The Processing and Acquisition of Reference

The Processing and Acquisition of Reference The Processing and Acquisition of Reference edited by Edward Gibson and ...

Author: Edward A. Gibson | Neal J. Pearlmutter

27 downloads 781 Views 2MB 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

The Processing and Acquisition of Reference


edited by Edward Gibson and Neal J. Pearlmutter

A Bradford Book The MIT Press Cambridge, Massachusetts London, England

© 2011 Massachusetts Institute of Technology All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from the publisher. For information about quantity discounts, email [email protected]. Set in Times New Roman and Syntax on InDesign by Asco Typesetters, Hong Kong. Printed and bound in the United States of America. Library of Congress Cataloging-in-Publication Data The processing and acquisition of reference / edited by Edward A. Gibson and Neal J. Pearlmutter. â•… p.â•… cm. This volume presents papers from the special session at the CUNY Sentence Processing conference that was hosted by MIT and Northeastern University in 2003. Includes bibliographical references and index. ISBN 978-0-262-01512-7 (alk. paper) 1. Reference (Linguistics) — Congresses.â•‡ 2. Language acquisition — Congresses.â•‡ I. Gibson, Edward, 1962–â•‡ II. Pearlmutter, Neal J., 1967– P325.5.R44P76â•… 2011 2010022680 401'.456 — dc22 10â•… 9â•… 8â•… 7â•… 6â•… 5â•… 4â•… 3â•… 2â•… 1

Contents

1

Introductionâ•…

Edward Gibson and Neal J. Pearlmutter

1

I

Children’s Acquisition and Processing of Reference

2

Cues Don’t Explain Learning: Maximal Trouble in the Determiner Systemâ•… 15

Ken Wexler

3

Children’s Use of Context in Ambiguity Resolutionâ•…

Luisa Meroni and Stephen Crain

4

Referential and Syntactic Processes: What Develops?â•…

John C. Trueswell, Anna Papafragou, and Youngon Choi

5

Parsing, Grammar, and the Challenge of Raising Children at LFâ•…

Julien Musolino and Andrea Gualmini

6

A Cross-Linguistic Study on the Interpretation of Pronouns by Children and Agrammatic Speakers: Evidence from Dutch, Spanish, and Italianâ•… 133

Esther Ruigendijk, Sergio Baauw, Shalom Zuckerman, Nada Vasić, Joke de Lange, and Sergey Avrutin

7

Processing or Pragmatics? Explaining the Coreference Delayâ•…

Tanya Reinhart

II

Adults’ Processing of Reference: Evidence from the Visual-World EyeTracking Paradigm

8

Disfluency Effects in Comprehension: How New Information Can Become Accessibleâ•… 197

Jennifer E. Arnold and Michael K. Tanenhaus

43 65 109

157

vi

Contents

9

It’s Not What You Said, It’s How You Said It: How Modification Conventions Influence On-Line Referential Processingâ•… 219

Jodi D. Edwards and Craig G. Chambers

10

The Effect of Speaker-Specific Information on Pragmatic Inferencesâ•…

Daniel Grodner and Julie C. Sedivy

11

Referential Processing in Monologue and Dialogue with and without Access to Real-World Referentsâ•… 273

Simon Garrod

III

Adults’ Processing of Reference: Evidence from Corpora and Reading Experiments

12

Noun-Phrase Anaphor Resolution: Antecedent Focus, Semantic Overlap, and the Informational Load Hypothesisâ•… 297

H. Wind Cowles and Alan Garnham

13

Investigating the Interpretation of Pronouns and Demonstratives in Finnish: Going beyond Salienceâ•… 323

Elsi Kaiser and John C. Trueswell

14

Not All Subjects Are Born Equal: A Look at Complex Sentence Structureâ•… 355

Eleni Miltsakaki

15

Complement Focus and Reference Phenomenaâ•…

Anthony J. Sanford and Linda M. Moxey

16

The Binding Problem for Language, and Its Consequences for the Neurocognition of Comprehensionâ•… 403

Peter Hagoort

Indexâ•… 437

381

239


1â•…

Introduction Edward Gibson and Neal J. Pearlmutter

This volume presents papers from a special session at the Sixteenth Annual CUNY Conference on Human Sentence Processing, hosted by the Massachusetts Institute of Technology and Northeastern University in 2003. The goal of the special session was to bring together researchers in the fields of language processing and language acquisition to discuss topics of common interest: how people refer to objects in the world, how people comprehend such referential expressions, and how children acquire the abilities to refer and understand reference. Linguistic reference is particularly well suited to connecting the fields of acquisition and processing because it is an active area of research in both linguistics and psycholinguistics, because questions related to producing and understanding reference are already under investigation in both adults and children, and because it is particularly amenable to investigation using headmounted eye-tracking methods (Tanenhaus, Spivey-Knowlton, Eberhard, and Sedivy 1995; Trueswell, Sekerina, Hill, and Logrip 1999), which have enabled direct investigation of referential processes using identical designs with adults and children. The papers in the volume are divided into three parts. Part I focuses on Â�issues related to children’s acquisition and processing of reference. Parts II and III are dedicated to issues related to adults’ processing of referential information, part II focusing on work using the visual-world paradigm and part III Â�focusing on work based on corpora and reading experiments. Part I: Children’s Acquisition and Processing of Reference

In chapter 2, Ken Wexler provides a novel account of the pattern of data from the eye-tracking study by Trueswell et al. (1999), which showed that children aged 5 behave very differently from adults when following instructions like Put the frog on the napkin into the box. In particular, children seem unable to use the presence of multiple elements of the same type in the context (e.g., two

2

Gibson and Pearlmutter

frogs) in order to disambiguate the prepositional phrase on the napkin as a modifier of the frog. Wexler proposes that the difficulty that children have in this task stems from their lack of knowledge of the meaning of the word the. He summarizes evidence from the literature that suggests that young children don’t know what the means, and then goes on to show how this lack of knowledge can account for the pattern of data that Trueswell et al. observed. Wexler also discusses the proposal of Grodzinsky and Reinhart (1993) — another account of children’s difficulties in the acquisition of referential expressions which postulates processing complexity as a cause — and suggests that it cannot explain the available acquisition data from the literature either. And then, in a spirited discussion, Wexler argues that the proposals that Trueswell et al. put forward to account for their data are inadequate in other ways. The result is a thoughtful challenge for processing-based theories of aspects of language development. In chapter 3, Luisa Meroni and Stephen Crain also focus on theories of the data pattern from the eye-tracking study by Trueswell et al. (1999). Meroni and Crain argue against aspects of the interpretation of these findings by Trueswell et al. In particular, although they agree with Trueswell et al. that children overcommit to early parsing decisions, they take issue with the claim that children are less able than adults to use the referential context to guide their online parsing decisions. Rather, Meroni and Crain argue that the behaviors shown by the children are completely explainable in terms of early commitment together with an adult-level processing system, where referential information and lÂ�exico-syntactic information are used to resolve ambiguity. In order to demonstrate this, Meroni and Crain present four experiments in which children’s actions are tabulated with respect to displays like the ones that Trueswell et al. first tested. Meroni and Crain first show that children, when faced with the instruction Put the frog on the napkin in contexts with one frog on a napkin and a second not on a napkin, overwhelmingly interpret this instruction to mean that they should put the frog that is not on a napkin onto a napkin. Meroni and Crain therefore argue that the context that Trueswell et al. had used in their experiments contains a pragmatic bias to interpret the frog that is not on a napkin as the one that the experimenter is referring to. They then show that when the context is changed so that both frogs are already on napkins (one on a blue napkin, one on a red napkin) the instruction Put the frog on the red napkin is infelicitous, so that children want to know which frog the experimenter is talking about. In these contexts, it is then demonstrated that children can easily use the prepositional phrase on the red napkin as a modifier in the instruction Put the frog on the red napkin into the box. The results of these experiments are useful additions to the literature on children’s processing of referential information.

Introduction

3

In chapter 4, John Trueswell, Anna Papafragou, and Youngon Choi discuss theories of adult and child sentence processing as constrained by the available data. They focus some of their attention on the aforementioned finding by Trueswell et al. (1999) that 5-year-old children did not seem to be able to use contextual cues to resolve ambiguity in sentences like Put the frog on the nÂ�apkin into the box. Trueswell et al. (ibid.) explained this finding in terms of development of processing resources or working memory: Children’s limited working memory capacity does not allow them to use information from the context to disambiguate linguistic input. In the current paper, Trueswell et al. have refined this claim to suggest that children’s gradual improvement in their ability to use referential context is related to developmental changes in general executive-function processes, specifically the ability to select among competing representations ( Novick, Trueswell, and Thompson-Schill 2005; Trueswell and Gleitman 2004). Trueswell et al. relate this proposal to evidence from the literature on adult language processing, in which it has been shown that normal adults sometimes hold on to meanings consistent with an initially considered but eventually abandoned interpretation of a temporarily ambiguous linguistic signal (see, e.g., Christianson, Hollingworth, Halliwell, and Ferreira 2001). In addition, Trueswell et al. provide evidence and arguments against the view offered here by Wexler and they respond to the data presented here by Meroni and Crain. In chapter 5, Julien Musolino and Andrea Gualmini summarize their investigation of how sentences with ambiguous scope are resolved by children. They focus on an ambiguity involving the universal quantifier every in subject position along with a negated main predicate, as in Every horse didn’t jump over the fence. In one interpretation of this sentence — the “surface-scope” iÂ�nterpretation — the expression every horse refers to all the horses in the set,Â€ such that for each horse in the set it is true that it did not jump over theÂ€ fence. In the other interpretation of this sentence — the “inverse-scope” iÂ�nterpretation — the expression every horse refers to a subset of the horses, such that it is not true that each horse jumped over the fence (i.e., there were some horses that did not jump over the fence). Whereas adults appear to be able to obtain the inverse-scope interpretation in relevant contexts, Musolino and Gualmini summarize work showing that children often seem to prefer the surface-scope interpretation, and cannot obtain the inverse-scope interpreÂ� tation. This finding suggests a stage of development where children have a strong bias in interpreting these kinds of ambiguities that disappears by adulthood. Musolino and Gualmini speculate on factors that might determine the initial preference and might make it disappear by adulthood. This interesting data pattern — children’s inability to interpret every in these constructions as

4


referring to only a subset of entities in the set — remains a puzzle for accounts of acquisition of reference. Chapters 6 and 7 evaluate different theories of binding and coreference in human language. In an influential paper published in 1983, Tanya Reinhart proposed a division between variable-binding aspects of anaphoric dependencies (such as between himself and every bear in the sentence Every bear washed himself.) and coreference between referring expressions (such as between him and John in a sentence like John wants Mary to love him.). Chien and Wexler (1990) later showed that young children’s understanding of vÂ�ariable-binding was much better than their understanding of coreference, thus offering additional empirical support for Reinhart’s proposed dissociation between the two kinds of anaphoric dependencies. Chien and Wexler’s observation that young children can understand variable-binding but not coreference has since been replicated many times. There are several theories of what kind of knowledge is acquired by young children so that they eventually achieve adult performance on coreference. Chapters 6 and 7 present research programs and data relevant to distinguishing among some of the theories of coreference development from the literature. In chapter 6, Esther Ruigendijk, Sergio Baauw, Shalom Zuckerman, Nada Vasić, Joke de Lange, and Sergey Avrutin present data from children and agrammatic aphasic patients whose native languages were Dutch, Spanish, and Italian that are relevant to distinguishing among some theories of coreference. In particular, they tested transitive and Exceptional Case Marking constructions (e.g., The boy saw him dance.) in order to distinguish between CÂ�homsky’s (1981) binding hypothesis, on the one hand, and Reinhart and Reuland’s (1993) and Reuland’s (2001) computational-complexity-based theories, on the other. The results showed that the interpretation of pronouns was more difficult in Exceptional Case Marking constructions than in transitive constructions in both children and agrammatic aphasics. Ruigendijk et al. argue that their results are more compatible with Reinhart and Reuland’s (1993) and Reuland’s (2001) hypotheses than with Chomsky’s (1981) binding theory. In chapter 7, Tanya Reinhart describes a research program that seeks to discover the nature of coreference in human language by using evidence from adult acceptability judgments on possible sentences and evidence from child language acquisition. In 1983, Reinhart had proposed that children are delayed in acquiring a pragmatic principle relevant to understanding appropriate coÂ� reference. In more recent work, Grodzinsky and Reinhart (1993) proposed that the computation involved in coreference is more complex than that involved in variable-binding, and that this computational complexity may explain children’s relative difficulties in the relevant tasks. Thornton and Wexler (1999)

Introduction

5

then provided arguments against the processing account. They maintained that the coreference delay reflects a pragmatic deficiency, and they developed an analysis of the relevant pragmatic factors. In the current paper, Reinhart summarizes this earlier work and presents further arguments in favor of the computational complexity account. On a very sad note, Tanya Reinhart passed away in the spring of 2007. We will all miss her greatly. Part II: Adults’ Processing of Reference: Evidence from the Visual-World Eye-Tracking Paradigm

In chapter 8, Jennifer Arnold and Mike Tanenhaus use a novel method in order to measure the relative accessibility of different referents to a particular speaker during language comprehension. Rather than relying on production data or reading times to different kinds of noun-phrase referents, Arnold and Tanenhaus recorded people’s eye movements to different physical objects in a scene. They manipulated how fluently the referent was pronounced in the utterance and found strong evidence that the fluency of production of a referent affected interpretation of that referent. In particular, listeners preferred to look at dÂ�iscourse-given objects when the productions were fluent, but they looked at discourse-new objects when the productions were slower with suggestions of disfluency. The results from these experiments suggest that the notion of discourse “accessibility” is complex. In particular, people may expect reference to non-given referents when they notice that their conversation partner is having difficulty accessing the words for the target referent. If the accessibility of a referent from a speaker’s point of view is determined by whether the speaker has that referent activated in his or her mind, we can measure that accessibility by cues in their speech, such that non-fluent speech may indicate a shift to talking about something that has not been talked about recently. Furthermore, the results suggest that naturalistic productions can be very informative: speech “errors” such as disfluent speech can be used to understand how the human production and comprehension mechanisms work. In chapter 9, Jodi Edwards and Craig Chambers present two head-mounted eye-tracking experiments investigating the use of probabilistic information in resolving noun-phrase reference in the comprehension of spoken language. They focus on comprehenders’ ability to use information about the likelihood of prenominal versus postnominal modification to pick out a referent in visual arrays. The first experiment used simple spoken instructions with a pÂ�ostnominal noun-phrase modifier (e.g., Click on the square with the diamonds) and compared visual displays with a target object that required a postnominal modifier

6


and one of four competitor objects. The competitor could be a different category from the target (e.g., a circle instead of a square), or could be of the same category as the target (e.g., a square) but marked with a different property that is typically expressed with a prenominal modifier (e.g., the green square), marked with a different property that could be specified either postnominally or prenominally (e.g., the starred square / the square with the star), or marked with a property that required a postnominal modifier (e.g., the square with the happy face). In the experiment, comprehenders showed clear sensitivity to the likelihood that a postnominal modifier would be used: The proportion of looks to the competitor object from the onset of the postmodifier increased steadily with the likelihood that a postnominal modifier would be used for the particÂ� ular contrast present in the environment. This is the first demonstration of sensitivity to this rather subtle property of the relationship between the internal syntactic structure of the referring noun phrase and specific properties of the potential referents in the environment. In the second experiment, however, Edwards and Chambers demonstrate at least one limit on this sensitivity: Within a paradigm similar to the original examination of contextual disambiguation ofÂ€a syntactic prepositional-phrase-attachment ambiguity by Tanenhaus et al. (1995), comprehenders showed no sensitivity to the relative likelihood of prenominal versus postnominal modification; difficulty with the ambiguity was not affected by this potentially useful constraint. Edwards and Chambers conclude with a discussion of a variety of possible explanations for the limitation on the use of this likelihood information, and they expand on how information about modifier form might fit with other sources of constraint on reference related to common ground, conversational inference, and pragmatics. In chapter 10, Daniel Grodner and Julie Sedivy continue a research program that investigates the time course of the use of pragmatic information in language comprehension. Sedivy’s earlier research demonstrated that when comprehenders encounter a referential form including a modifier that typically indicates contrastiveness (e.g., the tall glass in a set of glasses), they assume that the referential form is being used contrastively, and consequently they look to the relevant object in the scene very early when processing the referÂ� ential expression ( before the noun glass is encountered). Sedivy has argued that people are implicitly following Grice’s (1975, p. 26) maxim of quantity: “Don’t make your contribution more informative than is required for the purposes of the present exchange.” In the current volume, Grodner and Sedivy explore whether speaker-specific information affects listeners’ expectations about how speakers will refer to an item in the context. One prediction of the Gricean aÂ�pproach is as follows: If comprehenders think that a speaker is not using referring expressions in an appropriate (contrastive) way, they may adapt

Introduction

7

their responses to productions from such speakers. Grodner and Sedivy find evidence consistent with this prediction in an experiment manipulating the iÂ�nformativeness of the speaker. In particular, they demonstrate that comprehenders stop using contrastive information early in resolving reference when speakers consistently do not use contrastive elements appropriately. GrodnerÂ€and Sedivy then speculate on how a Gricean approach to understanding reference — in which perceivers must reason counterfactually about what the speaker could have said but did not — might be implemented in the human mind. In the final chapter in part II, Simon Garrod provides an overview of how the role of the situation model in referential processing can link seemingly incompatible results from studies of monologue and dialogue and from studies of reading and visual-world eye tracking. Garrod notes how results from studies of pronoun resolution in reading suggest a two-step model, in which candidate antecedents for an anaphor are first identified on the basis of gender matching and number matching, then evaluated with respect to the overall situation model. In similar visual-world studies, these stages appear to be collapsed together, and Garrod suggests an explanation in terms of differences in how the situation model is related to the text versus the visual display of potential referents. Similarly, in the case of monologue versus dialogue, Garrod focuses on how the relationship among the situation model, the preceding discourse, and the current utterance enforces different constraints on reference processes, leading to different patterns of results. He thus shows how a comparison among these different types of studies can shed new light on the processes of reference resolution in general. Part III: Adults’ Processing of Reference: Evidence from Corpora and Reading Experiments

In chapter 12, Wind Cowles and Alan Garnham examine the behavior of full noun-phrase anaphors in discourse contexts in an attempt to determine how focus status and semantic overlap between an anaphor and its antecedent interact. Following up on Almor’s (1999) Informational Load Hypothesis, which predicts that an anaphor’s semantic overlap with its antecedent should trade off against the accessibility of the antecedent, Cowles and Garnham report three experiments using noun-phrase category anaphors and manipulating the focus status of the antecedent and its typicality within the anaphor’s category (e.g., oak vs. palm as antecedent for tree), where greater typicality yields more sÂ�emantic overlap. The first experiment shows a pattern supporting Almor’s hypothesis, the anaphor being read more quickly after typical than atypical

8


antecedents when the antecedent was not focused, and the reverse pattern when the antecedent was focused. However, the latter difference is relatively weak, and Cowles and Garnham attempt to obtain clearer patterns by adjusting the instructions, the presentation, and the comprehension questions in the second and third experiments. Surprisingly, however, not only do these manipulations not strengthen the focus by typicality interaction, they eliminate it. Noting that these results raise more questions than they answer, Cowles and Garnham go on to discuss possible explanations for the apparent instability of the focustypicality relationship and to consider what other factors may need investigation within the Informational Load Hypothesis. Antecedent accessibility and salience in reference are also examined in chapters 13 and 14. In chapter 13, Elsi Kaiser and John Trueswell argue that the salience notion commonly employed in the accessibility/salience hierarchy for referential forms is oversimplified. They compare two third-person anaphoric forms referring to humans in Finnish: the gender-neutral pronoun hän (s/ he) and the demonstrative tämä (this). Pronominal and demonstrative forms are located at different points in the accessibility hierarchy, with pronominals hypothesized to refer to entities more salient than demonstratives. But Kaiser and Trueswell show in two off-line sentence-completion experiments and an on-line eye-movement monitoring experiment that hän and tämä depend differently on factors that affect saliency — specifically, word order (informationstructure constraints) and grammatical role (subject vs. object). In particular, hän is used largely to refer to subjects regardless of the previous sentence’s word order, whereas tämä typically refers to the object in a preceding subjectverb-object sentence but refers about equally often to the subject and the object when the preceding sentence’s word order is object-verb-subject. Thus reference for the two anaphoric forms cannot depend on a single measure of salience, because that measure would sometimes have to depend on word order and sometimes not. Kaiser and Trueswell suggest instead that pronominals like hän refer to the most salient element in the available syntactic/semantic representation, while demonstratives like tämä refer to the most salient element in the current discourse representation or mental model. These data not only provide a specific result of interest to understanding reference but also illustrate the importance of investigating such questions cross-linguistically. Eleni Miltsakaki, in chapter 14, makes use of Centering Theory as a framework for another examination involving antecedent salience. She assumes the traditional accessibility/salience hierarchy for referential forms in investigating the effect of subordinate clauses on topichood and salience. In particular, she examines whether subordinate clauses form independent utterance domains and whether entities referenced within such clauses interact with those

Introduction

9

in main clauses for the purposes of determining salience. She discusses several experiments on adverbial subordinate clauses, the results of which suggest that entity topic status is updated sentence-by-sentence rather than clause-by-clause and that entities in subordinate clauses are relatively less available as candidate topics. She then investigates relative clauses in a pair of corpus studies. She contrasts use of pronouns versus full noun-phrase referring expressions to refer to entities within preceding relative clauses, in light of the accessibility hierÂ� archy. The data show that pronouns are rarely used, and that when they are used the salience of an antecedent entity is based on a mention outside the relative clause in addition to within it. In the second corpus study, Miltsakaki computes Centering-Theory-based topic shifts from sentences containing a sentence-final non-restrictive relative clause into the following sentence and shows that these topic shifts are less coherent if the relative clause is treated as an independent utterance than when it is treated as a part of the main clause. Both of these corpus results are consistent with the salience results from adverbial subordinate clauses. They suggest that entities in subordinate clauses are treated differently with respect to salience and topichood than those in main clauses, and that subordinate clauses ought not to be treated as independent utterances for purposes of computing these properties. In chapter 15, Tony Sanford and Linda Moxey describe a range of experimental data that they have gathered over the course of almost two decades of research investigating how different determiners and quantifiers make different sets of entities available for reference. For example, in the sentence Every student passed the physics test., the quantifier every has two sets as its semantic arguments: the set of students and the set of physics-test-passers. Under one analysis of the semantics of every, the complete set of students has to be a subset of the set of test-passers in order for the sentence to be true. We can then refer to this subset in later sentences using a pronoun like they (e.g., as in They were very happy.). A strong claim from the linguistics literature is that possible meanings of determiners and quantifiers are constrained such that the only kinds of sets that determiners and quantifiers make available cross-lÂ�inguistically are subset relationships, but never complement sets. This claim is made explicit in Discourse Representation Theory (Kamp and Reyle 1983). If this generalization is right, there should never be a case in which we can use a pronoun to refer to the students who did not pass the test after a sentence like student passed the physics test. However, Sanford and Moxey demonstrate that negative quantifiers like few seem to permit reference to a complement set, contrary to Discourse Representation Theory. For example, in the following sequence of sentences, the pronoun they refers to the set of students that did not pass the test:

10


Few students passed the physics test. They couldn’t be bothered to study the night before. This is reference to the complement set. Sanford and Moxey discuss possible objections to their analysis and provide further empirical evidence in support of their claim that complement-set relationships can be made available by cÂ�ertain determiners and quantifiers. They then speculate on what possible cÂ�onstraints there are for reference with respect to quantifiers, if any, in the semantics of discourse more generally. In the closing chapter, Peter Hagoort outlines a broad-coverage model of language comprehension ( based on Vosse and Kempen 2000) and examines how it fits with neuropsychological data on sentence processing. He first summarizes the main language-relevant waveforms identified in the literature on recording of event-related potentials (ERPs), then points out the contrast in the literature on syntactic processing between syntax-first and more interactive models of sentence processing, then discusses ERP data that present difficulties for syntax-first models. Arguing for an interactive model in which information from multiple sources can be used to guide processing, he shows how cases in which syntactic processing fails in the Vosse and Kempen model can be mapped onto different ERP waveforms. With respect to semantic processing, Hagoort focuses on the question of whether computing the fit of newly processed material to existing sentential versus discourse context involves two separate stages. He describes ERP data that show largely identical patterns for violations in the two cases and argues for a model in which linking to a discourse model ( built up both within and across sentences) is crucial. Finally, he links elements of the cognitive model he proposes to particular regions of the brain. He suggests that the left posterior superior temporal gyrus may be responsible for storage and retrieval of lexical information (including syntactically relevant information stored with lexical items), while the left posterior inferior frontal cortex is responsible for integrating the retrieved information. Whereas Hagoort’s paper does not focus specifically on reference (and, indeed, the “binding problem” he discusses is not the linguistic one that arises for referentially dependent forms, but rather the more general problem of how the brain is able to link together various pieces of an utterance which arrive spread out over time), anaphora is of course a paradigmatic case of the requirement that language imposes on a comprehension system to track and combine information across time; consequently, a better understanding of the psychological and neural bases of human cognitive processes involved in combining linguistic elements more generally should inform models of reference resolution in particular.

Introduction

11

References Almor, A. 1999. Noun-phrase anaphora and focus: The informational load hypothesis. Psychological Review 106, 748–765. Chien, Y. C., and Wexler, K. 1990. Children’s knowledge of locality conditions in binding as evidence for the modularity of syntax and pragmatics. Language Acquisition 1, 225–295. Chomsky, N. 1981. Lectures on Government and Binding. Foris. Christianson, K., Hollingworth, A., Halliwell, J., and Ferreira, F. 2001. Thematic roles assigned along the garden path linger. Cognitive Psychology 42, 368– 407. Grice, H. P. 1975. Logic and conversation. In P. Cole and J. Morgan (eds.), Syntax and Semantics 3: Speech Acts. Academic Press. Grodzinsky, Y., and Reinhart, T. 1993. The innateness of binding and coreference. Linguistic Inquiry 24, 69–101. Kamp, H., and Reyle, U. 1993. From Discourse to Logic: Introduction to Model-Â� Theoretic Semantics of Natural Language, Formal Logic and Discourse Representation Theory. Kluwer. Novick, J. M., Trueswell, J. C., and Thompson-Schill, S. 2005. Toward the neural basis of parsing: Prefontal cortex and the role of selectional processes in language comprehension. Journal of Cognitive, Affective, and Behavioral Neuroscience 5, 263–281. Reinhart, T. 1983. Anaphora and Semantic Interpretation. Croom Helm. Reinhart, T., and Reuland, E. 1993. Reflexivity. Linguistic Inquiry 24, 657–720. Reuland, E. 2001. Primitives of binding. Linguistic Inquiry 32, 439– 492. Tanenhaus, M. K., Spivey-Knowlton, M. J., Eberhard, K. M., and Sedivy, J. C. 1995. Integration of visual and linguistic information in spoken language comprehension. Science 268, 1632–1634. Thornton, R., and Wexler, K. 1999. Principle B, VP Ellipsis, and Interpretation in Child Grammar. MIT Press. Trueswell, J. C., and Gleitman, L. 2004. Children’s eye movements during listening: Developmental evidence for a constraint-based theory of sentence processing. In J. M. Henderson and F. Ferreira (eds.), The Interface Of Language, Vision, and Action: Eye Movements and the Visual World. Psychology Press. Trueswell, J. C., Sekerina, I., Hill, N. M., and Logrip, M. L. 1999. The kindergartenpath effect: Studying on-line sentence processing in young children. Cognition 73, 89–134. Vosse, T., and Kempen, G. A. M. 2000. Syntactic structure assembly in human parsing: A computational model based on competitive inhibition and lexicalist grammar. Cognition 75, 105–143.

Iâ•…

Children’s Acquisition and Processing of Reference

2â•…

Cues Don’t Explain Learning: Maximal Trouble in the Determiner System Ken Wexler

Development of the Computational System of Language

This paper is devoted to a theory (and discussion of some relevant data) of the development of the definite and indefinite determiner in children, with special attention paid to the distinction between competence ( knowledge) and performance ( processing in this case). In psycholinguistics, the study of performance is dependent on the study of competence. Ever since Chomsky formulated the competence/performance distinction in modern generative grammar, empirical research has yielded a large amount of evidence for the fruitfulness, even necessity, of such a distinction. The reason it is so easy to forget this distinction in adult psycholinguistics is that the competence basis underlying research comes largely by assumption — competence is determined by linguistic theory, and psycholinguists routinely start from a linguistic description (e.g. a description of the structure of a relÂ� ative clause) and try to understand performance factors (e.g. difficulties, or errors) by adding performance assumptions to the linguistic theory. Of course, this is an oversimplification — sometimes the correct linguistic description is unknown, or at issue, sometimes performance considerations help in determining the competence, etc. Nevertheless, it is a truism to say that, by and large, the competence assumed in psycholinguistic studies is the competence as determined by linguistic theory. Such an easy assumption does not pass over to child psycholinguistics. For children, it is an empirical issue whether or not the competence described by linguistic theory applies to them, especially whole cloth. This is why most studies in the generative tradition (maybe most studies in all traditions) have started by asking what the competence of children is. For only when we have some notion of competence can we begin to investigate performance. In many areas of syntactic development, a field that has seen great progressÂ€ in the last 15 years (see Guasti’s 2002 textbook for an introduction),

16

Wexler

pÂ�sycholinguists have established many results about children’s competence. In broad outline, children know Universal Grammar (UG) and much about the particular properties ( parameters being the most central case) of their own language. But there are properties in which children’s knowledge (computational systems, CS) differs from adults’. In particular, there are slight differences between children’s representation of UG and that of adults. The idea is that the child’s CS grows into the adult’s. For example, there is general agreement that children pass through the Optional Infinitive stage of grammatical development because of a restriction on their grammars; proposals include the Tense Omission Hypothesis of Wexler (1990, 1994), the Truncation Hypothesis of Rizzi (1994), the ATOM model of Schütze and Wexler (1996), and the Unique Checking Constraint (UCC) of Wexler (1998). The latter predicts many syntactic phenomena through about age 2;6 or 3;0, including which languages go through the optional infinitive stage, properties of case, etc. The child has set the correct relevant parameter values, but the restriction causes the CS to take finite sentences as ungrammatical. All the papers I’ve just referenced state a syntactic (maturational) restriction on the CS and attempt to explain the phenomena via this restriction. Another example is the conclusion that verbal passive and unaccusative sentences are taken as ungrammatical by the child’s CS up to about age 5; again, this is not due to a parameter mis-setting, but rather to a restriction on the child’s grammar, for example the A-Chain Deficit Hypothesis of Borer and Wexler (1987, 1992), its re-statement by Babyonyshev, Ganger, Pesetsky, and Wexler (2001), or the Full Phase Hypothesis of Wexler (2004). There are several other examples, but this is not the place to review them. The point to remember is simply that in syntactic development, a field that has made great strides, the conclusion is that the child’s competence, her CS, has developmental restrictions that are not present in adult grammar. The assumption is that the growth of the child’s brain eliminates these restrictions, expanding the range of syntactic competence.1 In contrast, the study of the acquisition of semantics has for the most part not undergone such a development. The major approach within generative studies, at least, has been to assume that the child knows everything about semantics but that pragmatic development is delayed (e.g. Wexler and Chien 1985; Chien and Wexler 1990; Wexler 1999; Thornton and Wexler 1999). Stephen Crain has argued likewise that children know the universal properties of semantics; see, for example, Crain and Thornton 1998. The paradigmatic example is the argument by Wexler and Chien (1985) and Chien and Wexler (1990) that the Apparent Delay of Principle B effect is due to pragmatic difficulties with reference, rather than to lack of knowledge of syntactic or semantic principles. Briefly, Wexler and Chien (1985) showed that children to about age 5 accept

Cues Don’t Explain Learning

17

sentences like Mama Bear pointed to her as meaning that Mama Bear pointed to herself. This appears to be a violation of Principle B, but Wexler and Chien (1985) proposed that Principle B was not violated by the children. Rather, a pragmatic principle of reference was violated. Chien and Wexler (1990) showed that indeed it was a pragmatic/referential principle that was at stake by experimentally demonstrating that in bound variable contexts like every bear pointed to her (where the referential principle was irrelevant) children did not violate Principle B.2 We can say that the standard “working strategy” of semantic development (to date) is the External Hypothesis: Semantic properties of UG are known to the child, but “external” properties may be delayed. That is, the External Hypothesis says that all aspects of the computational system of semantics are known to the child but that linguistic properties external to the computational system (e.g. pragmatic properties, or properties of processing) may not be fully developed in the child. In this paper, I challenge the External Hypothesis, at least for the development of determiners. In this regard, I seek to bring syntactic development and semantic development closer together, and to provide a motivation for semantic development other than the “confirmation” of linguistic theory in the child. The title of this paper is a pun — it uses maximal in two senses. On the one hand, the working strategy of the paper is to cause maximal trouble to the standard External Hypothesis, by arguing that children actually have a different representation of determiners than do adults; there is a difference in the semantic representations of determiners. On the other hand, maximal trouble also refers to the difficulties that I will claim children have with determiners, namely with the assumption of maximality for the definite determiner. I will discuss the classic empirical problems with use of determiners, discuss the standard “pragmatic” solution, illustrate why this solution is not complete, propose a new solution based on difficulties with the semantic feature of maximality, and show how this analysis captures the acquisition facts. I will then discuss a potential processing/memory explanation and show that it fails to capture the data. Lastly, I will show how “on-line” data using eye tracking cannot be explained using the hypotheses suggested in the literature, but follow from the hypothesis of this paper. The the → a Error in Child Language

There is a large empirical literature on the development of definite and indefinite determiners; I will mention only a few of the studies, but the generalizations I discuss are widely accepted, if not always explicit.

18

Wexler

Karmiloff-Smith (1979) (KS subsequently) is a major study of the development of French determiners. These experiments involve many children and age groups and give us a very clear idea of the overall patterns of development. Results for English (e.g. Warden 1974, 1976) confirm the French results, and I’ll discuss some English results also. For convenience, I’ll use the and a instead of the French translations. Consider when a N or the N is used, for the restrictor noun N. Intuitively, when there is a unique salient instance of N in the discourse (overt or assumed discourse), the definite determiner the is required. Otherwise, a is used.3 I will provide the standard formal definition later in the paper. KS’s Experiment 1 varies whether the object to which the child has to refer is a singleton (requiring the Xâ•ƒ), one of several different-colored objects (requiring the plus adjective, e.g. the blue Xâ•ƒ), or one of several identical objects (requiring a Xâ•ƒ). Here is a sketch of the set-up: The child is told (and sees) that there is a girl-doll’s playroom and a boy-doll’s playroom, with various toys in the playrooms. Here is one relevant instance: girl-doll’s playroom one blue book three multicolored balls one baby-bottle

boy-doll’s playroom three books (each a different color) one multicolored ball one car

The experimenter touches an object (e.g. the blue book in the girl-doll’s playroom) and the child is told to “ask the girl to lend you that.” The child replies and the form the child uses is noted; e.g. the child might say “lend me the blue book” (the definite response; in fact the correct adult response, since the girldoll has exactly one blue book) or “lend me a blue book” (the indefinite response). On the other hand, if the experimenter touched one of the three identical multicolored balls in the girl-doll’s playroom, we would expect an adult to use the indefinite — that is, to say something like “lend me a ball.” Demonstratives (e.g. that book) were discouraged. The age range of participants was 3;0 –9;11, with six to eight children in each year (five in the last year). For most age ranges, the total number of rÂ�esponses over the age range was 50 –70, so there is a substantial amount of data. For simplicity, I’ll summarize results here for only three age groups, ignoring the case of several different-colored objects (which supports the conclusions drawn), and I’ll add together KS’s separate counting of whether the referring expression has a modifier or not, simply counting definite or indefinite. The percentages of indefinite and definite responses for a particular sÂ�timulus (single object or several objects) for the 3-year-olds do not total


19

Table 2.1

Percentage of definite and indefinite determiners (based on Karmiloff-Smith 1979). Age

Type of determiner

Singleton object

Several identical objects

3;0–3;11

Definite Indefinite

â•‡ 62 â•‡â•‡ 0

â•‡ 48 â•‡â•‡ 3

6;0–6;11

Definite Indefinite

â•‡ 92 â•‡â•‡ 8

â•‡ 56 â•‡ 44

9;0–0;11

Definite Indefinite

100 â•‡â•‡ 0

â•‡â•‡ 0 100

100Â€because, despite efforts, these children produced some demonstrative responses, which make up the remainder of the 100 percent. In table 2.1, we see that 9-year-olds are “perfect” — i.e., they agree with adult judgments, always using the definite (the) for the singleton and the iÂ�ndefinite (a) for the non-singleton. So adult controls aren’t necessary; the 9-year-olds produce behavior that agrees with adult judgments (and the results of linguistic theory). They can serve as controls, showing that in fact the tÂ�heory’s predictions (and adult judgments) are perfectly replicated in the behavior of children who are old enough. Turning to the younger children, we see that they almost never produce an indefinite for the singleton situation (0 percent for the 3-year-olds, 8 percent for the 6-year-olds), so they know that singletons take the. But younger children (including 6-year-olds) very often use the definite for the “several identical objects” case, where the indefinite is required; e.g., asking the girl-doll for one of the balls, the child will say “lend me the ball” instead of “lend me a ball” (48 percent for the 3-year-olds, 56 percent for the 6-year-olds). This is the classic aâ•–→â•–the child response (still in existence in this experiment at age 6, completely resolved by age 9). That is, a required a is instead produced as the. Crucially, the error is not a random response; children know to use the for the singleton. And it’s not lack of knowledge of a. Children use a correctly when there is no context set, as other experiments show. Now let’s briefly consider KS’s Experiment 5: girl / boy acting. There are three identical objects, e.g. pigs. A girl-doll knocks over one of them and the child is asked “what did the girl do?” Alternatively, there is a singleton object (e.g. a pig), the girl-doll knocks it over, and the child is asked “what did the girl do”? Five-year-olds and 6-year-olds responded 100 percent with the X to the singleton case (e.g. saying the girl knocked over the pig). For the three identical objects, they used the definite about 60 percent of the time, showing overuse of

20

Wexler

the definite article. Ten-year-olds used the definite article only about 13 percent of the time in the case of three identical objects. Again, we see the aâ•–→â•–the eÂ�rror for the 5-year-olds and 6-year-olds, with perfect responses when the was required, i.e. the non-existence of a theâ•–→â•–a error. In this same experiment, the girl then knocked over another X (e.g. another pig), and the child was asked what she had done. Ten-year-olds used the X here only about 11 percent of the time and used various kinds of adequate expressions (mostly another X, sometimes one of the X’s) about 83 percent of the time. Five-year-olds used the X about 56 percent of the time and the indefinite article about 36 percent of the time. Again, we see the aâ•–→â•–the error. To emphasize: Five-year-olds (and 6-year-olds too) very often used the X to refer to a different identical item; e.g. they had just talked about a pig, and then they referred to a different pig as the pig. It seems as if the first use of the pig would have established that pig as the context set, disallowing use of the to refer to another pig (as the 10-year-olds showed), but the 5-year-olds and the 6-year-olds referred to this other pig as the pig. Clearly something is amiss. There are many other experiments in the literature confirming the asymmetry of the errors (i.e., the great prevalence of the aâ•–→â•–the error) in children through at least age 5, but I don’t have the space to review the literature here. Let me just point out some of the evidence that the error isn’t due to lack of knowledge or to inability to use a. In KS’s Experiment 2, Hide and Seek, children have to talk about what is in a bag under various conditions. Five-year-olds overuse the in the case of several identical objects under conditions similar to the previous experiments. But when the experimenter asked children “What’s in the bag?” under conditions in which a discourse set wasn’t established, they almost always used an X (not the Xâ•ƒ) (rates of the were 9 percent, except 1 percent for one group and 3 percent for one other group). Sometimes the younger children omitted the article, but the correct use of a was never less than 80 percent for any of the groups. So children know a, and use it in many situations, despite the fact that they often use the X in the case of a context involving several objects X. This particular situation was a kind of naming situation, but there are also other kinds of situations in which a is used correctly, namely situations in which no context set of X’s has been set up, in what are called de dicto contexts. In Maratsos’s (1976) stories experiment (examining English), children have to say or guess something about a category X. In some particular conditions of the experiment, situations are set up such that no specific example of X has been introduced, but the child has to say that any instance of X is what is being sought. In these situations, a X is the correct way to introduce X. For example, in the lookingfor story, it is made clear to the child that a man is looking for a lion or a zebra,


21

but no particular lion or zebra is introduced (essentially these are de dicto contexts). The man is looking for any old lion or zebra. There is no context set of lions or zebras. The child has to say (guess) what the man is looking for. Two groups of 4-year-olds answered correctly with the indefinite a lion or a zebra 91 percent and 98 percent of the time, respectively. Clearly these children (who overuse the in contexts where there is a salient discourse set of N’s) know and use a appropriately in other contexts. We can conclude, then, that there is a specific problem with the choice of a or the in particular semantic situations — the problem is not a lexical gap in the child. The Egocentric Theory

Why do children make the aâ•–→â•–the error? The traditional explanation for the error is that it is an “egocentric response.” Maratsos (1976) used this term, following Piaget (1955), although Karmiloff-Smith has essentially the same explanation. The idea is that children have trouble taking the listener’s point of view. Because the X is used when the discourse context makes exactly one X salient, this discourse context must be known to both the speaker and the listener. The child, though, might use the so long as the child has one salient referent in mind, ignoring whether the listener does also. Piaget (1955, p. 115, as quoted by Maratsos, p. 11) writes: “.â•–.â•–. the explainer always gave us the impression of talking to himself, without bothering about the other child [to whom he or she was telling a story or explaining mechanical workings.] Very rarely did he succeed in placing himself at the latter’s point of view.” Maratsos ( p. 63) writes: “Our analyses have uncovered a developmental stage where egocentric definite responding is quite common. The children fail to take into account that even if they have established for themselves a particÂ� ular boy or girl, or monkey or pig that does something, that referent is not yet uniquely specified for their listener, and must be introduced to the listener with an indefinite expression.” Similarly, Karmiloff-Smith ( p. 72) writes: “It is suggested that .â•–.â•–. the use of definite referential expression [by young children] is in fact deictic. Like the demonstrative, the definite article points to an object under focus of attention and is not yet an exophoric reference to the singleton characteristic of the referent in relation to its current context.” Maratsos’s explanation is essentially pragmatic; the child ignores whether or not the listener knows which object is under discussion. Karmiloff-Smith’s explanation is similar; so long as there is a single object in the focus of attention of the child, the is used no matter whether that single object is in the “current context,” i.e., is known to the listener.

22

Wexler

Wexler and Chien’s explanation of the Principle B effect, as discussed earlier, took the error to be due to the child’s not paying attention to what the listener knew. Avrutin and Wexler (1992) explicitly related this property to the overuse of the and the pragmatic (egocentric) explanation for that use. Vinnitskaya and Wexler (2001) related exactly the same property to overuse of imperfect aspect by children speaking Russian. So the standard view in developmental studies of semantics has come to involve the External Hypothesis, discussed earlier, that all semantic properties ( properties of the computational system of language) are known to the child but that external properties (e.g. paying attention to the listener) may not have developed. The Egocentric Theory Doesn’t Account for All the Data

Thus, although the egocentric view has been standard in both generative-based and non-generative-based developmental psycholinguistics, we may have to awake from a dogmatic slumber. Can the pragmatic view explain all the developmental results on determiners? Let’s try to apply the egocentric theory to Maratsos’s stories experiment, which Karmiloff-Smith replicated with many more participants and age groups. The experimenter tells a child a story, eliciting a response. There were either several boys and girls (in the I (Indefinite) version Xsâ•–→â•–a Xâ•ƒ) or one boy and one girl (in the D (Definite) version, a Xâ•–→â•–the Xâ•ƒ). For example, one of Maratsos’s stories, the Making Noise story, is as follows: Once there was a lady. She had (I version: lots of girls and boys, about four girls and three boys; D version: a boy and a girl ). They were very noisy, and they kept her awake all the time. One night she went to bed. She told them to be very quiet. She said, ‘If anyone makes any noise, they won’t get any breakfast tomorrow.’ Then she went to bed. But do you know what happened? One of them started laughing and giggling. [If the I version was being told, the experimenter said something like ‘Now let’s see, there were four girls and three boys.’] .â•–.â•–. Who was laughing and giggling like that?

The nineteen 4-year-olds were divided into 4-High and 4-Low groups (with roughly same number of children in each) on the basis of an imitation task. Their mean accuracies are presented in table 2.2. When there was a single boy and a single girl, all the children gave close to 100 percent correct answers, i.e. the X. When there were several boys and girls, the 4-Low group used the X (incorrectly) about 58 percent of the time. This is the classic aâ•–→â•–the error; whereas when the is required, it is used correctly. WeÂ€ can take the 4-High group to have already developed the adult pattern. Karmiloff-Smith replicated Maratsos’s study (in French), but with a much


23

Table 2.2

Percent correct (modified from Marastsos 1975). Age group

a X expected

the X expected

4-Low 4-High

42 98

94 97

Table 2.3

Percent correct (based on Karmiloff-Smith 1979). Expected response definite

Expected response indefinite

Age

Definite

Indefinite

Other

Definite

Indefinite

Other

6 years 10 years

â•‡ 83 100

9 0

8 0

63 â•‡ 0

37 89

â•‡ 0 11

larger population (68 children, from 3 to 11 years old) and a large number of items. A sampling of results is presented in table 2.3. In KS’s study, the 6-year-olds give 63 percent definite responses — a much larger percentage of error than Maratsos’s 4-year-old group.4 Nevertheless, both studies agree on the strong overuse of the in the story experiment. Maratsos and Karmiloff-Smith take it that the egocentric theory explains these results. But on inspection, there seems to be a problem for this view. I thank Irene Heim for pointing out (in our seminar on Semantics and Acquisition; also see the talk she gave at the Sixteenth Annual CUNY Conference on Human Sentence Processing) that the egocentric/pragmatic hypothesis doesn’t seem capable of explaining the results. Consider the Making Noise story. The child hears a story about “lots of boysÂ€and girls” but isn’t shown the boys or the girls. No discussion of any of the boys or girls takes place; no individual properties are discussed. In short, no individuation of the boys or the girls takes place at all. The child has no boyÂ€or girl to concentrate on, or, in KS’s words, to put “in focus.” So when theÂ€child answers the boy, there can’t be any boy the child has focused on; at least, there can be no boy that the child can infer from what has been presented. For the Egocentric Theory to explain what the child does by saying that the child is concentrating on a particular boy in the set, the child would have to have invented this boy from whole cloth. While not necessarily impossible, it seems unlikely, and it is at least against the spirit of the explanation that says the child sees or has a concept of a particular boy, one of many, perhaps, but the one the child is concentrating on. This idea might explain, say, Karmiloff-

24

Wexler

Smith’s playroom data (table 2.1), but not the story experiments (tables 2.2 and 2.3). The Child’s Problem with Maximality

Let’s go back to square one and rethink what the child might be missing, so that the overuse of the can be understood both when the child has an object in focus and when she doesn’t. It seems natural to think that the error the child makes is to think that the X is to be used if there are any salient elements in the dÂ�iscourse; there doesn’t have to be exactly one. Such an analysis is natural as soon as one looks at the actual semantic-theory definition of the definite determiners, and why the and a are used in particular contexts. The standard (Fregean) analysis of the definite determiner in semantic theory (for now concentrating only on the case of singular NP’s) posits that the use of the N presupposes that there is exactly one N in the relevant context. If there are no N’s in the context, or if there are multiple N’s, the expression is infelicitous. The formal definition of the Fregean analysis in (1) below is taken from (47) of Heim (1991, English translation). The index i refers to the situational context.5 (1)â•‡Regardless of the utterance context, [the x] P expresses that proposition which is: true at an index i, if there is exactly one x at i, and it is P at i; false at an index i, if there is exactly one x at i, and it is not P at i; and truth-valueless at an index i, if there isn’t exactly one x at i. How can a x be used, e.g. a lecturer arrived ? The standard analysis of a x is simply existential, as in the formal definition in (2) below, taken from Heim’s (112): (2)â•‡A sentence of the form [a x] P expresses that proposition which is true if there is at least one individual which is both x and P, and false otherwise. Note that there is no “context” index specified in the definition of a; it is simply existential, i.e. at least one x with property P must exist. A classic question in semantic theory is “Why can’t a X be used when it is known that there is exactly one x?” For example, (3) (Heim’s (118)) is not acceptable; we have to say the father, not a father. Yet (2) would seem to make (3) perfectly acceptable. (3)â•‡ ?I interviewed a father of the victim. Semantic theory concludes that (3) is true, but that there is a rule of the form (4) (Heim’s (123)).


25

(4)â•‡In utterance situations where the presupposition for [the x] P is already known to be satisfied, it is not permitted to utter [a x] P. Since it is generally known that a person has only one father, the father of the victim is appropriate ( presuppositions are satisfied), so (4) prevents a father from being used. Heim ( p. 28) suggests that the general principle from which (4) can be deduced is (5) ( her (62)), a pragmatic principle that, Heim argues, is different from Gricean maxims (and different from whatever is responsible for scalar implicature). (5)â•‡ Make your contribution presuppose as much as possible. I’ll call this principle Maximize Presupposition. Thus, the adult use of the definite and indefinite has three parts: the semantic analysis of the (3), the semantic analysis of a (4), and the pragmatic principle Maximize Presupposition (5). So what is the child missing? Crucially, [the x] P has no truth-value (i.e. can’t be used) if there are no x’s at i, or if there are multiple x’s at i. We know from the experiments that children use the X when there are multiple X’s. On the other hand, when there are no X’s in the context (as in the naming experiment of KS, and as in the looking for (de dicto) type story of Maratsos), the child does not use the. We want the child to correctly use a X where there are no X’s in the context. Children seem to know that there must be a context set of X’s in order for the X to be used, but they don’t seem to know that there has to be exactly oneÂ€X; they seem to think that the X can refer to an X even if there are multiple X’s. Here is the idea for my analysis: Children take the N to presuppose the existence of an N. But they don’t have the uniqueness condition on N. So in Maratsos’s Making Noise story, the lady has lots of boys and lots of girls and one of them cries. The child answers the boy because the child presupposes that there is a boy (at least one) and asserts that a boy is crying. In other words, the child has a presupposition of existence of an N for the N to be useable, but doesn’t have a presupposition of uniqueness. The major proposal of my analysis is that the child has a different lexical entry for the, which I will call theC (the child’s analysis of adult the) and which is pronounced the for the child.6 A formal definition is given in (6). (6) Regardless of the utterance context, [theC x] P expresses that proposition which is true at an index i, if there is an x at i, and it is P at i;

26

Wexler

false at an index i, if there is an x at i, and there is no x such that x is P at i; and truth-valueless at an index i, if there is no x at I. This is modeled directly on the definition of the adult analysis of the x in (1), but presupposes only that there exists at least one x at i; the x doesn’t have to be unique. If there is at least one x that is P, the sentence is true. If there is at least one x but no x’s are P, the sentence is false. If there are no x’s, the sentence is truth-valueless. In terms of truth-value, theC Xâ•–=â•–one of the X’s when there are multiple X’s in the context, and theC Xâ•–=â•–the X when there are no X’s or exactly one X in the context. The child’s analysis of the definite and indefinite determiner system has three parts. In contrast to the adult analysis in (1, 2, and 5), the child has the semantic analysis of theC (6), the semantic analysis of a (2) and the pragmatic principle Maximize Presupposition (5). The child differs only in the semantic analysis of the; otherwise the analysis is identical to the adult’s; a receives the existential definition and Maximize Presupposition holds. The consequences of this analysis are as follows. When there are multiple X’s in the context, the child will use the X, in accordance with the experiments discussed above. This follows from the definition of theC plus Maximize Presupposition, which induces use of theC instead of a. This is the aâ•–→â•–the error. When there are no X’s in the situation, the child will (correctly) use a X, since the presupposition of existence of X for theC is not satisfied. This too is in accordance with experimental results. When there is exactly one X in the situation, the child will (correctly) use theC X, in accordance with the experimental results on singletons. If there is no context set at all, and the child is simply expressing a wish, the prediction is that the child will say, e.g., I want a cone (answering a question, such as What would you like to eat?, in which no choices are given or assumed). This is the right empirical result, as in Maratsos’s (1976) looking for story. (The prediction follows from the definition of theC cone, which presupposes at least one cone in the context.) The assumed difficulties with uniqueness capture all the facts I have discussed. In semantic theory, uniqueness generalizes to maximality when plural sets are considered (Heim 1991). Informally, the oranges are large is true if the maximal group of oranges in the context set of oranges (the ones in the situation) has the property that they all are large. The sentence is false if the oranges in this maximal set aren’t all large. And the sentence lacks truth-value if there is no maximal set (e.g. there are no oranges at all in the context). It can be


27

shown that in the singular case, the property of maximality reduces to existence and uniqueness (see Heim 1991). The natural move for the current developmental theory is to generalize the definition of theC to maximality, as is done in Wexler 2003. This predicts errors by the child on plural uses of the, which can be shown to occur empirically. There are many more results in the empirÂ� ical literature, including comprehension experiments. In Wexler 2003, I discuss many of these results and show that troubles with maximality underlie these errors in exactly the way we have shown for the experiments discussed in the present paper. For completeness I should mention that there is at least one other theory in the literature that claims that the aâ•–→â•–the error is due to a semantic error rather than to egocentrism. Matthewson and Schaeffer (2000) suggest that children assume that the refers to a specific entity rather than to a definite entity. That is, children do not have a presupposition of existence. Matthewson had earlier argued that Salish contains an article with this property, so Matthewson and Schaeffer suggest that children learning English might mis-set the parameter to the Salish value. For a discussion of predictions made by this hypothesis, see Ionin, Ko, and Wexler (2004), which investigates whether this property might be relevant to second-language acquisition of determiners. Most of the empirÂ� ical tests appropriate to this hypothesis have not been carried out for children. I don’t see how the hypothesis can explain Maratsos’s stories experiment, but it is worth noting that there might be a variety of hypotheses concerning semantic properties of determiners in children. A Possible External Hypothesis Analysis

On the theory I proposed in the preceding section, the semantic analysis of the is problematic for the child; the pragmatic principle of Maximize Presupposition must be in place. There is no processing difficulty, for example, with respect to Maximize Presupposition. So the External Hypothesis is wrong, on this analysis. Is there an External Hypothesis alternative? Let us consider one idea in the literature about the child’s processing difficulties that one might useÂ€to attempt to explain the empirical results on determiners. Grodzinsky and Reinhart (1993) attempt to modify Wexler and Chien’s (1985) and Chien and Wexler’s (1990) pragmatic explanation (difficulty with the referential prinÂ� ciple) of the Apparent Principle B Delay discussed above. They accept the claim of the previous authors that there is a basic split between the syntactic and referential behavior. But instead of the pragmatic difficulty with referenceÂ€(ability to infer what other minds know; Wexler (1999)), Grodzinsky and

28

Wexler

Reinhart claim that there is a processing difficulty with reference. In particular, they claim that the poor referential behavior stems from the difficulty of keeping two representations in mind. When two representations have to be compared (necessary on their account for Principle B contexts), the child breaks down and guesses (see Reinhart, this volume). Call this No Comparisons. Thornton and Wexler (1999) present a discussion of difficulties with this processing account for Principle B effects. Here let us just see what happens when we apply this idea to determiners. Let us accept the standard semantic theory of determiners as embodied in (1), (2), and (5) above. The only aspect of this theory that demands comparison of representations is (5), Maximize Presupposition. Suppose that a should be used — that is, the presuppositions of (2) are satisfied (there are no presuppositions in (2)); and a unique entity doesn’t occur in the context set, so the presuppositions of (1) aren’t satisfied, and thus the can’t be considered. A child who has the adult semantic theory will not make an eÂ�rror on a given No Comparisons, because there is no situation in which two representations have to be compared. Now suppose that there is a unique entity in the context set. This means that the presuppositions of (1) are satisfied and the can be used. But the presuppositions of (2) (none) are also satisfied. So a can be used. The choice between (1) and (2) is determined by Maximize Presupposition (5), which chooses the. But a child subject to No Comparisons will not be able to always correctly compare the two representations (for a and the) to see which has greater presuppositions. Thus, if we apply Grodzinsky and Reinhart’s analysis, the child should guess in these situations and should use a half the time and the half the time. In other words, applying Grodzinsky and Reinhart’s analysis predicts that children will make no errors in contexts where a is correct, but will make eÂ�rrors in contexts where the is correct. This predicts an asymmetry, with many theâ•–→â•–a errors but no aâ•–→â•–the errors, exactly the opposite of the pattern that holds. Thus one well-known version of the External Hypothesis for grammatical representations (a particular version of a processing difficulty, No Comparisons) makes the wrong predictions for the determiner case. I don’t know of any other version of the External Hypothesis that might work, but of course any proposals would have to be considered on their empirical merits. How does Maximality develop in children? The most reasonable explanation seems to lie along the lines of brain-based linguistic maturation (Borer and Wexler 1987; Wexler 1999, 2002; Babyonyshev et al. 2001). If it were a question of learning, Borer and Wexler’s Triggering Problem would exist: Why


29

doesn’t the evidence that allows for triggering at a later age not work at the much younger age? Similarly, Babyonyshev et al. (2001) present their Argument [for maturation] from the Abundance of the Stimulus: Since there is so much evidence available to the child, why does it take so long for the child to learn if there is no maturational factor?7 It is possible that maximality itself is a notion that develops maturationally, in which case we would expect any construct that uses maximality to be delayed in the same age range as the. There is some evidence (Chien and Wexler 1990) that children until age 5 will apply every X to a subset of all the instances of X in the situation, rather than limiting its application to the entire set. It is possible that this error is due to an error on maximality. Do Children Fail to Process Interpretations On-Line? Some Problems with a Multiple-Cue Theory

Trueswell et al. (1999) claim that they have uncovered a special problem in children around age 5 that prevents them from accessing interpretation on-line in processing. They call this the “Kindergarten-path Effect.” Briefly, the argument goes as follows: Sentences like (7) are temporarily ambiguous. (7)â•‡ Put the frog on the napkin in the box. Before in the box is heard and processed, it is possible that on the napkin is the Destination ( VP attachment) or on the napkin is a Modifier ( N attachment) (â•ƒ frog on the napkin). Of course, once in the box is processed, only the Modifier analysis is possible; the sentence is not ambiguous. Adults (and children) have a famous processing tendency to take the Destination reading at first in (7), therefore garden-pathing. There are many explanations for this in the literature. (For some examples, see Frazier and Fodor 1978; Frazier 1987; Trueswell and Tanenhaus 1994.) Choosing between them is not relevant to the present discussion; we simply assume the best eÂ�xplanation. It is known that in adults the Destination Preference can be suppressed by the referential situation (Tanenhaus et al. 1995). Trueswell et al. replicate the result with an eye-tracking experiment with the following set-up: One-frog situation: one frog on a napkin, one empty napkin, one cow, one box Two-frog situation: one frog on a napkin, another frog (not on a napkin), one empty napkin, one box In the one-frog situation, adults very often garden-path (following Destination preference); their eyes look at the wrong destination (the napkin) very often. In

30

Wexler

the two-frog situation, adults garden-path hardly at all; they take the Modifier reading, frog on the napkin. For standard semantic theory, as I described it earlier, the reason that the two-frog situation eliminates the garden path is clear, if we assume only that the participant uses his or her semantic knowledge while responding. Use of the frog, via maximality, presupposes that there is a unique frog in the context set. Taking the relevant N to be frog on the napkin ensures that this presupposition of maximality (uniqueness) is met, since there is only one frog on a napkin. If instead, the Destination reading were taken, and the relevant N was frog, since there are two frogs in the situation, the presupposition of uniqueness for the is not satisfied. In order to interpret the sentence in accordance with their semantic knowledge (concerning the), adults overcome their processing tendency to take the Destination reason. In the one-frog situation, on the other hand, the presupposition of maximality (uniqueness) is satisfied even if the N is frog, unmodified by on the napkin. Thus the natural processing tendency leads to the Destination reading, a garden path. Clearly, semantic properties (the definition of the) are computed on-line and interact with processing tendencies to produce adults’ interpretations and parsing patterns. Trueswell et al. show that 5-year-olds, on the other hand, garden-path in both the one-frog and two-frog situations. Varying the referential context (number of frogs) does not influence their parsing decisions with respect to garden-pathing; this is seen in both eye-tracking results and behavioral (i.e. how they obey the command in (7)) results. Trueswell et al. do not provide an explanation for why children are different, but they seem to suggest that it has something to do with verb biases used in on-line processing. But verb biases do not explain why children garden-path in the two-frog situation. The question is why children don’t use the relevant semantic information provided by the and the context (two frogs) to determine that there is likely a modifier coming after frog. Trueswell et al. suggest that perhaps children can’t consider two alternative syntactic analyses, and thus they simply use the probabilistic cue associated with the likely Destination phrase. But this would be a really major assumption about the processing abilities of a child — that they can’t consider two alternatives. The theory, if believed seriously, would be likely to make major mispredictions about the ability of a child, though there is no room to discuss this here. Moreover, the conceptual idea is hardly well grounded in the processing theory that the authors advocate, the integration of “multiple cues.” What does it mean to not be able to consider alternatives in such a theory? Presumably only one cue could


31

be considered at a time by the child. If this idea could be worked out, it seems likely that it would lead to a total breakdown of the child’s ability to comprehend a sentence. Another potential explanation that Trueswell et al. seem to be offering is thatÂ€children don’t make use of referential or semantic properties in the course of processing. This is extremely unlikely to be true, given many experiments inÂ€ which children of the relevant age interpret sentences using contextual iÂ�nformation. For example, Wexler and Chien (1985) showed that children at age 5, when presented with two pictures, one of Cinderella’s sister washing herself and one of Cinderella’s sister washing Cinderella, and told Cinderella’s sister is washing herself, would select the correct picture (the one in which the sister is washing herselfâ•ƒ) more than 90 percent of the time. The point of the experimentÂ€was to test whether children know the property that a reflexive must be c-commanded by its antecedent. But the children could attain success only if they not only knew the appropriate syntactic property but also understood and took account of context. If children didn’t take account of context in parsing the sentence, they couldn’t appropriately match the sentence to the picture. There are hundreds of experiments in the literature, on many topics, that would make the same point — that children pay attention to context. Trueswell et al., working within an undifferentiated statistical “cue” model, attribute the difference between adults and children to many different properties, but not to the child’s knowledge or use of the. They (at least implicitly, to some extent explicitly) assume that children know all the relevant properties ofÂ€the and behave perfectly on the in off-line experiments. But of course, the classic developmental literature shows that children make many errors on the, as I have briefly summarized. To their credit, Trueswell et al. (this volume) do see (in response to this paper) that knowledge of the is centrally involved, and suggest that children have difficulty learning the properties of the because of unclear input. (See note 6 to the present paper for a comment on this sÂ�uggestion.) Consider another attempt to explain the effect: Hurewitz et al. (2000, p. 263) suggest that the child first makes an attempt to go with the most “reliable” cue (the preference of put for taking a Destination) and “considers this cue first .â•–.â•–. He never returns from the end of the garden path, because of the inability to maintain alternative analyses over an extended period of time.” It isn’t clear what isn’t reliable about the uniqueness property of the. The input to the child is quite clear on the fact that the X presupposes a unique X inÂ€the context set (where context set is partially determined by internal goals and purposes). So it isn’t clear that there is an objective relevant sense in which the preference of put for taking a Destination is any more “reliable” than the

32

Wexler

uniqueness of the entity named by the X (singular). Furthermore, since adults use the uniqueness property of the (which they do, in the experimental results), this must be a reliable “cue.” Perhaps what Trueswell et al. mean is not that the cue isn’t “reliable,” but that it isn’t “accessible.” Perhaps they think that children have difficulties with referential/interpretive properties of certain kinds, but not with syntactic properties. If so, this would be getting close to the “egocentric” theory that the kindergarten-path papers have ignored. Basically, no explanation along these lines seems to me to be possible because of what I see as the fundamental difficulty of the underlying model — the multiple-cue model. What the authors in this tradition (in particular the authors of the frog experiments) argue is that the “reliability” of a cue is the determining factor in whether it is “learned” early. Very reliable cues are learned very early, and less reliable cues are learned later. I know of no evidence for this fundamental assumption in the developmental literature, and it isn’t clear that a definition of “cue” that is empirically valid could be created in order to test the notion. To illustrate in just one relevant domain, let us try to predict the result on de dicto contexts in Maratsos’s stories experiment. In input to the child, does the object exist in the context or not? The answer is “sometimes.” Consider the de dicto verb look for. Suppose we say I’m looking for a pretty gift. This might be used when there are no gifts around ( just telling somebody about the situation). Or it might be used when there are several objects which are potential gifts in a store, and the speaker is looking at them. (Though of course the speaker isn’t referring to one of these objects. But that doesn’t matter from the standpoint of cue theory. So far as I understand, cue theory is concerned only with what is in the environment). So from the standpoint of cue theory, look for doesn’t provide a reliable cue concerning the existence of the referent of the direct object. Sometimes the referent object is there, sometimes it isn’t. The prediction would then follow that the ability to use look for as a de dicto verb would have to be quite late in children. How could they learn this from a non-reliable cue? (The analogy is to the “explanation” that the properties of the are late because of the “unreliability” of the relevant cues.) But recall that Maratsos (1976), in the looking-for story, showed that when it was made clear to 4-year-olds that a man was looking for a lion or a zebra, but no particular lion or zebra was introduced, the children answered correctly with the indefinite a lion or a zebra 91 percent to 98 percent of the time (two groups). Clearly these children had learned the “cue,” in the sense of this theory, despite the fact that the cue isn’t “reliable” in the relevant sense.


33

Other examples could be constructed at will by anybody familiar with the developmental language acquisition literature. The knowledge of the child is a complex internal system, not a system of learned “cues.” There is learning to be sure: parameter setting, learning of lexical properties, learning how to map phonetic forms of words to syntactic and semantic properties. But there is no reliable sense of “cue” in the external world that can be used to construct a theory along the lines of multiple-cue theory. Such a theory can make no predictions; it is in the position of Utility Theory in the social sciences, which allows one to always say that there must be some utility that is being maximized (yet no theory of utility is given). When a child learns something, multiple-cue theory will say “Well, that’s a reliable cue.” And if the child doesn’t learn something, it will say “Well, that’s not a reliable cue.” There is no theory of cues that has an empirical testing domain in the child. In my opinion, no such theory can be given. The kindergarten-path effect follows from the theory in this paper, with no recourse to special difficulties in child processing. If children are missing the presupposition of maximality from their analysis of the (or if they can’t compute maximality), then in the two-frog situation, the child can take the frog to refer to one of the frogs (either of the frogs). Thus there is no semantic reason to prefer the modifier analysis, no presupposition to be satisfied by taking a particular analysis. The only presupposition of the frog for the child is that there be at least one frog, and this is true in both the one-frog situation and the two-frog situation. Thus there is no reason for the child to prefer the modifier analysis in the two-frog situation. The natural tendency for the child (as for the adult) to prefer the Destination analysis wins out. None of the papers on the kindergarten-path effect refer to the rather extensive and useful experimental literature on the child’s competence concerning the,8 and for the most part the papers don’t even recognize that the difficult iÂ�ssue for the child concerns the. This brings us back to the point made at the beginning of this paper. The study of natural language processing in adults starts from a competence model; the competence model is derived from linguistic theory. It is this competence model that allows processing studies to proceed. We also need a competence model for children before we can start processing studies. Research on language acquisition has provided, in many areas, this competence model. Work on processing in children can proceed only if it starts from this competence model. In conclusion, with respect to kindergarten-path studies, children process language just like adults; the large differences in behavior follow from one small difference in analysis of the determiner, a difference for which there is extensive evidence. The processing studies are quite valuable; they confirm, using

34

Wexler

on-line methods (eye tracking), that children have essentially the same processing strategies or abilities as adults (e.g. whatever causes the garden path), and that they have the difficulties with the computation of the properties of the (maximality, as I have claimed) that have been discovered in off-line studies. It is very nice to have such an integration of results from different methods. The Revised Kindergarten-Path Explanation of Trueswell, Papafragou, and Choi

Elsewhere in this volume, Trueswell, Papafragou, and Choi ( henceforth TPC) respond to my paper and modify their theory accordingly. The editors asked me to take their paper into consideration in preparing the final version of my paper. In this section I consider the modified ideas presented by TPC. My paper is an investigation of the child’s knowledge of and use of determiners. I pointed out that “none of the papers on the kindergarten-path effect refer to the rather extensive and useful experimental literature on the child’s competence concerning the, and for the most part the papers don’t even recognize that the difficult issue for the child concerns the.” TPC accept this point and refer extensively to the developmental literature on the. Basically, the authors now invoke the “egocentric” hypothesis as part of the explanation for the kindergarten-path effect, which, as I noted, is the standard explanation for overuse of the. Trueswell and his colleagues had never referred to these studies or ideas in their earlier papers on the topic, attempting to explain the results as some kind of cue-learning, and assuming that children had no problems with the. Thus the TPC paper is an advance over earlier versions that Trueswell and his colleagues presented, bringing the kindergarten-path effect into line as simply one more instance of the child’s pattern of errors that I called the aâ•–→â•–the error. As I have argued, there is nothing special about the kindergarten-path effect; one can simply assume that children have the same processing abilities as adults (that is, they produce the garden-path effect for the same reasons as adults do), but they have difficulties with the. That was the main thrust of my discussion of Trueswell’s studies in the preceding section, and to their credit TPC have basically accepted the point. In my view, the point provides a major new (really old) perspective on the kindergarten-path effect, so it is good to have agreement. However, TPC go on to make other points, and these are important to comment on. My paper presented a new theory of children’s difficulties with deÂ� terminers: troubles with maximality. I pointed out that egocentricity was the standard idea, that I had often used the idea, and that I had attempted in several papers to put it together with many other phenomena in the development of language (for example, problems with certain kinds of binding situations).


35

However, I argued that there were phenomena (especially as reflected in Heim’s observation) that it didn’t seem possible for egocentrism to handle, but that the new theory could handle. My main motivation was to attempt to find a theory that could handle the phenomena. It is possible that egocentrism and troubles with maximality are both true, although of course it would be a more elegant theory if only one of them had to be invoked. TPC concentrate on trying to show that the Maximality Troubles (let’s call it MT) theory that I suggest is wrong. It is important to realize that even if that is true, it doesn’t vitiate my basic argument about Trueswell’s earlier attempts to explain the kindergarten-path effect. But TPC have conceded on that major point, so now we need only discuss the interesting empirical issue: Is egocentrism or MT more correct? TPC attempt to vindicate the “pragmatic” account against my “semantic” account. In this regard, it is worth observing that the brunt of my argument in the preceding section was that Trueswell had missed the pragmatic account, so we are now proceeding from a completely different space of ideas than Trueswell’s earlier papers, one starting off from the conclusions in my paper. But let’s see what the particular arguments are. TPC’s first and major point is logically false. They say that “Wexler’s semantic account appears to miss intuitions about referential errors.â•–.â•–.â•–. On his account, in a referential domain of several boys (e.g. one crying, one laughing, one sleeping) the sentence The boy is crying is false, rather than simply infelicitous. That is, if there are multiple boys in the domain of reference (and the other boys are not crying), the predicate is false because it is not true of all members of the set boy. However, the present authors all have the strong intuition that the utterance in this situation is true but infelicitous; that is, it should have been said in a different way.” Their claim — that my account predicts that the sentence is false — is itself false. “My” account of the determiner the is the standard Fregean account; I gave Heim’s formalization of this in (1). The Fregean theory has three values of truthfulness: true, false, and truth-valueless. The standard interpretation of truth-valueless in semantic theory is that such a value for an utterance results in an “infelicity.” (1) says that false is not the value in the current situation, since to derive false there has to be exactly one boy at index i — that is, exactly one boy in the current situation. Since there is more than one boy, false is not the value of the sentence the boy is crying. In fact, (1) predicts quite straightforwardly that in this situation the boy is crying is truth-valueless. Therefore the sentence results in an infelicity.9 This account of the semantics of the has been well known for more than 100Â€ years — ever since Russell’s famous the present king of France is bald

36

Wexler

example. Russell considered this sentence false, but Frege argued that it had no truth-value, introducing a crucial concept to semantic theory. Russell’s example involves what we would now call the presupposition of existence (not satisfied since there is no present king of France). TPC’s example involves the presupposition of uniqueness. I explicitly invoked the Fregean (not the Russellian) theory, so there is no way the theory would predict false for this sentence in this situation. The main point of my small comments on the Trueswell studies was that they didn’t take into account a large body of developmental psycholinguistics. Taking this work into account, I claimed, one could well maintain that there was nothing special about the Trueswell processing results; they would simply follow from the child’s difficulties with the plus the child having the same processing abilities as the adult. Nothing had to be said about statistics of locative phrases following put, etc., at least nothing different than what was said for the adult. So I offered a quite parsimonious account, which seems to stand up well. As I pointed out TPC have accepted the relevance of this literature. I now have one other suggestion, something I didn’t think I would have to argue for. In addition to taking into account a large experimental and analytic literature on psycholinguistic development, it would be very useful to take account of standard semantic theory with regard to the constructions under development. Never in the earlier papers did Trueswell and his colleagues seem to talk about the properties of the or their development. Now that they take the developmental properties to be relevant, it would be a further advance to inÂ� corporate the adult properties of the. From a semantic point of view, the FÂ�regean theory does this. If TPC were to accept such a theory (or give us a well-reasoned alternative), the discussion could be carried out on a much surer footing. In view of TPC’s misunderstanding of standard semantic theory, it is difficult to know what to make of their further points about my theory. Their second point is that there seem to be pragmatic issues involved in the use and development of the. I didn’t deny that; in fact, I pointed it out and mentioned that Trueswell and colleagues weren’t appreciating them. I gave an empirical argument (Heim’s observation) in favor of the troubles with Maximality theory. I believe that theory does account for the empirical literature that I know about (quite a large literature). Furthermore, it is possible that pragmatic difficulties also have to be added to the semantic problem. I will only make two points here. First, TPC have ignored much of the relevant literature, for example the de dicto experiments that I discussed above. Second, I don’t have to be reminded of the importance of the development of pragmatic issues of just this kind. It is a standard theme in my research (see the title and discussion from


37

Chien and Wexler 1990). Avrutin and Wexler (1992) tied together the standard empirical results on determiners (Karmiloff-Smith’s especially) and the egocentric analysis with results on the Delay of Principle B. Wexler 1999 and Thornton and Wexler 1999 contain extensive, detailed discussions of the relevant pragmatic issues. I am writing in the spirit of seeing if there is an alternative to the standard egocentric theory, since there appear to be some empirical difficulties with that hypothesis. I am not sure what TPC’s primary goal is. Is itÂ€to say that there is something special about the kindergarten-path effect, in addition to the standard understanding in the literature, a claim that I deny? Is it to defend a theory? The multiple-cue theory? It is an advance that they recognize that children have some difficulties other than those related to the probability that a locative phrase follows put, but I am not sure what else they intend to say, other than to accept the fact that there is pragmatics as well as semantics. TPC’s third point is that the data in one of the two studies I referred to (Maratsos 1976) concerning Heim’s observation weren’t unequivocal. I’ll leave it up to others to read the studies. TPC report the response of two out of four adults on a translation of Karmiloff-Smith’s (1979) materials. Six 10-year-olds produced 0 percent definite articles in the experiment on stories in which there were many X’s (Karmiloff-Smith 1979, table 27, p. 144). Eight 9-year-olds produced 14 percent definite articles. But children of ages 4 to 7 (eight to twelve children in each group) produced between 48 percent and 63 percent definite articles. There is a huge “error” in the younger children, going down to zero in the older children. The fact that two of four adults in anecdotal reporting asked a question like “I don’t know, Mary” doesn’t mean anything other than that they guessed when they thought more information was required. It has nothing to do with the fact that young children ( but not older children, nor, presumably, adults) will use the. This returns us to my main criticism, in the preceding section, of the work by Trueswell et al. on the kindergarten-path effect: that it ignores the experimental developmental literature in favor of preformed ideas. There is a large literature on this topic. TPC’s fourth point is about an experiment by Papafragou and Tantolou (a work not available to me at the time of writing) that claims to show (as I understand it from TPC) that children know that the oranges must refer to the entire set of oranges. Karmiloff-Smith (1979) reports an experiment that shows that isn’t so, that children of this age (5 years) in fact will refer to a subset of the entire set of oranges in the situation as the oranges. It will be interesting to see what the difference between the two experiments is. My first thought is that the Papafragou and Tantolou result is due to the child’s being told by the elephant “I ate some.” To understand the result, we need a precise theory of the sÂ�emantics

38

Wexler

of some and of the relevant pragmatics (including scalar implicature and /or Maximize Presupposition). I don’t understand TPC’s fifth and final point. The theory I suggested was that 5-year-olds have difficulties with the (of the type that are standardly uÂ�nderstood) plus the same processing capacities (at least the relevant ones) as the adult. This yields the results of the kindergarten-path experiments so far as I can tell. I didn’t claim (quite the contrary) that the children have no processing limitations. I suggested that they could be taken to be whatever the limitations were in the adult, whatever caused the garden path. TPC mistakenly say that my claim that children have difficulties with the is related to the results presented in Chien and Wexler 1990. I do not know what to make of this. In that paper, we claimed that children had pragmatic difficulty with reference of a particular kind involving intensions of participants, the type of difficulty that was familiar from the egocentric theory of children’s determiner errors. There was nothing in Chien and Wexler about children’s lack of full semantic knowledge of the. I’ll lay out my bias here. I find the “multiple cue” idea either empty or false, depending on how it is interpreted. On the one hand it is a truism; how else can a sentence be analyzed and interpreted than by integrating lots of different kinds of information? But if it is interpreted, as I fear Trueswell and his colleagues interpret it, to mean that there is nothing in the mind except a set of undifferentiated probabilistic cues, dictated by the “environment,” then I find it false on its face. To be blunt, I don’t understand what is claimed to be the “complicated” theory in TPC. To be concrete, in TPC’s own words, their idea is that “there exist multiple probabilistic sources of evidence for constraining the grammatical structure of an utterance, and the child must discover, weigh, and combine this evidence.” Under one interpretation, as discussed, this is a truism, not a theory. Under another interpretation, if the statement is taken to mean, as I fear the authors intend, that there are no structural conditions, no well-defined syntax and semantics, no genetically programmed properties that the child is looking for, then the statement is false and they provide no reasons to revise this assessment. I hope TPC will think about the following: How could their theory be shown to be false? What kind of result would show that? I would like to stress that I find the kindergarten-path experiment lovely and useful; it confirms what we know about the development of determiners (the overuse of the in certain types of contexts, although so far it hasn’t helped us in differentiating accounts of that phenomenon), and it offers evidence that 5-year-olds have the same type of processing capacity and limitations as do adults. I welcome the results. But statements about uniqueness and specialness


39

of the results aren’t optimum when instead the results can fit into a broader and more experimentally supported set of ideas. We need understandable precise theories; this is what at least some of the developmental psycholinguistic literature has attempted to provide. If the kindergarten-path literature joins this attempt at clear analysis, and integration with a broader literature, it might be able to contribute substantially to our understanding. I look forward to such an outcome. Notes 1.â•‡ “Learning” explanations are unlikely, since the empirical results are very strong and clear: children learn properties that vary across languages almost always right from the start (e.g., Wexler 1998). The kinds of syntactic properties that develop late seem to be part of UG, not variable properties. We’ll try to study semantic development in similar terms. 2.â•‡ Grodzinsky and Reinhart (1993) follow in essence the development of Chien and Wexler, with a revision concerning the details of the pragmatics, and claim that the child difficulties are due to processing/memory overload in comparing representations rather than to pragmatics. Thornton and Wexler (1999) argue that this processing account can’t explain the empirical results. Wexler (2003) shows that this processing account mispredicts the empirical results concerning determiners that will be the central issue studied in this paper. But see Reinhart (this volume) for an alternative perspective. 3.â•‡ Here I’m only discussing singular nouns. I’ll briefly discuss plurals later. 4.â•‡ KS provides a discussion of why she gets somewhat more such responses than Maratsos. KS’s study seems more reliable, since it is much larger. Also, Maratsos simply didn’t test older ages. It might be worth noting that Maratsos’s participants were recruited from ads in such publications as the Harvard Crimson, suggesting that the academic family nature of his participants might have resulted in a more advanced verbal profile than the average population. 5.â•‡ The situational index i refers to the world in which the utterance is evaluated; for example, if that world is one in which the singleton pig that is being talked about is fat, the pig is fat is true at i; if the single pig in the world is not fat, the sentence is false. 6.â•‡ An alternative is that the child has the same lexical entry as the adult, but has trouble computing the uniqueness (maximality more generally) requirement specified in the presupposition of the in (1). This could account for why the child has poor but not completely reversed behavior from the adult in contexts in which uniqueness doesn’t hold, e.g. table 1, in which the 6-year-olds provide 56 percent (incorrect) use of the in the several identical objects condition rather than 100 percent. (I am assuming that the 56 percent figure means that each child has a figure quite different from 0 percent or 100 percent, which might not be the case. I’m not sure if enough individual subject analysis has been done to confirm this versus the alternative that each individual child is close to 0 percent or 100 percent.) The lexical entry theory given in the text (6) could also account for the alternating behavior under the assumption that uniqueness (more generally maximality) is a weak /variable property for the child. There is no space here to

40

Wexler

consider what empirical measures could distinguish the two alternatives. It is important to note that either alternative relies on the assumption that the difficulty has to do with uniqueness (more generally, maximality), not with egocentrism or with other semantic conditions or with the computation of Maximize Presupposition. 7.â•‡ Trueswell et al. (this volume) suggest that the evidence for uniqueness for the is unclear, since the X can be used, for example, when there are two X’s, but one is much more accessible (e.g. much closer to the hearer). But their argument assumes that the child doesn’t know that context situations are determined by internal states, purposes and goals. If the child does know this, then uniqueness is apparent in the input. If the child doesn’t know it, then the question is why not? And we would need a maturational explanation to explain how that knowledge is attained. I know of no reason to believe that children think that context sets are whatever is available in the visual input, rather than being partially determined by goals and purposes. 8.â•‡ To their credit, Trueswell et al. (this volume), responding to this paper, are an exception. They add the egocentric hypothesis that I reviewed earlier to their set of factors that help to cause the kindergarten-path effect, referring to the experimental literature. See the next section for a discussion of this paper. 9.â•‡ When the authors claim that their intuitions are that the sentence is “true but infelicitous,” they refer only to their intuition; they don’t mention any semantic theory to support this intuition nor say what the analysis would be. The standard Fregean theory gives an explicit analysis of the intuition. There is presupposition failure. Perhaps TPC want to somehow return to a kind of pre-Fregean Russellian analysis (see the next paragraph in the text for what Russell said), but they give no details of what they have in mind. References Avrutin, S., and Wexler, K. 1992. Development of principle B in Russian: Coindexation at LF and coreference. Language Acquisition 2, 259–306. Babyonyshev, M., Ganger, J., Pesetsky, D., and Wexler, K. 2001. The maturation of grammatical principles: Evidence from Russian unaccusatives. Linguistic Inquiry 32, 1– 44. Borer, H., and Wexler, K. 1987. The maturation of syntax. In T. Roeper and E. Williams (eds.), Parameters in Language Acquisition. Reidel. Borer, H., and Wexler, K. 1992. Bi-unique relations and the maturation of grammatical principles. Natural Language and Linguistic Theory 10, 147–189. Chien, Y. C., and Wexler, K. 1990. Children’s knowledge of locality conditions in binding as evidence for the modularity of syntax and pragmatics. Language Acquisition 1, 225–295. Crain, S., and Thornton, R. 1998. Investigations in Universal Grammar: A Guide to Experiments in the Acquisition of Syntax and Semantics. MIT Press. Frazier, L. 1987. Sentence processing: a tutorial review. In M. Coltheart (ed.), Attention and Performance XII. Erlbaum.


41

Frazier, L., and Fodor, J. D. 1978. The sausage machine: A new two-stage parsing model. Cognition 6, 291–325. Grodzinsky, Y., and Reinhart, T. 1993. The innateness of binding and coreference. Linguistic Inquiry 24, 69–101. Guasti, T. 2002. Language Acquisition. MIT Press. Heim, I. 1991. Articles and definiteness. In A. Stechow and D. Wunderlich (ed.), Handbook of Semantics: An International Handbook of Contemporary Research. Walter de Gruyter. Hurewitz, F., Brown-Schmidt, S., Thorpe, K., Gleitman, L., and Trueswell, J. 2000. One frog, two frog, red frog, blue frog: Factors affecting children’s syntactic choices in production and comprehension. Journal of Psycholinguistic Research 29, 597–626. Ionin, T., Ko, H., and Wexler, K. 2004. Article semantics in L2 acquisition: The role of specificity. Language Acquisition 12, 3–69. Karmiloff-Smith, A. 1979. A Functional Approach to Child Language. Cambridge University Press. Maratsos, M. 1976. The Uses of Definite and Indefinite Reference in Young Children: An Experimental Study of Semantic Acquisition. Cambridge University Press. Matthewson, L., and Schaeffer, J. 2000. Grammar and pragmatics in the acquisition of article systems. In J. Gilkerson, M. Becker, and N. Hyams (eds.), UCLA Working Papers in Linguistics: Language Development and Breakdown. Piaget, J. 1955. The Language and Thought of the Child. Meridian Books. Rizzi, L. 1994. Some notes on linguistic theory and language development: The case of root infinitives. Language Acquisition 3, 371–393. Schütze, C., and Wexler, K. 1996. Subject case licensing and English root infinitives. In A. Stringfellow, D. Cahma-Amitay, E. Hughes, and A. Zukowski (eds.), Proceedings of the 20th Annual Boston University Conference on Language Development. Cascadilla. Tanenhaus, M. K., Spivey-Knowlton, M. J., Eberhard, K. M., and Sedivy, J. C. 1995. Integration of visual and linguistic information in spoken language comprehension. Science 268, 1632–1634. Thornton, R., and Wexler, K. 1999. Principle B, VP Ellipsis, and Interpretation in Child Grammar. MIT Press. Trueswell, J. C., and Tanenhaus, M. K. 1994. Toward a lexicalist framework for cÂ�onstraint-based syntactic ambiguity resolution. In C. Clifton, K. Rayner, and L. Frazier (eds.), Perspectives on Sentence Processing. Erlbaum. Trueswell, J. C., Sekerina, I., Hill, N. M., and Logrip, M. L. 1999. The kindergartenpath effect: Studying on-line sentence processing in young children. Cognition 73, 89–134. Vinnitskaya, I., and Wexler, K. 2001. The role of pragmatics in the development of Russian aspect. First Language 21, 143–186. Warden, D. A. 1974. An Experimental Investigation into the Child’s Developing Use of Definite and Indefinite Referential Speech. Ph.D. thesis, University of London.

42

Wexler

Warden, D. A. 1976. The influence of context on children’s use of identifying expressions and references. British Journal of Psychology 67, 101–112. Wexler, K. 1990. Optional infinitives, head movement and the economy of derivation in child grammar. Paper presented at annual meeting of Society of Cognitive Science. Wexler, K. 1993. Optional infinitives, head movement and the economy of derivations. In D. Lightfoot and N. Hornstein (eds.), Verb Movement. Cambridge University Press. Wexler, K. 1998. Very early parameter-setting and the Unique Checking Constraint: A new explanation of the Optional Infinitive stage. Lingua 106, 23–79. Wexler, K. 1999. Maturation and growth of grammar. In W. C. Ritchie and T. K. Bhatia (eds.), Handbook of Language Acquisition. Academic Press. Wexler, K. 2002. Lennenberg’s Dream: Learning, Normal Language Development and Specific Language Impairment. In J. Schaeffer and Y. Levy (eds.), Towards a Definition of Specific Language Impairment. Erlbaum. Wexler, K. 2003. Maximal Trouble. Talk given at Maryland Mayfest: Semantics and Acquisition. Wexler, K. 2004. Theory of phasal development: Perfection in child grammar. MIT Working Papers in Linguistics 48, 159–209. Wexler, K., and Chien, Y. C. 1985. The development of Lexical Anaphors. Papers and Reports on Child Language Development 24, 138–149.

3â•…

Children’s Use of Context in Ambiguity Resolution Luisa Meroni and Stephen Crain

Many experimental investigations of human sentence processing have shown that listeners do not wait until they reach the end of a sentence before they begin to compute an interpretation. Rather, listeners incrementally make commitments to an interpretation as the linguistic input unfolds in real time. A consequence of this feature of sentence comprehension is that it sometimes gives rise to so-called garden-path effects. In the presence of a temporary ambiguity, listeners may assign an interpretation that later turns out to be unworkable and must, therefore, be abandoned in favor of an alternative interpretation. Various explanations have been proposed to account for the garden-path effects that have been documented in certain experimental contexts (e.g. Frazier and Rayner 1982; Trueswell and Tanenhaus 1994; MacDonald 1994). One line of research has claimed, however, that the referential contexts in which sentences ordinarily appear often mitigate, or even eliminate, gardenpath effects. This is the Referential Theory proposed by Crain and Steedman (1985) and extended by Altmann and Steedman (1988). According to the Referential Theory, listeners experience garden-path effects primarily when sentences are interpreted outside any referential context or in infelicitous contexts. If the Referential Theory is correct, garden-path effects are largely experimental artifacts. Recent work by Trueswell, Sekerina, Hill, and Logrip (1999) suggests that children, in resolving temporary syntactic ambiguities, may not be as sensitive to features of the referential context as adults are. If so, children would be expected to experience garden-path effects that are not experienced by adults. This finding would represent a serious setback to the Referential Theory, as it pertains to children; it would also be a setback to the Continuity Assumption, which maintains that children and adults engage the same cognitive mechanisms in language processing (see, e.g., Pinker 1984). In this paper we re-examine the conclusion that children suffer from gardenpath effects to a greater extent than adults do in (on-line) sentence processing.

44

Meroni and Crain

We offer an alternative account of children’s non-adult responses in previous research, and we provide experimental evidence for the conclusion that children actively use the referential context. The Role of Context in Resolving Ambiguity

A central question in recent studies of on-line sentence processing is whether or not children rely on the same parsing strategies as adults do in resolving temporary syntactic ambiguities. In particular, several studies have focused on the ability of children and adults to resolve ambiguities involving two prepositional phrases (PPs) in sentences like (1) (see Trueswell and Gleitman 2004). (1)â•‡ Put the frog on the napkin into the box. Upon encountering the first PP ‘on the napkin,’ the listener is not sure whether additional linguistic material will follow. A temporary ambiguity arises at this point, because the PP ‘on the napkin’ can be attached ‘high’ to the verb, specifying the destination of the action (i.e., where to move the frog), or can be attached ‘low’ to the noun phrase, indicating which frog to move (i.e., the frog that is on the napkin). Previous research in adult sentence processing has demonstrated that, in the absence of context, the psychological parser prefers high attachment, which yields the interpretation in which ‘on the napkin’ is the destination of the action. This conclusion is supported by evidence of telltale signs of garden-path effects at the second PP ‘into the box’ (Ferreira and Clifton 1984; Britt 1994). Different explanations have been proposed for the source of this attachment preference outside of context.1 One possibility, which we endorse, is that the VP-attachment preference follows from a parsing principle that instructs the parser to assign the theta roles associated with a verb as soon as possible, all other things being equal. Let us call this the Theta Assignment Principle (e.g. Gibson 1991; Gorrell 1995; Weinberg 1992). Adherence to the Theta Assignment Principle creates the garden-path effects witnessed in sentences like (1), according to the following scenario. The English verb put assigns two theta roles to its internal arguments: a Theme role and a Destination role.2 These theta roles are usually realized by a noun phrase and a prepositional phrase, respectively. In accordance with the Theta Assignment Principle, the parser immediately assigns the Theme theta role to the NP ‘the frog’ and assigns the Destination theta role to the PP ‘on the napkin’. These theta roles are assigned before the parser encounters the second PP, ‘into the box’. The theta roles associated with the verb having been discharged, the subsequent appearance of the PP ‘into the box’ informs the parser that it has been led down a garden path.

Children’s Use of Context

45

Figure 3.1

The one-referent context.

In response, the parser attempts to revise its initial decision about the Destination theta role. The necessary revisions include a reanalysis of the first PP, ‘on the napkin’, as a modifier of the NP ‘the frog.’ 3,4 The effects of the Theta Assignment Principle can also be observed when sentences are presented in certain referential contexts. One such context is depicted in figure 3.1, where there is one frog on a napkin, a chick, an empty napkin, and a box. Following Trueswell et al. (1999), we call this the one-Â� referent context. In response to the instruction in (1), both children and adults are expected to experience a garden path in this context when they encounter the second PP, ‘into the box’ (Tanenhaus et al. 1995; Spivey and Tanenhaus 1998). In other referential contexts, according to the Referential Theory, gardenpath effects are not anticipated. In such contexts, the temporary ambiguity involved in attaching the PP ‘on the napkin’ (as a modifier) should be resolved without recourse to the Theta Assignment Principle. One such context is depicted in figure 3.2, where there is one frog on a napkin, a second frog not on a napkin, an empty napkin, and an empty box. Following Trueswell et al. (1999), we call this the two-referent context. In response to the instruction in (1), in the two-referent context the parser should immediately respond by analyzing the PP ‘on the napkin’ as a modifier. According to the Referential Theory, the elimination of garden-path effectsÂ€in the two-referent context follows from the modular architecture of the language apparatus. A parsing principle, called the Principle of Referential Success, is hypothesized to take precedence over the Theta Assignment Principle in the two-referent context (Crain and Steedman 1985; Altmann and

46

Meroni and Crain

Figure 3.2

The two-referent context.

Steedman 1988). The Principle of Referential Success directs the parser to abandon structural analyses that do not refer to entities represented in its current model of the domain of discourse, and to maintain analyses that succeed in referring to entities in the model. In the two-referent context, the presence of two frogs in the domain of discourse prevents the application of the Theta Assignment Principle, which, as we have seen, would otherwise direct the parser to analyze the PP ‘on the napkin’ as the destination of the ‘putting’ event. This attachment of the PP, however, would prevent the parser from successfully identifying the referent of the initial definite NP ‘the frog’ in the discourse. The Principle of Referential Success preempts the application of the Theta Assignment PÂ�rinciple and leads the parser to attach the PP ‘on the napkin’ as a modifier because, on this analysis, the parser succeeds in referring to an entity in the discourse (the frog on the napkin), thus satisfying the presupposition of uniqueness triggered by the definite NP ‘the frog’. Once the PP ‘on the napkin’ has been attached low, the Theta Assignment Principle becomes operative again and guides the parser’s assignment of the Destination theta role to the PP ‘into the box’. On this analysis, the second definite NP, ‘the napkin’, succeeds in referring, despite the presence of the empty napkin; it refers to the napkin associated with the frog under consideration. On the Referential Theory, no worries are anticipated for the parser in the two-referent context for either children or adults. Children’s Lack of Sensitivity to the Referential Context

In a recent series of studies using a free-head eye-tracking system, Trueswell etÂ€ al. (1999) examined the parsing strategies used by children and adults in


47

resolving the attachment ambiguity in sentences like (1) (see also SpiveyKnowlton and Sedivy 1995 for adults). Trueswell et al. considered contexts that supported the early application of the Theta Assignment Principle (the one-referent context) and contexts in which the Principle of Referential Success was expected to forestall the application of the Theta Assignment PrinÂ� ciple (the two-referent context). In the one-referent context, the principle of Referential Success and the Theta Assignment Principle can both be satisfied, because these principles are not in conflict when there is a unique referent for the initial NP ‘the frog’. The parser should experience a garden-path effect, however, upon encountering the second PP, ‘into the box’, in the one-referent context. In the two-referent context, by contrast, the Principle of Referential Success should obviate the application of the Theta Assignment Principle at the initial PP, ‘on the napkin’. Hence, no garden-path effects are expected to occur in the two-referent context. The results for adults were exactly as expected in the Trueswell et al. (1999)Â€study.5 In the one-referent context, adults showed a preference for VPattachment of the PP ‘on the napkin’. Although adults typically performed the correct actions (e.g., moving the frog directly into the box) in the one-referent context, they nevertheless moved their eyes to the “empty” napkin as they were processing the PP ‘on the napkin.’ Taken together, the findings are interpreted by Trueswell et al. (1999) as evidence that adults initially mis-analyzed the PP ‘on the napkin’ as filling the Destination theta role ( hence the glances to the empty napkin), but adults were able to revise their analysis ( hence the correct actions). The findings for adults in the two-referent context were also consistent with the Referential Theory. Adults not only performed the correct actions in this context; they did not even move their eyes to the empty napkin. By inference, adults immediately interpreted the first PP, ‘on the napkin’, as a modifier of the NP ‘the frog,’ thereby avoiding garden-path effects. This pattern of behavior conforms to the prediction of the Referential Theory. The responses of the children in the Trueswell et al. study tell a different story. In the one-referent context, children’s responses were consistent with the Referential Theory, but children were less able than adults to overcome the garden-path effects they experienced. Not only did children look at the empty napkin, as did adults; they also moved the frog onto that napkin on more than half of the trials. From this, one difference between children and adults is aÂ�pparent: children are less able than adults to recover from their initial misanalysis of temporary ambiguities in sentences like (1). A second difference between children and adults emerged in the two-Â�referent context. In contrast to adults, children continued to make “errors” in performance, just as they had in the one-referent context.6 Children performed “incorrect” actions on more than half of the trials. In particular, on trials where

48

Meroni and Crain

children produced non-adult responses, instead of moving the frog that was on the napkin to the empty napkin (for example), they moved the “wrong” frog — the one that was not on the napkin — 90 percent of the time. Presumably, on these trials, children interpreted the PP ‘on the napkin’ as the destination of the putting event, rather than as a modifier of the NP ‘the frog’. If so, it would seem that for children the Principle of Referential Success did not effectively block the application of the Theta Assignment Principle.7 And once again children were unable to revise their initial attachment of the PP ‘on the napkin’ as the destination of the putting event. These findings led Trueswell et al. (1999) to conclude (a) that children are less able than adults to use the referential context to guide their on-line parsing decisions, at least under certain processing conditions, and ( b) that children are unable to retrace their footsteps once led down a garden path.8 What Lies Beneath

In this section we present our own analysis of the experimental finding reported by Trueswell et al. (1999). Many of the features of our analysis were discussed by Trueswell et al., but the conclusions they reached were quite different from ours. On our account, not only do children consistently adhere to the Principle of Referential Success in resolving temporary structural ambiguities, but they adhere to all the other parsing principles that adults do, and they schedule these principles in the same way as adults do. Obviously children differ from adults, but we would emphasize the limits on these differences, on learnability grounds (see Crain and Wexler 1998). We pursue two features of the Trueswell et al. experimental findings. We wish to acknowledge, first, that children are less able than adults to revise their initial interpretation when they are confronted with linguistic input that cannot be incorporated into the structural analysis they are entertaining. This difference between children and adults is well documented in the literature as the source of children’s non-adult responses to several linguistic constructions. In the study by Trueswell et al., it is readily apparent that once children begin to execute a response, in both the one-referent and the two-Â� referent contexts, they are unable to revise the cognitive algorithms (i.e., plans for acting out the meanings of linguistic expressions), which they had already begun to execute. On the basis of previous research, we assume that children only gradually develop the performance routines needed to delay the immediate execution of plans. As for children’s non-adult responses, Trueswell et al. conclude that the vast majority of children resort to a guessing paradigm in the two-referent context.


49

This claim is based on an analysis of the responses of all children, not just the responses of children who committed errors. According to Trueswell et al., children who produced the “right” sequence of actions did so for the wrong reasons. Their argument goes roughly as follows: Most children initially guessed which of the two frogs to move, and sometimes a child guessed “right,” choosing to move the frog that was on a napkin. Once a child had selected the “right” frog, however, he or she was reticent to put it on the empty napkin, because the frog the child had selected was already on a napkin. So children ignored the initial PP on these trials and responded only to the second PP, ‘into the box’. We propose another account of children’s adult-like responses, according to which children draw upon the same parsing principles that account for adult responses. To see this, we need to compare children’s responses in the tworeferent context with their responses in the one-referent context. Trueswell etÂ€ al. 1999 observed that whenever children selected the “right” frog in the two-referent context — the one on the napkin — they performed an “incorrect” sequence of actions on only 10 percent of trials. By contrast, whenever children selected the “wrong” frog — the one that was not on the napkin — 90 percent of their responses were “incorrect.” Clearly, this is a significant difference in behavior. To explain the difference, Trueswell et al. argue that children who guessed “right” ignored the PP ‘on the napkin’ because the frog they had selected was already on a napkin. Although this sounds plausible at first, there is ample evidence against it. If it were true, then children should have made the same inference in the one-referent context. Recall that in that context there was a single frog on a napkin, and there was an empty napkin. By parity of reasoning, the child subjects should have ignored the PP ‘on the napkin’ in the onereferent context just as often as the child subjects did in the two-referent context, putting the frog directly into the box, rather than on the empty napkin, on the grounds that it was already on a napkin.9 However, children moved the frog (from the napkin it was on) to the “empty” napkin on 56 percent of the trials in the one-referent context. This casts doubt on the proposal by Trueswell et al. that children ignored the PP ‘on the napkin’ in the two-referent context as a result of a pragmatic inference. For this reason, we disagree with the account proposed by Trueswell et al. We maintain that children who performed the right actions in the two-referent context did so for the right reasons, not because of a pragmatic inference. This brings us to the children who chose the “wrong” frog. The critical observation is that children who chose the “wrong” frog — the one that was not on the napkin — performed “incorrect” sequences of actions on 90 percent of the trials. We contend that in this case children’s non-adult responses were due to

50

Meroni and Crain

a pragmatic inference, and that the pragmatic inference was the children’s attempt to comply with the Principle of Referential Success. The story is simple. As children heard the sentence fragment “Put the frog on the napkin,” they (implicitly) reasoned as follows: “The experimenter is asking me to put the frog on the napkin. But there are two frogs. Which one is she referring to? I know: the one that is not already on a napkin.” In our view, the children were making a reasonable pragmatic inference that the experimenter’s intended referent for the NP ‘the frog’ was clearly discernable in the context — it was the frog that was not already on a napkin.10 If this is correct, then children were complying with both the Theta Assignment Principle and the Principle of Referential Success concurrently. By applying the pragmatic inference, children were able to interpret the PP ‘on the napkin’ as the destination of the action (Theta Assignment Principle) and, at the same time, were able to identify a uniquely salient frog in the context (Principle of Referential Success). In sum, the application of a pragmatic inference together with children’s inability to revise their initial interpretation explains children’s non-adult actions without supposing that children’s language apparatus vÂ�iolates any tenet of the Referential Theory. Plans and the Structure of Linguistic Behavior

The process of acting out instructions like “Put the frog on the napkin into the box” draws upon a cognitive algorithm, or plan. Plans are created and used in concert with other aspects of language processing. These include the outputs of the syntactic and semantic components, as well as the components of the language apparatus that are responsible for accessing linguistic principles and coordinating their use, e.g., verbal working memory. Devising a plan and executing it may be separated in time, or they may be interleaved in time. These two possibilities are analogous to the distinction between compiling and interpreting computer languages. The computer analogy has implications for understanding the differences that children experience in responding to various linguistic structures. Children are less “automated” or “compiled” than adults, and children may have less verbal working memory capacity. As a consequence, children sometimes act out the meanings of sentences in an order-ofmention fashion (Amidon and Peter 1972; Crain 1982), whereas adults act them out in an order that is conceptually correct (e.g., “given” information first). Because children interleave planning and execution, they tend to act out parts of the plans they generate, before all planning has been completed. The interpret-mode behavior of children can be rendered infeasible in certain situations — for example, by withholding the experimental display until


51

Figure 3.3

the child has listened to the entire instruction. In such circumstances, the child can (mentally) devise a plan in the compile mode, as adults do. If so, children are expected to perform conceptually correct sequences of actions, just as adults do. (Later in this paper, we present an experiment demonstrating that children do not commit errors in responding to instructions like “Put the frog on the napkin into the box” in the “phrase-and-then-display” condition, presumably because this experimental manipulation prevents children from prematurely executing the plans they generate on-line.) There are several examples of children’s premature execution of linguistic instructions in the literature on child language (i.e., relative clauses, sentences with before/after) . The example we focus on here involves the interpretations of prenominal modifiers, such as second striped ball. In a study of children’s command of phrase structure, Matthei (1982) reported systematic non-adult behavior by children when they were asked to perform the instruction in (2) using the array of objects depicted in figure 3.3. (2)â•‡ Point to the second striped ball. When adults were given the same instruction (counting from the left), theyÂ€consistently pointed to the second of the striped balls, i.e., the third object in the array. By contrast, children 4 to 6 years old often pointed to the second ball in the array, which happens to be striped. Hamburger and Crain (1984) raised the possibility that “the child subject .â•–.â•–. might start to plan and even act while the sentence is being uttered, possibly making a premature and incorrect dÂ�ecision.” Consistent with this hypothesis, children’s “errors” vanished if the experimenter produced the test sentence before the child had a chance to execute any action. This feature of experimental design is called phrase-and-then-display. Returning to the findings of Trueswell et al., here is what we are suggesting: Children’s inability to revise their initial interpretation in both the one-referent context and the two-referent context may be the result of interpret-mode behavior. In particular, children may have formulated a plan and started to execute it before the entire instruction “Put the frog on the napkin into the box” had been uttered by the experimenter. Specifically, children may have planned a specific course of action that interpreted the first PP, ‘on the napkin’, as referring to the destination. Having committed themselves to putting the frog that

52

Meroni and Crain

was not on a napkin on the empty napkin, it proved untenable for children to revise their interpretation of the instruction, because the action they were committed to perform could not be integrated into any appropriately revised plan. We call this the bird-in-the-hand problem. It is time to take stock. In our view, the experimental findings reported by Trueswell et al. do not support the conclusion that children are less sensitive than adults to contextual factors in the resolution of ambiguities. But children do differ from adults, because children assemble plans and begin to execute them in the interpret mode, rather than (as adults do) in the compile mode. However, in the discussion of children’s interpretation of phrases like second striped ball we have seen how to render the interpret mode inoperative: The experimental maneuver of phrase-and-then-display sufficed to elicit compilemode behavior from children. It is expected that children will perform like adults, then, if the phrase-and-then-display methodology is implemented in studies of PP attachment. In a later section, we report the findings of an experiment using the phrase-and-then-display methodology to encourage children to process linguistic input with a temporary ambiguity in the compile mode. But we are only halfway home. It remains to substantiate the hypothesis that children who selected the “wrong” frog did so because of a pragmatic iÂ�nference. A Pragmatic Inference

In this section we scrutinize children’s non-adult responses. As we noted in anÂ€earlier section, when children made “incorrect” responses they chose the “wrong” frog 90 percent of the time. We hypothesized that children’s non-adult behavior was due to a pragmatic inference that they made to satisfy the Principle of Referential Success. The pragmatic inference enabled children to iÂ�dentify a uniquely salient referent for the NP ‘the frog’ in the experimental workspace. Upon hearing the fragment “Put the frog on the napkin,” children inferred that the unique referent of the NP ‘the frog’ was the one that was not already on a napkin. Experiment I was designed to evaluate this hypothesis. Experiment I

This experiment used contexts like the one depicted in figure 3.4, which includes two frogs. One frog is on a napkin, the other is not. In addition, there is an empty napkin in the experimental workspace. We call this the on /off context, because the two frogs can be distinguished in virtue of being on or off a napkin. Children were presented with contexts similar to the one depicted in


53

Figure 3.4

The on/off context.

figure 3.4 and were asked, by means of instructions like (3), to perform relevant actions. (3)â•‡ Put the frog on the napkin. The experimental hypothesis was that, in response to (3), children’s predominant response would be to move the frog that was not already on a napkin. Twelve children participated in the experiment. They ranged in age from 3;10 to 5;9, with a mean age of 4;7. The experimenter arranged the toys in the experimental workspace as illustrated in figure 3.4. Before each trial, the experimenter ensured that the child knew the name of each object. If the child had a non-standard name for an object, the child’s name for that object was used by the experimenter thereafter. The task was presented to the child as a game in which she had to help the experimenter in getting everything ready for a story. The experimenter instructed the child subject to act out (“Do what I say”) a sentence using the toys available in the experimental workspace. Each child was presented with four target trials, four fillers, and one warm-up. All and only the target trials were instructions containing the verb put as illustrated in (3). As predicted, children moved the frog that was not already on a napkin on 92 percent of trials (44 out of 48). It is worth noting that children never asked the experimenter which frog they should move. Moreover, when they were asked, on the very last trial, to explain why they had moved the frog not on a napkin, they consistently answered “because that one was already on a napkin.” We interpret the findings as confirming our hypothesis that children could make a pragmatic inference in the two-referent context in previous research in

54

Meroni and Crain

order to single out a salient frog, as dictated by the Principle of Referential Success. The preceding discussion suggests that the experimental results reported by Trueswell et al. can be understood without abandoning either the Referential Theory or the Continuity Assumption. If a pragmatic inference is responsible in part for children’s non-adult behavior, the Principle of Referential Success is not at issue. It also follows that, if the pragmatic inference that allows the hearer to identify a “salient” frog can be blocked (or at least delayed), the Principle of Referential Success will become operative in children’s processing. Blocking the Pragmatic Inference

If the present account of children’s non-adult behavior is on the right track, then children should refrain from making the pragmatic inference under discussion if, in the context, neither frog is more salient than the other, i.e., if both frogs are “equally” salient. In such a circumstance, children would be compelled to attach the PP ‘on the napkin’ as a modifier, just as adults do in the two-referent context. This attachment decision about the PP ‘on the napkin’ would enable children to identify a uniquely salient frog, in accordance with the Principle of Referential Success. The second step towards evoking adult-like performance from children, then, is to identify a referential context in which neither of the frogs is more salient than the other. We pursued this by constructing a context in which the two frogs could be distinguished only by a non-contrastive expression. Color adjectives fit the bill. Previous research has shown that color adjectives (e.g., blue), as compared to scalar adjectives (e.g., tall, big), are not interpreted contrastively (Sedivy, Tanenhaus, Chambers, and Carlson 1999; Sedivy 2003). For example, in a series of eye-tracking experiments (Sedivy et al. 1999), subjects were shown displays with four objects and were instructed to move various objects. One set of instructions contained scalar adjectives (example: “Pick up the tall glass”). The finding was that subjects identified the target object (the tall glass in the display) from a competitor object (an object that could have been described using the same adjective tall ) more quickly when a contrasting object of the same category of the target object (a short glass in the display) was present than when a contrasting object was not present. The findings were interpreted as evidence that participants systematically used the contrastive interpretation of prenominal scalar adjectives. Color adjectives were not found to exhibit the same contrastive function as scalar terms in similar contexts (Sedivy 2003). For example, when a subject was presented with a display containing different objects and asked to “Pick up the blue cup,” the presence in


55

the display of a contrastive object (a red cup) did not facilitate the identification of the target object. With this in mind, we return to the pragmatic inference that we claim children made (in identifying the referent of the NP ‘the frog’) in the two-referent context of the study by Trueswell et al. In the previous section, we showed that children presumably identified the unique referent of ‘the frog’ using a pragmatic inference: it was the one that was not on a napkin. Having satisfied the Principle of Referential Success, children were then able to apply the Theta Assignment Principle. Inhibiting the pragmatic inference, however, should prevent application of the Theta Assignment Principle, forcing children to useÂ€the PP ‘on the napkin’ as a modifier to satisfy the Principle of ReferentialÂ€Success. Recent work by Sedivy and her colleagues serves the stated goal of finding a situation that blocks the pragmatic inference in the two-referent context. Experiment II

We conducted an act-out task to investigate the hypothesis that the use of a non-contrastive expression in the instructions, such as a color adjective, would suffice to inhibit the pragmatic inference. On a typical trial, children were pÂ�resented with a context with two frogs, each one on a napkin of a different color — one red and one blue. In addition, there was an empty red napkin, as illustrated in figure 3.5. Children were given instructions such as (4). (4)â•‡ Put the frog on the red napkin.

Figure 3.5

The color adjective context.

56

Meroni and Crain

The two frogs could, in principle, be distinguished by the colors of the napkins on which they were placed. This could foster the application of a pragmatic inference similar in spirit to the one illustrated in experiment I. That is, in response to “Put the frog on the red napkin,” children could reason as follows: “The experimenter is asking me to put the frog on the red napkin. But there are two frogs. Which one is she referring to? I know: the one that is not already on a red napkin.” The question is whether children perform this reasoning to the same extent as they did without the color adjective, in experiment I. Twelve children participated in the experiment. They ranged in age from 3;11 to 5;8 years, with a mean age of 4;08.11 The procedures were similar to those in experiment I, with two differences: the two frogs were placed on napkins of different colors, and the target instructions contained a color adjective, as in (4). Children were tested using an act-out task. Each child was presented with four target trials, in addition to four fillers and one warm-up trial. All and only the target trials were instructions containing the verb put. As expected, all the children experienced difficulty in identifying the frog that was the intended referent of the initial NP, “the frog.” In the first two trials, children consistently displayed confusion. They queried the experimenter about which frog to move on 18 of 24 trials. Across all the trials, moreover, children behaved at chance in selecting one of the two frogs to move. Interestingly, when asked in the last trial to explain the reason for selecting the frog they had moved on that trial, children answered “it doesn’t matter” or “it’s the same.” This contrasts with the findings of experiment I, in which children experienced no difficulty in identifying a salient frog in the experimental workspace, presumably owing to a pragmatic inference. Children did not make a similar inference in the present experiment, when such an inference would have been contingent on the use of a color adjective to distinguish between the two frogs. To recap: On the basis of observations from the study by Trueswell et al., we proposed that the majority of children’s “errors” in the two-referent context involved a pragmatic inference. In the preceding sections we focused on two features of the findings presented by Trueswell et al., namely (a) children’s inability to revise their initial interpretation (i.e., interpret-mode behavior) and ( b) children’s selection of the “wrong” frog on the vast majority of trials in which they produced non-adult behavior (i.e., a pragmatic inference). In discussing these findings, we considered two experimental maneuvers that could lead to improved behavior by children by erasing the effects of these two impediments to children’s successful performance. In particular, we discussed how to render the interpret mode inoperative. For example, we hypothesized, on the basis of previous empirical evidence, that the experimental maneuver of


57

phrase-and-then-display should suffice to elicit the compile mode from children in response to sentences like “Put the frog on the napkin into the box.” Similarly, we discussed how we might go about inhibiting children’s pragmatic inference in response to sentences like “Put the frog on the napkin” by using color adjectives. This brings us, at last, to an experiment designed to encourage children to process linguistic input with a temporary ambiguity, as in “Put the frog on the (red) napkin into the box.” Experiment III

To follow up on these observations, we conducted an experiment similar to theÂ€two-referent context experiment employed by Trueswell et al. (1999) but introducing two changes in the design. One change was made to inhibit the pragmatic inference, using color terms; the second was to implement the phrase-and-then-display condition, to impede children’s premature execution of a response. The experimental hypothesis was that these experimental procedures would result in “correct” adult interpretations. To inhibit the pragmatic inference, both the relevant objects were positioned on a platform of the same kind but of a different color, so that only a color adjective could be used to distinguish between them. For example, one trial was about two frogs, each of which was on a napkin (one red, one blue), and there was an empty red napkin (figure 3.6). To implement the phrase-andthen-display maneuver, children were introduced to the scene depicted in fÂ�igure 3.6 in the most neutral way, avoiding leads-in that could have biased

Figure 3.6

The two-referent context.

58

Meroni and Crain

children toward one interpretation over the other. For example we never mentioned that one object was on a platform while the other was misplaced. We also avoided the use of the preposition on in describing the context, in order to block earlier inferences. After being introduced to the scene, children were asked to turn away while they were listening to the instructions, which contained a temporary ambiguity, as in (5). After hearing the entire sentence, children were aÂ�llowed to look at the scene again and perform the action they were asked to perform. We interviewed 22 children in the same age range of those tested in the study by Trueswell et al. (The children who participated in our experiment ranged in age from 3;09 to 5;09, with a mean age of 4;09.12) Each child was presented with four target trials, four fillers, and one warmup. All and only the target trials were instructions containing the verb put, as illustrated in (5). (5)â•‡ Put the frog on the red napkin into the box.13 As predicted, the children performed the correct action on more than 92 percent of trials (81 of 88). The findings suggest that children’s failure to useÂ€ referential information in previous research was due to their tendency toÂ€ execute response plans on the fly, which in turn exploited their ability to make a pragmatic inference in order to satisfy the Principle of Referential SÂ�uccess. Before drawing firm conclusions about the results of this experiment, we need to rule out the possibility that the experimental maneuvers we introduced by using the two-referent context might have improved children’s responses even in the one-referent context. That is, the changes that were introduced ( phrase-and-then-display, use of prenominal adjectives) might have increased children’s willingness to interpret the initial PP as a modifier regardless of the specific context. If so, then the present study simply shows that children are capable of interpreting PPs as modifiers. To control for this possibility, we conducted an experiment incorporating these features, but using a one-referent context. Experiment IV

The one-referent context employed in the present experiment is similar to the two-referent context in experiment III. The only difference is that a frog has been replaced by a different animal in the experimental workspace. For example, one trial was about one frog and one chick, each of which was on a napkin (one napkin was red, one blue), and there was an empty red napkin (see


59

Figure 3.7

The one-referent context.

figure 3.7). After children were introduced to the scene, they were asked to turn away. While unable to see the display, they were given a verbal instruction that contained a temporary ambiguity, as in (6). Having heard the entire sentence, they were allowed to turn back to the scene and to comply with the instruction. (6)â•‡ Put the frog on the red napkin into the box. In the study by Trueswell et al. (1999), children were found to experience garden-path effects in responding to instructions like “Put the frog on the napkin into the box” in the one-referent context. By hypothesis, the changes introduced in experiment III were not expected to increase children’s tendency to interpret PPs as modifiers in the one-referent context exemplified in figure 3.7. To the contrary, we expected children to be led down the garden path to (more or less) the same extent in the present experiment as they had been in previous research. We interviewed twelve children (ages 3;05–5;04; mean age 4;10).14 Each child was presented with four target trials, four fillers, and one warm-up. All and only the target trials were instructions containing the verb put followed by the preposition on, as in (6) above. As predicted, children manifested the signs of having been led down a garden path: they moved the frog onto the empty napkin and then into the box on 52 percent of the trials (25/48).15 The findings of experiment IV suggest that the changes introduced in experiment III do not simply facilitate children’s interpretation of the initial prepositional phrase as a modifier. If they had, children would have performed as successfully in the one-referent context as in the two-referent context. This

60

Meroni and Crain

invites the following conclusion about experiment III: When the relevant pragmatic inference was blocked, and when on-line execution of children’s response plans was blocked, children demonstrated adherence to referentially based parsing principles. We interpret the findings as indicating that children generate a structural analysis that conforms to the Principle of Referential Success, so long as they are prevented from immediately executing the response dictated by the Theta Assignment Principle. Conclusion

Based on the findings of experiment IV, we feel confident in maintaining our default hypothesis. First, children and adults share all the core properties of the performance system in which linguistic knowledge is embedded, as well as sharing knowledge of the linguistic principles themselves. We conclude, in particular, that both children and adults adhere to fundamental parsing principles that have been proposed in the literature: the Theta Assignment Principle and the Principle of Referential Success. The difference between children and adults is that the processing of linguistic input is less automatic for children than it is for adults. This conclusion enables us to maintain the Continuity AÂ�ssumption. Acknowledgments

We wish to thank Andrea Gualmini, Carson Schütze, and Amy Weinberg for helpful discussion, and Megan Gilliver, Martin Hackl, Utako Minai, Beth Rabbin, and Eileen Rivers for help in conducting the experiments. We also thank John Trueswell for his detailed comments on an earlier draft. Most of all, we thank the children, teachers, and staff at the Center for Young Children at the University of Maryland. Notes 1.â•‡ See Spivey-Knowlton and Sedivy 1995 and Trueswell et al. 1999 for a complete review and references. 2.â•‡ The verb put assigns a third theta role to its external argument, the subject NP. However, this argument has no empirical consequences for the process of ambiguity resolution, so we do not consider it further (cf. Ni, Crain and Shankweiler 1996). 3.â•‡ The example is open to another interpretation, in which the PP ‘in the box’ is interpreted as a modifier of the preceding NP ‘the napkin’. This meaning can be paraphrased as “Put the frog on the napkin that is in the box.” Although this interpretation may be


61

accessed in certain contexts, it is ruled out in the one-referent context by the Principle of Referential Success, since there is no napkin in the box. 4.â•‡ Trueswell et al. adopt the Lexicalist Parsing model to explain the VP-attachment preference (MacDonald et al. 1994; Trueswell and Tanenhaus 1994). According to the model, the preference for VP-attachment is based on the lexical properties of the linguistic input (i.e., the verb put) and the frequency with which particular syntactic constructions are used. The verb put frequently associates the first PP with the Destination theta role, so the Lexicalist Parsing model predicts a VP-attachment preference for the sentences under consideration. 5.â•‡ This is our interpretation of the study by Trueswell et al. We refer the reader to the original paper for further discussion of a particular version of the Referential Theory, and for presentation of the Lexicalist Parsing Model. 6.â•‡ We use quotation marks to refer to the pattern of children’s non-adult behavior to avoid committing to the view that these responses were errors. 7.â•‡ For the purposes of this paper, we focus on the actions performed by children in both contexts under consideration. We refer the reader to Trueswell et al. 1999 for a complete and detailed discussion of the on-line eye-movement patterns. 8.â•‡ The results of Trueswell et al. (1999) have been replicated by Hurewitz et al. (2000). 9.â•‡ The logic behind this argument was sketched but not pursued by Trueswell et al. (1999). One qualification about our account is needed, however. We do not think that a visual difference in the context alone determines the salience of one frog over the other. Rather, what is relevant is how that difference is encoded in the linguistic instructions. 10.â•‡ In a later section, we present experimental evidence that children do make a pragmatic inference to identify a salient frog. 11.â•‡ The children who participated in experiment II did not participate in experiÂ� ment Â€I. 12.â•‡ The children who participated in experiment III did not participate in experiment I or experiment II. 13.â•‡ Despite the occurrence of a color adjective in the instruction, a temporary ambiguity emerges in (5) after the first PP, as in the study by Trueswell et al. 14.â•‡ The children who participated in experiment IV did not participate in experiment I, experiment II, or experiment III. 15.â•‡ Trueswell et al. (1999) reported roughly 60 percent incorrect actions. References Altmann, Gerry, and Steedman, Mark. 1988. Interaction with context during human sentence processing. Cognition 30, 191–238. Amidon, Arlene, and Carey, Peter. 1972. Why five-year-olds cannot understand before and after. Journal of Verbal Learning and Verbal Behavior 11, 417– 423. Britt, Anne M. 1994. The interaction of referential ambiguity and argument structure. Journal of Memory and Language 33, 251.

62

Meroni and Crain

Clark, Eve V. 1971. On the acquisition of the meaning of before and after. Journal of Verbal Learning and Verbal Behavior 10, 266 –275. Crain, Stephen. 1982. Temporal terms: Mastery by age five. Papers and Reports on Child Language 21, 33–38. Crain, Stephen. 1991. Language acquisition in the absence of experience. Behavioral and Brain Sciences 14, 597– 650. Crain, Stephen. 2002. The continuity assumption. In I. Lasser (ed.), The Process of Language Acquisition. Peter Lang. Crain, Stephen, and Steedman, Mark. 1985. On not being led up the garden-path: The use of context by the psychological parser. In D. Dowty, L. Kartunnen, and A. Zwicky (eds.), Natural Language Parsing. Cambridge University Press. Crain, Stephen, and Wexler, Kenneth. 2000. Methodology in the study of language acquisition. In W. C. Ritchie and T. K. Bhatia (eds.), Handbook on Language Acquisition. Academic Press. Ferreira, Fernanda, and Clifton, Charles. 1986. The independence of syntactic processing. Journal of Memory and Language 25, 348–368. Frazier, Lyn, and Rayner, Keith. 1982. Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology 14, 178–210. Gibson, Edward. 1991. A Computational Theory of Human Sentence Processing: Memory Limitations and Processing Breakdowns. Ph.D. thesis, Carnegie Mellon University. Gorrell, Paul. 1995. Syntax and Parsing. Cambridge University Press. Hamburger, Henry, and Crain, Stephen. 1982. Relative acquisition. In S. Kuczai (ed.), Language Development II. Erlbaum. Hamburger, Henry, and Crain, Stephen. 1984. Acquisition of cognitive compiling. Cognition 17, 85–136. Hamburger, Henry, and Crain, Stephen. 1987. Plans and semantics in human processing of language. Cognitive Science 11, 101–136. Hurewitz, Felicia, Brown-Schmidt, Sarah, Thorpe, Kirsten, Gleitman, Lila, and Trueswell, John. 2000. One frog, two frog, blue frog: Factors affecting children’s syntactic choices in production and comprehension. Journal of Psycholinguistic Research 29, 567– 626. MacDonald, Maryellen C. 1994. Probabilistic constraints and syntactic ambiguity resolution. Language and Cognitive Processes 9, 157–201. MacDonald, M. C., Pearlmutter, N., and Seidenberg, M. 1994. The lexical nature of syntactic ambiguity resolution. Psychological Review 101, 676 –703. Matthei, Edward M. 1981. The acquisition of prenominal modifier sequences. Cognition 11, 301–332. Ni, Wejia, Crain, Stephen, and Shankweiler, Donald. 1996. Sidestepping garden paths: Assessing the contributions of syntax, semantics and plausibility in resolving ambiguities. Language and Cognitive Processes 11, 283–334.


63

Pinker, Steven. 1984. Language Learnability and Language Development. Harvard University Press. Sedivy, J. C. 2003. Pragmatic versus form-based accounts of referential contrast: Evidence for effects of informativity expectations. Journal of Psycholinguistic Research 32 (1), 3–23. Sedivy, J. C., Chambers, C., Tanenhaus, M., and Carlson, G. 1999. Achieving incremental semantic interpretation through contextual representation. Cognition 71, 109– 147. Spivey, Michael, and Tanenhaus, Michael. 1998. Syntactic ambiguity resolution in discourse: Modeling the effects of referential context and lexical frequency. Journal of Experimental Psychology: Learning, Memory and Cognition 24, 1521–1543. Spivey-Knowlton Michael, and Sedivy, Julie. 1995. Resolving attachment ambiguities with multiple constraints. Cognition 55, 227–267. Tavakolian, Susan L. 1981. The conjoined-clause analysis of relative clauses. In S. Â�Tavakolian (ed.), Language Acquisition and Linguistic Theory. MIT Press. Tanenhaus, Michael K., Spivey-Knowlton, Michael J., Eberhard, Kathleen M., and Â�Sedivy, Julie. 1995. Integration of visual and linguistic information in spoken lanÂ� guageÂ€comprehension. Science 268, 1632–1634. Trueswell, John C., and Gleitman, Lila. 2004. Children’s eye movements during listening: Developmental evidence for a constraint-based theory of sentence processing. In J. M. Henderson and F. Ferreira (eds.), The Interface of Language, Vision, and Action: Eye Movements and the Visual World. Psychology Press. Trueswell, John C., Sekerina, Irina, Hill, Nicole M., and Logrip, Marian L. 1999. The kindergarten-path effect: Studying on-line sentence processing in young children. Cognition 73, 89–134. Trueswell, John, and Tanenhaus, Michael K. 1994. Towards a lexicalist framework for constraint based syntactic ambiguity resolution. In Charles Clifton, Lyn Frazier, and Keith Rayner (eds.), Perspectives on Sentence Processing. Erlbaum. Weinberg, Amy. 1992. Parameters in the theory of sentence processing: Minimal commitment theory goes east. Journal of Psycholinguistic Research 22, 339–364.

4â•…

Referential and Syntactic Processes: What Develops? John C. Trueswell, Anna Papafragou, and Youngon Choi

Introduction

A primary purpose of language is to permit individuals to communicate their perceptions and conceptions of the world. The linguistic system that underlies this communication must therefore be designed for fairly intricate interactions with the human perceptual and conceptual machinery. The study of sentencecomprehension abilities in adults shows quite clearly that this is the case. For instance, it has been found that the recognition of a word includes accessing detailed linguistic information about how that word is likely to combine syntactically and semantically with the current representation of the sentence. In addition, the referential implications of these analyses are computed in real time and appear to exert a simultaneous influence on the ongoing structural analyses, allowing the listener to pursue referentially plausible parses and exclude referentially implausible ones. This rapid “dance” among syntactic, semantic, and referential factors over the course of interpreting a sentence has led researchers to the conclusion that the recognition of a word in a sentence exerts immediate effects on multiple tiers of linguistic and nonlinguistic representation, i.e., phonological, syntactic, semantic, and referential. These representational systems, though distinct, mutually constrain each other in a dynamic fashion (Jackendoff 2002; MacDonald, Pearlmutter, and Seidenberg 1994; Trueswell and Tanenhaus 1994). In this paper, we explore the question of how this sentence-processing machinery develops in children. The study of the development of sentence processing has attracted renewed interest with the advent of a methodology for studying children’s comprehension abilities in real time. In these studies, children’s eye-gaze patterns to objects in the world are recorded as they hear sÂ�poken utterances about this world, with the measure providing a moment-bymoment window into their interpretation process. We will suggest from these and other data that the system is incremental and interactive at a relatively

66

Trueswell et al.

early stage in development, showing sensitivity to a variety of constraints on computing sentential meaning. But at the same time we will argue that there are systematic changes over developmental time in the reliance on certain sources of linguistic and nonlinguistic evidence — changes that depend upon the validity and reliability of the evidence as derived from the learner’s experience. The dynamic abilities of this processing system are also found to change and mature in time; this interface system, like many others, is subject to developmental changes in the control of information processing, especially changes in cognitive and attentional control. In order to begin our discussion of this theory of parsing development, we must first discuss what is known about the adult end state, that is, what we believe to be true about the sentence-comprehension abilities of adults and how they use multiple constraints to shape their on-line structuring of the input. With this account in hand, we will use it to motivate a developmental theory of sentence processing. We will then turn to experimental evidence that we believe justifies our claims. Real-Time Sentence Processing in Adults

Let us first consider what are likely to be inescapable truths about sentence comprehension in adults. First, given the way natural languages work, it is almost certainly the case that listeners must recover much or all of the intended syntactic structure of an utterance. This is because the structural characteristics of an utterance, when combined with the semantics of verbs and other lexical items, convey essential role assignments (i.e., who is doing what to whom). This, of course, has been the bread and butter of not only linguistics but also of most psycholinguistic research on sentence processing carried out over the last thirty years. Less often discussed, but arguably equally important, is the fact that the structure of an utterance simultaneously conveys intended discourse operations (e.g., discourse status and focusing). Grammatical choices made by a speaker (whether to passivize, whether to include a restrictive modifier, etc.) reflect discourse considerations and are designed to communicate what the speaker is referring to. Listeners therefore need this syntactic information to infer the intended meaning of an utterance and its reference to the world. Thus listeners must look for evidence in the linguistic input about the syntactic operations that gave rise to the utterance. Exactly how syntactic and semantic structure is recovered by a listener or a reader has been a topic of considerable debate (e.g., Altmann and Steedman 1988; Frazier and Fodor 1978; Frazier 1987; MacDonald et al. 1994; Trueswell and Tanenhaus 1994). In this paper, we will sketch only our own view — the

Referential and Syntactic Processes

67

view that motivates our developmental account. Specifically, we assume that during the comprehension of a sentence, listeners are engaged in the recovery of phonological, syntactic, and semantic characterizations of the input, each of which is maintained within partially independent representational systems (representational modularity). These representational systems dynamically constrain each other as the sentence unfolds (dynamic interactive processing). These three characterizations of the input ( phonological, syntactic and semantic) should be thought of as interim representations whose primary use is to allow listeners to update their mental model of the world (including what they believe speakers are trying to communicate). Importantly, we assert that the recovery of these interim representations is done in real time via probaÂ� bilistic mechanisms. The process of recognizing a word within a sentence acÂ� tivates probable phonological, syntactic, and semantic structures in parallel, including if necessary multiple alternatives within each subsystem. In turn, interface mechanisms act in real time as the sentence is unfolding to converge on the most consistent and probable solution across these domains (see Trueswell and Tanenhaus 1994; Kim, Srinivas, and Trueswell 2002). It follows from this account that the frequency-based accessibility of structural alternatives will play an important role in a comprehender’s ability to converge on the intended meaning of an utterance. Perhaps the best evidence for this claim comes from adult studies of temporary ambiguity during reading and listening. For instance, in the fragments below, temporary ambiguities arise as readers or listeners attempt to rapidly structure the input: (1)â•‡ The man sliced the loaf with .â•–.â•–. (2)â•‡ The child believed the doctor .â•–.â•–. In (1), with could be linked to the verb sliced, thereby denoting an instrument (e.g., with the sharp knife), or be linked to the noun phrase the loaf, thereby denoting possession of a property (e.g., with the burnt crust). In (2), the noun phrase the doctor could be structured as the direct object of believed (e.g., so that the sentence could end there) or as the start of an embedded sentence that is a complement of the verb (.â•–.â•–. believed the doctor was lying). A wealth of experimental findings now suggest that the structural and semantic analyses that comprehenders assign at the point of ambiguity are determined in part by detailed lexical factors. These factors include the probability that a verb takes particular complements, as well as the semantic fit of constituents into the intended roles assigned by the verb (e.g., Britt 1994; Garnsey, Pearlmutter, Myers, and Lotocky 1997; Trueswell, Tanenhaus, and Kello 1993; Trueswell, Tanenhaus, and Garnsey 1994). For instance, the tendency for the

68

Trueswell et al.

verb slice to include an Instrument role in the form of a PP and the tendency forÂ€ believe to include a Patient in the form of an NP predict initial parsing pÂ�references by readers and listeners encountering these phrases (e.g., Garnsey et al. 1997; Taraban and McClelland 1988).1 All of this suggests that wordrecognition processes often drive the structuring of input — indeed, some studies show that covert priming of a verb with different syntactic and semantic tendencies can unconsciously affect comprehenders’ parsing preferences for ambiguous phrases ( Novick, Kim, and Trueswell 2003; Trueswell and Kim 1998). This probabilistic recovery of structure is sensitive to other contingencies as well. In particular, the referential implications of these interim representations are computed in real time and can serve as an important top-down constraint on sentence processing. For instance, Altmann and Steedman (1988) found that readers structure phrases like sliced the loaf with the .â•–.â•–. differently depending on the contents of the story leading up to this sentence. For example, when there are two different loaves in the story, readers prefer to interpret the with-PP as a modifier of the preceding NP (the loafâ•ƒ ). The idea here is that a definite NP must uniquely specify a referent within the current referential domain (Crain 1980; Crain and Steedman 1985). If the simple NP the loaf fails to do this, further linguistic information is expected in the form of a post-NP modifier. Indeed, indefinite NPs (a loafâ•ƒ ) alter this parsing preference (SpiveyKnowlton and Sedivy 1995), and even other referential factors contribute to parsing decisions (Trueswell and Tanenhaus 1991). It seems, at least from this evidence, that comprehenders must also be dynamically tracking what is under discussion and what is within the current referential domain, since these factors rapidly influence the structuring of the input.2 Importantly, however, reading studies of this sort also indicate that the effectiveness of this contextual factor depends on the availability of the structural options at issue. For instance, in (1) above, the effectiveness of the twoloaf story in supporting a modifier interpretation depends on the kind of verb that is used in the stimuli: verbs that often include an Instrument role show substantially delayed and reduced contextual effects (Britt 1994; SpiveyKnowlton and Sedivy 1995; see also Garnsey et al. 1997 and Trueswell 1996 for lexically determined accessibility in other structures). In the picture that emerges from these data, the recognition of a word within a sentence automatically triggers linguistic representations at multiple levels. But this triggering is probabilistic in nature: given the evidence at hand, a listener is engaged in a “guessing game” in which the linguistic procedures that gave rise to the utterance are recovered. The referential implications of these representations are also computed in real time and, when possible, used to constrain the representational hypotheses that the listener is considering.


69

It cannot be emphasized enough that local ambiguity is ubiquitous in realtime language comprehension. Indeed, computational linguists have recognized the pervasiveness of ambiguity, especially since they began to implement “wide-coverage” parsers and interpreters that were designed to handle naturally produced text (e.g., Marcus, Santorini and Marcinkiewicz 1993). It has even been claimed that local ambiguity of the sort found in highly lexicalized formalisms can provide a processing advantage because it permits greater flexibility in recovering structure and meaning (e.g., Srinivas and Joshi 1999; Steedman 2000; see also Kim, Srinivas, and Trueswell 2002). The implication here is that the accessibility of structure is an unavoidable issue in the study of language comprehension, whether one is interested in syntax, in semantics, or in reference. We strongly suspect the same is true for the study of language comprehension in children. A Developmental Account of Sentence Comprehension

Let us now turn to how the child learns to implement these dynamic sentenceprocessing abilities, focusing especially on how referential contingencies are learned and used. To begin, we must spell out some basic assumptions of our account. First, we will assume a great deal of processing continuity over development. That is, the types of processes used for language comprehension remain constant throughout language learning and into adulthood. In particular, we will assume the following. 1.â•‡ Real-time processing continuity: From the outset, a language learner or a listener is attempting real-time incremental processing of the input speech stream. 2.â•‡ Probabilistic processing continuity: From the outset, the detection from the speech stream of already acquired linguistic elements (including syntactic and phrasal elements) is achieved via probabilistic pattern-recognition and patterncompletion processes. There is good reason to believe that these assumptions hold, especially as they pertain to the processing of sub-lexical and lexical elements. In particular, experimental results from Aslin, Newport, and Saffran indicate that 8–12-montholds are sensitive to the distributional properties of syllables, which allows them to discover likely lexical or morphological candidates from continuous speech (e.g., Aslin, Saffran, and Newport 1998; Saffran 2001, 2002; Saffran, Aslin, and Newport 1996). These results suggest that from the outset, language learners are attempting to extract potentially relevant linguistic elements from the input via probabilistic mechanisms. In turn, these elements serve as candidates for word learning. That is, language learners attempt to map these newly discovered elements onto known conceptual representations.

70

Trueswell et al.

As language learners build up a repository of word-meaning pairs, they are faced with temporary ambiguity from the start. They deal with this ambiguity in an adult-like manner — i.e., in real time, as the speech unfolds. For instance, the eye-tracking research of Fernald, Swingley, and colleagues shows that 18–24-month-olds process phonological word cohorts (dog/doll; tree/truck) in much the same way as adults, the major difference being that adults know more words (Swingley, Pinto, and Fernald 1999; Allopenna, Magnuson, and Tanenhaus 1998). Upon hearing doll in a sentence like Look at the doll, 18–24 month-olds will look to cohort referents of doll, such as a picture of a dog, but not to non-cohort referents, such as a picture of a mouse. Suggestive evidence also exists indicating that children in this age range are beginning to engage in real-time syntactic and semantic structuring of these utterances. For instance, 27-month-olds hearing Let’s roll the ball restrict looks to a ball, excluding a non-rollable object in view, as the word ball is being heard (Fernald 2004, as reported in Fernald, Zangl, Thorpe, Hurtado, and Williams 2008; see also Nation, Marshall, and Altmann 2003 for similar studies with older children). All these patterns have been observed in real-time studies of adult listeners (e.g., Allopenna et al. 1998; Altmann and Kamide 1999), suggesting considerable continuity of processing abilities over development. The question of interest here, though, is whether children’s sentence-Â� processing abilities show the same sort of continuity over development, which is exactly what we wish to assert. One obvious obstacle, of course, is that sentence comprehension is arguably orders of magnitude more complex than word recognition. We therefore must make the following further assumptions about the development of language processing. 3.â•‡ Representational modularity: The language-processing system is innately predisposed to organize linguistic input into three partially independent representational domains: phonological, syntactic, and semantic. 4.â•‡ Representational interfacing: The language learner expects systematic correspondences between these representational systems. For instance, the number and type of phrasal constituents present in an utterance will have a systematic mapping onto the number and type of participants denoted in the conceptual representation of an event (Gleitman 1990; Gleitman, Cassidy, Nappa, Papafragou, and Trueswell 2005). 5.â•‡ Interactive processing: From the outset, the language system is detecting and taking advantage of probabilistic tendencies among phonological, syntactic, and semantic elements computed from the input stream so as to constrain possible analyses of this input.


71

6.â•‡ Assume reference: The language learner is innately predisposed to assume that communicative acts refer to the world. Thus, from the outset, the language learner attempts to compute the referential implications of the linguistic characterizations of the input. These assumptions, when combined with what we believe are properties of the adult comprehension system, allow us to derive some predictions about how child listeners ought to resolve temporary syntactic ambiguity during sentence comprehension. First, like the adult system, the child sentence-comprehension system is engaged in the recovery of known syntactic and phrasal categories from the input, which is accomplished via pattern-recognition processes. These higher-order syntactic and phrasal elements are likely to be discovered via distributional and statistical mechanisms similar to those proposed for lexical discovery by Saffran, Newport, and colleagues (i.e., Mintz, Newport, and Bever 2002; Gerken 2002; Gómez 2002; Gómez and Gerken 2000; cf. Harris 1957). Crucially, though, we assert that certain sorts of categories are preferred by the linguistic processing system and are assumed to map onto semantic and conceptual representations in systematic ways (assumptions 3 and 4 above). Once a repository of syntactic representations has been learned, we would expect a processing situation somewhat similar to the one characterized inÂ€early lexical processing (and documented by Fernald, Swingley and colleagues): namely, a situation in which the child parsing system must also deal with syntactic ambiguities and must resolve these ambiguities in real time. Since the adult syntactic parsing system is a probabilistic device that weighs multiple contingencies, it follows that the child processing, though organized and operating in the same way, must gradually discover and learn these contingencies. For illustration, consider the Prepositional Phrase (PP). The PP in English is associated with a range of semantic functions, often ambiguously for any particular lexical head. PPs can be used for temporal and spatial specification of events or entities (e.g., sang on Tuesday, sang on the stage, the book on your left). PPs are also used as arguments of events and entities (e.g., given to Susan, put on the table, governor of California). Consider a child who has already learned that PPs serve these semantic functions.3 As the child detects a particular instance of a PP in the input stream (e.g., I really like your doll with the .â•–.â•–.), how does he or she decide which of these semantic (and syntactic) operations to compute? Based on the adult literature outlined above, we can list the possible sources of evidence: •â•‡ •â•‡

lexical head (e.g., with, on, in) and lexical semantics/syntax syntactic structure up to that point (e.g., verb-syntactic-projections)

72

Trueswell et al.

Figure 4.1

The syntactic alternatives associated with the ambiguous word with, shown using the LTAG formalism (Srinivas and Joshi 1999). •â•‡

semantic structure up to that point (e.g., verb-thematic-projections) referential operations up to that point (e.g., currently insufficient referential specificity) •â•‡ prosodic structure up to that point (e.g., currently open intonational phrase) •â•‡

First, the presence of a particular lexical head (with, on, ofâ•ƒ ) constrains the range of syntactic and semantic alternatives. But this is only probabilistic evidence for the listener. For instance, with is an ambiguous word that denotes, roughly, either ‘accompaniment’ or ‘instrumental-manner,’ and is associated with several syntactic operations (see figure 4.1). Like any ambiguous word, the probability of activating any of these linguistic characterizations will depend on the word’s dominant and subordinate meanings. Second, the current syntactic structure and semantic structure ( points 2 and 3) constrain alternatives probabilistically as well: Let’s color with .â•–.â•–. requires VP-attachment but either as an accompaniment (Let’s color with your friends) or as an instrument (Let’s color with your crayons). Also, the presence of a NP object permits other syntactic-semantic options (e.g., Let’s color the book with the torn cover). The probability of these alternatives is determined largely by the subcategorization and thematic tendencies of the verb, again as discussed above. Third, referential implications up to that point will play a role in resolving a particular PP occurrence, since, as was discussed above, the lack of referential success of a definite NP supports modification (e.g., the book with .â•–.â•–. when there is more than one book under discussion).


73

Finally, although not discussed above, prosody serves as probabilistic evidence, since the presence or absence of a major prosodic break has a correspondence to phrasal breaks — though again probabilistically. Now for the developmental question. Given these possible constraints on parsing and interpreting PPs, which of these constraints are going to be more valid predictors of semantic/syntactic choice, and hence appear to dominate child parsing and interpretation processes? If the literature on adult sentenceprocessing is any guide, we should expect that lexical constraints on structural analyses ( points 1–3) will play an early and potent role developmentally. Adults track subcategorization and thematic preferences to such a great extent that they immediately constrain parsing options. If children build such databases as they learn words, it follows that this information will appear as an early determinant of child parsing. Beyond the adult literature, there is additional evidence for this conclusion. Research on verb learning strongly suggests that children track the number and types of phrases that occur with verbs so as to assist in learning the meaning of these verbs (e.g., Fisher, Hall, Rakowitz, and Gleitman 1994; Gillette, Â�GleitÂ�man, Gleitman, and Lederer 1999; Gleitman 1990; Yuan and Fisher 2009). Said another way, children from an early age track subcategorization and aÂ�rgument-taking properties of verbs as they learn them. Our assertion here is that this probabilistic evidence, which was tracked and developed so as to discover the meanings of verbs, isn’t “thrown away” after the verb is learned. Rather it is used to recognize the intended structure of an utterance every time that particular verb is encountered again later in life. Moreover, children, like adults, deploy this knowledge of probabilities as a sentence unfolds. What about potential top-down influences of referential implications ( point 4)? Will children show similar early influence of such factors? Given the assumptions sketched above, the answer should be yes, as long as the particular contextual evidence is valid, accurately computed by the child, and reliably constrains the structural analysis. After all, we argue that it is a great advantage for a listener (adult, child, or infant) to discover the referential conditions uÂ�nder which instances of particular linguistic elements are occurring. However, we have reason to believe from the above-mentioned literature on adult parsing that the referential constraints on PP modification are substantially weaker than relevant lexical predictors of these same structures (e.g., Britt 1994). Moreover, it seems reasonable to assume that the ability to track what is currently relevant in the contextual setting of an utterance could be hindered in children, since this often requires dynamically tracking what the speaker is thinking about (e.g., Clark 1993). From children’s difficulties in taking other iÂ�ndividuals’

74

Trueswell et al.

perspectives, as studied in the theory of mind literature, it follows that the referential evidence children build for the purposes of sentence comprehension is going to be noisy and even contaminated. As a result, it should be expected, somewhat counter-intuitively, that lexical predictors of structure will exert an early and potent influence on child processing, whereas the sorts of contextual factors thus far studied in the adult literature will be developmentally delayed in children’s ambiguity-resolution abilities. The Kindergarten-Path Effect

Much of the initial impetus for developing this account comes from the results of a study of child sentence processing by Trueswell, Sekerina, Hill, and Logrip (1999). In that study, 5-year-olds, 8-year-olds, and adults were given spoken instructions to move objects around on a table while their eye movements were tracked.4 Eye position was used to infer listeners’ ongoing referential commitments, which are believed to be derived from provisional syntactic and semantic analyses of the spoken utterance (Tanenhaus, Spivey-Knowlton, Eberhard, and Sedivy 1995; Spivey, Tanenhaus, Eberhard, and Sedivy 2002). On critical trials, participants were presented with one of two types of visual scenes, examples of which are shown in figure 4.2. For both scene types, the child was given an instruction that was designed to refer to a particular stuffed animal — in this case, the frog that was sitting on a napkin, as in (3). (3)â•‡ Put the frog that’s on the napkin into the box. ( Unambiguous Modifier)5 Consider hearing this sentence in the context of the scene in figure 4.2a, where there is only one frog but there are two napkins, one of which is under the frog. If children were simply looking to possible referents based on the individual

Figure 4.2

One-referent (a) and two-referent ( b) scenes from Trueswell, Sekerina, Hill, and Logrip 1999.


75

words they heard, we might expect them to look to the frog upon hearing frog but to look to either napkin upon hearing napkin. However, if referential commitments (and the resulting eye movements) are derived from real-time structural analyses, and children of all ages are engaged in these real-time processes (assumptions 1 and 6 above), we should expect that both children and adults would not consider the empty napkin as a referent, since the word napkin is part of a relative clause that must modify the NP the frog. Indeed, this latter pattern was exactly what was found: children and adults looked to the frog within a few hundred milliseconds of hearing frog, and continued looking at the frog (and the napkin under it) when hearing napkin, rarely if ever looking over to the empty napkin. Notice however that when sentence (3) is heard in the context of the scene in figure 4.2b the structural position of the frog in the sentence permits reference to either frog, and it is only upon hearing that’s on the napkin that a listener could compute the correct referent. Indeed, eye movements in two-frog scenes supported this expectation. Upon hearing frog, listeners, regardless of age, launched an eye movement to a frog, but they were at chance as to which frog they looked at. This state of affairs remained until hearing napkin, upon which participants shifted gaze to the frog on the napkin if they were at that moment looking at the wrong frog. Again, participants rarely if ever considered the empty napkin. It should also be noted that participants were essentially flawless at carrying out the instruction, regardless of age: they moved the intended frog into the empty box. Thus the data from syntactically unambiguous trials provide compelling evidence that syntactic and semantic analyses are engaged in real time for the ages we have looked at here, and that the referential implications of these structural analyses are reflected in eye movements. It is possible that the pattern could in principle be explained as children’s engaging in some sort of intersecting-set strategy ( hearing frog causes looks to frogs, hearing napkin causes looks to frog-plus-napkin). However, data from another variant of the instruction rule out this possibility and confirm other assumptions and hypotheses sketched above. In particular, on certain trials participants heard instructions like (4). (4)â•‡ Put the frog on the napkin into the box. (Temporary Ambiguity) Here the PP on the napkin becomes temporarily ambiguous. It could be interpreted as a modifier of the NP the frog or as a Goal for the verb put, i.e., where to put the frog. If participants are engaged in real-time probabilistic estimation of the intended structure regardless of age (assumption 2) and they use lexical evidence

76

Trueswell et al.

to compute these estimations (as hypothesized above), we should expect a strong initial preference to interpret on the napkin as a goal (where to put the frog) rather than a modifier. This is because the semantics of the verb put requires a goal role, and the goal role is typically realized in the form of a PP headed by on or in. Even 5-year-olds are expected to have ample linguistic experience to know these structural tendencies for this verb.6 If in fact the NP the napkin is interpreted as part of a goal PP, the empty napkin in the scene now becomes a possible referent, since one could move a frog to it. Thus we should expect increased looks to the empty napkin upon hearing the word napkin in temporarily ambiguous sentences like (4). Indeed, for one-frog scenes (figure 4.2a) looks to the frog were immediately followed by looks to the empty napkin upon hearing napkin. This occurred for all ages, suggesting that the syntactic and semantic tendencies of verbs are used to estimate structure and compute referential implications in real time. Two-frog scenes, however, were designed such that referential implications of hearing the frog discouraged interpreting on the napkin as a goal. This is because the frog could refer to either frog. If listeners can use this referential analysis in real time to constrain the structuring of further linguistic input, we ought to expect them to prefer to interpret on the napkin as a modifier of the NP the frog, allowing them to specify a unique referent. This top-down constraint, however, must battle against the lexical biases that support the goal analyses of that same phrase. Thus, two-frog scenes provide the potential for a top-down constraint on structuring on the napkin that ought to reduce looks to the empty napkin in the scene, rendering the pattern more like what was seen for unambiguous modifiers like that’s on the napkin. However, we hypothesized above that less reliable evidence for estimating the structure of the input should result in a developmental delay in using this evidence. We also suggested above that this particular sort of referential constraint on structure is less reliable than lexical constraints, in part because referential factors appear to be less effective in adult parsing behaviors. Moreover, we suggest that in order to discover this sort of referential-syntactic contingency children must be fairly skilled at tracking what is currently under discussion (i.e., what is relevant to the speaker). Any difficulty in computing the referential domain of an utterance would contaminate the ability to discover this contingency. Thus, if any developmental progression is expected, it would be that younger children would fail to take into account that the two-frog scene supports a modifier analysis. Indeed, only older children and adults interpreted the ambiguous on the napkin as a modifier in two-frog scenes. That is, they looked randomly to either frog in the scene upon hearing frog but rapidly converged on the intended frog


77

upon hearing napkin, rarely if ever looking at the empty napkin. In contrast, 5-year-olds showed a strong preference to interpret on the napkin as a goal, just as in one-frog scenes. Upon hearing frog, they were at chance looking at either frog. And upon hearing napkin they were still at chance looking at either frog, which suggests that they often failed to realize that on the napkin could be a modifier of the NP the frog. Instead, hearing napkin triggered increased looks to the empty napkin just as much as one-frog scenes. Thus, 5-year-olds tended to think on the napkin was the goal of put rather than a modifier, despite the presence of two frogs. This was in striking contrast to how these same children behaved on unambiguous sentences like (3), where they realized the modifier analysis. We are fairly confident of our account of 5-year-olds’ interpretations bÂ�ecause of the resulting overt actions that were made by these children. In particular, 5-year-olds were so strongly committed to the goal analysis on ambiguous tÂ�rials that they often carried out actions that involved moving a frog to the empty napkin. On over 60 percent of the trials, they performed such an action, regardless of the type of referential scene. On unambiguous trials, they were nearly flawless, making errors on about only 5 percent of trials. Thus the second PP, into the box, did not always block the goal analysis of on the napkin. It did, however, for older children and adults: they were nearly flawless on all trials regardless of ambiguity or referential scene, though, as expected, onereferent scenes did induce some confusion and errors on ambiguous trials, even for adults. Questions about Our Account of the Kindergarten-Path Effect

Although the data presented in Trueswell et al. (1999) are consistent with our account, certain aspects of this account are as of yet inadequately motivated. We articulate these concerns here by posing two questions. The first question pertains to children’s insensitivity to the referential-scene manipulation: •â•‡ Why

are two-referent scenes less reliable and less effective at resolving this modifier-argument ambiguity? We have explained developmental differences in the ability to use two-referent scenes by appealing to the poorer reliability of this evidence. In particular, we suggested that this situation (i.e., hearing a definite NP that could refer to mÂ�ultiple visually co-present objects) is not as good a predictor of modifier useÂ€as the local lexical evidence (e.g., put .â•–.â•–. on) which favors an argument analysis. Moreover, we suggested that the poor reliability of this referentialscene evidence arises in part from the difficulty of discovering that it can be informative.

78

Trueswell et al.

However, we have not offered much in the way of an explanation for why such a referential setting would be difficult for a child to discover as iÂ�nformative. Moreover, the claim that this information is less reliable than certain lexical information leaves partially unexplained the end state of this developmental pattern. After all, if this referential situation only weakly predicts the need for linguistic modification, why did Trueswell et al. (1999) and Tanenhaus et al. (1995) find that adults could use the implications of a two-referent scene to override the strong lexical bias to interpret on the napkin as a goal of put? Trueswell et al. (1999) did note that adults had some difficulty with temporary ambiguities in two-referent scenes, but this difficulty was surprisingly small in view of our claims. We will address this issue below in the section titled Referential Scenes, Definite Reference, and Restrictive Modifiers. Specifically, we will look to the literature on definite reference and discuss findings from a recent referential communication study, all of which suggests that a speaker’s choice about linguistic specification (e.g., saying “the frog” vs. “the frog on the napkin”) is not strongly determined by the mere presence of multiple identical objects of the relevant type (e.g., multiple frogs). With this knowledge in hand, we will take a second look at the adult listener’s ability to use these referentialÂ€ situations to resolve syntactic ambiguity. We discuss findings indicating that this contextual evidence is not as effective as the original put studies mightÂ€suggest (Snedeker and Trueswell 2004; Novick, Thompson-Schill, and Trueswell 2008). Moreover, these findings show that young children, though exquisitely sensitive to experimental manipulations of the lexical biases, are insensitive to the referential scene manipulation even under conditions of weakly biased lexical evidence. Taken together, the data paint a very reasonable picture about the listener’s use of referential situations to constrain the structuring of linguistic input. Reference and parsing will also be discussed in relation to a recent alternative account of the findings of Trueswell et al. (Wexler, this volume). Based on earlier data from young children’s misuse of the definite determiner the in their own productions (Maratsos 1976; Karmiloff-Smith 1979), Wexler proposes that children in the relevant age range lack a complete understanding of the semantics of the. In particular, he proposes that children lack the notion of maximality: the requirement that the definite determiner must apply maximally to the current referential domain. In the section of the present paper titled Pragmatic vs. Semantic Accounts, we evaluate this hypothesis with respect to a broader range of data from Maratsos and Karmiloff-Smith and suggest that the real problem lies in a child’s understanding of what the referential domain isÂ€ atÂ€ any given moment (roughly, the original conclusions of Maratsos and Karmiloff-Smith). Thus, Wexler is correct in concluding that the data of Mar-


79

atsos and Karmiloff-Smith shed light on the studies of child parsing, though perhaps not in the way he proposes. A second important question that we wish to address here involves the large difference Trueswell et al. observed between younger children and adults in their final interpretation of these temporary ambiguities: •â•‡ Why

did younger children fail to revise initial parsing commitments?

Trueswell et al. (1999) offered a developmental reason for why 5-year-olds failed to revise their goal interpretation upon hearing into the box. In particular, the authors suggested that children’s difficulty with revising was the result of limited processing resources or limited working memory, both of which expand with age. Since then, our research group has refined this claim to suggest specifically that this change in revision abilities is related to developmental differences in general executive-function processes, specifically the ability toÂ€ inhibit competing representations (Trueswell and Gleitman 2004, 2007; Novick, Trueswell, and Thompson-Schill 2005). In the section titled Revision and Lingering Garden-Paths, we discuss this proposal. We begin by pointing out that children are not the only ones who sometimes fail to revise their interpretation of a temporarily ambiguous phrase. Christianson, Hollingworth, Halliwell, and Ferreira (2001) have found that normal adults can also hold onto beliefs consistent with a rejected interpretation. And Mendelsohn (2002, 2003) has found that individual differences in measures of general inhibition and executive control correlate with this ability to revise parses. Referential Scenes, Definite Reference, and Restrictive Modifiers

Given the issues discussed above, a crucial question becomes whether a listener can deduce from scene information alone the need for restrictive modification, specifically, the need for a speaker to provide restrictive modification of a definite NP. Can a person look out into the visual world and anticipate from this information alone a speaker’s need to utter the little star (and not the star), the toy closest to you (and not the toy), or the frog on the napkin (and not the frog)? The pragmatics and psycholinguistics literature on linguistic reference indicates that this is not the case. For instance, Lyons (1980) sketches a hypothetical situation in which two people are working on a motorcycle and one says “Pass me the spanner” (spanner being British English for wrench). In a situation in which there are two wrenches present, one near the speaker and one near the listener, the listener is likely to infer that the wrench closest to the listener is the one intended. And it seems unlikely that the speaker would have said “Pass the spanner that is closer

80

Trueswell et al.

Figure 4.3

to you” (see also Lyons 1999). Similarly, consider figure 4.3 (adapted from Stone and Webber 1998). A person could refer to a particular rabbit in this scene by saying “Pull the rabbit out of the hat,” but it would be ludicrous to say “Pull the rabbit that’s in the hat out of the hat.” Thus, the referential domain, when applied to a scene, is ultimately determined by what is relevant given the goals of the interlocutors. As a result, a referentially ambiguous NP can be disambiguated by these factors rather than linguistically via restrictive modification. In fact, a recent study (BrownSchmidt, Campana, and Tanenhaus 2002) suggests that these “Lyons-esque” situations are relatively common in conversations about visually present objects. Brown-Schmidt et al. observed that adults do not utter restrictive modifiers every time there are multiple potential referents; nearly half of all definite NPs uttered (48 percent) did not have a unique referent in the scene (e.g., “Okay, pick up the square” might be uttered in the presence of multiple squares.) However, listeners’ eye movements, actions, and vocal responses all showed that they routinely achieved referential success under these conditions (e.g., picking up the correct square). Obviously, this success isn’t evidence for psychic abilities on the part of the subjects. Rather, success occurred because the shape of the discourse and the goals of the task had narrowed the field of possible referents down to one (e.g., only one of the squares was currently a plausible referent). Definite NPs containing restrictive modifiers were uttered only when multiple potential referents were currently under discussion. Now consider the child, who is trying to learn how deictic reference works for definite and indefinite NPs. If the child has difficulty understanding how the goals of an interlocutor restrict referential domains (a reasonable assumption), the child should have specific difficulty with these Lyons-esque situa-


81

tions. Perhaps not surprisingly, developmental studies of definite reference show that young children (3– 6 years) tend to behave egocentrically in these situations. In particular, in the absence of information that might guide a child’s referential domain to the one intended by the speaker, young children’s comprehension and productions suggest that what they assume to be the referent must also be what their interlocutor assumes to be the referent (Maratsos 1976; Karmiloff-Smith 1979). We will return to this issue in more detail when we discuss Wexler’s (2003) recent reinterpretation of this data, but here we will simply relate the egocentric account to the current put-instruction data. In particular, we now have better reason to believe our account of the child put-study. That is, children receive only sporadic ( probabilistic) evidence that a definite NP (the frog) that deictically refers to a member of a set of objects of the same type will require restrictive modification (e.g., the yellow frog, the frog on the napkin). Moreover, children’s discovery of the actual contribution of restrictive modifiers requires an understanding of common ground, the dynamics of discourse, and a decent model of shared goals with interlocutors (see Clark 1993; Tanenhaus, Hanna and Chambers 2004). The findings of Maratsos (1976) and Karmiloff-Smith (1979) suggest that a child understands that a definite NP must apply to the current domain of reference, but his or her estimation of the domain of reference can become misaligned with the interlocutors’, certainly more often than an adult’s estimation in a similar situation. With this in mind, it becomes easier to understand the child-adult differences found in the put study. Adults, when presented with a set of objects that have been labeled (a frog, a napkin, another frog, a box, another napkin), uÂ�nderstand that this set of objects reflects the current referential domain. Young children in such contexts should be expected to behave egocentrically, thinking when hearing the frog that the frog they are thinking of is likely to be the referent. After all, this is true in the common deictic referential situation of one entity, and in situations where other factors allow the child to be guided toward an understanding of a referential domain that is the same as the speaker’s, i.e., the correct subset. Indeed, as Trueswell et al. (1999) noted, children’s eye-Â� fixation patterns show this egocentricity; the frog they looked to first is a fairly good predictor of which frog they return to, and act upon, in their action of putting. Lexical and Referential Evidence in Interaction in Adults and Children

We have suggested that the referential-scene manipulations were inadequate: simply placing multiple objects of the same type within view doesn’t mean that reference to a particular member of this set will be referentially disambiguated via linguistic means (such as adding a restrictive modifier). Unfortunately, this

82

Trueswell et al.

account does not completely explain the behavior of adults in the put study. That is, even in two-referent scenes we should have expected adults to show some temporary consideration of the goal interpretation of on the napkin, because the semantics and syntax of put supports this interpretation (contra the referential context). We know from years of research in adult sentence comprehension that strong lexical biases that run against contextual /plausibility biases result in temporary misanalysis, or garden-pathing (e.g., Britt 1994; Garnsey etÂ€ al. 1997; Spivey and Tanenhaus 1998; Trueswell 1996). Indeed, this is a fundamental prediction of constraint-satisfaction accounts of comprehension, which motivated our account of the developmental patterns. One possibility worth considering quite seriously is that adults in fact did experience some difficulty with “Put the frog on the napkin into the box” in two-frog scenes, but other factors conspired to reduce this difficulty and perhaps even measurement of this difficulty. The presence of a second prepositional phrase, into the box, increases the likelihood that the first PP is an NP modifier and not a goal. Adults may be especially good at using late post-Â� ambiguity information to revise parses. Also, the primary measurement of garden-pathing in adults (looks to the empty napkin) occurs during and after the perception of this second prepositional phrase, into the box. These factors alone could have reduced signs of considering the goal interpretation across the board in ambiguous trials (in both one-referent and two-referent scenes), with two-referent scenes approaching the floor of the measure: essentially no looks to the empty napkin. Indeed, Novick, Thompson-Schill and Trueswell (2008) have found evidence to support these conclusions. In a slightly modified visual setting in which one frog was on a napkin and another frog was in aÂ€bowl, Novick et al. observed clear signs of consideration of the goal interpretation of on the napkin in two-referent contexts during the ambiguous phrase itself, which disappeared immediately upon hearing into the box. The results of Novick et al. (2008) suggest that even adults may have trouble using two-referent visual scenes to completely determine a modifier interpretation of a PP when disambiguating evidence is not provided in the sentence itself. The prediction is that when lexical information supports a particular parsing preference of a globally ambiguous sentence (as in 5a below), adults ought to show difficulty and possibly even a total failure to take the referentialscene implications into account. Snedeker and Trueswell (2004) addressed this question. In the study, college-age adults heard sentences like those in (5). (5) a. Tickle the pig with the fan. (Instrument-biased Verb) b. Feel the frog with the feather. (Equi-biased Verb) c. Choose the cow with the stick. (Modifier-biased Verb)


83

Figure 4.4

Clip-art illustrations of one-referent and two-referent scenes from Snedeker, Thorpe, and Trueswell 2001 and Snedeker and Trueswell 2004. Physical objects were used.

Here with the X could attach to the verb as an instrument or could attach to theÂ€noun phrase as a modifier. Verbs were selected on the basis of a separate sentence-completion study that evaluated how often a with-phrase would be used for these verbs as an instrument. As a result, verbs were operationally defined as likely to mention an instrument (Instrument-bias), unlikely to mention an instrument (Modifier-bias), or somewhere in between (Equi-bias). The nouns in the PP (â•ƒ fan, feather, stick) were normed in advance for thematic fit as an instrument for the corresponding verb. Nouns were selected that were rated as poor-to-adequate instruments in each verb class, such that average thematic fit was the same across these verb comparisons. The type of visual scene was also manipulated: one-referent vs. two-referent scenes, as shown in figure 4.4.7 The Instrument-biased sentences are most like the put items above. This is because the verb biases support a VP-attach assignment of the PP. Just as the empty napkin was a potential goal in the put-studies, the large stand-alone object (e.g., the feather) serves as the potential instrument here; looks to and use of this object provided a measure of the VP-attach (Instrument) interpretation of the PP (e.g., with the feather). The off-line action and eye-gaze data from 24 adults are presented in fÂ�igureÂ€4.5. On the left is plotted the proportion of trials in which an instrument action was performed by the subject—for example, the proportion of trials in which the subject picked up the potential instrument (e.g., the large feather) and used it to act on one of the animals (e.g., the frog wearing a party hat). The right bar graph shows the proportion of trials in which subjects looked at the potential instrument during the course of the trial, regardless of whether they picked itÂ€up. Consider first the action data from two-referent scenes. If these scenes had required an NP-modifier interpretation of ‘with the X ’, and adult listeners were

84

Trueswell et al.

Figure 4.5

Data from adults (Nâ•–=â•–24). Source: Snedeker and Trueswell 2004.

aware of this fact, we should have seen essentially no use of instruments in these scenes. That is, subjects should have opted to use their hand to act upon the target animal, e.g., the frog holding the feather. However, we see this occurs only for Equi-bias and Modifier-bias verbs (0 –10 percent instrument actions). For Instrument-biased verbs in two-frog scenes, adults showed nearly 70 percent instrument actions, far greater than the expected 0 –10 percent. It appears that most adults didn’t realize that ‘with the x’ could be an NP modifier in this condition, precisely because the verb so strongly suggested the Instrument analysis. If this is so, it means that the definite NP (e.g., the frog) in this condition was referentially ambiguous. Indeed, on these instrument action trials, subjects half the time acted on the target animal (e.g., the frog holding the feather) and half the time acted on the other animal (e.g., the frog wearing the party hat). Subjects who were given these items would sometimes ask “Which one?” in response to the first two-referent target trial (to which the experimenter said “Please do your best.”). Moreover, post-experiment interviews showed that subjects were behaving in a pragmatically appropriate manner in response to this referential ambiguity, often in the way Lyons (1980) suggested. When explaining (at the end of the experiment) why they acted on a particular frog, they offered the spanner strategy (“I thought you must have meant the frog closer to me”) or some other strategy (“I picked the frog not holding the feather just to be symmetrical” or “I picked randomly”). Averaging across all observations, subjects went for either animal with equal frequency.


85

Adults are not oblivious to the scene constraint on the need for a restrictive modifier. If the linguistic evidence (the verb information) isn’t heavily biased in favor of the instrument interpretation, listeners consider and choose the modifier interpretation (as evidenced by two-referent Equi and Modifier-Â�biased conditions). And of course, one-referent scenes increased the rate of using and looking at the potential instrument in all verb types. This should be expected, since singular definite NP reference within a scene containing multiple potential referents only partially predicts whether the NP will be contrastively marked with a modifier. What if the data we just presented had come from 5-year-olds rather than from adults? One might have concluded that 5-year-olds do not fully understand definite reference — e.g., they don’t understand the meaning of the definite determiner the (Wexler, this volume). But these are adults, not children. The logical conclusion is that syntactic accessibility of forms, as determined by lexical factors, has a potent effect on interpretation and reference. Of course, the question worth considering is how 5-year-olds behave in these very settings. Given that definite reference in these settings poorly constrains the structuring of an upcoming PP, it follows the 5-year-olds might be even less influenced by these contextual manipulations than adults. In addition, these same children ought to be quite sensitive to the lexical manipulations, since this information is highly predictive of structure and is hypothesized to be easy to track. The data from 5-year-olds (as reported in Snedeker and Trueswell 2004) are plotted in figure 4.6. Indeed, children perfectly matched the verb biases (they know which verbs invite an instrument interpretation) but the referential-scene manipulation had no effect on their actions and a small (non-significant) effect on their eye fixations. These data bode well for the developmental account sketched above. In adults, hearing definite reference in these two visual settings only partially contributes to parsing commitments, and verb biases strongly constrain structuring of the linguistic input. Children have trouble using this sort of deictic referential information, but have little trouble using the lexical information.8 Notice also from the data in figure 4.6 that one gets a very different pictureÂ€of what the child finds referentially ambiguous depending on the lexicosyntactic properties of the utterance. This clearly indicates that the study of what children “know” and “don’t know” about language must be embedded within a theory of how linguistic information is dynamically processed by children. This point is often overlooked in the traditional study of language acquisition (see, e.g., Wexler, this volume).

86

Trueswell et al.

Figure 4.6

Data from 5-year-olds (Nâ•–=â•–48). Source: Snedeker and Trueswell 2004. The Effects of Discourse and Pragmatics on Child Parsing

The linguistic observations and experimental findings sketched in the previous section strongly suggest that the shape of the discourse and the goals of the interlocutors ought to be a far better predictor of the referential domain of a referential expression and hence a better predictor of the level of specificity needed for that expression. For instance, recall that Brown-Schmidt et al. (2002) found that the conversational content, when combined with the scene, helps to shape the referential domain for a listener and predicts quite well the specificity of definite reference. If a discourse guides a child listener toward conceiving of the situation in the same way as the speaker, we might expect a child to use this contextual information to guide parsing commitments (egocentrically or otherwise). To this end, Hurewitz (2001; see also Trueswell and Gleitman 2004, 2007) asked whether potentially potent evidence from a spoken discourse can influence 5-year-olds’ parsing decisions. In the study, a preceding discourse, established by two conversing puppets, provided the goal to isolate one referent from among multiple referents in the scene before hearing an ambiguous PP. If these discourse goals provide a strong constraint on the need for the otherwise ambiguous PP to be a modifier, one might expect even 5-year-olds to be sensitive to this fact, and to combine it with lexical constraints on structure. In other words, Hurewitz (2001) asked if she could turn children into adults with this manipulation. In the study, children (Nâ•–=â•–24; age 4;0 –5;6) were tested in a modified version of the Truth Value Judgment task (Crain and Thornton 1998). On each


87

Figure 4.7

Illustration of objects presented to child in example stimuli from Hurewitz et al.

trial, the child heard a story acted out in the presence of a puppet (Mr. Walrus, known to be not terribly bright). At the end of the story, a second puppet (the clever Ms. Rabbit, who had been hiding under the table listening to the story), appeared and asked Mr. Walrus questions about the story. The child’s job was to evaluate and if necessary correct Mr. Walrus’s answers to her questions. On critical trials, each child was always presented with a two-referent scene (as in figure 4.7, two cats, one on a book, one on a fence, a toy barn, another fence, and a turtle; again, real-world objects, not clip-art images, were used). The story, pre-recorded and acted out by the experimenter (E), deictically referred to each animal and established the pair of cats in distinct events. It is paraphrased here: This cat [E grabs the cat on the book] and this turtle [E grabs turtle] decided to go for a walk, and met up on top of the barn [E moves animal to barn]. Suddenly, the turtle tickled the cat. ‘Tickle, tickle tickle!’ ‘Hee! Hee! Hee!’ [E performs appropriate actions with the animals.] And then they went home. [E returns each animal to original starting place.] And, this cat [E grabs cat on fence] saw all this and laughed and laughed as well.

With each animal back in its original location, Ms. Rabbit returns to ask Mr. Walrus a question. In all conditions, Mr. Walrus’s answer contains an attachment ambiguity: I know, the turtle tickled the cat on the fence. Here on the fence can indicate where the tickling happened (Locative VP-attachment) or can indicate a particular cat (Locative NP-attachment). Mr. Walrus’s utterance, however, was preceded by a question from Ms. Rabbit that either supported theÂ€need to contrast the cats (the Contrastive Question condition, Which cat did the turtle tickle?) or did not support that goal (the Non-Contrastive Question

88

Trueswell et al.

condition, Can you tell me something about the story?). In all cases, both interpretations of the ambiguous sentence are false because the story actually involved the cat on the book being tickled by the turtle in a different location, i.e., when they both had been on the barn. Hence, however the child parsed the sentence, he or she must still correct Mr. Walrus. It is the child’s particular correction of Mr. Walrus that can reveal the implicit parse choice (No! It happened OVER HERE on the barn! or No! THIS CAT was tickled, the one on the book!). The Question-Type factor (Contrastive vs. Non-Contrastive) was crossed with a verb manipulation. Half the trials involved eventive verbs (such as tickle), which easily allow for locative ( VP-attach) modifiers such as on the barn. The other half involved stative verbs, where the story and the critical sentence involved, e.g., liking (The turtle liked the cat on the fence.). Stative verbs do not usually permit locative modifiers, because states typically are not connected to a particular place. Given the importance of conversation constraints in modifier use (see above), the multiple-constraint account predicts that this sort of discourse manipulation ( here, Q-type) and lexical information ( V-type) should both influence parsing preferences, even in 5-year-olds. That is, Contrastive questions and Stative verbs should both induce greater modifier interpretations and resulting corrections by the child. Figure 4.8 plots the proportion of NP modification interpretations exhibited by children in the four conditions. The pattern of these corrections across conditions supports our account. In particular, Contrastive questions led to many more NP modification corrections than when the question was not Contrastive. Also, there was an effect of Verb type. Stative verbs led to more NP modification interpretations than when the verb was eventive. This resulted in reliable effects of Question type and Verb type. And, interestingly, adult controls exhibited an even stronger reliance on the discourse needs of the questions, with adults even coercing stative verbs into eventive readings. It should be emphasized that our account of child parsing is that it is the product of a system that automatically combines multiple evidential sources, all to serve the purpose of coming up with an estimate of the intended meaning of an utterance. If this is the case, we should expect to observe striking differences in children’s use of structures in production and comprehension tasks under particular comprehension conditions. That is, children might correctly produce a particular structure in the right referential context, but then fail to understand this same structure during comprehension, when the lexical sources support an alternative analysis.


89

Figure 4.8

Proportion of trials in which 5-year-olds gave NP Modification response (e.g., “No, it was this cat, the one on the book”) (data from first block only.) Source: Hurewitz et al.

To test these claims, Hurewitz, Brown-Schmidt, Thorpe, Gleitman, and Trueswell (2000) examined 5-year-olds’ production and comprehension abilities in two-referent scenes. Children heard a story, acted out by the experimenter, that introduced salient differences between the two frogs by having them doing different things. Afterwards, they were tested by asking them a Specific question: Which frog visited Mrs. Squid’s house? Answering this question required (a) understanding the story, ( b) understanding that the question requires an answer that distinguishes the frogs via locative modification, as these frogs were otherwise identical, and (c) producing in the answer a Â�restrictive modifier; namely The frog/one on the napkin. Immediately after Â�answering the question, the same child was asked to perform the put-task of Trueswell etÂ€al. (1999): Very good. Touch the Smiley Face. Now put the frog on the napkin into the box. As a control, another group of children were asked a General question (Can you tell me something about the story?) before doing the put task. Children’s production performance on the Specific question showed that they were able to perform all the relevant non-linguistic and linguistic acroÂ� batics necessary to specify a unique referent through restrictive modification:Â€72 percent of all answers to the Specific question were correct, providing answers like The frog on the napkin. In striking contrast, these same children’s

90

Trueswell et al.

responses to the put instruction showed the same mis-analysis effects as those reported in Trueswell et al. 1999. They performed incorrect actions, which involved the incorrect destination, in over 70 percent of the trials. And children were at chance in selecting between the two frogs. That is, the very same child who had just correctly responded to the story question by producing a PPmodified NP (the frog on the napkin) might now, in response to Put the frog on the napkin into the box, pick up the other frog, move it over to the empty napkin, and put it into the box. The sheer differences in complexity between the two sentences cannot account for the findings, as we know from earlier experimentation (the same children have no difficulty with unambiguous control sentences of equal complexity, e.g., Put the frog that’s on the napkin into the box). A further experiment in this line (Hurewitz et al. 2000, experiment 2) investigated the possibility that children just weren’t inclined to notice napkins as salient components of scene description. Making the platforms on which frogs were ensconced more salient (frilly umbrellas and royal thrones) generally increased performance in production (87 percent restrictive modifiers in production), but still the striking asymmetry between production and comprehension was preserved (60 percent errors in comprehension). In addition, in this version of the experiment we eye-tracked the young subjects, and the on-line results replicated those of Trueswell et al. (1999). So, in both of these experiments, we observe, as in the Rabbit-Walrus study, children understanding how the discourse can specify the need for an NP restrictive modifier. In particular, in the case of the Rabbit-Walrus study, we see this discourse-syntax knowledge at work in comprehension: a Contrastive question generates an increased chance of interpreting an ambiguous PP as an NP modifier, though this knowledge must battle against lexical evidence that may support an alternative interpretation (e.g., Eventive verbs generated some Question / Discourse-inappropriate responses in children). The experiments in the present section demonstrate this discourse-syntactic knowledge in children’s own productions: Contrastive questions generated a need for referential specificity in the form of a modifier (The frog/one on the napkin), which the children often uttered in response to this question type. However, when we then pull out a put sentence from our lexical arsenal, we see that we can tip the scales back to VP attachment, even in the same child who had just a moment ago demonstrated knowledge of the discourse-syntax facts in his or her own productions. It should be noted, though, that the discourse conditions are indeed slightly different between our production and comprehension test conditions. The distinction between the frogs had just been made by the child in his or her utter-


91

ance, and thus the discourse goal of contrasting the frogs had been achieved by the time we tested for comprehension abilities in our put instruction. We strongly suspect, however, that put was exerting detrimental effects, since unpublished work from our lab has examined put sentences as part of an answer to a contrastive question. (Rabbit: Which frog should I move? Walrus: I know, put the frog on the napkin into the box.) Here we still find strong VP-Â�attachment preferences despite the immediately preceding contrastive question (Hurewitz, 2001). Thus, the data strongly support the automatic use of verb preferences in the young child parsing system. Pragmatic vs. Semantic Accounts

Our account of the child parsing data depends heavily on the pragmatics litÂ� erature on deictic reference and on the conclusions drawn by Maratsos and Karmiloff-Smith regarding the development of definite reference. Specifically, faced with a definite NP, children have difficulty calculating the relevant referential domain and frequently do so egocentrically, thinking that what they take to be the referent of a definite NP must also be what their interlocutor believes to be the referent. Wexler (2003 and this volume) has offered a reinterpretation of the data of Maratsos and Karmiloff-Smith and has gone on to apply his explanation to the kindergarten-path phenomena. For motivation of his position, Wexler points to a specific study by Maratsos (replicated later by Karmiloff-Smith in French). In this experiment, children heard a story that introduced either multiple potential referents of the same type (i.e., several boys and several girls) or singletons (i.e., one boy and one girl). The story ended with an intentionally vague assertion (e.g., Someone started giggling and laughing) and a question (Who was giggling and laughing?). It was found, especially in the Karmiloff-Smith version of the study, that younger children (ages 3–7) overused definite NPs in answering the question. That is, even in a story containing multiple, undifferentiated boys and girls, these children tended to answer “The boy” or “The girl” rather than the expected “I don’t know.” On the surface, these data seem consistent with an egocentric account: a younger child has a particular boy or a particular girl in mind, and, without realizing that the addressee might not have a particular boy or a particular girl in mind, utters “the boy” or “the girl.” Wexler questions this account because the stories used in these particular studies did not introduce individual boys and girls but simply established an undifferentiated set of boys and an undifferentiated set of girls. This makes it unlikely that the children could have focused on a particular boy or girl in their mental model.

92

Trueswell et al.

Wexler proposes an alternative explanation of these findings, according to which the child’s deficit is semantic rather than pragmatic. He suggests that children in this age range (3–7 years) lack a full understanding of the meaning of the, and hence systematically misuse (and misinterpret) it. In particular, it is proposed that children lack the notion of maximality, which is a crucial part of the linguistic meaning of the definite determiner. Maximality as it applies to the semantics of the definite determiner can be expressed logically as follows: Given a predicate P that takes an argument X, P must be true for each and every member of the set of type X that is in the current referential domain. For instance, in order for the sentence The boy laughed to be true, each and every boy in the current referential domain must have laughed. In case there is more than one boy in this set, the definite NP must be pluralized (The boys laughed ) and all the boys in the set must have laughed. Wexler claims that the child version of the lacks this notion of maximality: for children, the only requires that the predicate applies to some element in the referential domain. Thus, in the Maratsos and Karmiloff-Smith stories, children say that the character who laughed was “the boy,” even though only one of several boys in the referential domain laughed. In the case of Put the frog on the napkin into the box, the number of frogs doesn’t matter to 3–7-year-olds, and the NP the frog on its own can refer to any member of the set of frogs. Wexler’s proposal has some merit because it draws explicit connections between developmental data and semantic theories of definiteness and aims at providing a detailed account of referential development. However, we believe that the theory is unable to capture key intuitions about referential errors and does not handle the range of developmental data already collected about definite reference and resolution of syntactic ambiguity.9 First, Wexler’s semantic account appears to miss intuitions about referential errors of the sort we are interested in. On his account, in a referential domain of several boys (e.g., one crying, one laughing, one sleeping) the sentence The boy is crying is false, rather than simply infelicitous. That is, if there are multiple boys in the domain of reference (and the other boys are not crying), the predicate is false because it is not true of all members of the set boy. However, the present authors all have the strong intuition that the utterance in this situation is true but infelicitous; that is, it should have been said in a different way. Even though adult intuitions about truth vs. felicity are not always robust, it is an advantage for an independently motivated theory to be able to capture such adult judgments. Second, even though Wexler’s account moves away from pragmatic issues in the calculation of the referential domain for definite descriptions, such issues seem inescapable for anyone trying to account for the interpretation of


93

definites in both children and adults. (For pointers to a large linguistic and philosophical literature, see Larson and Siegal 1995.) From a psycholinguistic viewpoint, we have already seen numerous cases in which the referential domain is calculated not only from the discourse context but also from the pragmatic implications of the semantic content of the utterance itself. For instance, consider again the scene in figure 4.3. What is the maximal domain for Pull the rabbit out of the hat, and why is it different for Pet the rabbit ? In the latter case, we respond Which one? — why not in the former case? Furthermore, there is now good evidence that the referential domain of an utterance changes for an adult listener based on contextual demands as that utterance unfolds. In an online comprehension study with adults, Chambers, Tanenhaus, Eberhard, Filip, and Carlson (2002) showed that adult listeners found the sentence Put the duck inside the can to be felicitous in the presence of more than one can so long as one and only one can was large enough to hold the duck — so felicitous in fact, that they applied this pragmatic restriction in real time, moving their eyes to the appropriate sized can upon hearing .â•–.â•–. inside the.â•–.â•–.â•–. Chambers et al. demonstrated that this effect changes with the goals of the speaker. In the same scene, the sentence Are you able to put the duck inside the can? resulted in looks to all the cans, often accompanied by a vocal response of Which one? In order to capture this flexibility in interpreting referential expressions, any account of reference must include what is plausible given the goals of the interlocutors as a matter of course. Such an account should be able to explain how mismatches of referential domain may give rise to puzzlement or misunderstandings between interlocutors, even when no semantic problems with definite determiners exist (as in the adult case). A reasonable (and parsimoÂ� nious) interpretation of children’s errors with definite descriptions, then, would be to attribute such errors mostly to pragmatic-referential factors. After all, even the youngest children do not consistently fail in referential studies of production and comprehension; sometimes they have calculated the correct referential domain. Third, the full range of developmental data on definite reference do not seem to be explained by the semantic-deficit account, even for the specific study Wexler used to motivate his account. The stories in Maratsos 1976 enumerate the members of the set, which may encourage the establishment of individuals of these sets of boys and girls. This makes it more likely that children had a particular entity in mind when using definite descriptions (e.g. the boy) in their responses. There are also important puzzles in the data: only a subset of the 4-year-olds behaved in the way predicted by Wexler (those who performed poorly on a separate sentence repetition task). This correlation with memory abilities was found only in this age group, and younger children performed

94

Trueswell et al.

better, correctly uttering a boy in the context of several boys 83 percent of the time. Karmiloff-Smith (1979), in a very similar study in French, finds a different developmental pattern, more in line with Wexler’s account: when asked to indicate a particular boy or girl in a response to a question about a story containing several boys and several girls, young children (ages 3–7) answered le garcon /la fille (the boy/the girl) whereas older children (ages 8–10) responded predominantly with un garcon /une fille (a boy/a girl). But interestingly, two out of four adults we asked reacted to translations of these materials by answering with a specific referent (“I don’t know, Mary?”). These subjects volunteered a name despite the fact that there was no correct answer, presumably because the instructions (Guess who it was) required a single, specific referent. Indeed, Karmiloff-Smith noted that children and adults sometimes responded this way (e.g., giving their own name or the name of one of their school mates such as Juliette). This shows that even adults may make quite specific guesses about the identity of referents they cannot truly individuate. Fourth, under Wexler’s view, we would expect children to make more errors than they do with definite descriptions. In recent work that looked at children’s comprehension of implicatures, Papafragou and Tantalou (2004) presented children with stories in which a character was supposed to perform a certain action to win a prize. In one case, an elephant was given three oranges and was told that he had to eat the oranges. When the elephant came back and was asked if he had eaten the oranges, he answered I ate some. A vast majority of 5-year-olds decided the elephant should not get a prize and justified their response by indicating, e.g., that the elephant had not eaten all the oranges. This reveals that children’s preferred interpretation for the oranges obeys maximality, otherwise the contrast with some would be lost. The full range of data on syntactic ambiguity resolution is difficult to capture under a purely semantic account. In particular, Wexler suggests that a sÂ�emantic deficit accounts for the 5-year-olds’ failure to realize that a scene containing multiple frogs requires a modifier interpretation of an ambiguous phrase. However, if this is true, adults too seem to have this semantic deficit. Recall that we showed that adults frequently failed to realize that a modifier interpretation was needed in Tickle the frog with the feather in two-frog scenes (Snedeker and Trueswell 2004). Syntactic accessibility of the alternative meanings of with the feather also drives reference patterns, in both adults and children. Another way of stating this is that the semantic-deficit account must assume that very similar patterns in adults and children result from performance factors in one case and from competence factors in the other, and that the similarity between the adult and child data sets is accidental.


95

Instead, the picture emerging from this literature on reference and ambiguity resolution is quite different: success on reference assignment requires, among other things, rapid tracking of shifting and flexible referential domains. Children by age 4 or 5 have trouble aligning their referential domain with that of their interlocutor even though they understand the semantics of definiteness, including the maximality assumption. Indeed, such mismatches of referential domains can occur in adults (we return to this issue, and to individual differences in the adult population, in the next section). We should be clear here; we are not claiming that maximality is innate and pragmatic factors mask this fact. The pragmatic facts themselves suggest that learning how definite reference behaves requires some work. We do claim, however, that by age 5 children understand the semantics of definiteness, including the maximality aÂ�ssumption. Success and failure in child reference appears to be driven by successful match or mismatch in calculations of referential domain. Conflict Resolution and Garden-Path Lingering

Here we consider briefly one of the most striking findings reported in Trueswell et al. 1999 which we have not discussed so far. In the original put study and the follow-up studies, children appear to behave impulsively, not revising their initial referential and syntactic commitments. More concretely, after hearing the sentence Put the frog on the napkin into the box, children often act upon the frog that they looked to first upon hearing the frog. Moreover, the tendency to initially interpret on the napkin as a goal is often not rescinded upon hearing the PP into the box. Five-year-olds tend to carry out actions in which a frog goes to the empty napkin. Adults rarely do this, and neither do 5-year-olds when the ambiguity is removed (Put the frog that’s on the napkin .â•–.â•–.). Other labs have replicated this developmental change in the ability to revise parses (Weighall 2008), and the same pattern has been observed in a language that has a substantially different grammar, Korean, which is a verb-final language (Choi and Trueswell 2010). We believe that this developmental change in revision abilities is related to developmental changes in executive-function, specifically the ability to select a subordinate analysis under conditions of representational conflict (Trueswell and Gleitman 2004; Novick, Trueswell, and Thompson-Schill 2005). In particular, Novick et al. (2005) propose that these patterns represent changes in executive-function abilities generally over the course of development, especially those associated with the selection of representations under conditions of conflict (e.g., Diamond and Taylor 1996; Zelazo and Reznick 1991). Frontallobe regions (e.g., Left Prefrontal Cortex) appear to be implicated in this ability

96

Trueswell et al.

(e.g., Thompson-Schill, Jonides, Marshuetz, Smith, D’Esposito, Kan, Knight, and Swick 2002). And indeed, these regions are known to be late developing anatomically, well into years 5 and 6 (e.g., Huttenlocher and Dabholkar 1997). It seems quite plausible then to consider that the re-ranking of interpretations in garden-path phenomena (inhibiting an initial interpretation, and promoting a new interpretation) would involve these very systems. And indeed, perÂ� severation in response is the hallmark of patients with severe frontal-lobe dÂ�amage — and normally developing children. For instance, in the Wisconsin Card Sorting task, participants first sort cards by one criterion (e.g., color) but then are asked to switch to sorting by another criterion (e.g., shape). Children and frontal-lobe patients have great difficulty switching between criteria, presumably because one sort of evidence was developed (color) as relevant and then must be overridden by some other evidence (shape). To address this account of garden-path recovery, we begin by noting that children are not unique in this failure to rescind garden-path interpretations. Christianson, Hollingworth, Halliwell, and Ferreira (2001) have shown this in comprehension questions about garden-path sentences. The questions were specifically designed so that the correct answer would be No for the intended meaning of the sentence as a whole, but would be answered Yes if based (eÂ�rroneously) on the temporarily considered but rejected interpretation. These questions took significantly longer to answer and showed more errors than the same questions about unambiguous versions of the sentences. That is, the garden path appears to linger, as if subjects sometimes failed to completely inhibit or reject the intended interpretation (Christianson et al. 2001). One might expect individual differences in this ability to be related to executive function generally. Indeed, Mendelsohn (2002, 2003) has offered evidence in favor of this claim. In her study, she found that the size of subjects’ lingering garden-path effect correlated with several linguistic and nonlinguistic measures that arguably involve inhibition /selection mechanisms. Perhaps the most compelling observation was that a completely non-linguistic task, the so-called anti-saccade task, correlated with lingering garden-path measures. In this task, subjects were to look in a direction opposite of a flash of light (thus inhibiting the reflex to look to the light); difficulty on this task correlated with the ability to reject garden-path interpretations of sentences. Our own research group, in collaboration with Sharon Thompson-Schill, has been exploring similar issues ( Novick, Kahn, Trueswell, and Thompson-Schill 2009; January, Trueswell, and Thompson-Schill 2009). For instance, Novick etÂ€al. (2009) report on an adult patient with focal damage to the Left Inferior Frontal Gyrus (LIFG), who not only has trouble with executive function tasks


97

but also fails to recover from parsing commitments in the put-task of Trueswell et al. (1999). That is, in response to Put the frog on the napkin into the box, he will move the frog over to the incorrect goal (the empty napkin) and then into the box, just like five-year-old children. Likewise, he is perfectly normal in his actions for unambiguous trials, suggesting his deficit is associated with inhibiting the initial analysis of an ambiguous phrase. Indeed, he also has similar trouble with other ambiguities, such as lexical ambiguity. January et al. (2009) have found converging evidence in the form of fMRI data in normal adults (see also Ye and Zhou 2009, for similar findings) and other labs have recently rÂ�eported correlations between individual differences in executive-function abilities and the ability to resolve ambiguities, for both children and adults (Brown-Schmidt 2009; Khanna and Boland 2010; Nilson and Graham 2009). (See Mazuka, Jincho, and Oishi 2009 for a review of the relationship between executive function and child language; see Novick, Trueswell, and ThompsonSchill, in press, for the literature on adults.) We believe executive function issues are deeply related to some interesting experimental findings that point to differences in deictic and anaphoric reference in young children. We mentioned above that in some studies by Maratsos (1976) young children behaved much better in their use of definite and indefinite NPs. In these studies, Maratsos again compared referential situations in which either multiple entities were present (several boy dolls and several girl dolls) or singletons (one boy and one girl). All the dolls were placed at the top of a slide, and children played a game with the experimenter in which the experimenter could send a doll down the slide in a toy car. The child had to decide which doll to send down, and tell the experimenter. Two conditions were compared: one in which the dolls could be seen by the child and the experimenter, and one in which the dolls were shown but then placed out of view of the child. When the dolls were in view, many younger children made definite NP errors of the familiar sort (e.g., Okay, send down the boy! in the presence of multiple boy dolls). Maratsos reported that when children used a definite NP they clearly had a particular doll in mind and were looking directly at it. They also made explicit comments on the experimenter’s choice of referents, such as That’s the one! or That one’s all right, I guess. However, when the dolls were out of sight the use of indefinites increased significantly (Okay, send down a boy!). Maratsos took this as a sign that children were acting egocentrically, though it was not made entirely clear why hiding objects would discourage egocentric behavior. We suggest there may in fact be an underlying connection between visual attention and discourse focus that better explains the Maratsos out-of-sight

98

Trueswell et al.

phenomena. In particular, if one takes seriously the informal analogy that the current referential domain in a discourse model is like an attentional mechanism, these patterns are expected. In the case of focusing attention on entities in a discourse model, attention is not spatially restricted: we can think of two boys even if they are not near each other spatially. However, looking at a particular boy in the world necessarily focuses our attention on that boy and not other boys that are spatially distant (the fovea subtends 2–3 degrees visual angle, and material that is foveated is typically what is being attended to). Deictic reference reflects an interface between a mental model of the world and these co-present objects, whose perception and visual recognition point to these same entities in the mental model. Thus, these findings could be recast as arising from differences in executive function: visual attention and discourse attention can conflict; younger children may not easily deal with this sort of conflict. Visual inspection of an object should be expected to activate this entity in the discourse model, but in many cases the current referential domain is larger or different than this attentional space, and hence such effects on the model must be inhibited. Maratsos noted that even some adults behaved in a childlike fashion when the dolls were visually co-present, saying “send down the boy” in the presence of multiple boys. This is to be expected since this sort of disambiguation by eye gaze or other means can occur. However, we would predict that some of this behavior could very well correlate with individual differences in executive function, as described above for ambiguity resolution. Interestingly, Meroni and Crain (this volume) report a related out-of-sight phenomenon in a kindergarten-path put study. They report that children in twofrog scenes do quite well (around 90 percent correct actions) when they are first asked to close their eyes before hearing the utterance. The authors suggest a processing explanation at the motor-planning level: closing one’s eyes gets one out of an “interpret-mode” of language comprehension and children may have trouble inhibiting motor plans made incrementally during speech perception. We are skeptical of these results and their interpretation, especially since Meroni and Crain did not compare their results to children who had their eyes open during utterance interpretation. Indeed, Weighall (2008), in a study of 60Â€children, reports replicating all the findings of Trueswell et al. and none of the findings of Meroni and Crain; instead, children showed signs of gardenpathing whether or not they were initially looking at the scene. We suspect that the high number of successful interpretations in the study by Meroni and Crain reflect specific properties of their study; the discourse context and the goals for the children were different in the Meroni and Crain study, and, we believe, further supported the correct interpretation.


99

Closing Remarks

On the basis of the work reviewed here, it should be clear that young children who are trying to comprehend language are faced with a processing problem of considerable complexity. From a sequence of words, a child must rapidly glean detailed grammatical information in order to determine not only who is doing what to whom but also how an utterance relates to their current conception of the world. There exist multiple probabilistic sources of evidence for constraining the grammatical structure of an utterance, and the child must discover, weigh, and combine this evidence. We have suggested elsewhere (e.g., in Trueswell and Gleitman 2004, 2007) that the child tracks and builds detailed syntactic and semantic representations of words that allow for the efficient recovery of the structure of the sentence as a whole. However, this gets a child or adult listener only so far. The referential implications of these analyses must also be considered, since some analyses are very unlikely given the referential setting. Here the developmental theorist and the child face the same problem: they must grapple with the fact that the interlocutors’ goals and shared conception of the world guide what is considered to be the relevant referential domain. Within the realm of deictic reference, the scene itself only partially constrains the choice of referential expressions (i.e., the level of specificity or linguistic economy offered in an utterance). What matters more is which aspects of this world are under discussion and are relevant to the intended goals of the utterance. We discussed numerous instances of adult conversation and adult comprehension in which this was shown to be the case. We then turned our attention to the implications these observations might have for a child who is learning how deictic reference works, and how reference might impact relevant syntactic choices during comprehension. The story we have offered is long, and admittedly complicated, but how could it be any other way? After all, reference is amazingly sensitive to a vast array of considerations, from the linguistic to the nonlinguistic, and the child must learn these facts presumably bit by bit. Indeed, we believe the complexity of the referential process explains in part why “bottom-up” sources of structure (i.e., the lexical evidence) so strongly constrain syntactic ambiguity resolution at earlier stages of development. Even after a child understands how reference works within a referential domain, the computation of that domain, which changes rapidly, is expected to be frequently misaligned with the domain currently entertained by his or her interlocutor. Thus, fairly systematic errors can be observed in definite reference both in production and comprehension, especially with regard to the syntactic concomitants of reference.

100

Trueswell et al.

Nevertheless, we have documented several instances in which the child sÂ�uccessfully computes the referential domain, and understands the level of specificity needed for reference within this domain. Most of these successes, we would argue, arise from circumstances in which there is a clear and reliableÂ€indicator from the conversation about the shape of the current referential domain and the goal of the utterance (e.g., questions beginning with “Which .â•–.â•–.” need to contrast members of a set). Under a view in which comprehension is a “guessing game” in which listeners are continuously guessing the intentions of the speaker, we might expect such a pattern: highly constraining, easyto-discover sources of evidence trump, in developmental time, less reliable andÂ€more complex sources. We strongly suspect that other demonstrations of non-egocentric referential behaviors in young children also arise out of simplerÂ€mechanisms that unambiguously predict the attentional state of the interlocutor (e.g., eye gaze and physical constraints on perception; see Baldwin 1991 and Nadig and Sedivy 2002). Indeed, there is now growing evidence thatÂ€ “theory of mind” abilities emerge out of converging cues of this sort, andÂ€ that simpler use of some of these cues arises in part in other species andÂ€grows during development in children (Leslie 2000; Call and Tomasello 2005). Finally, we have considered attentional development in children, and have suggested that during the presence of physical objects under discussion children are overly sensitive to their own visual /perceptual attentional state when it comes to calculating what to attend to in their discourse model. We have concluded that children between the ages of 4 and 6 years understand that a definite NP ( produced or heard) must apply maximally to the current referential domain, but this domain is skewed by their attentional state. This explains a range of phenomena in their own use of definite and indefinite reference, but it also explains their difficulty applying referential facts to parsing procedures. Moreover, developmental changes in a listener’s ability to rescind syntactic commitments may reflect this same processing change, in which competing structure(s) must be rapidly inhibited and the correct structure must be activated. This points to an explanation of individual differences in adults regarding errors in production and comprehension, which are skewed in the direction of the child-like behaviors. It is clear that we are only at the early stages of understanding how the processing demands of language production and language comprehension iÂ�mpinge on the language-learning process. However, methods like those discussed here, which examine how children interpret speech in real time, offer new insight into these developing processing abilities. The data suggest that there is hope for building a more unified theory of language acquisition and language use


101

over a lifetime — a theory in which we recognize that multiple sources of evidence must be discovered (and sometimes built) by the child, in a way that aÂ�llows for immediate integration of this evidence into comprehension and production mechanisms. Acknowledgments

Some of the research reported herein was supported by grants from the National Institutes of Health (R01-HD37507 to John Trueswell and Lila Gleitman and 5F32-MH6502 to Anna Papafragou). We would like to thank Peter Gordon of Columbia University’s Teachers College and Colin Phillips of University of Maryland at College Park for insightful comments, specifically about the development of definite reference and Wexler’s semantic account of these phenomena. We also thank David January, Samantha Crane, and Alon Hafri for their assistance in the preparation of this paper. Finally, we thank Lila Â�Gleitman for her support and helpful comments on earlier drafts. Notes 1.â•‡ The explanation of lexically specific effects on parsing is more complicated than this. For instance, one could alternatively say that listeners (and language learners) use known semantic properties of the verb slice to compute the likelihood of an instrument phrase, rather than saying that listeners and learners track the syntactic contingencies given the word itself. It is very difficult to disentangle these two explanations since semantic and syntactic properties of verbs are so intimately related (e.g., Fisher, Gleitman, and Gleitman 1992). Also, one might argue that an individual’s compiled ‘corpus’ of their native tongue is too small to offer lexically based biases. However, even a conservative estimate of language exposure (two words a second for 5 hours a day) yields 13.1 million words per year (see Kelly and Martin 1994 for a similar estimate). This estimate is clearly conservative when one considers that television exposure alone isÂ€anÂ€average 3.6 hours/day by age 3 for American children (Christakis, Zimmerman, DiGiuseppe, and McCarty 2004). Nevertheless, it is an obvious necessity for any theory of language comprehension to include semantic-contingent predictors of structures given the way that language works creatively. Another way of stating this is that the learned semantic classes and properties of known words are also used to constrain parsing and are especially important when a word is uncommon. That is, the parser must “back off ” to larger categories when subcategories have a small N and even infer a likely category when the word is novel to the language user (i.e., Nâ•–=â•–1). For further discussion and experimentation, see Juliano and Tanenhaus 1994; Naigles, Gleitman, and Gleitman 1993; Naigles, Fowler, and Helm 1992, 1995; Trueswell, Tanenhaus, and Kello 1993; Trueswell and Gleitman 2004. 2.â•‡ Research suggests that adults, when engaged in conversations, coordinate and track the referential domain of utterances, especially when this domain is determined by such

102

Trueswell et al.

factors as common knowledge/ground. Debates exist however about the timing of these effects (e.g., Hanna et al. 2003; Keysar et al. 2000; Keysar et al. 1998). 3.â•‡ This of course is a miracle in our theory as laid out here. See Gleitman et al. (2005) for a theory of learning “hard words,” among which we include on, in, under, of, and within. See also Nappa, Wessel, McEldoon, Gleitman, and Trueswell (2009) for how parsing and visual attention may combine to aid in the learning of such words. 4.â•‡ We realize that the decision to study 5-year-olds skips over much of language learning. And indeed the initial choice of this age range was driven largely by mÂ�ethodological limitations (younger children were less willing to wear the eye-tracking visor). Nevertheless, several findings suggest that important developmental changes in parsing and interpretation extend well through this age range (e.g., Chomsky 1969) — findings which have not been adequately studied or explained. 5.â•‡ Trueswell et al. (1999) used in the box rather than into the box. The use of the preposition in introduces another ambiguity, which we wish to avoid. All follow-up studies (e.g., Hurewitz et al. 2000) have used into and have replicated the results reported here. For simplicity we adopt in our examples the use of into. 6.â•‡ Put is a very common verb in child-directed speech and is frequently accompanied by a Goal PP (see Trueswell et al. 1999 for corpus evidence). 7.â•‡ Verb type was manipulated between subjects such that one group of subjects got Instrument-biased stimuli, another group got Equi-biased stimuli, and a third got MÂ�odifier-biased stimuli (embedded in numerous filler trials). For each verb-type, half the trials were two-referent scenes and half were one-referent scenes (with item-Â� condition pairings counterbalanced across subjects). 8.â•‡ It appears that the earlier put studies may have had additional factors that reduced garden-pathing generally in adults but not in children: specifically post-ambiguity information and prosodic information. Consistent with this, Choi and Mazuka (2003) provided evidence that prosodic information has little effect on children’s interpretationÂ€of syntactic ambiguity in Korean, and Snedeker et al. (2003) found significant but weak effects of prosody on PP-attachment ambiguity resolution in children. Also, postambiguity (i.e., disambiguating) information might be especially effective for adults in parsing, an issue we return to later in the paper. Put the X on the .â•–.â•–. and Tickle the X with the .â•–.â•–. also differ in the type of semantic properties that are competing with each other in the ambiguity. It is possible that the put ambiguity is easier to revise because both interpretations involve the location of an object (see Snedeker and Trueswell 2004). 9.â•‡ Also, in a recent test of Wexler’s maximality account, Munn, Miller, and Schmidt (2006) offer some experimental data that do not support the maximality account and instead support a more traditional analysis of child definite reference. References Allopenna, P. D., Magnuson, J. S., and Tanenhaus, M. K. 1998. Tracking the time course of spoken word recognition: Evidence for continuous mapping models. Journal of Memory and Language 38, 419– 439.


103

Altmann, G., and Kamide, Y. 1999. Incremental interpretation at verbs: Restricting the domain of subsequent reference. Cognition 73, 247–264. Altmann, G., and Steedman, M. 1988. Interaction with context during human sentence processing. Cognition 30, 191–238. Aslin, R. N., Saffran, J. R., and Newport, E. L. 1998. Computation of conditional probability statistics by 8-month old infants. Psychological Science 9, 321–324. Baldwin, D. A. 1991. Infant contribution to the achievement of joint reference. Child Development 62, 875–890. Britt, M. A. 1994. The interaction of referential ambiguity and argument structure in theÂ€ parsing of prepositional phrases. Journal of Memory and Language 33, 251– 283. Brown-Schmidt, S. 2009. The role of executive function in perspective-taking during on-line language comprehension. Psychonomic Bulletin & Review, 16, 893–900. Brown-Schmidt, S., Campana, E., and Tanenhaus, M. K. 2002. Reference resolution inÂ€ the wild: On-line circumscription of referential domains in a natural, interactive problem-solving task. In Proceedings of the 24th Annual Conference of the Cognitive Science Society. Erlbaum. Call, J., and Tomasello, M. 2005. Reasoning and thinking in nonhuman primates. In K. Holyoak and B. Morrison (eds.), Cambridge Handbook of Thinking and Reasoning. Cambridge University Press. Christianson, K., Hollingworth, A., Halliwell, J., and Ferreira, F. 2001. Thematic roles assigned along the garden path linger. Cognitive Psychology 42, 368– 407. Chambers, C. G., Tanenhaus, M. K., Eberhard, K. M., Filip, H., and Carlson, G. N. 2002. Circumscribing referential domains during real-time language comprehension. Journal of Memory and Language 47 (1), 30 – 49. Choi, Y., and Mazuka, R. 2003. Young children’s use of prosody in sentence parsing. Journal of Psycholinguistic Research 32 (2), 197–217. Choi, Y., and Trueswell, J. C. 2010. Children’s (in)ability to recover from garden-paths in a verb-final language: Evidence for developing control in sentence processing. Journal of Experimental Child Psychology 106 (1), 41– 61. Chomsky, C. 1969. Acquisition of Syntax in Children from 5 to 10. MIT Press. Christakis, D. A., Zimmerman, F. J., DiGiuseppe, D. L., and McCarty, C. A. 2004. Early television exposure and subsequent attentional problems in children. Pediatrics 113 (4), 708–713. Clark, H. H. 1993. Arenas of Language Use. University of Chicago Press. Crain, S. 1980. Pragmatic Constraints on Sentence Comprehension. Ph.D. dissertation, University of California, Irvine. Crain, S., and Steedman, M. 1985. On not being led up the garden path: The use of context by the psychological parser. In D. Dowty, L. Karrttunen, and A. Zwicky (eds.), Natural Language Parsing: Psychological, Computational, and Theoretical Perspectives. Cambridge University Press.

104

Trueswell et al.

Crain, S., and Thornton, R. 1998. Investigations in Universal Grammar: A Guide to Experiments on the Acquisition of Syntax and Semantics. MIT Press. Diamond, A., and Taylor, C. 1996. Development of an aspect of executive control: Development of the abilities to remember what I said and to “Do as I say, not as I do.” Developmental Psychobiology 29, 315–334. Fernald, A. 2004. The search for the object begins at the verb. Talk presented at 29th Annual Boston University Conference on Language Development. Fernald, A., Zangl, R., Portillo, A. L., and Marchman, V. A. 2008. Looking while lisÂ� tening: Using eye movements to monitor spoken language comprehension by infants and young children. In I. Sekerina, E. Fernandez, and H. Clahsen (eds.), DevelopmentalÂ€ Psycholinguistics: On-Line Methods in Children’s Language Processing. John BÂ�enjamins. Fisher, C., Hall, G., Rakowitz, S., and Gleitman, L. 1994. When it is better to receive than to give. Lingua 92, 333–375. Fisher, C., Gleitman, H., and Gleitman, L. 1992. On the semantic content of subcategorization frames. Cognitive Psychology 23 (3), 331–392. Frazier, L. 1987. Sentence processing: A tutorial review. In M. Coltheart (ed.), Attention and Performance XII: The Psychology of Reading. Erlbaum. Frazier, L., and Fodor, J. D. 1978. The sausage machine: A new two-stage parsing model. Cognition 6, 291–325. Garnsey, S. M., Pearlmutter, N. J., Myers, E., and Lotocky, M. A. 1997. The contributions of verb bias and plausibility to the comprehension of temporarily ambiguous sentences. Journal of Memory and Language 37, 58–93. Gerkin, L. A. 2002. Early sensitivity to linguistic form. Annual Review of Language Acquisition 2, 1–36. Gillette, J., Gleitman, L., Gleitman, H., and Lederer, A. 1999. Human simulations of vocabulary learning. Cognition 73, 153–190. Gleitman, L. 1990. The structural sources of verb learning. Language Acquisition 1, 3–35. Gleitman, L. R., Cassidy, K., Nappa, R., Papafragou, A., and Trueswell, J. C. 2005. Hard words. Language Learning and Development 1 (1), 23– 64. Gómez, R. L. 2002. Variability and detection of invariant structure. Psychological Science 13, 431– 436. Gómez, R. L., and Gerkin, L. A. 2000. Infant artificial language learning and language acquisition. Trends in Cognitive Sciences 4, 178–186. Hanna, J., Tanenhaus, M. K., and Trueswell, J. C. 2003. The effects of common ground and perspective on domains of referential interpretation. Journal of Memory and Language 49, 43– 61. Harris, Z. 1957. Co-occurrence and transformation in linguistic structure. Language 33, 283–340. Hurewitz, F. 2001. Developing the Ability to Resolve Syntactic Ambiguity. Ph.D. dissertation, University of Pennsylvania.


105

Hurewitz, F., Brown-Schmidt, S., Thorpe, K., Gleitman, L., and Trueswell, J. 2000. One frog, two frog, red frog, blue frog: Factors affecting children’s syntactic choices in production and comprehension. Journal of Psycholinguistic Research 29, 597– 626. Huttenlocher, P. R., and Dabholkar, A. S. 1997. Regional differences in synaptogenesis in human cerebral cortex. Journal of Comparative Neurology 387, 167–178. Jackendoff, R. 2002. Foundations of Language. Oxford University Press. January, D., Trueswell, J. C. and Thompson-Schill, S. L. 2009. Co-localization of Stroop and syntactic ambiguity resolution in Broca’s area: Implications for the neural basis of sentence processing. Journal of Cognitive Neuroscience 21 (12), 2434–2444. Juliano, C., and Tanenhaus, M. 1994. A constraint-based lexicalist account of the sÂ�ubject /object attachment preference. Journal of Psycholinguistic Research 23 (6), 459– 471. Karmiloff-Smith, A. 1979. A Functional Approach to Child Language: A Study of Determiners and Reference. Cambridge University Press. Kelly, M. H., and Martin, S. 1994. Domain-general abilities applied to domain-specific tasks: Sensitivity to probabilities in perception, cognition, and language. Lingua 92, 105–140. Keysar, B., Barr, D. J., Balin, J. A., and Brauner, J. S. 2000. Taking perspective in conversation: The role of mutual knowledge in comprehension. Psychological Science 11, 32–37. Keysar, B., Barr, D. J., and Horton, W. S. 1998. The egocentric basis of language use: Insights from a processing approach. Current Directions in Psychological Science 7, 46 –50. Khanna, M. M., and Boland, J. E. 2010. Children’s use of language context in lexÂ� icalÂ€ambiguity resolution. Quarterly Journal of Experimental Psychology 63 (1), 160 – 193. Kim, A., Srinivas, B., and Trueswell, J. C. 2002. The convergence of lexicalist perspectives in psycholinguistics and computational linguistics. In P. Merlo and S. Stevenson (eds.), Sentence Processing and the Lexicon: Formal, Computational and Experimental Perspectives. John Benjamins. Larson, R., and Siegal, G. 1995. Knowledge of Meaning. MIT Press. Leslie, A. M. 2000. Theory of mind as a mechanism of selective attention. In M. Gazzaniga (ed.), The New Cognitive Neuroscience, second edition. MIT Press. Lyons, C. G. 1980. The meaning of the English definite article. In J. Van der Auwera (ed.), The Semantics of Determiners. Croom Helm. Lyons, C. G. 1999. Definiteness. Cambridge University Press. MacDonald, M. C., Pearlmutter, N. J., and Seidenberg, M. S. 1994. The lexical nature of syntactic ambiguity resolution. Psychological Review 101, 676 –703. Marcus, M. P., Santorini, B., and Marcinkiewicz, M. A. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics 19, 313–330. Maratsos, M. P. 1976. The Use of Definite and Indefinite Reference in Young Children: An Experimental Study of Semantic Acquisition. Cambridge University Press.

106

Trueswell et al.

Mazuka, R., Jincho, N., and Oishi, H. 2009. Development of executive control and language processing. Language and Linguistics Compass 3 (1), 59–89. Mendelsohn, A. 2002. Individual Differences in Ambiguity Resolution: Working Memory and Inhibition. Ph.D. dissertation, Northeastern University. Mendelsohn, A. 2003. The role of inhibition in sentence processing as demonstrated by a new task. Paper presented at CUNY Conference on Human Sentence Processing, Cambridge. Mintz, T. H., Newport, E. L., and Bever, T. G. 2002. The distributional structure of grammatical categories in speech to young children. Cognitive Science 26, 393– 424. Munn, A., Miller, K., and Schmitt, C. 2006. Maximality and Plurality in Children’s Interpretations of Definites. In D. Bamman, T. Magnitskaia, and C. Zaller (eds.), Boston University Conference on Language Development (BUCLD) Proceedings 30. pp. 377– 387. Nadig, A. S., and Sedivy, J. C. 2002. Evidence of perspective-taking constraints in children’s on-line reference resolution. Psychological Science 13 (4), 329–336. Naigles, L. G., Gleitman, L. R., and Gleitman, H. 1993. Children acquire word meaning components from syntactic evidence. In E. Dromi (ed.), Language and Cognition: A Developmental Perspective. Ablex. Naigles, L. G., Fowler, A., and Helm, A. 1992. Developmental shifts in the construction of verb meanings. Cognitive Development 7 (4), 403– 427. Naigles, L. G., Fowler, A., and Helm, A. 1995. Syntactic bootstrapping from start to finish with special reference to Down syndrome. In M. Tomasello and W. Merriman (eds.), Beyond Names for Things: Young Children’s Acquisition of Verbs. Erlbaum. Nappa, R., Wessell, A., McEldoon, K. L., Gleitman, L. R., and Trueswell, J. C. 2009. Use of speaker’s gaze and syntax in verb learning. Language Learning and Development 5 (4), 203–234. Nation, K., Marshall, C., and Altmann, G. T. M. 2003. Investigating individual differences in children’s real-time sentence comprehension using language-mediated eye movements. Journal of Experimental Child Psychology 86, 314–329. Nilsen, E. S., and Graham, S. A. 2009. The relations between children’s communicative perspective-taking and executive functioning. Cognitive Psychology 58, 220 –249. Novick, J. M., Kahn, I., Trueswell J. C., and Thompson-Schill, S. 2009. A case for conflict across multiple domains: Memory and language impairments following damage to ventrolateral prefrontal cortex. Cognitive Neuropsychology 26 (6), 527–567. Novick, J. M., Kim, A., and Trueswell, J. C. 2003. Studying the grammatical aspects of word recognition: Lexical priming, parsing and syntactic ambiguity resolution. Journal of Psycholinguistic Research 32 (1), 57–75. Novick, J. M., Trueswell, J. C., and Thompson-Schill, S. 2005. Toward the neural basis of parsing: Prefontal cortex and the role of selectional processes in language comÂ� prehension. Journal of Cognitive, Affective, and Behavioral Neuroscience 5 (3), 263– 281. Novick, J. M., Thompson-Schill, S., and Trueswell, J. C. 2008. Putting lexical constraints in context into the visual-world paradigm. Cognition 107 (3), 850 –903.


107

Novick, J. M., Trueswell, J. C., and Thompson-Schill, S. In press. Broca’s area and language processing: Evidence for the cognitive control connection. Language and Linguistic Compass. Papafragou, A., and Tantalou, N. 2004. Children’s computation of implicatures. Language Acquisition 12, 71–82. Saffran, J. R. 2001. Words in a sea of sounds: The output of statistical learning. Cognition 81, 149–169. Saffran, J. R. 2002. Constraints on statistical language learning. Journal of Memory and Language 47, 172–196. Saffran, J. R., Aslin, R. N., and Newport, E. L. 1996. Statistical learning by 8-month old infants. Science 274, 1926 –1928. Snedeker, J., Thorpe, K., and Trueswell, J. C. 2002. On choosing the parse with the scene: The role of visual context and verb bias in ambiguity resolution. In Proceedings of the 22nd Annual Conference of the Cognitive Science Society, Edinburgh. Snedeker, J., and Trueswell, J. C. 2004. The developing constraints on parsing decisions: The role of lexical-biases and referential scenes in child and adult sentence processing. Cognitive Psychology 49 (3), 238–299. Spivey-Knowlton, M., and Sedivy, J. 1995. Resolving attachment ambiguities with multiple constraints. Cognition 55, 227–267. Spivey, M. J., and Tanenhaus, M. K. 1998. Syntactic ambiguity resolution in discourse: Modeling the effects of referential context and lexical frequency. Journal of Experimental Psychology: Learning, Memory and Cognition 24 (6), 1521–1543. Spivey, M. J., Tanenhaus, M. K., Eberhard, K. M., and Sedivy, J. C. 2002. Eye movements and spoken language comprehension: Effects of visual context on syntactic ambiguity resolution. Cognitive Psychology 45, 447– 481. Srinivas, B., and Joshi, A. K. 1999. Supertagging: An approach to almost parsing. Computational Linguistics 252 (2), 237–265. Steedman, M. 2000. The Syntactic Process. MIT Press. Stone, M., and Webber, B. 1998. Textual economy through close coupling of syntax and semantics. In Proceedings of INLG. Swingley, D., Pinto, J. P., and Fernald, A. 1999. Continuous processing in word recognition at 24 months. Cognition 71, 73–108. Taraban, R., and McClelland, J. 1988. Constituent attachment and thematic role assignment in sentence processing: Influences of content-based expectations. Journal of Memory and Language 27, 1–36. Tanenhaus, M. K., Spivey-Knowlton, M. J., Eberhard, K. M., and Sedivy, J. C. 1995. Integration of visual and linguistic information in spoken language comprehension. Science 268, 1632–1634. Tanenhaus, M. K., Hanna, J., and Chambers, C. 2004. Referential domains in spokenÂ€language comprehension: Using eye movements to bridge the product and action traditions. In J. M. Henderson and F. Ferreira (eds.), The Interface of Language, Vision, and Action: Eye Movements and the Visual World. Psychology Press.

108

Trueswell et al.

Thompson-Schill, S. L., Jonides, J., Marshuetz, C., Smith, E. E., D’Esposito, M., Kan, I. P., Knight, R. T., and Swick, D. 2002. Effects of frontal lobe damage on interference effects in working memory. Cognitive, Affective and Behavioral Neuroscience 2, 109– 120. Trueswell, J. C. 1996. The role of lexical frequency in syntactic ambiguity resolution. Journal of Memory and Language 35, 566 –585. Trueswell, J. C., and Gleitman, L. 2004. Children’s eye movements during listening: Developmental evidence for a constraint-based theory of sentence processing. In J. M. Henderson and F. Ferreira (eds.), The Interface of Language, Vision, and Action: Eye Movements and the Visual World. Psychology Press. Trueswell, J. C., and Gleitman, L. R. 2007. Learning to parse and its implications for language acquisition. In G. Gaskell (ed.), Oxford Handbook of Psycholinguistics. Oxford: Oxford Univ. Press. Trueswell, J. C., and Kim, A. E. 1998. How to prune a garden-path by nipping it in the bud: Fast-priming of verb argument structures. Journal of Memory and Language 39, 102–123. Trueswell, J. C., Sekerina, I., Hill, N. M., and Logrip, M. L. 1999. The kindergartenpath effect: Studying on-line sentence processing in young children. Cognition 73, 89–134. Trueswell, J. C., and Tanenhaus, M. K. 1991. Tense, temporal context and syntactic ambiguity resolution. Language and Cognitive Processes 6 (4), 303–338. Trueswell, J. C., and Tanenhaus, M. K. 1994. Toward a lexicalist framework for cÂ�onstraint-based syntactic ambiguity resolution. In C. Clifton, K. Rayner, and L. Frazier (eds.), Perspectives on Sentence Processing. Erlbaum. Trueswell, J. C., Tanenhaus, M. K., and Garnsey, S. 1994. Semantic influences on parsing: Use of thematic role information in syntactic ambiguity resolution. Journal of Memory and Language 33, 285–318. Trueswell, J. C., Tanenhaus, M. K., and Kello, C. 1993. Verb-specific constraints in sentence processing: Separating effects of lexical preference from garden-paths. Journal of Experimental Psychology: Learning, Memory and Cognition 19 (3), 528–553. Weighall, A. R. 2008. On still being led down the kindergarten path: Children’s processing of structural ambiguities. Journal of Experimental Child Psychology 99, 75–95. Wexler, K. 2003. Maximal trouble. Paper presented at CUNY Human Sentence Processing Conference, Boston. Ye, Z., and Zhou, X. 2009. Conflict control during sentence comprehension: fMRI evidence. NeuroImage 48, 280 –290. Yuan, S., and Fisher, C. 2009. “Really? she blicked the baby?”: Two-year-olds learn combinatorial facts about verbs by listening. Psychological Science 20 (5), 619– 626. Zelazo, P. D., and Reznick, J. S. 1991. Age-related asynchrony of knowledge and action. Child Development 62 (4), 719–735.

5â•…

Parsing, Grammar, and the Challenge of Raising Children LF at Julien Musolino and Andrea Gualmini

In this paper, we tie together two lines of psycholinguistic research represented by work on sentence processing and language acquisition. Until very recently, these two lines of inquiry have mostly proceeded independently from one aÂ�nother. That is, investigators concerned with the operation of the languagecomprehension system have rarely asked when or how the processing mechÂ� anisms uncovered in adults develop in the course of language acquisition ( butÂ€see Trueswell, Sekerina, Hill, and Logrip 1999; Meroni and Crain, this volume). Conversely, researchers concerned with the study of grammatical development, who often rely on language-comprehension paradigms, have traditionally focused on children’s grammatical competence, overlooking the study of linguistic performance in young children ( but see Crain, Ni, and Conway 1994; Crain and Thornton 1998). In this paper, we observe that when the study of language development is brought together with the study of sentence processing, the picture that emerges is substantially more interesting for both sides. First, we show that the study of grammatical development cannot make progress without a careful consideration of the child’s developing processing abilities. Moreover, we argue that differences between children and adults in sentence comprehension can be used to uncover important aspects of children’s grammatical knowledge. Second, we show that data from child language can be used to broaden the empirÂ� ical basis upon which hypotheses on adult sentence processing can be evaluated. In order to accomplish these goals, we follow a traditional approach in the field of psycholinguistics which consists in studying ambiguity resolution as a means to investigate the operations underlying language comprehension (e.g., Fodor, Bever, and Garrett 1974; Frazier and Rayner 1982; Frazier and Fodor 1978). Here, we extend this approach to the study of child language by focusing on the case of ambiguous sentences containing multiple scope-bearing operators, i.e., negation and quantified noun phrases (e.g., some dogs, every boy). The discussion is organized as follows. The second section introduces scope

110

Musolino and Gualmini

ambiguity as it pertains to negation and quantified NPs. The third section preÂ� sents the results of previous studies by Musolino (1998) and Musolino, Crain and Thornton (2000) designed to investigate the way preschoolers (and adults) resolve the scope ambiguities discussed in the second section. The main developmental finding emerging from the work of Musolino and colleagues is the observation that children systematically rely on the surface position of the relevant quantificational elements in their calculation of scope relations. In other words, if a quantificational NP occurs within the scope of negation in the surface syntax, it will correspondingly be interpreted in the scope of negation, and vice versa for an NP occurring outside the scope of negation. This is described by the so-called Observation of Isomorphism (Musolino 1998). The remainder of the chapter is devoted to an exploration of the causes, limits, and implications of this observation. The fourth section discusses whether children’s interpretation is constrained by the linear order between the relevant quantificational elements or the c-command relations holding between them. The fifth section is concerned with the causes of isomorphism. First, we review previous accounts of the phenomenon suggesting that isomorphism reflects a grammatical difference between children and adults. Then we turn to more recent experimental evidence which challenges the grammatical accounts of isomorphism. The sixth section investigates the limits of the phenomenon further, by showing that not all quantificational NPs give rise to isomorphism effects, and in particular that preschoolers are sensitive to the presuppositional status of the NPs involved. Finally, the seventh section considers the implications of the research and findings discussed throughout the paper for both theories of language acquisition and theories of language processing. Scope Relations: The Interaction of Negation and Quantified NPs

An often-discussed property of quantificational expressions is their ability to interact with one another and with other scope-bearing elements to create scope ambiguity (e.g., Horn 1989; Jackendoff 1972; May 1977). Consider for example the sentence in (1). (1) Every student can’t afford a new car. a. ∀x [student (x)â•–→â•–¬â•–can afford a new car (x)] b. ¬∀x [student (x)â•–→â•–can afford a new car (x)] On one reading, (1) can be paraphrased as Every student is such that he or she cannot afford a new car. In this case, the universally quantified subject is interpreted outside the scope of negation (abbreviated ‘everyâ•–>â•–not’), as indicated by the representation in (1a). We call this an isomorphic interpretation, since in

Parsing, Grammar, and Children

111

this case the scope relation between every student and negation can be directly read off their surface position. On another reading, (1) can be paraphrased as Not every student can afford a new car. Here, every student is interpreted within the scope of negation (abbreviated ‘notâ•–>â•–every’), as shown in (1b). We call this a non-isomorphic interpretation, since in this case the surface position of every student and negation does not coincide with their relative position at the level of semantic interpretation. The example in (2) shows that the availability of non-isomorphic interpretations depends in part on the lexical nature of the quantificational element involved. That is, replace every student in (1) with some students and the sentence is no longer perceived to be ambiguous. (2) Some students can’t afford a new car. $x [students(x)â•–∧â•–¬â•–can afford a new car (x)] The most natural — if not the only — interpretation of (2) is an isomorphic interpretation on which the sentence can be paraphrased as There are some students who cannot afford a new car (i.e., ‘someâ•–>â•–not’). Another factor determining the availability of non-isomorphic interpretations is the syntactic position of the quantified expression. This can be seen by comparing (1), which finds the universally quantified NP in subject position and is perceived to be ambiguous, and (3), which finds the universally quantified NP in object position and is not perceived to be ambiguous. (3) The professor didn’t talk to every student. ¬∀x [student (x)â•–→â•–talked to ( professor, x)] Indeed, the most natural interpretation of (3) is one on which the professor talked to some of the students but not to others — in other words, every student is interpreted within the scope of negation (i.e., ‘notâ•–>â•–every’), an isomorphic interpretation. As in the case of quantified subjects, the lexical nature of a quantified object affects its scopal properties with respect to negation. If every student in (3) is replaced by some students, as in (4), the most natural interpretation becomes a non-isomorphic one on which (4) can be paraphrased as There are some students to whom the professor didn’t talk (i.e., ‘someâ•–>â•–not’). (4) The professor didn’t talk to some students. $x [students (x)â•–∧â•–¬â•–talked to ( professor, x)] Finally, a numerally quantified object gives rise to scopal ambiguity and hence to both an isomorphic and a non-isomorphic interpretation, as shown by the indefinite two students.

112


(5) The professor didn’t talk to two students. a. ¬$2x [students (x)â•–∧â•–talked to ( professor, x)] b. $2x [students (x)â•–∧â•–¬â•–talked to ( professor, x)] The example in (5) can be paraphrased as meaning that it is not the case thatÂ€ theÂ€ professor talked to two students (for instance, the professor may haveÂ€talked to only one student). In this case, two students receives a narrowscope interpretation with respect to negation (abbreviated ‘notâ•–>â•–two’), which corresponds to an isomorphic interpretation. Alternatively, (5) can be paraphrased as meaning that there are two particular students to whom the proÂ� fessor didn’t talk. Here, two students receives wide scope with respect to negation (abbreviated ‘twoâ•–>â•–not’), which corresponds to a non-isomorphic interpretation. In sum, quantified expressions interact with other logical operators such as negation. These interactions result in complex interpretive patterns determined in part by the nature of the quantificational expressions involved and by their syntactic position. In order to investigate children’s (and adults’) interpretation of sentences containing negation and quantified NPs, the studies presented in this paper employ an experimental technique known as the Truth Value Judgment Task (TVJT) (see Crain and Thornton 1998). The TVJT was specifically developed to investigate the meanings that children assign to the sentences of their language. The TVJT typically involves two experimenters. The first experimenter acts out short stories in front of the subjects using small toys and props. The second experimenter plays the role of a puppet who watches the stories alongside the subjects. At the end of the story, the puppet makes a statement about what he thinks happened in the story. The subjects’ role is to determine whether the puppet is “right” or “wrong” about the story. Finally, the subjects are asked to justify their answers by explaining why they think the puppet was right or wrong.1 In all the experiments described in this paper, children witnessed stories that were acted out in front of them using small toys and props. Participants typically heard four test sentences (i.e. four different examples of the construction under investigation) and four filler/control sentences. Adult controls were shown the same stories as the ones witnessed by children, but instead of having the stories acted out in front of them they were shown video recordings of the stories. The Observation of Isomorphism

In order to determine when and how preschool children become aware of theÂ€complex mappings between form and meaning involved in sentences containing quantified expressions and negation, Musolino (1998) and Musolino


113

etÂ€al. (2000) tested children’s (and adults’) comprehension of sentences like (6)–(9). (6)â•‡ Every horse didn’t jump over the fence. (7)â•‡ The Smurf didn’t buy every orange. (8)â•‡ Some girls didn’t ride on the merry-go-round. (9)â•‡ The detective didn’t find someone. One of the stories used to test subjects’ comprehension of sentences like (6) involved three horses trying to jump over a fence. Two of the horses jumped over the fence but the third one didn’t. At the end of the story, a puppet described the situation as in (6). Notice that (6) is true on the non-isomorphic (i.e. ‘notâ•–>â•–every’) interpretation, since it is true that not all of the horses jumped over the fence. However, the puppet’s statement is false on the ‘everyâ•–>â•–not’ (isomorphic) interpretation, since two horses did jump over the fence. A ‘yes’ response to the puppet’s statement (along with appropriate justification) would therefore indicate that subjects are accessing the ‘notâ•–>â•–every’ (non-iÂ�somorphic) interpretation; a ‘no’ response (along with appropriate justification) would indicate that they are accessing the ‘everyâ•–>â•–not’ (isomorphic) interpretation. Musolino et al. (2000) tested a group of 20 English-speaking children between the ages of 4;0 and 7;3 (mean age 5;11) and a control group of adults and found that, whereas adults always accepted the puppet’s statement (showing that they could easily assign these sentences a non-isomorphic interpretation), children accepted the puppet’s statements only 7.5 percent of the time. When asked to justify their answers, children typically said that the puppet was wrong because two of the horses did jump over the fence. Children, therefore, unlike adults, systematically accessed the isomorphic interpretation of sentences like (6).2 One of the stories used to test subjects’ interpretation of sentences like (7) involved a Smurf who went to the grocery store to buy some fruit.3 The Smurf considered buying apples and oranges and ended up buying one of three oranges. At the end of the story, the puppet described the situation as follows: “The Smurf didn’t buy every orange.” Notice that this statement is true on an isomorphic (‘notâ•–>â•–every’) interpretation, since the Smurf didn’t buy all the oranges. However, the statement is false on a non-isomorphic (‘everyâ•–>â•–not’) interpretation, since it is false that the Smurf bought none of the oranges. Therefore, a ‘yes’ response to the puppet’s statement would indicate that subjects interpret every orange in the scope of negation, an isomorphic interpretation. A ‘no’ answer, on the other hand, along with appropriate justification, would indicate that subjects assign the sentence an ‘everyâ•–>â•–not’ interpretation.

114


What Musolino et al. found in the case of sentences like (7) is that a group of 20 English-speaking children ranging in age between 3;11 and 6;0 (mean age 4;10) accepted the puppet’s statements 85 percent of the time, and therefore systematically assigned sentences like The Smurf didn’t buy every orange a ‘notâ•–>â•–every’ interpretation. Moreover, when asked to justify their answers, children pointed to the fact that the Smurf had bought only one of the three oranges.4 In one story used to test sentences like (8), each of three girls considered riding on a merry-go-round but only one of them ended up doing so. At the end of the story, the puppet described the situation as follows: “Some girls didn’t ride on the merry-go-round.”5 In this case, the puppet’s statement is true on an isomorphic interpretation (‘someâ•–>â•–not’), since there were indeed some girls who didn’t ride on the merry-go-round. On the other hand, a ‘notâ•–>â•–some’ (nonisomorphic) interpretation falsifies the puppet’s statement, since one of the girls did ride on the merry-go-round. Musolino (1998) tested a group of 20 Englishspeaking children ranging between the ages of 4;0 and 6;2 (mean age 4;10) and found that they assigned sentences like Some girls didn’t ride on the merry-goround a ‘someâ•–>â•–not’ (isomorphic) interpretation 100 percent of the time. One story used to test sentences like (9) involved a detective playing hideand-seek with two of his friends. As the story unfolded, the detective found one of his friends but failed to find the other. At the end of the story, the puppet described the situation as The detective didn’t find someone. Notice that this statement is true on a non-isomorphic interpretation since there is indeed someone that the detective didn’t find. However, the puppet’s statement is false on an isomorphic interpretation, since it was not the case that the detective didn’t find anyone. Therefore, a ‘yes’ response to the puppet statement would indicate that subjects access a non-isomorphic interpretation, while a ‘no’ response would indicate an isomorphic interpretation. The subjects were 30 English-speaking children ranging from 3;10 to 6;6.6 The experiment also involved a control group of adults. Adult subjects always accepted statements like (9), showing that they were systematically assigning such sentences a nonisomorphic interpretation. By contrast, children accepted the puppet’s statements significantly less often (i.e., 65 percent of the time for the 5-year-olds and 35 percent of the time for the 4-year-olds). When asked to justify their negative answers, children typically explained that the puppet was wrong because the detective had indeed found someone. Thus, children, unlike adults, often assigned statements like (9) an isomorphic interpretation. Children’s interpretation of sentences like (6)–(9) led Musolino (1998) to propose “The Observation of Isomorphism” as a descriptive generalization. This observation states that children, unlike adults, have a strong tendency to


115

interpret negation and quantified NPs on the basis of their surface syntactic position. See (10). (10) Sentence type7 Every horse didn’t jump over the fence The Smurf didn’t buy every orange Some girls didn’t ride on the merry-go-round The detective didn’t find someone

Children ∀¬ ¬∀ $¬ ¬$

Adults ¬∀ ¬∀ $¬ $¬

The Observation of Isomorphism and (more generally) the existence of any systematic difference in the linguistic behavior of children and adults, raise a number of questions. The first question involves the traditional Chomskyan distinction between linguistic competence and linguistic performance (Chomsky 1965). In the case at hand, this question asks whether children’s overly isomorphic interpretations reflect a stage in linguistic development during which children do not have implicit knowledge of the fact that the grammar of their language can generate non-isomorphic interpretations (a competence account). Alternatively, the Observation of Isomorphism may be due to limitations on the computational resources that otherwise grammatically competent children deploy during language comprehension (a performance account). This is what we call the competence question. A second question raised by the findings of Musolino et al. is whether Isomorphism obtains as a consequence of the linear arrangement of the scopebearing elements involved, or whether children’s interpretation is constrained by the c-command relations holding between these elements.8 This is what Lidz and Musolino (2002) call the structural question. This question arises because linear and hierarchical (i.e., c-command) relations are systematically confounded in the constructions investigated by Musolino et al. To be sure, owing to the canonical SVO order of English, the object position follows and falls within the c-command domain of sentential negation, whereas the subject position precedes and falls outside the c-command domain of negation. A third question, related to the scope of the phenomenon observed by Musolino et al., is whether the isomorphism effect can be observed in the acquisition of languages other than English, provided of course that such languages manifest the same kind of scope phenomena as English with respect to negation and quantified NPs. Following Lidz and Musolino (2002), we call this the crosslinguistic question. In the following sections, we present a series of experiments designed to refine our understanding of children’s developing quantificational competence by addressing the questions raised above.

116


The Structural Basis of Isomorphism

The role of structure-dependent notions in child grammar has been mostly discussed with respect to syntactic principles (see, e.g., Crain and Nakayama 1987). More recently, however, students of child language have also turned to semantic phenomena. For example, Crain, Gardner, Gualmini, and Rabbin (2002) have reported evidence that preschoolers compute scope relations between logical operators on the basis of c-command relations, rather than linear order. These researchers studied children’s interpretation of the disjunction operator or in negative sentences. As Partee, ter Meulen, and Wall (1990) observed, when the disjunction operator occurs within the scope of negation (or within the scope of any downward entailing operator) one can infer the validity of each disjunct. For example, (11) entails both sentences in (12). (11) The girl who stayed up late will not get a dime or a jewel. (12) a. The girl who stayed up late will not get a dime. b. The girl who stayed up late will not get a jewel. Interestingly, these inferences can be drawn only if disjunction is c-commanded by negation. Thus, a sentence like (13), in which negation precedes but does not c-command the disjunction operator or, does not entail either of the sentences in (14). (13) The girl who did not go to sleep will get a dime or a jewel. (14) a. The girl who did not go to sleep will get a dime. b. The girl who did not go to sleep will get a jewel. To determine if English-speaking children know that the sentences in (11) and (13) yield different patterns of entailment, Crain et al. (2002) conducted a Truth Value Judgment Task. Thirty children participated in the experiment, ranging in age between 3;11 and 5;9 (mean age 5;0). In one of the trials, children were told a story about two girls who had lost a tooth and were waiting for the tooth fairy. One of the girls decided to stay up late (i.e., not to go to bed) to see what the tooth fairy looked like. The tooth fairy was very disappointed by this girl’s behavior and decided to give her a jewel but no dime. Children accepted sentences like (13) 87 percent of the time in a context in which the girl who did not go to sleep received only a jewel, which suggests that they did not interpret (13) as entailing (14a). By contrast, children rejected (11) in the same context 92 percent of the time, which suggests that they did interpret (11) as entailing (12a).


117

Returning to isomorphism effects, the question is whether structural relations are also relevant for children’s interpretation of quantified noun phrases in negative sentences. In order to address the structural and the cross-linguistic question, Lidz and Musolino (2002) tested speakers of English and also went to India to test child and adult native speakers of Kannada (a Dravidian language spoken by approximately 40 million people in the state of Karnataka in southwestern India). The canonical word order in Kannada is Subject-ObjectVerb (SOV), and Kannada displays the same kind of scope ambiguities as English with respect to negation and quantified NPs (Lidz 1999). The crucial difference between Kannada and English for our purposes is that in Kannada linear order (qua precedence) and c-command relations are not confounded. Consider the tree diagrams in (15). (15)â•‡

In English, negation both precedes and c-commands the object position, as discussed earlier. In Kannada, however, negation c-commands the object but does not precede it. This means that in Kannada, a c-command account of isomorphism would predict a preference for the narrow-scope reading of the object with respect to negation whereas a linear account of isomorphism would predict a preference for the wide-scope reading of the object NP. To test the predictions of precedence vs. c-command, Lidz and Musolino (2002) tested 4-year-olds and adults on their interpretation of sentences like (16) in both English and Kannada. (16) Cookie Monster didn’t eat two slices of pizza. a. ¬$2x [slices of pizza (x)â•–∧â•–eat (Cookie Monster, x)] b. $2x [slices of pizza (x)â•–∧â•–¬â•–eat (Cookie Monster, x)] The experiment conducted by Lidz and Musolino had two conditions. In the first, called the narrow-scope condition, sentences like (16) were true on the narrow-scope interpretation (i.e., ‘notâ•–>â•–two’) but were false on the widescope interpretation (i.e., ‘twoâ•–>â•–not’). One of the stories used in this condition

118


involved Cookie Monster and two slices of pizza. As the story unfolded, Cookie Monster ate one of the slices but not the other. At the end of the story, the puppet described the situation by stating that Cookie Monster didn’t eat two slices of pizza. In this case, the puppet’s statement was true on the narrowscope interpretation, since Cookie Monster ate only one slice of pizza (and not two), and was false on the wide-scope interpretation, since there weren’t two slices of pizza that Cookie Monster didn’t eat (there was in fact only one such slice). In the second condition, called the wide-scope condition, sentences like (16) were true on the wide-scope interpretation but were false on the narrow-scope interpretation. One of the stories used in this condition involves a situation in which Cookie Monster tried to eat four slices of pizza but managed to eat only two. As before, the puppet described the situation by stating that Cookie Monster didn’t eat two slices of pizza. In this case, the wide-scope interpretation was true, since there were indeed two slices of pizza that Cookie Monster didn’t eat; however, the narrow-scope interpretation, which stated that it was not the case that Cookie Monster ate two slices of pizza, was false, since he ate exactly two slices. The English-speaking participants were 24 children between the ages of 3;11 and 4;11 (mean age 4;4) and 24 adults. The Kannada-speaking subjects were 24 4-year-olds between the ages of 4;0 and 4;11 (mean age 4;5) and 24 adult native speakers of Kannada. Lidz and Musolino (2002) found that adult subjects overwhelmingly accepted sentences like (16) in both the narrowscope and the wide-scope condition, and displayed no significant preference for one scope interpretation over the other (i.e., 97 percent acceptance rate in the narrow-scope condition vs. 93 percent in the wide-scope condition for EÂ�nglish-speaking adults, 87.5 percent acceptance rate in the wide-scope condition vs. 85.4 percent in the narrow-scope condition for Kannada-speaking adults). Four-year-olds, on the other hand, displayed a significant preference for the narrow-scope interpretation of sentences like (16) in both languages (i.e., 81 percent acceptance rate in the narrow-scope condition vs. 33 percent in the wide-scope condition for English-speaking 4-year-olds and 75 percent vs. 22.9 percent for Kannada-speaking 4-year-olds). When asked to justify their answers, children who rejected the puppet’s statements in the wide-scope condition typically explained that the puppet was wrong because Cookie Monster had indeed eaten two slices of pizza. By contrast, in the narrow-scope condition, children almost always said that the puppet was right, and they justified their answers by pointing out that Cookie Monster had eaten only one slice of pizza.


119

These results provide an answer to both the cross-linguistic question and the structural question. First, the Observation of Isomorphism is not limited to the acquisition of English. Second, isomorphism does not follow from a one-toone mapping between linear order and semantic scope, but rather from a oneto-one mapping between surface c-command relations and semantic scope.9 The Status of Isomorphism

The competence question as it pertains to the Observation of Isomorphism was addressed by Musolino (1998) and Musolino et al. (2000). In their discussion of the competence/performance distinction, Musolino et al. (2000), following Musolino (1998), write: “The available evidence favors the grammatical hypothesis.â•–.â•–.â•–. We argue, moreover, that isomorphism can be derived from the interaction of a learning principle, the subset principle, along with fundamental properties of UG.â•–.â•–.â•–.” ( p. 20) To illustrate the argument of Musolino et al., let us first consider sentences like Every horse didn’t jump over the fence. Musolino et al. observe that, although such sentences are ambiguous between an isomorphic (i.e., ‘everyâ•–>â•– not’) and a non-isomorphic (i.e., ‘notâ•–>â•–every’) interpretation in some languages (e.g., English), they only permit an isomorphic (i.e., ‘everyâ•–>â•–not’) interpretation in others (e.g., Chinese). Languages like Chinese therefore allow only a subset of the interpretations available in languages like English (i.e., {everyâ•–>â•–not} vs. {everyâ•–>â•–not; notâ•–>â•–every}, respectively). Following the logic of the subset principle proposed by Berwick (1985), Musolino et al. aÂ�rgue that English-speaking children go through a “Chinese-speaking” stage during which their grammar generates only the ‘everyâ•–>â•–not’ interpretation of sentences like Every horse didn’t jump over the fence. Similarly, Musolino et al. argue that the subset principle can also be used to explain children’s isomorphic interpretations of sentences like The detective didn’t find someone. Their approach is based on a distinction, originally due to Hornstein (1984), between two kinds of quantified NPs (QNPs). According to Hornstein, some QNPs are assigned scope via a movement-based mechanism, i.e. type 1 QNPs (see Hornstein 1995 for a specific proposal). In addition to the movement-based mechanism, other QNPs can also be assigned scope via a mechanism that does not involve syntactic movement, i.e., type 2 QNPs (see Reinhart 1997 for a specific proposal in terms of choice functions). Moreover, according to Hornstein’s proposal, QNPs can receive non-isomorphic interpretations with respect to negation only when they are not assigned scope via movement. In other words, when the scope of a QNP is derived via (covert)

120


syntactic movement, the resulting interpretation is always an isomorphic one. Thus, the options available to type 1 QNPs (i.e., taking scope via a movementbased mechanism) are a subset of those available to type 2 QNPs (i.e., taking scope via a movement-based or a non-movement-based mechanism). The logic of the subset principle therefore dictates that children initially hypothesize that all QNPs are of type 1, and hence that they must be interpreted isomorphically with respect to negation. Since Musolino’s original findings, children’s interpretation of quantifiers in negative sentences has been the subject of numerous studies. One line of research has focused on the properties of negative sentences discussed by research in formal pragmatics and on the role of “felicity conditions” associated with the use of negative statements. We can introduce the role of felicity conditions for the interpretation of negative sentences through the examples in (17) and (18), which are from Wason 1972. The sentence in (17) is true. Still, (17) appears to be more difficult for people to evaluate than a positive sentence like (18), although (17) and (18) seem to express the same proposition. (17)â•‡ 5 is not an even number. (18)â•‡ 5 is an odd number. Interestingly, the difficulty associated with (17) is mitigated considerably if the sentence is preceded by a positive statement, as in (19). (19)â•‡ 4 is an even number, and/but 5 is not an even number. Similar considerations about the difficulty associated with negative sentences have been offered by De Villiers and Tager-Flusberg (1975). Consider (20). (20)â•‡ I didn’t drive to work. As De Villiers and Tager-Flusberg (1975, p. 279) point out, the statement in (20) “is more plausible, and consequently easier to comprehend, if it is made by someone who normally drives rather than by someone who commutes by train.”, and this property of negative sentences results from the fact that “negÂ� ative statements are generally used to point out discrepancies between a listener’s presumed expectations and the facts.” The observation that negative sentences present a specific set of felicity conditions led De Villiers and Tager-Flusberg (1975) to an experimental investigation of children’s comprehension of negation. In the experiment, children were asked to complete negative questions (i.e., ‘This is not a â•…â•… ?’) in contexts in which the use of a negative sentence was or was not plausible (e.g., in the plausible context, the experimenter had pointed to various instances of a particular object). The results showed that children as young as two responded signifi-


121

cantly faster when the negative question was presented in the plausible context. De Villiers and Tager-Flusberg interpret this finding as evidence that very young children know that the use of negative sentences is subject to the satisfaction of specific felicity conditions. To sum up, previous studies investigating the difficulty of negative sentences show that subjects experience a difficulty in processing negative sentences. The same studies, however, suggest that the difficulty associated with negative sentences can be mitigated if the target sentence is preceded by a positive lead-in or used to point out that an expectation went unfulfilled. Interestingly, recent investigations have studied the role of both factors for children’s interpretation of quantified NPs in negative sentences. The role of felicity conditions in children’s interpretation of negative sentences containing some was investigated by Gualmini (2003). In particular, Gualmini attempted to determine whether children’s difficulty with negative sentences containing some would persist when those negative sentences were used to point out a contrast between what happened and what was expected to happen. Thirty children ranging in age from 4;01 to 5;08 (mean age 4;11) participated in a Truth Value Judgment Task. The children were divided into two groups. In a typical trial, children were told a story about a firefighter playing hide-and-seek with four dwarves. During the story, it was repeatedly pointed out that the firefighter was supposed to find all of the dwarves. In the end, the firefighter managed to find only two of the four dwarves. The two groups of children were asked to evaluate (21) and (22) respectively. (21)â•‡The firefighter didn’t find some dwarves. (22)â•‡The firefighter didn’t miss some dwarves. Both sentences were true, but only (21) pointed out a discrepancy betweenÂ€what happened and what was expected to happen. Consistently, children as young asÂ€4;01 accepted sentences like (21) 90 percent of the time but accepted sentences like (22) only 45 percent of the time. In short, the results documented by Gualmini (2003) show that 4-year-olds can indeed access the non-isomorphic interpretation of negative sentences containing some in object position.10 In a similar vein, the predictions of Musolino, Crain, and Thornton’s (2000) grammatical account of isomorphism were tested in an experiment by Musolino and Lidz (2006) based on Musolino (2000). The experiment had two conditions. In both conditions, children were presented with a context in whichÂ€two horses jumped over a fence and a third horse did not jump over the fence. The first condition was designed to replicate Musolino’s (1998) finding regarding children’s isomorphic interpretation of sentences like (23). In the second condition, children were presented with sentence (23) but, in order to satisfy the

122


felicity conditions on the use of negative statements described above, this sentence was preceded by an affirmative statement that differed from (23) only in the object NP, as shown in (24). (23)â•‡ Every horse didn’t jump over the fence. (24)â•‡Every horse jumped over the log but every horse didn’t jump over the fence. In the story corresponding to (24), the three horses all jumped over the log, and then two of them also proceeded to jump over the fence. Notice that if children interpret Every horse didn’t jump over the fence in (24) to mean that none of the horses jumped over the fence, an isomorphic interpretation, they should reject the puppet’s statement and should explain, as they did in the Musolino etÂ€al. (2000) study, that the puppet was wrong because two horses did jump over the fence. On the other hand, if children assign Every horse didn’t jump over the fence in (24) a non-isomorphic (i.e., ‘notâ•–>â•–every’) interpretation, they should accept the puppet’s statement, since it is true that not all the horses jumped over the fence. The subjects were 20 English-speaking children between the ages of 5;0 and 5;11 (mean age 5;4) and 20 adult native speakers of English. Musolino and Lidz found that, whereas adult subjects accepted sentences like (23) and (24) 92.5 percent and 100 percent of the time respectively, children, who accepted sentences like (23) only 15 percent of the time, accepted sentences like (24) significantly more often, i.e. 60 percent of the time. Moreover, when asked to justify their affirmative answers to (24), children typically said that the puppet was right because one of the horses didn’t jump over the fence while the two others did. Such justifications show that children were indeed accessing a nonisomorphic interpretation. In addition to replicating Musolino’s original finding, these results demonstrate that children’s ability to access the non-isomorphic interpretation of sentences like Every horse didn’t jump over the fence improves dramatically when such a sentence is preceded by an affirmative statement. That children’s ability to access the non-isomorphic interpretation of sentences like The firefighter didn’t find some dwarves or Every horse didn’t jump over the fence improves under certain contextual manipulations casts serious doubt on the grammatical account proposed by Musolino (1998). Such a competence account makes a strong, falsifiable prediction: If children’s non-adult interpretations reflect lack of grammatical knowledge, contextual support should have no effect on the child’s ability to access non-isomorphic interpretations. Clearly, the facts pÂ�resented here indicate otherwise, and so the grammatical account must be abandoned.


123

The results presented in the preceding paragraphs provide compelling evidence against a grammatical account of isomorphism. The strategy so far has been to show that under certain conditions children can show adult-like beÂ� havior. Another way to minimize the differences between children and adults would be to show that adults can be “turned into children” (Freeman, Sinha, and Stedmon 1982). This is what Musolino and Lidz (2003) did with respect to sentences containing negation and numerally quantified NPs. First, Musolino and Lidz showed that although adult speakers of English can easily access Â�either scope interpretation of sentences like (25) they display a preference for the isomorphic interpretation in a forced choice situation. (25)â•‡ Cookie Monster didn’t eat two slices of pizza. Moreover, Musolino and Lidz showed that the isomorphism effect typically displayed by children can also be seen in adults in the case of sentences like (26). (26)â•‡ Two frogs didn’t jump over the rock. Notice that this example can receive an isomorphic interpretation, i.e. ‘twoâ•–>â•– not’, on which the sentence can be paraphrased as There are two frogs that didn’t jump over the rock. Alternatively, (26) can receive a non-isomorphic interpretation, i.e. ‘notâ•–>â•–two’, on which it can be paraphrased as It is not the case that two frogs jumped over the rock. Musolino and Lidz tested a group of 20 adult native speakers of English on their interpretation of sentences like (26) in a design similar to the one described in the fourth section of the present paper (i.e., one condition in which the isomorphic interpretation is true and the non-isomorphic condition is false and another condition in which this pattern is reversed). Musolino and Lidz found that adults displayed a significant preference for the isomorphic interpretation of sentences like (26), i.e., 100 percent acceptance in the isomorphic condition vs. only 27.5 percent in the non-isomorphic condition. Thus, the isomorphism effect typically seen in children can also be observed in adults, which invites us to consider whether that preference for surface scope could also account for the child data. Beyond Isomorphism

In the previous section, we discussed two lines of evidence against the grammatical account of isomorphism originally proposed by Musolino (1998) and Musolino et al. (2000). Yet, despite having shown that isomorphism is unlikely to reflect a grammatical difference between children and adults, we haven’t really said anything about what gives rise to the effect in the first place. In other

124


words, why do children — whose grammars we now know can generate both isomorphic and non-isomorphic interpretations — nevertheless so often resort to isomorphic interpretations? In this section, we focus on the case of sentences containing indefinite objects and negation. We propose an account of children’s isomorphic behavior based on independently motivated principles regulating the interpretation of indefinite NPs. This account, in turn, will allow us to make specific predictions regarding the limits of the phenomenon, i.e. situations in which isomorphism is expected to occur and situations in which it isn’t. These predictions are then tested — and supported — by new experimental evidence. Our account draws on theoretical work on the interpretation of iÂ�ndefinites. In his seminal work on existential constructions, Milsark (1977) observes that whereas strong determiners (e.g., every, the, most) can only have a quantificational reading, weak determiners (e.g., some, many, two) are ambiguous between a “cardinal reading” and a “quantificational reading” (see also Postal 1966). To see this, consider this example from Milsark (1977): (27)â•‡ Some salesmen walked in. This sentence can mean either that an undetermined number of people have the property of (a) being a salesman and ( b) having walked into the room (the cardinal reading) or that, given a set of salesmen, some members of this set (and not others) walked into the room (the quantificational reading). On the first reading, the property of being a salesman is attributed to the entities under discussion; on the second reading, the existence of entities who have that property is presupposed. More recently, Diesing (1992) proposed that the cardinal and the quantiÂ� ficational interpretation of indefinites are associated with two different logÂ� icalÂ€ forms. According to Diesing, an indefinite used on its cardinal reading isÂ€ not quantificational, in the sense that it only introduces a variable, which isÂ€ then bound through the existential closure operation proposed by Heim (1982). On this view, the cardinal interpretation of (27) can be represented as in (28). (28)â•‡ $x [salesmen (x) and walked in(x)] By contrast, the quantificational reading, which Diesing (1992) refers to as a presuppositional reading, involves a tripartite structure. For Diesing, the presuppositions associated with quantified sentences are represented in a reÂ� strictive clause (RC). This means that the use of a quantifier that triggers a presupposition will prompt the tripartite structure illustrated in (29). (29)â•‡ $x [salesmen (x)]RESTRICTIVE CLAUSE [walked in(x)]NUCLEAR SCOPE


125

Coming back to sentences containing negation and indefinite objects, we can view the wide-scope reading of indefinites with respect to negation as associated with the projection of a tripartite structure, (30b). On this interpretation, the indefinite is treated quantificationally. By contrast, the narrow-scope reading of the indefinite with respect to negation corresponds to the cardinal reading, (30a), on which the indefinite is treated as a variable. (30) The Smurf didn’t catch two birds. a. ¬ [$x [Smurf caught 2(x) and bird(x)]]NS (indefiniteâ•–=â•–variable) b. $2x [bird(x)]RC ¬[Smurf caught (x)]NS (indefiniteâ•–=â•–quantifier) On this view, one can hypothesize that children’s preference for the (isomorphic) narrow-scope reading of the indefinite in sentences like The Smurf didn’t catch two birds correlates with a preference for the representation in (30a). In other words, children’s “isomorphic” behavior corresponds to the selection of a representation of indefinite objects that does not involve the projection of a tripartite structure. A consequence of this line of reasoning is that the wide-scope interpretation of indefinites in negative sentences should be facilitated if a tripartite structure is independently required. In turn, this suggests that the wide-scope interpretation of indefinites in negative sentences should be facilitated if the indefinite NP also triggers a presupposition. One linguistic expression that fits this description is given by partitive constructions, as in (31). (31)â•‡ The Smurf didn’t catch two of the birds. The strongly preferred interpretation of (31) is the presuppositional, widescope reading. An explanation consistent with Diesing’s view would point to the presupposition triggered by the definite article the. On Diesing’s view, the need to represent the presupposition of existence triggered by the definite article will lead to a tripartite structure for (31), thereby generating the structure associated with the wide-scope interpretation of the indefinite object. Thus, if children’s preference for the narrow-scope reading of sentences like The Smurf didn’t catch two birds arises when they do not treat the object NP asÂ€being presuppositional, we predict that the presence of an inherently preÂ� suppositional NP, such as the partitive in (31), should lead children to assign such sentences a wide-scope interpretation much more often than in the former case. In order to test this prediction, Musolino and Gualmini (2004) tested 22 English-speaking children between the ages of 3;9 and 4;11 (mean age 4;4) and 22 English-speaking adults on their interpretation of sentences like (32) and (33) in contexts that made the wide-scope reading true and the narrow-scope reading false.

126


(32)â•‡ The Smurf didn’t catch two birds. (33)â•‡ The Smurf didn’t catch two of the birds. Musolino and Gualmini found that, whereas adults overwhelmingly accepted the puppet’s statements in both the partitive and the non-partitive condition (96 percent and 95 percent of the time, respectively), children’s tendency to assign such sentences a wide-scope, non-isomorphic reading was significantly higher in the partitive condition (33) than in the non-partitive condition (32) — 75 percent and 25 percent, respectively. As predicted, the presence of a partitive in sentences like (33) yielded a significant increase in children’s ability to access the non-isomorphic interpretation. In order to see whether this result would extend to sentences containing some, Musolino and Gualmini tested a group of 15 English-speaking children between the ages of 3;6 and 5;1 (mean age 4;4) and 15 adult native speakers of English on their comprehension of sentences like (34).11 (34)â•‡ The fireman didn’t find some of the guys. Sentences like (34) were used in contexts in which the wide-scope, non-Â� isomorphic reading (i.e., ‘someâ•–>â•–not’) was true but the narrow-scope, isomorphic reading (i.e. ‘notâ•–>â•–some’) was false. Musolino and Gualmini found that children and adults did not differ significantly in how often they accepted the puppet statements, i.e., 83.3 percent and 73.3 percent of the time, rÂ�espectively.12 Implications for Language Development and Sentence Processing

The research discussed in the present paper indicates continuity of representation as well as continuity of processes between children and adults, thereby lending support to the Continuity Assumption proposed by Pinker (1984, p. 9): “In sum, I propose that the continuity assumption be applied to accounts of children’s language in three ways: in the qualitative nature of the child’s abilities, in the formal nature of the child’s grammatical rules, and in the way that those rules are realized.” Moreover, the findings discussed here concerning children’s interpretation of sentences containing negation and quantified NPs can be brought to bear on recent theoretical proposals in the literature on sÂ�entence processing. A question that has begun to attract the attention of researchers in the field of sentence processing is what we might call “semantic processing” or, more specifically, how the parser constructs LF representations on-line from available syntactic representations (e.g., Kurtzman and MacÂ� Donald 1993; Tunstall 1997; Frazier 1999; Villalta 2003). In order to address such questions, psycholinguists have often turned to the phenomenon of scope


127

ambiguity, for reasons which should by now be clear. As Villalta (2003) observes, several recent studies have converged on the notion that the parser prefers to build LF representations that differ minimally from the corresponding surface syntactic representations: Indeed, the few recent studies that have been explicit about how corresponding LF representations are associated to surface representations, argue that the parser first chooses to construct the LF that requires minimal changes from the surface representation (e.g., the Principle of Scope Interpretation by Tunstall 1997 and the Minimal Lowering Principle by Frazier 1999). Of particular interest is the proposal made by Tunstall (1997), which explicitly addresses the question of how LF representations for sentences with multiple quantifiers are constructed. Her Principle of Scope Interpretation states that the preferred interpretation of a sentence corresponds to the LF that differs minimally from the surface structure. (2003: 123)

This captures the essence of the notion of isomorphism that we have discussed throughout the present paper. In other words, the findings presented here regarding children’s interpretation of sentences containing multiple quantificational operators (i.e. negation and quantified NPs) seem to comport rather well with the predictions of principles such as Tunstall’s Principle of Scope Interpretation and Frazier’s Minimal Lowering Principle. The challenge for such accounts, of course, is to explain why this preference isn’t always present in adults (e.g. sentences like Every horse didn’t jump over the fence, in which children and adults assign different interpretations). Moreover, such accounts would also need to explain why in certain cases (e.g., sentences containing partitives, as discussed in the previous section), children (and adults) arrive at interpretations that differ from the surface structure. Finally, processing accounts of the kind described above would also need to explain why and how certain contextual manipulations lead to significant improvements in performance and hence to a relaxation of isomorphism (see fifth section). In sum, although processing accounts relying on principles minimizing differences between surface structure and LF capture the broad pattern described in this pÂ�aper (i.e. the observation of isomorphism), many of the details remain to be accounted for. Thus, data from child language broaden the empirical basis upon which hypotheses on adult sentence processing can be evaluated. Let us now consider the main implications of our findings for theories of language acquisition. A theorist concerned with accounts of children’s lÂ�inguistic competence, and initially intrigued by the observation of isomorphism, might be disappointed by the conclusion that we ultimately reached, namely that isomorphism does not reflect a grammatical difference between children and adults. Seen from this perspective, the discovery of isomorphism may look like an unnecessary detour en route to a proper characterization of preschoolers’

128


grammatical knowledge. However, such a pessimistic view misses two important points. First, recall from the fourth section that we were able to determine that children compute scope relations on the basis of c-command relations — and hence that they know c-command — precisely because they behaved isomorphically. Thus, a “mere” performance difference between children and adults nevertheless has important implications for accounts of children’s grammatical competence and, consequently, for theories of language acquisition. Notice that the same point can be made regarding children’s ability to distinguish between partitive and non-partitive NPs (see the preceding section).Â€Here too, we were able to find that children have knowledge of a subtle linguistic property having to do with presuppositional status of NPs, precisely because of their isomorphic behavior in the case of plain NPs and lack thereof in the case of partitive NPs. Second, recall that the observation of isomorphism allowed us to demonstrate, in the case of sentences containing numerally quantified NPs (e.g. two frogs) and negation (see fourth section), that the effect observed in children represents an exaggerated interpretive preference also observable in adults (see Musolino and Lidz 2006). Furthermore, Musolino and Lidz showed that under certain conditions, adults can be “turned into children” and can display the isomorphism effect seen in children. Finally, Musolino and Lidz showed that the same contextual factors that enable children to override their isomorphic preferences have a similar effect on adults “turned into children.” These results underscore the fact that although the sentence-processing aÂ�bilities of children and adults may differ quantitatively, they do not differ qualitatively — at least within the confines of the phenomena under investigation. Notes 1.â•‡ For a more detailed description of the TVJT, see Crain and Thornton 199. 2.â•‡ All the differences between children and adults reported in this paper as significant are significant at the .05 level. 3.â•‡ Since children’s responses to sentences like (7) and (8) comport so well with grammatical judgments (which are extremely clear in these cases), adult data were not collected for these sentences. 4.â•‡ Recall that when the universally quantified NP occurs in subject position, as in sentences like Every horse didn’t jump over the fence, children assigned a ‘notâ•–>â•–every’ interpretation only 7.5 percent of the time. 5.â•‡ In fact, the puppet’s statements in this condition used the future tense instead of the past tense — e.g., Some girls won’t ride on the merry-go-round instead of Some girls didn’t ride on the merry-go-round. This difference, however, is not relevant for the purposes of the present discussion.


129

6.â•‡ In fact, one group of children (nâ•–=â•–18) was tested on their comprehension of expressions like something/someone and a second group (nâ•–=â•–12) was tested on their comprehension of expressions like some N. 7.â•‡ For sentences like Every horse didn’t jump over the fence (which is ambiguous) and sentences like The detective didn’t find someone (which may be ambiguous), the readings that adults accessed are reported in Musolino (1998). 8.â•‡ A c-commands B if and only if (i) the first branching node dominating A also dÂ�ominates B, (ii) A does not dominate B, and (iii) A is different from B (see Reinhart 1976). 9.â•‡ For further evidence that this preference is based on c-command relations and for a discussion of the implications of this result, see Lidz and Musolino 2002 and Lidz and Musolino 2003. 10.â•‡ It is worth observing that children’s ability to access non-isomorphic readings of negative sentences that point a discrepancy between an expected outcome and the actual outcome also emerged from an experiment investigating sentences containing three scope-bearing elements, such as Every farmer didn’t clean some animal (see Gualmini 2003). 11.â•‡ The effect of the partitives in minimal pairs such as (32) and (33) having been established, controls of the form The fireman didn’t find some guys were not used in the case of (34). Another reason to not include such controls is that previous findings (e.g. Musolino 1998) demonstrate that children behave isomorphically in such cases. 12.â•‡ The findings are consistent with the hypothesis that different scope assignments correlate with different semantic representations. However, we would not necessarily conclude that the difference in semantic representation dictates what scope assignment children select. It is possible that the projection of tripartite structure is too costly for young children and that, when they can, children resort to the semantic representation that does not involve such a structure. It is also possible, however, that children are perfectly capable of constructing a semantic representation involving a tripartite structure, but for independent reasons they do so only in limited circumstances. References Berwick, R. 1985. The Acquisition of Syntactic Knowledge. MIT Press. Chomsky, N. 1965. Aspects of the Theory of Syntax. MIT Press. Crain, S., and Nakayama, M. 1987. Structure dependence in grammar formation. Language 63, 522–543. Crain, S., Ni, W., and Conway, L. 1994. Learning, parsing and modularity. In C. Clifton, L. Frazier, and K. Rayner (eds.), Perspectives on Sentence Processing. Erlbaum. Crain, S., and Thornton, R. 1998. Investigations in Universal Grammar: A Guide to Research on the Acquisition of Syntax and Semantics. MIT Press. Crain, S., Gardner, A., Gualmini, A., and Rabbin, B. 2002. Children’s command of negation. In Proceedings of the Third Tokyo Conference on Psycholinguistics.

130


De Villiers, J., and Tager-Flusberg, H. 1975. Some facts one simply cannot deny. Journal of Child Language 2, 279–286. Diesing, M. 1992. Indefinites. MIT Press. Fodor, J. A., Bever, T., and Garrett, M. 1974. The Psychology of Language. McGraw-Hill. Frazier, L., and Fodor, J. D. 1978. The sausage machine: A new two-stage parsing model. Cognition 6, 291–328. Frazier, L., and Rayner, K. 1982. Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology 14, 178–210. Frazier, L. 1999. On Sentence Interpretation. Kluwer. Freeman, N. H., Sinha, C. G., and Stedmon, J. A. 1982. All the cars — which cars? From word to meaning to discourse analysis. In M. Beveridge (ed.), Children Thinking Through Language. Edward Arnold. Gualmini, A. 2003. The Ups and Downs of Child Language. Ph.D. dissertation, University of Maryland, College Park. Heim, I. 1982. The Semantics of Definite and Indefinite Noun Phrases. Ph.D. dissertation, University of Massachusetts, Amherst. Horn, L. R. 1989. A Natural History of Negation. University of Chicago Press. Hornstein, N. 1984. Logic as Grammar. MIT Press. Hornstein, N. 1995. Logical Form: From GB to Minimalism. Blackwell. Jackendoff, R. 1972. Semantic Interpretation in Generative Grammar. MIT Press. Kurtzman, H. S., and MacDonald, M. C. 1993. Resolution of quantifier scope ambiguities. Cognition 48, 243–279. Lidz, J. 1999. The morphosemantics of object case in Kannada. In Proceedings of WÂ�CCFL 18. Lidz, J., and Musolino, J. 2002. Children’s command of quantification. Cognition 84, 113–154. Lidz, J., and Musolino, J. 2003. Continuity in linguistic development: The scope of syntax and the syntax of scope. Manuscript, Northwestern University and Indiana UÂ�niversity. May, R. 1977. The Grammar of Quantification. Doctoral dissertation, Massachusetts Institute of Technology. Milsark, G. L. 1977. Toward an explanation of certain peculiarities of the existential construction in English. Linguistic Analysis 3 (1), 1–29. Musolino, J. 1998. Universal Grammar and the Acquisition of Semantic Knowledge: An Experimental Investigation of Quantifier-Negation Interaction in English. Doctoral dissertation, University of Maryland. Musolino, J. 2000. Universal quantification and the competence/performance distinction. Presented at Boston University Conference on Language Development. Musolino, J., Crain, S., and Thornton, R. 2000. Navigating negative quantificational space. Linguistics 38 (1), 1–32.


131

Musolino, J., and Gualmini, A. 2004. The role of partitivity in child language. Language Acquisition 12, 97–107. Musolino, J., and Lidz, J. 2003. ‘The scope of isomorphism: Turning adults into children. Language Acquisition 11 (4), 277–291. Musolino, J., and Lidz, J. 2006. Why children aren’t universally successful with quantification. Linguistics 44 (4), 817–852. Partee, B., ter Meulen, A., and Wall, R. 1990. Mathematical Methods in Linguistics. Kluwer. Postal, P. 1966. On so-called pronouns in English. In F. Dineen S. J. (ed.), Report of the 17th Annual Round Table Meeting on Languages and Linguistics. Georgetown University Press. Reinhart, T. 1976. The Syntactic Domain of Anaphora. Doctoral dissertation, Massachusetts Institute of Technology. Reinhart, T. 1997. Quantifier scope: How labor is divided between QR and choice functions. Linguistics and Philosophy 20: 335–397. Trueswell, J. C., Sekerina, I., Hill, N., and Logrip, M. 1999. The kindergarten-path effect: Studying on-line sentence processing in young children. Cognition 73, 89–134. Tunstall, S. 1997. Interpreting Quantifiers. Doctoral dissertation, University of Massachusetts, Amherst. Villalta, E. 2003. The role of context in the resolution of quantifier scope ambiguities. Journal of Semantics 20, 115–162. Wason, P. 1972. In real life negatives are false. Logique et Analyse 15, 17–38.

6â•…

A Cross-Linguistic Study on the Interpretation of Pronouns by Children and Agrammatic Speakers: Evidence from Dutch, Spanish, and Italian Esther Ruigendijk, Sergio Baauw, Shalom Zuckerman, Nada Vasic´, Joke de Lange, and Sergey Avrutin

Introduction

Young children and agrammatic aphasic speakers show atypical interpretation of pronouns, but not of reflexive elements. When confronted with sentences such as (1), English and Dutch children four to six years of age exhibit chance level performance on (1a) but have no problems with sentences like (1b) (e.g., Chien and Wexler 1990; Koster 1993). (1) a. The boy washes him. (chance performance) b. The boy washes himself. (good performance) The same pattern has been found for English-speaking agrammatic aphasics (Grodzinsky et al. 1993). Apparently, the interpretation of pronouns is more difficult than the interpretation of reflexives for both populations. In the literature on language acquisition, this phenomenon is known as the delay of Principle B effect (Chien and Wexler 1990; Philip and Coopmans 1996). Studies on the comprehension of sentences like (1) demonstrate that the interpretation of pronouns is non-adult-like for children and is disturbed in agrammatic aphasia. Nevertheless, there are also studies which show that pronouns are not always difficult to interpret for these populations, and that, importantly, the interpretation seems to be governed by linguistic principles (Avrutin and Thornton 1994; Avrutin et al. 1999; Grodzinsky and Reinhart 1993; Grodzinsky et al. 1993; Thornton and Wexler 1999). In this paper, we explore further the linguistic principles that underlie the interpretation of pronouns and reflexives by examining two sentence types: simple transitive sentences (as in (2a)) and Exceptional Case Marking (ECM) constructions, as in (2b). (2) Transitive versus ECM a. *John said that the boyi touched himi. b. *John said that the boyi saw himi dance.

134

Ruigendijk et al.

We use these two different structures to distinguish between important linguistic theories: Government and Binding Theory (GB, Chomsky 1981) on the one hand and Reflexivity (REFL, Reinhart and Reuland 1993) and Primitives of Binding (PoB, Reuland 2001) on the other hand. The paper is organized as follows. First, the relevant aspects of the linguistic approaches to binding are summarized. Next, relevant data from earlier studies on language acquisition and agrammatism are presented. This is followed by a description of the methods used in the study. Finally, the results are reported and discussed. We present an alternative account of data from earlier studies and the new data from the present study. Linguistic Background

According to Government and Binding theory (Chomsky 1981), the inÂ� terpretation of pronouns and reflexives is governed by two complementary pÂ�rinciples: Binding Principles (GB, Chomsky 1981) Principle A: An anaphor is bound within its governing category.1 Principle B: A pronoun is free in its governing category. Principle B rules out the interpretation in which the pronoun ‘him’ in (1a) (and (2a) and (2b)) refers to the subject ‘the boy’. Conversely, the reflexive anaphor ‘himself’ must refer to the subject ‘the boy’ in (1b), following Principle A. Reinhart and Reuland (REFL 1993), among others, have pointed out that there are some problems with this description. One problem for GB arises when it is applied to languages with more than one reflexive element. Dutch, Spanish, and Italian have both simple and complex anaphors, or SE- (Simplex Expression) and SELF-anaphors (Reinhart and Reuland 1993). These elements do not have the same distribution. For example, Dutch zich and zichzelf cannot always occur in the same position: (3) a. De jongeni gedraagt zichi/*zichzelfi The boy behaves SE/*himself b. De jongeni ziet *zichi/zichzelfi The boy sees *SE/ himself In order to account for the distribution of pronouns and both SE- and SELFanaphors, Reinhart and Reuland (1993) propose an alternative approach to Binding theory: Reflexivity. In this model, the referential behavior of pronouns and reflexives is constrained by two interacting modules: the (alternative)Â€binding theory, consisting of two binding conditions, and a (generalized) A-Chain condition.

Interpretation of Pronouns

135

Binding Conditions (REFL, Reinhart and Reuland 1993) Condition A: A reflexive-marked predicate must be interpreted reflexively. Condition B: A reflexive predicate must be reflexive-marked. General Condition on A-Chains A maximal A-Chain (α1â•–.â•–.â•–.â•–αn) has exactly one link: α1, which is both +R and marked for structural Case (where an element is +R when it is referentially independent and specified for all phi-features). The distribution of SE- and SELF-anaphors in (3) follows from Condition B: a reflexive predicate must be reflexive-marked. According to Reinhart and ReuÂ� land, reflexive marking can take place at two different levels: in the lexicon, as a result of particular lexical-semantic properties of the verb, or in the syntax, with the help of SELF-anaphors in object position. Verbs like gedragen ‘bÂ�ehave’ (3a) belong to a class of inherently reflexive verbs whose lexical-Â� semantic properties allow them to be reflexive-marked. This means that SELFanaphors are not needed to reflexive-mark the predicate.2 Verbs like zien ‘see’, on the other hand, are not inherently reflexive. Therefore, if they are interpreted reflexively, they require the presence of a SELF-anaphor in the object position to reflexive-mark the verb, as in (3b). The ungrammaticality of (2a), repeated here as (4), can be accounted for in the same way: ‘touch’ is not an inherently reflexive predicate; therefore Condition B is violated when (4) is interpreted reflexively. (4)â•‡ *John said that the boyi touched himi. However, Condition B cannot account for the ungrammaticality of (2b), repeated here as (5). (5)â•‡ *John said that the boyi saw himi dance. Here ‘the boy’ and ‘him’ are not coarguments of the same verb, ‘the boy’ being the subject of ‘saw’ and ‘him’ being the subject of ‘dance’. This means that no reflexive predicate is formed and therefore Principle B is irrelevant. Reinhart and Reuland (1993) claim that (5) is ungrammatical because it violates the A-Chain Condition. They argue that pronominal elements establish an A-Chain with local binders. The Chain Condition states that the tail of an A-Chain must be [−R], i.e., referentially deficient. According to Reinhart and Reuland, pronouns and DPs are [+R], whereas SE-anaphors (and SELF-anaphors) are [−R]. The [+R] specification is dependent on the specification of phi-features and structural case. In his Primitives of Binding, Reuland (2001) restates the Reflexivity approach in terms of the Minimalist Program. There are two important dÂ�ifferences

136

Ruigendijk et al.

between Reflexivity and Primitives of Binding. First, Reuland shows how the different conditions on binding can be derived, rather than postulated as was done in GB and REFL; second, he proposes to replace the Chain Condition by an economy hierarchy for establishing referential relationships. Reuland (2001) proposes that Condition B should be viewed as a filter on arity reduction in syntax. Reflexivization involves the reduction of one theta role by identifying it with another theta role, i.e., it turns a two-place predicate (λxλy (xRy)) into a one-place predicate (λx (xRx)). Concretely, arity reduction is not possible in the computational system or in the course of the interpretational process, unless the verb is reflexive-marked in the lexicon. In some languages this is done by adding a reflexive-marking morpheme to the verb; in other languages, including Dutch, reflexive marking is tied to specific lexicalsemantic properties of the verb. Reuland’s (2001) additional contribution in fine-tuning the concept of binding concerns the nature of Chains. He argues that Chain formation is a byproduct of a feature-checking operation between a pronominal element and a local DP. This operation leads to the deletion of the phi-features of the pronominal element and to the recovery of these features by the local DP. As a result of this operation, a referential dependency — an A-Chain — is established between the pronominal element and the local DP. Reuland further claims that [number], unlike [ person], cannot be recovered by the local DP, owing to its different ontological status.3 This means that no Chain can be created between a pronoun, such as hem ‘him’, and a local DP (as in (4) and (5)), since, unlike SE-anaphors such as zich, pronouns such as hem ‘him’ are specified for both [ person] and [number]: zich can be singular or plural, whereas hem ‘him’ is only singular. In our view the most important modification of PoB is the introduction of an economy hierarchy on referential dependencies, since this hierarchy relates theoretical linguistic notions to ( psycholinguistic) processing aspects of establishing dependencies. According to Reuland (2001), referential dependencies can be established at different linguistic levels: (i)â•‡ Narrow syntax (feature checking) (ii)â•‡ C(onceptual)-I(ntentional) interface ( bound-variable interpretation) (iii)â•‡ Knowledge based (discourse) (coreference)4 Referential dependencies established at different grammatical levels have different computational costs attached to them. The cost depends on the number of cross-modular operations necessary for interpreting a referential element. The most economical dependency is formed by checking features in narrow syntax, thus establishing an A-Chain between an element underspecified for


137

[number], for example an SE-anaphor such as Dutch zich, and a local DP, as in (6). (6) Het meisje wast zich The girl washes SE ‘The girl washes herself’ This possibility is not available to pronouns, because they are specified for [number] and cannot be involved in a feature-checking operation. However, pronouns can in principle enter into dependencies formed outside narrow syntax, at the C-I interface (a bound variable interpretation) or at the discourse level (coreferential interpretation). Both types of dependencies are assumed to be less economical than dependencies formed in narrow syntax, since a boundvariable interpretation requires one extra cross-modular operation and a coreferential interpretation requires two, thus demanding extra computational resources. These other types of dependencies are possible only if no cheaper options are available, otherwise the economy hierarchy would be violated. Syntactic chain formation is therefore the only possible way of establishing a referential dependency between two local elements, if the feature-checking operation is available in a language, and the alternative options ( bound variable and coreferential interpretation) do not yield a different interpretation. Thus, according to Reuland (2001), (7a) is out, because it violates the economy hierarchy that favors Chain formation (as in 7b) over referential dependencies created outside narrow syntax. (7) a. *Mariai voelde haari wegglijden (Bound-variable/coreference: Mary felt her slide-away MÂ�aria — haar) b. Mariai voelde zichi wegglijden (Chain formation: Maria — zich) Mary felt SE slide-away To conclude, REFL and PoB differ in one important aspect from the GB approach to binding. REFL and PoB treat the ungrammaticality of (8a) and (8b) as having a different source, whereas the GB approach considers them both Principle B violations in that the pronoun hem ‘him’ should be free in its governing category. (8) Pronouns in transitive versus ECM constructions a. *De jongeni heeft hemi aangeraakt The boy has him touched *‘The boyi touched himi’ b. *De jongeni zag hemi dansen The boy saw him dance *‘The boyi saw himi dance’

138

Ruigendijk et al.

According to REFL, only (8a) involves a Condition B violation: aangeraakt ‘touched’ is interpreted reflexively, but it is not reflexive-marked. In addition, (8a) violates the chain condition, since the pronoun hem is fully referential. The sentence in (8b) only violates the chain condition; Condition B does not apply, since de jongen ‘the boy’ and hem ‘him’ are not coarguments of the same verb. PoB also makes a distinction between (8a) and (8b), but it differs from REFL in one important respect: the economy hierarchy of establishing referential dependencies. According to Reuland (2001), (8a) and (8b) are ruled out since they violate the economy hierarchy: dependencies created in narrow syntax are cheaper than dependencies created outside narrow syntax. The difference between (8a) and (8b) under PoB is that in (8a) (and not in (8b)) de jongen and hem are coarguments and thus there is an arity-reduction violation. Arity reduction is not permitted here, since the Dutch verb aanraken ‘to touch’ is not reflexive-marked in the lexicon. The different status of pronouns in ECM sentences (such as (8b)) and transitive sentences (such as (8a)) makes it possible to investigate the roles that different modules of our language system (lexicon, narrow syntax, C-I interface, discourse) play in the interpretation of pronominal anaphora. Importantly, the cross-modular approach to pronominal reference assignment as proposed by Reuland (2001) also has implications for language acquisition and breakdown. This approach enables us to pinpoint the linguistic level or levels at which children’s and agrammatic aphasics’ problems with the interpretation of pronouns occur. Language Acquisition and Agrammatic Aphasia

It has repeatedly been shown that children make more mistakes in interpreting pronouns than in interpreting reflexive elements (e.g. Chien and Wexler 1990; Koster 1993; Philip and Coopmans 1996; Thornton and Wexler 1999). At first sight, children’s acceptance of the local subject as the antecedent of the pronoun could be interpreted as the result of children’s failure to acquire Principle B. Alternatively, both Principle A and Principle B may be innate, but binding domains may be parameterized (Wexler and Manzini 1987). Concretely, children may initially consider the VP as the relevant binding domain of pronouns, allowing them to be bound by the local subject freely. Children stop allowing this after they extend the binding domain of pronouns to IP (McKee 1992). Grodzinsky et al. (1993) studied the interpretation of pronouns and reflexives by agrammatic aphasic patients and found the same pattern in their experiment with aphasics that Chien and Wexler (1990) had found with children: the interpretation of reflexives did not cause any problems, whereas the inter-


139

pretation of pronouns in transitive sentences was at chance level. Interestingly, both Chien and Wexler (1990) and Grodzinsky et al. (1993) showed that children and agrammatic aphasics did not violate Principle B of the (standard) Binding Theory. They found that when the local subject was quantified, as in (9a), both populations scored well above chance level. (9) a. Every boy touched him (above chance performance) b. The boy touched him (chance performance) This suggests that children and agrammatic aphasics do not have difficulties with binding but with coreference.5 Local coreference rests on the interpretation of pronouns as free variables. It is not completely impossible in languages such as English and Dutch, but it is limited to special contexts.6 Apparently children and agrammatics allow local coreference in contexts where nonbrain-damaged adults reject it. Grodzinsky and Reinhart (1993) have proposed that local coreference is constrained by a syntax-discourse interface principle called Rule I. This principle states that local coreference is possible only if it yields a different interpretation than the bound-variable construal. Grodzinsky et al. (1993) further claim that children and agrammatic aphasics allow local coreference in sentences like (9b) because they have trouble applying Rule I. This rule requires the speaker to keep two different possible interpretations of the same sentence in short-term memory (the coreference and the bound variable interpretation) in order to check whether applying coreference yields a different interpretation than variable binding. Grodzinsky et al. (1993) argue that children and agrammatic patients often fail to do so because of their more limited processing resources. As a result, they resort to guessing in order to determine the reference of the pronoun. Unlike referential DPs, quantified DPs cannot establish coreference relations with pronouns; they can only bind them. Since a binding construal of (9b) leads to a Principle B violation, children and agrammatic aphasics reject a reflexive interpretation. This account of the errors with pronouns is not complete, since it is mainly based on the interpretation of pronouns in simple transitive sentences. Recently, Philip and Coopmans (in Dutch and English, 1996) and Baauw (in Spanish, 2002) also investigated children’s interpretation of pronouns in ECM constructions. Philip and Coopmans (1996) found that 5-year-old Dutch children performed at chance level on sentences such as (10a), but far below chance level (only 20 percent correct) on sentences like (10b). (10) a. De jongen heeft hem aangeraakt The boy has him touched ‘The boy touched him’

140

Ruigendijk et al.

b. De jongen zag hem dansen The boy saw him dance ‘The boy saw him dance’ Baauw (2002) also found a strong contrast in performance between simple transitive sentences and ECM sentences for Spanish children. The difference between the two types of sentences is that in ECM sentences only the Chain condition (Reinhart and Reuland 1993)7 prevents the pronoun from being bound by the main clause subject. Condition B is irrelevant. Philip and Coopmans (1996) proposed that as a result of an incomplete morphosyntactic acquisition Dutch children often misanalyze third-person pronouns as [−R] elements, treating them as SE-anaphors such as Dutch zich (see Baauw 2002 for a similar but not identical proposal). When the pronoun is interpreted as a [−R] element, the reflexive interpretation of (10b) becomes grammatical, but not the reflexive interpretation of (10a), since this interpretation would still violate Condition B. To summarize, the errors that children and agrammatic speakers make with pronouns in simple transitive sentences have been attributed to (i) a problem in the syntax-discourse interface, not in syntax proper, and have been proposed toÂ€ result from (ii) a language-processing problem, not from incomplete acÂ� quisition or loss of knowledge. The former explains why neither children nor agrammatic aphasics violate syntactic principles such as Principle B, as is shown by their good performance on sentences with quantified subjects. The higher error rate on ECM constructions could be related to the incomplete acquisition of morphosyntactic features and thus have a morphosyntactic source. We would like, however, to propose an alternative account, and to show that it is also possible to explain these data (and our own data) assuming a lack of processing resources for both children and agrammatic aphasic speakers. The aim of the present study is to investigate the nature of the difficultiesÂ€ with the interpretation of pronouns in both simple transitive and ECM sentences. We will present experimental results of children and agrammatic speakers of Dutch, Spanish, and Italian. The reason for this comparative approach is that the acquisition data show interesting cross-linguistic differences. It has been reported that Spanish and Italian 5-year-olds do not show a delay in Principle B effect in simple transitive sentences like (11), performing almost like adults (Baauw, Escobar, and Philip 1997; McKee 1992). (11) La niña la señala the girl her touches ‘The girl is touching her’ Baauw (2002) argues that the absence of the delay in Principle B effect in the Romance languages is due to the fact that Romance weak pronouns, being


141

syntactic clitics, cannot establish local coreference relations (see Baauw 2002 and Baauw and Cuetos 2003 for details). This entails that the reflexive interpretation of (11) always involves binding, which then leads to a violation of Principle B. If the cause of the difficulties with pronouns is similar in children and agrammatic aphasics (i.e. as has been proposed: a lack of processing resources, Grodzinsky et al. 1993), it is expected that the same cross-linguistic dÂ�ifferences found in children will also show up in agrammatic aphasics. In other words, as reported in earlier studies, Spanish and Italian children are not expected to show a delay in Principle B effect, and Spanish agrammatic speakers are expected to perform above chance level with the interpretation of pronouns in transitive sentences such as (11). Crucially, we expect to find a difference in the performance of children and agrammatic speakers of all three languages on pronouns in transitive and ECM constructions, where the latter will be more problematic (in line with Philip and Coopmans 1996 and Baauw 2002). These data will be used to distinguish between Government and Binding on the one hand and Reflexivity and Primitives of Binding on the other hand and to show which of the theories, in our view, can account best for the experimental rÂ�esults. Methods Participants

We examined 28 Dutch children (ages 4;3– 6;2, mean age 5;7), 15 Dutch nonbrain-damaged adults (ages 35–84, mean age 56;6) and 8 Dutch agrammatic aphasic speakers (ages 38–82, mean age 56,8). We also examined 38 Spanish children (ages 5;1–5;11, mean age 5;6), 19 non-brain-damaged adults (ages 19;0 –22;0, mean age 218), and 3 agrammatic aphasic speakers (ages 58;7– 68;6, mean age 63;10), as well as 24 Italian children (ages 4;0 –5;7, mean age 4;7) and 24 Italian adults (ages 21;8–23;6, mean age 23). The Dutch aphasic sÂ�peakers were examined with the standard Dutch aphasia examination battery, the Aachen Aphasia Test (AAT, Graetz et al. 1992), before this study. All the patients except one (a non-fluent, non-classifiable patient) were classified as Broca’s aphasics. The Spanish patients were examined with a Spanish version of the Boston Diagnostic Aphasia Examination (BDAE) and were all classified as agrammatic Broca’s aphasics. All aphasic patients spoke agrammatically as determined by the characteristics of agrammatism proposed by Menn and Obler (1990): low number of (finite) verbs, omission of pronouns and determiners, short utterances. Only patients for whom both the speech therapist and an experienced aphasiologist agreed on the diagnosis of agrammatism were included in the study.

142

Ruigendijk et al.

Materials and Procedure

The Dutch and the Italian participants were tested with a picture-selection task, whereas the Spanish participants were examined with a truth-value judgment task (TVJT). The reason for using TVJT in the Spanish study was that one of the goals of the experiment was to investigate whether Spanish agrammatic speakers show the same high performance on (11) as was found (with TVJT) for Spanish children by Baauw (2002). Since there is evidence that pictureselection tasks lead to better performance across the board (Baauw et al., submitted), the use of this task would make it impossible to determine whether possible near perfect performance on (11) is due to similar performance of children and agrammatics or to properties of the task. Both tasks consisted of more conditions than discussed here, eight in all. The other conditions included filler sentences as well as sentences with pronouns in subject or object position, with or without contrastive stress. Results on these conditions are discussed in Zuckerman et al. 2002 and in Ruigendijk et al. 2002. Both methods are described below. Dutch and Italian Picture-Selection Task

The picture-selection task was designed as follows: Each item consisted of two conjoined sentences, always starting with a sentence like ‘First the boy and the man ate something and.â•–.â•–.â•–.’ This first conjunct was followed by a sentence of one of the conditions. An example of the Dutch and Italian conditions (with an English translation) is given in (12) and (13). Note that for Dutch we included reflexives in transitive sentences (12a) so as to be able to compare our data to patterns found in earlier studies,9 such as that of Grodzinsky et al. (1993). In a later study with four agrammatic speakers, we also tested reflexives in ECM sentences such as (12e) in order to control for a possible sentence complexity effect (see Ruigendijk et al. 2006).10 This condition is reported here as well. (12) Dutch Conditions a. reflexives in transitive sentences: .â•–.â•–. daarna heeft de man zichzelf geknepen .â•–.â•–. then has the man himself pinched ‘.â•–.â•–. then the man pinched himself’ b. pronouns in transitive sentences: .â•–.â•–. daarna heeft de man hem geknepen .â•–.â•–. then has the man him pinched ‘.â•–.â•–. then the man pinched him’ c. pronouns in ECM sentences: .â•–.â•–. daarna zag de man hem voetballen .â•–.â•–. then saw the man him playing soccer ‘.â•–.â•–. then the man saw him playing soccer’


143

d. filler sentences: .â•–.â•–. daarna hebben ze gevoetbald .â•–.â•–. then have they played soccer ‘.â•–.â•–. then they played soccer’ e. reflexives in ECM sentences: .â•–.â•–. daarna zag de man zichzelf voetballen .â•–.â•–. then saw the man himself playing soccer ‘then the man saw himself playing soccer’ (13) Italian Conditions a. pronouns in transitive sentences: .â•–.â•–. poi l’uomo l’ha pizzicato .â•–.â•–. then the man him has pinched ‘.â•–.â•–. then the man pinched him’ b. pronouns in ECM sentences: .â•–.â•–. poi l’uomo l’ha visto giocare a calico .â•–.â•–. then the man him has seen play soccer ‘.â•–.â•–. then the man saw him playing soccer’ c. reflexives in ECM sentences: .â•–.â•–. poi l’uomo si è visto giocare a calico .â•–.â•–. then the man himself is seen play soccer ‘.â•–.â•–. then the man saw himself playing soccer’ d. filler sentences: .â•–.â•–. poi hanno giocato a calico .â•–.â•–. then have played soccer ‘.â•–.â•–. then they played soccer’ The examiner, who was a native speaker of the language of the task, read the sentences, while the subject was presented with four pictures on two A4-sized sheets of paper for each item. (See figure 6.1 for some examples of the pictures.) One picture on the left side represented the first conjunct of the sentence, introducing the possible antecedents. On the right side, three pictures were presented, one depicting the action in the sentence, one depicting a direct distracter (a reflexive interpretation in the case of a pronoun condition and vice versa), and one depicting an indirect distracter with the same actors but a different action. The subject was asked to listen carefully and choose the picture that fit the sentence best. For the Dutch control group and the Dutch aphasic group we included 15 items per condition. Together with the four other conditions (included for aÂ�nother study on the effect of contrastive stress) this resulted in a battery of 120Â€items. To be able to examine this high quantity of items, the agrammatic speakers were tested in two sessions with one or two weeks between them. For

144

Ruigendijk et al.

Figure 6.1

An example of the picture-selection task. Testing ECM: “First the boy and the man ate and then the man saw him playing soccer.” The middle picture on the right depicts the correct response.

the Dutch children we included six items and they were tested twice with the same test, resulting in 12 testing points per condition. The Italian children and controls were tested with six items per condition, and could unfortunately be tested only once (due to limited availability of the participants). For the Dutch and the Italian children the task was presented as a game, using a puppet. (See figure 6.1.) Spanish Truth-Value Judgment Task

The truth-value judgment task differed only in the way the items were presented. Again each item was introduced by a sentence (Primero la mujer y la


145

niña rieron .â•–.â•–. ‘First the woman and the girl were laughing .â•–.â•–.’), with a picture. The experimental sentence was presented with another picture. The participant was asked to judge whether the sentence that was presented described the picture.11 We included as many ‘yes’ items as ‘no’ items, and each sentence occurred in both ‘yes’ and ‘no’ conditions. Six items were included per condition (see (14) for an example of each condition). The test consisted of 60 items in total. (14) Spanish Conditions a. reflexives in transitive sentences: .â•–.â•–. la niñase pellizcó .â•–.â•–. the girl SE pinched ‘.â•–.â•–. the girl pinched herself’ b. pronouns in transitive sentences: .â•–.â•–. la niña la pellizcó. .â•–.â•–. the girl her pinched. ‘.â•–.â•–. the girl pinched her’ c. reflexives in ECM sentences: .â•–.â•–. la mujer se vio reir .â•–.â•–. the woman SE saw laughing ‘.â•–.â•–. the woman saw herself laughing’ d. pronouns in ECM sentences: .â•–.â•–. la niña la vio bailar .â•–.â•–. the girl her saw dancing ‘.â•–.â•–. the girl saw her dancing’ e. Filler .â•–.â•–. la niña pellizcó al hombre .â•–.â•–. the girl pinched acc.-the man ‘.â•–.â•–. the girl pinched the man’ For the children the experiment was presented as a game (with a puppet that “produced” the experimental sentences), and the examiner was a native speaker of Spanish. During testing it was noted whether the participant said ‘yes’ or ‘no’ to the experimental sentence. For additional details on this method, see Baauw and Cuetos 2003. Results

Table 6.1 shows the results of the experiments. Both children and agrammatic aphasics made significantly fewer errors with the interpretation of pronouns inÂ€ simple transitive sentences than in ECM constructions. This held for all

146

Ruigendijk et al.

Table 6.1

Percentages correct on the different conditions. n.a. = not available. For the Spanish data, only scores on the ‘no’ conditions are reported.

Condition

Transitive sentence

ECM construction

Reflexive

Pronoun

Reflexive

Pronoun

Fillers

Dutch

Controls Children Aphasics

â•‡ 98.7 â•‡ 92.3 â•‡ 93.3

98.2 75.0 91.7

100a n.a. â•‡ 96.7b

97.8 46.7 56.7

100 â•‡ 99.7 â•‡ 99.2

Spanish

Controls Children Aphasics

100 â•‡ 88.6 â•‡ 83.3

96.3 90.4 88.7

â•‡ 94.4 â•‡ 86.0 â•‡ 83.3

96.3 59.7 11.3

100 â•‡ 94 100

Italian

Controls Children

n.a. n.a.

98.6 78.5

100 â•‡ 75.7

99.3 57.6

100 â•‡ 97.2

a.â•‡ Data from another control group of 13 Dutch speakers (mean age 42;6). b.â•‡ Data from four of the eight agrammatic subjects.

languages that were examined (Wilcoxon Signed Ranks Test, Dutch agrammatic aphasics: Zâ•–=â•–−2.379, pâ•–

The Processing and Acquisition of Reference

The Processing and Acquisition of Reference

Language acquisition: knowledge representation and processing

Language Acquisition: Knowledge Representation and Processing

Image Acquisition and Processing with LabVIEW (Image Processing Series)

Language acquisition: knowledge representation and processing

Guide to preservation in acquisition processing

The Acquisition of Swahili

Language Processing and Acquisition in Languages of Semitic, Root-based, Morphology (Language Acquisition & Language Disorders)

Grammar Acquisition and Processing Instruction: Secondary and Cumulative Effects (Second Language Acquisition)

The acquisition of the lexicon

The Acquisition

The Acquisition of the Lexicon

The Acquisition

Acquisition

The Acquisition of Aspect and Modality: The Case of Past Reference in Turkish

Processing Instruction: Theory, Research, and Commentary (Second Language Acquisition Research)

Data Acquisition and Signal Processing for Smart Sensors

Data Acquisition and Signal Processing for Smart Sensors

The Acquisition of Swedish Grammar

The acquisition of complex sentences

The Acquisition of Complex Sentences

The Acquisition of Syntactic Knowledge

The Process and Politics of Defense Acquisition: A Reference Handbook (Contemporary Military, Strategic, and Security Issues)

The Acquisition of French: The development of inflectional morphology and syntax in L1 acquisition, bilingualism, and L2 acquisition (Language Acquisition and Language Disorders)

The Acquisition of Swedish Grammar (Language Acquisition & Language Disorders)

Language acquisition and learnability

Language Acquisition and Learnability

Language Acquisition and Learnability

Ferengi Rules Of Acquisition

Ferengi Rules of Acquisition

The Processing and Acquisition of Reference