Text representation: linguistic and psycholinguistic aspects

Text Representation is a forum for interdisciplinary research on the nature and organization of the cognitive systems...

Author: Ted Sanders | Joost Schilperoord | Wilbert Spooren

202 downloads 835 Views 8MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

Text Representation

is a forum for interdisciplinary research on the nature and organization of the cognitive systems and processes involved in speaking and understanding natural language (including sign language), and their relationship to other domains of human cognition, including general conceptual or knowledge systems and processes (the language and thought issue), and other perceptual or behavioral systems such as vision and nonverbal behavior (e.g. gesture). 'Cognition' should be taken broadly, not only including the domain of rationality, but also dimensions such as emotion and the unconscious. The series is open to any type of approach to the above questions (methodologically and theoretically) and to research from any discipline, including (but not restricted to) different branches of psychology, artificial intelligence and computer science, cognitive anthropology, linguistics, philosophy and neuroscience. It takes a special interest in research crossing the boundaries of these disciplines. HUMAN COGNITIVE PROCESSING

Editors Marcelo Dascal, Tel Aviv University Raymond W. Gibbs, University of California at Santa Cruz Jan Nuyts, University ofAntwerp

Editorial address Jan Nuyts, University of Antwerp, Dept. of Linguistics (GER), Universiteitsplein 1, B 2610 Wilrijk, Belgium. E-mail: [email protected] Editorial Advisory Board Melissa Bowerman, Nijmegen; Wallace Chafe, Santa Barbara, CA; Philip R. Cohen, Portltmd, OR; Antonio Damasio, Iowa City, IA; Morton Ann Gernsbacher, Madison, WI; David McNeill, Chicago, IL; Eric Pederson, Eugene, OR; Fran~ois Recanati, Paris; Sally Rice, Edmonton, Alberta; Benny Shanon, Jerusalem; Lokendra Shastri, Berkeley. CA; Dan Slobin, Berkeley, CA; Paul Thagard, Waterloo, Ontario

Volume8 Text Representation: Linguistic and psycholinguistic aspects Edited by Ted Sanders, Joost schilperoord and Wilbert Spooren

Text Representation Linguistic and psycholinguistic aspects Edited by Ted Sanders University of Utrecht

Joost Schilperoord University ofTilburg

Wilbert Spooren Free University, Amsterdam

John Benjamins Publishing Company Amsterdam I Philadelphia

,(;)i;j~

\ _ __/

The paper used in this publication meets the minimum requirements of American National Standard for Information Sciences - Permanence of Paper for Printed Library Materials, ANSI z39.48-1984.

Library of Congress Cataloging-in-Publication Data Text Representation : Linguistic and psycholinguistic aspects I edited by Ted Sanders, Joost Schilperoord, Wilbert Spooren. p. em. (Human Cognitive Processing, ISSN 1387-6724; v. 8) Based on papers presented at a conference held July 1997, Utrecht University. Includes bibliographical references and indeL l. Discourse analysis--Psychological aspects--Congresses. I. Sanders, Ted. II. Schilperoord, Joost. III. Spooren, Wibert. IV. Series. P302.8.T49 2001 401 ~41--dc21 ISBN 90 272 2360 2 (Eur.) /1 58811 077 X {US) {Hb; alk. paper)

2001035506

C 2001- John Benjamins B.V. No part of this book may be reproduced in any form, by print, photoprint, microfilm, or any other means, without written permission from the publisher. John Benja.mins Publishing Co.· P.O. Box 36224 · 1020 ME Amsterdam· The Netherlands John Benja.mins North America · P.O. Box 27519 · Philadelphia PA 19118-0519 · usA

Table of Contents

Preface 1.

vu

Text representation as an interface between language and its users Ted Sanders & Wilbert Spooren 1. ACCESSIBIUTY IN TEXT AND Accessibility theory: an overview 29

SECTION

2.

TEXT PROCESSING

1

27

Mira Ariel 3·

4

The influence of text cues on the allocation of attention during reading 89 Michelle L Gaddy, Paul van den Broek & Yung-Chi Sung Lexical access in text production: On the role of salience in metaphor resonance 111 Rachel Giora & Noga Balaban 2. RELATIONAL COHERENCE IN TEXT AND TEXT PROCESSING Semantic and Pragmatic relations and their intended effects 12.7

SECTION

5.

Alistair Knott 6.

7.

On the production of causal-contrastive although sentences in context 153 Leo G. M. Noordman Beyond elaboration: The interaction of relations and focus in coherent text 181

Alistair Knott, ]on Oberlander, Michael O'Donnell, Chris Mellish 8. Unstressed en I and as a marker of joint relevance 197 Henk Pander Maat 9.

Argumentation, explanation and causality: an exploration of current linguistic approaches to textual relations 2.31

Francisca Snoeck Henkemans

12.5

V1

Text Representation

SECTION

3.

fROM TEXT REPRESENTATION TO KNOWLEDGE

247 Constructing inferences and relations during text comprehension

REPRESENTATION

10.

249

Arthur C. Graesser, Peter Wiemer-Hastings & Katja Wiemer-Hastings IL

Thinking about bodies of knowledge: Tests of a model for predicting thoughts 273 Bruce K. Britton, Peter Schaeffer, Michael Bryan, Stacy Silverman &

Robert Sorrells SECTION

4.

SEGMENTATION IN TEXT AND TEXT REPRESENTATION

307

12. Conceptual and linguistic processes in text production; interactive or autonomous? 309

Joost Schilperoord 13. Subordination and discourse segmentation revisited, or: Why matrix clauses may be more dependent than complements 337

Arie Verhagen Subject index 359

Preface

The chapters of this book volume are all based on papers presented at the

International workshop on text representation: Linguistic and psycholinguistic aspects, held at Utrecht University in July 1997. We are indebted to the Centre for Language and Communication (CLC) and the Faculty of Arts of Utrecht University, the Center for Language Studies (CLS) and the Discourse Studies Group of Tilburg University, the Netherlands Organization for Scientific Research NWO and the Royal Academy of Sciences KNAW for their financial support of this workshop and to the Utrecht Institute of Linguistics UIL OTS (which now includes CLC) for logistic and financial support in the preparation of the manuscript. We are very grateful to Gerben Mulder, Annelous van Rijn and Martien Kroon for their invaluable help in editing this manuscript. We would also like to thank the many colleagues who acted as reviewers of the chapters: Yves Bestgen Diane Blakemore Lucille Chanquoy Liesbeth Degand Jack DuBois Charles Fletcher Bruce Fraser David Galbraith Morton Ann Gemsbacher Paul van den Hoven EdHovy

Jan van Kuppevelt Fons Maes Bonny Meyer Joanna Moore Gisela Redeker Manfred Stede Mark Torrance Carel Van Wijk Luuk Van Waes Jos van Berkum

Finally, we want to thank the editors of the Benjamins series Human Cognitive Processing, especially Jan N uyts, for publishing this book volume in their series. Ted Sanders, Joost Schilperoord, Wilbert Spooren 's-Hertogenbosch I Utrecht, December 2000

CHAPTER

1

Text representation as an interface between language and its users Ted Sanders Wilbert Spooren University of Utrecht/Free University, Amsterdam

1.

From meaning to processes, from sentence to discourse

The theme of this volume is text representation, or more specifically the linguistic and psycholinguistic aspects thereof. In our view, a text representation is a cognitive entity: a mental construct that plays a crucial role in both text production and text understanding. In text production it is the basis for lexical retrieval and for producing and combining the discourse units. In text understanding it is the result of the decoding of the linguistic information in a discourse. This book characterizes a field of study in which the two disciplines, linguistics and psycholinguistics, are growing together. Traditionally, the linguists' task was to connect linguistic forms with meanings. This was usually done within a simplex model of communication, in which isolated sentences were connected to interpretations. Gradually the insight grew that such an enterprise cannot be undertaken without concern for the language user that produces and interprets an utterance. Consequently, the cognitive representations and actions of producers and interpreters entered the linguistic model. This development was stimulated by the rise of cognitive psychology in the early 1960s, at the expense of behaviorism. Cognitivists regarded the cognitive processes proper as their object to be explained, rather than the product of such processes, that is, the behavior in which these processes resulted. And since the end of the 1970s the view on linguistic communication has been extended in such a way that not only sentences, but also extended discourse is the object of study. We can describe the present situation in the following schematic view on communication through language, although we realize that it is an oversimpli-

~

Ted Sanders and Wilbert Spooren

~ 0

Q,o 0

• Figure 1. Communication through written text: From cognitive representation to text, to cognitive representation (as seen by Laura van Beek, 9 years old)

fied model, that in fact reflects the 'conduite metaphor of communication' (Reddy 1979). In this view, there is a producer who has a cognitive representation of what she intends to communicate; this is formulated in a linguistic code, called the text, and this text is decoded by the interpreter who can be said to understand a text once he has made a coherent representation ofit. This view fits theories that describe the link between the structure of a text as a linguistic object, its cognitive representation and the processes of text production and understanding. This schematic view also shows very clearly that research on the representation of language and text should by its very nature be an interdisciplinary enterprise. As such, one might expect a strong connection between linguistic analyses and psycho linguistic research. However, we believe it is fair to say that researchers have not always appreciated this interdisciplinary aspect, and have often worked in isolation, thereby maintaining the traditional borders between the two disciplines: linguists describe language structures, psycholinguists study mental representations and processes. In recent years, this situation has improved significantly (see, among others, Gernsbacher & Giv6n 1995). Topics like interclausal and inter-sentence relations (for instance, the contributions to Fayol & Costermans 1997), information distribution (Chafe 1994) and the structure of complete texts (Mann & Thompson 1992), have received serious attention in descriptive studies and, to a lesser extent, in studies of text processing. Furthermore,

Text representation as an interface between language and its users

linguists pay more attention to the cognitive aspects of language use (compare the emergence of cognitive linguistics), and psycholinguists give more serious consideration to the linguistic complexity of the object under study. This volume is intended to contribute to the information exchange among researchers from these disciplines. Before giving an overview of the content of the different contributions to this volume, we first highlight themes we believe to be central to the cognitive and linguistic study of text representation. We start with three general tendencies in research on text representation. Then we introduce two linguistic characteristics that constitute a text, which will be the object of study in this volume.

2.

Three major themes in research on text representation

2.1

Multiple representations

So far, we have been using the term 'the text representation'. This term is in fact an oversimplification. Both in linguistics and in psychology text representations are taken to be composite. In psycholinguistics we find this idea very explicitly present in the work ofKintsch and associates (Kintsch 1998, but also in Kintsch & van Dijk 1978 and in Van Dijk & Kintsch 1983; for an overview see also Singer 1990). Their research focuses on the receptive side of communication and states that readers make multiple representations of the sentences of a text: a surface code (a short-lived representation of the exact linguistic material in the sentences), a text base (containing the propositions expressed by the sentences and their interrelations), and a situation model (in which the linguistic material is integrated with the background knowledge of the reader). In linguistics, the concept of multiple representations has also been developed. In formal semantics in the work ofKamp (see Kamp 1981; Heim 1982; Kamp & Reyle 1993) in cognitive linguistics most explicitly in the work by Fauconnier ( 1985, 1994) on mental spaces. In general, the idea is that linguistic expressions are instructions to update the current mental representation, that is based on previous discourse, background knowledge and inferencing. Thus, expressions are considered to have a procedural meaning. Or, as Fauconnier ( 1994, p. xviii) has put it: Language does not itself do the cognitive building - it "just" gives us minimal, but sufficient, clues for finding the domains and principles appropriate for build-

3

4


ing in a given situation. Once these clues are combined with already existing configurations, available cognitive principles, and background framing, the appropriate construction can take place, and the result far exceeds any overt explicit information.

2.2

Underspecification of mental representations

The quote from Fauconnier brings us to the second important theme of recent research on text representations, the underspecification of mental representations. Contrary to what is maintained in the standard coding theory of meaning, it is fairly generally accepted in many branches of linguistics and psychology that what an utterance means cannot in any easy, transparent and compositional way be connected to the meaning of the individual elements in the utterance and their interrelation. An utterance explicitly codes only part of the meaning of the utterance into explicit linguistic material, the rest having to be provided by inferencing. In linguistics this point has been formulated forcefully by Sperber and Wilson ( 1992, Chapter 1) and Fauconnier ( 1994, Introduction), but it is in fact prominent in most descriptive accounts of coherence (see Bublitz, Lenk & Ventola 1999, for a recent overview). It is even one of the major tenets, in a much more radical 1 form, in, e.g., conversation analysis (cf. Pomerantz & Fehr 1997). The same linguistic items will be interpreted differently in different situations and contexts, and hence any adequate theory of meaning will have to allow for rich inferential mechanisms. The Gricean program (e.g., Grice 1975) is one such attempt to provide the necessary inferential power: Under the assumption of cooperativeness, participants in a conversation will generate implicatures in order to extend the literal meaning of an utterance and be able to make a coherent representation of the discourse. This inferential mechanism has been explored systematically by Levinson (1991 ), for instance to account for the binding properties of anaphors. In the field of text linguistics, Spooren ( 1997) uses a similar mechanism to account for so-called underspecified uses of connectives (cases where coherence relations are not (completely) matched by the meaning of the connectives that occur in the text). In psychology there are numerous findings demonstrating the underspecification of linguistic material. For instance, Graesser, Singer, and Trabasso ( 1994) have suggested that in narrative texts causal connections function as a default, thus allowing for coherence relations to remain unmarked. Noordman and Vonk (1998) and Sanders and Noordman (2000),


suggest that such a view is promising for expository texts as well. In the latter study it is found that causal relations lead to faster processing of the connected information, but still lead to a more integrated representation. Taking the role of non-linguistic factors, such as reader's characteristics, into account, Noordman and Vonk (Noordman, Vonk & Kempff 1992; Noordman & Vonk 1992) have shown that depth of processing of causal relations is dependent on the reader's goals and knowledge. 2..3

Dynamic representations

A third recurrent theme is that text representations are constructed dynamically: The effect of a language element on a representation is dependent on the current state of that representation, which is updated incrementally. 2 Most of the current formal semantic systems (Situation Semantics, File Semantics, Discourse Representation Theory, etc.) specifically incorporate this aspect. In psycholinguistic text production and reception systems, incrementation and dynamicity are omnipresent (see Andriessen, De Smedt & Zock 1996; Garnham 1996; Sanders & Van Wijk 1996; Schilperoord 1996). The vocabulary used to capture this aspect of text representation is usually that of connectionism. Although the status of connectionist models as theories of natural language production and interpretation is under dispute (cf. Levelt 1989, p. 20), they have the promise to quite easily capture the flexibility needed to model the temporal course of discourse comprehension and production. The insight that the cognitive processes of text production and interpretation can be modeled as dynamic processes in which activation fluctuates, and that this process is influenced, or even to a large extent determined by the linguistic characteristics of the text, raises one of the most important challenges at the intersection of linguistic and psycholinguistic studies ofdiscourse. Fortunately, there is also an ongoing use of sophisticated empirical methods which enable researchers to ask and find answers to quite precise questions. For instance, a dynamic view on the process of discourse comprehension leads to the expectation that while a reader proceeds through the text, the activation of concepts, facts and events as parts of a discourse representation fluctuates constantly. So, hypotheses considering activation patterns can be tested with on-line methods like reading time registration, naming tasks or eye movement registration (see HaberIandt 1994, for an overview of such methods). Eventually, the fluctuating activation patterns settle into a relatively stable memory representation of the text. Several discourse comprehension models are based on these insights and

S

6


empirical findings, such as the Structure Building Framework (Gernsbacher 1990) the Landscape model of Reading (Van den Broek, Young, Tzeng & Linderholm 1998) and the Construction-Integration model (Kintsch 1998). Questions of how exactly this activation fluctuates, and how the activation is influenced by the linguistic characteristics of the text are currently major research questions, that are partly addressed in contributions to this volume, among others by Gaddy, Van den Broek & Sung in Chapter 3. It is important to note that there is a similar tendency in research on production, even though there is in general less attention for production studies, as Levelt ( 1989) notices in his handbook Speaking, because of the bias in psycholinguistics towards perception research, at the cost of production research. Along the same lines, Kintsch, presenting an overview of discourse psychological work ( 1994, p. 728) remarks that "many psychological studies have concerned themselves with this problem in the past few years, although overwhelmingly with the comprehension rather than the production side." Where empirical findings such as longer or shorter reading times of segments are taken to indicate the level of activation of a concept being processed during text understanding, the on-line registration of pause times opens a promising route to gain further insight in the on-line processes of text production. Schilperoord ( 1996) has used the method of analyzing the location and duration of pauses during written discourse production in an attempt to open up the •black box' of a discourse producer's cognitive representation. He found that text producers tend to pause longer before segments located high in a structural hierarchy of the text under production, than before segments located low in such a hierarchy. If we assume that differences in pause time reflect differences in cognitive effort needed to retrieve information from Long Term Memory, then it can be hypothesized that the hierarchical structure of discourse is a crucial factor in determining the on-line level of accessibility of information (Schilperoord & Sanders 1997, 1999). This line of work, in which a cognitively inspired text-analysis (Sanders & Van Wijk 1996) is combined with on-line psycholinguistic research methods, is an example of how the combination of linguistic and psycholinguistic methods contributes to the development of integrated theories of language structure and language processes. Similar tendencies can be found in research in Cognitive Linguistics, a research paradigm not specifically aimed at the discourse level, on issues like the polysemy of prepositions (Sandra & Rice 1995), epistemic modality (J. Sanders & Spooren 1996) or the study of metaphor and figurative speech (see among many others, Gibbs 1994, 1996 ). In this volume, Chapter 4 by Giora


and Balaban, is another example of such a study on methaphor.

3· Two constituting principles of text: referential and relational coherence Now that we have discussed general properties and tendencies in research on text representation, an obvious question emerges: What are the exact characteristics of the text, the object of representation? We discuss the ones we believe to be most important, i.e. those that determine a set of sentences to be a text rather than a loose set of sentences. 3.1 What makes a text a text?

A constituting characteristic of texts is that they show connectedness. The question of how to characterize this connectedness is generally considered crucial in discourse studies. A dominant stance is that coherence explains best for this connectedness, where coherence is a characteristic of the representation rather than of the text itself. In other words, coherence is considered a mental phenomenon; it is not an inherent property of a text under consideration. Language users establish coherence by relating the different information units in the text. The notion of coherence has a prominent place in both (text- )linguistic and psycholinguistic theories of text and discourse. Although this is not a particularly new view of coherence- it is a dominant thesis in most recent work (see, among many others, Van Dijk & Kintsch 1983; Garnham & Oakhill 1992; Hobbs 1990; Noordman & Vonk 1997; Sanders, Spooren & Noordman 1992)- it is a crucial starting point for theories that aim at describing the link between the structure of a text as a linguistic object, its cognitive representations and the processes oftext production and understanding. And in our view it is this type of theory, located at the intersection of linguistics and psycholinguistics that could lead to significant progress in the field of discourse studies. Generally speaking, there are two respects in which texts can cohere: 1. Referential coherence: units are connected by repeated reference to the same object; 2. Relational coherence: text segments are connected by establishing coherence relations like CAUSE-CONSEQUENCE between them.

7

8


In this volume the role of both types of coherence in text processing is discussed. Now that we have identified the types of representation and the constituting principle of text, we can state a central theme of this volume in different terms: A major issue is the relationship between the linguistic surface code (what Giv6n 1995, calls 'grammar as a processing instructor') and the meaning representations. Both coherence phenomena under investigation - referential and relational coherence - have dear linguistic indicators that can be taken as processing instructions, which will typically affect the surface representation. For referential coherence these are anaphoric devices such as pronouns, and for relational coherence these are connectives and (other) lexical markers of relations. 3.2.

Referential coherence

The relevant linguistic indicators for referential coherence are pronouns and other devices for anaphoric reference. Ever since the seminal work of linguists like Chafe (1976) and Prince (1981), both functional and cognitive linguists have argued that the grammar of referential coherence can be shown to play an important role in the mental operations of connecting incoming information to the existing mental representations. For instance, referent NPs are identified as either those that will be important and topical, or as those that will be unimportant and non-topical. Hence, topical referents are persistent in the mental representation of subsequent discourse, whereas the non-topical ones are non-persistent. Recently, more and more empirical data from corpus studies become available which underpin this cognitive interpretation of referential phenomena, following a route guided by functional linguists such as DuBois (1980). In a distributional study, Giv6n ( 1995), for instance, shows that in English, the indefinite article a( n) is typically used to introduce non-topical referents, whereas topical referents are introduced by this. In addition, there is a clear interaction between grammatical subjecthood and the indefinite article this: most this-marked NPs also appear as grammatical subjects in a sentence, while a majority of a(n)-marked NPs occurred as non-subjects. Across languages there appears to be a topic persistence of referents; in active-transitive clauses the topic persistence of subject NPs is systematically larger than that of object NPs. In several publications Ariel (1988, 1990) has argued that regularities in grammatical coding should indeed be understood to guide processing. She has studied the distribution of anaphoric devices and she has suggested that zero


anaphora and unstressed pronouns co-occur with high accessibility of referents, whereas stressed pronouns and full lexical nouns signal low accessibility. This co-occurrence can easily be understood in terms of cognitive processes of activation: high accessibility markers signal the default choice of continued activation of the current topical referent Low accessibility anaphoric devices such as full NPs or indefinite articles signal the terminated activation of the current topical referent, and the activation of another topic. Ariel ( 1988, 1990) has argued that binding conditions on the distribution and interpretation of pronominal and anaphoric expressions actually are the 'grammaticalized version' of cognitive processes of attention and accessibility of concepts that are referred to linguistically. Speakers encode the degree of accessibility of mental structures in several ways. Each referential expression is a kind of retrieval device for the listener. In fact, Ariel proposes to define referring expressions in terms of a processing procedure: zero-anaphors and pronominal expressions encode highly accessible concepts, where lexical anaphors refer to less accessible referents (see also Chafe 1994, for a processing view on linguistic structures of'given' and 'new'). In experimental research on text processing, quite some work has been done which can be taken to demonstrate the 'psychological reality' of linguistic indicators of referential coherence. On-line studies of pronominal reference have resulted in the formulation of cognitive parsing principles for anaphoric reference (cf. Garrod & Sanford 1994; Sanford & Garrod 1994). For instance, it is easier to resolve a pronoun with only one possible referent, and it is easier to resolve pronouns with proximal referents than distant ones. As for the time course, eye fixation studies have repeatedly shown that anaphoric expressions are resolved immediately (e.g., Carpenter & Just 1977; Ehrlich & Rayner 1983 ). Consider an example like (1). ( 1) a. The guard mocked one of the prisoners in the machine shop. b. He had been at the prison for only one week. When readers came upon ambiguous pronouns such as he in (lb), they frequently looked back in the text. More than 50% of these regressive fixations were to one of the two nouns in the text preceding the pronoun, suggesting that readers indeed attempted to resolve the pronoun immediately. As for the meaning representation, it has been shown that readers have difficulty to understand the text correctly when the antecedent and referent are too far apart and reference takes the form of a pronoun.

9

10


On a more global text level, research on the exact working of accessibility markers as processing instructions is rare, but a good example is Vonk, Hustinx and Simons (1992). They show the relevance of discourse context for the interpretation of referential expressions. Sometimes anaphors are more specific than would be necessary for their identificational function (for instance, full NPs are used rather than pronominal expressions). The authors convincingly argue that this phenomenon can be explained in terms of the thematic development of discourse: if a character is referred to by a proper name after a run of pronominal references, then the name itself serves to indicate that a shift in topic is occurring. Readers process the referential expressions differently, as becomes apparent from reading times. Whereas anaphoric reference modulates the availability of previously mentioned concepts, cataphoric devices change the availability of concepts for the text that follows. Gemsbacher ( 1990) and her colleagues have demonstrated the reader's sensitivity for this type of linguistic indicators of reference. They contrasted cataphoric reference byway of the indefinite a(n) versus definite this to refer to a newly introduced referent in a story. So the new referent egg was introduced either as 'an egg' or as 'this egg'. It was hypothesized that the cataphor this would signal that a concept is likely to be mentioned again in the following story and that therefore the this-cataphor results in a higher activation. Subjects listened to texts and where then asked to continue the text after the critical concept. They appeared to refer sooner and more often to a concept introduced by this than by an. These and other results show that concepts that were marked as a potential discourse topic by this are more strongly activated, more resistant to being suppressed in activation, as well as more effective in suppressing the activation of other concepts (Gemsbacher 1990; Gernsbacher & Shroyer 1989). It is this type of findings that provides the psycholinguistic underpinning for the idea of'grammar as a processing instructor'. 3.2

Relational coherence

So far, we have discussed examples of the way in which linguistic signals of referential coherence affect text processing. We now move to signals of relational coherence. In many approaches to discourse connectedness, coherence relations are taken to account for the coherence in readers' cognitive text representation (cf. Hobbs 1979; Mann & Thompson 1986; Sanders et al 1992; Sanders, Spooren & Noordman 1993 ). Coherence relations are meaning relations which connect two text segments (i.e. minimally clauses). Examples are relations like


CAUSE-CONSEQUENCE, LisT and PROBLEM-SOLUTION. These relations are conceptual and they can, but need not, be made explicit by linguistic markers. Below, we will first focus on the second aspect of relational coherence, that on the level of the surface code: linguistic markers such as connectives and signaling phrases. After that we will move to the level of the meaning representation: the nature of the relations themselves. Ever since Ducrot (1980) and Lang (1984), there have been linguistic accounts of connectives as operating instructions. The basic idea is that a connective has the function of relating the content of the connected segments in a specific type of relationship. Anscombre and Ducrot ( 1977), for instance, analyze but as setting up an argumentative scale (for instance, the (un)desirability ofJohn), with one segment tending towards the negative side of the scale and the other towards the positive side: (2) John is rich, but dumb. In Chapter 9, Snoeck Henkemans explores the relationship between this type of argumentative approaches to connectives and the text-linguistic ones that were considered earlier above. In his influential work on Mental Spaces, Fauconnier (1985, 1994) treats connectives as one of the so-called space-builders, linguistic expressions that typically establish new mental spaces. Mental spaces are mental constructs set up to interpret utterances, "structured, incremental sets [ ... ] and relations holding between them [ ... ], such that new elements can be added to them and new relations established between their elements, (Fauconnier 1994, p. 16). Other examples of space builders are prepositional phrases (In 1929, From her point of view), adverbs (really, probably). A connective acting as a space-builder is an if-then conditional, as in 'Ifl were a millionaire, my VW would be a Rolls• or 'If he had listened to his mother, this criminal would be a Saint". Such expressions if p then q set up a new mental space H in which a p and q hold. So, if p is the space builder and in this new space my VW from the initial space is identified with the Rolls in the new space (for the detailed analyses see Fauconnier 1994, Chapters 3 and 4 and Sweetser 1996). Fauconnier argues that the solution to some of the problems of traditional semantics, such as opacity, presupposition and the like, falls out naturally from these mechanisms. In the same vein, Spooren (1989) has argued that but-coordinations typically function to contrast conflicting information coming from different perspectives, and that this may even affect the truth-conditional level For instance, (3a) is possible, but (3b) is a contradiction (Spooren 1989, p. 69).

n

12


(3)

Cassius Clay was shy, but Muhammed Ali wasn't. b. Muhammed Ali was shy, but Muhammed Ali wasn't

a.

This type of dynamic approach to connectives 'as processing instructors' is becoming more and more important, not in the least because of the rise of Cognitive Linguistics as a new branch on the linguistic tree. Is there any psycholinguistic work showing the relevance of ideas like this? Indeed, in various on-line processing studies the function of linguistic markers is examined. These studies have primarily aimed at the investigation of the processing role of the signals per se, rather than on more sophisticated ideas like the exact working of 'space building'. The experimental work typically includes the comparison of reading times of identical textual fragments with different linguistic signals preceding them. Recent studies on the role of connectives and signalling phrases show that these linguistic signals affect the construction of the text representation (cf. Deaton & Gernsbacher in press; Millis & Just 1994; Noordman & Vonk 1997; Sanders & Noordman 2000). Millis and Just (1994), for instance, investigated the influence of connectives like because immediately after reading a sentence. When participants had read two clauses that were either linked or not linked by a connective, they judged whether a probe word had been mentioned in one of the clauses. The recognition time to probes from the first clause was consistently faster when the clauses were linked by a connective. The presence of the connective also led to faster and more accurate responses to comprehension questions. These results suggest that the connective does influence the representation immediately after reading. Deaton and Gernsbacher (in press) combined on-line and off-line measures to investigate readers' use of because. They found that two causally related clauses connected by because were read more rapidly than when they were presented without the conjunction. When the clauses were conjoined by because, the second clauses were also recalled more frequently in a prompted recall test. Generally speaking, studies on the influence of linguistic markers on text representation show a rather inconsistent pattern. Sometimes linguistic markers give rise to better structure in free recall (Meyer, Brandt & Bluth 1980), to faster and more accurate reactions on a probe task, to faster and more accurate responses to comprehension questions (Degand, Lefevre & Bestgen 1999; Millis & Just 1994), and to better recall in a prompted recall task (Deaton & Gemsbacher, in press), but they do not lead to more information recalled


(Britton, Glynn, Meyer & Penland 1982; Meyer 1975). At the same time, online data suggest that the presence of linguistic markers facilitates processing (Britton et al. 1982; Deaton & Gemsbacher, in press). Thus far, we have discussed the role of connectives and signaling phrases in discourse processing. A preliminary conclusion might be that they can be treated as linguistic markers which instruct readers in how to connect the new discourse segment with the previous one (Britton 1994). In the absence of such instructions readers have to determine for themselves what coherence relation connects the incoming segment to the previous discourse. Such an inference process requires additional cognitive energy and results in longer processing times. If this idea has any validity, it implies that the coherence relations themselves would have a major influence on discourse processing as well. One might expect that the type of relation that connects two discourse segments, be it causal, additive, contrastive etc., affects the discourse representation. Here we move into another area where the combination of text linguistic and discourse psychological insights has lead to significant progress: the categorization of coherence relations. In the last decade, a significant part of research on coherence relations has focused on the question how the many different sets of relations should be organized (Hovy 1990; Knott & Dale 1994; Pander Maat 1998; Redeker 1990; Sanders 1997a). Sanders et al. ( 1992, 1993) have started from the properties common to all relations, in order to define the 'relations among the relations', relying on the intuition that some coherence relations are more alike than others. For instance, the relations in ( 4), (5) and (6), all express (a certain type of) causality, whereas the ones in (7) and (8) do not. Furthermore, a negative relation is expressed in (7), as opposed to all other examples, and (8) expresses an enumeration or addition. (4) The buzzard was looking for prey. The bird was soaring in the air for hours. (5) The bird has been soaring in the air for hours now. It must be a buzzard. (6) The buzzard has been soaring in the air for hours now. Let's finally go home! (7) The buzzard was soaring in the air for hours. Yesterday we did not see it all day. (8) The buzzard was soaring in the air for hours. There was a peregrine falcon in the area, too. A dominant distinction in existing classification proposals is that between socalled content, ideational, external or semantic relations on the one hand, and

13

14


presentational, internal and pragmatic relations, on the other hand. In the first type of relations, segments are related because of their propositional content, i.e. the locutionary meaning of the segments. They describe events that cohere in the world. The relation in (9) can be interpreted as semantic because it connects two events in the world; our knowledge allows us to relate the segments as coherent in the world. A relation like (9) could be paraphrased as "the cause in the first segment (Sl) leads to the fact reported in the second segment (52)" (Sanders 1997a). (9) The neighbours suddenly left for Paris last friday. So they are not at home. (10) The lights in their living room are out. So the neighbours are not at home. In (10) however, the two discourse segments are related because we understand the second part as a conclusion from evidence in the first, and not because there is a causal relation between two states of affairs in the world: It is not because the lights are out that the neighbours are not at home. The causal relation (10) could be paraphrased as "the description in 51 gives rise to the conclusion or claim formulated in the 52." Hence, in the second type of relation the discourse segments are related because of the illocutionary meaning of one or both of the segments. The coherence relation concerns the speech act status of the segments. If this distinction is applied to the set of examples above, the causal relation (4) is semantic, whereas (5) and (6) are pragmatic. This systematic difference between types of relations is noted by many students of discourse coherence. Still, there is quite a lot of discussion about the exact definition of a distinction like this (see e.g., Bateman & Rondhuis 1997; Degand 1996; Hovy 1990; Knott & Dale 1994; Knott 1996; Knott & Sanders 1998; Martin 1992; Moore & Pollack 1992; Oversteegen 1997; Pander Maat 1998; Sanders 1997a; Sanders & Spooren 1999 and several contributions to Couper-Kuhlen and Kortmann (2000)). At the same time, several researchers have come up with highly similar distinctions, and there seems to be basic agreement on the characteristics of the prototypical relations (Sanders 1997a). Moreover, very similar distinctions have been shown useful in describing the differential meaning of conjunctions (Sweetser 1990). Also, the saliency of categorizations like these has been shown in experiments in which, among other tasks, language users were asked to judge the similarity of relations. Still, the discussion on this issue is clearly continued in this volume, especially in Section 2 (Chapters 6 to 9).


One of the emerging tendencies in recent linguistic research on the classification of coherence relations is the relevance of the notions perspective and subjectification. In several influential publications, Ducrot has stressed the diaphonic nature of discourse. Even in monologual texts traces can be found of other 'voices', information that is not presented as fact-like, but from a particular point-of-view, either the current speaker's (subjectified information, in the terminology of}. Sanders & Spooren 1997) or another cognizer's (perspectivized information). Langacker has contributed much to the study ofthe linguistic effects of notions like subjectification (e.g., Langacker 1990) and this notion also seems valid for the study of coherence relations and connectives (Pander Maat & Sanders 1999; Verhagen 1995, and some other contributions to Stein & Wright 1995). Although perspective remains an elusive notion for linguistics and psycholinguistics alike, Fauconnier's mental space framework seems adequate in capturing this intriguing aspect of language (J. Sanders 1994; J. Sanders & Redeker 1996, Verhagen 2000). If categorizations of coherence relations and connectives indeed have cognitive significance, they should show relevant in areas like language development, both diachronic (grammaticalization, cf. Sweetser 1990; Traugott 1988; Traugott & Heine 1991) and synchronic (language acquisition), and discourse processing. In all three areas, substantial studies are under way (Evers-Vermeul 2000; Spooren & Sanders in prep.), and there already exists suggestive evidence. Research on first language acquisition shows that the order in which children acquire connectives shows increasing complexity, which can be accounted for in terms of the relational categories mentioned above: ADDITIVES before CAUSALS, POSITIVES before NEGATIVES (see Spooren, Sanders & Visser 1994; Spooren 1997). And in text processing, there is work on the role of (different types of) coherence relations in the construction of a meaning representation. However, at first sight, results of experimental studies on that issue provide a less dear picture than the one for the role of linguistic markers, especially when it concerns expository rather than narrative texts. Perhaps this situation is largely due to the fact that it is difficult to design reading experiments in which coherence relations or text structures are varied in a succesful and independent way, while at the same time avoiding the use of ill-formed texts (what Graesser, Millis, and Zwaan 1997, have called "textoids"). Nevertheless, the idea that coherence relations affect text processing does get support from results of processing studies. Several studies suggest a processing difference between CAUSAL and non-CAUSAL relations. For instance, causally related events in short narratives are recalled better (Black & Bern

15

16


1981; Trabasso & Van den Broek 1985; Trabasso & Sperry 1985). Keenan, Baillet and Brown ( 1984) and Myers, Shinjo and Duffy ( 1987) demonstrated that the effect of causal connectedness on memory for sentences is greatest for moderate levels of causality. Also, causally related sentences are read faster (Haberlandt & Bingham 1978), and the reading time decreases when the causality increases (Keenan et al. 1984, Myers et al. 1987). Fewer studies exist for expository text. One example is a study by Meyer and Freedle (1984). They claimed that differences exist in the amount of organization of different types of text structure. The better organizing types are CoMPARISON, CAUSATION and PROBLEM-SOLUTION, whereas COLLECTION is a weaker organizing type. These structure types are rather similar to the often distinguished types of coherence relations. In a free recall experiment, Meyer and Freedle expected readers to reproduce more information from better organized types than from less organized ones. The results show that recall of the CAUSATION and COMPARISON passages was indeed superior to the recall of the coLLECTION passage. 3 However, there are some problems with the Meyer and Freedle study, see Horowitz (1987) and Sanders & Noordman (2000) for further details. Sanders and Noordman (2000) embedded a similar text segment in two different contexts. In one case it was a Solution to a Problem, in the second case the same segment was part of an addition. It was found that PROBLEM-SOLUTION relations lead to faster processing, better verification and superior recall. The authors conclude that the processing of a text segment depends on the relation it has with preceding segments. Perhaps the most interesting finding in this experiment is the contrast in the effect of the two independent variables: linguistic markers (implicit or explicit) and type relations (PROBLEM-SOLUTION or LIST). Explicit marking of the relations resulted in faster processing, but did not affect recall. However, verification data concerning the representation immediately after reading, show an effect of the linguistic marker. This finding is quite similar to the effect Millis and Just (1994) found for the influence of because. Hence, it can be concluded that the relational marker has an effect during on-line processing, but that its influence decreases over time. This contrasts with the effect of the coherence relation, which is also manifest in the recall. This contrast is similar to another frequently observed finding in language comprehension: initially a reader or listener constructs the surface representation of a sentence, but after a short time interval only the meaning or gist of the message is retained. Sachs ( 1967) found this effect for the form and meaning of


sentences which participants had to identify as identical or different. Results like these strongly support the idea that coherence relations are an indissoluble part of the cognitive representation itself, whereas linguistic markers like connectives and signaUing phrases are merely expressions of these relations, which guide the reader in selecting the right coherence relation. This conclusion is highly compatible to a view on coherence in which linguistic markers, as part of the surface code, 'guide' the reader towards a coherent text representation (cf. Gemsbacher & Giv6n 1995; Graesser et al. 1997; Noordman & Vonk 1998). So far, we have presented an overview of text-linguistic and psycholinguistic work on referential and relational coherence. This overview might suggest discourse processing to depend entirely on text characteristics such as the linguistic markers of referential and relational coherence. That is not the case. It is rather plausible that the role of coherence and its linguistic markers interacts with 'reader factors' like interestingness of materials (Spooren, Mulder & Hoeken 1998), domain knowledge (Birkmire 1985; McNamara & Kintsch 1996), topic complexity (Spyridakis & Standal 1987) reader's goals (Noordman et al. 1992) and verbal ability (Meyer, Young & Bartlett 1989). We are convinced that the interaction of such reader's characteristics with textstructural properties are prime issues for further research on text processing (see also Kintsch 1998). However, the focus of this chapter, and in fact, of this entire volume, is on gaining further insight into the role of coherence and text structure itsel£ In many discourse psychological studies on the interaction of textual and reader's factors, the coherence of passages is varied by manipulating many different textual aspects of coherence at the same time, such as adding elaborative information, identification of anaphoric references, and even supplying background information. As a result, this type of research has often conflated coherence per se and various other textual aspects that potential influence coherence. As we have argued above, we think it is crucial for the further progress of the field, that we get a better grip on the linguistic factors that determine the cognitive discourse representation. This can only be achieved by a further cross-fertilization of the fields of text linguistics and discourse psychology. A good iUustration of this point concerns the role of relational markers of text structure. It has been argued above that signaling phrases and connectives make existing relations explicit. This also implies that the use of the markers is bound to restrictions: not every connective can express every relation. In recent text-linguistic work we are beginning to understand what

17

18


these restrictions are and how they interplay with the meaning expressed by the connected segments {cf. Knott & Dale 1994; Pander Maat & Sanders 2000 and several contributions to Risselada & Spooren 1998). It is this type of insights that underlines the importance of further cooperation of text linguists and psycholinguists working on discourse (Sanders 1997b).

4-

The coherence of this volume

What can readers expect of this volume? The collected chapters typically present a cross-disciplinary account of text representation, by both linguists and psycholinguists. This implies that linguistic analyses of textual characteristics ultimatdy aim at accounting for the cognitive interpretation they can receive. At the same time, psycholinguistic studies focus on the relevance of text characteristics for theories of text processing, where text processing concerns both production and interpretation. An important benefit of this combination of text linguistics and psycholinguistics, and of production and understanding is that we will encounter various methodologies, which are complementary: linguistic analysis, text analysis, corpus linguistics, computational linguistics, argumentation analysis, and the experimental psycholinguistic study of text processing. A final focus of this book is the comparison and further testing of linguistic and processing theories of text representation. The following 12 chapters are divided in four sections. Section 1 deals with referential coherence in text and text representation, and especially with accessibility: how can the notion of varying accessibility explain for different referential forms, and what is the evidence for such a dynamic account of the cognitive representations language users have? In Section 2 focus shifts from referential to relational coherence in text and text representation, when the classification of coherence rdations and connectives is discussed in a closely connected cluster of chapters, combining various theoretical approaches (from Relevance Theory to Argumentation Theory and cognitive accounts of coherence relations) and different empirical methods (from text-analysis to reading experiments). Section 3 focuses on the cognitive representations of discourse and its relation to knowledge representations: how are they related? How large is the role of linguistic factors? Finally, Section 4 discusses an issue typically neglected in the previous sections: when coherence is said to exist, it exists between something, for


instance, discourse units or text segments. But how are these segments defined? And when we distinguish between different linguistic levels of representation (word, clause, sentence, paragraph), do we know that these levels have any psychological validity? Together, the chapters in these four sections present an overview of a growing field of interest, at the intersection of linguistics and psychology, the study of a phenomenon that is crucial to our behavior because it is the mostly used vehicle of communication: that of text and its cognitive representation.

Notes More radical in that radical conversation analysists deny the possibility of attributing any meaning to an utterance without regard of its context.

1.

Stricdy speaking there is no logical connection between dynamic systems and incremental systems, in that dynamic systems can be constructed that are not incremental and incremental systems that are not dynamic. Yet in every serious language interpretation system that we know of the two go together.

1..

3· However, there are some problems with the Meyer and Freedle study, see Horowitz

( 1987) and Sanders and Noordman (2000) for further details.

References Andriessen, J., De Smedt, K., 8c Zock, M. ( 1996). Discourse planning: Empirical research and computer models. InT. Dijkstra, 8c K. De Smedt (Eds.), Computational linguistics. AI and connectionist models ofhuman language processing(pp. 247-278). London etc.: Taylor 8c Francis. Anscombre, J., & Ducrot, 0. ( 19n). Deux MAIS en Fran~ais? Lingua, 43, 23-40. Ariel, M. (1988). Referring and accessibility. Journal ofLinguistics, 24,65-87. Ariel, M. {1990). Accessing noun-phrase antecedents. London: Roudedge. Bateman, }.A., 8c Rondhuis, K. J. (1997). Coherence relations: towards a general specification. Discourse Processes. 24, 3-49. Birkmire, D. P. {1985). Text processing: The influence of text structure, background knowledge and purpose. Reading Research Quarterly, 20, 314-326. Black. J. B., & Bern, H. {1981). Causal coherence and memory for events in narratives. Journal of Verbal Learning and Verbal Behavior, 20, 267-275. Britton, B. K., Glynn, S.M., Meyer, B.}. F., & Penland, M. J. (1982). Effects of text structure on use of cognitive capacity during reading. Journal of Educational Psychology, 74, 51-61. Britton, B. K. ( 1994). Understanding expository text: Building mental structures to induce insights. In M.A. Gemsbacher {Ed.), Handbook of psycholinguistics (pp.

19

~o


640-674). San Diego etc.: Academic Press. Bublitz, W., l..enk, U., & Ventola, E. (Eds.). (1999). Coherence in spoken and written

discourse. Amsterdam etc.: John Benjamins. Chafe, W. L. ( 1976). Givenness, contrastiveness, definiteness, subjects, topics and point of view. In C. N. U (Ed.), Subject and topic (pp. 25-55). New York: Academic Press. Chafe, W. L. (1994). Discourse, consciousness, and time. The flow and displacement of conscious experience in speaking and writing. Chicago: Chicago University Press. Couper-Kuhlen, E. & Kortmann, B. (Eds.). (2000). Cause, Condition, Concession, Contrast. Cognitive and discourse perspectives. Berlin: Mouton de Gruyter. Deaton, J. A., & Gemsbacher, M.A. (in press). Causal conjunctions: Cue mapping in sentence comprehension. Journal ofMemory and lAnguage. Degand, L. ( 1996 ). A situation-based approach to causation in Dutch with some implications for text generation. Dissertation, Universit~ Catholique de Louvain, Belgium. Degand, L., Lefevre, N., & Bestgen, Y. ( 1999). The impact of connectives and anaphoric expressions on expository discourse comprehension. Document Design, 1, 39-51. DuBois, J. W. ( 1980). Beyond definiteness: The trace of identity in discourse. In W. L. Chafe (Ed.), The Pear Stories: Cognitive, cultural and linguistic aspects of narrative production (pp. 203-274). Norwood, NJ: Ablex. Ducrot, 0. {1980). Essai d'application: MAIS - les allusions a l'monciation d~locutifs, perfonnatifs, discours indirect In H. Parret (Ed.), Le langage en context: etudes philosophiques et linguistiques de pragmatique (pp. 487-575). Amsterdam: Benjamins. Ehrlich, K., & Rayner, K. (1983 ). Pronoun assignment and semantic integration during reading: Eye movements and immediacy of processing. Journal of Verbal Learning and Verbal Behavior, 22,75-87. Evers-Vermeul, J. {2000). De complexiteit van connectief-verwerving [The complexity of connective acquisition]. Nederlandse Taalkunde. Fayol,M., &Costennans, J. (Eds.) (1997). Processinginterclausal relationships in production and comprehension of text Hillsdale, NJ: Erlbaum. Fauconnier, G. (1994). Mental spaces: Aspects of mtJJning construction in natural language (2nd ed.). Cambridge, MA: Bradford. Gamham, A., & Oakhill, J, (Eds.). ( 1992). Discourse representation and text processing. A special issue of Language and Cognitive processes. Hove, UK: Lawrence Erlbaum Associates. Gamham, A ( 1996 ). Discourse comprehension models. In T. Dijkstra, & K. De Smedt (Eds.), Computational Linguistics. AI and connectionist models of human language processing(pp. 221-246). London etc.: Taylor & Francis. Garrod, S. C., & Sanford, A. J. ( 1994 ). Resolving sentences in a discourse context: How discourse representation affects language understanding. In M. A. Gernsbacher (Ed.), Handbook ofpsycholinguistics (pp. 675-698). San Diego etc.: Academic Press. Gernsbacher, M.A., & Giv6n, T. (Eds.) (1995). Coherence in spontaneous text. Amsterdam etc.: John Benjamins. Gemsbacher, M.A., & Shroyer, S. (1989). The cataphoric use of the indefinite this in spoken narratives. Memory 6- Cognition, 17, 53fr540.


Gemsbacher, M. A. ( 1990). Language comprehension as structure-building. Hillsdale, NJ: Erlbaum. Gibbs, R. W. ( 1994). Figurative thought and figurative language. In M. A. Gernsbacher (Ed.), Handbook ofpsycholinguistics (pp. 411--446). San Diego etc.: Academic Press. Gibbs, R. W. (1996). What's cognitive about cognitive linguistics? In E. Casad (Ed.), Cognitive Linguistics in the Redwoods (pp. 27-54). Berlin: Mouton de Gruyter. Giv6n, T. ( 1995 ). Coherence in text vs. coherence in mind. In M.A. Gemsbacher, & T. Giv6n (Eds.), Coherence in sponttmeous text (pp. 59-115). Amsterdam etc.: John Benjamins. Typological Studies in Language, 31. Graesser, A. C., Millis, K. K., & Zwaan, R. (1997). Discourse comprehension. Annual ReviewofPsychology, 48, 163-189. Graesser, A. C., Singer, M., & Trabasso, T. (1994). Constructing inferences during narrative text comprehension. PsychologiaU Review, 101,371-395. Grice, H. P. ( 1975). Logic and conversation. In P. Cole, & P. L. Morgan (Eds.), Syntax and semantics 3: Speech acts (pp. 41-58). New York: Academic Press. Haberlandt, K., & Bingham, G. (1978). Verbs contnbute to the coherence of brief narratives: Reading related and unrelated sentence triples. Journal of Verbal Learning and Verbal Behavior, 17,419-425. Haberlandt, K. ( 1994). Methods in reading research. In M. A. Gemsbacher (Ed.), Handbook ofpsycholinguistics (pp. 1-31 ). San Diego etc.: Academic Press. Heim, I. (1982). The semantics of definite and indefinite noun phrases. Unpublished Thesis, Univ. ofMassachussetts. Hobbs, J. R. ( 1990). Literature and cognition. Menlo Park, CA: CSLI. Hobbs, J. R. (1979). Coherence and coreference. Cognitive Science, 3, 67-90. Horowitz, R. ( 1987). Rhetorical structure in discourse processing. In R. Horowitz, & S. J. Samuels (Eds.), Comprehending oral and written language (pp. 117-160). San Diego, CA: Academic Press. Hovy, E. H. (1990). Parsimonious and profligate approaches to the question of discourse structure relations. Proceedings of the 5th International Workshop on Natural Language Generation. Just, M. & Carpenter, P. (Eds.), Cognitive processes in comprehension. Hillsdale NJ: Erlbaum. Kamp, H., & Reyle, U. (1993). From discourse to logic. Introduction to model-theoretic

semantics ofnatural language, formal logic and discourse representation theory. Kluwer: Dordrecht. Kamp, H. ( 1981 ). A theory of truth and semantic representation. In J. Groenendijk, T. Janssen, & M. Stokhof (Eds.), Truth, interpretation, and information (pp. 1-41 ). Dordrecht: Foris. Keenan, J. M., Baillet, S.D., & Brown, P. (1984). The effects of causal cohesion on comprehension and memory. Journal of Verbal Learning and Verbal Behavior, 23, 115-126. Kintsch, W., & van Dijk, T. A. (1978). Toward a model of text comprehension and production. Psychologiatl Review, 85, 363-394. Kintsch, W. (1994). The psychology of discourse processing. In M.A. Gemsbacher

21

n


(Ed.), Handbookofpsycholinguistics(pp. 721-739). SanDiegoetc.:Academic Press. Kintsch, W. ( 1998 ). Comprehension. A paradigm for cognition. Cambridge: Cambridge University Press. Knott, A. ( 1996). A d~Jta-driven methodology for motivating a set of coherence relations. Doctoral dissertation. University of Edinburgh, Edinburgh, Scotland. Knott, A., & Dale, R. ( 1994). Using linguistic phenomena to motivate a set of coherence relations. Discourse Processes, 18, 35-62. Knott, A., & Sanders, T. (1998). The classification of coherence relations and their linguistic markers: An exploration of two languages. Journal ofPragmatics, 30, 135175.

Lang, E. ( 1984). The semantics ofcoordination. Amsterdam: Benjamins. Langacker, R. W. ( 1990). Subjectification. Cognitive Linguistics, 1, 5-38. Levett, W. J. M. ( 1989). Speaking. From intention to articulation. Cambridge (Ma) etc.: Bradford/MIT Press. Levinson, S. C. ( 1991 ). Pragmatic reduction of the binding condition revisited. Journal ofLinguistics, 27, 107-161. Mann, W. C., & Thompson, S. A. (Eds.). (1992). Discourse description. Diverse analyses ofa fund-raising text. Amsterdam: Benjamins. Mann, W. C., & Thompson, S. A. ( 1986 ). Relational propositions in discourse. Discourse Processes, 9, 57-90. Martin, J. R. ( 1992). English text. System and structure. Philadelphia: John Benjamins. McNamara, D. S., & Kintsch, W. (1996). Learning from texts: Effects of prior knowledge and text coherence. Discourse Processes, 22, 247-288. Meyer, B. J. F. (1975). The organization ofprose and its tffects on memory. Amsterdam: North-Holland. Meyer, B. J. f., & Freedle. R. 0. (1984). Effects of discourse type on recall Ameriam EduaJtional Research Jouma~ 21, 121-143. Meyer, B. J. F., Brandt, D. M., & Bluth, G. J. (1980). Use oftop-level structure in text: Key for reading comprehension of ninth-grade students. Reading Research Quarterly, 16, 72-103. Meyer, B. J. F., Young, C. J., & Bartlett, B. J. (1989). Memory improved: Enhanced reading comprehension and memory across the life span through strategic text structure. Hillsdale, NJ: Erlbaum. Millis, K. K., & Just, M.A. ( 1994). The influence of connectives on sentence compr~ hension. Journal ofMemory and Language, 33, 128-147. Moore, J.D., & Pollack, M. E. (1992). A problem for RST: The need for multi-level discourse analysis. Computational Linguistics, 18, 537-544. Myers, J. L., Shinjo, M., & Duffy, S. (1987). Degree of causal relatedness and memory. Journal ofMemory and Language, 26, 453-465. Noordman, L. G. M., & Vonk, W. (1997). The different functions of a conjunction in constructing a representation of the discourse. In M. Fayol, & J. Costermans (Eds. ), Processing interclausal relationships in production and comprehension oftext (pp. 7593 ). Hillsdale, NJ: Erlbaum. Noordman, L. G. M., & Vonk, W. ( 1998). Memory-based processing in understanding


causal information. Discourse Processes. 26, 191-212. Noordman, L. G. M., & Vonk, W. (1992). Readers' knowledge and the control of inferences in reading. Language and Cognitive Processes, 7, 373-391. Noordman, L. G. M., Vonk, W., & Kempff, H. J. (1992). Causal inferences during the reading of expository texts. Journal ofMemory and Language, 31, 573-590. Oversteegen, L E. ( 1997). On the pragmatic nature of causal and contrastive connectives. Discourse Processes. 24, 51-85. Pander Maat, H. ( 1998). The classification of negative coherence relations and connectives. Journal ofPragmatics, 30, 177-204. Pander Maat, H., & Sanders, T. (2000). Domains of use and subjectivity. On the distribution ofthree Dutch causual connnectives. In Couper-KuhlenandKortmann (Eds.), Cause, Condition Concession, Contrast. Cognitive and Discourse perspectives,

57-83. Berlin: Mouton de Gruyter. Pomerantz, A., & Fehr, B. J. (1997). Conversation analysis: An approach to the study of social action as sense making practices. InT. A. Van Dijk (Ed.), Discourse as social interaction (pp. 64-91 ). Thousand Oaks, CA: Sage. Prince, E. (1981). Toward a taxonomy of Given-New information. In P. Cole (Ed.), Radical Pragmatics. New York: Academic Press. Reddy, M. (1979). The conduit metaphor. In A. Ortony (Ed.), Metaphor and thought. Cambridge: Cambridge University Press. Redeker, G. ( 1990). Ideational and pragmatic markers of discourse structure. Journal of Pragmatics, 14,305-319. Risselada, R., & Spooren, W. (Eds.). (1998). Discourse Markers. Special issue of Journal of Pragmatics, 30. Sachs, J.D. (1967). Recognition memory for syntactic and semantic aspects of connected discourse. Perception and Psychophysics. 2, 437-442. Sanders, J. M. (1994). Perspective in narrative discourse. Dissertation KU Brabant, Tilburg. Sanders, J. M. & Redeker, G. (1996). Perspective and the representation of speech and thought in narrative discourse. In G. Fauconnier & E. Sweetser (Eds.), Spaces. worlds and grammars (pp. 290-317). Chicago: University of Chicago Press. Sanders, J., & Spooren, W. ( 1996). Subjectivity and certainty in epistemic modality: A study of Dutch epistemic modifiers. Cognitive Linguistics. 7, 241-264. Sanders. J., & Spooren, W. ( 1997). Perspective, subjectivity and modality from a cognitive linguistic point of view. In W. A. Liebert, G. Redeker, & L. Waugh (Eds.), Discourse and perspective in cognitive linguistics. (pp. 85-112). Amsterdam: Benjamins. Sanders, T. ( 1997a). Semantic and pragmatic sources of coherence. On the categorization of coherence relations in context. Discourse Processes, 24. 119-147. Sanders, T. ( 1997b). Psycholinguistics and the discourse level: Challenges for cognitive linguistics. Cognitive Linguistics, 8, 243-265. Sanders, T. f. M., & Noordman, L G. M. (2000). The role of coherence relations and their linguistic markers in text processing. Discourse Processes, 29, 1, 37-60. Sanders, T., & Spooren W. ( 1999). Communicative intentions and coherence relations.

23

24


In W. Bublitz, U. Lenk, & E. Ventola (Eds.), Coherence in text and discourse (pp. 235-250). Amsterdam etc.: John Benjamins. Sanders, T. J. M., Spooren, W. P.M., & Noordman, L. G. M. (1992). Toward a taxonomy of coherence relations. Discourse Processes, 15, 1-35. Sanders, T. }. M., Spoorcn, W. P.M., & Noordman, L. G. M. (1993). Coherence relations in a cognitive theory of discourse representation. Cognitive Linguistics, 4, 93133. Sanders, T ., & van Wijk, C. ( 1996). PISA- A procedure for analyzing the structure of explanatory texts. Text, 16,91-132. Sandra, D., & Rice, S. (1995). Network analyses of prepositional meaning: Mirroring whose mind- the linguist's or the language user's? Cognitive Linguistics, 6, 89130. Sanford, A.}., & Garrod, S.C. (1994). Selective processes in text understanding. In M.A. Gemsbacher (Ed.), Handbook ofpsycholinguistics (pp. 699-720}. San Diego etc.: Academic Press. Schilperoord, }. (1996). It's about time. Temporal aspects of cognitive processes in text production. Amsterdam: Rodopi. Schilperoord, }., & Sanders, T. ( 1997). Pauses, cognitive rhythms and discourse structure. An empirical study of discourse production. In A. Liebert, G. Redeker, & L. Waugh (Eds.}, Disourse and Perspective in Cognitive Linguistics (pp. 247-249). Amsterdam: Bcnjamins. Schilperoord, }., & Sanders, T. (1999). How hierarchical structure affects retrieval processes; implications of pause and text analysis. In D. Galbraith, & M. Torrance (Eds.), Conceptual processes in writing (pp. 13-30). Amsterdam: Amsterdam University Press. Singer, M. (1990). Psychology of language: An introduction to sentence and discourse processes. Hillsdale, NJ: Erlbaum. Sperber, D., & Wilson, D. (1992}. Relevance: Communication and cognition (2nd ed.). Oxford: Blackwell. Spooren, W. P.M. S. ( 1989). Some aspects of the form and interpretation ofglobal contrastive coherence relations. Ph.D. Dissertation Nijmegen University. Spooren, W. (1997). The processing of underspecified coherence relations. Discourse Processes, 24, 149-168. Spooren, W., Mulder, M., & Hoeken, H. (1998). The role of interest and text structure in professional reading. ]oumal ofResearch in Reading, 21, 109-120. Spooren, W., & Sanders, T. (in prep.). What does children's discourse tell us about the nature ofcoherence relations? Paper VU Amsterdam I UiL OTS UtrechL Spooren, W., Sanders, T., & Visser, J. (1994). Taxonomic van cohcrcnce-rclaties: Evidentie uit taalverwervingsonderzoek? (Taxonomy of coherence relations: Evidence from language acquisition research?) GrammafiTI, 3, 33-54. Spyridakis, f. H., & Standal, T. C. ( 1987). Signals in expository prose. Reading Research Quarterly, 22, 3, 285-298. Stein, D., & Wright, S. (Eds.). (1995). Subjectivity and subjectivisation. Cambridge, Cambridge University Press.


Sweetser, E. E. (1990). From etymology to pragmatics. Cambridge: Cambridge University Press. Sweetser, E. ( 1996). Mental spaces and the grammar of conditional constructions. In G. Fauconnier, & E. Sweetser (Eds.), Spaces, worlds and grammar (pp. 318-333). Chicago: The University of Chicago Press. Trabasso, T., & Sperry, L L ( 1985). Causal relatedness and importance of story events.

Journal ofMemory and Language, 24, 595-611. Trabasso, T., & Van den Broelt, P. (1985). Causal thinking and the representation of narrative events. Journal ofMemory and Language, 24, 612-630. Traugott, E. C. ( 1988 ). Pragmatic strengthening and grammaticalization. Proceedings of

the Berkeley Linguistics Society, 14,406-416. Traugott, E. C., & Heine, B. (Eds.). (1991). Approaches to grammaticalization. Amsterdam: Benjamins. Van den Broek, P., Young. M., Tzeng, Y., & Linderholm, T. (1998). The landscape model of reading: Inferences and on-line construction ofa memory representation. In H. Van Oostendorp, & S. R. Goldman (Eds.}, The construction of mental representations during reading (pp. 71-98). Mahwah, NJ: Erlbaum. Van Dijk, T. A., &Kintsch, W. (1983). Strategiesofdiscoursecomprehension. New York: Academic Press. Verhagen, A. (1995). Subjectification, syntax and communication. In D. Stein, & S. Wright (Eds.), Subjectivity and subjectivisation. Cambridge, Cambridge University Press. Verhagen, A. (2000). Concession implies causality, though in some other space. In Couper-Kuhlen & Kortmann (Eds.), Cause, Condition, Concession, Contrast. Cognitive and Discourse perspectives. 361-380. Berlin: Mouton de Gruyter. Vonk, W., Hustinx, L. G. M. M., & Simons, W. H. G. (1992). The use of referential expressions in structuring discourse. Language and Cognitive Processes, 7, 301-333.

2.5

SECTION

1

Accessibility in text and text processing

Both in modeling discourse structure and in processing studies, accessing referents and the linguistic coding of accessibility are central issues. The three chapters in this section show that these phenomena need to be studied from a linguistic and a psycholinguistic angle. In Chapter 2, Ariel discusses the linguistic means of reference to discourse entities. Her central claim, based, among others, on corpus analysis, is that language users do not arbitrarily switch between different referential forms, such as pronouns and full NPs, but that they show a systematic pattern. As in her earlier work, Ariel argues that the form of referential expressions can be explained by means of accessibility theory: the less accessible a referent is, the more elaborate the referential marker used by the language user. She gives an overview of her own Accessibility Theory, re-explaining aspects that have sometimes been misunderstood, and elaborating it with new findings. Then, her theory is compared to other accounts of reference. She explicitly addresses the issue of the cognitive motivation behind the theory, and discusses the relationship with psycholinguistic work on anaphoric reference. In Chapter 3, Gaddy, Van den Broek and Sung use a typically psychological framework to model allocation of attention in what they call the Landscape Model of Reading. The model addresses the issue how various text characteristics (linguistic, discourse-structural) guide the reader's attention during reading and how they affect the mental representations readers construct of the discourse. The referential forms studied by Ariel (see Chapter 2) are one of these textual devices determining the workings of the model. The authors claim that theirs is an adequate model of the on-line reading process. The chapter once again underlines the importance of the notion of 'activation' as an explanatory concept in understanding the reading process and its result: a coherent mental representation of the information expressed in the text. Activation is also a key concept in Chapter 4, by Giora and Balaban. This chapter deals with accessing literal and non-literal (or metaphorical) lexical meaning in text production, such as the boys' fight in the schoolyard (literal) vs. the union's fight against the government (non-literal). On the basis of experi-

2.8

Section 1: Accessibility in text and text processing

mental research, the authors defend a modified version of the view that lexical processes are autonomous, namely that salient meanings of a word (i.e., those that are coded into a language) are always activated, regardless of whether the activated meaning is contextually appropriate. Empirical evidence for the socalled 'graded saliency hypothesis' comes from a rating experiment in which participants were asked to indicate how they had understood a particular word. It is shown that even if the surrounding discourse strongly evoked a 'figurative' meaning, participants activated the coded meaning of the target word as well. The chapter relates to Gaddy et al. 's chapter in its dynamic conception of linguistic meaning. Furthermore it relates to Schilperoord's claim (Chapter 12) that in text production the planning takes place in a modular way.

CHAPTER

2

Accessibility theory: an overview Mira Ariel Tel Aviv University

Accessibility theory (Ariel 1985a, 1990 and onwards) describes how human language, specifically, the referential system, is responsive to facts about human memory, where memory nodes are not equally activated at any given time. Some are highly activated, others are only mildly activated, and in between, the range of activation is infinite in principle. Most memory nodes are of course not at all activated. Yet, speakers may wish to refer to Given (i.e., familiar) pieces of information, regardless of their current degree of activation for the addressees. Accessibility theory offers a procedural analysis of referring expressions, as marking varying degrees of mental accessibility. The basic idea is that referring expressions instruct the addressee to retrieve a certain piece of Given information from his memory by indicating to him how accessible this piece of information is to him at the current stage of the discourse.• To be sure, most referring expressions simultaneously also contain some conceptual content which contributes to the retrieval process. For example, she simultaneously means 'highly accessible', and 'female and singular', and the friend implies that the entity is 'of a relatively low degree of accessibility' because it is a definite description, and also that it is a 'close acquaintance', etc. But some linguistic entities (e.g., zeroes) are purely procedural, namely lacking any content, only marking a specific interpretative procedure. 2 Yet others do carry a conceptual meaning, but are undistinguishable from other expressions in terms of the concept they convey (e.g., it and this/ that). These, I have argued, are undistinguishable with respect to the description they provide for the intended referent (an inanimate object). They can only be distinguished from each other in terms of the processing instruction they mark: personal pronouns mark a higher degree of accessibility than demonstrative pronouns. Here is an example where the speaker repairs an it to a that, not because that better describes or identifies the entity referred to, and not because that is accompanied by some deictic gesture. Rather, it codes too high a degree of

30

Mira Ariel

accessibility for the word awakened: 3 ( 1) Melissa:

Ron: Frank:

Well, I'll say awakene~, cause that1's what I have written down . . . . (Sniff) ... Just watch, He'll put a note by it;.. note by that1• . . . I really like that word1 Melissa (Household).

Although speakers mark as Given information units packaged as NPs, as whole propositions, as VPs, and as verbs (see Ariel1985b, 1988b), and although all of these Given pieces of information may vary in degree of accessibility, I have concentrated on the intricate retrieval process involved in referential acts performed by NPs (Ariel1985a, 1988a, 1990 and onwards)." The structure of this paper is as foUows. Section 1 presents the basic claims and findings of accessibility theory, emphasizing aspects of the theory which have sometimes been misunderstood ( 1.2, 1.3). Section 2 sums up recent research which corroborates and further develops accessibility theory. In Section 3, I argue that while accessibility theory is cognitively motivated, accessibility marking constitutes a linguistic proper phenomenon. Section 4 compares accessibility theory with other theories of reference, and Section 5 lists open questions for linguistic and psycholinguistic research of anaphora.

1.

Accessibility theory: Basic claims and findings

1.1

Introducing accessibility theory

Accessibility theory argues that context retrievals of pieces of information from memory are guided by signaling to the addressee the degree of accessibility with which the mental representation to be retrieved is held This assumption entails that speakers do not guide addressees' retrievals by referring them to the correct "geographic" source which serves as the basis for assuming that the information is Given. In other words, languages do not provide us with conventional codes specialized for ( 1) information retrievable from our general encyclopedic knowledge (e.g., there existed an entity by the name of Simone de Beauvoir), for (2) information extractable from the immediately available physical context (e.g., there exists a table between us), or for (3) information

Accessibility theory

previously mentioned in the discourse (e.g., that the speaker has a dear friend). I have argued against Clark and Marshall (1981) that proper names (e.g., Simone de Beauvoir) are not specialized for retrieving general encyclopedic information, that demonstrative pronouns (e.g., this table) are not specialized for retrieving physically salient objects, and that personal pronouns (e.g., she) are not specialized for retrieving from the preceding linguistic context. All of these referring expressions can and do retrieve from all three "geographic" contexts (see Ariel 1988a, 1998b ). Instead, each referring expression codes a specific (and different) degree of mental accessibility (Ariell988a), and referring expressions are actually accessibility markers, i.e., expressions cueing the addressee on how to retrieve the appropriate mental representation in terms of degree of mental accessibility. Based on distributional findings re such distinctions I have suggested the following accessibility marking scale (see Ariel 1990, p. 73), which proceeds from low accessibility markers to high accessibility markers: (2) Full name+modifier >full name> long definite description> short definite description 5 > last name > first name > distal demonstrative+ modifier > proximate demonstrative+ modifier > distal demonstrative+ NP >proximate demonstrative+ NP> distal demonstrative (-NP) >proximate demonstrative (-NP) >stressed pronoun+gesture > stressed pronoun > unstressed pronoun > diticized pronoun 6 > verbal person inflections > zero A point that needs clarification is the relevant domain in which degree of accessibility is assessed. What is the basis for our determining that a specific mental representation is of high or oflow degree of accessibility (in the absence of direct tapping of the brain)? One potential source which determines degree of accessibility is the physical context of the speech situation. Another is the discourse world, where the discourse topics and other entities mentioned or reliably predicted to be relevant to the discourse at hand can receive high or low degrees of accessibility according to their discourse role. I have argued that it is the discoursal rather than the physical salience of the entities involved which determines the degree of accessibility assigned to particular mental representations (Ariel 1998b; see also Webber 1991). Although the physical context does affect the discourse model of the speakers, mental representations are a direct product of our discourse model only. One piece of evidence for this claim comes from references to the speaker. Although a (Hebrew) speaker can (almost) always refer to herself by a personal pronoun, she can cliticize the

31

32

Mira Ariel

pronoun, or she can use verbal person agreement (with a zero subject) or even zero subject alone (all are higher accessibility markers than a full pronoun), provided she is maximally accessible within the current discourse. In other words, whereas the physical accessibility of the speaker in the real world does not change in the course of the conversation, her discourse role and prominence in it may. It is the latter which determines whether the speaker can be referred to by higher accessibility markers than a pronoun. Indeed, Rieder and Mulokandov's (1998) analysis of two television interviews corroborates my initial findings (Ariel1990, 1998b ): Even turn-initial position shows an accessibility distinction re the speaker: The fuller form is preferred turn-initially, the shorter forms are preferred when immediately preceded by at least two previous mentions. I have argued that the form-function correlations on the accessibility marking scale (namely, which referring expressions code which degree of accessibility) are not arbitrary. Three partially overlapping criteria are involved: Infonnativity (the amount of lexical information); rigidity (the ability to pick a unique referent, based on the form); and attenuation (phonological size). The prediction is that the more informative, rigid and unattenuated an expression is, the lower the degree of accessibility it codes, and vice versa, the less informative and rigid and the more attenuated the form is, the higher the accessibility it codes. Thus, "true" zero subjects (as in Chinese), verbal person agreement (as in Italian and Hebrew), cliticized pronouns (as in Hebrew and English), pronouns, stressed pronouns, demonstrative and definite NPs and proper names (of all kinds) are each specialized for (slightly) different degrees of accessibility, which accounts for their different discourse distributions. Based on previous work (most notably Sanford & Garrod 1981; Giv6n 1983) and my own, I have suggested that we can tap the degree of accessibility associated with a given mental representation at a given moment by considering properties of the mental representation/ antecedent (not necessarily a linguistic one), as well as the relationship between the antecedent and the anaphor (the accessibility marker). Thus, the more salient the antecedent the more highly accessible it is deemed. I distinguished between global discourse topics (highest degree of accessibility), local discourse topics (relatively high degree of accessibility) and non-topics (relative low degree of accessibility) in this connection, as well as between the speaker and the addressee (high degree of accessibility) versus a referent which is neither (a 3rd person- a relatively low degree of accessibility). Another salience distinction depends on the automaticity/stereotypy of the inference required in generating a Givenness status


for an entity. Inferred entities come in different degrees of accessibility (see Sanford & Garrod 1981; Ariel 1985a, 1990, 1996; Oakhill, Gamham, Gemsbacher & Cain 1992; Gundel, Hedberg & Zacharski 1993; Garnham, Traxler, Oakhill & Gernsbacher 1996; Matsui 1998). Frame induced entities (e.g., waiters in restaurants) are more accessible than inferable entities which are not salient or necessary in a specific frame (e.g., umbrellas in restaurants). In fact, some inferential information is indistinguishable from explicitly mentioned pieces of information (see Beeman & Gernsbacher, Ms., and references cited therein). This accounts for the difference between initially referring to the waiter without any anchoring, versus to Maya's umbrella (the necessity to anchor the umbrella to a Given entity) in the context of a restaurant story. Another factor influencing the relative degree of accessibility of an antecedent is competition on the role of antecedent (see Clancy 1980; and see O'Brien & Albrecht 1991, for experiments establishing that we initially access multiple antecedents). The more potential antecedents there are, the lower the degree of accessibility each is entertained with. The relationship between the antecedent and the anaphor, the degree of their unity, or cohesion (Ariel1990), can be tight, in which case the degree of accessibility of the relevant mental representation is higher, or it can be loose, in which case degree of accessibility is lower. Such a relationship exists between linguistic units primarily (an antecedent and an anaphor ). The distance between a previous mention of the same referent and the current mention is an obvious measure of an accessibility distinction. The larger the distance separating different mentions of the same mental entity, the lower the degree of accessibility with which the mental representation is entertained. But distance is not necessarily measured by words. Paragraphs and episode boundaries create a distance, despite the linear continuity (see Ariel1990; Clancy 1980; Sanford & Garrod 1981; Tomlin 1987). At episode boundaries, people have difficulties accessing prior information (Beeman & Gernsbacher, manuscript; Sanford & Garrod 1981). Similarly, units (clauses) more cohesively linked entail more dependency in their interpretation, so that material from one clause is more readily available for the interpretation of another. Such constructions create higher degrees of unity and hence, accessibility. Looser connections, on the other hand, entail more independence in the processing of each clause, in which case there is less availability (accessibility) of material of one clause for the interpretation of the other. Such differences account for the different anaphoric patterns observed for subordinations (higher degree of accessibility- repeated proper names are clearly dispreferred) versus coordinations (a lower degree of

33

34

Mira Ariel

accessibility- repeated proper names are at least possible), nontensed (infinitival) versus tensed clauses (the former show a preference for zero subjects), and restrictive versus nonrestrictive relative clauses (the latter favor resumptive pronouns more often than the former). In sum, I have argued that referring expressions are chosen according to the assessed degree of accessibility of the mental entities corresponding to them. Degree ofaccessibility depends on factors related to the inherent salience of the entity and on the unity between the antecedent and the anaphor. In addition, the conventional degree of accessibility coded by referring expressions is motivated by their relative informativity, rigidity and attenuation. 1.2

Accessibility as a complex concept

I have tried to emphasize that assessing degree of accessibility is a complex matter, since multiple factors are involved. It is the complex concept of accessibility which determines referential form, and not any single factor. This is why when we examine any one factor of accessibility, the results are significant, but far from absolute (see Ariell999, on resumptive pronouns and Garcia 1996, on reflexives and pronouns in Spanish). Accessibility factors may converge on pointing to a high (or low) degree of accessibility, as when the speaker or the addressee (highly accessible) is also the global discourse topic (highly accessible), or when the discourse topic has been recently mentioned (high accessibility), and/or has been mentioned numerous times (high accessibility). However, although distance, for instance, is a crucial factor determining degree of accessibility, it cannot be taken to perfectly represent the overall degree of accessibility involved. For instance, pronouns (high accessibility markers) can sometimes (over25% in mydata-seeAriell990, p.l8 for sources), refer to mental entities last mentioned in a previous rather than a current paragraph (entailing a lower degree of accessibility in terms of distance). The reason is that these distant references are mostly references to the global discourse topic (92%). Discourse topics can maintain a relatively high degree of accessibility despite the larger distance. The same clash would explain why it is that when two entities are introduced (e.g., Maya; kissed Rachel). the first mention topical but relatively distant NP (Maya) is later coded as an unstressed pronoun (a higher accessibility marker), whereas the more recent, nontopical NP (Rachel) is coded by a stressed pronoun (a lower accessibility marker) (e.g., and then she/SHEj..). Similarly, Brennan (1995) found that nonprominent entities (less salient) were referred to by full NPs (low accessibility markers) rather than by


pronouns (high accessibility markers), despite the recency of their mention (high accessibility). The more previous mentions an antecedent has enjoyed, the higher its accessibility. Still, discourse topics can usually be referred to by high accessibility markers despite a low count of previous mentions. Perhaps this is due to the fact that some entities, discourse topics more than others, are inferred to be present even when explicit mention is lacking (see also Grosz, Joshi & Weinstein 1995; O'Brien &Albrecht 1991;).7 Maes andNoordman (1995) find that54.2% (682) of their demonstrative+ NP expressions ( 1259, in Dutch) are a second mention reference to an entity (or proposition) just mentioned (in the same or previous sentence, for the most part). Since a modified demonstrative pronoun is an intermediate accessibility marker, but the distance factor points to a high degree of accessibility, this finding appears to be a counter-example to accessibility theory. However, a newly introduced discourse entity (and even more so when the antecedent is complex- a proposition) is not instantaneously highly accessible enough for further references by high accessibility markers (as they note themselves in note 15; see Ariel1990; Clancy 1980; Du Bois 1980; Du Bois & Thompson 1991; Tao 1996).8 Similarly, tum-initial positions are expected to contain lower accessibility markers (they form a discourse break). Rieder and M ulokandov ( 1998) then explain the surprisingly high occurrence ofzeroes and cliticized pronouns for first person references in initial turn position (in Hebrew) by noting that the preceding tum was a question addressed to the speaker in an overwhelming majority of these cases. We must not therefore mistake the individual factors contributing to degree of accessibility (e.g., tum-initial position) for criteria} conditions on linguistic usage. A study of resumptive pronouns versus gap usage in conversational Hebrew relative clauses (Ariel 1999) revealed that nonrestrictive relative clauses (relative lower degree of accessibility of the head when the relativized position is processed) contain many more resumptive pronouns than restrictive relative clauses (relative higher degree of accessibility). Still, two thirds of the nonrestrictive relative clauses contained gaps rather than resumptive pronouns. This might again seem a counter-example to accessibility theory, since an extremely high accessibility marker (the gap) is used when the degree of accessibility hypothesized between the head and the relativized position is relatively low. However, in dose to three fourths of these gapped cases, the relativized position is a subject (rather than an indirect object, for example). Subject position is assigned to prominent entities, ones which are of a higher degree of accessibility.

35

36

Mira Ariel

In general, I found that a combination of accessibility factors (head complexity, distance, grammatical role of the relativized position, and restrictiveness) predicts the occurrence of gaps and resumptive pronouns better than any one of the above factors. This is so because any particular instance may involve values for both high and low accessibility, and it is only the assessed combination of these factors (as well as others) which determines the overall degree of accessibility dictating the form chosen by the speaker. Note also that it is only the general concept of degree of accessibility which can account for why a variety of factors which seem unrelated to each other (e.g., the restrictiveness of the relative clause, grammatical role, whether the head is long or short, etc.) all encourage or discourage the usage of a resumptive pronoun. Garcia ( 1983, p. 203) similarly shows how the higher accessibility marker is chosen for cases where the antecedent is a nearby grammatical subject, as well as for cases in which the antecedent is contextually salient, although grammatically speaking, there is no similarity between these two conditions. The form-function correlation for reflexive pronouns (e.g., herself> also demonstrates how degree of accessibility cannot automatically be determined. I had probably mistakenly classified reflexive pronouns as higher accessibility markers than pronouns (Ariel 1990), because their antecedents are (for the most part) very local (within their C-command domain). Note, however, that reflexive pronouns are less attenuated than pronouns (in English), and should have therefore been relatively lower accessibility markers under accessibility theory. Now, this marking exception could be explained by reference to the high frequency of pronouns versus reflexives (unmarked forms tend to be short), but I think an alternative explanation, one based on the historical development of(English) reflexives (see Paltz 1977; Keenan 1994), shows them to be lower accessibility markers than pronouns, despite the fact that they are locally bound, whereas pronouns are locally free. Reflexive pronouns within the C-command domain are basically pronouns referring to unexpected entities. Their antecedents are quite accessible in some absolute sense, but they are not expected (and therefore accessible) in the specific role they actually occur in (which is coded by the reflexive). The same applies to the contrastive pronouns in Dutch, as analyzed by Comrie ( 1994). It is illuminating to compare locally bound coarguments of the same verb (accusatives and objects of prepositions) with locally bound nonarguments in Old English (Keenan 1994). The latter are referentially dependent on the subject, and they mark a high involvement of the referent in the event (e.g., The king walked him to London). These high involvement nonarguments are coded


by pronouns. But coarguments which are objects of verbs of serious personal harm (e.g., threaten, kill) are invariably coded by pronouns + self). The reason is that coarguments of the same verb (especially of the above kind) are expected to be disjoint in reference, since we are expected not to hit or threaten ourselves (Faltz 1977; Farmer & Harnish 1987). Hence, the argument is expected notto be co referent with the subject. The referent of the subject in such contexts is then of a lower degree of accessibility in the object role, and a lower (less attenuated) accessibility marker (a pronoun+ self) is used. We see that it is indeed the degree of expectation (accessibility) for subject coreference that matters (rather than argument versus nonargument role) when we compare the accusatives of bodily harm with those of verbs of grooming (e.g., dress). Unlike the former, the latter do create a high expectation for subject coreference (we are expected to dress ourselves), hence the accessibility of subject coreference for the object is high, and indeed they take regular pronouns, rather than reflexives in Old English (e.g., She dressed her). What is crucial for accessibility theory at this point is that we realize that a relatively lower accessibility marker (the reflexive) can grammaticize for a syntactic context that other things being equal is considered to be a very high accessibility context, namely the C-command domain. When the entity (although highly accessible) is not predicted to appear in a certain role, its degree of accessibility is (relatively) low, despite the short distance from the previous mention, and despite the fact that its previous coding marks it as highly accessible (subject). For the same reason, namely the complexity involved in accessibility assessment, I believe Giv6n (1992) was too hasty in his conclusion that accessibility is reducible to a binary distinction in language. Giv6n finds that the definite descriptions in his data retrieve antecedents which occurred at a variety of distances, unlike zeroes and pronouns, which retrieve discourse entities mentioned 1-2 clauses back for the most part. We should remember, however, that accessibility cannot be established on the basis of one factor (distance in this case) and that definite descriptions do not actually constitute a homogenous category of referring expressions in terms of degree of accessibility (see Almor 1999; Ariel1990, 1996}. First, the fact that definite descriptions retrieve antecedents from many distances can be explained by reference to other factors involved in assessing degree of accessibility: grammatical role (i.e. subject versus nonsubject}, degree of discourse salience (topicality), paragraph and frame boundaries, the number of previous mentions, etc. In other words, one should examine the "exceptional" short-distance definite descriptions and establish that they do not actually code a low degree of accessibility despite the

37

38

Mira Ariel

short distance. Only highly accessible entities (as measured by a comprehensive assessment of degree of accessibility) which are coded by low accessibility markers nonetheless constitute counterexamples to the theory. Toole (1992, 1996) checked such cases and found that the majority of these can be explained within a more complex assessment of degree of accessibility (plus intended divergences produced in order to generate special implicatures - see 2.1 below). The accessibility factors considered by Toole were: (1) Distance (whether anaphoric expressions corefer with antecedents in the same or in the last proposition, in the same or in a previous episode); (2) Topicality (how many times the antecedent was mentioned in the last four propositions); (3) Competition (how many matching intervening entities there were between the last mention of the antecedent and the anaphor). Second, when I divided definite descriptions into lower and higher accessibility markers according to their degree of informativity, consistent distributional differences were discernible. In Ariel ( 1990, p. 44), I presented data which showed that whereas the majority (78.2%) of definite descriptions composed of 1-2 content words were discourse anaphoric (higher accessibility), the majority (65.3%) of the definite descriptions which contain 3+ content words were first mentions (lower accessibility). In Ariel (1996) I distinguished between definite descriptions of 1, 2, and 3+ content words, as well as definite NPs + relative clauses, all as introducers of new discourse entities. Indeed, the lower the accessibility of the entity introduced, the more informative the referring expression was. 9

1.3

On the nonexclusivity of accessibility considerations

While accessibility considerations are a central aspect of referential choices, they by no means exhaust the selection process. Contextual assumptions must be relied upon in order to ascertain that a referential (rather than attributive or generic) use has been made (see Mueller-Lust & Gibbs 1991 ). Relevance-based considerations help select among equally accessible potential referring expressions, such as my neighbor, the mayor, Mark and the idiot Inferential processes are also crucially involved in determining reference, as in Horrified, she snatched the mea~from the do~ and threw i~ into the fire (from McEnery & Thomas 1992; see Ariell990, Part III; Comrie 1988b; Gundel et al. 1993; Gundel & Mulkern 1998; Matsui 1998, inter alia Sanford & Garrod 1981 ). Such considerations may sometimes even dictate violations of accessibility theory for special effects (see Ariel 1990, Part Ill). 10 Thus, I have argued that women and minorities are consistently referred to by higher accessibility markers than are called for given


the specific (relatively low) accessibility assessment (initial references), because speakers are not careful enough in making sure their addressees actually identify the referents intended. The dearest example for this phenomenon is the wide-spread use of first names for women and minorities (for data and analysis of references to the 'other', see Ariell990, 9.2; and see Mulkern 1996, p. 247). Such violations are very much socially and culturally bound. Indeed, there is a rich anthropological linguistics literature on naming patterns in different societies, which shows how the universal accessibility marking scale is embedded in social norms. I here mention one such example of a cultural difference, from the Nayaka, a hunter-gatherer group in India. Bird-David ( 1995) finds that names are not the rigid designators we usually think of. They function quite differently, and are rarely used in this society. Children are referred to as 'girl' or 'boy', or in relational terms, e.g., 'daughter', even by non-parents. Adults are mostly referred to by kin terms, which is a mark of intimacy (rather than their names). It is mostly adolescents that are referred to by nicknames. But names are not even necessarily constant in a person's life. People may have a number of names simultaneously. Bird-David notes that her informants sometimes asked people what they are called these days, even though members of the community live in extreme proximity to each other, and are in constant contact, even in their home-huts.u In contrast, Downing (1996, p. 130) argues that bare proper names are co-recognitionals, and are used when "present in the territory of information ofboth participants" (of the conversation). Only when the referents do not meet this condition other referring expressions are used (e.g., this+ proper name). It remains to be seen how general this principle is (I had independently made similar claims for Hebrew- Ariell990, pp. 203-206), in view ofthe Nayaka pattern. Note that although accessibility theory defines the relevant degree of accessibility to be that of the addressee as assessed by the speaker, a speaker may pretend to speak for another, and she then has to assume the degree of accessibility of the entity as it would have been assessed by that speaker. This is what happens in example ( 1) above when Frank is assuming the teacher's identity who will be reading Melissa's paper. He refers to awakened by a lower accessibility marker, that word, since the teacher will have no basis for assuming that the word is highly accessible to Melissa when reading his comments. Following Kuno ( 1987), I have also claimed that higher accessibility markers are used to code the character whose point of view is reflected in the discourse (Ariel 1990, pp. 203-204). The examples below demonstrate this

39

40

Mira Ariel

clearly. (3) and (4) are excerpts describing the same rape by two Hebrew newspapers (Haaretz and Maariv). Both newspapers relied on the same source: the police record, and hence the extraordinary similarity. Note that Haaretz and Maariv differ in that only Haaretz clearly adopts the victim's point of view. This can be seen from: a. the choice of subject role in describing the meeting between the rapist and the victim (Haaretz chose the victim, Maariv chose both the rapist and the victim), b. the choice of verb for the rapists expressing interest in having sex with the victim (demand in Haaretz, asked in Maariv), c. the addition of'as a result' adverbial in Maariv, making the rapes appear to be the result of the victim's refusal: (3) i.

ii

In the complaint the woman; claimed that on May 2, 0; met Roter1 ••• Then the twoi+k demanded from her; to have sex with themi+k" According to her,, when o, refused, Roter1started punching htrj ... (Haaretz, 5.17.1995).

(4) i.

According to the police, Roter1and the rape victim; met in the beginning of May ... ii. . .. and at a certain point 01+k asked the rape victim; to have sex with themi+k" This one; refused, and as a result, the twoi+k cruelly raped

hetj ... (Maariv, 5.17.1995). Note that in both papers the victim and the rapist are initially introduced (in this narrative) by a low accessibility marker, as is appropriate. It is in (ii) that the difference in point of view clearly shows itself in referential forms as well: In Haaretz the victim is coded by 0, and the rapist by a proper name (a low accessibility marker); In Maariv the rapists are coded by 0, and the victim is referred to by a demonstrative pronoun (a mid accessibility marker). When comparing all the zero versus pronoun ratios in the two papers, Haaretz has 6 zeroes versus 2 pronouns for the victim, but Maariv has precisely the opposite ratio: 6 pronouns versus 2 zeroes for her (pronouns here include demonstrative pronouns). The newspapers do not differ with respect to the zero/pronoun ratio for the rapists (1.4 and 1.3 times more zeroes respectively). Similarly, Giv6n (1998) finds that in his own unpublished novel, for which he has two versions, one from the perspective of one character and one from the other's, all the full NPs were reserved for references to the character whose point of view was not being represented. 12 A few additional examples of why accessibility theory can only account for default referential choices follow. Du Bois ( 1991) discusses what he terms


analogue reference, namely, cases where the speaker refers to X, but intends the addressee to derive from it also a reference toY as a conversational implicature. Such references may not only violate the requirement to select referring expressions according to the degree of accessibility of the mental representation at hand, they may ignore the accessible/inaccessible distinction, referring by an indefinite NP to a referent previously identified by a definite NP (One of his examples has: The cook. followed by a cook.>· Hakulinen ( 1987) argues that Finnish speakers avoid personal references, and thus, generic zero references are by now almost completely conventionalized as first person references. The next, originally Hebrew, example shows the speaker oscillating between 1st and 3rd person references to himself: (5) But Ii insisted then. A person j? i? devoted two months and a half, O;w built a whole program, 0;1 1.! took care of a budget, it is not as if Minister Katzav gave mei, I took care, I went ... (Mudai, TV interview, 2.11. 1998, from Mulokandov and Rieder 1998). The speaker here is dearly understood to be speaking of himself, but he is trying to create the impression that he refers to himself from an objective, "other", rather than "self' point of view. Hence the 3rd person "inappropriate" generic references to an indefinite person, which combine with predicates which unequivocally describe his and only his actions. Sanford and Moxey ( 1995) show that despite the theoretical (high) accessibility of some discourse entities, they are not easily referred to: (6) In the garden, I saw a young girli kicking a tree;. 1I looked at themi+i for a while (Sanford & Moxey's example 17, 17').

The ability to refer to even highly accessible entities is relevance-based. In a different context, the above example is perfectly acceptable, as they show. Indeed, accessibility theory takes it for granted that speakers have already decided on who to refer to, even though it is not at all a cut and dry decision, simply depending on "who did what to whom". This is clearly seen in the following example, where the speaker switches from his initial we to a he. It is certainly not because he has suddenly realized that the string is Yo Yo Ma's, rather than the whole dance group's: (7) Morris: We broke a STRING, Or HE broke a string (A TV interview with Yo Yo Ma and Mark Morris, a choreographer, Israeli TV, 7.9.1998).

41

43

Mira Ariel

Such choices of referential forms are made independent of accessibility theory, and they generate a whole array of conversational implicatures (in this case, how Mark Morris sees Yo Yo Ma as integral to his show). Maes and Noordman ( 1995) argue that a combination of a demonstrative pronoun and a noun phrase is used when the NP serves a predicational rather than an identification function. Such expressions are actually used in order to modify the addressee's representation of the intended entity. A lower accessibility marker is then used for such a marked purpose. The marked accessibility marker (there is a mismatch between the high degree of accessibility and the accessibility marker chosen) conveys information which the speaker directs the addressee to access in connection with the referent. For example, when the expression This Reagan follows the sentence Ronald Reagan is clearly suffering from memory failure, it is interpreted as 'the Reagan suffering from memory failure' (see their example 17). Taken together, Sections 1.2 and 1.3 argue that accessibility theory is not reducible to any one linguistic principle, because degree of accessibility is a complex psychological concept, and at the same time, that accessibility theory cannot exclusively account for referential choice and interpretation.

2.

Corroborating and enriching accessibility theory

I have presented many pieces of evidence for the applicability of accessibility theory in Ariel ( 1985a, 1988a, 1990 ). The reader is referred to those sources for original analyses of mine and for extensive references to other works which support the accessibility claim. More recent research has corroborated and enriched the applicability of accessibility theory. I here restrict myself to citing works I did not have access to when initially presenting accessibility theory (even though some had been published before). Much of the work to be mentioned was conducted independently of accessibility theory, some is a direct product of it. 2.1

General accessibility predictions

This section is dedicated to supporting general accessibility theory points: I quote works which argue for the replacement of formal conventions with what I would term degree of accessibility codings (Comrie, Garcia). I present distributional analyses of a variety of referring expressions (Saadi, Toole, Dolman),


all of which point that accessibility considerations (e.g., distance, competition, antecedent prominence etc.) are at work. Some also argue that genre differences do not refute the validity of the accessibility proposal (Toole, Dolman, Saadi, Kronrod and Engel). I present findings which confirm that degree of accessibility is a dynamic and complex notion which cannot be reduced to single factors (Gernsbacher et al., McKoon et al., Toole, Kibrik), and that it is not the only factor determining referential form (Kronrod and Engel, Almor). Finally, I present research which develops accessibility theory beyond my original proposals (Almor, Epstein). Comrie (1994) shows how Dutch contrastive pronouns refer to the less expected antecedent (lower accessibility on my account). While in most cases this means a nonsubject, it is not invariably so. Comrie argues that it has to be the non topic actually. Garcia ( 1983, 1996) argues that what seems to be a difference between subject versus nonsubject antecedent (for si versus el in Spanish) is a difference in "contextual obviousness", what I would term degree of accessibility. Indeed, in many cases, antecedents of high accessibility are also subjects, but Garcia demonstrates how non subject antecedents can take si anaphors, provided they are highly accessible (e.g., discourse topics). Assuming an accessibility distinction between si and el can also explain the higher frequency of si (the higher accessibility marker) with human versus nonhuman antecedents. In addition, Garcia (1996) specifically relates the fact that si does not distinguish for gender to its marking a higher degree of accessibility (my terminology) than the pronominal forms, which do. This corresponds to my Informativity criterion for accessibility marking. She also examines the role of competing antecedents and determines that the more salient the nonantecedent competitor, the lower the accessibility marker required for the antecedent. Finally, Garcia finds that the higher accessibility marker is preferred when the argument is governed. I have suggested the same for gap versus resumptive pronoun usage (Ariel 1999). Governed arguments are more predictable and hence more accessible. Saadi ( 1997) examined the English and Hebrew versions of one children's story and one adult story. Her findings support the accessibility predictions. All four sources shifted from a predominant use of zeroes and pronouns to lexical NPs as the distance from the last mention of the same entity was larger. The same applies to the factor of competing (intervening) referents: the more intervening characters mentioned, the higher the likelihood for a lexical NP to occur. All sources also distinguished between the main character (more salient, mentioned more times) and a secondary character, so that the main character

43

44

Mira Ariel

was referred to by high accessibility markers much more often than the secondary character. In both languages there was also a difference between the adult and the children's stories (more so in Hebrew), in that the children's stories contained more lexical NPs. Saadi suggests that this difference is due to the fact that adults writing for children assume that children's short term memories are not fully developed. McKoon, Ward, Ratcliff, and Sproat ( 1993) testify to the fact that degree of accessibility is a complex concept, the components of which may work in opposite directions, as when the antecedent is part of a compound (accessibility is lower) but the entity is topical (higher accessibility). Kibrik (1996) shows that degrees of activation dictate referential forms in Russian narratives. He also underscores the importance of the multiplicity of factors involved in determining degree of accessibility. Toole (1992, 1996) has convincingly shown how degree of accessibility, when measured by a few criteria (see 1.2) can explain the distribution of referring expressions in four discourse genres. Her conclusion is that despite the statistically reliable differences in referring expressions in different genres (see Fox 1987), accessibility theory can account for referential choices in all the written and spoken genres she examined. The statistical differences found stem from contextual factors which determine what types of discourse entities (in terms of degree of accessibility) tend to occur in discourses of different genres. In other words, a case by case analysis ofthe referring expressions used in all the genres revealed the same accessibility form-function correlations. Toole found that accessibility marking violations are only performed in order to achieve special objectives, e.g., low accessibility markers to clarify at the addressee's request, to define a term etc (see also Ariell990: Part III; Maes & Noordman 1995; Vonk, Hustinx & Simons 1992). Dolman ( 1998) too found no differences in referential choice between children from high and low socioeconomic backgrounds.' 3 Both groups complied with accessibility theory (degree of accessibility measured as a combined function of distance from last mention and the importance of the character to the story). Kronrod and Engel (1998) reached similar conclusions in their examination of referential forms used in newspaper headlines (see also Jucker 1996 ). They found no genre differences between the high brow subscription paper and the news stand popular papers, and between the different sections within the papers (front page, other news items, stories and sports). All the headlines showed a clear preference for intermediate accessibility markers (first names, last names and short definite descriptions). The fact that intermediate accessi-


bility markers predominate, despite the initial retrieval status, where low accessibility markers are expected, is explained by reference to Du Bois' (1985) notion of competing motivations. Headlines must be short and vague (in order to save space and arouse curiosity). High accessibility markers would have served that function best. But because the referents are also initial retrievals, and hence of a rather low degree of accessibility, a compromise is struck, and most of the referential forms are of an intermediate degree of accessibility. Here is (my own, originally Hebrew) illustrative example. Compare the referring expressions in the headline (a) with their counterparts in the opening sentence of the article (b): (8)

Arafati invited Kadafi, to pray in ]erusaletf\. when(\ will be the Palestinian capital b. The Palestinian authority chair, Yassir Arafati, invited the Lybian leader, Muamar Kadafi;• to pray in Emtern ]erusalemk, when this on'it will become the capital of the Palestinian state (Haaretz, 7.14.1998). a.

Gemsbacher, Hargreaves, and Beeman (1989) show how and why degree of accessibility of concepts shifts in the course of discourse. Sentential first mention entities are later entertained at a relatively high degree of accessibility due to comprehenders' assumption that first mentions are the discourse topic. 14 But mention in the last clause also facilitates retrieval, due to the high accessibility associated with the last clause processed (dark & Sengul 1979). As Gemsbacher et aL (1989) noted, these two facilitating conditions sometimes contradict each other. In a series of experiments measuring accessing speed at different processing stages, Gernsbacher et al. were able to establish that degree of accessibility is a dynamic phenomenon. Thus, an entity mentioned clauseinitially is less accessible than a more recently mentioned entity at first, but later, it gains in accessibility, as the units in which the two entities appear are integrated into one whole. In other words, recency is a short-term accessibility booster, whereas sentence-initial mention is a long-term accessibility booster. That the time in which we measure degree of accessibility is of the essence can also be seen in an experiment by Gemsbacher (1989). Almor ( 1999) embeds my initial proposal that referring expressions are "price tags" on processing effort in a more comprehensive system of processing assessment. Other things being equal, low accessibility markers take longer to process than high accessibility markers. Anaphors with a high informational load (roughly low accessibility markers) are easier to process when the antecedent is of a relatively low degree of accessibility (a nonfocussed antecedent).

4S

46

Mira Ariel

The same low accessibility anaphors are harder to process if the antecedent is highly accessible. This seems to echo the "repeated noun penalty" (see Gordon, Grosz & Gilliom 1993; Gordon & Scearce 1995). At the same time, I have argued (Ariell990, chap. 9) that intended divergences from appropriate accessibility marking are possible, but limited to cases where specific conversational implicatures are sought, above the referential function. I reasoned that the extra contextual implications justify the extra-processing cost. Almor formulates this intuition into a principle whereby "additional cost must serve some additional discourse function" (p. 5), such as adding some new information about the referent (and this is contra the "repeated noun penalty"). In this way, Almor integrates the cognitive approach with the pragmatic approach. Thus, high accessibility contexts can accommodate relatively low accessibility markers, provided increased contextual effects result. Almor then underscores the fact (convincingly illustrated also by Maes & Noordman 1995) that we cannot really account for the distribution of referring expressions by reference to the referential function of NPs alone. In fact, his experiments demonstrate that low accessibility markers are relatively easily processed, despite the high accessibility of the antecedents, provided they add some new information about the referent. Aimor ( 1999) is mainly interested in processing effort: He wants to calculate the ease of processing anaphors as an interaction of three factors: discourse focus (i.e., degree of accessibility), the amount of new information contributed by the anaphor, and the information load differential between the antecedent and the anaphor. Informational load is not equivalent to my Informativity criterion. It is calculated as the conceptual distance between the anaphor and the antecedent. In order to assess this difference, Almor draws distinctions among antecedents and anaphors not previously made by accessibility theory, such as between more general (and less informative) terms (e.g., the bird, a creature) and more specific (and informative) ones (e.g., the robin, an ostrich), between more versus less typical instances of a category (e.g., 'robin' vs. 'ostrich'). His experiments show different response times to different pairs of antecedent/anaphor according to these distinctions, some of them, even somewhat counter-intuitive (e.g., that the same anaphoric expression, e.g., the bird will be processed faster when anaphoric to an antecedent which is a less typical member in its category (e.g., an ostrich) than when it is typical (e.g., a robin), when both are focused. Almor thus adds another dimension to the antecedentanaphor relation that I did not discuss, that of conceptual difference. Finally, Epstein (l998b) extends the concept of (low) accessibility to in-


dude the accessibility of new discourse entities as well. His claim is that low accessibility characterizes the appropriate use of definite descriptions referring to entities which lack a previously stored mental representation, so that the addressee is instructed to construct a new representation, the definite article marking that the knowledge required for the construction is accessible. Such accessibility for non-Given entities can derive from the high prominence of the entity, from the fact that it is a frame-appropriate inferred role, or from the accessibility it has for a noncanonical narrator. 15 2..2.

Accessibility predictions pertaining to the type of antecedent

Recall that accessibility theory predicts that the higher the accessibility with which the mental representation is entertained, the higher the accessibility marker used to retrieve it (and vice versa for low accessibility). I present below recent findings corroborating this claim. Gemsbacher {1990) proposes the structure building framework, according to which when comprehenders are engaged in constructing mental representations for incoming information, their strategy is to build coherent structures, by first laying a foundation and then incorporating information that coheres with the foundation into it. Less coherent information makes comprehenders shift to a new constructed structure. According to Gemsbacher, two very basic cognitive processes are enhancement and suppression. These bear direct relevance for accessibility theory. Enhancement mechanisms elevate the degree of accessibility of memory nodes, suppression mechanisms reduce it. Enhanced entities "overshadow" and suppress the activation of other discourse entities. They are also more resistant to being suppressed by other discourse entities (see Gemsbacher and Jescheniak ms). One example of an enhanced entity is a "cataphoric" NP. Gemsbacher and Shroyer (1989) distinguish between NP forms as to degree of"cataphoricity", namely how marked they are for potential further references (see also Downing 1996; Sanford, Moore & Garrod 1988). 16 The assumption is that the way in which discourse entities are introduced (e.g., by an indefinite article versus by an indefinite this, in English) give rise to different expectations re further mentions (see also Giv6n 1992, about the interaction of grammatical role with marking by an indefinite this; Mueller-Lust and Gibbs 1991, on proper names; and Sanford et al. 1996, and Paterson et al. 1998, about quantified NPs as antecedents). Translated into accessibility terminology, "cataphorically" marked discourse entities become relatively more salient antecedents, because they occupy a privileged position among mental representations. They should therefore be

47

48

Mira Ariel

referred to by relatively higher accessibility markers, and they are. 17 However, NPs are not simply classified into+ versus- "cataphoric", i.e., as+ versusself-enhancing and other-suppressing. Some (e.g., contrastive stress) are more "cataphoric" than others (indefinite this), that is, they trigger a higher degree of activation, so their antecedents are more highly accessible. Gemsbacher ( 1989) also shows how the introduction of different, even new discourse entities suppresses the accessibility of current discourse entities, even if these have been established as topics before. I have referred to this phenomenon as competition, which, I argued, lowers the accessibility of all discourse entities (see Garcia 1983, p. 200, for why pronouns rather than short reflexives are sometimes used for the accessible discourse topic for this reason; Halmari 1996, p. 172, and Keysar et al. 1998, for why some competition is or is not a problem after all). Other nominal forms may be distinguished as to discourse prominence and hence to degree of cataphoricity. Halmari ( 1996) shows how zeroes, pronouns, demonstrative pronouns, proper names and definite descriptions signal different degrees of accessibility, by examining the grammatical role of their antecedents. The assumption is that subjects are used for more highly accessible entities than other grammatical roles are. And indeed, 98% of the zeroes she found had subject antecedents. The same applied to 72.5% of the pronouns, but the antecedents for demonstrative pronouns, for example, were evenly distributed among all grammatical roles. About 30% of the proper names and the definite descriptions refer to genitive antecedents (as opposed to 1% of the zeroes and 13o/o of the pronouns). Indeed, Gordon et al. (1993) and Gordon and Chan (1995) found that the "repeated name penalty" (using too low an accessibility marker) applies to subjects, but not to other syntactic statuses. Stebbins ( 1997) shows how some languages use number marking cataphorically (my term) only or preferentially for establishing new discourse entities, linking this usage to the high Informativity involved in nouns marked for number. Such languages may omit number marking in subsequent mentions. The same goes for noun classifiers and noun particles (see my interpretation of Hinds' (1983) findings re Japanese in Ariel1990, p. 90). Sproat and Ward (1987) and McKoon et al. (1993) (see also Greene, Gerrig, McKoon & Ratcliff 1994; Ward, Sproat & McKoon 1991) present similar findings. Sproat and Ward and McKoon et al. show how the way we present a concept in the discourse affects its degree of accessibility, even if it is not actually introduced as a discourse entity. This in tum affects referential options and ease of processing, as measured by reading times. For example, McKoon et al. compare anaphoric references to the non referential 'deer' in


deer hunting versus the referential 'deer' in hunting deer. Indeed, when the discourse creates a high degree of accessibility, an "illicit" (nonreferential) antecedent is properly referred to even by the high accessibility pronoun. McKoon et al. then conclude that syntactic factors contribute to the determination of degree of accessibility (and the same could be claimed for subject position proved crucial by Halmari (1996)). I tend to think the other way round, namely that it is degree of discourse prominence which influences both syntactic role and degree of accessibility (see also Gundel et al. 1993). In other words, more important entities will be introduced as referential, rather than as nonreferential, as subjects rather than as nonsubjects. Oakhill et al. 1992 show how depending on the antecedent, conceptual anaphors (e.g., I need a plate. Where do you keep them?) are appropriate, though at some processing cost. In general, they show that depending on the degree of the accessibility of the antecedent, different referring expressions are appropriate. Garrod and Sanford ( 1982) and Albrecht and Clifton ( 1998) find that an entity coded as an NP conjunct constitutes an inferior antecedent (less accessible) so references to it take longer to process. Alrnor ( 1999) demonstrates the role that focus plays in raising the degree of accessibility of an antecedent. Referents coded by focussed NPs and later referred to by anaphoric expressions were read faster than referents coded by nonfocussed NPs. Conversely, Alzheimer Disease damages working memory. Almor (in press) then explains why Alzheimer patients prefer references by lower accessibility markers (repetitive definite descriptions) over the more context-appropriate high accessibility markers (pronouns). Arnold (1997, to appear) corroborates Almor's (1999) findings, but then seeks to explain the apparent puzzle of why topic (old information) and focus (new information) both facilitate reference interpretation. The reason is that despite the differences in the nature of the information they themselves convey, both elevate the degree of accessibility of the entity they are associated with. Arnold also finds that the global topic has a stronger effect than focus or local topic. This is important in that it shows that we cannot substitute the complex concept of degree of accessibility with simple rules such as "if anaphoric with a subject, or with a focussed NP, or if a sentence topic, then the mental representation intended is to be coded by a high accessibility marker" (see also Arnold to appear).

49

so

Mira Ariel

2..3

Accessibility predictions pertaining to the type of anaphor

Accessibility theory predicts that accessibility markers which are relatively uninformative, nonrigid and attenuated retrieve highly accessible mental representations (the opposite holds for low accessibility markers). The researchers here mentioned support this claim by pointing to the correspondence between degree of antecedent accessibility and the informativity, rigidity or attenuation of the anaphor. Fowler, Levy, and Brown ( 1997) note that the same conditions which encourage the usage of pronouns (high accessibility markers) also encourage the shortening of the pronunciation of proper names (thereby making them signal a higher degree of accessibility). Brennan ( 1995) found that subjects lengthened their pronunciation of pronouns (thereby turning them into slightly lower accessibility markers) when the antecedents were nonsubjects (a lower degree of accessibility). Downing (1986) argues that Japanese classifiers are used as anaphoric expressions, the degree of accessibility (my terminology) they mark being in between pronouns and lexical NPs. Because of their high informativity, classifiers can refer to relatively distant antecedents, and in contexts where there are intervening antecedents. Both contexts are indications of an intermediate degree of accessibility. 18 Garnham et al. ( 1994), Rinck and Bower ( 1995), and Cacciari et al. ( 1997) present evidence for the importance of gender marking (even if it is arbitrary gender)- my informativity criterion. 19 Mithun ( 1996) shows how prosodic cues affect the degree of accessibility coded by the same accessibility marker, a definite NP. She distinguishes between lexical NPs which occur in separate intonation units, those that do not, and those that occur in the more Given syntactic position (postverbally in Central Porno) with a specific intonation.20 Baker (1995) presents data showing that discourse prominence and contrast determine the appropriate usage of English free reflexive forms (i.e., unbound reflexives). Although he does not note this, a superficial count of the data he quotes reveals a difference between bare reflexives (relatively attenuated) and reflexives combining with pronouns and lexical NPs (less attenuated expressions). The former mark a higher degree of accessibility than the latter. 21 Thus, languages can utilize very many additional formal markings than I have originally listed (see Ariel1990, pp. 69-93, on the universality of the accessibility marking scale). Obviation, logophoric forms and switch reference systems come to mind (on the latter see Ariel 1990). Mulkern (1996, p. 245) notes how full names function differently from partial names (and see Ariel1990, pp. 36-46). The latter mark a higher degree


of accessibility because they are less informative. Lichtenberk's ( 1996) data can be adduced in support of my claim that proximate and distal demonstratives show an accessibility distinction, and not just a deictic distinction. When tracking discourse entities, the distal demonstrative + NP retrieved entities mentioned more than twice the distance of the antecedents of the proximate demonstrative+ NP. Brizuela (1997) shows that a demonstrative NP codes a higher degree ofaccessibility than a demonstrative pronoun+ a definite marker. Interestingly, the same distinction in Hebrew is merely a register difference. Once again, we see that length of expression, not necessarily accompanied by additional content, determines a lower degree of accessibility. Onishi and Murphy ( 1993) note that metaphoric references (too low accessibility markers) to the topic slow subjects down, even though the same metaphors do not slow them down when they do not serve as referring expressions. Beun and Cremers ( 1998) find that speakers use redundant information (making their expressions code a lower degree of accessibility) when referring to physically available objects, especially when the objects are out of focus (of a lower degree of accessibility). Mehudar ( 1996) analyzes the differences between proximate and obviative references in terms of degree of accessibility (see also Arnold, to appear). She corroborates my proposal (Ariel1990, pp. 7~91) that all languages distinguish between some degrees of accessibility in their referential system, although the distinctions need not be uniform. Thus, in some languages, the proximate is reserved for humans only, in Fox the entity has to be a human with a high social status even. In some languages, the sentence is the relevant unit for determining the choice between the obviative and the proximate (proximates refer intra-sententially, obviatives extra-sententially). Crucially, what remains constant across languages is that the proximate refers to the more highly accessible entity than the obviative, the different researchers referring to it alternatively as the one in the focus of attention, the central focus of the discourse, the focus of consciousness etc. 2..4 Accessibility predictions pertaining to the antecedent-anaphor relationship (unity) Accessibility theory predicts that higher accessibility markers should be used when the connection between the antecedent (unit) and the anaphor (unit) are tight (and vice versa for a loose connection). Recent work has supported this claim. Halmari (1996) presents data showing how paragraph-initial position

51

S1

Mira Ariel

creates a lowering of the accessibility marking for continuing discourse entities. The 90 cases where she found too low accessibility markers were also paragraph-initial. In a psycholinguistic experiment, Fowler et al. ( 1997) found that episode boundary was crucial for choosing longer anaphoric expressions (of a lower degree of accessibility). Khan (1999, p. 330) finds that in conversational Jewish Neo-Aramaic (of Arbel), the use of the grammatically optional subject pronoun marks the "clause as being separated from the preceding context by some kind of discontinuity or disjunction,. It is then relatively more frequent when there is a change of subject or in grounding, when the events are percieved as separate, and at the beginning of speech. I have argued that the way we refer to initial retrieval entities (loose connection to an antecedent) is also crucially dependent on degree of accessibility (Ariel 1996). Even initial retrievals, which are brand new to the discourse, can be more or less accessible. For example, frame-induced entities are highly accessible. They are coded by relatively higher accessibility markers, then. Chafe (1996, pp. 42-46) distinguishes between two types of inferred entities. He mentions in this connection a contrast between a stressed definite description and an unstressed one. The latter was used when the inferred entity was more automatically accessible. Ziv (1996) shows how when the inferred entity is stereotypically accessible (i.e., highly accessible) even pronouns can be used for initial retrievals. Maes and Noordman ( 1995) find that Dutch definite NPs refer to more remote antecedents than demonstrative NPs, initial retrievals for the most part. Section 2 has described recent research which supports the main tenets of accessibility theory, namely that referential choice is made by assessing the degree of accessibility of the mental representation retrieved, by considering the salience of the antecedent and the degree of unity between the antecedent and the anaphor.

3·

Accessibility theory and the grammar-pragmatics division of labor

3.1

The grammatical status of the accessibility principles

Accessibility theory correlates between specific referring expressions and their usage by reference to a cognitively motivated principle. In this respect, accessibility theory resembles recent attempts to reduce some anaphora phenomena to pragmatic principles, such as Reinhart (1983), Kempson (1984), Levinson (1987, 1991), Huang {1994), Gundel et al. (1993), and Ward et al. (1991). One


could then suggest that accessibility theory should be formulated as a set of extralinguistic inferences, connecting between linguistic forms and proper contexts on the basis of common sense inferences from their semantic meanings, rather than based on conventional form-function correlations (see Reboul 1997; Bach 1998). Such a move would minimize the contribution ofaccessibility theory to predicting referential form usage. Alternatively, one could maximize the role of accessibility theory, by arguing that the accessibility principles actually replace formal rules. Thus, while I made no attempt to replace the Ccommand domain by a cognitive concept (although I view it as the grammaticization of a highly accessible context), van Hoek ( 1995, 1997) uses accessibility theory to reduce C-command to a discourse concept which is sensitive to the prominence of the antecedent and the degree of unity between the antecedent and the anaphor. She thus reformulates Reinhart's C-command restrictions against full NPs being in the C-command domain of pronouns co indexed with them as an accessibility marking violation, where a low accessibility marker is used in a high accessibility context. It remains to be seen whether van Hoek's accessibility restrictions can actually replace the grammatical principle. For example, for the most part, the subject is indeed the most highly accessible entity, discourse-wise as well, which explains why the entities under its domain can be dependent on it for interpretation but not vice versa. However, what if a nondiscourse topic happens to be the grammatical subject (as in 11 below)? Can such a subject, since it is not so salient, be pronominal and coindexed with a full NP in its domain? I doubt that. I therefore see van Hoek's intriguing development of accessibility theory within the sentence (see Ariel1990, Part II originally) more as testifying to a plausible grammaticization path of accessibility considerations into grammatical rules. I myself have opted for a nonminimal nonmaximal position (Ariel 1990, 1994). In general, I have argued that the linguistic-extralinguistic division of labor does not neatly divide utterance interpretation according to the topics identified by linguists (see Ariel1998c, Ms). Such a division would posit that all aspects of reference interpretation belong either in the grammar or in pragmatics. Instead, most probably each and every linguistic form undergoes a dual interpretation procedure, whereby some aspects of its interpretation are linguistically derived, and others are associated with it extra-linguistically (i.e., inferentially). This is certainly the case for referring expressions, where a pragmatic theory (such as Sperber and Wilson's (1986) Relevance theory) has a major role to play (Ariel1990). Moreover, I have argued that while the formfunction correlations established by accessibility theory are cognitively well

53

S4

Mira Ariel

motivated for the most part (by the criteria of informativity, rigidity and attenuation mentioned above), some aspects of the accessibility scale (which expressions code which degree of accessibility) need to be grammatically stipulated nonetheless (see Ariel1990, pp. 76-87). Reference interpretation then is modularized between a linguistic (formal rules and accessibility degree lexically specified) and an extralinguistic inferential competence (see also Farmer and Harnish 1987).22 This much is perhaps obvious. What is less obvious is that the linguisticextralinguistic division does not coincide with the sentential-extrasentential division either, nor with the obligatory/optional dichotomy. Garcia ( 1983) and Ariel ( 1987, 1990) have emphasized that imposing on grammatical principles a sentential domain misses generalizations that hold both within and across sentences. Aissen ( 1997) confirms that the same principles account for obviation within and outside the clause. The span within which one third person referent must be proximate and all others obviative can be indefinitely large. The same applies to logophoric markers (marking the character whose point of view is conveyed) (see Hyman and Comrie 1981). In fact, in Plains Cree, the constraint that there must at least be one proximate marker is imposed on a stretch of discourse and not on the sentence, which may well not contain any (Comrie 1994). A switch reference system can also involve a relationship between nonadjacent clauses (Comrie 1994). Degree of accessibility, I have argued, is crucial both within and across sentences, and this is why when extremely high accessibility obtains, a zero can be used, whether its antecedent is sentential (e.g., a matrix antecedent in a control context) or extra-sentential (the discourse topic, for the most part). This is why Spanish si refers to the subject for the most part (a sentential highly accessible antecedent), but when it does not, it refers to the discourse topic (an extrasentential highly accessible antecedent). I also suggest that the grammarians' division into a grammatical (i.e., obligatory) versus a pragmatic (i.e., optional) "avoid pronoun principle" (for different languages- see Bouchard 1983; Hermon 1985) is unnecessary. Pronoun "avoidance" corresponds straightforwardly to avoiding too low an accessibility marker when the antecedent is extremely highly accessible. Precisely such variability between languages is expected if we assume that cognitive principles apply in all language, but they grammaticize only in some of the cases (see also Comrie 1994),23 The position I have adopted is that while there is a universal cognitive basis for referential form and usage, specific grammars translate the cognitive generalization somewhat differently (see Levinson 1987, 1991 for a similar point rea


pragmatic universal). There is then a role to the specific grammar of the language in determining referential forms and interpretations (see also Gundel et al. 1993). This division of labor between extragrammatical and grammatical principles explains the differences among languages (Cf. the use of zero subjects in English and Chinese, high accessibility markers in both languages) despite my assumption that mental representations are similarly accessible to speakers of different languages. Since each language only draws a certain number of accessibility distinctions, the choice of actual forms (to have or not to have a definite article, for example) and the precise accessibility domain carved for each referring expression (e.g., what to count as extremely high accessibility licensing a zero in Chinese and in English) may vary. Many languages allow (or dictate) zero for second person references in imperatives (i.e., where the entity referred to is highly accessible), as well as in control verb contexts, where depending on the type of verb, a high cohesion between the clauses creates a high degree of accessibility for the matrix antecedents (as in I didn't want to see him, or Like he wanted me to look at him- Jury). But these are grammaticized conventions, rather than directly motivated tendencies, in that they do not absolutely have to occur in each language. Greek and Sakapulteko Maya, for instance, do not have zero subjects in control contexts, and the latter does not force a zero in imperatives (DuBois, personal communication).lndeed, accessibility markers even show dialectal variability (see Garcia 1996, and Cameron 1997, on the variability of Spanish referential forms). Similarly, when we examine the usage of accessibility markers, we can see how formal and cognitive factors work in tandem in conditioning their occurrence. Hyman and Comrie (1981) argue that Gokana logophoric suffixes can always be anaphoric to subjects (a formal condition), but they can be anaphoric to an object, provided it is the source of the information (a pragmatic condition). Aissen (1997) claims that the (obligatory) choice of the argument to be coded as proximate depends on grammatical function, semantic properties and discourse salience (a mixture of formal and pragmatic conditions). In Ariel ( 1987) I proposed a scale ofaccessibility contexts, showing that formally defined contexts (e.g., where there is an obligatorily and uniquely determined antecedent, as in obligatory control contexts, or in wh- extractions) are on a par with cognitivelydefined contexts (e.g., the discourse topic) in that both may require or encourage the use of the same referring expression. Indeed, accessibility markers can be properly used by either fulfilling a formal criterion, or by fulfilling a pragmatic condition. For example, reflexive pronouns in English have an obligatorily formally defined condition: they have to be bound within

SS

S6

Mira Ariel

their C-command domain. But they can also be used without a sentential antecedent at all, when they are the subject of consciousness. Are these syntactic and pragmatic contexts really different? Note that within the sentence, reflexive pronouns have some contrastive residue. Not so outside the sentence. I suggest that what these uses have in common is an intermediate degree of accessibility. In the minimal (C-command) domain (high accessibility), only a contrastive (relatively low accessibility) entity is of an overall intermediate degree of accessibility. Across the sentence boundary (low accessibility), the subject of consciousness (high accessibility) is also entertained at an overall intermediate degree of accessibility (see Zribi-Hertz 1989 re long distance reflexives). Perhaps we can say that at some deep level these two contexts are cognitively the same. This will allow us to distinguish between potential grammaticizations (where similar degrees of accessibility get coded by the same accessibility marker) versus impossible ones (where different degrees of accessibility get coded by the same referring expression). Note, however, that languages may differ with respect to these two contexts. There may be languages which allow their reflexive pronouns in one but not in the other context. While the degree of accessibility associated with (long) reflexives may be intermediate for all languages, we need to specify for each language what mid accessibility translates into for the specific marker. 24 3.2

The grammaticization of accessibility markers

Grammaticization often entails a transition from a pragmatic, extralinguistic tendency to a grammatical, often obligatory rule. In Ariel ( 1998a, 2000) I have outlined such a historical path of change, leading from free pronouns to verbal person agreement inflections (more attenuated than pronouns, hence marking a higher degree of accessibility), arguing that such a change occurs for the forms referring to highly accessible discourse entities. Since speakers tend to shorten the forms referring to highly accessible entities (the criterion of attenuation), and since the speaker and the addressee (but not 3rd persons) are consistently highly accessible, it would be first/second person pronouns which consistently get shortened (as a pragmatic tendency). Shortening may lead to cliticizations and eventually to obligatory inflection (a grammatical rule). This is why most of the languages which manifest verbal person agreement markers restrict them to lst and 2nd persons. Person agreement development is a case where accessibility theory directly motivates bona fide grammatical morphemes (i.e., person agreement markers which are shortened free pronouns). I believe


that the creation of reflexive pronouns from independent pronouns and independent adjectival selfin English (see Keenan 1994) can be similarly motivated. Pronouns and modifier self were independent forms which consistently cooccurred in Old English in contexts where subject co reference was unexpected. A bare pronoun was then modified by selfin order to mark the relatively lower degree of accessibility of the subject by a longer referring expression. In fact, the same process can be seen in the current example: (9) Frankly, I'm torn my own selfas to which way to raise hell (Clark Reed, as quoted in The International Herald Tribune, Jan 2-3, 1999). In other cases, accessibility theory can motivate grammatical, even obligatory constraints on the distribution of various referring expressions. I have mentioned in this connection the binding conditions (Ariell987, 1990; see Keenan 1994; Levinson 1987, 1991 ). In Ariel (1999), I argue that whereas the distribution of zeroes and resumptive pronouns in relative clauses seems quite diverse among the languages of the world (e.g., some languages make zeroes obligatory with subject relativized positions, some allow or encourage resumptive pronouns only with nonrestrictive relative clauses etc.), accessibility theory can motivate the variability in grammatical patterns we actually find. These stem from frequent discoursal patterns which reflect the usage of zeroes and resumptive pronouns according to the degree of accessibility of the antecedent (the relative clause head) when the relativized position is processed. Zeroes are an option, or preferred, or grammaticized for extremely high accessibility contexts, and resumptive pronouns for relatively low accessibility contexts (e.g., syntactic islands). The precise use conditions are language-dependent, of course. It is important to note that while grammaticizations are often merely the freezing of specific realizations of accessibility distinctions into obligatory linguistic rules (e.g., for gaps and resurnptive pronouns in relative clauses), once some rule is part of the language, it may interact with other linguistic facts, and generalizations of patterns may then even obliterate the originally pragmatically motivated distribution (see Comrie 1983, 1988a). This is how I interpret Keenan's (1994) explanation for why English lost its high involvement pronouns (as in the king walked him to London). In Old English there were two contexts where anaphors were locally bound to their subjects: nonarguments of the high involvement type and contrastive coargurnents of the verb. As I have mentioned before, the latter are less accessible than the former, even though both pick the subject as antecedent. A situation where (locally

57

sa

Mira Ariel

bound) pronouns are used for high involvement subjects and reflexive pronouns are used for (locally bound) contrastive arguments is quite compatible with accessibility theory (see 1.2 above). However, once local binding becomes a characterizing feature of distribution, it is harder (less general) to have two types of referring expressions in the same, by now grammatically defined context (local binding with the subject). Perhaps this is why English dropped the "exceptional" use of high involvement pronouns. Hebrew did not. The same explanation applies to the less even spread of reflexive forms to objects of preposition (as in you making positive choices for yourself in your life- Death, versus that's his way ofdrawing your attention to him- Jury- see Bouchard 1985; Faltz 1977; Zribi-Hertz 1980, 1989). For objects of prepositions, especially ungoverned ones, coreference is not so unlikely as for accusatives. Hence, a pronoun could have been acceptable (indeed it was in Old English, and still is in some cases, as in Do you have any sharp objects on you?- Risk; Cf. with they brought it upon themselves- Cutiepie).25 Note also that whereas genitives are pronominal in many languages, they are reflexive in some (e.g., Swedish and Turkish). Accessibility expectations allow them to be pronouns (due to the high accessibility of their referents), but a formally defined generalization may force a reflexive in this context. Arnold (to appear) argues that Mapudungun subjects code the most accessible entity of the clause. However, this choice has been frozen into an animacy scale, whereby first/second person references are automatically higher on the scale than third person references. The result is that on the rare occasions when third person referents are more accessible, it is still the first/second persons which are selected for subject position. This is another case where a formal rigid distinction replaces the more variable, cognitive one. Rieder and Mulokandov (1998) find a surprising fact: Hebrew first person plural pronouns (anaxnu) contract more often (2.5 times more) than singular first persons (am). 'We' is also coded as zero more often than 'I' (1.5 times more). These are seemingly unexpected under accessibility theory, since surely the degree with which speakers are accessible to their addressees exceeds the accessibility ofthe speaker plus another or others (the referents of 'we'). However, once we take into consideration that Hebrew 'we' is three syllables long, the findings are no longer unexplained. Recall that long forms (i.e. least attenuated) code a relatively low degree of accessibility. Since speakers must choose between zero and a three syllable NP for 'we' (in modem Hebrew), they would tend to opt for the high accessibility forms more often. Such findings demonstrate the interaction of accessibility theory with specific facts of particular languages, in this case the


lexical options available. Indeed, in Saadi ( 1997), the number of high accessibility markers (pronouns and zeroes calculated together) was identical for the adult story in Hebrew and English, but Hebrew showed more zeroes than English, and English showed more pronouns than Hebrew. These differences are obviously motivated by the freer zero options available in Hebrew. The grammaticization of specific anaphoric expressions in certain syntactic structures can also be motivated by accessibility considerations. Ziv (1994) is explicit about it. She argues that one should not simply treat left and right dislocations as whole syntactic constructions used in specific (and different) pragmatic contexts. Rather, she shows that the facts of their pragmatic distribution match the referential forms they employ (an initial NP for left dislocations, an initial pronoun for right dislocations), which, in turn, are governed by the degree of accessibility associated with the entity coded by the dislocated NP. In other words, the frozen referential forms in left and right dislocated sentences are no different from their free occurring counterparts. Montgomery ( 1989) discusses it versus that left dislocations. He finds that that dislocations occur with the more complex (and clausal) NPs, they establish a contrastive focus in 26% of the cases (as opposed to 7o/o ofthe it dislocations), and they initiate an "oral paragraph". Note that these are all features which characterize entities of a relatively lower degree of accessibility, and in this respect the findings for it and that left dislocations are parallel to the ones presented in Unde (1979), Grosz (1981), and Schiffman (1984) for anaphoric it and that (all quoted in Ariel 1990). Finally, Heller {1998) argues that the Hebrew demonstrative ze 'this', when functioning as a copula, forces an extended refernce interpretation for the subject. Again, this directly mirrors the referential properties of the intermediate accessibility marker ze (see Ariel1998b). Giora and Lee {1996) also show that an initially motivated accessibility finding can develop into a (partially) grammaticized fact of a somewhat different functional nature (see also Marslen-Wilson et al. 1982). Giora and Lee argue that while accessibility theory can account for the fact that paragraphinitially, accessibility markers tend to be lower (pronouns instead of zeroes in Chinese), it cannot account for the fact that paragraph-final referring expressions also tend to be lower accessibility markers. It is possible that this distribution is due to the fact that lower accessibility markers are better cataphoric devices. Alternatively, since lower accessibility markers naturally occur paragraph-initially, they may be reanalyzed (also) as discourse segmentation markers. This is what Giora and Lee argue for. 26 In a similar vein but more radically, Vonk et al. ( 1993) argue that overspecified referring expressions (too low

59

6o

Mira Ariel

accessibility markers) affect discourse structure, rather than merely reflect it, as I originally argued. Lower accessibility markers instruct the addressee to shift from the current global discourse topic, even if the protagonist herself remains the same. In their experiments, the decision to shift to a new unit of information was determined by the choice of a too low accessibility marker, rather than the other way round. I believe that the high correlation between segment-initial position and low accessibility markers is originally motivated by the default strategy of emptying short term memory at the end of segments. However, this correlation may then be used in the other direction, namely to aid addressees in segmenting the discourse, especially when other means (such as time and place shifting expressions - see Gernsbacher 1991) are not available. In Section 3 I have argued that referential choice and interpretation is partly governed by grammatical principles and partly by extragrammatical accessibility considerations. However, because of grammaticization processes, the grammar-internal/external division of labor is not rigid across languages, nor within languages.

+ Competing theories of reference Accessibility theory is not the only theory which seeks to anchor referential forms in a broader, less than fully linguistic system. Chafe ( 1976 and onwards), Giv6n (1983), Levinson (1987, 1991) (and Huang 1994), Gundel et al. (1993) and Centering theorists (Grosz et al. 1986, 1995) have also offered such theories.27 It is important to note, however, that these theories do not clash with accessibility theory on a few important points: all theories offer some version of a scale on which referring expressions are arranged; all agree that additional, pragmatic factors can override the principles they propose. Crucially, the theories converge on the predictions re gross distinctions between zeroes, pronouns and lexical NPsl8; indeed, no counter-examples to accessibility theory have been shown to be better accounted for by these theories. Still, there are conceptual as well as empirical differences between these theories, and I will here briefly mention why I think that accessibility theory provides a better account for referential form use and interpretation. 29 Chafe (1976, 1994, 1996) was the first to argue for a direct connection between referential forms and cognitive statuses. In fact, accessibility theory can be seen as an extension of his (and later Giv6n's 1983) basic insight. Chafe recognizes that activation states are not categorical (discrete), but for language,


he distinguishes between three types of activation states only: activated, semiactive and inactive. Referential forms are chosen according to the estimated cognitive status of the referent: unstressed pronouns retrieve activated referents, and stressed nouns and noun phrases retrieve semiactive and inactive referents. Chafe then has to attribute many distinctions that I attribute to degree of accessibility to other distinctions which are partially orthogonal to degree of activation (identifiability, familiarity, contrastiveness). Although I believe that identifiability and contrastiveness are orthogonal to degree of accessibility, I think that Chafe is attributing distinctions to these concepts that are better treated as accessibility distinctions. For example, one wonders why stressed forms are consistently used for both lower accessibility and contrastiveness in many languages. I have argued (Ariel1990) that a contrastive form is used for an entity not predicted to occur (in the particular role). Hence the connection between contrastiveness and a relatively lower degree of accessibility. Also, Chafe claims that demonstrative pronouns identify better than pronouns (he contrasts it with this, see also Vonk et al. 1992, p. 303 ), in order to distinguish between them, since the three-way activation division is not enough for that. It is not dear how this identifies anything better than it (except by marking a lower degree of accessibility). Although it is widdy believed that this is more informative than it (presumably because of its deictic component), the actual distribution of it versus this shows that spatial deixis is very marginal in discourse. And this does not provide more information about the intended antecedent than it(see ex. 1 again). The main problem with Chafe's proposal is that a three-way distinction cannot account for the range of data I have referred to. In fact, Chafe himsdf presents a counterexample to his own three-way distinction: stressed pronouns which are not contrastive. They are an intermediate category. Finally, Chafe (1996, p. 40) anticipates that more degrees of activation may need to be recognized. Levinson's basic intuition is that coreference is preferred over noncoreference, and that minimal forms (e.g., zero, pronoun) should be used, unless fuller forms (e.g., lexical NPs) are specifically required (i.e., if the grammar does not allow the use of a minimal form). If, however, a fuller form is found where a more minimal form is licensed by the grammar, the addressee draws an implicature that the speaker did NOT intend a coreference reading (see originally Reinhart 1983 ). I have argued against Levinson's ( 1987, 1991) theory at length (see Ariel1994, 1996, and see also Blackwell, 2000). I have presented many counterexamples to his predictions, most of which stem from his insistence on the (grammarian's) coreference-disjoint reference distinction. Thus,

61

62.

Mira Ariel

Levinson can indeed motivate why certain anaphoric expressions are disjoint in reference from certain antecedents (when a low accessibility marker is used instead of a high accessibility marker). But he cannot explain why the same low accessibility markers must be interpreted as co referent with other antecedents, which are undistinguishable from the "illegitimate" antecedents on his account. I will here mention one example (His and the zero refer to the sexual abuser, who is the discourse topic): (10) REBECCA: RICKIE: REBECCA: RICKIE: REBECCA:

.. put the newspaperi on his lap, Y[eah], 0 [mas Iturbated, and then lifted the paperi [up I, [Yeah], .. for her to see (Jury).

Note that grammatically, 'the newspaper' could have been referred to by an it in the second mention. Cross-sentential pronouns are quite frequent in discourse (in fact, see the use of his and her in this example). But it was not. Still, no disjoint reading is generated, and we understand the two expressions as coreferent NPs. The reason is, I have argued, that the mental representation of 'the newspaper' is not highly accessible enough to merit a pronoun, but that does not at all rule out a coreference reading. Levinson seems to equate between high accessibility marking and coreference and between low accessibility marking and disjointness. I have argued that (non) coreference and degree of accessibility are orthogonal to each other. Another insensitivity of Levinson's (and others) is manifest in this example, namely the lack of attention paid to the difference between types of full NPs, here the newspaper and the paper. The latter is a shorter referring expression, therefore marking a higher degree of accessibility. It is a full lexical NP on Levinson's account, therefore undistinguishable from the longer alternative. But the shorter low accessibility form is not accidentally used. One of the most important claims of accessibility theory is that accessibility comes in a rich array of degrees, and any attempt to reduce it to a binary (coreference/disjointness) distinction is doomed to fail. Next, consider Gundel et al. (1993). A superficial look at Gundel et al.'s theory reveals an important advantage over accessibility theory. Whereas accessibility theory claims that degree of accessibility is responsible for the distribution of referring expressions, no attempt is made to specify a one-to-one cognitive correlate for each referring expression beyond the claim that a repre-


sentation is supposed to be relatively more or relatively less accessible given a specific referring expression. No cognitive status is described in the absolute. Gundel et al. 's Givenness hierarchy proposes precisely that. Their theory maps mental representations referred to onto six implicationally related cognitive statuses (each status implies that the statuses to its right hold as well): ( 11)

In focus > activated > familiar > uniquely identifiable > referential > type identifiable.

Unfortunately, the list of statuses specified looks suspiciously compatible with the distribution of just those referring expressions linguists have tended to focus on (i.e., some but not even all the referential forms in English+ zero). Now, I agree that linguists must absolutely set their goal at explicating linguistic forms, but the result in this case is that the cognitive aspect of the explanation is severely compromised. The cognitive basis of referential forms is drastically reduced if cognitive statuses are actually defined as a disjunction of statuses. Consider the status of'uniquely identifiable'. This status actually comprises two rather different cognitive activities: the addressee is either to retrieve an existing representation for a specific entity, or else to immediately generate such a representation. Now, I am not denying that definite descriptions (most prominently) trigger both of these cognitive processes. But are we really justified in claiming that these two are one and the same cognitively? The status of'referential' is also a hi-cognitive status according to Gundel et al.'s definition: "the addressee must either retrieve an existing representation of the speaker's intended referent or construct a new representation" (276). In fact, it is hard to see how the characterization of 'referential' differs from that of 'uniquely identifiable'. 30 In addition, it is not dear how the first disjunct of 'uniquely identifiable' and of 'referential' differs from the status of 'familiar', i.e., "The addressee is able to uniquely identify the intended referent because he already has a representation of it in memory" (p. 278). Moreover, it is not only that cognitive statuses are disjunctive, so is the relationship between referring expressions and cognitive statuses. That, this, IT (stressed) and this N are all said to mark one and the same 'activated' status in English (Japanese has 6 expressions for this cognitive category); Russian and Spanish have two expressions for the 'familiar' status. These show a many forms-one function relationship. Mulkern (1996), using Gundel et al.'s (1993) theory, finds that partial proper names are either 'familiar' or 'activated', whereas full names are either 'uniquely identifiable' or 'familiar'. These show a

63

64

Mira Ariel

one form-many functions relationship. In other words, there is no one-to-one correspondence between forms and cognitive status in any direction. Another problem with Gundel et al.'s Givenness hierarchy is raised by Ziv {1996): pronouns ('in focus') are predicted to always be 'uniquely identifiable' according to the Givenness hierarchy (because the hierarchy is implicational), but they are definitely not always so. In Ziv's examples they are unidentified inferred role players. This seriously undermines the potential explanatory superiority of the Givenness proposal. Also, while there is psychological evidence for the scalar relationship between 'in focus' and 'activated' (see the references above), there is no psychological evidence for the scalar distinctions between the other four categories on the scale. Finally, Gundel et al.'s (1993) theory (as well as the other theories here discussed) are far too restricted as to the referring expression types they recognize. The problem of a one-to-one correlation between an absolutely defined cognitive status and each referring expression is aggravated once we take into consideration the actual rich array of expression types. For example, how would one distinguish between zeroes and pronouns in a language which uses both as very high accessibility markers (e.g., English, and even more so, Hebrew)? Both must be classified as 'in focus' markers, but they each have a distinct distributional pattern. How can we distinguish between full and cliticized pronouns? Between first/second and third person pronouns, between more and less informative definite descriptions and proper names? Between longer and shorter definite descriptions and names where length does not affect the degree of informativity (Cf. the newspaper with the paper above)? I doubt that Gundel et al. can offer as many coherently defined cognitive statuses as there are distinguishable referring expression types. The only way they (as well as Levinson, Chafe and Centering theorists -see below) can handle such different distributions is by incorporating additional explanations. Accessibility theory handles most of these distributional patterns by one and the same generalization, although it does not completely replace identifiability and contrastiveness. Centering theory has focussed on an important factor in referential form choice: text coherence and its effect on the prominence of potential antecedents. Centering theorists (Grosz, Joshi & Weinstein 1986, 1995; Walker, Iida & Cote 1990) distinguish between antecedents as to their likelihood ofbecoming the focus of the next clause: topics, entities empathized with and subjects are expected to be the next clause topic more than non-topics, non-empathized entities and non-subjects respectively. They themselves are ordered as above with respect to potential future focussing on. Centering theorists then cor-


rectly predict that the more prominent discourse entity will be coded by zero or pronoun (depending on the language). They also emphasize that the salience of a discourse entity is determined by a combination of syntactic, semantic and pragmatic factors. All this is of course quite compatible with accessibility theory. However, Centering theory cannot be taken as a theory about referring expressions in general. Its proponents cannot even be said to characterize the usage of pronouns really, which they purport to. Their formulation of the 'pronoun rule' (or the zero rule forlanguages like Japanese) is that ifsome entity is realized as a pronoun, then so must be the highest ranking entity. This is for the most part the discourse topic, and hence, it will indeed be coded by at least as high an accessibility marker as the less accessible discourse entities. 31 But note that this formulation is far from a complete picture of anaphora (Centering theory does not even consider referential nonanaphoric cases). How does one decide whether she can refer to the other, less salient antecedent by a pronoun? And how does one decide on referring expressions other than pronouns? Note that Centering theory predictions are not violated if the highly salient discourse topic is coded by a full NP, provided the other, lower-ranking centers are too (or if they are not mentioned). This is a serious problem in view of the discourse findings presented in the literature. In fact, "the repeated name penalty", which has been presented as support for Centering theory is not actually predicted by it, just because the pronoun rule is formulated in such a way that it does not rule out lexical anaphors. What are the predictions for the highest ranking entity if a lower one is coded by a demonstrative pronoun? Probably because their main interest lies in coherence relations between clauses, Centering theorists, do not address these questions. Their rule determines what the coding should be for the one, most prominent discourse entity, and even this is not stated absolutely, but rather, in comparison to other discourse entities. I suggest that what the pronoun rule boils down to is 'use a high accessibility marker for highly accessible entities, subject to what the selection of high accessibility markers is in the language'. Since some languages can distinguish between zero and pronoun and/or between cliticized and full pronouns, the most highly accessible discourse entity may actually be required to be coded by a higher accessibility marker than the next one in ranking (a stronger requirement than the Centering one). Further research is required. The following example shows that Centering theorists focus too much on local connections. While in most cases the discourse topic is also the sentence subject of the sentence, in this case it is not. Thus, despite the fact that the higher officer is mentioned as a subject in two

65

66

Mira Ariel

consecutive clauses (not to mention the hair) it is still the discourse topic (officer Feil) which is pronominal and not the higher officer in both instances: (12) It is not as if he; looks like a hippie really, or anything like that. ... Fell's; grey-brown hair ... covers his; collar from behind. But one day in 93' the officer; in charge of him; demanded from him; to have a haircut ... The officer; accused him; ... (Haaretz 1.21.99).

In sum, all the theories discussed in Section 4 correctly predict some of the distributional patterns of referring expressions, but none, I believe, can account for the full range of data as well as accessibility theory.

s.

Directions for further research

While anaphora has been extensively researched by both linguists and psycholinguists, many questions are still unresolved. I list below a series of open questions pertaining to referential forms and linguistic and psycholinguistic research. I have divided them into linguistic (5.1) and psycholinguistic (5.2) questions.

s.t

Linguistic proper questions

Kirsner (1979, 1990; Kirsner & Van Heuven 1988) has data which contradicts accessibility theory predictions re proximate and distal demonstratives used anaphorically. Accessibility theory predicts that the demonstrative used for proximate physical pointing will also code a higher degree of accessibility when used anaphorically, as compared with the distal demonstrative. However, in Kirsner's Dutch data, it is the distal one (die) which refers to the less distant antecedents (70% of the demonstratives which had an antecedent in the same sentence were distal); It is the proximate demonstrative (deze) which refers to the more distant antecedents (89% of the demonstratives referring to an antecedent 2-3 sentences away were proximate). However, other languages pattern as predicted by accessibility theory (see Ariell998b; Lichtenberk 1996), and 87o/o of deze's do find their antecedents in the same or the previous sentence. Perhaps because the proximate demonstrative is originally used for highly accessible entities which are marked in the specific role they occur in, they are reinterpreted as "greater urging that the hearer find the referent" (Kirsner 1979, p. 358), leading to their reclassification as relatively lower


accessibility markers. Further research is needed. 32 Potential complications for accessibility theory are introduced by Arnold (to appear). Arnold first establishes that Mapudungun subjects code the most accessible entity of the clause. This should mean that when the subject is referred to in a subsequent clause, the anaphoric expression used should be a high accessibility marker. This is generally true, but Arnold shows that there is an additional factor at work: parallelism. The probability for an anaphoric object to be nonovert (a high accessibility marker) is higher when the antecedent is an object (i.e. representing an entity of a relatively lower degree of accessibility) than when it is a subject (coding the most highly accessible entity of the previous clause). Arnold attributes this phenomenon to the effect of parallelism. Rosen ( 1996) discusses a similar phenomenon. Zero subjects are interpreted not necessarily on the basis of the previous clause zero subject, but rather, on the basis of a possibly nonadjacent previous clause where the verb has a similar argument structure. Chambers and Smyth (1998) provide psycholinguistic evidence for the preference for pronouns to be coreferent with antecedents of the same structural status (subjects with subjects, and crucially, objects with objects). It remains to be seen whether parallelism is a separate factor working orthogonally to accessibility theory or whether the findings can be motivated within accessibility theory, by incorporating expected grammatical role as an accessibility factor. Du Bois ( 1980) had in fact argued that there is a separate tracking mechanism for objects. Degree of accessibility is a feature that characterizes Given information. It should then be fruitful to re-examine all those forms argued to code Given information (e.g., presuppositions, ergative nominals), and see whether it is general Givenness which determines their proper use or whether it has to be a specific degree of accessibility. For example, DuBois ( 1987) found that the rate of new entities is significantly lower than the rate of lexical NPs. This gap is probably explained by the presence of Given discourse entities which are of a relatively low degree of accessibility, and hence, are coded by lexical NPs despite their Givenness. If the motivation he proposes for ergative and accusative markings is based on the lexical versus nonlexical distinction, then it is probably based on the consistently high degree of accessibility of agents versus the inconsistent degree of accessibility associated with intransitive subjects and objects, rather than on the Given-New distinction between them. 33 The same applies perhaps to the pragmatic principle Du Bois proposes as underlying his "preferred argument structure", namely,"Avoid more than one new argument per clause" (p. 826). It should perhaps be replaced by "Avoid more than one

67

68

Mira Ariel

argument of a low degree of accessibility". Asswning that this is true, it remains to be seen what counts as a too low a degree of accessibility. 34 Gemsbacher (1989) demonstrates how different referring expressions enhance the accessibility of the mental representations associated with them. More explicit expressions (lower accessibility markers, proper names, for example) boost the activation of their mental representations faster and more than higher accessibility markers (pronouns, for example, see also Clifton & Ferreira 1987). In effect, the same accessibility marking scale reflects accessibility enhancing (and suppressing): the lower the accessibility marker used, the more enhanced the discourse entity coded by it will become (and the more suppressed other discourse entities will become). This means that the same accessibility markers code a specific current degree of accessibility (say, low), but at the same time, they contribute (at least partly) to the opposite degree of future accessibility (high). This can explain why speakers shift to lower accessibility markers from time to time, even when they continue to discuss the same discourse entity. These shifts combined with results obtained by Sanford and Garrod ( 1981) and Almor ( 1999), point to conflicting motivations in referential expression choice: "Live for today" versus "Live for tomorrow". Sanford and Garrod's results show that using too low an accessibility marker (e.g., a definite NP when the antecedent is a repeated discourse topic) slows subjects down. Gernsbacher's results, on the other hand, show that lower accessibility markers boost future degree of accessibility. In other words, in some cases the speaker has to choose whether she wishes to attend to her addressee's needs by choosing her accessibility marker in accordance with the current degree of accessibility (e.g., high), or by" ensuring the future", that the entity at hand remains/regains a high degree of accessibility (by choosing an accessibility marker which is relatively too low). Such competing motivations are rampant in natural language (see DuBois 1985), and further research is called for in order to find out when it is that speakers opt for reflecting current degree of accessibility and when they opt for establishing or maintaining a high degree of accessibility for future references. An interesting open question that needs researching into is the question of the correlation between referentiality and degree of accessibility. It seems that some forms are not only very high accessibility markers (e.g., zeroes), they are also more compatible with nonreferential readings (see Cameron 1997; Doron 1982; Garcia 1996; Sells 1984). However, as antecedents, nonreferential entities (e.g., 'whoever', generic NPs, impersonal 'you') are on the whole less accessible, and hence should have prompted relatively lower accessibility anaphoric


expressions. I have argued against a referential/nonreferential marking dichotomy, showing that even nonreferential arguments in relative clauses are not restricted to gaps and may take resumptive pronouns (Ariel1990, pp. 153155). But the fact remains that the preference for gaps is stronger when relative clause heads are nonreferential, despite the relative low accessibility of nonreferential antecedents. Perhaps this is related to the future accessibility marking of NPs. Nonreferentials are typically noncataphoric, hence the avoidance of low accessibility markers. 3s Grammaticization raises interesting questions too, not limited to referring expressions. How do we determine that a certain interpretative process is grammaticized, as opposed to being merely a common-sensical choice? (see also Kirsner & Van Heuven 1988). McDonald and MacWhinney (1995) show that when there is a clash between a 1st mention antecedent and an antecedent compatible with the semantics of the verb, the latter wins out. Is this a grammatical fact or only an e:xtragrammatical strategy (because violating the latter would cause incoherence)? When do we say that a discourse pattern has become grammaticized? Is a certain statistical percentage sufficient? Do we require lOOo/o? If we do, we will hardly be able to establish any obligatory grammatical rules. But then, if we do not impose such a high requirement, it is hard to tell the difference between the discourse profile of some form and the linguistic convention dictating its distribution (see also Ariel 1999). For example, most of the antecedents of Spanish si are subjects. In this case, Garcia ( 1983) argues against this being a grammatical rule. Instead, si is taken to demand highly accessible antecedents. But Garcia herself (as all functionalists, in fact) is not committed to "all or nothing" principles. Rather, she presents her theory as a set of principles generating discoursal preferences. These by definition are not lOOo/o correlations. The question then arises as to when we decide that a certain high percentage represents a formal rule and when we posit an extralinguistic generalization, which, as Garcia says, may not show a lOOo/o correlation either. Perhaps we should after all impose a requirement for (a near) lOOo/o correlation for grammatical principles, provided we recognize their complexity (see 1.2 above), as well as the fact that sometimes competing factors may block full compliance with the generalization. Dahl and Fraurud (1996) and Fraurud (1996) argue that we need to recognize the importance of animacy in referential choice. In their Swedish data, pronouns retrieved some human but no nonhuman antecedents which were not in the immediately preceding sentence. All the nonhuman referents coded by pronouns had a nearby antecedent. In general, whereas over a third of

69

70

Mira Ariel

the human definite 3rd person NPs were coded by pronouns, only 8% of the nonhuman NPs were coded by pronouns. Now, is this a discourse profile, or a grammatical convention? Should we say that pronouns cannot refer to nonhuman antecedents which do not occur in the same or the previous sentence? It seems that this is true in 100% of the cases, after all. I would rather not impose such a grammatical rule in this case. Fraurud's findings can be seen as reflecting the discourse profile of pronouns with nonhuman antecedents: Nonhuman entities are not as salient to us as humans are. If so, we should only expect to find that nonhumans in the same (large) distance as human antecedents are of a lower degree of accessibility. Hence the inability of pronouns to refer to them. In order to distinguish between these two options one should examine cases where nonhumans are very salient, as when they are the continuing discourse topic. If they cannot be referred to by nonimmediate pronouns even in such cases, then Fraurud's findings should be incorporated as a grammatical (semi-arbitrary) convention. 5.2 Questions pertaining to the connection between psycholinguistic research and grammar

The cognitive psychologists' findings so far seem to me to corroborate all existing theories, although they are presented as supporting either Centering theories (e.g., Gordon & Chan 1995; Kennison & Gordon 1997), Giv6n's (1983) topic continuity theory (Gernsbacher & Jescheniak, Ms.; Gemsbacher & Shroyer 1989) or accessibility theory (Almor 1999, in press; Arnold 1997). The reason is that the psycholinguistic findings support any theory which posits some scale of referential forms (Gundel et al. 1993 included). It would be interesting to think of psycholinguistic experiments which would test the different predictions of these different theories in order to establish whether one is possibly superior to others. 36 Recall that Almor ( 1999) argues very forcefully that when low accessibility markers are justifiably used in high accessibility contexts, processing is not slowed down. This finding contradicts my claim that proper accessibility marking can and is violated for special effects at a processing cost. It is not clear to me that Almor has actually proved that this is not the case. In order to do that, he would have to compare contextually informative low accessibility markers with pronouns when the antecedent is in focus. The comparisons he presents only compare justified versus unjustified low accessibility markers, but not high accessibility markers. I expect pronouns to take less time than


informative low accessibility markers. I would then view his current findings as showing that unjustified low accessibility markers merely slow addressees down more than justified low accessibility markers do, and not that justified low accessibility markers do not absolutely slow processing down. Gernsbacher and Faust ( 1991) explain the problem of less skilled comprehenders by reference to their less efficient suppression mechanism. Their experiments concern ambiguous words, where they find that the less skilled comprehenders have no problem making use of contextual cues for the appropriate interpretation. They also have no problem enhancing contextually appropriate information, but they find it relatively difficult to reject contextually inappropriate meanings that were generated automatically. Two questions come to mind. If suppression mechanisms are crucial for reference interpretation, then these same comprehenders are expected to also have problems in interpreting referring expressions where reliance on suppression is required (i.e., when there are competing antecedents). A more radical research goal is to look into the possibility that comprehenders may have different suppression and enhancement problems in different tasks, specifically, in reference determination versus ambiguity resolution. Gernsbacher assumes that enhancement and suppression are general cognitive skills, and indeed shows that the same less skilled comprehenders have difficulties suppressing non-verbal stimuli. Still, we should ascertain that this generalization holds across different linguistic interpretative processes as well. Psychologists have worried in the past about the ecological validity of their laboratory experiments, namely about the applicability of their experimental findings to the natural activities of their subjects. This problem still exists, of course, but I would like to point to a related problem. Suppose we grant that the discourses recently tested in many psycholinguistic experiments are real enough. Based on the psycholinguistic findings, we could easily establish a very rich scale of degrees of mental accessibility for concepts in various contexts. An important research goal then awaits linguists in trying to understand which of these psychologically real distinctions translates into a possible grammatical distinction (i.e., one that occurs at least in some language). Is it just the frequency of the cases in which the accessibility-related processing distinction is crucial for communicative purposes that determines that a linguistic distinction is to be instituted? If so, can we prove that this is the case? Alternatively, it could be that what is universal is not that specific "fundamental" (essential) accessibility distinctions are to be drawn by each grammar, but rather, that some accessibility distinctions be drawn. In other words, perhaps it is not so

71

7'l.

Mira Ariel

important what precise contexts are to be declared as bearing a distinct (high/ low/intermediate, etc.) degree of accessibility so much, as it is crucial that each language should have at least a minimal number of linguistically marked accessibility distinctions, which it then maps on to various contexts in a motivated yet somewhat language-dependent way. The universal could also be some combination between these two alternatives, namely, that some specific essential distinctions are "obligatory", and others have to occur, but with no restrictions as to where they are drawn. Some processing distinctions may simply be uncodable by (semi-)formal rules, and thus can only constitute laboratory results. Some distinctions may be mutually exclusive (obviation and logophoricity perhaps), because they are too similar. For example, Gemsbacher ( 1989) finds that when pronouns are relatively quite informative (when the gender distinction can distinguish between the intended antecedent and an unintended one), they suppress other discourse entities more than when they are not as informative (because the competing antecedents are of the same gender, see also MacDonald & MacWhinney 1990). No grammaticized consequence is expected in this case, however, since languages do not usually offer a choice between gender-marked versus gender unmarked pronouns; they have one or the other option. Similarly, Gernsbacher et aL (1989) find that first mentions in conjoined (new) NPs are later more highly accessible than second mention NPs. Non subject initial participants are also more accessible (Gemsbacher 1991). Yet, no language is known to have grammaticized the notion of clausal first mentions. It is usually the subject that serves as a locus of grammatical conventions, e.g., the restriction of reflexives in some languages to subject antecedents. No language, to the best of my knowledge, restricts reflexives to clausal first mentions, even though many languages allow nonsubjects in sentence initial position quite freely. It is still possible, of course, that a clausal first mention discoursal preference will be found, regardless of grammatical role. Karmiloff-Smith ( 1985) points out that children tend to refer to main characters by pronouns, and to marginal characters by lexical NPs. While this would be a well-motivated grammaticization path, since main characters are consistently more salient than marginal ones, I do not expect it to be an adult grammaticization path: We need to refer to main characters by lower accessibility markers sometimes (e.g. following episode boundaries) and we need to refer to temporarily highly accessible marginal characters by high accessibility markers. 37 Gernsbacher' s ( 1989) findings raise an interesting question pertaining to the nonautomatic connection between psycholinguistic and linguistic facts.


Gemsbacher's subjects first read a sentence which introduced two characters. They were then presented with a participial phrase which biased them towards one of the two potential antecedents, followed by a pronoun. Reaction time measurements revealed no difference in the accessibility of the two antecedents at the stage where the pronoun was encountered, despite the biasing adverbial. At the end of the clause, however, the degree of accessibility associated with the appropriate antecedent was higher. 38 Such findings raise a question about the processing stage relevant for measuring degree of accessibility. If accessibility theory is correct, we should expect that the relevant time is the stage at which the anaphor is processed (whenever that may be). However, it is not dear that grammaticizations can be sensitive to fine-tuned accessibility fluctuations over very short spans of time, so perhaps we must not expect a perfect fit between degree of accessibility as measured by psycholinguistic experiments and degree of accessibility as it is reflected in linguistic conventions. Further research is needed to settle this question. Last, some linguistic and psycholinguistic results that we have are actually in conflict with each other. Thus, dark and Sengul (1979) compare the retrieval of antecedents in previous clauses of the same sentence versus across sentence boundary. Based on subjects' reaction times, they conclude that it is the clause rather than the sentence that makes the significant difference in the processing time of anaphors. However, Clancy (1980), examining English and Japanese narratives, finds that it is the sentence rather than the clause that better accounts for the distribution of fuller versus attenuated referential forms. We need to find out whether there is one relevant unit (either the clause or the sentence), or whether under different conditions, or for different forms, one of them may be the more relevant unit (a more plausible possibility). 39 In this article I have tried to describe the main claims and findings of accessibility theory, emphasizing that the notion of accessibility is complex and that it is by no means the only factor determining referential form. I corroborated the accessibility claims relying on more recent research which also further develops the theory. I have argued that accessibility theory is at least partly linguistic, despite the fact that it is well motivated cognitively, and that it accounts for the rich data better than other theories of discourse reference. Finally, I have raised several open questions regarding discourse references. I hope linguists and psycholinguists will be prompted to explore them.

73

74

Mira Ariel

Notes

Accessibility theory is very much in tune with many recent psycholinguistic proposals, where mental accessibility of various referents has been experimented with. In fact, these experiments (together with discourse data) form the empirical basis of accessibility theory. However, psycholinguists and linguists approach the accessibility of mental representations with different goals in mind. Whereas the psycholinguists are interested in learning about human memory, the linguists are interested in learning about naturallangauge expressions. Hence, the psycholinguists use pronouns in order to draw conclusions about working memory, and any definite lexical NPs (definite descriptions and first names) to learn about the reinstatement process from memory. In contrast, the linguist must establish a formfunction correlation for each referring expression type. The psycholinguists see anaphora in general as a coherence device, and do not pay careful attention to minute differences between different anaphoric devices. They ignore nonanaphoric referential uses. In addition. psycholinguists are interested in how the processing of anaphora is performed, in how speakers assess the degree of accessibility of mental representations to their addressees (Morton Ann Gernsbacher, personal communcation). They want to define processing cues, which are different from linguistic codes (see Garnham, Oakhill & Cruttenden 1992; Garrod, Freudenthal & Boyle 1994; MacDonald & MacWhinney 1990; McDonald & MacWhinney 1995; Rinck & Bower 1995, inter alia), and to find when the accessing is performed, (e.g. Cacciari, Carreiras & Cionini 1997; Garnham, Traxler, Oakhill & Gernsbacher 1996; Lucas, Tanenhaus & Carlson 1990; McDonald & MacWhinney 1995). All of these are not of direct interest to the linguist. Last, many psycholinguists are committed to a dichotomy between working and long term memory, and therfore invariably compare two accessibility contexts or two referring expressions at a time. I find that unacceptable from a linguist's point of view, since the impression cretaed is that language poses a binary decision, parallel to the shortlong term memory division, where in reality, referring expression must be selected from a large variety of referring expressions. 2. Zeroes are empty argument slots, as in '0 (=you) wanna gor'. 1.

3· I thank Jack Du Bois (personal communication) for providing me with this example. All the examples in this text are taken from DuBois (2000), unless otherwise specified. 4- See Schilperoord ( 1996) for an argument that degree of accessibility (resulting from the

hierarchical structure of the text) determines pause lengths.

s.

Note that definite descriptions count as quite low accessibility markers here. Other researchers, however, have sometimes had to say that definite descriptions refer to the most •satient" or contextually uniquely identified referents (e.g.Chafe 1994, 1996; McCawley 1979 ), in order to make sure that addressees interpret the expression as referring to the immediately relevant entity. See Walker and Prince ( 1996, ex 1) where the guy is preferentially understood to refer to the non-topic 'guy', rather than to the topical'guy', and ex. 13, where in a discourse about two sisters, her sister changes its reference to whoever is not the sister in focus.


6. Oiticized pronouns are shortened pronouns, e.g., 'ya'.

7· Note the following example (Jury), where the molester is first referred to as he, then, strictly speaking, he is not referred to for a few intonation units I marked with •. Still, he is later referred to again with a pronoun:

RICKIE:

" '"REBECCA: '"RICKIE: '"REBECCA: '"RICKIE: '"REBECCA: '"RICKIE:

,.

You know like, (H) but he was making, I don't know how you describe it, you know how you can be like a nuisance to someone?

[Mhm). [Or)_ you may smell or some[thi]ng.

[Yeah]. you know like that you [know,

[Yeah). or] moving around, you know like, ... as he wanted her to move.

Indeed, Mauner, Tanenhaus, and Carlson (1995) found that missing agents in agentless passive sentences were processed nonetheless. 8. Independently, Terken and Nooteboom (1988) found that one previous mention was not sufficient for subjects to treat an entity as Given. In fact, Maes and Noordman (1995) also argue for special functions of these second mentions (see below).

9· In addition, in order to reduce all distinctions to one binary distinction, Giv6n simply ignores many referential devices, e.g., names (and first, last and full names each code a different degree of accessibility), and agreement markers. He also lumps together referential fonns which have different distributional patterns, e.g., zero and pronoun.

Nurit Assayag (personal communication) brought to my attention the following example (from her originally Hebrew conversational data), where the speaker refers to herself by too low an accessibility marker (a full pronoun rather than zero in the second mention) in order to maintain a syntactic parallelism with the preceding clause: So I began and nobody said anything. So I continued and nobody said anything.

10.

u. But note that Bird-David's research is on naming rather than on referential forms per se. This is true for all the anthropological work on names. However, Giv6n's examination of another novel which alternates between the perspectives of the two main characters (Cold mountain) revealed a different pattern: The zero/ pronoun versus NP distribution for references to the two characters is either similar, or else, there are more full NPs for the other than for the set£ At first blush, these findings seem contradictory, but actually, once the author relinquishes the narrator's role to some character, that character's consciousness is at work. Of course, normally, that entails the centrality of the self ( Giv6n' s novel), but at other times, the other is so central to the self that the other merits a higher rate of high accessibility markers (Cold mountain). This last point deserves further checking.

12.

13. This finding contradicts Bernstein's ( 1970) conviction that speakers of lower classes

7S

76

Mira Ariel

only use the restricted code. J.4. Note, however, that first mentions are more highly accessible only when all other factors are equal. When a non-first mention is marked as focus, as in wh-clefts, it is the second mention entity, coded by the focussed NP that is more highly accessible (see Almor 1999). 15. In fact, I have already briefly argued that the accessibility markers used to access inferred entities manifest accessibility differences (Ariel1990, pp. 184-190). 16. See Van den Broek (1990) for the importance of predictive inferences in general. 17. We can similarly establish that some NPs are marked for low or no catapboricity.

Quantified NPs, for example, are known to serve as antecedents for intra-sentential anaphora, but not for extra-sentential anaphora. 18. But first, note that referring expressions are not only measurable along one dimension. Thus, as Downing points out, classifiers also exempt speakers from marking the social status of the (human) referent. Also, as Kirsner (1990) argues, while the Dutch definite article and the distal demonstrative are sometimes interchangeable, the latter are used when the entity referred to bas been distinguished from others, while the demonstratives (both the distal and the proximate) are used when the speaker needs to alert the addressee to seek out the referent. Similarly, Epstein (1998a) argues that the has additional functions to reference establishing, e.g., marking the referent as prominent. Second, expressions commonly used to refer are not always used referentially, and as such are also otherwise classified (most notably, definite descriptions, which are sometimes used attributively or generically- see Mueller-Lust & Gibbs 1991). 19. But although Cacciari et al. ( 1997) found that gendered anapboric expressions speeded

up interpretations even when there was no competition over antecednthood, Gamham et al. 1992 suggest that the gender cue is not always used by subjects. However, we need to also examine the informativity and length of the lexical NPs involved. 21. In fact, C. L. Baker (personal communication) agreed with me on this point. 20.

22. I completely reject Reboul's (1997) assumption of an 'all or none' grammatical/extragrammatical status for reference interpretation. The fact that some aspects of referentiality are better accounted for by a pragmatic theory does not mean that all must be accounted for by pragmatic principles.

23. I actually believe that the 'avoid pronoun' principle is superfluous (see Ariel1990, pp.

100-105). In this case, then, I suggest to replace a grammatical principle with a functional principle. 24- One should however remember to distinguish between long and short reflexives. Accessibility theory predicts that they would be used differendy, and indeed they are (see Reinhart & Reuland 1993).

25- I thank Jack Du Bois for giving me the PP reflexive examples. 26. See Mithun

(1996, p. 231) for a similar finding.

27. Reboul (1997) argues against accessibility theory, but in effect against all attempts to

offer a linguistic theory for extrasentential referential forms. Although she herself does not


propose a specific account, she believes that with Relevance theory (Sperber & Wilson 1986) "one can account for the use of referring expressions, if one considers the semantic content of such expressions and the relationship between their semantic content and their referring ability" (p. 91, emphasis added). I had in fact argued against the first part of such a proposal in Ariel {1990, p. 83-86). I have shown that many referring expressions do not differ with respect to their semantic content, but they signal a different degree of accessibility nonetheless (e.g., it/that; name/shortened name; full pronouns/reduced pronouns/verbal person agreement markers). Degree of accessibility could be seen as the relationship between the semantics of the expression and referring ability, but it is not a transparently inferred relationship. Differences between languages which have the same referential forms (e.g., English, Hebrew and Otinese all have pronouns and zeroes, but they use them quite differently) are also left unaccounted for under an exclusively pragmatic theory. 28. In fact, Tao ( 1996) is the only one who claims to have different findings, where zero (in Chinese) is used to shift, rather than to maintain reference. 19. The theories also differ in scope of application. Only Levinson has argued that his principles actually replace the binding rules (and see also Garcia, 1996).

30. Gundel et al. ( 1993) claim that unlike 'referentials', 'uniquely identifiables' are identified based on the referring expression alone without reference to the rest of the sentence. I doubt that context is ever ignored. In any case, it is hard to know how one could check whether or not sentential (or other) context was actually used in the interpretative process. 31. In fact, Otambers and Smyth (1998) point out that Centering theory also cannot account for the acceptability of examples such as: Josh criticized Paul and then Marie insuhed him, where the pronoun does not refer to the most prominent forward looking center, nor is it the subject (and topic?) of either clauses. For other arguments against centering theory, see Chambers and Smyth (1998). 31. Kirsner argues that deze (+NP) codes HIGH DEIXIS, which often translates to relative low accessibility (in terms of referential distance and antecednt complexity). Note, however, that Kirsner's own attempt to incorporate the higher effort required in HIGH DEIXIS with references to entities physically near, rather than far from the speakers is unconvincing. Also, if important entities require HIGH DEIXIS, does that mean that pronouns coding continuing discourse topics are HIGH DEIXIS too? In other words, Dutch poses a puzzle as to why its proximate demonstrative marks higher accessibility for physical pointings but lower accessibility for discoursal refernces, when compared with the distal demonstrative (but see Piweck et al., as cited in Beun & Cremers 1998, for a different claim re the deictic usage of the proximate and distal demonstratives in Dutch). 1 tentatively suggest that this bas to do with the markedness of the proximate demonstrative (by far the rarer form in spoken Dutch). Thus, there is a potential conflict between demonstratives (in general) and definite descriptions. In terms of accessibility coding. the demonstrative should be the shorter form, but in terms of frequency it is the definite (or the distal demonstrative) which is predicted to be the shorter form. However, once length is established via markednmess (i.e., demonstratives are longer than definites) this formal difference in attenuation may affect the degree of accessibility later attributed to them.

n

78

Mira Ariel

33· If we replace Givenness with degree of accessibility, we can perhaps also explain why proper names pattern with 3rd persons in split ergative systems (both are not extremely highly accessible), rather than with 1st/2nd persons, even though they are (almost) equally Given (see Dixon 1979, p. 87). 34· In fact, DuBois (to appear) proposes that the deeper generalization behind the distribution of agents versus intransitive subjects and objects is sensitive to low versus high processing costs. It is the highly demanding NPs which are restricted in distribution. Indeed, other things being equal, high accessibility marking entails a low processing cost because the entity is highly accessible, and low accessibility marking entails a high processing cost because the entity is not so easily retrievable. However, pragmatically motivated exceptions to accessibility theory do occur, where highly accessible entities are referred to by relatively low accessibility markers (e.g., epithets), or vice versa (less common), where entities of a relatively low degree of accessibility are referred to by high accessibility markers. Accessibility theory predicts that both cases entail a high cost of processing, and hence, they should pattern as high processing cost entities, rather than according to either their marking or their real cognitive accessibility. This hypothesis requires testing. I thank Jack Du Bois for discussing this point with me. 35· It is also possible that using high accessibility markers (usually zeroes or pronouns) promotes the dependence of the interpretation based on another linguistic marker, which is required for nonreferentials. 36. As Morton Ann Gernsbacher (personal communication) reminds me, a huge task stiU remains of finding psycholinguistic lab evidence for the continuum of accessibility. 37· In fact, the children tested initially referred to the protagonists with indefinte NPs (and not pronouns), and they did from time to time refer to secondary characters by pronouns. Unfortunately, Karmiloff-Smith does not provide actual numbers. Also, the opportunity to refer to secondary charaters by pronouns was quite limited, since they were mentioned twice at most. )8. Unlike McDonald and MacWhinney (1995), Garnham et al. (1996) too find that

relevant semantic information takes effect only at the integration stage. Cacciari et al. ( 1997) suggest that the different findings re when semantic information is used in reference tracking may actually point to differences between different languages. 39· The psycholinguists also have contradicting results sometimes, e.g., Garrod and Sanford (1982) versus Albrecht and Clifton (1998) re anaphoric references to a conjoined NP antecedent when the anaphor is a subject.

References Aissen, J. (1997). On the syntax of obviation. Language. 73,705-750. Albrecht, f. E., & Clifton, C. Jr. (1998). Accessing singular antecedents in conjoined phrases. Memory and Cognition, 26, 599-610. Almor, A. (1999). Noun-phrase anaphora and focus: The informational load hypoth-


esis. Psychological Review 106,748-765.. Almor, A. (in press). Constraints and mechanisms in theories ofanaphor processing. In M. Pickering, C. Clifton, & M. Crocker (Eds.), Architectures and mechanisms in sentence comprehension. Cambridge: Cambridge University Press. Ariel, M. ( 1985a). Givenness marking. PhD thesis, Tel-Aviv University. Ariel, M. ( 1985b). The discourse functions of Given information. Theoretical Linguistics, 12, 99-113. Ariel, M. (1987). The grammaticalization of accessibility. Ms. Tel Aviv University (A much abridged version of this appeared as Part U of Ariel1990). Ariel, M. ( 1988a). Referring and accessibility. Journal of Linguistics, 24, 6~7. Ariel, M. (1988b). Retrieving propositions from context: Why and how. Journal of Pragmatics, 12, 567-600. Ariel, M. ( 1990). Accessing Noun Phrase antecedents. London: Routledge. Ariel, M. (1994). Interpreting anaphoric expressions: A cognitive versus a pragmatic approach. journal ofLinguistics, 30, 3-42. Ariel, M. (1996). Referring expressions and the +/- coreference distinction. In T. Fretheim, & J, K. Gundel (Eds.), Reference and referent accessibility (pp. 13-35). Amsterdam: John Benjamins. Ariel, M. ( 1998a). Three grammaticalization paths for the development of person verbal agreement in Hebrew. In Jean-Pierre Koenig (Ed.), Discourse and cognition: Bridging the gap (pp. 93-112). Stanford: CSLI/Cambridge University Press. Ariel, M. (1998b). The linguistic status of the 'here and now'. Cognitive linguistics, 9, 189-237. Ariel, M. ( 1998c). Mapping so-called 'pragmatic' phenomena according to a 'linguisticextralinguistic' distinction: The case of propositions marked 'accessible'. In M. Darnell, E. Moravcsik, F. Newmeyer, M. Noonan, & K. Wheatley (Eds.), Functionalism and formalism in linguistics. Amsterdam: John Benjamins. Ariel, M. ( 1999 ). Cognitive universals and linguistic conventions: The case of resumptive pronouns. Studies in language 23, 217-264. Ariel, M. Ms. Linguistic pragmatics. Ariel, M. (2000). The development of person agreement markers: From pronouns to higher accessibility markers. In M. Barlow and S. Kemmer (Eds), Usage-based models oflanguage (pp. 197-260). Stanford: CSLI. Arnold, J. ( 1997). What is salience?: The role of topic and focus in processing reference. Unpublished Ms. Stanford University. Arnold, J. (to appear). Multiple constraints on choice of reference: Null, pronominal, and overt reference in Mapudungun. In f. W. DuBois, L Kumpf & W. Ashby (Eds.), Preferred argument structure: Grammar as architecture for function. Amsterdam: John Benjamins. Bach, K. (1998). A review ofThorstein Fretheim and Jeanette K. Gundel (Eds.) Reference and referent accessibility. Pragmatics and cognition, 6, 335-57. Baker, C. L. (1995). Contrast, discourse prominence, and intensification, with special reference to locally free reflexives in British English. Language, 71,63-101. Beeman, M., & Gernsbacher, M.A.. Structure building and coherence inferendng during comprehension. University of Oregon manuscript

79

So

Mira Ariel

Bernstein, B. A. ( 1970). A sociolinguistic approach to socialization: With some reference to educability. In F. William (Ed.), Language and poverty. Chicago: Markam. Beun, R. J., & Cremers, A. H. M. (1998). Object reference in a shared domain of conversation. Pragmatics and cognition. 6, 21-52. Blackwell, S. E. (2000). Anaphora interpretations in Spanish utterances and the neoGricean pragmatic theory. Journal ofPragmatics, 32,289-424. Bouchard, D. (1983). The avoid pronoun principle and the elsewhere principle. In P. Sells, & C. Jones (Eds.), NELS 13. University of Massachusetts, Amherst. Bouchard, D. (1985). The binding theory and the notion of Accessible SUBJECT. Linguistic inquiry, 16, 117-33. Brennan, S. E. (1995). Centering attention in discourse. LangMge and cognitive processes, 10, 137-67. Brizuela, M. ( 1997). The selection ofdefinite expressions in Spanish: A note on the notion of processing effort. Paper presented at the International cognitive linguistics conference at Amsterdam. Broek, P. van den, (1990). The causal inference marker: Toward a process model of inference generation in text comprehension. In D. A Balota, G. B. Flores d'Arcais, & K. Rayner (Eds.), Comprehension processes in reading (pp. 423-445). Hillsdale, NJ: Lawrence Erlbaum. Cacciari, C., Carreiras, M., & Cionini, C. (1997). When words have two genders: Anaphor resolution for Italian functionally ambiguous words. Journal of memory and language, 37, 517-32. Cameron, R. ( 1997) Accessibility theory in a variable syntax of Spanish. Journal of pragmatics, 28, 29-67. Chafe, W. L. ( 1976). Givenness, contrastiveness, definiteness, subjects, topics and point of view. In C. N. Li (Ed.), Subject and topic(pp. 25-55). NewYork:Academicpress. Chafe, W. L. (1994). Discourse, consciousness, and time. Chicago: The University of Chicago Press. Chafe, W. L ( 1996). Inferring identifiability and accessibility. InT. Fretheim, & J. K. Gundel (Eds.), Reference and referent accessibility (pp. 37-46). Amsterdam: John Benjamins. Chambers, C. G., & Smyth, R. (1998). Structural parallelism and discourse coherence: A test of centering theory. Journal ofmemory and language, 39, 593-608. Oancy, P.M. ( 1980). Referential choice in English and Japanese narrative discourse. In W. L. Chafe (Ed.), The pear stories: cognitive, cultural, and linguistic aspects of narrative production (pp. 127-202). Norwood, NJ: Ablex. Clark. H. H., & Sengul, C. J. (1979). In search for referents for nouns and pronouns. Memory and cognition. 7, 35-41. Clark H. H., &Marshall, C. (1981). Definite reference and mutual knowledge. InA. K. Joshi, B. L. Webber, & I. A. Sag (Eds.), Elements ofdiscourse understanding(pp. 1063 ). Cambridge: Cambridge University Press. Clifton, C., & Ferreira, F. ( 1987). Discourse structures and anaphora: Some experimental results. In M. Coltheart (Ed.), Attention and performance Xl1 (pp.635-653). Hove, East Sussex: Lawrence Erlbaum.


Comrie, B. ( 1983). Form and function in explaining language universals. Linguistics, 21, 87-103. Comrie, B. (1988a). Topics, grammaticalized topics, and subjects. BLS, 14,265-79. Comrie, B. ( 1988b). Coreference and conjunction reduction in grammar and discourse. In J. A Hawkins (Ed.), Explaining language universals (pp. 18~208). Oxford: Blackwell. Comrie, B. (1994). Coreference: Between grammar and discourse. Proceedings of the eighteenth annual meeting of the Kansai Linguistic society ( 1993 ), 1-10. Dahl, 0., & Fraurud, K. (1996). Animacy in grammar and discourse. InT. Fretheim, & J, K. Gundel (Eds.), Reference and referent accessibility (pp. 47-64). Amsterdam: John Benjamins. Dolman, R. (1998). Comparing referential forms in narratives ofchildren aged 5-6 from low and high socioeconomic classes. A seminar paper, Tel Aviv University. Do ron, E. ( 1982 ). On the syntax and semantics of resumptive pronouns. Texas linguistic forum, 19, 1-48. Downing, P. A. ( 1986). The anaphoric use of classifiers in Japanese. In C. Craig (Ed.), Noun classes and categorization (pp. 345-375). Amsterdam: John Benjamins. Downing, P. A. ( 1996). Proper names as a referential option in English conversation. In B. Fox (Ed.), Studies in anaphora (pp. 95-143). Amsterdam: John Benjamins. DuBois, J, W. (1980). Beyond definiteness: The trace of identity in discourse. In W. L. Chafe (Ed.), The pear stories: cognitive. cultura~ and linguistic aspects of narrative production (pp. 203-274). Norwood, NJ: Ablex. DuBois, J. W. (1985). Competing motivations. In J, Haiman (Ed.), Iconicity in syntax (pp. 343-365). Asterdam: John Benjamins. DuBois, J.W. ( 1987). The discourse basis of ergativity. Language, 63, 805-855. DuBois, J. W. (1991). Definiteness, reference, and analogues. Paper presented at the cognitive linguistics symposium, UC San Diego. DuBois, J. W. (2000). Santa Barbara Corpus ofSpokenAmerican English. 3 CO-ROMs. Linguistic Data Consortium, University of Pennsylvania. DuBois, J.W. (to appear). Discourse and grammar. In M. Tomasello (Ed.), The new

psychology of language: Cognitive and functional approaches to language structure, Vol. 2. Erlbaum. Du Bois, J. W., & Thompson, S. A. (I 991 ). Dimensions of theory of information flow. Unpublished manuscript, University of California, Santa Barbara. Epstein, R. ( 1998a). Reference and definite referring expressions. Pragmatics and cognition, 6, 189-207. Epstein, R. ( 1998b ). Definiteness and the construction ofdiscourse referents. Unpublished MS. Rutgers University-Camden. Faltz, L. M.( 19n). Reflexivization: A study in universal syntax. PhD thesis, University of California, Berkeley. Farmer, A., & Harnish, H. (1987). Communicative reference with pronouns. In J. Verschueren & M. Bertuccelli-Papi (Eds), The pragmatic perspective: Proceedings of the international pragmatics conference (pp. 547-566). Amsterdam: John Benjamins.

81

h

Mira Ariel

Fowler, C. A., Levy, E. T., & Brown, J. M. (1997). Journal of memory and language, 37, 2~0.

Fraurud, K. {1996). Cognitive ontology and NP form. InT. Fretheim, & J. K. Gundel (Eds.), Reference and referent accessibt1ity (pp. 65-87). Amsterdam: John Benjamins. Garcia, E. C. ( 1983). Context dependence of language and linguistic analysis. In F. Klein-Andreu (Ed.), Discourse perspectives on syntax (pp. 181-207). New York: Academic Press. Garcia, E. C. ( 1996). What ..reflexivity" is really like. Linguistics, 34, 1-51. Gamham, A., Oakhill, J., & Cruttenden, H. ( 1992 ). The role of implicit causality and gender cue in the interpretation of pronouns. Language and cognitive processes, 7, 231-255.

Garnham, A., Oakhill, J., & Ehrlich, M. F. ( 1995 ). Representations and processes in the interpretation of pronouns: New evidence from Spanish and French. Journal of memory and language, 34,41-62. Garnham, A., Traxler, M ., Oakhill, J., & Gernsbacher, M.A. (1996). The locus of implicit causality effects in comprehension. Journal of memory and language, 35, 517-543.

Garrod, S., Freudenthal, D., & Boyle, E. ( 1994). The role ofdifferent types of anaphor in the on-line resolution of sentences in a discourse. Journal ofmemory and language, 33,39-68.

Garrod, S., & Sanford, A. J. ( 1982). The mental representation of discourse in a focused memory system: Implications for the interpretation of anaphoric noun phrases. Journal ofsemantics, 1, 21-41. Gernsbacher, M.. A. ( 1989). Mechanisms that improve referential access. Cognition, 32, 99-156.

Gemsbacher, M. A. ( 1990 ). Language comprehension as structure building. Hillsdale, NJ: Erlbaum. Gemsbacher, M. A. ( 1991 ). Cognitive processes and mechanisms on language comprehension: The structure building framework. The psychology oflearning and motivation, 27, 217-263. Gernsbacher, M.. A., & Faust, M. E. (1991). The mechanism ofsuppression: A component of general comprehension skill. Journal of experimental psychology: Learning, memory and cognition, 17, 245-262. Gemsbacher, M.A., Hargreaves, D., & Beeman, M. ( 1989). Building and accessing clausal representations: The advantage of first mention versus the advantage of clause recency. Journal ofmemory and language, 28, 735-755. Gemsbacher, M.A., & Jescheniak, J. D. Ms. Cataphoric devices in spoken discourse. Gemsbacher, M.A., & Shroyer, S. (1989). The cataphoric use of indefinite this in spoken narratives. Memory and cognition, 17, 536-540. Giora, R, & Lee, C. (1996). Written discourse segmentation: The function of unstressed pronouns in Mandarin Chinese. InT. Fretheim, & J. K. Gundel (Eds.), Reference and referent accessibility (pp. 113-140). Amsterdam: John Benjamins. Giv6n, T. (1983). Topic continuity in discourse: An introduction. InT. Giv6n (Ed.),


Topic continuity in Discourse: A Quantitative Cross-Language Study (pp.1-42). Amsterdam: John Benjamins. Giv6n, T. ( 1992). The grammar of referential coherence as mental processing instructions. Linguistics, 30, 5-55. Giv6n, T. ( 1998). The usual suspects: The grammar of perspective in narrative fiction. U Diversity of Oregon Institute of Cognitive and decision sciences: Technical report no.9H6. Glenberg, A. M., & Kruley, P. ( 1992). Picture and anaphora: Evidence for independent processes. Memory and cognition, 20,461-471. Gordon, P. C., & Davina, D. (1995). Pronouns, passives, and discourse coherence.

Journal of memory and language, 34, 21~231. Gordon, P. C., Grosz, B. J., & Gilliom, L. ( 1993 ). Pronouns, names, and the centering of attention in discourse. Cognitive Science, 17,311-348. Gordon, P. C., & Scearce, K. A. (1995). Pronominalization and discourse coherence, discourse structure and pronoun interpretation. Memory and cognition, 23, 313323. Greene, S. 8., Gerrig, R. J., McKoon, G., & Ratcliff, R. (1994). Unheralded pronouns and management by common ground. Journal of memory and language, 33, 511526. Grosz, B. J. (1981). Focusing and description in natural language dialogues. In A. K. Joshi, B. L. Webber, & I. A. Sag ( Eds. ), Elements ofdiscourse understanding (pp. 84105 ). Cambridge: Cambridge University Press. Grosz, B. J., Joshi, A., & Weinstein, S. (1986). Towards a computational theory of discourse interpretation. Unpublished Ms. Grosz, B. J., Joshi, A., & Weinstein, S. (1995). Centering: A framework for modeling the local coherence ofdiscourse. IRCS Report 95-01. The Institute for research in cognitive science, University of Pennsylvania. Gundel,}. K., Hedberg, N., & Zacharski, R. (1993). Cognitive status and the form of referring expressions in discourse. Language, 69, 27 4-307. Gundel, J. K., & Mulkern, A. E. ( 1998). Quantity implicatures in reference understanding. Pragmatics and cognition, 6, 21-45. Hakulinen, A. (1987). Avoiding personal reference in Finnish. In J. Verschueren & M. Bertuccelli-Papi (Eds ), The pragmatic perspective: Proceedings of the international pragmatics conference (pp. 141-153). Amsterdam: John Benjamins. Halmari, H. (1996). On accessibility and coreference. InT. Fretheim, & J. K. Gundel (Eds.), Reference and referent accessibility (pp. 155-178). Amsterdam: John Benjamins. HeUer, D. (1998 ). The demonstrative set-A second pronominal copula in Hebrew. Unpublished Ms. Tel Aviv University. Hermon, G. (1985). Syntactic modularity. Dordrecht: Foris. Hinds, J. (1983). Topic continuity in Japanese. ln T. Giv6n (Ed.), Topic continuity in Discourse: A Quantitative Cross-Language Study (pp.43-95). Amsterdam: John Benjarnins. Hoek, K. van (1995). Conceptual reference points: A cognitive grammar account of

83

84

Mira Ariel

pronominal anaphora constraints. Language, 71,310--40. Huang, Y. (1994). The syntax and pragmatics of anaphora. Cambridge: Cambridge University press. Hyman, L. M., & Comrie. B. ( 1981 ). Logophoric reference in Gokana. Journal ofAfrican languages and linguistics, 3, 19-37. Jucker, A. H. ( 1996). News actor labelling in British newspapers. Text, 16, 373-90. Karmiloff-Smith, A. (1985). Language and cognitive processes from a developmental perspective. Language and cognitive processes, 1, 61-85. Keenan, E. L. ( 1994). Crtt~ting anaphors: An historical study of the English reflexive pronouns. Unpublished MS, UCLA. Kempson, R. M. ( 1984). Wtt~k crossover, logical form and pragmatics. Paper delivered at GLOW, April1984. Keysar, B., Barr, D.}., & Balin, J. A. ( 1998). Definite reference and mutual knowledge: Process models of common ground in comprehension. Journal of memory and language, 39, 1-20. Khan, G. ( 1999). A grammar of neo-Aramaic: The dialed of the Jews in Arbel. Leiden: Brill Kibrik, A. A. (1996). Anaphora in Russian narrative prose: A cognitive calculative account. In B. Fox (Ed.), Studies in anaphora (pp. 255-303). Amsterdam: John Benjamins. Kirsner, R. S. (1979). Deixis in discourse: An exploratory quantitative study of the modem Dutch demonstrative adjectives. In T. Giv6n (Ed.), Discourse and syntax (pp. 355-375). New York: Academic Press. Kirmer, R. S. ( 1990). Grappling with the ill-defined: Problems of theory and data in synchronic grammatical description. In R Amacker, & R Engler (Eds. ), Presence de Saussure: Actes du CoUoque lntemationale de Geneve (21-23 Mars 1988) (pp. 187201 ). Publications du Cercle Ferdinand de Saussure 1. Librairie Droz, Geneva. 187-201. Kirsner, R. S., & Heuven, V. J. van ( 1988 ). The significance of demonstrative position in modem Dutch. Lingua, 76, 209-48. Kronrod, A., & EngeL 0. (2001). Accessibility theory and referring expressions in newspaper headlines. Journal ofPragmatics, 33, 683-699. Kuno, S. ( 1987). Functional syntax. Chicago: The Chicago University Press. Levinson, S. C. ( 1987). Pragmatics and the grammar of anaphora: A partial pragmatic reduction ofbinding and control phenomena. Journal oflinguistics, 23, 379-434. Levinson, S. C. ( 1991 ). Pragmatic reduction of the binding conditions revisited. Journal oflinguistics, 27,107-161. Lichtenberk, F. ( 1996). Patterns of anaphora in To'aba'ita narrative discourse.ln B. Fox (Ed.), Studies in anaphora (pp. 379-4ll). Amsterdam: John Benjamins. Linde, C. ( 1979). Focus of attention and the choice of pronoun in discourse. InT. Giv6n (Ed.), Syntax and semantics 12: Discourse and syntax. (pp. 337-354). New-York: Academic Press. Lucas, M. M., Tanenhaus, M., & Carlson, G. N. (1990). Levels of representation in the interpretation of anaphoric reference and instrument inference. Memory and cog-


nition, 18,611-631. MacDonald, M., & MacWhinney, B. ( 1990). Measuring inhibition and facilitation from pronouns. journal of memory and language. 29, 469-92. Maes, A. A., & Noordman, L. G. M. (1995). Demonstrative nominal anaphors: A case of nonidentificational markedness. Linguistics, 33, 255-282. Marsico-Wilson, W., Levy, E., & Komisarjevsky Tyler, L. ( 1982). Producing interpretable discourse: The establishment and maintenance of reference. In R. J. Jarvella & W. Klein (Eds), Speech, Place and Action (pp. 339-380). Chichester: John Wiley and sons. Matsui, T. (1998). Pragmatic criteria for reference assignment A relevance-theoretic account of the acceptability of bridging. PragmtJtics and cognition. 6, 47-97. Mauner, G., Tanenhaus, M. K., & Carlson, G. N. ( 1995). Implicit arguments in sentence processing. journal of memory and language. 34. 357-382. McCawley, J.D. (1979). Presupposition and discourse structure. In C. K. Oh, & D. A. Dinneen (Eds.), Syntax and semantics voL 11: Presupposition (pp. 371-388). New York: Academic press. McEnery, T. & Thomas, J. (I 992). The pragmatic basis ofautomated pronoun resolution. A paper presented at the 4th IPrA conference, Kobe Japan. McDonald J. L., & MacWhinney, B. (1995). The time course of anaphor resolution: Effects of implicit verb causality and gender. journal of memory and language, 34, 543-566.

McKoon, G., Ward, G., Ratcliff. R., & Sproat, R. (1993). Morphosyntactic and pragmatic factors affecting the accessibility of discourse entities. journal ofmemory and language, 32,56-75. Mehudar, 0. (1996). The phenomenon of obviation in third person pronouns in Indian languages in light ofaccessibility theory. A Tel Aviv University seminar paper. Mithun, M. ( 1996 ). Prosodic cues to accessibility. InT. Fretheim, & J. K. Gundel (Eds. ), Reference and referent accessibility (pp. 223-233 ). Amsterdam: John Benjamins. Montgomery,M. (1989). Choosing between that and it. In R. W. Fasold, & D. Schiffrin (Eds. ), Language change and variation (pp. 241-254). Amsterdam: John Benjamins. Mueller-Lust, R. A. G., & Gibbs, R. W. ( 1991). Inferring the interpretation of attributive and referential definite descriptions. Discourse Processes, 14. I07-31. Mulkern, A. E. (1996). The game of the name. InT. Fretheim, & J. K. Gundel (Eds.), Reference and referent accessibility (pp. 23~250). Amsterdam: John Benjamins. Oakhill, }., Gamham, A., Gernsbacher, M.A., & Cain, K. (1992). How natural are conceptual anaphors? Language and cognitive processes, 7, 257-80. O'Brien, E. J& Albrecht, J. E ( 1991 ). The role of context in accessing antecedents in text. journal ofexperimental psychology: Learning, memory and cognition, 17, 94-102. Onishi, K. H., & Murphy, G. L. ( 1993 ). Metaphoric reference: When metaphors are not understood as easily as literal expressions. Memory and cognition, 21, 763-772. Paterson, K. B., Sanford, A. J., Moxey, L. M., & Dawydiak, Eugene (1998). Quantifier polarity and referential focus during reading. journal of memory and language, 39, 290-306.

Reboul, A. ( 1997). What (if anything) is accessibility? A relevance-oriented criticism

Bs

86

Mira Ariel

of Ariel's accessibility theory of referring expressions. In J, H. Connolly, R. M. Vismans, C. S. Butler, & R. A. Gatward (Eds.), Discourse and pragmatics in functional grammar (pp. 91-108). Berlin: Mouton de Gruyter. Reinhart, T. ( 1983). Anaphora and semantic interpretation. Chicago: Chicago University Press. Reinhart, T., & Reuland, E. (1993). Reflexivity. Linguistic inquiry, 24,657-720. Rieder, G., & Mulokandov, N. (1998). The cliticization and deletion of first person pronouns in Hebrew. A seminar paper. Tel Aviv University. Rinck, M., & Bower, G. H. ( 1995). Anaphora resolution and the focus of attention in situation models. Journal of memory and language, 34, 11~ 131. Ro~n. V. (1996 ). The interpretation of empty pronouns in Vietnamese. InT. Fretheim, & J, K. Gundel (Eds.), Reference and referent accessibility (pp. 251-261). Amsterdam: John Benjamins. Saadi, I. (1997). Untitled seminar paper. Tel Aviv University. Sanford, A. J,, &Garrod, S.C. (1981). Understanding written language. Chichester: John Wiley and sons. Sanford, A.}., Moore, K., & Garrod, S.C. (1988). Proper names as controllers of discourse focus. Language and Speech, 31,43-46. Sanford, A.}., & Moxey, L. M. (1995). Notes on plural reference and the scenariomapping principle in comprehension. In G. Rickheit, & C. Habel (Eds.), Focus and coherence in discourse processing (pp. 18--43). Berlin: Walter de Gruyter. 18-34. Sanford A. J,, Moxey. L. M., & Paterson, K. B. ( 1996). Attentional focusing with quantifiers in production and comprehension. Memory and cognition, 24, 144-155. Schiffman, R. J. (1984). The two nominal anaphors it and that. In}. Drogo, V. Mishra, & D. Testen (Eds.), Papers from the twentieth regional meeting of the Chicngo Linguistic Society (pp. 344-357). Chicago: Chicago Linguistic Society. Schilperoord, J. (1996). It's about time: Temporal aspects of cognitive processes in text production. Amsterdam/Atlanta: Rodopi. Sells, P. ( 1984). Syntax and semantics ofresumptive pronouns. PhD thesis, University of Massachusetts. Sperber, D., & Wilson, D. ( 1986). Relevance. Oxford: Blackwell. Sproat, R., & Ward, G. ( 1987). Pragmatic considerations in anaphoric island phenomena. CLS, 23,321-35. Stebbins, T. (1997). Asymmetrical nominal number marking: A functional account. Sprachtypologie and Universalienforschung, 50, 5-47. Tao, L. ( 1996). Topic discontinuity and zero anaphora in Chinese discourse: Cognitive strategies in discourse processing. In B. Fox (Ed.), Studies in anaphora (pp. 487513 ). Amsterdam: John Benjamins. Terken, J,, & Nooteboom, S. }. (1988). Opposite effects of accentuation on verification latencies for Given and New information. Language and cognitive processes, 2, 145163.

Tomlin, R. S. (1987). Linguistic reflections of cognitive events. In R. Tomlin (Ed.), Coherence and grounding in Discourse (pp. 455-480). Amsterdam: John Benjamins. Toole, J. (1992 ). Local or global: An investigation ofthe effect ofgenre on referential choice.


An MA thesis submitted to Monash University. Toole, J. (1996). The effect of genre on referential choice. InT. Fretheim, & J. K. Gundel (Eds.), Reference and referent accessibility (pp. 263-290). Amsterdam: John Benjamins. Vonk, W., Lettica, G., Hustinx, M., &W. Simons (1992). The use of referential expressions in structuring discourse. Language and cognitive processes, 7, 301-333. Walker, M. A., Iida, M., & Cote, S. (1990). Centering in Japanese discourse. In COLING

90 : Proceedings of the 19th international conference on computational linguistics, 251-261.

Walker, M.A., & Prince, E. F. (1996). A bilateral approach to Givenness: A hearerstatus algorithm and a centering algorithm. InT. Fretheim, & J. K. Gundel (Eds.), Reference and referent accessibility (pp. 291-306). Amsterdam: John Benjamins. Ward, G., Sproat, R, & McKoon, G. ( 1991 ). A pragmatic analysis of so-called anaphoric islands. Language, 67,439-74. Webber, B. L. ( 1991). Structure and ostension in the interpretation of deixis. Language and cognitive processes, 6, 107-35. Ziv, Y. ( 1994 ). Left and right dislocations: Discourse functions and anaphora. Journal of pragmatics, 22, 629-645. Ziv, Y. (1996). Pronominal reference to inferred antecedents. The Belgian journal of linguistics, 10, 55-67. Zribi-Hertz, A. (1980). Coreference et pronoms ref1echis: Note sur le contraste lui/luimeme en Fran~. Linguisticae Investigationes, 4. 131-179. Zribi-Hertz, A. ( 1989). Anaphor binding and narrative point of view: English reflexive pronouns in sentence and discourse. Language, 65,695-727.

87

CHAPTER

3

The influence of text cues on the allocation of attention during reading Michelle L. Gaddy* Paul van den Broek Yung-Chi Sung University of Minnesota

When attempting to comprehend text, readers must devise ways to condense the wealth of information that is presented to them. Limitations on working memory or attentional resources make this task particularly important, for without the ability to selectively attend to particular textual information, readers would be faced with the burden of trying to retain every piece of information that is presented in text throughout the reading process. Successful comprehension of text therefore involves a reader's ability to appropriately allocate his or her attention to the most important aspects of a text, and at the right times, during reading. The purpose of this chapter is to describe how linguistic, typographical, and text-structure cues guide readers' attention during the process of reading. Linguistic cues are verbal devices that "include short signal phrases that pave the way for important items of information" (Glynn, Britton & Tillman 1982, p. 196). Examples of linguistic cues include function and relevance indicators, anaphors, and cataphors. Typographical and text-structure cues are devices that are used by authors to help readers determine the main points of a text (Leon & Carreteras 1992; Lorch 1989). Examples of typographical cues are boldface type, italic font, and underlining. Examples of text-structure cues are titles and headings. In this chapter we illustrate how linguistic, typographical, and text-structure cues direct the allocation of attention during reading and how, in doing so, they affect the comprehension process as well as its outcome, the mental representation of the text as a whole. In the first section we provide a brief overview of the interplay between attention, comprehension processes, and the construction of a memory representation. In the second section we

90

MicheUe L. Gaddy, Paul van den Broek and Yung-Chi Sung

discuss the role of the various types of text cues and review relevant findings in cognitive studies of discourse comprehension.

1.

The role of attention allocation in reading: The Landscape Model

Cognitive research on reading has made clear the central role that attention plays in the comprehension process (e.g., Hidi 1995; Just & Carpenter 1980, 1992; Kintsch & Van Dijk 1978 ). A reader's allocation of attention to particular elements in a text has a direct impact on the way in which the text is interpreted, as is evident in recent theoretical models of reading (Goldman & Varma 1995; Kintsch 1988; Langston & Trabasso 1998; Van den Broek, Risden, Fletcher & Thurlow 1996; Van den Broek, Young, Tzeng & Linderholm 1998). The importance of attention allocation, and hence of the role that text cues play in reading, is best appreciated when one considers in some detail the cognitive processes that occur during reading. We will use the Landscape model (Van den Broek et al. 1996, 1998) to lay out the most recent theoretical insights into these processes. In the Landscape model, as well as in other recent models, reading is conceptualized as a cyclical process. With each consecutive reading cycle a new text segment and its constituent concepts are processed: 1 New textual information enters the reader's working memory or attentional buffer (i.e., is "activated") and, because working memory capacity is severely limited (e.g., Just & Carpenter 1992), information that was in working memory during the preceding cycle is at least in part erased. Moreover, in the course of processing and interpreting the new information the reader may activate background knowledge or reactivate text information from earlier reading cycles. This results in further competition for the limited attentional resources. Thus, during each reading cycle four potential sources determine the activation values of concepts: the text that is currently being processed, the immediately preceding reading cycle, reading cycles that occurred even earlier than the immediately preceding cycle, and the readers' own background knowledge (Van den Broek et al. 1996, 1998 ). As we will discuss below, various factors, including text cues, influence which of these sources will contribute in a particular cycle. Over the course of reading, individual concepts fluctuate in their activation as the reader proceeds from cycle to cycle: Some concepts come into the focus of attention, others fade, and yet others remain in working memory but fall and rise in the level of their activation.2 Together, the fluctuations in activations of all concepts form a landscape. Figure 1 provides an example of

Influence of text cues

Table l.

Example Expository Text Fragment (Numbers indicate processing cycles)

Why American Songbirds are Vanishing ( 1) The steep declines in waterfowl, shoreline birds, and grassland birds over the past several decades generally are well understood (2). What is not as obvious is why forest-dwelling migratory songbirds also are vanishing- especially the so-called Neotropical migrants that breed in northern latitudes but migrate to winter homes in the tropics (3). As decreases in their populations accumulated (4), it was widely noted that the missing

species could still be found in large continuous tracts of forests but not in isolated tracts(S). This observation was dubbed the forest fragmentation effect (6 ). What possible explanations might be given for the forest fragmentation effect (7)? One simple hypothesis to explain the effect is that they generally prefer larger forest plots as nesting sites and so avoid isolated plots because they tend to be small (8). It is important to note that this hypothesis predicts that the density of these birds' nests in a forest will increase as the size of the forest increases (9). However, when they set about documenting the presence or absence of songbird species in forest fragments of different size ( 10), researchers obtained mixed and- sometimes - contradictory results ( 11). (Adapted from Scientific American)

such a landscape, based on the text in Table 1. The horizontal dimensions reflect the reading cycles and the major concepts relevant to the text, respectively, whereas the vertical dimension represents the degree of activation of concepts in each cycle. Activation levels are relative so any numerical scale could be used. Here, we use a scale of'O' (no activation) to '8' (high activation). Later (also see footnote 3) we will describe the particular patterns of activation in this example, but for now the important point is that concepts fluctuate in their activation across reading cycles. The patterns of activations across reading cycles form the basis for the memory representation that the reader constructs of the text. At each cycle in Figure 1, a cross-section of the landscape shows the different concepts that are activated, to various degrees, in that cycle. When two concepts are activated simultaneously, a connection is built between them in the reader's episodic memory for the text. The strength of this connection is a function of the amount of activation of each of its constituent concepts: The more strongly each is activated, the stronger the ensuing connection. Furthermore, if two concepts repeatedly are co-activated the memory connection between them is updated with each recurrence, again in proportion to the activations of the two

91

9~

Michelle L. Gaddy, Paul van den Broek and Yung-Chi Sung

ACTI\IAnON

f

0

R&11DINO CYCLES

Figure 1.

"

Landscape of activations for the Songbird passage

concepts. In each cycle, multiple connections are created or updated. In this fashion, the reading process results in the gradual building of a representation of the text. The evolving representation, in turn, affects subsequent reading cycles: Concepts that are activated in the current cycle elicit partial activation of other concepts that have become associated to them in the reader's current memory representation or background knowledge. As a consequence of this process, called cohort activation, the incoming information in each cycle is interpreted against the backdrop of the interpretations of the information in preceding cycles (Van den Broek et al. 1998).3 This overview of the cognitive processes that take place during reading makes clear the central importance of attention allocation. The dynamic distribution of attention over the course of reading determines which concepts enter into the comprehension process and how they connect to each other. A change


in the attentional distribution at any point during reading reverberates throughout the comprehension process and the outcome of this process, the memory representation of the text. Many aspects of the Landscape depiction of the reading process are based on prior research by numerous investigators of discourse processing (for reviews, see Gemsbacher 1994; Balota, Flores d'Arcais & Rayner 1990). However, in the majority of these earlier studies the focus has been on one or two aspects at a time, ignoring or holding constant the others. The Landscape model extends these earlier conceptualizations by including additional components, amongst them notions from models of memory representation and access (e.g., connectionist models such as McClelland & Rumelhart 1988), and by simultaneously capturing all components and their interactions. The depiction of the reading process in all its dynamic detail leads to a considerable improvement in explanatory power and to an increased ability to predict human behavior (Goldman & Varma 1995; Kintsch 1988; Langston & Trabasso 1998; Van den Broek et al. 1996, 1998). At its most general level the model provides a conceptual framework of the reading process, which allows one to understand the complex interplay of factors and to construct well-grounded hypotheses about the impact of a specific factor. At a more detailed level, the model can be used to test empirically a specific theory or hypothesis, by implementing the theory or hypothesis and by comparing the predicted landscape of fluctuating activations and accompanying memory representation to empirical data. A good example is the investigation of the various possible sources of activation. Most investigators agree that concepts in the current reading cycle and, to a lesser extent, from the preceding cycle influence the activation of concepts. There is less agreement, however, about the contribution of the other two potential sources, reinstatement of concepts from prior cycles and the recruitment ofbackground knowledge (e.g., Graesser, Singer & Trabasso 1994; McKoon & Ratcliff 1992; cf. Van den Broek, Fletcher & Risden 1993). Under what circumstances are these sources accessed, if at all? Competing views about source access during reading of narratives have been implemented in the Landscape model and their predictions for on-line activation and for the eventual memory representation were compared to empirical data (Risden 1996; Van den Broek et al. 1998). The results indicate that readers have standards for coherence, by which they gauge their comprehension and determine whether to engage in further processing (Van den Broek, Risden & Husebye-Hartmann 1995). Although standards of coherence vary depending upon the reader and the circumstances (e.g., read-

93

94


ing goal, see Linderholm, Gustafson, Van den Broek & Lorch 1997), two standards of coherence that are almost universally applied by readers in normal reading situations involve establishing referential and causal coherence. H the information in the current sentence and that which is carried over from the preceding sentence do not provide referential and causal coherence, then the reader will access the other two sources, prior cycles and background knowledge.4 This example illustrates how the Landscape model can help us understand and test specific hypotheses about factors that may affect the reading process and the emerging memory representation. In the past two decades, considerable knowledge has been gained about how characteristics of the reader (comprehension processes, attentional capacity, background knowledge, etc.), in interaction with text content, affect activation patterns during reading. In contrast, very little is known about the role of the textlinguistic form of the reading materials (Lorch 1989). In the remainder of this chapter we will explore how textual cues may affect the distribution of readers' attentional resources and how such effects would reverberate throughout the reading process and the construction of the final representation of the text.

1.

The role of textual cues in attention allocation

1.1

Linguistic cues

Meyer ( 1975) describes linguistic cues as a class of textual devices that give emphasis to certain aspects of the semantic content of text. These linguistic cues serve as verbal devices that assist readers in deciding where attention should be allocated during the reading process (Glynn 1978). We will focus on three particular types of linguistic cues: function and relevance indicators, anaphors, and cataphors.

Function and relevance indicators Function and relevance indicators, or pointer words as they are sometimes called, are used in text to "explicitly inform the reader of the author's perspectives of a particular idea" (Meyer 1975, p. 80). These types of cues also can be considered importance indicators, for they indicate to the reader which content in text is most important (Lorch 1989). Examples of function and relevance indicators include phrases such as "it is important to note that ... ", "let


me stress that ... ", "fortunately... ", and "in summary ... ". In the Songbird passage, the following sentence segment contains an example of this type of linguistic cue: (1)

It is important to note that this hypothesis... (cycle 9)

Function and relevance indicators usually precede the content that they signal, but occasionally succeed it (e.g., "The British government in particular has been keen on this idea"). In terms of the Landscape model, concepts signaled by function and relevance indictors will be highly activated as a result of the increased attention that readers pay to textual information that is cued as important. Furthermore, these concepts will not only be more strongly activated in the current reading cycle, they also will be carried over more frequently into subsequent cycles. As a result, their connections to other textual information will be both stronger and more numerous. This, in tum, will result in increased memory for textual content preceded by function and relevance indicators. In example ( 1), the information following the relevance indicator, the prediction that nest density increases with forest size, is highlighted and will receive more activation. Figure 2 displays the change in activations for the text concepts as a result of function and relevance cues as well as of the cues to be described below. 5 As a result of these changes in activation, some concepts will accumulate more activation over the course of the entire text and, also, enter in more or stronger connections with other concepts that are activated in the same cycle or, through carry over, with concepts in immediately following cycles. In example ( 1), the concepts in the sentence marked by the functional marker (e.g., 'predict', 'density', 'forest size') nearly double in the strength with which they are encoded in the memory representation and more than double in their overall connectedness to other concepts in the text. 6 The effects of function and relevance indicators on the eventual representation of a text are fairly well documented. For example, readers recall text information more often when it is signaled than when it is not signaled (Leon & Carreteras 1992; Loman & Mayer 1983; Lorch & Lorch 1986; Mayer, Dyck & Cook 1984). Less is known, however, about the way in which function and relevance indicators direct readers' attention during the reading process itself. In one of the few studies in which both on-line reading behaviors and off-line memory were investigated, Lorch and Lorch ( 1986) observed that marking text segments with a summary indicator (e.g., "To summarize, ... ") prompted readers to slow their reading significantly. In contrast, marking text segments

9S

96

Michelle L. Gaddy, Paul van den Broek and Yung-Chi Sung

CONCEPTS

.,

READING CYCLES

Figure 2. Change in landscape of activations for the Songbird passage as a result of the inclusion of textual cues

with importance signals (e.g., "One argument is particularly notable") did not result in slower reading. Both types of markers led to significant improvements in recall of the signaled text segments. The absence of reading-time effects for importance signals led Lorch and Lorch to speculate that these signals may prompt readers to tag the information as important as soon as it is encountered, thereby leaving reading time unaffected. These findings, and the possibility that the two types of signals evoke different kinds of processing, are readily understood in the context of the Landscape model There are two ways in which a concept can receive increased activation. First, it can be processed for a longer period of time, in essence extending the number of processing cycles, or by increasing the activation it receives in a particular cycle (see Figure 2). A summary signal prompts the


reader to connect the current information with concepts that had been activated in preceding cycles. This requires reinstatement of the earlier concepts, a process which has been shown to be time consuming resulting in slower reading times (e.g., Albrecht & O'Brien 1993; Van den Broek & Thurlow 1990; cf. Van den Broek 1994).ln contrast, an importance signal prompts the reader to allocate more of the available activation in a particular cycle to the signaled concept, a process which does not alter the length of the reading cycle. Thus, both types of signals result in a stronger memory representation of the signaled information. In the case of summary signals it is through stronger connections with concepts from prior cycles, and in the case of importance signals it is through higher activation of the current concepts and hence through stronger connections to concepts in subsequent cycles.

Anaphors Anaphoric devices alert readers to the fact that an object or event mentioned in one sentence is identical to an object or event mentioned in a preceding sentence (Albrecht & O'Brien 1993; Gemsbacher 1990; Van den Broek 1994). The foUowing sentence, taken from the passage in Table 1, contains an example of an anaphor. (2)

One simple hypothesis to explain the forest fragmentation effect is that they generally prefer larger forest plots as nesting sites ... (cycle 8)

The word "they" serves as an anaphor by referring to "songbird populations", which is mentioned three sentences earlier. Anaphora relate sentences by means of anaphoric reference, and readers establish basic coherence in text by resolving anaphoric references. In the Landscape model an anaphoric device improves access to previously mentioned concepts in text by eliciting their re-activation. As a result these concepts will participate in additional reading cycles and hence form more and/or stronger connections to other concepts. As a result, they wiU feature more strongly in the memory representation of the text. Again, Figure 2 illustrates the effects of anaphors on activation patterns in the Songbird passage. There is ample evidence that anaphors lead to reactivation of their referents. For example, McKoon and Ratcliff ( 1980) asked participants to read short paragraphs in which the final sentence of the paragraph either did or did not contain an anaphoric reference to a concept mentioned in the first sentence of the paragraph. After each paragraph was read, participants were given a recog-

97

98


nition test for the referent mentioned in the first sentence. When the last sentence contained an anaphor, participants recognized the referent more quickly than when the last sentence did not contain an anaphor, demonstrating that anaphors activate their referents. In a similar study, Dell, McKoon, and Ratcliff ( 1983) demonstrated that an anaphor not only reactivates its referent but also other concepts that were in the same clause or sentence as the referent. However, the reactivation of other concepts is weaker than that of the referent: Unlike the referent, the 'other' concepts quickly decline in activation level once the anaphor is processed. Similar findings on the role of anaphors in directing activation through reinstatement have been obtained under different circumstances and using a variety of measures (e.g., Clifton, Kennison & Albrecht 1997; Ehrlich & Rayner 1983; O'Brien, Duffy & Myers 1986; Vonk, Hustinx & Simons 1992).7 These findings support the Landscape's predictions concerning the activation patterns that result from anaphoric references. As predicted, anaphors result in the reactivation of referents who thus enter in additional reading cycles. Moreover, the observation that concepts that co-occurred with the referents also will be reactivated, albeit weakly, is readily understood in the context of cohort activation. By virtue of their co-occurrence, the referent and the other concepts in that cycle have become connected. When the referent is reactivated, cohort activation will lead to partial reactivation of the other concepts as well. The implications of anaphoric references for the eventual memory representation are less well investigated. According to the Landscape model, the fact that anaphoric referents and, to a lesser extent, their cohorts participate in the current cycle and, through carry over, to some extent in subsequent cycles leads to the a stronger memory trace of the reactivated concepts (through repeated activation) and to the establishment of additional connections in the memory representation (through additional co-occurrences with other concepts). Such effects have been observed with respect to concepts that are reactivated to establish causal coherence (e.g., O'Brien & Myers 1987; Van den Broek & Lorch 1993; Van den Broek & Thurlow 1990; Van den Broek, Lorch & Thurlow 1996) but as yet the predictions with respect to the effect of anaphoric references on the memory representation remain untested.

Cataphors Cataphors (sometimes called 'anticipatory anaphora') are linguistic cues that refer to subsequently mentioned concepts (Matthews 1997). An example of a


cataphoric device is the unstressed, indefinite article this (Gernsbacher & Shroyer 1989). Thus, the concept 'egg' receives more activation if it is first introduced in a text by "And then the man picks up this egg" than when introduced by "And then the man picks up an egg." An illustration of a cataphoric device in the Songbird passage is the use of 'they' to signal the upcoming referent: ( 3)

However, when they set about documenting the presence or absence ... , researchers obtained... (cycles 10 & 11)

Cataphoric devices signal to readers that the concept mentioned in or after the cataphor (In example 3, the as yet unnamed people who were documenting the absence or presence of songbirds) will become important or be mentioned again at a later point in the text. The cognitive role that cataphoric devices play during the reading process can be understood in terms of readers' attention allocation. In the Landscape model, concepts that occur in or immediately after cataphors will be highly activated. These concepts will remain activated for several cycles due to the fact that cataphors signal to readers that particular concepts will be mentioned again or become important at a later time. In contrast, concepts that are activated by non-cataphoric devices, such as a or an, will not remain activated over succeeding cycles because readers will not assume that such concepts will become relevant again. Because concepts marked by cataphors will be activated over many cycles, they will be strongly connected to other concepts in text. These connections will influence memory for the concepts signaled by cataphors. Figure 2 illustrates the impact of the cataphoric construction in the Songbird passage. It is interesting to contrast the cataphoric construction of the text in cycles 10 and 11 with the non-cataphoric form, "However, when the researchers set about documenting the presence or absence ... , they obtained ... ". In the cataphoric version the contents of the entire cycle 10 ("the people who are documenting..") is marked because its exact interpretation is not clear yet. As a result, all concepts are carried forward with increased activation (over and above carry over and cohort activation) and hence interact with the subsequent cycle ( 11 ). In contrast, in the non-cataphoric version only the concept "researchers" receives increased activation in cycle 11. Thus, the net effect of the cataphor in this example is to increase the activation of all concepts from cycle 10 during processing of the next cycle and hence to create a more tightly

99

100


interwoven memory representation of the concepts in cycles 10 and 11. As always, such changes will reverberate throughout the processing and representation of the text Relative to anaphors, cataphors have received little attention in the psycholinguistic literature. However, there is indirect evidence that cataphoric devices influence attention allocation. For example, Gemsbacher and Shroyer ( 1989) asked participants to listen to a series of short narratives and to provide endings for each narrative. Half of the narratives included a sentence in which a concept was preceded by a cataphoric device, the indefinite this, whereas the other half of the narratives included a sentence in which the same concept was preceded by a or an. Concepts preceded by the indefinite this were mentioned more frequently in participants' continuations of the narratives than were concepts preceded by a or an. Thus, cataphoric devices increase subsequent referential access to the concepts that they cue. This, in tum, suggests that the cataphor directed the way in which the listeners allocated their attention. According to the Landscape model, cataphors influence the allocation of attention during reading in a similar way: Signaled concepts receive extra activation in the current cycle and are more strongly carried over into subsequent cycles. This leads both to an overall increase of the signaled concept over the course of reading and to the strengthening or construction of connections in predictable ways. To test these on-line and off-line predictions, studies on the effects of cataphoric devices in written discourse are in order. 2..2.

Typographical cues

Just as the semantic form of text influences comprehension, so does its physical layout. Boldface type, italic type, underlining, capitalization, and color variation are examples of typographical cues that are often used by textbook authors to make certain words or phrases within text visually distinct In Table 1, the italicized phrase "forest fragmentation effect" serves as a typographical cue. The purpose of using such cues is to draw readers' attention to particular information in text, such as topic sentences or key words. Indeed, there is considerable evidence that readers tend to remember cued information more than non-cued information (e.g., Cashen & Leicht 1970; Crouse & Idstein 1972; Foster & Coles 1977; Fowler & Barker 1974; Glynn & DiVesta 1979; Golding & Fowler 1992; Klare, Mabry & Gustafson 1955; Leon & Carreteras 1992; Lorch, Lorch & Klusewitz 1995; Nist & Hogrebe 1987). As Lorch ( 1989) has pointed out, there is no reason to believe that certain typographical cues are


more effective than others, and researchers have not demonstrated the superiority of particular cues. Apparently, all varieties of typographical cues are processed in a similar manner and the memory representations that result from such processing will be similar, although it is plausible that their visual distinctiveness affects their efficacy in attracting attention. Typographical cues direct readers' attention to specific textual content, much as function and relevance signals do. According to the Landscape model this greater attention to cued textual information results in a larger and stronger memory trace for this information, both because of the total amount of activation it receives and because of new or stronger connections between cued concepts and other, co-occurring concepts in the text. To illustrate, Figure 2 shows the effects of italicizing "forest fragmentation effect" on its activation pattern across reading cycles. The change in activation pattern, in turn, results in a 33% increase in the strength of this concept in the final memory representation and in a 150% increase in its connections to other concepts. 8 The effect of typographical cues on memory for textual information is well-documented. There is ample evidence that the presence of typographical cues in text will facilitate recall of cued textual information, with little or no effect on recall of non-cued information (see Lorch, Lorch & Klusewitz 1995). The effects of typographical cues on the reading process and allocation of attention itself and, hence, how cued information is connected to other information in the text are not as well established. Apparently, typographical cues affect readers' memory representation of text by increasing the amount of attention that readers pay to cued portions of text (Lorch et al1995, Exp. 2). When cues are present in text, readers take longer to read cued than non-cued parts, and the increased processing time allotted to cued information results in that information being recalled more frequently than non-cued information. These findings provide insight in the mechanisms which may lead cued information to attain a prominent position in readers' memory representations of text, but it should be noted that measures of reading time are only indirect indicators of attention allocation. Longer reading times for particular text segments are suggestive that readers are attending more to those segments but they do not allow a firm conclusion about what actually is activated. The Landscape model allows predictions about the relative distribution and time course of activation as a result of these cues. The predictions are consistent with the reading time and memory findings but converging evidence is needed to fully grasp the way in which typographical cues exert their effects on attention allocation during reading.

101

102


2.3 Text-structure cues

Typographical and linguistic cues both operate on the content contained within the body of a text. In contrast, text-structure cues, such as titles and headings, operate by means of structuring the text as a whole. In terms of function, text-structure cues are similar to typographical cues, for both types of cues convey information about the relative importance of text elements to readers (Lorch 1989). The main difference is one of scope, with typographical cues operating primarily on the word or phrase that is being signaled and textstructure cues operating on the entire body of text.

Titles and headings In many instances, the main body of a text is preceded by a title that contains information about the content of the text. Due to variations in spatial location and typeface, titles are often easily distinguishable from the text that succeeds them (Lorch 1989), and they can be thought of as "umbrellas" under which content pertaining to a particular topic or theme can be found. Thus, titles act as advance organizers (Kozminsky 1977). The title of the passage in Table I, "Why American Songbirds are Vanishing", serves this function. Like titles, headings usually differ from the text that they signal by means of spatial location and typeface variation. Headings are text-structure cues that delineate subsections of text (Lorch 1989). They signal the organization of the text and provide the reader with information about the content of the immediately following text. The main difference between headings and titles is that titles usually apply to text as a whole whereas headings apply to subsections within text. In terms of the Landscape model, titles and headings direct readers' attention to particular content in text and bias readers' comprehension strategies. They do so in several ways. First, titles and headers contain content information that is entered into a processing cycle just like information in other text segments and, thus, influences the interpretation and activation patterns of subsequent text segments. Second, the content of titles and headings will be processed more deeply than that of regular text segments. To some extent this is due to the fact that titles and headings are physically distinct from the remainder of the text (much like segments within the text that are signaled by function or relevance cues), but, more importantly, titles and headings provide information about the theme of the text. As a result of their theme-carrying function, titles and headings may be processed more extensively, for example by


prompting the reader to recruit background knowledge, and they are more likely to be re-instated -or carried over from cycle to cycle- during the reading of the text that follows. Thus, the theme itself will be activated more strongly and more frequently, and this high level of activation will influence how the reader processes the subsequent text and the kind ofinformation that the reader remembers. Third, by providing readers with information about the theme of the text, titles and headings may prompt readers to attempt to detect congruence between later segments and the theme and to selectively allocate attention to that information that is congruent. In this fashion, titles and headings may alter the readers• comprehension processes. Figure 2 depicts how the title in the Songbird passage influences activation patterns throughout reading. This, in tum, results in dramatic changes in the memory representation. For example, the strength of the concept 'songbirds• in the representation increases by SSo/o and the strength of its connections to other concepts by 150%.9 The results of several studies on the effects of titles on comprehension support the Landscape view of the function of text-structure cues. Kozminsky ( 1977) demonstrated the potential biasing effect of titles. Participants read one of three texts that differed only with respect to their titles, each of which emphasized different ideas in the text. The results of a free-recall test administered after completion of the reading task indicated that recall was biased toward the theme emphasized in the title. The role that theme information can play in directing attention and memory is reflected in the finding that readers process text differently based on their beliefs about the content of the text. In a set of classic studies, Pichert and Anderson ( 1977) asked participants to read a story from one of two perspectives, that of a home-buyer or that of a burglar. Participants recalled different details from the passage depending on their particular perspective. Thus, readers use information about the content or theme of a text to assist them in deciding on appropriate strategies to employ during text-processing. Similarly, Zwaan ( 1994) demonstrated that readers process text differently depending on their belief about the genre of the text. Two groups of participants read the same texts, with half being told that the texts were excerpts from newspapers and the other half being told that the texts were literary stories. Expectations about text genre affected the way in which readers processed the text. Participants who thought they were reading literary stories took longer to read the texts, recalled significantly more surface information from the texts, and recalled significantly less situational information from the texts. Titles prompt readers to establish an appropriate context within which to

103

104


interpret the text (Lorch 1989). Titles achieve this by activating the appropriate background knowledge or schema that readers need in order to understand the text. When readers are given text without a title, they experience difficulty in comprehending the text, especially if the main idea of the text is rather ambiguous (Bransford & Johnson 1972; Dooling & Mullet 1973). In such cases, the presence of a title serves to "facilitate the assembly of previous knowledge that will incorporate the 'new' textual information" (Kozminsky 1977, p. 482). The effects of headings on text-processing and comprehension have received less attention than those of titles (Lorch 1989). Moreover, the extant research has been focused primarily on the influence of headings on readers' memory for textual content. In general, headings, like titles, provide readers with an explicit representation of the organization and theme of text as evidenced by the fact that readers recall more from text when supplied with a heading than when a heading is absent (Lorch 1989). The above research shows that text-structural cues such as titles and headings play an important role in readers' memory representations for text. Little is known, however, about how titles and headings actually are processed by readers and how exactly they affect the processing of subsequent text. The Landscape view holds that titles and headings will direct readers' attention during the reading process. By providing readers with information about the structure and theme of a text, these cues will be highly activated during the reading process and they will be reactivated or carried over throughout the reading process. Moreover, they prompt readers to attempt to connect subsequent text to the theme and thus direct attention allocation during later cycles. The ability to do so effectively is likely to differ among individuals. Thus, one would expect that good readers show stronger effects of text-structural cues such as titles and headings in their allocation of attention and eventual memory representation than do poorer readers.

3·

Concluding remarks

Attention allocation plays a very important role in text comprehension. In order for comprehension to be successful, readers must selectively attend to the most important parts of text, at the appropriate times during the reading process. When attention is allocated in this manner, those textual elements that receive increased attention will secure a more prominent position in readers' memory representations of text.


In this chapter we have attempted to illustrate how the content and form of text influence both the way in which text is processed and the eventual memory representation. The effect of content and form on comprehension is mediated by various linguistic, typographical, and text-structure cues that authors use to direct readers' attention to particular textual information. Some text cues have been studied primarily in terms of their effect on the memory representation that readers form of text, whereas others have been investigated mostly in terms of their effects on the allocation of attention during reading. The Landscape model provides an integrated view of reading process and the construction of a memory representation (Van den Broek et al. 1996; Van den Broek et al. 1998). In doing so, it allows us to understand the effects that have been observed and to make informed predictions about as yet unexplored effects. The Landscape model highlights the fact that text comprehension involves an interaction between reader and text characteristics. Text characteristics, such as the content and form of the text, signal to readers that attention must be allocated to certain aspects of the text. The Landscape model allows us to trace the impact of these characteristics, in the context of readers' limited attentional resources and standards for coherence, on the allocation of attention during reading and on the memory representation that gradually emerges. In doing so, it helps us understand the way in which textual cues act as attention-focusing instructions to the reader which, if properly followed, result in efficient and proficient reading and comprehension of the text.

Notes ,. We would like to thank Yuhtsuen Tzeng for his assistance in modeling the phenomena described in this chapter. The work presented here was made possible by the Guy Bond Endowment for Reading Research and by the Golestan Foundation and The Netherlands Insitute for Advanced Study. In principle any type of text segment (e.g., sentences, clauses, even paragraphs) can be selected as input unit. Most researchers adopt sentences or clauses as units of analysis, however. The reason is that, although processing occurs continuously, there is ample evidence that major processing occurs at the end of clause and sentence boundaries and thus that they constitute processing units. Likewise, input segments can be decomposed into different types of concepts (e.g., propositions, nouns/verbs). In the current chapter we use sentences as input units for the cycles and clauses (Trabasso, Van den Brode & Sub 1989) or major propositions (Kintsch 1998) as constituent concepts within a cycle.

1.

z. In older models of working memory and reading it was assumed, mostly for reasons of

105

to6 Michelle L Gaddy, Paul van den Broek and Yung-Chi Sung

simplicity, that at any particular moment in time concepts either were or were not activated in working memory (e.g., Kintsch &Van Dijk 1978; Kintsch & Miller 1980).As models have become more sophisticated and as computational power has increased, this dichotomous view has been abandoned in favor of a continuous one, in which concepts fluctuate along a dimension of degrees of activation (e.g., Goldman & Varma 1995; Just & Carpenter 1992; Kintsch 1988; Myers & O'Brien 1998; Van den Broek et al. 1998). ). We are providing only a summary of the main aspects of the Landscape model. Details of these and other aspects, as well as of their computational properties, can be found in Van den Broek et al (1998).

Songbirds passage in Figure 1. Concepts that are explicidy mentioned in a cycle receive high activation (S-7), depending on their importance in the propositional structure of the sentence (see Kintsch & Van Dijk 1978), concepts that are activated from background knowledge (e.g., the superordinate category "Birds") or reinstated (e.g., because they establish causal or referential coherence) receive intermediate activation (4), and concepts that provide spatial coherence receive some activation (2). In addition, concepts can be activated to various degrees in later cycles as a result of carry over from the preceding cycle or of cohort activation.

4- These principles form the basis for the landscape for the

s.

In this illustration all cues increase the activation value of the signaled content by 2.

6. The strength with which a concept is encoded can be calculated easily using the com-

puter implementation of the Landscape model (Van den Broek et al1998 ). For example, on a scale ofO (not encoded) to 5 (strongly encoded) the concept 'density' will be encoded with strength .9 in the unmarked version and with strength 1.6 in the marked version. On a scale of 0 (no connections) to 60 (strong/many connections), 'density' is connected to other concepts with an overall value of 11 in the unmarked version and of 26 in the marked version. 7· Different anaphors (e.g., pronouns, demonstratives, repetitions) may have slightly different effects on processing and hence on attention allocation (cf. Garrod, Freudenthal, & Boyle 1994). Such differences can be expected to lead to subde differences in attention allocation and memory representation. 8. The strength of the individual concept changed from 2.5 to 3.3 (on a 0-5 scale) and its total connections changed from 15 to 38 (on a 0-60 scale).

9· The strength of the individual concept changed from 2. to 3.7 (on a 0-5 scale) and its total connections changed from 18 to 48 (on a 0-60 scale).

References Albrecht, J. E., & O'Brien, E. J. (1993). Updating a mental model: Maintaining both local and global coherence. Journal of Experimental Psychology: Learning, Memory 6- Cognition, 19, I 061-1070. Balota, D. A., Flores d'Arcais, G. 8., & Rayner, K. (Eds.) ( 1990). Comprehension processes in reading. Hillsdale, NJ, USA: Lawrence Erlbaum Associates, Inc.


Bransford, J. D., & Johnson, M. K. ( 1972). Contextual prerequisites for understanding. Journal of Verbal Learning and Verbal Behavior, 11,717-726. Cashen, V. M., & Leicht, K. L. ( 1970). Role of the isolation effect in a formal educational setting. Journal ofEducariont~l Psychology, 61,484-486. Clifton, C., Jr., Kennison, S.M., & Albrecht, J. E. (1997). Reading the words her, his, him: Implications for parsing principles based on frequency and on structure. Journal ofMemory & Language, Vol36(2), 276-292. Crouse, J. H., & ldstein, P. ( 1972). Effects of encoding cues on prose learning. Journal of Eduaarional Psychology, 63, 309-313. Dell. G. S., McKoon, G., & Ratcliff, R. ( 1983). The activation of antecedent information during the processing of anaphoric reference in reading. Journal of Verbal Learning and Verbal Behavior, 22, 121-132. Dooling, D. J., & Mullet, R. L. ( 1973). Locus of thematic effects in retention of prose. Journal ofExperimental Psychology, 97, 404-406. Ehrlich, K., & Rayner, K. ( 1983 ). Pronoun assignment and semantic integration during reading: Eye movements and immediacy of processing. Journal of Verbal Learning & Verbal Behavior, 22,75-87. Foster, J., & Coles, P. (1977). An experimental study of typographic cueing in printed texL Ergonomics, 20, 57-66. Fowler, R. L., & Barker, A. S. ( 1974). Effectiveness of highlighting for retention of text material Journal ofApplied Psychology, 59, 358-364. Garrod, S., Freudenthal, D., & Boyle, E. A. ( 1994). The role of different types of anaphor in the on-line resolution of sentences in a discourse. Journt~l of Memory & Language, 33, 39-68. Gemsbacher, M.A. ( 1990). Language comprehension as structure building. Hillsdale, NJ: Erlbaum. Gemsbacher, M.A. (Ed.). (1994). Handbook ofpsycholinguistics. San Diego, CA, USA: Academic Press, Inc. Gemsbacher, M.A., & Shroyer, S. (1989). The cataphoric use of the indefinite this in spoken narratives. Memory & Cognition, 17, 536-540. Glynn, S. M. ( 1978). Capturing readers' attention by means of typographical cueing strategies. Educational Technology, 18,7-12. Glynn, S. M., Britton, B. K., & Tillman, M. H. ( 1982). Typographical cues in text: Management of the reader's attention. In D. H. Jonassen (Ed.), The technology of text (Vol2, pp. 192-209). Englewood aiffs, NJ: Educational Technology Publications. Glynn, S.M., & DiVesta, F. J. (1979). Control of prose processing via instructional and typographical cues. Journal ofEducational Psychology, 71,595-603. Golding, J. M., & Fowler, S. B. (1992). The limited facilitative effect of typographic signals. Contemporary Educational Psychology, 17,99-113. Goldman, S. R., & Varma, S. (1995). CAPping the construction-integration model of discourse comprehension. In C. A. Weaver, S. Mannes, & C. R. Fletcher (Eds.), Discourse comprehension: Essays in honorofWalter Kintsch (pp. 337-358). Hillsdale, NJ: Erlbaum.

107

1o8 MicheUe L. Gaddy, Paul van den Broek and Yung-Chi Sung

Graesser, A. C., Singer, M., & Trabasso, T. (1994). Constructing inferences during narrative text comprehension. Psychological Review, 101,371-395. Hidi, S. E. ( 1995). A reexamination of the role of attention in learning from text. Educational Psychology Review, 7, 323-350. 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. ( 1992 ). A capacity theory of comprehension: Individual differences in working memory. Psychological Review, 99, 122-149. Kintsch, W. ( 1988). The role of knowledge in discourse comprehension: A construction-integration model Psychological Review, 95, 163-182. Kintsch, W. (1998). Comprehension: A paradigm for cognition. New York, NY, USA: Cambridge University Press. Kintsch, W., & Van Dijk. T. A. (1978). Toward a model of text comprehension and production. Psychological Review, 85, 363-394. Klare, G. R., Mabry, J. E., & Gustafson, L M. ( 1955). The relationship of patterning (underlining) to immediate retention and to acceptability of technical material. The Journal ofApplied Psychology, 39, 40-42. Kozrninsky, E. ( 1977). Altering comprehension: The effect ofbiasing tides on comprehension. Memory and Cognition, 5, 482-490. Langston, M. C., & Trabasso, T. ( 1998). Identifying causal connections and modeling integration of narrative discourse. In H. van Oostendorp, & S. R. Goldman (Eds.), The construction ofmental representations during reading(pp. 29-69). Mahwah, NJ: Erlbaurn. Leon, J. A., & Carretero, M. (1992). Signal effects on the recall and understanding of expository texts in expert and novice readers. In A. J. M. Oliveira (Ed.), Structures ofcommunication and intelligent help for hypermedia courseware (pp. 97-111 ). New York: Springer. Linderholm, T., Gustafson, M., Van den Broek. P., & Lorch, R. F. (March, 1997). The effect of reading goals on inference generation during reading expository text. Paper presented at the Annual Meeting of the American Educational Research Association, Chicago, IL. Loman, N. L., & Mayer, R. E.( 1983 ). Signaling techniques that increase the understandability of expository prose. Journal of Educational Psychology, 75,402-412. Lorch, R. F., Jr. (1989 ). Text-signaling devices and their effects on reading and memory processes. Educational Psychology Review, 1, 209-234. Lorch, R. F., Jr., & Lorch, E. P. (1986). On-line processing of summary and important signals in reading. Discourse Processes, 9, 489-496. Lorch, R. F., Jr., & Lorch, E. P. ( 1996 ). Effects of organizational signals on free recall of expository text. Journal ofEducational Psychology, 88, 38-48. Lorch, R. F., Jr., Lorch, E. P., & Klusewitz, M.A. (1995). Effects of typographical cues on reading and recall of text. Contemporary Educational Psychology, 20, 51-64. Matthews, P. H. (1997). The Concise Oxford Dictionary of Linguistics (p. 48). Oxford University Press, Inc.

Influence of text cues 109

Mayer, R E., Dyck, )., &Cook, L. K. (1984). Techniques that help readers build mental models from scientific text: Definitions pretraining and signaling. Journal ofEducational Psychology, 76, I 089-1105. McClelland, J, L., & Rumelhart, D. E. ( 1988). Explorations in parallel distributed processing: A handbook of models, programs, and exercises. Computational models ofcognition and perception. Cambridge, MA: MIT Press. McKoon, G., & Ratcliff, R. (1980). The comprehension processes and memory structures involved in anaphoric reference. Jounral ofVerbal Learning and Verbal Behavior, 19, 668-682. McKoon, G., & Ratcliff, R ( 1992). Inference during reading. Psychological Review, 99, 440-466. Meyer, B. J. F. (1975). The organization of prose and its effects on memory. New York: Elsevier. Myers, J. L., & O'Brien, E. J, (1998). Accessing the discourse representation during reading. Discourse Processes, 26, 131-157. N ist, S. L., & Hogrebe, M. C. ( 1987). The role of underlining and annotating in remembering textual information. Reading Research and Instruction, 27, 12-25. O'Brien, E.}., Duffy, S. A., & Myers, J. L. ( 1986). Anaphoric inference during reading. Journal ofExperimental Psychology: Learning, Memory & Cognition, 12, 346-352. O'Brien, E. J., & Myers, J. L. (1987). The role of causal connections in the retrieval of text. Memory 6- Cognition, 15, 419-42 7. Pichert, ). W., & Anderson, R. C. (1977). Taking different perspectives on a story. Journal ofEducational Psychology, 69,309-315. Risden, K. A. C. ( 1996). Causal inference in narrative text comprehension. Dissertation

Abstracts International Section A: Humanities 6- Social Sciences, VoL 57(6-A). Trabasso, T ., Van den Broek, P ., & Sub, S. Y. ( 1989). Logical necessity and transitivity of causal relations in stories. Discourse Processes, 12, 1-25. Van den Broek, P. (1994). Comprehension and memory of narrative texts: Inferences and coherence. In Gernsbacher, M.A. (Ed.), Handbook of psycholinguistics (pp. 539-588). San Diego, CA: Academic Press. Van den Broek, P., Fletcher, C. R., & Risden, K. (1993). Investigations of inferential processes in reading: A theoretical and methodological integration. Discourse Processes, 16, 169-180. Van den Broek, P., & Lorch, R F. (1993). Network representations of causal relations in memory for narrative texts: Evidence from primed recognition. Discourse Processes, 16,75-98. Van den Broek, P., Lorch, E. P., & Thurlow, R. ( 1996). Children's and adults' memory for television stories: The role of causal factors, story-grammar categories, and hierarchical level Child Development, 67, 3010-3028. Van den Broek, P., Risden, K., Fletcher, C. R, & Thurlow, R (1996). A "landscape" view of reading: Fluctuating patterns of activation and the construction of a stable memory representation. In B. K. Britton, & A. C. Graesser (Eds.), Models of understanding text(pp. 165-187). Mahwah, NJ: Erlbaum.

no MicheUe L. Gaddy, Paul van den Broek and Yung-Chi Sung

Van den Broek. P., Risden, K., & Husebye-Hartmann, E. {1995). The role of readers' standards for coherence in the generation of inferences during reading. In R F. Lorch, Jr., & E.}. O'Brien (Eds.), Sources for coherence in retJding (pp. 353-374). Hillsdale, NJ: Erlbaum. Van den Broek, P., & Thurlow, R (1990). Reinstatements and elaborative inferences during the reading of narratives. Paper presented at the annual meeting of the Psychonomic Society, New Orleans, LA. Van den Broek, P., Young, M., Tzeng, Y., & Underholm, T. (1998). The landscape model of reading: Inferences and on-line construction of a memory representation. In H. van Oostendorp, & S. R. Goldman (Eds. ), The construction ofmental representations during retlding (pp. 71-98). Mahwah, NJ: Erlbaum. Vonk, W., Hustinx, L G., & Simons, W. H. (1992). The use of referential expressions in structuring discourse. Language 6- Cognitive Processes, 7, 301-333. Zwaan, R. A. (1994). Effect of genre expectations on text comprehension. journal of

Experimental Psychology: Letlming. Memory, and Cognition, 20, 920-933.

CHAPTER4

Lexical access in text production On the role of salience in metaphor resonance* Rachel Giora and Noga Balaban Tel Aviv University

1.

Introduction

What kind oflexical processes are involved in text production initially? Do they mirror the processes involved in text comprehension? On the face of it, it seems plausible to assume that, at least, in one respect they should not: Since authors know what they have in mind, and since they know their intended meaning prior to the production of the word( s) they select to convey that meaning, they would access only that intended meaning. Text production, it might be assumed, then, would not involve accessing unintended senses of a word selected for expressing a certain meaning or concept (as has been shown for text comprehension, see later). Rather, the word's intended meaning- the one compatible with the context- would be accessed directly. However, since the text producer is also her own comprehender (Levelt 1989, p. 13), text production may very well resemble text comprehension. In this study, we will test this hypothesis with regard to metaphor production in discourse. Lexical access involved in the very early stages of comprehension has been largely shown to be insensitive to contextual information (see Peleg, Giora & Fein 2001, but see Vu, Kellas & Paull998). The main thrust of lexical research into initial processes involved in comprehension and disambiguation has found ample evidence in favor of an exhaustive access model, or a variation of it, exhibiting sensitivity to meaning frequency. According to the exhaustive access model, lexical access is modular: Lexical processes are autonomous and impervious to context effects; all the word's coded meanings are accessed automatically upon its processing, regardless of contextual information or frequency (Fodor 1983; Onifer & Swinney 1981; Rayner, Pacht & Duffy 1994;

tu

Rachel Giora and Noga Balaban

Swinney 1979). Upon the ordered access version of the model, lexical access is exhaustive but frequency sensitive: The more frequent meaning is accessed first, and search for the intended meaning proceeds only in case the more frequent meaning is incompatible with contextual information (for a review, see Gorfien 1989; Rayner et al. 1994; Simpson 1994). Context, then, affects interpretations only at a later stage and suppresses incompatible meanings (Fodor 1983; Swinney 1979). The modular view of lexical access has been challenged by a direct access, interactionist hypothesis. According to the direct access model, lexical access is selective. Context directs access completely, so that only the appropriate meaning (of words) is made available for comprehension (Simpson 1981; Glucksberg, Kreuz & Rho 1986; Jones 1991; Martin, Vu, Kellas & Metcalf 1999; Tabossi 1988; Vu, Kellas & Paul1998). In the field of figurative language comprehension, the evidence adduced so far apparently supports a direct access model. The prevailing hypothesis is that in a rich and supportive context, figurative and literal interpretation should be accessed directly, without recourse to irrelevant interpretations. Particularly, the intended figurative meanings of metaphors, ironies, and idioms should be tapped directly without having to process the sentence literal meaning at all (see Gibbs 1994 for a review). Similarly, in a context biased toward the literal meaning, only that meaning should be made available for comprehension. Literal and figurative interpretations, then, should involve equivalent processes sensitive to contextual interpretation. They should be processed automatically (Keysar 1989; Gildea & Glucksberg 1983; Glucksberg, Gildea & Bookin 1982), involve the same categorization procedures (Giucksberg & Keysar 1990; Shen 1997), and take equally long to read (Kemper 1981; Inhoff, Lima & Carroll 1984; Ortony, Schallert, Reynolds & Antos 1978). The picture, however, has not been monolithic. Tapping online processes by measuring reading times at the end of figurative phrases rather than at the end of sentences showed that even when embedded in a context a few sentences long, metaphoric phrases required longer processing times than the same phrases used literally (Janus & Bever 1985 ). In addition, figurative referring expressions were found to take longer to read than their literal equivalents (Gibbs 1990). Familiar metaphors were found to be processed initially both literally and metaphorically, regardless of contextual information (Williams 1992). 1 They were further shown to retain their contextually incompatible, literal meaning in contexts biasing their interpretation toward the metaphoric

Lexical access in text production

meaning (Giora & Fein 1999b; Williams 1992). More recently, ironic utterances were found to take longer to read when embedded in ironically than in literally biasing context (Dews & Winner 1997, 1999; Giora, Fein & Schwartz 1998; Schwoebd, Dews, Winner & Srinivas, 2000), and to be interpreted only literally initially (Giora et al.1998; Giora & Fein 1999a). (For a reinterpretation of Gibbs's ( 1986) findings suggesting that ironies are interpreted literally rather than ironically first see Giora 1995). In contrast, conventional language was processed faster than less conventional language. For instance, idioms were found to take longer to read in literally than in idiomatically biasing contexts (Gibbs 1980), and faster than their variant versions (McGlone, Glucksberg & Cacciari 1994). Conventional ironies took equally long to respond to in ironically and literally biased contexts. Similarly, they were processed initially both literally and figuratively (Giora & Fein 1999a). In Pexman, Ferretti and Katz (2000), novel metaphors took longer to read than familiar metaphors, and in Turner and Katz ( 1997) familiar proverb were faster to read than less familiar ones. Which processing model then can best account for this array of inconsistent findings? Recently Giora ( 1997, in press) proposed that comprehension of figurative and literal language be viewed as governed by a general principle of salience, according to which salient meanings should always be accessed upon encounter. A meaning of a word or an expression is salient if it is coded in the mental lexicon. Salience, however, is a graded notion. Factors affecting degree of salience are conventionality, frequency, familiarity, and prototypicality. Thus, the institutional meaning of bank would be salient, that is, foremost on our mind, if we interact with commercial banks more often than with riverbanks. It would be less salient if the reverse holds. Conversational implicatures constructed on the fly, however, would be nonsalient, because they are not coded in the mental lexicon. The graded salience hypothesis thus predicts paralld access for similarly salient meanings (the figurative and literal meanings of conventional metaphors), sequential processes when salience imbalance is involved (e.g., novel metaphors whose literal but not figurative meaning is salient). Prior context may affect comprehension immediately. It may be predictive and avail the compatible meaning very early on. However, it is not sensitive to linguistic information and does not interact with lexical accessing. Consequently, it is not effective in blocking salient but contextually incompatible meanings (see also Pdeg et al. 2001). As a result, salient but contextually

U3

114


incompatible meanings that have been involved initially would be suppressed postlexically if they interfere with comprehension. They would not, if they are conducive to the compatible interpretation ('The retention hypothesis, Giora, in press; Giora & Fein 1999b). According to the graded salience hypothesis, then, when the most salient meaning is compatible (for instance, the figurative meaning of conventional idioms), it would be accessed directly (as shown by e.g. Gibbs 1980; Vander Voort & Vonk 1995) and integrate with contextual information. However, when a less rather than a more salient meaning is invited by context (e.g., the figurative meaning of novel metaphors, the literal meaning of conventional idioms, or a novel interpretation of a highly conventional literal expression), contextual information would not inhibit the salient meaning. Rather, that meaning would be accessed upon encounter regardless of context, and would be suppressed and replaced by the appropriate meaning, or retained for further processing, depending on the role it plays in comprehension (Giora in press; Giora & Fein 1999b). This holds even when context is strong and highly predictive (Gerrig 1989; Gibbs 1980; Giora et aL 1998; Peleg et al. 2001; Turner & Katz 1997). The graded salience hypothesis, thus, differs from the modular view in that it is salience sensitive and does not posit automatic suppression of contextually incompatible meanings (as assumed by Gernsbacher, Keysar & Robertson in press; Grice 1975; Swinney 1979). In sum, the direct access view pairs with the plausible assumption that authors know what's on their mind and predicts that speakers and authors would access appropriate meaning selectively. In contrast, the modular view and the graded salience hypothesis predict that lexical access in production may resist context effects. Diverging from the modular view, the graded salience hypothesis further predicts that contextually incompatible meanings would be retained if they do not obstruct the comprehension process.

2.

Lexical access in production -the case of metaphoric language

As a working hypothesis, we assume here that the processes involved in text

production mirror, at least partially, those involved in text generation. Given that speakers have access to both their internal and overt linguistic products, functioning both as producers and comprehenders of their own text (Levett 1989), the factors found to be crucial for text comprehension are assumed here to govern text production as well.


2.1

Predictions

On this assumption, the (received) direct access view regarding understanding figurative language would predict that generating metaphors should not involve activating their incompatible literal interpretation, provided the context is strong and supportive of a metaphoric interpretation. This claim has been weakened recently only with regard to novel and less familiar metaphors, suggesting that their interpretation may involve some recourse to underlying conceptual metaphors (Keysar, Shen, Glucksberg & Horton 2000, Shen & Balaban 1999 ),2 and hence to the literal meaning as well. For example, comprehension of the novel metaphor The microbe of a unity government (see Appendix) may involve accessing the conceptual metaphor POLITICS IS A DISEASE. Such view thus implies that processing novel metaphor involves activation of some aspects of its literal meaning ('disease'). According to the modular view, contextually inappropriate meanings that have been activated upon production should be suppressed after a short delay (of 300-500 msec). In processing figurative language this means that the literal meaning of nonliteral utterances which has been activated initially (about 0300 msec) should be disposed of after it has been utilized. In contrast, the graded salience hypothesis predicts that the literal meaning of familiar as well as less of unfamiliar metaphors would be activated upon text production, because its components are salient and would be accessed automatically. It further predicts that it would be retained, because it supports the metaphoric interpretation (cf. the retention hypothesis above. For a different view, see Gemsbacher et al., in press). To reject the direct access view, it is sufficient to show that either (a) familiar and less or unfamiliar metaphors involve their contextually inappropriate literal meaning more or less indistinguishably; or alternatively (b) that conventional metaphors involve their figurative and contextually incompatible literal meaning indistinguishably. While such findings contest the direct access view, they are also inconsistent with the modular view, which predicts immediate suppression of such incompatible meanings. They are, however, accountable by the retention hypothesis supplementing the graded salience hypothesis. Absence of traces of the literal meaning, however, will not contest the graded salience hypothesis (since speakers do not have to use available information), though they will be more consistent with the direct access and modular views.

us

116


Given that online measures of text production are hardly available, one way to test these predictions is to study the ecology of metaphors in naturally occurring discourses. Mention of a meaning of a metaphor in the metaphor's immediate neighborhood, i.e., in the metaphor's clause or in the next two or three clauses, may count as evidence that that meaning was active in the producer's mind and was, therefore, made manifest. Rather than discarded as incompatible, that meaning was retained for further processes. Note that the various approaches in question (with the exception of the modular view) differ only insofar as unintended (i.e., literal) meanings of familiar metaphors are concerned. Thus, if familiar and less familiar metaphors both prime and retain their unintended literal meaning indistinguishably, this would support the graded salience hypothesis, but challenge both the modular and direct access views, the former on the basis of the suppression hypothesis and the latter on its selectiveness hypothesis. Given that in a metaphor-inviting context, familiar metaphors are expected to be processed only metaphorically, findings indicating that familiar metaphors were processed literally would be problematic for the direct access view: They will attest that these meanings have been accessed regardless of context. However, if only less familiar metaphors activate and retain their unintended literal meanings, this would be consistent with the weaker version of the direct access view proposed by Keysar et al. (2000) and only partly consistent with the graded salience hypothesis and modular view. To test these hypotheses we collected naturally occurring metaphors that were either elaborated on following their mention, or were not. We then looked into (a) whether the set of metaphors elaborated on differed in terms of familiarity/novelty from the set which received no elaboration. We further looked into (b) whether metaphors judged as highly familiar exhibit no elaboration, as would be predicted by the various versions of direct access view. Findings to the contrary would favor the graded salience hypothesis. 2..2.

Method

Materials Our materials were metaphors appearing in newspaper articles. We randomly collected 60 metaphors from the columns' section of Ha'aretz- an Israeli daily- during the months of August and September 1997. Thirty involved some mention or echo of their (unintended) literal interpretation, i.e., a word or an expression semantically related to their literal meaning, in the same or next dause(s) (e.g., la, lb), and 30 did not (e.g., 2a, 2b, see also Appendix).


(1) a.

The strikes in the Education system took place when the Union was putting up a fight against the government. In this fight, threats, sanctions and even a general strike were the weapons. (Ha'aretz, 4.9.97: B1) b. In this situation, the Treasury looks like an island of sanity in a sea of unconstrained demands. (Ha'aretz, 12.9.97: Bl)

(2) a. He lost his health, and his spirit broke. (Ha'aretz, 1.9.97: B1) b. Every honest and benevolent person should have given a shoulder to the minister of Treasury so that he can succeed in implementing his plan. (Ha'aretz, 4.9.97: Bl)

Participants Forty native speakers of Hebrew {30 females, 10 males) participated in the experiment. They were all undergraduates in the department of Poetics and Comparative Literature, Tel Aviv University, aged 21-40. They participated in the experiment voluntarily.

Procedure The participants were presented with the 60 metaphors and another five, contrived, novel metaphors. They were asked to rate them on a 1-7 familiarity scale: from the least unfamiliar ( 1) to the most (7) familiar metaphor. 2..3 Results The mean familiarity rate of each metaphor (ranging from 2-6.95) was the basic datum for the analysis. Findings showed that, as predicted by the graded salience hypothesis, but contra the direct access approach (see (a) above), metaphors followed by a mention of their literal meaning did not differ familiarity-wise from those that were not, the difference between the means was insignificant (t=0.96, p=0.34, two tail). That is, the metaphors whose literal meaning was retained - echoed and elaborated on - in the immediate or next clause{s) (e.g., (la-b) above) were not evaluated as more or less familiar than those that received no literal extension (e.g., (2a-b) above). Moreover, a check of the number of metaphors which received the highest familiarity rates {6-7) reveals that 15 of them belonged in the group of (30) metaphors which had literal extensions (e.g., (Ia -b) above) and 17 belonged in the other group of (30) metaphors whose literal meaning was not elaborated on (e.g., (2a-b) above). Thus, as predicted by (b) above, and in accordance with

U7

118


the graded salient hypothesis, but contra the direct access and standard pragmatic models, highly familiar metaphors retained both their compatible and contextually incompatible (literal) meaning indistinguishably, suggesting that even highly conventional metaphors involve processing their salient though contextually incompatible meaning.

3·

Discussion and conclusions

Since we did not use an on-line measure, our findings cannot directly support an autonomous view of lexical access in production. However, they are certainly consistent with it, suggesting that context does not block activation of contextually incompatible but salient (i.e., coded) meanings. Even highly familiar metaphors, whose metaphoric meaning may be processed directly, avail their salient, literal meaning upon their production alongside the metaphoric meaning, though this meaning may be incompatible with contextual information. These findings are consistent with the graded salience hypothesis, which is (only) partially congruent with a modular view of lexical access in production. 3 They suggest that the processes involved in text production are similar to those involved in text comprehension. Activation of salient meanings is automatic, and does not interact with the context at an early stage: Context does not pre-select only the appropriate meaning; that is, it does not inhibit activation of salient incompatible meanings (cf. section 1, and see also Honeck 1997, p. 467), though it can of course predict it. Illustrative is the following example (cited in Honeck 1997, p. 47), where the salient meaning of leaf(a part of a plant) is activated in the idiom 'turn over a new leaf even though this is not the meaning of leaf (page) on which the idiom is based. This 'error', we propose, attests to the involvement of salient (though incompatible) meanings in text production, as opposed to less salient (though more compatible) meanings (such as the 'page' meaning of leaf): (3)

Fred is cutting his lawn earlier now. I guess he's turning over a new leaf.

Our findings corroborate findings by Nayak and Gibbs ( 1990) quoted in Gibbs ( 1994, p. 301 ), which show that readers are sensitive to the salient, contextually incompatible meanings of figurative language when they are asked to rate the appropriateness of an incoming text-segment (and act as text producers, to a certain extent). For example, when asked to rate the appropriateness of idioms

Lexical access in text production 119

for a particular paragraph ending, subjects tended to select as more appropriate the idiom stemming from the source domain enhanced by the paragraph. Thus, out of (4a) or (4b) in the context of (4), they selected (4a) whose (incompatible) literal meaning 'coheres' or 'resonates' with the salient literal meanings of the chain of metaphors instantiated in (4): (4) Mary was very tense about this evening's dinner party. The fact that Bob had not come home to help was making her fume. She was getting hotter with every passing minute. Dinner would not be ready before the guests arrived. As it got closer to five o'clock the pressure was really building up. Mary's tolerance was reaching its limits. When Bob strolled in at ten minutes to five whistling and smiling, Mary (a) blew her top (b) bit his head off These findings show that text progression is not only sensitive to the figurative meanings accessed by the text's comprehender/producer (which should have resulted in no appropriateness difference between the two possible idiomatic continuations), but that it is even more sensitive to salient though incompatible meanings. Though this experiment is only partially relevant to our research, involving comprehension and text appreciation rather than text production, its findings are suggestive of the same processes alluded to here. Our findings can also be viewed as an instantiation of a more general phenomenon of"dialogic syntax" (DuBois 1998, 2000a,b and see also Coates 1966; Levelt 1989). Dialogic syntax occurs when a speaker constructs an utterance based on an immediately co-present utterance. Du Bois discloses the ubiquity of"dialogic syntax", showing that a vast array of linguistic elements such as syntactic, semantic, pragmatic, lexical, and even phonetic patterns in one speaker's discourse can be traced back to an immediately co-present utterance. This suggests that activation of any linguistic element makes it available for the same or next speaker to elaborate on. Activation of that element enhances it, which, in turn, suppresses other possible alternatives. Enhancement makes it accessible and, hence, a preferable candidate to be selected and elaborated on in the next utterance (Gernsbacher 1990). Our findings show that metaphors, not least familiar metaphors, are processed (also) literally in the mind of the discourse producer, thereby allowing reoccurrence of the salient/literal meaning in the next discourse segment. Evidence of similar effects of a given utterance on adjacent ones ( cf. Du Bois 1998,

120


2000b) suggests that salience of meanings (and constructions), which is subject to local manipulation (enhancement/suppression) in discourse through dialogic resonance, is a major factor in discourse production.

Appendix Translated sample items, all taken form Ha'aretz, August-September 1997, Bl: I. Metaphors with an extension of their literal meaning (in the Hebrew original):

(I) Of all the viruses and microbes which call upon our country, the worst is the microbe of a unity government. This plague, which usually breaks out abruptly... (2) This is a story that has begun but has not yet ended. One important scene ended when Yitzhak Rabin was assassinated [ ... 1. The second scene ended when Peres was Prime Minister ... (3) The Palestinian affair is a time bomb that has to be dismantledl(disassembled) before it explodes. (4) Israel needs ... not only those who flirt with the capital market, but those who marry it, for better or worse. in poverty and in wealth, until "a purchase proposal" do they part. (5) Politically the present Croatian leaders' wishing to blur the impression of the "Ostasha" rehabilitation is understandable. But they will find it difficult to erase the moral stain of their attempt to rehabilitate the murderers and their accomplices. (6) (Addressed to Albright): If you sink(/delve) into [meaning preoccupy yourself with I Dahania, or into the port, or into the security arrangements and into the days on which a bulldozer will pass or will not pass, you will drown in a St!ll of details. (7) All this happened when the civil servants were drowsy while on duty. If they were alert/ (wide awake), the deterioration [ ... 1might have been prevented. II. Metaphors with no extension of their literal meaning: (I) In her position as the mother of the future king, [DianaI was stuck as a bone in the throat of the British monarchy. And form this position and being so bright, she opened a window into the inhumanity of the royal family. (2) Ninety percent of the property in Israel bas now been turned into state property [ ... 1 regardless of whether the laws passed the elementary test of justice and equality. (3) They don't talk of small fish, they talk of terrorists with blood on their hands. (4) Soon the patients rolled up their sleeves and their lawyers entered the battlefield.

Notes • This research was supported by a grant from The Israel Science Foundation. We have also benefited from discussions with Mira Ariel, Jack Du Bois and Ray Gibbs. Thanks are also extended to Sharon Himmelfarb and two anonymous reviewers for their comments.


Williams (1992) actually studies polysemous words but most of them were metaphorically based.

1.

Inconsistently with the direct access view, the consensus is that comprehension of novel metaphors differs from comprehension of familiar metaphors and should involve activating their contextually incompatible salient meaning. i.e., their literal meaning (e.g., Keysar et al. 2000, and see Gibbs 1994 for a review).

2..

3· The modular view of lexical access in speech production distinguishes between meaningrelated and form-related processes (see, e.g., Levett 1993; Dell & O'Seaghdha 1992). However, this view is only marginally relevant to our discussion here.

References Blasko, G. D., & Connine, C. ( 1993 ). Effects of familiarity and aptness on metaphor processing. Journal of Experimental Psychology: Learning, Memory, and Cognition. 19,295-308. Coates, J. (1996). Women talk. Oxford: Blackwell. Dell, S. G., & O'Seaghdha, P. G. ( 1992). Stages oflexical access in language production. Cognition, 42,287-314. Dews, S., & Winner, E. ( 1997). Attributing meaning to deliberately false utterances: The case of irony. In C. Mandell, & A. McCabe (Eds.), The problem of meaning: Behavioral and cognitive perspectives (pp. 377-414). Amsterdam: Elsevier. Dews, S., & Winner, E. (1999). Obligatory processing of the literal and nonliteral meanings of ironic utterances. Journal of Pragmatics, 31, 1579-1599. Du Bois, W. J. (1998). Dialogic syntax. Paper presented at The cognitive theories of intertextuality conference. Tel Aviv university. DuBois, W. J. (2000a). Santa Barbara corpus of spoken American English. CD-ROM. Philadelphia: Linguistic Data Consortium.[ www.ldc.upenn.edu/Publications/ SBC/]. DuBois, W. J. (2000b). Reusable syntax: Socially distributed cognition in dialogic interaction. Paper presented at CSDL 2000: Fifth Conference on Conceptual structure, discourse, and language. University of California, Santa Barbara. Fodor, J. (1983). The modularity ofmind. Cambridge: MIT Press. Gernsbacher, M.A. (1990). Language comprehension as structure building. Hillsdale, N.J.: Erlbaum. Gemsbacher, M.A., Keysar, B., & Robertson, R. W. (in press). The role of suppression and enhancement in understanding metaphors. Journal ofMemory and Language. Gerrig. R. J. (1989). The time course of sense creation. Memory & Cognition, 17, 194207. Gibbs, R. W. Jr. ( 1980). Spilling the bean on understanding and memory for idioms in conversation. Memory 6- Cognition, 8, 449-456. Gibbs, R. W. Jr. ( 1986). On the psycholinguistics of sarcasm. Journal of Experimental Psychology: General, 115, 3-15.

Ul

12.2


Gibbs, R. W. Jr. (1990}. Comprehending figurative referential descriptions. Journt~l of Experimental Psychology: Learning. Memory and Cognition, 16, 56--66. Gibbs, R. W. Jr. ( 1994}. The poetics of mind. Cambridge: Cambridge University Press. Gildea, P., & Glucksberg, S. (1983}. On understanding metaphor: The role of context. Journal of Verbal Learning and Verbal Behavior, 22, 577-590. Giora, R. ( 1995}. On irony and negation. Discourse Proasses, 19, 239-264. Giora, R. (1997}. Understanding figurative and literal language: The graded salience hypothesis. Cognitive Linguistics. 7, 183-206. Giora, R. (1999}. On the priority of salient meanings: Studies of literal and private figurative language. Journal of Pragmatics, 31, 919-929. Giora, R. (in press}. Ott our mind: Salience. context and figurative language. New York: Oxford University Press. Giora, R. & Fein, 0. ( 1999a). Irony: Context and salience. Metaphor and Symbo~ 14, 241-257. Giora, R. & Fein, 0. (1999b). On understanding familiar and less-familiar figurative language. Journal ofPragmatics, 31, 1601-1618. Giora, R. & Fein, 0. ( 1999c). Irony comprehension: The graded salience hypothesis. Humor, 12,425-436. Giora, R., Fein 0., & Schwartz, T. ( 1998).1rony: Graded salience and indirect negation. Metaphor and Symbo~ 13, 83-101. Glucksberg, S., Gildea, P., & Bookin, H. G. ( 1982). On understanding nonliteral speech: Can people ignore metaphors? Journal of Verbal Learning and Verbal Behavior, 21, 85-98. Glucksberg, S., & Keysar, B. ( 1990). Understanding Metaphorical Comparisons: Beyond Similarity. Psychological Review, 97, 3-18. Glucksberg, S., Kreuz, R., & Rho, S. H. (1986). Context can constrain lexical access: Implications for models of language comprehension. Journal ofExperimental Psychology: Learning. Memory, and Cognition, 12, 323-335. Gorfein, S. D. (Ed.). ( 1989). Resolving semantic ambiguity. New York: Springer Verlag. Grice, P. H. (1975). Logic and Conversation. In P. Cole, & }.Morgan (Eds.), Speech acts. Syntax and semantics. 3 (pp. 41-58). New York: Academic Press. Hogaboam, T. W., & Perfetti, C. A. ( 1975 ). Lexical ambiguity and sentence comprehension. Journal ofVerbal Learning and Verbal Behavior, 14,265-274. Honeck, R. P. (1997). A proverb in mind: The cognitive science of proverbial wit and wisdom. Mahwah, NJ: Erlbaum. lnhoff, A. W., Lima, S.D., & Carroll, P. J. (1984). Contextual effects on metaphor comprehension in reading. Memory & Cognition. 12, 558-567. Janus, R. A., & Bever, T. G. (1985). Processing of metaphoric language: An investigation of the three stage model of metaphor comprehension. Journal of Psycholinguistic Research, 14, 473-487. Jones, f. L. ( 1991). Early integration of context during lexical access of homonym meanings. Current Psychology: Research 6- Reviews, 10, 163-181. Kemper, S. ( 1981 ). Comprehension and interpretation of proverbs. Journal of Psycholinguistic Research, 10, 179-183.

Lexical access in text production 123

Keysar, B. ( 1989}. On the functional equivalence of literal and metaphorical interpretations in discourse. Jourmd ofMemory and LIJnguage, 28, 375-385. Keysar, B., Shen, Y., Glucksberg. S., & Horton, S. W. (2000). Conventional language: How metaphorical is it? Journal ofMemory and Language, 43, 57~593. Levett, W. J. M. ( 1989}. Speaking: From intention to articulation. Cambridge Mass.: MIT Press. Levett, W. J. M. (Ed.). (1993). Lexical access in speech production. Cambridge and Oxford: Blackwell. Martin, C., Vu, H., Kellas, G., & Metcalf, K. ( 1999}. Strength of discourse context as a determinant of the subordinate bias effect. The Quarterly Journal of Experimental psychology, 52, (2). McGlone, M.S., Glucksberg, S., & Cacciari, C. (1994). Semantic productivity and idiom comprehension. Discourse Processes, 17,167-190. Nayak, N., & Gibbs, R. W. Jr. (1990). Conceptual knowledge in idiom interpretation. Journal ofexperimental Psychology: GeneraL 116, 315-330. 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. Ortony, A., Schallert, D. L., Reynolds, R. E., & Antos, S. J. (1978). Interpreting metaphors and idioms: Some effects of context on comprehension. Journal of Verbtll Learning and Verbal Behavior, 17,465-477. Peleg, 0., Giora, R., & Fein, 0. (200 1}. Can context selectively activate the contextually appropriate meaning of an ambiguous word? Metaphor and SymboL Pexman P., Ferretti, T., & Katz, A. (2000). Discourse factors that influence irony detection during on-line reading. Discourse Processes, 29,201-222. Rayner, K., Pacht ). M., & Duffy, S. A. ( 1994). Effects of prior encounter and global discourse bias on the processing oflexically ambiguous words: Evidence from eye fixations. Journal ofMemory and Language, 33, 527-544. Rttanati, F. ( 1995). The alleged priority of literal meaning. Cognitive Science, 19, 207232.

Schwoebel, }., Dews, S., Winner, E., & Srinivas, K. (2000). Obligatory Processing of the Literal Meaning of Ironic Utterances: Further Evidence. Metaphor and SymboL 15, 47~1.

Shen, Y. (1997). Metaphors and global conceptual structures. Poetics, 25,1-17. Shen. Y.• & Balaban, N. (1999). Metaphorical (in)coherence in discourse. Discourse Processes, 28, 139-154. Simpson, G. B. ( 1981). Meaning dominance and semantic context in the processing of lexical ambiguity. Journal ofVerbal Learning and Verbal Behavior, 20, 12~136. Simpson, G. B. (1994). Context and the processing of ambiguous words. In M A. Gernsbacher (Ed.). Handbook of Psycholinguistics (pp. 359-374). San Diego: Academic Press. 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.

124 Rachel Giora and Noga Balaban

Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. journal of Verbal Learning and Verbal Behavior, 18, 645659. Tabossi, P. ( 1988 ). Accessing lexical ambiguity in different types of sentential contexts. journal ofMemory and Language, 27, 324-340. Turner, N. E., & Katz, A. (1997). Evidence for the availability of conventional and of

literal meaning during the comprehension of proverbs. Pragmatics and Cognition, 5,203-237.

VandeVoort, M. E. C., &Vonk, W. (1995). You don't die immediately when you kick an empty bucket: A processing view on semantic and syntactic characteristics of idioms. In M. Everaert, E-J van der Linden, A. Schenk & R. Schreuder (Eds.), Idioms: Structural and psychological perspectives (pp. 283-299). Hillsdale NJ: LEA Vu, H., Kellas, G., & Paul, S. T. {1998). Sources of sentence constraint in lexical ambiguity resolution. Memory & Cognition, 26,979-1001. Williams, J. N. ( 1992). Processing polysemous words in context. Evidence from interrelated meanings. Journal ofPsycholinguistic Research, 21,193-218.

SECTION

2

Relational coherence in text and text . processmg

In this second section we switch from referential coherence and the activation of concepts to relational coherence. Over the last decade, coherence relations, or rhetorical relations, like CAUSE-CONSEQUENCE, CoNTRAST and PROBLEM-SoLlTTION have been studied intensively. One key issue has turned out to be the categorization of coherence relations and of their prototypical expressions, connectives: CAUSE-CONSEQUENCE relations are more similar to CLAIM-ARGUMENT relations than to LisT relations, just like the causal connective because has more in common with since than with and. The Chapters 5-9 can be read as contributions to a discussion about these categorization issues. What all these text-linguistic chapters have in common is that none of them deals with connectives and markers as linguistic items per se, but that the analyses are taken to be relevant to mental representations of these linguistic elements. In Chapter 5, Knott argues that the distinction often made between socalled pragmatic and semantic or content, epistemic and speech act coherence relations is not always clearcut. As an alternative, he proposes an intentionbased definition of pragmatic relations. In Chapter 6, Noordman deals with the same distinction, now with regard to causal (because) and concessive (although) relations. He uses corpus research to analyze, first, how semantic and epistemic concession relations are realized linguistically, and, second, how such realizations affect the surrounding discourse. Noordman provides evidence from reading time studies that suggests that semantic concessions are easier to process than epistemic relations. Moreover, he shows how various ordering variants (e.g. although p, q versus q, although p) can be accounted for by the thematic development within the discourse surrounding these variants, thus establishing a clear link between the mental representation of discourse and its surface characteristics.

tUi

Section 2: Relational coherence in text and text processing

Research on relational coherence is strongly influenced by Mann and Thompson's Rhetorical Structure Theory (RST; Mann & Thompson 1988). One of the relations in RST, the Elaboration relation, is debated in Chapter 7, by Knott. Oberlander, O'Donnell and Mellish because it does not fit in with other discourse relations and is often used as a waste-paper basket, to be used in text analysis when no other relation fits. As an alternative, the authors propose to account for coherence either by coherence relations (for local dependencies) or by entity-based moves (for non-local dependencies). The two remaining chapters in this Section introduce other frameworks to extend the theoretical basis of the analysis of rdational coherence. In Chapter 8 Pander Maat uses Sperber & Wilson's Relevance Theory (Sperber & Wilson 1986) to analyze the meaning of the conjunction and. He makes a number of controversial claims on the proper analysis of and, suggesting among others that this connective does not link two segments directly, but makes them jointly rdevant to the surrounding context. To support his analysis Pander Maat makes ample use of corpus analysis. In Chapter 9 Snoeck Henkemans continues the discussion about different categories of coherence relations and connectives from the viewpoint of Argumentation Theory. She explores the rdationship between, on the one hand, text-linguistic accounts of connectives like those present in Chapters 5 through 8, with the tradition of Argumentation Theory on the other hand. More specifically, she discusses the starting-points of the pragma-dialectical account of argumentation, and the way in which the speech acts of argumentation and explanation are characterized. Then, she compares this account with the textlinguistically inspired approach. One of the conclusions is that it is sometimes unclear how propositional and illocutionary levels can be separated as different hierarchical levels in the analysis of explanation and argumentation. Snoeck Henkemans argues that the pragma-dialectical approach creates a good basis for interpreting linguistic cues like connectives in a well-founded and systematic way.

References Sperber, D. & Wilson, D. ( 1992). Relevance: Communication and cognitiotL (2nd ed.). Oxford: Blackwell. Mann, W. C. & Thompson, S. A. (1988). Rhetorical Structure Theory: A functional theory of text organization. Text, 8, 243-281.

CHAPTER

5

Semantic and pragmatic relations and their intended effects Alistair Knott University of Otago

1.

Introduction

This paper takes as its starting point an assumption that has received much attention in both computational and psycholinguistic treatments of discourse: that the coherence of an extended text can be explained in terms of an account of the relations which hold between its component spans. Naturally, when it is stated as baldly as this, the assumption is almost vacuously true. However, it is useful in providing the beginnings of a vocabulary for discussing the issue of text coherence. It means that the empirical question of what makes a text coherent can be re-expressed as two more specific questions. Firstly, how can we define the set of relations which are permitted within coherent text? And secondly, what are the structural constraints on the configuration of these relations within coherent text? In the present paper, I will concentrate on the first of these questions. Accounts of the semantics of coherence relations have often suggested that they should be thought of as composite entities, defined in terms of a number of different dimensions. This assumption is implicit in systemic approaches to conjunctive relations (Halliday & Hasan 1976; Martin 1992), in psycholinguistic/text-linguistic approaches (Van Dijk 1979; Redeker 1990; Sanders, Spooren & Noordman 1992), and in computationaUsemantic approaches (Hobbs 1985; Elhadad & McKeown 1990; Knott & Mellish 1996; Oversteegen 1997). In all of these approaches, specifying the semantics of a coherence relation involves specifying a number of different values, which position it within a multidimensional space. Furthermore, if we assume that intersententiaUinterclausal conjunctions can be used to signal coherence relations in surface text, then the semantics of these conjunctions can be specified in a similar fashion. There is a

12.8 Alistair Knott

large measure of variation in the above studies as to the closeness of the mapping between the set of coherence relations and the set of connectives. Likewise, the studies differ as regards the number of dimensions they propose and the degree of interdependence they envisage between relations. However, the principle is the same in each case. One of the fundamental dimensions to have been proposed distinguishes between relations that hold between the content of the text spans they link and those that hold between the utterances of the text spans themselves, or the beliefs which underlie them. This distinction is most clearly illustrated with reference to two kinds of 'causal' relations: those that simply describe a cause and effect occurring in the world (such as example (I) below), and those that have an argumentative or rhetorical force (such as example (2) ). ( 1} Bill was starving, so he had a sandwich. (2} Bill had five sandwiches, so he was/must have been starving.

In each case, so is taken to link a cause and effect. But while in example (1) the cause and effect are taken to be the eventualities described by the respective clauses, in (2), the speaker's beliefin the eventuality described in the first clause is taken to cause her conclusion about the eventuality described in the second. Nearly all the theories that advocate a decomposition of relations include a dimension which reflects this difference. The main reason for this is the apparent productivity of the distinction, in particular in capturing alternative senses of sentence and clause connectives. There are many connectives, of several different types, which can be analyzed as ambiguous with respect to this dimension. As well as causal/inferential connectives like so and because, there are conditional connectives such as if, contrastive connectives such as but, disjunctive connectives such as or and otherwise, and temporal connectives such as then. In each case, a similar ambiguity seems to be discernible; the challenge is to find a formulation of the distinction which covers all the cases. Several different terms have been used for the two types of relation, and slightly different definitions have been given. Halliday and Hasan (1976) and Martin ( 1992) refer to EXTERNAL and INTERNAL relations; Van Dijk ( 1979) refers to SEMANTIC and PRAGMATIC relations; and Redeker ( 1990) refers to IDEATIONAL and PRAGMATIC relations. I will adopt van Dijk's formulation in the remainder of this paper. In this paper, I begin in Section 2 by outlining the two types of relation in more detail. In Section 3, I then consider some problems with this account

Semantic and pragmatic relations U9

noted by Sweetser ( 1990) 1 which led to her proposing a new tripartite distinction in place of the bipartite one. In the new classification, the class of PRAGMATIC relations is divided in two: ( 1) the class of EPISTEMIC relations, which hold at the level of premises and conclusions about what is the case in the world and (2) the class of SPEECH-ACT relations, which hold between the utterances themselves. I also present Sweetser's extension of the set of relations from causal to conditional, disjunctive and sequential relations. In Section 4, I consider the advantages and disadvantages of the proposals previously outlined, and also note some problems which apply equally to each proposal. In Section 5, I then propose an alternative, intention based definition of PRAGMATIC relations which overcomes a number of these problems. An assessment of the new definition, noting advantages relating to the scope of its applicability and to its explanatory force, as well as cases which it cannot handle, is given in Section 6. I conclude in Section 7 by considering some possible solutions to these problems, by introducing a new bipartite distinction, orthogonal to the SEMANTIC/PRAGMATIC distinction, between CAUSE-DRIVEN and RESULT-DRIVEN relations.

1.

The bipartite semantic/pragmatic distinction

Consider again the following sentences: ( 1) (2)

Bill was starving, so he had a sandwich. Bill had five sandwiches, so he was/must have been starving.

The distinction between SEMANTIC relations such as example ( 1) and PRAGMATIC relations such as example (2) has been captured by many researchers by interpreting the latter type of relation as containing an implicit performative. In the example above, the performative could be made explicit as follows: (3)

Bill had five sandwiches, so I conclude he was starving.

Generalizing from this example, we can state that SEMANTIC relations hold directly between the propositional content of the two related utterances, while PRAGMATIC relations hold between the utterances themselves, interpreted as speech acts. The dearest expression of this idea comes from Van Dijk ( 1979, p. 449): "Pragmatic connectives express relations between speech acts, whereas semantic connectives express relations between denoted facts".

130

Alistair Knott

Many other theorists have advocated a similar distinction - although typically qualified in some way - including Halliday and Hasan (1976), Martin (1992), Redeker (1990), and Sanders et al. (1992). (Different qualifications will be mentioned in Sections 3 and 4.2.) Interpreting PRAGMATIC relations as holding between speech acts allows us to give a connective like so a single denotation, which applies equally to PRAGMATIC contexts and to SEMANTIC ones. Rather than assuming a lexical ambiguity in so (as signaling either a cause in the world, or an argumentative relation), we can assume that the relation signaled is simple causality in both cases, with the difference between the cases stemming from a systematic variation in which propositions are being linked by this relation. The effect of this analysis is essentially to reduce relations at the PRAGMATIC level to relations at the SEMANTIC level. SEMANTIC relations are ones in which information about a relation between two propositions in the world is conveyed from the speaker to the hearer. PRAGMATIC relations, according to the definition just proposed, are exactly the same; it is just that the information which the speaker is communicating to the hearer, is (partly) about her own speech acts. A point to note about the definition of PRAGMATIC relations is that it extends in a useful way to relations involving sentences which are not in the indicative mood. Consider the following two sentences, for instance: (4) Bill's starving. So why isn't he eating? (5) Bill's starving, so give him something to eat. The point is that some transformation on the latter sentence in each example is necessary to render it directly comparable to sentences in the indicative mood. Moreover, we want the transformation to reflect the real differences between the different kinds of sentences. Moving to the level of the speech act which underlies the sentence is a way of doing this. The speech act behind a sentence can always be expressed as an indicative sentence, which gives us the possibility of making direct comparisons. Thus we can gloss (4) and (5) respectively as follows: (6) Bill's starving. So I ask 'why isn't he eating'? (7) Bill's starving, so I instruct you to give him something to eat. In summary, the distinction between SEMANTIC and PRAGMATIC relations can be seen as bearing on two separate problems: the problem of identifying the ambiguity of certain connectives, and the problem of finding a uniform repre-

Semantic and pragmatic relations

sentation for sentences in different moods. It is fair to say that it is the former problem which originally motivated the distinction, and that it is an advantage of the distinction that it goes some way towards addressing the latter problem. However, as we will see, the distinction does not enable a complete solution to the latter problem. In fact, this should already be obvious from the fact that PRAGMATIC relations are to be analyzed simply as providing information about the world; a complete treatment of imperatives and interrogatives is bound to require something more than this.

3·

The tripartite CONTENT/EPISTEMIC/SPEECB-ACT distinction

Sweetser ( 1990) notes an important underspecification in a gloss such as given in example (3). Is the implicit act of 'concluding' to be thought of as a speech act? There are good reasons for maintaining that this is inappropriate. For Sweetser, (t)here is a class of causal-conjunction uses in which the causality is that between premise and conclusion in the speaker's mind( ... ) and there is another class of uses in which the causality actually involves the speech act itself.

Sweetser draws a sharp distinction between the act of drawing a conclusion and the act of stating it. A speaker is only obliged to state a conclusion she has reached "inasmuch as the rules of conversation make it incumbent on us to say things we believe to be true"; it is therefore inappropriate to gloss an argumentative relation such as that in example (2) in terms of speech acts. Essentially, what the speaker decides to say and what the speaker decides to believe are two quite different things, and it is not possible to accurately express one in terms of the other. Note that we can still analyze (2) as containing an implicit performative, but the performative verb should be clearly understood as describing a theorem-proving act on the part of the speaker rather than a linguistic one, thereby nailing an important ambiguity in a word like conclude, which can be understood in both senses. 2 Sweetser proposes a division of the class of PRAGMATIC relations into two categories: SPEECH-ACTrelations, which hold between the utterances themselves, and EPISTEMIC relations, which hold at the level of premises and conclusions about what is the case in the world. Her overall classification is tripartite; there is a third category of relations, termed CONTENT relations, which effectively cover the same ground as those termed SEMANTIC in the bipartite classification.

131

132.

Alistair Knott

Note that the distinction between EPISTEMIC and SPEECH-ACT relations is to some extent independent of the question of whether a relation contains any non-indicative sentences. There can be SPEECH-ACT relations between two indicative sentences as well as EPISTEMIC ones. Example 8 is a case in point: (8) The answer's on page 200, since you'll never find it for yourself.

Here we would certainly not want to suggest that the speaker is reaching a conclusion about where the answer is, on the basis of knowing that the hearer will not find it for himself. Glossing with an implicit speech-act performative ('I tell you that the answer's on page 200') is much more appropriate. However, there is still some degree of dependence on mood in Sweetser's definitions. As she says (1990, p. 78): "If an utterance is imperative or interrogative in form, then it cannot reasonably be causally conjoined to another utterance except at the speech-act level". In practice, it seems likely that many advocates of the bipartite position are aware of the variations which Sweetser points out within the class of PRAGMATIC relations. Definitions of PRAGMATIC relations tend in fact to be hedge the issue to some extent, frequently containing disjunctive elements or appealing to somewhat ill-defined higher-level constructs. For instance, Sanders ( 1997, p. 126) defines PRAGMATIC relations as "applying between the content of one span and the speaker's 'claim/advice/conclusion' about the content of the other", while for Redeker ( 1990, p. 369) "they hold between the 'beliefs and intentions' which underlie the two spans". Others appeal to rather nebulous higher-level constructs; for instance, for Halliday and Hasan ( 1976, p. 240) the relationship in question is not so much a relationship between speech acts (though it may take this form) [ ... ] as a relationship between different stages in the unfolding of a speaker's COMMUNICAnoN ROLE [ ••• ] his choice of speech role and rhetorical channel, his attitudes, his judgments and the like.

dearly, theorists are looking for a way of generalizing across the various different types of PRAGMATIC relations. Sweetser, by dividing the class in two, has to some extent abandoned this goal, but by the same token her definitions are much more intelligible. 3.1

SPEECH-ACf and EPISTEMIC

disjunctions

As mentioned in Section 1, one of the attractions of the SEMANTidPRAGMATIC


distinction is its productivity: it seems to find application not only for causal relations, but also for other types of relations. A strong recommendation for Sweetser's finer-grained distinction is that it too seems productive across this range of relations. Some of the interesting distinctions between EPISTEMIC and SPEECH-ACI' relations are outlined in the following three sections; I will omit discussion of the corresponding CONTENT relations, which should in each case be quite dear. Examples of EPISTEMIC and SPEECH-ACI' disjunctive relations are given in (9) and (10) respectively. (9) John is home, or somebody is picking up his newspapers. (10) Would you like to come round tonight? Or is your car still in the shop?

Sweetser analyses example (9) as conveying that the alternative propositions presented are the only two possible conclusions that one can reach. On the other hand, she analyses ( 10) as containing a pair of alternative speech acts: the speaker is asking to be understood by the hearer as performing either one or the other. The distinction between the two analyses seems useful here; it is inappropriate to gloss example (9) in terms of speech acts, for the same reason as it is inappropriate to gloss example (2) in such terms, as discussed at the beginning of Section 3. 3.2.

SPEECH-ACf and EPISTEMJC conditionals

Examples of EPISTEMIC and SPEECH-Acr conditional relations are given in examples (11) and (12) respectively. ( 11) If}ohn went to that party, he was trying to infuriate Miriam. ( 12) How old are you, ifit is not a cheeky question?

Example ( 11) is to be analyzed as expressing an implication relation between the speaker's beliefs: the speaker is really informing the hearer that 'if I [the speaker] believe that John went to the party, I believe that he was trying to infuriate Miriam'. On the other hand, example ( 12) is to be understood as the conditional performance of a speech act: the speaker only wants to ask about the hearer's age if it is not a cheeky question. Again, for the same reasons, it is not appropriate to gloss ( 11 ) in terms of speech acts, and thus there seems to be good reason for giving different analyses for the two cases.

133

134

Alistair Knott

3·3

SPEECH-ACT and EPISTEMIC temporal sequences

Finally, examples of EPISTEMIC and SPEECH-ACT temporal relations are given in examples (13) and (14) respectively. ( 13) A: Why don't you want me to take basket weaving this summer? B: Well, Mary took basket weaving, and she joined a religious cult. (14)

Go to bed now! And no more backtalk!

In each of these examples, according to Sweetser, and is to be interpreted sequentially. However, while the temporal sequence in (14) relates to the order in which the speech acts are performed, the sequence in example ( 13) relates to the order of events in the epistemic world. The idea in the latter case is that the two propositions are both to be interpreted as premises in an argument that A should not take basket weaving, but that their ordering is significant. While this latter analysis is not completely clear to me, the fact that explicitly sequential conjunctions like to begin with and next can be used to link multiple premises in an argument has often been noted, and provides all equally good rationale for temporal EPISTEMIC relations. For instance, to echo examples given by Halliday and Hasan (1976): (15)

John's unsuitable for the job. To begin with, he is too young. Next, he is too hotheaded. Finally...

Sweetser's objections against speech-act glosses of argumentative relations are as telling in this case as in the others; an epistemic interpretation of the sequential relation is again preferable.

4-

Some problems with the existing distinctions

While the bipartite and tripartite classifications are both extremely useful in distinguishing between different types of coherence relation and uses of connective phrases, there remain some outstanding problems in each case. In this section, I will review some of these.

4-1

Generalizations across SPEECH-ACT and EPISTEMIC relations

One problem with the tripartite account is that EPISTEMIC and SPEECH-ACT


relations do seem to share a lot in common. For instance, as Sweetser points out, many connectives (such as French puisque) are appropriate for both EPISTEMIC and SPEECH-ACT relations, but not for CONTENT relations. It is useful to be able to capture this by positing a single category of relations into which they both fall. Indeed, this is an important reason for thinking of PRAGMATIC relations as a single class. On the other hand, as we have seen, the bipartite distinction hardly makes matters clearer. To my mind, those definitions of PRAGMATIC relations which are truly general (in the sense that they propose a single general characterization of all relations in this class) are open to the objections which Sweetser has leveled at them, while those definitions that allow for the distinction which Sweetser notes are essentially disjunctions, and hence tripartite in every respect except perhaps a terminological one. And so the problem of defining the commonalities between SPEECH-ACT and EPISTEMIC relations is still outstanding for both types of account. 4-2

Problems in defining the level at which relations hold

A related problem with both proposed classifications concerns another type of disjunction which is present, more or less implicitly, in the definition of PRAGMATidSEMANTIC/SPEECH-ACT relations. In each case, the general idea is to suggest that a basic relation, such as cause, disjunction or temporal sequence, applies between the two related spans, and that the different classes of relation are the result of this basic relation applying at different levels: between the propositions which are expressed by the spans, or between the speaker's beliefs in these propositions, or between the speaker's utterance of the propositions. To make this more concrete, we can imagine three functions that operate on an utterance U: one function c(U) which returns its propositional content, one function b(U) which returns the speaker's belief in its content (or some event relating to the genesis of this belief), and one function u(U) which returns the utterance itself (i.e. the identity function). We then could state, generally, that the coherence relation between two utterances UJ and U2 is to be analyzed as R (f(Ul),f(U2))3

where R denotes a 'basic' relation (cause, sequence or whatever), and f denotes one of the above functions. The definitions of the three categories of relation would then be straightforward:

135

136 Alistair Knott

for SEMANTIC (coNTENT) relations, f = c; for EPISTEMIC relations, f =b; for SPEECH-ACT relations, f = U. However, while these definitions would be nice, they do not always result in the analyses that have been suggested as suitable for the examples we have seen thus far. In fact, getting the analysis right frequently depends on us applying different functions to the two utterances in the relation. For instance, consider the following SEMANTIC, EPISTEMIC, and SPEECH-ACT relations: (16) Bill was starving, so he ate loads offood. ( 17) Bill is starving, so he'll want loads of food. (18) Bill is starving. So why isn't he eating?

In each case, the basic relation can be taken to be cause. In (16) and (17), the functions seem to work out: the relation applies between the propositional contents of the two spans in the former case, and between the beliefs in these propositions in the latter case. But in ( 18) we would not want to say that the relation applies between the two utterances. The cause of my asking the question 'why isn't Bill eating?' is simply the fact that Bill is starving, or possibly my belief in this fact; it is certainly nothing to do with my utterance of it. In fact, the case of SPEECH-ACT relations is even more problematic. While some SPEECH-ACT relations only involve one of the utterances, others involve both. The following relation is a case in point: (19) Would you like to come round tonight? Or are you busy?

In this case, as we have seen, we must hold that the basic relation (here, disjunction) applies between the two interrogative speech acts to get the desired analysis. To sum up: in the accounts we have seen so far, it does not seem possible to state in a general way 'which variable changes' when we move from one level to another. Again, we have to rely on disjunction to get things right. For instance, Sanders et al. ( 1992) define the class of PRAGMATIC relations as those where the segments are related "because of the illocutionary meaning of one or both of the segments". What we really want is a formula with a free variable in it, and different value for this variable for CONTENT, EPISTEMIC and SPEECH-ACT relations; but it is far from clear how this can be achieved.


4·3

Problems with imperative sequences

A general problem we have already mentioned is that the classifications of relations considered so far have a certain amount to say about relations involving non-indicative sentences, but seem unlikely as the basis of a complete account, either of these sentences or of the relations between them. To take a specific example of this problem, consider the case of a temporal sequence of imperatives. For instance: (20) Peel the onions. Then chop them. Interestingly, there are no examples in Sweetser of and or then being used to relate two imperatives in a temporal sequence. This is because the account of speech-act relations is clearly not suitable in such cases. We certainly do not want to interpret the temporal sequence here as holding between the speaker's utterances: the important thing, of course, is that the actions themselves need to be performed in the right sequence. The problem, then, is to come up with an account which does justice to cases such as (20) while still giving the kind of analysis of sequences of premises or speech acts that we have already seen in Section 3.3. It should be quite clear that any complete account of non-indicative sentences would involve a departure from the current definitions of PRAGMATidSPEECH-ACT relations, in the sense that these definitions are all expressed in terms of the information which relations convey to a hearer, over and above any information which is contained in the individual utterances. It would obviously be wrong to analyze a simple imperative proposition in this way. There is clearly more to the imperative do X than the information I [the speaker1 tell you to do X; we have to make reference in the definition to the conditions under which the imperative has the desired effect on the hearer. The same seems uncontroversially true for the sequence of imperatives given above: the relation of sequence has to be seen as part of the desired effect of the utterance. What is needed is a way of unifying this analysis with the analyses we have seen so far of imperatives, as they feature in other types of relation. 4·4

Problems of explanatory adequacy

The problem just mentioned for SPEECH-ACT relations has a parallel in the definitions proposed so far for EPISTEMIC or argumentative relations. Again,

137

138 Alistair Knott

this stems from the fact that the content of these relations is essentially taken to be informative. Consider the following example: (21)

Bill is starving, so he11 want loads of food.

According to the analyses seen so far, the relation in this text basically provides the hearer with information about the causes of one of her beliefs. This seems plausible enough in some contexts. However, a common- even prototypical - function of argumentative relations is to achieve a rhetorical effect on their hearers, causing them to believe a proposition in circumstances where simply stating it would not. (Indeed, the category of PRAGMATIC relations has long been associated with the effect of increasing some desire or positive regard in the hearer; see for instance Mann and Thompson (1988), Bateman and Rondhuis (1997).) Analyzing an argumentative relation as providing information about why the speaker believes something does not by itself explain how it can have this effect, at least not directly. It requires in addition the stipulation that when the hearer learns about the causal processes amongst the speaker's own beliefs, this will prompt similar processes amongst his own beliefs. This stipulation is not part of the relation definitions we have seen so far. Moreover, the idea that this kind of analogy is responsible for the effect of an argumentative relation is not particularly plausible: an argument that is compelling for one person is not always compelling for someone else. 4·S

Problems with 'conditional speech acts'

A final problem for the definitions we have seen so far relates to their application in conditional or disjunctive contexts. The problem is specific to the class of SPEECH-Acr relations. Consider, for instance, Sweetser's analysis of the following texts: (22) Would you like to come round tonight? Or is your car still in the shop?

(23) How old are you, if it's not a cheeky question? As we have seen, (22) is analyzed as a case of SPEECH-ACT disjunction: the disjunction is between whether the speaker is taken to perform the interrogative speech act associated with the first utterance, or that associated with the second. Example (23) is a SPEECH-Acr conditional.That is to say, the question 'how old are you?' is only to be understood as being asked if it is not a cheeky one. Both of these analyses rely on the notion of conditional speech acts. But

Semantic and pragmatic relations 139

this notion is not unproblematic, as Sweetser recognizes. Is it really up to the hearer to decide whether the speaker has performed a speech act? Clearly, conversation involves a great deal of cooperation between speakers and the assumption of certain shared goals. But is the issue of whether a speaker has performed a speech act really a matter for negotiation between the two participants? It seems that to claim this is at least to diverge from what is normally meant by a speech act or utterance. Moreover, there are some cases where a speech-act analysis would require a fairly radical divergence from what is normally meant. Consider the following example: (24) If you run out of money, come to me for a loan. I should point out that (24) is not the kind of text that Sweetser uses to exemplify SPEECH-ACT conjunctions. However, as already stressed, it would be useful to have an account which was general enough to cover this case as well. The alternative of having distinct treatments for the two kinds of conditional imperative seems to involve a measure of redundancy. In any case, if we were to attempt a SPEECH-ACT reading of this example ('If you ever run out of money, I tell you to come to me for a loan'), we would certainly encounter problems: we would have to envisage the speech act as occurring some way into the future, at some as-yet unknown moment. Moreover, if we interpret the text as expressing a habitual conditional (where we could substitute if with whenever), the speech act must also be allowed to occur arbitrarily often. Clearly, in either case we are stretching the notion of'speech act' beyond its original meaning.

s.

A new bipartite definition

s.t

Intention-based definitions of coherence relations

The problems mentioned in Sections 4.4 and 4.3 both point towards the need to include reference to the intentions which underlie all utterances in the definitions of SPEECH-ACT and EPISTEMJC relations. In fact, reference to intentions in relations has long been advocated by those interested in discourse relations from the point of view of computational linguistics, in particular from text generation. For Mann and Thompson ( 1988 ), a specification of the effect achieved on a reader by juxtaposing the two spans in a relation is a central component of its definition. This idea in turn gave rise to a procedural concep-

140 Alistair Knott

tion of relations, as planning operators in discourse structuring systems; for a review, see Hovy (1993). More recently, Moore and Paris (1993) and Moore and Pollack (1992) have also emphasized the need for a representation of intentions in discourse, though in their model it is proposed that intentional relations should be represented alongside informational relations, rather than as an alternative to them. While this work provides us with many of the concepts needed to think of relations in terms of intentions, none of it directly addresses the issue we have been concerned with, namely the problems with and tensions between the definitions of PRAGMATIC, EPISTEMIC and SPEECH-ACf relations. An intentionbased approach to the problems is proposed in the next section. s.2

Intention-based definitions for SEMANTIC and PRAGMATIC relations

The idea I am proposing is quite simple. Let us assume we are characterizing the relation between two utterances Ul and U2. Again, we assume that we can identify a basic relation R (cause, disjunction, or whatever) that is to be used in this characterization, and that the task is to identify what this relation applies between. I am proposing that two classes of relation are sufficient to capture much of the data so far considered; retaining the bipartite terminology, these will be termed SEMANTIC and PRAGMATIC. In each case, a relation between Ul and U2 is expressed in terms of the intended effect of the complex utterance Ul + U2. We also make use of the functional notation introduced in Section 4.2. Two functions are needed: one function c(U) returns the propositional content of an utterance U as before; the other function ie(U) returns the intended effect of U. We can then define ie(Ul + U2) in terms of R, Ul, U2, and these two functions: Relation-type SEMANTIC

PRAGMATIC

Definition

ie (Ul + U2) =believes(hearer, R(c(Ul), c(U2))) ie (Ul + U2) =R(ie(Ul ), ie(U2))

In English: the intended effect of a SEMANTIC relation is that the hearer believes that the basic relation R holds between the propositional contents of the two related utterances, while the intended effect of a PRAGMATIC relation is that the basic relation actually holds between the intended effects of the two related utterances. In each case, the relation definition evaluates to a proposition, which can be thought of as the goal state which the complex utterance is


expressed to achieve. In the case of SEMANTIC relations, this proposition is that the hearer believes a second, embedded proposition, namely that the relation applies. For PRAGMATIC relations, the proposition that the relation applies is directly asserted, rather than scoped within a hearer belief. The other difference between SEMANTIC and PRAGMATIC relations concerns which function is applied to Ul and U2 to obtain the propositions between which the relation is asserted to apply. Before we proceed to some examples of these definitions in action, some axioms need to be specified for determining the intended effect ie(U) of an atomic utterance U. While the notion of the intentions behind an utterance is notoriously ill-defined, I will be using it in the following technical, and fairly concrete, sense: a. The intended effect of an indicative sentence is that the hearer believes its propositional content. b. The intended effect of an imperative sentence is that the hearer performs the action in question. c. The intended effect of an interrogative sentence is that the hearer answers the question. Note that in each case, the intended effect evaluates to a proposition. We can now illustrate how the definitions work in a range of cases. 5·3

Causal relations

The following two texts are SEMANTIC and PRAGMATIC causal relations respectively: (25) Bill was starving, so he ate five sandwiches. (26) Bill ate five sandwiches, so he was [must have been) starving.

For (25), the propositional content of the first utterance is that Bill was starving; the propositional content of the second utterance is that Bill ate five sandwiches; and the intended effect of the complex utterance is that the reader believes that Bill's being starving caused Bill's eating of five sandwiches. For ( 26 ), the intended effect of the first utterance is that the hearer believes that Bill ate five sandwiches; the intended effect of the second utterance is that the hearer believes that Bill was (must have been) starving; and the intended effect of the text as a whole is that the hearer's belief that Bill ate five sandwiches

141

142 Alistair Knott

actually causes the hearer's belief that Bill was starving. This analysis seems to provide a good account of an argumentative relation. In describing the cause in such a case as something which the speaker intends to occur between two hearer beliefs, it provides the basis for an explanatory account of how argumentative relations achieve their persuasive impact; this is something which in Section 4.4 we argued was independently needed anyway. Sweetser would classify (26), being argumentative, as EPISTEMIC. Now note that the new definition of PRAGMATIC relations applies equally well for SPEECHACT causal relations. For instance: (27) We're late, so hurry up. In this case, the intended effect of the first utterance is that the hearer believes that we are late; the intended effect of the second utterance- an imperativeis that the hearer hurries up, and the intended effect of the whole text is that the hearer's belief that we are late actually causes the hearer to hurry up. Again, this seems a plausible account of the text in question. In fact, it seems more direct than an account in terms of speech-act glosses: ifwe assume the speech-act gloss ('so I tell you to hurry up') there is still a step to take between the hearer's understanding the cause ofthe speaker's statement and his actually hurrying up. In comparing (26) and (27) the main thing to note is that the work which in Sweetser's account is done by differences between the definitions of EPISTEMIC and SPEECH-ACT relations is done in the new account simply by the differences between the intended effects of an imperative and an indicative sentence. As we have argued in Section 4.3, these are differences which any theory of sentence semantics needs to represent anyway. Moreover, in framing a single definition which covers EPISTEMIC and SPEECH-ACT relations, we are addressing the generalizations between the two classes of relation which, as noted in Section 4.1, are problematic for the previous accounts.

5·4

Disjunctive relations

Now consider some disjunctive relations. (28) The milk is in the fridge, or it is on the sideboard. (29) John is home, or somebody is picking up his newspapers. (30) Would you like to come round tonight? Or is your car still in the shop? Example (28) is a SEMANTIC relation. The analysis here is straightforward: the


propositional content of the first utterance is that the milk is in the fridge; the propositional content of the second utterance is that the milk is on the sideboard; and the intended effect of the complex utterance is that the hearer believes that either the milk is in the fridge or it is on the sideboard. Example (29), which is EPISTEMJC in Sweetser's terms, is PRAGMATIC in ours. Accordingly, the intended effect of the first utterance is that the hearer believes that John is at home, the intended effect of the second utterance is that the hearer believes that someone is picking up John's newspapers; and the intended effect of the complex utterance is that either the hearer believes John is home, or the hearer believes someone is picking up his newspapers. This analysis seems at least as good as that given by the definition of EPISTEMIC relations. Example (30) is also PRAGMATIC for the present account. Thus, the intended effect of the first utterance is that the hearer answers the question (i.e. says whether he would like to come round tonight). The intended effect of the second utterance is also that the hearer answers the question (i.e. says whether his car is still in the shop); and the intended effect of the complex utterance is that either the hearer answers one question, or answers the other. Again, this analysis seems to be perfectly good. In fact, it improves on the speech-act account: note that we do not have to introduce the problematic notion of 'conditional' speech acts.

5·5

Conditional relations

Now consider how conditional relations fare under the new definition: IfJohn goes to a party, he gets drunk. (32) IfJohn went to the party, he was trying to infuriate Miriam. (33) How old are you, if it's not a cheeky question?

(31)

Example (31) is a SEMANTIC relation. The propositional content of the first utterance is that John goes to a party; the propositional content of the second utterance is that John gets drunk; and the intended effect of the complex utterance is that the hearer believes that if John goes to a party he gets drunk. Example (32), which is EPISTEMIC in Sweetser's terms, is PRAGMATIC in ours. Accordingly, the intended effect of the first utterance is that the hearer believes that John went to the party; the intended effect of the second utterance is that the hearer believes that John was trying to infuriate Miriam; and the intended effect of the complex utterance is that if the hearer believes that John went to

144 Alistair Knott

the party, the hearer believes that John was trying to infuriate Miriam. Again, this seems as good as the EPISTEMIC account. Example (33), which is SPEECH ACf in Sweetser's terms, is also PRAGMATIC for the new definition. Thus, the intended effect of the first utterance is that the hearer answers the question (i.e. says how old he is); the intended effect of the second utterance is that the hearer believes that the question is not cheeky; and the intended effect of the complex span is that if the hearer believes the question is not cheeky, he answers it. This seems to me to be an improvement on the SPEECH-ACT account. For one thing we avoid the problematic notion of conditional speech acts, as before. Moreover, the analysis specifies that it is the hearer's judgment as to whether or not the question is cheeky on which something is conditional- the SPEECH-ACf account does not say anything about the perspective from which this is to be judged. Finally, consider another type of conditional imperative, the kind that Sweetser's account does not treat: (34} If you run out of money, come to me for a loan.

The new definition of PRAGMATIC works just as well for this example as for the others: the intended effect of the complex span here is that if the hearer believes he has run out of money, the hearer comes to the speaker for a loan. 4

s.6

Sequential relations

Finally, consider some sequential relations. (35} John peeled the onions. Next he chopped them. (36} John's unsuitable. To begin with, he's too young. Next, he's hotheaded ... (37} Peel the onions. Next, chop them.

Example (35) is a SEMANTIC relation: the intended effect ofthe complex utterance is that the hearer believes that John peeled the onions and then chopped them. Example (36) is a PRAGMATIC relation, and would presumably be taken as EPISTEMIC by Sweetser. On the new analysis, the intended effect of the complex utterance is that the reader first believes that John is too young, and then believes that John is hotheaded. Again, this analysis seems as good as the EPISTEMIC analysis; we are still talking about the ordering of events in the epistemic world, but it is the hearer's world rather than the speaker's. Example (37) is also PRAGMATIC in our terms. It is not the kind of example


of a speech-act sequence that gets discussed in the literature, because as we saw in Section 4.3, it would be wrongly analyzed in this case. However, by the new definition it is unproblematic: the intended effect of the first utterance is that the hearer peels the onions; the intended effect of the second utterance is that the hearer chops them, and the intended effect of the complex utterance is that the hearer first peels the onions and then chops them.

6.

An assessment of the new definitions

6.1 Advantages of the new definitions

I will begin by briefly summarizing the advantages of the new definitions of SEMANTIC and PRAGMATIC. Firstly, the definitions permit some useful generalizations across EPISTEMIC and SPEECH-ACT relations. Because the intended effect of all imperative speech acts is different from that of an assertive one, expressing the definition of PRAGMATIC relations in terms of intended effects allows a single characterization of these two kinds of relation. Secondly, the new definitions provide a clearer picture of compositionality. They are compositional in two senses. Firstly, they allow a consistent account of the variable that changes when we move from SEMANTIC to PRAGMATIC relations- namely, the identity of the function (c(U) or ie(U)) that is applied to the two related utterances. Moreover, it is compositional in a more conventional sense. When we think of compositional semantics, we say that the meaning of a sentence can be derived from the meaning of its constituents, plus their manner of combination. The new definition allows us to make a similar statement about intended effects, for PRAGMATIC relations: we can say that the intended effect of the complex text span formed by a PRAGMATIC relation is derivable from the intended effects of its two constituent spans, together with the relation which links them. Thirdly, the new definitions provide an explanatory story for argumentative relations. The new definition of PRAGMATIC relations, as applied to an argumentative text, does not simply represent it as a description ofthe speaker's own thought processes, but specifies how it is that it can, if successful, have a rhetorical impact on the hearer. Fourthly, the new definitions provide for a better treatment of conditional speech acts. As we have seen, by making the conditional events the intended effects of utterances, rather than the utterances themselves, we avoid a number

145

146 Alistair Knott

of problems in analyzing texts where ifand or are used to conjoin imperative or interrogative sentences. Finally, the new definitions appear to provide the basis for a more general treatment of imperatives. We have seen that the new definition of PRAGMATIC relations allows them to apply more widely, to sequences of imperatives and conditional instructions, as well as to the set of cases dealt with under the rubric of SPEECH-ACT relations. 6.2.

Some problems for the new definitions

While there are many advantages with the new definitions, I should also note two cases where they do not apply straightforwardly, and where some modification of the account appears necessary. Both of these cases involve relations which Sweetser has analyzed as SPEECH ACT; and in each case there are aspects of this definition which the new definition of PRAGMATIC does not apparently capture.

Problematic SPEECH-ACT causes Consider the following case of a causal SPEECH-ACT relation. (38) Since you asked nicely, I am 78.

In this text, the definition of SPEECH-ACT relations seems more natural: it does indeed seem as though the primary effect here is informational, with the hearer being told why the speaker is making a certain utterance. If we used the new conception of PRAGMATIC relations, we get a strange analysis, where the intended effect of the complex utterance is that the hearer's belief that he asked nicely should cause the hearer to believe that the speaker is 78. This seems off the mark. There are other cases where the new PRAGMATIC definition is possible, but not particularly attractive. For instance: (39) The rules can't be broken, so "no". (40) Since we're on the subject, when was George Washington born? In each case, the original definition of SPEECH-ACT relations is a lot more natural. In summary, it must be acknowledged that there are cases where the new definition of PRAGMATIC relations does not work. So we have not eliminated the need for a class of SPEECH-ACT relations altogether; we have just made it smaller, and possibly more homogeneous. The question therefore remains as to how to define the class of SPEECH-ACT


rdations. It might be thought that, now the class is smaller, we could frame a definition for it on the same terms as those given for SEMANTIC and PRAGMATIC relations. We could imagine re-introducing the function u(U), which for an utterance U returns the utterance itself, and suggesting that for SPEECH-ACT rdations, the intended effect of the complex utterance ie(Ul + U2) given some basic relation R is something like believes(hearer, R(u(Ul), u(U2))). But this does not work. The problem is the same as before: we want to apply different functions to Uland U2. (It is not because the speaker says that we are on the subject of George Washington that he asks about when he was born, but because he is on the subject, or believes he is.) The attractive feature of the new definitions for SEMANTIC and PRAGMATIC is that the same function is applied to both utterances.

Problematic SPEECH-ACT conditionals Another class of problematic cases includes a group of conditionals of the foUowing form: (41) If you're hungry, there's some soup in the fridge.

Sweetser would analyze this as a SPEECH-ACT conditional: 'if you are hungry, then [let us consider that) I tell you that there is soup in the fridge'- the idea being that the assertive speech act is not understood to have been made unless the hearer is hungry. This certainly seems strange, for the reasons already discussed, but in this case it is not clear that the new definition of PRAGMATIC is any easier to understand. On the PRAGMATIC account, the intended effect of the complex utterance is that if the hearer believes that he is hungry, then he also believes that there is soup in the fridge. But surdy, given the text in question, the hearer is going to believe there is soup in the fridge regardless of whether he believes he is hungry? The notion of conditional hearer actions (the obeying of orders, the answering of questions) is perfectly intelligible, but the notion of conditional hearer beliefs in this context seems at least as strange as that of conditional speech acts.

7·

Future directions: generalizing over the intentions of participants

and protagonists Before concluding, I will sketch an extension (still very speculative) to the current definitions of SEMANTIC and PRAGMATIC rdations, which points to

148 Alistair Knott

further possible generalizations using the concept of intentions. This involves the introduction of a second bipartite classification of relations, into CAUSEDRIVEN and RESULT-DRIVEN relations, which I have discussed elsewhere (Knott & Mellish 1996). 7.1

Cause-driven and result-driven relations

The new distinction can be introduced by considering the following pair of sentences: (42) John had been up all night, but he looked fresh as a daisy. (43) John was hungry, but he couldn't find anything to eaL

But is typically analyzed as signaling a violated expectation of some kind. For instance, in (42), the fact that John has been up all night leads to the expectation that he will look tired, which is then not forthcoming. The analysis is often formalized in terms of presupposed defeasible rules (Knott & Mellish 1996; Lagerwerf 1998). In (42), for instance, the presupposed rule would include 'X has been up all night' on its left-hand side, and 'X looks tired' on its right: the but indicates that this rule is defeated in the situation being described. While this analysis works for (42), it does not work for (43), as is noted by Spooren (1989) and Knott and Mellish (1996). To conform to the proposed representation, we would have to assume a rule of inference whose premise is that X is hungry and whose conclusion is that X finds something to eat. More generally, to deal with cases like (43 ), we would need a rule of inference whose premise is the fact that an agent has a goal, and whose conclusion is that the goal is achieved. However, this rule of inference is clearly not always a valid one: agents in the world frequently have goals that cannot be achieved. Of course, if we also know that the conditions in the world are such that the goal is achievable, then the inference is perfectly possible to make. But that is precisely what is not known in an example like (43)- until the clause introduced by but is processed. 5 Knott and Mellish (1996) propose that instead of attempting to reduce cases like (43) to a violated expectation, we should distinguish between two kinds of but: one, termed CAUSE-DRIVEN, which signals a violated expectation, and one, termed RESULT-DRIVEN, which signals a frustrated plan. The notion of a protagonist's plan is thus accepted as a primitive in the account. A generalization between the two types of but is then proposed, which draws on the fact that a defeasible rule can be used in two ways: deductively, to generate


an expectation of its right-hand side based on knowledge of its left-hand side, or abductively, to plan an action which achieves a state of affairs which features on its left-hand side in pursuance of a goal to achieve the state of affairs on its right-hand side. See Knott and Mellish ( 1996) for details of this distinction. The crucial point to note about the distinction for present purposes is that the definition of RESULT-DRIVEN relations makes explicit reference to the goals of a protagonist.

7.2

Speaker and protagonist goals

The presence of goal-based primitives in the definitions of PRAGMATIC and RESULT-BASED relations opens up interesting possibilities for further generalizations. Specifically, there seem to be some interesting similarities between the goals a speaker pursues by making utterances, and the goals a protagonist being described in a discourse pursues by taking actions. Compare these two texts, for instance: (44) Bill told Jim to go to bed, but Jim wasn't tired. (45} Bill [to Jim]: Go to bed! Jim [to Bill]: But I'm not tired! In each case, we can talk about Bill seeking to satisfy a goal that Jim goes to bed by telling him to do so, and thereby discovering a circumstance that indicates that this action on its own will not be enough. In each case we can also use the conjunction but. But the two texts are nevertheless quite different in (44), the intention is that of the protagonist in a narrative monologue, while in (45) it is that of the first speaker in a conversational exchange. Moreover, (44) contains only indicative clauses, while (45) contains an imperative clause too. It seems quite likely that the definition of PRAGMATIC which is required to account for the imperatives in (45) will also suffice to introduce the goals necessary for the analysis of the but in this example as RESULT-DRIVEN. This possibility is something which will be pursued in further work.

8.

Summary

This paper began by noting some difficulties with existing notions of PRAGMATIC, EPISTEMIC and SPEECH-ACT relations. A solution to some of these prob-

ISO

Alistair Knott

lems was then proposed, in the form of a new definition of PRAGMATIC relations framed in terms of the intended effects of a speaker's utterances. This new definition has a number of advantages over its predecessors, although there are still some remaining problems with it. Finally, a possible extension to the account was presented, by introducing a new category of RESULT-DRIVEN relations: there appear to be some interesting generalizations in prospect relating to speaker and protagonist goals. Whether these can be exploited in improved definitions of SEMANTIC and PRAGMATIC relations is a matter for future research.

Notes Sweetser in fact adresses two early versions of the bipartite distinction formulated by Ross ( 1967) and Davison ( 1973 ). However, her criticisms apply to the more recent versions too.

1.

2. In fact, Sweetser ( 1990, p. 92) is unwilling to commit to the view that explicit performative should be understood as contributing to the semantics ofa sentence. But the rationale for this is mainly concern for a proper demarcation the domains of'semantics'and 'pragmatics'; the only relevant implication for our purposes is that the performative glosses are to be understood as falling in the latter domain rather than the former.

3· Or perhaps more precisely, as stating that the proposition R(f (Ul), /(U2)) is asserted by the speaker.

+

Note that the account works equally well if the conditional is interpreted as a habitual: there are none of the problems associated with conditional, future or repeated speech acts that would arise if we extended the SPEECH-ACT accounL Note that the protasis has to be evaluated from the perspective of the hearer in this case. This reading certainly seems possible, but it might be thought that it could also be evaluated from the speaker's perspective (i.e. the hearer comes to the speaker for a loan if the hearer believes the speaker believes he has run out of money). If this is considered a problem, it could be borne in mind that the intended effect of an indicative utterance with content p can also be taken to be that the hearer believes that the speaker believes that p. With this alternative interpretation, both of the required perspectives are available.

s.

Spooren ( 1989) apparently suggests that a violated expectation does arise in a case such as (43), as the necessary additional premise (in this case, that there is food to be found) is an implicature of the utterance expressing the protagonist's goal (i.e./ohn was hungry). While we must certainly implicate the fact that John has a goal if this is not explicitly stated, there seems no good reason to suggest that the speaker's informing the hearer about this goal should lead to an implicature that the goal is satisfiable. If anything. hearers are used to hearing about situations in which a protagonist's goals are not satisfied: this is certainly the predominant case in narratives.


References Bateman, J., & Rondhuis, K. J. (1997). Coherence relations: Towards a general specification. Discourse Processes, 24, 3-49. Davison, A. ( 1973). Performative verbs, adverbs and felidty conditions: an inquiry into the nature ofperformative verbs. Ph.D. thesis, University of Chicago. Elhadad, M., & McKeown, K. R (1990). Generating connectives. In COLING-90, 97101.

Halliday, M., & Hasan, R. ( 1976). Cohesion in English. London: Longman. Hobbs, J. R. ( 1985 ). On the coherence and structure ofdiscourse. Technical Report CSU85-37, Center for the Study of Language and Information, Stanford University. Hovy, E. ( 1993 ). Automated discourse generation using discourse structure relations. ArtifidaiinteUigence, 63,341-385. Knott, A., & Mellish, C. ( 1996 ). A feature-based account of the relations signalled by sentence and clause connectives. Language and Speech, 39, 143-183. Lagerwerf, L. ( 1998). Causal connectives have presuppositions. Ph.D. thesis, Tilburg, The Netherlands: Katholieke Universiteit Brabant. Mann, W. C., &Thompson, S. A. (1988). Rhetorical structure theory: A theory of text organization. Text, 8, 243-281. Martin, J. (1992). English Text: System and Structure. Amsterdam: Benjamins. Moore, J., & Pollack, M. ( 1992). A problem for RST: The need for multi-level discourse analysis. Computational Linguistics, 18,537-544. Moore, J.D., & Paris, C. L. (1993). Planning text for advisory dialogues: Capturing intentional and rhetorical information. Computational Linguistics, 19, 651-694. Oversteegen, L ( 1997). On the pragmatic nature of causal and contrastive connectives. Discourse Processes, 24,51-86. Redeker, G. (1990). Ideational and pragmatic markers of discourse structure. Journal of Pragmatics, 14,367-381. Ross, J. R. ( 1967). Constraints on variables in syntax. Ph.D. thesis, MIT. Now reprinted as Infinite Syntax. Norwood, NJ: Ablex. 1986. Sanders, T. ( 1997). Coherence relations in context; on the categorization of positive causal relations. Discourse Processes, 24, 119-147. Sanders, T. J. M., Spooren, W. P.M., & Noordman, L. G. M. (1992). Towards a taxonomy of coherence relations. Discourse Processes, 15, 1-35. Spooren, W. ( 1989). Some Aspects Qf the Form and Interpretation ofGlobal Contrastive Coherence Relations. Ph.D. thesis, Catholic University of Nijinegen, the Netherlands. Sweetser, E. ( 1990). From Etymology to Pragmatics: Metaphorical and Cultural Aspeas of Semantic Structure. Cambridge: Cambridge University Press. Van Dijk, T. A. (1979). Pragmatic connectives. Journal ofPragmatics, 3, 447-456.

CHAPTER

6

On the production of causal-contrastive although-sentences in context Leo G. M. Noordman* Katholieke Universiteit Brabant

In our thinking and reasoning, concepts of causation and negation play an important role. For example, on the basis of observations in the world, we are able to predict the consequences of the facts we observe. We are able to anticipate future consequences of observed causes. We are able to reason about events that do not exist, and about hypothetical events. We can generate expectations about future events, but also negations of expectations. These processes may be rather complex; they involve causality, including expectations, and negation. The notions of causality and negation are pervasive in our thinking and reasoning processes. In general, we express our thinking and reasoning in language. Language is the expression of our train of thought and it reflects the discursive processes in our thinking. Consecutive sentences or utterances reflect in general the thinking process that we perform as speakers and writers and that we want our listeners and readers to perform as well. This chapter deals with the train of thought that occurs in our mind when we deal with causality and negation. This train of thought is referred to as a thinking or reasoning process in this chapter. These processes are observable in the language. What kinds of strategy do the speaker and writer use in formulating their complex thought? What kinds of decision do they make? The way in which we try to discover the reasoning processes is by analysing the product of these processes: the sentences that express the train of thought of the speaker and writer. Similarly, we are interested in the processes in understanding complex reasoning. How do listeners and readers understand the complex thought of their interlocutor? This article restricts itself to the analysis of concessive sentences expressed by the conjunction although. An example is sentence ( 1).

154

Leo G.M. Noordman

(1) Although John had worked hard, he failed the exam.

A concession expresses a complex thought. According to Quirk, Greenbaum, Leech and Svartvik (1985, p. 1098) "concessive clauses indicate that the situation in the matrix clause is contrary to expectation in the light of what is said in the concessive clause". That a concessive relation includes a causal and a contrastive relation is dearly expressed by Helwig and Buscha (1991): "an expected causal connection is ineffective. The ground that is mentioned in the subordinate clause does not have the effect that is expected on the basis of the law of cause and effect" (p. 591; translation LN); "A concessive relation includes a causal relationship as well as an adversative relationship" (p. 562; translation LN). And Grevisse ( 1986, p. 1667; translation LN) argues that "a concessive proposition indicates that the logical relation that one expects between the concession and the main verb does not hold. A concessive proposition expresses a cause that is ineffective and counteracted and that did not have the effect that one could expect". But now compare sentences (1) and (2). (2) Although John failed the exam, he had worked hard. Both sentences express a causal and a contrastive relation. The subordinate clause evokes an expectation on the basis of causal world knowledge. The expectation is contrasted with and denied by the content of the main clause. Although both sentences contain the same words, the expectation in the two sentences is quite different. In a sense, the sentences express different relations. So, the first question we have to address concerns the kinds of relation that are expressed by although-sentences. If the although-sentences have an underlying causal relation, it seems that there are different causalities involved Since the focal point of this chapter is how causality is expressed in language, the analysis of the although-sentences focuses on the underlying causality. The second question deals with the way in which the concessive relation affects the organization of the discourse. How do the although-sentences function in their context? If a concession expresses a denied expectation, one may wonder to what extent the expectation is evoked by the preceding context. If the expectation is denied, does the speaker/writer give in the subsequent text an explanation for the denial or does (s)he give an alternative reason or motivation for the event described in the main clause? How does the subsequent context follow up on the thoughts expressed in the subordinate clause and the main clause, i.e., how does the reasoning process develop? Several

On the production of causal-contrastive although-sentences in context

authors have argued that the two clauses have different status. This presumably has consequences for the function of the clauses in their contexts. The conceded part has little prominence (Grote, Lenke & Stede 1997). Elhadad and McKeown (1990) claim that the main clause in an although-sentence has directive status and the subordinate clause subordinate status. The argumentative orientation (Ducrot 1984) of the complex sentence as a whole is the argumentative orientation of the main clause (Elhadad & McKeown 1990). The same asymmetry is expressed by Mann and Thompson ( 1988) in terms of the notions nucleus and satellite. In an although-sentence that expresses a concessive relation, the main clause expresses the nucleus; the subordinate clause expresses the satellite. Nuclei are more central in a text than satellites. If one deletes the satellites from a text, what remains is still a coherent text; if one removes the 'most-nuclear' unit of the text, the text looses its significance. The nucleus and satellites have also different functions and effects for the writer/ reader. "The nuclear portion realizes the primary goal of the writer and the satellite provides supplementary material" (Thompson & Mann 1987, p. 360). A concessive "relation holds when a writer chooses to strengthen a point by affirming that point in the face of a potentially opposing point" (Thompson & Mann 1987, p. 363) . The nucleus is the point that is affirmed; the satellite is the opposing point. The writer "has positive regard for the situation presented in the nucleus" (Mann & Thompson 1988, p. 254). The nucleus expresses the belief, the approval or the intention of the writer. The writer intends that the reader's positive regard for what is described in the nucleus increases. A similar analysis is given by Spooren ( 1989) in his discussion of but-sentences. Spooren demonstrates that the but-conjunct dominates the other conjunct in the sense that the but-conjunct reflects the opinion of the writer more than the other conjunct. If the main clause expresses the directive act and the subordinate clause a subordinate act, or if the main clause expresses the nuclear information and the subordinate clause the satellite, this may have consequences for the thematic development of the text. The directive act and the nuclear information reflect the purpose of the writer. Then one may expect that the subsequent context will be a continuation of the main clause. The research by Spooren (1989) is relevant in this respect. He demonstrated that the but-clause reflects the opinion of the writer, and that, accordingly, the subsequent text is more related to the but-conjunct than to the other conjunct. This asymmetry was not found for clauses connected by the conjunction and.

•ss

1;6

Leo G.M. Noordman

The question about the role of the context is the more obvious if one realizes that the same relation can be expressed in different ways by changing the order of the clauses. For example, the information in sentence ( 1) can be expressed in sentence (3). (3)

John failed the exam, although he had worked hard

What will be a likely continuation of this sentence? Sentences ( 1) and (3) have the same nuclei. Spooren (1989) demonstrated that the subsequent context after a but-sentence is a continuation of the second clause, i.e. the but-conjunct. The question arises whether the subsequent context continues on the butconjunct, because the but-clause is the last conjunct, or because it expresses the opinion of the writer. When sentences were connected by the conjunction and there was no tendency for the subsequent context to follow up on the second conjunct. This suggests that the asymmetry found by Spooren is due to the conjunction but, and not to the linear order. On the basis of these considerations, one could expect that the context following sentence (I) and (3) is a continuation of the main clause, that expresses the nuclear information and the directive act and not a continuation of the last clause. But then the question arises whether there are differences between sentence (I) and (3 ). Aie there contextual constraints for these different kinds of the denial of expectation? It is likely that differences in clause order may be subject to thematic development in the text. One remark should be made with respect to the scope of the analysis of although-sentences in their discourse context in this chapter. Discourse reflects cognitive processes. These processes include the generation of the information by the writer, the expression of the information in such a way that the writer expresses his/her viewpoint and accommodates the viewpoint of other people including the potential reader. That these processes are involved is particularly clear in concessive sentences. Concessive sentences have in general an argumentative function in a discourse. The speaker/writer expresses his or her opinion in relation to other opinions. The nucleus has 'positive regard' and realizes the goal of the writer; the satellite is the opposing point. This chapter does not claim to analyse the argumentative richness of concessive sentences. That goes beyond the scope of this chapter. The analysis deals with the argumentative function of although-sentences only to the extent that we investigate how the main and subordinate clauses function in the development of the thought the writer expresses in the preceding and subsequent context. In that

On the production of causal-contrastive although-sentences in context 157

sense the present analysis can be considered as an empirical test of the notions nucleus and satellite: do the nucleus and the satellite behave in a different way with respect to the context? What emerges is "that the role of connectives is not only to indicate a logical or conceptual relation, but also to indicate the structural organization of discourse" (Elhadad & McKeown 1990, p. 98 ). This chapter will deal with these two aspects of concessive although-sentences. The first issue is: how is the logical or conceptual relation expressed in the although-sentence? The conceptual relation is an expectation, based on causality, that is denied. How is this relation expressed? We will analyse the underlying causality in although-sentences. The second issue is: how does the concessive relation fit in the organization of the discourse?

1.

Different kinds of although-relations

1.1

Text analytic differences between the relations

The reasoning underlying sentence ( 1) can be represented as a syllogism in which the conclusion is contradicted by the statement in the main clause: 'If one works hard, one normally passes the exam. John worked hard. Therefore, John would normally pass the exam. But in fact, John failed the exam'. The major premise in this syllogism is a general expectation that is generated on the basis of the subordinate clause. The expectation is contradicted by the main clause for the specific situation described by the minor premise. The major premise in (1) is different from the premise in (2). The premise in (2) is: 'If one fails the exam, one normally has not worked hard'. Consider the premise in sentence ( 1): 'If one works hard, one normally passes the exam.' This clause expresses a contingency between working hard and passing the exam. Underlying sentence ( 1) is the assumption that working hard is a cause for passing the exam. The relation between the two propositions in the premise is a relation between two events in the world: 'Working hard' of somebody (John) leads to 'passing the exam' of that person. There is a causal relation between the two events stated in the premise and, consequently, there is a causal relation underlying the events expressed by sentence ( 1). The analysis of sentence ( 2) is formally similar- in the sense that a premise is derived in the same way as in sentence ( 1) - but the premise is different. The derived premise, 'If one fails the exam, one normally has not worked hard', does not describe a contingency

158

Leo G.M. Noordman

between failing the exam and not working hard. It is not the case that failing the exam is a cause for not working hard. However, that sentence does express a contingency: failing the exam is the cause for the speaker's conclusion that the person in question probably did not work hard. So, the relation between the two propositions in the premise is not a relation between two events in the world, but between an event and a conclusion of the speaker. One can say that the relation between the propositions in ( 1) is a relation between locutions and in (2) between illocutions. It should be noted that the validity of the premise in sentence (2) rests on the validity ofthe premise in sentence ( 1): 'If one fails the exam, one normally has not worked hard' is justified because 'If one works hard, one normally passes the exam.' In the expectation underlying sentence ( 1) the consequence (passing) is deduced from the cause (working hard). I will refer to this as reasoning from cause to consequence. In the expectation underlying sentence (2) the reasoning is in the inverse direction: from the consequence one infers the cause. There is a kind of backward reasoning. The consequence is the sign for the cause. I will refer to this as reasoning from consequence to cause. One can argue that there is an incongruence in sentence (2) between what is conceptually the condition and the consequence and what is expressed in the sentence as antecedent and consequent. What conceptually is the consequence (failing the exam) is expressed as the antecedent in the sentence (if one fails, one has probably not worked hard). On the other hand, in sentence ( 1) there is a congruence between what is conceptually the cause and the consequence and what is linguistically expressed as antecedent and consequent. The expectation that is generated in ( 1) can be represented as 'if p then not-q', where p stands for the subordinate clause in (1), and q for the main clause in (1) (representing sentence (1) as: Although p, q). The expectation generated in (2) can be represented as 'if p then conclude not- q'. Relations as exemplified by ( 1) are called 'semantic' relations by Sanders, Spooren and Noordman ( 1992), since the relation between the propositions in the language express a relation between events in the world. In contrast, relations as exemplified by (2) are called 'pragmatic' relations, since an illocution, or a speech act is involved in the relation. Sweetser (1990) identifies the relations as content relations and epistemic relations resp. 'Epistemic' refers to the fact that a conclusion by the speaker is involved in the relation. It should be noted that content and epistemic relations are defined in terms of(il)locutions and not in terms of (reversed) causality. Reversed causality will in general imply an epistemic relation. But non-reversed causality does not necessarily imply a

On the production of causal-contrastive although-sentences in context 159

content relation. For example: 'I hear that it is raining, so the streets are wet' may mean that I conclude that the streets are wet and therefore is an epistemic relation ('so the streets must be wet'). Noordman (1979) investigated the psycholinguistic processing of these relations and called them 'condition-consequence relations' and 'inference relations' resp. The term 'inference relation' expresses that the condition is inferred from consequence. The sentence expresses an act of concluding. Traxler, Sanford, Aked, and Moxey ( 1997) used the labels'causal' and' diagnostic', where diagnostic expresses the inverse reasoning process from the consequence to the cause. In a more discourse semantics approach, Oversteegen ( 1997) described the relations involved in because-sentences and althoughsentences in terms of presuppositions and operators involved in the presuppositions. As an example: sentence (2) contains the presupposition that in general if x fails the exam, one may conclude that x did not work hard. The phrase 'one may conclude' is called a belief operator. One may identify a third kind of although-relation. An example is: 'Although I sympathize with your problems, get the paper in tomorrow' (Sweetser 1990, p. 79). The underlying meaning is 'I command you ...... , in spite of my sympathy'. The underlying premise is: if I sympathize with your problem, I will not command you to get the paper in tomorrow. The expectation can be represented as 'if p then say not-q'. The relation here involves an act of saying. Therefore, Sweetser calls these relations speech act relations (or rather speech act conjunctions). Oversteegen discusses these relations in terms ofa speech act operator. The three kinds of although-relation discussed so far contain an underlying expectation that is denied by the main clause. Although-sentences expressing a denial of expectation form the topic of this chapter. There is a fourth kind of although-sentence that is slightly different: the concessive opposition. As an example, consider the sentence 'Although that fiscal regulation yields much money, it is not fair' in the context of a discussion whether a particular fiscal regulation has to be maintained or not. There is no expectation that if a fiscal regulation yields much money, the regulation is fair. It is not the case that an expectation is evoked that is denied by the main clause. Rather, the subordinate clause and the main clause are expressions of opposite arguments. The subordinate clause is an argument in favor of the fiscal regulation and the main clause is an argument against it. This can be represented as p ~ r, and q ~ -r, where p is the subordinate clause; q is the main clause and r is a proposition that expresses the issue that is being discussed in the concessive opposition. In this

16o Leo G.M. Noordman

case, r is the issue whether that fiscal regulation should be maintained or abolished. In general r is a proposition that is introduced in the previous context (see Spooren 1989 for a similar analysis of concessive but-sentences). A concession is more symmetric than an denial of expectation, because both the main and the subordinate clauses in a concession refer to a proposition r that is derived from the preceding context. I prefer to distinguish between denials of expectation and concessive opposition, although Grote, Lenke and Stede ( 1997) correctly claim that "all concessions seem to share a common structure on an abstract level of knowledge representation" (p. 114). That abstract representation of concessions is: "On the one hand, A holds, implying the expectation of C. On the other hand, B holds, which implies Not-C, contrary to the expectation induced by A." (p. 95). But they have to make a distinction depending on whether not-Cis explicitly mentioned or not. IfA is the although clause and note is the main clause, then the concession is a denial of expectation. If A is the although-clause and B is the main clause, the concession is a concessive opposition (Spooren 1989). In order to decide whether an although-sentence expresses a denial of expectation or a concession, one has to find out whether the subordinate clause is directly related to the negation of the main clause, or whether both the main clause and the subordinate clause are related to a proposition in the context. Another way to distinguish between denial of expectation and concession is the following heuristic, that is based on the fact that concessions are more symmetric than denials of expectation: If one reverses the clauses without the conjunction ('Although p, q' into 'Although q, p') and if the difference between the two versions corresponds to the difference between a non-reversed and a reversed causal order, the original sentence was a denial of expectation; if there is no difference, the original sentence was a concession. The main focus in this chapter is on although-sentences that express a denial of expectation. The analysis of the concessive relations restricts itself to the underlying causal expectation: does the underlying expectation express a causeconsequence order or a consequence-cause order? I will use the notions default order causal relation and reversed order causal relation for this distinction. 1.2

Evidence for the differences in processing default order and reversed order causal relations

This section focuses on the processing of causal sentences that describe the relation between events in default order and in reversed order. Reversed order

On the production of causal-contrastive although-sentences in context

causals are more complex than default order causals. In a reversed order causal, there is an underlying 'you may conclude that' and that conclusion is based on a default order causal relation: 'If you fail the exam, one may conclude that you probably have not worked hard, because if you work hard, you pass the exam'. To test this hypothesis, Noordman (1979) had participants read conditional sentences that expressed causal relations. There were two kinds of sentence, exemplified by 'I{John is ill, he is not going to his work' and 'IfJohn is not going to his work, he is ill'. The first kind of sentence corresponds to a default order causal relation and the second kind to a reversed order causal relation. In these examples, the reasoning in the default order relation is from cause to consequence; in the reversed order relation it is from consequence to cause. One may hypothesise that reasoning from cause to consequence is easier than from consequence to cause, since there is a congruence between what is cause and consequence in the real world and what is linguistically expressed in the sentence as antecedent and consequent. In the sentences expressing a reversed relation, there is no such correspondence: "there is an incongruence between what is the condition and the [consequent] according to the structure of the sentence and what is the condition and the consequence according to the knowledge of the listener and speaker. Handling this incongruence requires extra processing time. This incongruence can be solved in fact by processing the embedding proposition: 'it can be inferred [concluded] that ... "' (Noordman 1979, p. 97) and by processing the default order causal relation.

Method Sixteen different item types were constructed by combining four dichotomies, only one of which is relevant for the present study: whether the sentence expressed a default order relation or a reversed order relation. There were eight topics of sentences, yielding 128 sentences altogether. Half of the sentences expressed a true situation in the world and half of the sentence expressed a situation that was not true in the world. There were two groups of sentences that were in all respects comparable. Each of the 20 participants in the experiment was presented with one group of 64 sentences. The sentences were presented by means of projectors with electronic shutters. The participants were required to judge as quickly as possible whether the sentence was true or false. The answers were given by pressing one of two response buttons. The time was measured from the moment the sentence was presented until the participant pressed a button.

161

162.

Leo G.M. Noordman

Results and discussion The average verification times for the correct answers are presented in Table 1. Table I. Mean verification times (ms) for the two relations. default order causal reversed order causal

3547 3833

Sentences expressing a default order causal relation were processed 286 ms faster than sentences expressing a reversed order causal: minF' (1,23)=10.19, ptpAR + P2>tSEN + 133*CLA + P4*CON+ 135*WOR

( + tulk>tPA~ + t2iJk>tSEN,Jk + t 31Jk·c~Jk + t 411k>tCON~lt + tsijt >two~ik + J.l 1o;t >tPA~lk + ~o;k >tSENiik + J.t30Jk*CLAiik + J.l40ik*CONiik + J.lsOjk*WORiik)

The variable Yiik represents a pause at location type ~ in the j-th text of the k-th subject. So, model ( 1) allows us, for each individual pause in the corpus, to estimate the overall mean for pauses at each location type (the so-called fixed parameters J3 1_5), and both within-text and between-text variances for pauses at each location type (the so-called random parameters, given between parentheses). The random parameters denote location-specific residuals (tiik) and text-specific residuals (J!Ojk). The first random parameter allows us to estimate variances for all location types (the so-called level 1 variances). In classical test theory this term indicates measurement error and it is assumed that tiik =N(O, crf). The second random parameter allows us to estimate text-

Conceptual and linguistic processes in text production 319

specific variances for each location type (the so-called level2 variances). 3·4

Results and discussion

Table 1 shows the results of the estimates of the mean length of pauses at the distinct location types and the within and the between-text variances. Table 1.

Estimates of mean times of pauses at five location types, estimates of within text and between text variances (ms). estimates locations fixed effects•

paragraphs sentences clauses constituents words

8472

2676 1336 1185 1046

within texts

between texts

19240 4673 1359 676

12050

468

406

162 019 003

• chi-squares for simultaneous contrasts= 212.5, df = 5, p = < .001

Table 1 shows that, on average, longest pauses occur between paragraphs, than between sentences, than between clauses, and so on. These results thus suggest that the selected location types are sensitive to pause time differences, all differences reaching the level of statistical significance. Furthermore, pause time variances, represented in the right-hand columns in Table 1, show that there are indeed considerable deviations from the estimated averages, both within and between texts, although between-text differences gradually loose their impact on low-level locations. In the remainder of this paper, these residual estimates will be of special concern. For now it suffices to note that the five selected location types account for over 50% of the variances observed at the various location types. The dean-cut correlation found between a pause's location and its duration warrants the assumption that levels of processing can indeed be inferred from the data. Granting the assumptions discussed in Section 2, conceptual processes appear to take place predominantly at the level of paragraphs, and perhaps at the level of sentences, indicating that text producers tend to engage in conceptual processing especially before a new paragraph and to a somewhat lesser extent before a new sentence. Linguistic processes such as grammatical shaping and lexical choice, appear to occur mainly at the 'lower' levels of

320

Joost Schilperoord

clauses, constituents and words, indicating that decisions as to syntactic shaping and wording are being made especially before clauses, constituents and words. This is not to imply that we can, for each distinct location, infer a uniquely determined kind of process. It may be so that planning at the sentence level differs from planning at the paragraph level in terms of the range, or the level of conceptual abstraction. For example, paragraphs are planned as to their 'gist' whereas sentences are planned in more conceptual detail. But this is off course mere speculation, as such differences cannot be discerned directly from differences in pause length. So, in conclusion, only a gross distinction between conceptual and linguistic processes appears to be observable from the data.

4- Interactions between conceptual and linguistic processes in text production 4-1 Introduction We now turn attention to the main question in this paper, concerning the interaction between conceptual and linguistic processes in text production. Section 2 addressed the interaction-issue from a theoretical point of view. Here, we start by making explicit what the presence of interaction would stand for empirically, notably, in terms of pause patterns. Consider the data model presented in Figure 2.

Figure 2.

Data model

The box in the middle of Figure 2 represents a simplified version ofa production model in which only conceptual processes (CP) and linguistic processes (LP) are distinguished. The left hand side depicts a 'production-external' explanatory variable X. For X it can be argued that, someway or another, it will have some effect on the performance of both CP and LP, so that (2) can be assumed.

Conceptual and linguistic processes in text production 3D

(2)

X~ X~

CP LP

In addition, these effects of X on CP and LP will be reflected by particular pause time parameters, for example, means and variances. Let us refer to these as PPcp and PPlp then the entire system of relationships can be given as (3 ). (3)

X ~ CP ~ PPcp X~

LP

~

PP1p

Since CP and LP are covert processes, the analyses to be reported will focus on exploring relationships (4). (4)

X~

( ... ) ~ PPcp pplp

X~( ... ) ~

As a first approximation of 'interactions between CP and LP', we may say that there are no interactions if there is no way of predicting PPcp on the basis ofPP1P or vice versa. In other words, interactions are assumed to be present if (4) can be rewritten as the covariation model (5). (5)

X

~

( ... )

~

ppcp H pplp

The assumption to be tested is that conceptual and linguistic processes run more or less interactively, so therefore (5) represents the hypothesis to be tested. The general structure of the next three subsections will be, first, to introduce X as a relevant external variable that has alleged effects on CP and on LP, then to specify how model (4) can be cast in terms of X and in terms of hypothesized pause patterns, and, finally, to test the hypothesis.

4-2

Differences between texts

The first external variable tested is differences between texts. Such differences may be of various sorts -conceptual complexity, or text length, but for now the precise nature of these differences need not bother us, as long as it is reasonable to assume that they exist, and may indeed have a certain impact on cognitive processes, especially at the conceptual level. Therefore, we hypothesize that, if indeed conceptual processes and linguistic processes interact, then conceptual difficulties are solved at the level of conceptual processing, but will have marked effects on 'lower' levels of processing as well. In terms of pause

3:1.2

Joost Schilperoord

patterns, we expect text-rdated differences to influence pause time variances at the level of paragraphs, but also at lower levels. To test such a hypothesis, we must gain insight as to the amount of pause time variance for all distinct location types that can be attributed to between-text factors. To put this in other words, the question is to what extent differences between texts carry over to pause time variances for the various location types. In terms of the level 1 and level 2 sources of variance discussed in Section 3, the analytical question thus boils down to estimating the relative contribution oflevel2 variance to the overall variance in pause time (level 1 + level 2) at all distinct locations. One way of obtaining such information is to estimate the so-called intra-class correlation coefficient, called p (cf. Cochran 1977). The factor p can be estimated by calculating the ratio of between-text residues and within-text residues for each location type (see (6)). (6)

Px = ~k ·~jk I

Eiik ·~ik

+ ~Dik•xiik

Equation (6) thus depicts the level2/level1 + level2 ratio for each pause xijk' Both the level 1 and levd 2 estimates were already given in Table 1. The interaction-hypothesis predicts the intra-class coefficients for paragraphs, sentences, clauses, constituents and words to be more or less similar. Applying (6) to these estimates results in the estimates shown in Table 2. Table 2.

p - estimates for location types

Location

Estimate

paragraphs sentences clauses constituents words

.385 .079 .106 .028 .005

Table 2 shows the influence of text-related differences to be dearly present at the paragraph level (.385). Hence, differences between texts have marked effects on the pause time variances at paragraph transitions. Lower, but still meaningful p-coefficients were obtained for transitions between sentences and clauses, indicating that these levds are also sensitive to differences between texts. On the other hand, there is almost no effect on pause time variance for the 'lower• textual categories 'constituents• and 'words·. In other words, PP1p cannot be predicted on the basis of PP cp' Therefore it must be concluded that the intra-class data do not provide any support for the covariance-model (5).

Conceptual and linguistic processes in text production 3:13

Granting location types to mark the difference between 'conceptual' and 'linguistic' processing, these results thus suggest that from a temporal point of view linguistic processing goes on independently from text-specific factors, such as text length or complexity, which contradicts the interaction-hypothesis. The clear break between the level of clauses on the one hand and constituents on the other, indicates that text-related problems are solved primarily at the paragraph-level, and to a somewhat lesser extend at the sentence- and clause-level, but they leave processing at the constituent and word level unaffected. At this stage, however, there is not much of a base to be anymore specific about what goes on at the intermediate levels of sentences and clauses. This issue will be addressed in more detail in the next section. 4·3 Interactions between conceptual and linguistic processes within texts

The former section informed about how pause times at different levels vary along with between-text differences. This section considers a second kind of external factor: how do pause times at different levels covary or correlate with each other within texts? Let us refer to these possible covariations as inter-level relations between various levels of processing. There are three logical possibilities as to how pauses at different locations may covary/correlate: They may correlate positively, negatively, or there may be no significant correlation at all. Let us try to cast these possibilities in processing terms. The first possibility entails that if, relative to the average length for some superordinate level Lx' pause time at X increases, so will pause times at a subordinate level Lx-I.l, ... n· If a text producer uses more time than average to plan a paragraph, he will also use more time to plan the sentences within that paragraph, the clauses within that sentence, and so on. The second possibility represents the reversed situation: if, relative to the average length, pause times increase at level L"' pause times will decrease at a subordinate level Lx-I.l•... ,· If a paragraph is thoroughly planned there will be less need to plan the constituing sentences and so on. So these first two possibilities imply interactions between conceptual and linguistic planning, for in both cases some aspect ofPP1P (its variance) is predictable on the basis of the same aspect of PPcp' The third logical possibility entails that there is a zero or non-significant correlation between variances in pause times at levels L" and L" _ I.l•... n·= deviations from the fixed parameters for different levels are randomly distributed. Cast in processing terms: planning load at the paragraph level may vary, but such variances will have no predictable effect on planning load at lower levels. This statistical possibility is assumed to be the

324

Joost Schilperoord

counterpart of the interaction-hypothesis which predicts either positive or negative correlations. To test the hypothesis, we first need to estimate the inter-level covariances/ correlations. Covariances between each pair of levels X and Y can be estimated by using the equation (7). (7) Covxy = I(Xojk- Xoo.c) {yOjk- Yook}/n

Equation (7) states that covariances can be estimated by subtracting the overall means for pause times at levels X and Y from the text-specific means for the same levels for all120 texts, then to multiply the differences, then to summerize the multiplied differences, and finally to divide them by the number of observations (i.e. 120). Correlations between X andY can then be estimated by using equation (8). (8) Corxy

=Covxy I lV11X ,.lP...,.

Table 3 summarizes the results of the covariance/correlation analysis. Table 3. Correlations (in italics) and covariances between levels of processing. In order to be significant covariances must be 1.96 as large as the associated standard error estimates (given between brackets) Estimates par

sen 1.77 (.43) .801

sen

cia

con

.978 (.25) .701 .241 (.053)

.201 (.074) .418 .071 (.016)

.942

cia

.808

.162 (.039) .825

con.

wor .025 (.035) .145 .014 (.007) .467 .010 (.004) .497 .007 (.002) 1.00

Note first that only positive correlations are found for each pair oflevels X and Y, no matter how 'far' apart these levels are. For example, variances at the level of paragraphs correlate positively, although only slightly (.145), with fluctuations at the word-level. Generally speaking, if pause times at some superordinate level increases with respect to the overall mean, so will pause times at all subordinate levels. Correlations increase if the levels under consideration become closer to each other. Second, Table 3 shows correlations to be highest at adjacent levels of processing - the matrix therefore represents a so called a


simplex structure (Schilperoord 1996). However, although positive correlations between all levels have been obtained, it cannot be immediately concluded that they support the interaction hypothesis. In order to test this hypothesis, we have to explore more precisely the amount of variance at lower levels that can be explained on the basis of variances at higher levels. To see why, consider a simplified model of processing with only three levels, A, B and C. Let it further be assumed that conceptual processes take place at level A, and linguistic processes at level C, and that level B represents some intermediate level of processing. If conceptual processes and linguistic processes would indeed interact, then the interaction hypothesis predicts a 'jwnp' between levels in that processing at level A affects processing at level C, regardless ofwhat happens at level B. This situation would be the case if it can be shown that a considerable amount of level-C variance can be explained by level-A variance. If, on the other hand, conceptual processes have no impact on linguistic processes, the relative contribution of A-variances in explaining C-variances will be insignificant. This would indicate that the impact of A-level processes on lower levels is subdued, so to say, at the intermediate level. Hence, the following two models were tested: Model (1): A+ B-+ C Model (2): B-+ C Model ( 1) represents the statistical counterpart of the interaction hypothesis, whereas model (2) represents the autonomy hypothesis. Analysing these models can be done by constructing a path-model which specifies the variables of interest (that is, pause time variances at five location types) and all possible relations that may hold between them. However, both the simplex structure shown in Table 3, and the theoretical considerations given so far, indicate that only the maximally restrictive model should be analysed, that is, the model that allows significant relations to exist only between adjacent levels of processing.5 The statistical test itself proceeds by estimating the so-called maximwn likelihood parameters, i.e. the set of regression weights holding between the variables. The test aims at fitting the hypothesized model with the data (i.e. the covariance-matrix shown in Table 3). If differences between the model and the observed data are small, the model is said to fit the data. The fit of the model is represented as a goodness-of-fit index (gfi), expressing the amount of explained covariance. Obviously, since the minimally restrictive model will fit the data completely (because all possible relations between levels are allowed), the

326

Joost Schilperoord

analytical question boils down to an estimation of the maximally restricted set of parameters which nevertheless fit the observed scores reasonably. Therefore, we started analysing a model with only 4 regression parameters. The results of the data-fitting test are presented in Table 4. Table 4. Estimates of model parameters for adjacent levels Xn' Xn- 1, and residual covariances Estimate par

sen

cla

con

wor

1.0

.708 .498

.891 .205

.683 .533

.732 .478

p1,1-l 'l'x

gfi = .985

The gfi estimate in Table 4 indicates that the maximally restricted model fits the data extremely well as it accounts for .985 of the covariance between location types. Enlarging the model with six additional parameters only adds an extra .015 explanatory power to the gfi, which is next to negligible. Let us therefore concentrate on the regression weights ~x.x-l" As the processing direction can be assumed to be top-down6 , the variance at the paragraph level is entirely determined by the residual term 'l'x· The regression weights ~x.x-l (with x = 2, .., 5) express the strength of the connections between estimates for all pairs of adjacent levels. Residual variances for all levels are indicated by 'l'x (with x = 2, .., 5). These parameters express the amount of unexplained variance at each level granted the values for regression weights. As Table 4 shows, the variance at the clause level is particularly well-explained, as there is only .205 residual variance (hence, only 21 o/o remains unaccounted for). Taken together, the regression weights and the residual parameters allow us to estimate the strength of remote connections by applying equation (9). (9) ~x.x-n = ~n-l,n''l'n ·~x.n- t•'l'n- I

Equation (9) allows us to estimate the amount of explained variance at, say, the level of constituents by considering the joint variances at the level of paragraphs and sentences, and oppose this estimate with the one concerning the amount of explained variance at the constituent level by considering the intermediate level of clauses. It then turns out that paragraphs- and sentence-variances only account for 10% variance at the constituent level, whereas clause-variances account for over 36% for constituent-variances. 7 Phrased in terms of processing, what happens at the conceptual level (paragraphs, sentences) does not tell


us much about what happens at the lower, linguistic levels. In other words, processing at these levels seems to go on relatively independently, which can be taken to support the no-interaction hypothesis. However, it would be wrong to assume a sudden switch between levels of processing. As the data indicate, the level of clauses seems to occupy an 'intermediate' level of processing: The clause may constitute such an intermediate level of processing because its structure interacts with the process of assembling conceptual fragments, whereas at the same time it provides increasing constraints on linguistic processing as the process proceeds. Therefore, it represents a level of processing 'below' which influences of conceptual processes are subdued, so to speak, that is, no longer present at the levels of linguistic processing. Hence, the correlation and regression data tested here strongly contrast the validity of the interaction hypothesis. As it seems, processing at the conceptual level goes on more or less independently from processing at the 'lower' linguistic level. The data imply an 'adjacency model' of processing: no 'jumps' across levels occur. Moreover, it seems that the level of clauses acts as an intermediate level 'between' conceptual and linguistic processes. Processing at this level is affected by higher level processes, and affects lower level processes (see also Schilperoord 1997a). 4·4

The time-dependent nature of conceptual and linguistic processes

This section considers yet another factor accounting for pause time variances: the variable 'processing time'. In research on writing and discourse production, the variable 'Time' has long been neglected as a source of variance in processing. Only recently its importance has been acknowledged (Van den Bergh & Rijlaarsdam 1999; Schilperoord 1995, 1996). Van den Bergh and Rijlaarsdam mention two ways in which the variable 'time' might aVect our current theoretical view on text production. First, they stress that cognitive activities might serve widely dNerent purposes if occurring at diVerent points in (real) writing time. Furthermore, they have shown that the relationship between cognitive activities and text quality is not Wxed, but in stead timedependent. A process such as 'reading a writing assignment' correlates positively with text quality if it occurs during the early stages in writing, but negatively if it occurs during later stages Although such a Wnding runs the risk of belabouring the obvious, it does quite nicely illustrate the relevance of taking the time dimension seriously. Moreover, this approach has revealed timedependent patterns in cognitive processing that are less obvious.

3:z.8

Joost Schilperoord

The real time approach taken here deviates slightly from that of Van den Bergh and Rijlaarsdam in that we only take into account pause patterns at various levels of processing, whereas in the study referred to above, the data base consisted of tagged thinking-aloud protocols. The strategy will be to find out how the variable 'time' might affect processes at different levels, and how such affects might differ from each other. If 'time' is the variable X of model (4) (see Section 4.1), what effects on conceptual and linguistic processes can then be expected, and how can these effects be tracked? The crucial assumption behind the time-approach to the study of text production is that in the course of production the cognitive 'circumstances' change gradually in time as more text is already produced. This might lead to different purposes served by otherwise identical processes, but also to varying interactions between different processes. How does this flesh out in the case of routine letter production? Arguably, routine letters are called 'routine' because different exemplars share much of their contents with each other. This will be the case especially at the beginnings and endings of such letters. These parts often contain formulaic language, as they contain standardized opening and closing statements, and fixed pragmatically oriented formulae. What makes a letter unique, on the other hand, is its central part, the part that contains the actual message to the addressee, and the part that constitutes the ultimate reason for the writer to produce the letter. Modelled as a function of production time, we then see a alternation between 'fixed' and 'new' parts within these letters. If the above assumptions cut some ice, this might have an impact at the level of conceptual processing, but not necessarily at the level of linguistic processing. For example, we might expect pause times at paragraph junctures to increase towards the central letter parts, and then to decrease towards the end of the letter. Such an assumption is based on the idea that producing the 'new' parts ofletters will be cognitively more effortful than producing the more or less fixed parts. However, former analyses have pointed out that such effects do not carry over to the 'lower', linguistic levels of processing. Content-related problems are mainly solved at the 'upper' levels and they do not significantly affect lower-level processing (see Section 4.3). So, if some way can be found to model the influence of the 'time' variable, we expect time-bounded influences to be present at the conceptual level, but not at the linguistic level. That is, time-bounded influences at this level cannot be predicted on the basis of timebounded influences at the conceptual level. On the other hand, the interaction hypothesis would, somehow, predict

Conceptual and linguistic processes in text production

time-bounded effects at all levels to correlate. Although the aforementioned considerations, together with the results of previous analyses seem to provide good reasons not to assume such correlations, they may in fact appear after all, that is: it may be the case that time-bounded effects at the level of conceptual processes allow us to predict the nature of such effects at the lower levels. The hypothesis tested here therefore specifically entails predictable effects of the variable time to be present at level X, given the observed effects at some superordinate level Y. In order to test this time-version of the interaction hypothesis, for each pause the number of words was counted that was produced in one spurt immediately following that pause. These counts were subsequently combined with the location ofthe pause, the latter indicating the level of processing. These so-called p(ause)/a(rticulation) rates will undoubtedly display considerable variation, and the analytical enterprise now is:

1. to see to what extent such variances can be explained away by the variable 'time'; 2. to see whether these variances covary (if indeed they do, this would support the interaction hypothesis). We started estimating the p/a variances that occurred between paragraphs, between sentences, between clauses, between constituents and between words respectively. Figure 3 shows the results.

Figure 3. Mean and residual pause/articulation ratio's for five type oflocations (1 =par, 2 =sen, 3 =cia, 4 =con, 5 = wor)

329

330 Joost Schilperoord

As could be expected from earlier endeavours, the p/a rates for all location types indeed show considerable variation. In terms of cognitive processing, the data moreover indicate that the 'ease of processing' varies per level. Variances at the level of paragraphs by far exceeds variances at the levels of sentences, clauses, constituents and words respectively. Now, what happens if these variances are modelled as a function of (real) production time? This can be estimated by aligning the various paragraphs, sentences, clauses, constituents and words according to their linear order of production (and according to the order in which they appeared in the final text). The results of this analysis are represented graphically in Figure 4.

'·' 3.1

•a

'

"a a a

a

"

Figure 4. Residual pause/articulation ratio's for five type oflocations, modelled as a function of real production time

Real production time is projected on the horizontal axis of Figure 4, whereas the estimated variances in p/a ratios for each location type are projected on the vertical axis. The (somewhat stylized) curves are formed by projecting each individual pause at a particular location type on the time track. As extremely long texts (over 600 seconds real production time) are somewhat exceptional, the curves for paragraphs and sentences gradually become less dense because fewer observations could be made here. The upper regions of Figure 4 shows the p/a rates for paragraph pausing as a function of (real) production time. The curve represents almost precisely what we had expected to find: processing efforts increase towards the middle sections of the process/text, in order then to decrease towards the end of the process/text. It can therefore be concluded that the variable 'time' does have an

Conceptual and linguistic processes in text production

impact on processing at the level of paragraphs. Assuming this level to concern mainly conceptual processes, it can be concluded that at this level processing is indeed influenced by the time variable. The same tendency, although to a much slighter extent, can be observed for p/a rates at the sentence level, displaying a slight, but still significant curve. However, as can be easily gleaned from Figure 4, time-bounded effects are markedly absent at the clause-level, the constituent-level and the word-level. These curves are hardly affected by the variable 'time' and only show a nonsignificant curve: they therefore ought to be taken as representing straight lines. This means that the p/a ratios are stable regardless of their position on the time-track. These results imply that no correlations exist as to the effect of time on conceptual processes on the one hand, and linguistic processes on the other (although, admittedly, this conclusion is based on observations done 'by the eye'). Time has an effect on conceptual processing, but there is no way of predicting, on that basis, what the effects will be at the linguistic levels of processing. All in all, the results do not provide any kind of support for the interaction hypothesis, and are therefore consistent with the results of the previous analyses.

s.

Discussion

This paper has sought to find out whether empirical evidence can be produced, based on an analysis of pausing patterns in text production, supporting the claim that in text production conceptual and linguistic processes interact, i.e. the interaction-hypothesis. The research strategy that we employed was to select three production-'external' circumstances for which it could be argued that they would cause variation in conceptual and linguistic processes. We first argued that temporal parameters such as 'pause length' and 'pause location' are responsive to the distinction between the two kind of processes. The three external variables explored were: differences between texts, variation in pause lengths for different location types in texts, and the variation due to the variable 'real production time'. For all three variables it could be demonstrated that they produce distinct pause patterns for both kinds of processes. Distinct in that one particular pattern could not be predicted on the basis of the other caused by the same external variable. These results thus do not provide empirical evidence favouring the interaction hypothesis.

331

332

Joost Schilperoord

Does this mean the no-interaction hypothesis to be the correct one? Obviously, given theoperationalization of the notion of'interaction' adopted in this study, it is altogether plausible to conclude that this hypothesis best accounts for the data obtained. For example, given the assumption that cognitive load will increase towards the middle parts of the texts produced, and decrease towards the concluding parts, it could be clearly shown that this only affects conceptual processes, but leaves linguistic processes more or less unaffected. Hence, the latter kind of process, arguably concerned with lexical search and syntactic alignment, does not respond in any predictable way to increased cognitive load. This is not to imply that no processing trouble pops up at this level, but these troubles cannot be predicted on the basis of higher-level conceptual processing features. Granting this to be the main lesson that can be learned from the analyses reported, we should, however, bear in mind some reservation. As is always the case with empirical data- they hardly ever allow one to draw ali-or-nothing conclusions. It should be stressed, first of all, that the conclusions drawn here are necessarily restricted to the kind of writing processes that was studied: highly routinized writing by expert adults. Flower and Hayes' multi-representation thesis might very well typify writing processes requiring much more problem-solving and planning activity. Here it seems plausible to assume that the ease oflocal processing is facilitated to the extent that content and structure of the to-be-written text is pre-planned. This line of thought is also suggested by Welschen's study on the 'cognitive balance' between higher and lower order processes in children's writing (Welschen 1982). Second, we have seen that the no-interaction hypothesis particularly holds for the highest and lowest level of processing: paragraphs and words. At the intermediate levels, however, we do find some interaction between levels. Especially the analysis reported in Section 4.3 points out that the level of clauses appears to take an immediate position in between 'higher' and 'lower' levels of processing. The general impression one gets is that conceptual and linguistic processes 'merge' so to speak at the clause level. Clauses represent a level of processing that is affected by higher level conceptual processes, and that affects lower level linguistic processes. Clause structures interact with the process of retrieving content from memory and assembling conceptual fragments as its structural characteristics provide increasing constraints as the process proceeds, whereas, on the other hand, clear interactions between the clause level and lower levels of processing were obtained. Both text-related differences, and differences between paragraphs may cause a text producer to put extra effort


into planning the contents and structure of texts. But, these conceptual problems are largely solved at the levels of paragraphs and sentences: They do not percolate to the 'lower' processing levels of constituents and words. This does not mean that finding lexical items to express ones thoughts, or aligning them within syntactic frames goes on effortlessly- it only means that problems that emerge at these levels of processing are solved independently from problems encountered at higher levels.The intermediate position occupied by the level of clauses is further substantiated by an analysis of the development of pause times within clauses, reported in Schilperoord (1997a, b). Here it is demonstrated that both the average pause length and pause time variances gradually decrease within clauses, that is, by each successive pause within a clause structure. This phenomenon does not show up within, say, sentences. We can account for these findings by assuming that, at the conceptual level, producing texts basically proceeds in a clause-by-clause manner (see also Ford 1982; Ford & Holmes 1978 ). Decreasing pause times within clauses then indicate conceptual processing to get increasingly constrained by grammatical factors, such as the obliged verb-final position in dependent clauses in Dutch. Admittedly, these are rather speculative statements. Indeed, what happens at the level of paragraphs has little to do with what happens the level of words, rightly so predicted by the no-interaction hypothesis. But much remains to be learned about the intermediate levels, and it seems wise not to assume a strict division between conceptual and linguistic processes.

Notes • I would like to thank Ted Sanders and Wilbert Spooren and three anonymous reviewers for their comments on an earlier version of this paper. I especially thank one of the reviewers for her or his highly valuable advice with regard to the presentation of the materials in this paper, and for preventing me from a major theoretical pitfall. I can only hope my revisions are as good as her or his advice. If I have understood his book correctly, Fodor ( 1983) would say that these two modules can no longer be considered to be informationally encapsulated.

1.

Although this may be a mere artefact of the fact that the Flower and Hayes model lacks the detailed precision of Levett's model.

2.

3· See for a detailed summary: Schilperoord (1996).

4· Apart from this analysis, we also estimated the average durations of pauses excluding zero-pauses. In ordinal terms, the results of this analysis were similar to those including zero-pauses (see Table I), that is the order of ranking remained in tact (Schilperoord 1996,

334

Joost Schilperoord

92) . For this reason, Table 1 only reports the 'true' estimates, hence, pause time averages

without taking zero-pauses into account.

s.

For details concerning the construction of path-models, see Schilperoord ( 1996, pp. 100

ft). 6. This does not appear to me to be a far reaching claim. It would, after

all, be odd to assume that a writer begins retrieving lexical elements before he has decided on what information should be expressed in, say, a paragraph. 1· The full computations can be found in Schilperoord (1996).

References Ariel, M. ( 1988). Referring and accessibility. Journal ofLinguistics, 24, 65-87. Bereiter, C., & Scardamalia, M. ( 1987). The psychology ofwritten composition. Hillsdale NJ: Erlbaum. Bergh, H. van den, & Rijlaarsdam, G. (1999). The dynamics of idea generation during writing: an on-line study. In M. Torrance & D. Galbraith (Eds.), Knowing Mtat to Write. Conceptual Processes in Text Production (pp. 99-121). Amsterdam: University Press. Boomer, D. S. (1965). Hesitation and grammatical encoding. Language and Speech, 8, 148-158. Butterworth, B. ( 1975). Hesitation and semantic planning in speech. Journal ofPsycholinguistic Research,~ 75-87. Chafe, W. (1987). Cognitive constraints on information flow. In R. Tomlin (Ed.), Coherence and grounding in discourse (pp. 21-51 ). Amsterdam: Benjamins. Chafe, W. ( 1994). Discourse, consciousness, and time. The flow and displacement of conscious experience in speaking and writing. Chicago: Chicago University Press. Cochran, W. G. (1977). Sampling techniques, third edition. New York: John Wiley & Sons. Flower, L., & J, R. Hayes ( 1984). Images, plans and prose; the representation of meaning in writing. Written Communication, 1,120-160. Fodor, J. A. ( 1983). The modularity ofmind. Cambridge, MA: MIT Press. Ford, M. ( 1982). Sentence planning units: implications for the speaker's representation of meaningfull relations underlying sentences. In J, Bresnan (Ed), The mental representation ofgrammatical relations (pp. 797-827). Cambridge MA: MIT Press. Ford, M., & Holmes, V. M. (1978). Planning units and syntax in sentence production. Cognition, 6, 35-53. Garrett, M.P. (1980). Levels of Processing in Sentence Production. In B. Butterworth (ed.) Language Production. Vol I: Speech and Talk. (177-220). London: Academic Press. Gee, J. P. ( 1986). Units in the production of narrative discourse. Discourse Processes, 9, 391-422.

Conceptual and linguistic processes in text production 33S

Goldman-Eisler, F. (1958). Speech production and the predictability of words in context. Quarterly Journal ofExperimental Psychology, 10,96-105. Goldman-Eisler, F. (1968). Psycholinguistics: experiments in spontaneous speech. New York: Academic Press. Hayes, J. R., & Flower, L. ( 1980). Identifying the organization of writing processes. In L. Gregg, & E. Steinberg (Eds.), Cognitive Processes in Writing (pp. 3-29). Hillsdale, New Jersey: Earlbaum. Henderson, A. I. (1974). Time patterns in spontaneous speech- cognitive stride or random walk? A reply to Jaffe, et al. 1972. Language and Speech , 17, 119-125. Henderson, A. 1., Goldman-Eisler, F., & Skarbek, A. (1966). Sequential temporal patterns in spontaneous speech. Language and Speech. 9, 207-216. Holmes, V. M. (1988) Hesitations and sentence planning. Language and Cognitive Processes, 3, 323-361. Just, M.A., & Carpenter, P. A. ( 1987). The psychology of reading and language comprehension. Newton, MA: Allyn and Bacon. Kowal, S., & O'Connell D. C. (1987). Writing as language behaviour: myths, models, methods. In A. Matsuhashi (Ed.), Writing in real time. Modeling production processes (pp. 108-133). Norwood NJ: Ablex Publishing Corporation. Levelt, W. J. M. (1983). Monitoring and self repairs in speech. Cognition, 14,41-104. Levelt, W. J. M. ( 1989). Speaking: from intention to articulation. Cambridge, MA: MIT Press. Longacre, R. (1979}. The paragraph as a grammatical unit. InT. Givon. (Ed.), Syntax and Semantics, voL 12: discourse and syntax (pp. 115-134). New York: Academic Press. Lounsbury, F. G. (1954). Transitional probability, linguistic structure, and systems of habit-family hierarchies. In C. E. Osgood, & T. Sebeok (Eds.), Psycholinguistics: a survey oftheory and research problems (pp. 88-126}. Bloomington: Indiana University Press. Maclay, H., & Osgood, C. E. (1959) Hesitation phenomena in spontaneous English speech. Word, 15, 19-44. Matsuhashi, A. ( 1981 ). Pausing and planning: the tempo of written discourse production. Research in the TeachingofEnglish, 15, 113-134. Pool, E. van der (1995). Writing as a conceptual process. A text-analytical study of developmental aspects. Doctoral dissertation, University ofTilburg. Sanders, T., Janssen, 0., Pool, E. van der, Schilperoord, J., & Wijk, C. van (1995). Hierarchical structures in writing products and writing processes. In G. Rijlaarsdam, H. van den Bergh, & M. Couzijn (Eds.), Theories, models and methodology. Current trends in research on writing (pp. 473-493). Amsterdam: UVA-press. Schilperoord, J. (1995). The distribution of pause time in written text production.ln G. Rijlaarsdam, H. van den Bergh, & M. Couzijn (Eds.), Theories, models and methodology. Current trends in research on writing (pp. 263-275 ). Amsterdam: UVA-press. Schilperoord, J. ( 1996). It's about time: temporal aspects of cognitive processes in text production. Amsterdam/Atlanta: Rodopi.

336 Joost Schilperoord

Schilperoord, J. ( 1997a) Temporele modificatie in clauses; een pauze-analytische studie naar tekstproductie (Temporal modification in clauses; a pause-analytical study on text production). In H. van den Bergh, D. Janssen, N. Bertens, & M. Damen (Eds.), Taalgebruik ontrafeld (Unraveling language use) (pp. 263-275). Dordrecht: Foris Publications. Schilperoord, J. ( 1997b ). Eenheden van tekstproductie; de clause als domein van temporele modifietltie (Units in text production; the clause as domain for temporal modification). Lecture presented at the founding colloquium of the Utrecht institute of Linguistics OTS, 13-2-1997. Schilperoord, )., & Pool, E. van dcr (1997). Schrijfpauzes en tekststructuur; naar ecn cognitief model van real-time tekstproductie (Writing pauses and text structure; towards a cognitive model of real-time text production). Gramma/TIT, 5, 119137.

Schilperoord, J., & Sanders, T. (1997). Pauses, Cognitive Rhythms and Discourse Structure; An empirical study of discourse production. A. Liebert, G. Redeker, & L. Waugh (Eds.), Discourse and perspective in Cognitive Linguistics (pp. 247-269). Amsterdam/Philadelphia: Benjamins. Welschen, A. (1982). Formulecrvaardigheid en de cognitieve balans bij bet schrijven (Formulating skills and the cognitive balance in writing). Tijdschrift voor Taalbeheersing, 4, 131-162. Wijk, C. van (1987). Spetrking, writing and sentence form. Doct. Diss., University of Tilburg.

CHAPTER

13

Subordination and discourse segmentation revisited, or: Why matrix clauses may be more dependent than complements Arie Verhagen University of Leiden

1.

Introduction

All approaches to the structural analysis of texts and discourse have to make assumptions about the smallest units out of which larger pieces of discourse are constructed. A plausible first candidate for the status of "minimal discourse segment" is the grammatical clause, so it comes as no surprise that from the start of an approach such as Rhetorical Structure Theory (Mann & Thompson 1988), this assumption has actually been put forward. A simple text consists of a series of simplex clauses, connected by particular conceptual relations making the series of clauses into a coherent text. Thus one naturally wants to take at least all main clauses of a text as minimal segments. Complications arise when other clauses than main ones are also taken into consideration; sometimes one wants to assign such a clause the status of segment, sometimes not. This has also been evident since the beginning of RST. The problem I want to address in this paper is how to give a principled account of the relationship between grammatical subordination on the one hand, and the segmentation of texts into their minimal units on the other. Let me begin by reviewing explicitly the motivation for denying certain non-main clauses the status of discourse segments. Consider the following example: ( 1) They left early; they absolutely wanted to be on time. This mini-text consists of two segments. The conceptual relation connecting them and holding them together as a textual whole is some sort of causality; a

338 Arie Verhagen

competent reader will know that the desire to reach a destination on time provides a motivation for leaving early, and thus interpret the contents of the second segment as actually providing the cause of the event described in the first. Obviously, the number of discourse segments corresponds exactly to the number of finite grammatical clauses in ( 1). 1 Now consider example (2). (2) They left early(,) because they absolutely wanted to be on time.

If (1) is considered to be a text, consisting of two segments, then (2) is one as well. There are also two separately identifiable propositions, and the conceptual rdationship between them is the same as in ( 1). The difference is that the relationship is explicitly marked as causal in (2), whereas (1) lacks such a marking; thus although the interpretation of (2) can therefore be said to be more constrained than that of ( 1), there is no reason to assign it a fundamentally different status as a text. Calling ( 1) a text consisting of two segments and (2) a single clause text, for example, would dearly miss a generalization. As a matter of fact, it is intuitions like these that motivated the idea that this type of 'clause combining' can actually be regarded as the grammaticalization (conventionalized structural expression) of discourse relations (Matthiessen & Thompson 1988). However, the same kind of considerations (concerning conceptual interclausal rdationships) also leads to the conclusion that not all clauses should be considered to constitute discourse segments. Consider examples (3) and (4). (3) They left early; it is essential that they be on time. ( 4) They left early; they think that in that way they will definitely be on time. With respect to cases like these, one also wants to make it possible to state a generalization: there is a conceptual rdationship of causality in ( 3) connecting its component sentences in the same way as is the case in (1) and (2), and the same is true for (4). This requires one to assume that both (3) and (4) contain two segments, but each contains three finite clauses (the part following the semicolon consisting of a main and a subordinate clause). Therefore, as early as in Mann and Thompson's original RST -proposal, clauses that functioned as subjects (cf. (3)) or complements (cf. (4)) were denied the status of discourse segments. Matthiessen and Thompson ( 1988) called the type of subordination exemplified in (2) "clause combining", while the type of subordination in (3) and (4) was characterized as "embedding"; only the former cases are to be considered as grammaticalizations of discourse relations, while the latter are

Subordination and discourse segmentation revisited 339

properly viewed as actual constituents of their host clauses. Although the notions are not really defined in a fully explicit manner, it is intuitivdy clear what the authors are trying to get at, and this distinction also turns up in later approaches to discourse structure (Pander Maat 1994, p. 3036; Sanders 1992, p. 115/6; Sanders & van Wijk 1996, p. 126/7). However, it should also be noted that the exceptional status of subject and complement clauses is not really explained in this way. This becomes even more problematic when one realizes that there is minimally one more exceptional type of clause: restrictive relatives. Again, the motivation for assigning these a different status is not formulated very explicitly, but it can be made sufficiently clear. Consider examples (5), containing a restrictive relative clause, and (6), with a nonrestrictive one. (5) These schools all appear to have relatively many students who grew up in culturally deprived families. (6) They shouted at the waiter, who so far did not seem to have noticed them. One does not want to view (5) as a text consisting of two segments primarily because there does not seem to be a conceptual relationship between the two clauses making them into a textual whole (i.e. the clause just specifies some property of its head noun, restricting its denotation; cf. below). In other words, one does not want to divide (5) into two segments as indicated in (5)'. (5)'

~a.

These school all appear to have relatively many students b. who grew up in culturally deprived families.

On the other hand, the non-restrictive relative clause in (6) does have some conceptual relationship with the matrix clause (beyond mentioning a property of its referent): a plausible interpretation could be that the situation mentioned in the relative clause specifies the reason for their shouting at the waiter. Thus one would want to divide (6) into two discourse segments between which a textual relation (in this case, of causality) may be construed; cf. (6)': (6)' a. They shouted at the waiter, b. who so far did not seem to have noticed them. What is it that restrictive relative clauses, subject and complement clauses have in common which makes them exceptions to the 'rule' that discourse segments correspond to grammatical clauses? This is the question that has to be an-

340 Arie Verhagen

swered in order to make a start with an explanatory account of the relationship between the two. In the remainder of this paper, I want to propose a number of hypotheses intended as steps in that direction; as my point of departure I will take the analysis proposed in Schilperoord and Verhagen (1998).

2.

Conceptual independence and discourse segmentation

Working on the basis of analyses of {non- )restrictive relatives that were developed independently from the issue of discourse segmentation (Daalder 1989; Verhagen 1992, 1996a), Schilperoord and Verhagen (1998) propose a condition on discourse segmentation that can briefly be stated as follows: (7)

Condition on discourse segmentation (conceptual independence): "If a constituent of a matrix-clause A is conceptually dependent on the contents of a subordinate clause B, then B is not a separate discourse segment" (cf. Schilperoord & Verhagen 1998, p.lSO ).

This condition utilizes the idea that a matrix structure may be dependent for its conceptualization on some subordinate structure {cf. Langacker 1991, p. 436), and that as a consequence, the subordinate structure involved cannot be a separate discourse segment. Thus, it is not so much conceptual dependence of the subordinate structure that makes it inappropriate as a discourse segment, but rather its role in making its matrix structure conceptually independent. This 'shift' is crucial, as we will see shortly. But let me first illustrate the condition by showing how it applies to relative clauses. Consider the restrictive relative in {5) once more: (5) These schools all appear to have relatively many students who grew up in culturally deprived families. Notice that the conceptualization of the referent of students is crucially dependent on the contents of the relative clause. The sentence does not say that the schools have relatively many students {and that these grew up in culturally deprived families), but rather that relatively many students grew up in such families. In (6) on the other hand, the conceptualization of the referent of waiter is not crucially dependent on the contents of the relative clause: (6) They shouted at the waiter, who so far did not seem to have noticed them.


In the non-restrictive interpretation, the denotation of the waiteris determined independently of the relative clause, which then provides some additional information; this sentence does mean that they shouted at the waiter, and that he did not seem to have noticed them so far. Thus the explanation is that a restrictive relative clause is required to complete the conceptualization of some part of another clause, and hence cannot function as a separate discourse segment. Schilperoord and Verhagen ( 1998) claim that the same condition explains the exceptional role of subject and complement clauses, i.e. in so far as it is exceptional. The point is that the usual formulation of the exception is not fully adequate. In Mann and Thompson's (1988) original formulation, the claim was that a subject or complement clause was to be considered as "part of its host clause". As long as we take only relatively simple cases of embedding into consideration, that procedure gives the desired segmentation, but problems arise when we apply it to more complicated cases. Such complications actually abound in the material used for the research reported in Schilperoord (1996); (8) is a typical example. Te uwer informatie merk ik nog op dat client voorziet dat het niet eenvoudig zal zijn om snel ander werk te vinden. b. Daarbij komt dat zijn echtgenote zwaar gehandicapt is en dat hij een gezin heeft te onderhouden. a. For your information I .mzg that my client anticipates that it l!d1l not be easy to find another job fast. b. To this mmml. be added that his wife u seriously disabled and that he }uy a family to care for".

(8) a.

Fragment (8) consists of two sentences (marked a and b); it contains 6 finite clauses (indicated by the underlined finite verbs). Sentence (a) actually consists of a series of embedded clauses, and it is quite conceivable that recursive application of "Mann & Thompson's rule" could handle it, producing the identification of (a) as a single segment, as seems desirable. The real problem is exemplified by sentence (b). Straightforward application of Mann & Thompson's rule would result in it being characterized as a single segment: each of the two subordinate clauses is a subject clause, to be taken as a part of its host clause. However, the segmentation of(8) into two segments (corresponding exactly to the two full sentences a and b) seems undesirable because it makes it impossible to capture the fact that the writer of this fragment, a

342. Arie Verhagen

lawyer, adduces three arguments in favor of his clients position: the problem of finding another job, the health condition of his wife, and the fact that there is a family to be cared for. The essence of the first of these is contained in the single right-most embedded clause of (a), while the other two points are presented in the two subject clauses of (b); as a result of the way Mann & Thompson's rule is formulated, it is not possible to recognize the fact that there are actually two points being made in (b). 2 The condition proposed in (7) actually is capable of making the relevant distinction. The reason is that the relevant property of conceptual dependence is attributed to the matrix clause rather than to the subordinate one. Notice that the phrase (in Dutch) Daarbij komt ("To this should be added"), as an instruction to add certain pieces of information to a previously established one, is not conceptually complete, in a sense not even interpretable, without the information provided in the subordinate clause. The point is not that this information could only be provided by a clause (the ·subject slot' could also be filled by a noun phrase, for example nog iets anders, "something else"); rather, the point is that in this case, it is a subordinate clause that fulfills this necessary function of making the matrix conceptually independent, so that the subordinate clause does not constitute a separate discourse segment. Now by the same token, one complete clause always suffices for creating a conceptually complete message; the unit of a matrix and the first subordinate clause is never conceptually dependent on a second one. Consequently, all further subordinate clauses can be properly characterized as separate discourse segments,1 so that fragment (8) may be divided into the three segments indicated in (8)": (8)" a.

Te uwer informatie merk ik nog op dat clil!nt voorziet dat het niet eenvoudig zal zijn om snel ander werk te vinden.

For your information I note that my dient anticipates that it will not be easy to find another job fast. b. Daarbij komt dat zijn echtgenote zwaar gehandicapt is

To this should be added that his wife is seriously disabled c.

en dat hij een gezin heeft te onderhouden.

and that he has a family to care for. Interestingly, this does not only provide us with a principled account of a segmentation that is in accordance with our intuitive understanding of such fragments, it also appears to specify the boundaries of planning units in actual language production (see Schilperoord 1996, Chap. 6, and Schilperoord 1997).


Returning to the relationship between condition (7) and Mann and Thompson's original procedure, we can note that certain clauses that fulfill the syntactic function of subject or complement must in fact be allowed to be assigned the status of separate discourse segments. Mann & Thompson's rule could not accommodate such cases, but condition (7) does, while preserving the idea that a matrix forms a discourse unit together with a single complement or subject clause; it furthermore provides a generalization over these clauses and the restrictive relatives. I therefore consider it a substantial part of a more explanatory account of the relationship between grammatical and discourse structure: Only a relationship of conceptual dependence between syntactically related clauses is a sufficient condition preventing them from constituting separate discourse segments. Still, there are some remaining questions, in particular: (a) What is the reason that certain matrix clauses are not conceptually independent? Do such constructions have anything in common that relates to a specific discourse function, distinct from the discourse function of adverbial clauses (the cases of "clause combining" in terms of Matthiessen & Thompson 1988)? (b) How do subject and complement clauses differ from restrictive relatives, such that the latter never constitute separate discourse segments? (c) How can we avoid the grammatically impossible conclusion, suggested by the segmentation in (8(, that in fragment (8) a main clause (segment b) and a subordinate one (segment c) are being coordinated? I believe that there are interesting, interrelated answers to these questions, which will allow us to further deepen our understanding of relationships between grammar and discourse, in particular of the discourse function of the grammatical phenomenon known as complementation.

3·

Dimensions of text interpretation: (lnter)subjectivity

A fundamental aspect of the human capacity for using language is the ability to recognize other entities as essentially like oneself, and to take another person's perspective as one that could be one's own. For one thing, the whole idea of intentionally producing utterances to be recognized as such and to be thereby understood (i.e. linguistic communication), would not make sense without that. 4 More importantly for my present purposes, this ability is manifested in

344 Arie Verhagen

the interpretation of linguistic utterances in a very general sense: As soon as some observable phenomena (sounds, marks in stone or on paper, gestures) are recognized as instances of language, this implies that their content is attributed to some subject of consciousness, possibly unknown, but by implication seen as capable of linguistic communication just like the interpreting person him/herself; if an interpreter would not take the signals observed as having been produced as such, they would simply not count as language (possibly still as signs, but then non-intentional ones, i.e. symptoms). Thus the interpretation of discourse may always be seen as not just constructing some understanding of the events and situations depicted in it, but also as coordinating with some subject of conceptualization; the interpretation of linguistic discourse necessarily has both a "content-dimension" and a (intersubjective) "coordination-dimension".5 In view of this inherent, general feature of discourse interpretation, it should come as no surprise that there are several kinds of linguistic elements and constructions that serve to indicate particular features of this coordination dimension, for example modal expressions of different types. In the present context, it seems that this idea is also highly relevant for the semantic characterization of complementation constructions. A natural description of the function of matrix clauses such as My client anticipates ... and It should be added... is precisely that they do not provide information in the contentdimension, but rather in the coordination dimension of the interpretation of the discourse. The first explicitly instructs the reader to construe the informational content (e.g. "finding a new job will be hard") as an anticipation, of a particular person. The second provides an instruction by the writer to construe the content information (e.g. "He has a family to care for") as an additional point, paralleling a previous one. The (more implicit) latter case thus invokes intersubjective coordination between writer and reader, whereas the former expression invokes coordination between the reader and a specific individual mentioned. 6 Suppose now that we distinguish discourse segments not just linearly, in one dimension, but in two, taking this discussion into account. Then fragment (8) may be represented (somewhat abbreviated) as in Figure 1: coordination dimension

content dimension

I note that client anticipates:

not easy to find other work fast

Add to this that

his wife is severely disabled

and that7

he has a family to look after

Figure 1. Text segmentation in two dimensions


In such a representation, the matrix clauses are not part of segments in the content dimension. For one thing, this immediately provides an answer to question (c) mentioned above: In this dimension there is no coordination of a matrix and a subordinate clause, which allows us to avoid the suggestion to that effect in segmentation (8)". However, a more important question at this point is: Is this just an incidental property of the particular matrix-structures in this particular fragment, or is this a manifestation of a more general phenomenon? How general can the procedure be of assigning the content of matrix clauses to the coordination dimension of discourse interpretation? As a matter offact, I think such a procedure can actually be fairly general. I would like to suggest that, whereas constructions with adverbial clauses ('clause combining•, see section 1) may be viewed as grammaticalized expressions for rhetorical relations (cf. Matthiessen & Thompson 1988}, complementation constructions may be viewed as general grammaticalized expressions for intersubjective coordination (with the lexical content of the matrix clauses and the complementizers providing the specifics). To start, it is interesting to have a look at the set of complement-taking verbs, for example as listed for Dutch in the comprehensive reference grammar Algemene Nederlandse Spraakkunst (ANS, both in the first and in the recent second edition), and especially to see what kind of concepts these verbs express; the subtypes distinguished by the ANS are presented, with a few examples, in table 1. Table 1.

a b c

d e

f

Semantic types of verbs taking 'direct object clauses' according to the ANS Verbs expressing a statement, question, command, promise, etc., i.e. having a communicative meaning Verbs expressing some form of knowing, believing. supposing. etc. Verbs expressing evaluation [including constructions of the type "find it a pity/strange", etc.) Verbs expressing wish or desire Verbs expressing a way of perceiving Verbs expressing causation

(ANS 1, 1984, p. 840-842; ANS2, 1997, p. 1156-1158).

There is dearly a generalization to be made over cases (a) through (e): Such predicates all evoke some mental state or process of a subject of consciousness, and the content of the complement is to be attributed to this subject of

346 Arie Verhagen

consciousness. In other words, these predicates are all "mental space builders" in the sense ofFauconnier (1994). We could express this generalization in the form of a "constructional schema" (in the sense of Langacker 1991, p. 546 or Goldberg 1995): a construction consisting of a mental space building predicate and a clausal complement means that the contents of the subordinate clause is to be attributed to the subject of conceptualization referred to in the matrix clause:8 (9)

Complement Construction: construction form: construction meaning:

ls.. NP. [MentaJSpacePredicate····l datlofls.b·····l I AlTRIBtrrE CONTENTS OF

S-B TO REFERENT OF NP,.

Category (0 is different: these predicates indicate causality, with the complement denoting the result; I believe these can be integrated into the account in a motivated way, but as this is only indirectly related to the issue of segmentation, I will not pursue that matter further here. 9 In any case, it is clear that evoking, in some specific respect, a mental space for the contents of another clause is a very general function of matrix clauses of complements. With respect to segmentation it is important to ask if this is also true for other matrix clauses, especially those taking subject clauses (another subtype denied segment-status by Mann & Thompson's rule). In fact, I think it is not difficult to see that is. First of all, one important category of matrix predicates of subject clauses are the passive forms of the predicates mentioned in Table 1 (It was argued... , It has been claimed... , It can be seen ... , in which exactly the same relation between matrix and subordinate clauses holds as in the active voice. Another class consists of matrix clauses in which a predicate nominal phrase evokes some subjective point of view, i.e. adjectives as in It is clear/ puzzling... , or noun phrases as in It is a problem/question .. .. 10 Expressions of this type are specifications of a cognitive state with respect to the proposition expressed in the subordinate clause, and thus evoke a conceptualizer entertaining this cognitive state. Thirdly, grammatical subject clauses may be embedded under 'connecting phrases' such as Daarbij komt... (lit. There-to comes ... ; "It should be added", "Additionally") in (8), or Hier staat tegenover (lit. Opposite to this stands ... ; "On the other hand", "Conversely"). This type can be considered as evoking subjectivity too, albeit in a way that is more implicit than the other ones: such expressions are instructions on how to handle information and therefore imply a subject providing them, rather than that they explicitly mention some cognitive state with respect to a proposition.


The matrix of a complement clause always explicitly specifies a source of subjectivity, while this is not necessary in the matrix of subject clauses. As a consequence it seems that the subjectivity of subject clause constructions is usually interpreted as relating to the producer of the discourse, rather than to some other entity. Consider (10), for example. ( 10) Er is echter dringend behoefte aan nieuwe modellen. De twee-relatie is weliswaar een ideaal voor zeer veel homofielen, maar het is duidelijk dat dat dan heel iets anders is dan het traditionele huwelijk. However, there is an urgent need for new models. It is true that the tworelationship is an ideal for many homosexuals, but it is clear that this will be entirely different from the traditional marriage. In interpreting this fragment, a reader will normally ascribe responsibility for the claim that something is clear to the writer of the text. In other words: the matrix of a subject clause is usually taken as a manifestation of speaker/writer subjectivity (i.e. that of a speech act participant), rather than as character subjectivity (cf. note 6) as the clause itself contains no reference to a participant who is the source of the subjectivity. It should be pointed out though, that this is not an obligatory semantic feature of the construction as such, but a default option given the fact that the construction does not mark a source of subjectivity and the subjective roles of speech act participants are always available for use in interpretation. If the context contains an explicit reference to another subject of conceptualization, then the attribution of responsibility for the claim is easily changed. Suppose that this fragment was a report about someone delivering a speech on types of homosexual relationships; then it might well have been formulated as in ( 10)': ( 10 )' Er is volgens de spreker echter dringend behoefte aan nieuwe modellen. De twee-relatie is weliswaar een ideaal voor zeer veel homofielen, maar het is duidelijk dat dat dan heel iets anders is dan het traditionele huwelijk. However, according to the speaker there is an urgent need for new models. It is true that the two-relationship is an ideal for many homosexuals, but it is clear that this will be entirely different from the traditional marriage. Note that the second sentence in this fragment is actually fully identical to the one in ( 10), but that the opinion expressed in it is now naturally ascribed to the referent of"the speaker" in the previous sentence. But it is a difference between

348 Arie Verhagen

the matrix of a subject clause and that of a complement clause that identification of the latter's subject of conceptualization is constrained linguistically, whereas the matrix of a subject clause does not necessarily provide such constraints, and is thus the only type that allows for speaker/writer subjectivity without any special markings (complement clause constructions requiring some form of first person marking). The interpretive 'freedom' for the matrix of subject clauses is, in my view, a manifestation of the general property of any instance of language use mentioned at the beginning of this section: it being taken as language implies it being taken as having been intentionally produced as meaningful, and therefore implies the projection of some other cognitive entity like the interpreter. Whatever entity is available for attributing a particular thought in a text to can function as such in the case of subject clauses, but complement clause constructions have the special property that their matrix predicate provides a specific constraint on this attribution. Still, it is clear that a generalization over the discourse function of the matrix of complement and subject clauses can and should be formulated; the contents of such clauses is attributed to some subject of consciousness, explicitly or implicitly specified in the conceptualization of the matrix clause. This can be represented by means of a generalization of the construction in (9), which I will call the "embedding constructionn: 11 (II) Embedding Construction:

construction form: construction meaning:

(S-a ( Predica1e • • • .j

dat/ of( S-b • • • •• ) j

ATTRIBtrrE CONTENTS OF CONCEPTUAUZER IN

S-b TO

S-a

The idea now is that it is the construction's meaning that is the basis for the conceptual dependence of the matrix clause on a subordinate one. It specifies that the reader/hearer should engage in cognitive coordination with some subject of conceptualization, and such coordination always takes place with respect to some piece of information; cognitive coordination is never 'void': there is no illocutionary act without propositional content, no assessment without an object of evaluation, no instruction to handle incoming information without such information, in general: no coordination between subjects of conceptualization without some object of conceptualization. If interpreting a text involves the alignment of one's cognitive state with that of another, then such alignment necessarily takes place with respect to some informational content. That is a matter of conceptual necessity; what is a matter of linguistic


convention, on the other hand, is the degree to which this relationship between dimensions of discourse interpretation is 'encoded' in one or more specific words (for example modal adverbs) or constructions (such as ( 11 )). All in all, we now have completed the line of argumentation that aUows us to provide an answer to two of the questions formulated at the end of Section 2. First, as regards question (a), the above analysis contains an account of what it is that complement and subject clause constructions have in common, and that explains why the matrix clauses are not conceptually independent: They provide specifications of the coordination dimension, which must be completed by some specification in the content dimension. In a sense, we have thus turned the traditional notion of' dependent clause' upside down, by showing that it is the matrix clause that is actually conceptually dependent on a subordinate one. Whereas the original rule formulated by Mann & Thompson seemed to imply that it was the subordinate clause that was not independent, we now have reached the conclusion that it is actually the matrix that should be denied the status of separate discourse segment (along with, of course, one subordinate clause). This does not have to conflict with a functional interpretation of the notion of subordination, as soon as it is recognized that matrix clauses function in a dimension of discourse interpretation (that of cognitive coordination with a subject of conceptualization) that is functionally different from the content dimension (that of providing information). Viewing the embedding construction as a grammatical instrument (certainly not the only one) for indicating relationships between the coordination and content dimensions of discourse interpretation allows us to say simultaneously that structurally embedded information is subordinated to something else (viz. a mental space, usually in some specific way), and that it often still provides the most important information, especially new information. Also, several pieces of information can be subordinated to the same mental space (recall the string of embedded clauses in (8)), without them becoming just constituents of a single discourse segment. Secondly, this approach also provides a basis for an answer to question (b) at the end of Section 2: What is the difference between subject and complement clauses on the one hand and restrictive relative clauses on the other, such that the latter never constitute separate discourse segments? The answer can be formulated in terms of the distinction between the two dimensions introduced in this section: Whereas the specific character of embedding constructions precisely consists in a relationship between these two dimensions, restrictive

350 Arie Verhagen

relatives by definition always function in the same dimension as their head noun, and thus in the same dimension as their matrix clause.

4·

Thematic continuity in the content dimension

Before concluding I would like to present an additional piece of evidence suggesting that discourse analysis may profit from a segmentation procedure that takes the distinction between the two dimensions introduced in the previous section into account. 12 This evidence involves certain phenomena of thematic continuity in texts, i.e. indications of how the topic or 'theme' of a particular discourse segment is connected to previous segments. In Onrust, Verhagen and Doeve (1993, Ch. 2), two ways are distinguished in which the initial and final positions of sentences (in Dutch) may contribute to the thematic cohesion of texts. Given two adjacent sentences S1 and S2 in a text, then: a. when the sentence initial constituents ofS 1 and S2 refer to the same piece of information, we have a so-called "constant pattern" (about the same topic, two statements are being made); b. when the initial constituent of S2 refers to the same piece of information as a constituent that is (more or less) final in S2, we have a so-called "chaining pattern". This is indicated schematically in Figure 2. Constant pattern:

I5 1 A ...... B) 152 A ...... C) I

I

Chaining pattern:

151 A ...... B) 152 B ...... C) t.:...=...J

Figure 2. Two patterns ofthematic cohesion.

These patterns are not obligatory ones, but when used they do contribute to the cohesion of texts. As they have been defined these notions only apply to the initial and final positions of (complete) sentences. This sometimes restricts the applicability of the notions, giving rise to conflicts with language users' intuitions about textual cohesion. Students, when applying the analytic method of

Subordination and discourse segmentation revisited

which these definitions form a part, quite generally treat cases like the following as instances of chaining: b. Het gevaar bestaat dat uw klan ten door de aanhoudende vertragingen ontevreden worden oyer uw bedrijf. Wij denken dat dit voorkomen kan worden door te zorgen voor een snellere informatiestroom naar de bezorgafdeling. Een mogelijkheid hiertoe wordt gevormd door ... "The danger exists that because of the continuing delays. your customers will become djssatisjied wjth your company. We think that llW can be prevented by accelerating the flow of information to the delivery department. One possible option in this respect is... , According to the definitions in Onrust et al. ( 1993) however, this cannot be an instance of any pattern, because the demonstrative anaphor dit is not in an initial position of a sentence. Now note that between the final position of the first sentence in (6) - the underlined part "become dissatisfied with your company" - and the demonstrative in the second sentence is the matrix "We think that"- i.e. information relating to the coordination dimension. If we take this into account, and segment the text as in Figure 3, it is immediately apparent that in the content dimension, the demonstrative is adjacent to its antecedent, so in this dimension we actually do have a chaining pattern. coordination dimension

content dimension

Het gevaar bestaat dat

uw klanten ... ontevreden worden oyer uw bedrjjf.

The danger exists that

... your customers wiU become dissam,fied with your compaey dit voorkomen kan worden door te wrgen voor een

Wij denken dat

We think that

sneUere informatiestroom ...

.r1m can be prevented by accelerating the flow of information ...

Figure 3. Thematic cohesion in content dimension. What we may want to propose is that the conditions for patterns of thematic continuity should apply within one specific dimension of discourse representation. Given the distinction between the two dimensions as suggested here, we may say that material that is linearly intervening but relates to another dimension than the antecedent and the anaphor, is 'invisible' to the formation of

3il

3.52 Arie Verhagen

patterns of thematic cohesion. In order to see if this adaptation of the patterning conditions would account for the actual use of discourse anaphors in spontaneously produced texts, a search was undertaken in a corpus with text fragments from different genres, 13 collecting all instances of the complementizers dat and of that were immediately followed by a demonstrative with an antecedent elsewhere in the text (i.e. not in the same sentence). In this corpus, there were 62 instances satisfying this criterion - thus all of them 'violating', as it were, the thematic continuity conditions as formulated by Onrust et al. ( 1993 ). However, taking the distinctions proposed here into account, 39 of them turn into straightforward examples of the chaining pattern (perhaps even 42), and 9 (possibly 10) into examples of the constant pattern. So there are at least 48 out of 62 'exceptional' cases that turn out to be regular ones as an immediate consequence of distinguishing the coordination and content dimensions in the representation of discourse. An example of the most frequent pattern, that of chaining, is given in (13). At the end of one sentence the idea is expressed of the government taking over the entire production machinery. In the linear text, we then get a matrix clause opening a mental space assigned to some economists who used to believe something on theoretical grounds, thus belonging to the coordination dimension (which is indicated by small capitals), and then, as the first element of a new segment in the content dimension, we get the anaphor, referring to the idea at the end of the previous content segment. Such pieces of text are indeed completely natural and unproblematic. ( 13) [... ] Wanneer wij, in de rug gesteund door de moderne economie, bet laissez faire afwijzen, dan staan wij voor de keus tussen twee altematieven. In de eerste plaats kan de oyerbeid bet gehde produktieapparaat oyememen. SOMMIGE ECONOMEN MEENDEN VROEGER OP rnEORETISOfE GRONDEN, DAT di1 niet tot gevolg kon hebben dat de welvaart op gunstige wijze zou worden verdeeld, maar dit standpunt is thans door de meeste economen verlaten. "[ ... ]When we, with thesupportofmoderneconomy, decline the principle of'laissez-faire', we face a choice between two alternatives. On the one hand, the government could take oyer the enhrepraduction machinery. SOME ECONOMISTS USED TO BEUEVE ON 111EORE11CAL GROUNDS, THAT 1hU could not lead to a advantageous distribution of richness, but this opinion has now been abandoned by most economists."


Fragment ( 14) contains an example of a constant pattern, that can be analyzed in a similar way. ( 14) De EEG-raad van ministers van landbouw heeft maandag in Luxemburg in beginsel overeenstemming bereikt over de methodiek van een regeling voor vlas: er zal een forfa.itaire toeslag per hectare worden gegeven [... ]. Over het bedra& van die toesla& zal de Europese Commissie nog een voorstel doen. MINISTER L\RDINOIS VERWACHTIE WEL OAT dczc iets boger zal worden dan de huidige Nederlandse toeslag [ ... ].

"On Monday, the European council of ministers ofagriculture has reached agreement in Luxembourg about the method ofa regulation for flax: a standard surcharge per acre will be given [... ].As to the amount Q/the surcharte, the European Committee wiU produce a proposaL MINISTER LARDINors DID EXPECT THAT lhU will tum out somewhat higher than the present surcharge in the Netherlands [... ]. n Thus there is not only evidence from readers' intuitions, but also from the distribution of discourse anaphors in spontaneously produced texts, that language users treat these devices for cohesion across sentence boundaries in a way that takes the distinction between the coordination and content dimensions into account. This finding thus provides independent support for the proposal to systematically use this distinction in the segmentation of texts.

5.

Conclusion

The central claim in this paper is that it necessary, for an adequate segmentation procedure for natural language texts, to take into account two distinct dimensions of discourse interpretation with respect to which textual fragments may be interpreted. The nature of these dimensions is an immediate consequence of an intrinsic property of interpreting language, viz. that it by definition implies not only processing informational content, but also engaging in cognitive coordination with some entity projected to be responsible for that information. As cognitive coordination in tum presupposes some information to function as object of coordination, the interpretation of expressions in the coordination dimension is not conceptually independent from information in the content dimension. In these terms, constructions with complement and subject clauses have been analyzed as grammatical means for systematically distributing information over these two dimensions; the claim is that this

353

3S4 Arie Verhagen

provides a functional explanation for the condition that in such constructions, the matrix is not conceptually independent, therefore does not constitute a separate discourse segment, but needs at least one subordinate clause to make it conceptually independent (allowing further subordinate clauses to be added as separate segments). This view of conceptual independence as a condition on discourse segmentation is also empirically superior to previous formulations of the conditions on discourse segments.

Notes 1.

Under certain analyses, the second segment of ( 1) might be said to properly contain an

infinite clause (the complement of want), but I will not consider that issue in this paper,

though I believe that the present approach can ultimately be helpful in clarifying that as well. See Verhagen (1995) for some suggestions. This is not meant to imply that the three arguments are necessarily to be taken as equal. Recall that the issue here is just segmentation, not the assignment of (hierarchical) structure. Thus the grammatical structure of (b) could very well be taken as an indication that the last two arguments are to be taken as constituting a set to be added to the single argument in sentence (a) (cf. below). The point here is simply that the question of segmentation precedes the assignment of structure. 2.

3· Again: only segmentation is the issue here, not the assignment of structural relationships; cf. note 2. A dear exposition of the view of linguistic communication as influencing another person's cognition by displaying the intention to do so can be found in Keller (1995, p. 153ff., 1998, p. 136ff.). It is crucially related to Grice's (1957) notion 'meaningNN' and also occurs, in slightly variable forms, in several other approaches to pragmatics.

4-

s.

This distinction is related, but not identical, to distinctions between different domains of use, for example the distinction between epistemic and content domains as proposed in Sweetser ( 1990), or that between pragmatic and semantic sources of coherence as proposed in Sanders (1992). Cf. Foolen (1996) and Verhagen (1996b, p. 274/5) for some discussion. This difference consists in the distinction between what I called "speaker-hearersubjectivity" and "character-subjectivity" (cf. below; also J. Sanders 1994). In certain areas, such as that of language change, this difference is very important (cf. Verhagen 2000, for an example).

6.

7· In this representation the coordinating conjunction is taken to be an element in the coordination dimension, but this is not crucial. At the moment I have no principled considerations to offer on this point, but I find this representation useful for expository purposes. 8.

In a different terminology, a similar insight has been formulated for that-clauses, the


most prototypical subclass of complements, by Wierzbicka ( 1988: 132-140): " ... reference to knowledge is present in all sentences with THAT" (p.l37; Wierzbicka in turn cites a few other linguists who have proposed partly similar analyses, notably Bolinger). I believe that the mental space cum construction approach provides a generalization over these and other types of complements (linking the 'space building' feature to the construction and leaving other aspects of the semantics to the lexical specifications of the verbs and complementizers involved), as well as one that allows for integration into a more general theory of perspectivization. 9· The direction of the generalization I would like to propose is that causation is also attribution of the situation denoted by the complement clause to something else, but then to an objective factor (ie. the cause) rather than to a subjective one. In that perspective, the complementation construction would constitute an example of a particular kind of constructional polysemy. See Verhagen (1996b) for some discussion of this idea, and Foolen ( 1996) for some criticism.

In Dutch, the matrix clauses do not have to contain the pro-form it, for neither category of predicate nominal. Thus Dutch does not only have matrix clauses of the type Een probleem is dat ... ("A problem is that ... "), but also Duidelijk is dat ... (lit. "Clear is that ... "). The parallel between these two types is one reason why in some grammatical traditions, the initial noun phrase in a clause of the type Em/lret probleem is dat.. is analyzed as a preposed predicate nominal rather than a subject. 10.

u. As with complement clauses, the matrix of a subject clause may also have a causal relationship with the subordinate clause ( Cf. The result/reason is ... ). As mentioned in note 9, I think a further generalization is possible, so that we actually have constructional polysemy here, but I will not pursue that issue in this paper. There should also be independent grammar-internal arguments for positing a construction such as ( II) as part of the grammar of a language. I think such arguments can indeed be provided, at least for Dutch (cf. Verhagen 1996b and Foolen 1996, for somewhat different views). Furthermore, this analysis has consequences for the grammatical characterization of subordination as such. Again, these issues are only indirectly related to the matter of discourse segmentation, so I will not go into them here.

12..

13. The Eindhoven Corpus, in the version available from the Free University in Amsterdam; it is described in Uit den Boogaart (1975) and Renkema (1981).

References ANS 1 (1984). Algemene Nederlandse Spraakkunst. Onder redactie van G. Geerts, W. Haeseryn, J. de Rooij, M. C. van den Toom. Groningen/Leuven: WoltersNoordhoff. ANS2 (1997). W. Haeseryn, K. Romijn, G. Geerts, J, de Rooij, M. C. van den Toorn. Algemene Nederlandse Spraakkunst. Tweede, geheel herziene druk. Groningen/ Deume: Martinus Nijhoff/Wolters Plantyn.

3SS

356 Arie Verhagen

Boogaart, P. C. uit den ( 1975). Woordfrequenries in geschreven en gesproken Nederlands. Utrecht Oosthoek, Scheltema 8c Holkema. Daalder, S. (1989). Continuative relative clauses. In Norbert Reiter (Ed.), Sprechen und Hl1ren. Akten des 23. Linguistischen Kolloquiums (pp. 195-207}. Tobingen: Niemeyer. Fauconnier, G. (1994). Mental Spaces. Aspects of Meaning Construction in Natural Language. Cambridge: Cambridge University Press. [First edition, 1985, Cambridge, MA: The MIT Press.) Foolen, A. ( 1996). Tekstsegmentatie, onderschikking en subjectiviteit. Commentaar op Arie Verhagen. Gramma/TTI', 5, 269-272. Goldberg, A. E. ( 1995 ). Constructions. A Construction Grammar Approach to Argument Structure. Chicago/London: The University of Chicago Press. Grice, H. P. ( 1957}. Meaning. The Philosophiad Review, 66, 377-388. Keller, R. ( 1995). Zeichentheorie.Tilbingenl Basel: Francke Verlag. Keller, R. ( 1998 ). A Theory ofLinguistic Signs. Oxford: Oxford UDiversity Press. Langacker, R. W. ( 1991 ). Foundations ofcognitive grammar, 2 Descriptive application. Stanford: Stanford University Press. Mann, W. C., 8c Thompson, S. A. (1988). Rhetorical Structure Theory: Toward a functional theory of text organization. Text, 8, 243-281. Matthiessen, C., 8c Thompson, S. A. ( 1988). The structure of discourse and 'subordination'. In J. Haiman 8c S. A. Thompson (Eds.), Clause Combining in Grammar and Discourse (pp. 275-329). Amsterdam: John Benjarnins, 275-329. Onrust, M., Verhagen, A., 8c Doeve, R. (1993). Formuleren. Houten/Zaventem: Bohn Stafleu Van Loghum. Pander Maat, H. (1994). Telcstanalyse. Groningen: Martinus Nijhoff. Renkema, J. (1981). De taal van "Den Haag": een kwantitatief-stilistisch onder:zoek naar aanleiding van oordelen over taalgebruik. 's-Gravenhage: Staatsuitgeverij. Sanders, J. ( 1994). Perspective in narrative discourse. Tilburg (diss. KUB). Sanders, T. J. M. (1992). Discourse structure and coherence: Aspects ofa cognitive theory of discourse representation. Tilburg (diss. KUB). Sanders, T. J. M., &, van Wijk, C. ( 1996 ). PISA- A procedure for analyzing the structure of explanatory texts. Text. 16,91-132. Schilperoord, J. (1996). It's about time. Temporal aspects of cognitive processes in text production. Amsterdam/Atlanta: Rodopi. Schilperoord, J. ( 1997). Temporele modificatie in clauses; een pauze-analytische studie naar tekstproductie. In H. van den Bergh, D. M. L. Janssen, N. Bertens, M. Damen (Eds.), Taalgebruik ontrafeld. Bijdragen van het zevende VIOT-taalbeheersingscongres gehouden op 18, 19 en 20 december 1996 aan de Universiteit van Utrecht (pp. 263-274). Dordrecht: ICG Publications. Schilperoord, J., 8c Verhagen, A. ( 1998 ). Conceptual dependency and the clausal structure of discourse. In Jean-Pierre Koenig (Ed.), Discourse and Cognition. Bridging the Gap (pp. 141-163 ). Stanford: CSLI Publications. Sweetser, E. E. (1990). From etymology to pragmatics. Metaphorical and cultural aspects ofsemantic structure. Cambridge: Cambridge University Press.


Verhagen, A. {1992). Patroonsplitsing en zinsstructuur. In H. Bennis & J. W. de Vries (Eds.), De Binnenbouw van het Nederlands. Een bundel artikelen voor Piet Paardekooper(pp. 373-382). Dordrecht: ICG Publications. Verhagen, A. (1995). Subjectification, syntax, and communication. In D. Stein & S. Wright (Eds.), Subjectivity and subjectivisation: linguistic perspectives (pp. 103-128). Cambridge: Cambridge University Press. Verhagen, A. (1996a). Sequential conceptualization and linear order. In E. H. Casad (Ed.), Cognitive Linguistics in the Redwoods. The Expansion of a New Paradigm in Linguistics (pp. 793-817). Berlin: Mouton de Gruyter. Verhagen, A. (1996b). Tekstsegmentatie, onderschikking en subjectiviteit. Grarnma/ TIT, 5, 249-268, 273-275. Verhagen, A. (2000). 'The girl that promised to become something": An exploration into diachronic subjectification in Dutch. InT. F. Shannon & J.P. Snapper (Eds.),

The Berkeley Conference on dutch Linguistics 1997: the Dutch Language at the Millennium (pp. 197-208). Lanham, MD: University Press of America. Wierzbicka, A. {1988). The Semantics of Grammar. Amsterdam/ Philadelphia: John Benjamins.

Subject index

A Accessibility 6, 9-10, 27-fJ7 Activation 27, 125 Additive relations 13, 15-16, 185, 198, 223 Adverbial clause 345 Adversative relations, see also Contrastive relations 154,212,215 Allocation of attention 94 Anaphor4,10,29-87,94,97,249,352353 Animacy 69-70 Antecedent-anaphor relation 29-87 Argumentation 231-246 Argumentation theory 126, 233 Argumentative discourse 231-232 Argumentative relations 130, 142, 145, 209,216,241 Argumentative scale 11 Argumentative orientation 155,214 Attention 313 Autonomous processes 112, 115, 118, 309-334

c Catapbor 98-100 Categorization ofcoherence relations, see Classification of coherence relations Causal relations 5, 13, 15-16,94, 130, 132-133, 141, 146, 148, 153-178, 189, 199,210-211,215,220,223-224,231, 234, 238,243,256-257,258,260, 337-338,346,355 Causal-contrastive relations 153-178

Centering theory 64-65, 189 Classification of coherence relations 13, 15, 125, 127-150,153-178, 181, 198, 221,223-224,239,260 Clause combining 338, 345 Clauses 33,307, 309-334,337-357 complement 338-339, 341,348,353, 355 elliptical217-220 finite338 full217-220 main 156, 164-167, 169-170,174176,337,343,346 relative 339,341,343,349-350 subject 339,341,346, 348,353 subordinate 156, 158-159, 164-167, 169-170,174-176,307,337,340, 342-343,345,355 Coherence relations 10-16, 125, 127150,153-178,181,235,247,260,261, 264,337-338,345 Cohesion 33, 350, 353 Comparative relations 198, 234 Conceptual dependency 336, 340, 342343,353-354 Conceptual entities 250, 273 Conceptual processes and production 310,312-313,315,319,321,323,327328,232-233 Concessive relations 153-178, 187 Conclusion relations 131,200 Concomittance relations 234, 237 Conditional relations 11, 128, 133, 138, 139, 143, 147, 159 Conjunctions, see Connectives

360 Text Representation

Conjunctive relations 181 Connections in network 274-275, 284, 299,302 excitatory 275 inhibitory 275 Connectives, see also Linguistic markers of coherence relations 4, 8, 11-12, 125, 190,194,197-227,235 al (Dutch) 177 although 153-180 and126,134, 155,197-227 because 12, 16, 125

besides224 but11,148-149,155-156,168, 177, 200,213-214,226 en (Dutch) 197-227 if133, 138,143-144,147 or133, 136,138,142 since 146 so 14, 128-130, 136, 138, 141, 146 then 136 though 177 Constituents316, 319,322,329-331 Construction Integration Model 6, 250 Constructionist theory 250, 252, 256, 258 Content relations 13, 125, 158,308,344, 353 Con~tivepronouns43

Con~tive

relations, see also Adversative relations 13, 128, 260 Conversational implicature 222 Coordination 308, 343, 354 Coordination dimension 344, 353 Coreference 57, 61 Corpus research, see Corpus analysis Corpus analysis 73, 125-126, 162, 166, 169,197,208-227 Cues for coherence relations, see Linguistic markers of coherence relations

D Deixis 193 Denial of expectation relations, see Causal-contrastive relations Discourse production 5 Discourse comprehension, see Text comprehension Discourse relations, see Coherence relations Discourse Representation Theory (DRT)

5 Discourse structure 247, 343 Discourse units 307, 316, 337-357 supra-sentential316 Disjunctive relations 132, 138 Domain knowledge, see World knowledge

E Ecological validity 71 Elaboration relations 126,181-195, 199200, 211, 216 Embedding 187 Episodic memory 91 Epistemic modality 6 Epistemic relations 125, 127-150,158, 223-224,238 Event indexing model 250, 261 Explanation relations 187, 232-234, 241 External relations 13, 128 Eye movement registration 5 F Figurative speech 6 Focus (of attention) 46, 90, 181, 189190, 192,205-206 Forgetting 303 Folk-psychology 281 Full nouns 316 G

Given/new 30, 67 Givenness hierarchy 63 Global coherence 10, 32, 192-193, 260

Subject index 361

Graded Salience Hypothesis 28, 113, 115, 118 Grammar 53 Grammatical structure 343

H Headinp, see Titles and headinp Hierarchy 192

I Ideational relations 13, 128, 308 Idioms 112 Indefinite article 8, 10,316 Inference 3, 4, 13, 32, 148, 159, 198-199, 2l2,249,252,254 Intentional relations 260 Intentions 139-150, 182 Interactive processes 112,309-334 Internal relations 14, 128 Intraclausal and interclausal relations 54, 204-208 Irony 112

K Knowledge representation 247, 273-306 of experts 276, 286-287 Knowledge structure 274 -306 Knowledge Telling model312

L Language production, see also Text production 342 Landscape model of Reading 6, 27, 90110,250 Lexical access 111-124 Lexical nouns 9 Linguistic markers of coherence relations 8,12,27,89-110,190,192,235,250, 261,264,338 Linguistic markers of referential coherence 10,89-110,205 Linguistic, processes and production 311,312-313,315,319,321,323,327328,332-333

List relations 219

Local coherence 260 Local operations 310

M Matrix of concept relatedness 2n-278 Memory extemal255 long term 285, 302 storage and retrieval296, 301 working 89-90, 260, 285, 301, 317 Memory representation 5 Mental model, see Situation model Mental space 3, 11, 15, 346 Metaphor 6, 27, 111-124, 307 Minimalist hypothesis 250, 152 Modularity, see Autonomous processes Multilevel model of analysis 318-334 Multiple representations 3

N Negative relations 13, 15, 153, 223-224 Negative causal relations, see Causalcontrastive relations Network representation 274-275, 285286,299 Nucleus/Satellite 155-156, 183, 185, 187

0 On-line measures 256, 258 Order of relations 160, 162, 164-167, 176, 178 p Paragraphs 33, 316, 319, 322, 323, 326, 333 Pause locations 309-334 Pauses307,309-334 variance in 318,326-327 Pause times, see Pauses Polarity 223-226 Positive relations 15, 223-224 Pragma-dialectical approach, see Argumentation theory

362 Text Representation

Pragmatic inference, see Inference Pragmatic principles 52 Pragmatic relations 14, 125, 127-150, 158,22l-224,236-237,239-241,308, 354 Pragmatics 53, 197,233 Presentational relations 14 Processing 45, 73,89-110, 159,207 Pronouns 8-9, 29-87,221, 249,316 Proposition 3,14,133,135,137,141,157,178,187189,191,193-194,200,215,234,338, 348 Psycholinguistic vs. linguistic data 72

Q Quantitative models of thinking 247, 271-306

R Readingtnnesl0,12, 16,125,258,261 Reaction tUnes 73 Reca1ll2, 16, 227 Referential choice 69-70 Referential coherence 7, 10, 17, 94, 125, 218,317 Referential distance 38, 46 Referential expressions 10,17, 29-87, 316 Referential function 46 Reflexives 34 Relational coherence 7, 17, 247 Relations between concepts 277-278 Relevance 38, 41 indicators of94-95, 102 Relevance theory 126, 197-227 Resonance model250 Resumptive pronoun 34, 57 Rhetorical Structure Theory (RST) 126, 181-195,236,337 Rhetorical relations, see Coherence relations

s Salience 55 Segmentation 19, 337-357 Segmentation marker 226 Semantic relations 13, 125, 127-150, 158,22l-224,236,239-241,308,354 Sentences 316,319,322,329-331,333 Simulation model274, 281, 283 Situation model 3, 249, 253, 258 Speech acts, see also Speech act relations 232 Speech act relations 125, 127-150, 158159,164,238 Speech production 316 Spreading activation 273, 279-284, 296297

Spreading activation model247 Structure building framework 250 Subject of consciousness 15, 343, 345, 348 Subjectivity 15, 343, 347, 354

T Temporal sequence relations, see Temporal relations Temporal relations 128, 134-135, 137, 144, 211, 220, 260 Text comprehension 5, 89, 105, 163,227, 249, 254, 264, 274, 307 Text generation 182 Text planning 307,316,342 Text production 6, 27, Ill, 114, 118, 153,307,309-334 Text representation 10, 12,247,307 Text segments, see Discourse units Thematic continuity, see also Topic continuity 167-178,316, 350-351 Think aloud protocol256 Thinking process 273-306 Thought product 301 Thoughtstruehrre288,294,301 Titles and headings 102-104 Topic 8, 10, 32, 186, 193, 197,201-202,

Subject index 363

205-206,209,216,219,316-317 Topic continuity, see also Thematic continuity 70,202,218 Topicality 38 Truth-conditional properties, see Truthfunctional properties Truth-functional properties 221 Typographical cues 89, 100-101

v Verbal protocols 256 Verification 16, 161-162

w Working memory 89-90, 285 World knowledge 17, 164, 250,256,261, 264,273-306

u

z

Units of analysis 307, 337-357

Zero-anaphora 8-9,57,317

In the series HUMAN COGNITIVE PROCESSING (HCP) the following titles have been published thus far or are scheduled for publication: 1. NING YU: The Contemporary Theory ofMetaphor. A perspective from Chinese. 1998. 2. COOPER, David L.: Linguistic Attractors. The cognitive dynamics of language acquisition and change. 1999. 3. FUCHS, Catherine and Stephane ROBERT (eds.): Language Diversity and Cognitive Representations. 1999. 4. PANTHER, Klaus-Uwe and Ganter RADDEN (eds.); Metonymy in Language and Thought. 1999. 5. NUYTS, Jan: Epistemic Modality, Language, and Conceptualization. A cognitivepragmatic perspective. 2001. 6. FORTESCUE, Michael: Pattern and Process. A Whiteheadian perspective on linguistics. 2001. 7. SCHLFSINGER, Izchak, Tamar KEREN-PORTNOY and Tamar PARUSH: The Structure ofArguments. 2001. 8. SANDERS, Ted, Joost SOIILPEROORD and Wilbert SPOOREN (eds.): Text Representation. Unguistic and psycholinguistic aspects. 200 I.

Text representation: linguistic and psycholinguistic aspects

Introduction to Linguistic Annotation and Text Analytics

The linguistic basis of text generation

Cross-Linguistic Aspects of Processability Theory

Gender Across Languages: The Linguistic Representation of Women and Men

History, Theory, Text: Historians and the Linguistic Turn

History, Theory, Text: Historians and the Linguistic Turn

Explorations in Pragmatics: Linguistic, Cognitive and Intercultural Aspects

Representation and Reality (Representation and Mind)

Representation and Reality (Representation and Mind)

Representation and Reality (Representation and Mind)

The Grassmannian Variety, Geometric and Representation-Theoretic Aspects

Representation and Reality (Representation and Mind)

Handbook of Bilingualism: Psycholinguistic Approaches

Cross-Linguistic Influence in Third Language Acquisition: Psycholinguistic Perspectives (Bilingual Education and Bilingualism)

Representation

Children's Explanations: A Psycholinguistic Study

Children's Explanations: A Psycholinguistic Study

Handbook of Bilingualism: Psycholinguistic Approaches

Cross-Linguistic Aspects of Processability Theory (Studies in Biligualism)

Diagrammatic Representation and Inference

Urban Space And Representation

Knowledge representation and reasoning

Fair and Effective Representation?

Representation and Realism

Diagrammatic Representation and Inference

Representation and Reality

Meaning and Mental Representation

Knowledge Representation and Reasoning

Lexical representation and process

Representation And Reality

Text representation: linguistic and psycholinguistic aspects

Introduction to Linguistic Annotation and Text Analytics

The linguistic basis of text generation

Cross-Linguistic Aspects of Processability Theory

Gender Across Languages: The Linguistic Representation of Women and Men

History, Theory, Text: Historians and the Linguistic Turn

History, Theory, Text: Historians and the Linguistic Turn

Explorations in Pragmatics: Linguistic, Cognitive and Intercultural Aspects

Representation and Reality (Representation and Mind)

Representation and Reality (Representation and Mind)

Representation and Reality (Representation and Mind)

The Grassmannian Variety, Geometric and Representation-Theoretic Aspects

Representation and Reality (Representation and Mind)

Handbook of Bilingualism: Psycholinguistic Approaches

Cross-Linguistic Influence in Third Language Acquisition: Psycholinguistic Perspectives (Bilingual Education and Bilingualism)

Representation

Children's Explanations: A Psycholinguistic Study

Children's Explanations: A Psycholinguistic Study

Handbook of Bilingualism: Psycholinguistic Approaches

Cross-Linguistic Aspects of Processability Theory (Studies in Biligualism)

Diagrammatic Representation and Inference

Urban Space And Representation

Knowledge representation and reasoning

Fair and Effective Representation?

Representation and Realism

Diagrammatic Representation and Inference

Representation and Reality

Meaning and Mental Representation

Knowledge Representation and Reasoning

Lexical representation and process

Representation And Reality

Recommend Documents