overall gist representation of a text (global coherence). We reasoned that readers with poor WMS who were trying to process difficult or ambiguous prose would face a tradeoff between maintaining local coherence and global coherence. Our data showed an interesting parallel with the ambiguity research at the lexical level: Multiple interpretations are often activated and the larger context is used to narrow the interpretation. However, the extent to which knowledge directs the process of building a representation varies with the processing capacity of the reader. The data consisted of verbal protocols generated by subjects as they read successive events of several ambiguous narratives. These “thinking-out-loud” (TOL) protocols were collected after every two or three sentences of each narrative. Of course, caution should be exercised in using verbal reports as data. However, there have been a number of studies that have shown that if care is taken to avoid post hoc rcconstructions of thinking then TOL protocols can yield very rich data concerning on-line processes (e.g.. Ericsson & Simon,

The Role of Knowledge in Comprehension

211

1984; Fletcher, 1986; Olson, Duffy, & Mack, 1984).

The subjects (n=24) were given a reading span test patterned after the measure developed by Daneman and Carpenter (1980). In this test, subjects read a series of sentences for comprehension and must also maintain the last word of each sentence in memory. Subjects from the upper and lower third of the distribution were used in the study. These subjects returned for a session in which they read through stories that were ambiguous as to what was taking place. The stories were taken from previous research by Bransford and Johnson (1973) and Collins et al., 1983). Each story was divided into major events, with each event typed on a separate page. After each event, subjects described their thoughts about what was happening in the story in a procedure similar to that used by Olson, Mack, and Duffy, (1981). The verbal protocols were recorded and later transcribed. Each TOL protocol was divided into sentence-like idea units. Trained judges classified the idea units based on categories adapted from Olson et al. (1981). Over 80% of all of the idea units fell into one of three categories: general elaborations, specific elaborations, or restatements of the text. We were most interested in the pattern of use of elaborative inferences by memory span. The main difference between the two span groups was that the low span readers were much more concrete. The low span readers made specific inferences in 23% of their statements. The high spans made such inferences in only 12% of their statements. The percentage of general elaborations was almost identical for the two groups (32% and 33% for the low spans and high spans, respectively). Obviously, the low span readers made more use of background knowledge to process these texts. Additional analyses on the inferences gives striking insights into how the strategies differed between the two groups of readers. Table 3 shows the mean percentage of the subjects’ specific elaborations that occurred in the upper, middle, and lower third of each protocol. For the high span subjects there was a significant increase in specific elaborations as each passage was read. Evidently, these subjects felt more comfortable in making specific elaborations when they had more information to go on toward the end of each passage. There was no reliable trend in the data for the low span readers. None of the means shown differ from the 33% expected if the inferences were evenly distributed throughout each protocol. The data seem to suggest an open-ended hypothesis testing strategy by the high span readers but a different strategy by the low span readers. Finally, we analyzed the number of new overall interpretations introduced into each protocol. Two raters judged how many of each subject’s elaborations introduced a new thematic interpretation for a passage (with inter-rater reliability of r= .69). Interestingly, the proportion of each subject’s total inferences that introduced new thematic interpretations was about the same for the two groups: 0.35 for the low span readers and 0.31 for the high span readers. However, the standard deviation of the low span group was almost three times

P. Whitney and D.A. Waring

212

that of the high span’s, SDs= 0.14 and 0.05, respectively. The variability among the low span readers seems suggestive of a mixture of processing strategies. A large number of new themes introduced is consistent with an emphasis on local coherence and a failure to form a coherent overall representation. Making very few such inferences is consistent with a overly knowledge-driven style. Both strategies appear to be present in our sample. Half of the low span readers had proportions at or beyond the upper extreme of the high span group. The other low span readers had proportions close to or beyond the lower extreme of the high span group. In other words, although the span groups had similar means in their proportion of new thematic inferences, there was very little overlap in the distributions of scores.

Table 3 Percentage of Specific Elaborations by Position in the Protocol for High and Low Span Readers Position in Protocol Reader Group

Upper

Middle

Lower

High Span

19.6

35.3

45.1

Low Span

26.7

39.8

33.5

These data suggest that high span readers keep several interpretations in mind and test them against the text as more information is given. This strikes an adaptive balance between local and global coherence. The low span readers seem less capable of this balance so they interpret each event in relative isolation or form an overall interpretation quite early and process the remaining text in an overly top-down fashion. CONCLUSIONS

We have now examined the role of knowledge in directing comprehension at three levels - lexical, sentence, and passage, as a function of three dimensions of context - materials, comprehension tests, and processing capacities of the subjects. In each case, the data are consistent with the view that the construction of a mental representation is a very flexible process. Depending on the conditions, the construction of the representation may rely heavily on prior

The Role of Knowledge in Comprehension

213

knowledge or may proceed in a relatively bottom-up fashion. To further flesh out the details of this process we are conducting studies to examine how the cohesiveness of the background knowledge relevant to a passage affects the strategies of readers at different ability levels. We hope this line of research will forge stronger links with the data on context effects at the lexical and sentence levels. The unifying theme across these levels is that readers adjust their processing to fit the context. Studying these adjustments is the essence of the cognitive control perspective.

References Abbott. V., Black, J.B., & Smith, E.E. (1985). The representation of scripts in memory. Journal of Memory and Language, 24, 179-199. Anderson, J.R. (1983). The architecture of cognition. Cambridge, MA: Harvard University Press. Anderson, R.C. (1978). Schema-directed processes in language comprehension. In A. Lesgold, J. Pellegrino, S.Forkkema, and R. Glaser (Eds.), Cognitive psychology and instruction. New York: Plenum. Anderson, R.C., Pichert, J.W., Goetz, E.T., Schallert, D.L.,Stevens, K.W., & Trollip, S.R. (1976). Instantiation of general terms. Journal of Verbal Learning and Verbal Behavior, 15,667-679. Black, J.B. (1981). The effects of reading purpose on memory for text. In J. Long and A. Baddeley (Eds.), Attention & Performance IX. Hillsdale, NJ: Erlbaum. Bransford, J.D., & Johnson, M.K. (1973). Consideration of some problems of comprehension. In W.G. Chase (Ed.), Visual information processing. New York: Academic Press. Collins, A., Brown, J.S., & Larkin. K.M. (1980). Inference in text understanding. In R.J. Spiro, B.C. Bruce and W.F. Brewer (Eds.), Theoretical issues in reading comprehension. Hillsdale, NJ: Erlbaum. Corbett, A.T.. & Dosher, B.A. (1978). Instrument inferences in sentence encoding. Journal of Verbal Learning and Verbal Behavior, 17,479-491. Daneman, M., & Carpenter, P.A. (1980). Individual differences in working memory and reading. Journal of Verbal Learning and Verbal Behavior, 19,450466.

Dosher, B.A., 8i Corbett, A.T. (1982). Instrument inferences and verb schemata. Memory & Cognition, 10, 531-539. Ericsson, K.A., & Simon, H.A. (1984). Protocol analysis. Cambridge, MA: Bradford Books. Fletcher, C.R. (1986). Strategies for the allocation of short-term memory during comprehension. Journal of Memory and Language, 25,43-58. Gough, P.B. (1972). One second of reading. In J.F. Kavanagh and I.G.Mattingly (Eds.), Language by ear and eye (pp. 331-358). Cambridge, MA:

214

P. Whitney and D.A. Waring

MIT Press. Gumenik, W.E.(1979). The advantage of specific terms over general terms as cues for sentence recall: Instantiation or retrieval? Memory & Cognition, 7,240-244. Jenkins, J. (1979). Four points to remember: A tetrahedral model of memory experiments. In L.S. Cermak and F.I.M. Craik (Eds.), Levels of processing in human memory. Hillsdale, NJ: Erlbaum. Johnson, M.K., Bransford, J.D., & Solomon, S.K. (1973). Memory for tacit implications of sentences. Journal of Experimental Psychology, 98, 203205. Just, M.A., & Carpenter, P.A. (1980). A theory of reading: From eye-fixations to comprehension. Psychological Review, 87, 329-354. Kieras, D.E. (1981). Component processes in the comprehension of simple prose. Journal of Verbal Learning and Verbal Behavior, 20, 1-23. Kintsch, W. (1988). The role of knowledge in discourse comprehension: A construction-integrationmodel. Psychological Review, 95, 163-182. Kintsch, W., & Mross, E. (1985). Context effects in word identification. Journal of Memory and Language, 24, 336-349. Kintsch, W., & van Dijk, T.A. (1978). Toward a model of text comprehension and production. Psychological Review, 85, 363-394. Lachman, R., Lachman, J.L., & Butterfield, E.C. (1979). Cognitive psychology and information processing: An introducrion. Hillsdale, NJ: Erlbaum. Lorch, R.F., Lorch, E.P., & Mogan, A.M. (1978). Task effects and individual differences in on-line processing of the topic structure of a text. Discourse Processes, 10, 63-80. McKoon, G., & Ratcliff, R. (1981). The comprehension processes and memory structures involved in instrumental infcrcnces. Journal of Verbal Learning and Verbal Behavior, 20,671-682. McKoon, G., C Ratcliff, R. (1986). Inferences about predictable events. Journal of Experimental Psychology: Learning, Memory, and Cognition, 1 2 , 82-9 1. McKoon, G., & Ratcliff, R. (1989). Semantic associations and elaborative inference. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 326-338. Neill, W.T., Hilliard, D.V.,& Cooper, E.A. (1988). The detection of lexical ambiguity: Evidence for context-sensitive parallel access. Journal of Memory and Language, 27,279-287. O’Brien, E.J., Shank, D.M., Myers, J.L., Rayner, K. (1988). Elaborative inferences during reading: Do they occur on line? Journal of Experimental Psychology: Learning, Memory, and Cognition, 1 4 , 410-420. Olson, G.M., Duffy, S.A., & Mack, R.L. (1984). Thinking-out-loud as a method for studying real-time comprehension processes. In D.E. Kieras and M. Just (Eds), New methods in the study of immediate processes in compre-

The Role of Knowledge in Comprehension

215

hension (pp. 253-286). Hillsdale, NJ: Erlbaum. Olson, G.M., Mack, R.L., Duffy, S.A. (1981). Cognitive aspects of genre. Poetics, 10, 283-315. Paris, S.G., & Lindauer, B.K. (1976). The role of inference in children’s comprehension and memory for stories. Cognitive Psychology, 8,217-227. Perfetti, C.A. (1985). Reading ability. New York: Oxford University Press. Rumelhart, D.E., & McClelland, J.L. (1986). PDP models and general issues in cognitive science. In D.E. Rumelhart and J.L. McClelland (Eds.), Parallel distributed processing. Explorations in the microstructure of cognition. Vol 1:Foundations (110-146). Cambridge, MA: MIT Press. Rumelhart, D.E., Smolensky, P., McClelland, J.L., & Hinton, G.E. (1986). Schemata and sequential thought processes in PDP models. In J.L. McClelland and D.E. Rumelhart (Eds.), Parallel distributed processing. Explorations in the microstructure of cognition. Vol. 2 : Psychological and biological models. Cambridge, MA: MIT Press. Schank, R.C. (1978). Predictive understanding. In R.N. Campbell and P.T. Smith (Eds.), Recent advances in the psychology of language -formaland experimental approaches. New York: Plenum Press. Schank, R.C. (1982). Dynamic memory: A theory of reminding and learning in computers and people. New York: Cambridge University Press. Schwanenflugel, P.J., & LaCount, K.L. (1988). Semantic relatedness and the scope of facilitation for upcoming words in sentences. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 344-354. Sharkey, N.E. (1986). A model of knowledge-based expectations in text comprehension. In J.A. Galambos, R.P. Anderson, and J.B. Black (Eds.), Knowledge structures (pp. 49-70). Hillsdale, NJ: Erlbaum. Sharkey, N.E., & Mitchell, D.C. (1981). Match or fire: Contextual mechanisms in the recognition of words. Paper presented to the Experimental Psychology Society, Oxford. Sharkey, N.E., & Mitchell, D.C. (1985). Word recognition in a functional context: The use of scripts in reading. Journal of Memory and Language, 24, 253-270.

Simpson, G.B. (1984). Lexical ambiguity and its role in models of word recognition. Psychological Bulletin, 96, 3 16-340. Simpson, G.B., & Kellas, G.(1989). Dynamic contextual processes and lexical access. In D.S. Gorfein (Ed.), Resolving semantic ambiguity (46-56). New York: Springer-Verlag. Stanovich, K.E. (1980). Toward an interactive-compensatory model of reading fluency. Reading Research Quarterly, 16,32-7 1. Stanovich, K.E. (1984). The interactive-compensatory model of reading: A confluence of developmental, experimental, and educational psychology. Remedial and Special Education, 5 , 11-19. Stanovich, K.E. (1986). Cognitive processes and the reading problems of learn-

216

P. Whitney and D.A. Waring

ing disabled children: Evaluating the assumption of specificity. In J. Torgesen and B. Wong (Eds.), Psychological and educational perspectives on learning disabilities (pp. 87-131). New York: Academic Press. Swinney, D.A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18,645-659. Tabossi, P. (1988). Accessing lexical ambiguity in different types of sentential context. Journal of Memory and Language, 27,324-340. Tanenhaus, M.K., Dell, G.S., & Carlson, G. (1987). Context effects in lexical processing. A connectionist perspective on modularity. In J. Garfield (Ed.), Modularity in knowledge representation and natural language understanding. Cambridge, MA: MIT Press. Tanenhaus, M.K., Leiman, J.M., & Seidenberg, M.S. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18,427-440. Till, R.E., Mross, E.F., & Kintsch, W. (1988). Time course of priming for associate and inference words in a discourse context. Memory & Cognition, 16,283-298. Whitney, P. (1986). Processing category terms in context: Instantiations as inferences. Memory & Cognition, 14, 39-48. Whitney, P. (1987). Psychological theories of elaborative inferences: Implications for schema-theoretic views of comprehension. Reading Research Quarterly, 22, 299-310. Whitney, P., & Clark, M.B. (1988). Working memory and the use of inferences in comprehension. Paper presented at the Psychonomic Society meeting, Chicago, IL. Whitney, P., & Clark, M.B. (1989). Disambiguation and cognitive control. In D.S. Gorfein (Ed.), Resolving semantic ambiguity. New York: SpnngerVerlag. Whitney, P., & Kellas, G. (1984). Processing catagory terms in context: Instantiation and the structure of semantic catagories. Journal of Experimental Psychology: Learning, Memory, and Cognition, 10, 95- 103. Whitney, P., & Waring, D.A. (1989). Task effects on reader's use of elaborative inferences. Paper presented at the Psychonomic Society meeting. Atlanta, GA. Whitney, P., Williams-Whitney, D.L. (in press). Toward a contextualist view of elaborative inferences. In A.C. Graesser & G.H. Bower (Eds.), The psychology of learning and motivation. New York: Academic Press. Yekovich, F.R., & Walker, C.H. (1987). The activation and use of scripted knowledge in reading about routine activities. In B.K. Britton and S.M. Glynn (Eds.). Executive control processes in reading. Hillsdale, NJ: Erlbaum.

Understanding Word and Sentence G.B. Simpson (Editor) Q Elsevier Science Publishers B.V. (North-Holland), 1991

Chapter 9 Understanding Idiomatic Expressions: The Contribution of Word Meanings Cristina Cacciari Universith di Bologna Bologna, Italy Sam Glucksberg Princeton University Princeton, New Jersey U.S.A.

Idioms pose interesting problems for standard theories of language comprehension. Some idioms, such as by and large, appear to be nothing other than long words. The meaning of by and large is roughly equivalent to the word generally, and the idiom itself behaves as if it were a single word. Its meaning is opaque in the same way that most word meanings are opaque. Word meanings cannot be discovered; they must be learned because there is no systematic relationship between the sound of a word and its meaning, or between the individual elements of single-morpheme words and their meanings. The meanings of phrases and sentences, in contrast, can be discovered from the meanings of their individual elements. Phrase and sentence meanings are thus compositional, while most word meanings are not. If idioms are indeed simply long words, then idiom meanings must also be noncompositional. This view of idioms as noncompositional extends to transparent idioms as well as to opaque ones (Katz, 1973). Indeed, the issue of idiom transparency is moot once non-compositionality is assumed. Even idioms that may have originally been metaphors, such as carrying coals to Newcastle, are viewed as noncompositional strings. no different in kind from such opaque idioms as by and large or spic and span. The issue of compositionality has been central to competing theories of idiom comprehension. On the one hand are those theories that treat idiom comprehension as unique and different from ordinary language processes. These include the idiom list hypothesis (Bobrow & Bell, 1973), the lexicalization hypothesis (Swinney & Cutler, 1979). and the direct access hypothesis (Gibbs, 1980). Each of these approaches to idiom comprehension assumes that idioms are not compositional. Thus, the meanings of the individual elements of an idiom contribute nothing to the meaning of the idiom itself. An

218

C.Cacciari and S . Glucksberg

alternative view, expressed in Gibbs & Nayak’s (1989) decompositionality hypothesis (also see Nunberg, 1978) and in Cacciari & Tabossi’s (1988) configuration hypothesis, treats idioms as continuous with ordinary forms of language use. On this view, the meanings of an idiom’s elements may play important roles in idiom interpretation and use, depending on the particular type of idiom involved. Before considering the issue of idiom types, we will briefly review the evidence relating to the two major classes of idiom comprehension models: non-compositional, which assumes that idioms are a unique form of language, and compositional, which assumes that idioms may range from the non-compositional word-like phrase to fully compositional metaphor-like constructions (cf. Gibbs & O’Brien, in press). Idioms as Noncompositional Strings Bobrow and Bell (1973) proposed that idioms are represented in a mental idiom list, separate from and independent of the mental lexicon. When an idiom is encountered, the literal meanings of the words are first examined. Then, if the literal meanings are not interpretable in context, the idiom list is searched. If the word string is found in the list, then the listed idiom meaning is taken to be the intended meaning. Aside from the implausibility of such a cumbersome literal-first strategy, the evidence easily rejects this hypothesis. Because literal meanings are always considered first, the idiomatic senses of idioms should take longer to understand than their literal senses. This clear-cut prediction of the idiom list hypothesis is false. Familiar idioms are understood as quickly or quicker than their literal counterparts (Swinney & Cutler, 1979; Gibbs, 1980). For example, the idiomatic meanings of such idioms as kick the bucket are understood more quickly in their idiomatic sense (to die) than in their literal sense (i.e,, boot the pail). Swinney & Cutler tried to account for rapid idiom comprehension by assuming that (1) idioms are stored directly in the mental lexicon as “long words” and (2) word meanings and the compositional meanings of word strings are processed concurrently. When an idiom is encountered, its literal, compositional meanings are derived, and, if the word string matches a “long word” in the mental lexicon, then that long-word idiom is also activated. Because word recognition is usually faster than phrase comprehension, idiomatic meanings may be identified more quickly than literal phrase meanings, e.g., the idiomatic meaning of kick ihe bucker may be accessed directly before its verb phrase meaning of booting the pail. A major problem for the lexical representation hypothesis is that many idioms behave like ordinary phrases, not like words. Many idioms can undergo syntactic operations such as tense marking, e.g., one can kick the bucket now, one will kick the bucket tomorrow, or one may have kicked the bucket last week. If the string kick the bucket is merely a long word, then the element kick

Idiom Meanings

219

should not be syntactically productive. In addition, many idioms are lexically productive. Gibbs, Nayak, Bolton and Keppel (1989) found that some idioms retained their figurative meaning even when one word was substituted for another, as in crack fhe ice instead of the more usual break the ice. These phenomena - the syntactic and lexical flexibility of at least some types of idioms - effectively reject the lexical representation hypothesis as a general model of idiom comprehension. One final form of the lexical representation hypothesis merits consideration. Gibbs (1980) found that idioms in context could be understood more easily than comparable literal expressions. This basic finding led Gibbs to propose the direct access hypothesis: People can bypass literal meanings entirely, and so the most familiar, conventional meaning of a word or a phrase will be the first meaning that people will arrive at. Hence, if the idiomatic meaning of a word string is highly conventional, then that meaning will be accessed before any literal meanings. This position. of course, faces the same problems as Swinney and Cutler’s (1979) lexicalization hypothesis. Even if literal meanings could be entirely bypassed, the direct access hypothesis requires some mechanism for accessing idiomatic rather than literal meanings. What cues lead listeners to seek idiomatic meanings in the first place? The direct access hypothesis also requires an exact match between an input string and a stored idiom. Yet it is clear that different patterns of words can yield the same idiomatic meanings, e.g., “Mary was just letting off some steam,” “Some steam was let off by Mary” and “Mary let some steam off’ can all be recognized as the same idiom (Gibbs et al., 1989). Such syntactic variants arc not confined to the psycholinguistic laboratory. A recent newspaper article on the refusal of a traditional military college to admit women quoted the college dean as saying “We will continue to march to the drums we’ve been marching to” (New York Times, 1989). Activation of Literal Meanings During Idiom Processing In addition to the problems raised above, the direct access hypothesis is also inconsistent with the notion that language comprehension is non-optional. Understanding “...occurs automatically without conscious control by the listener. [We]...cannot refuse to understand...” (Miller & Johnson-Laird, 1976, p. 68). Automatic comprehension may not apply to discourse levels of language use, but it certainly applies at the word level. Stroop, for example, demonstrated that people could not ignore the meanings of printed words, even when the task was to ignore what the words meant and to report only the color of the ink they were printed in (1935). When words have more than one meaning, then the most frequent meaning of a word is always activated (Simpson, 1981; Tabossi, 1988; Tabossi, Colombo lk Job, 1987; Rayner & Frazier, 1989). Thus, even when a word string has a highly conventional meaning, as in “kick the

220

C . Cacciari and S . Glucksberg

bucket,” the most frequent meanings of the individual words would be activated in addition to the idiomatic meaning of the word string. This implies that people should not be able to suppress the literal meanings of these words, no matter how conventional an idiom may be. In addition, models that posit complete bypass of literal meanings have difficulty with the segmentation problem: What cues might be available to enable people to suspend or inhibit ongoing linguistic processing? Whether word meanings that are automatically activated play a functional role in idiom understanding is, of course, a separate question. Direct evidence on the activation of literal meanings during idiom comprehension was reported by Cacciari and Tabossi (1988). People listened to stones that could end either with a literally intended expression or an idiomatic expression. The idiom expressions could not be recognized as such until the very last word of the idiom string, e.g., an expression could be literal as in “he was in seventh place” or idiomatic as in “he was in seventh heaven.” What meanings of this expression are activated in this context? Using a cross-modal priming paradigm, Cacciari and Tabossi found that the literal meaning of heaven is activated immediately. The idiomatic meaning, related to the concept of happiness, is not activated until 300 msec later, when both the literal and idiomatic senses are active. In a later study, Tabossi and Cacciari (1988) provided contexts that were appropriate to the idiomatic meanings of target phrases. Under this condition, both the idiomatic and literal meanings were available immediately. At first glance, this finding would seem to parallel the findings for ambiguous words in general. Both the most frequent, dominant meaning and the contextually appropriate meanings are activated. However, there is one important difference between literally ambiguous words and words that are used idiomatically. In the literal ambiguous case, the dominant but contextually inappropriate meanings do not remain active for more than 200 msec or so, perhaps for as little as 120 msec (Onifeer & Swinney, 1981; Simpson, 1981). In the idiom case, both the idiomatic and literal meanings may persist for 300 msec (Cacciari & Tabossi, 1988) and in certain cases much longer than that (see below). These findings suggest that word meanings are accessed and may play important roles in the use and comprehension of idiomatic expressions. Two models of idiom comprehension are consistent with this general idea: the configuration hypothesis proposed by Cacciari and Tabossi (1988). and the decompositionality hypothesis proposed by Gibbs and Nay& (1989).

Idioms as Recognizable Configurations The configuration hypothesis treats idioms as no different from any other kind of familiar, memorized string of words. When such a string is encountered, word meanings are activated and, at the same time, the string itself can be recognized as a unit, or configuration. Very familiar word sequences, such as lines of poetry or snatchcs of songs, are prototypical configurations. Some of

Idiom Meanings

22 1

these configurations can be recognized upon hearing the first few words, so that when someone utters the words “Oh beautiful for spacious....,” most U.S.residents would immediately recognize the sequence as the opening line of the song, “America, the Beautiful.” How this utterance might be interpreted is, of course, context-dependent. If uttered while flying over the Rocky Mountains, it would probably be interpreted as a comment on this continent’s natural beauty. If uttered while walking by an unsightly, noxious landfill, it would probably be interpreted as an ironic comment on the environment. Other familiar strings may not be recognized as such until the phrase is completed, as in “Oh say, can you .....,” which can be an ordinary request, as in “....can you make it for dinner next week?”, or the opening line of the U. S. National Anthem, “...can you see.” In either case, the meanings of the individual words are activated while the configuration is being perceived and recognized. The word “take,” for example, is a lexical entry that is activated when someone utters the sentence “the boy took the book,” and this lexical entry is also activated when someone utters idioms such as “take the bull by the horns” or “take to heart” (Cacciari t Tabossi, 1988). The meanings of the individual words in configurations, whether they form idioms or not, can thus play important roles in discourse. First, they can play a role in immediate idiom comprehension. Second, they may be involved in the syntactic and lexical flexibility of idioms, i.e., they might enable people to produce syntactic and lexical variants of familiar idioms, and, of course, to understand such variants. They can play a role in a third important idiom phenomenon, semantic productivity, as when people use the semantics of an idiom’s elements to create an idiom with a new meaning, e.g. “no matter how terribly they tortured him, he didn’t spill a single bean!”. Finally, the internal semantics of an idiom’s elements should be involved in a fourth and relatively unexplored idiom phenomenon, discourse productivity. When idioms are used in conversation, responses to idioms may well play on the semantics of the words in the idiom, as in: Tom: Did the old man kick the bucket last night? Joe: Nah, he barely nudged it! In this case, the verb “nudge” can only be understood by reference to the verb “to kick.” The discourse meaning of “barely nudge” can be understood as “not even close to dying” only by analogy with “kick”: Not being close to dying is to dying as barely nudge is to kick. In cases such as this, both the idiomatic and the literal meanings of the verb “to kick” must be available for (a) Joe’s response to Tom, and (b) Tom’s ability to understand that response. Viewed in this light, Cacciari and Tabossi’s finding that both literal and idiomatic meanings are activated and remain activated during idiom comprehension makes sense. Both the literal and the idiomatic meanings can contribute to idiom interpretation and use in discourse.

222

C . Cacciari and S. Glucksberg

These examples suggest that idioms are not monolithic, nondecomposable units. Instead, they are no different from any other kind of familiar word sequence whose interpretation is context dependent and whose individual elements may play important roles in ongoing discourse. The extent to which the meanings of individual words in a sequence contribute to the meaning of the sequence itself can, of course, vary. As we saw earlier, some brief word sequences, such as the idiom by and large, seem to behave like a long word. Such sequences cannot be varied, either syntactically or lexically. In addition, responses that play on the individual words wouldn’t make sense, e.g., it is difficult to imagine a context where responses such as by and small or by or large might be interpretable. Other familiar sequences, such as pop the queslion, do seem to depend, at least in part, on the meanings of the words in the sequence. Such idioms can be varied, and can also be responded to in ways that can use the meanings of their individual elements, as in Susan: When do you think he’ll pop the question? Evelyn: I don’t know, but when he does I’ll pop the answer! Clearly, word sequences, including idioms, may differ widely in the extent to which their elements may contribute to their meanings. How shall these differences be characterized? Semantic Decomposition and Idiom Flexibility Nunberg (1978) proposed that idioms vary in the extent to which they are semantically analyzable, or decomposable. A normally decomposable idiom is one whose parts map directly onto their idiomatic referents. In the idiom pop the question, for example, the verb pop and the noun phrase the question map directly onto their respective idiomatic referents “suddenly ask” and “marriage proposal.” A nondecomposable idiom such as spic and span, in contrast, has no parts that map onto the idiomatic meaning of “perfectly neat, clean and orderly.” On this logic, idioms such as kick the bucket are nondecomposable because there is no semantic relation between any of the words in the idiom and the idiomatic meaning of die. Between these two extremes are so-called abnormally decomposable idioms, such as carry a torch and spill the beans. In idioms such as these, the idiom’s individual components relate to their referents metaphorically instead of directly, i.e., there is no clear semantic relation between beans and secrets as there is between the question and marriage proposal. In an extensive series of studies, Gibbs and his colleagues found that college students could reliably categorize idioms in terms of degree of semantic decomposition. When people were asked to classify idioms as decomposable or not, they did so reliably. The decomposable idioms were then presented a scc-

Idiom Meanings

223

ond time for grouping into normally and abnormally decomposable, and again people could classify these idioms reliably (Gibbs & Nayak, 1989). These data demonstrate that people can reliably discriminate among the three idiom types. Can differences in idiom flexibility, both syntactic and lexical, be attributed to differences in semantic decomposition? Decomposable idioms are more syntactically and more Semantically flexible than less decomposable idioms, supporting Gibbs and Nayak’s decomposition hypothesis (Gibbs & Nayak, 1989; Gibbs, Nayak & Cutting, 1989; Gibbs, Nayak, Bolton & Keppel, 1989). For decomposable idioms, syntactic variants such as “the law was laid down by John” are judged as sensible. For nondecomposable idioms, variants such as “the bucket was kicked by John” are not acceptable. Similarly, decomposable idioms retain their meanings when synonyms are substituted, as in burst the ice for break the ice. Nondecomposable idioms tend to lose their meaning when such substitutions are made, e.g., boot the bucket for kick the bucket. Phrasal idioms thus seem to vary along a continuum of compositionality (or analyzability), and the more analyzable they are, the more flexibly they behave, both syntactically and lexically. These phenomena challenge the notion that idioms have a single semantic representation that is unrelated to the meanings of their components. Even more challenging are two additional phenomena, semantic productivity and discourse productivity. Each of these has yet to be accounted for in theories of idiom representation and idiom comprehension, and it is to these two phenomena that we now turn.

Semantic Productivity Semantic productivity is the use of lexical and syntactic operations to create new idiomatic meanings from old ones. Studies of lexical flexibility, in contrast, have only examined the extent to which idioms retain their original meanings when one word is substituted for another. Without exception, studies of lexical flexibility have used lexical substitutions without any contextual or communicative motivation. For some idioms, substituting a synonym can have little if any effects on either meaning or acceptability, as in hit the hay vs. hit the sack. For other idioms, synonym substitution can render an idiom either uninterpretable, unacceptable, or both, as in kick the bucket vs. kick the pail. The functional determinants of lexical flexibility have not been studied in detail, other than Gibbs’ finding that analyzable idioms tend to be more flexible both syntactically and lexically than non analyzable idioms. Semantic productivity, as we noted above, is the ability of people to create new idiomatic meanings by changing various aspects of an idiom’s individual elements. In contrast to simple and unmotivated synonym substitutions, semantically productive operations serve communicative functions: They are

224

C . Cacciari and S . Glucksberg

motivated by communicative intentions and so they should be informative. Some relatively simple productive operations have been subsumed under the rubric of syntactic or lexical flexibility. Among these are: 1. Adjectival modification, as in “When drugs are involved. it’s time to

speak your parental mind.” 2. Adverbial modification, as in “Did hefinally speak his mind?” 3. Quantification, as in “As a diverse but purposeful group, you should speak your minds.” 4. Tense marking, as in “He spoke his mind.” 5 . All of the above, as in, “The tenants’ association finally spoke their collective minds.” What is noteworthy about this example is not only that an idiom can be semantically productive, but that this particular idiom is one of a group of nondecomposable idioms used by Gibbs, Nayak and Cutting (1989). Recall that nondecomposable idioms should tend to be both lexically and syntactically frozen, yet this idiom seems to be quite productive. This example suggests that semantic productivity may be independent of both syntactic and lexical flexibility, and it may be independent of semantic analyzability as well. Before examining these issues, however, we wanted to know whether people could understand variants of idioms that are intended to mean something different than the original idioms. To examine this issue, one of our students, William Cohen, generated a set of 27 one-paragraph stories that provided contexts to motivate a semantically modified idiom. Three kinds of semantic modifications were examined: 1. Quantification, as in “he has three left feet.” 2. Negation or antonymy, as in “he always bit off a bit less than he could chew.” 3. Miscellaneous, as in “She tended to burn her bridges ahead of her” to describe someone who insured her own future failures. Ten college students read each story with its modified idiom and provided a paraphrase of that idiom. After completing the entire set, the students provided ratings of how understandable each of the modified idioms was. Finally, the students were given the 27 idioms in their original form and asked to rate these for familiarity. One student apparently misunderstood the instructions and provided no interpretable paraphrases of the idiom variants. The other nine subjects seemed to have no difficulty and provided quite sensible paraphrases, indicating that the idiom variants were easily understood. One of these variants provides a good example:

Idiom Meanings

225

Roger always signed up for the easiest courses on campus, even though he was very smart. On hiking trips, he always chose the safest and easiest trails, even though he was in terrific shape and had years of hiking experience. He was the sort of person who always bit offmuch less than he could chew. The nine paraphrases we obtained were: He did less than he could; Roger never pushed himself; He did not challenge himself; He always did less than his potential best; He took the easiest way out; He always took the easiest way out, although capable of more; He was a person who always did less than he was capable of; He never took on challenges; and He does less than he could, just to be safe. The idiom variant (much less than he could chew) was clearly understood in the same way by all nine subjects. The understandability ratings reinforce this conclusion: On a scale of 1 to 5 (where 1 indicates that the meaning is very clear and 5 that the sentence makes no sense) this item received a mean rating of 1.3. When asked to rate how familiar the original idiom was (bite off more than one could chew), the subjects rated it as quite familiar, with a mean rating of 1.1 on a scale of 1 (very familiar) to 5 (very unfamiliar). Overall, the mean understandability rating of the idiom variants was 1.9 (SD= 0.53). indicating that people had little difficulty interpreting these idiom variants. The mean familiarity rating of the original idioms was 1.6 (SD = 0.56), indicating that the original idioms were, in general, quite familiar. As one might expect, the more familiar idioms were rated as easier to understand in their variant form. The correlation between rated familiarity and comprehension was +.36. Despite the restricted range, this correlation is reliably greater than zero (p c .05). However, even those idioms that were not familiar were generally interpreted sensibly. The idiom that was rated as least familiar by our group of subjects was To carry a torch for (someone). The variant of this idiom appeared in the following context: From the moment Jenifer met Clyde, she had been infatuated with him. Even though he seemed uninterested, she would wait outside his classes, hoping to talk to him for even a moment. She even tried to learn Italian when she heard that he liked European women. It seemed that no matter how cold and distant Clyde acted towards her, she couldn’t stop her infatuation. Finally, Clyde stopped Jenifer outside the student union and told her very directly that he could never love her. Jenifer cried for days, but finally put down her torch. The mean understandability rating for this idiom variant was 2.8; familiarity with original was rated 3.1. The paraphrases that were provided for this variant were (with the subject’s individual comprehensibility and familiarity ratings):

226

C . Cacciari and S . Glucksberg 1. Stopped thinking about him 2. no response 3. She gave up her attempt 4. She gave up (sort of a confusing analogy) 5. She stoppd being infatuated 6. She stopped liking him 7. Gave up 8. Eventually, she stopped feeling love for him 9. no response

(4; 4)

(5; 5 ) (4; 5 ) (3; 4) (1; 1)

(1; 2) (1; 1) (1; 1)

(5; 5 )

For those subjects who were familiar with the original idiom, comprehension posed no problems (Subjects 5, 6, 7 and 8). But for those subjects who were relatively unfamiliar with the original, comprehension was not that difficult. Only two subjects failed to provide any response, even though the idiom’s meaning could be inferred from context. The conclusion that we draw from this exploratory study is that idioms can be semantically productive, at least in context. This is consistent with the view that the semantics of an idiom’s elements are available, and may play a role in understanding idioms both in their original form and in contextually appropriate variants. How easy is it to understand such contextually appropriate variants? Can they be interpreted as easily as ordinary, literally intended language? To deal with this issue, another student, Matthew McGlone, asked subjects to read stories with both original and variant idioms one line at a time. Reading rate was timed, and this provided a measure of comprehension difficulty. We expected that original idioms would be understood more quickly than their variants when the context was relatively general. In these contexts, both the original and the variant would be plausible and appropriate, but there would be no particular reason to use the variant over the original. When both the original and variant were specifically appropriate, then we expected comprehension to be quite easy. Examples of general and specifically-appropriate contexts are shown in Table 1. The results of this exploratory study were clear. People can easily interpret variants of idioms, particularly when those idioms are specifically contextappropriate. With more general contexts, where the idioms were not motivated, idioms in their original form were read more quickly than their variants (mean reading time 2150 rnsec and 2360 msec, respectively). When, however, context was specifically appropriate, then both the original and variant forms were facilitated (1540 and 1780 msec, respectively). Note that the original is always faster than the variant, as would be expected on the basis of sheer familiarity, but also that the variant in the specifically appropriate condition was read at least as quickly as the original in the generally appropriate condition (2150 and 1780 msec, respcctivcly). More interestingly, both the original and variant idi-

Idiom Meanings

227

oms were read more quickly than ordinary “literal” filler sentences in comparable discourse contexts; the mean reading time for literal filler sentences in context-approriate discourse locations was 2270 msec. This is consistent with Gibbs’ (1980) finding that idioms can be understood more quickly than comparable literal expressions. What is new here is that motivated variants of idioms can also be understood more rapidly than comparable literal expressions. Apparently, the familiarity of an idiomatic expression, even in variant form, can facilitate comprehension.

Table 1

Examples of Original and Variant Idioms with General- and Specific-Appropriate Contexts. General While Sam was strolling through the city park, he happened to see his old friend Vince feeding the ducks. He and Sam had been closest friends during college, and through the years they had found time for chats in the local coffee shop. They stood and talked by the pond while Vince doled out the rest of his bread to the hungry ducks. Vince asked him about his plans for the summer.

Specific Lieutenant Sam Murphy was a pilot during the war. While conducting a reconnaissance mission he was shot down over enemy territory and captured. He was presented before one of the enemy commanders and was interrogated for details of his squadron’s attack strategy. He knew the entire battle plan, but he acted ignorant. The enemy commander threatened to kill him if the plans weren’t disclosed. Original: Sam spilled the beans. Variant: Sam didn’t spill a single bean.

These results suggest very strongly that idiom comprehension involves more than simply recognizing a string as an idiom. In addition to idiom recognition, the words comprising the idiom are processed and can be used to generate interpretations of idiom variants. In the next section, we examine how the internal semantics of idioms can play an important role in how idioms are used and responded to in ongoing discourse.

C.Cacciari and S. Glucksberg

228 Discourse Productivity

Once an idiom has been used in a conversation, the semantics of its elements may both constrain its further use and provide the bases for elaborations and responses. Recall the example of the conversation between Tom and Joe: Tom: Did the old man kick the bucket last night? Joe: Nah, he barely nudged it. The verb phrase “kick the bucket” permits responses and elaborations that play on the semantics of the verb “to kick.” Accordingly, we would expect that different types of idioms might function differently in conversation depending on the nature of the elements comprising the idiom and how those elements contribute to the idiom’s meaning. We have already seen that Nunberg’s (1978) classification of idioms into three types - normally decomposable, abnormally decomposable and nondecornposable - accounts in part for some aspects of idiom comprehension and use. Gibbs, Nayak, Bolton and Keppel (1989), for example, found that decomposable idioms - i.e., those that were analyzable in terms of literal relations between an idiom’s elements and idiomatic meaning - were more lexically flexible than nondecomposable idioms. Similarly, Gibbs and Nayak (1989) found that decomposable idioms were more syntactically flexible than nondecomposable idioms. Finally, decomposable idioms are easier to understand than nondecomposable (Gibbs, Nayak & Cutting, 1989). Analyzability, however, accounts only partially for these three phenomena. In each of the studies cited above, abnormally decomposable idioms do not differ substantially from normally decomposable idioms. In addition, we have already seen that even nondecomposable idioms, such as speak one’s mind can be semantically productive. Such idioms can also be productive in discourse, as in: Mary: Did Harry speak his mind on the bond issue? Sally: He can’t speak his mind if he doesn’t even know it yet!

In this example, the concept mind is clearly related to both the idiom’s meaning and Sally’s response, yet Nunberg’s idiom typology fails to capture these relations. At the same time, idioms do seem to differ markedly in their lexical and syntactic flexibility and, more importantly, in their semantic and discourse productivity. What differentiates among idioms that vary in flexibility and productivity?

TOWARDS A FUNCTIONAL TYPOLOGY OF IDIOMS We begin with the assumption that idioms, like other forms of natural

Idiom Meanings

229

language, are processed semantically and syntactically. This assumption fits nicely with Cacciari and Tabossi’s (1988) finding that the literal meanings of an idiom’s elements are activated and remain activated during idiom comprehension (see also Tabossi & Cacciari, 1988). This assumption also accounts for Gibbs, Nayak and Cutting’s (1989) finding that nonanalyzable idioms take longer to understand than analyzable ones. If people semantically and syntactically analyze all word strings whether they are idiomatic or not, then comprehension should be interfered with when the results of those analyses fail, i.e., have no relation to the word string’s meaning. This argument suggests an initial classification of idioms into two general types: those in which there is no apparent relation between an idiom’s elements and its meaning, and those where there is such a relation. This is basically the distinction between non-analyzable and analyzable idioms. For non-analyzable idioms (type N). semantic and syntactic analyses are nonfunctional. Accordingly, idioms such as by and large and spic and span should be neither semantically nor lexically flexible, and they should also be nonproductive in discourse. Such idioms can, in essence, be treated as special lexical entries, no different than ordinary long words where the individual morphemes do not comprise the meaning of the word itself. For analyzable idioms, some relationships between an idiom’s elements and its meaning can be discerned. In such idioms, the particular relationship may affect how an idiom might be productive. There are several ways that the “literal” meanings of an idiom’s words can map onto the meaning of the idiom itself, and idioms may be classified accordingly. Analyzable-opaque, Type A 0 In this type of idiom, the relations between an idiom’s elements and idiom meaning may be opaque, but the meanings of individual words can nevertheless constrain both interpretation and use. For the idiom kick the bucket, for example, the semantics of the verb to kick constrains interpretation as well as semantic and discourse productivity (see examples above).

Analyzable-transparent, Type AT In these idioms, there are clear semantic relations between the elements of the idiom and components of the idiom’s meaning, usually because of metaphorical correspondences between the idiom’s elements and components of the idiom’s meaning. In the idiom break the ice, for example, the word “break” corresponds to the idiomatic sense of changing a mood or feeling, and the word “ice” corresponds to the idiomatic sense of social tension. The elements of the idiom spill rhe beans similarly map onto the components of the idiom’s meaning. “Spill” corresponds to the act of revealing or letting out, and “beans”

230

C . Cacciari and S . Glucksberg

corresponds to the material that had been heretofore concealed or otherwise unknown. Included in this class are both the normally and abnormally decomposable idioms of Nunberg’s (1978) and Gibbs’ et al. (1989) classification. We see no compelling reason to distinguish between these two subtypes. Recall that normally decomposable idioms have quasi-literal relations between elements and meanings, as in pop the question, while in abnormally decomposable idioms these relations are more or less metaphorical, as in spill the beans or break h e ice. There seems to be no principled reason why this distinction should be functional either in idiom comprehension or idiom productivity, just as the general distinction between literal and metaphorical usage seems to be nonfunctional in discourse processing contexts (see Gibbs, 1984; Glucksberg, 1989; Keysar, 1989; Rumelhart, 1979). Quasi-metaphorical, Type M

In these idioms the literal referent of the idiom is itself an instance of the idiomatic meaning, e.g., giving up the ship is simultaneously an ideal or prototypical exemplar of the act of surrendering and also a phrase that can refer to any instance of complete surrender. Other examples of this idiom type include carry coals to Newcastle to refer to any instance of bringing something to a place that already has a surfeit of that something, count your chickens before they are hatched to refer to any instance of premature confidence in an outcome, and so forth. Related to these idioms are such metonymic phrases as bury the hatchet, where the action of burying a hatchet was once an actual part of the ritual of making peace, but is now used to refer to any instance of peace making in its entirety. Quasi-metaphorical idioms may serve the same communicative function as do metaphor vehicles in expressions such as “my lawyer was a shark” or “my job is a jail.” In these metaphors, vehicles such as shark or jail serve as ideal exemplars of their metaphoric categories - cutthroat people and confining, unpleasant situations, respectively - and simultaneously as names for those categories (Brown, 1958; Glucksberg & Keysar, in press). These metaphors may be used to characterize their referents by assigning them to categories that are diagnostic and often evaluative, as in “Margaret Thatcher is a bulldozer’’ (Glucksberg & Keysar, in press). Quasi-metaphorical idioms function in precisely this way. They simultaneously refer to an ideal exemplar of a concept (e.g., total surrender) and also characterize some event, person or object as a new instance of that concept, as in: Nick: Things look so bad I think I’ll drop the whole project. Alice: Don’t give up the ship. There’s too much at stake In this interchange, Alice identifies dropping the project as an instance of sur-

tdiom Meanings

23 1

render by using the idiom, don’t give up the ship. This is accomplished by implicitly grouping the two actions into the same category: Dropping the project and giving up a ship are analogues of one another, and both belong to the concept of total surrender. Total surrender, in turn, is referred to by means of an ideal exemplar of surrender, giving up a ship. This analysis suggests that idiom flexibility and idiom productivity will be governed by the functional relations between an idiom’s elements and the idiom meaning. Lexical substitutions, syntactic operations and discourse productivity should be possible whenever those functional relations are preserved. In addition, there should be some communicative or discourse purpose that is served by using an idiom in some form other than the original; a listener or reader must be able to infer a reason for the change. It thus follows that no typology, whether structural or functional, will be fully sufficient to account for idiom flexibility or productivity. The internal semantics of the idiom and the discourse context will always be the functional determinants of idiom use and variation. To illustrate this approach, we consider some variants of each of the four idiom types outlined above.

DETERMINANTS OF IDIOMFLEXIBILITY AND PRODUCTIVITY

The idiom by and large is considered non-analyzable because a semantic and syntactic analysis of the idiom and its elements fails to produce anything that is relevant to the idiom’s meaning. Hence, word substitutions are not possible because there are no similarities between an original word and its substitute that could be relevant. Syntactic operations would also be unrelated to any relevant syntactic properties of the idiom, and so no meaning-preserving syntactic operations should be possible. To the extent that there is any semantic property of an idiom’s elements and the idiom meaning, some minimal semantic productivity is possible. The word “large” bears some relation to the idiom meaning of “generally,” and so relevant modifications should be possible, as in by and not-so-large in a context that would support this qualification. Finally, discourse productivity should also rely on the semantic properties of idiom elements, and so it would be possible to use the semantics of the word “large” to reply as follows:

Ned:By and large, people are well off these days. Mark: By and not-so-large! Have you seen the figures on homelessness in America? The productive use of negation in this idiom points up another problem for the view of idioms as non-compositional strings. If an idiom is truly an

232

C. Cacciari and S . Glucksberg

unanalyzable whole, then the scope of negation - or, more generally, scope of modification - must be limited to the entire string. A negation or an adjective cannot be used to modify a semantically empty element or constituent within a string (Cruse, 1986). In some cases, modification of idiom constituents, as in break the proverbial ice can be treated as a metalinguistic comment on the expression as a whole. Nevertheless, there is a clear and important theoretical difference between such metalinguistic comments and true semantic modification, as in the by-and-large example (above), or in such cases as he broke the really frigid ice, where the concept of social lack of warmth is intensified, not merely commented on. These examples suggest that the idiom by and large i: sot a pure type N: The semantics of “large” do bear some functional relation to the idiom’s meaning, and so this idiom is partially, if minimally, compositional. Indeed, pure type N idioms may not exist at all. To the extent that a constituent of an idiom may be modified independently of the idiom as a whole, it is compositional and so can be used productively in discourse.

Type A 0 We have already provided some examples of how the analyzable-opaque idiom, kick the bucket may function in discourse. There may be no discernible semantic relation between the verb “to kick” and the idiomatic meaning of to die, yet the semantics of “kick” do constrain the idiom’s use. Consider how the semantics of “kick” and the communicative conventions of discourse govern this idiom’s use.

Lexical Flexibility Although lexical substitutions might be understood, they would not be considered apt or informative. The variants boot the bucket and kick the pail might be recognized as meaning to die (Gibbs et al. 1989). but people would be at a loss to understand why someone would use these variants. Neither the substitution of boot for kick nor pail for bucket seems motivated by any communicative purpose, and so would not be considered acceptable, unless used by a non-native speaker. In this latter case, the usage would be understood but recognized as a mistake. When near synonyms are substituted for both the verb and the noun, as in boot rhe pail. then the idiomatic meaning of 10 die is not recognized (Gibbs et al., 1989). For opaque idioms, then, lexical substitution by near synonyms is, generally, unacceptable for the following reason. To begin with, there is no relation, metaphorical or olherwise, between the semantics of the idiom’s words and components of the idiom’s meaning. Near-synonym substitutions, then. would not change or shade the idiom’s meaning. Therefore, synonym substitutions, even though understandable, would not be acceptable

Idiom Meanings

233

because there would be no communicative reason to use synonyms in place of the original wording. This joint result of opacity and communicative purposes makes opaque idioms lexically rigid.

Syntactic FlexibilitylProductivity The constraints on syntactic operations should be largely semantic and pragmatic. These constraints stem jointly from the semantics of the verb “to kick” and the concept of dying. For example, kicking is a discrete action, and so even though one can lie dying for a week, one cannot say “he lay kicking the bucket for a week.” One can, for the same kinds of reasons, say “almost, will, can, might, may, should, or didn’t kick the bucket...” The operation of joint constraints can be seen in two examples of adjectival modification, one acceptable, the other not. It would be acceptable to say “he silently kicked the bucket” because both kicking and dying can be accomplished silently. It would not be acceptable to say “he sharply kicked the bucket” because there is no way clear way to understand how anyone could die “sharply” (cf. Wasow, Sag & Nunberg, 1983). The operation of pragmatic constraints is illustrated by the non-acceptability of the passive voice for this idiom. People tend to reject “the bucket was kicked by John” as a paraphrase of “John kicked the bucket.” The communicative role of the passive form provides a good reason for not using it for such idioms. Passives are used to put focus on the object of a clause or sentence, usually when there is some prior topicalization, as in: The woman turned the corner when she was hit by a car. What happened to John? He was hit by a truck.

No such communicative purpose can be served by topicalizing “bucket,” and so the passive form is uninterpretable, i.e., the use of the passive would not be motivated. The general principle we propose is: A syntactic operation on an idiom will be acceptable if and only if it produces a comprehensible difference in interpretation, i.e., a reasonable communicative intention for the syntactic operation can be inferred. This principle, of course, applies to all idiom types. For opaque idioms, it means the passive form will not be acceptable because there would be no good reason to topicalize or focus on a grammatical or logical object. Tense markings, in contrast, would be acceptable and interpretable provided that those tense markings would be interpretable for the idiomatic meaning itself, e.g., one can die in the future and so one can also kick the bucket in the future.

Discourse Productivity The kinds of constraints illustrated above also control how idioms are

C . Cacciari and S . Glucksberg

234

used and responded to in discourse. All the constraints on lexical and syntactic operations apply for discourse. If, for example, the idiom in question is lexically rigid, then word substitutions in subsequent responses to that idiom would be unacceptable, e.g., George: Did Andy kick the bucket yet? Richard: He kicked the pail last night. Beyond the lexical and syntactic constraints on the original idiom, knowledge of the world will also govern discourse use. The permissible sequence of events in the world, for example, can control idiom use: One cannot go to heaven and then die, and so one cannot go to heaven and then kick the bucket. At this discourse level, then, all the constraints of semantics, syntactics and pragmatics will govern idiom use and comprehension.

Comprehension and use of transparent idioms, such as break the ice and spill the beans should be governed by the same principles that govern opaque idiom use. The central difference, however, is that the elements of transparent idioms can be mapped onto the components of the idiom’s meaning. Any operations that (a) respect the semantics of each element, (b) preserve the relationship between the idiom’s elements and meaning components, and (c) respect the idiom meaning itself should be acceptable and interpretable provided that a reasonable communicative intent can be inferred. Lexical substitutions should be acceptable if they satisfy these conditions, and so the following variants of break the ice should be acceptable: crack lhe ice, break the frost, break the chill. In each of these cases, the concept of abrupt breaking is preserved and the metaphorical relation between physical temperature and interpersonal warmth/ coolness is also preserved. However, why anyone would choose to say “crack the ice” instead of “break the ice” is unclear, and so most people would regard the former as, perhaps, a slip of the tongue. In contrast, using the words “frost” or “chill” does seem to imply a more personal form of coldness, not just an impersonal social awkwardness, and so these variants would be considered more apt and acceptable. Lexical variants that violate the conditions specified above should be considered unacceptable. To say “crush the ice” would be unacceptable, primarily because the kind of metaphorical ice involved in this idiom is not the kind that can be crushed: It is, metaphorically speaking of course, thin and brittle, capable of being cracked or perhaps even shattered. This example illustrates again how the semantics of an idiom’s elements can govern idiom use and productivity at the lexical level. At the syntactic level, the same principles apply: Any syntactic operations

Idiom Meanings

235

that satisfy both the semantics and pragmatics of the idiom’s elements and the idiom’s meaning would be appropriate - again with the proviso that a communicative purpose can be served (or at least inferred by listeners). Accordingly, passive transforms would be acceptable if it would be appropriate to focus on a grammatical object, as in “the ice was finally broken” or “despite days of intensive questioning not a bean was spilled.” Note that in this latter expression “bean” was used in the singular. Pluralization operations will also be acceptable if they would be appropriate for the idiom’s meaning. “Beans” in this context can be either singular or plural because “secrets” can be singular or plural: The pragmatics of the idiom’s referent are the governing factor. The “ice” in “break the ice” cannot be pluralized, because the social tension referred to by the term “ice” is a singular, momentary state. In other contexts, of course, both ice and social tensions can be pluralized. One final example makes this principle clear. Pop the question is a transparent idiom and is also considered normally decomposable (Gibbs & Nayak, 1989) because each of this idiom’s elements maps directly and literally onto the idiom meaning of utter a (marriage) proposal. Despite the “literal” nature of the idiom-meaning relationships, the term “question” cannot be pluralized. People cannot realistically “pop the questions” because people realistically do not make more than one marriage proposal at a time. Finally, transparent idioms function in discourse exactly as do opaque idioms, with the added productivity that transparency provides. Because the idiom’s elements are semantically related to the idiom’s meanings, the semantics of those elements can be exploited in discourse. The following example illustrates this kind of productivity: William: David is really weak; I bet he spilled the beans. Alice: Spill? He literally poured them all over the place! In this case the meaning of “spill” and its relation to the act of revealing permit the use of “pour” as an intensifier or emphasizer. Similarly, secrets can be “spilled” all over the place. and so it’s appropriate to spill beans that way too.

Quasi-metaphorical idioms pose an interesting set of test cases. Because they are ideal exemplars of what they represent, lexical flexibility is highly constrained. Consider carrying coals to Newcastle. Ordinary paraphrases that use arbitrary lexical substitutes should typically fail, e.g., carrying wood to Birmingham communicates nothing even close the original meaning. When, however, a communicative intent can be inferred, then well-chosen paraphrases can be effective. A recent newspaper article on the dismal failures of a nuclear

236

C . Cacciari and S . Glucksberg

generating plant at Shoreham on Long Island used the apt headline: “Carrying coals to Shoreham.” The opening phrase served to remind readers of the original idiom, and familiarity with the Shoreham nuclear dilemma made the innovative idiom’s meaning clear. Within the constraints imposed by the metaphorical nature of the idiom, lexical substitutions and variants are freely available. The verb “to carry” can be replaced by another verb so long as the action involved is consistent with the intended meaning and context. Thus, bringing to or from, sending, selling, offering, can be used as appropriate. Like any metaphor, quasi-metaphorical idioms can be tailored to suit discourse purposes. One striking example appeared in a discussion of the Democratic party’s difficulties during the presidential primaries of 1988. Jesse Jackson had done very well in the popular vote, particularly in some industrial midwestern states such as Michigan, but had absolutely no chance to win the nomination. The journalists discussed how the democrats had not adequately planned how to handle this issue at the upcoming convention, and were acting, in general, in a rather self-destructive manner. To the question “how will the Democrats handle the Jackson issue at the convention?”, one commentator quipped “Oh, I guess they’ll jump offthat bridge when they come to it” (emphasis added). Two things are worth noting about this example. First, the idiom cross the bridge when one gets to it is productive because one can substitute “jump off” for “cross” and produce, essentially, a new metaphor. Second, the reality of the literal meaning of bridge is clearly revealed by the reference to jumping - the bridge in question is not just a symbol but is real enough to be jumped from. It still is, of course, a symbol of a future event or hurdle, and so unlike a real bridge it cannot be painted, repaired, or photographed, among other things. But so long as an action is consistent with bridge as symbol of future event, that action can be used to generate a novel metaphor that is based on the original idiom. The principles that govern syntactic operations in general also apply to quasi-metaphorical idioms. Syntactic operations must be communicatively motivated, and so any changes they make in an idiom’s meaning must be interpretable in context. Consider the passive form. For most quasi-metaphorical idioms, no purpose would be served by focussing on the grammatical object. For this reason, it would make no sense to say “Newcaslle was where the coals were carried to.” This constraint, however, is not a general one. There can be metaphorical idioms that would make sense in the passive form, as in “after intensive discussions among the warring parties, the hatchet was finally buried once and for all.” In this case, the grammatical object, “hatchet,” can be the focus of the expression. The applicability of any syntactic operation will be governed by such communicative considerations. Discourse productivity will, as before, also be governed primarily by pragmatic considerations. Returning to the bridge example, one can easily imagine a context for the following interchange:

237

Idiom Meanings

Ken: Don’t worry, I’ll cross that bridge when I come to it. Ann: By that time they will have burnt it down! Here, as in earlier examples, the semantics of an idiom element (in this example, “bridge”) can be used to generate appropriate conversational responses to the original idiom. While retaining its role as symbol, “bridge” can still be treated as a real bridge so long as its symbolic function is preserved.

CONCLUSIONS Idiom use and comprehension is an integral part of everyday conversation, and so it should not be surprising to find that it is also an integral part of discourse processing. In discourse processing, the meanings of words and the compositional meanings of phrases and sentences are routinely generated and then used to infer what a speaker intends to convey. Utterances that are intended non-literally must be treated in exactly the same way as are those that are literally intended. After all, rarely could a listener (or reader) know in advance how any given utterance might be intended. The optimal strategy for discourse comprehension, then, is to treat all utterances in the same way. One strategy that could be universally employed would be to analyze all utterances semantically and syntactically, and to retain the results of such analyses at least until an intended meaning is inferred. In the context of idiom comprehension, this implies that “literal meanings” are always generated, and then used where applicable. For non-analyzable idioms, i.e., for those idioms whose elements do not contribute to idiom meaning, the search for literal meanings is non-productive and may interfere with comprehension. This would account for Gibbs, Nayak and Cutting’s (1989) finding that frozen opaque idioms (a) tend to be non-analyzable, and (b) take longer to understand than flexible, transparent idioms. The general processing strategy of applying semantic and syntactic analyses to all word strings is consistent with the general observation that people cannot ignore word and phrase meanings (cf. Miller & Johnson-Laird, 1976; Stroop, 1935). When applied to idioms, this strategy yields two products: the meaning(s) of the word string itself, and the idiomatic meaning. This, in turn, enables people to use idioms generatively in conversation. One additional source of evidence for the role of word meanings in idiom use comes from production phenomena. Slips of the tongue can reveal important language processing mechanisms. Word substitutions occur quite often in discourse, and when they do the substitutions are not random. Substituted words .are often semantically related to the “correct” or intended word (Fromkin, 1971). Examples of such substitutions in idioms include scarce as pig’s teeth for scarce us hen’s teeth, and swallow the bullet instead of bite the bullet. In

238

C.Cacciari and S . Glucksberg

each of these cases the speaker intended the meaning of the original idiom and did not notice the error. More interestingly, the listeners involved understood the intended meaning in each case, and some did not even notice the errors either, even though the substitutions were, if taken completely literally, paradoxical. Pigs have teeth, so saying that something is as scarce as pig’s teeth is to say that it is not scarce at all. Yet, because of the close semantic relation between pigs and hens (both barnyard animals), the intended and contextually appropriate idiomatic meaning of scarcity came through. Similarly, bite the buffet is a quasi-metaphorical idiom in which the action of biting a bullet is symbolic of all actions that require stoicism in the face of pain. Swallowing a bullet can only be understood by reference to the original idiom, and in the particular incident observed, some of the people involved in this interaction failed to notice the speaker’s error. Everyone, including the speaker, recognized the error when it was pointed out. These anecdotal observations should be followed up by systematic studies of idiom production, but their import is clear. People cannot isolate or ignore the meanings of words or the meanings of phrases when engaging in discourse. At the same time, people rely on familiar, memorized “chunks” of speech whose meanings derive not only from the language itself, but from their role in everyday experience. Included in this category of language are all those strings of words that we have learned, such as movie and book titles, song titles and song lyrics, poetry, proverbs, cliches, morals, and so forth. All of these have “literal” meanings, and all of these have other meanings besides. Fluent speakers of a language in a culture must be able to deal simultaneously with the language itself, and with the use of language in that culture. Like such memorized word strings as songs and poems, idioms are recognized and identified as having their own meanings, but are also treated as linguistic entities and analyzed as such. We are just beginning to understand how people can integrate these various levels of meaning, An important first step is to recognize that people do, indeed must, be able to perform such integrations in discourse. Acknowledgments

We are grateful for the financial support provided by the National Science Foundation, #BNS 8519462 and #BNS 8819657, and by the Public Health Service, #HD25826-01 to Princeton University, and to the Council for International Exchange of Scholars for a Fullbright travel grant to C. Cacciari. We thank Kay Deaux, Boaz Keysar and Susan Sugarman for their valuable comments and suggestions, and William Cohen and Matthew McGlone for their creative materials development and data collection. Correspondence can be sent to Crislina Cacciari, Dipartimento di Psicologia, Universita’ di Bologna, Italia, or to Sam Glucksberg, Department of Psychology, Princeton University, Princeton, NJ 08544-1010. USA.

Idiom Meanings

239

References Bobrow, S., & Bell, S. (1973). On catching on to idiomatic expressions. Memory & Cognition, 1 , 343-346. Brown, R. (1958). Words and things. New York: The Free Press. Cacciari, C., & Tabossi, P. (1988). The comprehension of idioms. Journal of Memory and Language, 27,668-683. Cruse, D. A. (1986). Lexical semantics. New York: Cambridge University Press. Fromkin, V. A. (1971). The non-anomalous nature of anomalous utterances. Language, 47,27-52. Gibbs, R. W. (1980). Spilling the beans on understanding and memory for idioms in context. Memory & Cognition, 8 , 149-156. Gibbs, R. W. (1984). Literal meaning and psychological theory. Cognitive Science, 8,275-304. Gibbs, R. W., & Nayak, N. (1989). Psycholinguistic studies on the syntactic behavior of idioms. Cognitive Psychology, 21, 100-138. Gibbs, R. W., Nayak, N., & Cutting, C. (1989). How to kick the bucket and not decompose: analyzability and idiom processing. Journal of Memory and Language, 2 8 , 576-593. Gibbs, R. W., Nayak, N. P., Bolton, J. L., & Keppel, M. E. (1989). Speakers’ assumptions about the lexical flexibility of idioms. Memory & Cognition, 17, 58-68. Gibbs, R. W., & O’Brien, J. E. (In press). Idioms and mental imagery: The metaphorical motivation for idiomatic meaning. Cognition. Glucksberg, S. (1989). Conversational metaphors: How are they understood? Why are they used? Metaphor and Symbolic Activity, 4, 125-143. Glucksberg. S., & Keysar, B. (In press). Understanding metaphoric comparisons: Beyond similarity. Psychological Review. Katz, J. (1973). Compositionality, idiomaticity, and lexical substitution. In S. Anderson & P. Kiparsky (Eds.), A Festschrift for Morris Halle. New York: Holt, Rinehart & Winston, pp. 357-376. Keysar, B. (1989). On the functional equivalence of literal and metaphorical interpretations in discourse. Journal of Memory and Language, 28. 375385. Miller, G. A., & Johnson-Laird,P. N. (1976). Language and perception. Cambridge, MA: Harvard University Press. New York Times, July 2, 1989. Nunberg, G. (1978). The pragmatics of reference. Indiana University Linguistics Club. Onifer, W., & Swinney, D. A. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory & Cognition, 9,225-236. Rayner, K., & Frazier, L. (1989). Selection mechanisms in reading lexically

240

C. Cacciari and S.Glucksberg

ambiguous words. Journal of Experimental Psychology: Learning, Memory and Cognition, 15,779-790. Rumelhart, D. (1979). Some problems with the notion of literal meanings. In A. Ortony (Ed.), Metaphor and thought. London: Oxford University Press, pp. 78-90. Simpson, G. B. (1981). Meaning dominance and semantic context in the processing of lexical ambiguity. Journal of Verbal Learning and Verbal Behavior, 20, 120-136. Stroop. J. R., (1935). Studies of interference in serial verbal reactions. Journal of Experimental Psychology, 18,643-662. Swinney, D., & Cutler, A. (1979). The access and processing idiomatic expressions. Journal of Verbal Learning and Verbal Behavior, 18, 523534. Tabossi, P. (1988). Accessing lexical ambiguity in different types of sentential contexts. Journal of Memory and Language, 27,324-340. Tabossi, P., & Cacciari, C. (1988). Context effects in the comprehension of idioms. Proceedings of the Tenth Annual Conference of the Cognitive Science Society, Montreal, Canada, pp. 90-96. Tabossi, P., Colombo, L., & Job, R. (1987). Accessing lexical ambiguity: Effects of context and dominance. Psychological Research, 49. 161-167. Wasow, T., Sag, I., & Nunberg. G. (1983). Idioms: An interim report. In S. Hattori & K. Inoue (eds.), Proceedings of the XIIIth International Congress of Linguistics, Tokyo.

Understanding Word and Senience

G.B. Simpson (Editor) Ca Elsevicr Science Publishers B.V. (Norh-Holland), 1991

Chapter 10 On the Combinatorial Semantics of Noun Pairs: Minor and Major Adjustments to Meaning Edward J . Wisniewski University of Michigan Ann Arbor, Michigan U.S.A.

Dedre Gentner University of Illinois Champaign, Illinois U.S.A.

The process of conceptual combination involves accessing two or more concepts and determining how they fit together to form a new concept. In a sense, conceptual combination is very broad in scope, involved in many situations in natural language understanding. For example, understanding a story probably includes the combining of concepts of the individual sentences. Understanding a sentence, in turn, probably includes combining the meanings of noun, verb, and prepositional phrases. To understand a noun, verb, or prepositional phrase, we combine the meanings of individual words. In this chapter, we will focus on how people combine concepts when they attempt to understand complex noun phrases (i.e., noun phrases other than those consisting of a noun or a determiner and a noun). For example, to understand a phrase like “elephant tie,” one might combine the concepts elephant and tie in such a way to mean, “a tie worn by circus elephants” or “a tie with a picture of an elephant on it.”l These arc possible interpretations of the phrase “elephant tie.” Recently, there has been a fair amount of psychological research on this kind of conceptual combination (e.g., Osherson & Smith, 1982; Smith & Osherson, 1984; Medin & Shoben, 1988; Murphy, 1990). Several models have been proposed to account for this process (Hampton, 1987; Smith, Osherson, Rips, 81 Keane, 1988; Cohen & Murphy, 1984; Murphy, 1988). Besides models of understanding complex noun phrases, there has also been research on how people combine the meanings of nouns and verbs in understanding sentences (Gentner & France, 1988). This chapter is organized into three parts. In the first part, we introduce

242

EJ. Wisniewski and D . Gentner

the process of conceptual combination and discuss its importance as a topic of investigation. We will outline some empirical results of studies on conceptual combination and general challenges that any theory of conceptual combination must address. In the second part, we present three psychological models of conceptual combination and evaluate them in terms of the specific psychological findings and the general challenges outlined in the first part of the chapter. The models are the attribute inheritance model (Hampton, 1987). the selective modification model (Smith et al., 1988) and the concept specialization model (Cohen & Murphy, 1984; Murphy, 1988). In the third part of the chapter, we examine one of the major assumptions of two of the conceptual combination models. Both the selective modification and concept specialization models propose a type of slotfilling as the primary mechanism for combining concepts. (A process called elaboration is also very important in the concept specialization model.) In these models, concepts are composed of slots and tillers (as in frames and schemata). One combines a pair of concepts by filling a slot in one concept (which we will call the head concept) with that of the other concept (which we will call the predicate concept)? A slot in the head concept is restricted to having the predicate concept as its filler or value. For example, to interpret a combination like red box, one finds a slot in box (i.e., the color slot) that can be filled by the concept red. The concept red box is thereby restricted to having red as the filler of its color slot. In this section, we will suggest that slot filling may be a common default strategy for combining concepts. To address this hypothesis, we discuss some results from a preliminary study that examines the kinds of descriptions that people give for noun-noun concepts. In this study, people defined novel combinations of count and mass nouns that were either artifacts or natural kinds. While we found evidence for slot filling, a number of examples from this study appear to be exceptions to the slot-filling view of conceptual combination. In order to account for the full range of results, we will propose some other mechanisms that may be involved in conceptual combination. In addition, we will argue that these mechanisms operate on much richer conceptual representations than those typically emphasized in the literature on conceptual combination. In particular, the relational structure in concepts plays an important role in how they are combined.

THE CONCEPTUAL COMBINATION PROBLEM Studying how people combine concepts is important for several reasons. First, the use of novel complex noun phrases is a very common, natural, and creative way to fill a vocabulary gap. People often introduce new terms into a language by combining existing words rather than inventing new words. For example, to denote a particular kind of table that supports computers, one might introduce the phrase “computer table,” instead of inventing a new word. Novel

Combinatorial Semantics of Noun Pairs

243

complex noun phrases are common in newspaper headlines, where they conciseh convey important information (e.g.. “Record Pentagon Procurement Overcharges Cited,” which appeared in the Washington Post on New Year’s day, 1989). The coining of novel noun phrases is evident in children at an early age (E. Clark, Gelman, & Lane, 1985). Indeed, many researchers have speculated that conceptual combination provides a route to language change and growth (Downing, 1977; H. Clark, 1983; Gentner & France, 1988). Downing (1977, page 823) suggests that the creation of novel concept combinations, “serves as the back door into the lexicon.” Second, any general-purpose natural language processing system will have to interpret complex noun phrases. Building systems that can understand such combinations is very difficult (e.g., Brachman, 1978; Cottrell. 1988; Finin, 1980; Hirst, 1983). Knowledge of how people combine concepts might assist in the development of such systems. Third, studying how concepts combine can provide a way lo constrain theories about how concepts are represented. In fact, work by Edward Smith and Daniel Osherson on conceptual combination presented a serious challenge to theories of concepts that were derived from research on single concepts (Osherson & Smith, 1981, 1982; Smith & Osherson, 1984). In particular, they demonstrated that prototype theories augmented with fuzzy set theory accounts of conceptual combination could not predict a number of findings on how people combine concepts. This research provided clues to conceptual structure that one may not have been able to discover by just studying single concepts. Based on our own studies of how concepts combine, we will also suggest how concepts should be structured. For these reasons and others, there has been increasing interest in conceptual combination. Below, we describe some recent studies. as well as a number of general characteristics of conceptual combination that make it a challenging problem. Representational Assumptions Crucial to any discussion of how concepts are combined is some notion of how they are represented. Researchers in the field have used different representations for concepts as well as different terminology for the same representation. To keep our discussion of representational issues explicit and clear, we will briefly define some terms. The term “attribute” or “feature” will refer to any property of an object that is represented in the concept of that object. So, for example, “has a pair of wings.” “is colored red,” and “flies” might be attributes or features that are represented in the concept of robin. In describing his model, Hampton (1987) uses attribute in this manner. Many researchers, however, distinguish between slots and fillers when discussing properties of objects that are represented in concepts. In slot and filler notation, the attributes

244

E.J. Wisniewski and D . Gentner

of robin would be represented as the tuples (robin part wings), (robin color red), (robin locomotion flies). The first term in each tuple is the concept name, the second term is a slot of the concept, and the third term is a filler or value of the slot. One can view slots and their fillers as dimension-value pairs of a concept. Researchers often discuss slots and fillers in the context of frames. A frame is a knowledge structure that represents one’s concept of a stereotypical situation or object (Minsky, 1975). It consists of a list of tuples that are generally true of the stereotypical situation or object. (A frame instance would represent a specific example of that situation or event-e.g., a specific r ~ b i n . )So, ~ the frame for robin would include the tuples above (as well as others). With some slight modifications (discussed later), Smith et al. (1988) use frames to represent concepts in their model. One can view a concept as being more than a list of slots and fillers, however. A structured frame consists of a structured list of slots and fillers. The list is structured in the sense that it captures various relationships between a slot and its filler and between different slots and fillers. For example, in the concept pie, the slot made-of might indicate that its filler must be something edible, capturing the fact that pies are made of edible things. As a second example, in the concept rectangle, the slot area might indicate that its filler is the product of two other fillers (namely, the fillers of the height and widrh slots). With some slight modifications, Murphy (1988) uses structured frames to represent concepts in his model. Ambiguity

Combining concepts involves three kinds of ambiguity-syntactic, lexical, and relational ambiguity. In syntactic ambiguity, the concept that a constituent modifies is ambigous. When the number of constituents is more than two, there is the “who modifies whom” problem of combining concepts. In a concept pair in English, the first concept almost always modifies the second concept. However, when there are more than two concepts, determining who modifies whom is not straightforward. The combination is syntactically ambiguous. Often, combinations are nested within other combinations. So, in solid state RCA color television, a system or person must recognize that solid modifies state and that this combination modifies television. In water meter cover adjustment screw, water modifies meter and the combination modifies cover, which in turn forms a new combination water meter cover that modifies the final concept. These examples also suggest that a model of conceptual combination must have a mechanism for the recursive processing of nested combinations. In lexical ambiguity, one or more of the meanings of the constituent words of the combination is ambiguous. Many words of English have more than one meaning and many common words have a very large number of meanings (Hirst, 1983). As an example of a phrase that has ambiguous constituents,

Combinatorial Semantics of Noun Pairs

245

consider “ball bat.” The word “ball” could mean (among other things) any of the balls used in playing sports or it could mean the kind of ball that one dances at. The constituent “bat” could refer to a type of animal or sports equipment. In relational ambiguity, the relation between the constituents is ambiguous. In this case, it is neither the individual constituents nor their syntactic relationship that is ambiguous but rather, how they fit together-their conceptual relationship. Consider the example of elephant tie above. Assume that the meanings of the constituents are unambiguous (e.g., that elephant means a type of animal and that tie means a type of clothing). Given an appropriate context, elephant and tie could be related to each other in many possible ways, as in, “a tie worn by circus elephants,” “a tie that is large like an elephant,” “a tie that has pictures of elephants on it,” and so on. In fact, constituents can be related to each other in an arbitrary number of ways that can be very plausible, given the appropriate context (Kay & Zimmer, 1976; H. Clark, 1983). Concept-dependent Combination Processes How a particular predicate concept is combined with the head concept often depends on the head concept. To see why, consider a simple. straightforward system in which concepts are combined in a way that is independent of the head concept. A particular predicate concept would be combined with any head concept in the same way. For example, the concept red might be combined with any head concept X, to mean “an X that has the color red.” Therefore, for each concept, a straightforward rule would describe how the concept combines with all others when it functions as the predicate. However, there is increasing evidence that, in human language, predicate concepts do not combine with all concepts in the same way (e.g., Halff, Ortony, & R. C. Anderson, 1976; Rips & Turnbull, 1980; Murphy, 1988; Medin & Shoben, 1988). As an example, aflower man is a man who sells flowers, a flower garden is a garden that contains flowers, aflower painting is a painting that depicts a flower, aflower necklace is a necklace made out of flowers. and so on. In these cases, how the predicate conceptflower combines with the head concept varies as the head concept varies. Medin and Shoben (1988) provide evidence from typicality ratings for this claim. Results of their second experiment suggest that, for example, gold is combined with coin to mean “a coin made out of gold” but that it is combined with railing to mean “a railing with the color of gold.” Murphy (1988) also showed that the meaning subjects gave for a simple adjective varied with the noun that it was combined with. These findings suggest that specifying the combinatorial rules of conceptual combination will not be straightforward. Typicality effects Smith and Osherson (1984) describe several findings involving the typi-

246

E.J. Wisniewski and D.Gentner

cality of adjective-noun concepts. A complete theory of conceptual combination needs to specify what it is about the structure of adjective-noun concepts that accounts for these effects. The first finding, called the conjunction effect, is that the typicality of an instance to an adjective-noun concept (i.e., a conjunction) exceeds its typicality to the noun concept. So, a particular red apple is more typical of the concept red apple than it is of apple. The second finding is the compatible-incompatible conjunction effect-that is. the conjunction effect is greater for incompatible than compatible conjunctions. Here, an incompatible conjunction is one in which the adjective denotes an unlikely filler for a slot of a noun (e.g., as in blue apple) and a compatible conjunction is one where the adjective denotes a likely filler for a slot of a noun (e.g., as in red apple). So, the extent to which a blue apple is judged more typical of blue apple than apple is greater than the extent to which a red apple is judged more typical of red apple than apple. The third finding, called the reverse conjunction effect, is that the typicality of a noninstance to an adjective-noun concept is less than its typicality to the noun concept. So a blue apple is less typical of red apple than it is of apple.

Relative Ease of Combining Concepts Some concepts are easier to combine than others. Reaction time studies suggest that several factors affect how easy it is to understand complex noun phrases. One factor is the form class of the predicate concept-in particular, whether the predicate is an adjective or noun. Murphy (1990) found that subjects understood adjective-noun pairs faster than noun-noun pairs. For example, people understood “pleasant punishment” more quickly than “bear punishment.” It is unlikely that this result could be explained by differences in the familiarity of the objects named by the phrases or in word frequency. Noun-noun pairs were used that had been previously judged as interpretable. Furthermore, the predicate nouns actually had higher word frequencies than the adjectives. One possible reason for the difference is related to the different roles that form classes play in language. In general, adjectives function as operators whose role in language is to pick out a particular slot of a noun to fill. Often, an adjective picks out the same slot of many different nouns. For example, “green” picks out the color slot of the nouns in green apple, green table, green grass, and so on. In contrast, a noun primarily serves to establish reference to individual objects or categories (Gentner & France, 1988). Using a noun as an operator violates its preferred use as a referent. An adjective-noun phrase might be easier to understand than a noun-noun phrase because both of its constituents are playing their primary roles whereas in a noun-noun phrase. the predicate noun is playing a role that violates its preferred role. A second factor that may affect ease of understanding is conceptual complexity. Murphy (1990) prefers this explanation for why noun-noun pairs are

Combinutorial Semantics of Noun Pairs

247

more difficult to understand than adjective-noun pairs. Nouns are more conceptually complex than adjectives. In general, compared to adjectives, there is more knowledge represented in nouns. To understand a noun-noun phrase, people must combine two relatively complex representations whereas to understand an adjective-noun phrase, they must combine one simple and one complex representation. Therefore, combining a pair of nouns should involve more computation. Within the class of adjective-noun concepts, Murphy (1990) has identified three factors that affect comprehension: typicality, relevance, and predication. People more quickly understand a combination containing an adjective that is typical of the noun than one containing an adjective that is atypical (e.g.. “red apple” is easier to understand than “blue apple”). (Familiarity of adjectives was controlled for by having each adjective serve both as a typical and atypical adjective;when paired with different nouns.) A combination containing an adjectivc that is relevant to the noun is easier to understand than one containing an adjective that is irrelevant. According to Murphy, an adjective is relevant if it picks out a slot that is present in the noun (i.e.. is part of the representation of the noun). So, the adjective “cold” is relevent to beer because it picks out temperature-a slot that is part of the concept of beer. In contrast, “cold” is irrelevant to garbage because temperature is not part of the concept of garbage. One must infer the fact that garbage has a temperature, presumably by inheritance from its superordinate. Therefore, the phrase “cold beer” is easier to understand than “cold garbage.**Relevance is independent of typicality. In this experiment, typicality was measured by the proportion of objects in a noun category that had the adjective property. In the example above, Murphy found that the proportion of objects in the category beer that had the attribute cold was judged to be about the same as the proportion in the category garbage that had this attribute. Finally, people more quickly understand combinations containing an adjective that has a predicating relationship to a noun than one having a nonpredicating relationship. An adjective is predicative if the combination can be mapped onto a sentence of the form noun be adjective that makes sense and reflects the meaning of the combination (e.g., Levi, 1978). So, “ugly” is a predicating adjective in “ugly painting” because the phrase can be mapped onto “The painting is ugly” - a sentence that makes sense and reflects the meaning of “ugly painting.” In contrast, “rural” is a nonpredicating adjective in the phrase “rural policeman.” The sentence “The policeman is rural” does not make sense. Within the class of noun-noun compounds, several studies have investigated differences in ease of understanding. Murphy (1990) found that context can speed the interpretation of noun-noun phrases. In one study, novel nounnoun phrases were preceded by either helpful or neutral contexts. A helpful context was one that plausibly indicated how the predicate noun was related to

248

EJ. Wisniewski and D . Gentner

the head noun. The neutral context mentioned the predicate noun and the head noun but did not indicate a plausible relation. When subjects actually read sentences containing the noun-noun phrases, they understood those embedded in the helpful contexts more quickly than those embedded in the neutral contexts. This finding suggested that when the context specified the relation between the predicate noun and the head noun, the combination process was faster. Wisniewski (1990) identified another factor that predicts how easy it will be to combine concrete, artifact nouns-namely, the functional scope of the object named by the head noun. The functional scope of an artifact refers to the range of objects that can enter into the artifact’s function. The range can be relatively unconstrained such that many objects can participate in the object’s function. For example, the functional scope of soap is unconstrained because many objects can enter into its function (i.e., many objects can be cleaned). The range can also be relatively constrained such that few objects can participate in the function. For example, the functional scope of comb is constrained because few objects can enter into its function (i.e., few objects can be straightened or styled with a comb). Qpically, functions whose scopes are unconstrained are those for which achieving the function depends on a nearly universally-applicable characteristic of objects, i.e., a characteristic that almost all concrete objects can possess. For example, the functional scope of a box is relatively unconstrained (i.e., a box can be used to contain many things). This is because the function of box depends for its achievement on a nearly universally applicable characteristic4.e.. the characteristic of occupying finite volume so that being contained is possible. Wisniewski found that people more quickly understood noun-noun phrases involving head nouns with unconstrained scopes (e.g.. “jacket box”) than those involving head nouns with constrained scopes (e.g., “jacket fork”). The results suggested that for artifact nouns, people often interpret compounds by trying to relate the predicate noun to the function of the object named by the head noun. In the case of a head noun with unconstrained functional scope, people can easily relate the predicate noun to the head noun’s function and thus interpret the noun-noun pair. But, in the case of a head noun with constrained functional scope, people must seek other ways to meaningfully relate the constituents, thus increasing comprehension time. Differential Mutability Gentner (1981) suggested that words vary in terms of their mutability. Specifically, she formulated the verb mutability hypothesis: the meanings of verbs and other predicate terms are more likely to be altered to fit the context than are the meanings of object-reference terms. In support of this hypothesis, Gentner and France (1988) examined how people paraphrased noun-verb com-

Combinatorial Semantics of Noun Pairs

249

binations that were either semantically natural or semantically strained. An example of a combination that is semantically natural is “The lizard limped.” The combination meets the requirement that an animate object (i.e., lizard) be the agent of the verb (i.e., limped). In contrast, “The lantern limped’’ is semantically strained-the agent of “limped” is not animate. Gentner and France found that when a combination was semantically strained, people paraphrased the combination by altering the meaning of the verb more than the noun. So, people might paraphrase “The lantern limped” as “The lighting device gave off a flickering light” (which alters the meaning of “limped” and preserves the meaning of “lantern”). They seldom paraphrased such combinations by altering the meaning of the noun and preserving the meaning of the verb (e.g., as in “The bright person walked lamely”). Gentner and France (1988) suggested that the nouns in a noun is a noun sentence show a similar pattern of differential mutability (though less extreme). In these sentences, the predicate noun (which functions as operator) typically adapts its meaning to the subject noun (which functions to establish object reference). Thus, “The acrobat is a hippopotamus” conveys a clumsy acrobat, whereas “The hippopatomus is an acrobat” conveys an agile hippopotamus. The interpretation of noun-noun concepts should parallel this finding. That is, the first noun should function as an operator whereas the second noun should function to establish object reference (we will discuss some exceptions to this rule later). Emergent and Interacting Features Features often emerge in concept combinations that are not present (or at least not salient) in the constituents of those combinations. Murphy (1988) found that subjects judged certain features to be typical of adjective-noun concepts but atypical of the noun or adjective concept alone. For example, people believe “lose money” is a typical feature of empty store but an atypical feature of store or empty. Murphy (1988) argued that such a feature was not a conceptual part of either the constituent empty or store but rather emerged through an interaction of the constituents and people’s general world knowledge. Gentner and France (1988) also suggested that features emerge in concept combinations that are not present in the constituents. They found that when a noun-verb combination was semantically strained, people often altered the verb’s meaning by invoking a novel meaning of the verb. For example, in one case, the sentence “The lizard worshipped,” was paraphrased as “The small gray reptile lay on a hot rock and stared unblinklingly at the sun.” In this example, the feature “stared unblinkingly at the sun” is not highly typical of either concept. Gentner and France (1988) argued that when paraphrasing the meaning of verb in such combinations, people often go beyond simply selecting from a range of prestored aspects of verb meanings. Instead, they adapted the

250

E.J. Wisniewski and D . Gentner

meaning of the verb to fit the noun’s meaning. The implication of all of these findings is that conceptual combination is not a closed operation and involves knowledge other than the lexical entries of the constituents. Features in a concept combination also interact. For example, people believe that wooden spoons are large spoons whereas metal spoons are small spoons (Medin & Shoben, 1988). In this example, the made of dimension of spoon interacts with the size dimension. One interpretation of this finding is that a combination inherits correlations contained in the head concept that are made relevant by the pralicate concept. So, the representation of spoon might contain information about the correlation “spoons that are made of wood are also large.” This correlation would influence the interpretation of wooden spoon.’ The implication of this finding is that concepts cannot be represented simply as lists of features, as implied by many past theories (e.g., Rosch & Mervis, 1975; Smith, Shoben, & Rips, 1974). Instead, concepts also capture dependencies and relations between features. For example, features may be statistically correlated (as in the spoon example above), causally connected, (e.g., has wings andflies in the concept bird), functionally related (e.g., the legs of a table typically support its top), mathematically related (e.g.. the volume of a cube is the product of its height, width, and length), and so on. A model of conceptual combination must take into account such dependencies. Concept Combinations as a Heterogeneous Class We began this chapter by suggesting that conceptual combination is very broad in scope. Restricting conceptual combination only to noun-noun and adjective-noun pairs, there are still a number of psychologically important dimensions along which such combinations can vary. First, as previously noted, there can be form class differences-the predicate term of a combination can be either a noun or an adjective. In addition, an adjective can have a predicating or nonpredicating relationship to the head noun (as described above). Combinations can be conjunctive or nonconjunctive. A conjunctive concept designates a category whose members belong to both constituent categories (Hampton, 1987). For example, the members of pet iguana are both pets and iguanas. The members of red truck are both red things and trucks. In contrast, the members of a nonconjunctive category are members of only one constituent category (that named by the head noun). So, apartment dogs are dogs but not apartments. Combinations also vary in their degree of familiarity-from well-known, lexicalized terms (e.g., “apple pie”) to novel phrases coined by eccentric writers. (For example, the counterculture author Richard Brautigan, 1967, titled the last chapter of Trout Fishing in America, “The Mayonnaise Chapter,” a reference to the fact that the chapter ended with the word “mayonnaise.”) It is assumed that novel terms are interpreted by combining the meanings of the

Combinatorial Semantics of Noun Pairs

25 1

individual constituents to form a new meaning (Murphy, 1988). A lexicalized term, on the other hand, is assumed to be interpreted by directly accessing its meaning, rather than by deriving it from the meanings of its constituents. As support for this claim, Lees (1968) suggested that as a combination is frequently used in a language, it may lose components of its initial meaning or gain aspects of meaning that are not derivable from its constituents. For example, the lexicalized term “marshmallow” originally named a type of plant that lived in marshes, then came to mean a confection made from the root of this plant and today means a soft, spongy confection made of sugar and corn syrup, roasted over camp fires. Here, the term marshmallow has lost its initial meaning (based on the constituents “marsh” and “mallow”) and gained a meaning that is not even derivable from its constituents. Summary Taken together, these phenomena suggest that an adequate theory of conceptual combination will be rather complex. In noun-noun pairs, the predicate noun can be related to the head noun in arbitrary ways. Therefore, it does not look promising that a theory can be constructed out of a set of rules that maps constituents onto a small set of relations between them. The theory must also postulate some way of choosing the appropriate meanings of constituents (i.e., resolving lexical ambiguity) as well as the appropriate relations between them (resolving relational ambiguity). For combinations that contain more than two constituents, the theory must have a mechanism for selecting which constituents to combine (i.e., resolving syntactic ambiguity). Importantly, these mechanisms will have to interact with the context surrounding a concept Combination. Finally, such a theory must also be able to represent not only the rich, internal structure of concepts but the general, world knowledge that lies outside those concepts, since this knowledge is often used to combine those concepts.

MODELSOF CONCEPTUAL COMBINATION In this section, we describe the representational and processing assumptions of three models of conceptual combination. All of these models are inrensional: A combination X Y is formed by using representations of X and Y. These theories can be contrasted with extensional theories of conceptual combination (e.g., Osherson & Smith, 1981; Zadeh, 1965) in which a combination XY is formed by intersecting the sets of the members corresponding to X and Y. Strong evidence suggests that the psychological validity of extensional theories is untenable (e.g., Osherson & Smith, 1982; Murphy, 1989). These intensional models have focused chiefly on the interpretation of adjective-noun compounds and/or noun-noun compounds (although Smith et al. have extended their model to account for adverb adjective-noun compounds

252

E.J. Wisniewski and D. Gentner

like “very red apple”). We will restrict our evaluation of these models to how well how they account for the range of psychological findings on adjectivenoun and noun-noun combinations that were reviewed in the first section. To describe each model, we first note the scope of combination phenomena that it has explicitly addressed, then discuss the model’s representation of concepts and the combinatorial processes that it postulates. Finally, we summarize evidence for and against the model. The Attribute Inheritance Model

Hampton (1987) proposed a model for how people interpret novel conjunctive concepts, such as machine vehicle, tool weapon, and sport game. These concepts are a subset of noun-noun combinations. As mentioned, the members of a conjunctive category are members of both constituent categories. So, a member of the category pet shark is both a pet and a shark. Specifically, Hampton’s model has been applied to conjunctive concepts of the form “X that is a Y” (i-e., “pet that is a shark”) rather than conjunctive noun-noun concepts. (It is not clear whether this syntactic difference would significantly affect how people process the two types of conjunctions.) In Hampton’s model, concepts are represented as lists of independcnt attributes, weighted by importance. As a default, a conjunctive concept is formed by taking the union of the attributes belonging to its constituents, and reweighting them in the resulting conjunctive concept. The weighted importance of an attribute for the conjunctive concept is a rising monotonic function of the attribute’s importance weights associated with its constituents. There are several cases in which an attribute of a constituent concept will fail to be inherited by the conjunctive concept. First, attributes that are true of one constituent but impossible or highly implausible for the other will not be included. For example, the attribute “is warm and cuddly” which is generally true of pets is a highly implausible attribute for sharks. Therefore, the conjunctive concept pet shark would not contain this attribute. Second, if the average importance of an attribute for the constituent concepts is low then it may be correspondingly low for the conjunctive concept and fail to be inherited. Third, attributes from each constituent may be incompatible or conflict with each other such that the conjunctive concept may contain one but not both. So, “lives in a domestic environment” (for the constituent per) and “lives in the ocean” (for the constituent shark) are incompatible and only one of them would be contained in pet shark. Hampton argues that the attribute that is chosen is the one that is most compatible with the other attributes of the conjunctive concept. In addition to these attribute inheritance failures, there are also situations that strongly predict that a attribute will be inherited. Specifically, an attribute that is necessary or highly probable for either constituent will also be included in the conjunctive concept, So, the attribute “has gills” would be included in pet

Combinatorial Semantics of Noun Pairs

253

shark since it is a necessary or highly probable attribute of sharks. Hampton also hypothesized that a constituent of a conjunctive concept would often dominate the other constituent in the sense that it would have more salient and important features. As a consequence, combining the constituents would result in a conjunctive concept that included more important attributes of the dominant concept. This hypothesis was based on previous studies in which Hampton (1988) noted that constituent concepts often contributed unequally to the determination of the typicality of items in their conjunctive concept. That is, for some pairs of constituents, X and Y,Hampton (1988) found that the typicality of items to the constituent X carried more weight in predicting the typicality of those items to X that is a Y than the typicality of those items to Y did. On this basis, Hampton categorized a number of concepts as dominant.

Evidence for the Model

Hampton obtained evidence supporting his model from subjects’ listings of attributes and ratings of their importance In general, attributes that belonged to (i.e., were rated as important for) either or both constituents were inherited by the conjunctive concept, supporting the model’s assumption that a conjunctive concept includes the important attributes of its constituents. Not all attributes belonging to the constituents were inherited by the conjunctive concept, however. But, in more than half of these inheritance failures, the attribute had a low average importance rating for the constituent concepts. Therefore, as predicted by the model, these attributes should not be inherited. Hampton also showed that attributes that were necessary for defining a constituent were inherited by the conjunctive concept whereas those that were impossible for a constituent were not inherited. To show this, Hampton classified an attribute as necessary for a constituent if subjects had rated it as ‘‘necessarily true of all possible examples of the constituent” and impossible for a constituent if subjects had rated it as “necessarily false of all possible examples of the constituent.” In virtually all cases, necessary attributes for one or both of the constituents were also necessary attributes for the conjunctive concept. Impossible attributes for one or both of the constituents were also impossible attributes for the conjunctive concept. To investigate whether the importance of an attribute in the conjunctive concept was a rising function of its importance in each constituent, Hampton performed regression analyses. In general, a weighted average of the constituent scores best predicted the importance of an attribute in the conjunct. Finally, Hampton showed that dominant concepts did in fact have more attributes that were important than nondominant concepts. Furthermore, in regression analyses, the importance of an attribute for the dominant concept carried more weight in predicting the importance of that attribute in the conjunctive concept.

254

E.J. Wisniewski and D . Gentner

Evaluation of the Model

Hampton’s work has shown that attributes defining a conjunctive concept are, to a large extent. reasonably predicted by a model that specifies the inheritance of attributes from the individual constituents of the conjunctive concept. Hampton has identified a number of possible factors (incorporated into the model) that determine which attributes are inherited and what their importance will be in the conjunctive concept, as well as which attributes will be excluded in the conjunctive concept. These factors include the necessity. impossibility, and importance of the features in the constituents and the dominance of a concept. Given the predictive success of the model, future work might explicitly specify the processes that underly these factors. Currently, such factors are primarily determined by subjective ratings. One example of an unspecified process in the model concerns inheritance failures of impossible attributes. Given an important attribute for one constituent, exactly what process determines that the attribute is impossible for the other constituent, and therefore impossible for the conjunctive concept? As another example, subjective ratings determine dominance of one concept with respect to another concept. What are the underlying reasons for why a concept dominates over another? The most important limitation of the attribute inheritance model is its lack of generality. Currently, it only applies to a very small set of conceptual combination phenomena of the form, “X that is also a Y,” where “X” and “Y” are nouns. Many - perhaps most - noun-noun combinations are not conjunctive (Murphy, 1988). For example, dog sleds are sleds but they are not dogs, apartment dogs are dogs but not apartments, and so on. As formulated, Hampton’s inheritance rules will not correctly apply to thcse noun-noun concepts. For example, in the case of dog sled, inheritance of necessary attributes predicts that “breathes” should be an attribute of dog sled (because dogs must breathe) whercas noninheritance of impossible attributes predicts that it should not be (because sleds cannot breathe). Inheritance of necessary attributes predicts that “has walls” should be an attribute of apartment dog whereas noninheritance of impossible attributes predicts that it should not be. and so on. In general, nonconjunctive combinations are characterizcd by attribute inheritance asymmetry between the predicate noun and head noun. Almost all of the attributes of the combination come from the head noun. For example, the attributes of apartment dog are almost all taken from dog just as the attributes from dog aparfment are almost all taken from apartment. Therefore, feature necessity and impossibility depends primarily on the head noun. Other research suggests that h e focus of combining nonconjunctive concepts is on determining a plausible relation between the conslituents, rather than on selecting attributes from the constituents. For example, slot filling models (see next section) would predict that a plausible meaning of apartment dog is “a dog that lives in apartments.” This meaning captures a relation be-

Combinatorial Semantics of Noun Pairs

255

tween the constituents, rather than an attribute that is inherited from one or the other constituent. It does not appear possible to extend the attribute inheritance approach to nonconjunctive combinations without considerable alteration. It is also unclear whether the model could be extended to encompass adjective-noun concepts. One could view many adjective-noun concepts (e.g., blue apple) as conjunctive concepts of the form “X that is a Y” (e.g., “apple that is blue colored”). However, at least in some cases, it appears that Hampton’s model incorrectly predicts that the conjunct will fail to inherit important attributes. In the blue apple example, it seems obvious that the attribute “has the color blue” should be a salient feature of blue apple. However, this attribute will be unimportant in the apple constituent (since almost all apples are red, yellow, and green) and important in the concept blue. If apple is viewed as the dominant constituent, then a weighted average of the importance of this feature for the two constituents (giving more emphasis to the importance of the attribute for apple) would predict that “has a blue color” will not be inherited by blue apple. Also, certain adjective-noun concepts are not conjunctive. As with nounnoun combinations that are not conjunctive, the focus of conceptual combination is not on selecting which features from the constituents become inherited. Consider the combination square bicycle. One might interpret this combination as a conjunction of the form “bicycle that is also square shaped.” However, our intution suggests that this interpretation does not make sense (one could not ride a bicycle that was literally square-shaped). In contrast, compare this combination to square box which probably does make sense when interpreted as box that is also square shaped. A more plausible interpretation of square bicycle is “bicycle with a square frame.” In this case, the adjective applies to part of the object named by the head concept. The feature has a square frame appears to be an emergent rather than inherited property of square bicycle (see discussion in last section). Selective Modification Model Smith, Osherson, Rips and Keane (1988) proposed a model for constructing adjective-noun concepts from individual adjective and noun concepts. More specifically, it was developed to account for typicality judgments involving adjective-noun concepts. The model has also been applied to adverb adjectivenoun concepts (e.g., very red apple). In general, the model postulates that the adjective directs the formation of the adjective-noun concept by restricting the filler of a noun slot to the adjective concept and by increasing the diagnosticity of this slot. To take a simple example, the adjective “green” would direct the formation of the concept green apple by restricting the color slot of green apple to the filler green and by increasing the diagnosticity of the color slot? The model has three major characteristics. First, nouns and adjectives are

E.J. Wisniewski and D . Gentner

256

viewed as simple frames which consist of the typical properties associated with their instances. For example, a frame for apple would contain typical properties such as “has a round shape,” “has a smooth texture,” and “has the color red.” More specifically, the frame consists of a list of slots and a set of fillers associated with each slot. Each slot has a diagnosticity value and each filler of the slot has a salience score. The diagnosticity of a slot measures how useful the slot is in discriminating examples of the category from examples of contrast categories. The salience score of a filler of a slot reflects its subjective frequency among examples of the category as well as its perceptibility. For ease of explaining the model, the salience score is viewed as the “number of votes” for a particular slot filler. Figure 1 (adapted from Smith et al., 1988) shows examples of the apple, brown. and brown apple frames. Second, the model proposes a mechanism for operating on adjective and noun frames to produce an adjective-noun frame. The mechanism operates as follows. The slots of the adjective frame select the corresponding slots of the noun frame. For each of these slots in the noun frame. there is an increase in the salience of the filler indicated by the adjective and a decrease in the salience of other fillers (i.e.. votes get shifted to the filler from the other fillers). In addition, there is an increase in the diagnosticity of the slot.

I

Bmva Appl. Diag. Slot Filler Salienca 1.5 Color: red 0 grccn 0 bmvn 30 0.5

shape: round g-

cyllndrlcal

0.5

Texhus: s m ~ t h

mugh bumpy

Figure 1 . Using slot selection (or slot filling) the selective modification model.

LO

15 0 5 5 5

0

combine brown and apple to form brown apple, in

Combinatorial Semantics of Noun Pairs

257

Figure 1 also illustrates the operation of this mechanism for the frames brown and apple. Here, the brown frame contains a single slot (color) and it selects that slot in the apple frame, increasing the salience of the filler brown and the diagnosticity of the slot color, in brown apple. The votes for the other fillers of color are shifted to the filler brown. Although the model is not stated in these terms, one can view selection as slot filling. By selecting a slot in the noun frame, the adjective frame essentially restricts or limits the slot to having the adjective concept as its filler. The third characteristic of the model is a process for computing the typicality (or similarity) of an instance with respect to its frame, using a slightly modified version of Tversky's (1977) contrast rule. According to this rule, typicality is an increasing function of the votes for an slot filler that are common to the instance and frame and a decreasing function of the votes for a slot filler that are distinct to the frame and distinct to the instance: Typicality (I, F) = Xi [af, (I

+ F)- bf, (F - I) - cfi (I - F)

where I is the instance, F is the frame, i indexes the slots, I + F is the set of votes of the slot fillers of slot i that overlap in the instance and frame, F - I designates the set of votes of the fillers of slot i that are distinct to the frame, and I - F designates the set of votes of the fillers of slot i that are distinct to the instance. The parameters a, b, c determine the relative contributions of these sets. Figure 2 shows an example of how the typicality of an apple instance to the apple frame has been computed (a, b, and c all have the value 1). In this example, for the slot color, the instance and frame have 25 overlapping votes for the filler red. The frame, in turn, has 5 votes for the filler green that are distinct, and the instance has 5 votes of the filler red that are distinct. These sets are then multiplied by the diagnosticity value of the color slot. The equation is applied to the other slots in a similar manner. Apple Instance Sbt

Color:

Filler

Apple

Salience

red

enen

brow

30 0 0

1.0

20

0.50

Shape. round g-

0

Texture' SmoOrll rough

30 0

cyllndncal

bumpy Typicality

* (25 - 5 - 5 ) * C

Color:

Franc Filler

Salience

ad enen

brow

25 5 0

Shape: roand

15

0 cyllndncal5 SQ-

0

0 =1

Diw. Slot

0.25

TeXhm: S m o o a rough bumpy

25 5

0

;0*(15-5-5) +0.25*(25-5-5)=21.15 Figure 2. Computing the typicality of instance of apple to apple in the selective modification model.

258

E.J. Wisniewski and D . Gentner

The selective modification model was developed primarily to account for typicality findings involving adjective-noun concepts. As discussed in the first section of this chapter, there were three major findings. The first finding was the conjunction effect. In some cases, the typicality of an instance to an adjective-noun concept exceeds its typicality to the noun concept (e.g., a brown apple is more typical of brown apple than of apple). The second finding is the compatible-incompatible conjunction effect. That is, the conjunction effect is greater for incompatible than compatible conjunctions. So, the extent to which a blue apple is judged more typical of blue apple than apple is greater than the extent to which a red apple is judged more typical of red apple than apple. The third finding was the reverse conjunction effect. In some cases, the typicality of a noninstance to an adjective-noun concept is less than its typicality to the noun concept. So, a blue apple is less typical of red apple than it is typical of apple. The model accounts for these findings in a straightforward way. For example, to account for the conjunction effect, consider the apple and brown apple frames (shown in Figure 1). Notice that both frames are identical except for the color slot. This slot has been modified in brown apple (during conceptual combination) such that the number of votes for the filler brown (Le., its salience) and the diagnosticity of the slot color are greater in brown apple than in apple. Also, the votes for the other color fillers are less in brown apple than in apple. Note that more votes of the color fillers for a particular brown apple will match those of brown apple than apple, and the color slot will have greater diagnosticity in brown apple than in apple. Furthermore, more votes of the color slot for the particular brown apple will mismatch those of apple than brown apple. Therefore, using the equation above, a particular brown apple will be more typical of brown apple than apple. As a second example, to account for the reverse conjunction effect, note that a particular brown apple will mismatch a highly salient filler (i.e., red) in red apple and that this mismatch will be increased (relative to apple) by the increased diagnosticity of color for red apple. Therefore, the particular brown apple will be less typical of red apple than it will be of apple. The selective modification model has also proposed a mechanism for combining adverbs with adjectives and nouns. In particular, the model has examined adverbs that intensify or diminish aspects of concepts. For example, an adverb like “very” appcars to increase the filler of a slot. In “very red fruit,” “very” increases the redness of red fruif. Other adverbs like “slightly” and “non” appear to decrease a slot’s filler, In “slightly red fruit” and %on red fruit,” the adverbs decrease the redness in red fruit. In the model, these adverbs function as scalars that multiply the salience scores of slot fillers. The adverb “very” is a scalar greater than 1, “slightly” is a scalar between 0 and 1, and “non” is a scalar less than or equal to 0. For example, “very” would multiply the votes for red in red apple by some scalar greater than 1. Thus, very red apple would have more votes on red than red apple would.

Combinatorial Semantics of Noun Pairs

259

Evidence for the Model Smith et al. empirically tested their model by collecting data that allowed them to derive the frames of various concepts and to predict typicality ratings of instances to those frames. They then compared the predicted typicality ratings of the model with actual typicality ratings of subjects. Specifically, Smith et al. had subjects list slot and filler pairs (e.g., coforred) for various examples of vegetables and fruits (e.g., onion, carrot, peach, apple). Using these data, they constructed frames for vegetable and fruit. The number of subjects who listed a particular slot-filler pair was taken as the salience or number of votes for that slot-filler pair in the concept. Diagnosticity values for each slot were estimated by measuring the extent to which the fillers of that attribute were associated with vegetable but not fruit (or vice versa). Smith et al. then calculated the typicality of the examples to fruit, vegetable and to the eight adjective-noun combinations of the four adjectives “red,” “white,” “round,” and “long” with “fruit” and “vegetable.” Next, they compared these typicality ratings predicted by the model with those given by another group of subjects. The predicted values were highly correlated with the subject ratings for most of the adjective-noun concepts that were tested. Smith et al. also calculated the model’s typicality ratings of examples to adverb adjective-noun concepts that paired the adverbs, “very,” “slightly,” and ‘‘non” with the adjective-noun concepts above, and compared them to subjects’ ratings. They generally found reasonable correlations between obtained typicality ratings and ratings predicted by the model.

Evaluation of the Model The selective modification model has a number of strengths. First, it postulates a clear, well-specified combinatorial mechanism. Second, Smith and his colleagues have carefully tested the model’s predictions by comparing them to psychological studies of how adjectives and nouns are combined, as well as how some adverbs, adjectives, and nouns are combined. One limitation of the current model is that in general, it cannot be applied to noun-noun concepts. To understand why the model would have difficulty with constructing a noun-noun frame from two noun frames, consider one of the major processes in the model: the slots of the adjective select the corresponding slots of the noun and modify them in the combination. To use one of Smith et al.’s examples, in forming the combination shriveled apple. the slots of the adjective shriveled (texture and shape), would select the corresponding texture and shape slots in the noun apple. These slots would (and probably should) be modified in the shriveled apple frame. If we apply an analogous process for combining two noun frames, the slots of the predicate noun should select the corresponding slots of the head noun and modify them in the combi-

260

E.J. Wisniewski and D . Gentner

nation. Clearly, however, many of the slots of the head noun frame that correspond to those in the predicate noun should not be modified. For example, in dog house, likely corresponding slots might be texture. shape, and size (to name a few). We would not want the dog house frame to contain dog’s fillers for texture, shape, and size. Dog houses are not furry and shaped like dogs, and they are certainly larger than dogs. Thus, it does not appear possible to extend the selective modification model to noun-noun combinations without considerable alteration. This limitation may not be not overly important in evaluating the model, however. One might argue that there is form-class specificity in how concepts are combined and that the selective modification model addresses how adjective-noun concepts are formed. Other processes are involved in noun-noun combinations. This view is reasonable, given that the primary roles of adjectives and nouns are different. Adjectives inherently are predicate terms whereas nouns primarily are used as object referents and only secondarily used as predicate terms. More seriously, the generality of the model’s assumptions about adjective-noun combination has been called into question. First, the model assumes that adjectives and nouns are combined using a closed operation-an adjective modifies slots within the noun or adds slots to the noun that are within the adjective. Such an assumption is contradicted by Murphy’s (1988) finding that subjects judged certain attributes to be considerably more typical of adjectivenoun concepts than of either the noun or adjective concept alone. These results imply that such attributes may not be present in either the adjective or the noun concept but rather emerge through an interaction of the noun and adjective concepts and general, world knowledge. (Gentner and France (1988) found similar results with noun-verb combinations.) Second, the selective modification model treats the attributes of a concept as independent. Hence, it predicts that modifying one attribute of a concept should not affect other attributes. However, recall Medin and Shoben’s finding that attributes in concept combinations can be correlated (e.g., wooden spoons tend to be large whereas metal spoons tend to be small). This finding suggests that an adjective may affect more than one slot in a noun. For example, in large spoon, “large” not only determines the filler of the size slot of spoon but also determines the filler of its made-of slot. (Some of the Medin and Shoben results are open to an alternative interpretation, provided by Smith and Gray (1990)see footnote 3). Besides this specific finding, there is substantial evidence that relations between attributes are very important and that in general, many concepts are best represented as having complex, relational structure (Gentner, 1975, 1981, 1983, 1989; Gentner & Clement, 1988; Gentner & France, 1988; Goldstone, Gentner, & Medin, 1989; Malt & Smith, 1984; Medin, Altom, Edelson, & Freko, 1982; Medin & Shoben, 1988; Medin & Wattenmaker, 1987; Murphy & Medin, 1985; Murphy & Wisniewski, 1989a, 1989b; Palmer, 1978;

Combinatorial Semantics of Noun Pairs

26 1

Norman & Rumelhart, 1975). Smith et al. also note that the model does not apply to several kinds of adjective-noun concepts. First, in some combinations, the adjective indicates a slot that is normally not a slot of the noun. For example, “upside-down” cues the slot orientation which would not normally be associated with fruit. Thus, upside-downfruit could not be formed by selecting the appropriate slot in fruit. Presumably, some additional mechanism would add the slot to the noun (along with its diagnosticity, values, and the values’ saliences). Second, they note that some adjectives can have complex effects on the nouns that they combine with. For example, “fake” in “fake apple” leaves some slots of apple intact (e.g., shape, color) but negates others (e.g., taste, edibleness). In the next section, we will make a related claim that mass nouns (e.g., glass, chocolate) have complex effects on their head nouns when they play the role of the predicate term in a combination. For all its admirable explicitness, the selective modification model does not present a complete picture of adjective-noun combination. It may be more accurate to say that Smith and his colleagues have identified one important process that operates when adjectives and nouns combine. Namely, when an adjective combines with a noun, it may select one or more slots of the noun, changing their diagnosticity and the saliences of their possible fillers. It appears that other mechanisms or representational assumptions are needed to specify how general knowledge affects the combination process and how filling a slot affects other slots in the concept. There is some suggestive evidence that selective modification may be a first stage in adjective and noun combination, with other processes operating later (Smith & Gray, 1990). Concept Specialization Model The concept specialization model was explicitly formulated to account for both the interpretation of adjective-noun and noun-noun concepts. As with the selective modification model, slot filling is an important mechanism in the concept specialization model (Cohen & Murphy, 1984; Murphy, 1988). Importantly however, the model hypothesizes a second process that operates in conceptual combination. This process, called elaboration, is driven by people’s general background knowledge that lies outside the concepts being combined. The model hypothesizes a richer representation for concepts than the first two models. The formulation of the concept specialization model was influenced by an A1 model called KL-ONE (Brachman, 1977; 1978; 1979). This model represented concepts as structured sets of slots and fillers and it proposed slot filling as one of the primary mechanisms for combining concepts. (Brachman called the mechanism slot restriction rather than slot filling.) The concept specialization model also incorporates thcse characteristics but differs in that it is being

262

E.J. Wisniewski and D.Gentner

proposed as a psychological model of conceptual combination. Therefore, some characteristics of the model reflect findings in the psychological literature on concepts. In this model, a concept is a structured set of slots. A slot specifies a default filler and a list of other possible fillers that may fill the slot, weighted by typicality. The slots are structured in the sense that relations between slots and between slots and their possible fillers are also represented. Although this aspect is not spelled out in detail, fillers of different slots may be statistically correlated, and may have causal, numerical, functional, or logical dependencies between them. Concepts are organized into hierarchies and can inherit slots from concepts higher in the hierarchy. Several of the representational aspects of the model are designed to capture concept typicality. In particular, the possible fillers of a slot and the subconcepts of a concept are ordered by typicality. The typicality of a subconcept to its parent concept (e.g., the typicality of robin to bird) is measured by computing the degree of family resemblance, after Rosch and Mervis (1975). According to Cohen and Murphy, family resemblance is calculated by counting the number of slots and slot fillers that a subconcept shares with the other subconcepts of the parent and subuacting the number it shares with non-subconcepts of the parent. Thus, subconcepts that have many slots and fillers in common with each other and few in common with non-subconcepts will be more typical of the parent concept. Concepts are combined using a two-stage process. The first process is similar to that outlined in the selective modification model. Here, a combination is created by filling one of the slots of the head concept with the predicate concept. For example, to interpret elephant box, one would fill a slot in box (e.g., a slot like contains) with the predicate concept elephant. So, elephant box might be interpreted as “a box that contains elephants.” Figure 3 illustrates the slot filling process for elephant box. There are several ways of determining which slot to fill. First, a predicate term may be listed in the head concept as one of the possible fillers for a slot. If the predicate term is a possible filler of more than one slot, then presumably the slot of which it would be more typical is selected. Second, context (e.g., a discourse setting) can drive the slot-selection process, by activating a slot in the head concept. So, during a discussion of washing and the mention of a phrase like “finger cup,” a slot will be activated that reflects a “cup used for washing fingers” interpretation of finger cup. Third, one may use general knowledge to determine the best slot. Especially for novel combinations, the predicate term may not be listed as a possible filler for a slot and discourse may be insufficient for selecting an appropriate slot. The second process is called elaboration and it involves refining and augmenting the combination, using world knowledge (Murphy, 1988). This knowledge is used to infer other likely characteristics of the combination. To continue the example above, one might reasonably conclude that an elephant box is

Combinatorial Semantics of Noun Pairs

263

larger than the usual box and augment the combination with this fact (probably by filling the sire slot of elephant box with large). One also might conclude that an elephant box is sturdier than the usual box and therefore is made of wood rather than cardboard. In Figure 3, elephanl box has been elaborated to reflect these conclusions. Elaboration may be based on some type of plausible reasoning process (Collins, 1978; Collins & Michalski, 1989). It may also involve recalling examples of the head concept. So, one might recall examples of boxes that contained elephants and use them to refine and augment the meaning of elephant box. Hampton (1985) calls this process extensional feedback (see also Cohen & Murphy, 1984). It involves accessing knowledge of actual objects in the world.

Elephant Slot: .Filler(s): Color. P Y Shape: elephant-like size large park, tusks, trunk Habitat: ZOO, savannzd

+

Elephant Box

BOX

Slot. Color

Piller(s) brown

Shape

cube-like

Contains Made-of. cardboard Slze medium

Color: Contans.

Size. Made4f:

I: Figure 3. Using slot filling and elaboration concept specialization model.

10 combine

elephant

brown

I

large (elaborat vood

I

and box to form elephant box. in

the

Besides the processing assumptions of slot filling and elaboration, the model also assumes that conceptual combination is typically asymmetrical: a combination of the form XY is not at all the same as one of the form YX. So, for example, an apartment dog is not the same as a dog apartment. This assumption has been emphasized in order to contrast the model with extensional models of conceptual combination, in which an XY combination would be formed by intersecting the sets corresponding to X and Y . By commutativity of set intersection, this view would predict that an apartment dog is the same as a dog apartment. The asymmetry of conceptual combination may be due to the different roles that the predicate and head nouns play (Gentner & France, 1988). In an XY combination, the meaning of X is more mutable (because X functions as an operator) and the meaning of Y is more stable (since it serves to designate the referent of an object). The reverse is true for a YX combination. Therefore, the meaning of XY will be different from YX . Evidence for the Model

Murphy (1988, 1990) details several studies that provide support for the general assumptions of the model. One line of support for the use of general

264

E.J. Wisniewski and D . Gentner

knowledge comes from a previously mentioned study suggesting that features appear in a combination which are not present in either the adjective or noun concepts but which emerge through an interaction of these concepts, mediated by general knowledge. In addition, Murphy’s finding that irrelevant-adjective noun concepts were more difficult to understand than relevant-adjective noun concepts suggests that one needs to access concepts outside of the constituents (specifically, superordinate concepts) to understand the former. So, in trying to understand cold garbage one will not find an appropriate slot in garbage that cold can fill (because a slot like temperature is not relevant to garbage). Instead, one must determine an appropriate slot via inheritance from one of the superconcepts of garbage. In contrast, understanding cold beer should be easier because such a slot is represented in the concept beer. Murphy’s (1988) finding that a helpful context speeds up the interpretation of noun-noun concepts suggests that context may activate slots in a concept, thus speeding the combination process. Context may suggest a plausible slot for the predicate term, making the combination process easier.

Evaluation of the Model The concept specialization model provides a unifying account of how adjective-noun and noun-noun concepts are interpreted. In this respect, i t is more general than the selective modification model (which applies only to adjective-noun concepts) and the attribute inheritance model (which applies only to the small subset of noun-noun concepts that are conjunctive). It is also the only model that has attempted to account for the important role of context in conceptual combination. On the other hand, while the model’s notion of world knowledge (used in the elaboration process) seems necessary to capture emergent features and to determine which slots to fill, it is a vague principle. Murphy (1988) has noted that the model refers to people’s knowledge in a rather unconstrained manner and that its use of knowledge is not spelled out to any degree. Moreover, the model has not been empirically evaluated as carefully as either the attribute inheritance model or the selective modification model. Further development of the model will need to take these issues into account. Nevertheless, the concept specialization model is extremely plausible. Indeed, we suspect that it is the default modcl for combining nouns. However, in the next section, we will argue that in some cases, it is necessary to go beyond the model’s processing and representational assumptions. In particular, while the modcl assumes structured representations for nouns, the importance of such structure for combining concepts has not bcen demonstrated. We will suggest that in a number of noun-noun combinations this structure (which includes relations between slots) plays a very important role in the combination process. Besides an emphasis on structured representations, we will also sug-

Combinarorial Semantics of Noun Pairs

265

gest that a complete model of noun-noun combination must employ other processes besides slot filling. Summary We have described a number of models of conceptual combination. It is clear that no model is complete. There are two major reasons why this is the case. First, psychological research on conceptual combination is relatively recent compared to work in other areas (memory retrieval, attention, structure and processing of single concepts, etc). There is not yet a large base of empirical studies on how people combine concepts. We need to learn more about this cognitive process that people do naturally and easily. Second, as argued in the last section, any complete model of conceptual combination will have to be complicated and extensive. The approach taken in all of these models is to carve out a piece of the problem and first attempt to understand that well. In doing so, these models have made a number of implicit, simplifying assumptions which make them incomplete at this point in their development. They avoid the problem of “who modifies whom” by assuming that combinations are composed of only two constituents. The models also implicitly assume that it is clear which meanings of the constituents are being combined (thus avoiding the problem of lexical ambiguity). The models also limit the types of combinations that they address. Smith et al.3 selective modification model has been applied to a subset of adjective-noun concepts. Hampton’s attribute inheritance model has been applied to the subset of nounnoun concepts that are conjunctive. Murphy’s concept specialization model is the most general model-specifying how people interpret both noun-noun and adjective-noun concepts. However, the model has not explicitly addressed the important role of conceptual structure (e.g., relations between slots) in combining nouns. In the next section, we will suggest that this structure sometimes is involved in combining noun meanings. We will also describe other processes besides slot filling that operate noun-noun combinations.

How

DO

PEOPLEDEFINENOVELCOMBINATIONS -WHAT IS A PONYCHAIR?

In this section, we address the generality of slot filling in conceptual combination. Our goal is to examine people’s descriptions of novel combinations to see how well thcy fit this view. We are especially interested in determining other strategies that people use to combine concepts as well as the kinds of noun representations that would be needed to accommodate these strategies. As mentioned, slot filling is a major component of both the selective modification model and the concept specialization model. The authors of these models imply that slot filling typically occurs when people combine concepts. The

266

E.J. Wisniewski and D . Gentner

concept specialization model also postulates a second process (elaboration) that follows slot filling. We will focus here on noun-noun concepts and will not examine the generality of the process for adjective-noun concepts. In terms of evaluating the two models, our data are more directly applicable to the concept specialization model (since it explicitly addresses how noun-noun concepts are combined). Assumptions and Plausibility of the Slot Filling Process There are three underlying assumptions involved in slot filling. First, the process is applied to the head concept and to a slot that the head concept contains or can inherit from its superordinate concepts. Second, the resulting combination XY is basically a Y with an additional restriction on one of its slots. Third, the process involves restricting the filler of the slot to the predicate concept (and not other concepts). So, when forming a combination XY, people restrict the filler of a slot in the head concept Y to the predicate concept X. For example, consider a very plausible meaning of book box: “box that contains or holds books.” Assume that the concept box has a number of slots that can be filled by other concepts. When people interpret a phrase like “book box” they search for a slot in the head concept box that can be filled by the predicate concept book. In this case, people interpret book box by filling a slot of box (that corresponds to “contains” or “holds”) with book. This slot is restricted to having book as its filler. A book box is a box except that it contains books and not other things. (Of course, such a representation does not rule out the possibility that a book box could contain other things. We will ignore this subtle distinction.) Intuitively, it seems that slot filling is a very natural strategy for combining concepts. There may be several reasons for why people prefer this strategy. First, it allows one to use the predicate noun as a predicate while preserving the integrity or cohesiveness of its meaning, as well as the meaning of the head noun. That is, slot filling may involve minor adjuslmenls to noun meanings. Gentner (1981, 1982) has suggested that concrete nouns, relative to other parts of speech, have highly coherent, internally constrained meanings and that people prefer to preserve those meanings whenever they can. Simple nouns typically refer to objects in the world and their meanings incorporate a large amount of perceptual information that is determined by those objects. Other parts of speech, especially verbs, are less tightly constrained by the perceptual world. As previously noted, verbs are more likely to change their meanings than nouns (Gentner & France, 1988). Also, compared to nouns, languages vary more in terms of which meaning components they conflate into verbs (Gentner, 1981; 1982; cf. Talmy, 1978). Slot filling amounts to asscrting a relation between the head noun and the predicate noun (e.g., “box that contains books”) and does not disrupt the basic

Combinatorial Semantics of Noun Pairs

267

meanings of the nouns. Of course, slot filling may alter the meaning of either noun to some extent. For example, assume that “pear pie“ means “a pie madeof pears” and that pear fills the made-of slot of pie. In this case, the predicate noun “pear” probably refers to sliced and peeled pears rather than the typical pear. As another example, the head concept soap in tank soap is probably different than the typical soap (e.g., more abrasive, more concentrated). It is likely that the meanings of these nouns have been altered. However, they probably retain enough of their original meanings so that people would agree that they still refer to pears and soap. Later, we will present examples suggesting that this is not always the case. Second, in terms of computation, slot filling may be an easier strategy than others. In general, one only has to check the meaning of the predicate noun rather than to alter its structure. Specifically, slot filling may require that one check whether the predicate noun fits certain constraints on the slot. In the example of pear pie, filling the made-of slot of pie with pear might involve checking whether pear fits a constraint on the made-of slot such as being edible. In contrast, we will suggest that other strategies require one to dismantle and significantly alter the meaning of one or both nouns. That is, some strategies involve major adjustments to noun meanings. Presumably, these adjustments are more computationally complex than those involved in slot filling.

A N EXPERIMENT The study that we will describe was largely exploratory in nature. We were interested in assessing the generality of slot filling as well as discovering other combinatorial strategies and the corresponding representations that they operate upon. One way to examine such strategies is to collect a large number of descriptions of many novel combinations. The obvious problem with this approach is that one needs a way to meaningfully sample from the huge number of possible noun combinations. To introduce some constraints, we varied nouns along three conceptually important dimensions: predicate versus head noun position, artifact versus natural kind, and count noun versus mass noun. Intuitively, we also believed that nouns varying along these dimensions might interact in interesting ways when they were combined. These interactions might result in situations where slot filling was more or less preferred as a combinatorial strategy. For example, intuitively, the “predicate versus head noun position” probably cues whether a noun is an operator or a referent. On the other hand, count nouns and mass nouns may differ in terms of how natural it is to use them as referents and operators. Objects (particularly artifacts) are often composed of mass quantities (e.g., windows made of glass, vases made of clay, etc). One uses count nouns to refer to such objects rather than the mass terms of which they are composed of. Therefore, in a novel combination, one might prefer to use a count noun as an object referent and a mass noun as an

E.J. Wisniewski and D.Gentner

268

operator. A mass/artifact-count combination preserves the preferred roles of the constituents whereas an artifact-countlmass combination violates those roles. As a result, it might be more straightforward to interpret a mass/artifact-count combination by slot filling than an artifact-countlmass combination. In fact, in the predicate position, mass nouns may function like adjectives, picking out a particular slot (i.e.. composition) to fill. On the other hand, subjects might use some other strategy to interpret count-mass terms (e.g., use the predicate noun as the referent and the head noun as an operator). We used three groups of nouns to create the noun-noun phrases. One group consisted of 10 count nouns and a second group consisted of 10 mass nouns. A third group of nouns consisted of 10 count nouns, as in the first group. Half of the nouns in each group were artifacts and half were natural kinds. The three groups of nouns are shown in Figure 4. To form noun phrases, we first paired each noun from group 1 with each noun from group 3 and paired each noun from group 2 with each noun from group 3. This procedure resulted in 200 pairings. It also resulted in a hierarchy of combination types, shown in Figure 5 . For each pairing, we then formed the two noun-noun phrases that were possible (e.g., for the pairing of “robin” and “clock,” the two phrases “robin clock” and “clock robin” were possible). This procedure resulted in 400 noun-noun phrases. Each of 20 subjects defined 20 of these noun-noun phrases.

aroup 1

Group 2 I

[Count N o w

Figure 4. The three groups of nouns used in the experiment.

Subjects read the novel noun-noun phrases and were asked to write down descriptions of their most likely meanings. They were told to pretend that they

Combinatorial Semantics of Noun Pairs

269

I Noun-Noun Combinations I

Icount-count

PlsSsJ

I MMS-COU~]

J

I

copper car

+

clay elephant

+

robin snake

+

pan squurel

box clock

papetbook

+

+

fish ~

sMkc candy 0

3

~

*

pencil sand

+.

car plaspc

Figure 5. A hierarchy of combination types (with examples of each type) used in the experiment.

had just heard a phrase during a conversation and that they should think of a meaning that seemed most natural to them. Subjects were instructed to try to arrive at meanings that were specific and clear and to define every phrase. Generality of Slot Filling

To look at the generality of slot filling. we asked two questions about a description that would provide evidence for the slot filling view. The first question was what is the referent of a subjects’ description: Slot filling predicts that the referent will be a type of the head concept. For example, if a subject described book box as above, “a box that contains or holds book,” then the referent would clearly be a type of box. In most of the subjects’ descriptions, the referent could be determined syntactically. Typically, as in this example, it

270

E.J. Wisniewski and D . Gentner

is the first noun that is mentioned in the subject’s description. However, this is not always the case. For example, consider the description for chocolate pony, “a pony made of chocolate.” Syntactically, a pony is being described. However, conceptually, the referent is not really a pony, but rather chocolate in the shape of a pony (according to our intuitions). The second question is whether all or part of a description can be characterized as Y slot-relation X , where Y is the head concept, X is the predicate concept, and slot-relation is some relation being asserted between Y and X that corresponds to a slot contained in the head concept. The description of book box can be characterized as a Y slot-relation X (“box contains books”). In contrast, a description such as “squirrel with a black stripe down its back” for squirrel skunk, cannot be characterized in this manner, as no relation is being asserted between skunk and squirrel. Rather, it appears that a property of squirrel is being asserted of skunk. To address the first question (i.e., to determine the referents of the subjects’ descriptions), we gave the descriptions to a group of undergraduate judges. Each of 20 judges read half (200) of the definitions. For each description, they determined whether the referent was a type of the predicate concept, a type of the head concept, both, or some other object. Specifically. subjects were asked to answer the following question about each description: What is the object that is being described? That is, what would be the best name for the object that would let someone know what it really is. For a given description, this procedure resulted in 10 judgements about the identity of the referents. The referent of a description was determined by the consensus of the judges. In general, the head noun was the referent, as predicted by slot filling. The judges believed that a majority of the descriptions described types or kinds of the head concept. Interestingly, however, for 151 (38%) of the 400 descriptions, the head concept was not the referent. Two examples of this violation were chair ladder, which was described as “a chair that for necessity is used as a ladder,” and paper elephant, which was described as “paper in the shape of an elephant.” (Notice that in the both cases, the predicate noun functions as the referent and head noun as the operator.) Two examples of descriptions in which the head concept was judged as the referent were “a tiger that preys on horsed ponies, etc” for pony tiger, and “glass for holding pencils’’ for pencil glass. To address the second question (i.e., to determine whether a description could be characterized as a Y slot-relation X ) , two graduate students from the University of Michigan read each definition and decided which of two categories it belonged to. If a description included a relation between the two objects named in the phrase, it was placed in the relation category. Otherwise, the description was categorized as other. We gave the raters several examples of the relation and other categories, using descriptions that were not from the

Combinatorial Semantics of Noun Pairs

271

experiment. For example, they were told that “a factory that is smelly and processes fish” belonged to the relation category, as it asserts a “processes” relation between factory andfish. As another example, they were told that “a dangerous man” belongs to the orher category (no relation is being asserted between a pair of nouns). Also, some descriptions specified a relation between the head and predicate noun even though one of the nouns was not explicitly mentioned. An example of such a description was “a tiger that likes to read a lot” for the phrase book tiger. In this description, a relation between book and tiger is strongly implied even though book has been omitted from the description. Judges were instructed to place these descriptions in the relation category. Note that this procedure provides a liberal test of the generality of slot filling. The raters only determined that a relation was being asserted between the head concept and predicate concept. They did not have to judge which relations,corresponded to slots contained in the head noun. This leads to a generous count, for it includes relations that may not be part of the head noun’s frame. For example, one might argue that in the description for ladder skunk, ‘‘a skunk that climbs ladders, “ the slot being filled in skunk (i.e., climb) originates in ladder. (Certainly, climbing is much more typical of ladders than of skunks.) The procedure also does not distinguish those descriptions based solely on slot filling from those that included other strategies in addition to slot filling. The two raters initially agreed on 87% of their judgments about the descriptions. Differences in scoring were discussed and resolved. The raters judged only 40% of the descriptions as stating a relation. Two examples of descriptions that were categorized as relation were “a pan for frying fish” for fish pan, and “car made out of copper” for copper car. Two examples of descriptions that were judged as other were “a square box” for box clock and “a ladder whose rungs are far apart” for frog ladder. The results of these analyses suggest that nouns are not always combined by slot filling. Indeed, the majority of the descriptions in our corpus were not classified as Y slot-relation X . An examination of those descriptions that did not conform to slot filling suggests two general conclusions. First, there are other important processes besides slot filling that are used to combine nouns. Some of these processes involve major adjustments to meaning (relative to slot filling). Second, some of these processes operate on noun representations that are more complex than those currently proposed in the literature. We will argue that noun representations must include more than slots and fillers. Importantly, they must include relations between slots within a noun (i.e.. internal relations) as well as relations between slots of different nouns (i.e., external relations). In addition, fillers of slots can themselves be complex structures (i.e.. nested s m c tures). Below, we describe some of these conjectured processes and the representations that they operate upon. At this point, we will make no claim about their generality, except to say that the noun-noun descriptions that suggest these processes were not rare occurrences in our data.

E.J. Wisniewski and D . Gentner

272

OTHERSTRATEGIES FOR COMBINING NOUNMEANINGS As we have suggested. slot filling is an example of a process that preserves the basic meanings of the predicate and head nouns. The combination that results from slot filling is a type of the head noun with some relation (designated by the slot) to the predicate noun. In general, this process does not significantly alter the meaning of either noun. We will now suggest that many of the descriptions that do not conform to slot filling reflect processes in which only part of the meaning of the predicate noun is involved in the resulting combination. We will also suggest that under some conditions (for example, often when mass and count nouns combine) only part of the meaning of the head noun is involved in the resulting combination. A large number of the subjects’ descriptions (approximately 30%) had the form property Y or Y with property, where Y is the head concept (e.g., “a large frog” was the definition for elephant frog). In these descriptions, a property is being asserted of the head concept, rather than a relation between the predicate concept and the head concept (as in slot filling). That is, the predicate noun is not participating as a whole in the resulting combination. It is not playing the role of a slot filler. What role then does it play in such combinations? We suggest that the predicate noun plays at least two other roles besides being a slot filler. First, in a process called property mapping, thefiller of a slot in the predicate concept is used as a filler in the corresponding slot of head concept. Second, in a process called structure mapping: the complex structure of the predicate noun guides the creation of new structure or the transformation of existing structure in the head noun. We illustrate these processes using some examples taken from our data. Besides property mapping and structure mapping, we will also discuss complex effccts that occur when mass nouns in the predicate position combine with count nouns in the head noun position.

Property Mapping To illustrate this process, consider the description, “a red snake,” that was given by a subject for robin snake.This description docs not fit the slot filling view since a slot in snake is not being filled with the predicate concept robin. (A description for robin snake that would involve slot filling is “snake that eats robins”). Instead, it appears that the filler red of the slot color in robin becomes the filler for the color slot in snake. as illustrated in Figure 6. (In this figure and those that follow, we have added unspecified connections between slots to emphasize the importance of relations.) Here, the color slot of robin is aligned (or put into correspondence) with the color slot of snake. The filler of color (red) is then mapped across and becomes the filler of the color slot in snake. As in the standard view of slot filling, a slot in the head concept is affected and the

Combinutorial Semantics of Noun Pairs

273

resulting combination X Y is basically a Y, with an additional restriction on the slot. However, property mapping restricts the slot to thefiller of a slot in the predicate concept rather than the predicate concept. Notice that this process leaves the meaning of the head noun basically intact. However, the predicate concept contributes only a small part of its meaning to the combination. As described, one could incorporate property mapping into the concept specialization model without altering the the model’s basic assumptions. The slot filling mechanism would consider both slot fillers in the predicate concept and the predicate concept itself as potential fillers for head noun slots. Moreover, the mechanism could still operate successfully on a list of slots and fillers

Figure 6. Using property mapping to combine robin nnd s ~ k Cto form robin S

M ~ .

(the representation for nouns that is emphasized in the model) as long as an alignment of slots could be made to guide the property mapping. However, the next process that we consider is quite different from this augmented view of slot filling and requires more complex representations.

Structure Mapping We will illustrate this process using three examples taken from our data. The first example is pony chair, which was defined as “a small chair.” This description does not fit the slot filling view since a slot in chair is not being filled with the predicate concept pony. One might be tempted to classify this description as an an example of property mapping. Here, the filler small of the slot size in pony fills the size slot in chair, yielding the interpretation of pony chair as a small-sized chair. However, note that the typical pony is actually larger than the typical chair. If one literally interprets pony chair as a chair similar in size to a pony, then paradoxically, a pony chair will be larger than most chairs! The resolution rests on noting that “small” is a relative adjective and that ponies are small relative to other horses. We suggest that pony chair literally means a chair that is small relative to other chairs. How then were pony and chair combined to yield this meaning? To interpret pony chair in this manner, we suggest that the representations of pony and chair must be more complex

EJ. Wisniewski and D . Genmer

274

than lists of slots and fillers. In particular, the representation of pony includes a relation (i.e., less-than) between its size slot and the size slot of horse which represents the fact that ponies are small relative to other horses. To combine pony and chair, a similar relation is created between the size slot of pony chair and the size slot of chair. The relation represents the fact that pony chairs are

.

Pony

Pony Chair

slot:

Piller(r):

Slot:

Cobr: Tam: sound:

brown, vhi@, bkh

Cobr: Texture:

fw

neighs fm,mnch. 54

-

(tltgnment)

Slot:

Piller(r ):

Cobr Tsxm

brown, wtu , b L I

Locatbn:

4Size.

Piller(r): b m m , bbck, m... slmoth vwd, mtal, phstic. Mng m m ,klchen

e

:t1d!g3

(struciure

Mada-of

Filler(r): bmm, black, BX... rmoolh

M o d , mcral, p h t k

Samd

Location: S$:

I

ltving room, WLn 3

.

1

(alignment) ?

Figure 7.Using structure mapping to combine pony and chair to form pony chair.

small relative to other chairs. This example illustrates a structure mapping which involves aligning (or putting into correspondence) the size slot of pony with the size slot of pony chair and the size slot of horse with the size slot of chair (see Figure 7). Then, a less-than relation is mapped across between the size slots of pony chair and chair, leading to the notion of a small chair.’ There are several important differences between this process and those of property mapping and slot filling. First, neithcr the predicate concept or a filler of one of its slots functions as a slot filler. Rather, a structural relation between one of its slots and the slot of another (closely associated) item guides the creation of a new, similar structural relation in the head concept. Second, the process operates on and creates representations that are more complex than a list of slots and fillers. In this example, the representations include relations between slots of different concepts (is., external relations). A second example is snake glass, which was described as a “tali, very thin drinking glass.” Once again, this description does not fit the slot filling view. How then were snake and glass combined? First, note that a snake glass resembles the shape of a snake in some way. We suggest that in general, the

Combinatorial Semantics of Noun Pairs

275

shape slot specifies a complex structure that relates various aspects of shape to each other (cf. Palmer, 1975; Marr & Nishihara, 1978; Biederman, 1985). In the concept snake, this structure might (among other things) indicate that the typical snake is much longer than it is wider. In the concept glass, this structure might (among other things) indicate that the typical glass is somewhat taller than it is wider. In this example, we believe that the shape of glass is modified in a way that is analogous to the shape of snake. Just as a snake is much longer than it is wider, a snake glass is much taller than it is wider. Figure 8 outlines how structure mapping might operate to produce snake glass. Notice the various parts of structure that have been aligned (e.g. the length slot of snake has been aligned with the height slot of glass.) The process results in the height of snake glass being increased relative to its widrh. Unlike the example of pony chair, existing structure in glass is being transformed to create snake glass

Shap

’3 x-1

e : width length z-1 height 0 0

Shape: length y-2 width 2-2 height x-2 (alignment) I

F i g w e 8 . Using structure mapping to combine S

I M

I2

0 0

*

and ~ gloss to form S

M

gloss. ~

rather than new structure being added. On the other hand, interpreting snake glass in the manner described might result in the new knowledge that snake glasses are longer than typical glasses. Therefore, one would need to augment snake glass with a longer-than relation between the length slot of snake glass and the length slot of glass. (i.e., an external relation). A final example is ladder rake which was defined as “utensil which is elongated so as to use to reach high places.” As in the examples before, this description is not a case of slot filling. We present one possible interpretation of how ladder and rake are combined. First, assume that the function of ladder rake actually shares aspects of both the function of ladder and of rake. Like ladders, ladder rakes are used to reach high places. Although not specified, one might also surmise that they are used to collect or gather things from high places, thus preserving aspects of the function of rakes. Figure 9 sketches how ladder and rake might be combined. Notice that the fillers of the function slots point to complex structures which we have represented using notions derived from case grammar. To combine ladder and rake, the function of ladder is

E.J. Wisniewski and D. Gentner

276

aligned with that of rake. The function of ladder rake is created by modifying the function of rake. Specifically, the fillers of the source and destination slots of thefunction of ladder are mapped across and become the fillers of the source and destination slots of thefunction of rake (see Figure 9). As a result, a ladder rake has a function that is similar to that of a rake except that one uses a ladder rake to rake in a vertical direction rather than a horizontal direction. Notice also

1

Raking

Climbing

LadderPake Slot

Piller(r):

Raking

1

Slot:

I

I

2 Filler(#):

I

(structure transformatton)

Figure 9. Using structure mapping to combine ladder and rake to form ladder rake.

that the filler of the shape slot of ladder rake is different from that of rake (ladder rakes are longer than rakes because they are used to reach high places). The change in the shape slot is an example of interacting properties (see the first section). In particular, thefunction and shape slots must include relations between them that capture these interactions. A variation on the structure-mapping process is that a slot that is filled may actually be one that is inherited by the head concept from the predicate concept. Two examples are clock tiger which was defined as “a tiger that can tell time” and pony frog which was defined as “a frog that is trained to ride ponies.” In both of these (somewhat strange) examples, the relation being asserted is strongly associated with the predicate concept rather than the head concept. These examples illustrate structure creation as a novel slot is being added to the head concept.

Combinatorial Semantics of Noun Pairs

277

Complex Eflects of Mass Terms Mass terms often refer to substances that things are made of. For example, the mass noun “glass” refers to a substance that many objects are made of-windows, vases, bottles, plates, and so on. Therefore, people may be biased to interpret a combination of the form mass termcount term as naming an object that is made of the mass term. In fact, according to our judgment, 59% (59 out of 100) of the descriptions of mass-count terms described an object that was made of that mass term. Some of these descriptions explicitly stated this relation, as in ‘‘a clock made out of copper” for copper clock. Other descriptions were less explicit, as in “statue” for stone snake. However, it also appears that interpreting a combination in this manner has important effects beyond just indicating the composition of an object. These effects are related to semantic differences in the head nouns that we used in this study. Recall that a head noun was either an artifact or an animal natural kind (see Figure 5). In general, an artifact can be made of a variety of substances (often named by mass nouns) whereas a given animal is generally believed to be composed of one kind of substance. One can often assert that an artifact is made of a variety of different substances arid still preserve the identity of the artifact-what appears important though is that the function of the artifact be preserved (Gelman, 1988; Keil, 1986; 1987). A plate for example, can be made of wood, metal, plastic, glass, and so on. and still be a plate (being made of such substances does not affect its function). On the other hand, a dog can’t be made of wood or plastic and still be a real dog. This difference between artifacts and animals suggest different conditions under which head nouns will lose their referential priveleges. When a massanimal term describes an object made of the mass term, the referent will be some object other than the animal. For example, the description of chocolate snake (“chocolate in the shape of a snake”) names an object that is made of the mass term. The referent is also the mass term (“chocolate”) rather than the animal (“squirrel”). In this example, the head noun has given up its referential priveleges to the predicate noun. Of course, squirrel still confers its shape on the referent. (In fact, in many contexts, shape may be an important property for determining reference. Even though the head noun loses its referential priveleges, it may in a sense, “effect a compromise,” by contributing an important property for determining reference.) In contrast, one can usually intepret a mass-artifact term as an “artifact made of mass term” if doing so would preserve the artifact’s function. In these cases, the artifact retains its referential privileges as the head noun. However, if such a description would fail to preserve the artifact’s function, then the referent will be some object other than the artifact. Two examples from our data illustrate these different cases. The referent of clay ladder (“a ladder made of clay”) was judged to be ladder. This description also appears to preserve the

278

E.J. Wisniewski and D. Genrner

function of ladder-i.e., one could use a clay ladder for climbing. On the other hand, the referent of candy ladder (“a long strip of candy”) was judged to be candy. It appears that function of ladder would not have been preserved if the term was literally interpreted as “a ladder made of candy.” As evidence for these hypotheses, recall that there were 59 descriptions of mass-count terms that referred to objects made of those mass terms: 27 of these descriptions involved mass-animal terms and 32 descriptions involved massartifact terms. A group of undergraduate judges had determined the referents of these descriptions. In 96% (26 of 27) of the mass-animal terms, the animal term failed to retain its referential privilege. In 66% (21 of 32) of the mass-artifact terms, the artifact retained its referential privilege. For the mass-artifact terms, we have not systematically evaluated whether retaining versus giving up referential privilege corresponds to preserving versus violating an object’s function. However, in those combinations in which the artifact term gave up its referential privilege, it did appear that interpreting them as “artifact made of mass term” would have violated their functions.

SUMMARY These preliminary results suggest that there may be a variety of mechanisms that operate in conceptual combination. Although we did find evidence for slot filling, it was by no means the only strategy that people used. Most notably, another very common strategy was aligning the structures of the two nouns and mapping part of the predicate noun’s structure onto the structure in the head noun. Either a filler from the predicate noun was mapped to fill a slot in the head noun (property mapping) or a relation between slots in the predicate noun (or between a predicate slot and a slot in a related concept) was mapped to the head noun (structure mapping). In either case, the predicate noun is (in a sense) dismantled: instead of filling a slot in the head noun it yields part of its meaning in forming a combination. Finally, we also showed that people combine mass and count terms in complex ways. They often interpret a mass-count phrase as naming an object whose composition is indicated by the mass term but whose referent is not always a type of the head concept. In particular, if the head noun names a natural kind, it will lose its referential priveleges although it may contribute an important referential property (i.e.. its shape) to the combination. If the head noun names an artifact, it generally retains its referential priveleges unless the composition of the combination violates the function of the artifact. One possible objection to the present study is that we collected just a single definition for each combination and that our results reflect idiosyncratic responding in our subjects. For example, if we asked a large number of people what a pony robin was, would the majority actually respond “a robin with a tail” (as the subject in our experimcnt did)? Probably not. However, while a

Combinatorial Semantics of Noun Pairs

279

description for a given combination might be idiosyncratic, the different strategies discussed above were not rare occurrences. For example, property mapping and structure mapping were common strategies across subjects. A second possible objection is that our results are based on unusual or even bizarre noun-noun compounds that one rarely encounters in typical natural language contexts. For example, how often does one encounter a phrase like “chair pony,” consisting of a natural kind and an artifact? We have two answers to this objection. First, individual constituents of the compounds are not unusual. Therefore, studying how meanings of common words interact (even if their occurring in the same context is unlikely) could shed light on the nature of the meanings themselves. We believe that conceptual combination, like analogy and metaphor, forces words to reveal aspects of their meanings that may not become apparent in more usual contexts. In fact, the interactions between word meanings that we found suggest that the representations of individual constituents need to be structured and complex. Second, a number of the combination types that we used in this experiment (see Figure 5 ) do appear in our language, as indicated by lexicalized entries in the dictionary. Some examples of naturalkind pairs include: tiger salamander, sparrow hawk, moose bird, dog salmon, and gopher snake. Natural-kind artifact pairs include: whale boat, monkey jacket, book scorpion, carpet beetle, and oyster rake. Artifact natural-kind pairs include: trumpet flower, guitar fish, chimney swallow, razor clam, and pill bug. Mass-count terms include: paper knife, clay pigeon, stone fly, coal fish, and plastic bomb. The findings raise a number of interesting issues. For example, we have suggested that slot filling is the default strategy for combining noun meanings. An obvious question is when do people adopt other strategies like property mapping and structure mapping? At this point, we can only speculate on the answer to this question. We can think of at least two conditions which might promote the use of property mapping and structure mapping. First, the more similar two objects are, the easier it should be to map properties or structures from one object to the other. In this context, similarity specifies the degree to which one frame can be aligned with another. In a combination such as zebra horse, high similarity may bias people to map properties from zebra to horse. Informally, when asked to describe a zebra horse, people typically responded, “a horse with stripes” (which is an example of property mapping). Although high similarity may facilitate the mapping of properties and structure, people may still interpret some combinations by slot filling. For example, two very plausible descriptions of dolphin shark are “shark with a dolphin-like nose” (property mapping) and “shark that eats dolphins” (slot filling). It may be that a highly salient, plausible relation between two objects can override mapping processes. A second condition which might encourage mapping processes is the difficulty of finding a plausible relation between objects. As a result, people are

280

E.J. Wisniewski and D . Gentner

not able to meaningfully combine nouns by slot filling (their default strategy) and must consider other strategies. For various reasons, a plausible relation between two objects may not exist. Factors like high dissimilarity of two objects and their low co-occurrence in the environment may rule out such a relation. This condition may apply to some of our examples of structure mapping (such as snake glass and pony chair). There are many other interesting questions that future research might address. For example, at some level, can we reliably predict the meanings that subjects will construct for novel combinations? Can we reliably predict which combinations are more difficult to understand than others? Are some meanings of novel combinations “better than others” (as judged by subjects) and why? At this time, we have little to say about these issues. The major goal of the current work has been to determine the strategies that people use and the nature of noun representations required for those strategies. Once we have a better appreciation of these issues, we can begin to address the more difficult questions.

Acknowledgments This research was supported in part by the National Institute of Child Health and Human Development under Fellowship Award 1 F32 HD07279-01 to the first author. It was also supported by the National Sciencc Foundation under Grant BNS 87-20301 and by the Office of Educational Research and Improvement under Cooperative Agreement No. G0087-C1001-90 with the Reading and Education Center. We have benefited from discussions of this work with Arthur Markman, Douglas Medin, Gregory Murphy and Edward Smith.

Notes We use italics to indicate concepts or parts of concepts, and reserve quotes for the words that denote those concepts. We use the terms “head concept” and “predicate” concept to capture an important point. First, as discussed below, a noun-noun combination of the form XY typically refers to a Y and not an X - e.g., a dog apartment is an apartment and not a dog. Therefore, Y determines more of the meaning of XY. In this sense, Y is the head or main concept. In fact, one can view conceptual combination as an abstract function X(Y) in which X acts as an operator on Y. In this sense, it is a predicate. There are a number of subtleties about framcs and frame instanccs that we will ignore in this paper. For example, the semantics of fillers are different in frames and frame instances. In a frame instance, a slot and its fillcr specify a fact that is actually true of a particular object. So, (robin-I7 color red-13) roughly means that the color of a particular robin is a particular red. (Here, numbers are appended to robin and red to distinguish them from other instances

Combinatorial Semantics of Noun Pairs

28 1

of robins and their instances of color.) In contrast, a slot and filler of a frame specify a default fact about an object. So, (robin color red) roughly means that the color of robins is typically red. Smith and Gray (1990) provide an alternative interpretation for some of these feature interactions. Specifically, they suggest that people already may be familiar with some combinations and that feature interactions do not result from a combination process but rather from experience with examples of the familiar combination. So, people may already be familiar with the combination wooden spoon, and may have acquired their belief that wooden spoons are large from experience with examples of wooden spoons. In this view, knowledge that wooden spoons are large is not derived from combining wooden and spoon but rather from examining examples of wooden spoons after the combination process. While acknowledging that their representations are much like frames, Smith et al. actually use the terms attribute and value for slot and filler and the term prototype for frame. Structure mapping is the mapping of relational structure from the predicate noun’s meaning to the head noun’s meaning (as in Gentnet’s (1983. 1989) discussion of analogy) We have assumed that the meaning of pony.was accessed in order to interpret pony chair as a small chair. It is plausible that one may have accessed the meaning of pony keg instead. Nevertheless, we would claim that similar processing and representational assumptions would still hold. The basic difference would be that structure mapping would operate on a relation between slots of pony keg and keg instead of between slots of pony and horse.

‘

References Biederman, I. (1985). Human image understanding: Recent research and a theory. Computer Vision, Graphics, and Image Processing, 32, 29-73. Brachman, R. J. (1977). What’s in a concept? Structural foundations for semantic networks, International Journal of Man-Machine Studies, 9, 127-152. Brachman, R. J. (1978). A structural paradigm for representing knowledge. BBN Report No. 3605, Cambridge, MA Brachman, R. J. (1979). On the epistemological status of semantic networks. In N. V. Findler (Ed.), Associative networks: Representation and use of knowledge by computers. New York: Academic Press. Brautigan, R. (1967). Troutfishing in America. New York: Dell Publishers. Clark, E. V., Gelman, S. A., & Lane, N. M. (1985). Compound nouns and category structure in young children. Child Development, 56, 89-94. Clark, H. H. (1983). Making sense of nonce sense. In G. B. Flores d’Arcais & R. Jarvella (Eds.), The process of understanding language. New York: Wiley.

E.J. Wisniewski and D . Gentner

282

Collins, A. (1978). Human plausible reasoning. BBN Report # 3810, Cambridge, MA. Collins, A. & Michalski, R. S. (1989). The logic of plausible reasoning: A core theory. Cognitive Science, 13, 1-50. Cohen. B., & Murphy, G. L. (1984). Models of concepts. Cognitive Science, 8 , 27-58.

Cotmll, G. W.(1988). A model of lexical access of ambiguous words. In S. L. Small, G. W. Cottrell, & M. K. Tanenhaus (Eds.), Lexical ambiguity resolution: Perspectives from psycholinguistics, neuropsychology & artificial intelligence. san Mateo, CA: Morgan Kaufmann Publishers. Downing, P. (1977). On the creation and use of English compound nouns. Language, 53,810-842. Finin. T. (1980). The semantic interpretation of nominal compounds. Proceedings of the First Annual National Conference on Artificial Intelligence, Stanford, CA. Gelman, S. (1988). Development of induction within natural kind and artifact categories. Cognitive Psychology, 20, 65-95. Gentner, D. (1975). Evidence for the psychological reality of components: The verbs of possession. In D. A. and D. E. Rumelhart (Eds.), Explorations in cognition. San Francisco: Freeman. Gentner, D. (1981). Some interesting differences between verbs and nouns. Cognition and Brain Theory 4 , 161-178. Gentner, D. (1982). Why nouns are learned before verbs: Linguistic relativity versus natural partitioning. In S. Kuczaj (Ed.), Language development: Language, cognition. and culture. Cambridge: Cambridge University Press. Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7, 155-170. Gentner, D. (1989). The mechanisms of analogical learning. In S. Vosniadou and A. Ortony (Eds.), Similarity and analogical reasoning. New York: Cambridge University Press. Gentner, D., & France, I. M. (1988). The verb mutability effect: Studies of the combinatorial semantics of nouns and verbs. In S. L. Small, G. W. Cottrell, & M. K. Tanenhaus (Eds.), Lexical ambiguity resolution: Perspectives from psycholinguistics, neuropsychology & artificial intelligence. San Mateo. CA: Morgan Kaufmann Publishers. Gentner, D., & Clement, C. A. (1988). Evidence for relational selectivity in the interpretation of analogy and metaphor. In G. H. Bower (Ed.), The psychology of learning and motivation. New York: Academic Press. Goldstone, R. L., Medin, D. L., & Gentner, D. (1990). Relations, attributes, and the non-independence of features in similarity judgments, Cognitive Psychology, in press. Halff, H. M., Onony, A., & Anderson, R. C. (1976). A context-sensitive repre-

Combinatorial Semantics of Noun Pairs

283

sentation of word meanings. Memory & Cognition, 4, 378-383. Hampton, J. A. (1987). Inheritance of attributes in natural concept conjunctions. Memory & Cognition, 15, 55-71. Hampton, J.A. (1988). Overextension of conjunctive concepts: Evidence for a unitary model of concept typicality and class inclusion. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 12-32. Hirst, G. (1983). Semantic interpretation against ambiguity. Technical Report CS-83-25, Computer Science Department, Brown University. Kay, P., & Zimmer, K. (1976). On the semantics of compounds and genitives in English. Sixth California Linguistics Association Proceedings. San Diego: Campile Press. Keil, F, (1986). The acquition of natural kind and artifact terms. In W. Demopoulos & A. Mamas (Eds.), Language learning and concept acquisition. Norwood, NJ: Ablex. Keil F. (1987). Conceptual development and category structure. In U. Neisser (Ed.), Concepts and conceptual development: Ecological and intellectual factors in categorization. Cambridge: Cambridge University Press. Lees, R. B. (1960). The grammar of English nominalizations. Bloomington, IN: Indiana University. Levi, J. N. (1978). The syntax and semantics of complex nominals. New York: Academic Press. Malt, B. C., & Smith, E. E. (1984). Correlated properties in natural categories. Journal of Verbal Learning and Verbal Behavior, 23, 250-269. Marcus, M. P. (1980). A theory of syntactic recognition for natural language. Cambridge, MA: MIT press. Marr, D, & Nishihara, H. K. (1978). Representation and recognition of the spatial organization of three-dimensional shapes. Proceedings of the Royal Society of London, B 200,269-294. Medin, D. L., Altom, M. W., Edelson, S. M., & Freko, D. (1982). Correlated symptoms and simulated medical classification. Journal of Experimental Psychology: Learning, Memory, and Cognition, 8.37-50. Medin, D. L., & Shoben, E. J. (1988). Context and structure in conceptual combination. Cognitive Psychology, 20, 158-190. Medin, D. L., & Wattenmaker, W. D. (1987). Category cohesiveness, theories, and cognitive archeology. In U. Neisser (Ed.), Concepts and conceptual

development: The ecological and intellectual factors in categorization. Cambridge: Cambridge University Press. Minsky, M. (1975). A framework for representing knowledge. In P. H. Winston, (Ed.), The psychology of computer vision. New York: McGraw-Hill. Murphy, G. L. (1988). Comprehending complex concepts. Cognitive Science, 12, 529-562. Murphy, G. L. (1990). Noun phrase interpretation and conceptual combination, Journal of Memory and Language, in press.

284

E.J. Wisniewski and D . Gentner

Murphy, G. L., & Medin, D. L. (1985). The role of theories in conceptual coherence. Psychological Review, 92, 289-3 16. Murphy, G. L., & Wisniewski, E. J. (1989a). Categorizing objects in isolation and in scenes: What a superordinate is good for. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15,572-586. Murphy, G.L., & Wisniewski, E. J. (1989b). Feature correlations in conceptual representations. In G. Tiberghien (Ed.), Advances in cognitive science: Theory and applications, (Vol. 2). 23-45, Chichester: Ellis Horwood. Norman, D. A.. & Rumelhart, D. E. (1975). Memory and knowledge. In D. A. Norman and D. E. Rumelhart, (Eds.). Explorations in cognition. San Francisco: Freeman. Osherson, D. N., & Smith, E. E. (1981). On the adequacy of prototype theory as a theory of concepts. Cognition, 9,35-38. Osherson, D. N., & Smith, E. E. (1982). Gradedness and conceptual combination. Cognition, 12, 299-318. Palmer, S. E. (1975). Visual perception and world knowledge: Notes on a model of sensory-cognitive interaction. In D. A. Norman and D. E. Rumelhart, (Eds.), Explorations in cognition. San Francisco: Freeman. Palmer, S. E. (1978). Fundamental aspects of cognitive representation. In E. Rosch & B. B. Lloyd, (Eds.), Cognition and categorization. Hillsdale, NJ: Lawrence Erlbaum Assoc. Rips, L. J., & Turnbull, W. (1980). How big is big? Relative and absolute properties in memory, Cognition, 8, 145-174. Rosch, E., & Mervis, C. G.(1975). Family resemblances: Studies in the internal structure of categories. Cognitive Psychology, 7, 573-605. Smith, E. E., & Gray, K.C. (1990). Mechanisms of conceptual combination, Unpublished manuscript. Smith, E. E. & Osherson. D. N. (1984). Conceptual combination with prototype concepts. Cognitive Science, 8, 337-361. Smith, E. E., Osherson, D. N., Rips, L. J., & Keane, M. (1988). Combining prototypes: A modification model. Cognilive Science, 12,485-528. Smith, E. E., Shoben, E. J. & Rips, L. J. (1974). Structure and process in semantic memory: A featural model for semantic decisions. Psychological Review, 81, 214-241. Talmy, L. (1976). Semantic casative typcs. In M. Shibatani (Ed.), Syntax and semantics, (Vol. 6). New York: Academic Press. Tversky, A. (1977). Features of similarity. Psychological Review, 84, 327-352. Tversky, B., & Hemenway, K. (1984). Objects, parts, and categories. Journal of Experimental Psychology: General, 113, 169-193. Wisniewski, E. J. (1990). Functional biases in understanding complex noun phrases, Unpublished manuscript. Zadeh, L. A. (1965). Fuzzy sets. Infomation Control, 8, 338-353.

Understanding Word and Sentence G.B. Simpson (Editor) 0 Elsevier Sciencc Publishers B.V. (Noh-Holland). 1991

Chapter 11 Making Sentences Make Sense, or Words to that Effect

Gregg C. Oden University of Iowa Iowa City, Iowa U.S.A. Jay G.Rueckl Harvard University Cambridge, Massachusetts U.S.A.

Thomas Sanocki University of South Florida Tampa, Florida U.S.A.

Meaningful glances exchanged between friends are meaningful for precisely the Same reasons that sentences are meaningful when they are: because the participants have a thorough understanding of what each other knows and how each other thinks. A communicative act is performed with the belief that you know what your partner will think you intended to communicate by performing that particular act instead of some other. You know that your partner knows this about you and will interpret what you say in light of this belief.’ And so on and so on, relentlessly Thus, an effective act of communication requires the full intellectual power of two capable minds working in sync. This ‘two-mind’s worth’ requirement is often forgotten when only the speaker or only the listener is considered, or when the sentence is taken to be an entity with independent existence. A listener observed understanding a sentence may appear to do it all by himself but this is never so. The speaker of sentences needs to have selected and produced a string of words in such a way as to have the required effect: namely, to lead the listener to think that he means by that sentence the specific thing he hopes to mean. This word sequence will usually comprise considerably less than a complete, literal statement of the intended meaning (if indeed such a thing is even possible; see Gibbs, 1984). The listener then can count on the utterance containing just what he needs to recover the intended meaning2 but only if he does his part in making all of the invited inferences. As a result, both

G.C. Oden, J.G. Rueckl and T. Sanocki

participants make sentences make sense: the speaker by making them be sensibly interpretable by the listener, the listener by doing whatever is necessary for making sense out of what the speaker said. In our research, while focussing on the job of the listener (or reader) we have tried to stay mindful of the work the speaker/writer has done. Thus, we see the listener being presumptuous because the speaker expects him to be. Seen in this light, it is not surprising that words taken from comprehensible conversations are as likely as not to be incomprehensible by themselves (Pollack & Pickett, 1964). After all, the speaker knows that the listener will be able to make up for shoddy sensory information by using the information provided by higher linguistic structure (and vice versa). As one of us has put it (Oden, 1984b), “the information needed to recognize words is no less than that needed to comprehend larger phrases, yet that information appears to be ‘spread’ thinly but evenly over the entire phrase or utterance.” The challenge, then, for an account of lexical contributions to sentence comprehension is to specify how such thinly spread information is laken up and accumulated in order to accomplish word recognition and sentence understanding in one fell swoop, as it were. This is the aim of the work reported here. The theoretical framework that has guided and informed the research we will decribe was developed to be a solution to a fundamental problem of cognitive psychology. The problem is to make modcls perform interesting cognition without being overly sensitive to perturbations in the input. To achieve this goal depends on two basic design principles: robustness requires compensatory integration. complex computation requires structural diversity (and hence, non-linearity). 9

Compensatory Integration Noise would quickly bring most artificial language processing systems to their knees (if they had knees) under the sorts of conditions in which people communicate successfully evcryday. One way to overcome noise is to boost signal strength; for example, the speaker could be requircd to speak very, very clearly, but this is not our normal approach because speakers know that listeners don’t need it. Instead, listeners must rely on the other sort of recourse: to ‘average’ over an overwhelming amount of data and let the noise be washed out through cancellation of errors. That is, given that listeners can count on the speaker having provided a consistent (though noisy) signal, it is justifiable to integrate multiple features and compute the best fitting overall interpretation. However, in order for this to work, the featural intcgration must be compensatory; that is, it must be the case that a small change in one featural value can be offset by an appropriately opposite changc in another. Otherwise, it wouldn’t be

287

Making Sentences M a k e Sense

possible for positive and negative errors to cancel. Compensatory integration, in turn, all but requires that feature values be continuous. Structural Diversity The classic form of continuous integration is addition. It is also the most completely compensatory. Unfortunately, it is too undifferentiated to support real cognition. The problem is that addition is completely uniform: Every bit of information is combined with every other bit in exactly the same way.3 As a result, the integration is structurally ‘flat’; that is, for example, there is no relevance to an organization of featural information corresponding to ((a + b) + c + (d + e + f) + g) because it is just the same as (a + b + c + d + e + f + g) or any other configuration. With linear systems, there is no reflection of featural necessity or sufficiency, of coocurrence or exclusion, or of any other such structural relationship. What is needed is something with at least the structural richness of propositional logic with its ands, ors, nots, ifs, and the like.

FUZZY PROP Our solution to this problem is the fuzzy propositional approach (Oden & Massaro, 1978 and many sequelae). In this approach (which we will call FuzzyProp, for short) knowledge -whether of how words are spelled, what letters look like, or how sentences can be constructed- is taken to be propositional in character. By this is meant that complex thoughts are made up of simpler ones combined by various conceptual operators. However, unlike traditional propositional models, FuzzyProp allows (a) that the fundamental or primitive concepts/features are fuzzy predicates that may hold more or less in a given situation and (b) that the connective operators are also fuzzy in that they preserve the fuzziness introduced by the primitives. As a result, each propositional expression, whether basic or composite, will be true of any given input utterance to a degree depending on the specifics of the component terms and of the logical ‘recipe’ given by the connectives. The evaluation of fuzzy propositional expressions with respect to input constitutes a form of structurally rich continuous feature integration, thereby simultaneously satisfying both design criteria listed above. In the course of trying to comprehend a sentence, FuzzyProp presumes that the propositions corresponding to all possible respective patterns are evaluated at the letter and word levels. At higher linguistic levels, the propositions representing all alternative interpretations must be constructed and evaluated on the fly. In each case, the degree to which each candidate proposition matches the relevant aspects of the input is determined. Goodness of match is computed by means of the structurally appropriate function of the degrees to which the components of

G.C.Oden, J.G. Rueckl and T. Sanocki the proposition are present at the other levels feeding into that level. The goodness of match values are taken to directly determine identification and interpretation; that is, the degree to which a string of letters is identified as one word rather than some other or the degree to which a sentence is interpreted one way rather than another is presumed to be a direct function of the respective relative degrees of match, all things considered. In the end, this amounts to a process of finding the interpretation that provides the best fit semantically, syntactically, phonologically etc. with the sensory input. For present purposes, suffice it to say that all of this has been worked out in exquisite quantitative detail; for overviews, see Oden (1984b) and Massaro (1987). Early work established that FuzzyProp is able to provide quantitative accounts of phenomena at each of several linguistic levels. It also demonstrated that it can serve as a useful framework within which to address many important theoretical issues such as whether multiple meanings are necessarily entertained in the course of resolving syntactic ambiguities, whether word envelope contributes to word recognition independently of letter identities, and whether or in what sense categorical perception is really categorical. Building upon this base, we have in recent years investigated a number of interesting topics to be found all up and down the linguistic hierarchy. In keeping with the theme of the present volume, we will here discuss how listeners and readers make words out of sensory (featural) input and then make sentences out of these words.

MAKINGWORDS Handwriting Considering the vast quantities of research that have been performed on the questions of speech perception and on the reading of printed text, the comprehension of handwritten language has been strangely neglected. Barely a handful (so to speak) of studies have been carried out addressing this ubiquitous human ability. Not only is it a common and important activity, but it is also one that ideally exemplifies structurally-governed continuous variation. Perhaps this is why it has been so neglected: it is impossible to ignore the inherent continuous variation from token to token. With speech, one can pretend that ‘special-to-speech’ (magical) perceptual devices effectively remove the variability; with text, one can pretend that the unvarying physical stimulus is perceived unvaryingly. With handwriting, the variability is there to be seen; there is no denying it. Thus, it would not be surprising for someone who operated within the context of traditionally discrete cognitive models to consider handwriting to be unusual and unnatural and, as a result, uninteresting. But, of course, this is far from the case. With no substantial body of research to build on, we began our investiga-

Making Sentences M a k e Sense

289

tion of this topic (Oden & Rueckl, in preparation) with a systematic exploration of feature structure in handwriting. Partly this involved taking a new look at naturally produced handwriting samples such as student exam answers and at transcriptions of hundreds of words we obtained from subjects specifically for this purpose. But mostly it was a matter of generating many tokens of the 262 letter pair sequences in our own hands in order to get some overall sense as well as specific information about features and about possible featural dependencies. As is usually the case, this observational phase provided a rich trove of knowledge about the character of handwritten words; the experiments are to a fair degree simply rigorous demonstrations of what we learned. Nevertheless, the experiments do additionally address a specific important question not answerable on the basis of observation alone: Are letter features perceptually dependent on the word context in which they occur? To address this issue, we began by collecting numerous tokens of words specifically chosen because they differed from matched mates in exactly two letter positions with the letters in those positions differing along some physical dimension that seemed perceptually unitary.’ From our collection, we selected instances of various words, choosing tokens that were legible but not excessively neatly written. Each token was digitized as a continuous two-dimensional waveform. This waveform was then successively edited to produce separate continuous manipulations of the two relevant features through seven steps. For example, for the word “day”, one manipulation involved raising the drooping line that constitutes the transition between the “a” and the following letter, gradually turning the “a” into a well-formed “0.”In FuzzyProp terms, we say that each level of this stimulus factor corresponds to a letter that is to some degree an “a” and to some degree an “o”, with the particular degrees in each case varying regularly from level to level. The other manipulation involved closing in a loop over the final letter to turn it from a “y” to a “g.” Together, these two manipulations change the original word from “day” to “dog” through two independent factors, yielding a 7 x 7 matrix of stimuli. A 3 x 3 subset of this matrix is shown in Figure 1 for illustration of the technique.

Figure 1. A subset of a matrix of synthetic handwritten words.

290

G.C.Oden, J.G. Rueckl and T. Sanocki

Inspection of the resulting matrix of stimuli indicates the success of the stimulus generation process and also vindicates the reasoning that led to choosing to work with handwritten words in the frrst place. Each stimulus appears entirely natural, and, even in the extremely ambiguous cases, could well have been produced by a subject. Furthermore, each stimulus obtained from a particular original token clearly preserves the individuality of the original writer - they are all of his or her handwriting style. Several such matrices were constructed in this way to get at different aspects of the featural dependency issue. The stimuli from these matrices were then presented in a random, intermixed order for identification by other subjects. The results were graded identification probabilities that are systematic joint functions of the separate feature manipulations in each case. The FuzzyProp model of pattern identification provided a very good account for the quantitative pattern of the data without requiring the incorporation of crossletter feature dependency components (that is, for example, the proposition for “day” was essentially “d AND a AND y”; there was no term for an “ay” component predicate or for “y” conditional on “a” etc.). This indicates that the features were in fact evaluated independently without regard for each other. Featural independence has been regularly found in our earlier work, but it was, of course, a good deal less obvious that it would also hold with handwriting for which the possibilities of co-articulation between letters are so clear. To focus more specifically on the issue of co-articulation. we used two matrices that involved featural manipulations that seemed especially dependency prone: (1) The band/fund matrix - we had observed that the letter “f‘ was exceptional (perhaps the only case) in that a specific property clearly depended on the letter following; namely. the elevation of the terminating point of the upper loop ‘floats’ with the beginning height of the next character, which is high for ‘‘u” and low for ‘‘a’’.s (2) The pad/pace matrix - here, one feature manipulation involving the degree of closure of two loops not only determines letter identities but also letter segmentation; that is, whether there is one letter (either a “d” or an ‘‘a” depending on the other feature manipulation) or two (a “c” followed by either an “e” or an “I”). Even in these extreme cases, however, the FuzzyProp model was able to account for the data with only component feature predicates, indicating that the features are evaluated in an independent fashion. In another study involving quite a different experimental approach (Rendeiro & Oden, in preparation), we have obtained evidence using handwriting that directly supports the principle of cooperativity between communicative participants. This study, modeled after a classic study using spoken materials (Lieberman, 1963), demonstrated that words that have been excised from semantically constraining sentences are more identifiable than when the same words have been excised from relatively unconstraining sentences. Thus, with unconstraining sentences, handwrifers apparently take more care in articulating

Making Sentences Make Sense

29 1

the word forms to offset the paucity of information provided by higher linguistic levels. In so doing, they maintain a level of total information sufficient to allow handreaders to understand the sentences taken as a whole (which we then perversely precluded by yanking the word out of its sentence home). Although this study is not able to illuminate the featural characteristics that are involved in the sort of detail that was possible in the previously reported one, it does clearly show the role that the precisely modulated featural information observed there would play in the process of using word information at the sentence level. Information Accrual

Our earlier work on reading printed text was successful in establishing important characteristics of word and letter features and the processes by which such features are evaluated and integrated. But that work did not directly address the taking up of such featural information over the course of word identification. This is a topic that has been much studied using masking and reaction time procedures, but nearly all of the previous research was conducted within the traditional theoretical framework presuming functionally discrete features. As a result, it does not speak directly to the question we feel to be of the greatest importance: How is continuous featural information accrued? To study this issue, we developed a new technique that is complementary to traditional procedures but more amenable to providing the fine-grained information we require (Oden & Sanocki, in preparation). This technique employs a succession of degraded presentations of a word. The word is degraded by randomly changing (from black to white or from white to black as the case may be) a certain proportion of the pixels that define the word as figure and its background as field. As this proportion varies from 0 to .5,the bitwise predictability of the image varies from complete to nonexistent6 and the quality of the image varies from unblemished to uniformly random, featureless ‘snow.’ The successive presentations of a word began with relatively high degrees of degradation and proceeded to lower degrees so that the word effectively ‘emerged’ from the haze. Figure 2 illustrates part of this progression from 35% to 25% changed pixels. We were concerned with how the information that is available at each point in the progression is used by the subject and how the character of the word identity evolves as a result. Our initial studies using this technique made use of a large number of confusable word pairs (usually words differing in a single similar letter) in order to validate and calibrate the method with both free identification and forced choice response tasks. The results indicated that identifiability is a fairly continuous function of degree of degradation, that there are only small carry-over effects from one frame to later frames in a presentation sequence, and that the effective range from 0% to 100% identification covers roughly half of the logically possible range of degradation - enough to allow

292

G.C. Oden, J.G. Rueckl and T. Sanocki

Figvrc 2. Degraded words varying in bitwise predictability.

for it to be effectively manipulated for a wide array of stimuli and display types. In the key study in this series, we combined the successive haze removal technique with the continuous manipulation of letter features in a manner analogous to that used in the handwriting studies discussed above. We used two matrices, the watch/water and chase/erase matrices, that had been used in an earlier study (Oden, 1984a). These matrices had been designed to examine whether word envelope plays an independent role in word identification. The critical comparison hinges on the fact that, in addition to its effect on letter identity, the h/r feature manipulation also affects the word envelope and does so differentially in the two word contexts (wat- vs. -ase). This is in contrast to the e/c manipulation, which has no effect on envelope and thereby provides a baseline for the comparison. The results of that study indicated that word envelope does not play a role in word identification, at least not to any appreciable extent in terms of the final identification. Nevertheless, it is possible that such effects may be evident at early stages during the growth of identification. One could well imagine that envelope would be of most use when conditions are poor and letter feature information is scarce. Whether or not this is actually the case is one of the questions that the presently considered study was meant to address. The results of this study

Making Sentences M a k e Sense

293

revealed (1) that even with only partial information, the pattern of identification probabilities exhibits the form expected by the FuzzyProp model (indicating that from early in the identification process, featural information is employed in a continuous, independent, compensatory fashion) and (2) that there is little or no indication of word envelope effects even at early stages of processing. (The complete story, of course, is a good deal more complicated.) These conclusions are highly compatible with our view of sentence processing as comprising a succession of overlapping, parallel processes of component pick up and analysis at feature, letter, and word levels. It is also compatible with our view of identification at each of these levels as being an independent but integral part of the larger process of sentence comprehension. Thus, as each word is being identified, information about its form and meaning are combined with that of neighboring words and with the emerging interpretation of the sentence. Given that the reader or listener is developing interpretations at multiple levels simultaneously, it is important that experimentalists wishing to investigate this process do their part as writer or speaker. That is, if the experimentalist wishes to examine sentence processing in its full glory, he must cooperate with the subject by providing a stimulus for which the appropriate constraints apply. Some popular experimental paradigms are not cooperative in this sense. Consider, for example, the widely-used sentence-priming paradigm, in which the focus is on responses to a single target word preceded by either an incomplete sentence context or a non-informative baseline context (e.g., Stanovich & West, 1983; Tulving & Gold, 1963). To gauge the magnitude of contextual influences on word identification in this paradigm, the congruity of the context with the target word is often varied. However, the appearance of incongruous contexts reduces the validity of contextual information in the experimental environment. As a result, facilitory effects of congruous contexts may be reduced or eliminated, relative to when only congruous contexts appear in the experimental environment (Sanocki & m e n , 1984). This can occur both with measures of lexical decision latency (Sanocki & Oden, 1984) and naming latency (Norris, 1987). Even if the validity of contextual information is preserved in an experiment, there is another more subtle problem with paradigms in which a single target word is presented apart from its context. In normal reading, the processing of a word may be spread out over time in a cascaded manner (Sanocki et al., 1985). For example, the processing of a word in a sentence could begin as the reader picks up information about it parafoveally. Processing continues as the word is fixated and, in some cases, is not completed until subsequent words have been fixated. Sanocki et al. (1985) review evidence supporting these claims; see also, e.g.. Ehrlich & Rayner (1983) and Inhoff (1989). Recently, investigators using a gating paradigm to study speech perception have demonstrated that in a significant portion of cases, words in spoken sentences are

294

G.C.Oden, J.G. Rueckl and T. Sanocki

identified with the aid of information provided by subsequent words in the sentence (Bard, Shillcock, & Altmann, 1988). In our view, readers or listeners process words in a cascaded manner because they can expect that the word identification problem will be solved together with the sentence comprehension problem. Advantages stemming from cascaded processing would not be observed for single target words presented separately from their context. Therefore, we presented phrases, whole sentences, or scrambled sentences for subjects to scan for nonwords (Sanocki et al., 1985). In the first of these experiments, there was a large difference in scanning time for sensible sentences as compared with scrambled versions of the same sentences indicating contextual facilitation. The second experiment involved two comparisonsof the time needed to scan phrases preceded by different types of contexts. The phrases formed complete sentences when combined with meaningful contexts. In the syntactic comparison, scanning times were contrasted for phrases preceded by a meaningless baseline context or non-predictive but meaningful contexts. In the semantic comparison, scanning times were contrasted for phrases preceded by meaningful contexts that were either highly or moderately predictive. Consistent with the idea that both general syntactic or semantic information and specific semantic or pragmatic information contribute to word identification, we obtained facilitation for non-predictive information in the syntactic comparison and for highly predictive information in the semantic comparison. In the present view, sentence level constraints are combined with word level constraints to speed word identification beyond what could be expected if responses were determined by a race of autonomous word level and sentence level information (Forster, 1979). In the third experiment of this series, we used conditions in which responses were determined by word level information only, sentence level information only, or by information at either level. Consistent with the idea of the combination of mutual constraints, we found that there was a large advantage for processing in the condition in which both levels of information could be used. Moreover, this advantage was larger than would be predicted by a race of autonomous word level and sentence level processes (see Sanocki et al., 1985).

Font Specifics The emphasis within FuzzyProp on multiple continuous constraint satisfaction has led us well away from the traditional stance of expecting pattern specifications to be confined to the minimum of features, that is, to the smallest collection of stimulus properties that are adequate to discriminate the pattern from all relevant others. Instead, we expect that a n y stimulus property that is reliably and systematically associated with a letter or word will be taken advantage of in achieving robust identification. In essence, the old view presumed

Making Sentences M a k e Sense

295

that it would be easiest to identify a pattern if you only had to extract a few features whereas the new view is that identification is easiest when there are multiple, partially redundant features so that success does not depend on accurately picking up each of the few critical features. One line of research that dramatically exemplifies this shift in view is that of Sanocki (1987, 1988) regarding the importance of font-specific features in the identification of letters. This topic represents a particularly vivid contrast of the views, since the necessity of being able to identify letters across widely different fonts has traditionally been taken to be a prime motivation for throwing away all but the most essential features of a letter’s identity. Sanocki (1987) proposed that the letter identification process becomes modified for particular fonts, in order to efficiently use some of the detailed, partially redundant information that is specific to a font. As a result, letter identification should be more efficient with a single font than with a mixture of fonts. In a number of experiments, performance identifying short strings of unrelated letters was compared for blocks in which all target letters were from a single font and for blocks in which the target letters were from two or more fonts. Performance was higher in consistent font conditions, both in a reaction time task (Sanocki, 1987). and in a backward-masking task in which items were presented briefly and accuracy was measured (Sanocki, 1988). Further, in the reaction time task, additivity of font-consistency and string length effects implied that font consistency affected early stages of processing. Time course data in the backward masking task provided converging support for this conclusion. These results support the idea that perceivers “tune into” and use font-specific details of letters. The issue of effects of font variation in reading is an interesting one. It is clear that reading is slowed by large variation in type such as that produced by mixing letters that differ in case, or that differ substantially in size (e.g., Rudnicky & Kolers, 1984). However, to our knowledge there are no reports of reading being affected by more subtle manipulations such as variation in font with size or case constant. We suspect that, in proportion to the totality of information available from all sources and levels, the effects of font variation become vanishingly small in the context of reading.’ A reduction in effects of font variation with a sensible context would be consistent with the notion that comprehension results from compensatory integration processes. A meaningful context provides constraints that can compensate for difficulties at lower levels, and this would reduce and perhaps eliminate effects of font variation.

MAKING SENTENCES Syntactic Analysis

In a set of studies that we feel will soon be seen as having been seminal in their area, Beach (submitted) examined the role of prosodic features on syntac-

296

G.C. Oden, J.G. Rueckl and T. Sanocki

tic ambiguity resolution. This work was provoked by a current example of discrete thinking within the field of syntactic parsing, the Minimal Attachment Hypothesis of Frazier (1978; see also Frazier and Rayner, 1982). This hypothesis is a modem instance of the long tradition within linguistics and psycholinguistics of trying to uphold the claim that syntactic analysis operates relatively autonomously. The hypothesis is that the syntactic parsing mechanisms will at least initially analyze sentences as having one grammatical form rather than another on the basis of purely syntactic factors (namely, which yields the simplest arrangement of ‘attachments’ in the underlying syntactic structure). During this stage of processing, non-syntactic sources of information are not considered.* Such a position is, of course, anathemic to our massively multiple constraint satisfaction view. Indeed, it was i n response to an earler instantiation of the autonomous syntactic processing view, then in the form of ‘perceptual strategies’ (Bever, 1970; Kaplan, 1972), that the FuzzyProp approach was originally developed. This was in the course of showing that continuous semantic constraints play a critical role ifi determining syntactic ambiguity resolution even for sentences for which both interpretations are sensible9 (Oden, 1978; see also Oden, 1983). Beach set out to address this issue from a different angle by demonstrating that prosodic factors influence syntactic analysis af the point of syntacfic choice. To illustrate the issue, consider the sentence fragment The city council argued the mayor’s position... which could be continued by eithcr of the following:

forcefully. (The city council argued the mayor’s position forcefully). or

was incorrecl. (The city council argued the mayor’s position was incorrect). These two sentence continuations correspond to two possible syntactic analyses for the sentence fragment. In the former case, the noun phrase “the mayor’s position” is the direct object of the main verb “argued” whereas in the latter case, it is the subject of a relative clause (i.e., “the mayor’s position was incorrect”). According to the Minimal Attachment hypothesis, the former analysis should be preferred since it involves a simpler attachment structure. Thus, if syntax is autonomous, the parser should attempt to use this direct object interpretation until forced to abandon it, which means that it should always be the interpretation that is held when only the words of the fragment shown above have been received. In Beach’s experiment, subjects heard such fragments read

Making Sentences Make Sense

297

under various prosodic conditions established by an earlier study to be relevant to syntactic structure. Her results showed that segment duration and pitch contour systematically determined the interpretation obtained for the ambiguous sentence fragments in an independent, jointly continuous fashion. That is, both duration and pitch rise and fall influenced the synactic interpretation that was obtained and the contributions of these two factors was well accounted for by the FuzzyFVop model. Thus, this kind of prosodic information is apparently integrated with that from syntactic biases during the course of parsing to determine the syntactic structure that is assigned to a sentence. In a related series of studies, Lustgarten (1985) examined the processing of another sort of ambiguity, that of reference and anaphora. As in the studies described above, he made use of FuzzyRop to ascertain the way in which various sources of information cooperatively but independently determine the assignment of reference. For example, in one of the studies, the stimuli were sentence pairs such as “Cindy hosted a big party after sealing an important business deal. It had taken a lot of planning.” Whether the pronoun “it” was taken to refer to the event of the main clause or that of the subordinate clause depended in the predicted smoothly systematic way on the plausibility of each of the respective events requiring a lot of planning. This was manipulated by independently varying the two verb phrases, e.g. “Cindy bought an ice cream cone after quitting her job.” or “Cindy asked Fred out on a date after eating dinner,” etc. Another study involved an interesting but seldom studied type of ambiguity such as that exemplified by “Anne wanted to buy a piano.” In this case, we presume that “a piano” refers to no particular piano, in contrast to the case with sentences like “Anne wanted to sell a piano.” or “Anne managed to buy a piano.” Changing the two verbs in subtle but important ways such as this determined the degree to which “a piano” was taken to have a specific as opposed to generic reading and did so in the manner expected according to the model. That is, “sell” pushes one toward a specific reading to such and such a degree, “wanted” pushes one toward a generic reading to its own degree, etc., and the net ‘leaning’ toward one reading versus the other depends on the balance of the respective forces of the ‘pushes’. Cross-level Integration Within the FuzzyProp framework, the overarching motivation for examining evaluation and integration processes at each of the several linguistic levels is to uncover the kinds of continuous information that would be available to the language processing system in determining the overall best-fitting interpretation of a spoken or written utterance. Perhaps the keystone of this research program has been the one set of studies (Rueckl & Oden, 1986) that have explicitly taken a look at the process of integrating such information from widely different linguistic levels, namely, information derived from continuous

G.C.Oden, J.G. Rueckl and T. Sanocki semantic constraints and from continuous featural values. As in many of the studies discussed above, featural information was manipulated by continuously varying a feature that distinguishes between two otherwise identical letters. For example, in the r/n continuum, the length of the line that distinguishes “r” from “n” was varied. These letter tokens were then embedded in letter strings, devised so that on either interpretation of the letter token, the string formed a word (e.g.. the r/n tokens were embedded in the strings “pai-” and “bea-s”). The letter strings were in turn embedded in sentence contexts that vaned in the degree to which they supported the alternative interpretations of the strings. For example, in the sentence frame The -has a (pair/pain) in his hand. the blank was replaced by “cardplayer,” “shoemaker,” “piano player,” and “arthn tic.” Note that the relative sensibleness of the alternative interpretations of the target (“pair”/“pain”) differs across these contexts. The “cardplayer” context is more sensibly completed by the target “pair,” whereas the “arthritic” context is more sensibly completed by “pain.” The “shoemaker” and “card player” contexts are less strongly biased in either direction. Subjects were presented with a number of such sentences, and were asked to either read the sentence aloud or to choose which of the alternative interpretations of the target word had been in the sentence. The results indicated that both the featural and the contextual manipulations influenced the outcome of the identification process. Moreover, the FuzzyProp model provided a good quantitative account of the results, suggesting that, as proposed by the model, the featural and semantic support for a given interpretation are evaluated independently of one another. The results of these evaluations are then integrated in a continuous, compensatory fashion in order to select the interpretation of the stimulus that simultaneously best satisfies the constraints at all linguistic levels.

FUZZYPROP A N D CONNECTIONISM In recent years, we have been exploring connectionist models as a grounding for FuzzyProp constructs. Unlike most other folks working with such models, our aim has not been to supplant our old model. Indeed, one of the main motivations other people have for turning to connectionism - to achieve the robustness that results from continuousness and multiple soft constraint satisfaction - is something that we already have with FuzzyProp. Instead, we are interested in the twin aims of (1) extending FuzzyProp down to the level of (more nearly) realistic brain mechanism implementation and (2) to illustrate how FuzzyProp may effectively serve as the symbolic levello description of connectionistically formulated cognitive accounts.

Making Sentences Make Sense

299

These aims seem to us to be of obvious value to achieve and reasonable to pursue, but this sort of enterprise is not without critics. Some seem to feel that it is trivial since FuzzyProp and connectionism share so many basic properties. Others feel it is impossible because of some fundamental incompatibility between symbolic and subsymbolic approaches generally. Naturally our feeling is that it is somewhere between trivial and impossible; that is, not too easy and not too hard, but just right. Of course, there are indeed trivial solutions to be had for some posings of this problem. For example, one might rename “proposition” to be “node” and “degree of truth” to be “level of activation” and so on and then proclaim FuzzyProp to be a connectionist model. Or, a more formal minded person might translate both FuzzyProp and connectionism into something like Turing machine equivalents and then try to use that as an interlanguage basis for compiling one into the other. But neither of these would count as attaining our goal of providing a direct relating of one level to the other in a way that is true to the spirit of each. On the other hand, it is also possible to define the problem in such a way that it is impossible, as Fodor and Pylyshyn (1988) appear to do by ruling out (as philosophically uninteresting) solutions that involve “mere implementation” of symbolic functioning in connectionist terms. In our view (Oden, 1988, in press; Rueckl, 1986, in press), there is no such thing as mere implementation; that is, much of science and also of engineering (whether cognitive or otherwise) is concerned with exactly how one level of reality is implemented in terms of another. To say how minds are instantiated in brains should be neither trivial nor impossible. From our attempts so far, we can attest that it is hard, but progress is being made and whether or not it will prove to be doable is ultimately an empirical question. Presuming that this program is at least worthy and feasible, we can consider what we might expect to gain from it in practical terms. On the one hand, connectionism can be expected to provide FuzzyProp with the computational primitives needed to address several difficult issues. For example, connectionism can help provide FuzzyProp with a principled account of how knowledge is acquired, especially in situations where this happens in a gradual, incremental fashion over repeated experiences. Similarly, connectionist primitives could help explain how fuzzy propositions - such as those describing the semantic structure of a sentence or the orthographic structure of a pseudoword - are constructed “on-line.” On the other hand, FuzzyProp can be expected to provide to connectionism a means for representing and dealing with recursive processes and highly compositional (e.g. self-referential) knowledge structures. In each case, the account will remain naturally rooted in its ‘home’ level; if translated into the terms of the other level, it will undoubtably be at least somewhat awkward and ad hoc. Thus, we believe that any complete, natural cognitive story will necessarily include both propositional and connectionist levels.

300

G.C.Oden, J.G. Rueckl and T. Sanocki

CONCLUSION We have described a chunk of an ongoing research program and, not incidentally, a frame of mind. As is undoubtably clear from this description, the program has not even tried to pretend to proceed in the incremental deductive fashion taught in experimental psychology textbooks. The operative metaphor has been not so much that of assembling an edifice brick by brick as it has been that of making a sheet take on the form of what it covers by tacking it down here and there. Or - to use an analogy that is more apt within the present context - we consider the scientist to be in collaborative conversation with nature: To have any realistic chance of success, the scientist must presume that nature has provided a coherent message and must use available clues, paltry as they may be, to make as much sense out of the message as he can. In this way, we see ourselves as just like the language comprehender who must use all he knows about words to achieve the effect of making sentences make sense. Acknowledgments The research reported here was supported by grants to the first author from the National Science Foundation (BNS83-10870) and from the Wisconsin Alumni Research Foundation. Notes I None of this is novel, of course. Indeed, it is widely accepted among those who have thought most about it, that is, sociolinguists and sociolinguistically-minded psycholinguists; see, for example, Clark (1985) and references therein. An probably only what he needs to do is this; that is, the conversants mutually subscribe to a ‘principle of least collaborative effort’ according to Clark and Wilkes-Gibbs (1986). Indeed, doing more than necessary is taken to be a meaningful act in itself (Grice, 1975) and therefore to be avoided unless the resulting inference is desired. This is also the case with other popular continuous integratio functions such as distance measures in multidimensional spatial formulations. It doesn’t matter that the dimension correspond to only a single perceptual feature provided that it doesn’t unnaturally ‘cut across’ any feature. Every letter has characteristic starting and ending elevations, but ordinarily the discrepancy that may exist between the ending height of one and the beginning height of the next is absorbcd in the transition from one to the other rather than in any substantial systematic perturbation in the features of the letters themselves. If half of the pixels are changed, then whether one is black or white is

Making Sentences Make Sense

301

totally uninformative about its original state. If such effects were easy to obtain, one might expect them to have been discovered by Tinker (1963), who seems to have systematically examined every variable that might be relevant to reading speed. * The hypothesis allows that subsequent ‘verification checks’ may force a re-analysis in the case, for example, where the preferred parse leads to a nonsensical interpretation. When both are sensible, the syntactic autonomy position cannot resort to verification check failure as the recourse for obtaining the alternative interpretation. lo The term “level” is used here in a different sense from that in the phrase “linguistic level” and also differently from the way in which it was used by Marr (1982). The sense here is of levels of abstraction or identity of some entity as when the same electrical charge within a computer is accurately described as +5v, as an ‘on’ bit, and as a minus sign at physical, machine, and program levels, respectively.

’

References Bard, E. G.,Shillcock, R. C., & Alunann, G. T. M. (1988). The recognition of words after their acoustic offsets in spontaneous speech: Effects of subsequent context. Perception & Psychophysics, 44,395-408. Beach, C. M. (submitted). Prosody and parsing spoken language. Bever, T. G. (1970). The cognitive basis for linguistic structures. In J. R. Hayes (Ed.) Cognition and the development of language. New York: Wiley. Clark, H. H. (1985). Language use and language users. In G. Lindzey & E. Aronson (Eds.) Handbook of social psychology (3rd Ed.). New York: Harper and Row. Clark, H. H., & Wilkes-Gibbs, D. (1986). Referring as a collaborative process. Cognition, 22, 1-39. Ehrlich, S. F., & Rayner, K. (1981). Contextual effects on word perception and eye movements during reading. Journal of Verbal Learning and Verbal Behavior, 20, 641-655. Fodor, J., & Pylyshyn, Z. (1988). Connectionism and cognitive architecture: A critical analysis. Cognition, 28, 3-71. Forster, K. I. (1979). Levels of processing and the structure of the language processor. In W. E. Cooper & E. Walker (Eds.). Sentence processing: Psycholinguistic studies presented to Merrill Garrett. Hillsdale, NJ: Erlbaum. Frazier, L. (1978). On comprehending sentences: Syntactic parsing strategies. Unpublished doctoral dissertation, University of Connecticutt. Frazier, L., & Rayner, K. (1982). Making and correcting errors during sentence comprehension: eye movements in the analysis of structurally ambiguous

302

G.C. Oden, J.G. Rueckl and T. Sanocki

sentences. Cognitive Psychology, 14, 178-210. Gibbs, R. W. (1984). Literal meaning and psychological theory. Cognitive Science, 8, 275-304. Grice. H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.) Syntax and semantics 3: Speech acts. New York: Academic Press. Inhoff, A. W. (1989). Lexical access during eye fixations in reading: Are word access codes used to integrate lexical information across interword fixations? Journal of Memory and Language, 28,444-461. Kaplan, R. M. (1972). Augmented transition networks as psychological models of sentence comprehension. Artificial Intelligence, 3,77-100. Lieberman, P. (1963). Some effects of semantic and grammatical context on the production and perception of speech. Language and Speech, 6, 172-187. Lustgarten, P. C. (1 985). Decision processes in understanding English discourse anaphora. Unpublished doctoral dissertation, University of Wisconsin. Marr, D. (1982). Vision. San Francisco: W. H. Freeman. Massaro, D. W. (1987). Speech perception by ear and eye: A paradigm for psychological inquiry. Hillsdale, NJ: Erlbaum. Norris, D. (1987). Strategic control of sentence context effects in a naming task. Quarterly Journal of Experimental Psychology, 39A, 253-275. Oden, G . C. (1978). Semantic constraints and judged preference for interpretations of ambiguous sentences. Memory & Cognition, 6,26-37. Oden, G. C. (1983). On the use of semantic constraints in guiding syntactic analysis. International Journal of Man-Machine Studies, 19, 335-357. Oden, G. C. (1984a). Dependence, independence, and emergence of word features. Journal of Experimental Psychology: Human Perception and Performance, 10,394-405. Oden, G . C. (1984b). Integration of fuzzy linguistic information in language comprehension. Fuzzy Sets and Systems, 14,294 1. Oden, G. C. (1988). FuzzyProp: A symbolic superstrate for connectionist models. Proceedings of the IEEE International Conference on Neural Networks, Vol. I , 293-300. Oden, G. C. (in press). Connectionism: Self-abuse is improper treatment. (Commentary on “On the proper treatment of connectionism” by P. Smolensky.) Brain and Behavioral Sciences. Oden, G . C., & Massaro, D. W. (1978). Integration of featural information in speech perception. Psychological Review, 85, 172- 19 1. Oden, G. C., & Rueckl, J. G. (In preparation). Taking language by the hand: Reading handwritten words. Oden, G. C., & Sanocki, T. (In preparation). Information accrual and integration in word identification. Pollack, I., & Pickett, J. M. (1964). The intelligibilty of excerpts from conversation. Language and Speech, 6, 165-171. Rendeiro, T., & Oden, G. C. (In preparation). Some effects of semantic and

Making Sentences Make Sense

303

grammatical context on the production and perception of handwriting. Rueckl, J. G. (in press). The connectionist research program: Something old, something new. T. Horgan & J. Tienson (Eds.), Connectionism and the philosophy of mind.Norwell, MA: Kluwer Academic Press. Rueckl, J. G., & Oden, G. C. (1986). The integraton of contextual and featural information during word identification. Journal of Memory and Language, 25,445-460. Rudnicky. A. I., & Kolers, P. A. (1984). Size and case of t y p as stimuli in reading. Journal of Experimental Psychology: Human Perception and Performance, 10,231-249. Sanocki, T. (1987). Visual knowledge underlying letter perception: Font-specific, schematic tuning. Journal of Experimental Psychology: Human Perception and Performance, 13,267-278. Sanocki, T. (1988). Font regularity constraints on the process of letter recognition. Journal of Experimental Psychology: Human Perception and Performance, 14.472-480. Sanocki, T., & Oden, G. C. (1984). Contextual validity and the effects of lowconstraint sentence contexts on lexical decisions. Quarterly Journal of Experimental Psychology, 36A, 145- 156. Sanocki, T., Goldman, K., Waltz, J., Cook, C., Epstein, W.,& Oden, G. C. (1985). Interaction of stimulus and contextual information during reading: Identifying words within sentences. Memory & Cognition, 13, 145-157. Stanovich, K. E., & West, R. F. (1983). On priming by a sentence context. Journal of Experimental Psychology: General, 112, 1-36. Tinker, M. A. (1963). Legibility of print. Ames, IA: Iowa State University Press. Tulving, E.. & Gold, C. (1963). Stimulus information and contextual information as determinants of tachistoscopic recognition of words. Journal of Experimental Psychology, 6 6 3 19-327.

This Page Intentionally Left Blank

Understanding Word and Sentence G.B.Simpson (Editor) Q Elsevier Science Publishers B.V. (North-Iiolland), 1991

Chapter 12 How is Verb Information Used During Syntactic Parsing? Fernanda Ferreira and John M . Henderson

University of Alberta Edmonton, Alberta Canada

Until recently, explorations into language processing in cognitive psychology have consisted of attempts to understand how words are identified (at one end of the language processing sequence) and how the relations among sentences are computed (at the other). Little emphasis has generally been given to the sentential level of analysis, and the research that has been conducted at this level has traditionally focused on semantics: establishing a sentence’s propositional structure and identifying the semantic roles (e.g., agent, patient) played by its constituent phrases. Recently, however, psycholinguists have argued that language processing also includes a stage at which the syntactic structure of a sentence is created. For example, in the sentence “The boy bit the dog,” it is the syntax or constituent structure that determines which actor is the biter and which is the bitee. Following others, we will refer to the process of assigning syntactic structure to a sentence as pursing, and that aspect of the language processing system responsible for parsing an input string as the parser. One way to explore the operation of the parser is through the use of sentences that are temporarily ambiguous. For example, consider a sentence such as “Mary knew the answer was correct.” The noun phrase the answer is ambiguous when it is first encountered, because it could function either as the direct object of knew. or as the subject of an upcoming embcdded clause (the answer was correct). I n this example, the syntactic context following the verb forces the latter interpretation. There are a number of ways that the parser might deal with such a sentence. The parser could delay making any decision about how to attach the ambiguous phrase when it is first encountered, and wait until clearly disambiguating information arrived ( h e word was constitutes such information). Alternatively, the parser could immediately attach the noun phrase, either by considering only one analysis (which might sometimes lead to errors), or by considering both analyses in parallel (which might be computa-

306

F. Ferreira and J.M. Henderson

tionally costly). Entwined with the question of how the parser deals with syntactic ambiguity is the question of whether non-syntactic sources of information can be used to guide the parser’s decisions about which structure to adopt or retain. On the one hand, the parser might use syntactic rules that refer only to syntactic categories (e.g., noun, verb), and apply those rules “blindly,” without regard for semantic or plausibility constraints. On the other hand, the parser might use non-syntactic information during initial parsing either to select just the most plausible analysis (on the view that the parser operates serially), or to increase the strength or activation level of one structure at the expense of another (on the view that the parser constructs alternative structures in parallel). This question of whether non-syntactic information is consulted by the parser when making initial parsing decisions has recently been phrased in terms of the modularity thesis (J.A. Fodor, 1983; Frazier, 1985). A modular parser is one that consults only a limited domain of information (presumably syntactic information) when performing its task of assigning syntactic structure (Forster, 1979). A nonmodular or interactive parser would be one that had access to syntactic as well as more general cognitive information (e.g., Kurtzman, 1984; Marslen-Wilson, 1975; McClelland, 1987; Taraban & McClelland, 1988). Frazier (1978, 1987, 1989) has proposed a parsing model that takes a position on each of the above issues. First, words are immediately assigned to a syntactic structurc as they are received by the parser (except when the syntactic category of a word is ambiguous; Frazier & Rayner, 1987). Thus, the model proposes immediate rather than delayed assignment of syntactic structure. As Frazier has pointed out, words arrive rapidly during normal sentence processing (about 3 to 4 per second in reading and in speech), and so must be rapidly integrated into an overall representation if the sentence is to be understood. Second, the parser constructs only a single syntactic structure as each new word is encountered. Thus, Frazier’s model is serial rather than parallel. The single structure initially constructed is always the syntactically simplest structure. For example, in the sentence “Mary knew the answer was correct,” the parser would attempt to incorporate the answer as the direct object of knew. because this structure is simpler than one where the answer is the subject of a new clause (i.e., the direct object structure requires fewer syntactic nodes). Third, the simpler structure is always chosen regardless of the presence of higher level information concerning semantic plausibility, even when the plausibility information is more consistent with the less simple syntactic analysis. The model is therefore modular rather than intcractivc. Notice that because the verb was in the example sentence cannot bc integrated into the simplest syntactic structure consuuctcd to that point, the initial analysis of the sentence will be incorrect, and the sentence will have to be reanalyzed towards the more complex analysis. Because Frazier’s proposed parser will misanalyze a sentence when the simplest analysis turns out to be wrong (i.e., will be led down the “garden-path”), it

Verb Information During Syntactic Parsing

307

has come to be known as the garden-path model of parsing (Frazier & Rayner. 1982). A good deal of evidence has been accumulating in favor of the gardenpath model. In a seminal study, Frazier and Rayner (1982) monitored the eye movements of readers while they read temporarily ambiguous sentences. Like our example sentence above, these sentences contained an ambiguous noun phrase that either ended up being a direct object of the main verb or the subject of an embedded clause. Frazier and Rayner found relatively long fixation times following the ambiguous region of the sentence, but only in the syntactically more complex condition. This result indicated that readers initially adopted the simpler structure. If the sentence continued in a manner consistent with the simpler interpretation, there was no elevation in reading time. However, when the initial simple analysis turned out to be incorrect (as indicated by a disambiguating word such as was in the example sentence), readers had to revise their initial analysis. This revision caused a discernible increase in reading time on the disambiguating word. Thus, Frazier and Rayner argued that the parser initially follows a minimal attachment principle-the parser attempts to construct the simplest structure that is grammatically licensed. The Frazier and Rayner (1982) study supported the hypothesis that only one syntactic analysis (the simplest) is initially constructed for a given input string (for further evidence, see also Ferreira & Henderson, 1990a). However, in that study, readers saw single sentences isolated from any constraining semantic context. It could be the case that given a biasing semantic context favoring the nonminimal analysis, the more complex nonminimal analysis would be initially constructed (Alunann & Steedman, 1988; Crain & Steedman, 1985). On the other hand, Frazier (1978, 1985) argued that non-syntactic information such as that provided by semantic context would be unable to override the minimal attachment principle, because the parser is a processing module and initially consults only syntactic information. This view leads to the prediction that semantic information should not influence how a word is attached into the syntactic structure under construction. Rayner, Carlson, and Frazier (1983) found evidence consistent with this prediction: The meaning of the sentence up to the current word did not bias the parser to initially adopt a nonminimal attachment structure. For example, subjects took more time to read a sentence such as “Mary played the record with the scratches” (nonminimal attachment) than a sentence such as “Mary played the record with the needle” (minimal attachment). However, it could be argued that Rayner et al.’s within-sentence semantic context was neither strong enough nor available fast enough to influence the parser, and that a more powerful manipulation of semantic context would cause the parser to construct a nonminimal attachment structure. In order to test this hypothesis, Ferreira and Clifton (1986) created discourse contexts that strongly favored the nonminimal attachment analysis. Consistent with the modularity assumption of the garden-path model, they found evidence for

308

F. Ferreira and J.M.Henderson

strong garden-pathing effects even with these highly biased discourse contexts. Alternative views concerning the influence of non-syntactic knowledge, however, have been proposed. For example, Altrnann and Steedman (1988) took issue with Ferreira and Clifton’s interpretation of their results, and argued that appropriate control contexts (against which the minimal attachment effect was assessed) were not used. Altrnann and Steedman conducted a study employing these controls. The target sentences were sentences such as “He blew up the safe with the lock,” where the ambiguous phrase was the prepositional phrase with the lock. The phrase could either be a locative phrase (the minimal attachment interpretation), or a modifier of the safe (the nonminimal attachment interpretation). Using a self-paced phrase-by-phrase reading time task with a cumulative display (i-e.. successive phrases remained visible as subjects paced through the sentence), Almann and Steedman found that nonminimal attachment sentences in an appropriate context were as easy to parse as their minimal attachment counterparts-in fact, they were easier. However, Clifton and Ferreira (1990) criticized the Altmann and Steedman study on a number of grounds, including their use of the cumulative self-paced reading task. (We will discuss the limitations of this method later in this chapter.) In addition, Clifton and Ferreira did not replicate the nonminimal attachment superiority once a confound present in the Altmann and Steedman materials was eliminated. Minimal and nonminimal attachment versions were about equally difficult, a result which Clifton and Ferreira attributed to the use of the self-paced reading task. (See Steedman and Altmann, 1990, for further debate on these issues.) Similarly, Taraban and McClelland (1988) have argued that the minimal attachment principle can be overridden in sentences that are sufficiently semantically biased. Like Altmann and Steedman, Taraban and McClelland conducted a study employing sentences with an ambiguous prepositional phrase. The sentences were semantically biased eilher towards a minimal attachment or a nonminimal attachment reading. Using a word-by-word self-paced reading task, they found no increase in reading times for nonminimal attachment sentences that were appropriately biased. However, as was the case with the Altmann and Steedman (1988) study, the Taraban and McClelland results were based on the use of a self-paced reading task. Although some variants of this task (e.g., the noncumulative display version used by Taraban and McClelland) can reveal garden-path effects, particularly when the reinterpretation of the sentence is difficult (Ferreira & Clifton, 1986; Ferreira & Henderson, 1990b; Taraban & McClelland, 1988). the task may not be sensitive enough to distinguish the effects of initial analysis from effects of reanalysis. The bulk of the evidence, then, suggests that semantic context derived from overall sentence or discourse meaning cannot overcome the parser’s bias to assign initially the simplest syntactic structure to a sentence (see also Clifton & Ferreira, 1990, for a more complete review of the evidence for this position). However, there is a another source of infoxmation that could potentially bias the

Verb Information During Syntactic Parsing

309

parser away from the simplest syntactic structure. It has been proposed that information about the syntactic properties of verbs can be used to constrain the structure given to a sentence. Consider the example “Mary knew the answer was correct.” Intuitively, it seems that if a verb such as realized were substituted for knew, the sentence would be easier to process. The reason for this intuition is that the verb know is ambiguous between two subcategorization frames: It permits both direct objects and sentential complements. A verb such as realize, however, does not take direct objects as easily, and therefore the parser is less tempted to try to make the answer a direct object. According to a parsing model that permitted the use of verb information in initial parsing (which we will refer to as a verb-guidance model), upon receipt of the verb realize, the parser would have an expectation that a sentential complement will follow (Holmes, 1987; Holmes. Stowe. & Cupples, 1990). The phrase the answer would thus be analyzed as the subject of the embedded sentence, and so the sentence would not be initially misanalyzed. According to a verb-guidance model, then, a principle such as minimal attachment is not a real operating principle of the parser. Evidence for the principle is obtained only because verbs such as know are used, which prefer to be followed by direct objects rather than sentential complements. If a verb such as realize were used instead, the sentential complement structure would not be problematic for the parser. In short, according to a verb-guidance model. ease of parsing is based on the extent to which the structure of the sentence is consistent with the verb’s most frequent use. It is important to note that parsers which use verb information are modular parsers. According to verb-guidance models, the syntactic properties of verbs are used to influence the process of constructing a syntactic tree. Upon access of a verb such as know, its meaning, syntactic category, and subcategorization information would all be activated. This information would then be used in a bottom-up fashion to determine the characteristics of the rest of the syntactic tree. Thus, the issue between the garden-path and verb-guidance models is not modularity per se, but rather the extent of modularity. The possible use of verb subcategorization information during online parsing is still a use of syntactic information. If. however, it turns out that such information is not used during initial parsing, it would suggest that not all syntactic information is uniformly consulted during parsing. Instead, such a finding would suggest that there is modularity within the syntactic processor itself. Evidence for modularity within the syntactic processor has been obtained by Clifton and Frazier (1986, 1989) and Freedman and Forster (1985): They found that syntactic information about phrase structure was consulted at a different point in processing from syntactic information about binding constraints (Chomsky, 1981; 1986). (For a contrary view of these results, see J.D. Fodor, 1988.) Similarly, evidence that verb subcategorization information is consulted at a different stage from phrase-structure information would constitute evidence that the syntactic module actually

3 10

F. Ferreira and J.M.Henderson

consists of a number of smaller, more specialized submodules. In the remainder of this chapter, we will compare the following two models: the garden-path model, so-called because the parser is presumed to construct an initial analysis based on contentless parsing principles, and the verbguidance model, according to which verb subcategorization information is used during the initial construction of a phrase structure me.

THEGARDEN-PATH MODELvs. THEVERB-GUIDANCE MODEL Distinguishing between the garden-path model and the verb-guidance model of parsing requires both the careful construction of materials and the use of a task that is sensitive to the difference between initial and later stages of syntactic analysis. On the first point, most experiments designed to examine the use of verb information in parsing take the following form: Sets of sentences are constructed that differ in syntactic complexity according to the minimal attachment principle. In most of the studies we will discuss here, contrasting sentences are used that are either syntactically ambiguous and end up being nonminimal attachment sentences (garden-path sentences), or are syntactically unambiguous, and so do not pose any problems for the parser. The unambiguous conditions serve as a baseline against which to compare reading times for the ambiguous sentences. Then, for the ambiguous and unambiguous versions, the sentences are varied so that their main verb is either consistent with minimal or nonminimal attachment. The quadruple below illustrates the contrasts: (1)

a. The girl knew that the man cheated on the exam. b. The girl realized that the man cheated on the exam. c. The girl knew the man cheated on the exam. d. The girl realized the man cheated on the exam. The verb knew is consistent with minimal attachment of the ambiguous noun phrase, because it permits a direct object; realize is more consistent with nonminimal attachment, because it rarely takes a direct object. On the gardenpath model, garden-pathing should occur with both nonminimal attachment sentences (c and d): That is, there should be an elevation in reading times on the disambiguating word cheated regardless of verb bias. If the elevation in reading time is less in the biased version (d) than the unbiased version (c), the result is still consistent with the garden-path model. This interaction would indicate that initially, the parser was fooled by both nonminimal sentences, but recovered from its error faster with the sentence containing a verb biased towards the ultimately-correct interpretation. In other words, such a pattern would provide evidence for the claim that reanalysis processes operate more efficiently with a biased verb, but initial parsing is unaffected. The verb-guidance model predicts

Verb Information During Syntactic Parsing

311

that garden-pathing will only be found in the (c) version, because only in that condition is the bias of the verb inconsistent with the ultimate syntactic form of the sentence. Reading times on the disambiguating word cheated should be equivalent in (d), (a), and (b), since the verb bias in (d) should prevent the parser from misanalyzing the previous noun phrase. To distinguish between initial and later stages of parsing, it is necessary to use an appropriately sensitive task. To begin, it should be noted that virtually all of the studies that have been done on phrase structure parsing have been done visually, and so little is known about the effects of intonation, stress, and other auditory factors on parsing processes, or the effects of the time course of information availability in auditory versus visual processing. This fact is lamentable, but such is the state of the field (exceptions include Slowiaczek, 1981; Speer & Slowiaczek, 1990). Thus, all the tasks we will discuss are reading tasks of one sort or another. Second, a task is required that can distinguish reading times in each of the different regions of the sentence. For example, one wants to know if reading times are specifically elevated on the word cheated in (lc and Id), since that is the word that disambiguates the analysis of the previous noun phrase. Thus, three different tasks have been widely used: self-paced region-byregion reading with cumulative display, self-paced region-by-region reading with noncumulative display, and eye movement recording. The two self-paced tasks differ in their sensitivity to initial analysis processes (Ferreira & Henderson, 1990b). Intuitively, it is apparent that subjects reading in the cumulative paradigm might adopt a strategy of simply pushing the pacing button until the entire sentence is visible, and then reading the sentence. This intuition has been experimentally verified by Just, Carpenter, and Woolley (1982), and Ferreira and Henderson (1990b). Thus, this task is less than ideal for revealing the on-line parsing of sentences. The non-cumulative self-paced task (in which only one region is visible at a time) is far superior, and in some experiments, yields results that are similar to those obtained using eye movement recording, particularly if the region is only one word long (Ferreira & Clifton, 1986; Ferreira & Henderson, 1990b). It has been our experience that the smaller the regions that are presented, the more the results from selfpaced reading resemble those obtained from eye movement recording, so that the best results are obtained with word-by-word presentation. For example, in the sentences in (1) above, it would not do to have a region as large as, say, cheated on the exam. Consider sentence (Id). If the parser is initially gardenpathed on the disambiguating word cheated but then recovers from the misanalysis quickly because of helpful information from the verb, the reading time effect could easily be washed out over such a large region, but would be apparent if the region comprised only the disambiguating word itself. Despite the fact that the word-by-word self-paced reading task with noncumulative display yields results similar to eye movement monitoring, the latter has advantages that make it perhaps the best task to use. First, as argued by a

312

F. Ferreira and J.M. Henderson

number of researchers (e.g., Henderson & Ferreira, 1990; Henderson, Pollatsek. & Rayner, 1989; McConkie, 1979; Momson, 1984). the attentional and eye movement systems are functionally coupled, so that the point at which subjects are fixating is generally also the locus of processing (see also, Just & Carpenter, 1987; Rayner & Pollatsek, 1989). In contrast, attention is not linked so closely to button-pressing, and subjects have to learn during an experimental session how to make button presses in the service of comprehension. In addition, subjects who are reading while having their eye movements recorded are free to make both forward and backward eye movements, and so can re-read sections that have been syntactically misanalyzed (Ferreira & Clifton, 1986; Ferreira & Henderson, 1990b; Frazier & Rayner, 1982). In contrast, the word-by-word non-cumulative self-paced reading task forces subjects either to adopt an unnaturally careful initial reading so that reanalysis will not be necessary, or to reanalyze misunderstood text from memory. Finally, the subtler the garden-path induced by a syntactically difficult sentence, the more likely it is that the effect will be washed out in self-paced reading. Having dealt with these methodological preliminaries, we will now turn to a discussion of studies designed to assess the effects of verb information on parsing processes. Holmes (1987; Holmes et al., 1990) has been a proponent of the verb-guidance hypothesis. In one set of experiments, Holmes (1987) claimed to find that garden-pathing did not occur with biased verbs. Holmes used a word-by-word grammaticality judgment task with a cumulative display, in which subjects paced through each sentence and indicated at each word whether the sentence could continue in a grammatical fashion. Subjects were shown sentences such as the ones in (1) above. Holmes found that when the sentence was unambiguous, grammaticality judgment times were uniform regardless of verb bias. But for syntactically ambiguous sentences, verb bias influenced judgment times. When the verb was biased towards the nonminimal attachment reading, judgment times on the disambiguating word were less elevated than when the verb was biased towards the minimal attachment reading. The garden-path effects with the verbs biased toward a minimal attachment interpretation were large; the garden-path effects with the verbs biased towards a nonminimal attachment interpretation were smaller, although still significant. As can be deduced from the discussion earlier, this pattern of results is perfectly consistent with the garden-path model. As discussed above, any significant elevation in reading times on the disambiguating word relative to reading times on that word in the unambiguous control condition indicates that subjects initially misanalyzed the preceding noun phrase, and had to reanalyze it. The finding that reading times were less elevated indicates not that the parser avoided being garden-pathed altogether, but rather that the parser could reanalyze the ambiguous noun phrase more easily when the sentence contained a biased verb. Note also that the task used in this experiment was less than ideal. The display used was a cumulative display, which has the problems outlined

Verb Information During Syntactic Parsing

313

above. In addition, a word-by-word grammaticality judgment task is bound to induce subjects to engage in abnormal reading strategies, or at least to spend inordinate amounds of time on each word of the sentence. Consistent with this argument, the time spent on each word of the sentences was unusually high. Reading times for the critical sentences always averaged more than 600 msec per word, which is almost three times the duration of an average eye fixation (Rayner & Pollatsek, 1989). In another set of studies, Holmes et al. (1990) replicated the study from Holmes (1987), but eliminated the part of the task that required subjects to make a grammaticality judgment on each word. Subjects simply read through the sentence word-by-word at their own pace, with the words remaining visible (cumulative display). Holmes et al. found no evidence of garden-pathing with nonminimal attachment verbs. However, there was no effect of ambiguity in this experiment either; sentences with the complementizer that were no easier than those without the complementizer.This result is contrary to the findings of several other studies (Ferreira & Henderson, 1990b; Rayner & Frazier, 1987). and again indicates that the cumulative display paradigm is not appropriate for examining on-line parsing processes. Holmes et al. conducted an additional experiment using the non-cumulative version of the self-paced reading task and found evidence for garden-pathing even with nonminimal attachment-biased verbs, although the effects occurred only when the ambiguous noun phrase was long. Ferreira & Henderson (1990b) examined the same issues, but in addition, explicitly compared the effects obtained with three different reading tasks: eye movement recording, self-paced word-by-word reading with a non-cumulative display, and self-paced word-by-word reading with a cumulative display. The sentences that were contrasted were similar to the ones shown in (1). In the experiment using eye movement monitoring, Ferreira and Henderson found that verb bias did not prevent garden-pathing with the ambiguous, nonminimal attachment sentences. Further, the presence of helpful verb information did not influence total reading times for the sentences. The only effect of verb bias was on number of regressive (backward) eye movements: Fewer regressions were made when the sentences contained a verb biased towards the nonminimal attachment interpretation. The experiment using self-paced reading with a noncumulative display showed a similar pattern. Longer reading times were observed on the disambiguating word of the sentences in the syntactically ambiguous conditions, indicating that subjects were garden-pathed. Verb bias had no effect on the extent of garden-pathing observed on the disambiguating word. On the word following the disambiguating word, garden-pathing was still evident, but here verb bias had an effect: The garden-path effect was reduced with the appropriately biased verb. Finally, the cumulative display reading task was not sensitive to garden-path effects, and only slightly sensitive to the effects of verb bias.

3 14

F. Ferreira and J.M.Henderson

From these results, Ferreira and Henderson argued that verb bias does not override the minimal attachment principle. Instead, subjects are garden-pathed by any nonminimal attachment sentence. Verb information becomes available at a later stage of parsing, the stage at which the initially produced syntactic structure is reanalyzed following detection of the parsing error. Further, this study may indicate that the more difficult reanalysis is, the more useful verb information will be. (The experiment we will report in detail below further supports this point.) Ferreira and Henderson argued that the non-cumulative self-paced reading task makes reanalysis more difficult than it is during natural reading (or when eye movements are monitored) because in the former case subjects cannot look back to earlier material. Given that in the non-cumulative task subjects must reanalyze the sentence from memory, they are more likely to use any and all sources of information at their disposal, including the information provided by the verb. Finally, we want to consider briefly some work that does not directly bear on the issue of whether verb subcategorization information can modify gardenpathing, but does address the question of whether phrase structure parsing and the activation of verb information are modular with respect to each other. In one study, Shapiro, Zurif, and Grimshaw (1987) found that the more argument structures associated with a verb, the harder that verb was to process (but see Schmauder, 1989). (Argument structure and subcategorization information were separated in their study, and only the former affected ease of processing. Our arguments are neutral with respect to this distinction. See Shapiro et al. for details.) In a later study, Shapiro, Zurif, and Grimshaw (1989) demonstrated that all possible argument structures of a verb are activated when a verb is encountered, even when the preceding syntactic context makes it clear that only one argument structure is syntactically appropriate. This result indicates that the phrase structure parse that has been constructed up to the point of encountering a verb does not affect the way the verb is processed. In our research, we find the complementary effect: Information stored with particular verbs cannot influence the initial parsing of a sentence. The common thread in both of these studies is that, at least during initial sentence processing, verb information and phrase structure information are encapsulated from each other, and thus may constitute separate processing modules.' In summary, the weight of the evidence supports the claim that verb information does not influence initial parsing decisions. Verb information is not used initially to prevent a garden-path. These results are consistent with the garden-path model, and inconsistent with the verb-guidance model. As mentioned earlier in this chapter, the verb information that is at issue here is a kind of syntactic information. Verb subcategorization preferences are stated over a syntactic vocabulary. Nevertheless, this syntactic information is not used in the initial stages of constructing a constituent structure for a sentence. It appears, then, that the syntactic processor does not treat different types of syntactic

Verb Information During Syntactic Parsing

3 15

information equivalently. Instead, the language system is modularized. Not only is the syntactic module encapsulated from nonsyntactic information, but the syntactic module itself is modularized, so that information about subcategorization preferences is not consulted in the initial stages of constructing a phrase structure tree.

SYNTACTIC REANALYSIS: OVERCOMING THE GARDEN-PATH We have argued thus far that verb information is not used during initial parsing, but is used in later stages to aid in the reanalysis of a syntactically misanalyzed string. We will address the following question in this section: How are the processes that initially assign structure coordinated with the processes that later revise that structure, and that seem to have access to non-phrasestructural sources of information? In other words, how is it that this information can influence the later operation of the parser, given that the information cannot influence its initial operation? Mitchell (1987) has proposed a model of parsing that deals with both initial analysis and reanalysis. He argues that the parser consists of two processors, one that constructs phrase structure trees in accordance with its parsing principles, and a monitor or filter that immediately checks the output of the first processor to insure that it has created a lexically and semantically coherent string. This model predicts that in a sentence such as “The girl realized the man cheated on the exam,” the phrase the man will initially be analyzed as a direct object, as predicted by the minimal attachment principle, due to the operation of the first processor. The filter would then immediately kick in and reject this sequence, because a direct object after realized violates the verb’s subcategorization preferences, and so the sequence realized the man is anomalous on the direct object analysis. Therefore, reanalysis will be initiated on the ambiguous phrase the man. In contrast, with a more neutral verb such as knew, reanalysis cannot begin until the disambiguating word is encountered, because the sequence knew the man (where the man is a direct object) is not anomalous. This view predicts, then, that with a biased verb, reading times will increase on the ambiguous phrase; with a less biased verb, reading times will not increase until the disambiguating phrase is encountered. In other words, the effect of verb bias is to cause reanalysis processes to begin earlier than they would with a more neutral verb. Although this model seems plausible, the data available to this point do not support it. For example, consider again the studies by Ferreira and Henderson (1990b). They found that reading times on the ambiguous noun phrase and disambiguating word were equivalent regardless of verb bias. Verb bias only influenced the amount of garden-pathing on the word following the disambiguating word, and only when subjects read sentences in the noncumulative selfpaced reading task. It appears, then, that verb information does not cause re-

316

F. Ferreira and J.M.Henderson

analysis processes to begin early, on the ambiguous noun phrase. Even if the anomalous nature of the sequence realized the man is detected by the language processing system, this information does not initiate reanalysis. Instead, the parser begins reanalysis only after receiving information indicating that its initial analysis is syntactically ill-formed, as when a word requiring the more complex analysis is encountered. Thus, a more plausible model would be one where reanalysis processes are initiated most reliably by a syntactic error signal. That is, the parser does not assume that a new syntactic structure for a string is required until its initial analysis fails to yield a well-formed structure. After detecting the error, the parser would begin the work of restructuring the ambiguous, misanalyzed material, a task that is made easier when helpful verb information is available. One model of reanalysis assumes the existence of two processors: a phrase-structure processor or parser, and a thematic processor (Rayner et al., 1983). The first processor assigns phrase structure to a suing, using only phrase structure rules. It is responsible both for creating the initial syntactic structure, and for revising that structure if necessary. The second processor, the thematic processor, uses the lexical information stored with verbs to assign thematic roles to syntactic constituents. Every verb in the lexicon contains information about its subcategorization properties, that is, information about the kinds of syntactic phrases with which it can co-occur. These phrases are typically arguments of the verb. For example, the verb put must occur with both a direct object and a prepositional phrase (PP), while the verb kick takes only a direct object. Of course, kick could also occur with a PP, as in kick the bull into the air. However, here, the PP functions not as an argument, but as a modifier or adjunct phrase. Every argument of a verb must be assigned a thematic role (Chomsky, 1981, 1986). For example, the verb put assigns the role of theme to its direct object, and location to its PP. In addition to assigning thematic roles to these “internal arguments” (Williams, 1980), a thematic role is also assigned to the subject of the sentence (the external argument, so-called because the argument occurs externally to the verb phrase). In the case of put, the role of agent is assigned to the subject. The arguments to which thematic roles are assigned are the syntactic phrases created by the parser. In situations where the parser has run into difficulty parsing a sentence, the thematic processor must revise its thematic role assignments. Typically, the thematic processor will require that the parser construct a new structure for the phrase that has been assigned an incorrect thematic role, because the existing structure cannot support an alternative thematic role assignment. Williams (1980) proposes a set of “realization rules,” which state the correspondence between thematic roles and syntactic phrases. For example, the role of theme must be fulfilled by a noun phrase; the role of instrument must be fulfilled by a PP; and the role of proposition must be fulfilled by a clause. In this model of thematic role assignment, thematic roles are assigned to

Verb Information During Syntactic Parsing

317

phrases as soon as possible, so that the parser and the thematic processor operate in parallel. More precisely, we will assume that as smn as the thematic processor receives the head of a phrase, the thematic processor assumes that it has an argument, and assigns a thematic role to that argument, whether the phrase is complete or not. This view predicts that the thematic processor would not attempt to assign a role to the word the, nor would it attempt to assign a role to a sequence such as the ugly, because the head noun has not yet been encountered. The thematic processor would assign a role to the ugly boy, however, because it has encountered the head noun. Notice that the role would be assigned, even though the thematic processor has no way of knowing whether the phrase is complete; for example, the entire phrase could be the ugly boy with the large wart. Later, we will review the results of experiments that provide evidence for these claims. Let us consider once again the sentence “Mary knew the answer was correct.” The parser would begin by analyzing Mary as a noun phrase, and as the subject of the sentence. The thematic processor would immediately take this phrase as input, and assign Mary the most likely thematic role for a subject, namely agent.2 The word knew would be analyzed by the parser as the main verb of the sentence. This information would be output to the thematic processor, which would also receive (from the lexicon) and consider all possible thematic structures for that verb (Shapiro et al., 1987, 1989; Stowe, 1989; Tanenhaus & Carlson, 1989). A thematic structure is a list of the thematic roles that the verb assigns to each of its arguments. The verb know has two thematic structures: One structure states that the first argument is an agent, and the second argument is a theme; the second structure states that the first argument is an agent, and the sccond argument is a proposition. The thematic processor itself has no way to determine which of these two should be instantiated. Therefore, the two structures remain available, awaiting further information. Next, the noun phrase the answer is analyzed by the parser as a direct object, according to the minimal attachment principle. The parser outputs this phrase to the thematic processor. At this point, the thematic processor selects the agent-theme structure, because a direct object cannot be a proposition (Williams, 1980). At the same time, the other thematic structure begins to decay, or perhaps is actively suppressed (evidence of suppression in the domain of lexical processing has been obtained by Gernsbacher, 1989; and Simpson, 1989). The first processor, the parser, then encounters the embedded verb was, and its initial syntactic analysis breaks down. An error signal is sent to the thematic processor, indicating that its assignment of thematic roles is also likely to be incorrect. At this point, reanalysis begins. If the thematic processor still has access to the unchosen, decaying thematic structure, then this information can be used to facilitate reanalysis. This alternative thematic structure states that the second argument of the verb know is a proposition. The thematic processor proposes to the parser that it construct an alternative analysis, one which

318

F. Ferreira and J.M. Henderson

could support a propositional argument. The syntactic structure that can best support a proposition is a clause, so the direct object analysis of the sequence the answer is changed to one where the answer is the subject of a new clause. This model of reanalysis makes a number of predictions. The first prediction is that reanalysis processes will not be initiated until a syntactic error signal is received. The results obtained by Ferreira and Henderson (1990b) are consistent with this prediction, as are results obtained by Ferreira and Clifton (1986, Experiment 1). As discussed above, no increase in reading time is found on a region when it is ambiguous compared with when the region is made unambiguous by the presence of a complementizer. The increased reading time never appears before the disambiguating word, because that word marks the point at which the initial structure has broken down? The second prediction is that verb information will be used in parsing only to revise an incorrect syntactic analysis. This prediction follows from the division of labor between the two processors. The parser, the first processor, has primacy over the second, because thematic roles cannot be assigned until a phrasal head is encountered. The parser constructs phrases in accordance with the minimal attachment principle, without regard for the particular information stored with verbs. Verb information will not be used to prevent an initial misanalysis. Again, the results reviewed above support this prediction. The third prediction is that the longer the thematic processor has been committed to one thematic structure, the harder it will be to reanalyze the sentence. Reanalysis will be more difficult because once the thematic processor has committed itself to one thematic structure, the other structure(s) begins to decay (or is suppressed). When the error signal is received from the parser, the thematic processor must recover the decaying thematic structure. Clearly, the more that structure has decayed, the less likely it is that the thematic processor will be able to recover it. If the structure cannot be recovered, then reanalysis will be impossible. Earlier, we stated that thematic roles are assigned as soon as the thematic processor receives a head noun. A specific prediction, then, is that the greater the distance between the head noun of the ambiguous phrase and the syntactic error signal, the harder reanalysis will be, because the greater this distance is, the less available the alternative thematic structure will be. We examined this prediction in a recent study (Ferreira & Henderson, 1990a). We presented subjects with sentences that were ambiguous between an early closure and a late closure interpretation. Frazier (1978; Frazier & Rayner, 1982) has argued that late closure is another parsing principle the parser obeys, in addition to minimal attachment. The late closure principle states that phrases remain open as long as possible. The parser prefers to incorporate new material into the current phrase rather than opening a new one (see Frazier, 1978, and Frazier and Rayner, 1982, for a description of the relation between minimal attachment and late closure). Given a sequence such as “After the Martians invaded the town ...,” the parser initially attempts to interpret the town as a direct

Verb Information During Syntacric Parsing

319

object of the verb, because of the operation of the late closure principle. If, however, the sentence continues was evacuated, then this initial analysis must be revised: The phrase the town must be analyzed as the subject of a new clause. We varied the length of the ambiguous region so that it was either short (the town), or long (the town that bordered the ciry). The four versions of the sentences are given in (2). (2) a. After the Martians invaded the town was evacuated. (Early closure, short ambiguous region) b. After the Martians invaded the town that bordered the city was evacuated. (Early closure, long ambiguous region) c. After the Martians invaded the town the residents were evacuated. (Late closure, short ambiguous region) d. After the Martians invaded the town that bordered the city the residents were evacuated. (Late closure. long ambiguous region) Sentences were presented one word at a time at a constant rate using the Rapid Serial Visual Presentation (RSVP) technique, and subjects were asked to judge at the end of the sentence whether it was grammatical (see Warner & Glass, 1987). This task does not reflect on-line parsing processes, but does measure whether subjects are able eventually to parse these sentences correctly. We were interested in discovering the circumstances under which reanalysis is relatively difficult, and were not attempting to distinguish initial analysis from reanalysis. Consistent with previous research (Frazier & Rayner, 1982; Mitchell, 1987). we found in three experiments that the early closure sentences were judged grammatical less often than the late closure sentences, regardless of region length. More importantly, the longer the ambiguous region, the less often early closure sentences were judged grammatical. Length had a much smaller effect on late closure sentences. In other words, the longer the ambiguous region, the harder reanalysis was. How do these results address the model of reanalysis given above? These results are consistent with the prediction that the longer the thematic processor has been committed to an incorrect thematic role assignment, the harder reanalysis will be. For the early closure sentence with a short region (2a). the parser structures the sequence the town into a noun phrase, and the thematic processor then assigns this phrase the thematic role of theme. On the very next word an error signal is sent to the thematic processor, and so little time has

320

F. Ferreira and J . M . Henderson

elapsed between the assignment of a thematic role and the onset of syntactic reanalysis. Now consider sentence (2b). an early closure sentence with a long ambiguous region. The parser structures the sequence the town, and the thematic processor immediately assigns it the role of theme, as with the short region. The parser then continues to build the longer noun phrase the town rhat bordered the city, while the thematic role assignment remains the same. Thus, the thematic processor has been committed to the incorrect thematic role for five words before the receipt of the error signal caused by the word was. As a result, reanalysis of this region is much harder, because the longer the thematic processor has been committed to one thematic structure, the less likely it is that the alternative, correct thematic structure will be available to guide reanalysis. The view that we have outlined thus far makes a further prediction: If a sentence contained a long ambiguous phrase, but somehow thematic role assignment were delayed until the last word of the phrase, reanalysis would be easy. We tested this prediction by comparing sentences such as those shown in (3) (Ferreira & Henderson, 1990a).

(3) a. While the boy scratched the dog yawned loudly. (Short ambiguous region) b. While the boy scratched the dog that Sally hates yawned loudly. (Long ambiguous region, head early in phrase) c. While the boy scratched the big and hairy dog yawned loudly. (Long ambiguous region, head late in phrase) (All these sentences are early closure, so all should induce garden-pathing. Late closure sentences were also included in the study, but are omitted here for the sake of brevity.) The (b) version was judged grammatical less often than the (a) version, as expected. However, the (c) version was judged grammatical as often as the (a) version, even though it contained a long ambiguous region. The reason for the difference in results between (b) and (c) is due to the location of the head noun. In (b), the head noun occurred early in the phrase, and so thematic role assignment could begin immediately. But in (c), thematic role assignment could not begin until a head noun was received. Thus, the assignment was delayed until the final word of the phrase. The thematic role assignment was immediately followed by a syntactic error signal, both in (c) and in (a). As a result, the alternative thematic structure had decayed little in these conditions, and so reanalysis was comparatively easy.

EXPERIMENT: VERB BIAS AND SYNTACTIC REANALYSIS In our earlier discussion of the garden-path and the verb-guidance models,

Verb Information During Syntactic Parsing

321

we argued that verb bias did not affect initial analysis, but only had its effect during reanalysis. The reanalysis model we have outlined, then, must somehow be capable of accounting for the verb bias effect on reanalysis. Our assumption is that ease of reanalysis is determined by the availability of the alternative thematic structure, and availability is determined by the activation level of the alternative structure when it is needed. In turn, that activation level will be a function of the initial activation levels of the two structures and the length of time since the alternative structure has begun to decay. One obvious consequence of this argument is that reanalysis towards an early closure structure will be easier following a highly biased early closure verb (i.e., a highly intransitive verb) compared to a more equally biased verb, because in the former case, the alternative thematic structure will be more available than in the latter case, at the point where the signal that reanalysis is required has been received. This prediction has been confirmed in earlier studies. The additional prediction we make is that verb bias should interact with the length of the ambiguous region, because the longer one thematic role assignment has been instantiated, the more the alternative structure will have decayed. In the case of neutrally biased verbs, the alternative structure will sometimes be available following a short ambiguous region, but may be essentially unavailable following a long region. For verbs biased toward early closure, the alternative thematic structure should be more likely to be available regardless of region length because of its higher initial activation level. To test this prediction, we conducted the following experiment: We compared sentences in which the critical verb took two thematic structures, but in the condition we call the biased verb condition, one structure was highly preferred over the other due to the degree of transitivity of the first verb; in the neutral verb condition, the two structures were approximately equally preferred (equally transitive). Examples of such sentences are shown below in (4) and (5). (4) Biased verb conditions:

a. While the girls jog the boys like to watch. (Early closure, short region, highly intransitive verb) b. While the girls jog the kids that the boys beat like to watch. (Early closure, long region, highly intransitive verb) c. While the girls defeat the boys the coaches like to watch. (Late closure, short region, highly transitive verb) d. While the girls defeat the kids that the boys beat the coaches like to watch. (Late closure, long region, highly transitive verb)

F. Ferreira and J.M.Hendersorr

322 ( 5 ) Neutral verb conditions:

a. While the girls race the boys like to watch. (Early closure, short region, highly intransitive verb) b. While the girls race the kids that the boys beat like to watch. (Early closure, long region, highly intransitive verb) c. While the girls race the boys the coaches like to watch. (Late closure, short region, highly transitive verb) d. While the girls race the kids that the boys beat the coaches like to watch. (Late closure, long region, highly transitive verb) Within each verb set, sentences were varied so that they were either early or late closure (early closure sentences were expected to induce a garden-path), and so that the ambiguous region was either short or long. The length was created so that the head noun of the ambiguous noun phrase occurred early in the phrase. As a result, the thematic role assignment would also occur early, and thus we would expect that reanalysis would be difficult.

Method Subjects

The subjects in this experiment were 36 undergraduates from the University of Alberta. All were native speakers of Canadian English, and were naive with respect to the purposes of the experiment.

Materials and Design The experiment was conducted so that the verb bias factor was both between subjects and between items. The closure and length factors were within subjects and within items. Thirty-six sentence quadruples such as the one shown in (4) were constructed; this set of materials constituted the verb bias set. These sentences were intermixed with 72 filler sentences. Half of the filler sentences were grammatical, and the other half were ungrammatical. The ungrammatical items were sentences such as “The player hit the ball the man who saw it. Typically, they were missing obligatory constituents or contained too many constituents. These 108 items were presented to 18 subjects. Another set of thirty-six sentence quadruples such as the one shown in (5) were constructed; this set of materials constituted the neutral verb set. These 36 items were constructed so as to be as similar as possible to the verb bias set,

Verb Information During Syntactic Parsing

323

with the constraint that they be semantically plausible given the choice of verb. These sentences were intermixed with the same filler sentences as were used with the verb-bias set. The entire set of items (108 in total) was presented to 18 subjects, none of whom had seen the verb bias materials. For both the verb bias and neutral verb sets, presentation of the items was randomized individually for each subject, and was preceded by a short practice session.

Procedure Subjects were told that their task was to judge whether a sentence was grammatical. Care was taken to insure that subjects understood the intended meaning of the term “ungrammatical.” For example, in the instructions, subjects were shown a sentence with a dangling preposition and told that such a sentence was to be considered grammatical. Sentences appeared on a computer screen. A button panel was situated in front of the subject, with one button labelled “G”for grammatical, and the other labelled “U” for ungrammatical. Sentences were presented using the RSVP technique. Each word appeared in the center of the screen for 250 msec, with the current word replacing the previous word. At the end of the sentence, the subject was prompted for his or her judgment. The subject pushed the appropriate button to indicate whether the sentence was grammatical or ungrammatical. A Zenith 80286 computer controlled the experiment and recorded both the subject’s judgments and judgment times.

Results

No significant effects were obtained on judgment times, and so they will not be discussed further. The results we will present are the percentage of times sentences were judged grammatical in a given condition. Results were analyzed as a 2 x 2 x 2 factorial, both over subjects (F1)and items ( F 2 ) . The first variable refers to whether the sentence contained a biased or a neutral verb; the second variable refers to whether the sentence was early closure (garden-pathing) or late closure; and the third variable refers to whether the sentence contained a short or a long ambiguous region. For the subjects analysis, the first variable was between-subjects, and the other two variables were within-subjects. For the items analyses, the first variable was betweenitems, and the other two variables were within-items. The results of the experiment are shown in Figure 1. First, there was a three way interaction, significant by both subjects and items, Fl(1.34) = 22.8, MSe = 2.75, p < .001; F2(1, 70) = 21.8, MSe = 4.20, p c .001. In order to explore this interaction, separate analyses were conducted on the early and late closure sentences independently, treating each analysis as a 2 x 2 factorial with verb bias and ambiguous region length as factors. For the late closure sen-

F. Ferreira and J.M. Henderson

324

tences, there was a main effect of region length, F1(1,34) = 33.0, MSe = 3.67, p < .001; F2(1,70) = 33.6, MSe = 5 . 5 7 , ~< .001, no effect of verb bias, both F's < 1, and no significant interaction between the two factors, both p's > .lo. In contrast, for the early closure Sentences, there was a main effect of verb bias, F1(1,34) = 69.4, MSe = 4.60, p c .001; F2(1,70) = 67.2, MSe = 7.63, p c .001, a main effect of length, F1(1,34) = 60.8, MSe = 3.19, p c .001; F2(1,70) = 86.3, MSe = 4.70, p < .001, and a significant interaction between the two, F1(1,34) = 20.3, MSe = 3.19, p < .001; F2(1,70) = 25.6, MSe = 4.70, p < .001. Examining this interaction more closely, it is clear that the verb bias aided reanalysis much more when the ambiguous region was long, compared with when it was short. When the ambiguous region was long, sentences with neutral verbs were accepted only 18% of the time, and with biased verbs, these same sentences were accepted 79% of the time. In contrast, when the ambiguous region was short, sentences with neutral verbs were accepted 69% of the time, and with biased verbs, 93% of the time.

-0

.-0

100

I

verb bias 0neutral BDl biased

c,

E"

80 --

E IJ,

0 -

60 --

U Q)

CT -0

40 --

short

long

Early Closure

short

long

Late Closure

Figure 1 . ?he percentage of sentences judged grammatical as a function of verb bias, syntactic closure, and length of the syntactically ambiguous region. (See text for details.)

Discussion Because we expected from previous experiments that garden-pathing would not occur with late closure sentences, the lack of an interaction between length and verb bias for our late closure sentences is not surprising. The small

Verb Information During Syntactic Parsing

325

effect of length (compared to the effect with early closure sentences) can be accomodated by the assumption that memory load will have some effect on overall parsing efficiency, and the greater the number of words, the greater the memory load. The crucial point to note is that the effect of length was larger for early closure than for late closure sentences (see also Ferreira & Henderson, 1990a). For the early closure sentences, we found evidence of garden-pathing. Further, when the ambiguous region was short, recovery from the garden path was relatively easy, and most of the time it was successful regardless of the bias of the verb, though clearly verb bias had some facilitatory effect. On the other hand, when the ambiguous region was long, reanalysis given a neutral verb was rarely successful. But given a biasing verb, the success rate increased dramatically. In summary, then, we found that, as expected according to the gardenpath model, the late closure sentences did not garden-path the parser, and so verb bias had no effect. This result makes sense from the perspective of the Frazier and Rayner garden-path model and our own reanalysis model, because unless the parser is in a state of having been garden-pathed, there is nothing for the verb to assist with. Early closure sentences, on the other hand, did elicit an error signal from the parser. Reanalysis is consequently necessary, and was found in our experiment to be easier under some circumstances than others. When the ambiguous region was long, reanalysis was difficult, a result we expected based on our previous work (Ferreira & Henderson, 1990a). Further, if the main verb of the sentence was biased towards an early closure analysis (i,e., the verb was highly intransitive), reanalysis was far easier than if the main verb was neutral. Finally, verb bias had a much greater effect on reanalysis when the ambiguous region was long (and so reanalysis was difficult) compared to when the region was short. In summary, the main purpose of this experiment was to investigate the way in which the frequency of alternative thematic structures would affect the ease with which sentences could be reanalyzed in the context of our reanalysis model. In particular, we predicted that a long ambiguous region would make reanalysis far more difficult if the critical verb was neutral rather than biased towards the ultimately correct syntactic structure. This prediction was confirmed.

CONCLUSIONS In this chapter, we explored the effects of verb information on parsing. The first section outlined the garden-path model proposed by Frazier (1978, 1987, 1989) and Frazier and Rayner (1982), in which initial syntactic analysis is distinguished from reanalysis. In the second section, we argued that verb bias information does not influence initial parsing, but does make reanalysis easier.

326

F. Ferreira and J . M . Henderson

In the third section, we described a model of reanalysis. According to this model, the parser creates syntactic phrases. The phrases are then sent to a thematic processor which assigns them (more precisely, assigns arguments) thematic roles (Rayner et al., 1983). If more than one thematic role structure is possible, the one that was not initially chosen (due to the parse that was created) begins to decay. Ease of reanalysis (which depends on being able to retrieve the initially unchosen structure) differs depending on the the availability of the initially unchosen structure. In the fourth section, we expanded this model to account for the verb bias effect. We proposed that frequency of use affected the initial activation level of the two possible thematic structures of a verb, and that the higher the initial activation level of a structure, the longer it would take for that structure to become unavailable. We correctly predicted that verb bias information would be far more helpful when the ambiguous region was long compared with when the region was short. In sum, we would argue that verb information is not used by the language processing system to guide the initial syntactic parse of a sentence. Instead, information about a verb's thematic role structure(s) is used to aid syntactic reanalysis following a misparse. Most researchers have noted that thematic roles play a crucial role in sentence comprehension. However, models of comprehension have not specified how these roles are assigned to syntactic constituents (e.g., Kintsch & Van Dijk, 1978, point out that their model only assumes that thematic roles are assigned to arguments, but leaves unspecified the assignment process). Our model states that thematic roles are assigned based on the initial parse and the consequent choice of thematic structure. The thematic processor assigns these roles as soon as the head of a phrase which could be an argument of a verb is encountered by the parser. If a syntactic revision is required, thematic roles must be revised as well. The eventual grammatical interpretation of a sentence includes both a legitimate syntactic structure, and a complete and accurate assignment of thematic roles to the constituents of a sentence. Acknowledgments The research reported here was supported by grant OGP-37323 to the first author, and OGP-41792 to the second, from the Natural Sciences and Engineering Research Council of Canada. We would like to thank Charles Clifton for his comments on an earlier version of this chapter, and Steven Hayduk for his assistance conducting the experiment. Notes

'Schmauder (1989) has conducted a number of experiments designed to replicate Shapiro et al. (1987, 1989) with a variety of different measures, and

Verb Information During Syntactic Parsing

327

finds no effects of the verb complexity variable. At this point, it is unclear why Schmauder’s results differ from Shapiro et al.’s, and research designsd to resolve the discrepancies is in progress. 2Another possibility is that the assignment of agent is delayed until a verb which assigns that role is encountered. There are no data with which to distinguish between these alternatives, and at this point the distinction is irrelevant to the points we wish to make. 3 0 ~ claim r that reanalysis does not begin until syntactic disambiguation applies only to sentences that are locally, or temporarily, ambiguous. For sentences that are fully ambiguous, such as “Mary saw the cop with the glasses,” it is not clear how the semantics of the sentence triggers reanalysis. Even worse, because the sentences are globally ambiguous, it is difficult to know when or under what conditions readers change their initial interpretation. For these reasons, we will not consider globally ambiguous sentences in our reanalysis model. See Rayner, Carlson, & Frazier (1983) for further discussion of these issues, and some suggestions about how these sentences might be reanalyzed.

References Altmann, G. & Steedman, M. (1988). Interaction with context during human sentence processing. Cognition, 30, 191-238. Chomsky, N. (1981). Lectures on government and binding. Dordrecht: Foris. Chomsky, N. (1986). Barriers. Cambridge: MIT Press. Clifton, C.E., & Ferreira, F. (In press). Ambiguity in context. Language and Cognitive Processes. Clifton, C.E., & Frazier, L. (1986). The use of syntactic information in filling gaps. Journal of Psycholinguistic Research, 1.5, 209-224. Clifton, C.E., & Frazier, L. (1989). Comprehending sentences with long-distance dependencies. In G.N. Carlson and M.K. Tanenhaus @ids.), Linguistic structure in language processing. Dordrecht: Kluwer. Crain, S., & Steedman, M. (1985). On not being led up the garden path: The use of context by the psychological parser. In D. Dowty, L. Kartunnen, and A. Zwicky (Eds.), Natural language parsing. Cambridge: Cambridge University Press. Ferreira, F., & Clifton, C.E. (1986). The independence of syntactic processing. Journal of Memory and Language, 2.5, 348-368. Ferreira, F., & Henderson, J.M. (1990a). Effects of sentence length and complexity on syntactic reanalysis. Manuscript submitted for publication. Ferreira, F., & Henderson, J.M. (1990b). The use of verb information in syntactic parsing: A comparison of evidence from eye movements and word-byword self-paced reading. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 555-568.

328

F. Ferreira and J.M. Henderson

Fodor, J.A. (1983). The modularity of mind. Cambridge: MIT Press. Fodor, J.D. (1988). On modularity in syntactic processing. Journal of Psycholinguistic Research, 17, 125-168. Forster, K.I. (1979). Levels of processing and the structure of the language processor. In W.E. Cooper & E.C.T. Walker (Eds.), Sentence processing. Hillsdale, N.J.: Erlbaum. Frazier, L. (1 978). On comprehending sentences: Syntactic parsing strategies. Unpublished doctoral dissertation, University of Connecticut. Frazier, L. (1985). Modularity and the representational hypothesis. Proceedings of NELS 15. Amherst, MA: GLSA. Frazier, L. (1987). Sentence processing. In M. Colthean (Ed.), Attention and Performance X I I . Hillsdale, N.J.: Erlbaum. Frazier, L. (1989). Against lexical generation of syntax. In W. Marslen-Wilson (Ed.), Lexical representation and process. Cambridge: MIT Press. Frazier, L., & Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in h e analysis of structurally ambiguous sentences. Cognitive Psychology, 14, 178-210. Frazier, L., & Rayner, K. (1987). Resolution of syntactic category ambiguities: Eye movements in parsing lexically ambiguous sentences. Journal of Memory and Language, 26,505-526. Freedman, & S.E., Forstcr, K.I. (1985). The psychological status of overgenerated sentences. Cognition, 19, 101-132. Gernsbacher, M.A. (1 989). Mechanisms that improve referential access. Cognition, 32, 99-155. Henderson, J.M., & Ferreira, F. (In press). The effects of foveal processing difficulty on the perceptual span in reading: Implications for attention and eye movement control. Journal of Experimental Psychology: Learning, Memory, and Cognition. Henderson, J.M., Pollatsek, A., & Rayner, K. (1989). Covert visual attention and exuafoveal information use during object identification. Perception and Psychophysics, 45, 196-208. Holmes, V.M. (1987). Syntactic parsing: In search of the garden path. In M. Colthcart (Ed.), Attention and Performance X I I . Hillsdale, N.J.: Erlbaum. Holmes, V.M., Stowe, L., & Cupples, L. (In press). Lexical expcctations in parsing complement-verb sentences. Jcurnal of Memory and Language. Just, M.A., & Carpcntcr, P.A. (1987). The psychology of reading and language comprehension. Newton, MA: Allyn and Bacon. Just, M.A., Carpcntcr, P.A., & Woolley, J. (1982). Paradigms and processes in reading Comprehension. Journal of Experimental Psychology: General, 10, 833-849. Kintsch, W., & Van Dijk, T.A. (1978). Toward a model of text comprehension and production. Psychological Review, 85, 363-394. Kurtzman, H. (1984). Studies in syntactic ambiguity resolution. Unpublished

Verb Information During Syntactic Parsing

329

doctoral dissertation, M.I.T. Marslen-Wilson, W. (1975). Sentence perception as an interactive parallel process. Science, 189, 226-228. McClelland, J.L. (1 987). The case for interactionism in language processing. In M. Coltheart (Ed.), Attention and Performance XII. Hillsdale, N.J.: Erlbaum. McConkie, G.W. (1979). On the role and control of eye movements in reading. In P.A. Kolers, M.E. Wrolstad, and H. Bouma (Eds.), Processing of visible language (Vol.1). New York: Plenum Press. Mitchell, D.C. (1987). Lexical guidance in human parsing: Locus and processing. In M. Coltheart (Ed.), Attention and Performance XII. Hillsdale, N.J.: Erlbaum. Morrison, R.E. (1984). Manipulation of stimulus onset delay in reading: Evidence for parallel programming of saccades. Journaf of Experimental Psychology: Human Perception and Performance, 10,667-682. Rayner, K., Carlson, M., & Frazier, L. (1983). The interaction of syntax and semantics during sentence processing: Eye movements in the analysis of semantically ambiguous sentences. Journal of Verbal Learning and Verbal Behavior, 22, 358-374. Rayner, K., & Frazier, L. (1987). Parsing temporarily ambiguous complements. Quarterly Journal of Experimental Psychology, 39A, 657-673. Rayner, K. & Pollatsek, A. (1989). The psychology of reading. Englewood Cliffs, N.J.: Prentice Hall. Schmauder, A.R. (1989). Argument structure frames: A lexical complexity metric? Unpublished Master’s thesis, University of Massachusetts. Shapiro, L.P., Zurif, E., & Grimshaw, J. (1987). Sentence processing and the mental representation of verbs. Cognition, 27, 219-246. Shapiro, L.P., Zurif, E., & Grimshaw, J. (1989). Verb processing during sentence comprehension: Contextual impenetrability. Journal of Psycholinguistic Research, 18, 223-243. Simpson, G.B. (1989). Lexical access and meaning suppression. Paper presented at the Sylvia Beach Language Comprehension Conference, Newport, Oregon. Slowiaczek, M.L. (1981). Prosodic units as language processing units. Unpublished dissertation, University of Massachusetts. Speer, S.R., & Slowiaczek, M.L. (1990) The influence of prosodic structure on comprehending sentences with temporary syntactic ambiguities. Manuscript submitled for publication. Steedman, M. & Altmann, G. (In press). Ambiguity in context: A reply. Language and Cognitive Processes. Stowe, L. (1989). Thematic structurcs and sentence comprehension. In G.N. Carlson and M.K. Tanenhaus (Eds.), Linguistic structure in language processing. Dordrecht: Kluwer.

330

F. Ferreira and J . M . Henderson

Tanenhaus, M.K., & Carlson, G.N. (1989). Lexical structure and language comprehension. In W. Marslen-Wilson (Ed), Lexical representation and process. Cambridge: MIT Press. Taraban, R., & McClelland. J.L. (1988). Constituent attachment and thematic role assignment in sentence processing: Influences of content-based expectations. Journal of Memory and Language, 27,597-632. Warner, J., & Glass, A.L. (1987). Context and distance-to-disambiguation effects in ambiguity resolution: Evidence from grammaticality judgments of garden path sentences. Journal of Memory and Language, 26,714-738. Williams, E. (1980). Predication. Linguistic Inquiry, 11,203-238.

Understanding Word and Sentence G.B. Simpson (Editor) Q Elsevier Science Publishers B.V. (Noh-HoUmd), 1991

Chapter 13 The Role of Lexical Representations in Sentence Processing Julie E. Boland Michael K . Tanenhaus University of Rochester Rochester, New York U.S.A.

Most research on lexical processing has focused on how words are recognized and the role of sentential context on the recognition process. This chapter focuses on an equally important, though less thoroughly investigated question, namely, how information made available during the word recognition process is used in sentence processing. Consider the sentences in (1). (1) a. The girl walked the dog. b. *The girl slept the dog. Although these sentences have the same syntactic structure, sentence (la) is a perfectly normal English sentence, while sentence (1 b) is clearly unacceptable. The difference in acceptability reflects the specific knowledge we have about the verbs, slept and walked, which defines how they can be used and specifically, whether or not they can have a direct object. All words have a set of specifications which are stored as lexically specific knowledge. Thus, when you “know” a word, you know what it means, how to use it in a sentence, and perhaps how to spell and pronounce it. This body of knowledge forms the word’s lexical representation in long term memory. This chapter is primarily concerned with “combinatory knowledge,” the subset of lexical knowledge that specifies how words are used in sentences. The combinatory knowledge associated with verbs is particularly interesting because it provides information about both structural and conceptual aspects of the sentence. It is a rich resource which could be exploited by the processing system to allow for the immediate interpretation of linguistic input. While sentence processing may occur in stages, our subjective experience is of continuous and immediate interpretation. Investigation of the role of combinatory lexical information in sentence processing has taken place within the context of the ongoing controversy about modularity in the language processing system. On the one hand, some have argued that the processing system is divided into functionally autonomous proc-

332

J.E. Boland and M . K . Tanenhaus

essing modules, with each module consulting only a limited domain of information (e.g., Ferreira & Clifton, 1986; Fodor, 1983; Forster, 1979; Frazier, 1989; Tanenhaus, Carlson, & Seidenberg, 1985). Such a system would allow, at least in principle, for fast, efficient computation within modules at the cost of occasional errors because relevant, “extra-modular” information was ignored. Several current proposals suggest that initial syntactic decisions are made without reference to some relevant combinatory lexical information (Mitchell, 1989; Ferreira & Henderson, in press; Frazier, 1987a & 1989; Frazier, Clifton & Randall, 1983). Others, in contrast, have argued that the processing system has an “interactive” or “integrative” architecture which optimally coordinates semantic, syntactic, discourse, and pragmatic information (e.g., McClelland, 1987; McClelland, St. John, & Taraban. 1989; Marslen-Wilson & Tyler, 1987). Researchers working within this framework emphasize the immediate use of lexically-specific information in parsing and interpretation. The issue of how and when combinatory lexical information is used in processing has taken on central importance in this debate because lexical representations are perhaps the richest source of information that could be used to integrate different aspects of representation. As we will see, the lexical knowledge associated with verbs provides considerable information about the rest of the sentence, and importantly, it provides information across many levels of representation - from the syntactic to the conceptual. An integrative or interactive system would be expected to exploit this information during processing, whereas evidence that this information cannot be used, or is delayed would provide strong support for a modular system.’ There is considerable disagreement about how lexical representations are used in sentence processing, especially syntactic processing. Part of the reason for this is because it is only recently that psycholinguists have begun to explore the time course with which different types of lexical information are accessed and used (but see Fodor, Bever, & Garrett (1974) for some interesting early proposals. This lack of consensus is also due to methodological issues. As the research topic focuses on increasingly fine grained time course issues, the considerable variety of measures used often obtain different results. The organization of this chapter is as follows. First, we discuss the major types of lexical information associated wilh verbs. Thcn we review some recent research investigating when these types of information are accessed and when they are used in sentence processing, focusing on our recent studies investigating how readers make use of lexical information in processing filler-gap sentences. We will discuss work on phrase attachment and clause closure, and the role of argument structure and verb control information in sentence processing.

LEXICALREPRESENTATIONS All lexically specific information must be stored or indicated for each

Lexical Representation in Sentence Processing

333

lexical item. This includes non-combinatory information such as phonological and orthographic descriptions. It also includes combinatory information such as syntactic category, number, gender, meaning or sense, subcategorization frames, the thematic roles associated with arguments, control properties, and conceptual (event) structure. When we talk about lexical information we are simply referring to lexically specific information which must somehow be encoded for each word. What counts as a word, or lexical entry, is an open question. For example, many words have multiple senses. The two senses may be very different (as in rose - which is either a flower or an actiodevent) or basically the same (as in kick - which is either a way to move an object across space using the foot, or away of expressing displeasure by quickly bringing the foot in contact with a relatively immobile object). The same problem occurs with word-forms which are ambiguous with respect to tense and aspect (e.g., passive/past ambiguities: The horse walked past the barn ... ). At some level of representation each sense must have its own set of lexical information, but multiple senses are indistinguishable based on perceptual information (spelling, pronunciation) - thus the ubiquitous phenomenon of multiple access (Marslen-Wilson, 1987). The level at which combinatory information is used to either build or evaluate interpretations is not known, but this is an issue we would like to pursue. Word recognition, or lexical access, is believed to make available the information in the lexical entry, and there is evidence that syntactic and semantic information from the lexical entry is available early in the recognition process, even before a single candidate has been selected (Marslen-Wilson, 1987). This raises a number of questions about how word recognition interacts with the use of combinatory lexical information to interpret sentences. When more than one word sense is accessed does the corresponding combinatory information become available too? If so, what are the consequences of having multiple sets of combinatory information active? If not, why is some lexical information made available while other information is not? It should already be apparent that word sense is related to other sorts of lexical information, including combinatory information. For instance, rose will be used in a sentence very differently depending on which of the above senses is intended. Pushing this relationship further, Fisher, Gleitman, and Gleitman (1989) have argued that syntactic subcategorization frames are closely related to the meanings of verbs. That is, verbs with similar meanings tend to have similar sets of subcategorization frames, and many semantic distinctions between verbs can be captured in terms of distinctions between sets of subcategorization frames. Because different types of lexical information are interrelated (and sometimes appear redundant), oftentimes it will not be clear what sort of information has been used by the processing system. In this chapter, “subcategorization frame” will be used to refer to a syntactic representation containing the arguments of a verb. We distinguish argu-

J.E. Boland and M.K.Tanenhaus

334

ments from adjuncts, which are always optional and appear freely. A subcategorized argument must appear in order for the sentence to be grammatical.’ The subcategorization frame for a purely intransitive verb (e.g., sleep) contains only one argument, the subject N P (as in, John sleeps lightly). The dative verb, give, on the other hand, has a subcategorization frame containing three arguments, the subject NP, and two object NPs (as in John gave Mary a book). However, not all arguments are noun phrases. Remind takes three arguments: subject, object, and infinitive phrase (as in John reminded Mary to sleep). Many verbs, give and remind included, have multiple subcategorization frames. Several alternative subcategorization frames for give are illustrated in 2 below. (2)

a. NP - NP PP: John gave a book to Mary. b. NP - NP NP: John gave Mary a book. c. NP - PP: John gives to charity. d. NP - passive PP: The book was given by John.

The thematic roles of a verb are directly related to its subcategorized arguments. Each argument, whether it is a noun phrase, prepositional phrase, or some type of sentential complement, must have a thematic role associated with it; just as verbs have subcategorization frames, they also have corresponding sets of thematic roles. Thematic roles are motivated in part by an attempt to characterize the consistent relationship between a verb and its arguments under differing subcategorizations. In this way, thematic roles map syntactic structure onto conceptual structure (see Tanenhaus, Carlson, & Trueswell, 1989b). For example, the second NP in (2b) has the same relationship to the verb as the first NP in (2a), that of “Theme” - the object being moved. Likewise, when a verb is passivized its arguments appear in a different syntactic form, but maintain the same thematic roles. In addition. thematic roles can characterize semantic differences between verbs with the same subcategorization frames by giving the arguments different thematic roles. For example, rob and steal are different thematically because rob assigns a “Patient” role to its object while steal assigns a “Theme” role (thus, John robbed a bank, but He stole money). A real problem lies in determining the appropriate level of generalization. How many thematic roles there are and how they can be defined remain open questions. There is considerable debate about whether thematic roles are syntactic, semantic, or conceptual in nature, and whether they are even grammatically significant entities at all (see Dowty, 1988; Ladusay, 1988). The view we are adopting here is that thematic roles are a useful level of representation for sentence processing, lying between syntactic and conceptual levels of representation. We will refer to the block of information which includes the set of subcategorization frames and the thematic roles associated with them as the verb’s “argument structure.” The conceptual (event) structure describes a still higher level of abstrac-

Lexical Representation in Sentence Processing

335

tion. Every verb specifies a certain type of situation, including the number and type of thematic roles that persons or objects must play in the situation (Farkas, 1988). For example, consider what it means to have a giving situation.Such a situation requires three “participants:” a giver, a receiver, and the thin&) given. Certain restrictions apply as to what makes a good participant in each of these three roles; these restrictions are partially characterized by the thematic roles assigned by the verb, and further specified by the core meaning of the verb. A certain state of affairs is presupposed before the giving event (the giver has the thing) and another state of affairs afterwards (the receiver has the thing). The lexical representation of give must store or provide access to this information, so that it can be used to form and interpret sentences containing the word give. The conceptual structure can be thought of as the complete set of thematic roles, plus the lexically-specific information about the situation and its participants which is necessary to constitute an instance of the situation denoted by the verb. Finally, there is a type of lexical information which applies specifically to the class of verbs which take “subjectless” infinitive phrases as arguments. In these sentences, there is no overt subject of the infinitive verb. Rather, either the subject or the object of the matrix verb is understood as - or controls the interpretation of - the implicit subject of the infinitive. Some examples of verb control are shown in (3) below. Sentence (3a) illustrates the use of a”subject control” verb (promise), and sentence (3b). an “object control” verb (tell)? (3) a. The girli promised her brother to sing. b. The girl told her brotherj - to sing. In sentence (3a), the subject of promise is understood to be the subject of the infinitive, (i.e., the girl is promising her brother that she will sing), whereas in sentence (3b). the object of the matrix verb is understood as the subject of the infinitive, (i.e., the girl is telling her brother that he should sing). We will assume that infinitive phrases have ,”gap” in subject position. Interpreting an infinitive phrase requires the comprehender to coindex the subject gap with the appropriate antecedent noun phrase, which is said to be the “filler” for the gap: Gaps are represented by underlined blanks and subscripts coindex the gap with its antecedent noun phrase.Some accounts of verb control assume that the control properties of a verb are stored directly as part of the lexical entry of the verb (Bach, 1979; Bach & Partee, 1980; Ruzicka, 1983). However, linguistic evidence suggests that control is not arbitrary and can be derived from syntactic, semantic, or pragmatic pioperties of the verb (Chierchia, 1984; Comrie. 1984; Farkas, 1988; Jackendoff, 1974; Sag & Pollard, in press). Semantic and pragmatic theories of control typically focus on the type of situation denoted by the verb. These accounts emphasize that verbs with similar meanings invariably have the same control properties (for a recent review, see Sag & Pollard, in

336

J.E. Boland and M.K.Tanenhaus

press). Although the appropriate grammatical status of control remains unresolved, all approaches agree that information made available by the verb is central to determining the controlling noun phrase.

THEROLE OF LEXICAL INFORMATION IN SENTENCE PROCESSING In the material which follows, we will look at some specific types of lexical information with these questions in mind: What information is made available when a lexical entry is accessed? What evidence is there that the information has been used/ignored? How early is the information used? In thinking about how various types of lexical information might be used, it is useful to distinguish between parsing and interpretation (see Altmann, 1989). Parsing refers solely to the derivation of syntactic structure from an input string. Interpretation involves higher level processes and its endpoint is a meaningful representation of the situation denoted by the input. We will explore the role of lexical information in both of these domains, focusing first on parsing. Attachment and Closure Since the early 1960s. it has been commonly assumed that comprehending a sentence involves building a syntactic structure for it. Two major questions that arise in building a constituent structure are when to terminate or “close” a constituent, and where to attach a new constituent. These two interrelated procedures are the basic steps of structure-building, or parsing. Several types of combinatory information are relevant to attachment and closure. Subcategorization information could be used to find out if certain structures are likely to occur (or even possible). Thematic role information could be used to rule out structures based on the thematic characteristics of the part of the sentence that has been processed and the phrase to be attached. However, it is possible that no combinatory lexical information is used for the initial parsing decisions. As originally pointed out by Bever (1970), there are basic patterns in a language which can be reflected by some fairly simple parsing strategies. Use of such strategies could provide a means for selecting between alternatives when lexical information is ambiguous. They could also substantially decrease the amount of time and working memory required to build structural interpretations by minimizing the amount of information necessary to make structural decisions. In 1978. Frazier (see also Frazier & Fodor, 1978) proposed two parsing strategies, Minimal Attachment and Late Closure, which compute the simplest structure allowed by the phrase structure rules, taking into consideration only the syntactic category information of the words being processed. These strategies make powerful predictions about when garden-paths should and should not occur. The beauty and efficiency of these strategies is that lexical and contextual information can be ignored, or backgrounded, until the structure is built. If, in

Lexical Representation in Sentence Processing

337

fact, lexical information is not ignored, the power of these two strategies is greatly reduced. Relevant lexical information might be ignored because of strict limits on time and working memory, which force the processing system to quickly build a syntactic interpretation. In fact, these considerations provide the psychological motivation for Minimal Attachment and Late Closure. Mitchell (1989) and Ferriera and Henderson (in press) have argued that no lexical information except syntactic category is used to build structural interpretations. Decisions are made without relevant information such as subcategorization. According to this view, garden paths should occur whenever the parsing strategies fail to provide the correct structure. The garden paths should be reflected in measures of processing difficulty. Lexical and contextual information filter out inappropriate structures and repair garden paths. This type of model was called “lexical filter,’’ by Mitchell (1987). On the other hand, if combinatory information is available immediately, it could be used to build the initial structural interpretations. There would still be occasional garden paths because lexical information can be ambiguous, but the initial “blind” step would be eliminated. This type of model was called “lexical proposal” (Mitchell, 1987), although combinatory information checks/filters structures as well as proposing them. Both types of models are open to serial and parallel versions. Since the lexical proposal models would never predict a garden path when the lexical information is unambiguous, evidence that garden paths occur in such instances would be strong evidence against proposal models. In contrast, evidence that syntactic structure has been proposed or anticipated based solely on lexical information and in advance of direct syntactic evidence would be strong evidence against the filter models, since they claim that lexical information cannot be used until a structure is constructed from the input string. Syntactic Subcategorization Notice that there are actually two kinds of lexical information about subcategorization which would be relevant for parsing decisions: the “list” of subcategorization frames and information about the frequency of the various frames for a particular verb. If a verb has only a single possible subcategorization frame, it is obligatorily subcategorized. When there is more than one possible frame, but one is much more common than the other(s), the most common frame is the preferred subcategorization. (Subcategorization preferences are undoubtedly influenced by prior context, but how this works is largely unexplored.) The conservative parser might ignore preference information and use only obligatory subcategorization information since it will always be right. However, Mitchell (1987, 1989) and Ferreira and Henderson (in press) argue that even obligatory subcategorization is initially ignored. Mitchell (1987) conducted a self-paced, phrase-by-phrase reading study

J.E. Boland and M.K.Tanenhaus

338

which compared sentences like (4a) & (4b), where the slash divides the sentence into the two display segments Mitchell used for the sentences of interest. (All of the studies described in this section presented the stimuli visually, with no commas or other internal punctuation.) (4)

a. After the child visited the doctor / prescribed a course of

injections. b. After the child sneezed the doctor / prescribed a course of injections. According to Frazier's (1978) Minimal Attachment Strategy, a noun phrase immediately after a verb will always be attached to the verb phrase as an object rather than being taken as the subject of a new clause because the former attachment creates a simpler structure. However in the sentences above, the latter attachment is correct. Thus, the Minimal Attachment Strategy predicts that a garden path should occur, in which the doctor is mistakenly taken to be the direct object of the verb, whether it is visited or sneezed. Mitchell's results were consistent with this prediction. Subjects read the first part of (4a) faster than the fist part of (4b). but the last part of (4b) was read faster than the last part of (4a). Mitchell argued that a garden path occurs for both verb types but it is corrected quickly by the lexical filter of the intransitive verb in (4b). whereas in the transitive condition (4a) the [Subject [Verb Object]] structure passes through the lexical filter and isn't corrected until the unambiguous information is encountered in the second segment. If this explanation is correct. it provides evidence that lexical information becomes available fairly quickly since the effects were seen two words after the verb. It might be the case that the potential clause boundary provided access to the lexical knowledge. Of course, this pattern of data can be explained in another way. Suppose subcategorization information is used to build structure. Since visit is a transitive preference verb, Mitchell's garden path explanation would still account for the longer reading times in the last half of (4a). But since sneezed is obligatorily intransitive, a clause boundary would be posited after sneezed. However, the absence of a comma after sneezed and the line break after docror strongly bias a clause boundary after doctor. This is clearly incompatible with subcategorization information. The longer reading times for the first segment of (4b) could well have been caused by these conflicting cues. (Bates & MacWhinney, 1982), rather than initial blindness to subcategorization information. Ferreira and Henderson (in press) tracked eye movements and collected self-paced reading times for sentences like those in (5) below, that are temporarily ambiguous between a [Subject [Verb Object]] structure and a [Subject [Verb Sentence]] structure. Since the later structure is correct but more complex, here again the Minimal Attachment Strategy predicts a [Subject [Verb

Lexical Representation in Sentence Processing

339

Object]] garden path. (5)

a. He forgot (that) Pam needed a ride with him. b. He wished (that) Pam needed a ride with him.

The sentences used matrix verbs which were either transitive-preference(5a) or intransitive (5b)5 and half of the sentences contained the complementizer, that, which biases toward the sentential argument reading. The important comparison was between sentences of each verb-type with and without the complementizer. Ferreira and Henderson found that subject looked at needed longer and made more regressive eye movements if there was no complementizer for both verbtypes. The complementizer effect was also found using a word-by-word reading time measure. Ferreira and Henderson argue that if subjects had used subcategorization information, there would be no complementizer effect for the inuansitive verb - the complementizer would be redundant with the subcategorization information. Unfortunately, Ferreira and Henderson provide no evidence that the effect is not due simply to the fact that sententid complements are easier to process when preceded by a complementizer. Thus, the study does not provide definitive evidence that subcategorization information is ignored or unavailable. Let us now consider evidence which has been used to support the position that subcategorization information, both obligatory and preferred, guides phrase attachment and clause closure. In a word-monitoring experiment, MarslenWilson, Brown, and Tyler (1988) compared sentences with transitive and intransitive verbs. They found that response times to targets such as guitar were slower in sentences with obligatorily intransitive verbs (e.g., The young man slept the guitar) compared to sentences with transitive verbs (e.g.,The young man carried the guitar). This indicates that subcategorization information is, at minimum, available, and could potentially be used for parsing decisions. Stowe and Holmes (1989) tested this by comparing obligatorily intransitive verbs with intransitive preference and transitive preference verbs in sentences like those shown in (6) below. In control sentences, a disambiguating phrase (shown in parentheses) was inserted after the verb. (6)

a. Although Mr. McKenzie arrived (with his partner) his wife paid no attention. [intransitive verb] b. Although Mr. McKenzie cheated (his partner) his wife paid no attention. [intransitive-preference verb] c. Although Mr. McKenzie heard (his partner) his wife paid no attention. [transitive-preference verb]

Using a self-paced, word-by-word grammaticality detection task, they found increased reaction times at paid in both potentially transitive conditions, (6b and c), but not in the obligatorily intransitive condition, (6a). This suggests that

J.E. Boland and M.K.Tanenhaus

340

garden paths occur only when the lexical information is ambiguous, as predicted by the lexical proposal models. Further, reaction times at paid were significantly longer for the transitive preference versions compared to the intransitive preference ones, indicating that the garden path was modulated somewhat by lexical preferences. Unfortunately, these results also fail to distinguish between lexical projection and lexical filter models. In addition to the results described above, Stowe and Holmes found an interaction between the presence/ absence of the disambiguating phrase and verb transitivity. That is. subjects responded more slowly to his (wife) when there was no disambiguating phrase compared to when there was one in the obligatorily intransitive versions, while they responded slower to his (wife) when there was a disambiguating phrase in the optionally transitive versions.The difficulty of incorporating a noun phrase after an obligatorily transitive verb could indicate either a local garden path in which a [Subject[Verb Object]] structure was posited and filtered out, or it might reflect the inconsistency of a noun phrase in object position after an intransitive verb. Stowe and Holmes’ results are thus consistent with either a lexical proposal model or a lexical filter model in which the filter operates on a word-by-word basis. “Filler-gap” sentences provide another domain in which to examine the use of subcategorization information in parsing decisions. Example sentences are shown in (7). (7)

a. Which dog, did John see - i? b. Which dog, did John tiptoe past - ,?

Interpretation of each sentence requires associating the filler, which dog, with the structural position denoted by the gap. In Wh-questions, the filler can be identified immediately by the Wh-phrase, but the location of the gap may be uncertain until the end of the sentence. Positing a gap right after the verb is equivalent to positing a [Subject [Verb Object]] structure. The question is, does the subcategorization of the verb affect whether or not subjects posit a gap after it? There is a great deal of evidence for lexical effects in the interpretation of filler-gap sentences. Clifton, Frazier, and Connine (1984) found effects of both obligatory and preferred subcategorization on whole-sentence grammaticality judgments. Tanenhaus, Stowe, and Carlson (1985) found effects of preferred transitivity (“lexical preference effects”) using a word-by-word self-paced reading task. This effect was replicated using an evokcd potential measure by Garnsey, Tanenhaus and Chapman (1989a, 1989b). In addition, Nicol and Osterhout (1989) found evidence that transitive verbs prime fillcrs but obligatorily inuansitive verbs prime fillers Significantly less. What is not clear is where the lexical effects take place - during the initial parsing decisions, or later in the interpretation process. There is overwhelming support for the postulation and filling of gaps in

Lexical Representation in Sentence Processing

34 1

Wh-questions immediately after transitive verbs, but this can be explained by all the models, whether or not they use lexical information to propose structures. Given the tendency to posit gaps early, the crucial question is whether gaps are posited after all verbs as a general strategy in filler-gap sentences, or if gaps are only posited after transitive verbs. If gaps are posited after intransitive verbs, then that constitutes evidence that subcategorization information is ignored. The Tanenhaus et al. (1985) and Garnsey et al. (1989b) studies suggest that subcategorization preferences are, in fact, used. Kurtzman (1989) provides additional evidence that gaps are not posited after intransitive verbs. He presented sentences like those in (8) using an RSVP technique. Of interest here are the sentences that used intransitive preference ((a) and (b)) and pure intransitive ((c) and (d)) verbs either transitively ((a) and (c)) or intransitively ((b) and (d)). Subjects made grammaticality judgments when they heard a beep (represented by the asterisk). The percent of “grammatical” judgments is shown after each sentence.

(8)

a. What, did John escape - while * ... 70% b. What, did John escape from - * ... 90% c. What, did John escape - while * ... 35% d. What, did John crawl under - * ... 93%

,

,

Most sentences using intransitive-preference verbs were judged grammatical, but more intransitive sentences were judged grammatical than transitive sentences. In contrast, the sentences with pure intransitive verbs were judged to be grammatical only if the verb was used intransitively. In contrast, Frazier and Clifton (1989) argue that there is a general strategy to posit and fill a gap as soon as possible, regardless of verb-type. Frazier and Clifton had subjects read questions with intransitive preference verbs similar to those in (9) in a self-paced, phrase-by-phrase study. (9)

a. What did the man whisper - to his fiance during the movie? b. What did the man whisper to his fiance about - during the movie?

They found that subjects read the the phrases after the verb faster in the transitive versions (9a) compared to the intransitive versions (9b). This suggests that it is easier to use an intransitive-preference verb transitively than intransitively. However, these results are consistent with those of Kurtzman if the early-filling strategy operates within the restrictions imposed by obligatory subcategorization. There may be constraints that make processing difficult when too much structure separates a filler from its gap. If so, it would be easiest to posit and fill a gap as soon as possible, as long as the position was licensed by the verb.

J.E. Boland and M.K.Tanenhaus

342

Thus, we can envision a model in which both strategies and lexical information play a role. We have seen that the evidence for both lexical filter and lexical proposal is open to interpretation. While these studies provide evidence that lexical information is available and used fairly quickly in parsing, none of them provide conclusive evidence for or against the lexical proposal models. Mitchell (1987) and Ferreira and Henderson (in press) attempted to disprove the lexical proposal model by showing that garden paths occur even when there is unambiguous lexical information which should prevent the garden path. Unfortunately, the results are confounded by other factors which cast doubts on their interpretation. On the other hand, Stowe and Holmes (1989) showed that a similar garden path can be averted by unambiguous lexical information and modulated somewhat by lexical preference information. However, it is possible that the garden path occurred and was corrected before the disambiguating region of the sentence. Thus their results do not provide conclusive evidence either. The filler-gap studies, taken together, show that strategies and subcategorization information are both used in parsing, and there is some evidence that suggests that subcategorization information is used to propose structures. However, it is also possible that the tasks used by Kurtzman (1989). Tanenhaus et al. (1985) and Garnsey et al. (1989a, 1989b) reflect later stages in processing than the tasks used by Frazier and Clifton (1989). Thus, much disagreement exists, and the matter is still unsettled.

Thematic I n for ma tion Now we will turn to the use of thematic information in parsing. The next set of studies manipulates subject animacy, which limits the number of possible argument structures. This is another type of lexical information which could eliminate garden paths. In particular, we will be examining whether information about plausible thematic role assignment provides input to the parser about the plausibility of alternative syntactic analyses. The basic research strategy adopted here is to determine whether garden-path effects usually observed within locally ambiguous sentences are eliminated when biasing thematic information is provided. Stowe (1989) manipulated subject-animacy of causative/ergative verbs in sentences like those shown below in (10).

(10) a. Before the police stopped the driver was already getting nervous. b. Before the truck stopped the driver was already getting nervous. When causative/ergative verbs have an animate subject, as in (10a). the transitive subcategorization is “preferred,” and when they have an inanimate subject, as in (lob), the intransitive subcategorization is preferred. Another way to talk

Lexical Representation in Sentence Processing

343

about this is in terms of the thematic structure of causative/ergative verbs. Two possible thematic structures are [Agent -Theme] and [Theme -1, corresponding to [Subject [Verb Object]] and SV, respectively. Since the role of Agent is only consistent with an animate noun phrase, the first thematic structure (and corresponding syntactic frame) would not be available6after the verb in (lob) if such information can be used to limit proposed structures. In a self-paced, word-by-word reading task, subjects spent more time reading was in (10a) compared to (lob), indicating a local garden path in (10a). There is no evidence for a corresponding garden path in (lob) as would be predicted by the Late Closure Strategy. However, let us consider the possibility that the [Subject [Verb Object]] structure was initially posited in (lob) and corrected by the lexical filter before the second verb was read. We would then predict no garden path effect at was, but the cost of structural revision should be reflected in longer reading times at the driver,’ which Stowe did not find. Thus, the pattern of effects is most consistent with a lexical proposal model. Stowe interpreted these results to mean that thematic information was available to the parser, but the sample materials illustrate that animacy alone is often not a reliable way to manipulate argument structure. Trucks are mobile and often take roles similar to Agent (e.g., The truck passed the police car), or an Instrument role (e.g., The truck pulled the car out of the ditch). In Stowe’s study, these problems would only serve to minimize her effects (by making the inanimate versions more like the animate versions), so they do not compromise her results. However, they do raise the question of how the processing system could make use of thematic information in an efficient manner. The lexical properties of the subject noun phrase undoubtedly interact with the lexical properties of the matrix verb, and the resulting expectations may not be adequately captured by generalized thematic roles. A broad range of conceptual knowledge may be drawn on as well. Taraban and McClelland (1988) also concluded that conceptual information influences the parser. They argue that prepositional phrase attachment is not governed by the Minimal Attachment heuristic, but rather by the expectations set up by the content of the sentence frame. Specifically, they suggest that the sentence frame sets up an expectation for the prepositional phrase to fill a particular thematic role. For instance, in the sentences below (1la) sets up an expectation for an Instrument and (llb) sets up an expectation for a Modifier, the first being a verb phrase attachment and the second being a noun phrase attachment. (11) a. The janitor cleaned the storage area with ... b. The hospital admitted the patient with ... Thus, not all sentences predict verb phrase attachment, and not all verb phrase attachments (or Minimal Attachments) are equally good. A noun phrase corn-

344

J.E. Boland and M.K.Tanenhaus

pletion like cancer is much better than a verb phrase completion like apologies for (1lb). For (11a), the manager and the broom are both verb phrase completions, but the Instrument completion (broom) seems more congruent. These intuitions were confirmed by rating norms. Word-by-word reading times were consistent with the norms: A violation of expected attachment (to noun phrase or verb phrase) or expected role of prepositional phrase led to longer reading times immediately after the prepositional phrase. While this series of experiments provides evidence that conceptual information can guide (or perhaps bias) parsing decisions, Taraban and McClelland make no claims about what aspects of the sentence are responsible. Ferreira and Clifton (1986), like Stowe (1989), looked for effects of subject-animacy, but used eye movements to measure processing difficulty and reached quite different conclusions. Eye movement measures, unlike most reading time measures, have the advantage of measuring a process which normally occurs during reading (eye movements) rather than an additional task such as button pressing or judgments. Thus, reading proceeds at fairly normal rates.This could be especially important when studying parsing since lexical filter models claim that there is an initial syntactic stage of processing in which lexical information is not available. Ferreira and Clifton used verbs which had identical simple past and past participle forms, such as examined in (12) below. The entire sentence was presented at once, with one line break. (12) a. The defendant (that was) examined by the lawyer turned out to be unreliable. b. The evidence (that was) examined by the lawyer turned out to be unreliable. The simple past tense form of the verb assigns the role of Agent to its subject and the past participle form assigns Theme to its subject. The sentences are syntactically disambiguated at the by phrase, or at the words [hat was, which appeared in half of the sentences of each type. However, the sentences with inanimate subjects could be thematically disambiguated earlier since Agents are generally animate. To the contrary, Ferreira and Clifton found that subjects looked longer in the by-phrase region, by the lawyer, in sentences without that was, regardless of whether the subject was mimate or not. This result is consistent with a pure Minimal Attachment Strategy that does not consult subcategorization knowledge. This experiment is subject to criticisms which have been raised above. First, we should not be surprised that reduced relatives are more difficult than unreduced relatives (the “complementizer effect” again). Second, many of the “inanimate” subjects can serve Instrument or Agent-like roles with the verbs used (e.g.. The car towed ...). making the “inanimates” more like the “animates,” and in this case, that does compromise the data since Ferreira and

Lexical Representation in Sentence Processing

345

Clifton are arguing that there is no difference. Trueswell. Tanenhaus, and Garnsey (1990) tested the latter criticism. They adjusted the materials so that none of the inanimates could serve Instrument or Agent-like roles and re-ran the eye-tracking study with the new materials and no line-breaks. They found that animacy did interact with the presence/absence of the complementizer - there was a complementizer effect for sentences with animate subjects, but not for sentences with inanimate subjects. Thus, they argued that thematic information does influence phrase attachment. They ran a second set of experiments using a whole-sentence reading time measure to determine whether the difference between their results and those of Ferriera and Clifton could be explained by the difference in materials or the difference in display. Trueswell et al. found that both the materials and the display conuibuted to the difference in results. They hypothesized that recognition of the verb makes available both the past tense and past participle forms with their corresponding thematic structures. A single form is quickly selected based on frequency (the past tense form being more frequent than the past participle form) and contextual information, including animacy and information from the right periphery. Burgess (1990) comes to a similar conclusion in a self-paced reading study in which he compared one-word-at-a-time and two-word-at-a-time presentation using materials modeled on those used by Ferriera and Clifton and Trueswell et al. With one-word-at-timepresentation, longer reading times were found after the ambiguous verb for both sets of materials, regardless of subject animacy. With two-word-at-a-time presentation, in which the verb and the preposition by were presented together, longer reading times were found for both animate and inanimate versions of the Ferriera and Clifton materials, but only for the animate versions of the materials constructed using the Trueswell et al. criteria. These results would seem to rule out the strong claim made by Ferreira and Clifton (1986). that thematic information does not have an effect on phrase attachment. Trueswell et al. (1990) and Burgess (1990) have shown that sufficiently biasing context does have an effect, but of what nature? Their data cannot be naturally explained by a serial filter model: Since there was no complementizer effect for the sentences with inanimate subjects, it is unlikely that the past tense form was initially selected, then rejected, as in the sentences with animate subjects. More likely, the past tense form was not selected because it was inconsistent with the inanimate subject. Trueswell et al. do not make the strong claim that an inanimate subject rules out the possibility of the past tense form (we have already noted the occurrence of active verbs with Instruments, and there are other instances as well in which inanimate subjects take active verbs). Rather, they suggest that there is some parallelism in the system. They argue that multiple forms are accessed in parallel, just as for any ambiguous word. In addition, argument structure must be accessed with the word forms or there would be no basis on which to choose between them.

346

J.E. Boland and M.K.Tanenhaus

In summary, much of the evidence regarding use of lexical information in parsing is consistent with more than one position. It is very difficult to distinguish between evidence that combinatory lexical information has been used to project structure (lexical proposal) and evidence that lexical information has been used to check structures built without reference to it (lexical filter). The issue is further complicated by the various possible points in the sentence at which filtering could occur. What has been demonstrated clearly is that lexical information is available quickly. What can be ruled out are models in which lexical filtering does not occur on a word-by-word or nearly word-by-word basis. Building upon the Trueswell et al. conclusions, consider the hypothesis that all syntactic ambiguity can be reduced to lexical ambiguity. Lexical ambiguity may be morphological, as in the tense/aspect ambiguity of examined. Or it may be semantic or thematic, with slight or drastic variations in meaning tied to different argument structures. The point is. many verbs have multiple possible argument structures associated with them and, if argument structure information is accessed when a verb is recognized, then one might think of multiple argument structures for a particular verb in the same way that one thinks about multiple senses of ambiguous words like bank. However when the argument structure of a verb is ambiguous, a mistaken guess disrupts interpretation of the entire sentence since the verb’s argument structure specifies number and type of participants in the sentence and the relationship between them. If we take multiple access seriously, when a verb is recognized all of its “senses” are initially available until one is chosen based on context and frequency. Such a view predicts (1) all alternative argument structures for a verb are initially available, and (2) one is quickly selected based on information gathered from the leftcontext and perhaps frequency information. This is one possible parallel proposal model. The model presumes that lexical information is available immediately upon recognition of a word, and all information consistent with the orthographic or phonological string will be activated in parallel. Thus, lexical information is accessed in a bottom-up manner, but then becomes input for the parser. If it is right, what does it say about processing? One concern is that in the face of massive ambiguity, such a parallel proposal system might provide too much information to be processed efficiently. However, there is linguistic and psychological evidence that all argument structure ambiguities are not alike. Tanenhaus and Carlson (1989) “decomposed” the semantic representation of verbs into two parts, a “core meaning” and a set of thematic roles. According to them, some verb ambiguities are core ambiguities and others are thematic ambiguities. They found that sentences containing thematic ambiguities are easier to process - presumably garden path recovery is easier - than sentences containing core ambiguities. This could reflect a hierarchical representation of word meaning in which thematic variations are organized within core meanings which are organized within a

Lexical Representation in Sentence Processing

347

phonological or orthographic string. Thus, all the information could still be made available when a word is recognized, but a core meaning would be quickly selected, leaving available the thematic structures within that particular core meaning. Shapiro, Zurif, and Grimshaw (1987, 1989) have argued that alternative argument structures are made available when a verb is recognized. Further, they predict (and find) that the greater the number of alternative argument structures, the more processing complexity there is associated with that verb. However, Schmauder (1990) has been unable to replicate that effect. We would argue that the complexity prediction may be misguided. After all, there is no evidence that ambiguous words have more of a processing load than unambiguous words because multiple meanings are accessed (complexity differences, when they obtain, are most likely due to integration effects). We would argue that the same should hold for verbs with multiple possible argument structures. Argument Structure

If recognition of the verb in a sentence makes available the set of arguments/thematic roles associated with it, then this information would be useful for interpreting (and perhaps projecting) the syntactic structure of a sentence. Tanenhaus, Garnsey, and Boland (1990, see also Boland, Tanenhaus, Carlson, & Garnsey, 1989) have taken advantage of the properties of filler-gap questions to explore this issue. Consider the sentences in (13). The sentences are odd becausefood is not a very good direct object of read.

(13) a. The child read some food in school. b. Which food, did the child read - in school? People should find the sentences strange as soon as they realize that food is the object being read. In (13a), this doesn’t happen until the wordfood is read. But in (13b),fOOd is at the beginning of the sentence, way before the verb. At some point, the reader must realize thatfood is actually the direct object of read, and at that point, the sentence should seem odd. Thus, an oddity detection task will help ascertain the point at which the gap is “filled.” We have called this way of measuring filler-gap interpretation the “embedded anomaly technique.” In general terms, the embedded anomaly approach manipulates the plausibility of a fronted Wh-phrase (a salient filler) for a particular gap or potential gap in the sentence. Since the sentence cannot become implausible until the filler has been associated with the gap, the point where plausibility effects occur indicates when the filler has been associated with the gap. We have already noted the large body of evidence that post-verbal gaps are filled as soon as a transitive verb is encountered. It includes “filled-gap” studies (Crain & Fodor, 1985; Stowe, 1986; Tanenhaus, Boland, Garnsey &

J.E. Boland and M . K . Tanenhaus

348

Carlson, 1989a; and Frazier & Clifton, 1989). cross-modal priming studies showing priming to the antecedent of a trace beginning at the verb (e.g., Clifton & Frazier. 1988; Nicol & Swinney, 1989; Swinney, Ford, Frauenfelder & Bresnan, 1988). embedded anomaly effects beginning at the verb with evoked potential and behavioral response measures (Garnsey, Tanenhaus & Chapman, 1989a; Tanenhaus, Stowe & Carlson, 1985; Tanenhaus et al.. 1989a). and assorted reading time measures (Clifton, Frazier & Connine, 1984; Frazier & Clifton, 1989). The processor does not wait for direct evidcnce of a gap, but projects one when it encounters a verb. For instance, (Tanenhaus et al., 1989a) saw anomaly effects at the verb when the filler was implausible as an argument of the verb. However, if the system can use information about argument structure, one would not predict anomaly effects if the filler was plausible as some argument of the verb. Argument structure specifies the argumentdthematic roles associated with a verb, each of which is a potential gap site. As long as the filler can

Table 1 Sample sentence from Tanenhaus et al. (1990)

*‘lhc last4 wads famed eihu

U)

adverbial phrase (for Ihe IN Iwo vcrb-types) or

M

infinitive com-

pkmenl (for Ihe I m p V 6 l Y p ) .

plausibly fill one of the gap sites, the sentence should still be plausible. Tanenhaus, Garnsey, and Boland (1990) looked at verbs with thrcc types of argument structures: “simple transitives” (e.g.. watch) took only a subject and an object; datives (e.g., give) took a subject and two objects; and “infinitive complement verbs” (e.g., remind) took a subject, an object, and an infinitive argument. We predicted that simple transitive verbs would force the parser to posit and fill a direct object gap because there are no other argument positions which the filler could fill. Dative verbs should not require the parscr to posit and fill a direct object gap since there is another possible argument position. Likewise, a filler will not necessarily fill a direct object gap after an infinitive complement verb since the infinitive complcment is itself an argument with a thematic role, and it contains another verb which may provide further argument positions. Although both infinitive complement verbs and dative vcrbs can be used with just one object (e.g., Tom told Bill and John sent his assistant), we chose verbs where the role associated with the second internal argument seems to be a necessary part of the event described by the verb, even if it is not made

Lexical Representation in Sentence Processing

349

explicit in the sentence. An informal test for this is that, without context, a sentence using the verb with two internal arguments seems more natural than

Figure 1 . Data from Tanenhaus et al.. 1990. a) Cumulative percent of "no" responses at each word position. b) Mean reading times for positive responses at each word position.

one using the verb with only one internal argument. A sentence set, consisting of two sentences which were identical except for the fronted Wh-phrase, was constructed for each verb. In all cases, the Whphrase filled the direct object gap. In one version the filler was plausible as the direct object, and in the other version the filler was implausible. Sample sen-

350

J.E. Boland and M.K.Tanenhaus

tences are shown in Table 1. At the verb, there was no syntactic or semantic evidence to force a filler-gap assignment. That is, it was always possible to complete the sentence in such a way that the Wh-phrase would be a plausible filler for some later gap-position (e.g., Which Stone did the assistant watch the craftsman carve - ?). Plausibility effects at the matrix verb are evidence that the filler was immediately associated with the (then) potential gap position. Plausibility effects a word or two later indicate that the filler-gap association was not made until there was direct evidence of a gap in that position which needed a filler. Subjects controlled the word-by-word presentation rate by pressing a key for each word. If the sentence stopped making sense, they were to press another key as soon as they noticed the implausibility. This was called a “no” response. After a “no” response, presentation of the sentence was halted and a new trial began. Reading times were collected from the onset of each word. Thus, we had two measures of implausibility, increased reading times and an increased number of “no” responses at a particular word position. A summary of the results can be seen in Figure 1. There were markedly different patterns between sentences with simple transitive verbs and those with infinitive complement verbs. Simple transitive sentences showed increased numbers of “no” responses and slower reading times for the implausible versions, beginning at the verb. Infinitive complement sentences showed no effects until one word later. As predicted, dative sentences patterned with the infinitive complement sentences. The results support the hypothesis that argument structure is used for filler-gap assignment. Plausibility effects were seen at the simple transitive verbs but not at the dative or infinitive complement verbs. Plausibility effects for these verbs were only seen after syntactic information made it clear that the filler was the direct object of the verb. The effects can be seen most clearly in the “Percent No” data, but are reflected in the reading time trends as well. These results suggest that even complex lexical information is available early in the comprehension process and can influence structure assignment. On their own, these results don’t explain how argument structure influences structure assignment. We have argued that recognition of a verb makes available the set of roles associated with the subcategorized arguments of the verb. An “active filler” (Clifton & Frazier, 1988) is immediately associated with one of these roles, with preference given to the role associated with the direct object as long as the filler is consistent with that Further experiments suggest that this way of thinking about the role of argument structure is correct. According to our hypothesis, a filler that is semantically inappropriate for the direct object role but is appropriate for the indirect object role should be immediately assigned the indirect object role. We tested this by constructing question pairs with fillers that were not plausible direct objects, but were plausible indirect objects. In each pair, one of the fillers was implausible with respect to the direct object uscd. In this way, the implausibility was not appar-

35 1

Lexical Representation in Sentence Processing

ent until the direct object was read. Consider the sentences in (14). (14) a. Which uneasy pupils did Harriet distribute the science exams to

- in class? b. Which car salesmen did Harriet distribute the science exams to in class? It doesn't make sense to distribute science exams to car salesmen, although other things can be distributed to them. The question of interest is: When do people notice the implausibility - before or after the to, which marks loo

+ ,

no-

Which uneasy pupils did Harriet

...

---.hWhich car salesmen did Harriet ...

,...__.. &-.."'-""A ,,&-"

*.....,..'

M -

,Li40c1

e V

v

L

a"

h

E

v v)

.-;

700

b

M

.5

(00.

U cp 0)

N

500

1

I

I

1

Figure 2 . Data from Tanenhaus et al., 1990. a) Cumulative percent of "no" responses at each word position. b) Mean reading times for positive responses at each word position.

352

J.E. Boland and M . K . Tanenhaus

the position of the trace. As can be seen in Figure 2, anomaly effects were found beginning at the direct object noun phrase (science), before the to, suggesting that the filler was understood to be the indirect object before the end of the noun phrase (and before the syntactic trace marking the gap position). Thus, we found evidence for early “filling,” in a role-assignment sense. This raises the question of whether the semantic interpretation of the filler is mediated by a structural trace. One possibility is that when the verb is encountered, structure is projected based on argument structure information, and the thematic role comes from the projected trace. The other possibility is that fillers can be interpreted as satisfying the thematic rolcs specified in the argument structure without mediation from the structural projection of a trace. We believe the second possibility, the direct interpretation of thematic roles, is the more likely explanation. While we do not have definitive evidence in favor this hypothesis, there are two reasons why we favor it. First, the structural mediation explanation assumes the parser anticipates structure. While this might seem reasonable when there is a single, plausible structure, in many cases there are several, very different, possible structures. Why would the parser risk error by projecting structure rather than waiting? Secondly, we will argue that the intuitive and experimental data on implicit arguments provides evidence for the direct interpretation of thematic roles without structural mediation. Both experiments provide evidence that argument structure is made available when a verb is recognized, and that this information can immediately influence filler-gap assignment. The first experiment demonstrated that an implausible filler will not initially be assigned to the direct object role unless there is no other argument position in the verb’s argument structure. The second experiment demonstrated that if a filler is implausible as direct object, but plausible as indirect object, it will be assigned the indirect object role immediately at the verb. What of the parallel proposal model we outlined in the previous section? We suggested that when a verb is recognized, all possible argument structures are made available in parallel, Many of the verbs in these experiments have several possible argument structures. See an example in (15), below. (15) a. The book reminded the girl of the movie.

b. The woman reminded the girl about the movie. c. The woman reminded the girl to watch the movie. d. Which movie did the woman remind ... When there are multiple argument structures associated with a verb, how can the argument structure information guide filler-gap assignment? We have just suggested that recognition of the verb makes available its entire argument structure and if the filler is not plausible as the first argument, it will be interpreted

Lexical Representation in Sentence Processing

353

as a later argument. However, the precise structure of the later argument is unimportant, because this is not a “core ambiguity” (see Tanenhaus and Carlson, 1989). In all reminding situations there is a “reminder,” a “remindee,” and the object or action brought to mind by the act of reminding. It is possible to develop a partial interpretation for the filler which is consistent with all possible argument structures. Thus, in (15d). the reader knows that the woman isn’t reminding the movie, but is reminding some person something about a movie. The nature of these partial interpretations is an important topic for further exploration. We have suggested that our experiments in this section demonstrate that readers have access to thematic role information associated with indirect objects and infinitive complements as soon as they recognize a verb. These results are somewhat surprising because indirect objects and infinitive complements are typically optional arguments. This is in contrast to verbs like put that require all of their arguments to be syntactically realized. One cannot say John put the cards, or John put in the drawer. Verbs that can take an indirect object or an infinitival complement often need not do so, as is illustrated in (16). (16) a. Mary served drinks (to her guests). b. Bill invited Tom (to go to the party). In sentences like these, an argument that is not realized syntactically seems to be “understood” or necessarily implied. For example, if Mary served drinks, then she necessarily served the drinks to someone. Likewise, if Bill invited Tom, he invited him to go somewhere or to do something. One way of understanding these intuitions is to assume that some thematic roles that are not realized syntactically are, in fact, realized in the conceptual representation of the sentence. In order to make this proposal more explicit, it is necessary to make some detailed assumptions about the thematic structure of different verbs, and how the thematic structure maps onto the verb’s subcategorizations. Some preliminary proposals were presented in Carlson and Tanenhaus (1988) and Tanenhaus and Carlson (1989). For the present purposes, it will be sufficient to point out some cases where we have argued that implicit arguments (or “open” thematic roles) can function as a bridge. The implicit arguments allow the reader or listener to quickly integrate aspects of the sentence being processed with information in the discourse. (17) After the library had its budget slashed... a. Bill donated some books. b. Bill donated some books to the library. (18) Tom nearly forgot to go to the bank, but fortunately... a. Bill reminded him. b. Bill reminded him to go to the bank

354

J.E. Boland and M.K.Tanenhaus

Whether or not an optional argument is syntactically realized depends, in part, on context. Contrast the (a) and (b) sentences in (17) and (18). In isolation, the (a) sentences (with the open thematic roles) are less natural because they feel incomplete. However, when they are placed in a context that introduces the “missing” information, the (a) versions actually become more natural than the complete versions, which now seem somewhat awkward because the information in the optional argument feels redundant. This intuition makes sense if we assume that recognition of a verb makes available the thematic structure of the verb, and that optional thematic roles can function much like anaphoric pronouns and look back to the context for an appropriate antecedent. If the role is thusly “filled” just after the verb is recognized, then the role would already be taken before the explicit argument is encountered, explaining why the (b) versions feel redundant in context. Of course, these conjectures need to be tested experimentally. Halpern (1990) has recently found’evidencethat missing complements are linked to their antecedents using a probe recognition task. Halpern had subjects read two-part discourses like the one shown in (19), below. The condition of interest is the (a) ending, where there is an implicit infinitive argument, to sweep thefloor. If the implicit argument truly acts like an anaphor, then subjects should respond just as quickly to sweep after (a) as they do after (b), where the anaphor is explicit. This is in fact what Halpern found: Probe recognition times were faster to the verb, sweep. after both (a) and (b) endings compared to (c). Studies investigating the time course of “null complement anaphora” are now in progress. (19) I don’t know who should sweep the floor though I bet a. Carol will volunteer reluctantly. b. Carol will volunteer to, reluctantly. c. Carol made the mess. Carlson and Tanenhaus (1988) have experimental evidence that some optional arguments that are not specified by either the context or the sentence are encoded in the form of an unspecified discourse entity or an “open thematic role.” In one experiment they examined comprehension times to sentences beginning with definite noun phrases (see 2Oc) in the context of a single sentence that did not introduce an explicit antecedent for the noun phrase (see 20a & b). As Haviland and Clark (1974) demonstrated, comprehension time is longer for sentences with definite noun phrases when the context does not establish an antecedent than when it docs because in the absence of an explicit antecedent the reader needs to make a “bridging inference.” In the context (20a) there is an implicit Theme, the thing being unloaded, which could provide an antecedent for the suitcase in the target sentence (20c). In contrast, the context (20b) creates a plausible situation for a suitcase but does not introduce a possible thematic role. Reading times for the target sentences were significantly faster

Lexical Representation in Sentence Processing

355

following (20a) than following (20b). (20) a. Bill hurried to unload his car. b. Bill humed to catch his plane. c. The suitcases were very heavy. Mauner (1990) is investigating similar integration effects in sentences with short passives. In a preliminary study she used sentences such as those in (21) and (22),where the correct interpretation of (22) is that the person who sank the ship (whoever that may be) did it to collect the insurance money. (21) a. The ship was sunk. b. The ship sank. (22) The D.A. said it was - to collect the insurance. While the passive form of (21a) implies that someone or something sunk the ship (i.e., there is an implicit Agent), no such implication exists for (21b). Thus, (22) is felicitous after (21a). but not (21b). This demonstrates the “reality” of the implicit argument. The anaphor, it, refers to the event in (21a) in which someone sank a ship. That someone can then be the subject of collect, even though there is no explicit Agent in the previous sentence. (22) doesn’t make sense after (21b) because it must refer to ship, which then must be the subject of collect. Mauner is currently exploring the hypothesis that the implicit agent can provide an antecedent for a definite noun phrase using target sentences such as The culprit wus never caught after contexts like (21). Taken together, these intuitive and experimental data on implicit arguments provide clear examples of thematic interpretation without structural rnediation. Remember, we said at the outset that evidence that syntactic structure had been proposed via lexical information (before it could be constructed heuristically from the input string) would be evidence against lexical filter models. While these results might be interpreted as showing just that, we would argue that the lexical filter vs. lexical proposal debate misses the boat. That debate revolves around structure building, and we have just argued that interpretation can actually precede structure-building in some circumstances, and in many cases implicit arguments are not realized syntactically. Verb Control We have shown that argument structure plays an important and immediate role in sentence interpretation. This should not be surprising; after all, argument structure defines the structure of the sentence. Other elements in the sentence are interpreted in terms of the situation denoted by the verb and they are as-

356

J.E. Boland and M.K.Tanenhaus

signed structural positions in accordance with its argument structure. But we have argued that all lexical information becomes available when a verb is recognized, regardless of whether or not it is structural in nature. Verb control is a type of combinatory information that is arguably not structural in nature. As explained previously, control specifies whether the verb’s subject or object is to be understood as the implicit subject of the infinitive clause. While this is ultimately necessary for a full interpretation of the sentence, it is not an assignment on which other structural decisions are based. Thus, there is no pressure on the processor to use control information quickly as was the case for argument structure information. Frazier et al. (1983) presented evidence that verb control information is initially ignored in filler-gap sentences. The result is important because it suggests that some combinatory lexical information is not available to the processing system during the structure-building stage. Frazier et al. argued that we fill gaps using the heuristic of Most Recent Filler (the MRF strategy), then check the filler-gap assignments when lexical information becomes a~ailable.~ In contrast, Boland, Tanenhaus, and Garnsey (in press), again using the embedded anomaly technique, found not only that verb control was used to make the correct assignments, but that it was available early so that the implicit subject was interpreted immediately. Our results demonstrated that, (1) coindexing decisions are not made before control information becomes available, and (2) that verb control information is available for use very early during sentence processing. (23) a. The cowboy signalled the outlaw - to surrender his weapons quietly. b. The cowboy signalled the horse - to surrender his weapons quietly. Compare sentences (23a) and (23b). Sentence (23b) is less plausible because horses cannot surrender. The subject gap has to have been correctly interpreted before people can notice this, because one must realize that the horse (and not the cowboy) is surrendering for the sentence to become implausible. The MRF strategy predicts that horse will be correctly interpreted as the subject of surrender. However, if horse is fronted, as in the Wh-question, Which horse did the cowboy signal to surrender his weapons quietly?, the MRF strategy predicts that cowboy will initially fill the subject gap since it is closer to it. Thus, the reader would be unaware of the oddity until the first-pass coindexing was corrected. In the present experiment, if verb control is used immediately there should be a plausibility effect for the object control sentences, but not the subject control sentences. This is because “implausible” fillers were implausible under an object control interpretation, but not under a subject control interpretation.

Lexical Representation in Sentence Processing

357

Consider the sentences in Table 2. “Obj-Implaus” and “Obj-Plaus” contain object control verbs: “Subj-Implaus” and “Subj-Plaus” contain subject control verbs. Otherwise the two sets are identical. However, “Obj-Implaus” is implausible because horses cannot surrender; “Subj-Implaus” is plausible because cowboys can. We predicted more “no” responses at surrender and the immediately following words in “Obj-Implaus” compared to “Obj-Plaus.” In contrast, we expected no plausibility effects for the subject control sentences. Obtaining these results would be evidence for the early use of verb control information. because the difference in plausibility between”0bj-Implaus” and “Subj-Implaus” is due completely to the control relations.

Table 2 Sample sentences from Boland, et al. (in press, b).

The sense-monitoring procedure was again used. As predicted, a plausibility effect was seen for the object control versions, but not the subject control versions. In contrast to the MRF strategy, verb control was clearly used to make the correct coindexing assignments. Further, there was no delay in the use of verb control information.I0Evidence against the MRF strategy has also been obtained in a recent series of experiments reported by Nicol and Osterhout (1989). They tested for the reactivation of fillers for subject and object gaps using a cross-modal lexical decision paradigm. Although they found priming for the fillers of the object gaps, they did not find consistent patterns of priming for the correct fillers of the subject gaps. Instead, they found some evidence that correct and incorrect fillers were both primed after subject gaps. Nicol and Osterhout concluded that the MRF strategy was not being used, but neither was control information available immediately. They argue that control information becomes available fairly late, and the subject gap remains unresolved until control information can be used. While this conclusion differs from ours, it is difficult to evaluate their conclusions in the absence of clear priming patterns. However as mentioned above. Mauner (1990) has provided compelling examples (see (21) and (22)) which suggest that subject gaps are filled very quickly using conceptual information. If on-line data confirm the intuitions, they will provide clear evidence for the rapid interpretation of implicit subjects as argued for by Boland et al. (in press).

J.E. Boland and M.K. Tanenhaus

358

4001

'

N

I 1

I

v 1

to

I

I

I

V2

V 2 t 1 V2+2 V2+3

Figure 3. Data from Roland et al.. in press-b. a) Cumulative percent of "no" responses at each word position. b) Mean nzading times for positive responses at each word position.

In conclusion, our results show that control information allows the processing system to interpret lexically empty subjects relatively easily and without error as long as the control information is unambiguous. These results complement other recent studies (Boland et al., 1989; Tanenhaus et al., 1989a; Trueswell et al.. 1990) demonstrating that the language processing system fully exploits the combinatory lexical information associated with verbs to rapidly coordinate a range of different types of information during language comprehension.

Lexical Representation in Sentence Processing

359

CONCLUSION We have seen that combinatory lexical information is indispensable to the language processing system. This fact is uncontroversial. However, much debate surrounds the question of how combinatory information can be used by the processing system, particularly the syntactic processing system (or the parser). Mitchell (1987) outlined two polar positions regarding the use of lexical information in parsing, lexical proposal and lexical filter. On the first position, lexical information is used by the parser to build structure, while on the second position, lexical information is used solely to check the parser’s output. While there is a lack of clear evidence supporting one position over the other, a body of evidence has emerged which demonstrates that combinatory information is used very rapidly during the comprehension process. One of the current challenges facing psycholinguists is to sort through conflicting data gathered via different techniques in order to better understand the psychological processes of language comprehension and how they can be measured, or fail to be measured, by various experimental methodologies. Whereas much emphasis in the literature has been placed on parsing, we would place the focus on the interpretation process as a whole. In fact, we have reported evidence that non-syntactic combinatory information (in at least some circumstances) allows the processing system to develop interpretations without any corresponding structure. We have suggested that recognition of a verb makes available all of its forms in parallel, along with their corresponding sets of lexical information. We have discussed a substantial body of research which supports the view that this information is exploited in the interpretation process. However, many questions remain. One unexplored issue concerns the parallel activation of multiple argument structures. Some indirect support for this idea is provided by Trueswell et al. (1990). However, a direct test is provided by Boland in some work in progress. Using naming and modified lexical decision” responses to proper name probes, she is comparing sentences in which the matrix verb has a sense ambiguity, one sense having a single internal argument (e.g., toss as in tossing a salad) and one sense having two internal arguments (e.g., tossing a ball to someone). Consider the following sentences: (24) a. Which salad did Bill toss?

b. Which ball did Bill toss? As for any ambiguous word, multiple senses of toss will be accessed and the context will be used to quickly select the appropriate sense. The argument structure for the “salad” sense of toss is saturated in the above example, but the “ball” sense contains an open thematic role, that of Recipient. Responses to a proper name probe ought to be faster if there is a thematic and syntactic position for a person in the sentence. Thus, if multiple argument structures are

360

J.E. Boland and M.K.Tanenhaus

activated along with the multiple senses, then subjects ought to respond just as quickly to the probe, John, after (24a) as they do after (24b). as long as the senses remain unresolved. Preliminary results suggest that integration effects are found in naming, but not modified lexical decision. If this line of research proves to be fruitful, it will be another step in detailing the process by which lexical information interfaces with the sentence processing system. In summary, the information made available by word recognition is of profound interest in the study of sentence comprehension. Psycholinguistic theories are held accountable to explain the rapid and seamless nature of language comprehension, and we have argued that combinatory lexical information is uniquely capable of facilitating this process. Because word recognition provides access to such a rich body of knowledge, it may act as a bridge between the sensory input and the syntactic and conceptual knowledge associated with interpretation (following Marslen-Wilson, 1989). However, research exploring the interface between lexical representations and syntactic, thematic, and conceptual processing has only just begun. Much work remains before detailed processing models can be developed which are explicit about time course issues, parallel activation, partial interpretation, the relative availability of various types of lexical information, and many other issues. Acknowledgments This research was supported in part by NSF grant BNS-8617738 and NIH grant HD-2227 1. Notes ‘We will describe a view of processing in which structural, thematic, and conceptual information from the lexical entry of the main verb play a central role in the interpretation of sentences. This may seem to fit SVO languages such as English, in which the matrix verb typically occurs quite early, but as Mitchell (1989) points out, it seems strange when you consider languages/ constructions in which the verb occurs late in the sentence. This issue has been explored by Frazier (1987b; Frazier & Flores d’Arcais, 1989) who found that speakers of Dutch, a predominantly verb-final language, made structural commitments prior to encountering the verb. If, as we will argue, the syntactic, thematic, and conceptual information provided by the verb can be used by the processing system, verb-final languages might appear to be at a disadvantage. However in any language, considerable interpretation can occur prior to encountering the verb. The more specific the interpretation prior to the verb, the better an idea the reader will have of the verb’s syntactic, thematic, and conceptual description, so that the information provided by the verb will do less work. The fact that information is sometimes unavailable in no way suggests that

Lexical Representation in Sentence Processing

36 1

when it is available, it will be ignored. It is true, a strategy which allowed no processing to take place prior to the appearance of a verb would put unbearable stress on one’s memory in a long, verb-final sentence. But no one, not even the most devout lexicalist, would assume that processing occurs prior to the verb (see Steedman & Altmann, 1989, for a discussion of the lexicalist canon.). Naturally, words will be grouped into noun phrases, prepositional phrases, sentential clauses, and the like. If there is case information available it will be used to specify structural relationships. We argue only that relevant information is used when it becomes available. *We treat “optional arguments” as distinct subcategorization frames. ’We are following the standard linguistic convention and distinguishing “true” infinitive clauses from purpose and rationale clauses such as “John read Proust to better himself.” Purpose and rationale clauses are characterized by an infinitive clause which could well be paraphrased as “for the purpose of ...” or “in order to ...” 4Gap-fillingimplies that the empty category is replaced by the filler noun phrase. Although that is a reasonable way to think about the filling of trace, that isn’t quite what we mean here. When we say the subject gap is interpreted or filled, we mean that it is indexed with the same conceptual entity as the filler noun phrase. The filler noun phrase retains its own thematic role and syntactic position, and subject gap has its own thematic role and syntactic position. ?Some of the intransitive verbs had transitive readings with a limited set of NPs. All of the verbs were strongly biased towards the intransitive reading. 6Most likely, recognition of the verb would make available all thematic structures and evaluation of the subject NP would rule out the Agent-Theme structure. However, it is possible that an inanimate NP in subject position prohibits (or alternatively, ranks as low) all thematic structures in which Agent corresponds to the noun in subject position ’This “cost” is the explanation Mitchell (1987) gave for the long reading times in the first segment of (4a). If the parser is serial, any structural revision should invoke some cost. It is possible to hypothesize a parser which constructs ranked structures in parallel (see Gorrell, 1989) which is consistent with the Minimal Attachment Strategy (i.e., [Subject [Verb Object]] is ranked above [Subject [Verb [Subject [Verb]]]], but both are constructed in parallel). This could minimize the cost associated with a garden path. and rapid detection (of the garden path) could minimize it still further. Thus, Stowe’s results could still be consistent with a ranked parallel version of Minimal Attachment, but such a version would make much weaker predictions than are currently popular. 8Thisproposal is similar to Pritchett (1988), but unlike Pritchett, we would argue that reference is made to the content of the roles. 9Accordingto Frazier et a]., this strategy is used for all kinds of gaps, but is particularly crucial for implicit subjects. Object gaps must be filled by a noun phrase in a non-argument position (such as a fronted, Wh-phrase), so those

362

J.E. Boland and M.K.Tanenhaus

fillers are often clearly marked. However, implicit subjects are coindexed with a noun phrase in an argument position, so both subject and object noun phrases are possible fillers. ‘OIn Experiment 1, Boland et al. (in press) compared the timing of plausibility effects for Wh-question and declarative versions of object-control sentences. (In declarative versions, the most recent filler is the correct one, whereas in Wh-questions it is not.) They found no evidence that it took more time to fill a gap with a distant filler than with a recent filler. Plausibility effects were seen at the first point at which the sentences with implausible fillers could become implausible for both types of sentences. The timing of the plausibililty effects in the current experiment was consistent with the timing of those in Boland et al., Experiment 1. ”Subjects were told to respond “yes” if they recognized the string as a common first name, “no” if they did not.

References Altmann. G.T.M., (1989). Parsing and interpretation: An introduction. Language and Cognitive Processes, 4 , 1-19 Bach, E. (1979). Control in Montague grammar. Linguistic Inquiry, 10, 515531. Bach. E. & Partee, B. (1980). Anaphora and semantic structure. Papers from the parasession on pronouns and anaphora. Chicago, IL: Chicago Linguistic Society. Bates, E. & MacWhinney, B. (1982). Functionalist approaches to language acquisition. In Wanner, E. & Gleitman, L.R. (Eds.), Language acquisition: The state of the art. Cambridge: Cambridge University Press. Bever, T.G. (1970). The cognitive basis for linguistic structures. In J.R. Hayes (Ed.), Cognition and the development of language. New York: Wiley. Boland, J.E., Tanenhaus, M.K., Carlson, G., & Garnsey, S.M. (1989). Lexical projection and the interaction of syntax and semantics in parsing. Journal of Psycholinguistic Research, I8, 563-576. Boland, J.E., Tanenhaus, M.K.,& Garnsey, S.M. (In press). Evidence for the immediate use of verb control information in sentence processing. Journal of Memory and Language. Burgess, R.C. (1990). The role of lexical and semantic context in syntactic disambiguation. Unpublished doctoral dissertation, University of Rochester. Carlson, G.N., & Tanenhaus, M.K. (1988). Thematic roles and language comprehension. In W. Wilkens (Ed.), Syntax and semantics, Vol. 21. New York: Academic Press. Chierchia, G. (1984). Topics in the syntax and semantics of infinitives and gerunds. Unpublished doctoral dissertation, University of Massachusetts.

Lexical Representation in Sentence Processing

363

Clifton, C., Jr., & Frazier, L. (1988). Comprehending sentences with longdistance dependencies. In G.N. Carlson & M.K.Tanenhaus (Eds.), Linguistic structure in language processing. Dordrecht: Kluver Academic. Clifton, C., Jr., Frazier, L, & Connine, C. (1984). Lexical expectations in sentence comprehension. Journal of Verbal Learning and Verbal Behavior, 23, 696-708. Comrie, B. (1984). Subject and object control: Syntax, semantics, pragmatics. Berkeley Linguistics Society, 10,450-464. Crain, S., & Fodor, J.D. (1985). How can grammars help parsers? In D.R. Dowty, L. Kartunnen, & A.M. Zwicky (Eds.), Natural language processing: Psychological, computational, and theoretical perspectives. New York: Cambridge University Press. Dowty, D. R. (1988). Thematic proto roles, subject selection, and lexical semantic defaults. Unpublished manuscript. Farkas, D (1988). On obligatory control. Linguistics and Philosophy, 11, 27-58. Ferreira, F, & Clifton, C., Jr. (1986). The independence of syntactic processing. Journal of Memory and Language, 25, 348-368. Ferreira, F. & Henderson, J.M. (In press). The use of verb information in syntactic parsing: Evidence from eye movements and word-by-word selfpaced reading. Journal of Experimental Psychology: Learning, Memory, and Cognition.

Fisher, C., Gleitman, H., & Gleiunan, L.R. (1989). Relations between verb syntax and verb semantics: On [he semantic content of subcategorization frames. Submitted for publication. Fodor, J.A. (1983). Modularity of mind. Cambridge, MA: MIT Press. Fodor, J.A., Bever, T.G., & Garrett, M.F. (1974). The psychology of language: An introduction 10 psycholinguistics and generative grammar. New York: McGraw-Hi11. Forster, K. (1979). Levels of processing and the structure of the language processor. In W.E. Cooper & E.C.T. Walker (Eds.), Sentence processing: Psycholinguistic studies presented to Merrill Garrett. Hillsdale, NJ: Erlbaum. Frazier, L. (1 978). On comprehending sentences: Syntactic parsing sfrategies. Unpublished doctoral dissertation, University of Connecticut. Distributed by the Indiana University Linguistics Club). Frazier, L. (1987a). Theories of syntactic processing. In J.L. Garfield (Ed.), Modularity in knowledge representation and natural language processing.

Cambridge, MA: MIT Press. Frazier, L. (1987b). Syntactic processing: Evidence from Dutch. Natural Language and Linguistic Theory, 5 , 519-560. Frazier, L. (1989). Against lexical generation of syntax. In W.D. MarslenWilson (Ed.), Lexicul representation and process. Cambridge, MA: MIT Press. Frazier, L., & Clifton, C., Jr. (1989). Successive cyclicity in the grammar and

364

J.E. Boland a d M.K. Tanenhaus

the parser. Language and Cognitive Processes, 4 , 93-126. Frazier, L., Clifton, C., Jr., & Randall, J. (1983). Filling gaps: Decision principles and structure in sentence comprehension. Cognition, 13, 187-222. Frazier, L. & Flores d’Arcais, G.B. (1989). Filler driven parsing: A study of gap filling in Dutch. Journal of Memory and Language, 28,331-344. Frazier, L. & Fodor, J.D. (1978). The sausage machine: A new two-stage parsing model. Cognition, 6, 291-325. Garnsey, S.M., Tanenhaus, M.K., & Chapman, R.M. (1989a). Evoked potentials and the study of sentence comprehension. Journal of Psycholinguistic Research, 18. 51-60. Garnsey, S.M.,Tanenhaus, M.K., & Chapman, R.M. (1989b, November). Preferred verb argument structure in sentence comprehension: An ERP study. Paper presented at the annual meeting of the Psychonomic Society, Atlanta, GA. Gorrell, P. (1989). Establishing the loci of serial and parallel effects in syntactic processing. Journal of Psycholinguistic Research, 18, 61-74. Halpern, A. (1990). Priming in VP anaphora. Unpublished manuscript. Haviland, S. E., & Clark, H. H. (1974). What’s new? Acquiring new information as a process in comprehension. Journal of Verbal Learning and Verbal Behavior, 13,512-521. Jackendoff. R. (1974). A deep structure projection rule. Linguistic Inquiry, 5 , 481-506. Kurtzman, H.S. (1989). Locating Wh-traces. In C. Tenny (Ed.), The MIT parsing volume, 1988-89. Ladusay, W. A. (1988). Towards a non-grammatical account of thcmatic roles. In W. Wilkens (Ed.). Syntax and semantics 21: Thematic roles. New York: Academic Press. Marslen-Wilson, W.D. (1989). Access and integration: Projecting sound onto meaning. In W.D. Marslen-Wilson (Ed.), Lexical representation and process. Cambridge, MA: MIT Press. Marslen-Wilson, W.D. (1987). Functional paral!elism in spoken word recognition. Cognition, 25, 71-102. Marslen-Wilson, W.D., Brown, C., & Tyler. L.K. (1988). Lexical representations in language comprehension. Language and Cognitive Processes, 3, 1- 16. Marslen-Wilson, W.D. & Tyler, L.K. (1987). Against modularity. In Garfield, J.L. (Ed.), Modularity in knowledge representation and natural language processing. Cambridge, MA: MIT Press. Mauner, G. (1990). Implicit arguments. Informal presentation, University of Rochester. Mitchell, D.C. (1987). Lexical guidance in human parsing: Locus and processing characteristics. In M. Coltheart (Ed.), Attention and performance X I I : The psychology of reading. Hillsdale, NJ: Erlbaum.

Lexical Representation in Sentence Processing

365

Mitchell, D.C. (1989). Verb-guidance and other lexical effects in parsing. Language and Cognitive Processes, 4 , 123-154. McClelland, J.L. (1987). The case for interactionism in language processing. In M. Coltheart (Ed.), Attention and performance X U : The psychology of reading. Hillsdale, NJ: Erlbaum. McClelland, J.L., St. John, M., & Taraban, R. (1989). Sentence comprehension: A parallel distributed processing approach. Language and Cognitive Processes, 4 , 287-335. Nicol, J & Osterhout, L. (1989). Re-activating antecedents of empty categories during parsing. Manuscript submitted for publication. Nicol. J. & Swinney, D. (1989). The role of structure in coreference assignment during sentence comprehension.Journal of Psycholinguistic Research, 18, 5-20. Pritchett, B. L. (1988). Garden path phenomena and the grammatical basis of language processing. Language, 64,539-576. Ruzicka, R. (1983). Remarks on control. Linguistic Inquiry, 11, 97-102. Sag, LA. & Pollard (In press). A semantic theory of obligatory control. Language. Shapiro, L.P., Zurif, E., & Grimshaw, J. (1987). Sentence processing and the mental representation of verbs. Cognition, 27, 219-246. Shapiro. L.P., Zurif, E.B., & Grimshaw, J. (1989). Verb processing during sentence comprehension: Contextual impenetrability. Journal of Psycholinguistic Research, 18, 223-243. Steedman, M. & Altmann, G. (1989). Ambiguity in context: A reply. Language and Cognitive Processes. 4 , 105-122. Stowe, L.A. (1986). Parsing WH-constructions: Evidence for on-line gap location. Language and Cognitive Processes, I , 227-245. Stowe, L.A. (1989). Thematic suuctures and sentence comprehension. In G.N. Carlson & M.K. Tanenhaus (Eds.), Linguistic structure in language processing. Dordrecht: Kluwer Academic. Stowe, L.A. & Holmes, V.M. (1989). Verbal expectation and late closure: The nature of verb information in ambiguity resolution. Unpublished manuscript. Swinney, D., Ford, M., Frauenfelder, U., & Bresnan, J. (1988). On the temporal course of gap-filling and antecedent assignment during sentence comprehension. In B. Grosz, R. Kaplan, M. Macken, & I. Sag (Eds.), Language structure and processing. Stanford, CA: CSLI. Tanenhaus, M.K., Boland, J.E., Garnsey, S.M., & Carlson, G.N. (1989a). Lexical structure in parsing long-distance dependencies. Journal of Psycholinguistic Research, 18, 37-50. Tanenhaus, M.K. & Carlson, G.N. (1989). Lexical structure and language comprehension. In W.D. Marslen-Wilson (Ed.), Lexical representation and process. Cambridge, MA: MIT Press.

366

J.E. Boland and M.K.Tanenhaus

Tanenhaus, M.K., Carlson, G.N., & Seidenberg, M.S.(1985). Do listeners compute linguistic representations? In D.R. Dowty, L. Kartunnen, & A.M. Zwicky (Eds.), Natural language processing: Psychological, computational, and theoretical perspectives. New York: Cambridge University Press. Tanenhaus, M.K., Carlson, G.N., & Trueswell, J.C. (1989b). The role of thematic structures in interpretation and parsing. Language and Cognitive Processes, 4,211-234. Tanenhaus, M.K., Gamsey, S.M., & Boland, J.E. (1990). Argument-structure and the parsing of sentences with long-distance dependencies. Manuscript in preparation. Tanenhaus, M.K., Stowe, L.A., & Carlson, G. (1985). The interaction of lexical expectation and pragmatics in parsing filler-gap constructions. Proceedings of the seventh annual cognitive science society meetings. Taraban, R. & McClelland, J.L. (1988). Constituent attachment and thematic role assignment in sentence processing: Influences of content-based expectations. Journal of Memory and Language, 27,597-632. Trueswell, J., Tanenhaus, M.K., & Garnsey, S.M.(1990). Semantic influences on parsing: Use af thematic role information in syntactic disambiguation. Unpublished manuscript.

Understanding Word and Sentence G.B. Simpson (Editor) 0 Elsevier Science I’ublishers B.V. (North-Holland), 1991

Chapter 14 The Resolution of Interdeterminacy During Language Comprehension: Perspectives on Modularity in Lexical, Structural and Pragmatic Process David A. Swinney

The Graduate Center of the City University of New York New York, New York U.S.A.

The resolution of indeterminacy is the hallmark of the language comprehension process. Indeterminacy exists at essentially every descriptive level of word, sentence, and discourse analysis. It is found in conditions ranging from overt ambiguity (such as for polysemous words and structural ambiguity) which have been the focus of work in psycholinguistics since the inception of the field, to less obvious, but equally vexatious, problems involving for such things as phonetic identity, segmentation assignment, thematic role assignment, quantifier scope, and co-reference assignment, to name but a few. The problem is so pervasive that about the only generalization that appears possible is that the language processing device s e e m to resolve these uncertainties. Interestingly, models of language theory (such as linguistic grammars) are traditionally not expressed in terms that capture many of the problems of indeterminacy; uncertainties of this type are considered to be a problem to be dealt with by processing models, something viewed as performance perturbations ovcr which generalizations are to be made, not facts which are to be captured in models of ‘universal grammar’.’ The immediate focus of this chapter, then, will be on an issue central to processing models: that of how the language processor resolves several types of indeterminacy during comprehension. In what follows, evidence concerning the resolution of co-reference assignment (an issue that involves structural and semantic information processing) and the resolution of lexical ambiguity will be examined. In each case, recent evidence examining the role of different types of ‘context’ on the resolution of these types of indeterminacy will be examined in detail. Prior to presentation of such evidence, it is important to briefly discuss a fundamental issue that underlies the concern over indeterminacy resolution, the issue of the general nature of the cognitive architecture that supports language processing. One of the central contributions of the information processing ap-

368

D . Swinney

proach in cognition to the study of language has been the development of models of processes (and representations over which they operate) that are couched in terms of the nature of the temporal interaction of different information types during comprehension and production. Out of this approach have emerged two general models which, in their extreme, are totally apposite. In one, the Interactive Model (see, e.g., Marslen-Wilson & Tyler, 1987, for a modified version), any one type of information (say, for example, world knowledge) is deemed available to any other information type (say, for example, lexical representation) at any time during processing. That is, any type of information can be used to help constrain the processing of other types of information as soon as it is useful. The other, the Modularity Model (Fodor, 1983). holds that there are principled points of interaction between information sources and only at those points may different information types have access to each other. This modularity model has as an important component the concept that many processes are autonomous - that is, their internal operations are never affected by information from other sources. Rather, it is only upon completion of the internal operations of such an autonomous ‘module’ that the results of that process are available for use by other processes. In the study of language processing, examination of the effects of one putative information source upon the processing of another has become the empirical testing ground for the interactivity vs. modularity debate. An important issue related to the facts of methodology in this field follows from this: It has become axiomatic that only on-line (sometimes called real-time) methodologies are capable of making the fine-grain temporal distinctions that have come to represent acceptable tests capable of distinguishing these theories. The problem is, of course, that ‘on-line’ is not a categorical notion. Rather, tasks are only relatively more or less sensitive to the temporal course of information interaction. Two task that appear to ‘tap’ processing of a sentence at the same point in time may actually differ by a few hundredths of a second in terms of what their own task demands add to the process, and that difference may be sufficient to (in one case) capture a process or (in the other) just ‘miss’ capturing the process. The point here is that all ‘on-line’ tasks are not equal, and much of whether one believes that a task has ‘demonstrated’ interactivity or modularity relies on careful evaluation of whether or not that particular task is demonstrably sensitive to both the processing effects under investigation and their (exceedingly rapid) time-courses. In what follows, two processes - co-reference assignment and lexical access - are examined for evidence of the effects of a number of types of semantic and pragmatic contexts upon their operation. A task is used that has been demonstrated to be sensitive to the results of such processes and, as will be seen, the evidence leads to an overall conclusion that both of these processes are autonomous, independent subprocesses of the language comprehension system.

The Resolution of Indeterminancy

369

ASSIGNMENT OF CO-REFERENCE DURINGCOMPREHENSION One area of indeterminacy resolution that has particular significance for both ‘competence’ and ‘processing’ models of language lies in the area of coreference. This is the manner in which a referentially-dependentitem (such as a pronoun, reflexive, or ‘empty category’ element is linked to it’s referent). For example, in the sentence: When the butcher saw the baker he was quite intoxicated. the exact antecedent (referent) assigned to the word ‘he’ is indeterminate. However in order to understand the sentence (i.e., to have a meaning assigned to it) the listener must eventually assign one or the other of the two possible antecedents to the word ‘he’. Now, in this particular case, the assignment must be made on pragmatic or strategic grounds: If the listener knows something about this butcher or this baker, or butchers or bakers in general, he/she may be able to decide which one is more likely to be intoxicated. However, if such pragmatic knowledge is unavailable, the listener may simply use strategies (such as: The first-mentioned person is the likely sentence topic, and hence the likely subject of the main clause). Indeterminacy resolution for cases such as this involving overt pronouns is an interesting and important topic for study. However, there is an even slightly more intriguing (and structurally parallel) case of co-reference resolution, that associated with empty categories, which will serve as the focus of inquiry here. Empty categories are sentence constituents which have no phonological, acoustic or orthographic realization. They are theoretical entities which have roughly been a part of generative grammar at least since the appearance of Aspects of a Theory of Synfax (Chomsky, 1965; see Fodor, 1989, for an excellent extensive review of the empty category issue). They are entities postulated to exist when deletions or movement cause some piece of a sentence to be missing from its canonical position. Thus, for example, according to some approaches, the sentence:

Which book, did Ed read (e,)? derives from movement of a wh-phrase (‘which book’) that is the object of the verb ‘read’ in the canonical (underlying) sentence : Ed read which book? According to the theory, the wh-phrase is moved to the initial position in the sentence, and a ‘trace’ is left in the original position it was moved from. This trace is an empty category (e). It is important to add that, again according to

370

D . Swinney

some theories, co-indexation between the moved wh-phrase and the trace is what allows the correct assignment of the thematic role of object (patient) to the moved phrase; the thematic role is assigned to the position the trace occupies, and it must be somehow linked to the moved phrase - hence co-indexation. Note that it is not critical that this analysis be described in terms of current generative theory, as has been done above; many other theories could in principle accommodate an account of the apparent displacement of an object from its language-canonical position to some other place in a sentence. However, it is only the current generative models which have made the issue of empty categories a primary focus of linguistic inquiry. Hence, given that it is in terms of ‘empty categories’ that this concept has been most comprehensively elucidated, it makes sense to utilize this descriptive mechanism here. The processing question that is raised by empty categories (and, for now, the focus will be on wh-trace), is how the antecedent to that trace is linked to the empty category during comprehension; how co-reference is accomplished between the antecedent (which book) and the trace (e) during actual processing. There are a number of simple, preliminary hypotheses that one might make about how such relationships are established. To begin, one could imagine that while co-reference is established for overt pronouns, it may not be for empty categories; they may just be convenient tools of a linguist to keep grammatical theory consistent. Thus, under this model, there is no indeterminacy and thus one would expect to find no co-reference processing necessary. A second hypothesis is that the processor does nothing when some reference-seeking item is first encountered (be it a pronoun or an empty category), but rather simply waits until such relationships are clearly and uniquely disambiguated by later information. This Delayed Determination model has the apparent advantage of fitting with the fact that we aren’t typically aware of linking a pronoun to some antecedent when we first hear it, and to the possibility that if we wait long enough, context may provide some help in the assignment process. The final hypothesis that will be considered is one which holds that the comprehension device cannot delay establishing co-referential relationships, because they help constrain other on-going processes, and hence the system will do everything it can to establish co-reference immediately. There are several reasonable variants of this Immediate Determination Model. One, something that might be called the weakly-constrained variant, holds that the device activates all possible antecedents (e.g., all prior noun phrases in the immediate discourse) in order to then sift through them to find the correct antecedent to a pronoun or an empty category. A more strongly constrained variant would allow linguistic principles to constrain the choice of the co-referent ‘on-line’; failure of this to narrow the field to a single possibility might require guessing based on psychological processing principles (such as, for example, Frazier. Clifton & Randall’s, 1983, “most recent filler” hypothesis). Happily, there is at this point sufficient evidence to begin to choose among

The Resolution of Indeterminancy

37 1

these alternatives. A number of researchers, using different techniques, have provided evidence that for object-relative wh-traces the processing device appears to immediately re-activate the co-referent at the point where the trace appears; indeterminacy is resolved immediately. For example, Bever and McElree (1988) have used a probe-recognition task whereby they found that recognition of a probe word from an antecedent noun phrase (NP) is faster (with lower error rates) in cases theoretically containing these ‘empty’ whtraces than in control conditions with no wh-trace. In another study, Tanenhaus, Carlson, and Seidenberg (1985) have demonstrated that lexical decisions for words phonologically related to the antecedent are faster than those to a control at a point shortly after the ‘gap’ is encountered in a sentence. (See also recent papers by Tanenhaus, Boland, Garnsey & Carlson, 1989; Garnsey, Tanenhaus & Chapman, 1989; Nicol & Swinney, 1989, among others, for relevant findings.) One on-line study of co-reference resolution in wh-trace conditions by Ford, Frauenfelder, Bresnan and Swinney (1984; reported in Swinney, Ford, Bresnan, & Frauenfelder, 1988) will be presented here in a bit more detail, both because it employs one of the more temporally sensitive on-line tasks for examining this issue (cross-modal lexical priming), and because it leads directly to our current focus - the examination of the effects of contextual information on this process. In the Ford et al. study, subjects were presented auditorily (only the words were heard) with sentences such as: The policeman saw the boy that the crowd at the party *2 of the crime.

*, accused ( e )

‘*’ a string of letters appeared on a computer screen in front of the subjects and they made lexical decisions to those letter strings. In the critical conditions, those letter strings were semantic associates to a potential antecedent from the sentence. All conditions were counterbalanced, so that any one subject saw only one possible associate in only one of the two test positions. The logic of the experiment is that, if reaction time to make a lexical decision to a word related to the appropriate antecedent (coreferent) to the trace is faster than reaction time to an unrelated, but otherwise equivalent, control word, that ‘priming’ result constitutes evidence that reactivation of that antecedent must have taken place at that point in time. Note that because there is no actual physical stimulus present at the ‘trace’ (this is also called a ‘gap’) such an outcome would be a particularly striking example of evidence for the psychological reality of these ‘traces’. The results are easily summarized. When the letter string was related to one of the inappropriate potential antecedents (e.g., CROWD),reaction time to decide it was a word was no faster at test point *z than at test point *,.z However, when the probe word was related to the appropriate antecedent to the trace (in this case: BOY), priming was significantly larger at test point *2 (precisely where the trace is posited to exist) than after test point *,. In a follow-up experiment (Swirmey, At the points indicated by the

372

D . Swinney

Ford, Bresnan, Frauenfelder, 1986; reported in Swinney et al., 1988). the same materials were used with a so-called ‘naming’ task wherein reaction time for subjects to ‘say out loud’ the word that appears on the computer screen was made. Identical results were obtained with this naming paradigm as in the lexical decision paradigm. Thus, it was concluded that the appropriate co-referent was re-activated immediately after the verb (at the point where the trace is hypothesized to exist). Interestingly, there was no evidence that the appropriate co-referent was simply activated and kept active until the gap (trace) occurred in the sentence; the associate to that co-referent was not primed at test point *,). Equally interesting is the finding that other ‘potential’ co-referents to the trace (such as CROWD) are not activated at the gap. Rather, it appears that the strongly-constrained version of the Immediate Determination Hypothesis described above seems to hold for wh-trace processing; at the point the trace is posited (after the verb) the appropriate antecedent is calculated by the processing device? Under any interpretation, the immediate reactivation of the co-referent of the trace in these constructions constitutes one of the few documented examples of an on-line structural process in operation. That is, the reactivation of the antecedent of a trace can only be needed for structural processing (or, perhaps, the vague shadowy area combining structural and semantic processing), and it is one of the few such automatic, perceptual process in structural processing for which we have any evidence. As such, it makes an interesting testing ground for the issues of indeterminacy raised above, those concerning precisely what conditions control the discovery and activation of the appropriate co-referent to the empty category. Couched in terms of the interactivity/modularity hypotheses, the question becomes one of whether information that is external to structural processing can affect the gap-filling procedure. Under the modularity hypothesis, information external to a putative structural processor will not be able to affect the gap-filling process. However, the intcractionists would hold that, quite the opposite, any pragmatic or world-knowledge information that indicates the identity of the appropriate co-referent will aid in gap-filling. The two experiments reported below examine this controversy in this unique structural processing condition. In the first, plausibility of the potential co-referents is manipulated to determine whether world-knowledge (in the form of plausibility) can control the gap-filling process. In this, 86 subjects from Rutgers University were subjected to 106 auditory presented sentences, 48 of which represented sentence versions schematized in the following example: The crowd looked at the enormous heavyweight boxer that the small 12-year old boy on the comer had [beaten (e)* so badly /hugged (e)* so intensely] a few minutes earlier.

The Resolution of Indeterminancy

373

Subjects were required to understand the sentences and to make a naming responses to words that were visually displayed at the test points (*). Again, no subject saw or heard more than one variant of the materials: complete counterbalancing was performed. The visual words were either the pragmatically likely antecedent (BOXER for the verb ‘BEATEN’) or the syntactically correct antecedent (BOY), or ‘control’ words matched for a priori naming reaction times with each of the experimental words, but semantically unrelated to anything in the sentence. Note that a control verb was used (HUGGED in this example) for which either antecedent could have been pragmatically plausible. The logic of this experiment is that if plausibility can “penetrate” (control the internal operations of) a putatively modular syntactic process such as co-reference assignment, then we will find that following the verb ‘BEATEN’ there is ‘priming’ for the word BOXER even though it is not the structurally appropriate antecedent. If, on the other hand, modularity does exist for the structural processing device, then such plausibility information should have no such effect; one should only find priming for the correct co-referent (BOY) in both the context of the implausible verb (BEATEN) and the neutral verb (HUGGED). Reaction time to name the words in the conditions described above were collected, and are reported here in terms of ‘priming scores’ (reaction time to control minus related word). When the control verb HUGGED was used, there was a 31 millisecond priming effect for BOXER (pc.05). but only a non-significant 2 milliseconds effect for BOY at the test point. When the verb BEATEN was in the sentence, there was a 29 millisecond priming effect for BOXER (p