BILINGUAL SENTENCE PROCESSING
ADVANCES IN PSYCHOLOGY 134 Editor:
G. E. STELMACH
ELSEVIER A m s t e r d a m - L o n d o n - N e w Y o r k - O x f o r d - Paris - T o k y o Boston- San Diego - San Francisco - Singapore - Sydney
BILINGUAL SENTENCE PROCESSING
Edited by Roberto R. Heredia Department of Psychology Texas A &M International University Laredo, TX, U.S.A. Jeanette Altarriba Department of Psychology University at Albany State University of New York Albany, New York, U.S.A.
2002 ELSEVIER Amsterdam- London- New York- Oxford- Paris - TokyoBoston- San Diego - San Francisco- Singapore- Sydney
ELSEVIER SCIENCE B.V. Sara Burgerhartstraat 25 p.o. Box 211,1000 AE Amsterdam, The Netherlands 9 2002 Elsevier Science B.V. All rights reserved. This work is protected under copyright by Elsevier Science, and the following terms and conditions apply to its use: Photocopying Single photocopies of single chapters may be made for personal use as allowed by national copyright laws. Permission of the Publisher and payment of a fee is required for all other photocopying, including multiple or systematic copying, copying for advertising or promotional purposes, resale, and all forms of document delivery. Special rates are available for educational institutions that wish to make photocopies for non-profit educational classroom use. Permissions may be sought directly from Elsevier Science Global Rights Department, PO Box 800, Oxford OX5 IDX, UK; phone: (+44) 1865 843830, fax: (+44) 1865 853333, e-mail:
[email protected]. You may also contact Global Rights directly through Elsevier's home page (http://www.elsevier.nl), by selecting 'Obtaining Permissions'. In the USA, users may clear permissions and make payments through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA; phone: (+1) (978) 7508400, fax (+1) (978) 7504744, and in the UK through the Copyright Licensing Agency Rapid Clearance Service (CLARCS), 90 Tottenham Court Road, London W 1P 0LP, UK; phone (+44) 207 6315555; fax: (+44) 207 6315500. Other countries may have a local reprographic rights agency for payments. Derivative works Tables of contents may be reproduced for internal circulation, but permission of Elsevier Science is required for external resale or distribution of such material. Permission of the Publisher is required for all other derivative works, including compilations and translations. Electronic Storage or Usage Permission of the publisher is required to store or use electronically any material contained in this work, including any chapter or part of a chapter. Except as outlined above, no part of this work may be reproduced, stored in a retrieval system or transmitted in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without prior written permission of the Publisher. Address permissions requests to: Elsevier Science Global Rights Department, at the mail, fax and e-mail addresses noted above. Notice No responsibility is assumed by the Publisher for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions or ideas contained in the material herein. Because of rapid advances in the medical sciences, in particular, independent verification of diagnoses and drug dosages should be made.
First edition 2002 Library of Congress Cataloging in Publication Data A catalog record from the Library of Congress has been applied for.
ISBN: 0 444 50847 3 ISSN: 0166-4115 (Series) The paper used in this publication meets the requirements of ANSI/NISO Z39.48-1992 (Permanence of Paper). Printed in The Netherlands.
Acknowledgements We would like to thank the many people who helped us directly and indirectly in the completion of this book. First, we thank Fiona Barron of Elsevier for catching the vision of this unique work. Also, we would like to express our gratitude to Greg B. Simpson for providing a thoughtful and positive Preface to the current work. Our thanks to the contributors themselves who were cooperative in meeting our sometimes ambitious deadlines. I, Roberto, would like to acknowledge the Center for Research in Language (CRL), The Cognitive Science Department, and The Center for Information Processing at the University of Califomia, San Diego for sparking my interest in this very important field of bilingual sentence processing. I thank my excellent friend and colleague Jeffrey M. Brown for his unconditional support and his insightful comments and suggestions on every chapter of this volume. We deeply appreciate those dedicated students at Texas A&M International University for their support and assistance in the preparation of this book. Among them, we particularly thank Liza Michelle Garcia, Gretel Arredondo, Selma Nanette Maldonado, Patricia Gutierrez, Florentina Aguallo, Graciela Yafiez and Erika Rodriguez. I, Jeanette, feel grateful that the colleagues who contributed to this book acknowledge the importance of the topic in the field of bilingual research and continue to serve as role models for other scientists in this area. Finally, we wish to acknowledge our friends who are too numerous to name and our wonderful families. It is their love and support that always motivates and encourages us. Este libro est~i dedicado para todos ustedes.
Con Carifio, Roberto R. Heredia and Jeanette Altarriba
This Page Intentionally Left Blank
Preface
The presentation of a volume on bilingual sentence processing is timely (indeed, it seems long overdue) for several reasons. Perhaps the most obvious of these, though not necessarily what might first come to mind, is the recognition of the prevalence of multilingualism around the world. Except in the United States, bilingualism generally is the rule rather than the exception. Even in the United States, which has come slowly to this recognition, greater attention than ever before is being paid to bilingualism. This interest is reflected in studies of second-language acquisition and pedagogy. As it becomes clearer that multilingualism is increasingly important in a shrinking world, greater understanding of the processes of language learning and teaching becomes critical. Interest in bilingualism is reflected as well in the growing body of literature, exemplified by the present volume, concerning cognitive processes in bilingual speakers. Researchers and educators interested in issues of bilingualism will find the present treatment to be comprehensive in its consideration of language processes in bilingual speakers, from the representation of individual words in two languages, to the processing of figurative language. For many years, research in this area was dominated by questions regarding the representation in memory of lexical information in two languages. One can see from the chapters here that research and theory have progressed far beyond the word level. The topics span the entire range of issues commonly found in psycholinguistics texts, including syntactic and semantic processing, memory representation, and language acquisition. The present book, therefore, provides for the first time a comprehensive treatment of issues of bilingualism and their implications for issues of language processes in general, and for sentence processing, in particular. After consideration of methodological issues in the study of bilingual language processing, two connectionist models are presented. Like other models of this type, these two go far beyond consideration of a single level of representation or processing, and should prove fruitful for researchers studying interactions among processing domains by bilinguals. The third section of the book, on Memory Representation in Sentence Processing, represents what is perhaps the most traditional area in bilingual research. The chapters here, however, again exceed the questions of word representation that have been dominant in the past, and also introduce new procedures for researchers in bilingualism (such as the repetition-blindness paradigm). The use of context in the recognition of words in two languages is a well-traveled ground, but here these studies involve consideration of cross-language homographs and of context greater than the single word or sentence. In Psycholinguistic Theory and Research, the implications for issues outside bilingualism itself are perhaps the clearest. Hypotheses concerning, for example, parsing strategies are usually based on assumptions of universal aspects of language and language processing. Studies of bilingual speakers are therefore critical to tests of these hypotheses of syntactic processes. Similarly, the studies of figurative
viii
Preface
language that are reported here tell us as much about figurative language as they tell us about bilingual speakers, dealing with issues such as the processing of syntactic ambiguity in phrasal verbs, and visual imagery and idiom processing. Finally, the acquisition of two languages in children is given its proper due. Indeed, no topic could be more central to bilingualism. Here again, however, the discussion is not restricted to issues concerning characteristics of the learner and learning environment. Written language comprehension and production have not been widely considered in the past, and these chapters show how much research in bilingual language acquisition has progressed in recent years. Partly because of the promise of its educational applications, bilingualism is a legitimate field of study in its own right. This fact might be enough to capture the passion of psycholinguists generally, but alas, to date it seems not to have done so. What could capture this passion is the recognition that the bilingual language processor might offer a unique medium for understanding fundamental language processes that go beyond bilingualism itself. There is a disappointing tendency, certainly among many American researchers, to consider studies conducted in any language other than English to be about that language. Their own research, conducted in English, is naturally assumed to be concemed with the architecture of language processing in general. Similarly, research on bilingual processing may be neglected because it is perceived to shed light only on the issue of bilingualism per se. Obviously, the present authors are not guilty of such provincialism. Ultimately, promoting the recognition of the relevance of the issues discussed here in the context of the bilingual speaker to broader issues of language processing may be one of this volume's greatest contributions to the field.
Greg B. Simpson University of Kansas October 28,2001
Table of Contents
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
v
Preface Greg B. Simpson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
vii
Introduction and Overview Jeanette Altarriba and Roberto R. Heredia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
Part I. Methods in Bilingual Research Chapter 1. On-Line Methods in Bilingual Spoken Language Research Roberto R. Heredia and Mark T. Stewart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
Part II. Connectionist Models of Second Language Processing and Bilingualism Chapter 2. Extending the Competition Model Brian MacWhinney . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
Chapter 3. A Self-Organizing Connectionist Model of Bilingual Processing Ping Li and Igor Farkas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
Part III. Memory Representation in Sentence Processing C h a p t e r 4. Cross-Language Facilitation, Repetition Blindness, and the Relation Between Language and Memory: Replications of Altarriba and Soltano (1996) and Support for a New Theory Don G. MacKay, Lori E. James and Lise Abrams . . . . . . . . . . . . . . . . . . . . . . . 89
Chapter 5. The Use of Sentence Contexts in Reading, Memory, and Semantic Disambiguation Jeanette Altarriba and Jennifer L. Gianico . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Chapter 6. Exploring Language Asymmetries in Early Spanish-English Bilinguals: The Role of Lexical and Sentential Context Effects Arturo E. Hem~ndez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
137
Chapter 7. Text Comprehension in Bilinguals: Integrating Perspectives on Language Representation and Text Processing Gary E. Raney, Sharon M. Obeidallah and Timothy K. Miura . . . . . . . . . . . .
165
Part IV. Psycholinguistic Theory and Research Chapter 8. Relative Clause Attachment in Bilinguals and Monolinguals Eva M. Fem~indez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
187
Chapter 9. An On-Line Look at Sentence Processing in the Second Language Cheryl Frenck-Mestre . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 Chapter 10. Cross-Linguistic Aspects of Anaphor Resolution Dieter Hillert . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
237
Part V. Figurative Language Processing Chapter 11. Understanding Phrasal Verbs in Monolinguals and Bilinguals Teenie Matlock and Roberto R. Heredia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 Chapter 12. What Native and Non-Native Speakers' Images for Idioms Tell Us About Figurative Language Heather Bortfeld . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Part VI. Language Skill Development in Bilingual Children Chapter 13. The Role of Formal Definitions in Reading Comprehension of Bilingual Students Aydin Y. Durguno~lu, Zehra F. Peynircio~,lu and Montserrat Mir . . . . . . . . . 299 Chapter 14. Syntactic Structure, Grammatical Accuracy, and Content in Second-Language Writing: An Analysis of Skill Learning and On-Line Processing Wendy S. Francis, Laura F. Romo and Rochel Gelman . . . . . . . . . . . . . . . . .
317
Chapter 15. Code Switching in Preschool Bilingual Children Zehra Peynircio~,lu and Aydm Y. Durguno~lu . . . . . . . . . . . . . . . . . . . . . . . .
339
Authors' Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
359
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
363
Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
373
Introduction and Overview
Bilingual research has increased a great deal in the past ten years. Within the field of psycholinguistics, bilingual research has focused on a broad range of topics frtSm bilingual memory representation to language processing issues such as the nature of word representation, bilingual word recognition, and reading. Because of the universality of bilingualism, being much more the rule than the exception, interest in the field of bilingual sentence processing is growing significantly. Indeed, the study of bilingualism attracts a broad range of individuals from disciplines such as linguistics, psycholinguistics, cognitive science, communication, and artificial intelligence, just to name a few. A comprehensive volume such as the current one was intended to keep students and researchers interested in bilingual sentence processing issues and motivate them to pursue investigations into various related areas in bilingualism. To date, there are no other books or specialized texts devoted exclusively to the topic of bilingual sentence processing. The aim of the current work was to fill this void in the literature by collecting works that emphasize theoretical issues and the presentation of empirical findings or evidence in support of different theoretical positions. Research from both the visual and spoken modalities was included along with a comprehensive selection of paradigms and methods that have been applied in the cross-linguistic domain. A volume such as this one provides both the beginning researcher and the seasoned researcher with the latest on bilingual investigations. It serves as a technical handbook in the field as well as a tutorial reader that can provide a starting point for almost any direction in the bilingual language domain. We would be remiss if we did not cite some of the earlier books published on bilingual research as they helped to shape the direction and focus of the work reported in the current volume. Books such as Language Processing in Bilinguals edited by Vaid (1986), Cognitive Processing in Bilinguals edited by Harris (1992), and The Bilingual Lexicon, edited by Schreuder and Weltens (1993) introduced researchers and theoreticians to the basic issues and debates that structured the framework for exploring bilingual language issues that has been expanded upon in recent years. Most of these works contained papers that focused on the word level of language representation and its various attributes such as orthography, phonology, semantics, pragmatics and the like. They laid the groundwork for the investigations that are presented in the current work, as issues related to the processing of words in sentential contexts were informed by the previously published investigations. The central focus of Bilingual Sentence Processing was the presentation of an overview of the literature on bilingual sentence processing from a psycholinguistic and linguistic perspective. To do this, noted researchers in the field of bilingual language processing agreed to contribute works ranging from the development of connectionist models, to the representation of cross-linguistic figurative language, to investigations of reading and comprehension in bilingual children and adults. The volume is divided into six sections each describing work that can be viewed as a
2
Jeanette Altarriba and Roberto R. Heredia
major theme within the current thinking in the field. The work opens with a discussion of the methods used in bilingual research. This work provides an overview of the various on-line tasks that have been used to investigate language processing in bilinguals at the sentence level. The second section begins with the fundamental issue of mental representation and the nature of the structure of the two languages in memory, with the presentation and description of two connectionist models of bilingual language processing. Next, the interplay between language contexts, memory, and reading is explored by examining effects such as repetition blindness, lexical ambiguity, and sentence constraint. The fourth section has its strengths in the presentation of linguistic theory and research and examines those questions that surround the syntactic parsing and strategic processing that occurs specifically within sentence contexts. The nuts and bolts of the linguistics surrounding sentence comprehension are included here. A fifth grouping of chapters emphasizes an area that has often been overlooked in previous volumes on language processing--the representation and use of figurative language (i.e., metaphors, idioms and the like). This novel inclusion should draw the attention of individuals who are interested in the broader questions of language acquisition, the implicitness of language representation, and the interpretation of inferences and linguistic implicatures. They make for interesting reading, as well, as the examples provided should alert the reader to the fact that figurative language exists across all languages and is often an interesting link to the representation of thought within a particular language group or culture. Last but not least, a growing interest in second language acquisition prompted the inclusion of the final section of the current volume on the development of language skills in bilingual children. Questions regarding the representation of vocabulary in the two mental lexicons of bilinguals, to the extraction of semantic meaning from stories and narratives, to the implications of code-switching in verbal descriptions was included, rounding out a complete selection of topics in the area of sentence processing. Several debates abound in the area of bilingual sentence research. Quite notable is the issue regarding on-line versus off-line processing. Measures that record behavior in an on-line fashion are aimed towards examining language processing as it o c c u r s in a natural state of reading and listening, and word identification (cf. Chapters 1 & 6). Tasks such as those recorded by eyetracking systems and selfpaced reading tasks that may involve moving windows are an example of on-line procedures (cf. Chapters 5, 9, & 11). Other studies, particularly those interested in the processing occurring after encoded material is processed and integrated may rely on measures that are considered "off-line." Work in recall of words from mixed-language presentations or the measure of semantic comprehension after material is read are examples of this type of processing (cf. Chapters 4 & 7). It is clear that while some information can be gained within each paradigm, the ultimate decision as to which to employ should depend on the research question under investigation. Would one want to capture processes that are pre-lexical or those that occur early in the stages of language processing, or, would issues regarding memory retrieval and the time course of forgetting or interference in memory be of interest? Both types of approaches have value and research should continue to focus on existing tasks and the development of new tasks in both domains. Another concern seems to reside with the factors implicit in language learning that contribute to the
Introduction and Overview
3
mental representations of a bilingual's lexicons The relative significance of age of acquisition is not well understood and is often either overlooked or squarely emphasized as an important source of variance in language learning (see e.g., Chapter 2 vs. Chapter 9). Another intriguing focus that currently motivates quite a bit of research in the bilingual domain centers on the question of whether or not language use is selective or non-selective. That is, can an individual functionally process information in one language without interference from a second language? Work involving the Stroop paradigm in which words that name colors in a given language are presented in varying ink colors, indicates that languages may not be entirely separable when processing. The processing of ambiguity within and between languages also addresses the question as to the separateness of multiple languages (cf. Chapters 1, 5, 9, & 12). Finally, as in the monolingual domain, questions regarding the nature of the architecture of the mental lexicon(s) in the brain has led to the development of models of representation that are neurally inspired. The current volume includes two such models that challenge the current capacity models of information processing across languages by describing both localist and distributed properties of language. These connectionist models not only describe representation but also describe the processes involved in language learning. They can aptly predict findings for both priming effects and effects of interference across languages. Despite the varied approaches and abundance of data described in the current volume, the need ultimately remains to provide evidence regarding bilingual sentence processing in cases in which the exact same stimuli are used across a variety of experimental paradigms. Likewise, an investigation into the relative similarities and differences in processing as a function of languages from different historical origins is in order. Despite these issues in the field, a strength in the current volume lies in the fact that sentence processing has been reviewed at various levels from the most basic and theoretical to applications within education and specifically, within the classroom. Whereas researchers may typically focus on the level of theoretical constructs and their empirical counterparts, the current book also offers ways in which this work can be applied to a most pressing and timely topic-the acquisition of a new language. In this way, we hope this volume becomes a type of "handbook" for researchers wanting to know how to conduct work that will have practical implications as well as answer theoretical queries of general interest. Finally, we hope that this work will serve as a springboard for new research that moves this area of intrigue ahead by having summed up what we know up until the present.
References
Harris, R. J. (Ed.). (1992). Cognitive processing in bilinguals. Amsterdam: Elsevier Science Publishers. Schreuder, R., & Weltens, B. (Eds.) (1993). The bilingual lexicon. Philadelphia, PA: John Benjamins. Vaid, J. (Eds.) (1986). Language processing in bilinguals: Psycholinguistic and neuropsychological approaches. Hillsdale, NJ: Lawrence Erlbaum Associates.
This Page Intentionally Left Blank
Part I: Methods in Bilingual Research
This Page Intentionally Left Blank
Bilingual Sentence Processing- R.R. Heredia and J. Altarriba (Editors) 9 2002 Elsevier Science B.V. All rights reserved.
On-Line Methods in Bilingual Spoken Language Research Roberto R. Heredia Texas A&M International University Mark T. Stewart Willamette University
Abstract
In this chapter, we review several methodological approaches to studying spoken language comprehension in bilinguals. In particular, we focus on those tasks that allow for closer scrutiny of sentence level processing, including gating (Grosjean, 1980, 1996), cued shadowing (Bates & Liu, 1996), crossmodal lexical priming (Swinney, 1979), and the auditory moving-window (Ferreira, Henderson, Anes, Weeks, & McFarlane, 1996). Throughout the chapter, we offer suggestions as to how these techniques might be used to investigate a host of issues important to researchers in bilingualism, including but not limited to grammatical priming, ambiguity resolution, and contextual priming. We highlight the potential strengths and weaknesses of each paradigm and, whenever possible, offer suggestions as to future work that might be conducted using a particular experimental procedure. In the end, our hope is that the reader will come away with a stronger sense of the sorts of techniques that are available and are being used by researchers in the field of bilingual sentence processing. While research at the word level, and primarily in the visual modality, has examined the question of whether bilinguals organize their two languages into one or two memory systems (e.g., Durguno~lu & Roediger, 1987; Glanzer & Duarte, 1971; MacNamara & Kushnir, 1971), and most recently, the organization of the bilingual lexicon (e.g., de Groot, Dannenburg, & van Hell, 1994; Heredia, 1997; Kroll & Stewart, 1994), considerably less is known about how bilinguals process connected speech, both at the sentential and discourse levels, during the communicative process. This state of affairs is interesting given how much we rely on our auditory sense for our linguistic information in our everyday communication in our first and second languages (cf. Ferreira, Henderson et al., 1996). More important, however, is the fact that in most bilingual communities, communication between bilinguals is largely spoken, with comparatively little at the written level. For example, it is not unusual for Spanish-English bilinguals in the Southwest of the United States to show high language proficiency levels in understanding spoken language and in their ability to speak their two languages. However, these same bilinguals show limited ability in their written and
8
Roberto R. Heredia and Mark T. Stewart
reading skills (e.g., Heredia & Altarriba, 2001; cf. Hem~indez,/~vila, & Bates, 1996; see also Favreau & Segalowitz, 1982, 1983; Segalowitz, 1986). Moreover, it could be argued that certain linguistic forms of bilingualism occur at the spoken language modality and not necessarily at the written language level (e.g., Ferreira, Henderson et al., 1996). For instance, code switching or language mixing in which bilinguals substitute a word or a phrase in one language for a word or phrase in a second language (Li, 1996a) is mainly a spoken language phenomenon. This is not to say, that code switching can only be studied in the spoken language modality because there are impressive research findings that have used reading to study this phenomenon (e.g., Altarriba, Kroll, Sholl, & Rayner, 1996; see also Peynircio~lu & Durguno~lu, this volume). In this chapter, we review some of the main methodologies in the existing bilingual literature that have taken the sentence or discourse level information as their focus to examine bilingual language processing and semantic memory issues. Our discussion is limited to only those studies that have used the auditory modality (for an overview of other tasks using spoken language see, Guillelmon & Grosjean, 2001; Grosjean & Frauenfelder, 1996). We conclude by describing two research techniques that are currently being explored in our laboratories to study issues related to bilingual lexical access and the phenomenon of code switching. We now turn to our discussion of the various on-line psycholinguistic tasks used in examining bilingual sentence processing.
The Sentence Interpretation Task One of the first paradigms to be used to investigate bilingual sentence processing was the sentence interpretation task (e.g., Hem~indez, Bates, & Avila, 1994; Kilbom, 1987, 1989; Li, Bates, Liu, & MacWhinney, 1992; MacWhinney, 1987; MacWhinney & Bates, 1989; McDonald, 1987a; McDonald & Heilenman, 1991, 1992;Vaid & Pandit, 1991). In this task, participants listen to a sentence such as (1 a) below and are asked to report aloud as quickly and as accurately as possible who or what in the sentence is doing the action (Kilborn, 1989). Thus, in the example below, the correct verbal response would be dog. One variation has been to ask participants to make a timed choice between a pair of pictures (e.g., dog vs. cup) (Hern~indez et al., 1994). (1 a) The dog is chasing the cups This particular version of the task has been used predominantly in cross-linguistic studies (e.g., Bates, Devescovi, & Wulfeck, 2001; MacWhinney & Bates, 1989) seeking to understand how speakers utilize the various sources of information provided by a particular language, and how speakers incorporate and use these sources of information during the language comprehension process. For example, sentence (la) above provides important sources of information that the English speaker must consider during sentence processing. These sources of information are, (a) the position of the first noun that is doing the action (e.g., dog); (b) the agreement between subject and verb in person and number; and (c) a contrast in animacy between the subject dog,
On-line Methods in Bilingual Spoken Language
9
which is animate and the object cups, which is inanimate (Li et al., 1992). How these sources of information are utilized depends on the particular language. Languages such as English rely more on word order (e.g., SVO: subject-verb-object), than other languages such as Spanish that rely mainly on morphological information such as noun and verb agreement and less on word order. For instance, sentence (2a) is grammatically correct in English because it follows the SVO word order. Sentence (2b) on the other hand, is grammatically incorrect because it does not follow the English SVO word order, and without the subject/, it is unclear who wants the television. In Spanish, on the other hand, sentence (2b) is correct because the verb carries information about the subject. That is, in Spanish the subject is optional and word order is not restricted to SVO. (Please note that grammatically incorrect sentences are indicated by double asterisks "**"). (2a) I want the television (Yo quiero el televisor) (2b) **Want the television (Quiero el televisor) Hern~indez 9 et al. (1994) used a version of this task to examine sentence interpretation strategies in Spanish-English bilinguals compared to Spanish and English monolinguals. The question of interest was whether Spanish-English bilinguals would behave like English monolinguals during the comprehension of English sentences, or like Spanish monolinguals during the comprehension of Spanish sentences. SpanishEnglish bilinguals received both Spanish and English versions of sentences such as (3a3d) below that experimentally manipulated different variations of word order (e.g., NVN: noun-verb-noun; VNN: verb-noun-noun, and NNV: noun-noun-verb), agreement between noun and verb (sentences 3a-3d), and animacy, where the first noun was animate (e.g., 3a, 3b, and 3c) or inanimate (e.g., 3d), and the second noun was animate (e.g., 3a and 3d) or inanimate (e.g., 3b and 3c). (3a) The dog is chasing the cow (both nouns agree (3b) **The dog are chasing the cups (second noun (3c) The dog is chasing the cups (first noun agrees (3d) **The cup are chasing the dogs (second noun
with verb) agrees with the verb) with verb) agrees with the verb)
Spanish and English monolinguals, on the other hand, received the sentences in their respective language only. Participants were instructed to press a button corresponding to the side on which a picture of the noun appeared (e.g., for sentence 3a, a picture of a dog or a cow). In general, Hern~indez et al. (1994) found that English monolinguals were faster in sentence interpretation when sentences conformed to the SVO or NVN word order, followed by verb agreement and animacy. In other words, English monolinguals were faster to decide who was doing the action when the sentences followed the English SVO canonical word order. For the Spanish monolinguals, on the other hand, noun verb agreement produced the fastest sentence interpretation reaction times, followed by animacy and word order. More important, a comparison between bilinguals and English monolinguals showed remarkable similarities. B ilinguals, like English monolinguals were very sensitive to word order,
10
Roberto R. Heredia and Mark T. Stewart
and somewhat sensitive to verb agreement. A comparison between bilinguals and Spanish monolinguals showed that both groups were highly sensitive to verb agreement. However, unlike Spanish monolinguals, bilinguals were affected by word order. That is, bilinguals were faster in sentence interpretation than Spanish monolinguals when the sentences followed the SVO word order. In short, SpanishEnglish bilinguals and English monolinguals exhibited similar sentence processing strategies during the comprehension of English sentences. Both groups, in this case, were sensitive to word order and verb agreement, respectively. Although bilinguals showed similar processing strategies as Spanish monolinguals during the comprehension of Spanish sentences, unlike monolinguals, bilinguals were actually more sensitive to word order effects as opposed to animacy. These results suggest that at least for the bilinguals in this experiment, the second language (L2) influenced the processing of the first language (L1), in this case, the comprehension of Spanish sentences (see Liu, Bates, & Li, 1992 for similar results). This finding contrasts with Kilborn's (1989) results in which the participant's first language (German) influenced the processing strategies of the second language (English). This difference, we believe, may be due to language dominance. It is possible that Kilborn's (1989) bilinguals were more dominant in their first language and their sentence processing strategies reflected this dominance. In contrast, the bilinguals in Hern~indez et al.'s experiment were more dominant in their second language (cf. Heredia, 1997; Heredia & Altarriba, 2001), and this dominance is reflected by the influence of the second language in the processing of Spanish sentences. One important aspect to note here is that the bilinguals described here were to some extent more English-monolingual-like, than Spanish-monolingual-like in their sentence processing strategies. In addition to demonstrating the reliability of this task (for a review see Bates, Devescovi, & Wulfeck, 2001), this study underscores the importance of examining the interaction of the bilingual's two languages during the comprehension of language. The results reviewed here, in conjunction with findings from other bilingual language groups such as German-English (Kilborn, 1989), Chinese-English and English-Chinese (Liu et al., 1992), Hindi-English (Vaid & Pandit, 1991), Dutch-English and EnglishDutch (McDonald, 1987b), and English-French (McDonald & Heilenman, 1991) suggest that language processing by bilinguals can be influenced by the structural constraint of one or both of their languages, depending on such factors as language proficiency and language dominance. In relation to bilingual sentence processing, the usefulness of this task lies in its reliance on spoken language. As shown by Liu et al. and others (e.g., Kilborn, 1989) this task can be useful in studying the development of bilingualism and the various sentence processing strategies utilized throughout the second-language learning process. More important, this task can be used to study second-language developmental issues with children (A. E. Hern~indez, personal communication, May 12,2001; see also McDonald, 1986). Furthermore, this task may be suitable to examine the effects of gender in bilingual language processing (cf. Sera, Berge, & del Castillo Pintado, 1994). For example, in Spanish (and other Romance Languages) all objects are categorized in terms of "masculine" or "feminine." For instance, the word for flower in Spanish Oqor) is categorized as "female." Thus, a
On-line Methods in Bilingual Spoken Language
11
question of interest would be to examine whether or not Spanish-English bilinguals assign or transfer gender from Spanish to English objects. Anecdotal evidence suggests that when bilinguals code-switch between Spanish and English, they engage in this type of categorization. For example, in sentence (4a) below, peanuts take the masculine unos because cacahuates or manis are masculine in Spanish. In contrast, in sentence (4b) peanut butter takes the feminine la because crema de cacahuate is feminine for Spanish speakers. A possible experiment might involve English sentences such as I want peanuts in which participants determine if the noun peanuts is masculine or feminine (cf. Guillelmon & Grosjean, 2001). (4a) Quiero unos peanuts (I would like peanuts) (4b) Quiero la peanut butter (I want peanut butter) One prediction would be that if gender transfer were to take place, Spanish bilinguals would be faster (in milliseconds) deciding or judging that peanuts are masculine rather than feminine. Notice that it would also be possible to analyze responses in terms of choice data percentages (whether participants choose "masculine" or "feminine"). Of interest in this experiment would be a potential statistical interaction between gender and animacy, where things that are animate would be more likely to be given conceptual gender during language transfer.
The Phoneme-Triggered Lexicai Decision Task
The next methodology we consider is the phoneme triggered lexical decision task (PTLD). The PTLD has been used to study lexical access and semantic facilitation during sentence processing. This task is a combination of phoneme-monitoring (e.g., Connine & Titone, 1996; Foss, 1970) and lexical decision. In this task, listeners make a word or nonword judgment for a particular target occurring within the sentence. The target word begins with the sound specified for each particular sentence. The phoneme is used to trigger the lexical decision. For example, participants listening to sentence (5a) below may be asked to listen for a word or nonword that begins with the phoneme /k/. Once that target is identified, participants decide as fast and as accurately as possible whether the target is a word or a nonword. In this case, the participant identifying the phoneme/k/for the target word cake in sentence (5a) would respond that it is a word. Target phonemes for each sentence may be provided auditorily before the presentation of each sentence (Blank, 1980) or on a sheet of paper specifying the target phoneme and the sentence number. Response times are measured from the onset of the target phoneme until a response is recorded.
(5a) After lunch, the children asked for a piece of CAKE for dessert (5b) Depois do almogo os miudos pediram uma fatia de BOLO para sobremesa (5c) Depois do LUNCH os miudos pediram uma fatia de CAKE para DESSERT
12
Roberto R. Heredia and Mark T. Stewart
Soares and Grosjean (1984) utilized this task to examine lexical access between Portuguese-English bilinguals in comprehending Portuguese and English sentences, and English monolinguals listening to English sentences. Their objective was to address findings by Kolers (1966) and Macnamara and Kushnir (1971) suggesting that bilinguals were significantly slower to retrieve information from mixed-language sentences such as the Portuguese-English code-switched sentence in (5c) as opposed to monolingual sentences in English (5a) and Portuguese (5b). At issue was whether bilinguals would be faster at making lexical decisions to sentences (5a-5b) than to (5c). Another objective was to determine differences (if any) between English monolinguals and Portuguese-English bilinguals listening to English sentences. The results showed that when both monolinguals and bilinguals listened to English sentences, lexical decisions to real word targets were comparable for both groups. However, bilinguals were slower to respond to nonwords. When bilinguals listened to English and Portuguese monolingual sentences (5a and 5b) and code-switched sentences (5c), results showed no differences in the two monolingual sentence conditions. However, lexical decision times to word targets in the code-switched condition were slower, compared to both monolingual conditions. Lexical decisions by bilinguals for nonwords in the three language conditions were identical. In sum, Soares and Grosjean's results replicated the general finding that bilinguals are typically slower at retrieving information from one language, especially when this information is embedded in a second language, than accessing information in only one language (e.g., Kolers, 1966; Macnamara & Kushnir, 1971; see also Altarriba et al., 1996; Heredia & Altarriba, 2001). Their results also show that lexical access between monolinguals and bilinguals during sentence processing was quite similar. Lexical decisions to real words in English, while listening to English sentences, were identical for both groups. However, Portuguese-English bilinguals took longer to respond to English nonword targets, while listening to English monolingual sentences. As argued by Soares and Grosjean, it is possible that bilinguals took longer to respond to nonword targets because they had to search two mental lexicons instead of one. Thus, if in the initial search, while listening to English sentences and searching the English lexicon the target is not encountered, the bilingual defaults to the other language lexicons, in this case Portuguese, to identify the target. Although, this is a theoretically viable explanation, it is not clear as to why this same difference is not reflected in the real word data. Alternatively, this difference in the nonword data could be due to the manner in which the nonword targets were created and possible task demands. Perhaps, bilinguals were quick to respond to real words because these targets were clearly marked as belonging to English (cf. Grosjean, 1988; Li, 1996a). Nonwords, on the other hand, may have been too phonologically similar to English and Portuguese, thus making it difficult to decide whether the nonword was more like English or Portuguese, or both. This ambiguity in conjunction with the sensitivity of the PTLD to the sound unit (i.e., phoneme) may have caused the bilinguals to activate their second language lexicon. Indeed, this methodological issue cannot be rejected as a possible explanation. Overall, the PTLD appears sensitive to differences between lexical access in the monolingual (e.g., English or Portuguese) and bilingual (code-switched) mode. Given its reliance on the sound unit to trigger the lexical search during sentence
On-line Methods in Bilingual Spoken Language
13
comprehension, this task is ideal for studying issues related to how bilinguals process homophones (i.e., words that share phonological similarities, but whose meaning is different) across languages (cf. Li & Yip, 1998). For example, how do Spanish-English bilinguals process ambiguous sentences such as (6a-6b)? The English target word YELLOW is a homophone of the Spanish word HIELO which is related to ice. What are the effects of context and the interpretation of cross-language homophones? To what extent does language in conjunction with semantic context determine bilingual lexical search (6b)? That is, during the comprehension of sentence (6a), are bilinguals more likely to engage in exhaustive search in which both the English and Spanish lexicons become simultaneously activated? Is the preceding context of sentence (6b) more likely to re-direct the lexical search to the English lexicon only? In this example, we are assuming that participants would be in a bilingual mode. i (6a) She liked YELLOW because it was a pleasant color (6b) She liked [blue more than] YELLOW because it was a pleasant color An important aspect of this task is that it could be also employed simply as a phoneme-monitoring task (Connine & Titone, 1996) in which bilinguals would be asked to respond when they find the/y/sound. However, researchers utilizing this task should pay careful attention to the possibility that this task may be too sensitive to phonotactic structural factors (i.e., sound pattems) of a particular language (see discussion on the gating and cued shadowing sections for more information about the interaction of sound patterns and language). Clearly, empirical studies utilizing this task, both at the monolingual and bilingual domains are limited. More research is needed to determine the usefulness of this task in the field of bilingual research.
The Cross-Modal Naming Task
A task similar to the PTLD is the cross-modal naming task, used by Hern~indez et al. (1996, and this volume) and Hern~indez, Fennema-Notestine, and Udell (in press). It has been argued that this task is sensitive to both sentential and lexical priming. Participants in this task are presented with an auditory sentence such as (7a). At a predetermined point within the sentence, the sentence stops momentarily and a visual target appears in the middle of the computer screen. The participants' objective is.to name, as quickly and accurately as possible, the visually presented related or unrelated target word. For example, in sentence (7a) and (7b) the sentence stops at the offset of Actually, given the appropriate comparisons in which both English and Spanish sentences are utilized, we would predict little if any activation (as measured by lexical priming) of the Spanish meaning (e.g., ice) while listening to English sentences (e.g., yellow), and activation of the English meaning (e.g., color) while listening to Spanish sentences, in situations in which English is the bilingual's dominant language regardless of which language was learned first (for further details about the importance of language dominance see Heredia & Altarriba, 2001).
14
Roberto R. Heredia and Mark T. Stewart
the definite article the (depicted by "[**1]"), and the participant names the visual English street or Spanish calle target. Note that it would be possible to have participants do lexical decisions on the visually presented target. (7a) John was walking down thet**~1STREET when he remembered his appointment (7b) John was walking down thet**~1CALLE when he remembered his appointment (8a) Juan caminaba por lat,,~ 1CALLE cuando record6 su cita (8b) Juan caminaba por Iat,,~ISTREET cuando record6 su cita Hem~indez et al. (1996) investigated the effects of expectancy and predictability in cross-language priming. In cross-language priming, naming a word in an L2 (peace) is faster when preceded by a related word in the L 1 (guerra: war), than an unrelated word in the L1 (boteHa: bottle). Similarly, it is faster to name a word in the L 1 (guerra) when followed by a related word in the L2 (war) than an unrelated word in the L2 (bottle) (Altarriba, 1992). Hem~indez et al's. purpose was to determine the optimal conditions in which cross-language priming would be more likely to occur. Accordingly, during the comprehension of a sentence, for cross-language priming to occur, the bilingual must anticipate that a word in the second language will be presented in order to tune themselves into a "bilingual mode" and be able to access the lexical item in the second language (cf. Amrhein, 1999). Because cross-language priming may be strategic in nature, ifbilinguals are unable to tune themselves into this mode, cross-language priming would not occur (Hem~indez et al., 1996). Hern~indez et al. had Spanish-English bilinguals listen to sentences in English (7a) and Spanish (8a) and name visually presented words in both languages. For the English-Spanish bilingual condition, participants listened to sentences such as (7b) and then named a Spanish visual target. Participants in the Spanish-English condition listened to sentences such as (8b) and then named the visually presented target in English. All language conditions were either blocked (the predictable condition) or mixed (the unpredictable condition). To summarize, cross-language priming was obtained, but only when language presentation was blocked. That is, sentences (7b) and (8b) produced priming only when bilinguals were able to predict that a language switch would occur. When these bilingual sentences were presented in a mixed design, naming a related word in the second language, was no different than naming an unrelated target. However, when a delayed naming task was used, cross-language priming was obtained for both bilingual conditions in both blocked and mixed experimental conditions. However, sentences (7a, English sentence and English probe) and (8a, Spanish sentence and Spanish probe) exhibited priming regardless of whether they were presented blocked or mixed. In short, this pattern of results led Hern~indez et al. (1996) to conclude that "crosslanguage priming appears only when participants know what language to expect, when they have ample time to generate a response, or both" (p. 860). That is, switching from one language to the other takes time and access to a second language cannot occur
On-line Methods in Bilingual Spoken Language
15
unless the bilingual is in some type of "bilingual mode" (see also Hern~indez this volume). Two potential concerns with this task warrant further discussion. First, researchers choosing to use this task would do well to consider the point at which the visual target is presented during the spoken sentence. For example, notice that in (7a) above the sentence stops at the offset of the definite article the. This disruption and the importance of the definite article may give rise to what Grosjean and colleagues refer to as the "base-language effect" (Grosjean, 1988, 1997; L6wy & Grosjean, 2001; Li, 1996a). This effect refers to the bias toward the language of the sentence in which the listener expects or is primed for the next language item to be in the same language (Li, 1996a, p. 767). Bilinguals listening to sentence (7a) might immediately trigger or predict the correct word in English that is consistent with the preceding context and the grammatical structure of English. If true, when presented with sentence (7b), the bilingual must attempt to inhibit or develop a strategy to overcome the interference produced by the activation of the word that is consistent with the preceding language and contextual information. The general idea here is that the English word is more readily available, and to inhibit it would require extra time. If language presentation is blocked (the predictable condition) or response is delayed (delayed naming), there is sufficient time to bring under control this interference to successfully integrate the Spanish target word into the sentence (cf. Scarborough, Gerard, & Cortese, 1984). However, if no time is allowed to overcome this interference (the unpredictable condition), priming does not occur and the Spanish target word is not integrated and perhaps it is treated as a nonrelated word. In other words, the combination of the task processing requirements, the biasing context and the strong linguistic cues in effect mandates that the bilingual generate the lexical item most consistent with the preceding context and the linguistic structure of the sentence. The same reasoning applies to sentences (8a) and (8b). However, it could be argued that sentence (8b) would be more semantically felicitous at least for Spanish-English bilinguals who mix their languages during the communicative process. A second issue that applies to this paradigm and other on-line tasks concerns the "grammaticality" of the code-switched word. Linguistic evidence suggests that code switching follows a grammatical structure and that certain grammatical categories cannot be code-switched (Lederberg & Morales, 1985; Myers-Scotton, 1997). K. Pletsch de Garcia, (personal communication, December 5, 2000) suggests that care should be taken in constructing code-switched sentences because of the possibility of constructing unnatural linguistic groupings. Pletsch de Garcia argues that during codeswitching the natural tendency is not to break up linguistic categories such as noun phrases (e.g., the traffic) or infinitival phrases (e.g., to drive). It would be unlikely for a bilingual to code-switch the noun phrase the traffic into the trffico or the infinitive verb to drive into to manejar (cf. Lederberg & Morales, 1985). Inspection ofHern~indez et al's. stimuli suggests inconsistencies in relation to the position of the code-switched word and the preceding linguistic category. For example, in some instances the visually presented code-switched target is inserted following a word whose associate is highly related (e.g., traffic-LUZ), after a conjunction (e.g., when-TRAFICO) or breaking up an infinitival clause (e.g., to MANE JAR). Thus, the conclusion that cross-language
16
Roberto R. Heredia and Mark T. Stewart
priming is strategic and that it involves the use of expectations, deserves further consideration. Despite these concerns, the cross-modal naming task as implemented by Hem~ndez et al. (1996) has generally produced reliable cross-language priming effects similar to those reported in the bilingual associative priming literature (cf. Grainger & Beauvillain, 1988; see also, Keatley & de Gelder, 1992). Moreover, this task may prove useful in examining the effects of context on the comprehension of language in both bilingual (sentences 7b and 8b) and monolingual (sentences 7a and 8a; cf. Hem~ndez et al., in press) modes (for other possible manipulations such as visual degradation and speed naming, using this task, the reader is referred to the original paper).
The Gating Task The next task we will discuss is the gating task (Grosjean, 1980; 1988; Li, 1996a, 1996b; Li & Yip, 1998; Tyler, 1992). Participants in this task listen to a spoken word from left to right in segments of increasing duration, until the entire word has been presented. The first segment starts from the beginning of the word and has a duration of about 30-50 milliseconds (ms) and each successive gate may increase by about 30-50 ms and the last gate may correspond to the entire word. At each gate presentation, participants are required to identify or guess the most likely candidate based on the information being presented up to that particular gate. For example, in the Chinese sentence (9a) below, the first gate of the critical word FLIGHT may be Keoi ge FL. At this point, participants may generate such possible candidates as fiat, flood, and flop until they come up with the critical target or until they hear the entire word. These candidates are meaningful because they provide information into the underlying cognitive processes leading to the final identification of the critical target (Li, 1996a, p. 765). The general idea behind this task is to determine the amount (in milliseconds) or percentage of exposure needed to correctly identify the critical word. If a bilingual recognizes the critical target FLIGHT only after hearing the whole word, it can be said that 100% of acoustic information was needed to correctly identify the target. Research utilizing this task suggests that listeners recognize words very rapidly. Bilinguals and monolinguals need to hear about 200-250 ms of the beginning of a word in a constraining context and much more time in a neutral context in order to correctly identify it (Li 1996a; Marslen-Wilson, 1987). (9a) Keoi ge FLIGHTjin-ci (his/her FLIGHT is delayed) (9b) Keoi daap baan jin-ci ge FLIGHT (He/she boarded a delayed FLIGHT) Li (1996a) employed this task to examine the effects of three important psycholinguistic factors in the recognition of English targets embedded in Chinese sentences by Chinese-English bilinguals. Li's overall purpose was to replicate and extend earlier findings by Grosjean (1988). The first factor was a phonological variable concemed with the phonotactics or permissible sound sequences of both Chinese and
On-line Methods in Bilingual Spoken Language
17
English. For example, the English language allows both consonant-consonant (CC) and consonant-vowel (CV) clusters at the beginning of a word (e.g., BRide vs. Lion). Chinese, in contrast, allows CV clusters but lacks CC clusters. The objective of this manipulation was to determine the extent to which CC clusters, which are marked phonotactically as belonging to English, would be identified faster than CV clusters that are shared by both languages. In other words, because of its distinctiveness, lexical search of the target with the CC configuration would be directed specifically to the English lexicon without involving the Chinese lexicon. On the other hand, because both languages shared the CV cluster configuration, during lexical access, both lexicons would be searched to correctly identify the target word. The second factor concerned the phonetics or the manner in which the English critical target was pronounced. Li distinguished between "code-switchers" in which the critical target was pronounced using the appropriate English pronunciation, and "borrowers" in which the English target word was pronounced using the Chinese pronunciation. Code-switchers were predicted to be identified faster and to require less acoustic information because they are clearly English words. Borrowers would involve a search of both lexicons, and therefore would require more acoustic information during the identification process. The third factor examined the effects of context (prior context vs. neutral context) on the identification of code-switched words. To summarize, phonotactics made a difference in the amount of information needed to correctly identify the code-switched target. Contrary to the predictions, bilinguals needed more acoustic information to correctly identify English targets with CC than CV initial cluster configurations. More important, phonotactics interacted with phonetics. For the CC initial word configuration condition, borrowers needed more acoustic information of the critical target to correctly identify it. Code-switchers, on the other hand, needed significantly less information for identification. However, in the CV cluster configuration, both borrowers and code-switchers required the same amount of information to be correctly identified. Although context did not interact with the other two factors, the general trend was that critical targets in the prior context condition were recognized faster than targets in the neutral condition. In general, the results of this study are significant and underscore the importance of considering the influence of low-level linguistic information such as sound sequences in bilingual lexical access. Li's (1996a) results suggest that during lexical access, information that is shared by both Chinese and English (i.e., CV cluster configurations) is simultaneously activated and has priority over information that is not shared by the two languages (i.e., CC initial cluster configuration). Thus, it appears that the language preceding the code-switched target item determines which phonotactic configuration is to be activated (see also Grosjean's 1988). However, during the course of sentence processing, information that overlaps across languages has priority and is activated simultaneously. As a methodological tool, the gating paradigm is suitable to study ecologically valid issues that reflect real life situations such as the understanding of a "foreign accent," which is equivalent to what Grosjean (1988) refers to as "borrowing," and Li (1996a) as "borrowers." Another important aspect of this task is that it allows for the close examination of the bilingual's two languages and enables one to determine how each
18
Roberto R. Heredia and Mark T. Stewart
language's unique linguistic structure influences retrieval during the comprehension of code-switched information. The findings by Grosjean and colleagues are noteworthy when we consider that traditionally, these factors have been ignored in the bilingual literature. Moreover, research in the sentence-processing domain would benefit from future studies that systematically compare and contrast such tasks as the PTLD, the gating task, and the cross-modal naming technique described in the previous section. The first steps have been taken, and Li (1996a) has already shown remarkable similarities between the gating task and the cued shadowing task, to which we now turn.
The Cued Shadowing Task
The next task that we discuss is the cued shadowing task (Bates & Liu, 1996; Liu, Bates, Powell, & Wulfeck, 1997). It has been argued that this task is sensitive to lexical access, and has been used successfully in the monolingual literature to study the recognition of open- versus closed-class words (e.g., Herron & Bates, 1997), and code switching in bilingualism (Li, 1996a). Participants, in this task, are presented with an auditory sentence such as (9a). The target or code-switched word is recorded in a voice of the opposite sex from the voice in which the sentence was recorded. However, the code-switched word could be of the same voice as the rest of the sentence. The participant's objective is to repeat the target word aloud as quickly and as accurately as possible. Responses are recorded in milliseconds. Li (1996a) used this task to replicate and examine the generalizability of the gating task to other tasks. To summarize, the results of this task were remarkably similar to those obtained by the gating task. The cued shadowing technique proved to be sensitive to the effects of both phonotactic and phonetic factors. For example, bilinguals required significantly longer times to process English targets containing the CC cluster configuration than CV English targets. Again, this finding contrasted with Grosjean's (1988) results that found the opposite patterns. Although this task has replicated well-known psycholinguistic findings such as semantic priming and frequency effects, closer scrutiny is needed to determine whether the repeating or the shadowing of the code-switched word is performed without the activation of meaning (Bates & Liu, 1996). Moreover, researchers utilizing this task need to consider how the target word is to be presented. Will the target be recorded by the same or different gender voice? If recorded by the same voice, how are participants going to know which word from the sentence to repeat? In Li's (1996a) experiment, participants were told at the beginning of each block the specific location of the codeswitched word. For example, Chinese-English bilinguals were told that in certain blocks the code-switched word would always appear immediately after the phrase keoi g e _ in sentence (9a). Clearly, this issue calls for careful consideration. As an alternative to this potential methodological artifact, we propose a variant of this task in which regardless of voice shift, bilinguals are asked to listen for a target word that begins with a particular phoneme such as/f/for the target word in sentence (9a) above. Bilinguals can either press a computer key when the target sound is found, or repeat the word
On-line Methods in Bilingual Spoken Language
19
aloud (cf. Connine & Titone, 1996; see also Blank, 1980). Overall because the shadowing task does not require a high degree of metalinguistic knowledge, this task could be used to study language-processing issues with bilingual children, the elderly and with other bilingual populations that are unable to read and write in one or both of their two languages.
The Cross-Modal Lexical Priming (CMLP) Task The CMLP differs from the cross-modal naming task used by Hemfindez et al. (1996) in one critical way. During sentence presentation, the flow of the sentence is never interrupted, thus making it difficult for participants to engage in strategic processing (cf. Bates & Liu, 1996; Li, 1996a). This task has been widely used in monolingual studies and its sensitivity to semantic and associative relations, as well as contextual effects is well documented (e.g., Swinney, 1979; Swinney, Onifer, Prather, & Hirshkowitz, 1979; Tabossi, 1996; see also Ferreira & Anes, 1994). The procedure has been used repeatedly to demonstrate subject's abilities in resolving lexical ambiguities and multiple access during the comprehension of ambiguous words (e.g., Tabossi, 1996; but see Glucksberg, Kreuz, & Rho, 1986; Mckoon & Ratcliff, 1994; Prather, Swinney, 1988), syntactic related issues (e.g., Nicol & Swinney, 1989; Love & Swinney, 1996), figurative language (e.g., Blasko & Connine, 1993; Cacciari & Tabossi, 1988; Stewart & Heredia, in press; Titone & Connine, 1994) and stereotype processing (Heredia & Blumentritt, 1999,2001). Typically, participants are asked to listen to a sentence while watching a fixation point on a computer screen. At some point during presentation of the spoken stimulus (usually immediately following an ambiguous word), a test word replaces the fixation point and participants make a timed response to the item. The participants' task is to either make a lexical decision or name a critical target. Reaction times to test words and their matched controls are then compared to determine whether facilitatory priming exists. Li and Yip (1998) employed this task to examine the processing of Chinese-English homophones (words that share phonological similarities across languages, but whose meaning is different). Chinese-English bilinguals listened to sentences whose prior context was neutral or biased towards the meaning of a homophone embedded within a Chinese sentence. Homophones in the Chinese sentences were actual English words (e.g., lock) but pronounced in Chinese phonetics. The bilingual's objective was to name a visually presented English target that was the same word as the homophone (e.g., lock), and a Chinese target word that shared the similar phonological structure as the homophone in the sentence (e.g., lok). Additionally, bilinguals named two unrelated targets that shared no phonological similarities with the critical homophone in the sentence. The most significant result of Li and Yip's study was the interaction of homophony and language. Bilinguals were faster to name English targets (e.g., lock) than Chinese targets (e.g., lok). This facilitatory effect for the English visual target was surprising given that the auditory sentences were all in Chinese. Thus, it appeared as if the prior context was actually favoring the recognition of the English target and suppressing the
20
Roberto R. Heredia and Mark T. Stewart
recognition of the Chinese related target. However, this interpretation is only suggestive because context failed to interact with language and homophony. In fact, there was only a main effect of context, suggesting that prior context facilitated both language conditions equally. Critical to the results was the comparison between the unrelated controls and the related targets; however, this comparison was inappropriate because the unrelated targets were not appropriately controlled for frequency and word length relative to the related targets. Nonetheless, the possibility that semantic context can override language context is interesting and deserves further consideration. Overall, future work addressing this issue of bilingual lexical ambiguity and homophony would benefit from employing a more traditional priming paradigm. It would be possible to examine whether Chinese-English bilinguals activate one or both meanings of the Chinese-English homophone (e.g., lok vs. lock). For example, as bilinguals listen to the sentence and at the onset of the homophone, they name a target that is related to the English meaning of the homophone (e.g., key) and an appropriately controlled unrelated target (e.g., ice). If the English meaning of the homophone becomes activated, naming times for related targets would be faster (e.g., 600 ms) than response times for the unrelated controls (e.g., 740ms). In this case, one would say that there is a 140 ms priming effect. The logic would also apply for the other language condition. This task has also been used to examine the effects of context and lexical access across-languages (Heredia, 1998, 2000; Heredia & Stewart, 1998). Currently, we are employing the CMLP to address some of the assumptions of the revised hierarchical model of bilingual memory representation (Kroll & Stewart, 1994; cf. Hern~ndez this volume). This model suggests that bilingual lexical access is differentially affected by such factors as contextual and semantic information depending on whether bilinguals retrieve information from the first to the second language or vice versa. According to this model, access to the second-language from the first language (L l-L2) is sensitive to semantic or contextual information, whereas access to the first language from the second language (L2-L 1) is relatively insensitive to these factors (Kroll & Sholl, 1992; Kroll & Stewart, 1994). We believe that these assumptions can be addressed directly by systematically manipulating the contextual information prior to the critical prime within a sentence. We are studying this possibility in our ongoing research. For example, in one experiment, Spanish-English bilinguals listen to Spanish translations of contextually-unbiased sentences such as (10a) below. Notice that the information preceding the critical target (WAR) does bias the meaning of the critical prime. A second experiment involves sentences such as (10b). Notice that the preceding context of this sentence soldiers + combat is biased towards the meaning of the critical prime WAR. Participants in these two experiments are asked to name visually presented English or Spanish target words that are either related (peace, paz) or unrelated (light, luz) to the critical word. To be named targets are presented immediately after the critical prime denoted by subscript [* 1]. If the revised hierarchical model is correct and L l-L2 language directions are sensitive to contextual effects, cross-language priming should significantly increase from sentence (10a) to sentence (10b).
On-line Methods in Bilingual Spoken Language
21
(10a) It is difficult to admit that a WARt,~1sometimes brings more profits than losses (10b) Soldiers are trained for combat and WAR t*ll and that is why so much is invested in them Furthermore, using this task, our laboratory has been able to replicate the general observation that bilinguals in South Texas, a predominantly bilingual area, experience more interference from their second language (English) as they communicate in their first language (Spanish), and little or no interference from Spanish when they communicate in English. Our results using sentences such as (10a) in both Spanish and English show that bilinguals are generally faster in naming English target words while listening to Spanish sentences than to name Spanish words while listening to English sentences. That is, participants are faster at retrieving English words even when they are listening to Spanish sentences. When bilinguals listen to English sentences, as expected, English words are retrieved faster than Spanish words. More important is the finding that bilinguals are faster at naming English target words while listening to Spanish sentences than to name Spanish target words while listening to English sentences. Indeed, these results suggest a reliance on the second language rather than the first language. As argued by Heredia and Altarriba (2001), these results suggest that after a certain level of fluency and frequency of use is attained in a second language, a language shift occurs and the second language behaves as if it were the bilingual's first language. In other words, the second language becomes more readily accessible than the first language, and the bilingual comes to rely more on it. This language shift, as argued by Heredia and Altarriba, is due to frequency of language usage. Regardless of which language a bilingual leams first, the more active (dominant) language determines which lexicon is accessed faster (Heredia, 1997). Thus, unlike some theoretical accounts, bilingual lexical representation is not a static but a dynamic representational system in which the first language can fall in strength, while the second language becomes the dominant language. In short, the CMLP appears to be sensitive to measuring issues related to how bilinguals access information from their two language lexicons. More important, whether the critical prime (e.g., WAR) is in the same or different language, or it is recorded by the same or different voice is not an issue with this task because participants are responding to a visual target. However, empirical work should be carried out in which findings from other tasks (e.g., the PTLD and the cued shadowing technique) are systematically compared to the CMLP. In extending Li and Yip's (1998) original work, future work would benefit from using bilingualism as a tool to address issues related to multiple lexical access and ambiguity resolution. Cross-language homophones and homographs are the natural candidates to study these general psycholinguistic issues. Additional work on the examination of other psycholinguistic issues such as the comprehension and resolution of nonliteral language (e.g., Matlock & Heredia, this volume) will further our understanding of the interaction between the bilingual's two linguistic systems.
22
Roberto R. Heredia and Mark T. Stewart
The Auditory Moving-Window (AMW) The last task that we will discuss is the AMW. This task has been used to examine lexical frequency and syntactic complexity in monolinguals (e.g., Ferreira, Henderson, et al., 1996), the effects of prosody and language processing (e.g., Ferreira, Anes,& Horine, 1996), and most recently used to study the effects of aging and discourse processing (e.g., Titone, Prentice, & Wingfield, 2000), and hemispheric processing (e.g., Titone, Wingfield, Caplan, Waters, & Prentice, 2001). In this task, participants listen to sentences one or two words at a time and are required to press a button to receive successive segments as in sentence (11) (segments are depicted with the symbol "$"). Times between button presses or Interresponse times (IRT's) are recorded. Analysis can be performed on the IRT's or Difference times (DT's). DT's are computed by subtracting the IRT minus the duration of the segment during the digitizing of the word. For example, if the IRT for the target T E A C H E R in sentence (11) is 802 ms, and the time taken to record the target, (duration time), is 493 ms, then the DT would be 309 ms (Heredia, Stewart, & Cregut, 1997) (11) ,l,Erika ,l,estuvo,l, buscando a la,l,TEACHER,~ pero nunca,l, la encontr6,1,. (Translation: Erika was looking for the teacher but she never found her) Work underway in our laboratory employs the AMW task to investigate the effects of context and word frequency in the processing of code-switched sentences (e.g., Heredia et al., 1997). So far, our efforts have been successful and our results suggest interesting patterns depending on whether the target word is a code-switcher (pronouncing the target word with English phonetics) or a borrower (pronouncing the target word with Spanish pronunciation). For example, preliminary findings suggest that code-switchers are relatively unaffected by prior context. However, borrowers are actually processed faster under low-constraint than under high-constraint contextual conditions. Similarly, word frequency seems to affect borrowers the most. That is, high frequency borrowers seem to be processed faster than low frequency ones. In short, this task is an important altemative to some of the tasks discussed in this chapter, primarily because it is self-paced, and non invasive. That is, the AMW does not require the participant to produce a verbal or a decision response. This option is a valuable one, especially if the particular bilingual group of interest lacks academic training in one or both languages. However, researchers utilizing this task must assure that participants receive enough practice trials to familiarize themselves with the task. In addition to measuring the processing of a single target, like most of the tasks that we have reviewed, this task may be used to study the processing of linguistic units beyond the word level. For example, this task would be appropriate to study the comprehension of phrasal verbs such as turn on among bilinguals at different proficiency levels of the second language. Phrasal verbs represent a challenge to most speakers whose second language is English (for further details see, Matlock & Heredia, this volume). To illustrate, consider sentences (12a) and (12b).
On-line Methods in Bilingual Spoken Language
23
(12a) ,[,Mary,[, turned on,[,Colombo street,[, (12b) ,[,Mary,[, turned on,[, Colombo,[, In sentence (12a), turned on could be understood in terms of a literal interpretation in which it denotes some sort of physical motion and change in direction. Sentence (12b), on the other hand, may be interpreted in a nonliteral sense in which turned on may suggest that Colombo is sexually aroused by Mary (for further details, see Matlock & Heredia, this volume). Further experimental manipulations may involve biasing sentence (12a) with information congruent with the literal interpretation of the phrasal verb (e.g., Driving a new car). In this case, the preceding context would trigger the literal interpretation of the phrasal verb. The general idea in this possible experiment would be to measure the processing time for the phrasal verb under biasing and nonbiasing contextual conditions.
Final Thoughts In this chapter we have reviewed several of the experimental tasks used by researchers in bilingualism to examine spoken sentence comprehension. Our primary goal has been to describe each task and, where appropriate, to provide some insight into the usefulness of each of these procedures. We feel it is important that prospective users have a clear understanding of the potential strengths and weaknesses a particular paradigm may have. The final decision of which paradigm to use is left to the researcher and will be driven largely by the particulars of the empirical and theoretical issues under investigation. Moreover, there are technological constraints to consider; some tasks are more difficult to implement than others. Also, not all of these tasks are suitable for addressing the full range of empirical questions faced by researchers in bilingualism. Indeed, some are more focused than others and designed to examine very specific questions in greater detail. Regardless, our hope is that interested researchers will take from this chapter an understanding of the kinds of methodological issues and practices currently being developed for use in examining bilingual sentence comprehension, in the spoken domain.
References Altarriba, J. (1992). The representation of translation equivalents in bilingual memory. In R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 157-174). Amsterdam, Netherlands: Elsevier. Altarriba, J., Kroll, J. F., Sholl, A., & Rayner, K. (1996). The influence of lexical and conceptual constraints on reading mixed-language sentences: Evidence from eye fixations and naming times. Memory & Cognition, 24, 477-492. Amrhein, P. C. (1999). On the functional equivalence of monolinguals and bilinguals in "monolingual mode"" The bilingual anticipation effect in picture-word processing. Psychological Science, 10, 230-236.
24
Roberto R. Heredia and Mark T. Stewart
Bates, E., & Liu, H. (1996). Cued shadowing. Language and Cognitive Processes, 11, 577-581. Bates, E., Devescovi, A., & Wulfeck, B. (2001). Psycholinguistics: A crosslanguage perspective. Annual Review of Psychology, 52, 369-396. Blank, M. A. (1980). Measuring lexical access during sentence processing. Perception & Psychophysics, 28, 1-8. Blasko, D. G., & Connine, C. M. (1993). Effects of familiarity and aptness on metaphor processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 19, 295-308. Connine, C. M., & Titone, C. (1996). Phoneme monitoring. Language and Cognitive Processes, 11, 635-645. Cacciari, C., & Tabossi, P. (1988). The comprehension of idioms. Journal of Memory and Language, 27, 668-683. de Groot, A. M. B., Dannenburg, L., & van Hell, J. G. (1994). Forward and backward word translation by bilinguals. Journal of Memory and Language, 33, 600629. Durguno~;lu, A. Y., & Roediger, H. L. (1987). Test differences in accessing bilingual memory. Journal of Memory and Language, 26, 377-391. Favreau, M., & Segalowitz, N. (1982). Second language reading in fluent bilinguals. Applied Psycholinguistics, 3, 329-341. Favreau, M., & Segalowitz, N. (1983). Automatic and controlled processes in reading a second language. Memory & Cognition, 11, 565-574. Ferreira, F., & Anes, M. (1994). Why study spoken language? In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 33-56). San Diego, CA: Academic Press. Ferreira, F., Anes, M. D., & Horine, M. D. (1996). Exploring the use of prosody during language comprehension using the auditory moving window technique. Journal of Psycholinguistic Research, 25, 273-290. Ferreira, F., Henderson, J. M., Anes, M. D., Weeks, P. A., & McFarlane, D. K. (1996). Effects of lexical frequency and syntactic complexity in spoken-language comprehension: Evidence from the auditory moving-window technique. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 324-335. Foss, D. (1970). Some effects of ambiguity upon sentence comprehension. Journal of Verbal Learning and Verbal Behavior, 9, 699-706. Glanzer, M., & Duarte, A. (1971). Repetition, between and within language in free recall. Journal of Verbal Learning and Verbal Behavior, 1O, 625-630. Grainger, J., & Beauvillain, C. (1988). Associative priming in bilinguals" Some limits of interlingual facilitation effects. Canadian Journal of Psychology, 42, 261-273. Grosjean, F. (1980). Spoken word recognition processes in the g~.ting paradigm. Perception & Psychophysics, 28, 267-283. Grosjean, F. (1988). Exploring the recognition of guest words in bilingual speech. Language and Cognitive Processes, 3, 233-274. Grosjean, F. (1996). Gating. Language and Cognitive Processes, 11, 597-604. Grosjean, F. (1997). Processing mixed language: Issues, findings and models. In A. M. B. de Groot & J. F. Kroll (Eds.), Tutorials in bilingualism: Psycholinguistic
On-line Methods in Bilingual Spoken Language
25
perspectives (pp. 225-254). Mahwah, NJ: Erlbaum. Grosjean, F., & Frauenfelder, U. H. (Eds.). (1996). Spoken word recognition [Special issue]. Language and Cognitive Processes, 11(3). Grosjean, F., & Soares, C. (1986). Processing mixed language: Some preliminary findings. In J. Vaid (Ed.), Language processing in bilinguals: Psycholinguistic and neuropsychological perspectives (pp. 145-179). Hillsdale, NJ: Erlbaum. Glucksberg, S., Kreuz, R. J., & Rho, S. H. (1986). Context can constrain lexical access: Implications for models of language comprehension. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 323-335. Guillelmon, D., & Grosjean, F. (2001). The gender marking effect in spoken word recognition: The case of bilinguals. Memory & Cognition, 29, 503-511. Heredia, R. R. (1997). Bilingual memory and hierarchical models: A case for language dominance. Current Directions in Psychological Science, 6, 34-39. Heredia, R. R. (1998, August). Cross-modal approaches to the investigation of bilingual spoken language comprehension. Paper presented at the meeting of the International Association for Cross-Cultural Psychology. Western Washington University, Bellingham, WA. Heredia, R. R. (2000, May). Bilingual lexical access and code-switching. Paper presented at the Fifth Conference on Applied Linguistics, Universidad de Las Am6ricas-Puebla, Mexico. Heredia, R. R., & Altarriba, J. (2001). Bilingual language mixing: Why do bilinguals code-switch? Current Directions in Psychological Science, 10, 164-168. Heredia, R. R., & Stewart, M .T. (1998, November). Bilingual on-line sentence processing.'A moment-to-moment processing approach. Poster presented at the 39 th annual meeting of the Psychonomic Society, Dallas, TX. Heredia, R. R., & Blumentritt, T. L. (1999, November). On-line processing of stereotypes during spoken language: A moment-to-moment processing approach. Poster presented at the 40 th annual meeting of the Psychonomic Society, Los Angeles, CA. Heredia, R. R., & Blumentritt, T. L. (2001). On-line processing of stereotypes during spoken language. Manuscript submitted for publication. Heredia, R. R., Stewart, M. T., & Cregut, I. (1997, November). Bilingual on-line sentence processing." Frequency and context effects in code switching. Poster presented at the 38 t" annual meeting of the Psychonomi'c Society, Philadelphia, PA. Hern~ndez, A. E., Bates, E., & ,/Wila, L. X. (1994). On-line sentence interpretation in Spanish-English bilinguals: What does it mean to be "in between"? Applied Psycholinguistics, 15, 417-446. Hern~ndez, A. E., Bates, E., &/~vila, L. X. (1996). Processing across the language boundary: A cross-modal priming study of Spanish-English bilinguals. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 846-864. Hern~ndez, A. E., Fennema-Notestine, C., Udell, C., & Bates, E. (in press). Lexical and sentential priming in competition: Implications for two-stage theories of lexical access. Applied Psycholinguistics. Herron, D., & Bates, E. (1997). Sentential and acoustic factors in the recognition of open- and closed-class words. Journal of Memory and Language, 37, 217-239.
26
Roberto R. Heredia and Mark T. Stewart
Keatley, C., & Gelder, B. (1992). The bilingual primed lexical decision task: Crosslanguage priming disappears with speeded responses. European Journal of Cognitive Psychology, 4, 273-292. Kilborn, K. (1987). Sentence processing in a second language: Seeking a performance definition offluency. Unpublished doctoral dissertation, University of California, San Diego. Kilborn, K. (1989). Sentence processing in a second language: The timing of transfer. Language and Speech, 32, 1-23. Kolers, P. (1966). Reading and talking bilingually. American Journal of Psychology, 3, 357-376. Kroll, J. F., & Sholl, A. (1992). Lexical and conceptual memory in fluent and nonfluent bilinguals. In R. Harris (Ed.), Cognitive processing in bilinguals (pp. 157174). Amsterdam, Netherlands: Elsevier. Kroll, J. F., & Stewart, E. (1994). Category interference in translation and picture naming: Evidence for asymmetric connections between bilingual memory representations. Journal of Memory and Language, 33, 149-174. Lederberg, A. R., & Morales, C. (1985). Code switching by bilinguals: Evidence against a third grammar. Journal of Psycholinguistic Research, 14, 113-136. L6wy, N., & Grosjean, F. (2001). The computerized version of BIMOLA'A bilingual model oflexical access. Manuscript in preparation, University of Neuch~tel, Switzerland. Li, P. (1996a). Spoken word recognition of code-switched words by ChineseEnglish bilinguals. Journal of Memory and Language, 35, 757-774. Li, P. (1996b). The temporal structure of spoken sentence comprehension in Chinese. Perception and Psychophysics, 58, 571-586. Li, P., Bates, E., Liu, H., & MacWhinney, B. (1992). Cues as functional constraints on sentence processing in Chinese. In H-C. Chen & O. J. L. Tzeng. (Eds.), Language processing in Chinese (pp. 207-234). Amsterdam, Netherlands: Elsevier. Li, P., & Yip, M. C. (1998). Context effects and the processing of spoken homophones. Reading and Writing: An Interdisciplinary Journal, 10, 223-243. Liu, H., Bates, E., & Li, P. (1992). Sentence interpretation in bilingual speakers of English and Chinese. Applied Psycholinguistics, 13, 451-484. Liu, H., Bates, E., Powell, T., & Wulfeck, B. (1997). Single-word shadowing and the study of lexical access. Applied Psycholinguistics, 18, 157-180. Liu, H., Bates, E., & Li, P. (1992). Sentence interpretation in bilingual speakers of English and Chinese. Applied Psycholinguistics, 13, 451-484. Love, T., & Swinney, D. (1996). Coreference processing and levels of analysis in object-relative constructions: Demonstration of antecedent reactivation with the crossmodal priming paradigm. Journal of Psycholinguistic Research, 25, 5-24. Macnamara, J., & Kushnir, S. L. (1971 ). Linguistic independence ofbilinguals: The input switch. Journal of Verbal Learning and Verbal Behavior, 1O, 480-487. MacWhinney, B. (Ed.). (1987). Mechanisms of language acquisition. Hillsdale, NJ: Erlbaum. MacWhinney, B., & Bates, E. (Eds.). (1989). The cross-linguistic study ofsentence processing. New York, NY: Cambridge University Press.
On-line Methods in Bilingual Spoken Language
27
Marslen-Wilson, W. D. (1987). Functional parallelism in spoken word-recognition.
Cognition, 25, 71-102 McDonald, J. L. (1986). The development of sentence comprehension strategies in English and Dutch. Journal of Experimental Child Psychology, 41, 317-335. McDonald, J. L. (1987a). Assigning linguistic roles: The influence of conflicting cues. Journal of Memory and Language, 26, 100-117. McDonald, J. L. (1987b). Sentence interpretation in bilingual speakers of English and Dutch. Applied Psycholinguistics, 8, 379-413. McDonald, J. L., & Heilenman, L. K. (1991 ). Determinants of cue strength in adult first and second language speakers of French. Applied Psycholinguistics, 12, 313-348. McDonald, J. L., & Heilenman, L. K. (1992). Changes in sentence processing as second language proficiency increases. In R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 325-336). Amsterdam, Netherlands: Elsevier. McKoon, G., & Ratcliff, R. (1994). Sentential context and on-line lexical decision tasks. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 1239-1243. Myers-Scotton, C. (1997). Duelling languages: Grammatical structure in codeswitching. Oxford, England: Clarendon. Nicol, J., & Swinney, D. (1989). The role of structure in coreference assignment during sentence comprehension. Journal of Psycholinguistic Research, 18, 5-20. Prather, P. A., & Swinney, D. A. (1988). Lexical processing and ambiguity resolution: An autonomous process in an interactive box. In S. I. Small., G. W. Cottrell, & M. K. Tanenhaus (Eds.), Lexical ambiguity resolution: Perspectives from psycholinguistics, neuropsychology & artificial intelligence (pp. 289-310). San Mateo, CA: Morgan Kaufmann Publishers. Scarborough, D. L., Gerard, L., & Cortese, C. (1984). Independence of lexical access in bilingual word recognition. Journal of Verbal Learning and Verbal Behavior, 23, 84-99. Segalowitz, N. (1986). Skilled reading in the second language. In J. Vaid (Ed.),
Language processing in bilinguals: Psycholinguistic and neuropsychological perspectives (pp. 3-19). Hillsdale, NJ: Erlbaum. Sera, M. D., Berge, C. A. H., & del Castillo Pintado, J. (1994).Grammatical and conceptual forces in the attribution of gender by English and Spanish speakers. Cognitive Development, 9, 261-292. Soares, C., & Grosjean, F. (1984). Bilinguals in a monolingual and a speech model: The effect on lexical access. Memory & Cognition, 12, 380-386. Stewart, M.T., & Heredia, R. R. (2002). Moment-to-moment processing of reference metaphor during comprehension of spoken sentences. Journal of Experimental Psychology, 49, 1-11. Swinney, D. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18, 645-660. Swinney, D., Onifer, W., Prather, P., & Hirshkowitz, M. (1979). Semantic facilitation across modalities in the processing of individual words and sentences. Memory & Cognition, 7, 159-165.
28
Roberto R. Heredia and Mark T. Stewart
Tabossi, P. (1996). Cross-modal semantic priming. Language & Cognitive Processes, 11, 569-576. Titone, D. A., & Connine, C. M. (1994). Comprehension of idiomatic expressions: Effects of predictability and literality. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 1126-1138. Titone, D., Prentice, K. J., & Wingfield, A. (2000). Resource allocation during spoken discourse processing: Effects of age and passage difficulty as revealed by selfpaced listening. Memory & Cognition, 28, 1029-1040. Titone, D., Wingfield, A., Caplan, D., Waters, G., & Prentice, K. (2001). Memory and encoding of spoken discourse following right hemisphere damage: Evidence from the auditory moving window (AMW) technique. Brain and Language, 77, 10-24. Tyler, L. K. (1992). Spoken language comprehension: An experimental approach to disordered and normal processing. Cambridge, MA: MIT Press. Vaid, J., & Pandit, R. (1991). Sentence interpretation in normal and aphasic Hindi speakers. Brain and Language, 41, 250-274.
Part II: Connectionist Models of Second Language Processing and Bilingualism
This Page Intentionally Left Blank
Bilingual SentenceProcessing- R.R. Herediaand J. Altarriba(Editors) 9 2002 ElsevierScienceB.V. All rights reserved.
Extending the Competition Model Brian MacWhinney Carnegie Mellon University
Abstract Current models of language learning tend to focus on the outcome of the learning process. As a result, they fail to fully describe the detailed operations of learning mechanisms for particular target structures. Current models also underestimate the extent to which learners rely on a variety of complementary learning mechanisms. By extending the Competition Model to deal with the role of at least nine important learning mechanisms, we can derive a fuller picture of the actual process of language learning in the individual learner. Current models of second language learning divide into two basic types: nativist and emergentist. Nativist models hold that second language learning of grammar, even in adults, proceeds in accord with the principles and parameters of Universal Grammar (Chomsky, 1965). Emergentist models, such as the Competition Model (MacWhinney, 1987; MacWhinney & Bates, 1989), view grammatical learning as based on an interaction between the learner, the input, and the context (MacWhinney, 2001 c). Although the opposition between these two types of models has helped to clarify many issues, it has tended to obscure others. In part, this is because both types of models have not committed themselves to providing complete accounts of the learning process. Nativist models are responsible for only showing that principles and parameters have some influence on the learning process. Once linguists secure evidence of these effects (Epstein, Flynn, & Martohardjono, 1996), additional aspects of the language learning process are treated as outside of the core of syntax and become the responsibility of the psychologists and applied linguists. Until now, the Competition Model has assumed a similar limited explanatory responsibility. It measures the learner's overall command of grammatical markings at various points during second language learning. It then attempts to show how changes in overall level of attainment can emerge from the interaction between mental characteristics of the learner and properties of the input (MacWhinney, 2001 a). However, this level of description does not tell us how the learner moves through the process of language learning step by step. Thus, neither the Competition Model nor the nativist analyses have yet assumed responsibility for characterizing the details of the process of second language learning. This absence of attention to detail is particularly surprising when we consider how rich the second language learning literature is with studies of specific error types (Corder, 1983), learner overgeneralizations (Selinker, Swain, & Dumas, 1975), and learning strategies (O'Malley & Chamot, 1990). Why has the attention to
32
Brian Mac Whinney
detail that is so obviously an integral part of this literature failed to make its impact on models that attempt to account for the overall process? We can trace this failure to at least five sources: 1. Nativist models concern themselves with a restricted core of syntactic patterns. As a result, many important components of language learning have remained outside of the limelight generated by the conflict between nativism and emergentism. 2. The standards of experimental psychology require that we test hypotheses against group data. By studying groups, we can measure overall attainment at different stages in second language learning. Unfortunately, this requirement makes it difficult to track the details of individual patterns of language learning. 3. Because linguistic analysis is not subject to the same requirements, studies based on error analysis have produced some of the best material we have on the details of second language learning. However, these details have not been articulated yet in terms of processing models. 4. Much work on second language learning has focused on a particular level of linguistic structure, such as lexicon, phonology, orthography, syntax, morphology, or pragmatics. Within each area, researchers have been able to articulate models for learning. However, natural language learning occurs simultaneously across all of these levels. This means that, if we want to understand individual steps in the process of language learning, we need to understand how learning integrates across levels. 5. Accounts of individual strategies in second language learning (O'Malley & Chamot, 1990) ground themselves on the theory of individual differences, rather than accounts of the application of actual processes. If we are going to make progress toward a unified understanding of the process of second language learning, we will need to devise ways of overcoming these methodological restrictions. Of the limitations noted above, the most constraining is the requirement that evidence on second language learning should come from group data. By looking at individual subjects, we can obtain a fuller understanding of the details of language learning (Ericsson & Simon, 1984; Newell & Simon, 1972) and the ways in which individuals apply different configurations of basic learning strategies. My basic thesis in this paper is that language learning depends on a fairly large set of learning mechanisms. The role that particular mechanisms play in acquiring different levels of linguistic structure is dependent on the shape of the target structure and the nature of the learner. Understanding the interaction between these mechanisms can allow us to formulate an extended and more accurate version of the Competition Model for second language learning.
Extending the Competition Model
33
What Learning Mechanisms Are Available for Language Learning?
Cognitive psychology is rich in proposals for learning mechanisms. Although it might seem at first that one could easily get lost in this enormous literature, the relevant cognitive mechanisms can be counted on the fingers of two hands. Specifically, we are most interested in this set: 1. 2. 3. 4. 5. 6. 7. 8. 9.
Pattern Detection Predictive Association Selective Attention Transfer Chunking Resonance Error Detection Mental Model Construction Input Strategies
We can describe each of these mechanisms using either neural networks or symbolic production systems. Of course, all cognitive processing derives from neural function. For low-level processes such as (1) to (5), psychologists tend to prefer neural network modeling. However, for higher-level processes such as (6) to (9), it is often practical to abstract away from the details of neuronal functioning and to think in terms of higher-level symbolic systems such as production systems (Anderson, 1993). For our current purposes, the contrast between neural networks and production systems is largely irrelevant. Our focus now is simply on the mechanisms themselves and how they can provide explanations regarding the details of second language learning. Pattern Detection
Pattern detection is the most basic of all learning mechanisms. In the area of language learning, it functions most obviously when we first begin to listen to words in a new language. At this point, our auditory cortex stores the overall prosody and shape of words and phrases in the new language. Recent work (Aslin, Saffran, & Newport, 1999) has shown that, even at the age of 6 months, infants have become sensitive to a wide range of prosodic regularities in their native language. Moreover, they are able to quickly record co-occurrence patterns of syllable strings in the input. In fact, their ability to store such patterns in working memory is nearly as good as that of adults. This mechanism plays a crucial role in the learning of sounds and words. Predictive Association
Once we have stored away a set of auditory patterns, we can use them to predict other patterns. In classical conditioning, we use a stimulus to predict a response. This prediction can be unconscious and implicit. In response learning, we track the ways in which our actions bring about changes in the world. Listening to the results
34
Brian Mac Whinney
of our sound productions involves this type of prediction. In other cases, prediction operates between some external form, such as a word, and an internal interpretation. For language learning, this latter type of predictive association is crucial. Words represent predictive relations between sounds and their meanings. The Competition Model places a great emphasis on this basic association between a form and its function. In comprehension, the learner maps forms onto functions. In production, learning involves mapping functions onto forms. In accord with current theorizing in Construction Grammar (Kay & Fillmore, 1999), we view language as a collection of form-function mappings, some of which are lexical and some grammatical. Theories of predictive association all provide a central role for reinforcement learning. Work on the power law of practice (Newell & Rosenbloom, 1981) emphasizes the role of reinforcement and the fact that the best predictor of level of learning is time on task. Lexical learning illustrates the importance of this effect. For example, we are likely to learn high frequency words such as want or go in a new language, before we learn words like supervise or eradicate. In fact, frequency effects appear everywhere in language (Bybee & Hopper, 2001), affecting all aspects of phonology, syntax, and pragmatics and determining the course of grammaticalization. The Competition Model deals with frequency in terms of the construct of cue reliability. For example, in English, we would say that the cue of preverbal positioning is an extremely reliable cue to the choice of a noun as the subject of a transitive verb. On the other hand, the inherent animacy of the noun is less reliable as a cue, since we can have sentences such as "The ball hit the boy on the shoulder." Reliability is the conditional probability that an interpretation X should be selected given the presence of a cue Y (i.e., p(XIY)). If this probability is high, then Y is a reliable cue to X. The prediction from this initial analysis is that forms with a high conditional probability should appear early in development, should be resistant to loss in aphasia, should transfer strongly in second language (L2) learning, and should be the strongest determinants of processing in adults. Most Competition Model experiments use a simple sentence interpretation procedure. Subjects listen to a sentence with two nouns and a verb and are asked to say who was the actor. In a few studies, the task involves direct-object identification (Sokolov, 1988, 1989), relative clause processing (MacWhinney & Pl6h, 1988), or pronominal assignment (MacDonald & MacWhinney, 1990; McDonald & MacWhinney, 1995), but usually the task is agent identification. Sometimes the sentences are well-formed grammatical sentences, such as the cat is chasing the duck. Sometimes they involve competitions between cues, as in the ungrammatical sentence *the duck the cat is chasing. Depending on the language involved, the cues varied in these studies include word order, subject-verb agreement, object-verb agreement, case-marking, prepositional case marking, stress, topicalization, animacy, omission, and pronominalization. These cues are varied in a standard orthogonalized ANOVA design with three or four sentences per cell to increase statistical reliability. The basic question is always the same: what is the relative order of cue strength in the given language and how do these cue strengths interact? These studies have yielded a remarkably consistent body of results. The most important finding is that the order of cue strength in adults always corresponds with
Extending the Competition Model
35
the order of cue reliability yielded by text counts in the language. In different languages, we find different cue dominance patterns. In English, the dominant cue for subject identification is preverbal positioning. For example, in the English sentence the eraser hits the cat, we assume that the eraser is the agent. However, a parallel sentence in Italian or Spanish would have the cat as the agent. In Spanish, the prepositional object marker "a" is a clear cue to the object and the subject is the noun that is not the object. An example of this is the sentence el toro mat6 al torero (The bull killed to-the bullfighter). No such prepositional cue exists in English. In German, case marking on the definite article is a powerful cue to the subject. In a sentence such as der Lehrer liebt die Witwe (The teacher loves the widow), the presence of the nominative masculine article der is a sure cue to identification of the subject. In Hungarian, the subject is the noun not marked by any suffix or postposition. In Russian, the subject often has a case suffix. In Arabic, the subject is the noun that agrees with the verb in number and gender and this cue is stronger than the case-marking cue. In French, Spanish, and Italian, when an object pronoun is present, it can help identify the noun that is not the subject. There are many alternative formulations of the concept of predictive association. Neural network theory treats this in terms of the theory of classifier learning. A classifier is a mental structure (expressed in vector weights) that has learned to properly predict the assigriment of a given form to a category. The predictive associations between forms and functions in the Competition Model can be modeled as classifier learning using neural networks (Elman, 1991). Perhaps the clearest demonstrations of the applicability of neural network models are in the area of morphological learning (MacWhinney, Leinbach, Taraban, & McDonald, 1989). For example, when learning the English past tense, neural networks extract different phonological regularities to produce successful generalizations such as bent and fell as the past tenses of bend and fall. This type of learning clearly plays a major role in both first and second language learning. Like the other mechanisms we will discuss, there must be a role for predictive association in the updated Competition Model, particularly in the domains of morphology and syntax. However, predictive association, by itself, is certainly not the whole of the story. Selective Attention
We can distinguish two different ways in which attention modulates learning. First, it can operate like the lens of a camera to suddenly expose us to a new stimulus. This type of overall increase in sensitivity to exposure has been called "flashbulb memory." It operates primarily under conditions of high emotional arousal. Flashbulb memory may play some role in language learning, but a second type of attentional selection is probably more central. This is selective attention to particular components of a larger stimulus. Selective attention has clear influences on learning new words or characters. For example, when I was leaming the Cantonese phrase for "pivoting your foot inward," I initially encoded it as kau geu, instead of the correct form kau geuk. This is because there is a tendency in Cantonese to reduce final/k/. However, the reduced final/k/is not totally absent and has an effect on the quality of the preceding vowel. At first, I did not attend to this additional component or cue. However, after my encoding for kau geu became
36
Brian Mac Whinney
automated, my attentional focusing was then freed up enough so that I could notice the presence of the final/k/. This expansion of selective attention during learning is a very general process. The classic version of the Competition Model assumed that the learner paid simultaneous attention to all available cues. However, McDonald and MacWhinney (1991) and Matessa and Anderson (2000) have shown that, during the initial phases of language learning in miniature linguistic system experiments, subjects tend to focus on only one cue at a time. For example, in an experiment with four cues based on word order, animacy, case marking, and tone, a learner would decide to focus attention on only one of these as the predictor of sentence interpretation. Later on, after having tracked the use of this first cue, the learner will add a second cue to the mix and begin to use the two in combination. As development proceeds, the learner pays attention to additional cues. Selective attention is also important in early word learning. Smith (1999) has shown how the modulation of attention can lead a child to group objects into cognitive categories. Merriman (1999) has shown that one can use the Competition Model to account for aspects of lexical learning such as mutual exclusivity. However, to do this, the model must incorporate stronger assumptions regarding attentional selection. For example, at first a child might not focus on the fact that a cup is likely to have a handle, whereas a glass does not. Over time, attentional focusing sharpens the competition of words for referents in this way. Transfer
Psychologists think of transfer as something that happens when an animal moves between separate, but highly related learning tasks. For example, a rat that has learned to turn right when seeing a yellow gate in a maze should be able to transfer this learning to a new maze, as long as the two mazes have a similar overall pattern. This means that transfer is based on both pattern detection and predictive association. Associationist models rely heavily on similarity detection to learn the classifiers that produce predictive association. Because human languages are similar in so many ways, models that rely on predictive association predict massive amounts of transfer. The emphasis on the pervasiveness of transfer is a fundamental component of the Competition Model for second language learning. One of our first investigations of second language processing effects (Bates & MacWhinney, 1981) examined the comprehension of English sentences by some of our academic colleagues. One subject--a native speaker of German--had lived in the United States for 30 years. Married to an American, he had published several important textbooks in experimental psychology, all written in English. Remarkably, we found that this subject processed simple English sentences using the cue strength hierarchy of German. This is to say that he used agreement and animacy cues whenever possible, largely ignoring word order when it competed with agreement and animacy. We now have evidence for this preservation of a syntactic "accent" in comprehension from over a dozen studies (Bates & MacWhinney, 1981; de Bot & van Montfort, 1988; Gass, 1987; Harrington, 1987; Kilborn, 1989; Kilborn & Cooreman, 1987;
Extending the Competition Model
37
Kilborn & Ito, 1989; Liu, Bates, & Li, 1992; MacWhinney et al., 1989; McDonald, 1987a, 1987b; McDonald & Heilenman, 1991; McDonald & MacWhinney, 1989).
Chunking Learning a second language involves developing an automatic level of access to words and syntactic patterns that can guarantee fluency and accuracy. It is crucial to realize that fluency requires more than just reinforcement learning. Reinforcement may work to solidify small, highly determinant sequences of actions, such as those found in short words. However, when the learner tries to combine words into larger units, the strength of the individual components alone cannot guarantee fluency. Instead, the learner must work to build up larger components by chunking (Anderson & Lebiere, 1998). For example, in Spanish, learners can chunk together the plan for buenos with the plan for dias to produce buenos dias. They can then combine this chunk with muy to produce muy buenos dias "very good morning." Chunking allows the learner to get around problems with Spanish noun pluralization, gender marking, and agreement that would otherwise have to be reasoned out in detail for each combination. Although the learner understands the meanings of the three words in this phrase, the unit can function as a chunk, thereby speeding production. Chunking is the primary mechanism supporting the growth in fluency that is so important in learning a second language. Chunking occurs when we store a series of actions or percepts as a single unit. During production, this occurs as we say similar phrases repeatedly. During comprehension, it occurs as we hear separate items together. In both cases, chunking is dependent on reinforcement, since each repeated exposure to a chunk strengthens its unity. At first, we access the individual components in a slow, halting fashion. However, after we chunk them together, access is smoother and quicker. Once this occurs, resources are freed up that can allow us to expand our selective attention, thereby picking up additional syntactic cues, lexical meanings, phonological distinctions, and orthographic details. Within the word, chunking involves the increased automatization of the transitions between syllables and segments. Expanding on an account provided by Grossberg (1978), Gupta and MacWhinney (1997) present a neurologically based, computational model of the process of automatization of word production. The basic idea underlying this model is that words are stored as a chain in which the production of beginning segments triggers production of later segments. Learning involves the strengthening of the links in this chain until the learner can articulate the whole word smoothly as a single chunk. We often use studies of resource competition to measure the effects of chunking. There have been several Competition Model studies of this type. Kilborn (1989) has shown that even fully competent bilinguals tend to process sentences in their L2 more slowly than monolinguals. Moreover, when monolinguals listen to sentences under conditions of white noise, their reaction times are identical to those of bilinguals not subjected to noise. Similarly Blackwell and Bates (1995) and Miyake, Carpenter, and Just (1994) have shown that, when subjected to conditions of noise, normals process sentences much like aphasics not subjected to noise. In general, we can assume that the task of having to process two or more languages concurrently or
38
Brian MacWhinney
sequentially puts some load on the system. However, we also know that these effects decrease as one achieves improved fluency in the second language. At first, processing in L2 is so resource intensive, that it is impossible to carry out secondary tasks. However, as L2 chunking advances, resources are freed up for smoother processing. The ability of fluent simultaneous translators to cope with two different languages in parallel illustrates the extent to which growths in fluency can overcome resource limitations. However, even for this nearly miraculous example of bilingual fluency, there are limits. Gerver (1974) and Seleskovitch (1976) have shown that, during simultaneous translation, further imposition of a load through noise can result in marked deterioration in translation quality. Resonance
The classic version of the Competition Model relied entirely on the mechanisms of predictive association and transfer discussed above. However, if we wish to extend the model to cover the finer details of language learning, we need to consider the role of additional processes. Of these, the most important is resonance. Resonance is a modem formulation of a process that psychologists earlier called "spreading activation" and "interactive activation." The psychology of the 1970s (Lindsay & Norman, 1977) placed a great emphasis on spreading activation as the primary method for activating cognitive structures. McClelland and Rumelhart (1981) formalized the notion of spreading activation in their interactive activation model of word naming. In that model, words were composed of letters and letters were composed of graphemic segments. Recognition of a word occurred when the bottom-up activation from the segments and the letters spread more activation to the correct lexical item than to its nearest competitors. Systems of this type are still extremely useful in modeling reading processes in bilinguals (van Heuven, Dijkstra, & Grainger, 1998). Interactive activation provides a powerful, simple way of understanding the details of cognitive processing. However, it has one crucial feature that made it less useful as a general account of cognition. This is the fact that there is no learning rule for an interactive activation network. It is necessary to adjust all of the weights and connections inside the model by hand in order to achieve correct empirical predictions. To address these limitations, researchers have turned to more adaptive systems that use the concept of resonance. Primary among these systems is the adaptive resonance theory (ART) model developed by Grossberg (1987) and colleagues. Resonance occurs when several concepts are co-activated and their mutual excitation leads to the formation of a resonant pattern. A network can often store a few such patterns without additional attentional intervention. However eventually the network will need to rely on additional resources in terms of both computational units and attentional input. In ART, attentional activation occurs primarily as a result of an error signal. The error signal arises when the network cannot properly recharacterize the input pattem. This notion of recharacterization of a pattern is possible in an ART network because the system uses its resonant pattern to regenerate the shape of the input. In this sense, the model expresses concepts such as the neuronal group of Edelman (1987) or the convergence zones of Damasio and Damasio (1994). In ART attentional units are triggered by the operation of
Extending the Competition Model
39
motivational and affective processes. For second language learning, this could express the impact of limbic systems on learning (Pulvermtiller & Schumann, 1994). Resonance can be triggered passively through attention to error. However, learners can also control attention strategically to form resonant connections between already existing items. Gentner and Markman (1997), Hofstadter (1997) and others have formulated models of analogical reasoning that have interesting implications for language acquisition models. Analogies can be helpful in working out the first examples of a pattern. For example, a child learning German may compare steh aufi "stand up!" with er muff aufstehen "He must get up." The child can see that the two sentences express the same activity, but that the verbal prefix is moved in one. Using this pattern as the basis for further resonant connections, the child can then begin to acquire a general understanding of verbal prefix placement in German. Error Detection
The contrast between implicit and explicit learning has received a great deal of attention in recent work on second language learning. This interest stems originally from the contrast between learning and acquisition that was at the heart of much theorizing in the 1970s (Krashen, 1994). More recently, neural network models have placed a new focus on the role of implicit learning (Reder, 1996). It is clear that much language learning occurs through low-level pattern detection, perceptual chunking, and reinforcement. However, it is also true that some learners benefit more than others from explicit correction and instruction (Carroll & Swain, 1993; Ellis, 1994; Schmidt, 1994; Tomasello & Herron, 1988). The bulk of this discussion has examined learning from the viewpoint of the instructor, since it considers whether the provision of explicit corrective feedback has a detectable impact on language learning (MacWhinney, 1993). We can also examine the issue of error detection from the viewpoint of the learner. Learners may detect errors during either comprehension or production. During comprehension, they may notice that they cannot understand some important word, or they may notice that some default understanding they have does not actually make sense. When this happens, they can then store away this word as one they need to learn (MacWhinney, 1978). During production, learners can detect errors when they compare their own productions against stored images of correct productions from native speakers. When this happens, they can tag the problematic items as ones that they need to refine through selective attention and rechunk. Resonance can play an important role in the resolution of errors. For example, I recently noted that I had wrongly coded the stress on the Spanish word abanico "fan" as on the second syllable, as in ab6nico, perhaps under the influence of words like pl6stico. To correct this error, I spent time both rehearsing the correct stress pattern a few times and then visualizing the word as spelled without the stress mark or with the stress on the second syllable, which is normally not written in Spanish orthography. I also tried to associate this pattern in my mind with the verb abanicar "fan" and even the first person singular of this verb that has the form abanico. Having rehearsed this form in these various ways and having established these
40
Brian Mac Whinney
resonant connections, the tendency to produce the only incorrect form was somewhat reduced, although it will take time to fully banish the traces of the incorrect pattern.
Mental Model Construction Mental model construction probably plays a minimal role in child language acquisition. However, it can be of some marginal importance in adult learning. Although we know that learners derive little benefit from long hours of formal grammatical instruction in a foreign language, we know that they can benefit from some attention to grammatical pattems. Typically, adult learners store away these patterns in terms of formal structures such as charts of conjugations or grids of pronominal declension. For example, when learning Cantonese verbs, I have found it useful to think in terms of the verb as a pair of syllables separated by an aspect marker. Visually, this looks something like A - B - A', where A and A' are the two parts of the Cantonese verb and B is the aspect marker. Thus, a verb could be something like yau sdui "swim water" and the past tense is yau j6 sdui with the aspect marker j6 inserted between yau and s~ui. I have found that mental models of this type also work well for discontinuous verbal prefixes in Hungarian and German. To take yet another example, I find that, when in doubt about affixation, it is often useful to quickly color a Hungarian word in my mind as being either highvowel or low-vowel in terms of the vowel harmony pattems of Hungarian. I then use this internalized color to activate one of the correct allomorphs of the Hungarian suffixes that attaches to this stem. Finally, when trying to remember which Hungarian locative case suffix to select from the basic set of nine, I still tend to activate the grid of three positions by three movements that I leamed in my introductory Hungarian textbook. Of course, we only fall back on these mental models as crutches when we forget a form or have some gap in our vocabulary. However, mental models can serve an important role in the first phases of adult language learning, and they remain available later if needed.
Input Strategies Just as children seldom worry about constructing mental models, they seldom have to worry about acquiring methods for maximizing linguistic input. Typically, they derive excellent input from their family and friends (Snow, 1995). For the second language learner, however, the situation is very different. Often second language learners spend only a few hours or less each day using the second language. Worse still, these hours are often interspersed with ongoing first language (L1) input, L1 usage, and frequent code switching. In the L2 classroom, both child and adult learners have little control over the interactional context. In this context, successful learning strategies amount to little more than paying attention to the teacher, taking good notes, and keeping up with assignments. Given the difficulty involved in securing good L2 input, the learner's goal has to be to maximize input, while occasionally getting a chance to produce utterances. If there is a language-learning laboratory, the student can use videotapes, audiotapes, and computerized lessons as further sources of input. Even in the most favorable
Extending the Competition Model
41
naturalistic contexts, the L2 learner often does not have access to the rich system of social support that provides high quality language input to the child (Locke, 1995; Snow, 1995). In extreme situations, the L2 learner is overtly excluded from close personal interactions. In such cases, it becomes difficult to engage in the type of rich ongoing use of language that can maximally support learning. To compensate for this, the learner can develop a system of autosupport to improve the outcome of language learning. Concrete autosupport strategies include listening to television, radio, and movies, rehearsing taped dialogs, practicing new lexical items, and direct study of grammatical theory. These activities allow the adult learner to remain in contact with the input in ways that promote the functioning of neuronal loops for rehearsal, memory, and learning. These input strategies can be profitably linked to targeted use of resonance. Good learners realize that words and patterns are best learned when they are meaningfully related to other words, phrases, and orthographic patterns. This means that the process of maximizing input must be linked to an active process of resonant construction that provides the learner with time to construct and elaborate these resonant connections. Both L1 and L2 learners rely on resonance as a core learning mechanism. The major difference in the two cases is that the L2 learner must link the processing of resonant connections to the maximization of good input. For the child, there is no need to maximize input, since adequate input is always available.
How Do These Mechanisms Apply Across Linguistic Levels? Phonology Phonological learning involves two very different processes. Auditory acquisition is primary and begins even before birth (Moon, Cooper, & Fifer, 1993). It relies on inherent properties of the mammalian ear (Moon et al., 1993) and early pattern detection through statistical learning. This same statistical learning mechanism is operative in children, adults, and cotton-top tamarins (Hauser, Newport, & Aslin, 2001). The major challenge facing the learner is not the acquisition of perceptual patterns, but the development of articulatory methods for reproducing these patterns (Menn & Stoel-Gammon, 1995). The coordination of motor mechanisms for speech output is a relatively late evolutionary emergence (MacWhinney, 2001b) and it is not surprising that it is relatively difficult skill for the child to control. However, by age 5, most children have achieved control over articulatory processes. For the adult L2 learner and the older child, the situation is much different. Their learning begins with massive transfer of L1 articulatory patterns to L2 (Flege & Davidian, 1984; Hancin-Bhatt, 1994). This transfer is at first successful in the sense that it allows for a reasonable level of communication. However, it is eventually counter-productive, since it embeds L1 phonology into the emergent L2 lexicon. In effect, the learner treats new words in L2 as if they were composed of strings of L 1 articulatory units. This method of learning leads to short term gains at the expense of long-term difficulties in correcting erroneous phonological transfer. Older children acquiring a second language can rely on their greater neuronal flexibility to
42
Brian Mac Whinney
quickly escape these negative transfer effects. In doing so, they are relying on the same types of adolescent motor abilities that allow adolescents to become proficient acrobats, gymnasts, dancers, and golfers. Adults have a reduced ability to rewire motor productions on this basic level. However, even the most difficult cases of negative transfer in adulthood can be corrected through careful training and rehearsal (Flege, Takagi, & Mann, 1995). To do this, adults must rely on resonance, selective attention, and learning strategies to reinvigorate a motor learning process that runs much more naturally in children and adolescents. Lexicon
The second language learner comes to the task of learning a new language with a well-organized system of meanings and concepts. Unlike the child, who must develop a complete understanding of the components of his culture, the adult can readily transfer meaning, particularly if the cultures involved are similar. As in the area of phonology, the adult can achieve rapid initial progress in lexicon building by simply transferring this LI conceptual world en masse to L2. Young bilinguals can also benefit from this conceptual transfer. In the case of the simultaneous bilingual, the child only has to acquire one set of meanings. Although some words will be more closely linked to one language than another, the overall set of conceptual structures used to break up the world is essentially unified. For children who are picking up their second language somewhat later, lexical transfer is still highly positive. The case of Helen Keller is a dramatic illustration of this effect. Helen had begun to learn language as an infant, when she suddenly lost her hearing and sight. Later, when she came to learn language, she already had a well-established conceptual world that was ready to acquire new forms. As a result, after learning her first word water, she went on in the first day to learn 30 new words. When learners first acquire a new L2 form, such as silla in Spanish, they treat this form as simply another way of saying chair. This means that initially the L2 system has no separate conceptual structure and that its formal structure relies on the structure of L1. Kroll and Sholl (1992) emphasize the extent to which L2 relies on L1 forms to access meaning, rather than accessing meaning directly. In this sense, we can say that L2 is parasitic on L1, because of the extensive amount of transfer from L 1 to L2. The learner's goal is to reduce this parasitism by building up L2 representations as a separate system. They do this by strengthening the direct linkage between new L2 forms and conceptual representations. Given the fact that connectionism predicts such massive transfer for L1 knowledge to L2, we might ask why we do not see more transfer error in second language lexical forms. There are three reasons for this. First, a great deal of transfer occurs smoothly and directly without producing error. Consider a word like "chair" in English. When the native English speaker begins to learn Spanish, it is easy to use the concept underlying "chair" to serve as the meaning for the new word "silla" in Spanish. The closer the conceptual, material, and linguistic worlds of the two languages, the more successful this sort of positive transfer will be. Transfer only works smoothly when there is close conceptual match. For example, Ijaz (1986) has
Extending the Competition Model
43
shown how difficult transfer can be for Korean learners of English in semantic domains involving transfer verbs, such as "take" or "put." Similarly, if the source language has a two-color system (Berlin & Kay, 1969), as in Dani, acquisition of an eight-color system, as in Hungarian, will be difficult. These effects underscore the extent to which L2 lexical items are parasitic on L 1 forms. 2. Second, learners quickly correct some types of incorrect transfer. For example, when a learner tries to form a cognate in Spanish for the English noun "soap" the result is "sopa" which means "soup." Similarly, an attempt to translate the English form "competence" into Spanish as competencia will run into problems, since competencia means "competition." Once an error like this is made, it will be quickly detected. Moreover, the learner will become increasingly sensitive to possible errors in cognate transfer. .
Third, error is minimized when two words in L1 map onto a single word in L2. For example, it is easy for an L1 Spanish speaker to map the meanings underlying "saber" and "conocer" (Stockwell, Bowen, & Martin, 1965) onto the L2 English form "know." Dropping the distinction between these forms requires little in the way of cognitive reorganization. It is difficult for the L 1 English speaker to acquire this new distinction when learning Spanish. In order to control this distinction correctly, the learner must restructure the concept underlying "know" into two new related structures. In the area of lexical learning, these cases should cause the greatest transfer-produced errors.
Most current models of lexical development assume that the lexical link between a sound and a meaning is just a simple connection. This was a basic assumption of the classic Logogen Model (Morton, 1970). The problem with this model is that it treats words as if they were single neurons with easily located mental addresses. As MacWhinne~r (1994) has noted, neurons do not have addresses and there is no easy way to locate a word in neural space. Instead, we need to think of words as distributed patterns that have a maximum in a more mental processing space determined by their component sounds and meanings. This means that establishing a link between sounds and meanings is a highly constructive process. Although it is true that learners can develop a "fast mapping" (Carey, 1978) between a new word and a concept, they are unlikely to remember the word later on unless they have also established additional supportive relations between the sound and the meaning. Learners can use sound symbolism to establish a resonant connection between a sound and a meaning. It is not necessary that this symbolism be in accord with any established pattern. Because each learner-finds some idiosyncratic patterns of associations, it has been difficult to demonstrate the use of mnemonic connections in group studies of lexical learning. However, we do know that constructive mnemonics provided by the experimenter (Atkinson, 1975) greatly facilitate learning. For example, when learning the German word Wasser, we can imagine the sound of water running out of a faucet and associate this sound with the /s/ of
44
Brian Mac Whinney
Wasser. For this word, we can also associate the sound of the German word to the sound of the English word water. At the same time, we can associate Wasser with collocations such as Wasser trinken which themselves resonate with Bier trinken and others. Together, these resonant associations between collocations, sounds, and other words help to link the German word Wasser into the developing German lexicon. It is likely that children also use these mechanisms to encode the relations between sounds and meanings. Children are less inhibited than are adults in their ability to create ad hoc symbolic links between sounds and meanings. The child learning German as an L 1 might associate the shimmering qualities of Wasser with a shimmering aspect of the sibilant; or the child might imagine the sound as plunging downward in tone in the way that water comes over a waterfall. The child may link the concept of Wasser tightly to a scene in which someone pours ein Glas Waser and then the association between the sound of Wasser and the image of the glass and the pouring are primary. For the first language learner, these resonant links are woven together with the entire nature of experience and the growing concept of the world. The adult second language learner tends to rely on r~/ther less imaginative and more structured resonant linkages. One important set of links available to the adult is orthography. When an L2 learner of German learns the word Wasser, it is easy to map the sounds of the word directly to the image of the letters. Because German has highly regular mappings from orthography to pronunciation, calling up the image of the spelling of Wasser is an extremely good way of activating its sound. When the L2 learner is illiterate or when the L2 orthography is unlike the L 1 orthography, this backup system for resonance will not be available. L2 leaming of Chinese by speakers of languages with Roman scripts illustrates this problem. In Mainland China, use of Chinese characters is often accompanied by Romanized pinyin spellings. This allows the L2 leamer a method for establishing resonant connections between new words, their pronunciation, and their representations in Chinese orthography. However, in Taiwan and Hong Kong, characters are seldom written out in pinyin in either books or public notices. As a result, learners cannot learn from these materials. In order to make use of resonant connections from orthography, leamers must then focus on the learning of the complex Chinese script. This learning itself requires a large investment in resonant associations, since the Chinese writing system is based largely on radical elements that have multiple resonant associations with the sounds and meanings of words. The ongoing articulation and elaboration of these resonant connections between elements of L2 helps the learner establish a firewall against ongoing interference effects from L1. Consider the case of the English word "table" and the Spanish word "mesa." When learning this new word, the leamer links it to related Spanish words. During this process, the learner links "mesa" more closely to Spanish phrases and meanings than to English phrases and meanings. The more these two synonymous nouns link into separate worlds and to other words in the same language, the stronger will be the firewall that prevents interference. This type of separation must be achieved not only on the lexical level, but also on the phonological, syntactic, and semantic levels. In effect, this work undoes the early parasitic association of concepts that the beginning second language learner used to bootstrap the first phases of learning. The result of this process is the tightening of
Extending the Competition Model
45
within-language links in contrast to between-language links. In this way, linguistic modularity emerges from the formation of resonant connections. However, the establishment of a non-parasitic L2 lexicon does not produce full insulation between languages, even in the most advanced bilinguals (Grainger & Dijkstra, 1992). The interactive nature of language processing continues to promote transfer and interference, even when the lexicons and grammars of both languages are fully acquired. For those bilinguals and multilinguals who acquire their languages simultaneously during childhood, separate lexicons and grammars are constructed directly and there is no need to go through a process of undoing the initial connections formed through transfer (De Houwer, 1995; Grosjean, 1982). However, even in simultaneous bilinguals, some transfer and interference occurs, due to the interactive nature of cognitive processing. Orthography The United States Foreign Service has developed a system of classifying languages in terms of their difficulty for L1 English speakers. In this system, most of the Indo-European languages of Europe are on the easiest level of difficulty. The second level includes Indo-European languages with structured grammars such as Hindi, German, or Lithuanian. It also includes some non-Indo-European languages such as Bahasa Indonesia. Level 3 includes non-Indo-European languages like Finnish or Hungarian. The next level includes languages with relatively regular non-Roman orthographies such as Arabic. The most difficult are the Asian languages with truly complex non-Roman scripts. This order of difficulty reflects, in large part, the extent to which problems with orthography can determine difficult in learning a language. One might think that this can be overcome by simply teaching the language in Romanized form. This is certainly of some value. However, because the culture itself does not use Romanized script, this method has clear limitations. As in the case of the acquisition of the lexicon, the learning of a complex new orthography, such as the Chinese-Japanese-Korean system of characters, must rely on extensive construction of resonant connections. The learner must learn to write the characters in order. Each character must be analyzed into its component radicals and the meanings of the radicals must be learned. Once this is done, the characters must then be associated back to newly learned words and further resonant connections must be established.
Syntax By learning words in larger chunks, the child acquires data that facilitates the acquisition of syntactic categories. For example, the German child learns ein guter Mann and this provides good clues for the gender of the noun and the whole system of agreement. Same of the use of the definite article in phrases such as "I saw the Grand Canyon flying to New York." All of these rote phrases with their associated clues form a resonant ensemble.
46
Brian MacWhinney
Transfer also affects the acquisition of grammar. In the Competition Model, early syntactic learning involves the setting of weights between grammatical forms (word order, morphology, stress, lexical subclassification) and grammatical roles (agent, theme, recipient, coreferent). The L2 syntactic system is initially parasitic on L1 in the same way that L2 phonology and lexicon are parasitic on L 1. This means that the initial setting of weights in the L2 network is based on the weights in L1. However, learning of L2 involves more than simple retuning of these weights. In some cases, the second language requires the learner to seek out entirely new conceptual or discourse distinctions that were ignored in the first language, but which are now obligatory grammaticaJ contrasts in the new language. A prime example of this type of restructuring might be the foreigner's attempts to pick up the category structure underlying the two major verbal conjugations of Hungarian. Every time a speaker of Hungarian uses a verb, he or she must decide whether it should be conjugated as transitive or intransitive. Making this choice is not a simple matter. The intransitive conjugation is used not only when the verb is intransitive, but also when the direct object is modified by an indefinite article or by no article at all. It is also used when it is in the first or second person, when the head of the relative clause is the object within the relative clause, when the direct object is quantified by words like each, no, and so on. For example, the intransitive conjugation is used when a Hungarian says John runs, John eats an apple, John eats your apple, and John eats no apple. On the other hand, the transitive conjugation is used when the object is definite, when it is modified by a third person possessive suffix, when it is possessed by a third person nominal phrase, and so on. Thus, the transitive or definite conjugation is used when the Hungarian wants to say, John eats the apple or John eats Bill's apple, whereas the intransitive is used to say John eats an apple. There are some 13 conditions which, taken together, control the choice between the transitive and intransitive conjugations (MacWhinney & Pl6h, 1997). There is no single principle that can group together these 13 conditions. Instead, transitivity, definiteness, and referential disambiguation all figure in as factors in making this choice. Not surprisingly, L2 learners of Hungarian have a terrible time marking this distinction; errors in choice of the conjugation of the verb are the surest syntactic cue that the learner is not a native Hungarian. In order to acquire this new category, the L2 learner begins by attempting to transfer from L 1. To some degree this can work. The learner attempts to identify the intransitive with the English intransitive. However, the fact that transitive sentences also take the intransitive if the objects are indefinite tends to block the simple application of this conceptual structure. In the end, the leamer must be resigned to picking up the pieces of this new category one by one and restructuring them together into a working system. Here is an area where attempts at formal linguistic analysis on the learner's part only make matters worse. If the leamer had proceeded like a Hungarian child (MacWhinney, 1974), he or she would have learned the conjugations by generalizing from a rich database of collocations and phrases. The adult needs to amplify this case-based approach to learning with a way of focusing on contrastive structures in which cues are competing. For the adult, such focusing on particularly difficult parts of a grammatical system will increase the efficiency of acquisition.
Extending the Competition Model
47
In many cases, the transfer of syntactic patterns from L1 to L2 is structurally correct, but pragmatically inaccurate. For example, Tr6vise (1986) observes that French speakers make excessive use of topicalization structures in English in the form of structures corresponding to left-dislocations, right-dislocations, and c'est que in French. Although these structures are all permissible in English, the actual conditions on their usage are far more restrictive than in French. Similarly, Seliger (1989) notes that Hebrew learners of English tend to systematically underuse the passive. He attributes this underusage to the relatively tighter, genre-dependent conditions on the use of the passive in Hebrew. In general, it is clear that simple transfer of an L1 structure to L2 is not sufficient to guarantee correct usage, since both underutilization and overutilization can occur until the full conditions governing the use of a construction in L2 are learned. Competition Model studies use the interpretations of simple transitive sentences to estimate the strength of cues in L2 learners. Because the experimental design places cues into competition orthogonally, we can use mathematical techniques to estimate the strength of each cue in terms of its ability to determine the shape of our experimental data. We find uniformly that the learning of sentence processing cues in a second language is a gradual process. The process begins with L2 cue weight settings that are close to those for L I. Over time, these settings change in the direction of the native speakers' settings for L2. The pattern of results found in this research is clearest in the data from McDonald's studies of English-Dutch and Dutch-English second language learning (McDonald, 1987b). Figure 1 shows the decline in the strength of the use of word order by English learners of Dutch over increased levels of competence. In this graph the monolingual cue usage pattem for English is given in the first column and the monolingual Dutch pattern is given on the right. Between these two patterns, we see a declining use of word order and an increasing use of case inflection across three increasing levels of learning of Dutch. In Figure 2, we see exactly the opposite pattern for Dutch learners of English. These results and others like them constitute strong support for the application of the Competition Model to second language learning. Figure 2 depicts a set of connections between forms and functions for the processing of cues to sentence interpretation. It is possible to convert this intuitive analysis of the subject-marking system into an actual running network model. Janice Johnson and I have done this by giving the various surface cues specified in Figure 1 as input to a "recurrent" network of the type developed by Elman (1990). This network is able to simulate the learning of Dutch and English. The model takes as input a corpus of over 1000 sentences in each language. This corpus contains grammatical sentences from a wide variety of sentence types, including relative clauses, simple transitives, imperatives, subordinates, pronominalizations, and grammatical word order variations. After training in either language as L1, the model is given ungrammatical Competition Model sentences as test materials. At this point, we find cue interpretation patterns characteristic of L1 learners. In order to model second language learning, we then continue with a period of training with sentences from both languages. We then see exactly the pattern of cue development reported in Figures 1 and 2 from McDonald (McDonald, 1987b, 1989).
48
Brian MacWhinney 100
.-.,...
Noun animacy
80
Case inflection
T
....
Word order
!=t I=1
t-
w 6o 8
Y
C
~>
40
e-
89 2o
English
E/D 1
E/D 2
E/D 3
Dutch
Figure 1. Changes in cue strength as English speakers learn Dutch (McDonald, 1987b). 100
L
80
"ID tO
8
60
t~ e-t~ L_ t~ >
40
Noun Animacy
t~ r-
Case Inflection Word Order
20 A
v
I
Dutch
,
9
I
DIE 1
9
I
DIE 2
,
I
DIE 3
,
English
Group
Figure 2. Changes in cue strength as Dutch speakers learn English (McDonald, 1987b).
Extending the Competition Model
49
Pragmatics The acquisition of pragmatic patterns is also heavily influenced by L1 transfer. When we first begin to use a second language, we may extend our L1 ideas about the proper form of greetings, questions, offers, promises, expectations, turn-taking, topic expansion, face-saving, honorifics, presuppositions, and implications. If the two cultures are relatively similar, much of this transfer will be successful. However, there will inevitably be some gaps. In many cases, the L2 learner will need to eventually reconstruct the entire system of pragmatic patterns in the way they were learned by the child acquiring L1. Much of this learning is based on specific phrases and forms. For example, the L1 learners' understanding of greetings is tightly linked to use of specific phrases such as Guten Morgen or byebye. Learning about how and when to use specific speech acts is linked to learning about forms such as could you? listen, and why not? Learning these forms in a concrete context is important for both L1 and L2 learners. However, pragmatics involves much more than simple speech act units or pairs. We also need to learn larger flames for narratives, argumentation, and polite chatting. By following the flow of perspectives and topics in conversations (MacWhinney, 1999), we can eventually internalize models of how discourse represents reality in both L1 and L2.
Morphosyntax Learning of the morphosyntax of a second language is very different from learning of the other five areas we have discussed. This is because, in morphosyntax, it is typically impossible to transfer from L 1 to L2. For example, an English learner of German cannot use the English noun gender system as a basis for learning the German noun gender system. This is because English does not have a fully elaborated noun gender system. Similarly, a Spanish learner of Chinese, cannot use L1 knowledge to acquire the system of noun classifers, because Spanish has no noun classifiers. Chinese learners of English cannot use their L1 forms to learn the English contrast between definite, indefinite, and zero articles. This is because Chinese makes no overt distinctions in this area, leaving the issue of definiteness to be marked in other ways, if at all. The fact that morphosyntax is not subject to transfer is a reflection of the general Competition Model dictum that "everything that can transfer will." In the areas of phonology, lexicon, orthography, syntax, and pragmatics, we see attempts to transfer. However, in morphology we see no transfer, because there is no basis i~or transfer. The exception here is between closely related languages. For example, when learning French, a Spanish speaker can profitably transfer L1 noun gender assignments. Although this will lead to some mistakes, the overall effect of this transfer is positive. Learning of morphosyntax is heavily reliant on predictive association. Neural network models of the learning of the English past tense, German declension, Italian agreement, Arabic numerals, and other morphosyntactic systems have shown how closely this learning can be modeled using neural network simulations. However, it would be a mistake to think that L2 learning is identical to L1 learning for morphosyntax, as we will see in the next section.
50
Brian Mac Whinney Utterance-Based and Word-Based Learning
There are many sources of individual differences in language learning. The age of learning of L2 is certainly the largest single factor. After a certain age, it becomes difficult to block the transfer of an L I accent. Other central variables are motivation, opportunity for exposure, working memory span, motoric ability, and the presence of any organic neural limitations. Among these many factors, one that is particularly interesting is the contrast between utterance-based and word-based learning. In the child language literature, children who prefer word-based learning are said to be analytic and nominal. Children who prefer utterance-based learning are said to be gestaltist and phrasal. This sharp distinction between the two learning styles shows itself most clearly during the first stages of L1 learning and then vanishes. Presumably, the distinction evaporates because the nominal children find that they need to pay some attention to whole phrases and the phrasal children come to deal with individual words and morphemes. L2 learners are also influenced by this dichotomy. Some learning methods emphasize analysis of words and phrases into their components and others emphasize learning of large units, sometimes through pattern drill. Because L2 learning is driven so heavily by transfer, it is virtually impossible for the L2 learner to avoid analytic, word-based learning. Textbooks, grammars, and dictionaries are heavily biased toward word-based learning, and it is difficult not to rely on these resources. If the L2 learner is given the same type of input as the L1 child, utterance-based learning is possible. However, L2 learners almost never have access to this type of input. As a result, they fall back on word-based learning. This reliance on word-based learning has several consequences. The first is positive, since we know that, in terms of time spent in language learning, adult L2 learners are more effective than younger L2 learners in vocabulary learning (Snow & Hoefnagel-Hohle, 1978). If adult learners make good use of resonant learning and orthographic support, they can acquire good control over the L2 lexicon. With careful attention to phonological contrasts, adults can also acquire basic phonological competence, although they may never be able to fully block penetration of L 1 accent. However, L2 learners who rely on word-based learning will have problems at the phrasal and syntactic levels. One way in which this problem will reflect itself is in the acquisition of morphosyntactic patterns. Consider the example of learning of German gender. The analytic adult learner tends to acquire German nouns as single words. At best, the learner may try to associate Mann with der by learning the phrase der Mann. Without creating some particular resonant relation between Mann and der, even this type of learning is not very successful. A phrasal learner, on the other hand, will pick up units such as ein guter Mann, fiir diesen Manner, mein Mann, and so on. From these stored phrases, which are themselves linked to specific interactional scenes with rich expressive content, the learner can then extract the relational patterns of the German gender system. L1 learners rely on phrasal learning as input to a neural network method for acquiring the predictive associations in the pattern. L2 learners cannot do this effectively without focusing their attention on these phrasal units. Fortunately, good teaching and learning practices can remedy this problem and facilitate this aspect of L2 learning.
Extending the Competition Model
51
A further consequence of the heavy L2 reliance on word learning is a reliance on a particular translations strategy that maps words, rather than phrases. For example, a beginning L2 learner of-Spanish may want to say "I want to go to the store." Instead of saying "Quiero ir al mercado," the student says, "Yo querer a ir a el mercado." Here is a breakdown of the word-by-word translation strategy: I want to go Yo querer a ir
to the store a el mercado
Learners quickly repair the worst parts of this type of word-by-word translation. For example, the L2 Spanish learner will soon learn that the Spanish infinitive incorporates the article "a" and this can therefore be omitted. Then the learner realizes that the subject pronoun can be dropped, since person is marked on the verb. Finally, the learner would contract "a" and "el" into "al". For languages of markedly different structure, modifications must be even more radical. For example, when learning to form questions in Cantonese, the learner must acquire distinctions between a set of four sentence final particles, some of which differ only in tone. In addition, wh-movement has to be suppressed in favor of in situ wh. The actual system of wh-words also differs conceptually in several ways. Then, for yes-no questions, the learner must substitute aux-movement with a form that sounds something like know not know? Once all of these modifications of the basic translation strategy have been achieved, the learner starts to produce questions in a Cantonese format. In view of the overall word-based nature of L2 learning, it is probably fruitless for an instructor to fight against the translation strategy in sentence production. Instead, the instructor can provide the learner with simple and evocative whole phrases that illustrate the ways in which the translation strategy can be corrected.
A Unified Account
The search for a unified theory of cognition has been a major goal in models of psychological processes. Some accounts have sought to produce this unified theory by relying primarily on a single cognitive mechanism, such as chunking, predictive association, or analogy. Dual-process accounts try to limit the number of mechanisms to two. These attempts to reduce all learning to one or two mechanisms are ambitious and interesting. However, in the end, they will fail, since it is clear that learners rely on more than one or two learning mechanisms. This should not be too surprising, since the environment provides us with many different types of learning challenges. Evolution has had hundreds of millions of years to work out alternative mechanisms for different learning challenges. Many of these mechanisms share core components for processing and storage, but their use of these components may vary in accord with the learning task. Recent hominid evolution and the development of human language and representational systems has further extended the basic set of learning mechanisms. This means that a unified account of language learning will need to provide a role for more than just one or two mechanisms.
52
Brian Mac Whinney
The empirical exploration of this extended version of the Competition Model will place increased weight on the study of individual differences in language learning. Students of first language learning have often noted major individual differences. However, these differences tend to vanish by about age 5, since children of this age become uniformly competent in their L1, despite the initial differences in reliance on particular mechanisms. Because L2 learning is more protracted, we can see the effects of individual differences more clearly. For example, we can see how some learners rely heavily on reformulation and error correction, whereas others rely primarily on the less observable processes of resonant formation of associations. The extended version of the Competition Model emphasizes the links between particular learning mechanisms and particular linguistic tasks. It also emphasizes linkages between learning mechanisms. For example, as we have seen, the influence of transfer in L2 learning has major secondary impact on the use of all the other learning mechanisms. To deal with transfer, adult learners must then make particular use of resonance and error detection. This means that our models of second language learning must be quite specific about the details of the learning process, since the process shifts from one mechanism to another across the course of learning. If we want to study these changes using group study methodology, we will have to focus more on specific attempts to promote use of particular learning strategies for particular target linguistic structures at particular points in development. Using data of this type, we can begin to sketch out a more dynamic view of language learning.
References
Anderson, J. (1993). Rules of the mind. Hillsdale, N J: Lawrence Erlbaum Associates. Anderson, J., & Lebiere, C. (1998). The atomic components of thought. Mahwah, NJ: Erlbaum. Aslin, R. N., Saffran, J. RI,& Newport, E. L. (1999). Statistical learning in linguistic and nonlinguistic domains. In B. MacWhinney (Ed.), The emergence of language (pp. 359-380). Mahwah, NJ: Lawrence Erlbaum Associates. Atkinson, R. (1975). Mnernotechnics in second-language learning. American Psychologist, 30, 821-828. Bates, E., & MacWhinney, B. (1981). Second language acquisition from a functionalist perspective: Pragmatic, semantic and perceptual strategies. In H. Winitz (Ed.), Annals of the New York academy of sciences conference on native and foreign language acquisition (pp. 190-214). New York: New York Academy of Sciences. Berlin, B., & Kay, P. (1969). Basic color terms: Their universality and evolution. Berkeley: University of California Press. Blackwell, A., & Bates, E. (1995). Inducing agrammatic profiles in normals: Evidence for the selective vulnerability of morphology under cognitive resource limitation. Journal of Cognitive Neuroscience, 7, 228-257.
Extending the Competition Model
53
Bybee, J., & Hopper, P. (2001). Frequency and the emergence of linguistic structure. Amsterdam: John Benjamins. Carey; S. (1978). The child as word learner. In M. Halle, J. Bresnan, & G. Miller (Eds.), Linguistic theory and psychological reality (pp. 264-293). Cambridge, MA: MIT Press. Carroll, S., & Swain, M. (1993). Explicit and implicit feedback. Studies in Second Language Acquisition, 15, 357-386. Chomsky, N. (1965). Aspects of the theory of syntax. Cambridge, MA: MIT Press. Corder, S. P. (1983). A role for the mother tongue. In S. Gass & L. Selinker (Eds.), Language transfer in language learning. Rowley, MA: Newbury House. Damasio, A. R., & Damasio, H. (1994). Cortical systems for retrieval of concrete knowledge: The convergence zone framework. Cambridge, MA: MIT Press. de Bot, K., & van Montfort, R. (1988). 'Cue-validity' in het Nederlands als eerste en tweede taal. Interdisciplinair Tijdschrift voor Taal en Tekstwetenschap, 8, 111120. De Houwer, A. (1995). Bilingual language acquisition. In P. Fletcher & B. MacWhinney (Eds.), The handbook of child language (pp. 219-250). Oxford: Blackwell. Edelman, G. (1987). Neural Darwinism." The theory of neuronal group selection. New York: Basic Books. Ellis, R. (1994). A theory of instructed second language acquisition. In N. C. Ellis (Ed.), Implicit and explicit learning of language (pp. 79-114). San Diego: Academic. Elman, J. (1990). Finding structure in time. Cognitive Science, 14, 179-212. Elman, J. (1991). Distributed representations, simple recurrent networks, and grammatical structure. Machine Learning, 7, 195-225. Epstein, S., Flynn, S., & Martohardjono, G. (1996). Second language acquisition: Theoretical and experimental issues in contemporary research. Behavioral and Brain Sciences, 19, 677-758. Ericsson, K. A., & Simon, H. A. (1984). Protocol analysis: Verbal reports as data. Cambridge: MIT Press. Flege, J., & Davidian, R. (1984). Transfer and developmental processes in adult foreign language speech production. Applied Psycholinguistics, 5, 323-347. Flege, J., Yakagi, J., & Mann, V. (1995). Japanese adults can learn to produce English "r" and "1" accurately. Language Learning, 39, 23-32. Gass, S. (1987). The resolution of conflicts among competing systems: A bidirectional perspective. Applied Psycholinguistics, 8, 329-350. Gentner, D., & Markman, A. (1997). Structure mapping in analogy and similarity. American Psychologist, 52, 45-56. Gerver, D. (1974). The effects of noise on the performance of simultaneous interpreters: Accuracy of performance. Acta Psychologica, 38, 159-167. Grainger, J., & Dijkstra, T. (1992). On the representation and use of language information in bilinguals. In R. J. Harris (Ed.), Cognitive processing in bilinguals. (pp. 207-220). Amsterdam: Elsevier Science.
54
Brian MacWhinney
Grosjean, F. (1982). Life with two languages: An introduction to bilingualism. Cambridge, MA: Harvard University Press. Grossberg, S. (1978). A theory of human memory: Self-organization and performance of sensory-motor codes, maps, and plans. Progress in Theoretical Biology, 5, 233-374. Grossberg, S. (1987). Competitive learning: From interactive activation to adaptive resonance. Cognitive Science, 11, 23-63. Gupta, P., & MacWhinney, B. (1997). Vocabulary acquisition and verbal shortterm memory: Computational and neural bases. Brain and Language, 59, 267-333. Hancin-Bhatt, B. (1994). Segment transfer: A consequence of a dynamic system. Second Language Research, 10, 241-269. Harrington, M. (1987). Processing transfer: Language-specific strategies as a source of interlanguage variation. Applied Psycholinguistics, 8, 351-378. Hauser, M., Newport, E., & Aslin, R. (2001). Segmentation of the speech stream in a non-human primate: Statistical learning in cotton-top tamarins. Cognition, 78, B53-B64. Hofst~dter, D. (1997). Fluid concepts and creative analogies: Computer models of the fundamental mechanisms of thought. London: Allen Lane. Ijaz, H. (1986). Linguistic and cognitive determinants of lexical acquisition in a second language. Language Learning, 36, 401-451. Kay, P., & Fillmore, C. J. (1999). Grammatical constructions and linguistic generalization: The "what's X doing Y?" construction. Language, 75, 1-33. Kilborn, K. (1989). Sentence processing in a second language: The timing of transfer. Language and Speech, 32, 1-23. Kilborn, K., & Cooreman, A. (1987). Sentence interpretation strategies in adult Dutch-English bilinguals. Applied Psycholinguistics, 8, 415-431. Kilborn, K., & Ito, T. (1989). Sentence processing in Japanese-English and Dutch-English bilinguals. In B. MacWhinney & E. Bates (Eds.), The crosslinguistic study of sentence processing (pp. 257-291). New York: Cambridge University Press. Krashen, S. (1994). The Input Hypothesis and its rivals. In N. C. Ellis (Ed.), Implicit and explicit learning of languages (pp. 45-78). San Diego: Ac~idemic. Kroll, J., & Sholl, A. (1992). Lexical and conceptual memory in fluent and nonfluent bilinguals. In R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 191-206). Amsterdam: Elsevier Science. Lindsay, P., & Norman, D. (1977). Human information processing: An introduction to psychology. New York: Academic Press. Liu, H., Bates, E., & Li, P. (1992). Sentence interpretation in bilingual speakers of English and Chinese. Applied Psycholinguistics, 13, 451-484. Locke, J. L. (1995). Development of the capacity for spoken language. In P. Fletcher & B. MacWhinney (Eds.), The handbook of child language. Oxford: Basil Blackwell. MacDonald, M., & MacWhinney, B. (1990). Measuring inhibition and facilitation from pronouns. Journal of Memory and Language, 29, 469-492. MacWhinney, B. (1974). How Hungarian children learn to speak. Unpublished doctoral dissertation, University of California, Berkeley.
Extending the Competition Model
55
MacWhinney, B. (1978). The acquisition of morphophonology. Monographs of
the Society for Research in Child Development, 43, 1-123. MacWhinney, B. (1987). The Competition Model. In B. MacWhinney (Ed.),
Mechanisms of language acquisition (pp. 249-308). Hillsdale, NJ: Lawrence Erlbaum. MacWhinney, B. (1993). The (il)logical problem of language acquisition,
Proceedings of the Fifteenth Annual Conference of the Cognitive Science Society (pp. 61-70). Hillsdale, NJ: Lawrence Erlbaum Associates. MacWhinney, B. (1994). The dinosaurs and the ring. In R. Corrigan, S. Lima, & M. Noonan (Eds.), The reality of linguistic rules (pp. 283-320). Amsterdam: John Benjamins. MacWhinney, B. (1999). The emergence of language from embodiment. In B. MacWhinney (Ed.), The emergence of language (pp. 213-256). Mahwah, NJ: Lawrence Erlbaum. MacWhinney, B. (200 l a). The Competition Model: The input, the context, and the brain. In P. Robinson (Ed.), Cognition and second language instruction. New York: Cambridge University Press. MacWhinney, B. (200 l b). The gradual evolution of language. In B. Malle & T. Giv6n (Eds.), The evolution of language. Philadelphia: Benjamins. MacWhinney, B. (2001c). Models of second language acquisition. In P. Robinson (Ed.), Cognition and second language instruction. New York: Cambridge University Press. MacWhinney, B., & Bates, E. (Eds.). (1989). The crosslinguistic study of sentence processing. New York: Cambridge University Press. MacWhinney, B., & Pl6h, C. (1988). The processing of restrictive relative clauses in Hungarian. Cognition, 29, 95-141. MacWhinney, B., & P16h, C. (1997). Double agreement: Role identification in Hungarian. Language and Cognitive Processes, 12, 67-102. MacWhinney, B. J., Leinbach, J., Taraban, R., & McDonald, J. L. (1989). Language learning: Cues or rules? Journal of Memory and Language, 28, 255-277. Matessa, M., & Anderson, J. (2000). Modeling focused learning in role assignment. Language and Cognitive Processes, 15, 263-292. McClelland, J. L., & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part 1. An account of the basic findings. Psychological Review, 88, 375-402. McDonald, J. L. (1987a). Assigning linguistic roles: The influence of conflicting cues. Journal of Memory and Language, 26, 100-117. McDonald, J. L. (1987b). Sentence interpretation in bilingual speakers of English and Dutch. Applied Psycholinguistics, 8, 379-414. McDonald, J. L. (1989). The acquisition of cue-category mappings. In B. MacWhinney & E. Bates (Eds.), The crosslinguistic study of language processing (pp. 375-396). New York: Cambridge University Press. McDonald, J. L., & Heilenman, K. (1991). Determinants of cue strength in adult first and second language speakers of French. Applied Psycholinguistics, 12, 313348. McDonald, J. L., & MacWhinney, B. (1989). Maximum likelihood models for sentence processing research. In B. MacWhinney & E. Bates (Eds.), The
56
Brian MacWhinney
crosslinguistic study of sentence processing (pp. 397-421). New York: Cambridge University Press. McDonald, J. L., & MacWhinney, B. (1991). Levels of learning: A microdevelopmental study of concept formation. Journal of Memory and Language, 30, 407-430. McDonald, J. L., & MacWhinney, B. J. (1995). The time course of anaphor resolution: Effects of implicit verb causality and gender. Journal of Memory and Language, 34, 543-566. Menn, L., & Stoel-Gammon, C. (1995). Phonological development. In P. Fletcher & B. MacWhinney (Eds.), The handbook of child language (pp. 335-360). Oxford: Blackwell. Merriman, W. (1999). Competition, attention, and young children's lexical processing. In B. MacWhinney (Ed.), The emergence of language (pp. 331-358). Mahwah, NJ: Lawrence Erlbaum. Miyake, A., Carpenter, P., & Just, M. (1994). A capacity approach to syntactic comprehension disorders: Making normal adults perform like aphasic patients. Cognitive Neuropsychology, 11, 671-717. Moon, C., Cooper, R. P., & Fifer, W. P. (1993). Two-day infants prefer their native language. Infant Behavior and Development, 16, 495-500. Morton, J. (1970). A functional model for memory. In D. A. Norman (Ed.), Models of human memory (pp. 203-248). New York: Academic Press. Newell, A., & Rosenbloom, P. S. (1981). Mechanisms of skill acquisition and the law of practice. In J. R. Anderson (Ed.), Cognitive skills and their acquisition (pp. 1-55). Hillsdale, NJ: Lawrence Erlbaum Associates. Newell, A., & Simon, H. (1972). Human problem solving. Englewood Cliffs, N.J.: Prentice-Hall. O'Malley, M., & Chamot, A. (1990). Learning strategies in second language acquisition. Cambridge: Cambridge University Press. Pulvermialler, F., & Schumann, J. (1994). Affect in second language learning. Language Learning, 44, 681-734. Reder, L. (Ed.). (1996). Implicit memory and metacognition. Mahwah, NJ: Lawrence Erlbaum Associates. Schmidt, R. (1994). Implicit learning and the cognitive unconscious: Of artificial grammars and SLA. In N. C. Ellis (Ed.), Implicit and explicit learning of languages (pp. 165-210). San Diego: Academic. Seleskovitch, D. (1976). Interpretation: A psychological approach to translating. In R. W. Brislin (Ed.), Translation: Application and research. New York: Gardner. Seliger, H. (1989). Semantic transfer constraints on the production of English passives by Hebrew-English bilinguals. In H. Dechert & M. Raupach (Eds.), Transfer in language production (pp. 21-34). Norwood: NJ: Ablex. Selinker, L., Swain, M., & Dumas, G. (1975). The interlanguage hypothesis extended to children. Language Learning, 25, 139-152. Smith, L. B. (1999). Children's noun learning: How general processes make specialized learning mechanisms. In B. MacWhinney (Ed.), The emergence of language (pp. 277-304). Mahwah, NJ: Lawrence Erlbaum Associates. Snow, C. (1995). Issues in the study of input: Finetuning, universality, individual and developmental differences, and necessary causes. In P. Fletcher & B.
Extending the Competition Model
57
MacWhinney (Eds.), The handbook of child language (pp. 180-193). Oxford: Blackwells. Snow, C., & Hoefnagel-Hohle, M. (1978). The critical period for language acquisition: Evidence from second language learning. Child Development, 49, 1114-1128.s
Sokolov, J. L. (1988). Cue validity in Hebrew sentence comprehension. Journal of Child Language, 15, 129-156. Sokolov, J. L. (1989). The development of role assignment in Hebrew. In B. MacWhinney & E. Bates (Eds.), The crosslinguistic study of sentence processing (pp. 158-184). New York: Cambridge. Stockwell, R., Bowen, J., & Martin, J. (1965). The grammatical structures of English and Spanish. Chicago: University of Chicago Press. Tomasello, M., & Herron, C. (1988). Down the Garden Path: Inducing and correcting overgeneralization errors in the foreign language Classroom. Applied Psycholinguistics, 9, 237-246. Tr6vise, A. (1986). Is it transferable, topicalization? In K. Kellerman & M. Sharwood Smith (Eds.), Crosslinguistic influence in second language acquisition (pp. 186-206). New York: Pergamon. van Heuven, W., Dijkstra, T., & Grainger, J. (1998). Orthographic neighborhood effect in bilingual word recognition. Journal of Memory and Language, 39, 458483.
This Page Intentionally Left Blank
Bilingual Sentence Processing- R.R. Heredia and J. Altarriba (Editors) 9 2002 Elsevier Science B.V. All rights reserved.
3
A Self-Organizing Connectionist Model of Bilingual Processing
Ping Li and Igor Farkas University of Richmond
Abstract Current connectionist models of bilingual language processing have been largely restricted to localist stationary models. To fully capture the dynamics of bilingual processing, we present SOMBIP, a self-organizing model of bilingual processing that has learning characteristics. SOMBIP consists of two interconnected self-organizing neural networks, coupled with a recurrent neural network that computes lexical co-occurrence constraints. Simulations with our model indicate that (1) the model can account for distinct patterns of the bilingual lexicon without the use of language nodes or language tags, (2) it can develop meaningful lexicalsemantic categories through self-organizing processes, (3) it can account for a variety of priming and interference effects based on associative pathways between phonology and semantics in the lexicon, and (4) it can explain lexical representation in bilinguals with different levels of proficiency and working memory capacity. These capabilities of our model are due to its design characteristics in that (a) it combines localist and distributed properties of processing, (b) it combines representation and learning, and (c) it combines lexicon and sentences in bilingual processing. Thus, SOMBIP serves as a new model of bilingual processing and provides a new perspective on connectionist bilingualism. It has the potential of explaining a wide variety of empirical and theoretical issues in bilingual research.
Introduction Connectionism, parallel distributed processing (PDP) models, or neural networks have significantly influenced research in the cognitive sciences in the last fifteen years. Language, as one of the most important human cognitive components, has received in-depth treatments since the beginning of connectionist research. The acquisition of the English past tense (Rumelhart & McClleland, 1986), the recognition of speech (McClleland & Elman, 1986), and the processing of sentences (McClleland & Kawamoto, 1986) are among the earliest domains of connectionist
60
Ping Li and Igor Farkas
research in the original PDP models. Unfortunately, connectionist models or modeling have had very limited impact on the field of bilingualism as a whole. To this date, there are only a handful of connectionist models that are implemented to account for the processing or representation of the bilingual mental lexicon (see our review below). This lack of interaction between connectionism and bilingualism is lamentable, and it gives us a good reason to pursue research that would fill this gap. In this chapter, we present a connectionist model, more specifically, a selforganizing neural network model of bilingual processing and representation. Let us begin by reviewing a few important lines of research in connectionist bilingualism. In particular, we will examine the Bilingual Interactive Activation (BIA) model (Grainger, 1993; Dijkstra & van Heuven, 1998; van Heuven, 2000), the Bilingual Model of Lexical Access (BIMOLA; Grosjean, 1988, 1997; L6wy & Grosjean, 2001), and the simple recurrent network (SRN) model of bilingual memory (French, 1998). The first two belong to the so-called "localist" models, while the last one belongs to "distributed" models. In localist models, a word or a concept is represented by a single, unitary processing node in the network, whereas in distributed models, information about a word or a concept is distributed across several or many different units of processing. The localist models, that is, BIA and BIMOLA, are based on earlier interactive models of word recognition, that is, the interactive activation model (IA) of McClelland and Rumelhart (1981) and the TRACE model of McClelland and Elman (1986), respectively. In the IA model, there are three levels of nodes, with ascending complexity: (1) features of a letter such as curves, straight lines, or crossbars, (2) individual letters, and (3) words. Information at all three levels can interact with each other during the word recognition process, which may flow both "bottom-up" (features to letters to words) and "top-down" (words to letters to features). Within levels, nodes compete for activation (thus inhibiting each other); across levels, nodes either inhibit or excite each other. According to IA, it is these inhibitory and excitatory connections that give rise to the appropriate activation of patterns that correspond to the perception of words. The BIA model is a straightforward extension of the IA model. It consists of four levels of nodes: features, letters, words, and languages. As in IA, there are connections between nodes at each level and between nodes across levels. All nodes at the word level are interconnected with mutual inhibition. Feature units activate appropriate letters, and letter units activate appropriate words in the appropriate language. B IA uses the same parameters to regulate interactions within and across levels as in the original IA model. What is special to the BIA model (apart from the incorporation of two lexicons) are the language nodes (one for English and one for Dutch). Language nodes in B IA function as an important mechanism for the selection or inhibition of words in one or the other language, given that the model argues for and implements the language-independent access hypothesis, according to which words from different languages are represented in an integrated lexicon and are simultaneously contacted during word recognition. Results from B IA
Connectionist Model of Bilingual Processing
61
simulations suggest that the model is able to account for empirical results that directly support the language-independent access hypothesis, while at the same time is compatible with data previously thought to support the language-selective access hypothesis (Dijkstra & van Heuven, 1998; van Heuven, 2000). While the BIA model focuses on visual word recognition, the BIMOLA model (Grosjean, 1988, 1997; L6wy & Grosjean, 2001) aims at handling the recognition of spoken words. It was partly inspired by the TRACE model (McClelland & Elman, 1986), an IA-based model for acoustic input. Like TRACE, it consists of three levels of nodes, corresponding to features, phonemes, and words. There is no separate level of language nodes in BIMOLA, unlike in BIA. The feature level is common to both languages, but the phoneme and word levels are organized in subsets according to languages (still in the same extended system). Features activate phonemes that, in turn, activate words. Connections (both excitatory and inhibitory) are unidirectional (ascending) between features and phonemes, but bidirectional between phonemes and words. Descending connections from top down (global language activation and higher linguistic information, especially the "bilingual speech mode"; Grosjean, 1997) serve to activate words that, in turn, can activate phonemes. Language activation (or selection) takes place through these descending connections but also through within-language connections at the phoneme and word levels. Compared to the BIA model, BIMOLA can account for language-specific activation without the use of language nodes--it is yet unclear whether language nodes or language tags are necessary components of bilingual processing (Li, 1998a; see also General Discussion). The BIMOLA model is currently being implemented and evaluated against empirical data (L6wy, 2001; L6wy & Grosjean,
2001). Both BIA and BIMOLA can be said to be permanent (or stationary) models, despite their differences in input (visual vs. acoustic) and architecture (with or without language nodes). They differ from some connectionist models with a learning mechanism. In BIA and BIMOLA, the representations are fixed and manually coded, and are designed to capture proficient adult bilingual speakers' mental lexicon (but are not designed to evolve). In contrast, connectionist learning models dynamically develop representations from computing statistical characteristics of the input data (sentences). Along this direction, French (1998) presented a distributed model of bilingual memory based on SRN, a simple recurrent network (Elman, 1990) that learns representations online through sentence processing. As in the original SRN, the model takes in one input word at a time from a continuous stream of sentences and its task is to predict the next word of the current input in the sentence. The input stream is a series of artificially generated sentences of the N-VN structure, with English and French sentences intermixed at the sentence level. Elman (1990) showed that distinct grammatical and semantic categories (e.g., nouns and verbs, animate and inanimate) can emerge in the SRN's hidden-unit representations once the network has learned the next-word prediction task with a reasonable size of sentences. This is because the prediction task involves detecting
62
Ping Li and Igor Farkas
the context in which the current input occurs. With the SRN's exposure to mixed bilingual input, French showed that not only distinct categories emerge within each language in the hidden-unit representations (as in Elman, 1990), but also there are distinct patterns of the two languages: words from the two languages are separated in space on a hierarchical cluster tree of the hidden-unit activations. Note that the network incorporates no mechanism (nodes or tags) to explicitly label words in the two languages. This model provides support to the hypothesis that bilingual memory is organized as a single distributed lexicon rather than two separately stored lexicons according to language. Thus, the model can display distinct behaviors of the two lexicons without invoking separate mechanisms for each language, unlike the BIA model that uses language nodes to separate the bilingual lexicon. The connectionist model of bilingual processing that we present in this paper has the same spirit of the SRN model, but it differs from the above models in three important respects. First, our model combines both learning and representational properties. It is a learning model in the sense that lexical representations of both languages can emerge from the statistical learning of the input speech. This property is similar to that of the SRN, but is based on our network in explicitly modeling lexical co-occurrences in the acquisition of word meanings (Li, 1999, 2000, in press; Farkas & Li, 2001, in press). On the other hand, our model also has some of the representational features of BIMOLA: lexical forms are encoded by articulatory features of the phonemes of words (see also Li & MacWhinney, 2001). The representational characteristics of the lexical forms and meanings can also become clearly discernible on a 2-dimensional topological space, given the self-organizing maps used in our model. In addition, given both word meanings and word forms in the model, learning can occur in the associative links between meanings and forms via Hebbian learning, a biologically plausible mechanism of co-occurrence learning (see later discussion). Second, our model displays both distributed and localist properties. Unlike SRN or other connectionist models that use back-propagation as the learning algorithm, our model employs principles of self-organization, a type of unsupervised learning (Kohonen, 1995). Although the inputs in our self-organizing network are in the form of distributed representations, the 2-D topological map of the network (the output) bears significant similarity to localist representations in that each unit on the map tends to represent one lexical item. The localized patterns on the map allow an "explicit" representation of the lexicon, instead of "implicit" representations as in the hidden-unit activations of an SRN (MacWhinney, 2000a). At the same time, each unit is surrounded by neighboring units on the map that can become coactivated, simulating a distributed lexicon in which similar words or word properties are grouped together (see details below). Third, our model relies on the use of realistic linguistic data as input to the network, in particular, child-directed parental speech. In the SRN model of bilingual memory (French, 1998) as well as many current connectionist models, researchers
Connectionist Model of Bilingual Processing
63
have relied on the use of highly simplified, artificially generated input. Although such inputs are easy to construct and to control for, they are often isolated from realistic language uses, and run the risk of being successful just because of the use of certain properties in the input (see Lachter & Bever, 1988 for an earlier criticism of connectionist representations). For example, French (1998) structured the input data in such a way that the input has a fixed probability of 0.001 of switching from one language to another. In other words, on average, the network will first have learned 1000 sentences in one language before learning any sentences in the other language. We suspect that this artificially determined probability is what causes the network to display disparate behaviors for the two languages. To overcome potential problems associated with this approach to network modeling, we thus rely on corpus-based linguistic data that closely approximate the reality of language use (see also Li, in press, for discussion). In what follows, we first present an overview of our model, SOMBIP, a connectionist self-organizing model of bilingual processing. We then discuss some preliminary simulation results from the model. We conclude by showing how our model can shed light on important issues in bilingual lexical and sentence processing.
The SOMBIP Model
Background The design characteristics of the SOMBIP model are based on our selforganizing neural network model of language acquisition by young children (Farkas & Li, 2001, in press; Li, 1999, 2000, in press). In recent years, self-organizing neural networks have become increasingly important for cognitive and perceptual studies (Hinton & Sejnowski, 1999). Although significant progress has been made with models based on back-propagation, there are known limitations associated with these models, including catastrophic forgetting (inability to remember old information with new learning), scalability (inability to handle realistic, large-scale problems), and above all, its error-driven learning process, a procedure which propagates error signals from the discrepancy between desired and actual outputs to adjust weights. Some of these problems become most transparent when considered in the context of language acquisition (see Li, in press). Consideration of these problems lead us to look for models that bear more biological and psychological plausibility in the context of language learning, and we turn to the self-organizing models. One of the most widely used self-organizing models is Kohonen's (1982, 1989, 1995) self-organizing map (SOM). SOM belongs to the class of "unsupervised" neural networks, because learning in these networks does not require the presence of a supervisor or an explicit teacher; learning is achieved by the system's self-
64
Ping Li and Igor Farkas
organization in response to the input. During learning, the self-organizing process extracts an efficient and compressed internal representation from a highdimensional input space and projects this new representation onto a 2-D topological structure (Kohonen, 1982, 1989, 1995). Several important properties of SOM and related features make it particularly well suited to the study of language. (1) Self-organization. Self-organization in these networks typically occurs in a two-dimensional topological map, where each unit (or node, or neuron) is a location on the map that can uniquely represent one or several input patterns. At the beginning of learning, an input pattern randomly activates a group of the many units on the map, according to how similar by chance the input pattern is to the weight vectors (codevectors) of the units. Once a unit becomes active in response to a given input, the weight vectors of the unit and its neighboring units are adjusted so that they become more similar to the input and will therefore respond to the same or similar inputs more strongly the next time. In this way, every time an input is presented, an area of units will become activated on the map (the so-called activity "bubbles"), and the maximally active units are taken to represent the input. Initially, activation occurs in large areas of the map, but gradually learning becomes more focused so that only the maximally responding unit or units are active. This process continues until all the inputs have found some maximally responding units. (2) Representation. As a result of this self-organizing process, the statistical structures implicit in the high-dimensional input space are represented as topological structures on the 2-D space. In this new representation, similar inputs will end up activating the same units in nearby regions, yielding activity bubbles that become clearly visible on the map. Such a self-organizing process and its representation have clear implications for language acquisition: the formation of activity bubbles may capture critical processes for the emergence of semantic categories in the acquisition of the lexicon. In particular, the network organizes information first in large areas of the map and gradually zeros in onto smaller areas; this zeroing-in is a process from diffuse to focused patterns, as a function of the network's continuous adaptation to the input structure. This process allows us to model the emergence of semantic categories as a gradual process of lexical development. It naturally explains many generalization errors reported in the acquisition literature (elg., substitutions of put for give or fall for drop as reported by Bowerman, 1978, 1982). It also explains language disorders that result from the breakdown of focused activation or the inability to form focused representations (Miikkulainen, 1997; Spitzer, 1999). (3) Hebbian learning. Hebbian learning is not an intrinsic property of SOM, but several SOMs can be connected via Hebbian learning, such as in the multiple feature-map model of Miikkulainen (1993, 1997). Hebbian learning is a wellestablished biologically plausible learning principle, according to which the associative strength between two neurons is increased if the neurons are both active at the same time (Hebb, 1949). The amount of increase may be proportional to the level of activation of the two neurons. In the multiple SOM model developed by
Connectionist Model of Bilingual Processing
65
Miikkulainen, all units on one map are initially connected to all units on the other map. As self-organization takes place, the associations become more focused, such that in the end only the maximally active units on the corresponding maps are associated. Hebbian learning combined with SOM has strong implications for language learning: it can account for the process of how the learner abstracts relationships between word forms, meanings, and grammatical morphology, on the basis of how often they co-occur and how strongly they are co-activated in the representation. Because of these properties, SOM (a) allows us to track the development of the lexicon clearly as an emergent property in the network's self-organization (from diffuse to focused patterns or from incomplete to complete associative links); (b) allows us to model one-to-many or many-to-many associations between forms and meanings in the development of the lexicon; and (c) provides us with a set of biologically plausible and computationally relevant principles to study bilingualism without relying on corrective feedback. It is ideally suited for our task also because the bilingual mental lexicon is constructed for the most part without supervision and undergoes continuous self-organization over time.
Architecture
Our SOMBIP model has been inspired by the multiple self-organizing featuremap model of Miikkulainen (1993, 1997). Miikkulainen proposed an integrated model of memory and natural language processing, in which multiple SOMs dedicated to different levels of information are connected. A sub-component of this model is DISLEX, a SOM model of the lexicon, in which different maps correspond to different linguistic information (orthography, phonology, or semantics) and are connected through associative links via Hebbian learning. Our model has also been inspired by the Hyperspace Analogue to Language (HAL) model (Burgess & Lund, 1997, 1999). In particular, we derive our meaning representations of the lexicon through a word co-occurrence detector (WCD), a mechanism similar to the principle of HAL computation. HAL attempts to capture meaning by reference to global lexical co-occurrences--how many words co-occur with the target word, and how often, in a moving window that runs through a large-scale language corpus (Usenet texts of 300 million words). A co-occurrence matrix for any number of words in a given window is derived, and weighted by the frequency of co-occurrences among words. Thus, a word is represented in HAL as a vector of the column and the row in the matrix that encodes the co-occurrence constraints in a high-dimensional space of language use.
66
Ping Li and Igor Farkas
word form ) self..organizatilon
SOM 1 lexical (phono) map Hebbian I ] leaming
~vord meaning] self-organization
SOM 2 semantic map
Figure l a. SOM1 and SOM2 that self-organize on word forms and word meanings, respectively. They are interconnected via associative pathways, trained by Hebbian .. learning.
SOM2 semantic map
Tm I reduced word repr I
Iq t
representation of word meaning
l qzt~
layer A (''
i""
(w
9 ord localist cod
. ) -~"~,~
one-to--one copy
layer B
* * CI
j ---
= CN
Figure lb. WCD (lower panel) that computes lexical co-occurrence statistics. Layers A and B have full connectivity via modifiable i and r links. Other, one-toone links serve to feed the unit activity from A to B with discrete single time-step delay. SOM2 receives distributed word representations extracted from the modifiable links. The representations are reduced to vectors of uniform length (100 dimensions) as inputs to SOM2 (see Farkas & Li, 2001, in press, for details).
Figure 1 presents a diagrammatic sketch of the different components of SOMBIP. Figure l a depicts the two SOMs used in our model and their interconnections. During learning, a lexical form (phonological input) activates a unit or a group of units on SOM1, and simultaneously, its word meaning (semantic input) activates a unit or a group of units on SOM2. Note that in our current
Connectionist Model of Bilingual Processing
67
simulations we have used only phonological input in SOM1 to simulate the bilingual spoken lexicon; one can easily use orthographic input in SOM1 to simulate the bilingual visual lexicon. SOM1 and SOM2 are connected via associative links, such that the activation on one map can cause an activity to form in the other map. If the direction of the associative activity is from phonology to semantics, comprehension is modeled; if it goes from semantics to phonology, production is modeled. The associative links are trained by Hebbian learning, and the strengths of the connections are adjusted according to the form-meaning pairings in the input, which leads to adaptive associations between the two SOMs. Figure l b depicts WCD (lower panel), a special recurrent neural network that leams the lexical co-occurrence constraints of words. The WCD reads through a stream of input sentences (one word at a time), and given a lexicon sized N, it computes the transitional probabilities of words in the lexicon (see Farkas & Li, in press, for details). Assume that at time t the current word is i, and is represented by a localist vector o - [o 1, ..., ON] in layer A. Previous word j is represented by the vector c -- [c! . . . . , CN] in layer B. The adaptable connections (! and r links) between layers A and B serve to approximate the transitional probabilities between successive words, and as such, they are trained by Hebbian learning with weight decay so that they become normalized. Specifically, the link l/j is updated to approximate p(jt-l[it) (i.e., the probability that the word j precedes the word/). At the same time, the link rji is updated to approximate p(it~t-1), that is, the probability that i follows j. Word i is characterized by a concatenation of vectors !i = [lil, ..., liN ], and r i = [rli . . . . , rNi ], where !i approximates the probability distribution of words preceding i (left context), and r i the probability distribution of words following i (right context). The concatenated vectors, qi = [li, ri], then serve as distributed word representations to SOM2 (upper panel in Figure l b). Because the dimensions of the vectors are determined by the size of the lexicon (2N for any given vector), the vectors, before they are read by SOM2, are also submitted to a dimension-reduction mapping mechanism, which reduces the vectors to lower, fixed dimensions (e.g., 100 units, see Farkas & Li, 2001). As SOM2 takes representations from the WCD vectors qi, SOM1 also takes as its input the phonological representations of words. To represent the phonology o f the bilingual lexicon, we have followed a syllable-based template coding originally developed by MacWhinney and Leinbach (1991) and recently by Li and MacWhinney (2001). This coding scheme has the advantage over traditional phonemic representations in that it can accurately capture the phonological similarities of multisyllabic words (most previous connectionist models have used only monosyllabic words as input). In this scheme, the phonology of a word is made up by combinations of syllables in a metrical grid, and the slots in each grid are made up by bundles of features that correspond to phonemes, C's (consonants) and V's (vowels).
68
Ping Li and Igor Farkas
To code our Chinese-English bilingual lexicon, we used 4 C-slots and 4 V-slots that allowed for the representation of words of one and two syllables, in the template of CVVCCVVC (each CVVC is a syllable; the first C represents the initial consonant, and the last C the final consonant). Thus, the Chinese (Cantonese) word jat (one) is represented in the slots as jaVtCVVC and the English word about is represented as C@VCbaUt. Each C or V is represented by a set of 5 feature units, and the feature values (scaled between 0 and 1) are determined according to the articulatory features outlined by Ladefoged (1982) for English and by the Hong Kong Linguistic Society (1997) for Chinese (the 5 articulatory features are: Sound, Place, Manner, Chromacity, Sonority). For example, the phoneme/i/is represented in both languages as [. 1 . 0 . 0 . 2 . 3 ] , indicating [vowel none none high front] for the 5 features. A separate set of 12 units are used to represent lexical tones in Cantonese (6 tones for each syllable), whereas these units are left empty for English. Thus, each word in the bilingual lexicon is represented by a vector of 52 units in the phonological representation (5 feature units for 8 phoneme-slots plus 12 tonal units). 1 Note that in neither the phonological nor the semantic representations described above is there a label or tag that tells which lexicon (English or Chinese) a given word should belong to. Learning in the two SOMs is standard (Kohonen, 1989). Every SOM unit k has an array of connections in the form of a codevector m k = [ink1, ..., mk,2N ], which learns to approximate the inputs (semantic or phonological vectors) in such a way that every SOM unit becomes "specialized" for a given word, and that the neighboring units will become specialists ("winners") to similar words. During learning, both neighborhood radius and learning rate decrease over time. Task and Procedure
Upon training of the network, a phonological input representation of a word is presented to SOM1, and simultaneously, the semantic representation of the same word is also presented to SOM2. By way of self-organization, SOM1 forms an activity in response to the phonological input, and SOM2 an activity in response to the semantic input. As the two SOMs receive input and continue to self-organize, they also simultaneously form associations through Hebbian learning for all the active units in the two maps that respond to the inputs. The network's task is to create an ordered layout for all the input words in the bilingual lexicon and be able to make the appropriate form-meaning associations. Because our SOMBIP handles a bilingual lexicon, translation equivalents in the two languages are associated with each other in the following way: if the phonology of an English word is presented to SOM1, the semantics of the English word and that of its cross-language translation equivalent in Chinese are also presented to SOM2. Similarly, if the semantics of an English word is presented to SOM2, the phonology of the English word and that of its translation equivalent in Chinese are also presented to SOM1. For example, the word boat and syun are associated by the
Connectionist Model of Bilingual Processing
69
phonology of boat or syun co-occurring with the semantics of boat and syun, and vice versa. This procedure works in the same way for words in both languages, ensuring that translation equivalents have a chance to be associated. Because of the difficulty in determining what words should be counted as translation equivalents (especially with regard to English and Chinese), this procedure applied only to the nouns and verbs in our simulations (i.e., disregarding adjectives, adverbs, pronouns, prepositions, etc.). Multiple translation equivalents (e.g., English tell and Chinese gong and waa) were also associated through co-occurrences in the network, although such cases were rare in our data. As discussed earlier, artificially generated input data are often problematic in matching up with realistic language use. In this study, we used a realistic bilingual data set, the Hong Kong Bilingual Corpus from the CHILDES database (Yip & Matthews, 2000; MacWhinney, 2000b). This corpus contains transcripts of conversations between a child and his native English-speaking father and native Cantonese-speaking mother. The recordings were made when the child was between ages 1 and 3. The parents followed the one parent-one language principle when addressing the child. The language between the parents was mainly Cantonese with a great deal of English mixed in, as is characteristic of the speech of Hong Kong middle class families. Despite the "one parent-one language" principle, the quantity of input from the two languages was not all balanced. On the whole, the child received more Cantonese than English input in his first three years. Because of the relatively young age at which the recordings were made, there was not enough productive speech from the child. However, there was plenty of parental speech as input. We therefore extracted all of the parental speech plus the speech of the student assistants who made the recordings during the home visits, forming the bilingual input speech corpus that we used for our simulations. These speech data also effectively allow us to simulate what the leaming system (human or network) receives in a concurrent bilingual environment and how the system can, on the basis of the input, develop lexical representations from sentence processing (we used a similar procedure in modeling first language acquisition; see Farkas & Li, 2001, in press; Li, in press). A total of 5,154 word types and 185,279 word tokens are found in our bilingual parental corpus, according to the freq (frequency count) output of the CLAN (Child Language Analysis) program (MacWhinney, 2000b). For our purposes we trained our model on the 400 most frequent word types in this corpus, which effectively covers 56% of the entire data. These 400 most frequent words happened to contain 184 Chinese words, and 216 English words. During training, SOM1 received the phonological representations of the 400 words and self-organized on them. The WCD network of our model received the words in the input sentences one at a time, and built semantic representations from the lexical co-occurrence statistics. These representations were submitted to our dimension-reduction mapping so that all vectors had a uniform length of 100 units, and subsequently sent to SOM2 for selforganizing learning. SOM1 and SOM2 were accordingly linked by associative
70
Ping Li and Igor Farkas
links. Note that training in SOM1 and SOM2 was simultaneous, and therefore learning in SOM1 and SOM2 did not begin until the WCD network and the dimension reduction had completed their job. In principle, however, SOM2 need not wait until WCD is done, given that it can self-organize on early stages of semantic representations as WCD learns online. This method has also been implemented in our model as an incremental learning process (see Farkas & Li, 2001).
Results and Discussion In this section, we report results from several simulations with SOMBIP, and our analyses will focus on the network's performance with respect to distinct behaviors in the two languages, the formation of lexical categories, the interlingual priming and interference effects, and effects of proficiency and resource limitation.
Language Separation Without Language Nodes As discussed earlier, an important difference between different connectionist models of bilingualism is whether the model explicitly includes a separate level of language nodes. The BIA model does, whereas the BIMOLA and the bilingual SRN models do not. In B IMOLA, the feature level is common to both languages, and the phoneme and word levels contain subsets of units for each language. The differences between these subsets, coupled with global language activation and higher linguistic information from top down, allow the system to separate the two languages. The SRN approach of French (1998) is more radical, in that it makes no a priori assumptions about the differences between the two languages in the bilingual lexicon, but simply lets the system learn the bilingual (artificial) sentence data. However, the way input was structured in the SRN was problematic, as we discussed earlier. In our model, there is no place for explicit marking of languages, as in the SRN model. We trained our model on a realistic parental input corpus with bilingual speech intermixed between Chinese and English. Figure 2 presents a sketch of the phonological and semantic organizations of the bilingual lexicon in the two SOMs, after the network has been trained on the 400 target words for 500 epochs (the WCD network was trained on the input sentences for 3 epochs, i.e., 555,837 word tokens). 2 As can be seen, our network clearly displays distinct clusters of lexical representations of Chinese from those of English, on both the form (SOM1) and the meaning level (SOM2). Note that because of the topological nature of the maps and the "bubble-filling" nature of SOM learning (Kohonen, 1995), the border between the two languages is not a straight line in either case. Because the sketch leaves out the details, cases where the English and Chinese representations are intermingled are not shown here (but see our discussion in interlingual interference). In general,
Connectionist Model o f Bilingual Processing
71
the intermingled cases are uncommon, and the overall separation of the two lexicons is clear on the 2-D maps.
i-:-.:~::.-.:~:i~~-..--..--i :.~.~..i :-.: :. ~.~:"--.- - ~:,~!:'..:.::..7-::4..:-...::..: .,-".".",",..",.".--',-,-,,,--"-...,..-.".,'."-",,"-'
~---.--i......:i~i!i!~!ii~i. :-z:!i:::::~i!?!~i~!;5!i!~!ii~ii! ~iii~ii~............~!!.i~i-!iiiiii~!~-i-!iIii~i!~i!iii~iiii!~ :~"~:~ !~i-:-:;:....;:-:~!i!~iii
ii:i:"-"~!iiii:i!;'iNiii!!!i?!ii!N~iiii?ii ii!iilNiliiiiiii!Niiiiii . .!i. i.!i.i.i. . . . .ii!iNi . . . . . .iiiiiii!!!! . . . .!~i!!~!i!iiii~i!!iiii!~~~ii!:;i!~ii~!!ilN.........................................................................................
i~i~: ~::~.~:~:~`;i :.:;i:~:::::i`::::::;:..::::i.::..:.~;:.`:~.::::. ::::;`:::::i~::::::~;;;Z:::.:::: ~....... ::::::~`~. ~!!~.~:-~',~:~".~ :.~ ......:!~::~-::~::-~ii~~%~.,.~ ~i "'".,.'i~Z~'s :: ...... ; ...... .?" " " : ..*""" ':" "" ~....... : ..~'.~i~'':':"'"~'""'-:"".~'""':r"~'v"!-'"": 9 :','.'~..-~:;'.-. .;.~::~::-~'. -'.'" ~..'..~ .!~;i.'..'-:~ '~..':.'.~.~: 9
...... ~............. !-...... i....... i"--.~-~ ............. ~.-; ...... i ..... ~............ i..--'.". " .~...... ~t.~ ............ :~i; :'~g~i
,~'"'i ........ "-'~"i-'-~! ': ....... ~'~-':~'-"-,~---i.'"~ '-~-
!:'-,'-+",~:! ...... ='-~':-~
. ........................... . . . . . . . ............ . . , . ............... . . . . .
!~~....'~;~;i :"~#~i :"~:~.~g f~;.~ i'~ i:~g~ ~gg~.~ggi g~i :~;:~":;:i:~gi!ti..:g~g~ :;::::i:;:~.~g~ ::~~'~~i..":~
...... ~i---'"'i'.,'~:~!..;;-:-",,;~;-~a:~;;!,~;;~:~-~
....... i~g~-i
~!-!-i;:~.:-i;i~i.'~:~.:-'.:::;!.~$;~.'i;i~Z-'."i~.;i~i ~
.".-.:......!
~i~.~;~
~:~
:~
:~i~;~":~;~;:~'~i~
~i~:;:~.~;~.~.~.~.~
iZ:~.)5. ~--?,e~!~!2[-?i2--12-2.2..}~?~ .?--.)...<-~N:-<--2{--i.?-~:2. ?. i..? .i:1!-.-72.1~2ii-~2,i... . :~ii:.-i--....)??~2:1!.?~.?...i..-.-2~ .?.i--i.i-.~; -,- .i
Figure 2. A sketch of the phonological and semantic organizations of the bilingual lexicon in SOM1 and SOM2 after the network has been trained on the representations for 500 epochs. The upper panel represents the lexical form map (SOM1), and the lower panel the semantic map (SOM2). Shaded areas indicate the English phonological and semantic representations, in contrast to the Chinese representations.
Our network received no labels or tags for which language a given word should belong to, yet the self-organizing process was able to separate the two lexicons. On the phonological level (SOM1), it is clear that English and Cantonese differ greatly, even though some vowels may share the same quality. In our phonological representation scheme, we have tried to adapt the template in such a way that it could fit both languages (CVVCCVVC), cutting some comers in English (e.g., eliminating the representation of initial consonant clusters). This adaptation does not seem to have much effect on the cross-language distinction. Overall, the syllabic templates, the phonemic inventory, and the lexical tones have given the learning
72
Ping Li and Igor Farkas
system a great deal of differences to explore for the identification of the two lexicons. With respect to semantics, SOM2 shows an even clearer separation between the two languages. Note that our network performed essentially a sentence processing task on the speech corpus, in which the WCD reads through the streams of sentences, computes the co-occurrence statistics for each of the 400 words, disregarding words not in the target lexicon (treating them as one unknown type). The words that precede and follow each of the 400 words in the lexicon form the contextual history for that given word, and it is this contextual history that captures the use and meaning of the word. Thus, the words in the two lexicons will have very different contextual histories, because words in English are typically associated with English words, and words in Chinese are typically associated with Chinese words. This is true even with a mixed bilingual corpus like ours, in which English and Chinese sentences are intermixed. From the learner's perspective, the input contains both English and Chinese sentences, but it is relatively consistent whether a given sentence is English or Chinese, because the intermixing occurs at the sentence level, but rarely within the sentence level. Although it is common for Hong Kong Cantonese speakers to code switch, the most common code-switching situation is one in which the speaker uses an English guest word in a Chinese sentence (Li, 1996a). Given that the code-switched guest words are often isolated instances and their associations with the lexicon of the borrowing language are relatively weak, our network did not get confused about the identity of the two lexicons. Overall, our simulation results match well with empirical studies of the bilingual lexicon that argue for the representation of language-specific (or distinct) lexicons, but in an integrated network (e.g., Kirsner, Smith, Lockhart, King, & Jain, 1984).
Lexical Categories in the Bilingual Lexicon Given that our model was able to clearly differentiate the two lexicons, does it show any ability in developing categories within each language? In other words, is our network able to organize the phonological and semantic features into structurally meaningful clusters? Our previous research shows that SOMs are good at deriving semantic categories of the lexicon, even in cases where linguists fail to identify the categories with clear labels (e.g., Whorf's cryptotypes; Li, 1999, in press). A further examination of Figure 2 reveals that the network developed a considerable amount of structure on its semantic organization. 3 Figure 3 presents the same results as in Figure 2, with illustrative labels on the lexical categories in SOM2. In Figure 3, we can see that the SOM2 treated the nouns and verbs distinctly, in both Chinese and English. The network also grouped other words together as clusters that share the same grammatical categories (e.g., English prepositions and pronouns occurred in the same neighborhood). More interestingly, within each of the categories (assuming there are enough instances), semantically similar words
Connectionist Model of Bilingual Processing
73
also occurred together. For example, in English, the state verbs (e.g., know, like, have, want) were grouped together (the darkest shaded area under the label "Verbs"), in contrast to other activity verbs; 4 in Chinese, words that are related to cooking/eating (within the verb category) were grouped together (e.g., sik 'eat', jam 'drink', cit 'cut', gaau 'gnaw'). One can also see a category of verbs indicating perceptual/mental activities in Chinese (marked as "Verbsp", including teng 'listen', tai 'look', gin 'see', seong 'think'). These effects of categorical emergence from statistical learning match up with results from our previous monolingual simulations (Farkas & Li, in press; Li, 1999, 2000, in press).
~-~,r
~i~;,;
i,~#n
.
.
.
.
I.~I;C~IL
~
l i~!i~}}i~}}}~.:~.~}~(~}~({.}}~!~}~!i.~E(.}!~ii~}}}}}}!}~(.~!~i~!~!}}~(~}~!~}}}~i~(~}~}~{!{i~i:~{i!.}!~}}~(~}~.~.~.~!.~!~!(~.~!~{~i!!~}}~i{~}}}~}}~!~:`.}}~(({.}}}}~{~}~{.}ii~.}~i!.}}}}}~.~{.}}!}}~((.if..~!.}}~.~!i.~.}}}}
I~;i---.~ ...... i ''':~ . ................. ~,.a~. ...... :.....: ~...~'.,~-~-....:=.::..-......'. ........ :..~....:. .:...:-.:....~:.,:-..:.~...::i.~;.~.i......!..::.-!~.~'~.~..,;~ ..... : ............
:. . . . . . . :. . . . . . . :. . . . . . . : " ". . . . . . . . . . . . . . . .
J ......
:
~ : :
:
:.~.,~:
"
"
~
-
:
-
:
....... ....:.~.:.~z.`Lq.:.~:a..~i.i.~...::~.i~.::i~:~i``..:.i~?~.~it~i*.~.t.:~.?:~7?.!.i!~176162 :
.:,~:,
t
~
;~:~::: : ~ :
::
~
;
,,:::,;
,,;
;
~
::~::
:t
: ~
:"~:i :.
:
~i~:::::~i~*~.~:.~.::::.i~:::::~[~i.~:?.$ii~!~i~].iik~i].tiii~;t!~!.]i..::~!!~.t!~.~ .!~!~:~t~i~(~.~:~i!:!~i:i:i~/:~/.:.i~.~.!i~t::::~!~:::~.~i~:i~:~!~t:~:~.~f::!~.~!~:~i~i~.~.t!:.?.!..:i!~.~i!it~:~it!.!?:ii~:~?.:~5::i~!.~ii!it..:r
I ~...... i~...Z..~:F.....;.~....:.~.:..:.5......................~T.~:..~:...;~...*.~..~.....ZZ7;..Z
. .S?.':?.'I ~.....~s163
i....... i....... ...... ~...... !...... {.............~............. i.-... ~a:~/ 'i~:!:::ti::b"];"i
"'.Y'.'.'.T:.-'FTS~:
.;-..... ~.----.~-.*......-..';.,.-:.~.---.:i...... [.-.--i ...... i-..--.~:,:~-:ir
-:;:i~;:--i~-~-::::~i-:~:;~:,.'!~o~-};{-:ti~t -.~-q*:~--ff~r
i::x:............:.::;;........~':~:............:......: 7........:~......... : ;:~-:::-:.:-:i:-~-~i::.::~~i~
i-*i
;
:Verb.sp
:
i
i i
i
:
:: i i:~
i~i!!!:!ii~i~!;~!~i!!ii!!!~!r
~i~!~!i!~!iiii~i!~i~!!i~!!i~
ii!iiir162
Figure 3. Emergence of lexical categories (e.g., nouns, verbs, and subcategories in Chinese and English) in SOM2.
Interlingual Priming and Interference In the empirical literature, there are heated debates on whether bilingual lexical representation is language-specific or language-independent, and whether the two lexicons are stored separately or together (see Smith, 1997, for a summary). These debates determine how researchers look at well-known effects of cross-language
74
Ping Li and Igor Farkas
priming and interference in empirical studies (e.g., de Groot & Nas, 1991; Kirsner et al., 1984). In a way, our above simulation results already indicated that the two lexicons can reside in an integrated network, but separately in space. Given that the two lexicons exhibit distinct behaviors in our network, and that there is no language-independent conceptual representation in our network (as is the case in many other bilingual models, see Conclusions), how can we account for priming effects or interference that show that the two lexicons do interact with each other? Recall that our procedure involved training the mapping between word forms and meanings in such a way that the mapping is not one-to-one, but many-to-many (for nouns and verbs, at least). Thus, words that are considered as translation equivalents are linked together through associative pathways that are trained by Hebbian learning. After training the network on the bilingual speech corpus for 500 epochs on the 400 words, we tested the network's performance on comprehension (from phonology to semantics) and production (from semantics to phonology). Probing into the network's associative pathways is laborious, which involves examining each of the 400 words on SOM 1 or SOM2 to see what words it can activate on the other map. Our analyses indicate that there are clear effects of priming and interference. To illustrate with examples, the activation of the following words on SOM2 (let~) led to the activation of words on SOMI (right) via the associative links (items in parentheses are translations of Chinese words): SOM2
SOMI
bear --) bear, there, fish, this jat (one) --)jat (one), loeng (two), soeng (think), saam (three), san (new) bibi (baby) --) bi4bil (baby), maami (mummy), baby, maybe, mummy sik (eat) --) it, eat, sik (eat), maai (buy) These are some representative examples that indicate semantic as well as phonological priming effects in production. For example, bear activates fish because fish was semantically related to bear and SOM2 placed it next to bear. There and this were activated because of their phonological similarity to bear and fish, respectively (SOM1 placed there next to bear, and this next to fish). The second example in Chinese (/at) can also be similarly explained. These two examples show only within-language priming effects--no translation equivalents exist in our simulated lexicon. The next two examples show cross-language priming effects: mummy became activated because of its semantic similarity to baby, which in turn was activated by the Chinese word bibi. In addition, maybe was activated because of its phonological similarity with baby on SOM1. The last example, sik 'eat', is a similarly interesting case. Its strongest response on SOM 1 was it and not eat, a phonologically similar word. Note that these priming effects represent possible interference in language production. Because of strong associative
Connectionist Model of Bilingual Processing
75
pathways, the speaker might have said it instead of sik or eat, especially in a strong bilingual speech mode (Grosjean, 1997). Similarly, our model was also able to show cross-language interference effects in comprehension. For example: SOMI SOM2 look -9 lok (fall), fall, look, tai (look), ceot (go out). boat ") boat, dog, car, syun (boat), sau (hand) waan (play) --) waanx (play), wan (look for), play, read, find, see sik (eat) --) eat, cut, sik (eat), maai (buy)
Again, cross-language priming resulted from both semantic and phonological similarities as a function of the co-activation of forms (SOM1) and meanings (SOM2). For example, because of the phonological similarity of look and lok on SOM1, look activated lok in SOM2, which in turn activated fall as a result of semantic similarity (lok means "fall"). Li (1996a) found that bilingual ChineseEnglish speakers do not need more time to process cross-language homophones like look/lok than do monolinguals in processing monolingual materials, when the homophone occurs in a sentence context; in other words, there was no switching cost associated with the recognition of code-switched English words. This finding contradicts the notion of a language switch or monitor mechanism that says that switching languages takes time (Macnamara, 1967; Obler & Albert, 1978; see also Green, 1998). Li suggested that a parallel activation mechanism might be involved: multiple candidates in the bilingual lexicon are activated, and sentence context plays a crucial role in affecting the final activation of the target word (see also Altarriba, Kroll, Sholl, & Rayner, 1996, for a similar argument). The results shown in this simulation are consistent with the view that multiple candidates may be simultaneously activated and selected due to effects of cross-language priming.
Effects of Proficiency and Individual Differences
To extend the scope of the SOMBIP model discussed above, we modeled the effects of proficiency and individual differences in two separate simulations using the same bilingual lexicon. Proficiency was modeled by having SOM2 self-organize on "novice" word representations--representations that come out of the WCD network's learning of a smaller sample of English sentences (92,640 word tokens, as compared with 555,837 tokens in the above), simulating a beginning Chinese learner of English. Figure 4 presents a snapshot of the network's representation with the same learning parameters as in Figures 2 and 3. As is illustrated in SOM2 of Figure 4, the semantic representations of the English words were squeezed into the upper right-hand corner, and occupy only a very small portion of the map. Although there was some rudimentary structure in these vectors (e.g., good and bad were nearby on the map, so were say and tell), the
76
Ping
Li
and
Igor
Farkas
semantic vectors from the WCD network were still fuzzy and did not represent the word meanings accurately; thus the SOM could not fully differentiate them and accordingly organize them into meaningful structures. This picture shows that for novice bilingual speakers, the bilingual mental lexicon is still largely monolingual, and words in the second language are largely clamped together. It is also consistent with the view that the bilingual lexicon is a dynamically developing system, one that changes as a function of which language becomes the dominant language (Heredia, 1997). As the bilingual speaker becomes more proficient and fine-tunes the semantic system (amongst other developments, e.g., an expanding size of the L2 lexicon), this "dominance" situation may shift to an L2 instead of L1 dominance (Heredia & Altarriba, 2001).
~i~i}i.}iii}i}}i.i.i~i~i~i.}i-iii:!.i.i.ii:}i:~'i.~.~.~.~i}!~!~}}}}~}~.~.i.~.}}}}}i~:}~i.}~i~}}~.~.}~5~.}}}~.}!.}~.~?~!!!~.~}~!}}~.~.~!~.' !il .:.!~!.~.}}~.}}i~}}~.!!!!~i~:~.}}}~.}}}?~}!.~.E~}}~}~.~.~.~:i!...-.!!~!~.~.~.~}~.}~.~.}~}}~})~.:!!~ ~}}~.}?~!~!~!~}}~..}~}~}}i!}.-..!:.~!~.}}~.}~}~.~.~?~!f?~f~f!~l~}~}~}~iii~i~i~ii~i~ i":
..:
~ . ..... .. i i ~'.~, 9. ~ i . ~..~.7. i . :!:~,~: ..i. :i,,~: ~
..;...i 9 .i;... ::.:............:...i..:....... ..i.:...-i:...:.... ::.i...::.!.:..::.....;. ::.:...:..::..i,::.:..i....:..iii i ! ' i ~
.~
~ ~"~
,
~-~. ~:,~: ~
~
i
~
~i,~,
~ii~i,~.,'.:.~i~,,~,~.~,~~!,
..... i ..... ~""~. . . . . .
! z ......
~!~'~;~F~!~;!:~::i~i!...`~:!.~:~.~;!.~i.i!:;~i;:~:i;~;..:i ::
...... ~...................... ,~t~. ............. ! ...... ~.............. i. ................... ~............. ~ , . . , ~ ! i ~ ? : . ' ~ i i : ~ i ~ i ~ T J ~ ' ~ : - ' ! ~ ' ~ ; ~ : . . . : : i ~ } ~ : ~ t~::~7~'~i~i~it&~,,~ff~i~?~g2ff~-":~i~.:~22:.':~i:!:2 :'~i~::.:~::~ii;~::~-;"2~:~-:::.::~"!?.~!'-:'..;-:~;~i ..................... ~-.'.. ~ ...... i ............. !....... i......~., ............................ "............. : , ~ : ! ~ : i : ! ~ i ~ ; ~:i~ ~:~:~:::!~.'.:~:!~:!~i~:~ ~ ~:~;~:!'.`~.~;~:~:.~i~;i`~!:;~?~:.~::~!:i~!~!~:~::::i~i~::i~ ii !,~ii~:..:~ !~: i:ii*ti~i-:~'.-':i ~ :i!
....~:.~........ -,,~ ........ ,~,,~: ........:,.~.~............. i...... .............: ~ . . . : ` ~ i ~ : ~ r ~ ; ~ i ! ~ ; ~ : : ~ : ~ ` : . : : : : ~ ` : ~ : ~ ! ~ ` 7 ~ : ~ ! ~ ` . ` ~ i ! i ~ ~ ~ ! i ~ z ~ i ~ ` ~ : ~ ; ~ ;..:;:.i..::...~.:::.::::::::i'.:.::::i~;::.:~:...:::.i..;..:i:..::::::::i~:~:.:~:.,.~::::..i.:.:.~.i::..'...:~~.~:~.;~...:.~.'::.':'~::::..i.::.:..~.~.~ ~:~2~.~.~::C~.~:~.~!~~~g~i!i~ii;!~
',.,.:............:....!.......i,,,,,!.......
'\~!;!.!~~~'.ii! ~~!.!i!~!~,~;~i~
~'~}~.~.~'. ~ : i!ii~zr
i..~i...............~,~....... ,: i:~...............i2 ............i.~.~.i..~.................i::.:,i.......,!i~i~N!i~ii;iii~iN~}}}~ii!iN!ii~:!iiii~}}iiiiii!i~!!iii}i!~i;~ili
7g'~.i............;...........~.':.~.~:!..i..~..<2..~.........~....;.....~.~.~.~.......i.....:..:..~.;..~.~`.~..................:.....!......:......;`.~:.~........:...2.........~...~.:..:.~i:;.~`~...............`~:..... .,,.:~=~,. ,,,u.~, ~ ~ , . ~ .
Itt:~MilNIi~
~1-~
..
.............. i............. i.~..~.:- ............. i.'.~.~ ............... ~:~,.ei..-...:.~,,} ............ ~...... i ...... i-...--:-.-,.,.i. ............ :~,..--i ...... i.-----}-.....i ............ ~-.---.i ...... i ...... ;~.~-: ................ ,':-.~:i...... ::........ ~ ~ " , , i , ~ i ! ~ ............................................. : ..... :~,,~. . . . . . . . . . . . . . . . . . . . . . . . :...... :. . . . . . . . . . . . . . . . . . . . . . . . . . . . ~. . . . . . . . . . . . . -..-...~,~:,~...~ ..... ~ . ~ ~ ~ ~ ~: :~: ~ :,',-~ ~ ~ , ~ i ...... ' ...... :....... : : ~ ......... ...... }...... i ...... :....... ::............. ~...... i ............ :....... i................... ~ : ~ i : ........ i ....i......i......~!!t!~i~J.~;ii;:~r162
'~ ~...... ~...... ~...... i...... ............. i " ' ~ i ~ ...... iTM " ~. , i
i~i~,~'~ ............ ~.......! " i ..........
~:.
~ ~:;;~.~:~ ~
!~~!ii!iiiiiiiii!~ l~iiii~ iiii~ii~i~i~i~ii ~ i i ~
~::r-~::~,
,~ ~ ~ ~
~..,,2"i!ii~iiii~,?~ii~',2i~,?~i~.,,,:-.--:-i~....::,-
..-.:..-....-~'~..i
~ ~:~ ~ ~!~ ~a:: ~.~:'.::z~ ::.;~ .~ ~ ; : ~ . . . . . ~!~::~ii~ii~ iiii~i!~i~i~i~?~:.~i~-~i~
~}i}}iiii!~i~i~i~~!~~!!~?}?i~i~}ii}i!i~}ii:,iilN~}'~i!ill ii!~-i}~'~ii :~}'~i}N}~:::
..... i...... - - ...... i .................. i ..... i .......... i ..... i .. : .... i ..... i ............ !.. ~ ....: .:...i .... i........... i.......i. : i ~ , ~ i ! ~ 2 ~ } : ~ ! ~ i ~ : , ~ i ~ : i ~ : : ~ ~ i i 2 ~ ~ ~ 2 i : i ~ ~ i ~ i ~ i i i
.
.
.
:-..-..! ..-...--..:.
i:~
.
:
:
i..... i.-..........}.i~.~:.i...: %-"i:-"i:~4::---
..... 2 . . . . ..... ~...... < ......... ~..2.. -: ...~ . ...i...... !.... : . . . .
': ................. :..... : : : ~ ........................ :,~i:.:~i'~-~a~:;~:::~"-:i
i ..... i , i ~ ! ~ ! i ~ i ! ~ i ~ ! i ' : . ~ i i ~ i ' ~ i ; ~ : ~ i ! ~ i ~ i ~ ~ : ~ i ~
!""i:""' i..., i.......~;2;.i.-....i-....!.. ......... r...... ;........ ~....-.! ...... !:~i:~!~i~:N~i:~~ ....... . . - . . i
;...2. .~,;~,~-......~...............~.~. ~...... ;. ....-.~.-.....f.......~:.:..: .::... ~;~.....:~.:....~..;...Z
..... i ~ " i
.~:.~i.~..`.~sNii.~.~ii~..`..i~!~iii~!iii1iii~i~i!!i~i~!!i~i.i!i~5!i~!~i~:.~
...:..:.:.;:... ";"-:i:-::-"-.i....:..!...... :.. ::::~ ...:..~...... ~.:....; ......~. 2 . : ;.::.. 2 . . . . :~............. :.._ .i... . ' . ' : i i - " -~
F i g u r e 4. Effects of proficiency in lexical representations. English semantic representations in SOM2 were based on the WCD network's learning of a limited amount of input (92,640 word tokens).
Finally, to simulate individual differences in the representation of the bilingual lexicon, we constructed a smaller-in-size but otherwise identical network as in previous simulations. In the simulations reported above, we used 2,500 units each (a 50x50 grid) in SOM1 and SOM2. In the new simulation, we used 1,600 units (a 40x40 grid). This reduction in map size is meant to simulate a bilingual processor
Connectionist Model of Bilingual Processing
77
that has limited resources (e.g., working memory capacity) at disposal. There is evidence that individual differences in language processing are often due to working memory capacity (Just & Carpenter, 1992). Our network that used a smaller map size might thus show disrupted representational structure or processing abnormalities. Analyses of this manipulation show that the network was able to display language separation, lexical categories, and priming and interference effects as in the bigger network. However, given the same amount of training (500 epochs), the network had difficulty in distinguishing between many words that are either phonologically or semantically similar. In other words, the same units on the SOM often represented two or more similar lexical items. Recognition rate dropped to 57% for SOM1 (as compared to 74% with the 50x50 map) and to 54% for SOM2 (as compared to 74% with the 50x50 maps). 5 As a result of the decrease in recognition fidelity, the network's associative links were much more diffuse than those in the larger networks. Thus, there were stronger within- and cross-language priming and interference effects as compared with the bigger network, some of which could not be explained easily. For example, activation of let on SOM1 caused bei (let), let, show, and, mouse, and red to become activated on SOM2; activation of go on SOM2 caused say, waa (say), waai (bad), hal (be), gong (say), hoei (go), and know to become activated on SOM1, and so on. These results indicate that the unusual patterns in bilingual processing may be due to diffuse bilingual representations, which might in turn be due to limited processing capacity. In summary, manipulations of size of the learning corpus and size of the network extend the capabilities of our model in accounting for a variety of bilingual phenomena.
General Discussion and Conclusions
In this chapter we provided an overview of connectionist bilingualism, and presented a sketch of SOMBIP, a self-organizing connectionist model of bilingual processing. A number of theoretical and architectural considerations of the model are discussed, and preliminary simulation results from the model are analyzed. Our SOMBIP model consists of two interconnected self-organizing neural networks, coupled with a recurrent neural network that computes lexical co-occurrence constraints. Simulations with our model indicate that (1) the model can account for distinct patterns of the bilingual lexicon without the use of language nodes or language tags, (2) it can develop meaningful lexical-semantic categories through self-organizing processes, (3) it can account for a variety of priming and interference effects based on associative pathways between phonology and semantics in the lexicon, and (4) it can explain lexical representation in bilinguals with different levels of proficiency and working memory capacity. These capabilities of our model are due to its design characteristics in that (a) it combines
78
Ping Li and Igor Farkas
localist and distributed properties of processing, (b) it combines representation (through self-organization) with learning (recurrent neural network WCD, Hebbian learning of associative pathways), and finally (c) it combines the lexicon and sentences in bilingual processing. Most previous connectionist bilingual models (e.g., BIA and BIMOLA) are stationary models. To fully capture the dynamics in bilingual language processing, a model needs to incorporate the above mentioned properties, especially the learning characteristics such as in SOM and WCD. Green (1998) proposed an inhibitory control (IC) model of bilingual lexical processing. Although the IC model is not a connectionist model--on the contrary, it has quite some symbolic AI flavor, for example, with goal-oriented decision boxes and control schemas--it attempts to integrate activation-based accounts (e.g., interactive activation mechanisms). According to Green, there are multiple levels of control, with each level associated with a specific schema, from high-level event scripts to low-level articulatory controls. The particular level at which the IC model operates is an intermediate level, the lemma level, whereby an inhibitory mechanism suppresses the activation of lemmas that are tagged as belonging to the language other than the intended one. Crucial to the functioning of this mechanism are the language tags, tags that are believed to be part of the conceptual system of the lexicon. Green likened the language tags to the language nodes in the B IA model, which function to reinforce lexical activation of the currently activated language, while simultaneously decreasing lexical activation in the other lexicon. Interestingly, in a brief note, Dijkstra and van Heuven (1998) mentioned that the B IA language nodes might be thought to represent abstractions of very different statistical regularities in the two languages due to lexical co-occurrence differences. This view differs from what Green assumed, but is completely consistent with our argument here (in SOMBIP, lexical co-occurrence statistics are explicitly incorporated through WCD). In a commentary on the IC model, Li (1998a) questioned the necessity of language tags. One piece of evidence against the notion of language tags in IC is that some recent studies show that there need not be a cost associated with language switching, especially in natural speech situations (Grosjean, 1988, 1997; Grosjean & Miller, 1994; Li, 1996a)--according to IC, language switching takes time, since the process of switching to another language involves the inhibition of previous language tags. In the IC model, the inhibition of a particular stimulus shuts down the activation of all other related stimuli in the same language from the top down. This assumption seems to contradict several activation-based accounts that the bilinguals' two languages may be always activated, though the strength of the activation differs in specific linguistic situations, depending on the frequency of the target words, the sentence context, the speaker's proficiency in the two languages, and the speech mode (Grosjean, 1988, 1997; Li, 1996a). Thus, Li (1998a) argued that it is unclear whether the language tags are ad hoc constructs or necessary components of bilingual processing, just as it is unclear whether language nodes are necessary. Li suggested that the seemingly separate representations of the two
Connectionist Model of Bilingual Processing
79
lexicons, and the related interlingual priming/interference effects, might arise as a result of lexical and grammatical learning in a self-organizing neural network, in which no distinct labels are given to items of the two or more languages. Specifically, words from both languages may exist in the same topological map, but over time the network can develop localized patterns of activity in learning the mappings between phonology, orthography, morphology, and semantics. These localized patterns of activity may then correspond to the learner's internalized, distinct representations of the two lexicons. In this study, we have implemented precisely such a model, the SOMBIP, which nicely accounts for the independent (yet interactive) representations of the bilingual lexicon. A supra-lexical level of language nodes or language tags is absent in SOMBIP, but the effects of the language nodes or tags are captured in the system. In response to Li's commentary, Green (1998) agreed that the bilingual lexicosemantic system is composed of self-organized networks, as a presumption of the subset hypothesis (Paradis, 1989) adopted by the IC model. But he argued that language tags are necessary and that even in self-organizing networks such as the one used by Miikkulainen (1997) there is still room for language tags. Interestingly, Miikkulainen (1993) did mention a multiple SOM network that functions exactly as Green would argue: each language has its own separate lexicon, stored on different maps. A high-level control network would determine which language modules are active, switching between conversations in two languages. Such an approach was intended to model, according to Miikkulainen, distinct language behaviors in the representation of the bilingual lexicon in the brain. But Miikkulainen also speculated on a different (in our view a more plausible) model, one that is very similar in spirit to SOMBIP. This model would represent all word forms from the multilingual lexicon in a single lexical map, and all semantic representations in a single semantic map. These two maps could be associated with each other via the associative links. Cross-language priming and interference ("slips between languages" in his terms) could be modeled by noisy semantic mapping (like the diffuse map in our limited-resource network) or by associative propagation. In this version of the model, there would still be a high-level control network, but the task of the control mechanism merely modulates language output such that the same sentence parser could work on different language materials. In SOMBIP, however, this control network was unnecessary as the WCD network works equally efficiently in both languages. The ability of our network to show distinct patterns of the bilingual lexicon without distinct mechanisms (e.g., separate language tags or nodes) is yet another case of the classical connectionist argument: modular linguistic behaviors that are different or dissociated need not arise from'dual-mechanisms or dual-routes in the representational or processing system. They can naturally emerge from learning the input characteristics within a single system, a system that can be operationalized in terms of statistical or connectionist principles. This argument was advanced by the original PDP studies of the English past-tense acquisition (Rumelhart &
80
Ping Li and Igor Farkas
McClelland, 1986), and has since been supported by research in many domains of language studies, including language acquisition, normal and impaired word reading, and language organization in the brain (e.g., Elman, Bates, Johnson, Karmiloff-Smith, Parisi, & Plunkett, 1996; Plaut, McClelland, Seidenberg, & Patterson, 1996; Miikkulainen, 1997; Small, Hart, Nguyen, & Gordon, 1995). As a final important note, our model departs from the tradition that has been adopted by many bilingual researchers (see Kroll & de Groot, 1997, for discussion) in that it does not assume there is an integrated level of meaning-concept or meaning and concept (the "conceptual level"). The mixture of meaning with concept has created much confusion, in both first language and second language acquisition research (see Li & Shirai, 2000). This confusion is further complicated by the fact that many psycholinguists adopt a language-specific lexical-semantic level, the lemma level (Levelt, 1989). To avoid these confusions, thus, in our model there is no language-independent "conceptual" representation. Our semantic representation, or more accurately, lexical representation that contains both syntactic and semantic information, comes out the WCD network's analyses of global lexical cooccurrence information. This level of representation matches more closely with what linguists call "meaning", rather than "denotata" (Lyons, 1977)--entities of concepts that many bilingual researchers use in their picture naming tasks. To conclude, the SOMBIP model outlined in this chapter serves as a new model of bilingual processing and provides a new perspective on connectionist bilingualism. It has the potential of explaining a wide variety of empirical and theoretical issues in bilingual research. So far, we have tested the SOMBIP model only on spoken (transcribed) bilingual Chinese-English data. How the model generalizes to other bilingual situations should be further tested. Chinese differs significantly from English and other Indo-European languages in phonological, syntactic, and lexical structures and these differences are bound to affect language processing (Li, 1996b, 1998b). When one examines bilingual pairs of languages that are more similar (e.g., Dutch and German), it is likely that the model will show more integrated phonological and semantic representations of the bilingual lexicon. Moreover, if the mixed bilingual data contain more intra-sentential code-switches (rather than inter-sentential as in our data), the network that receives such data as input might also produce more integrated semantic representations. In addition, the current SOMBIP design could easily be extended to simulate orthographic-semantic relationships, and be evaluated against empirical findings and results from the B IA model. Other interesting extensions of the model would include simulating impaired semantic, phonological, or orthographic structures by adding noise to the SOM maps and by damaging the associative pathways that connect word form and word meaning. Such modeling efforts could shed light on bilingual representation and processing in the brain (Fabbro, 1999).
Connectionist Model of Bilingual Processing
81
Author Notes
This research was supported by a grant from the National Science Foundation (#BCS-9975249) to the first author. We would like to thank Fran~;ois Grosjean and Nicolas L6wy for comments and suggestions on an earlier draft of the paper, and Kathy Reid and Clint Schlenker for their editorial assistance and data preparation. The simulations were run on a Pentium PC with Linux OS. We used the DISLEX code (Miikkulainen, 1999) to run the SOMI and SOM2 and their associative links, and the WCD code was written by the second author.
Notes
1 The current scheme represents a scaled-down version of the full tri-syllabic template representation of English (see Li & MacWhinney, 2001, for the full version). As such, it is limited in certain respects; for example, it is not able to represent consonant clusters as in string/strlN/, which would require additional Cslots (e.g., CCCVVC for one syllable). 2 It is impossible to show all the details of the original figure here because each of the 2-D maps contains 400 words spread on a 50x50 grid of 2,500 units. 3 The organizations of the phonological representations are mainly based on phonemic and template similarities, and are relatively less interesting for our purpose here. 4 The distinction between state and activity verbs has attracted significant debates concerning the nature and origin of this distinction in language acquisition. Some researchers propose that this distinction is innately given (Bickerton, 1981, 1984). We have argued that this distinction can arise from children's statistical learning of the semantic properties of verbs, in the same way as it would occur distinctly in our connectionist networks (Li & Shirai, 2000). 5 The recognition rate is calculated as the rate of correctly mapped words in the SOM over the total number of words in the lexicon. Whether a word is mapped correctly is determined by a function that calculates the nearest distance between the input and-the codevectors in the map. Note that the recognition rate of 74% is not too impressive at first glance. This is because the network learns a many-tomany mapping function for the noun and verb translation equivalents. If the network learns a one-to-one mapping, the recognition rate would go up to 93% and 97% for SOM1 and SOM2, respectively.
References
Altarriba, J., Kroll, J. F., Sholl, A., & Rayner, K. (1996). The influence of lexical and conceptual constraints on reading mixed-language sentences: Evidence from eye fixations and naming times. Memory & Cognition, 24, 477-492.
82
Ping Li and Igor Farkas
Bickerton, D. (1981). Roots of language. Ann Arbor, MI: Karoma. Bickerton, D. (1984). The language bioprogram hypothesis. Behavioral and Brain Sciences, 7, 173-188. Bowerman, M. (1978). The acquisition of word meaning: An investigation into some current conflicts. In N. Waterson & C. E. Snow (Eds.), Development of communication (pp. 263-287). New York: John Wiley and Sons. Bowerman, M. (1982). Reorganizational processes in lexical and syntactic development. In E. Wanner & L. R. Gleitman (Eds.), Language acquisition: The state of the art (pp. 319-346). Cambridge, UK: Cambridge University Press. Burgess, C., & Lund, K. (1997). Modelling parsing constraints with highdimensional context space. Language and Cognitive Processes, 12, 1-34. Burgess, C., & Lund, K. (1999). The dynamics of meaning in memory. In E. Dietrich & A. Markman (Eds.), Cognitive dynamics: Conceptual and representational change in humans and machines (pp. 17-56). Mahwah, NJ: Erlbaum. Dijkstra, T., & van Heuven, W. (1998). The BIA model and bilingual word recognition. In J. Grainger & A. M. Jacobs (Eds.), Localist connectionist approaches to human cognition (pp. 189-225). Mahwah, NJ: Erlbaum. Elman, J. (1990). Finding structure in time. Cognitive Science, 14, 179-211. Elman, J. L., Bates, E. A., Johnson, M. H., Karmiloff-Smith, A., Parisi, D., & PlunkeR, K. (1996). Rethinking Innateness: A connectionist perspective on development. Cambridge, MA: MIT Press. Fabbro, F. (1999). The neurolinguistics of bilingualism: An introduction. East Sussex, UK: Psychology Press. Farkas, I., & Li, P. (2001). Modeling the development of lexicon with a growing self-organ.izing map. Manuscript submitted for publication. Farkas, I., & Li, P. (in press). A self-organizing neural network model of the acquisition of word meaning. In E. Altmann, A. Cleeremans, C. Schunn, & W. Gray (Eds.), Proceedings of the 4th International Conference on Cognitive Modeling. French, R. M. (1998). A simple recurrent network model of bilingual memory. In M. A. Gemsbacher & S. J. Derry (Eds.), Proceedings of the 20th Annual Conference of the Cognitive Science Society (pp. 368-373). Mahwah, NJ: Erlbaum. Grainger, J. (1993). Visual word recognition in bilinguals. In R. Schreuder & B. Weltens (Eds.), The bilingual lexicon (pp. 11-26). Amsterdam/Philadelphia: John Benjamins. Green, C. D. (1998). Mental control of the bilingual lexico-semantic system. Bilingualism, I, 67-82. de Groot, A. M. B., & Nas, G. L. J. (1991). Lexical representation of cognates and noncognates in compound bilinguals. Journal of Memory and Language, 30, 90-123. Grosjean, F. (1988). Exploring the recognition of guest words in bilingual speech. Language and Cognitive Processes, 3, 233-274.
Connectionist Model of Bilingual Processing
83
Grosjean, F. (1997). Processing mixed languages" Issues, findings and models. In A. M. B. de Groot & J. F. Kroll (Eds.), Tutorials in bilingualism: Psycholinguistic perspectives (pp. 225-254). Mahwah, NJ: Erlbaum. Grosjean, F., & Miller, J. (1994). Going in and out of languages: An example of bilingual flexibility. Psychological Science, 5, 201-206. Hebb, D. (1949). The organization of behavior: A neuropsychological theory. New York: John Wiley. Heredia, R. R. (1997). Bilingual memory and hierarchical models: A case for language dominance. Current Directions in Psychological Science, 6, 34-39. Heredia, R. R., & Altarriba, J. (2001). Bilingual language mixing: Why do bilinguals code-switch? Current Directions in Psychological Science, 10, 164-168. Hinton, G. E., & Sejnowski, T. (1999). Unsupervised learning: Foundations of neural computation. Cambridge, MA: MIT Press. Hong Kong Linguistic Society. (1997). Yueyu pinyin zibiao (Cantonese romanization scheme). Hong Kong Linguistic Society. Just, M. A., & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in working memory. Psychological Review, 99, 122-149. Kirsner, K., Smith, M. C., Lockhart, R. S., King, M-L., & Jain, M. (1984). The bilingual lexicon: Language-specific effects in an integrated network. Journal of Verbal Learning and Verbal Behavior, 23, 519-539. Kohonen, T. (1982). Self-organized formation of topologically correct feature maps. Biological Cybernetics, 43, 59-69. Kohonen, T. (1989). Self-organization and associative memory (3rd ed.). Berlin: Springer-Verlag. Kohonen, T. (1995). Self-organizing maps. Heidelberg: Springer. Kroll, J. F., & de Groot, A. M. B. (1997). Lexical and conceptual memory in the bilingual: Mapping from meaning in two languages. In A. M. B. de Groot & J. F. Kroll (Eds.), Tutorials in bilingualism: Psycholinguistic perspectives (pp. 169-200). Mahwah, NJ: Erlbaum. Lachter, J., & Bever, T. (1988). The relation between linguistic structure and associative theories of language learning: A constructive critique of some connectionist learning models. Cognition, 28, 195-247. Ladefoged, P. (1982). A course in phonetics (2nd ed). San Diego, CA: Harcourt Brace. Levelt, W. (1989). Speaking: From intention to articulation. Cambridge, MA: MIT Press. Ldwy, N. (2001). Computational psycholinguistics and spoken word recognition in the bilingual and the monolingual. Manuscript in preparation, University of Neuch~.tei, Switzerland. Ldwy, N., & Grosjean, F. (2001). The computerized version of BIMOLA: A bilingual model of lexical access. Manuscript in preparation, University of Neuch~tel, Switzerland.
84
Ping Li and Igor Farkas
Li, P. (1996a). Spoken word recognition of code-switched words by ChineseEnglish bilinguals. Journal of Memory and Language, 35, 757-774. Li, P. (1996b). The temporal structure of spoken sentence comprehension in Chinese. Perception and Psychophysics, 58, 571-586. Li, P. (1998a). Mental control, language tags, and language nodes in bilingual lexical processing. Bilingualism, 1, 92-93. Li, P. (1998b). Crosslinguistic variation and sentence processing: The case of Chinese. In D. Hillert (Ed.), Sentence processing: A crosslinguistic perspective (pp. 33-51). San Diego, CA: Academic Press. Li, P. (1999). Generalization, representation, and recovery in a self-organizing feature-map model of language acquisition. In M. Hahn & S. C. Stoness (Eds.), Proceedings of the 21st Annual Conference of the Cognitive Science Society (pp. 308-313). Mahwah, NJ: Erlbaum. Li, P. (2000). The acquisition of lexical and grammatical aspect in a selforganizing feature-map model. In L. Gleitman & A. Joshi (Eds.), Proceedings of the 22nd Annual Conference of the Cognitive Science Society (pp. 304-309). Mahwah, NJ: Erlbaum. Li, P. (in press). Language acquisition in a self-organizing neural network. In P. Quinlan (Ed.), Connectionism and developmental theory. Philadelphia/Briton: Psychology Press. Li, P., & MacWhinney, B. (2001). PatPho.'A phonological pattern generator for neural networks. Manuscript submitted for publication. Li, P., & Shirai, Y. (2000). The acquisition of lexical and grammatical aspect. Berlin/New York: Mouton de Gruyter. Lyons, J. (1977). Semantics (Vol. 2). Cambridge, UK: Cambridge University Press. Macnamara, J. (1967). The bilingual's linguistic performance: A psychological overview. Journal of Social Issues, 23, 59-77. MacWhinney, B., & Leinbach, J. (1991). Implementations are not conceptualiza-tions: Revising the verb learning model. Cognition, 40, 121-157. MacWhinney, B. (2000a). Lexicalist connectionism. In P. Broeder & J. M. Murre (Eds.), Models of language acquisition: Inductive and deductive approaches (pp. 9-32). Oxford, UK: Oxford University Press. MacWhinney, B. (2000b). The CHILDES project: Tools for analyzing talk (3rd ed). Mahwah, NJ: Erlbaum. McClelland, J. L., & Elman, J. (1986). Interactive processes in speech perception: The TRACE model. In J. L. McClelland, D. E. Rumelhart, & the PDP research group (Eds.), Parallel distributed processing: Explorations in the microstructure of cognition (Vol. 2, pp. 58-121). Cambridge, MA: MIT Press. McClelland, J. L., & Kawamoto, A. H. (1986). Mechanisms of sentence processing: Assigning roles to constituents. In J. L. McClelland, D. E. Rumelhart, & the PDP research group (Eds.), Parallel distributed processing: Explorations in the microstructure of cognition (Vol. 2, pp. 272-325). Cambridge, MA: MIT Press.
Connectionist Model of Bilingual Processing
85
McClelland, J. L., & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception. Part 1: An account of basic findings. Psychological Review, 88, 375-405. Miikkulainen, R. (1993). Subsymbolic natural language processing: An integrated model of scripts, lexicon, and memory. Cambridge, MA: MIT Press. Miikkulainen, R. (1997). Dyslexic and category-specific aphasic impairments in a self-organizing feature map model of the lexicon. Brain and Language, 59, 334366. Miikkulainen, R. (1999). The DISLEX simulator (new version) [on-line]. Available at ftp://ftp.cs.utexas.edu/pub/neural-nets/software/dislex.2.1.tar.Z. Obler, L., & Albert, M. (1978). A monitor system for bilingual language processing. In M. Paradis (Ed.), Aspects of bilingualism (pp. 105-113). Columbia, SC: Hornbeam Press. Paradis, M. (1989). Bilingual and polyglot aphasia. In F. Boiler & J. Grafman (Eds.), Handbook ofneuropsychology (Vol. 2, pp. 117-140). Amsterdam: Elsevier. Plaut, D. C., McClelland, J. L., Seidenberg, M. S., & Patterson, K. (1996). Understanding normal and impaired word reading: Computational principles in quasi-regular domains. Psychological Review, 103, 56-115. Rumelhart, D., & McClelland, J. (1986). On learning the past tenses of English verbs. In J. McClelland, D. Rumelhart, & the PDP research group (Eds.), Parallel distributed processing." Explorations in the microstructure of cognition (Vol. 2, pp. 216-271). Cambridge, MA: MIT Press. Small, S. L., Hart, J., Nguyen, T., & Gordon, B. (1995). Distributed representations of semantic knowledge in the brain. Brain, 118, 441-453. Smith, M. C. (1997). How do bilinguals access lexical information?. In A. M. B. de Gro0t & J. F. Kroll (Eds.), Tutorials in bilingualism: Psycholinguistic perspectives (pp. 145-169). Mahwah, NJ: Erlbaum. Spitzer, M. (1999). The mind within the net. Cambridge, MA: MIT Press. van Heuven, W. J. B. (2000). Visual word recognition in monolingual and bilingual readers: Experiments and computational modeling. Unpublished doctoral dissertation, NICI, University of Nijmegen, Netherlands. Yip, V., & Matthews, S. (2000). Syntactic transfer in a Cantonese-English bilingual child. Bilingualism, 3, 193-208.
This Page Intentionally Left Blank
Part III: Memory Representation in Sentence Processing
This Page Intentionally Left Blank
Bilingual Sentence Processing- R.R. Herediaand J. Altarriba(Editors) 9 2002 Elsevier Science B.V. All rights reserved.
Cross-Language Facilitation, Repetition Blindness, and the Relation Between Language and Memory: Replications of Aitarriba and Soltano (1996) and Support for a New Theory Don MacKay University of California, Los Angeles Lori E. James University of Colorado Lise Abrams University of Florida, Gainesviile
Abstract
Using improved materials, procedures, analyses, and experimental design, the present study replicated and extended several interesting effects reported in Altarriba and Soltano (1996). In two experiments, 48 highly proficient Spanish-English bilinguals read and immediately recalled rapidly presented (RSVP) sentences and three-word lists resembling those of Altarriba and Soltano: each stimulus contained a target and pretarget word that were related in three ways: complete identity (e.g. ants-ants), semantic identity across languages (i.e., cross-language translation equivalents such as hormigas-ants), and unrelated within or across languages (e.g., tie-ants). Results indicated that with complete identity, repetition blindness occurred in both lists and sentences (i.e., participants always recalled targets better when unrepeated than when repeated within the same language). In contrast, with semantic identity across languages, repeated concepts were recalled better than unrepeated concepts in lists, but this cross-language facilitation effect only occurred for ambiguous words with many translation equivalents: Neither RB nor facilitation occurred for unambiguous words with a single translation equivalent. These results support a new theory of repetition blindness and carry implications for theories of the relation between language and memory for lists versus sentences. MacKay and Miller (1994) presented mixed-language sentences via Rapid Serial Visual Presentation (RSVP) to Spanish-English bilinguals for immediate recall and found that when target words (T) were repeated in the same language within a sentence such as 1 (below; translation:."They saw horses, but horses were forbidden to enter"), T recall decreased compared to unrepeated Ts as in 2,
90
Don G. MacKay, Lori E. James, and Lise Abrams
replicating an effect known as repetition blindness (RB). They also demonstrated reduced recall when the T was a translation equivalent of the prior "pretarget" word (PT) as in 3 (T and PT underlined), rather than unrelated as in 4, an effect they called "semantic blindness" (SB). 1. They saw caballos, but caballos were prohibito to enter. (RB repeated T condition) 2. They saw ovejas, but caballos were prohibito to enter. (RB unrepeated T condition) 3. They saw horses, but caballos were prohibito to enter. (SB repeated T condition) 4. They saw sheep, but caballos were prohibito to enter. (SB unrepeated T condition) However, Altarriba and Soltano (1996) noted that some of the mixed-language sentences in MacKay and Miller (1994) were ungrammatical, and using grammatical sentences and other seemingly minor procedural changes in their Experiment 1b, they found RB for within-language repetition (as in 1 vs. 2), but did not find SB for conceptual or cross-language repetition (as in 3 vs. 4). MacKay, Abrams, Pedroza, and Miller (1996) next conducted detailed analyses of the procedures, analyses, and experimental design in Altarriba and Soltano, and noted seven factors that may have worked to diminish SB in their results, such as long lags between PTs and Ts in the sentences, relatively slow presentation rates, use of a dependent measure that failed to exclude trials where the PT was undetected, and use of split-language sentences with a single, predictable language switch in each sentence. In an attempt to resolve the conflicting results of these earlier SB studies, the present study tested whether SB would occur using the procedures, analyses, and experimental design of MacKay and Miller but with fully grammatical sentences as in Altarriba and Soltano. A second goal of the present study was to replicate an interesting difference that Altarriba and Soltano (1996) reported for lists vs. sentences: With h translation equivalent (manejar) as PT, a T such as drive became easier to recall in lists but not in split-language sentences such as "Mike leamed to drive and empes6 a manejar al trabajo" (translation: "Mike learned to drive and began to drive to work"). To illustrate the list procedure in Altarriba and Soltano (Experiment 2), each contained three words as in 5-8 (PT and T underlined), preceded and followed by masking stimuli (symbol strings) that participants (proficient Spanish-English bilinguals) were instructed to ignore. The dependent variable was recall of the T (e.g., drive in 5-8) as a function of whether its PT was identical (as in 5), semantically identical (i.e., a translation equivalent, as in 7), or unrepeated (as in 6 and 8). The interesting finding was that when drive had a translation equivalent (manejar) as PT, participants exhibited a statistically reliable facilitation effect (relative to the unrepeated condition illustrated in 8), but exhibited the usual inhibitory effect (RB) with an identical PT (as in 5).
Repetition Blindness in Bilinguals 5. 6. 7. 8.
91
drive-(intervening word)-drive (RB repeated T condition) steer-(intervening word)-drive (RB unrepeated T condition) manejar-(intervening word)-drive (SB repeated T condition) guiar-(intervening word)-drive (SB unrepeated T condition)
The present study tested an account of this list-sentence difference developed in detail in MacKay et al. (1996). This account involves three postulates for which there is already considerable support: (a) that the mechanisms for storing and retrieving verbal materials in lists are inseparable from mechanisms that have evolved for producing, comprehending, and representing sentences (see e.g., MacKay & Abrams, 1996); (b) that proficient bilinguals represent the meaning of translation equivalents via a single lexical node (see e.g., MacKay, 1982); and (c) that RB and SB occur when the same node must be activated and then reactivated in quick succession under RSVP (see e.g., Abrams, Dyer, & MacKay, 1996). Given these assumptions, the cross-language equivalents had different effects for sentences vs. lists in Altarriba and Soltano (1996) because of a fact noted in MacKay and Bowman (1969)" that a word such as thick has many possible meanings and translation equivalents when presented in isolation, but only a single meaning and translation equivalent (grueso) within a sentence such as, "Mike likes comer queso grueso encima de thick bread." For example, taken in isolation, thick has at least four distinct meanings corresponding to the Spanish translation equivalents grueso, espeso, denso, and bruto, and under the theory, a single lexical node represents each of these meanings for proficient Spanish-English bilinguals. These four shared nodes are illustrated in Figure 1 (in rectangles), together with the phonological representations for thick, grueso, espeso, denso, and bruto (in dashed rectangles). As shown, the phonological representation for thick connects bottomup with all four lexical nodes, which in turn connect top-down with phonological representations corresponding to grueso, espeso, denso, and bruto. This means that the occurrence of thick as a pretarget in lists such as thick horses thick, and thick caballos grueso would strongly prime the lexical nodes for all four meanings of thick, even though only one becomes activated and perceived, say the one that represents the meaning "not watery" (espeso). However, the phonological representations for grueso, espeso, denso, and bruto will all be primed or readied for activation (see Figure 1), so that when grueso occurs in this list, the phonological and lexical nodes for grueso, being already primed, are easier than normal to activate, giving rise to the facilitation effect observed in Altarriba and Soltano. Experiment 1 directly tested an implication of this account of cross-language facilitation within lists. The implication is that only lexically ambiguous PTs with many Spanish-English equivalents should facilitate recall of cross-language Ts in lists (e.g., 7). The facilitation effect that Altarriba and Soltano observed should not be obtained for PTs with only a single meaning and only a single Spanish-English equivalent in split-language lists.
92
Don G. MacKay, Lori E. James, and Lise Abrams
Lexical Concept Nodes (shared in Spanish and English)
//
//
/
Q I|
ss S
ss| s S
J
s ~
oS J
~s~ ~
ss
s I
I
..~'
s
J
s
S
I s SS
s I fs
stupid
dense
not watery
thick
~.1" | I |
1! |. . . . . .
! ! ! ! ! ! ! ! ! I
"1
,THICK
I
English' Phonological Form
,, I ! !
,, !
Spanish Phonological Forms
(5 ,, I
L
GRUESO
i, "= I
A
I I
ESPESO
~' I I
' I '
DENSO
='
I _,
,'
BRUTO
I
I
,
I a
Figure 1. Connections in NST between the phonological form for thick, a sample of
its (shared Spanish-English) lexical conceptual nodes, and phonological forms for corresponding Spanish words (grueso, ~ , denso, and bruto). Initial presentation of thick primes the unshaded nodes connected with broken lines (but does not activate them; see MacKay, 1990, for the priming-activation distinction), and activates the nodes connected with solid lines, resulting in self-inhibition (solid shading).
Experiment 1" Effects of Within- vs. Between-Language Repetition in Lists Experiment 1 assessed the degree of RB and SB (if any) in RSVP lists using improved procedures noted in MacKay et al. (1996), but was otherwise similar to Altarriba and Soltano (1996, Experiment 2) except that half of the PTs were ambiguous and allowed many translation equivalents and half were unambiguous and allowed only one translation equivalent. We expected no facilitation effect for
Repetition Blindness in Bilinguals
93
cross-language Ts with unambiguous PTs but we expected to replicate the facilitation effect of Altarriba and Soltano for Ts with ambiguous PTs (e.g., 7).
Participants Participants were 48 college students (28 females, 20 males, mean age = 19.7, mean years of education = 14.1) who participated for course credit or payment. All were Spanish-English bilinguals who had approximately equal proficiency in Spanish and English as assessed via a background questionnaire and conversations in English and Spanish with a bilingual experimenter. The participants provided Likert-scale ratings in the background questionnaire that suggested only slightly more daily use of English than Spanish: On a 5-point scale (1 = mostly English, 3 = equal use of Spanish and English, and 5 = mostly Spanish), their mean rating was 2.44. Answers to a question on how often they switch languages mid-sentence in everyday conversation also suggested flexible proficiency in both languages: 77% of the participants indicated that they often or sometimes switch languages, while only 21% indicated that they rarely or never do so. Three other questions also suggested extensive proficiency in both languages. For one question, 35% of the participants reported speaking Spanish more often in everyday life, 25% reported speaking English more often, and 17% claimed to speak Spanish and English equally often. For another question, 85% of the participants indicated Spanish as their native language with English acquired during early childhood (mean age - 5 years old), and 15% indicated English as their native language. Finally, mean ratings on five-point scales assessing relative competence in English vs. Spanish (where 1 = English much better, 2 = English somewhat better, 3 = Equal, 4 = Spanish somewhat better, and 5 = Spanish much better) were 1.91 for knowledge of grammar, 2.00 for vocabulary knowledge, 2.22 for spoken fluency, and 2.62 for spoken comprehension, indicating somewhat greater competence for grammar, vocabulary, and spoken fluency in English but almost equal competence for spoken comprehension in English and Spanish.
Materials Materials were 40 filler and 72 experimental lists. Experimental lists are illustrated in Appendix A and contained three words in the order PT, intervening word, and T. Ts and PTs were either identical within a language (e.g. ants-ants: the within-language repeated T condition), semantically identical across languages (i.e., non-cognate translation equivalents such as hormigas-ants" the cross-language repeated T condition), or unrelated within or between languages (e.g., tie-ants: the unrepeated T condition). As in these examples, Ts were identical across these three basic conditions but counterbalanced across subjects. For unrepeated conditions, Ts and PTs were equal in mean number of letters, and across all conditions, Ts and PTs were either English or Spanish translation equivalents selected from DuboisCharlier, Pritchard, Senerth, and Sola (1987) by a proficient bilingual (native Spanish-speaker). Language of PT and T was crossed to yield two within-language
94
Don G. MacKay, Lori E. James, and Lise Abrams
combinations (English PT-English T and Spanish PT-Spanish T) and two betweenlanguage combinations (English PT-Spanish T and Spanish PT-English T). The eight conditions that resulted were counterbalanced across participants so that every PT-T pair appeared in every condition. For between-language trials (English PTSpanish T and Spanish PT-English T), the word intervening between PT and T was English for half the lists and Spanish for the remaining half. For unrepeated conditions, PT and T were always semantically and phonologically unrelated, and all three words were the same language in within-language lists. Finally, we manipulated ambiguity of the English words, yielding 16 conditions overall. Ambiguous words had four or more meanings listed in Webster's College Dictionary (1995), and three or more Spanish translation-equivalents in DuboisCharlier et al. (1987). Unambiguous words had one distinct meaning in Webster's and one primary Spanish translation-equivalent in Dubois-Charlier et al. (see Appendix A). Because word forms increase in frequency with ambiguity (i.e., how many meanings dictionaries list for the word form; Zipf, 1949), word frequency was higher for ambiguous than unambiguous words in the present materials (M = 177 vs. 59 per million in Francis & Ku(,era, 1982, SD = 116). Filler lists served to make word repetition, language, and number of words per list less predictable: Fillers were either two or four words long and contained words that were unrepeated and unrelated (semantically and phonologically), and were either in English (N = 10), Spanish (N = 10), or an unpredictable mix of English and Spanish (N = 20). Procedure
Participants completed an informed consent form, a demographic questionnaire, a digit span test, and the language background questionnaire. Following general instructions regarding RSVP procedures and mixed-language sentences, participants received 8 practice trials with instructions to read and immediately recall each word in its presented position in the list. They were cautioned not to translate but to recall each word in the language presented. Responses were scored on-line by the experimenter and verified later from a tape recording of the session. Each trial began with a 2000 millisecond (ms) warning signal: "Get ready for the next list." Next came seven 90 ms frames: a row of percent signs, a row of asterisks, the PT, the intervening word, the T, a row of dollar signs, and a row of percent signs. A row consisted of five symbols of a given type. A string of question marks signaled the end of the list and called for recall, remaining on the screen until the participant pressed any key to trigger the next trial.
Results
As instructed, participants almost never recalled a list item in the wrong language, and overall, 64% of words were correctly recalled (SD = 13%), precluding floor and ceiling effects. Two lists were discarded due to stimulus flaws
Repetition Blindness in Bilinguals
95
discovered post hoc (one list contained homophones, and one contained a nonword), giving 70 experimental trials per subject in our analyses. Trials were scored for correct inclusion of T in recall under three scoring criteria (lenient, strict, and Altarriba & Soltano, described shortly), and Figures 2a and 2b show mean correct T recall by condition using lenient scoring criteria. Lenient scoring first removed trials containing incorrect or missing PTs and then required correct recall of both PT and T in the correct order, but otherwise ignored recall of the intervening word and did not penalize intrusions. For example, PT and T in the list "sky (PT) red dog (T)" were scored as correct given recall of"sky dog," "sky raw dog," "sky girl red dog," or "boy sky red dog." If PT and T were both recalled but in the wrong order relative to the intervening word, only the word in correct list position was scored as correct, for example, dog (T) in "raw sky dog" and sky (PT) in "sky dog red" and neither PT nor T was scored as correct in "dog red sky," or "red dog sky." A 2 (repetition: repeated, unrepeated) x 2 (ambiguity: ambiguous, unambiguous) x 2 (combination: within-language, between-language) x 2 (target language: English, Spanish) repeated measures ANOVA yielded no main effect of ambiguity, but a main effect of repetition, F(1,42) = 73.47, MSe = 0.09, p < .001, with better recall of unrepeated than repeated Ts, a main effect of combination, F(1,42) = 56.49, MSe = 0.13, p < .001, with better recall of between- than within-language Ts, and a main effect of target language, F(1,42) = 22.11, MSe = 0.06, p < .001, with better recall of English than Spanish Ts. These effects were moderated by an interaction of repetition and combination, F(1,42)= 142.25, MSe = 0.08, p < .001, with RB for within-language repetition, t(47) = -13.40, p < .001, but facilitation for between-language repetition, t(47) = 3.34, p < .01, and a repetition by combination by ambiguity interaction, F(1,42) = 4.34, MSe = 0.07, p <. 05. For ambiguous targets, there was an interaction of repetition and combination, F(1,47) = 110.46, MSe = 0.04, p < .001, due to occurrence of RB for withinlanguage repetition, t(47)=-12.11, p < .001, and facilitation for between-language repetition, t(47) = 3.83, p < .001. For unambiguous targets, there was a repetition by combination interaction, F(1,47) = 64.13, MSe = 0.04, p < .001, due to RB for within-language repetition, t(47) = -11.54, p < .001, but no effect for betweenlanguage repetition, t(47) = 0.43, p =.67. Within-language RB was equivalent for ambiguous and unambiguous words (47% vs. 45%; see Figure 2a), while betweenlanguage facilitation was 14% for ambiguous words versus only 2% for unambiguous words (see Figure 2b). Strict scoring of correct T recall required that PT, the intervening word, and T all be recalled in the correct order. Strict scoring yielded essentially the same pattern of results as lenient scoring. For example, a 2 (repetition: repeated, unrepeated) x 2 (ambiguity: ambiguous, unambiguous) x 2 (combination: within-language, between-language) x 2 (target language: English, Spanish) repeated measures ANOVA yielded no main effect of ambiguity, but a main effect of repetition, F(1,29) = 44.85, MSe = 0.11, p < .001, with better recall of unrepeated than
96
Don G. MacKay, Lori E. James, and Lise Abrams
repeated Ts, a main effect of combination, F(1,29) = 18.91, MSe = 0.16, p < .001, with better recall of between- than within-language Ts, and a main effect of target language, F(1,29) = 10.05, MSe = 0.06, p < .01, with better recall of English than Spanish Ts. However, these effects were moderated by an interaction of repetition and combination, F(1,29) = 61.47, MSe = 0.11, p < .001, with RB for withinlanguage repetition but facilitation for between-language repetition.
90% 80%
~o%
59%
n~ 60%
57% !"II Repeated Targets
50%
~6 L-I,.,.
I I Unrepeated Targets
40% 30%
~ 20% 10% 0% Ambiguous
Unambiguous
Figure 2a. Mean correct recall by condition (under lenient scoring) for within-
language lists in Experiment 1.
90% 80% 70% 60%
61%
57% ==o,,_
40% z,,,,. ,.,__
t~
Repeated Targets Unrepeated Targets
g 50% 30%
~ 20% lO% 0% Ambiguous
Unambiguous
Figure 2b. Mean correct recall by condition (under lenient scoring) for between-
language lists in Experiment 1.
Repetition Blindness in Bilinguals
97
Altarriba and Soltano scoring analyzed T recall without first removing trials
containing incorrect PTs following Altarriba and Soltano (1996). This analysis yieldedthe same pattem of results as lenient scoring. Subsidiary Results Using lenient scoring to test for possible effects of Spanish vs. English Ts, recall of cross-language Ts was comparable for English and Spanish Ts overall (54% vs. 52%), as well as for unrepeated Ts in English vs. Spanish (51% vs. 47%) and repeated Ts in English vs. Spanish (58% vs. 57%), yielding a comparable facilitation effect for cross-language lists with English vs. Spanish targets (7% vs. 10%), t(47) = -0.52, p > .60. However, for within-language Ts, the level of target recall was much higher for English than Spanish Ts (44% vs. 26%), whereas recall was roughly proportional for unrepeated Ts in English vs. Spanish (70% vs. 42%) and for repeated Ts in English vs. Spanish (18% vs. 10%), yielding a larger RB effect for English than Spanish lists (52% vs. 32%), t(47) = 4.76, p < .001. We currently lack a plausible explanation for this difference. We also tested for possible effects of word frequency using lenient scoring, categorizing our English Ts as low frequency (LF) vs. high frequency (HF) via median split at 40.5 per million in Francis and Ku~era (1982; LF = 40 or fewer per million, HF = over 40 per million). A 2 (repetition: repeated, unrepeated) x 2 (combination: within-language, between-language) x 2 (frequency: LF, HF) repeated measures ANOVA on trials with English T yielded main effects of repetition, F(1,42) = 44.56, MSe = 0.10, p < .001, of combination, F(1,42) = 12.94, MSe = 0.12, p <. 01, and repetition by combination interactions for both HF targets, F(1,46) = 79.09, MSe = 0.07, p < .001, and LF targets, F(1,43) = 56.92, MSe = 0.05, p < .001, but no main effect of frequency and no interactions involving frequency at p < .05, despite the fact that frequency covaried with ambiguity.
Discussion Using procedures advocated in MacKay et al. (1996), present results replicated the basic findings of Altarriba and Soltano (1996, Experiment 2): the usual inhibitory effect (RB) for within-language repetition, but a reliable facilitation effect for between-language repetition. This between-language facilitation effect poses problems for token individuation theory in both its original and newer forms. The original token individuation theory (e.g., Kanwisher, 1987; Kanwisher & Potter, 1990) viewed token individuation to be purely visual or restricted to the processing of visual forms rather than orthographic, phonological, or semantic concepts, and cannot explain why translation equivalents become especially individuated, whereas exact repetitions fail to become individuated. Newer token individuation theories (e.g., Bavelier, 1994) likewise cannot explain the semantic facilitation effect because they predict that orthographic, phonologic, or lexical repetition will lead to inhibition, but not facilitation under RB conditions.
98
Don G. MacKay, Lori E. James, and Lise Abrams
However, the interactions of repetition and combination with ambiguity in Experiment 1 refined the Altarriba and Soltano (1996) cross-language facilitation effect: Between-language facilitation occurred for ambiguous targets with many translation equivalents, but not for unambiguous targets with only a single translation equivalent. These interactions were predicted and support the new theory of RB outlined in MacKay et al. (1996). The connections and processes for explaining cross-language facilitation in lists in this theory has already been discussed in the introduction, and the same connections and processes explain RB in rapidly presented word lists such as thick shift thick: Initial presentation of thick in such lists primes and activates the phonological nodes for thick, which strongly primes the lexical nodes for all four meanings of thick, and activates the most primed one, say the one corresponding to the meaning grueso in Figure 1. As a result, phonological nodes for thick and the lexical node for the grueso meaning may be undergoing the self-inhibitory process that follows activation (represented via solid circles in Figure 1) at the time when the third word in the list appears. Since the third word in this example is thick, RB is a likely outcome because the self-inhibited phonological nodes for thick will be harder to activate than when the initial word is phonologically, orthographically, and semantically unrelated to thick.
Experiment 2: Effects of Within- versus Between-Language Repetition in Sentences
Experiment 2 had the same participants as Experiment 1 and always immediately followed Experiment 1, but the stimuli were grammatical mixedlanguage sentences. We expected to replicate the reliable SB effect observed in MacKay and Miller (1994). Materials
Materials were 24 filler sentences and 24 mixed-language experimental sentences (shown in Appendix B). Experimental sentences were 9.4 words long on average (range 6-13 words) with 1.7 words on average intervening between PT and T. PTs and Ts were similar to those in Experiment 1 except that ambiguity was not manipulated, PTs and Ts never occupied first or last position in a sentence, English PT-Spanish T was the only between-language combination, and with the T omitted, 88% of the sentences became ungrammatical. Language switched within the sentences 1.95 times on average (SD = 0.37; range 1-3), and at unpredictable points across sentences. To further reduce predictability of word repetition and the occurrence and locus of language switches, filler sentences contained no repeated words and many were entirely in English (N = 7) or entirely in Spanish ( N - 7).
Repetition Blindness in Bilinguals
99
Procedure
Experimental and scoring procedures resembled Experiment 1 with four exceptions: there were 12 experimental conditions (because there was no Spanish PT-English T condition); presentation rate was 70 ms/word for half the sentences and 90 ms/word for the other half, with rates for particular sentences counterbalanced across participants; there were only 6 practice trials; and "Get ready for the next sentence" was the prompt that immediately preceded each sentence.
Results
Analyses discarded two experimental sentences due to flawed computer assignment to conditions (yielding 22 experimental trials per subject), and ignored the rare instances where participants recalled words in the wrong language (mean probability per participant was .019 per word). Overall, 51% of the words in a sentence were correctly recalled (SD = 12%), precluding floor and ceiling effects. Lenient scoring required correct recall of PT and T in the correct order, but otherwise ignored recall or misrecall of any other word in the sentences. To illustrate, for the sentence "I live at one(PT) three six(T) Elm Street," PT and T were scored as correct given recall as "I have one and six elms," or "I live at one four three six South Elm Street." Figure 3 shows mean correct T recall by condition (collapsed across rate) under lenient scoring. A 2 (repetition: repeated, unrepeated) by 2 (combination: within, between) by 2 (rate: 70 ms, 90 ms) repeated measures ANOVA revealed a main effect of repetition, F(1,25) = 7.94, MSe = 0.16, p < .01, with better recall of unrepeated than repeated Ts, and a main effect of combination, F(1,25) = 38.48, MSe = 0.19, p < .001, with better recall of between- than withinlanguage Ts. These effects were moderated by a marginally significant interaction of repetition and combination, F(1,25) - 3.71, MSe = 0.12, p < .07. However, rate neither yielded a main effect nor interacted with any other variable, which allowed us to collapse across rate, thereby reducing empty cells and increasing our power to detect differences. The resulting 2 x 2 ANOVA revealed the same main effects as the prior ANOVA, together with a highly reliable interaction of repetition and combination, F(1,45) = 15.43, MSe - 0.05, p < .001, such that within-language repetition yielded RB, t(46)--5.60, p < .001, whereas between-language repetition yielded neither SB nor facilitation, t(45) = 0.11, p>.91. For within-language repetition, the level of recall was higher for English than Spanish Ts (24% vs. 18%), but recall was roughly proportional for unrepeated Ts in English vs. Spanish (36% vs. 28%) and for repeated Ts in English vs. Spanish (11% vs. 8%), yielding a slightly larger RB effect for English than Spanish sentences (25% vs. 20%). Strict scoring counted the PT as unrecalled (thus removing the trial from analysis) if one but not both T and PT were recalled and it was unclear which within the context of adjacent words. Strict scoring therefore biased against finding significant
100
Don G. MacKay, Lori E. James, and Lise Abrams
effects, but this more conservative procedure yielded the same pattern of results as lenient scoring except that the main effect of repetition was non-significant when collapsed across rate in a 2 (repetition: repeated, unrepeated) x 2 (combination: within-language, between-language) ANOVA, F(1,39) = 0.83, MSe = 0.12, p > .37. However, recall was again better for between- than within-language Ts, with a marginal repetition by combination interaction, F(1,39) = 3.21, MSe = 0.07, p = .08, due to occurrence of reliable RB for within-language repetition, t(39)=-2.11, p < .05, but neither SB nor facilitation for between-language repetition, t(45) = 0.11, p > .91.
90% 80% =
8
70% 55% 54%
~. 60% a~ 50% t~
~- 40%
I"11Repeated Targets
38%
I I Unrepeated Targets
~ 9 30%
i,,.. O
r
20% 10% 0% Within-Language
Between-Language
Figure 3. Mean correct recall by condition (under lenient scoring) for sentences in
Experiment 2.
Discussion
Like Altarriba and Soltano (1996), we observed reliable RB for within-language repetition but neither facilitation nor interference (SB) for cross-language repetition. Contrary to expectation, Experiment 2 failed to replicate the reliable SB effect observed in MacKay and Miller (1994). The present results therefore comport with the suggestion of Altarriba and Soltano that inclusion of ungrammatical sentences in MacKay and Miller may have increased the degree of SB. However, post hoc analyses of data for mixed-language sentences in Experiment 2 suggested a possible role for two additional factors: congruence of the language switches with the syntax of the sentences, and differences in the semantic scope of the pretarget and target words.
Repetition Blindness in Bilinguals
I Ol
We defined the congruence factor as follows: A mixed-language sentence was phrase-congruent if words in multi-word phrases or syntactic constituents in the sentence were all in the same language, as in [We] [added nuts] [a las nueces][in the salad], where the square brackets enclose coherent syntactic constituents. By contrast, a mixed-language sentence was phrase-incongruent if one or more multiword phrases or syntactic constituents contained words in different languages. For example, [I want to open la puerta] [pero la puerta] [est~i locked] is a phraseincongruent sentence since the phrase [est~i locked] contains words from different languages. For phrase-congruent sentences in Experiment 2 (N = 11), mean correct recall was 59% for unrepeated targets vs. 66% for repeated targets, a 7% difference indicating semantic facilitation. However, for phrase-incongruent sentences in Experiment 2 (N = 11), mean correct recall was 59% for unrepeated targets vs. 46% for repeated targets, a 13% difference suggesting SB. However, this 13% SB effect was not statistically reliable, and further research is needed to test whether SB occurs in mixed-language sentences if and only if language switches are incongruent with phrase boundaries. Turning to the semantic scope factor, the meaning or semantic scope of pretarget and target words in mixed-language sentences in Experiment 2 was either the same ( N - 16) (e.g., two and dos in "Yo vivo en two ocho dos Main Street") or different (N = 6). For example, drinks and bebidas in "We asked for drinks aunque las bebidas eran expensive" differ in scope since las bebidas can refer to the particular drinks ordered--a more restricted semantic scope than drinks in general. For same-scope sentences, mean correct recall was 45% for unrepeated targets vs. 68% for repeated targets, a 23% difference indicating semantic facilitation. However, for different-scope sentences, mean correct recall was 63% for unrepeated targets vs. 53% for repeated targets, a 10% difference suggesting SB. Although not statistically reliable, this 10% difference suggests that the semantic scope of pretarget and target words should be systematically examined in future studies of SB.
General Discussion
Having replicated the basic Altarriba and Soltano (1996) results and having noted how results in Experiment 1 support a new theory of RB and challenge current theories, we now address the relevance of our results to theories of the relation between language and memory. On the surface, differences between recall of lists versus sentences in Experiments 1 vs. 2 seem ready-made for the currently popular multi-store approach to relations between language and memory. Under this approach, list-sentence differences reflect a built-in dichotomy between systems for language versus memory. For example, in the multi-store theory of Gathercole and Baddeley (1993, pp. 8-32), the phonological loop is a memory subsystem that is separate and distinct from the system for processing words in sentences (the central executive) and specializes in processing and storing word lists for short time-periods in phonological form (see e.g., Shiffrin & Nosofsky,
102
Don G. MacKay, Lori E. James, and Lise Abrams
1994, and Zhang & Simon, 1985, for other multi-store theories that have postulated phonological, articulatory, or acoustic representations for short-term memory but not semantic representations). However, contrary to multi-store theories, effects of ambiguity in word lists (Experiment 1) and cross-language facilitation (Experiment 2) indicate that very short-term memory includes semantic representations in addition to phonological, articulatory, or acoustic representations. Present results therefore add to the problems that multi-store theories currently face. One general class of problems concerns cases where sentence variables influence list processing in ways that would not be expected if fundamentally distinct and separate memory systems process sentences versus lists (for reviews, see Caplan & Waters, 1990; and Saffran, 1990). For example, consider another syntactic/semantic factor that improves immediate recall within rapidly presented lists. MacKay and Abrams (1994) compared immediate recall of identical words in two types of RSVP lists: One type contained familiar two-word phrases located at unpredictable positions in the lists (e.g., 9, below); The other type contained many of the same words in identical positions, but no phrases (e.g., 10, below). The results showed that the identical words were better recalled as parts of phrases than as unrelated words. For example, night was better recalled as part of the phrase night gown in 9 than as an unrelated word in 10, but the unrelated word mind was recalled equally poorly in both lists. Because phrases are fundamentally syntactic/semantic entities, these findings indicate that syntactic/semantic factors influence short term memory within rapidly presented lists. The problem for multi-store theories is to explain the role of such syntactic/semantic factors in a supposedly separate store that has traditionally been viewed as purely phonological in nature. 9. phrase good faith mind night gown film (phrases underlined) 10. phrase people faith mind night hose film (unrelated word list) Another general class of problems that multi-store theories currently face is that some variables have parallel effects in immediate recall of both sentences and lists. The present study illustrates this problem for exact repetition within a language: If a distinct and separate memory system processes and stores lists, multi-store theories must explain why exact repetition causes RB in both lists and sentences (see Miller & MacKay, 1996, for other examples within this general class of multi-store problems). By contrast, parallel effects across sentences and lists are unproblematic within distributed memory theories of the sort illustrated in the introduction and elsewhere (e.g., MacKay, 1987; MacKay & Burke, 1990; MacKay, Miller, & Schuster, 1994). Under distributed memory theories, there are no distinct and separate memory stores for lists vs. sentences, and short term memory is not an isolable system that is separate from cognition in general, as in the multi-store approach. Instead, immediate memory represents "an umbrella term for a heterogeneous array (of) capacities for temporary storage.., distributed over diverse cognitive subsystems" (Monsell, 1984, p. 328). That is, a single set of memory capacities is involved in acquiring, comprehending, and producing words, whether
Repetition Blindness in Bilinguals
103
in lists or in sentences, and these memory capacities are related to the formation and strengthening of connections between nodes distributed throughout the cognitive system (see e.g., MacKay, 1990). Of course, further development of distributed memory theories will be required to specify in detail the exact structure and processing characteristics of connections for representing lists versus sentences in general. However, the effects of ambiguity in Experiment 1 suggest that words in rapidly presented lists and sentences share at least one of the same basic processing characteristics (semantics), consistent with distributed memory theories. We conclude with an important question for future research on SB in sentences and a methodological paradigm for addressing that question. The question is whether reliable SB occurs when language switches are incongruent with the syntax of mixed-language sentences and the methodological paradigm is a modified RSVP procedure developed by Abrams et al. (1996) to examine how phrase-congruence affects RB in English sentences. Using this procedure, each RSVP screen contains several words, and in the phrase-congruent condition, these words represent single complete phrases or syntactic constituents, as in [They wanted][to play sports] [but sports] [were not allowed] (where the square brackets represent individual screens). In the phrase-incongruent condition, each screen contains parts of several phrases, as in [They wanted to][play sports but]_[ sports were not] [allowed] . With number of screens and mean presentation time per word held constant in the phrase-congruent and phrase-incongruent conditions, Abrams et al. found exaggerated or greaterthan-normal RB with phrase-incongruent screens and no RB with phrase-congruent screens. Phrase-incongruent screens in the Abrams et al. paradigm therefore provide a means of magnifying interference effects due to repetition and may help resolve the issue of whether SB in sentences constitutes a reliable phenomenon or represents an artifactual and irreplicable result: If proficient bilinguals do not exhibit reliable SB in recalling mixed-language sentences presented in phraseincongruent RSVP screens using the Abrams et al. technique, this would suggest that SB does not occur in grammatical sentences. However, if reliable SB does occur in the phrase-incongruent condition, this would support the conclusion that like RB, SB is a real phenomenon that occurs under a restricted but theoretically interesting range of conditions (Abrams et al. discuss the significance of phraseincongruent effects for theories of RB in sentences, and Miller & MacKay, 1996, and MacKay & Miller, 1996, demonstrate and discuss the significance of analogous phrase-incongruent effects for theories of repetition deafness in sentences).
References
Abrams, L., Dyer, J. R., & MacKay, D. G. (1996). Repetition blindness interacts with syntactic grouping in rapidly presented sentences. Psychological Science, 7, 100-104. Altarriba, J., & Soltano, E. G. (1996). Repetition blindness and bilingual memory: Token individuation for translation equivalents. Memory & Cognition, 24, 700-711.
104
Don G. MacKay, Lori E. James, and Lise Abrams
Bavelier, D. (1994). Repetition blindness between visually different items: The case of pictures and words. Cognition, 51, 199-236. Caplan, D., & Waters, G. S. (1990). Short-term memory and language comprehension: A critical review of the psychological literature. In G. Vallar & T. Shallice (Eds.), Neuropsychological impairments of short-term memory (pp. 337389). Cambridge: Cambridge University Press. Dubois-Charlier, F., Pritchard, D. R., Senerth, D. T., & Sola, J. M. (1987). The American heritage larousse Spanish dictionary: English~Spanish (lst ed.). Boston, MA" Houghton Mifflin. Francis, W. N., & Ku~era, H. (1982). Frequency analysis of English usage: Lexicon and grammar. Boston: Houghton Mifflin. Gathercole, S. E., & Baddeley, A. D. (1993). Working memory and language. Hillsdale, NJ: Lawrence Erlbaum Ass. Ltd. Kanwisher, N. G. (1987). Repetition blindness: Type recognition without token individuation. Cognition, 2 7, 117-143. Kanwisher, N. G., & Potter, M. (1990). Repetition blindness: Levels of processing. Journal of Experimental Psychology: Human Perception and Performance, 16, 30-47. MacKay, D. G. (1982). The problems of flexibility, fluency, and speed-accuracy trade-off in skilled behavior. Psychological Review, 89, 483-506. MacKay, D. G. (1987). The organization of perception and action: A theory for language and other cognitive skills. New York: Springer-Verlag. MacKay, D. G. (1990). Perception, action, and awareness: A three body problem. In W. Prinz & O. Neumann (Eds.), Relationships between perception and action (pp. 269-303). Berlin: Springer-Verlag. MacKay, D. G., & Abrams, L. (1994, November). Chunking and repetition deficits challenge capacity theories and the single attachment principle. Paper delivered to the 35 th Annual Meeting of the Psychonomic Society, St. Louis. MacKay, D. G., & Abrams, L. (1996). Language, memory and aging: Distributed deficits and the structure of new-versus-old connections. In J. E. Birren & K. W. Schaie (Eds.), Handbook of the psychology of aging (4 th edition, pp. 251265). San Diego: Academic Press. MacKay, D. G., Abrams, L., Pedroza, M. J., & Miller, M. D. (1996). Crosslanguage facilitation, semantic blindness, and the relation between language and memory: A reply to Altarriba and Soltano. Memory & Cognition, 24, 712-718. MacKay, D. G., & Bowman, R. W. (1969). On producing the meaning in sentences. American Journal of Psychology, 82, 23-39. MacKay, D. G., & Burke, D. M. (1990). Cognition and aging: A theory of new learning and the use of old connections. In T. Hess (Ed.), Aging and cognition: Knowledge organization and utilization (pp. 213-263). Amsterdam: North Holland. MacKay, D. G., & Miller, M. (1994). Semantic blindness: Rapidly presented repeated concepts are difficult to encode and recall. Psychological Science, 5, 5255. MacKay, D. G., & Miller, M. (1996). Can cognitive aging contribute to fundamental psychological theory? Repetition deafness as a test case. Aging,
Repetition Blindness in Bilinguals
105
Neuropsychology, and Cognition, 3, 1-18. MacKay, D. G., Miller, M., & Schuster, S. P. (1994). Repetition blindness and aging: Evidence for a binding deficit involving a single, theoretically-specified connection. Psychology and Aging, 9, 251-258. Miller, M. D., & MacKay, D. G. (1996). Relations between language and memory: The case of repetition deafness. Psychological Science, 7, 347-351. Monsell, S. (1984). Components of working memory underlying verbal skills: A "distributed capacities" view. In H. Bouma & D. G. Bouhuis (Eds.), Attention and performance X." Control of language processes (pp. 327-350). Hove, UK: Erlbaum. Saffran, E. M. (1990). Short-term memory impairment and language processing. In A. Caramazza (Ed.), Cognitive neuropsychology and neurolinguistics." Advances in models of cognitive function and impairment (pp. 137-168). Hillsdale, N.J.: Erlbaum. Shiffrin, R. M., & Nosofsky, R. M. (1994). Seven plus or minus two: A commentary on capacity limitations. Psychological Review, 101,357-361. Webster's college dictionary (1995). New York, NY: Random House. Zhang, G., & Simon, H. A. (1985). STM capacity for Chinese words and idioms: Chunking and the acoustical loop hypotheses. Memory & Cognition, 13, 193-201. Zipf, G. K. (1949). Human behavior and the principle of least effort. Cambridge, MA: Addison-Welsey.
Appendix A English (and Spanish Translation Equivalents) in the Three-word Lists in Experiment 1 (see text for explanation).
Unambiguous Targets (and Repeated Pretargets) ants (hormigas) aphid (pulg6n) arrow (flecha) bee (abeja) blond (rubio) Broom (escoba) deaf (sordo) elm (olmo) eye (ojo) fist (pufio) flea (pulga) happy (alegre) hill (colina)
Unrelated Pretargets
Intervening Words
easy (fb.cil) bride (novia) brain (cerebro) beg (pedir) level (nivel) towel (toalla) belt (cinto) widow (viuda) law (ley) hair (pelo) soap (jab6n) music (m6sica) hand (mano)
tie (corbata) debt (deuda) tomato (tomate) son (hijo) narrow (angosto) father (padre) panther (pantera) blue (azul) bed (cama) harp (arpa) brother (hermano) mouth (boca) bull (toro)
106
Don G. MacKay, Lori E. James, and Lise Abrams
lake (lago) lettuce (lechuga) lung (pulm6n) mouse (rat6n) myth (mito) nanny (nifiera) oar (remo) onion (cebolla) quarrel (discusi6n) rain (lluvia) red (rojo) rice (arroz) shark (tibur6n) shrimp (camar6n) sky (cielo) spinach (espinaca) log (leno) thief (ladr6n) thigh (muslo) thinker (pensador) woman (mujer) wood (madera) yolk (yema)
Ambiguous Targets
(and Repeated Pretargets) bank (orilla) box (caja) brush (cepillo) clear (claro) couple (pareja) dark (oscuro) dirt (tierra) dock (muelle) face (cara) fast (r~ipido) fire (fuego) fly (mosca) free (libre) full (lleno) girl (chica)
rock (piedra) chapter (capitulo) crib (cuna) magic (magia) nape (nuca) snack (bocado) wig (peluca) tooth (diente) accused (acusado) wall (pared) lid (tapa) moist (hfmedo) realm (reino) pillow (almohada) art (arte) lobster (langosta) cup (taza) dwarf (enano) wind (viento) bottles (botellas) spark (chispa) song (canci6n) wine (vino)
delete (tachar) school (escuela) fork (tenedor) steak (bistec) film (tela) ticket (boleto) plane (avi6n) silver (plata) average (promedio) arm (brazo) cold (frio) north (norte) stick (palo) basement (sotano) life (vida) water (agua) bird (pb.jaro) finger (dedo) cake (pastel) neighbor (vecino) ear (oreja) great (gran) half (mitad)
Unrelated Pretargets
Intervening Words
wait (espera) day (dia) stain (mancha) hawk (halc6n) window (ventana) best (mejor) cord (cord6n) gray (gris) boat (barco) home (hogar) sott (suave) ball (pelota) focus (enfocado) baby (beb6) bomb (bomba)
youth (joven) cow (vaca) shade (sombra) green (verde) office (oficina) mail (correo) short (corto) nap (siesta) clock (reloj) white (blanco) bad (malo) tomb (tumba) turtle (tortuga) sauce (salsa) fall (otofio)
Repetition Blindness in Bilinguals groom (novio) hammer (martillo) high (alto) joke (chiste) key (llave) leaf (hoja) leak (gotera) nail (clavo) party (fiesta) picture (pintura) pink (rosa) play (obra) record (apuntar) shell (concha) sign (gesto) store (tienda) tape (cinta) thick (grueso) thing (objeto)
moose (alce) mirror (espejo) train (tren) harm (dafio) dye (tinte) fake (falso) name (nombre) clue (pista) bell (campana) contest (concurso) wolf (lobo) town (pueblo) socket (enchufe) death (muerte) luck (suerte) dream (suefio) lime (lima) shift (cambio) leash (correa)
107 fresh (fresco) sugar (az6car) soul (alma) beer (cerveza) dog (perro) honey (miel) farm (granja) game (juego) soup (caldo) tired (cansado) sour (agrio) age (edad) pasture (pastura) milk (leche) chicken (polio) cucumber (pepino) pen (pluma) neck (cuello) meat (came)
Appendix B Experimental sentences for the English-English, Spanish-Spanish, and English-Spanish conditions of Experiment 2. Targets and pretargets are underlined and unrelated pretargets appear in parentheses. (See text for explanation.) 1. Cuando Joe screams (fights) he screams like un loco When Jos6 grita (pelea) 61 grita como un maniac When Joe screams (fights) 61 grita como un maniac 2. Cuando we sell grapes (juice) the grapes are en temporada When we vendemos uvas (jugo) las uvas est~in in season When we sell grapes (juice) las uvas est~n in season 3. Tiraron this dough (bread) because the dough did not se levant6 They tiraron este masa (pan) porque la masa no se rise They threw out this dough (bread) porque la masa no se rise 4. Se fue de the church (chapel) when the church was pintada He left la iglesia (capilla) cuando la iglesia fue painted He left the church (chapel) cuando la iglesia fue painted 5. Nosostros asked for drinks (liquor) although drinks were cara We pedimos bebidas (lic6r) aunque las bebidas eran expensive We asked for drinks (liquor) aunque las bebidas eran expensive
108
Don G. MacKay, Lori E. James, and Lise Abrams 6. Queria wear purple (that skirt) because purole matched with sus zapatos She wanted to vestirse de morado (falda) porque el morado coordinaba con her shoes She wanted to wear purole (that skirt) porque el morado coordinada con her shoes 7. Quiero abrir the door (house) but the door is cerrada con llave I want to open la ouerta (casa) pero la puerta est~i locked I want to open the door (house) pero la puerta estfi locked 8. Those teachers (students) will be teachers of history el pr6ximo afio Esos maestros (alumnos) serb.n maestros de historia next year Those teachers (students.) ser~in maestros de historia next year 9. Her nephew (cousin) and my nephew look like gemelos Su sobrino (primo) y mi sobrino parecen ser twins Her nephew (cousin) y mi sobrino parecen ser twins 10. Ese hombre washes the wheels (cars) when the wheels are sucias That man lava las llantas (los carros) cuando las llantas estfin dirty That man washes the wheels (cars) cuando las llantas estfin dirty 11. The letters (oaoers) were letters from la segunda guerra mundial Las cartas (Los oaoeles) eran cartas de la second world war The letters (oaoers) eran cartas de la second world war 12. El cazador at__.ee(caught) fish and at___gecarne tambidn The hunter comi6 (pesc6) pescado y comi6 meat as well The hunter at__.ee(caught) pescado y comi6 meat as well 13. Alguien encontr6 the book (pencil) with the book que habifis perdido Someone found el libro (l~oiz) con el libro you had lost Someone found the book (pencil) con el libro you had lost 14. El gato duerme on the chair (table) by the chair en la cocina The cat sleeps en la silla (mesa)por la silla in the kitchen The cat sleeps on the chair (table) por la silla in the kitchen 15. La gente mira the bears (cubs) if the bears estb.n afuera del hfibitat People watch los osos (crios) si los osos are outside the habitat People watch the bears (cubs.) si los ososare outside the habitat 16. Me gusta your rabbit (hamster) but my rabbit es m~is lindo I like su conejo. (h~imster) pero mi conejo is prettier I like your rabbit (hamster) pero mi conejo is prettier 17. Ellos vieron the su_...nnas (.moon when) the su___n._nsali6 They saw el so_.._ll(la luna.) cuando el so__._!lrose They saw the sun (moo.n) cuando el so_.__!lrose
Repetition Blindness in Bilinguals 18. Tuve frogs (ducks) when frogs eran popluares I owned ranas (patos) cuando las ranas were popular I owned frogs (ducks) cuando las ranas were popular 19. Yo vivo en two (five) eight two de la Calle Mayor I live at do_.._fis(cinco) ocho do..__fisMain Street Yo vivo en two (five) ocho do___fisMain Street 20. Agregamos nuts (fruit) to the nuts en la ensalada We added nueces (fruta) alas nueces in the salad We added nuts (fruit) alas nueces in the salad 21. Necesitaba money (funds) but the money no era disponible He needed dinero (fondos) pero el dinero was not available He needed money (funds) pero el dinero was not available 22. Por favor pon 6ste dish (cup) and that dish al lado de la spoon Please put 6ste plato (vaso) y ese plato next to la cuchara Please put this dish (cup) y ese plato next to la cuchara
109
This Page Intentionally Left Blank
Bilingual Sentence Processing- R.R. Heredia and J. Altarriba (Editors) 9 2002 Elsevier Science B.V. All rights reserved.
The Use of Sentence Contexts in Reading, Memory and Semantic Disambiguation Jeanette Altarriba and Jennifer L. Gianico University at Albany, State University of New York
Abstract
Sentences provide the semantic constraints that are typically sufficient to guide the reader to an expected word or group of words. Studies conducted within our laboratories and those of other researchers will be reviewed that demonstrate how bilinguals use sentence contexts to predict specific lexical and conceptual features leading to a precise ending, how sentence contexts serve to facilitate the interpretation of code-switched information, and how those contexts facilitate in the disambiguation of cross-language homographs in processing. The contributions of bilingual research to understanding sentence processing in general for monolingual speakers is also discussed. The resulting bilingual findings help to inform both theory and research in the study of language and cognition.
Introduction
In sentence processing, there are various issues in the monolingual literature that have yet to be resolved. For example, do individuals process both lexical and conceptual information simultaneously and develop constraints as to upcoming words based on that information? How do individuals use previews of upcoming words to understand the current word being processed within a sentence? Finally, how do sentence contexts operate in the processing of lexically ambiguous words? Monolingual paradigms are limited in numerous ways. For example, across languages one can capitalize on the fact that a pair of translation equivalents can have the same meaning, but completely different forms. Though the withinlanguage comparison might be synonyms, it has been argued that synonyms may not overlap on the basis of their semantic features as closely as translations, for concrete, common objects (see e.g., Altarriba & Soltano, 1996). Cognates across languages are also useful to examine as the situation arises where there is an identical or almost identical visual overlap in features and again a strong overlap in semantic features (see e.g., Altarriba, Kambe, Pollatsek, & Rayner, 2001). For questions regarding the level at which a specific effect occurs in lexical processing (e.g., orthographic, semantic, etc.) cross-language stimuli offer a useful means of controlling for various features while manipulating others. Thus, the use of bilingual
112
Jeanette Altarriba and Jennifer L. Gianico
approaches in this area of sentence comprehension has led to new information regarding the processing of words within sentential contexts. The aim of the present paper is to review the current and past data regarding the above questions with an eye towards demonstrating the role of bilingual paradigms, tasks, methods, and techniques that serve to uncover the answers to these issues. The three main areas to be considered here include the processing of mixedlanguage sentences and the use of lexical and conceptual information, the influence of previews of upcoming words on sentence processing, and the resolution of lexically ambiguous words within sentential contexts. A review of the major themes in the monolingual literature in each area will be presented in preparation for considering the bilingual contributions to each individual area in solving basic theoretical issues in sentence processing.
Lexical and Conceptual Constraints in Sentence Processing Various aspects of language are processed when one encounters a sentence context. For example, conceptual constraints are operating to assist in the access of an acceptable meaning. Lexical constraints operate in building expectations for specific words or syntactic representations. Phonological aspects of words are tapped, and pragmatics may play a role in determining the overall semantics of a sentence construction. Indeed, the various parameters that are encountered within a sentence context lead to the ease or difficulty of processing of specific words within that context as opposed to cases in which words appear in isolation (see e.g., McElree & Griffith, 1998; Simpson, Peterson, Casteel, & Burgess, 1989; Stanovich & West, 1979, 1981, 1983; West & Stanovich, 1982; Whitney, McKay, Kellas, & Emerson, 1985). For example, Traxler, Foss, Seely, Kaup, and Morris (2000) showed that target words in sentences benefited from word primes that appeared earlier in the sentence if those primes were identical in terms of their lexical and semantic representations (e.g., The lumberjack greeted the lumberjack early this morning.). Somewhat less facilitation was exhibited for targets preceded by related primes (e.g., "The minister greeted the pastor yesterday at the post office."). Finally, unrelated primes conferred no benefit on the processing of subsequent targets (e.g., "The young man greeted the lumberjack early in the morning."). As will be shown later in this section, the second instance above is not the strongest test possible of the relative contributions of conceptual and lexical constraints on language processing as the words "minister" and "pastor" also differ somewhat on both dimensions simultaneously. The focus of the current section centers on the following question: Do individuals process both lexical and conceptual information simultaneously and develop constraints as to upcoming words based on that information? This area of research interest has been investigated primarily within the monolingual domain (see e.g., Duffy, Henderson, & Morris, 1989; Fischler, 1985; Fischler & Bloom, 1980; Morris, 1994; Schwanenflugel & LaCount, 1988; Schwanenflugel & Shoben, 1985). Researchers have questioned whether or not an individual's expectancy is
Sentence Contexts in Reading
113
based on the lexical form or representation of an upcoming word, or on its meaning or conceptual representation. Take for example, the sentence that follows below: (1) The wedding cake had a bride a n d . An individual who is knowledgeable about the English language might successfully predict that the completion of this sentence frame is the word "groom". How is that implication drawn? Do lexical constraints determine language comprehension when reading sentences? Or, is meaning the critical determiner of the successful completion of the comprehension process during reading? In the following section, various theories and evidence will be reviewed in the monolingual domain that address these questions concerning sentence processing. Following that review, data will be discussed for mixed-language sentences that contain code-switched words in an attempt to clarify the role of the various levels of language processing in sentence comprehension. A bilingual paradigm will be introduced as a powerful and sensitive measure of general language processing--a measure that represents an improvement over existing monolingual approaches. Monolingual investigations. While there is general agreement that contextual information facilitates sentence reading and comprehension, the source of that facilitation is currently under debate. The issue is whether the benefit in processing is due to lexical or to semantic/conceptual variables or to an interaction of these levels of representation. Even though arguments exist noting the contribution of both variables in reading, more recent findings indicate that the facilitatory effect of context is due to an interaction of lexical retrieval processes and processes due to higher level influences. These influences can be semantic or conceptual, and have been referred to as "message-level" variables (Duffy et al., 1989; Morris, 1994; Schustack, Ehrlich, & Rayner, 1987). One type of task that has been used to investigate the processing of sentential contexts involves tracking the movements of the eyes while reading (see e.g., Duffy et al., 1989; Schmauder, Morris, & Poyner, 2000). Although it appears that when reading, we smoothly glide along a page of text, actually, the eye is making a series of jumps or saccades across a line. In fact, our eyes move to a new location about four times every second (Rayner & Pollatsek, 1989). The movements of eyes during reading are called saccadic eye movements. The eyes are actually jumping across a line of text rather than flowing smoothly over a text. A saccade or "jump" takes about 10 or 20 ms (milliseconds) and crosses 10 letters on average. The jumps also occur either forward or backward. Each time the eye lands at a new location, the eye is said to fixate at that location. These pauses between jumps are called fixations. The reader is said to be fixating on a particular character or space. These fixations last about 250 ms, but this changes according to the difficulty of the text being read and the reading ability of the participant. Fixation duration is one index of the difficulty of information processing during reading. Evidence has been provided for the position that message-level representations interact with lexical retrieval through the use of the eyetracking paradigm (see e.g., Duffy et al., 1989; Morris, 1994). In
114
Jeanette Altarriba and Jennifer L. Gianico
Morris' work, for example, lexical and message-level information were independently manipulated. In a sentence including the nouns gardener and barber and the verb trimmed, eye fixation times on the target word mustache were facilitated only if barber was the agent of trimmed. In the case in which gardener was the agent of trimmed, no facilitation in fixation time was reported, although the lexical content of the sentence was similar. These results suggest that context effects may be the result of an interaction between lexical access and the message-level representation within a sentence. Schwanenflugel and LaCount (1988) suggested that sentence constraint effects likely operate at a semantic or conceptual level. They examined the influence of both sentence constraint and semantic relatedness on the processing of upcoming words in a sentence. In their work, participants showed difficulty in making lexical decisions (i.e., word-nonword judgments) to a word that was semantically related to the expected word in a highly constrained context, as in the following example: "The landlord was faced with a strike by the residents." The expected word in this case was tenants. They proposed a feature restriction model to explain the influence of sentence constraint on word processing. Sentence constraint is said to determine the number of feature restrictions that individuals generate when reading text. Readers, therefore, generate fewer feature restrictions for low-constraint sentences than for high-constraint sentences. Thus, given the low-constraint sentence, "She went to the store to buy a new ", a reader might generate few feature restrictions like [something carried in a store], [an article that can be purchased], [an actual object]. Many different concepts can form successful completions of this sentence frame. However, given a high-constraint sentence such as, "She took the cake out of the warm ", a reader might generate many more features that would ultimately lead to one or few final words. In this case, those features might include [an apparatus for baking], [something larger than a cake], [an item that becomes heated], etc. An individual would likely show facilitation to a word such as "oven" if it appeared for naming or a lexical decision. Facilitation will be shown only for those final words whose semantic representations do not mismatch any of the feature restrictions generated from the sentence context. Thus, related words, whose semantic descriptions mismatch on several features, will not show facilitation from the sentence context. One problem with the above account is that it relies most heavily on the semantic or conceptual features that drive the comprehension process when processing sentences. It is difficult to conclude whether or not lexical features play a decisive role in generating final completions for sentence frames such as the ones investigated by Schwanenflugel and LaCount (1988). What is needed is a paradigm that allows for the manipulation of lexical constraints while holding semantics constant across possible completions, that is, strictly constant, rather than simply manipulating degrees of semantic relatedness. Could a bilingual approach afford the ability to investigate this question? The following section describes a new approach in this area of investigation that allows for the manipulation of the relevant variables in ways that a monolingual approach could not provide for.
Sentence Contexts in Reading
115
If lexical constraints determine language Bilingual investigations. comprehension when reading sentences, then the presentation of a translation equivalent should disrupt processing. However, if meaning is the critical determiner of the successful completion of the comprehension process during reading, then the lexical format or actual visual characteristics of the words would not be important, and a conceptual equivalent would be processed equally well. For the statement, "The wedding cake had a bride and ," is an individual expecting the exact word "groom", or will a conceptual equivalent such as "novio" (its Spanish translation) suffice? Further, how do bilinguals process mixed-language sentences? We know that individuals often code-switch or perform language mixing when speaking to individuals who are also fluent in the speaker's languages (Altarriba & Santiago-Rivera, 1994; Heredia & Altarriba, 2001; Pfaff, 1979; Yoon, 1992). Do patterns of eye movements during reading reveal that Spanish-English bilinguals have a facility for reading mixed-language sentence? If so, what might this reveal about how the languages are interconnected in bilingual memory? In a series of studies conducted by Altarriba, Kroll, Sholl, and Rayner (1996) Spanish-English bilinguals who were equally fluent readers of both languages silently read sentences such as (2) that were either heterogeneous or homogeneous in terms of language. Sentences of low constraint such as (3) were also included for which a variety of completions was acceptable. (2) The wedding cake had a bride and groom figurine on the top layer. (3) He wanted the shirt to look like the one the groom was wearing at the wedding. The language change in the heterogeneous cases involved the insertion of a translation equivalent for the target word highlighted above. Participants did not see any highlighted words in the actual experiment. When the sentences were presented, participants were asked to merely read them silently. Eye movements were monitored using an eyetracking system. The results demonstrated that when bilingual readers are reading in a particular language, and a sentence builds up a particular expectancy for a specific word, that word is expected in a language that matches the entire sentence text. The presentation of a translation equivalent caused the detained examination of that foreign word as evidenced by longer fixation times. When the sentence is not that highly constrained, a word in either language seems not to disrupt processing. In a second experiment, the very same sentence contexts were used and were presented one word at a time using RSVP (rapid serial visual presentation). The participants' task was to name the capitalized target word in each sentence. For example, in sentences (2) and (3) the target word "groom" was capitalized. The pattern of naming latencies was identical to that for fixations in the eye movement data. Thus, it appears that in terms of bilingual language processing and bilingual memory, the languages are processed separately when expectations are strongly built up in a single language. When the conceptual information that a text or sentence implies is not strongly constrained or strongly contextualized, either
116
Jeanette Altarriba and Jennifer L. Gianico
language can provide a meaningful conclusion in processing. It appears then that bilinguals can process one language separately from another and can build expectations based on processing in a single language mode, but they can also easily process information across languages in less constrained settings, for example in informal conversations. Several mechanisms may be responsible for the effects of sentence constraint on the recognition of upcoming words reported in the bilingual work above. To review, Schwanenflugel and colleagues (Schwanenflugel & LaCount, 1988; Schwanenflugel & Shoben, 1985) suggested that sentence constraint determines the number of semantic featural descriptions that are generated by participants as they read a sentence. When sentences are of low constraint, participants may generate fewer feature restrictions, thereby leading to facilitation for a wide variety of potential completions. For high constraint sentences, participants can generate a large number of features leading to the most expected completion of a sentence. Facilitation will be shown for words that match the expected feature set but that do not mismatch any of the features generated from the sentence. Schwanenflugel and LaCount found facilitation only for the expected words in a sentence context; semantically related words did not produce facilitation. The current bilingual work described here extends this model by showing that participants generate feature restrictions that are both semantic and lexical in nature. By using a mixed-language paradigm, we were able to simultaneously manipulate both conceptual and lexical-level features of target words. We found that word recognition was facilitated to the degree that both the expected semantic features restrictions and the lexicai feature restrictions were matched by the target word. If the semantic or conceptual restrictions were met, but a mismatch occurred on a lexical level, interference occurred for both first fixation durations and naming latencies. These results are compelling given that the findings were consistent across different tasks (i.e., eyetracking and naming in RSVP) and different participants. The work of Altarriba et al. (1996) demonstrates how bilingualism can be used as a tool for exploring sentence constraint effects during reading. Because past research suggests that fluent bilingual speakers have overlapping conceptual representations across languages, experiments may be designed to take advantage of the fact that a single concept can be accessed by two distinct lexical forms. Like research on lexical ambiguity (which will be reviewed later in this chapter), bilingual research permits the separation of lexical and conceptual contributions to reading and comprehension.
Preview Effects in Sentence Processing
As noted earlier within this chapter, reading can be assessed by investigating the pattern of fixations and saccades that result when individuals engage in basic reading tasks. Fixation duration is one index of the difficulty of information processing during reading. The fact that words are sometimes skipped suggests that
Sentence Contexts in Reading
I 17
words can often be identified and processed parafoveally. A question arising from this observation is, "how is visual information integrated across successive fixations during reading?" Researchers have focused on the nature of the information that is integrated and how that information is selected and processed from a parafoveal preview. In the section that follows, research within the monolingual domain on the factors that moderate facilitation from parafoveal previews will be reviewed. Following this discussion, a bilingual approach will be presented that indicates the usefulness of a cross-language paradigm in solving a most difficult and intriguing problem. Monolingual investigations. To investigate the nature of the parafoveal processing of words, several researchers have used a technique whereby a word or letter string initially appears in parafoveal vision, followed by the participant's eye movement to the stimulus (see e.g., Balota & Rayner, 1983; McClelland & O'Regan, 1981; Rayner, McConkie, & Ehrlich, 1978; Rayner, McConkie, & Zola, 1980). During the movement or saccade, the initially displayed stimulus is replaced by a word that the participant is asked to read. For example, theword presented first might be "chart"; when movement occurs the word is replaced by "chest". These studies involve eye contingent display changes and are referred to as boundary paradigms. An initial word is present and when the eyes cross a boundary, a word change occurs. Participants are then asked to read the word, typically naming the word aloud. The participant never actually reports having seen a preview word. He or she only reports the fixated word. In general, the studies using this paradigm compare response times to the target word as a function of various types of preview words. For previews of the type noted in the example, the benefit in reading time is on the order of 30 to 50 ms. This type of study is typically conducted with the use of an eyetracking system. A participant is seated facing a computer screen, and the participant's eye is aligned with the apparatus that is interfaced with a computer, such that a recording of the eye movements can be made in terms of coordinates referring to locations on the screen. Studies have also been performed with words presented in sentences. For example, one might view the following sentence:
(4)
The policeman shouted at the thieves in the bank.
Participants are shown one of three items in the periphery during the saccade: shouted, sboutef, cdovtef. Fixation time typically increases as a function of the dissimilarity between the preview and the fixated target word. Studies conducted by Rayner (1978) and Rayner et al. (1978) indicated that naming and recognition time were influenced by the visual similarity between the letter string appearing in the preview, and the target or word in the fovea. They provided evidence to suggest that visual information obtained from the parafoveal preview can influence the time required to name a word presented on the fovea. A typical grouping of stimuli that were used includes the following:
118
Jeanette Altarriba and Jennifer L. Gianico
Target:
chest
Preview Conditions: Word-identical Word- 1 change Word-2 changes Nonword-2 changes Nonword
chest chart chovt (similarity is preserved) chfbt (dissimilarity is greater) ejovf
An interesting point to note is that there 1was no difference between the condition in which words (chart) were initially presented and the condition in which nonwords (chovt) were initially presented as long as both were similar in visual characteristics. A hypothesis was formed conceming these data suggesting that some abstract code about the letters of the parafoveal word is stored and integrated with the information available in the fovea after the saccade. Most of the research has indicated that this abstract code carries primarily lexical features--features relating to the visual appearance of the words. Similar results have also been reported by Briihl and Inhoff (1995) and Rayner, Well, Pollatsek, and Bertera (1982). Parafoveal preview effects have also been found when participants are not asked to verbalize their response, but rather semantically categorize the word in a yes-no task. If case is changed for every other letter in a display, that is, if visual characteristics are changed, participants respond just as rapidly as if no case changes were present. They are somewhat slower in reading the alternating case condition but it is assumed that this is because it is more difficult to read in that format. Thus, a general explanation involving just visual features is not really viable, gi~/en that you get similar effects, with or without alternating case changes (Rayner et al., 1980). The earlier hypothesis can therefore be modified by suggesting that certain initial letters are identified from the initially displayed stimulus. When the target word is consistent with this letter information, attention can be devoted primarily to the middle and latter parts of the base word during the fixation of that word. This reduces processing requirements hence reducing naming times. Rayner et al. (1980) investigated this issue using the following sample stimuli: Target:
write
Preview Conditions: Word-identical Same first letter Same initial sound Semantic associate Same final letters
write walks rough print trite
Sentence Contexts in Reading
119
Words beginning with the same phoneme but a different letter did not result in faster naming for the base word, thus leaving out any explanations based on verbalization. Semantic relatedness between the two words produced no facilitation. Also, certain similarities in the letter patterns produced no facilitation; neither having the first letter in common nor having all but the first letter in common resulted in significant facilitation. This finding rules out some possible versions of a preliminary letter identification hypothesis. It appears that the letters identified for the preview and used at the time of fixation are those at the beginning of a word, but include more than just the first letter. No facilitation occurs when the first letter alone is in common for two words (e.g., ghost-grave). Likewise, no facilitation occurs when all but the first letter are in common (e.g., brave-grave). Substantial facilitation occurs when the first two letters are in common (e.g., green-grave) or when three letters are in common (e.g., graingrave). The earlier work, therefore, of Rayner and colleagues (e.g., Rayner, 1978; Rayner et al., 1978; Rayner et al., 1980) indicated that increasing the similarity of the parafoveal stimulus to the target item decreased the time taken to pronounce that target item. If the first two or three letters in the initial parafoveal string and the target were identical, facilitation occurred. Two subsequent articles seemed to question the generalizability of these conclusions. McClelland and O'Regan (1981) and Paap and Newsome (1981) suggested that because there was a relatively small set of target words in the earlier Rayner studies (in most cases 30), which were repeated throughout a given experiment, there may have been sufficient contextual constraint to allow participants to generate expectancies about potential parafoveal targets. They argued that these expectancies allowed the participants in the Rayner studies to use partial parafoveal information. An interactive model would predict interactive effects of context and parafoveal information, since two weak sources of information may have little effect by themselves but when combined may produce facilitation. Perhaps the parafoveal effects found by Rayner and his colleagues were due to the extra activation fed into the logogens representing the constrained target set. To investigate these possible explanations further, Balota and Rayner (1983) examined the conjoint effects of semantic contextual information and parafoveal visual information. Participants were presented with either a word (e.g., reptile) or a row of Xs in foveal vision along with a parafoveal nonword (e.g., snckks). Individuals were asked to name the parafoveal stimulus aloud. During the saccade made to that stimulus, the nonword was replaced by one of several items: Semantically related to the foveal item and visually related to the preview: snakes
Semantically unrelated to the foveal item and visually related to the preview: sneaks
Semantically related to the foveal item and visually unrelated to the preview: lizard
Semantically unrelated to the foveal item and visually unrelated to the preview: limits
120
Jeanette Altarriba and Jennifer L. Gianico
The results yielded reliable effects of parafoveal visual information when only a neutral row of Xs was presented as the constraining context. Thus, these results clearly conflict with the arguments that had been made by McClelland and O'Regan and Paap and Newsome that parafoveal visual information is useful primarily when meaningful contextual constraints are placed on that information. It is interesting to note that even words that show strong priming effects in semantic priming paradigms but are visually distinct (e.g., crypt-grave) do not show facilitation in the typically used paradigm described earlier. Rayner and Morris (1992) argued against any hypothesis concerning semantic preprocessing. They argued that low-level visual information (primarily word length) is the main determinant of the patterns of fixations on succeeding words. Could it be that the duration of a saccade (approximately 250 ms) does not provide enough time for the extraction of semantic features? That is, is the task one in which only effects of visual similarity can be uncovered? It has been established, for example, that phonological codes are used in integrating information across saccades. Pollatsek, Lesch, Morris, and Rayner (1992) demonstrated that phonological information acquired on one fixation from a word in the parafovea is used to help identify that word when it is later fixated. A homophone of a target word, when shown as a preview in the parafovea facilitated processing of the target word seen on the next fixation more than a preview of a word matched with the homophone in visual similarity to the target word. But the question regarding possible semantic integration has not yet been well researched. One approach to studying the issue of semantic integration is to examine the processes that occur during normal reading. In a study by Inhoff, Starr, and Shindler (2000) the influence of various preview characteristics on reading processes was explored. Again when previews were visually distinctive or orthographically illegal, longer viewing times were reported for target words. Even though larger semantic contexts were included within the materials that were used, notable semantic preview effects were largely absent. The contexts used, though consistent with the specific items under investigation did not "overlap" as specifically as orthographic or visual features can overlap. The primary difficulty in examining the potential effects of semantically related parafoveal preview items is the fact that semantically-related words can vary on dimensions that are perceived to be more difficult to control than, say, orthography, word frequency, word length, etc. To adequately approach this issue, a better tool might be the use of cognates, homographic noncognates, and translation equivalents where the relative semantic features across language can be more highly controlled. Bilingual investigations. The use of a bilingual paradigm for the current issue allows us to more clearly separate the effects of purely visual characteristics and those that involve semantics as well. We can create the condition whereby items are visually similar and have the same meaning (cognates such as cream-crema) and compare those conditions to ones in which items share meaning but no physical features (e.g., witch-bruja). In other words, can we find evidence of a semantic parafoveal preview effect using cognates? It is important to note, as discussed
Sentence Contexts in Reading
121
elsewhere (cf. Altarriba & Soltano, 1996) that cross-language translations may be "closer" in meaning for fluent bilinguals than within-language synonyms for a monolingual speaker. Synonymous words (e.g., pupil-students; soil-dirt) may not share the same relative frequencies, may co-occur in the language with different levels of probability, and may carry connotations and denotations that are not easy to single out. On the other hand, for a fluent bilingual, it is likely that a word in one language (e.g., chair) that is concrete and commonly understood to have one primary meaning will share a large number of features, attributes, etc. with its otherlanguage translation (e.g., silla, in Spanish). For this reason, the use of these types of relations across languages can provide a more sensitive measure of the relative semantic "carry over" from a preview to a subsequent target word. Moreover, in the case of cognates, not only are semantic features shared, there is also a large degree of orthographic overlap. These items have been shown to produce priming in demonstrations of cross-language masked priming (see e.g., Cristoffanini, Kirsner, & Milech, 1986; de Groot & Nas, 1991). Only one study, to date, has examined the influence of semantic attributes on parafoveal processing in cross-language situations--Altarriba et al. (2001). In the first experiment reported within that work, fluent English-Spanish bilinguals viewed a preview word to the right of fixation. When they moved their eyes, the preview was replaced by either an English or Spanish target word having the same number of letters as the preview. Their task was to name the target word aloud. The preview was either identical to, a translation of, or different from the target word. Some of the translations were cognates and some were not. If the parafoveal preview is based solely on visual similarity, then both bilinguals and monolinguals should show equal facilitation for cognates (e.g., cream-crema) as opposed to identical controls. However, if semantic information is extracted, bilinguals should show an added benefit for cognate translations (e.g., cream-crema) and some benefit for noncognate translations (e.g., strong-fuerte). Perhaps most relevant to the present discussion was their second experiment in which participants read sentences containing the target word. A preview word was replaced by the target word when the reader's saccade crossed a boundary location. A sample sentence including a noncognate translation was the following: (5) The chocolate cake was very dulce and much too high in calories. When the participant crossed a specific boundary location that preceded the presentation of the Spanish word, that word became the English translation "sweet". (The Spanish word was not highlighted in any way, in the actual study.) In summary, the results of both experiments indicated that semantic codes are not integrated across fixations. The studies served as a stronger, more sensitive test, of the role of semantic features in the processing of parafoveally presented information. In both experiments, the preview benefit from the translation conditions was no greater than would be predicted by the orthographic similarity between preview and target. In summary, the data indicated that readers obtained no useful semantic information from words presented parafoveally that might enable
122
Jeanette Altarriba and Jennifer L. Gianico
quicker identification of subsequent, related target items. Evidence was clearly provided for the benefit of orthographic or lexical overlap across languages, but there was no added benefit of any additional semantic overlap. This is an area that is in need of further investigation, particularly as the limits and usefulness of a bilingual paradigm in this area of research are yet to be explored. More will be said in this vein, in the final section of this chapter.
Within-Sentence Resolution of Lexical Ambiguities
How do sentence contexts operate in the processing of lexically ambiguous words? In the monolingual literature, researchers worry about when or if the two meanings of an ambiguous word are accessed. In other words, is there a preferred meaning that is accessed first? Or, are both meanings activated, but only one, the appropriate one in the sentence, is actually accessed and retrieved? Within this section, an introduction to the basic concepts investigated in this area of research will be presented. Then a discussion of the monolingual literature focusing on the current problems that researchers are trying to address will be presented. Finally, the bilingual literature will be examined as it pertains to the issues that are of concern in this general area of research. Data will be reviewed that demonstrate the ways in which individuals who are bilingual access the various meanings of a crosslanguage, homographic noncognate--a specific case of lexical ambiguity. Introduction. Ambiguity in language processing has been a topic of great interest for both monolingual and bilingual researchers. An example of an ambiguous word in English is the word "toast". Toast can be read as 'crusty bread', or, as a speech of good will (usually with raised champagne glasses) given during a celebration (Martin, Vu, Kellas, & Metcalf, 1999). The investigation of lexical ambiguity is a way to understand how people process language in general. Within this area of study, researchers are concerned with the ways in which a word is disambiguated with regards to meaning. Accessing a contextually appropriate meaning for an ambiguous word in a sentence allows the reader to understand the concepts represented by that sentence much more easily. Readers may disambiguate a word based upon the frequencies of its various meanings. Meaning frequencies have been measured using word association tasks (Gorfein, Berger, & Bubka, 2000) and through studies involving rated frequency (Griffin, 1999). Within the monolingual literature, various models have been proposed to describe the ways in which readers access either or both meanings of an ambiguous word. Three such models have been outlined by Simpson (1984): contextdependent, ordered-access, and exhaustive access. A context-dependent model assumes that the surrounding context, a sentence for example, primes the selection of a particular meaning of a word. There is truly no ambiguity within the situation, as the appropriate meaning is the one that is accessed and retrieved in the process of comprehension. While this model has been supported with empirical data (see e.g., Schvaneveldt, Meyer, & Becker, 1976; Simpson, 1981), other researchers have
Sentence Contexts in Reading
123
provided evidence that a meaning that is inappropriate to a context is sometimes accessed as well (e.g., Onifer & Swinney, 1981; Seidenberg, Tanenhaus, Leiman, & Bienkowski, 1982; see also Li & Yip, 1998, for work in the spoken domain). An ordered-access model suggests that the meanings of an ambiguous word are accessed according to word frequency. The most frequent meaning of a word is accessed first and then evaluated according to the sentence context for appropriateness. Other meanings are activated in a serial manner until the most appropriate match is detected. Some would argue that this process is context independent--that is, the process of a serial search occurs automatically regardless of the nature of the related context (see e.g., Hogaboam & Perfetti, 1975). In general, support has been limited for this model stemming from sentence comprehension tasks (Holmes, 1979) and lexical decision tasks (Simpson, 1981). Finally, an exhaustive access model suggests that all of the meanings of an ambiguous word are activated simultaneously. Selection of one of those meanings rests on its appropriateness with regards to context. When the appropriate meaning has been determined, all other inappropriate meanings are discarded. The access of the lexical information is context-independent; however, the decision process involved in isolating a single meaning depends on the strength of the context (Onifer & Swinney, 1981; Seidenberg et al., 1982; Swinney, 1979). This model has received the strongest support thus far (however, see Simpson & Burgess, 1985, for evidence of selectivity) in the monolingual literature (see Simpson, 1984, for a review). Various bilingual researchers have used this model, and indeed, many of the monolingual paradigms to investigate cross-language ambiguity. Note, however, that some of the results that have been reported should be interpreted with reference to the methods that have been used as comparisons between "on-line" and "off-line" tasks may lead to different theoretical predictions. In order to understand sentence processing in bilinguals, researchers (Beauvillain & Grainger, 1987; Dijkstra, Grainger & Van Heuven, 1999; Gerard & Scarborough, 1989) have used lexically ambiguous words as a tool to explore issues of lexical access. Techniques such as eyetracking, sentence comprehension tasks, and priming (see e.g., Williams & Colombo, 1995) have been used to examine the time course involved in the access of the various meanings of an ambiguous, crosslanguage word. An example of an interlexically ambiguous word is the word "pie." In English, "pie" is a dessert, while in Spanish it is the word for foot. A representative selection of monolingual and bilingual literature concerning the processing of ambiguous items and factors influencing this processing within contexts and in isolation will be reviewed in turn below. The monolingual studies were conducted in English. Monolingual investigations. While the bulk of the literature involving monolingual participants supports a model in which multiple meanings are activated, the point at which one, appropriate meaning is selected has been under debate. In a study by Yates (1978), a priming technique was used in which participants received sentences of the kind "A(n) A is a(n) B" for verification in a timed task. "A" was an ambiguous word and "B" was either the dominant or
124
Jeanette Altarriba and Jennifer L. Gianico
subordinate meaning of that word. For example, "A mug is a cup," versus "A mug is a face". In cases in which the dominant meanings were most often represented within the task, participants showed greater priming for dominant over subordinate meanings. However, when B was equally likely to represent a dominant or a subordinate meaning of the word, priming was equivalent for both. Yates concluded that since both meanings of the ambiguous words were successfully primed in the latter demonstration, then clearly an exhaustive process concerning access was in operation. Yates further added that the processing of ambiguity appears to be affected by task demands and temporal parameters that are contained within a study. Other evidence of the multiple activation of meanings for ambiguous words comes from work by Swinney (1979) and Onifer and Swinney (1981). Swinney presented participants with sentences containing lexical ambiguities in an auditory mode (See Heredia & Stewart, this volume, for details of the cross-modal lexical priming (CMLP) technique.) Simultaneously, participants were shown words for a lexical (word-nonword) decision. These words corresponded to either the dominant or the subordinate meaning of the ambiguous words. Facilitation in response times occurred for both the dominant and the subordinate word targets as compared to unrelated controls. Moreover, when the sentence contexts were altered to strongly bias one or the other meaning, facilitation was still produced in both conditions. Again, it appears that this evidence favors a model in which both or all meanings of an ambiguous word are activated and available for further processing. Lexical access in this case, then, appears to be autonomous rather than directed by contextual constraints. Onifer and Swinney reported similar results as they controlled for sentence processing by following the above task with a sentence recognition task. In other words, participants were indeed processing the sentential material as they made their lexical decisions. A lexical decision task was also employed by Simpson (1981) first for pairs of words in isolation and then for targets that followed ambiguous words embedded within sentences. Unlike the previous work above, only responses for dominant meanings were facilitated as compared to unrelated controls. This was also the case with sentences. However, when sentence contexts were only weakly biased towards the subordinate meaning, evidence of the activation of both meanings was found. In other words, for sentential contexts there was ultimately evidence in favor of the activation of multiple meanings of ambiguous words only when bias occurred towards the least frequent meaning. Here the rationale is that the meanings are organized in memory on the basis of frequency. First, the most frequent meaning is activated and only when it is deemed inappropriate are the subsequent meanings considered for satisfactory comprehension of the sentence. Simpson argues that the overall pattern of his data do not directly support an exhaustive model. Sentence contexts appear to play an important role in determining the extent to which more than one meaning is activated for embedded, ambiguous words. These results, however, call into question the role of timing parameters and their influence on the access and selection of word meanings. Perhaps multiple meanings are initially accessed; however, within some period of time, perhaps prior to a response, a
Sentence Contexts in Reading
125
selection process occurs. This possibility has been explored and confirmed to some extent by others (see e.g., Seidenberg et al., 1982). Seidenberg et al. (1982) suggested that lexical access might be guided more by the properties of the words themselves and by presentation time than by sentential context, p e r se. When participants listened to sentences that contained a word that was highly semantically related to one meaning of a subsequent ambiguous target, naming responses were facilitated as compared to an unrelated control. However, when sentence contexts corresponded to a single meaning of the subsequent target through the use of other types of information, such as syntactic class, evidence appears for the access of multiple meanings of a word. This situation is somewhat similar to having a more weakly biased context with reference to an ambiguous target word. Moreover, when stimulus onset asynchrony (SOA; time from the presentation of a prime word to a subsequent target) was varied, it was reported that selection of a final reading of a word occurred within 200 ms. At shorter SOAs there is evidence for multiple access while at longer SOAs there is evidence of meaning selection (but see Van Petten & Kutas, 1987, for conflicting ERP (eventrelated potential) data, for the short SOA). It appears that there is a time course in the access and processing of multiple meanings of ambiguous words and that that time course can be examined and documented for the selection of one meaning over another (see e.g., Tanenhaus, Leiman, & Seidenberg, 1979). Perhaps a combination of exhaustive and context-dependent processes is at work in the resolution of lexical ambiguities. In keeping with an investigation of the role of time in the access of meaning, Lucas (1987) investigated the degree to which the access of multiple meanings is autonomous while the selection of one particular meaning is post-lexical, occurring at a later stage of processing. Using a visual, word priming task, similar to those reported before, the data indicated that multiple access of meanings occurred at about 100 ms after the presentation of an ambiguous word. In other words, lexical decision responses to targets representing both dominant and subordinate meanings as related to an ambiguous word embedded in sentences were equally facilitated. Before concluding this section, it is important to review the use of paradigms and techniques other than lexical or sentential priming. Eyetracking techniques have also been applied in the understanding of lexical ambiguity processing within sentences. Rayner and Frazier (1989) measured gaze durations (the amount of time a word is fixated before a movement occurs to a point outside of the word) while participants read sentences containing an ambiguous word. Sentence contexts were either biased towards the strongest or dominant meaning, or, were unbiased. When contexts were unbiased, gaze durations were longer for the ambiguous words than in the biased conditions. Reading was facilitated for the ambiguous targets when the sentences favored the dominant interpretation of the words. When the prior context aided in disambiguating a word in the formerly nonbiased contexts, gaze durations were reduced. However, if the disambiguating information was in favor of a subordinate meaning, then gaze durations were lengthened. These data favor a model in which the successful integration of a single meaning of an ambiguous word within the sentence context terminates-the access of other subordinate
126
Jeanette Altarriba and Jennifer L. Gianico
meanings. In fact, a context that is biased towards a dominant meaning elicits that meaning thereby eliminating the need to search and access a subordinate meaning. If a context is unbiased, there is a possibility that multiple meanings are accessed, in which case the more dominant meaning may still have an advantage over the immediate selection of the subordinate meaning. These data argue for a model that relies on the role of the strength of the context and the organization or structure of that context in the access of meanings for ambiguous words. These results conflict with those of Onifer and Swinney (1981) and others in that in the present case, evidence was produced favoring only the dominant meaning of an ambiguous word when context biased the interpretation of that meaning. Note, however, that there were task differences, timing differences, etc., that should be reviewed carefully before drawing conclusions regarding multiple vs. selective access. Moreover, one variable that is not that easy to compare is the stimulus sentences themselves, their construction, and the ways in which "bias" is manipulated. In a subsequent study also utilizing the eyetracking technique, Binder and Rayner (1998) reported findings similar to those reported above. Namely, gaze durations were lengthened for ambiguous target words when embedded in sentences that biased the subordinate meaning of the word. This was the case even though the context was very strongly biased. The pattern of data reported for gaze durations was also mimicked when the first fixation durations were used as the dependent measure (e.g., the amount of time during which the word is first fixated while reading). It is possible that the access of meanings for ambiguous words is exhaustive, however, the context and frequency of meanings may determine the order in which those meanings are accessed (see e.g., Simpson & Krueger, 1991). Other factors that may influence the degree to which multiple meanings of an ambiguous word are accessed and subsequently processed include individual differences such as working memory capacity (see e.g., Miyake, Just, & Carpenter, 1994). Miyake et al. presented high- and low-span readers with sentences containing an ambiguous word for verification. Information that would help to delimit the meaning of that word came after the word within the sentential context. It was assumed that high memory span participants could retain ihe multiple meanings of an ambiguous word longer in memory, in anticipation of disambiguating information, as compared to low memory span participants. For low-span participants, there was a larger ambiguity effect in verification times when the disambiguating information supported-the subordinate meaning of the word as compared to the dominant meaning. The authors inferred that these participants likely held the dominant meaning in memory and encountered difficulty in responding when the sentence favored the subordinate meaning. The same conflict was not experienced by the high-span readers. Apparently, this latter group was not hindered by the presence of information that would favor the secondary or less frequent meaning of the ambiguous target word. Bilingual investigations. Some studies that have examined ambiguity in the context of language processing across various languages have focused mostly on syntactic ambiguity and strategies engaged in parsing (e.g., Cuetos & Mitchell,
Sentence Contexts in Reading
12 7
1988; Gibson, Pearlmutter, Canseco-Gonzalez, & Hickok, 1996; Gilboy, Sopena, Clifton, & Frazier, 1995; MacDonald, Pearlmutter, & Seidenberg, 1994; Mitchell & Cuetos, 1991; Thornton, MacDonald, & Gil, 1999; Trueswell & Tanenhaus, 1994; see also Matlock & Heredia, this volume). The focus of the current discussion, however, is on ambiguities that are strictly lexical in nature. In bilingual studies of lexical ambiguity each meaning of a given word is language specific. For example, for Spanish-English bilinguals, the word "red" can be processed as a color word in English or as the Spanish translation for the English word "net." Assuming that Spanish is dominant for this group of bilinguals, then it is likely that the Spanish meaning has been more frequently encountered. (See Heredia, 1997, for a discussion of language dominance and language retrieval.) Therefore, that meaning may act as the so-called "dominant" meaning while the English meaning is akin to the "subordinate" meaning. Some of the very same questions considered in the monolingual literature are of interest here. Namely, are both meanings of a crosslanguage ambiguous word activated when the word is encountered? If both meanings are available, how is the correct one accessed? Could access be determined by word frequency within a given language? Most of the research on these issues has been conducted using lexical decision and naming tasks again examining either words in isolation or within biasing or neutral contexts (see e.g., Grosjean, 1988). In one of the classic studies in this area, Beauvillain and Grainger (1987) engaged English-French bilinguals in a priming task involving interlexical homographs. They were interested in examining the degree to which an instruction to process these words in a specific language mode was sufficient to block the alternative meaning in the other language. Could individuals process cross-language ambiguous words in a language-selective manner without any influence of their alternate language system (cf. Scarborough, Gerard, & Cortese, 1984)? Participants were shown prime-target pairs in which the prime was always in French and could also be an interlexical homograph (e.g., "coin" means corner in French). Lexical decisions in English were to be performed on the targets. In addition to finding facilitation for the targets that related to the French reading of the primes, targets that related to the English meaning were also facilitated. This occurred at a short SOA (i.e., 150 ms) indicating the availability of both meanings of the homograph. At a longer SOA (i.e., 750 ms) the targets related to the English meaning of the prime were not facilitated. Hence, the authors argued, as was the case in monolingual investigations, that longer SOAs allow for the selection of the meaning in the appropriate language. When word frequency was manipulated in a second study, priming occurred for word targets that were related to the higher frequency reading of their primes regardless of the suggested language for processing those primes. In other words, it appears that the most frequent meaning of a prime "pops out" while reading regardless of the language mode the participant has been instructed to follow. It appears that the access of the meanings of interlexical homographs depends quite heavily on the relative frequencies of those meanings in each language (see also Gerard & Scarborough, 1989, for a similar finding with a single-word lexical decision task). High frequency readings become more available
128
Jeanette Altarriba and Jennifer L. Gianico
than low frequency readings regardless of language of processing (cf. de Groot, Delmaar, & Lupker, 2000; Dijkstra, Van Jaarsveld, & Ten Brinke, 1998). Interestingly, this latter finding reported by Beauvillain and Grainger occurred at an SOA of 150 ms. While the authors do not relate their findings to the various models of lexical ambiguity processing, it would appear that the activation of a specific meaning takes place quite early in the processing of cross-language ambiguity (as in the monolingual case reported by Simpson & Burgess, 1985). It's not entirely clear from the pattern of data whether or not both meanings are initially accessed prior to the activation of meaning on the basis of word frequency. Also, note that within these experiments, a single prime word served as the biasing context. Perhaps the less frequent meanings of the ambiguous words would show priming under conditions in which the context is a larger group of words such as a sentence with a strong bias in favor of those meanings. A recent investigation by Dijkstra, Grainger, and Van Heuven (1999) examined the processing of interlexical homographs by Dutch-English bilingual speakers. Participants performed a lexical decision task in English with words that varied in their degree of semantic, orthographic and phonological overlap across languages. Reaction times were facilitated to the extent that items overlapped semantically and orthographically while phonological overlap hindered performance. That is, when a word that was to be judged for lexicality in English was phonologically similar to its Dutch counterpart, reaction times were slowed. Of interest for the current discussion is the finding that the activation of word meanings for interlexical homographic noncognates, in general, is non-selective. That is, there is clearly some effect of the non-target language on the processing of the current target language. This is especially true when the meaning in the non-target language is of relatively high frequency. Although again, the data point to a frequency account of the relative influence of an interlexical homograph's two meanings on word processing, the data do not rule out the possibility that both meanings are initially activated and available for future processing. Would bilingual participants operate in a language-selective manner when presented with interlexical homographs embedded in strongly biased sentence contexts? Altarriba, Carlo, and Kroll (1992) presented Spanish-English bilinguals with sentences in English and in Spanish. Sentences were chosen as contexts in order to increase the role of conceptual activation in processing. Participants were required to name a target word embedded in each sentence. The target words included homographic noncognates (e.g., "fin" means end in Spanish) or controls matched in frequency and in length. Sample sentences appear below. (6) We knew it was a shark because we saw its FIN as it approached the boat. (Ambigous/Congruous) (7) We knew it was a shark because we saw its RAM as it approached the boat. (Control/Incongruous) If lexical access for cross-language homographs is language selective (see e.g., Costa, Miozzo, & Caramazza, 1999), then performance should not be influenced by
Sentence Contexts in Reading
129
the alternate (or less frequent) language meaning supported by the context. However, if participants access both language meanings of the words, naming latencies for the homographs should be slower than those for matched controls. The results suggested that when the sentence context provided conceptual support for the dominant language interpretation, knowledge of the meaning of the word in the other language did not affect performance. However, when processing sentences in the language that matched the subordinate meaning, even with adequate contextual support, there was evidence that the dominant language interpretation of the homograph influenced performance. These results therefore provided evidence for both selective and multiple access of meaning (but see also, Beauvillain, 1992, who argues against language selectivity). When homographic noncognates were embedded in Spanish sentence contexts that biased the dominant language meaning of the homograph, it appeared that participants could selectively access the appropriate meaning. These results support those of Simpson (1981) and Tabossi (1988), in the monolingual literature. However, when homographs were embedded in Spanish or English sentence contexts that biased the alternate language meanings of the words, participants experienced interference resulting in increased naming times as compared to matched controls. This was also the case when homographs were embedded in English contexts that biased the non-dominant meaning. These results are consistent with those reported by Onifer and Swinney (1981) and Seidenberg et al. (1982). Previous work on homographic noncognates has included tasks in which words were presented in isolation. The tasks used may not have provided a situation in which conceptual information could also be accessed. The effects reported by Altarriba et al. (1992) support the idea that the processing of cross-language ambiguity involves the interaction of both lexical retrieval and contextual constraint--a conclusion most likely relevant to monolingual applications, as well. If lexical processes alone determine the pattern of access, then participants should have responded equally to Spanish homographs embedded in contexts that biased both the dominant and the non-dominant language meanings, for example. If context alone determines the pattern of access, then participants should have shown the standard congruity effect for homographs regardless of the relationship between the language context and the relative frequency of the item in either language.
Conclusions
The findings reported here indicate that bilingual measures can be used to investigate issues in language representation when it becomes necessary to distinguish between lexical and conceptual (or semantic) levels of processing. To accomplish this, researchers have capitalized on the fact that for cross-language items it is possible to either hold semantic features constant and vary orthography, or, hold orthography constant and vary semantic overlap. This possibility is somewhat limited in within-language comparisons as synonyms often do differ in terms of their connotative or denotative meanings making it difficult to assess the
130
Jeanette Altarriba and Jennifer L. Gianico
degree of featural overlap across items. In the current work, this approach resulted in the following conclusions: (1) individuals do not seem to benefit from a situation in which parafoveally presented information solely overlaps in terms of semantic features with foveally presented information; (2) sentence constraint effects operate at both a lexical and a conceptual level of representation to determine the nature of upcoming information; and, (3) the meanings of ambiguous words may be accessed in both a selective and a multiple manner depending on the strength of the surrounding context and the experimental methods chosen. The interaction between language, memory, and perception is a strong predictor of how both monolinguals and bilinguals process language inputs across varying modalities.
Future Research Directions
The implications of the work reviewed here are vast and exciting in terms of future research questions. It is clear that the use of a variety of measures and paradigms would increase the generalizability of the findings reported here and that perhaps future work in the spoken or cross-modal domains might best capture the representation of language for individuals who likely process differently in the auditory versus the visual domain. Language-mixing and language-switching in both written and spoken forms are areas of research that could benefit from the application of various on-line procedures as the dynamics of how and when they occur are just beginning to be understood (cf. Heredia & Altarriba, 2001). Finally, lexical ambiguities should be explored across languages that may share only some of their orthographic features, such as several Indian languages, to further disambiguate the roles of semantics and lexical processing in their mental representation. In summary, the current areas of research interest focused on the exploration of the use of sentential contexts in reading, memory, and semantic disambiguation for both monolinguals and bilinguals and emphasized the benefits of considering the ways in which levels of language representation can be separated or combined to answer interesting questions in the domain of language representation.
References
Altarriba, J., Carlo, M. S., & Kroll, J. F. (1992, March). Language dominance and the processing of cross-language ambiguity. Poster presented at the Fifth Annual CUNY Conference on Human Sentence Processing, New York, NY. Altarriba, J., Kambe, G., Pollatsek, A., & Rayner, K. (2001). Semantic codes are not used in integrating information across eye fixations: Evidence from fluent Spanish-English bilinguals. Perception & Psychophysics, 63, 875-890. Altarriba, J., Kroll, J. F., Sholl, A., & Rayner, K. (1996). The influence of lexical and conceptual constraints on reading mixed-language sentences: Evidence from eye fixations and naming times. Memory & Cognition, 24, 477-492.
Sentence Contexts in Reading
131
Altarriba, J., & Santiago-Rivera, A. L. (1994). Current perspectives on using linguistic and cultural factors in counseling the Hispanic client. Professional Psychology: Research and Practice, 25,388-397. Altarriba, J., & Soltano, E. G. (1996). Repetition blindness and bilingual memory: Token individuation for translation equivalents. Memory & Cognition, 24, 700-711. Balota, D. A., & Rayner, K. (1983). Parafoveal visual information and semantic contextual constraints. Journal of Experimental Psychology: Human Perception and Performance, 9, 726-738. Beauvillain, C. (1992). Orthographic and lexical constraints in bilingual word recognition. In R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 221-235). Amsterdam: Elsevier Science Publishers. Beauvillain, C., & Grainger, J. (1987). Accessing interlexical homographs: Some limitations of a language-selective access. Journal of Memory and Language, 26, 658-672. Binder, K., & Rayner, K. (1998). Contextual strength does not modulate the subordinate bias effect: Evidence from eye-fixations and self-paced reading. Psychonomic Bulletin & Review, 5, 271-276. Briihl, D., & Inhoff, A. W. (1995). Integrating information across fixations during reading: The use of orthographic bodies and of exterior letters. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 55-67. Costa, A., Miozzo, M., & Caramazza, A. (1999). Lexical selection in bilinguals: Do words in the bilinguals' two lexicons compete for selection? Journal of Memory and Language, 41, 365 397. Cristoffanini, P., Kirsner, K., & Milech, D. (1986). Bilingual lexical representation: The status of Spanish-English cognates. Quarterly Journal of Experimental Psychology, 38A, 367-393. Cuetos, F., & Mitchell, D. C. (1988). Cross-linguistic differences in parsing: Restrictions on the use of the late closure strategy in Spanish. Cognition, 30, 73105. de Groot, A. M. B., Delmaar, P., & Lupker, S. J. (2000). The processing of interlexical homographs in a bilingual and a monolingual task: Support for nonselective access to bilingual memory. Quarterly Journal of Experimental Psychology, 53, 397-428. de Groot, A. M. B., & Nas, G. L. (1991). Lexical representation of cognates and noncognates in compound bilinguals. Journal of Memory and Language, 30, 90123. Dijkstra, T., Grainger, J., & Van Heuven, W. J. B. (1999). Recognition of cognates and interlingual homographs: The neglected role of phonology. Journal of Memory and Language, 41, 496-518. Dijkstra, T., Van Jaarsveld, H., & Ten Brinke, S. (1998). Interlingual homograph recognition: Effects of task demands and language intermixing. Bilingualism, Language, and Cognition, 1, 51-66.
132
Jeanette Altarriba and Jennifer L. Gianico
Duffy, S. A., Henderson, J. M., & Morris, R. K. (1989). The semantic facilitation of lexical access during sentence processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 791-801. Fischler, I. (1985). Word recognition, use of context, and reading skill among deaf college students. Reading Research Quarterly, 20, 203-218. Fischler, I., & Bloom, P. A. (1980). Rapid processing of the meaning of sentences. Memory & Cognition, 8, 216-225. Gerard, L. D., & Scarborough, D. L. (1989). Language specific lexical access of homographs by bilinguals. Journal of Experimental Psychology: Learning, Memory and Cognition, 15, 305-315. Gibson, E., Pearlmutter, N. J., Canseco-Gonzalez, E., & Hickok, G. (1996). Recency preference in the human sentence processing mechanism. Cognition, 59, 23-59. Gilboy, E., Sopena, J. M., Clifton, C., Jr., & Frazier, L. (1995). Argument structure and association preferences in Spanish and English complex NPs. Cognition, 54, 131-167. Gorfein, D. S., Berger, S., & Bubka, A. (2000). The selection of homograph meaning: Word association when context changes. Memory & Cognition, 28, 766773. Griffin, Z. M. (1999). Frequency of meaning use for ambiguous and unambiguous words. Behavior Research Methods, Instruments, & Computers, 31, 520-530. Grosjean, F. (1980). Exploring the recognition of guest words in bilingual speech. Language and Cognitive Processes, 3, 233-274. Heredia, R. R. (1997). Bilingual memory and hierarchical models" A case for language dominance. Current Directions in Psychological Science, 6, 34-39. Heredia, R. R., & Altarriba, J. (2001). Bilingual language mixing: Why do bilinguals code switch? Current Directions in Psychological Science, 1O, 164-168. Hogaboam, T. W., & Perfetti, C. A. (1975). Lexical ambiguity and sentence comprehension. Journal of Verbal Learning and Verbal Behavior, 14, 265-274. Holmes, V. M. (1979). Accessing ambiguous words during sentence comprehension. Quarterly Journal of Experimental Psychology, 31, 569-589. Inhoff, A. W., Starr, M., & Shindler, K. L. (2000). Is the processing of words during eye fixations in reading strictly serial? Perception & Psychophysics, 62, 1474-1484. Li, P., & Yip, M. C. (1998). Context effects and the processing of spoken homophones. Reading and Writing: An Interdisciplinary Journal, 10, 223-243. Lucas, M. M. (1987). Frequency effects on the processing of ambiguous words in sentence contexts. Language and Speech, 30, 25-46. MacDonald, M. C., Pearlmutter, N. J., & Seidenberg, M. S. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101, 676-703. Martin, C., Vu, H., Kellas, G., & Metcalf, K. (1999). Strength of discourse context as a determinant of the subordinate bias effect. The Quarterly Journal of Experimental Psychology, 52, 813-839.
Sentence Contexts in Reading
133
McClelland, J. L., & O'Regan, J. K. (1981). Expectations increase the benefit derived from parafoveal visual information in reading words aloud. Journal of Experimental Psychology: Human Perception and Performance, 7, 634-644. McElree, B., & Griffith, T. (1998). Structural and lexical constraints on filling gaps during sentence comprehension: A time-course analysis. Journal of Experimental Psychology: Learning, Memory, and Cognition, 24, 432-460. Mitchell, D. C., & Cuetos, F. (1991). The origins of parsing strategies. In C. Smith (Ed.), Current issues in language processing (pp. 1-12). Austin: University of Texas Press. Miyake, A., Just, M. A., & Carpenter, P. A. (1994). Working memory constraints on the resolution of lexical ambiguity: Maintaining multiple interpretations in neutral contexts. Journal of Memory and Language, 33, 175-202. Morris, R. K. (1994). Lexical and message-level sentence context effects in fixation times in reading. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 92 103. Onifer, W., & Swinney, D. A. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency and contextual bias. Memory & Cognition, 9, 225 236. Paap, K. R., & Newsome, S. L. (1981). Parafoveal information is not sufficient to produce semantic or visual priming. Perception & Psychophysics, 29, 457-466. Pfaff, C. W. (1979). Constraints on language mixing: lntrasentential codeswitching and borrowing in Spanish/English. Language, 55, 291-318. Pollatsek, A., Lesch, M., Morris, R. K., & Rayner, K. (1992). Phonological codes are used in integrating information across saccades in word identification and reading. Journal of Experimental Psychology: Human Perception and Performance, 18, 148-162. Rayner, K. (1978). Foveal and parafoveal cues in reading. In J. Requin (Ed.), Attention and performance (Vol. 7, pp. 149-162). Hillsdale, NJ: Erlbaum. Rayner, K., & Frazier, L. (1989). Selection mechanisms in reading lexically ambiguous words. Journal of Experimental Psychology." Learning, Memory, and Cognition, 15, 779-790. Rayner, K., & Morris, R. K. (1992). Eye movement control in reading: Evidence against semantic preprocessing. Journal of Experimental Psychology." Human Perception and Performance, 18, 163-172. Rayner, K., McConkie, G. W., & Ehrlich, S. F. (1978). Eye movements and integrating information across fixations. Journal of Experimental Psychology." Human Perception and Peformance, 4, 529-544. Rayner, K., McConkie, G. W., & Zola, D. (1980). Integrating information across eye movements. Cognitive Psychology, 12, 206-226. Rayner, K., & Pollatsek, A. (1989). The psychology of reading. Englewood Cliffs, NJ: Prentice Hall. Rayner, K., Well, A. D., Pollatsek, A., & Bertera, J. H. (1982). The availability of useful information to the right of fixation in reading. Perception & Psychophysics, 31, 537-550.
134
Jeanette Altarriba and JenniJerL. Gianico
Scarborough, D. L., Gerard, L., & Cortese, C. (1984). Independence of lexical access in bilingual word recognition. Journal of Verbal Learning and Verbal Behavior, 23, 84-99. Schmauder, A. R., Morris, R. K., & Poyner, D. V. (2000). Lexical processing and text integration of function and content words: Evidence from priming and eye fixations. Memory & Cognition, 28, 1098-1108. Schustack, M. W., Ehrlich, S. F., & Rayner, K. (1987). The complexity of contextual facilitation in reading: Local and global influences. Journal of Memory and Language, 26, 322-340. Schvaneveldt, R. W., Meyer, D. E., & Becker, C. A. (1976). Lexical ambiguity, semantic context, and visual word recognition. Journal of Experimental Psychology: Human Perception and Performance, 2, 243-256. Schwanenflugel, P. J., & LaCount, K. L. (1988). Semantic relatedness and the scope of facilitation for upcoming words in sentences. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 344-354. Schwanenflugel, P. J., & Shoben, E. J. (1985). Tile influence of sentence constraints on the scope of facilitation for upcoming words. Journal of Memory and Language, 24, 232-252. Seidenberg, M. S., Tanenhaus, M. K., Leiman, J. M., & Bienkowski, M. (1982). Automatic access of the meanings of ambiguous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14, 489-537. Simpson, G. B. (1981). Meaning dominance and semantic context in the processing of lexical ambiguity. Journal of Verbal Learning and Verbal Behavior, 20, 120-136. Simpson, G. B. (1984). Lexical ambiguity and its role in models of word recognition. Psychological Bulletin, 96, 316-340. Simpson, G. B., & Burgess, C. (1985). Activation and selection processes in the recognition of ambiguous words. Journal of Experimental Psychology: Human Perception and Performance, 11, 28-39. Simpson, G. B., & Krueger, M. A. (1991). Selective access of homograph meanings in sentence context. Journal of Memory and Language, 30, 627-643. Simpson, G. B., Peterson, R. R., Casteel, M. A., & Burgess, C. (1989). Lexical and sentence context effects in word recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 88-97. Stanovich, K. E., & West, R. F. (1979). Mechanisms of sentence context effects in reading: Automatic activation and conscious attention. Memory & Cognition, 7, 77-85. Stanovich, K. E., & West, R. F. (1981). The effect of sentence context on ongoing word recognition: Tests of a two-process theory. Journal of Experimental Psychology: Human Perception and Performance, 7, 658-672. Stanovich, K. E., & West, R. F. (1983). On priming by a sentence context. Journal of Experimental Psychology: General 112, 1-36. Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18, 645-659.
Sentence Contexts in Reading
135
Tabossi, P. (1988). Accessing lexical ambiguity in different types of sentential context. Journal of Memory and Language, 27, 324-340. Tanenhaus, M. K., Leiman, J. M., & Seidenberg, M. S. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18, 427-440. Thornton, R., MacDonald, M. C., & Gil, M. (1999). Pragmatic constraint on the interpretation of complex noun phrases in Spanish and English. Journal of Experimental Psychology: Learning, Memory, and Cognition, 25, 1347-1365. Traxler, M. J., Foss, D. J., Seely, R. E., Kaup, B., & Morris, R. K. (2000). Priming in sentence processing: Intralexical spreading activation, schemas, and situation models. Journal of Psycholinguistic Research, 29, 581-595. Trueswell, J. C., & Tanenhaus, M. K. (1994). Towards a lexicalist framework of constraint-based syntactic ambiguity resolution. In C. Clifton, Jr., L. Frazier, & K. Rayner (Eds.), Perspectives on sentence processing (pp. 155-179). Hillsdale, NJ: Erlbaum. Van Petten, C., & Kutas, M. (1987). Ambiguous words in context: An eventrelated potential analysis of the time course of meaning activation. Journal of Memory and Language, 26, 188-208. West, R. K., & Stanovich, K. E. (1982). Source of inhibition in experiments of the effect of sentence context on word recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 25, 385-399. Whitney, P., McKay, T., Kellas, G., & Emerson, W. A., Jr. (1985). Semantic activation of noun concepts in context. Journal of Experimental Psychology: Learning, Memory, and Cognition, I 1, 126-135. Williams, J. N., & Colombo, L. (1995). Constraints on the range of contextindependent priming from ambiguous words. Psychological Research, 58, 38-50. Yates, J. (1978). Priming dominant and unusual senses of ambiguous words. Memory & Cognition, 6, 636-643. Yoon, K. K. (1992). New perspective on intrasentential code-switching: A study of Korean English switching. Applied Psycholinguistics, 13, 433-449.
This Page Intentionally Left Blank
Bilingual Sentence Processing- R.R. Heredia and J. Altarriba (Editors) 9 2002 Elsevier Science B.V. All rights reserved.
Exploring Language Asymmetries in Early Spanish-English Bilinguals: The Role of Lexical and Sentential Context Effects Arturo E. Hern~indez University of California, Santa Barbara
Abstract Recent research with bilinguals has uncovered a number of interesting effects of language dominance. Of particular interest has been the observation that lexical priming is larger from the dominant language to the subordinate language than vice-versa. This effect has also been found in previous studies using sentence priming, but only under visually degraded conditions. In the current study, participants were asked to name the target of a visual word pair that was presented instead of a word in an auditory context. In Experiment 1 (the within-language condition), the text and target language matched and the prime was in the opposite language. In Experiment 2 (the between-language condition), the text and prime language matched and the target was in the opposite language. Experiment 1 revealed no difference in the size of sentential and lexical priming as a function of target language. There was an interaction between sentential and lexical priming with English targets, and an additive pattern of sentential and lexical priming with Spanish targets. In Experiment 2, there was a reduction of both lexical and sentential priming for English targets when compared with Spanish targets. The results support a model in which fluency is reflected in both the speed of processing and the magnitude of priming effects. The implications of these results for current models of bilingual language processing are discussed.
Introduction In recent years, the timing involved in the activation of different sources of information has received a lot of attention in the psycholinguistic literature. This is clear from the large number of studies that have investigated the nature of language processing as it occurs in real time or "on-line." The importance of timing has also been a topic of study in the bilingual lexical-access literature. This paper describes a series of studies exploring how lexical and sentential priming interact in bilingual individuals. The inquiry will require consideration of a number of issues that have arisen in both the monolingual and bilingual word recognition literature, including the nature and timing of lexical and sentential context effects as well as these effects on within- and between-language priming. We will begin with a review of current
138
Arturo E. Hern6ndez
issues in monolingual word recognition before proceeding to issues that directly apply to priming in bilingual populations.
On the Nature of Lexical and Sentential Context Effects
A common paradigm for the investigation of on-line language processing is the semantic priming task. In this task, participants are shown a word pair that is either related (e.g., cat-dog) or unrelated (e.g., lock-dog). Numerous studies have found that college-aged adults are faster to pronounce or make a lexical decision about the second word in a related word pair than they are for the second word of an unrelated word pair (e.g., Neely, 1991). Despite the robustness of the semantic priming effect, there has been a debate about the locus and timing of the lexical priming effect (whether it is an effect that occurs during lexical access). In addition, a variety of possible frameworks have been proposed which account for a range of phenomena that are observed using the semantic priming paradigm. Traditionally, semantic priming was seen as a product of spreading activation within the lexicon. The strict dichotomy between lexical (automatic) priming and postlexical (controlled) priming, however, has met with a number of problems in the lexical priming literature (see Neely, 1991 for a review). This period has led to a number of altemative frameworks. These frameworks range from suggesting that lexical priming is the product of spreading activation and expectancy (Neely, 1991) or the product of the prime and a target forming a compound cue whose familiarity is assessed by accessing long-term memory (Hinton & Shallice, 1991; McNamara, 1992, 1994; Ratcliff & McKoon, 1994; Sharkey, 1989) to suggesting that priming is the product of a pattem of activation over a set of semantic primitives (Hinton, McClelland, & Rumelhart, 1986; Hinton & Shallice, 1991; Masson, 1991, 1995; Sharkey, 1989). The locus of sentence priming effects has also been the subject of debate. Traditionally, sentence context effects were thought of as being controlled postlexical effects that occurred after a lexical item was identified. This view has received mixed support in the sentence-priming literature and the lexical-ambiguity literature (Simpson, 1994; Stanovich & West, 1983). In the ambiguity literature, there is support in favor and against models in which lexical context effects (i.e., activation of an ambiguous word's two meanings) occur independently of sentential context effects (i.e., selection of only a sententially appropriate meaning) very early in processing (see Simpson, 1994 for a review). Thus, it is not clear that sentence context effects occur after lexical context effects. There is also little evidence that sentential contexts have different types of effects on word recognition or that they are dependent on lexical contexts. First of all, sentential priming and lexical priming show very similar effects with respect to facilitation, inhibition, and visual degradation (Stanovich & West, 1983). Second, the results from studies looking at lexical ambiguity and sentential priming suggest that sentential priming is not simply the product of lexical associates in a sentence, but instead lexical and sentential priming may be the product of the activation of overlapping semantic representations (Hem~indez, Fennema-Notestine, & Udell, in press; O'Seaghdha, 1989; Simpson, 1994; Simpson, Peterson, Casteel, & Burgess, 1989). This view has
The Role of Lexical and Sentential Context Effects
139
received support from studies using Event-Related Potentials (ERPs) (Van Petten & Kutas, 1991) and from behavioral studies which have found that sentences can "override" or alter the nature of lexical priming (Foss & Speer, 1983; Hem~indez, Bates, & ,~vila, 1996; Moss & Marslen-Wilson, 1993). Given the evidence above, it seems that lexical and sentential priming cannot be neatly divided into automatic vs. controlled, or lexical vs. postlexical processes. Nor is it clear that there is a moment in time in which lexical priming but not sentential priming has its effect on a target. This has led some to suggest that all of semantic priming (lexical or sentential) is postlexical. The only way to investigate "true" lexical access is by studying purely orthographic priming (Forster, 1990). However, it is not clear that lexical access is independent of contextual information even at the orthographic level (see Balota, 1989 for a review). Even if lexical and sentential contexts occur after lexical access, it is still not clear if these effects occur during different stages of processing. Consequently, the chronometry of sentential and lexical context effects and their interaction needs to be examined more carefully. We will return to this point later.
Lexical Access in Bilingual Speakers One of the fundamental questions in bilingual language processing is the nature of language representation. Does a bilingual have both sets of words in the same lexicon or does each language have a separate lexicon for each language? For many years, researchers addressed this issue by using traditional memory methodologies involving free recall or recognition (see Heredia & McLaughlin, 1992 for a review). More recently, investigations of the bilingual lexicon have shifted to the priming paradigm. Researchers have looked at whether cross-language priming is automatic and thus the product of strong intra-lexical priming. These studies have found that priming across languages may be automatic (Altarriba, 1991, 1992; Tzelgov & Eben-Ezra, 1992) or controlled (Grainger & Beauvillain, 1988) or both (Hern~indez, Bates, &/~vila, 1996; Keatley & de Gelder, 1992). Thus priming studies have not been able to clearly resolve the two-lexicon debate. One factor that has emerged as a possible explanation for the strength of withinor between-language priming is order of acquisition. Kroll and Sholl (1992) looked at the size of lexical priming from a bilingual's first language to their second language (L1 to L2) and from a bilingual's second language to their first language (L2 to L1) across a set of studies. In general, most studies suggest that priming "is larger from L1 to L2 than from L2 to L1. Kroll and colleagues (Kroll, 1994; Kroll & Sholl, 1992; Kroll & Stewart, 1994) suggest that the results from studies investigating within- and between-language lexical priming fit very well within their revised hierarchical model, a model that was initially designed to address the nature of second-language acquisition. First of all, it is posited that there are two lexical stores and a conceptual store. As can be seen in Figure 1, there are bidirectional links between the L1 lexicon and the L2 lexicon and between the conceptual store and each lexicon. Lexical links are stronger from L2 to L1 than from L1 to L2, a product of second-language acquisition. However, conceptual links are stronger to L 1 than to L2. This model has been confirmed through a series
140
Arturo E. Hern6ndez
of studies which have found that conceptual processing is stronger from L1 to L2 than from L2 to L1 while translation is faster from L2 to L1 than from L1 to L2 (i.e., translation asymmetry).
....... --,I 0 ,j,O"
CONCEPTS
Figure 1. The revised hierarchical model (adapted from Kroll & Stewart, 1994).
More recently some have presented evidence which suggests that modifications need to be made to the revised hierarchical model. For example, Heredia (1996) has found that concreteness may mediate the presence or absence of translation or naming asymmetries. Thus it is not clear that a bilingual's two lexicons can be so neatly divided, a view that has received support from de Groot and colleagues (de Groot, 1994; de Groot & Nas, 1991). In addition, he found that L1 and L2 are not equivalent to the more dominant (MDL) and less dominant language (LDL) for Spanish-English bilinguals in the United States. To account for these differences, Heredia proposed a Re-revised hierarchical model in which dominance and not order of acquisition determined the nature of language asymmetries (see also Heredia, 1997). Studies of lexical priming have also found that dominance may not be directly related to order of acquisition. For example, Altarriba (1992) found larger priming from a second language (L2) to a first language (L 1) than from L 1 to L2. Thus, studies of bilingualism should consider profoundly the age at which the second language was acquired and the particular usage of each language. In the case of Hispanics in southern Florida (and in southern California), the second language (English) ends up being the dominant language since this is the language in which most formal education occurs. Thus, studies of semantic priming in bilinguals must consider that in certain populations the second language is the dominant language. A final criticism of the revised hierarchical model is the distinction between lexical and conceptual processing of items in a second language. In fact, many have suggested that translation equivalents may be conceptually mediated both in acquisition (Altarriba & Mathis, 1997) and in usage (de Groot, Dannenburg, & Van Hell, 1994; de Groot & Hoeks, 1995). These findings suggest that dominance effects need careful consideration. While a substantial amount of work has been done with lexical priming and single-word naming, relatively little work has been done with sentence priming in
The Role of Lexical and Sentential Context Effects
141
bilinguals. Kroll and Boming (1987) looked at the effects of English sentences on lexical decisions to English and Spanish target words. Lexical decisions were slower to Spanish targets but the magnitude of priming did not vary systematically with the target language. A more recent study by Hem~indez, Bates, and Avila (1996) further explored the nature of within- and between-language sentence priming in a group of Spanish-English bilinguals. Like Kroll and Boming, Hem~indez et al. did not observe language asymmetries under normal viewing conditions. However, language asymmetries were found for targets which were visually degraded and for which the language was not predictable. In fact, these differences were larger both in the cross-language condition (English-Spanish priming > Spanish-English priming) and in the within-language condition (EnglishEnglish priming > Spanish-Spanish priming). The results from Hem~indez et al. (1996) extend findings in the single word priming literature with bilinguals by showing that sentence priming may be larger when a target is in L2 (which is the dominant language for these bilinguals). Most studies suggest that there are some differences between sentence and lexical priming. Altarriba, Kroll, Sholl, and Rayner (1996) found that under certain conditions sentence context effects can produce surprising results in bilinguals. Fixation durations and naming latency were recorded when bilinguals were reading sentences that contained high- and low-frequency Spanish and English words in high- and low-constraint sentences. Reaction times were faster and fixation durations were shorter for low-frequency targets in high-constraint sentences when compared to performance for these targets in low-constraint sentences across both languages. For high-frequency targets, however, there was an asymmetry. English high-frequency targets showed the same constraint effect observed for lowfrequency targets (faster in high-constraint sentences). Spanish high-frequency targets on the other hand showed slower reaction times and longer fixation durations in high-constraint sentences. Altarriba et al. suggest that high-frequency Spanish words in high-constraint English sentences led to slower reaction times because of violations of lexical expectations. That is, there may be some lexicallevel competition for a high-frequency Spanish word when the sentence provides a strong form of constraint for its English translation. The fact that this competition occurred even during first fixation durations suggests that this effect is a product of delays in word recognition. Studies of lexical priming in bilinguals have shown that asymmetries appear under normal reading conditions. As we mentioned earlier, the hierarchical model explains asymmetries in lexical priming by suggesting that a prime in L1 activates concepts more strongly than a prime in L2. The lack of asymmetries in normal sentence priming could be explained as the product of cumulative conceptual activation over the course of multiple words. A sentence might activate conceptual links to such an extent that differences between languages are not apparent (i.e., a ceiling effect). By the time the target is encountered there is sufficient conceptual activation even in the L2 or in the nondominant language. In this view, asymmetries appear only when sentence processing is made harder such that there is differential conceptual activation. Under normal processing conditions, sentences, even when they are in the weaker language will provide a strong form of context.
142
Arturo E. Hern6ndez
... was responsible enough to feed and care for any animal. His diligence paid off when he received a fluffy ~ for his birthday. 100 PRIME
100 PRIME
200 PRIME
0 ISI
100 ISI
0 ISI
200 PRIME 100 ISI
t
57056O 55O
dog- queen
..............,/7..
-5
/n
....
_;_
/7
.....
/ p e n /'king-
-5
/1
-5
queen I ( /
.
.
.
.
.
===
.
.
.
.
.
.
.
.
.
.
- g30 I= 520 cO
510
dog-cat
fT" 51111
w----"
490
-4
480
470
!
i!
I
I
I
i!
I
I
I
i!
I
I
I
i!
-ll-- LexicallyRelated Lexically Unrelated Figure 2. Sentential and lexical priming for Hern~indez et al. (in press).
The current study set out to investigate the effects of sentential priming and language dominance on lexical priming using an adaptation of the competitivepriming technique developed by Hem~indez, Fennema-Notestine, and Udell (in press). In the "competitive-priming" technique, participants are asked to listen to a multiple-sentence context and answer occasional questions that appear on a computer screen. During the penultimate or last sentence within each short context, participants are asked to read the second word of a word pair that appears in lieu of an auditory word in the sentence. Hem~indez et al. (in press) presented the prime (first word) for 100 milliseconds (ms) or 200 ms followed by a 0 ms or 100 ms Inter-Stimulus-Interval (ISI) between prime and target. The results which can be
The Role of Lexical and Sentential Context Effects
143
seen in Figure 2 revealed two very interesting findings: (1) Sentential priming was smaller at the longer duration than at the shorter duration; and (2) ISI interacted with lexical and sentential priming. The short ISI condition (0 ms) revealed an interactive pattem of results while the longer ISI revealed an additive pattern of results. ISI and prime duration did not interact. These results suggest that prime duration and ISI have slightly different roles. Prime duration changes the size of sentential priming while ISI dictates the pattern of lexical and sentential interaction (i.e., interactive or additive). The competitive-priming technique was adapted for use with bilinguals in the following manner. Translation equivalents that were not cognates were used in a block design with a 200 ms prime duration and a 0 ms ISI. Translation equivalents show robust priming for short SOAs (Altarriba, 1992) even when responses are speeded (Keatley & de Gelder, 1992). Using these particular timing parameters with translation equivalents will help us to investigate whether bilingual lexical and sentential context effects mirror those found in previous experiments with monolinguals. Using a blocked design will insure that within- and betweenlanguage sentential priming are both observed. In order to observe the strongest form of sentence priming, the sentences will all be high-constraint sentences. The participants will be a group of Spanish-English bilinguals who were native speakers of Spanish, learned English before the age of 8, and had received most or all of their education in English. The fact that these participants had received all of their education in English makes them English dominant, a fact that makes them different from other populations (see Altarriba et al., 1992 and Heredia, 1995 for studies with similar populations). Finally, like Hem~indez et al. (1996), naming will be used as a dependent variable. This should help to establish our effects for a dependent measure which is less susceptible to postlexical checking strategies (Balota, 1989; Balota & Lorch, 1986; Neely, 1991). In the competitive priming technique, subjects are asked to listen to a twosentence context "The other day my friend was skiing when she fell and broke her in two places. Since the accident she has refused to ski again". For the following examples, "sombrero" is the Spanish word for "hat" and "brazo" is the Spanish word for "arm." During the middle of the first or second sentence, subjects are shown a word pair that appears in one of six conditions: (1) Both words could fit the sentence (brazo-arm); (2) the target fits the sentence but the prime does not (sombrero-arm); (3) both prime and target do not fit the sentence but are translations (sombrero-hat); (4) the prime fits the sentence but the target does not (brazo-hat). In addition, two neutral controls were included (i.e., a neutral sentence); (5) where translations (brazo-arm) and (6) an unrelated word pair (brazo-hat) were embedded in a nonbiasing sentential context. Notice that in the last condition the prime is related to the sentence and the target is not. The within- and between-language conditions and the language of the prime and target will be crossed into four basic conditions: (1) English sentences, Spanish primes, and English targets, or EsE; (2) Spanish sentences, English primes and Spanish targets, or SeS; (3) English sentences, English primes, and Spanish targets, or EeS; and (4) Spanish sentences, Spanish primes and English targets, or SsE. Sample stimuli for English sentences can be seen in Table 1.
144
Arturo E. Hernfndez
Table 1. Sample Stimuli for Experiments IA and 2B (English Sentences) Sample Sentences
The other day my friend was skiing when she fell and broke her in two places ....
My cousin invented a game ... Her brother got confused w h e n she said too quickly,
Since my uncle is fromTexas ... He never leaves the house without putting on his boots and his ...
Sentential condition
Fits sentence
Neutral
Doesn't fit sentence
Translation
BRAZO-ARM
BRAZO-ARM
BRAZO-ARM
Unrelated
SOMBRERO-ARM
SOMBRERO-ARM
SOMBRERO-ARM
Translation
ARM-BRAZO
ARM-BRAZO
ARM-BRAZO
Unrelated
HAT-BRAZO
HAT-BRAZO
HAT-BRAZO
Experiment 1A EsE
Experiment 2B EeS
In using the revised hierarchical model, Kroll and colleagues (Kroll, 1994; Kroll & Sholl, 1992; Kroll & Stewart, 1994) have argued that fluency or dominance should be reflected in the magnitude of size of effects (for further discussion on dominance in early Spanish-English bilinguals see Heredia, 1997): Using the competitive priming technique gives us the opportunity to observe fluency in two ways. First, crossing the language of the sentence with language of the prime and target will help to further clarify when language asymmetries do and do not appear. The Revised Hierarchical model would .predict that priming from English to Spanish should be larger than priming from Spanish to English. For lexical priming, the question is whether there will be asymmetries and whether they will appear when the sentence and prime language match or when they do not match. In other words, do sentences enhance or reduce the size of the lexical priming asymmetry. Previous work with sentence priming shows that asymmetries appear when processing is made more difficult. In the current study, Experiment 2 is more difficult in that the target is not in the same language as the prime or sentence. Hence, sentence priming asymmetries should appear in Experiment 2. In addition, placing between-language lexical priming in competition with within- and betweenlanguage sentential priming should reveal how manipulations of timing are similar or different to manipulations of fluency. Faster processing of the prime should make
The Role of Lexical and Sentential Context Effects
145
the results appear as if the prime had been up for more time or more time had been allowed between prime and target. Slower processing of the prime should make the results appear as if the prime had been up for less time or less time had been allowed between prime and target. These bilinguals will be the prototypical bilingual that we have used in our experiments (i.e., more dominant in English than in Spanish) (Hem~indez & Kohnert, 1999; Kohnert, Bates, & Hern~indez, 1999; Kohnert, Hem~ndez, & Bates, 1998). For this group, English should be processed faster than Spanish. If English is processed faster than Spanish then when English is the prime (SeS & EeS) the results should look similar to those observed for monolinguals when the prime was presented for a longer duration or ISI. When Spanish is the prime (EsE & SsE) the results should look more like those observed for monolinguals when the shorter duration or ISI was used. In short, we should observe asymmetries in both the magnitude of lexical and sentential priming and the interaction between both forms of priming.
General Method Apparatus and Materials
The stimuli were presented on a Macintosh Quadra 660AV computer using the PsyScope experimental shell from Carnegie Mellon University (Cohen, MacWhinney, Flatt, & Provost, 1993). The auditory sentence contexts were recorded into a Sony Digital Audio Tape recorder in a soundproof booth. They were then digitized at 16-bit, 22k sampling rate using the SoundEdit 16 software package. A set of 240 noncognate translation equivalents was chosen with a mean frequency of 175 occurrences per million (SD -- 270) in English (Ku~era & Francis, 1967) and of 174 in Spanish (Juilland & Chang-Rodriguez, 1964), a mean familiarity rating of 5.95 on a scale from 1 to 7 (SD = 1.13), and a mean concreteness rating of 5.19 (SD = 1.14). For each of the word pairs, a two-sentence context was designed to bias the meaning of the translation pair (see Table 1). These sentences were designed to strongly bias the lexical item in both Spanish and English. The materials were split into two halves (an English version and a Spanish version). Each word pair had a corresponding sentence context. Each pair of translation equivalents was matched with another pair which was not related in meaning. The incongruent word pair was made by taking the prime of the first pair and replacing it with the prime of the second pair (i.e., brazo-arm/sombrero-hat sombrero-arm). Each congruent and incongruent word pair appeared in one of three sentential conditions: (1) the original sentence in which the target fits the sentence; (2) a neutral sentence which does not bias either the prime or the target; (3) an incongruent sentence in which the target does not fit the sentence. Note that for incongruent word pairs in incongruent sentences the prime but not the target fits the sentence. For both blocks, six different versions of the experiment were prepared. In every version, each sentence and each word pair was presented only once. All of the sentences were randomly assigned to a condition in the first version. For each subsequent version the word pair was presented in the next condition according to a
146
Arturo E. Hern(lndez
predetermined order. Thus each condition was rotated systematically across subjects. However, the order of presentation of sentences was random such that the subject was not able to predict in which condition the next sentence would appear.
Design and Procedure Each person participated in two separate sessions separated by at least a day. The order of test language was counterbalanced across subjects such that half heard English on their first day and half heard Spanish on their first day. The second day of testing was always at least twenty-four hours later. Each of these experiments included three sequential parts. First, participants saw 10 isolated word pairs on the screen and practiced saying the target word aloud. Second, a practice block of trials was administered in which participants were given three sentence contexts similar to those used in the subsequent experiment. The participants listened to each context and pronounced the target word when it appeared on the screen. Third, the subject participated in the 120 experimental trials. Neither the baseline nor practice items used any of the words or sentential contexts that appeared in the experimental trials. In the first practice condition, subjects were told they would be shown a row of eight X's for 750 ms, followed by 250 ms of a blank screen, followed by the prime for 200 ms which was immediately replaced by the target. The subjects were instructed to read the second word of the word pair. The target disappeared as soon as the subjects named the word. There was a 500 ISI between trials. Since the experiments were between subjects, the first practice session was necessary for participants to practice reading only the target words of the presented word pairs. After completing the first practice session, participants were told that they would be hearing a set of sentences presented auditorily over two speakers. They first saw a row of eight X's on the computer screen (750 ms) which served as a cue for the oncoming auditory context. Immediately after that cue, the text began. Participants were instructed that they should pay attention to the content of the passage because questions would be asked regarding the sentence context later. The participants were also told that, at a predetermined point during the sentence, the sound would stop and two visually presented words, one right after the other, would appear on the computer screen. The visual prime was presented immediately after the prior auditory word. Half of the visual word pairs appeared during the first sentence and half during the second sentence of the contexts. None of the targets appeared in sentence-initial or sentence-final position. The prime was presented for 200 ms and there was no ISI between prime and target. The participants were told to pronounce the target word as quickly as possible. Participants were informed that some words would fit with the contexts and others would not. A response window of 1200 ms was provided. If the participant responded before the 1200 ms had passed, the target would come off the screen and the context continued. Finally, a multiple-choice question was randomly presented at the end of some contexts (approximately one out of every five trials), to assure that participants attended to the meaning of the sentences. The participants were permitted as much time as needed to answer the question. Once the question was answered the next context would begin. For Experiments 1 and 2 the sentence contexts, prime and target words, design and procedure were identical. The only difference was that in Experiment 1 the
The Role of Lexical and Sentential Context Effects
147
sentence and the target were in the same language. In Experiment 2, the sentence and the target were in the opposite language. Each experiment was further subdivided by the language of the target. Thus, Experiment l a had English sentences, Spanish primes, and English targets and will be called the English within-language condition or EsE; Experiment l b had Spanish sentences, English primes, and Spanish targets and will be denoted the Spanish within-language condition or SeS; Experiment 2a had Spanish sentences, Spanish primes, and English targets and will be denoted the English between-language condition or SsE; Experiment 2b had English sentenCes, English primes, and Spanish targets and will be called the Spanish between-language condition or EeS. Finally, note that all related word pairs were translation equivalents. The terms translation priming and lexical priming will be used interchangeably in the results and discussion sections. Statistical analyses. For each experiment a 2 (related vs. unrelated lexically) x 3 (fits, doesn't fit, and neutral sentence) ANOVA will be conducted. In addition, a modified Bonferroni procedure will be used for planned comparisons in each experiment. Planned comparisons will test the significance of lexical priming for targets that fit the sentence, do not fit the sentence and for targets that are in a neutral context.
Experiment 1" Cross-Modal Lexical and Sentential Priming in the WithinLanguage Condition Participants This group consisted of eight men and 16 women (N = 24) with a mean age of 19.63 (SD = 1.79). Fourteen of the bilinguals tested considered themselves native Spanish speakers, whereas only three considered themselves to be native English speakers. Of those remaining, five considered both English and Spanish to be their native language and two did not provide this information. The mean number of years spent speaking Spanish was 19.09 as compared to 17.34 years spent speaking English. Furthermore, self-ratings revealed that subjects thought their usage was higher in English (11.50 in English vs. 13.52 in Spanish on a scale from 0 to 21; with lower numbers demonstrating more usage), and that their skill was higher in English (26.16 in English vs. 23.47 in Spanish on a scale from 0 to 28; with higher numbers demonstrating more skill). These characteristics suggest that these subjects were English-dominant. This is a characteristic profile for young Spanish-English bilinguals that have been tested before in our laboratory.
Results and Discussion Analyses over subjects (FI) and analyses All errors and all reaction times below 200 from the analyses (5.27% and 9.58 % of respectively). Error rates were below 3% and
over items (F2) will both be reported. ms or above 1000 ms were excluded all trials in Experiments l a and l b, were not subjected to further analyses.
148
Arturo E. Hern(mdez
In addition, error rates on the comprehension questions were very low (below 4%) and were not subjected to further analyses. A 2 (lexical priming) x 3 (sentence priming) ANOVA was performed for each language to look for lexical priming and sentential priming main effects as well as any interaction between them. In addition, for each experiment, a set of planned comparisons was used to test congruent and incongruent word pairs for facilitation and inhibition relative to their neutral control and to test for lexical priming for targets that fit the sentence, were in a neutral sentence, or did not fit the sentence. Finally, a 2 (Spanish/English) x 2 (related vs. unrelated lexically) x 3 (fits, doesn't fit, and neutral sentence) ANOVA across languages was used to compare the performance of subjects across languages.
Experiment la: Analyses of Reaction Times to English Within-Language Targets With Spanish Primes (EsE) The 2 (lexical priming) x 3 (sentential priming) ANOVA presented a significant main effect of sentential priming (Fj(2,46) = 57.31, p < .001, MSE = 332.30; /=2(2,238) = 43.583, p < .001, MSE = 2,645.52), lexical priming (F1(1,23) = 26.78, p < .007, MSE = 414.43; F2(1,119) = 6.16, p < .014, MSE = 5,397.60) and sentential by lexical interaction (F~(2,46) = 7.94, p < .001, MSE = 178.34; F2(2, 238) = 3.45, p < .033, MSE - 2,624.11). The results from a set of planned comparisons revealed significant translation priming for targets that did not fit the sentence (F~(1,23) = 98.10, p < .001, MSE = 178.34; F2(1,119) = 74.64, p < .001, MSE = 2624.11) and for targets in a neutral sentence (F~(1,23) - 48.58, p < .001, MSE = 178.34; F2(1,119) = 30.18, p < .001, MSE = 2624.11), but not for targets that fit the sentence (F/and F2 < 1). To summarize, Experiment 1a revealed significant sentential and lexical priming and a significant interaction. This pattern is very similar to the one observed by Hem~ndez et al. (in press) in monolinguals for semantic priming in the same temporal conditions (200 prime duration/0 ISI). Namely, there was an interaction between sentential and lexical priming, such that lexical priming did not appear for targets that do fit the sentence while lexical priming did appear in both the neutral sentences and in sentences where the target did not fit the sentence. The effects of lexical priming suggest that a Spanish interruption (i.e., the prime) can be analyzed quickly for semantic content and that this content can have an effect on word recognition but not when the target fits the sentence. However, note that the pattern is similar to the one observed for monolinguals (see Figure 2) when the prime was presented for shorter times and the ISI was shorter. Also note that sentential priming was robust across all lexical conditions.
The Role of Lexical and Sentential Context Effects
149
Experiment lb- Analyses of Reaction Times to Spanish Within-Language Targets With English Primes (SeS) The 2 (lexical priming) x 3 (sentential priming) A N O V A presented a significant main effect o f sentential priming (F1(2,46) - 85.73, p < .001, MSE = 823.08; F2(2,238) = 33.40, p < .001, MSE = 5266.80) and lexical priming (F1(1,23)= 17.42, p < .001, MSE = 823.09; F2(1,292) = 9.98, p < .002, MSE = 7306.24), but no significant interaction (FI and F2 < 1). The planned comparisons revealed lexical priming for targets that fit the sentence (F1(1,23) = 19.45, p < .001, MSE = 771.13; F2(1,119) = 4.04, p < .05, MSE - 5079.73), targets in neutral sentences (F1(1,23) = 21.04, p < .001, MSE = 771.13; F2(1,119) = 8.49, p < .01, MSE = 5079.73), but not for targets that did not fit the sentence.
Table 2. Mean Reaction Time (RT) and Percent (%) Errors in Experiment 1 Lexically related
Lexically unrelated
Lexical priming
Fits sentence
577 (.65 %)
582 (1.08 %)
5 n.s
Neutral
596 (.21%)
611 (.67 %)
16**
Doesn't fit sentence Sentence Priming
610 (1.25 %)
637 (1.48 %)
27**
33**
55**
Fits sentence
577 (2.36 %)
601 (2.58 %)
24 **
Neutral
604 (2.91%)
630 (3.84 %)
26**
Doesn't fit sentence Sentence Priming
640 (1.73 %)
649 (3.16 %)
9 ns
63**
48**
Experiment I a English withinlanguage (EsE)
Experiment I b Spanish withinlanguage (SeS)
** p < .05, n.s not significant at p < .05
150
Arturo E. Hern6ndez
Whereas the results from the ANOVA were consistent with an additive pattern of results, the results from the planned comparisons were not consistent with any of the previous findings. There was significant sentential priming (55 ms) and lexical priming (20 ms) with no interaction, consistent with an additive effect like the one observed in the 200/100 ms condition with monolinguals (e.g., see Figure 2). As predicted earlier, English primes should reveal pattems of results that were observed with longer duration or ISI's. Lexical priming appeared for targets that fit the sentence and for targets in the neutral condition. Surprisingly, however, there was no priming for targets that did not fit the sentence. Not only was there lexical priming for congruent targets but there was no lexical priming for incongruent targets. This pattern has not been observed in any of the monolingual competitivepriming experiments. Thus, it appears that under certain conditions bilinguals show a decrease in the size of lexical priming for words that cannot be integrated into the sentence. In other words, there may be some difficulty in viewing the current results as an exact extension of monolingual results. Some caution must be taken in interpreting these results since there was no interaction between lexical and sentential priming in the ANOVA. As can be seen in Table 2, Experiments 1a and lb suggest that within-language sentential priming is robust but cross-language lexical (translation) priming is variable. The interaction between lexical and sentential priming, however, had a different direction in each study. In the SeS condition, there was no lexical priming for targets that did not fit the sentence and in the EsE condition there was no lexical priming for targets that did fit the sentence. To further investigate these effects a three-way 2 (Spanish/English) x 2 (related vs. unrelated lexically) x 2 (fits, doesn't fit, and neutral sentence) mixed ANOVA was conducted in order to compare the relative magnitude of lexical and sentential priming across experiments. This revealed a main effect of sentential priming (F1(2,46) = 125.45, p < .001, MSE = 477.42; F2 (2,476) = 72.74, p < .001, MSE = 3956.16) of lexical priming (F1(1,23) = 35.96, p < .001, MSE = 637.42; F2 (2,476) = 16.10, p < .001, MSE = 6351.91) and a three-way interaction of language by sentential priming by lexical priming reliable across subjects only (F1(2,46) = 5.49, p < .001, MSE = 395.87; F2 (2,476) =2.13, p < 0.12, MSE - 3851.92). This three-way interaction can be seen in Figure 3. The results from Experiment 1 yielded a number of findings. First of all, there was a main effect of sentential and lexical priming. Thus, both the prime and the sentence speed the processing of the target. Second, there was no difference in the magnitude of lexical or sentential priming across languages, and no main effect of language. Much like in Kroll and Boming (1987) and Hern~indez et al. (1996), the blocked design under normal reading conditions does not lead to significant language effects in the within-language condition (i.e., SeS = EsE). In addition, there was no significant interaction between language and lexical priming. This suggests that in a sentential context when naming is used as a measure of lexical access, lexical priming does not show significant asymmetries when collapsed across experimental conditions, a critical difference with studies that have investigated lexical priming in isolation using lexical decision. Finally, there was a significant three-way interaction, consistent with the notion that an interactive or
The Role of Lexical and Sentential Context Effects
151
additive pattem of lexical and sentential priming depends on the language. This is consistent with the hypothesis that English is processed faster than Spanish. Hence, a Spanish prime reveals an interactive pattem of results whereas an English prime leads to a more additive pattern of results. Although there is no interaction in the SeS condition alone, planned comparisons revealed no lexical priming for targets that do not fit the sentence. Thus the asymmetries depend on the sentential condition of the word pair. When the target fits the sentence, priming is larger from English to Spanish than from Spanish to English, an asymmetry that would be expected given the English-dominant nature of the participants. In the neutral condition, lexical priming always appears. When the target does not fit the sentence, the asymmetry is reversed. This suggests that sentence contexts can change the nature of asymmetries, a point we will return to later. broke h e r
se rompio e l
Spanish text I English prime Spanish target (SeS)
English text/Spanish prime and English Target (EsE)
720 -
-720 -700
700
-680
680 arm-sombrero
660
brazo-hat
-660 -640
.=- 640 ID
E i-7- 620
~
600
~r
sombrero
,/,1"'~" ~ -~41
sombrero-
a rm
580 560
.//'~ ......"J"~
hat-brazo
c
.o_
./"~ ~
-620 -600
s o m b r e ro-
hat
-580
arm-brazo
brazo-arm
-560
540
-540
520
I
Fits Sen tence
--I--
I
neutral
I
Doesn't Fit Sen tence
I
I
Fits Sen tence
520
I
neutral
Doesn't Fit Sen tence
Translation
......O ..... Unrelated
Figure 3. Sentential and lexical priming for Experiment 1 (within-language
condition).
152
Arturo E. Hern6ndez
In order to further explore how lexical priming, sentential priming, and language interact, a between-language target condition was employed in Experiment 2. In this experiment, the text and the prime now belong to the same language and the prime and target continue to be from the opposite language. In this condition, processing should be more difficult given that participants are asked to name a target that is in the opposite language of both the prime and the sentence. The revised hierarchical model would predict that priming should be larger from English (L2) to Spanish (L1). This should appear at the lexical or sentential level or both. Finally, this experiment will reveal if indeed the presence of a Spanish prime leads to a more interactive pattern of results and the presence of an English prime leads to a more additive pattern of results.
Experiment 2: Cross-Modal Lexical and Sentential Priming in the BetweenLanguage Condition Participants This group consisted of 6 men and 18 women (N = 24) with a mean age of 18.88 (SD = 2.13). Fourteen of the bilinguals tested considered themselves native Spanish speakers, whereas only 2 considered themselves to be native English speakers. The remaining 8 considered both English and Spanish to be their native language. The mean number of years spent speaking Spanish was 18.88 as compared to 15.25 years spent speaking English. Furthermore, self-ratings revealed that subjects thought their usage was higher in English (10.71 in English vs. 13.29 in Spanish on a scale from 0 to 21; with lower numbers demonstrating more usage), and that their skill was higher in English (25.17 in English vs. 23.58 in Spanish on a scale from 0 to 28; with higher numbers demonstrating more skill). Note that their usage and skill suggests that English is their dominant language. This is a characteristic profile for Spanish-English bilinguals in Southern California.
Results and Discussion Again, analyses over subjects (FI) and analyses over items will be reported (F2). All reaction times below 200 ms or above 1000 ms were excluded from the analyses (5.63% and 8.61% of all trials for Experiments 2a and 2b respectively). Error rates were below 2% and were not subjected to further analyses. In addition, error rates on the comprehension questions were very low (below 4%) and were not subjected to further analyses. A 2 (related vs. unrelated) x 3 (fits, doesn't fit, and neutral sentence) ANOVA was performed for each language to measure lexical priming and sentential priming main effects as well as any interaction between them. In addition, for each experiment, a set of planned comparisons was used to test congruent and incongruent word pairs for facilitation and inhibition relative to their neutral control. Finally, a three-way ANOVA across languages was used to compare the performance of subjects across languages.
The Role of Lexical and Sentential Context Effects
153
Experiment 2a: Analyses of Reaction Times to English Cross-Language Targets With Spanish Primes (SsE) The 2 (lexical priming) x 3 (sentential priming) ANOVA presented a significant main effect of sentential priming (F~(2,46) = 23.02, p < .001, M S E = 443.29; /=:(2,238) = 30.57, p < .001, M S E = 1884.47), lexical priming (F1(1,23) = 9.27, p < .001, M S E = 271.01; F2(1,292) = 3.73, p < .056, M S E = 4524.49) and no interaction. Planned comparisons revealed lexical priming for words that were in a neutral sentence only (F~(1,23) = 15.47, p < .001, M S E = 397.08; F2(1,292) = 9.71, p < .001, M S E = 2231.69). To summarize, Experiment 2a revealed significant cross-language sentential and lexical priming. Sentential priming was about 30 ms (considerably smaller than that observed in Experiment 1). In addition, lexical priming only appeared reliably in the neutral sentence. Like Experiment 1a (EsE), priming was not present for targets that fit the sentence. However, priming was also not present for targets that did not fit the sentence. The fact that the prime was in the same language as the sentence led to a reduction in the size of translation priming. Thus, the pattern of results was neither additive nor interactive. Rather, there was very little lexical priming. The results from this experiment are reminiscent of those observed with monolinguals in the 100 prime duration/100 ISI condition (see Figure 2). In Experiment 2b, the same between-language condition will be used. However, the prime and sentence will be in EngliSh and the target in Spanish. The two questions being asked is whether there will be an increase in the size of lexical or sentential priming or both from English to Spanish. Furthermore, the results should reveal a more additive pattern of results as was observed when longer prime durations and ISI's were used.
Experiment 2b: Analysis of Reaction Times to Spanish Cross-Language Targets With English Primes (EeS) The 2 (related vs. unrelated lexically) x 3 (fits, doesn't fit, and neutral sentence) ANOVA presented a significant main effect of sentential priming (F1(2,46) = 19.56, p < .001, M S E = 1618.51; F2(2,238) = 20.11, p < .001, M S E = 4525.75) and lexical priming (F1(1,23) = 6.69, p < .001, M S E = 2429.85;/=:(2,238) = 10.96, p < .001, M S E = 4930.58) and no interaction between them. The planned comparisons revealed lexical priming for targets that fit the sentence (F~(1,23) = 7.77, p < .02, M S E = 893.22; /=:(1,119) = 4.74, p < .05, MSE = 5079.73), were in neutral sentences (F~(1,23) = 9.70, p< .02, M S E = 893.22; F2(1,119) = 8.61, p < .005, MSE = 5079.73) and did not fit the sentence (F~(1,23) = 18.16, p < .001, MSE = 893.22; 1=2(1, 119) = 3.83,p < .05, M S E = 5079.73). To summarize, Experiment 2b revealed significant sentential and lexical priming. The effects are very similar to those observed in Experiment l b (SeS) in that there was no interaction between lexical and sentential priming. Lexical priming effects were about 21 ms in magnitude and the sentential effects were about 51 ms in magnitude. Unlike Experiment l b, however, planned comparisons in Experiment 2b revealed lexical priming in all three sentential conditions. This
154
Arturo E. Hern6ndez
suggests that, very early in processing, English primes are processed very quickly and hence reveal a pattern of results that is equivalent to that observed with longer durations and ISI's in the monolingual experiments. The results from Experiment 2 can be seen in Table 3.
Table 3. Mean Reaction Time and Error Rates in Experiment 2 Lexically related
Lexically unrelated
Lexical priming
Fits sentence
587 (2.81%)
587 (2.12 %)
0 n.s
Neutral
597 (2.52 %)
613 (3.22 %)
16**
Doesn't fit sentence Sentence Priming
611 (2.74 %)
619 (4.43 %)
8 n.s
24**
32**
Fits sentence
620 (3.99 %)
637 (6.97 %)
17 **
Neutral
643 (6.42 %)
664 (6.93 %)
21'*
Doesn't fit sentence Sentence Priming
667 (3.41%)
692 (4.61%)
25
47**
55**
Experiment 2a English betweenlanguage (SsE)
Experiment 2b Spanish betweenlanguage (EeS)
** p < .05, n.s. not significant at p < .05
A three-way 2 (Spanish~nglish) x 2 (related vs. unrelated lexically) x 3 (fits, doesn't fit, and neutral sentence) mixed ANOVA was conducted in order to compare the relative magnitude of priming across experiments. This revealed a main effect of language (Fi(1,23) =17.42, p < .001, M S E = 10881.63; Fe(1,238) = 40.81, p < .001, M S E = 14354.66), sentential priming (F1(2,46) = 27.98, p < .001, M S E = 1381.12; F2(2,476) = 42.56, p < .001, M S E = 3365.66), and lexical priming (F1(1,23) = 10.26, p < .001, M S E = 1739.96; F2(1,238) = 14.13, p < .001, M S E = 4889.67) and an interaction between language and sentential priming (F1(2,46) =
The Role of Lexical and Sentential Context Effects
155
4.74, p < .013, MSE = 680.67; F2(2,476) = 1.12, p < 0.327, MSE = 3365.66). Both lexical and sentential priming were present overall in Experiment 2. These effects can be observed in Figure 4.
se rompioe l
broke h e r
Spanish text/prime English target (SsE)
English text/prime and Spanish Target (EeS)
720-
- 720
arm- sombre ro
700-
.,.,/.,.,.//.,.,0
680660=-9 640-
brazo-hat
E r
620-
._o "6 6 0 0 580-
ha t_ b ra zo . / . . / ' " ~ y sombrero
.................. 9
"
~
680 660 640 620
sombrero-j//.F .~ .....-.~ I I
arm i
- 700
arm-brazo
sombrerohat
600
580
brazo-arm
560-
560
540-
540
520
I
Fits Sen tence
--II-
I
neutral
I
Doesn't Fit Sen tence
I
I
Fits Sen ten ce
520
I
neutral
Doesn't Fit Sen ten ce
Translation
- - O ..... Unrelated
Figure 4. Sentential and lexical priming for Experiment 2 (between-language condition).
Of particular interest were the main effect of language and the interaction between language and sentential priming. Recall that in Experiment 1 there were no differences in the magnitude of priming or in the reaction times that were dependent on language. In Experiment 2, however, there were slower reaction times when
156
Arturo E. Hernfmdez
reading Spanish targets and more sentential priming with Spanish targets than English targets. These findings are consistent with the speed-of-processing account and with previous findings by Hern~ndez et al. (1996) in which there was larger priming when subjects were reading Spanish in the cross-language condition, under visually degraded conditions, when the target language was predictable. This experiment confirms that language asymmetries appear in sentence priming, but only when processing is made more difficult (i.e., the prime and text help each other but compete with the target). Although there were no interactions with lexical priming, planned comparisons revealed more significant lexical priming when the prime was in English and the target was in Spanish but not vice-versa. Thus, in this experiment there were language asymmetries at the sentential level and to a much weaker extent at the lexical level.
General Discussion
The current set of results provided support for both the revised hierarchical model and Heredia's (1996) Re-revised hierarchical model, and were consistent with the notion that speed of processing the prime may be correlated with fluency.
Lexical Priming
First, priming was not larger from English to Spanish than from Spanish to English for lexical priming in the main analyses for Experiments 1 and 2. Second, Experiment 1 revealed additional findings of interest with regard to priming asymmetries. For targets that fit the sentence, the asymmetry was in the expected direction. Lexical priming was larger from English to Spanish (stronger to weaker language) than from Spanish to English (weaker to stronger language). For targets that did not fit the sentence, however, the asymmetry was in the opposite direction. Lexical priming was larger from Spanish to English than from English to Spanish. Hence, it appears the planned comparisons revealed conditions in which lexical priming asymmetries could appear in the expected direction or the opposite direction depending on whether the target fit the sentence or not. Thus, in the same experiment it was found that asymmetries can appear in both directions depending on the language of the prime, the language of the target, and whether the target does or does not fit the sentence. This suggests that there may be a more complex relationship in which language may serve to increase or decrease lexical priming. For Experiment 2, planned comparisons revealed lexical priming in the neutral condition only in the SsE condition and in all three conditions in the EeS condition. However, there was no interaction between lexical priming and any of the other variables. Thus, the strongest effects of lexical priming were seen in Experiment 1.
The Role of Lexical and Sentential Context Effects
157
Sentence Priming Sentence priming was present in every experiment and across experiments as well. According to previous experiments by Hern~ndez et al. (1996) priming should be larger from Spanish sentences to English targets than from English sentences to Spanish targets. This was found in Experiment 2. In Experiment 1, there was no difference in the size of sentence priming depending on the language of sentence. Hence, the results were compatible with previous results and were also compatible with a modified version of the revised hierarchical model in which sentence priming asymmetries can appear under certain conditions.
Lexical Priming by Sentential Priming Interaction In the Introduction, it was predicted that English primes would be processed faster than Spanish primes. Hence, conditions in which the prime was in English would appear to be more additive whereas conditions in which the prime was in Spanish would appear to be more interactive. This appeared in Experiment 1 as a three-way language by lexical priming by sentential priming interaction in the predicted direction. Although there was no interaction of this type in Experiment 2, the pattern of results was also in the same direction. In fact, the lack of lexical priming appeared in monolinguals when the prime was presented for 100 ms and there was a 100 ms ISI. Thus, the fluency of language appears to be reflected in the speed with which information can be processed. In summary, it appears the presence or absence of lexical priming not only depends on the sentential condition of the target but also on the language of the text, the prime, and the target. Unlike studies of lexical priming in isolation with Spanish-English bilinguals (Altarriba, 1992) and by others in the literature (for a review see Kroll, 1993) there were differences in the size of lexical priming but these differences varied in an interactive fashion. The inability of our models to predict these findings suggests that there may be some new experimental and theoretical issues that need to be considered. One way in which current models could be improved is by incorporating speed as a dimension. That is, fluency may in fact result in a stronger semantic effect in one language. The way this has been traditionally seen is as an effect of links between the lexicon and the conceptual store. The stronger the links, the larger the priming effect seems to be. Another way in which this could be considered woul'd be in terms of speed. In other words, the faster an item is processed the faster conceptual information is activated and the larger the lexical priming effects. The current results, while reminiscent of a model that emphasizes speed, are not completely supportive of it. Furthermore, it reinforces the idea that monolingual models do not always account for bilingual language processing. Thus, it is important to consider that bilinguals are not two monolinguals in one head (Grosjean, 1989, 1992). A number of changes needs to be made to current models of bilingual language processing in order to account for the current results. First of all, models of bilingual language processing need to begin to consider that increasing the magnitude of priming is not always the optimal solution to a
158
Arturo E. Hern6ndez
problem. In the Introduction, it was predicted that there should be sentence priming asymmetry in Experiment 2 due to the increased processing load in this condition. The interesting part of this effect was that there was an asymmetry. However, the asymmetry was due to a reduction in the size of lexical priming in the SsE condition. This is reminiscent of a result that was previously observed by Hern~indez et al. (1996). In their mixed design, there was a reduction in the size of sentential priming across languages. Furthermore, there was also smaller sentential priming from Spanish to English when the targets were visually degraded. In order to account for these results, it was proposed that participants surface-read the words in order to "compensate" for the increased difficulty that is encountered for crosslanguage sentential targets. Interestingly, degrading the visual stimuli yielded an increase in every condition except for Spanish sentences with English targets. Apparently, degrading English words did not lead to the increase in cross-language sentential priming. In the current study, a similar reduction in priming was observed in the Ss E condition. Thus, in two different experiments bilinguals have shown a reduction in priming when processing of the sentence is difficult but processing of the target is easy. We suggest that these results may have implications for the bilingual lexical priming literature. So far, the presence of asymmetries has been explained as the product of the speed of processing or conceptual activation of the prime. What about the speed of processing of the target? That is, if the target is processed more slowly it could lead to more time for top-down conceptual information from the prime to influence the initial stages of word recognition. If the target is processed more quickly, there is less time for the top-down conceptual information from the prime to influence processing of the target. In other words, slower target recognition leads to more time for conceptual activation to accumulate from the prime. Larger priming from English to Spanish is due both to the strength of conceptual activation of the English prime and the weakness of the lexical signal of the Spanish target. Smaller priming from Spanish to English is due both to the weakness of conceptual activation from the Spanish prime and to the strength of the lexical signal of the English target. In this view, lexical priming is the product Of processing from both the prime and the target. In effect, participants may be using the context to compensate for bottom-up slowing (Perfetti, 1994; Stanovich & West, 1981; Stanovich & West, 1983; Stanovich, West, & Feeman, 1981), an effect which has been observed with other populations.
Future Directions
The current results highlight the importance of further understanding the nature of asymmetries and the factors that might influence lexical priming in isolation. This includes degradation, timing, and the use of other populations. Using degraded primes or degraded targets, or both might begin to clarify whether it is the strength of the prime or the weakness of the target, or both, that leads to larger lexical priming from the dominant to the subordinate language. Furthermore, it might begin to clarify whether surface reading also occurs in lexical priming. Looking at how
The Role of Lexical and Sentential Context Effects
159
degradation affects lexical priming under different timing constraints might also help to shed light on these issues. Finally, a few variations of the bilingual competitive priming technique could further elucidate how lexical and sentential contexts constrain bilingual word recognition. One variation would include comparisons of monolingual and bilingual processing varying ISI and prime duration. This might uncover if bilinguals show slower activation or faster decay of lexical priming in a sentential context in either of their languages. Additional manipulations like those suggested with lexical priming in isolation (i.e., changes in timing and degradation) could help to further clarify how language, sentential, and lexical information combine in real time in bilinguals. In short, there is room for a lot of experiments which explore how bilinguals integrate different informational types. Aside from contributions that new experiments would provide, it is also important to consider how to address the theoretical issues that the present study raises. The current results show that under certain conditions one can observe a reversal in the direction of asymmetries, when looking at three factors in combination. This is a variation of the three-body problem. Poincar6 concluded that one could easily calculate (mathematically) the influence of the sun and the moon on each other, the sun and the earth on each other, and the earth and the moon on each other. As soon as he attempted to calculate the effects of all three on each other it became mathematically impossible because of the nonlinear nature of the problem. The only way to do it is through simulations. Similarly, the current experiment manipulated three factors: target language, lexical priming, and sentential priming. The interaction between target language, sentential priming and lexical priming was only significant across subjects. It is important to note, however, that different types of context can interact in nonlinear ways. This clearly begs for simulations that can handle the nonlinear nature of these data. Because of their ability to handle nonlinear interactions, connectionist models provide the most natural method of exploring bilingual language processing. Connectionist models are able to handle a number of lexical-semantic phenomena, including dyslexia (Hinton & Shallice, 1991; Plaut & Shallice, 1993; Seidenberg, 1993), speech errors (Martin, Dell, Saffran, & Schwartz, 1994), semantic priming (Masson, 1991, 1995; Moss, Hare, Day, & Tyler, 1994), and lexical ambiguity (Kawamoto, 1993) among others. Furthermore, connectionist models also have the ability to self-organize, segregating informational types in a manner that optimizes the solution to a problem. Lastly, connectionist models use distributed representations which allows one to understand the nature of graded activation and graded representation. This is especially crucial in bilingualism where, as can be seen in the present study, there appears to be graded activation of different informational types. Where does this leave bilingual research in the interim? The revised hierarchical model has been and will continue to be one of the most influential models of bilingual semantic representation and language processing. Whereas the current results have not supported it completely, it is important to note that language asymmetries do exist. The revised hierarchical model does capture the general spirit of language asymmetries that was present in the current study. Future studies and models need to further elaborate the role of timing in bilingual language processing.
160
Arturo E. Hern6ndez
Note also that we observed asymmetries using naming. This suggests that our results are less likely to be the product of postlexical checking strategies. Finally, our results support the notion that the age of acquisition may be less crucial for the Spanish-English bilinguals that were tested. In fact, Heredia (1997) has suggested that age of acquisition may be a less important factor in bilinguals who have learned their languages at a very early age. It would be impossible to end without acknowledging the controversy surrounding the role of sentential and lexical contexts in lexical access. The present study suggests that sentential priming in bilinguals (as in monolingual populations) can alter the magnitude of lexical priming. Whether this is a product of processing during or after word recognition is an open question. The present study cannot solve this debate in a definitive manner. This study, however, does suggest that bilingual language processing involves a set of complex processing tradeoffs that are necessary in order to navigate through a number of potentially complex processing environments. These tradeoffs clearly depend on differential weighting of information which in turn depends on the environmental conditions. In other words, bilinguals may be using whatever information is available to them in order to arrive at a unique solution (see Hern~indez, Bates, & Avila, 1994 for discussion in a different domain). The nature of these differential weightings is one of the most interesting and intriguing areas of future research in bilinguals. It is perhaps in understanding how and when different informational sources interact that we can begin to get an accurate picture of bilingual language processing and finally start to understand how different sources of information interact in real time.
References
Altarriba, J. (1991). Is there semantic priming across a bilingual's two languages. Unpublished Manuscript, University of Massachusetts at Amherst. Altarriba, J. (1992). The representation of translation equivalents in bilingual memory. In R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 157-174). Amsterdam: North-Holland. Altarriba, J., Kroll, J. F., Sholl, A., & Rayner, K. (1996). The influence of lexical and conceptual constraints on reading mixed-language sentences: Evidence from eye fixation and naming times. Memory & Cognition, 24, 477-492. Altarriba, J., & Mathis, K. M. (1997). Conceptual and lexical development in second language acquisition. Journal of Memory and Language, 36, 550-568. Balota, D. A. (1989). The role of meaning in word recognition. In D. A. Balota, G. B. Flores d'Arcais, & K. Rayner (Eds.), Comprehension processes in reading (pp. 9-32). Hillsdale, NJ: Erlbaum. Balota, D. A., & Lorch, R. F. (1986). Depth of spreading activation: Mediated priming effects in pronunciation but not in lexical decision. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 336-345. Cohen, J., MacWhinney, B., Flatt, M., & Provost, J. (1993). PsyScope: A new graphic interactive environment for designing psychology experiments. Behavioral Research Methods, Instruments, & Computers, 25, 257-271.
The Role of Lexical and Sentential Context Effects
161
de Groot, A. M. B. (1993). Word-type effects in bilingual processing tasks: Support for a mixed-representational system. In R. Schreuder & B. Weltens (Eds.), The bilingual lexicon (pp. 27-51). Amsterdam: Benjamins. de Groot, A. M. B., Dannenburg, L., & Van Hell, J. G. (1994). Forward and backward word translation. Journal of Memory and Language, 33, 600-629. de Groot, A. M. B., & Hoeks, J. C. J. (1995). The development of bilingual memory: Evidence from word translation by trilinguals. Language Learning, 45, 683-724. de Groot, A. M. B., & Nas, G. L. (1991). Lexical representation of cognates and noncognates in compound bilinguals. Journal of Memory and Language, 30, 90123. Forster, K. I. (1990). Lexical processing. In D. N. Osherson & H. Lasnik (Eds.), Language: An invitation to cognitive science (Vol. 1, pp. 95-131). Cambridge, MA: MIT Press. Foss, D., & Speer, S. R. (1983). Global and local context effects in sentence processing. In R. H. D. Palermo (Ed.), Cognition and the symbolic processes: Applied ecologicalperspectives (pp. 115-139). Hillsdale, NJ: Erlbaum. Grosjean, F. (1989). Neurolinguists, beware! The bilingual is not t w o monolinguals in one person. Brain and Language, 36, 3-15. Grosjean, F. (1992). Another view of bilingualism. In R. Harris (Ed.), Cognitive processing in bilinguals (pp. 51-62). New York: Elsevier. Heredia, R. R. (1995). Concreteness effects in high frequency words: A test of the revised hierarchical and the mixed models of bilingual memory representations. Unpublished doctoral dissertation, University of California, Santa Cruz. Heredia, R. R. (1996). Bilingual memory: A re-revised version of the hierarchical model of bilingual memory. CRL Newsletter, 10, 1-3. Heredia, R. R. (1997). Bilingual memory and hierarchical models: A case for language dominance. Current Directions in Psychological Science, 6, 34-39. Heredia, R., & McLaughlin, B. (1992). Bilingual memory revisited. In R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 91-103). Amsterdam: NorthHolland. Hern~ndez, A. E., Bates, E., & Avila, L. X. (1994). Sentence interpretation in Spanish-English bilinguals: What does it mean to be in-between. Applied Psycholinguistics, 15, 417-466. Hem~ndez, A. E., Bates, E., & Avila, L. X. (1996). Processing across the language boundary: A cross modal priming study of Spanish-English bilinguals. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 846864. Hern~ndez, A. E., Fennema-Notestine, A. C., & Udell, C. (in press.). Selective or exhaustive access with unambiguous words?" A chronometric study of lexical and sentential context effects. Applied Psycholinguistics. Hern~ndez, A. E., & Kohnert, K. (1999). Aging and language switching in bilinguals. Aging, Neuropsychology and Cognition, 6, 69-83. Hinton, G. E., McClelland, J. L., & Rumelhart, D. E. (1986). Distributed representations. In D. E. Rumelhart & J. L. McClelland (Eds.), Parallel distributed processing (pp. 77-109). Cambridge, MA: MIT Press.
162
Arturo E. Hern6ndez
Hinton, G. E., & Shallice, T. (1991). Lesioning an attractor network: Investigations of acquired dyslexia. Psychological Review, 98, 74-95. Juilland, A., & Chang-Rodriguez, E. (1964). Frequency dictionary of Spanish words. The Netherlands: Mouton & Co. Kawamoto, A. H. (1993). Nonlinear dynamics in the resolution of lexical ambiguity: A parallel distributed processing account. Journal of Memory and Language, 32, 474-516. Keatley, C., & de Gelder, B. (1992). The bilingual primed lexical decision task: Cross-language priming disappears with speeded responses. European Journal of Cognitive Psychology, 4, 273-292. Kohnert, K., Bates, E., & Hem~ndez, A. E. (1999). Balancing bilinguals: Lexical-semantic production and cognititve processing in children learning Spanish and English. Journal of Speech, Language, and Hearing Research, 42, 1400-1413. Kohnert, K. J., Hem~indez, A. E., & Bates, E. (1998). Bilingual performance on the Boston naming test: Preliminary norms in Spanish & English. Brain & Language, 65, 422-440. Kroll, J. F. (1993). Accessing conceptual representations for words in a second language. In R. Schreuder & B. Weltens (Eds.), The bilingual lexicon (pp. 54-81). Amsterdam: Benjamins. Kroll, J. F., & Boming, L. (1987, November). Shifting language representations in novice bilinguals: Evidence from sentence priming. Paper presented at the 27 th Annual Meeting of the Psychonomic Society, Seattle, WA.. Kroll, J. F., & Sholl, A. (1992). Lexical and conceptual memory in fluent and nonfluent bilinguals. In R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 191-204). Amsterdam: Elsevier. Kroll, J. F., & Stewart, E. (1994). Category inference in translation and picture naming: Evidence for asymmetric connections between bilingual memory representations. Journal of Memory and Language, 33, 149-174. Ku~era, H., & Francis, W. N. (1967). Computational analysis of present-day American English. Providence, R.I.: Brown University Press. Martin, N., Dell, G. S., Saffran, E. M., & Schwartz, M. F. (1994). Origins of paraphasias in deep dysphasia: Testing the consequences of a decay impairment to an interactive spreading activation model of lexical retrieval. Brain and Language, 47, 609-660. Masson, M. J. (1991). A distributed memory model of context effects in word identification. In D. Besner & G. W. Humphreys (Eds.), Basic processes in reading: Visual word recognition (pp. 233-263). Hillsdale, NJ: Erlbaum. Masson, M. J. (1995). A distributed memory model of semantic priming. Journal of Experimental Psychology: Learning, Memory and Cognition, 21, 3-23. McNamara, T. P. (1992). Theories of priming: I. Associative distance and lag. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 1173ll90. McNamara, T. P. (1994). Theories of priming: II. Types of primes. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 507-520. Moss, H. E., Hare, M. L., Day, P., & Tyler, L. K. (1994). A distributed memory model of the associative boost in semantic priming. Connection Science, 6, 413427.
The Role of Lexical and Sentential Context Effects
163
Moss, H. E., & Marslen-Wilson, W. D. (1993). Access to word meanings during spoken language comprehension: Effects of sentential semantic context. Journal of Experimental Psychology: Learning, Memory, and Cognition, 19, 1254-1276. Neely, J. H. (1991). Semantic priming effects in visual word recognition: A selective review of current findings and theories. In D. Besner & G. W. Humphreys (Eds.), Basic processes in reading: Visual word recognition (pp. 264-336). Hillsdale, NJ: Erlbaum. O'Seaghdha, P. G. (1989). The dependence of lexical relatedness effects on syntactic connectedness. Journal of Experimental Psychology: Learning, Memory, and Cognition, I5, 73-87. Perfetti, C. A. (1994). Psycholinguistics and reading ability. In M. A. Gernsbacher (Ed.), Handbook ofpsycholinguistics (pp. 849-894). San Diego, CA: Academic Press. Plaut, D. C., & Shallice, T. (1993). Perseverative and semantic influences on visual object naming errors in optic aphasia: A connectionist account. Journal of Cognitive Neuroscience, 5, 89-117. Ratcliff, R., & McKoon, G. (1994). Retrieving information from memory: Spreading-activation theories versus compound-cue theories. Psychological Review, 101, 177-184. Seidenberg, M. S. (1993). A connectionist modeling approach to word recognition and dyslexia. Psychological Science, 4, 299-304. Sharkey, N. E. (1989). A connectionist model of text comprehension. In D. A. Balota, G. B. Flores d'Arcais, & K. Rayner (Eds.), Comprehension processes in reading (pp. 487-514). Hillsdale, N. J.: Erlbaum. Simpson, G. B. (1994). Context and the processing of ambiguous words. In M. A. Gernsbacher (Ed.), Handbook ofpsycholinguistics (pp. 359-374). San Diego, CA: Academic Press. Simpson, G. B., Peterson, R. R., Casteel, M. A., & Burgess, C. (1989). Lexical and sentence context effects in word recognition. Journal of Experimental Psychology: Learning, Memory and Cognition, 15, 88-97. Stanovich, K. E., & West, R. F. (1981). The effect of sentence context on ongoing word recognition: Tests of a two-process theory. Journal of Experimental Psychology: Human Perception & Performance, 7, 658-672. Stanovich, K. E., & West, R. F. (1983). On priming by a sentence context. Journal of Experimental Psychology: General, 112, 1-36. Stanovich, K. E., West, R. F., & Feeman, D. J. (1981). A longitudinal study of sentence context effects in second-grade children: Tests of an interactivecompensatory model. Journal of Experimental Child Psychology, 32, 185-199. Tzelgov, J., & Eben-Ezra, S. (1992). Components of the between-language semantic priming effect. European Journal of Cognitive Psychology, 4, 253-272. Van Petten, C., & Kutas, M. (1991). Electrophysiological evidence for the flexibility of lexical processing. In G. B. Simpson (Ed.), Understanding word and sentence (pp. 129-174). Amsterdam: Elsevier.
This Page Intentionally Left Blank
Bilingual Sentence Processing- R.R. Heredia and J. Aitarriba (Editors) 9 2002 Elsevier Science B.V. All rights reserved.
Text Comprehension in Bilinguals- Integrating Perspectives on Language Representation and Text Processing Gary E. Raney, Sharon M. Obeidallah, and Timothy K. Miura The University of Illinois at Chicago
Abstract
This chapter explores the relation between models of bilingual language representation and text comprehension. Two issues are addressed. The first is whether the processes used when reading in one language are the same as the processes used in a second language. The second is whether the verbatim wording, the meaning of words, and the situation described by the text are represented distinctly in memory. Based on research using bilingual and monolingual readers, we conclude that (a) similar processes are used when reading in each language, but reading strategies vary both as a function of the reader's fluency level and of the characteristics of the text, and (b) words and their meanings are represented separately in memory, and the meanings of words are interpreted in terms of the reader's background knowledge. Bringing together research on bilingual language representation and text comprehension enhances our understanding of bilingual reading. Determining how bilinguals comprehend text poses two distinct challenges. First, researchers must determine how bilinguals represent language in memory. Second, researchers must determine the processes involved in text comprehension. Each of these issues has been investigated extensively, but they are almost always treated as separate research topics. That is, researchers have explored (a) how bilinguals represent language in memory (see de Groot, 1993; Kroll & de Groot, 1997, for reviews), and (b) how readers (usually monolingual) comprehend text (see Gernsbacher, 1990; Graesser, Millis, & Zwaan, 1997; Kintsch, 1998, for reviews), but researchers have rarely merged the study of these two topics into one, that is, how bilinguals read and comprehend text. Our purpose in this chapter is to bring together research on text comprehension and bilingual reading. We will do this by evaluating bilingual reading using theoretical constructs from research on text comprehension and by evaluating text comprehension using theoretical constructs from research on bilingual reading and language representation. At first glance, one might wonder what is the difference between studying bilingual reading and studying text comprehension. After all, bilinguals must comprehend the texts they read. The difference is captured in the theoretical orientations of the research, the vocabulary used to express ideas, and the methodologies employed (Brown, 1998). For example, research on bilingual
166
Gary E. Raney, Sharon M. Obeidallah, and Timothy K. Miura
reading is often focused on understanding how a bilingual's languages are stored in memory (i.e., the organization of the lexicon) and how knowing two languages impacts the processes involved in reading (e.g., de Groot & Nas, 1991; Horiba, 2000; Kroll & Stewart, 1994). In contrast, text comprehension research is often focused on understanding how memory for a specific text is organized (i.e., the organization of a single memory) and how the text itself affects processing (e.g., how a text's structure influences comprehension) (e.g., Fletcher, 1994; Graesser, Kassler, Kreuz, & Mclain-Allen, 1998). There are also similarities in orientation. For example, individual differences in linguistic skill play an important role in research on bilingual reading and text comprehension (e.g., Zwaan & Brown, 1996). Although issues in text comprehension and bilingual reading overlap, we believe the differences are worthy of study. To limit the scope of our discussion, we will focus on two theoretical frameworks, one related to text comprehension and one related to bilingual lexical representation. The dominant theoretical framework for describing text comprehension is Kintsch and van Dijk's model (Kintsch & van Dijk, 1978; van Dijk & Kintsch, 1983), and the dominant framework for describing the organization of a bilingual's language in memory is the hierarchical model (de Groot, 1993; Kroll, 1993; Kroll & de Groot, 1997; Kroll & Stewart, 1994; Potter, So, Von Eckardt, & Feldman, 1984). We will adopt these models as the basis for our discussion; therefore, we begin with a review of each model that highlights points relevant to our subsequent discussion. After presenting an overview of these models, we illustrate how merging our knowledge of text processing and bilingual lexical representation will expand our understanding of bilingual reading. We next describe how Kintsch and van Dijk's model of text comprehension can be used to enhance our understanding of bilingual reading. We then discuss how the hierarchical model of lexical representation can be used to help explain the processes involved in text comprehension.
Kintsch and van Dijk's Model of Text Comprehension
Kintsch and van Dijk (Kintsch, 1988; Kintsch & van Dijk, 1978; van Dijk & Kintsch, 1983) suggested that texts are represented at three distinct levels: the surface form, the textbase, and the situation model (see Kintsch, 1998, for an overview). The surface form contains the wording used in the text. Surface-level representations result from lexical and syntactic analysis of the text (Fletcher & Chrysler, 1990; van Dijk & Kintsch, 1983). The surface level is the least durable of the three levels of representation; therefore, memory for surface features tends to be short lived (Graesser et al., 1997; Kintsch, Welsch, Schmalhofer, & Zimny, 1990). The textbase represents the meaning of a text as a network of propositions (Kintsch, 1974; Kintsch & van Dijk, 1978). The textbase is created based on the semantic content of the words in the text, but is independent of the words used in the text or of the background knowledge used to understand the text. In other words, the textbase depicts the meaning of the words in a text and how the words are related but does not necessarily include the exact words from the text. As such, the textbase can be thought of as a structure-preserving paraphrase of the original text
Text Comprehension in Bilinguals
167
(Fletcher & Chrysler, 1990). For example, the sentences "John bought a couch" and "John purchased a sofa" have different surface structures but essentially the same textbase (meaning). The textbase also provides a more durable form of memory than the surface form (Graesser et al., 1997; Kintsch et al., 1990). The situation model is an integration of the textbase and prior knowledge and can be thought of as a description of the events described by a text. Creation of the situation model reflects comprehension. The situation model includes information that is not part of the original text, such as general knowledge of the topic, inferences, and other byproducts of text comprehension (Fletcher, 1994; Kintsch, 1988; Kintsch et al., 1990). The situation model is the most durable level of text memory and has been shown to remain stable across intervals ranging from several hours to several days (Kintsch, 1998; Kintsch et al., 1990). Recent theories describe the construction of a situation model as one of the reader's primary goals (Gernsbacher, 1990; Graesser et al., 1997; Kintsch, 1988; van Dijk & Kintsch, 1983; see Zwaan & Radvansky, 1998, for a review of situation models in language.) Kintsch and van Dijk's model describes how texts are represented in memory, but it also defines the initial order of text processing (Kintsch, 1988). Reading necessarily begins by evaluating the surface form. After the words and their structure are evaluated, the textbase is constructed. The situation model is then formed by integrating the textbase with the reader's prior knowledge. This development cycle, which is an ongoing process at all three levels, is evident in Kintsch's construction-integration model (Kintsch, 1988). Subsequent processing is not, however, order dependent. For example, if the reader was asked to recall the general meaning of a text, the situation model could be retrieved without recovering the memory of the surface form. If the reader were asked whether a particular word was in a text, the reader would retrieve the surface form first.
The Hierarchical Model of Bilingual Lexical Representation
Current research supports a model of bilingual lexical memory in which word forms (lexical entries) from each language are stored in distinct lexical memories, and the lexical memories are linked to an independent conceptual memory containing the meanings of the words. This is referred to as a hierarchical model to reflect the fact that there are separate levels of representations for words and their meanings (de Groot, 1993; Kroll, 1993; Kroll & de Groot, 1997; Kroll & Stewart, 1994; Potter, So, Von Eckardt, & Feldman, 1984; but see Altarriba, 2000, and Altarriba & Mathis, 1997, for research that calls into question the assumptions of the hierarchical model). According to the hierarchical model, a bilingual has two lexical stores and one primary conceptual store, whereas a monolingual has one lexical store and one conceptual store. Recent versions of this model are referred to as the revised hierarchical model (e.g., Kroll & de Groot, 1997; Kroll & Stewart, 1994). A central issue is how words are linked across languages and to their meanings. According to the revised hierarchical model, words from each language are linked at the lexical level (lexical-to-lexical links), but the link from L2 to L1 is stronger than the link from L1 to L2. This asymmetric linkage reflects the fact that, when
168
Gary E. Raney, Sharon M. Obeidallah, and Timothy K. Miura
learning a second language, translations are made from L2 to L 1 to access meaning. Words from each language are also linked to their meanings (lexical-to-conceptual links). The strength of the lexical-to-conceptual links also varies across languages. The link between L1 words and their meanings is strong. The link between L2 words and their meanings is weak initially, but the strength of these links increases as L2 fluency increases (de Groot & Nas, 1991; Kroll & Stewart, 1994; Sholl, Sankaranarayanan, & Kroll, 1995; see Kroll & de Groot, 1997 for a review; see Altarriba & Mathis, 1997 for evidence that semantic links may form early). The strength of the lexical-to-lexical links and lexical-to-conceptual links depend on many factors, including fluency level, word frequency, and cognate status, to name a few (Altarriba, Kroll, Sholl, & Rayner, 1996; Beauvillain & Grainger, 1987; Caramazza & Brones, 1979; Grosjean, 1998). Recent views on bilingual memory divide the storage of meaning into subcomponents. For example, Pavlenko (1999) discusses the need to divide word meanings into semantic and conceptual components to distinguish between meanings tied to words and language-independent (multi-modal) representations, respectively, de Groot (1992) proposes that word meanings are stored as conceptual features, and others suggest the need for a lemma level that contains morphological and syntactic information and a conceptual level that represents the meanings of words (see Grosjean, 1998, for a discussion). Despite these variations, all of these models are hierarchical in nature, and all distinguish between the storage of words and their meanings. The hierarchical model describes how bilinguals organize their languages in memory. Unlike Kintsch and van Dijk's model (Kintsch, 1988; Kintsch & van Dijk, 1978; van Dijk & Kintsch, 1983), the hierarchical model does not define the order in which information is processed. The hierarchical model depicts the structure of lexical representation, not how words are processed. Processing can begin at the lexical level, the conceptual level, or lexical and conceptual processing can occur in parallel (Kroll & de Groot, 1997). Of course, the order of processing may change based on attributes of the reader and the task (Durguno~lu & Roediger, 1987; Grosjean, 1998). Although lexical and conceptual processing of words can occur in parallel, a particular reading situation may emphasize one of these levels. For example, if the reader does not recognize a word, conceptual processing will be emphasized as the reader tries to infer the meaning from the context. Furthermore, because lower-level decoding processes are more automatic when reading in L1 than L2, reading in L1 de-emphasizes decoding processes and emphasizes conceptual processes, whereas reading in L2 emphasizes decoding processes and draws attention away from conceptual processes (Koda, 1996; Zwaan & Brown, 1996).
Merging Text Comprehension with Bilingual Lexical Representation Let us begin with an example of how merging our knowledge of text processing and bilingual lexical representation will expand our understanding of bilingual reading. Assume a fluent Spanish-English bilingual reads a text once in English and then again in Spanish. Reading the English text should make the Spanish translation
Text Comprehension in Bilinguals
169
easier to read because the meaning of the Spanish text has already been encountered once (albeit in a different language). Whether reading the English text facilitates processing of the Spanish text depends on how strongly the two languages are connected within the lexicon (both lexically and conceptually) (Altarriba, 1992; Altarriba et al., 1996; Durguno~;lu, Mir, & Arifio-Marti, 1993; Kroll & de Groot, 1997; Pavlenko, 1999), the contents of memory after reading the English text (Durguno~,lu, 1997; Durguno~lu & Hancin-Bhatt, 1992; Horiba, van den Broek, & Fletcher, 1993; Raney, 1993; Raney, Atilano, & Gomez, 1996; Raney, Therriault, & Minkoff, 2000), and the strategies used to process the texts in each language (Alderson, 1984; Bernhardt & Kamil, 1995; Francis, 2000; Horiba, 2000; Langer, Bartomlome, Vasquez, & Lucas, 1990; Zwaan & Brown, 1996). But what mechanisms would allow the English passage to facilitate reading of the Spanish translation? If a bilingual's languages are linked at a lexical level (e.g., the English word dog is connected to the Spanish word perro), then words from the English text could be translated into Spanish and then the Spanish meanings could be activated. This would represent a surface-level (word-for-word) translation of the English passage. The English text might, however, contain words that the reader does not know in Spanish. The reader could use the developing textbase (i.e., the context) to help determine the meaning of unknown words. This context-based reading strategy would require a conceptual link between the lexicons of the two languages. In addition, inferences might be required to make the English text coherent. Generating inferences requires access to prior knowledge and the situation model. Fluent translation of inferences requires the languages to share links to general knowledge structures. What does the prior example illustrate? First, it illustrates how language is represented in memory at several levels, such as the word level, contextual level, and the meaning level. Second, it shows how translation of a text requires access to all three levels of a text representation (i.e., the surface form, the textbase, and the situation model). Third, the example shows that we must understand how bilinguals store their languages in memory, as well as how memory for text is organized, to develop a complete explanation of how bilinguals read and comprehend text (see Brown, 1998, for a discussion of related issues). This latter point defines the purpose of this chapter. Specifically, we explore how research on bilingual language representation can be used to enhance our understanding of text comprehension, and how research on text comprehension can be used to improve our understanding of how bilinguals process text. To focus our discussion, we concentrate on two issues. The first issue, which is a key topic in bilingual reading, is whether the processes used when reading in one language are the same as the processes used in a second language. We restate this issue in discourse processing terms and explore whether bilinguals create similar representations for the surface form, textbase, and situation model when reading in their L1 (first language) and L2 (second language). Whether or not the processes are the same across languages has important implications. For example, if the processes are not the same, then accurately translating a text would be difficult because different processes would be applied when reading the text in each language. Thus, the translation would have to be structured in a way that ensures that two different processes lead to the same interpretation or representation of the text. The second issue we discuss, which is a
170
Gary E. Raney, Sharon M. Obeidallah, and Timothy K. Miura
central topic in text comprehension, is whether the surface form, textbase, and situation model are distinct levels of representation. Whether or not the levels of representation are distinct also has important implications. For example, if the surface form is represented separately from the textbase, then it should be possible to translate a text without recognizing the relationship between elements of the text. This would lead to improper translations, and indeed this is often the case with word-for-word translations.
Using Models of Text Comprehension to Understand Bilingual Reading Many bilinguals are less proficient readers in their L2 than in their L1. This has led researchers to explore whether the processes used when reading in L1 are the same as when reading in L2. Specifically, researchers are interested in whether less proficient reading in L2 reflects a language fluency problem or a specific reading problem (Alderson, 1984; Bernhardt, 1986; Fitzgerald, 1995; Goodman, 1971; see Bernhardt & Kamil, 1995, for a review). The view that deficient L2 reading is a language fluency problem is captured in the linguistic threshold hypothesis (Alderson, 1984). According to this view, L2 reading ability is related to linguistic skills in L2, not to reading ability in L 1. That is, poor L2 skills will lead to poor L2 reading and good L1 reading skills cannot overcome poor L2 skills. If the processing of words and syntax in L2 is not efficient, there will be few resources available for comprehension processes. This leads to poor comprehension in L2. Consequently, a minimum level of L2 proficiency must be obtained before comprehension processes can be effectively activated (Bernhardt & Kamil, 1995). The view that deficient L2 reading is a general reading problem is captured in the linguistic interdependence hypothesis (Alderson, 1984; Bernhardt & Kamil, 1995), which is essentially similar to what Goodman called the Reading Universal Hypothesis (Goodman, 1971). According to the linguistic interdependence hypothesis, the cognitive processes used to comprehend texts in L1 transfer to L2 (Bernhardt & Kamil, 1995; Cummings, 1991; Tang, 1997). That is, reading skills mastered in L 1 are available to help processing in L2. Reading processes for L 1 are not separated from L2 reading processes. As a result, a good L1 reader has the potential to become a good L2 reader, but a poor L1 reader is unlikely to become a good L2 reader. The linguistic threshold hypothesis and the linguistic interdependence hypothesis are not mutually exclusive. Research supports the conclusion that L2 skill and language-independent reading skills both influence L2 reading ability (Alderson, 1984; Bernhardt & Kamil, 1995; Zwaan & Brown, 1996). One potential reason why both L2 skill and language-independent reading skills influence L2 reading ability is that L2 skill affects different aspects of L2 reading ability than language-independent reading skills. Skill in L2 influences the reader's ability to process lower-level information, such as lexical and syntactic analysis. This corresponds with processing at the surface and textbase levels. Languageindependent skills influence the reader's ability to complete higher-level comprehension processes, which roughly correspond with development of the situation model.
Text Comprehension in Bilinguals
171
This distinction leads to an alternative way to flame the question of whether reduced reading ability in L2 reflects a language problem versus a reading problem. Specifically, we can explore whether texts are processed similarly across languages by evaluating whether readers create representations based on the surface form, textbase, and situation model that are similar when reading in L1 and L2. If the representations are similar, we can take this as evidence that similar processes are used when reading in L1 and L2 (i.e., L1 processes transfer to L2). If the representations are different for L1 and L2, then we need to determine whether the differences reflect language-specific reading processes or language-independent reading processes. Certain aspects of reading, such as the reader's goals, as well as reader characteristics, such as linguistic fluency and background knowledge, naturally alter the representations of the surface form, the textbase, and the situation model (Horiba, 2000; Graesser et al., 1997; Zwaan & Radvansky, 1998). Evaluating the strategies used to comprehend texts in L1 and L2 might provide clues as to the source of reduced L2 reading ability. Several groups of researchers have evaluated the strategies used to comprehend texts in L 1 and L2 (Aaronson & Ferres, 1986; Barry & Lazarte, 1998; Chang, 1991; Dm:guno~;lu & Hancin-Bhatt, 1992; Fitzgerald, 1995; Francis, 2000; Gass & Selinker, 1983; Horiba, 2000; Horiba et al., 1993; Jim6nez, Garcia, & Pearson, 1996; Langer et al., 1990; Tang, 1997; Zwaan & Brown, 1996). The general conclusion that can be drawn from these studies is that the strategies used to read in L1 and. L2 are remarkably similar. Here we focus on research designed to explore how bilinguals comprehend texts and whether they construct representations corresponding to the surface form, textbase, and situation model when reading in L 1 and L2 (Horiba, 2000; Horiba et al., 1993; Tang, 1997; Zwaan & Brown, 1996). Tang (1997) conducted a detailed study of whether the processes used to comprehend texts in L1 and L2 are similar. Tang evaluated the comprehension strategies used by Chinese-English bilinguals who were skilled readers in their L1 and L2. Reading strategies were monitored in two ways: readers produced think aloud protocols while reading and readers completed strategy check-lists. The think aloud protocols and strategy check-lists were evaluated for signs of what Tang labeled "text-based strategies" (e.g., focus on vocabulary, relating one sentence to a prior sentence, using sentence or grammatical structure, summarizing), "text structure-based strategies" (e.g., look for key words and main ideas, recognizing the text structure, checking for coherence), and "text and prior knowledge combined strategies" (e.g., relate the text to prior experience, representing the meaning of the sentences, forming hypotheses). Tang found that the proportion of strategies of each type used was highly similar for both L1 and L2 texts. This finding supports two primary conclusions. First, the strategies used by skilled readers to comprehend texts are similar for L 1 and L2. Second, the processes used to read in L1 transfer to L2, which is consistent with the linguistic interdependence hypothesis. The strategies noted by Tang (1997) can also be evaluated in terms of Kintsch and van Dijk's model (Kintsch, 1988; Kintsch & van Dijk, 1978; van Dijk & Kintsch, 1983). It is clear that several of the strategies correspond with processes related to representing the surface form (e.g., focus on wording, using sentence, or grammatical structure), the textbase (e.g., relating one sentence to a prior sentence, or representing the meaning of the sentences), and the situation model (e.g., relate
172
Gary E. Raney, Sharon M. Obeidallah, and Timothy K. Miura
the text to prior experience, checking for coherence). Given that the proportion of strategies of each type were similar when reading in L1 and L2, we can conclude that representations of the surface form, the textbase, and the situation model were also similar across languages. Tang's (1997) study was based on skilled readers; therefore, no direct conclusions can be made about the influence of L2 skill on L2 reading ability. Zwaan and Brown (1996) also evaluated the strategies used to comprehend texts in L 1 and L2, but they were specifically interested in the effect of language fluency on situation model development. Participants were fluent English-speaking students who were enrolled in a third-semester college French class. Zwaan and Brown had readers engage in think-aloud protocols while reading in their L1 (English) or L2 (non-fluent French). The think-aloud protocols were evaluated for inference statements, paraphrases of the text, and metacomments (e.g., "I think") that reflected development of a situation model. Zwaan and Brown found that language fluency substantially affected situation model development. Readers generated more explanatory inferences, produced more accurate paraphrases, generated more metacomments, and created more complete situation models when reading texts in their L1 than in their L2. When reading L2 texts, readers concentrated on lowerlevel processes needed for the development of a surface structure and textbase, Zwaan and Brown suggested that a certain level of L2 ability is needed for accurate situation model development, which is consistent with the linguistic threshold hypotheses. At first glance, Zwaan and Brown's (1996) results seem to indicate different strategies were used when reading in L1 and L2. Their evaluation, however, shows that different strategies were used because of differences in skill level within each language and not because of language-specific processing. In essence, because processing, lower-level aspects of the texts consumed a large portion of the nonfluent French speakers' resources, they had limited capacity remaining to devote to development of a situation model. A similar finding has been produced in monolingual research paradigmsmskilled readers create more fully developed situation models than do beginning readers or poor readers (Gernsbacher, 1990; Zwaan & Radvansky, 1998). As in Zwaan and Brown's study, this finding reflects the fact that processing the surface features and the textbase is difficult for beginning readers and poor readers; therefore, they focus their attention on the surface features and the textbase at the expense of situation model development (Bourassa, Levy, Dowin, & Casey, 1998; Faulkner & Levy, 1999; Koda, 1996; Rashotte & Torgeson, 1985; Samuels, 1979). Zwaan and Brown's findings reveal that unskilled L2 readers produce the same deficiencies in situation model development as do beginning or poor L1 readers. Consequently, their findings support the conclusion that the processes used to read in L 1 and L2 are similar. (See Chang, 1991, for a similar conclusion based on an evaluation of the language training needs of English-as-a-second-language students.) Other research also supports the conclusion that differences between L 1 and L2 reading are based primarily on differences in reading skill. For example, based on a review of L2 word recognition research, Koda (1996) concluded that reading difficulties in L2 are often attributable to lexical-level processing difficulties, such as decoding surface features (which requires familiarity with orthographic patterns).
Text Comprehension in Bilinguals
173
This finding is consistent with recent work by Horiba (2000), who showed that the application of reading strategies varies based on available resources; when lowerlevel processes are demanding, attention is focused on lower-level processes and fewer resources are available for higher-level processing. Horiba suggested that nonnative (less skilled) readers have access to fewer strategies when reading in L2, but they do read in a strategic manner. For example, nonnative readers focus on lower-level processes because they do not have the skills or resources needed to complete higher-level comprehension processes. Further, because of inefficient processing, nonnative readers do not have the capacity to vary the application of strategies as much as native readers. Horiba's conclusions match those of Jim6nez et al. (1996), who found that successful (skilled) Latina/o readers adopted more strategies to enhance comprehension than less successful Latina/o readers. In each of these studies, a key variable for explaining variation in the application of reading strategies was reading skill (also see Graesser et al., 1998). This supports the linguistic threshold hypothesis (Alderson, 1984), and is consistent with the conclusion that a minimum level of L2 proficiency must be obtained before comprehension processes can be effectively activated (Bernhardt & Kamil, 1995; Horiba, 2000; Horiba et al., 1993; Zwaan & Brown, 1996). The importance of linguistic skill is reflected in studies based on good and poor monolingual readers. For example, Stanovich (1982) showed that beginning and poor readers have decoding problems, and this minimizes their ability to use contextual cues as comprehension aids. Also, Kletzien (1991) analyzed good and poor readers' reports of their comprehension strategies for expository texts of different levels of difficulty. She found that both good and poor readers used textbase-level strategies (focus on vocabulary, rereading) and situation model-level strategies (inferencing, incorporating prior knowledge) when reading easy passages. When reading difficult passages, poor readers used fewer strategies than good readers. This suggests that fluency (skill) level and text characteristics both affect situation model construction. If text difficulty is increased substantially, good readers behave like poor readers (Rayner & Pollastek, 1989). Within a bilingual paradigm, the influence of task demands is complicated because linguistic fluency varies across the bilingual's languages, and fluency level interacts with task performance (Durguno~lu & Roediger, 1987). The research cited above demonstrates that the strategies used to read in L1 and L2 vary, but this variation seems to reflect differences in reading skill across both L 1 and L2 and not language-specific processing strategies. The research also shows that readers form different representations when reading in their L1 and L2, but these differences seem to reflect the functioning of language-independent reading processes related to linguistic fluency. Readers form representations based on the surface form, the textbase, and the situation model when reading in L1 or L2, but the content of the representation varies as a function of fluency. When non-fluent bilinguals read in their L2, they perform essentially like less-skilled monolingual readers. Specifically, both non-fluent bilinguals and less-skilled monolingual readers form text representations that are biased toward the surface form and textbase at the expense of complete situation model development. In contrast, skilled readers focus on the development of the situation model. Although the outcome of reading in L1 and L2 might appear to be different, when examined in
174
Gary E. Raney, Sharon M. Obeidallah, and Timothy K. Miura
terms of how the surface form, textbase, and situation model are developed, we see that underlying comprehension processes are similar when reading in L 1 or L2.
Using Bilingual Reading to Understand Text Comprehension Thus far, we have focused on how models of text comprehension can be used to understand bilingual reading. Equally important is how research on bilingual lexical representation can be used to advance our understanding of text comprehension. Here we focus on whether the surface form, textbase, and situation model are distinct levels of representation in memory. Whether these three forms of representation are distinct has both theoretical and pedagogical implications, as will be evident from the discussion that follows. Given that the textbase builds on the surface form, and the situation model builds on the textbase, it seems likely that these representations are not independent. A broad range of research based on monolingual reading situations indicates, however, that the surface form, textbase, and situation model are distinct (Fletcher, 1994; Fletcher & Chrysler, 1990; Graesser et al., 1997; Kintsch, 1988; Kintsch et al., 1990; van Dijk & Kintsch, 1983; Zwaan & Radvansky, 1998). The challenge for researchers is finding appropriate methodologies for teasing apart these three levels of representation. Given that linguistic fluency alters the processes used to comprehend a text (Alderson, 1984; Bernhardt & Kamil, 1995; Fitzgerald, 1995; Horiba, 2000; Kleitzen, 1991; Raney et al., 1996; Zwaan & Brown, 1996), bilingual reading paradigms provide a natural procedure for exploring how the surface form, textbase, and situation model are represented in memory when reading L1 and L2 texts.
Distinguishing the Surface Form from the Textbase Distinguishing the surface form from the textbase is difficult because the textbase is based on the surface form. This problem can be partially overcome by evaluating how bilinguals process cognates and non-cognates when reading. Cognates share perceptual and conceptual representations across languages (e.g., the English word problem and the Spanish word problema), whereas non-cognates only share conceptual features (e.g., issues, asuntos) (Altarriba, 1992; Chen& Ng, 1989; de Groot, 1993; de Groot & Nas, 1991). In discourse terms, cognates share surface features and contribute the same meaning to the textbase, whereas noncognates contribute the same meaning to the textbase but do not share surface features. The fact that cognates share two levels of representation should make them easier to process when reading texts presented in both L I and L2. Raney et al. (1996) explored whether cognates and non-cognates are processed differently when read as part of a text. They had fluent Spanish-English bilinguals and non-fluent Spanish-English bilinguals (individuals with two or three years of high school Spanish classes) read texts twice in succession while their eye movements were recorded. The second reading was either the same passage in the same language (identical) or the translation. Embedded in each passage were sets of
Text Comprehension in Bilinguals
175
cognate and non-cognate target words. Overall reading times decreased during the second reading (i.e., a repetition effect) for identical passages and translations. Of importance for this discussion is the effect of repetition on the target words. If repetition benefits for translations occurred solely because the meaning of the words (i.e., the textbase) was repeated, then repetition benefits should be similar for cognates and non-cognates. If, however, repeating the orthographic aspects of words (i.e., surface features) also facilitated reading, then repetition benefits should be larger for cognates than for non-cognates. For fluent bilinguals, fixation times on cognates and non-cognates were similar during the second reading (i.e., equal repetition benefits). Thus, changing the surface form did not impact reading performance. For non-fluent bilinguals, fixation times on cognates and noncognates during the second reading were similar only when the second reading was in English (their dominant language). When the second reading was in Spanish, non-fluent bilinguals read cognates faster than non-cognates. Thus, when reading in their L2, processing difficulty for cognates was reduced because surface features were repeated. If the repetition was in L1, processing the meaning of words was automatic; therefore, no benefit from repeating surface characteristics occurred. Other researchers have also found that non-fluent bilinguals emphasize surface features when reading in their L2 (e.g., Koda, 1996). Durguno~lu and Hancin-Bhatt (1992) suggest that the relation between surface-level and textbase-level processing also varies as a function of task demands such as text difficulty. Further evidence for a distinction between surface-level and textbase-level representations can be found in a study by Altarriba et al. (1996). Altarriba et al. examined whether sentence contexts influence processing at the lexical level or a conceptual level. They gave readers English sentences that contained an occasional Spanish word and measured reading time for the Spanish target words. If sentence contexts lead readers to expect certain semantic properties, but not specific words within a specific language, then reading times for Spanish target words should not be longer than English target words because words from both languages satisfy the semantic constraints. If sentence contexts also specify lexical properties, such as the language of the expected word, then reading times for Spanish target words should be longer than English target words because only the English words match the lexical features of the sentence contexts. Altarriba et al. found evidence for both lexical and conceptual constraints, which supports a distinction between lexicallevel (surface) and conceptual-level (textbase) context effects (but see Costa & Caramazza, 1999, for evidence that lexical selection during speech production can be language specific). Altarriba et al.'s results are consistent with the separation bf surface-level constraints and textbase-level constraints in memory. The above research provides evidence that words are separated from their meanings, which is consistent with a distinction between surface-level and textbaselevel representations. Bilingual models of lexical organization, including the hierarchical model (Kroll & de Groot, 1997) and alternatives (e.g., de Groot, 1992; Grosjean, 1998; Pavlenko, 1999), are naturally consistent with this distinction because these models separate the representation of words from their meanings in the lexicon. It is interesting to note that current models of word processing, which are dominantly monolingual in orientation, also distinguish between the surface features of words (letters) and their meanings. For example, according to the
176
Gary E. Raney, Sharon M. Obeidallah, and Timothy K. Miura
interactive activation model, words are processed at the feature level, the letter level, and meaning level (McClelland & Rummelhart, 1981; Rayner & Pollatsek, 1989). Thus, from both a monolingual and bilingual perspective, mounting evidence supports the conclusion that words and their meanings are separated in the lexicon as well as in representations of text.
Distinguishing the Textbase from the Situation Model
Just as the textbase is separated from but dependent upon the surface form, research has shown that the situation model is distinct from but built upon the textbase. As mentioned earlier, a key difference between the textbase and the situation model is that the situation model contains information that is not part of the text, such as inferences and background knowledge. The situation model reflects the reader's comprehension of a text. As such, research methods that distinguish the meaning of the text itself from the reader's understanding of the text are needed to determine if the textbase is distinct from the situation model. Bilingual paradigms can provide this type of research method. A study by Durguno~,lu et al. (1993) provides evidence for distinguishing memory for the surface form and textbase from the situation model. In their study, bilinguals read a text in English or Spanish, reread the text in the same or different language, and then answered questions about the text. The second reading either immediately followed the first reading (massed repetition) or was delayed (spaced repetition). Past research has shown that comprehension is reduced following massed repetition relative to spaced repetition. Interestingly, Durguno~,lu et al. found that changing the language of the repeated text overcame the comprehension disadvantage of massed repetition, but only for comprehension questions that tapped surface-level and textbase-level information, that is, for questions that required recall of specific information from a text. Their results are consistent with research showing that processing a text in two languages (e.g., reading in one language while taking notes in a second language) enhances memory for some aspects of wording, but does not enhance comprehension (Kardash, Amlund, Kulhavy, & Ellison, 1988; also see Durguno~lu & Hancin-Bhatt, 1992). Durguno~;lu et al. (1993) also measured performance on questions that required the reader to make an inference, that is, generate answers that were not explicitly mentioned in the text. The number of inference questions answered correctly was similar for massed and spaced repetitions. Further, changing the language of the repetition had little effect on the number of inference questions answered correctly. This shows a distinction between comprehension of textbase-level information and situation model-level information. Reading a text once did not allow the reader to create a strong memory for the textbase. In contrast, reading a text once was sufficient to create a situation model; therefore, the language of the texts or the spacing of the repetitions did not influence situation-level memory. These results are consistent with a large body of research demonstrating stronger memory for the situation model than for the surface form or textbase (Fletcher, 1994; Fletcher & Chrysler, 1990; Graesser et al., 1997; Kintsch, 1998; Kintsch et al., 1990), and with research showing that attention switches from obtaining overall meaning during the
Text Comprehension in Bilinguals
177
first reading of a text to obtaining details about the text during the second reading (Raney, 1993). A distinction between the textbase and situation model can also be drawn based on the finding that vocabulary knowledge does not consistently support comprehension (Davis, 1989; Jacobs, Dufon, & Hong, 1994; Johnson, 1982). For example, Jacobs et al. (1994) found that providing vocabulary glosses (definitions within the margins of the text) did not improve high-level comprehension. This implies a separation between processing the word meanings (textbase) and integrating their meaning into the text (situation model). Specifically, providing the meaning of unfamiliar words does not automatically support processing at the situation model level.
Distinguishing the Surface Form from the Situation Model.
Given that the surface form is distinct from the textbase, and the textbase is distinct from the situation model, the surface form must be distinct from the situation model. Although this point seems obvious, the impact of this distinction is not necessarily obvious and warrants further discussion. One important consequence of separating the representation of surface form from the situation model (and the textbase) is that there is no direct link between the surface form and the situation model. As a result, the surface form cannot be precisely regenerated from the situation model because there is no index to differentiate what words were in the text originally from what words the reader added (Raney, King, & Therriault, 2000). To explore this phenomenon, Raney, King, et al. (2000) had subjects read a short text and then immediately recall (in writing) everything they could remember about the text. After a short delay, subjects were asked to recall the text a second time or to recall what they wrote in their first recall. The surface form of the second recalls was then compared to the original text and to the first recalls. Regardless of the recall instructions, subjects' second recalls matched their first recalls slightly better than they matched the original text. In essence, readers changed some of the words when storing the text in memory and then recalled those changes even if the task required them not to recall the changes. This finding is consistent with research demonstrating an inability to distinguish between an original text and a modified version of the text that matches the situation model of the text (Fletcher, 1994; Fletcher & Chrysler, 1990; Kintsch et al., 1990; Raney, Therriault, et al., 2000). A second consequence is that one text can be interpreted in more than one way based on the background knowledge applied to the text by the reader (Gumperz, 1982; Johnson, 1982; Wiley & Rayner, 2000). For example, Wiley and Rayner (2000) studied how individuals with low and high baseball knowledge interpret sentences containing ambiguous words that have baseball-related and non-baseballrelated meanings (e.g., bats, meaning baseball bats or the flying mammal). They found that readers with high baseball knowledge had more difficulty instantiating the non-baseball-related meaning of ambiguous words than did readers with low baseball knowledge even for sentences that were strongly biased towards the nonbaseball-related meaning. This shows that background knowledge operates at a very
178
Gary E. Raney, Sharon M. Obeidallah, and Timothy K. Miura
low level of language processing and influences the interpretation o f the words (i.e., the surface form). A parallel situation exists for bilingual readers. Readers who are not fluent in their L2 tend to have impoverished sociolinguistic knowledge about their L2. As a result, non-fluent speakers do not always realize the social interpretation of their words. For example, the phrase "you're crazy" in English means that the speaker believes the addressee has said something ridiculous. Calling someone crazy is not usually taken as a major insult. In at least one Palestinian Arab town, the Arabic phrase "you're crazy" ("inta madjnoon") is used only kiddingly with very close intimates; it is, otherwise, meant and taken literally to mean the addressee truly is or when the intention of the speaker is to strongly insult the addressee (see Gumperz, 1982, for the effects of misinterpretation due to cultural differences of seemingly straightforward messages). Hatim (1997) studied formal types of translation and the problem of apparently innocuous words and linguistic styles used culturally inappropriately. His explanation, the contrastive discourse model, is an attempt to understand how culturally determined and culturally idiosyncratic are word choice, word order, and text structure. Although the relation between knowledge and translation is a complex issue, the conclusion is straightforward: Individuals with different background knowledge can interpret a single surface form in different ways. A third consequence of separating the surface form from the situation model is that bilinguals are not always aware of code switching (language changes) in a text or in a conversation. What bilinguals are responding to, often unconsciously, are the linguistic and cultural cues that they have implicitly learned, just as monolinguals have implicitly learned when and how to use informal as opposed to formal linguistic modes of speech (see, Auer, 1995; Gardner-Chloros, 1995). This reflects the reader/listener's goal of comprehending the message, not storing the words. This would not be possible if the situation model was coded at the surface level.
Conclusions
The purpose of this chapter was to bring together research on text comprehension and bilingual reading. Our goal was to show how models of bilingual lexical representation can be applied to text comprehension research, and how models of text comprehension can be applied to research on bilingual reading. An important issue in bilingual reading research is the relation between reading in L1 and L2. Past research indicated that the strategies used to read in L1 and L2 are similar. We evaluated this issue by looking for evidence that readers create representations based on the surface form, textbase, and situation model that are similar when reading in L1 and L2. We found evidence that fluent bilinguals form similar representations for L1 and L2 texts. For non-fluent bilinguals, we found evidence that they form representations based on the surface form, textbase, and situation model when reading in L I or L2, but the content of their representations varies based on their fluency in each language. Non-fluent bilinguals focus on the surface form and textbase and create incomplete situation models. When examined in terms of how the surface form, textbase, and situation model are developed, the
Text Comprehension in Bilinguals
179
evidence supports the conclusion that comprehension processes are similar when reading in L 1 or L2. An important issue in text comprehension research is whether the surface form, textbase, and situation model are distinct representations. We examined this issue by looking for evidence in bilingual research that words are separated from their meanings, and that the meanings of words are interpreted in terms of the reader's background knowledge. This corresponds with a separation of the surface form from the textbase, and the textbase from the situation model, respectively. We also discussed the implications of separating surface form from the situation model. We concluded that separating words from their meanings, and meanings from linguistic knowledge is consistent with the hierarchical model of lexical representation. In conclusion, we believe that integrating research on bilingual reading and text comprehension produces a more complete understanding of bilingual text comprehension than can be obtained by studying either topic in isolation.
References
Alderson, J. C. (1984). Reading in a foreign language: A reading problem or a language problem? In J. C. Alderson & A. H. Urquhart (Eds.), Reading in a foreign language (pp. 1-24). London: Longman. Altarriba, J. (1992). The representation of translation equivalents in bilingual memory. In R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 157-174). Amsterdam: Elsevier Science Publishers. Altarriba, J. (2000). Language processing and memory retrieval in SpanishEnglish bilinguals. Spanish Applied Linguistics, 4, 215-245. Altarriba, J., Kroll, J. F., Sholl, A., & Rayner, K. (1996). The influence of lexical and conceptual constraints on reading mixed-language sentences: Evidence from eye fixations and naming times. Memory & Cognition, 24, 477-492. Altarriba, J., & Mathis, K. M. (1997). Conceptual and lexical development in second language acquisition. Journal of Memory and Language, 36, 550-568. Aaronson, D., & Ferres, S. (1986). Sentence processing in Chinese-American bilinguals. Journal of Memory and Language, 25, 136-162. Auer, P. (1995). The pragmatics of code-switching: A sequential approach. In L. Milroy & P. Muysken (Eds.), One speaker, two languages (pp. 115-136). New York: Cambridge University Press. Barry, S., & Lazarte, A. A. (1998). Evidence for mental models: How do prior knowledge, syntactic complexity, and reading topic affect inference generation in a recall task for nonnative readers of Spanish? Modern Language Journal 82, 176193. Bernhardt, E. B., (1986). Cognitive processes in L2: An examination of reading behaviors. In J. Lantolf & A. Labarca (Eds.), Research in second language acquisition in the classroom setting (pp. 35-50). Norwood, NJ: Ablex. Beauvillain, C., & Grainger, J. (1987). Accessing interlexical homographs: Some limitations of a language-selective access. Journal of Memory and Language, 26, 658-672.
180
Gary E. Raney, Sharon M. Obeidallah, and Timothy K. Miura
Bernhardt, E. B., & Kamil M. L. (1995). Interpreting relationships between L1 and L2 reading - consolidating the linguistic threshold and the linguistic interdependence hypotheses. Applied Linguistics, 16, 15-34. Bourassa, D. C., Levy, B. A., Dowin, S., & Casey, A. (1998). Transfer effects across contextual and linguistic boundaries: Evidence from poor readers. Journal of Experimental Child Psychology, 71, 45-61. Brown, C. M. (1998). L2 reading: An update on relevant L1 research. Foreign Language Annals, 31, 191-202. Caramazza, A., & Brones, I. (1979). Lexical access in bilinguals. Bulletin of the Psychonomic Society, 13, 212-214. Chang, F. R. (1991). An experimental investigation of text comprehension processes in bilinguals" Implications for training. In D. F. Ronna & J. W. Pellegrino (Eds.), Instruction: Theoretical and applied perspectives (pp. 71-81). New York: Praeger Publishers. Chen, H. C., & Ng, M. L. (1989). Semantic facilitation and translation priming effects in Chinese-English bilinguals. Memory & Cognition, 17, 454-462. Costa, A., & Caramazza, A. (1999). Is lexical selection in bilingual speech production language-specific? Further evidence from Spanish-English and EnglishSpanish bilinguals. Bilingualism: Language and Cognition, 2, 231-244. Cummings, J. (1991). Interdependence of first- and second-language proficiency in bilingual children. In E. Bialystok (Ed.), Language processing in bilingual children (pp. 70-89). Cambridge: Cambridge University Press Davis, J. N. (1989). Facilitating effects of marginal glosses on foreign language reading. Modern Language Journal, 73, 41-48. de Groot, A. M. B. (1992). Bilingual lexical representation: A closer look at conceptual representations. In R. Frost & L. Katz (Eds.), Orthography, phonology, morphology, and meaning (pp. 389-412). Amsterdam: Elsevier. de Groot, A. M. B. (1993). Word-type effects in bilingual processing tasks: Support for a mixed-representational system. In R. Schreuder & B. Weltens (Eds.), The bilingual lexicon (pp. 27-52). Amsterdam: John Benjamins Publishing Co. de Groot, A. M. B., & Nas, G. (1991). Lexical representation of cognates and noncognates in compound bilinguals. Journal of Memory and Language, 30, 90123. Durguno~,lu, A. Y. (1997). Bilingual reading: Its components, development, and other issues. In A. M. B de Groot & J. F. Kroll (Eds.), Tutorials in bilingualism: Psycholinguistic perspectives (pp. 255-278). Mahwah, NJ: Lawrence Erlbaum Associates. Durguno~,lu, A. Y., & Hancin-Bhatt, B. J., (1992). The role offirst language in the second-language reading process (Technical Report No. 555). Urbana, IL: University of Illinois Center for the Study of Reading. Durguno~lu, A. Y., Mir, M., & Arir'io-Marti, S. (1993). Effects of repeated readings on bilingual and monolingual memory for text. Contemporary Educational Psychology, 18, 294-317. Durguno~;lu. A. Y., & Roediger, H. L. (1987). Test differences in accessing bilingual memory. Journal of Memory and Language, 26, 377-391. Faulkner, H. J., & Levy, B. A. (1999). Fluent and nonfluent forms of transfer in reading: Words and their message. Psychonomic Bulletin & Review, 6, 111-116.
Text Comprehension in Bilinguals
181
Fitzgerald, J. (1995). English-as-a-second-language learners' cognitive reading processes" A review of research in the United States. Review of Educational Research, 65, 145-190. Fletcher, C. R. (1994). Levels of representation in memory for discourse. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 589-607). San Diego: Academic Press. Fletcher, C. R., & Chrysler, S. T. (1990). Surface forms, textbases, and situation models: Recognition memory for three types of textual information. Discourse Processes, 13, 175-190. Francis, N. (2000). The shared conceptual system and language processing in bilingual children: Findings from literacy assessment in Spanish and Nahuatl. Applied Linguistics, 2 I, 170-204. Gass, S. M., & Selinker, L. (1983). Language transfer in language learning. Rowley, MA: Newbury House. Gardner-Chloros, P. (1995). Code-switching in community, regional, and national repertoires: The myth of discreteness in linguistic systems. In L. Milroy & P. Muysken (Eds.), One speaker, two languages (pp. 68-89). New York: Cambridge University Press. Gernsbacher, M. A. (1990). Language comprehension as structure building. Hillsdale, NJ: Erlbaum. Goodman, K. S. (1971). Psycholinguistic universals in the reading process. In P. Pimsleur & T. Quinn (Eds.), The psycholinguistics of second language learning (pp. 135-142). Cambridge" Cambridge University Press. Graesser, A. C., Kassler, M. A., Kreuz, R. J., & Mclain-Allen, B. (1998). Verification of statements about story worlds that deviate from normal conceptions of time: What is true about Einstein's dreams? Cognitive Psychology, 35, 246-301 Graesser, A. C., Millis, K. K., & Zwaan, R. A. (1997). Discourse comprehension. Annual Review of Psychology, 48, 163-189. Grosjean, F. (1998). Studying bilinguals: Methodological and conceptual issues. Bilingualism: Language and Cognition, 1, 131-149. Gumperz, J. J. (1982). Discourse strategies. Cambridge: Cambridge University Press. Hatim, B. (1997). Communication across cultures: Translation theory and contrastive text linguistics. Exeter, UK: University of Exeter Press. Horiba, Y. (2000). Reader control in reading: Effects of language competence, text type, and task. Discourse Processes, 29, 223-267. Horiba, Y., van den Broek, P. W., & Fletcher, C. R. (1993). 2nd-language readers memory for narrative t e x t s - evidence for structure-preserving top-down processing. Language Learning, 43, 345-372. Jacobs, G. M., Dufon, P., & Hong, F. G. (1994). L1 and L2 vocabulary glosses in L2 reading passages: Their effectiveness for increasing comprehension and vocabulary knowledge. Journal of Research in Reading, 17, 19-28. Jim6nez, R. T., Garcia, G. E., & Pearson, P. D. (1996). The reading strategies of bilingual Latina/o students who are successful English readers: Opportunities and obstacles. Reading Research Quarterly, 31, 90-112. Johnson, P. O. (1982). Effects on reading comprehension of background knowledge. TESOL Quarterly, 19, 503-516.
182
Gary E. Raney, Sharon M. Obeidallah, and Timothy K. Miura
Kardash, C. A., Amlund, J. T., Kulhavy, R. W., & Ellison, G. (1988). Bilingual referents in cognitive processing. Contemporary Educational Psychology, 13, 4557. Kintsch, W. (1974). The representation of meaning in memory. Hillsdale, NJ: Erlbaum. Kintsch, W. (1988). The role of knowledge in discourse comprehension: A construction-integration model. Psychological Review, 95, 163-182. Kintsch, W. (1998). Comprehension: A paradigm for cognition. Cambridge, UK: Cambridge University Press. Kintsch, W., & van Dijk, T. A. (1978). Toward a model of text comprehension and production. Psychological Review, 85, 363-394. Kintsch, W., Welsch, D., Schmalhofer, F., & Zimny, S. (1990). Sentence memory: A theoretical analysis. Journal of Memory and Language, 29, 133-159. Kletzien, S. B. (1991). Strategy use by good and poor comprehenders reading expository text of differing levels. Reading Research Quarterly, 26, 67-85. Koda, K. (1996). L2 word recognition research: A critical review. The Modern Language Journal 80, 450-460. Kroll, J. F. (1993). Accessing conceptual representations for words in a second language. In R. Schreuder & B. Weltens (Eds.), The bilingual lexicon (pp. 53-82). Amsterdam: John Benjamins Publishing Co. Kroll, J. F., & de Groot, A. M. B. (1997). Lexical and conceptual memory in the bilingual: Mapping form to meaning in two languages. In A. M. B de Groot & J. F. Kroll (Eds.), Tutorials in bilingualism: Psycholinguistic perspectives (pp. 169-200). Mahwah, N J" Lawrence Erlbaum Associates. Kroll, J. F., & Stewart, E. (1994). Category interference in translation and picture naming: Evidence for asymmetric connections between bilingual memory representations. Journal of Memory and Language, 33, 149-174. Langer, J. A., Bartolome, L., Vasquez, O., & Lucas, T. (1990). Meaning construction in school literary tasks: A study of bilingual students. American Education Research Journal 27, 427-471. McClelland, J. L., & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part 1. An account of basic findings. Psychological Review, 88, 375-407. Pavlenko, A. (1999). New approaches to concepts in bilingual memory. Bilingualism: Language and Cognition, 2, 209-230. Potter, M. C., So, K. F., Von Eckardt, B., & Feldman, L. B. (1984). Lexical and conceptual representation in beginning and more proficient bilinguals. Journal of Verbal Learning and Verbal Behavior, 23, 23-38. Raney, G. E. (1993). Monitoring changes in cognitive load during reading: An event-related brain potential and reaction time analysis. Journal of Experimental Psychology: Learning, Memory, and Cognition, 19, 51-69. Raney, G. E., Atilano, R., & Gomez, L. (1996, November). Language representation and bilingual reading. Presented at the 37 th meeting of the Psychonomic Society, Chicago, IL. Raney, G. E., King, K. T., & Therriault, D. J. (2000, January). The stability of text memory: measuring memory for text vs. memory for memories of text. Presented at the Winter Conference on Discourse, Text, and Cognition, Jackson Hole, WY.
Text Comprehension in Bilinguals
183
Raney, G. E., Therriault, D., & Minkoff, S. (2000). Repetition effects from paraphrased text: Evidence for an integrated representation model of text representation. Discourse Processes, 29, 61-81. Rayner, K., & Pollatsek, A. (1989). The psychology of reading. Englewood Cliffs, NJ: Prentice Hall. Rashotte, C. A., & Torgeson, J. K. (1985). Repeated reading and reading fluency in learning disabled children. Reading Research Quarterly, 20, 180-202. Samuels, S. J. (1979). The method of repeated readings. The Reading Teacher, 32, 403-408. Sholl, A., Sankaranarayanan, A., & Kroll, J. F. (1995). Transfer between picture naming and translation" A test of asymmetries in bilingual memory. Psychological Science, 6, 45-49. Stanovich, K. E. (1982). Individual differences in the cognitive processes of reading: II. Text-level processes. Journal of Learning Disabilities, 15, 549-554. Tang, H. (1997). The relationship between reading comprehension processes in L1 and L2. Reading Psychology: An International Quarterly, 18, 249-301. van Dijk, T. A., & Kintsch, W. (1983). Strategies of discourse comprehension. New York: Academic Press. Wiley, J., & Rayner, K. (2000, November). Why baseball fans don't like lumpy batters: The influence of domain knowledge on lexical ambiguity resolution. Presented at the 41 st meeting of the Psychonomic Society, New Orleans, LA. Zwaan, R. A., & Brown, C. M. (1996). The influence of language proficiency and comprehension skill on situation-model construction. Discourse Processes, 21, 389-327. Zwaan, R. A., & Radvansky, G. A. (1998). Situation models in language comprehension and memory. Psychological Bulletin, 123, 162-185.
This Page Intentionally Left Blank
Part IV: Psycholinguistic Theory and Research
This Page Intentionally Left Blank
Bilingual SentenceProcessing- R.R. Herediaand J. Altarriba(Editors) 9 2002 ElsevierScienceB.V. All rights reserved.
8
Relative Clause Attachment in Bilinguals and Monolinguals
Eva M. Fern~indez Queens College, City University of New York
Abstract
This chapter presents evidence from a group of experiments comparing the perceptual strategies employed by Spanish-English bilinguals to those used by their monolingual counterparts. The data show that language history affects the way readers assign structure to ambiguous syntactic constituents, but in different ways off-line and on-line. Off-line, bilinguals depart from the monolingual model in their attachment of relative clauses in their nondominant language, and exhibit language-independent strategy use--with input in either language, bilinguals have behavior patterns similar to those of monolinguals of their dominant language. On-line, however, bilinguals appear to have no reliable attachment preferences, unlike monolinguals, who uniformly prefer attachment to the low site. The bilinguals' departure from the monolingual model is not taken to be indicative of a qualitative difference between monolingual and bilingual sentence processing. Instead, this result is attributed to a lack of sensitivity in the experimental procedure, which taps early processing with monolinguals, but misses the opportunity with bilinguals, whose reading times are overall slower.
Introduction
An intriguing question to pose about bilingual performance is whether fluent speakers of two languages, when processing input in each of their languages, behave in ways similar to those of their monolingual counterparts. There are two possible answers to this question, each with radically different implications regarding the architecture of the bilingual's system for decoding linguistic signals. However, this question can only be examined empirically given the availability of a phenomenon exhibiting attested cross-linguistic variation in monolingual populations. As will be argued below, looking for behavior in bilinguals that is language independent (same with input in either language) or language dependent (different with input in each language) requires an established behavioral difference in the two comparison groups, monolingual speakers of two languages. This paper summarizes some recent findings on the way bilinguals deal with input in their two languages, focusing on a phenomenon with which cross-linguistic differences in monolinguals have been observed. ~ The construction under
188
Eva M. Ferndndez
consideration is the relative clause attachment ambiguity, in which a relative clause modifies either of the two nouns in a complex noun phrase (NP), as in the example below. 1. Andrew had dinner yesterday with the nephew of the teacher that was in the communist party. 1." Andr6s cen6 ayer con el sobrino del maestro que estaba en el partido comunista. Studies of speakers of English and speakers of Spanish have reported crosslinguistic differences in the way such ambiguous relative clauses are preferably attached. In general, English speakers exhibit a preference for the relative clause, that was in the communist party, to refer to teacher, in contrast to Spanish speakers, whose preferred site for attaching the relative clause is sobrino ('nephew') when reading the Spanish translation-equivalent sentence, shown in (1"). These differences in monolingual performance permit the study of language (in)dependency in bilingual sentence processing: will Spanish-English bilinguals exhibit language-independent or language-dependent behavior when confronted with the relative clause attachment ambiguity? Language-dependent behavior would manifest itself as bilinguals using different strategies depending on the language of the current input (a low attachment preference in English, a high attachment preference in Spanish). In contrast, bilinguals exhibiting similar attachment preferences with linguistic input in both of their languages (for example, a low attachment preference in both English and Spanish) would be evidence of languageindependent strategy use. Current research on speakers of more than one language abounds with studies of bilingual competence, especially studies of acquisition history and the development of second language grammars (for a comprehensive review, see Ellis, 1994). There is also an extensive body of literature on lexical organization and bilingualism (see, e.g., Kroll & de Groot, 1997). To date, however, there has been relatively little work on how bilinguals assign structure to input when reading or listening to sentences (with the exception of the research collected in the present volume; see also Dussias, 2001; Frenck-Mestre, 1997; Frenck-Mestre &Pynte, 1997; Nicol, Teller, & Greth, 2001). The research reported in this paper was originally an attempt to fill this gap. Not only do the findings lead to some interesting answers regarding language (in)dependency in the processing of two languages by bilinguals, but at the same time the data uncover some promising avenues for future work. Before a discussion of the critical findings, the next two sections provide some necessary background. First the rationale for exploring the question of language (in)dependency in bilingual sentence processing will be outlined. Then some of the existing evidence on monolingual relative clause attachment preferences will be reviewed, along with the explanation of some of the theoretical assumptions as well as the identification of some of the questions that are still under debate in the monolingual processing literature.
Relative Clause Attachment in Bilinguals and Monolinguals
189
Studying Language (In)dependency in Bilingual Sentence Processing It is reasonable to assume that there is a high degree of interconnectedness between the two languages of a bilingual (for some formal proposals see de Bot, 1992; Dufour, 1997; Poulisse, 1997). Such critical links provide the necessary mechanisms to explain a number of phenomena in bilingual behavior--for instance, code-switching (Grosjean, 1997; Myers-Scotton, 1993; Ritchie & Bhatia, 1998), na'l've translation (Malakoff & Hakuta, 1991), and transfer of knowledge learned in one language into the other (Francis, 1999)--which any model of bilingual linguistic representation must account for. However, in addition to being able to communicate effectively when performing in the so-called bilingual mode (when the two languages are activated concurrently), bilinguals are also quite adept at restricting themselves to one code, when communicating with monolingual interlocutors, for example. This fact has implications regarding how critical aspects of the bilingual's linguistic architecture should be understood. In effect, the existence of this unilingual mode that bilinguals so effortlessly engage in (Green, 1998; Grosjean, 198.2, 1998) provides compelling evidence for assuming that there must exist separate lexical and grammatical components for Lx and Ly. 2 While this assumption is intuitively correct regarding grammatical competence and lexical encoding, the question raised here is whether the same is true for the systems involved in bilingual performance: are the bilingual's systems language-dependent (different, and which one is used depends on the language of the current input) or language-independent (the same, no matter the language of the current input)? The study of language (in)dependency in bilingual performance rests crucially on the question of whether cross-linguistic differences exist in monolingual sentence processing (Fernandez, 1998, 2000; see also Dussias, 2001; Nicol et al., 2001). Consider first a possible scenario in which monolingual speakers of all languages process input using a set of strategies which contain at least some language-specific components. In Figure 1, this is represented by the group of boxes labeled A. Monolingual speakers of Lx use a set of strategies, Sx, which differs (indicated by "~") at least in part from the set of strategies Sy used by monolingual speakers of Ly. In such a world we would be confronted with two possible (though not necessarily mutually exclusive) outcomes in bilingual behavior. Bilinguals may behave as monolinguals of their respective two languages, exhibiting languagedependent behavior, that is, using Sx for processing input in Lx, and Sy for processing input in Ly. Alternatively, bilinguals may rely on identical routines to process input in both of their languages. The language-independent set of sentence processing strategies employed by bilinguals might then be similar to the one used by monolingual speakers of the bilinguals' dominant language, or by monolinguals of the bilinguals' more frequently used language, or by monolinguals of the bilinguals' language acquired earlier in life. Additionally, we might speculate that whether a bilingual exhibits language-independent strategy use has to do with the bilingual's degree of balance in terms of proficiency in Lx and Ly, or frequency of use of Lx and Ly, such that the more balance, the greater the likelihood of language dependency.
190
Eva M. Fern6ndez
MONOLINGUALS Sx~Sy
MONOLINGUALS Sx = Sy
i
1
BILINGUALS Sx ~Sy language-dependent processing
BILINGUALS" Sx=Sy language-independent processing
I
BILINGUALS Sx-Sy language-independent processing
Figure 1. Correspondences between bilingual and monolingual behavior in sentence
processing, categorized as language-independent and language-dependent. Sx indicates set of strategies used to process Lx.
A second possibility exists in monolingual sentence processing, the case where monolinguals of all languages follow the same set of (universal, languageindependent) strategies. This is the case outlined by the group of boxes labeled B in Figure i. Given this scenario, one would logically deduce that bilinguals would follow the same set of (universal, language-independent) strategies, with linguistic input in any language. Any observed differences between monolingual and bilingual behavior would then necessarily be due to variables other than those related to the nature of the strategies employed to assign syntactic structure to linguistic input. If monolingual processing is found to be grounded on essentially a universal set of strategies that do not vary from language to language, then the study of language (in)dependency in bilingual sentence processing would be limited to exploring whether bilinguals have one copy or two of the universal device used to assign syntactic structure to linguistic input. Drawing an empirical distinction between the two alternatives not only would turn out to be a difficult exercise, but also might essentially be a question about neurolinguistic organization with little or no consequence at the level of linguistic organization. Distinguishing the two scenarios A and B in Figure 1, primarily rests on finding evidence of cross-linguistic differences in monolingual processing strategies. Differences in the behavior of monolinguals of two contrastive languages would be attributed to the use of different strategies, that is, to a case where Sx ~ Sy. On the other hand, similarities among monolinguals of two languages would be used as evidence that they employ similar (possibly universal) processing strategies, a case where Sx = Sy.
Relative Clause Attachment in Bilinguals and Monolinguals
191
The Bilingual "Performance Deficit" One of the well-documented characteristics of bilingual performance is the phenomenon one might call a "performance deficit" associated with bilingualism (see Cook, 1997, for a review; see Green, 1998, for discussion regarding the cognitive "side effects" of bilingualism; see Noel & Fias, 1998, for discussion of numerical cognition in bilinguals). Generally, bilinguals are relatively slower than monolinguals in processing linguistic input, especially if the input is in their nondominant language. Additionally, depending on the task, sometimes bilinguals' responses are not as accurate as those of monolinguals (for references and discussion, see Cook, 1997). We could speculate that the performance deficit is induced by a number of possible sources, including: differences in lexical access and retrievalbetween monolinguals and bilinguals (bilinguals require more time to find a lexical target in their lexical store, which contains items in both Lx and Ly); differences in the automaticity of the strategies used by monolinguals and bilinguals (bilinguals use routines that are less automatic than those of monolinguals, who have developed greater automaticity with their one language); and the existence of less rules to sort through in monolinguals than bilinguals (in bilinguals, there could be parallel activation of the two languages; see Altenberg & Cairns, 1983). The puzzle behind the bilingual performance deficit is determining exactly what it means that bilinguals are sometimes slower and sometimes less accurate than their monolingual counterparts, when performing in certain types of linguistic tasks. Crucially, none of the interpretations of the deficit presuppose differences in the types of processes (i.e., in the patterns of behavior) involved in bilingual versus monolingual perception. This is an important point in the context of this discussion, because it implies that any departure from the monolingual model in bilingual sentence processing (in the case where monolingual sentence processing is based on a universal set of principles) will result not in non-monolingual-like behavior, but rather in monolingual-like patterns, with overall performance being somewhat slower and somewhat less accurate (for findings with bilinguals in support of this last point, see Frenck-Mestre &Pynte, 1997). Even without a full understanding of the exact source for the performance deficit, we can still explore whether bilinguals will depart from the monolingual model in perceptual tasks, albeit not by looking at overall performance, but rather by exploring whether the patterns of bilingual behavior match those observed with monolinguals.
Processing Linguistic Input: The Monolingual Evidence We now turn to examine a specific aspect of processing that will allow us to pursue the question of language (in)dependency in bilingual sentence processing. Our focus is on the strategies employed by listeners and readers to assign structure to linguistic input, procedures that begin to apply as soon as the listener/reader ("perceiver") has decoded a linearly-ordered set of lexical items that need to be assigned syntactic structure.
192
Eva M. Fern6ndez
The device responsible for carrying out these operations is the syntactic parser which, according to most proposals (e.g., Mitchell, 1994), begins its operations early and quickly, without conscious control on the part of the perceiver. The parser's operations are grounded in strategies that are universal and based on the mental architecture. In a number of models of sentence processing, information from other components may play a role only after the parser has made an initial attachment (see Fern~indez, 2000, for additional discussion). Adjustments of the parser's initial decisions are made in later phases of processing based on considerations having to do with, for example, the prosodic structure eventually assigned to the sentence (Fodor, 1998, 2000), or the discourse saliency and real world plausibility of the constituents under consideration (De Vincenzi & Job, 1993; Frazier & Clifton, 1996; Gilboy, Sopena, Cliftori, & Frazier, 1995).
Early Processes We first consider some of the operations made in the earliest phases by the parser, which theoretically prefers to build the computationally simplest analysis of the string in the input. If the grammatical structure of the input mismatches the structural preferences of the parser, that is, if the parser needs to build a complex tree, then getting the correct interpretation will be difficult, because that structure is computationally complex. One of the principles followed by the parser is Late Closure (Frazier, 1979; Frazier & Fodor, 1978; for reviews, see Frazier & Clifton, 1996; Mitchell, 1994). According to Late Closure, the parser prefers to attach locally, low on the tree. The name "Late Closure" is derived from the original formulation of the principle as one by which the parser keeps open the phrase it is currently working on, preferring to attach inside it. Late Closure is ultimately a principle about recency or locality. Other proposals along the same lines include the principle of Recency Preference of Gibson, Pearlmutter, Canseco-Gonz~lez, and Hickok (1996), and the principle of Right Association of Kimball (1973). Let us examine an example illustrating the operation of Late Closure and its prediction of a preference for local attachments. The sentences below are translation-equivalents in English and Spanish. 2. John told us that Mary left yesterday. 2." Juan nos dijo que Maria se fue ayer. Both sentences in (2) contain a global ambiguity, since yesterday (ayer) could refer to the day when Mary left or to whenever it was that John was speaking about Mary's departure. The locality principle predicts a preference for low attachment of ambiguous constituents, so the initial and faster interpretation of the ambiguous sentence should be the one where yesterday attaches low (when it refers to Mary's leaving). An interpretation with yesterday attaching high is possible, but it is not as immediate or as easy to get as the low attachment alternative.
Relative Clause Attachment in Bilinguals and Monolinguals
193
The effect of the locality preference is also apparent if we consider disambiguated versions of the sentence: 3. a. John will tell us that Mary left yesterday. b. John told us that Mary will leave yesterday. 3. a." Juan nos dini que Maria se rue ayer. b." Juan nos dijo que Maria se ir~i ayer. In (3) the final constituent is forced either to attach locally, inside the lower verb phrase (VP) (3a), or to attach high, to the matrix VP (3b). Intuitively, the (a) versions of the sentences in (3) are easier than the (b) versions, in both languages. The explanation for this intuition lies in the violation of locality in the (b) sentences, where the attachment is forced to the site dispreferred by the parser. To summarize, the operation of Late Closure predicts a preference for local attachments given ambiguous input. Given disambiguated input, if the grammatical structure of the string mismatches the structural preferences of the parser (as in 3b), we expect evidence of difficulty. An Exception to Locality: The Relative Clause Attachment Ambiguity
The principle of Late Closure has been found to apply in a number of languages, with a wide variety of constructions (see discussion in Fern~indez, 2000). This fact might make it an impossible endeavor to pursue the question of language (in)dependency in bilingual sentence processing, since--as argued above--such an investigation requires the existence of cross-linguistic variation with monolinguals. However, there does exist a structure that so far has resisted the generalization that locality applies universally. The construction in question, shown earlier in (1) (repeated below) consists of a complex NP containing two nouns (N1 and N2), and a relative clause (RC). Crucially, N1 and N2 are each a possible host for attaching the relative clause. 1. Andrew had dinner yesterday with the [N~ nephew ] of the [N2 teacher ] [RC that was in the communist party ] 1. Andr6s cen6 ayer con el [Nl sobrino ] del [N2 maestro ] [RC que estaba en el partido comunista ] When instructed to decide between the two alternative interpretations (i.e., when asked Who was in the communist party?), English-speaking participants tend to choose the lower site, N2, teacher, while Spanish-speaking participants tend to choose the higher site, N1, sobrino ('nephew'). This contrast was first reported by Cuetos and Mitchell (1988), and has been replicated in a number of questionnaire studies, testing both monolingual and bilingual speakers of English and Spanish (Dussias, 2001; Ehrlich, Fernandez, Fodor, Stenshoel, & Vinereanu, 1999;
194
Eva M. Ferndndez
Fem~indez, 1998, 1999; Igoa, 1995; Igoa, Carreiras, & Messeguer, 1998; Quinn, in progress; Quinn, Fem~indez, de Almeida, Bradley, & Fodor, 2001; Walter, Clifton, Frazier, Hemforth, Konieczny, & Seelig, 1999). The attachment preferences of languages other than English and Spanish have also been examined (see Fem~indez, 2000, for further discussion and references), and the finding has been that languages fall into either the high attaching (Spanish) or low attaching (English) category. It is important to mention that the language-specific preferences in relative clause attachment are not as strong as one might expect them to be. Questionnaire studies have generally found rates of attachment to either the low or the high site hovering at around 60%, with a great deal of item-based and subject-based variation. It has been shown that manipulations in the target materials, either in the complex NP (Gilboy et al., 1995) or in the relative clause (Fem~indez, 2000; Lovri6, Bradley, & Fodor, 2000; Lovri6 & Fodor, 2000; Quinn, in progress; Quinn et al., 2001; Walter et al., 1999) result in variation in attachment preferences. Variation among participants (in terms of language history or reading ability) also affects their attachment preferences (Dussias, 2001; Fernandez, 1998, 1999; Frenck-Mestre, 1997; Mendelsohn & Perlmutter, 1999). Some explanations of the phenomenon claim that the ultimate preferences yielding language-specific behavior have to do with the initial attachments made by the parser. Mitchell and colleagues (Brysbaert & Mitchell, 1996; Cuetos, Mitchell, & Corley, 1996; Mitchell, 1994; Mitchell & Brysbaert, 1998; Mitchell, Brysbaert, Grondelaers, & Swanepoel, 2000; Mitchell & Cuetos, 1991; Mitchell, Cuetos, Corley, & Brysbaert, 1995) have proposed that the parser tunes its preferences based on the distribution of forced attachments it has encountered in the past. For example, the parser of an individual who exhibits an overall high attachment preference must have encountered more forced high than forced low attachments. This explanation, the Tuning Hypothesis, faces a serious explanatory challenge, given a set of evidence that contradicts its prediction that attachment preferences in perception and forced attachment distributions in production should match. Dutch has been confirmed to be a high attaching language in perceptual studies (Brysbaert & Mitchell, 1996). However, studies of the distribution of forced attachments in several Dutch corpora (both written and spoken) have found a preponderance of forced low attachments (Mitchell & Brysbaert, 1998; Mitchell et al., 2000). A second problem with the Tuning proposal of Mitchell and colleagues (Brysbaert & Mitchell, 1996; Cuetos et al., 1996; Mitchell, 1994; Mitchell & Brysbaert, 1998; Mitchell et al., 2000; Mitchell & Cuetos, 1991; Mitchell et al., 1995) has to do with its prediction that the preferences are due to the early attachments made by the parser. To examine this problem, we must consider some further facts about relative clause attachment preferences. The evidence collected by questionnaire studies has thus far shown that there are high attaching and low attaching languages, with convergence among a large group of researchers on which languages belong to which category. Studies using other methodologies, however, have so far resulted in a number of debates among various researchers. Questionnaire studies are generally understood as tapping late phases of processing. Since participants are not timed, and are allowed to consider their answers for as
Relative Clause Attachment in Bilinguals and Monolinguals
195
long as they need to, such methods are usually referred to as off-line. Theoretically, they reflect the operation of post-syntactic phases of processing, in which the initial decisions may be altered given considerations based on non-structural (discourse, prosody) principles. To examine what happens in the earliest phases of processing, a number of studies have used what are usually referred to as on-line tasks, which include self-paced reading measures and eyetracking measures, which theoretically reflect earlier decisions, thus possibly being more indicative of the parser's preferences. However, there is no guarantee that a given task labeled as being online will unequivocally tap the earliest phases of processing. An important difference between existing off-line and on-line methods (at least those considered here) is their divergent means for determining speakers' preferences. While questionnaire tasks typically query participants directly about the attachment, on-line procedures usually measure the difference in levels of "difficulty" associated with forcing the attachment one way or the other. For example, an on-line task might compare reaction times with items presented in three versions, ambiguous, forced high, or forced low (underlined nouns are compatible with the agreement features of the verb in the relative clause, was): 4. a .... the nephew of the teacher that was... b .... the nephew of the teachers that was... c .... the nephews of the teacher that was...
(ambiguous) (forced high) (forced low)
Speakers' preferences would manifest themselves as increased reading times with materials containing the dispreferred forced attachment, vis-a-vis reading times with either ambiguous materials or materials containing the preferred forced attachment. While a number of on-line studies have found that with languages like Spanish and languages like English, perceivers exhibit cross-linguistic differences in attachment preferences, the on-line data are not as straightforward as one might wish for them to be. Spanish speakers have been shown to prefer high attachment in a number of on-line experiments (Carreiras, 1992; Carreiras & Clifton, 1993, 1999; Mitchell & Cuetos, 1991; Mitchell, Cuetos, & Zagar, 1990). Some recent on-line studies with Spanish speakers, however, have reported no preference (Portolan, personal communication, April 2001) or a preference for low attachment (Carreiras, Betancort, & Messeguer, 2001; Fern~indez, 2000, see below). Similarly, studies using self-paced reading measures with English speakers have found either no reliable attachment preference (Carreiras & Clifton, 1993; Frazier & Clifton, 1996; Henstra, 1996), or a preference for low attachment (Clifton, 1988; Corley, 1995; Deevy, 1999; Frazier & Clifton, 1996); studies in English using eyetracking methodologies have consistently found a low attachment preference (Carreiras & Clifton, 1999; Henstra, 1996). This increasing lack of convergence in experimental findings is probably the result of incomparable methods, all labeled "on-line" but perhaps not uniformly tapping the same phases of processing, or inducing different behavior due to details in segmentation or disambiguation of the materials. We could interpret this to mean that the cross-linguistic differences observed with off-line measures may not be the
196
Eva M. Ferndmdez
result of early decisions made by the parser. Such an interpretation of the available evidence supports explanations claiming that the cross-linguistic differences are driven by the influence of non-syntactic information. According to this second class of proposals, after initial phases of processing (after the parser's routines have applied) there is a departure from low attachment, markedly so in Spanish, but to a lesser extent also in English. This implies that the cross-linguistic differences are not in the parser proper, but rather lie in other modules of the sentence processing machinery. Frazier and Clifton and colleagues (Frazier & Clifton, 1996; Gilboy et al., 1995) have proposed that in early phases of processing, there is no initial attachment made by the parser when the ambiguous constituent is a non-primary phrase (see Fermindez, 2000, for a review). The ultimate attachments are derived from the consideration of discourse principles. High attachment is driven by a preference to attach to the site that is more salient in the discourse, usually the higher noun, N 1, in the complex NP. De Vincenzi and Job (1993, 1995) have also invoked the use of discourse principles to determine ultimate attachment, while maintaining that early decisions are driven by the parser's routines, including Late Closure. Fodor (1998, 2000) has proposed that prosodic considerations influence attachment decisions. An attaching constituent may be more likely to seek a higher host if it is an independent prosodic unit. 3 This idea assumes that even when reading silently, a reader projects a prosodic structure for the input. Sometimes the syntax and the prosody project phrasings that coincide, but when they do not agree, the mismatch could lead to adjustments. To summarize, there exists evidence of differences in the ultimate relative clause attachment preferences of speakers of English (who prefer attachment to the low site) and speakers of Spanish (who prefer attachment to the high site). Explanations of the phenomenon include proposals that the differences originate in the syntactic parser proper. Other explanations attribute the differences to the operation of nonstructural principles (pragmatics, prosody).
Relative Clause Attachment in Bilinguals and Monolinguals: A Three-Way Comparison An examination of the strategies used by bilinguals and monolinguals in the attachment of relative clauses to complex NPs called for a set of experiments to carry out a three-way comparison. First, the experiments needed to test monolinguals and bilinguals, in order to consider the effects of language history and language dominance (the question of language dependency). Second, the participants needed to read materials in English and in Spanish, since those are two languages where the critical cross-linguistic differences had been documented. Finally, the data needed to be collected using at least two different types of tasks with different sensitivities. The goal was to distinguish early decisions, mostly influenced by the parser, from later decisions possibly influenced by non-syntactic components. Therefore, data were collected using a speeded task (a self-paced
Relative Clause Attachment in Bilinguals and Monolinguals
197
reading procedure), to get a sense of readers' initial preferences when they are processing input. Data were also collected using an unspeeded questionnaire, which would be indicative of readers' preferences after the post-syntactic components have had a chance to affect the attachment decision. We now tum to a more detailed description of the participants, materials, and procedures used in this three-way comparison.
Participants Monolingual participants were speakers of Castilian Spanish, tested in Madrid, and speakers of American (US) English, tested in New York City. Bilinguals were fluent speakers of both English and Spanish, and were tested in New York City.
Table I. Number and Mean Ages of Participants as a Function of Language History, Monolingual Speakers of US English (USENG) or Castilian Spanish (CSPA), and Bilinguals Dominant in English (EDOM) or Spanish (SDOM) Based on Their L2 Acquisition History (Simultaneous or Sequential Acquisition)for the Two Experimental Procedures (On-Line Self-Paced Reading and Off-Line Questionnaire) On-Line Self-Paced Reading CSPA
Off-Line Questionnaire
Monolinguals
USENG
N Age
40 25.0
40 19.4
24 23.1
24 19.8
Bilinguals
EDOM
SDOM
EDOM
SDOM
N Age
28 24.1
28 25.5
12 28.3
12 27.0
Simult. Lx & Ly: N (%)
9 (32.1)
2 (7.1)
2 (16.7)
0 (0.0)
L2-Spanish: N (~
5 (17.9)
0 (0.0)
3 (25.0)
1 (8.3)
Age L2
10.6
8.3
7.0
L2-English: N (%)
14 (50.0)
26 (92.9)
7 (58.3)
11 (91.7)
Age L2
4.4
13.0
8.7
13.!
m
USENG
CSPA
198
Eva M. Fern6ndez
Many of the bilinguals were born in the continental United States, but others were born throughout Latin America. Castillian Spanish and Pan-American Spanish do not differ regarding the construction under consideration. There is no experiment comparing the two varieties directly, but pilot data suggest this is the case, and comparing existing experiments also confirms the idea (for further discussion, and a report of the pilot data, see Fern~indez, 2000). Table 1 lists the total number of participants whose data are reported for the four experiments summarized below. Also listed are mean ages of participants in each category: monolingual speakers of English (USENG) and Spanish (CSPA), and Spanish-English bilinguals dominant either in English (EDOM) or Spanish (SDOM). Additionally summarized in Table 1 is the distribution of the bilinguals over three categories of L2 acquisition: simultaneous acquirers, L2-Spanish acquirers, and L2-English acquirers. Overall, the bilinguals tested for these experiments reflect the nature of the New York City bilingual college student population (the participants were predominantly undergraduate students), where Spanish is predominantly the L 1, and where English as a second language is learned relatively early (the average age of L2 acquisition for the sequential acquirers in the sample was 10.3). Based on earlier findings (Fern~indez, 1998, 1999), that language dominance appeared to play a role in determining the attachment preferences of bilinguals, the experiments here were designed to explore the operation of language dominance in detail. To this end, the bilinguals tested were each assigned to a languagedominance group, based on their answers to a series of questions administered in the form of a background (language history) questionnaire. The questions asked the bilinguals explicitly to indicate how well they spoke, understood, wrote, and read in each of their languages. Following a conventional procedure (Grosjean, 1982), language dominance was determined by the differential score between participants' self-reported proficiency in each of their languages. Two secondary sets of criteria for determining language dominance were additionally examined (but are not reported here, given space considerations), both deriving language dominance from bilinguals' preferred language used in special circumstances (e.g., when angry or happy, for basic arithmetic, and so on) or when faced with a life-saving decision to keep one but lose the other of their languages (see similar procedure to determine language dominance used by Cutler, Mehler, Norris, & Segui, 1992). These secondary criteria for establishing language dominance supported the primary criterion, self-reported proficiency differential. Further details are discussed in Fern~.ndez (2000). Table 2 lists overall self-rated proficiency differential scores for the two groups, showing that the two groups do not exhibit entirely symmetrical patterns of dominance. English-dominant bilinguals reported their literacy skills (reading comprehension and written production) to be much stronger in their dominant language, while Spanish-dominant bilinguals appear to have slightly less imbalance in their literacy skills between their dominant and non-dominant languages.
Relative Clause Attachment in Bilinguals and Monolinguals
199
Table 2. Mean Self-Reported Proficiency Differential Scores for English-Dominant (EDOM) and Spanish-Dominant (SDOM) Bilinguals
EDOM (N = 40)
SDOM (N = 40)
Oral Comprehension Oral Production
- 0.54 -0.95
+ 0.55 + 0.88
Reading Comprehension Written production
- 1.00 - 1.33
+ 0.70 +0.83
Self-Rated Proficiency
Note: Scores represent the difference between English and Spanish; a positive figure indicates Spanish is rated better than English, and a negative figure that English is rated better than Spanish.
Great care was taken to include in the sample only participants whose language history profiles matched the requirements for either monolinguality or bilinguality. Additionally, only participants whose performance was within the limits of normal behavior for the two tasks (see note 4) were included in the sample. The bilinguals were tested in each of their two languages on different testing sessions, separated by a minimum of two weeks. To induce bilingual participants into a unilingual mode, the experimenter (a Spanish-English bilingual) addressed them in the language of the test, and provided both oral and written instructions about the procedure in the language of the test. Materials
The entire set of experiments used an identical set of items, which included 24 target sentences and 48 distractor sentences. The 24 target sentences each contained a complex noun phrase like the nephew of the teacher, followed by a relative clause like that was in the communist party. The targets were disambiguated in the on-line task, and ambiguous in the off-line task, the reasons for which will be made clear in the next section. The disambiguation with the on-line materials was forced by using grammatical number agreement. For a given sentence, the verb was always the same, but had an unambiguous number, and could only agree with one of the two nouns in the noun phrase. Half of the items contained singular verbs in the relative clause (was, estaba), the other half, plural verbs (were, estaban) A sample item is provided below, in its six versions (attachment and language), each followed by the corresponding question used in the respective task. 5. a. Andrew had dinner yesterday with the nephew of the teacher that was in the communist party. (ambiguous)
200
Eva M. Ferndmdez
Who was in the communist party?
the nephew
the teacher
b. Andrew had dinner yesterday with the nephew of the teachers that was in the communist party. (forced high) Was the nephew in the communist party? YES/NO c. Andrew had dinner yesterday with the nephews of the teacher that was in the communist party. (forced low) Was the teacher in the communist party? YES/NO 5. a. Andr6s cen6 ayer con el sobrino del maestro que estaba en el partido comunista. (ambiguous) /,Qui6n estaba en el partido comunista? el sobrino el maestro b. Andr6s cen6 ayer con el sobrino de los maestros que estaba en el partido comunista. (forced high) /,Estaba el sobrino en el partido comunista? SI/NO c. Andr6s cen6 ayer con los sobrinos del maestro que estaba en el partido comunista. (forced low) LEstaba el maestro en el partido comunista? SI/NO The relative clauses in the target materials had two versions--one long, like that was in the communist party, and one short, like that was divorced. The length manipulation had to do. with an examination (discussed at length in Fern~.ndez, 2000, but omitted here given space considerations) of the proposal by Fodor (1998) that prosody plays a role in determining attachment decisions (see note 3). The fact that the target materials had lexically different relative clauses in each of their versions had an important advantage regarding the fact that the bilinguals were tested in both of their languages. Bilingual participants, tested in both of their languages, saw the long version of an item in Lx but the short version of the same item in Ly. The differences in the relative clauses served to reduce effects of repetition or episodic memory that might otherwise have arisen. For the cross-linguistic comparison to be justifiable, the materials were created in parallel in the two languages, as translation-equivalent sentences, which felt natural and were felicitous in both languages. (Extensive pre- and post-testing by expert judges of the target materials confirmed that this was indeed the case; see Fem~,ndez, 2000, for more details.) In each test, the 24 targets were interspersed among 48 filler sentences, containing a variety of other constructions, to distract the participants' attention away from the target construction. To reduce repetition effects with the bilinguals, the fillers were different in each of the two versions of the test that each bilingual saw. The monolinguals read the materials in their respective mother tongues, while the bilinguals read materials in both of their languages, but they were tested in each language on separate occasions.
Relative Clause Attachment in Bilinguals and Monolinguals
201
Procedure As noted earlier, these experiments seek to distinguish early versus late attachment decisions, on the assumption that earlier attachments are guided exclusively by the syntactic processor proper (i.e., the parser), while later adjustments may be prompted by the consideration of extra-syntactic information. To examine early processing, an on-line self-paced reading procedure was developed, in which participants were presented disambiguated versions of the materials. Reading times were recorded for the critical segment where the attachment was disambiguated. In addition, an off-line questionnaire was developed, in which participants were presented ambiguous materials on paper, and were asked to answer questions about the attachments, with no time limitations imposed.
On-Line Self-Paced Reading Task Early decisions were tested by analyzing mean reading times during a self-paced reading task. The participants were seated in front of a computer screen on which the materials were presented in two fragments, followed by a comprehension question. Participants began each trial by pressing a footswitch, which prompted the first part of the sentence to appear. When participants pressed a button on a response pad, the second frame replaced the first. In the experimental items, this second flame contained the relative clause in its entirety. An additional button press extinguished the second flame and revealed a YES/NO question, which the participants answered by pressing a green button for YES or a red button for NO. With respect to relative clause attachment preferences, the critical measure was the time it took participants to read the relative clause, in the two different attachment conditions. This measure of attachment is indirect, because the difference in reading times between the two conditions is used as an indicator of attachment preferences. A low attachment preference will manifest itself in the form of faster reading times with materials in which the relative clause is forced to attach low.
Off-Line Questionnaire Task To measure ultimate preferences, participants were tested using a pencil-andpaper questionnaire, in which they were not timed, but where they answered questions asking directly about the ambiguous attachment. The questionnaire test presented the target materials in their ambiguous versions. Each target item was followed by a question about the attachment of the relative clause, and participants were asked to circle what they thought the best answer was. With the targets, both answers were "correct"; an answer like the nephew (N1) for the example in (5) above would reflect a high attachment interpretation of the sentence, while an answer like the teacher (N2) would suggest a low attachment interpretation. The
202
Eva M. Fern6ndez
more frequent choice (N1 or N2) was taken to be indicative of overall attachment preferences. In both the on-line self-paced reading procedure and the off-line questionnaire task, the 24 target items were presented in a pseudo-random order, interspersed among 48 fillers containing a variety of other constructions. The fillers were designed to distract the participants' attention away from the target construction. The fillers played an additional role in that they served to identify inaccurate readers who were excluded from the analyses. 4
Results We begin the summary of the findings by outlining the results with monolingual participants. This will allow us to understand the monolingual model to which we will be comparing the bilingual data. Briefly, we expect monolinguals to have early low attachment preferences, on the assumption that cross-linguistic differences are the result of post-syntactic considerations. In contrast, we expect the standardlyobserved cross-linguistic differences to emerge in the off-line task, where postsyntactic information will have had a chance to play a role in the choices made by the participants. All subsequent analyses are performed by subjects (F~) and by items (F2).
On-Line Reading Times" Monolinguals In the self-paced reading experiment, monolingual participants read disambiguated versions of the target materials. The critical measure was the time it took participants to read the version of the sentences where the relative clause was forced to attach low versus the version where it was forced to attach high. The data are presented in Figure 2 the y axis represents the difference between reading times in the forced low and the forced high attachment conditions. A negative figure indicates faster reading times with forced low attachments, and a positive figure indicates faster reading times with forced high attachments. As Figure 2 illustrates, both groups of speakers had significantly faster reading times when the relative clause was forced to attach to the low site (F1(1,72)= 7.77, p < .01; F2(1,20) - 6.15, p < .05), with no difference in attachment preferences between English and Spanish speakers (the interaction of language and attachment site was not significant, FI and 1:2 < 1). This finding provides support for the proposal that initial choices in relative clause attachment are made by the parser using syntactic information exclusively. The lack of cross-linguistic differences in the behavior of monolingual English and monolingual Spanish speakers in this experiment narrows down the types of explanations that can be put forth regarding the origins of relative clause attachment strategies.
Relative Clause Attachment in Bilinguals and Monolinguals
=
15o
203
-
omu
@
@
.~ ~ 0
100-
~. ~
50-
E-~'
0
=~ ,~ .~
-50
oM
=o
0
-100 1
-97
-150 USENG
CSPA
Figure 2. Mean difference between relative clause reading times (ms) in forced low and forced high attachment conditions, as a function of language group (English (USENG) and Spanish (CSPA) monolinguals).
Off-Line Preferences: Monolinguals In the questionnaire task participants answered questions asking directly about the attachment. Responses were tallied as mean N1 choices, which are plotted in Figure 3. The analysis of the data indicated that the difference between N1 choices made by English monolinguals and N I choices made by Spanish monolinguals was reliable (main effect of Language, F1(1,44)= 5.48, p < .025; F2(1,10)= 56.05, p < .001). This experiment thus replicates the standard finding that Spanish speakers are more likely than English speakers to choose the high site when attaching a syntactically ambiguous relative clause to a complex NP.
204
Eva M. Fern6ndez
100 -
r~
75-
o j,,,~
57%
0
r~ Z
50-
43%
25-
USENG
CSPA
Figure 3. Mean percentage high attachment (N1 choices) as a function of language group (English (USENG) and Spanish (CSPA) monolinguals).
O n - L i n e R e a d i n g Times" Bilinguals
The monolingual results are critical in terms of determining how to proceed in examining the bilingual data. First, we expect to find a preference for low attachment in the early phases, since the language (in)dependency question cannot be examined there (given the uniformity in the behavior patterns of the monolinguals). Secondly, we will consider the language (in)dependency question with the data from the questionnaire task, where the monolinguals were found to differ, the likelihood for Spanish speakers to attach high being stronger than for English speakers. We turn first to the on-line bilingual data, which are plotted in Figure 4. The analyses performed on the data indicated that the effect of site was unreliable in both the subject- and item-based analyses (FI and F2 < 1), and did not interact with other factors, including language of the materials and language dominance (all values of p >. 15). That there was no effect of site found in this bilingual version of the on-line task could not mean that bilinguals do not use parsing strategies, like Late Closure, when processing linguistic input. A more viable explanation is that we have here a task
Relative Clause Attachment in Bilinguals and Monolinguals
205
which has different sensitivities with different types of readers. The difference between the readers essentially has to do with overall speed, decreased in the bilingual group because of the performance deficit experienced by bilinguals with certain types of linguistic tasks. (There is no simple solution to dealing with this difference, other than to devise a task less reliant to stimuli presented exclusively in written form. Auditory stimuli, with or without visual support in the form of images depicting the sentences, might overcome the differences in performance observed here between monolinguals and bilinguals. We return to this in the Concluding Remarks).
1EDOM E] SDOM r~
.~
150 1
0 .~ ~o
100
113
O
50
~
11
"-~
~N .~
21
o
-50
"~
-64
~-100 ~t
*~
-150 English
Spanish
Figure 4. Mean difference between relative clause reading times (ms) in forced low
and forced high attachment conditions, as a function of language history group (English-dominant (EDOM) and Spanish-dominant (SDOM) bilinguals) and language of the materials (English and Spanish).
Figure 5 plots overall reading times for the relative clause, for both monolingual and bilingual participants. On average, monolingual speakers of both languages had reading times of around 2500 milliseconds (ms), with no difference in overall reading times between the two language groups, FI < 1;/=2(1,20) = 2.56, p > . 10.
206
Eva M. Ferndndez
The figure illustrates the fact that the bilinguals were between 200 and 500 ms slower, on average, than the monolinguals. (No direct statistical comparison will be made of the monolingual and the bilingual data, given the uneven numbers in the samples--40 x 2 monolinguals versus 28 x 2 bilinguals. The fact that bilinguals participated in the experiment twice, once in each language, while monolinguals participated only once, also makes direct comparisons inappropriate.)
~
3300 -
3100-
0
_~
2900~ 2700"
USENG, CSPA
-"0"- SDOM -0-
EDOM
0"
2500r~
[-. 2 3 0 0 .o 2100~
1900 -
~
17001500 English
Spanish
Figure 5. Mean relative clause reading times, for English (USENG) and Spanish (CSPA) monolinguals and English-dominant (EDOM) and Spanish-dominant (SDOM) bilinguals, for materials in English and Spanish. The data are averaged over attachment site.
As can be seen, the performance deficit is overall associated with the minority language of the community, namely, Spanish, in addition to being associated with the bilinguals' non-dominant language. English-dominant bilinguals took longer to read Spanish than they did to read English (cf. the rising slope in the line corresponding to the English-dominant data). However, Spanish-dominant bilinguals, rather than exhibiting a straightforward mirror-image effect (i.e., English much slower than Spanish, which would be indicated by a sharply falling slope),
Relative Clause Attachment in Bilinguals and Monolinguals
207
experienced longer reading times than their monolingual counterparts in both their non-dominant and their dominant language, as shown by the fiat horizontal line. The analysis of the bilinguals' reading times for the relative clause indicated an interaction of language of the materials and language dominance, which in subanalyses by language of the materials showed no difference with Spanish materials (F/and F2 < 1), and a reliable difference with English materials, F1(1,48) = 3.95, p < .05; F2(1,20) = 43.80, p < .001. In the omnibus ANOVA of reading times for the relative clause, the two factors of language of the materials and language dominance engaged in a three-way interaction with relative clause length, F1(1,96) = 10.84, p < .001; F2(1,20) = 15.18, p < .001. (See discussion of the length manipulation in the Materials section, above. A detailed analysis and discussion of this interaction is provided in Fern~indez, 2000.) These findings are not at all surprising. Spanish-dominant bilinguals living in New York City have a reduced opportunity to maintain their literacy skills in their dominant language, since the written material they usually come in contact with is in English. Likewise, English-dominant bilinguals are very unlikely to have sufficient opportunity to develop reading (and writing) skills in their non-dominant language, for similar reasons. (Incidentally, these patterns correspond to the bilinguals' self-reported literacy--see Table 2 above--where there was an asymmetry in terms of the degree of imbalance, greater in the English-dominant group.) To make sense of the null findings with respect to the bilinguals' attachment preferences in the on-line task, we have noted that the procedure must have missed the opportunity of capturing the low attachment when it happened. The increase in reading times, most likely attributable to the additional time needed by bilinguals to decode the lexical material, allowed more time for the syntactic processor, running in the background, to make its initial attachment and pass its output on to the postsyntactic modules for adjustments based on non-syntactic information. Confirmation of the validity of this interpretation of the facts will only be possible once similar groups of bilinguals and monolinguals have been tested using a procedure that minimizes the performance deficit on the part of the bilinguals. Currently underway is a preliminary study using linguistic stimuli with minimal lexical difficulty, supported by visual contextual cues. Off-Line Preferences: Bilinguals We finally turn to the ultimate preferences exhibited by the bilinguals in the offline questionnaire task, where participants indicated directly which site they preferred for attachment of the relative clause. Having found cross-linguistic differences in this task with monolingual participants, we are licensed to explore the question of language (in)dependency in this bilingual version of the task. The results are plotted in Figure 6.
208
Eva M. Ferndndez
II EDOM 100 -
r~
17 SDOM
73%
75-
73%
e~
0
55%
0 Z
50-
a,
25
48%
English
Spanish
Figure 6. Mean percentage high attachment (N1 choices) as a function of language history g r o u p (English-dominant (EDOM) and Spanish-dominant (SDOM) bilinguals) and language of the materials (English and Spanish).
The critical finding in these bilingual off-line data was an unmodified effect of language dominance. The Spanish-dominant group was more likely to interpret the relative clause as attaching to the high site than the English-dominant group. This effect was significant, and did not interact with other factors (main effect of dominance, F1(1,40) = 9.04, p < .005; F2(1,20) = 59.36, p < .001; interaction of dominance and language, Ft < 1; F2(1,20) = 1.39, p > .25). This provides evidence of language independent processing strategies in bilingual sentence processing. Regardless of the language of the input, a bilingual exhibits attachment preferences that resemble those employed by monolingual speakers of the bilingual's dominant language.
L a n g u a g e Independency in Bilingual Sentence Processing: 1 + 1 - 1
We are now in a position to answer the question regarding language (in)dependency posed at the beginning of this discussion" are bilinguals like two monolinguals in one? The short answer is that they are not, that is, that bilinguals
Relative Clause Attachment in Bilinguals and Monolinguals
209
diverge from the monolingual model in their behavior in on-line and off-line tasks. However, this statement needs further clarification. First, rather than cross-linguistic differences in early monolingual processing, the data indicated that speakers of English or Spanish prefer low attachments in early phases of processing, as evidenced by their faster reading of relative clauses forced to attach low, in the on-line self-paced reading task. Given this lack of crosslinguistic differences among monolinguals, the language dependency question was inapplicable for the bilinguals. In the same self-paced reading task, bilinguals were slower readers than monolinguals (which was expected), but they did not reliably attach low (contrary to expectations). However, it was argued that to explain this null result we must consider the possibility that the experimental procedure taps different phases of processing with readers who have different reading profiles. There is no guarantee that a task is equally sensitive to processing with all types of readers. The lexical material might have slowed the bilinguals down in terms of when they pressed the button to extinguish the flame where the relative clause was presented, while the attachment routines continued running in the background. Thus the bilinguals' reading times might reflect processing not exclusively syntactic in nature (for a similar proposal, though based on somewhat different evidence, see Pynte, 1998; Pynte& Colonna, 2000). Where we did see cross-linguistic differences in the monolingual data was in the off-line questionnaire study. Spanish speakers tended to attach higher than English speakers. With the bilingual data, we witnessed an effect of differences in attachment preferences, not within the two dominance groups, but rather between the two dominance groups, which we took to be evidence of language independent processing. Bilinguals process linguistic input using the same type of strategies for both of their languages, and the set they use is the one associated with their dominant language. Independently of the language of the input, bilinguals have ultimate preferences in both of their languages that resemble those of monolingual speakers of their dominant language.
Concluding Remarks Regarding the question about relative clause attachment preferences raised at the beginning of this discussion, the evidence presented in this paper indicates that the observed cross-linguistic differences are not the result of language-specific processing within the parser proper. Instead, the data show that the departure from the universal locality preference comes from processing associated with postsyntactic components. In particular, monolingual speakers of English and Spanish exhibited uniform behavior in an on-line measure of their attachment preferences. Cross-linguistic differences were only evident when the task was off-line, both with monolingual and bilingual speakers. This narrows down the types of possible explanations that may be offered to account for the phenomenon of cross-linguistic differences in attachment preferences. In particular, these experiments support
210
Eva M. Fern6ndez
proposals in which the language-specific component is not in the parser proper, but has to do with post-syntactic considerations. Regarding sentence processing in bilinguals, we examined evidence of language independent processing, with language dominance playing a critical role. Bilinguals exhibit ultimate attachment preferences that are similar in both of their languages, and that resemble those of monolingual speakers of their dominant language. The sentence processing machinery in bilinguals thus should be understood as using one set of routines, not two, no matter what the language of the input may be. The crosslinguistic differences in relative clause attachment that appear here as unmodified dominance effects are apparently driven by a characteristic of the perceivers' processing routines, rather than by some property of the stimulus language. This final observation further constrains the set of possible theories to account for the phenomenon of cross-linguistic differences in relative clause attachment. A great deal of work remains to be done to clarify some of the claims made in the interpretation of the experimental findings reported in this paper. To begin with, empirical measures need to be devised that reliably tap the earliest phases of processing, regardless of variation in the backgrounds of participants or in the linguistic manipulations of the materials. This will require designing experiments where the stimuli are presented aurally or as images rather than as written sentences. Such procedures with auditory or visual stimuli will at the same time address the problem of evaluating the performance of bilinguals vis-a-vis that of their monolingual counterparts, possibly eliminating or at least reducing the bilingual performance deficit. Future investigations of bilingual sentence processing must also consider the origins of language dominance to determine whether language dominance is driven primarily by frequency of use, by education history, or by other psychological or sociological factors. Such work would allow for a better understanding of shifts in language dominance, as experienced by many immigrants to the United States (see, e.g., Heredia & Altarriba, 2001). In the context of shifting patterns of dominance, it is crucial to determine whether the sentence processing machinery also shifts its language independent preferences (from Sx to Sy) as Ly (here, L2) becomes dominant and Lx (L 1) is non-dominant or becomes attrited.
Notes
~This research was carried out as part of the author's doctoral dissertation, and is reported in detail in Fernfindez (2000). 2 Throughout, Lx and Ly refer to the bilingual's two languages, without reference to which one was learned first. The terms L1 and L2 are used to refer to the bilingual's two languages when order of acquisition is critical. 3A constituent is prosodically independent depending on its prosodic weight. A very short relative clause, like that was divorced, has very little prosodic weight, and is less likely to be its own prosodic constituent. Such a phrase would preferably be attached in a constituent where it could belong to a larger prosodic unit. A longer
Relative Clause Attachment in Bilinguals and Monolinguals
211
relative clause, like that was in the communist party, has a greater likelihood of becoming its own prosodic constituent, and may be freer to seek a higher host. The initial preference might be for local attachment, but considerations based on the prosodic structure eventually assigned to the relative, and to the rest of the sentence, might lead to adjustments (e.g., Fern~indez, 2000; Lovri6 & Fodor, 2000; Lovri6 et al., 2000; Quinn, in progress; Quinn et al., 2001). 4 The data from monolingual participants who made over 20% errors in the online task, 5% errors in the off-line task, were discarded and replaced. The error-rate cut-offs for bilingual participants' average error rate in both languages were set at 30% and 15% for the on-line and off-line task, respectively. For both tasks, the errors allowed for bilinguals were 10 percentage points more than for the monolinguals. This extra leeway provided for the bilinguals was necessary given the expectation that the performance deficit would decrease accuracy in these tasks, particularly in the bilinguals' non-dominant language. (Having maintained the same error-rate cut-offs for bilinguals as for monolinguals would have resulted in a biased sample due to the elimination of bilingual participants representative of the bilingual population.)
References
Altenberg, E. P., & Cairns, H. S. (1983). The effects of phonotactic constraints on lexical processing in bilingual and monolingual subjects. Journal of Verbal Learning and Verbal Behavior, 22, 174-188. Brysbaert, M., & Mitchell, D. C. (1996). Modifier attachment in sentence parsing: Evidence from Dutch. Quarterly Journal of Experimental Psychology, 4914, 664-695. Carreiras, M. (1992). Estrategias de an~ilisis sint~ctico en el procesamiento de frases: Cierre temprano versus cierre tardio. Cognitiva, 4, 3-27. Carreiras, M., Betancort, M., & Messeguer, E. (2001). Relative clause attachment in Spanish: Do readers use different strategies when disambiguating by gender and number? Poster presented at the 14th Annual CUNY Conference on Human Sentence Processing, Philadelphia, PA. Carreiras, M., & Clifton, C. (1993). Relative clause interpretation preferences in Spanish and English. Language and Speech, 36, 353-372. Carreiras, M., & Clifton, C. (1999). Another word on parsing relative clauses: Eyetracking evidence from Spanish and English. Memory & Cognition, 27, 826833. Clifton, C. (1988, August). Restrictions on late closure." Appearance and reality. Paper presented at the 6 th Australian Language and Speech Conference, Sydney, Australia. Cook, V. (1997). The consequences of bilingualism for cognitive processing. In A. M. B. de Groot & J. F. Kroll (Eds.), Tutorials in bilingualism." Psycholinguistic perspectives (pp. 279-300). Mahwah, NJ: Lawrence Erlbaum.
212
Eva M. Fern6ndez
Corley, M. (1995). The role of statistics in human sentence processing. Unpublished doctoral dissertation, University of Exeter. Cuetos, F., & Mitchell, D. C. (1988). Cross-linguistic differences in parsing: Restrictions on the use of the late closure strategy in Spanish. Cognition, 30, 73105. Cuetos, F., Mitchell, D. C., & Corley, M. M. B. (1996). Parsing in different languages. In M. Carreiras, J. E. Garcia-Albea, & N. Sebasti~n-Gall6s (Eds.), Language processing in Spanish (pp. 145-187). Mahwah, NJ: Lawrence Erlbaum. Cutler, A., Mehler, J., Norris, D., & Segui, J. (1992). The monolingual nature of speech segmentation by bilinguals. Cognitive Psychology, 24, 381-410. de Bot, K. (1992), A bilingual production model: Levelt's speaking model adapted. Applied Linguistics, 13, 1-24. Deevy, P. L. (1999). The comprehension of English subject-verb agreement. Unpublished Doctoral Dissertation, University of Massachusetts, Amherst, MA. De Vincenzi, M., & Job, R. (1993). Some observations on the universality of the late-closure strategy. Journal of Psycholinguistic Research, 22, 189-206. De Vincenzi, M., & Job, R. (1995). An inverstigation of late closure: The role of syntax, thematic structure, and pragmatics in initial and final interpretation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1303-1321. Dufour, R. (1997). Sign language and bilingualism: Modality implications. In A. M. B. de Groot & J. F. Kroll (Eds.), Tutorials in bilingualism: Psycholinguistic perspectives (pp. 301-330). Mahwah, NJ: Lawrence Erlbaum. Dussias, P. E. (2001). Sentence parsing in fluent Spanish-English bilinguals. In J. L. Nicol (Ed.), One mind, two languages: Bilingual language processing (pp. 159-176). Oxford, UK: Blackwell. Ehrlich, K., Fern~indez, E. M., Fodor, J. D., Stenshoel, E., & Vinereanu, M. (1999). Low attachment of relative clauses: New data from Swedish, Norwegian and Romanian. Poster presented at the 12th Annual CUNY Conference on Human Sentence Processing, New York, Nu Ellis, R. (1994). The study of second language acquisition. Oxford, UK: Oxford University Press. Fern~indez, E. M. (1998). Language dependency in parsing: Evidence from monolingual and bilingual processing. Psychologica Belgica, 38, 197-230. Femb.ndez, E. M. (1999). Processing strategies in second language acquisition: Some preliminary results. In E. C. Klein & G. Martohardjono (Eds.), The development of second language grammars: A generative approach (pp. 217-239). Amsterdam, NL: John Benjamins. Fernandez, E. M. (2000). Bilingual sentence processing: Relative clause attachment in English and Spanish. Unpublished doctoral dissertation, CUNY Graduate Center, New York, NY. Fodor, J. D. (1998). Learning to parse? Journal of Psycholinguistic Research, 27, 285-319. Fodor, J. D. (2000). Silent prosody? Paper presented at the COSWL Symposium, Linguistic Society of America, Chicago, IL.
Relative Clause Attachment in Bilinguals and Monolinguals
213
Francis, W. S. (1999). Analogical transfer of problem solutions within and between languages in Spanish-English bilinguals. Journal of Memory and Language, 40, 301-329. Frazier, L. (1979). On comprehending sentences." Syntactic parsing strategies. Unpublished doctoral dissertation, University of Connecticut. Frazier, L., & Clifton, C. (1996). Construal. Cambridge, MA: MIT Press. Frazier, L., & Fodor, J. D. (1978). The sausage machine: A new two-stage parsing model. Cognition, 6, 291-325. Frenck-Mestre, C. (1997). Examining second language reading: An on-line look. In A. Sorace, C. Heycock, & R. Shillcok (Eds.), Proceedings of the GALA 1997 conference on language acquisition (pp. 474-478). Edinburgh, UK: Human Communications Research Center. Frenck-Mestre, C., &Pynte, J. (1997). Syntactic ambiguity resolution while reading in second and native languages. Quarterly Journal of Experimental Psychology, 50A, 119-148. Gibson, E., Pearlmutter, N., Canseco-Gonz~lez, E., & Hickok, G. (1996). Recency preference in the human sentence processing mechanism. Cognition, 59, 23-59. Gilboy, E., Sopena, J. M., Clifton, C., & Frazier, L. (1995). Argument structure and preferences in the processing of Spanish and English complex NPs. Cognition, 54, 131-167. Green, D. W. (1998). Bilingualism and thought. Psychologica Belgica, 38, 251278. Grosjean, F. (1982). Life with two languages: An introduction to bilingualism. Cambridge, MA: Harvard University Press. Grosjean, F. (1997). Processing mixed language: Issues, findings, and models. In A. M. B. de Groot & J. F. Kroll (Eds.), Tutorials in bilingualism." Psycholinguistic perspectives (pp. 225-254). Mahwah, NJ: Lawrence Erlbaum. Grosjean, F. (1998). Studying bilinguals: Methodological and conceptual issues. Bilingualism." Language and Cognition, 1, 131-149. Henstra, J.-A. (1996). On the parsing of syntactically ambiguous sentences. Coordination and relative clause attachment. Unpublished doctoral dissertation, University of Sussex, UK. Heredia, R. R., & Altarriba, J. (2001). Bilingual language mixing: Why do bilinguals code-switch? Current Directions in Psychological Science, 10, 164-168. Igoa, J. M. (1995). Parsing decisions and the construal hypothesis." Attachment preferences in primary phrases in Spanish. Paper presented at the Second Symposium on Psycholinguistics, Tarragona, Spain. Igoa, J. M., Carreiras, M., & Meseguer, E. (1998). A study on late closure in Spanish: Principle-grounded vs. frequency-based accounts of attachment preferences. Quarterly Journal of Experimental Psychology, 51A, 561-592. Kimball, J. (1973). Seven principles of surface structure parsing in natural language. Cognition, 2, 15-47o Kroll, J. F., & de Groot, A. M. B. (1997). Lexical and conceptual memory in the bilingual: Mapping form to meaning in two languages. In A. M. B. de Groot & J. F.
214
Eva M. Fern6ndez
Kroll (Eds.), Tutorials in bilingualism: Psycholinguistic perspectives (pp. 169-200). Mahwah, NJ: Lawrence Erlbaum. Lovri6, N., Bradley, D., & Fodor, J. D. (2000). RC attachment in Croatian with and without preposition. Poster presented at AMLaP (Architectures and Mechanisms of Language Processing), Leiden, NL. Lovri6, N., & Fodor, J. D. (2000). Relative clause attachment in sentence parsing: Evidence from Croatian. Poster presented at the 13th Annual CUNY Conference on Human Sentence Processing, La Jolla, CA. Malakoff, M., & Hakuta, K. (1991). Translation skill and metalinguistic awareness in bilinguals. In E. Bialystok (Ed.), Language processing in bilingual children (pp. 141-166). Cambridge, UK: Cambridge University Press. Mendelsohn, A., & Perlmutter, N. (1999). Individual differences in relative clause attachment preferences. Poster presented at the 12th Annual CUNY Conference on Human Sentence Processing, New York, NY. Mitchell, D. C. (1994). Sentence parsing. In M. Gemsbacher (Ed.), Handbook of psycholinguistics (pp. 375-409). New York: Academic Press. Mitchell, D. C., & Brysbaert, M. (1998). Challenges to recent theories of cross linguistic variation in parsing: Evidence from Dutch. In D. Hillert (Ed.), Sentence processing: A cross linguistic perspective (pp. 313-335). San Diego, CA: Academic Press. Mitchell, D. C., Brysbaert, M., Grondelaers, S., & Swanepoel, S. (2000). Modifier attachment in Dutch: Testing aspects of construal theory. In A. Kennedy, R. Radach, D. Heller, & J. Pynte (Eds.), Reading as a perceptual process (pp. 493516). Oxford, UK: Elsevier. Mitchell, D. C., & Cuetos, F. (1991). The origins of parsing strategies. In C. Smith (Ed.), Current issues in natural language processing (pp. 1-12). Center for Cognitive Science, University of Austin, TX. Mitchell, D. C., Cuetos, F., Corley, M. M. B., & Brysbaert, M. (1995). Exposure-based models of human parsing: Evidence for the use of coarse-grained (non-lexical) statistical records. Journal of Psycholinguistic Research, 24, 469-488. Mitchell, D. C., Cuetos, F., &Zagar, D. (1990). Reading in different languages: Is there a universal mechanism for parsing sentences? In D. Balota, G. B. Flores d'Arcais, & K. Rayner (Eds.), Comprehension processes in reading (pp. 285-302). Hillsdale, NJ: Lawrence Erlbaum. Myers-Scotton, C. (1993). Duelling languages: Grammatical structure in codeswitching. Oxford, UK: Clarendon Press. Nicol, J. L., Teller, M., & Greth, D. (2001). Production of verb agreement in monolingual, bilingual, and second-language speakers. In J. L. Nicol (Ed.), One mind, two languages: Bilingual language processing (pp. 117-133). Oxford, UK: Blackwell Publishers. No~l, M.-P., & Fias, W. (1998). Bilingualism and numeric cognition. Psychologica Belgica, 38, 231-250. Poulisse, N. (1997). Language production in bilinguals. In A. M. B. de Groot & J. F. Kroll (Eds.), Tutorials in bilingualism: Psycholinguistic perspectives (pp. 201224). Mahwah, NJ: Lawrence Erlbaum.
Relative Clause Attachment in Bilinguals and Monolinguals
215
Pynte, J. (1998). The time course of attachment decisions: Evidence from French. In D. Hillert (Ed.), Sentence processing: A cross-linguistic perspective. syntax and semantics, Volume 31 (pp. 227-245). San Diego, CA: Academic Press. Pynte, J., & Colonna, S. (2000). Decoupling syntactic parsing from visual inspection: The case of relative clause attachment in French. In A. Kennedy, R. Radach, D. Heller, & J. Pynte (Eds.), Reading as a perceptual process (pp. 529547). Oxford, UK: Elsevier. Quinn, D. (in progress). The influence of implicit prosody on syntactic parsing preferences: Relative clause attachment in French and English. CUNY Graduate Center, New York, NY. Quinn, D., Fem~ndez, E. M., de Almeida, R. G., Bradley, D., & Fodor, J. D. (2001). Prosodic phrasing predicts RC attachment in French and English silent reading. Poster presented at AMLaP (Architectures and Mechanisms of Language Processing) 2001, SaarbrUcken, Germany. Ritchie, W., & Bhatia, T. K. (1998). Codeswitching, grammar, and sentence production: The problem of light verbs. In E. C. Klein & G. Martohardjono (Eds.), The development of second language grammars: A generative approach (pp. 269287). Amsterdam: John Benjamins Publishing Company. Walter, M., Clifton, C., Frazier, L., Hemforth, B., Konieczny, L., & Seelig, H. (1999). Prosodic and syntactic effects on relative clause attachment in German and English. Poster presented at AMLaP (Architectures and Mechanisms of Language Processing) 1999, Edinburgh, UK.
This Page Intentionally Left Blank
Bilingual Sentence Processing- R.R. Herediaand J. Altarriba (Editors) 9 2002 Elsevier Science B.V. All rights reserved.
An On-Line Look at Sentence Processing in the Second Language Cheryl Frenck-Mestre Centre National de Recherche Scientifique (CNRS) Laboratoire Parole et Langage Universit~ de Provence
Abstract
What are the factors that affect immediate syntactic processing in the second language of bilinguals? How might processing evolve with increasing skill in the second language? Herein, I will review these questions in the light of data from several on-line investigations of syntactic processing in the native and second language of bilinguals with varying levels of second language proficiency. Performance is examined, moreover, for bilinguals of different language backgrounds. The data from these experiments highlight the role of second language exposure, in line with models based upon linguistic experience. The present chapter is concerned with the acquisition of syntax in the second language by adult learners, a topic that has received considerable attention and fuelled many a debate. I will review the handful of recent on-line psycholinguistic studies of second language syntactic processing in adult, late bilinguals, in order to pinpoint the factors that influence immediate syntactic analysis. None would dispute the claim that the final comprehension of a sentence is affected by a myriad of factors (at the lexical, syntactic, discourse, and pragmatic levels). Our aim here, however, is to provide a clear picture of the factors that play an immediate role in second language syntactic analysis. We will examine the results obtained in studies of adult bilinguals at different levels of second-language proficiency, as well as different language pairs, to establish the role of exposure to the second language in acquiring L2 syntax and the influence of"forward transfer." As we will show, there is substantial evidence, first, that "exposure-based" models such as that proposed by Mitchell and colleagues (Cuetos, Mitchell, & Corley, 1996; Mitchell, 1994) may provide a viable framework for understanding how adults acquire the syntax of a second language, and second, that many common processes underlie native and second-language syntactic processing.
218
Cheryl Frenck-Mestre
Why is Second-Language Reading Slow? Quite often, on-line measurements obtained during the reading of single sentences reveal increased processing time for non-native readers (cf. Fem~indez, 2000; Frenck-Mestre, 1999; Frenck-Mestre & Pynte, 1997; Hoover & Dwivedi, 1998; Segalowitz & H6bert, 1990). The underlying cause for this relative slowness is not immediately apparent, given that it is not necessarily linked to a lack of knowledge at the lexical and/or syntactic level. It has been suggested that a lesser automatization of lower-level processes may be responsible. For example, a lesser ability to encode orthographic redundancies may slow lexical access in the second language (Favreau & Segalowitz, 1983; but see Frenck-Mestre, 1993). This would not only incur a general slowing but also a more specific difficulty, related to phonological activation. That is, given what we know about the time course of orthographic and phonological activation (see, for example, Ferrand & Grainger, 1993; Frost, Katz, & Bentin, 1987; Paap, McDonald, Schvaneveldt, & Noel, 1987), slower orthographic processing would allow for greater phonological activation and hence possible interference when accessing homophones (e.g., "fare" vs. "fair" and "great" vs. "grate," in English). Indeed, just such a result was reported by Segalowitz and H6bert (1990); bilinguals who read slower in their second than native language were more prone to interference from homophones when processing semantically anomalous sentences (e.g., "The weather was fare"), than were bilinguals who read equally fast in their two languages. While second language reading may indeed suffer from a lesser degree of automaticity as concerns such encoding processes, experiments performed in our laboratory provide evidence that other factors may also be at play. If one records the eye movements of (non-native or native) readers, it is possible to distinguish the "first pass" through the sentence from eventual re-readings, whether of the entire sentence or parts thereof (see Rayner & Pollatsek, 1987; Rayner, Sereno, Morris, Schmauder, & Clifton, 1989, for discussions of eyemovement recording and its use in psycholinguistic studies). This tool has enabled our research group to show that at a gross level, increased reading times for proficient non-native readers is often linked to a tendency to re-read sentences. This has been observed informally, in numerous experiments involving diverse syntactic structures, and reported in various talks and published studies (Frenck-Mestre, 1998; Frenck-Mestre &Pynte, 1995, 1997). It is worth restating that this general tendency to reread sentences is observed for proficient bilinguals, that is, in spite of a high level of lexical and syntactic knowledge in the second language. A different picture arises when we consider the oculomotor behavior of these bilinguals during the "first pass" through second language sentences (i.e., in the course of the first left-to-right pass through it). For first pass measures, advanced non-native readers and native readers do not differ notably in their performance at least insofar as we have observed (Frenck-Mestre, 1998; Frenck-Mestre & Pynte, 1995, 1997). That is, we have not found the second language performance of these speakers to differ notably from native performance as concerns the average length of a saccade, the mean duration of fixations, or the probability of making a regressive saccade during
On-Line Syntactic Processing
219
the first reading of the sentence. It is perhaps important to underline the differences, here, between the information provided by eye-movement recording as opposed to self-paced reading. While the latter technique allows one to track on-line processing, it does not allow one to distinguish between "first" vs. "second" readings of parts of the sentence, it may be that "rereading" of the segments presented in self-paced reading studies, rather than initial processing as concerns either lexical access or syntactic analysis, are also the reason behind slower reaction times reported in these studies. How then should one interpret "slow" non-native reading? Is it perhaps linked after all to differences and/or "deficits" in syntactic processing as compared to native readers' processing? Indeed, one could hypothesize that, despite their second language proficiency, even skilled bilinguals do not perform a full syntactic analysis during the first pass through the sentence, thus rendering necessary a second pass through it. As we report in greater detail later in this chapter, however, this hypothesis is untenable in the face of other on-line results. Indeed, proficient second-language readers demonstrate an immediate sensitivity to the same factors that influence native readers' initial progression through the sentence. Both structural ambiguity and lexical constraints influence the first pass reading times of proficient non-native readers, just as they do those of native readers (Frenck-Mestre, 1998; Frenck-Mestre & Pynte, 1997). It can be noted that the same immediate effect of syntactic ambiguity on second language processing has been reported using selfpaced reading (Hoover & Dwivedi, 1998; Juffs & Harrington, 1996). Hence, slower second language reading times, at least in skilled bilinguals, do not appear to be linked to differences in syntactic processing as compared to native readers (see also Segalowitz, 1986; Segalowitz & H6bert, 1990), as will be discussed shortly. Our own hypothesis, indeed which remains to be tested through systematic study, is that the tendency to re-read may be a "residual" of early difficulties experienced in second language reading. That is, at an early stage, re-reading in the second language may well be necessary due to lesser abilities (as concerns orthographic encoding, lexical access, and/or syntactic analysis). Later on, the more skilled bilingual may retain this "developed habit" of re-reading, despite it no longer being necessary. Although this remains to be demonstrated, it is in line with the results reported below showing similar syntactic processing for native speakers and skilled bilinguals in their second language, despite longer overall reading times in the latter.
Slower, but not Necessarily Different Processing in the Second Language Consider, first, three recent studies that examined the on-line processing of syntactically ambiguous sentences, in both native and second language readers. In two of these studies, the performance of proficient English-French bilinguals revealed immediate sensitivity to syntactic ambiguity quite independently of their overall reading speed. In the third, however, the pattern of results from ChineseEnglish second language (ESL) learners, who were both less proficient than the
220
Cheryl Frenck-Mestre
bilinguals who participated in the other two studies and quite slow as compared to native speakers, is more questionable as concerns their sensitivity to second language syntactic ambiguity. Hoover and Dwivedi (1998) looked at the on-line processing of a syntactic ambiguity that is present in romance languages, in the current case French, but absent in English. The ambiguity is related to the role of the clitic pronoun (highlighted in the example) in French causative sentences, as illustrated in (1 a) and (1 b), below. (1 a) Sarah le fera signer en pr6sence de son avocat (Sara it will have signed in the presence of her lawyer) (1 b) Sarah le fera demain en pr6sence de son avocat. (Sara it will do tomorrow in the presence of her lawyer) In both (la) and (lb), a full direct object noun phrase (NP) has been pronominalised and replaced by the clitic pronoun "le." However, in the first example (1 a), the clitic is ambiguous. This is because the verb "faire" can either be used as a thematic verb, as is the case in lb, or in a causative construction as illustrated in la. Therefore, upon reading the clitic in (1 a), the reader could initially treat it as an argument of the proceeding verb "faire" only to find this an erroneous assignment upon the reading of the second verb (in the example, "signer"). In (1 b), no ambiguity is present, as the clitic pronoun can only be assigned as an argument of the verb "faire." This ambiguity is known to cause difficulty for native French speakers, as was demonstrated experimentally in a self-paced reading study by Dwivedi and Hoover (1996, reported in Hoover & Dwivedi, 1998). In a subsequent self-paced reading study, Hoover and Dwivedi (1998) looked at the processing of this ambiguity for two groups of highly fluent L2 French readers (LI English). While both groups rated quite high as concerns level of proficiency in French, one group was classified as "fast" readers on a reading test and the other comparatively "slow." The main objective of the study was to determine whether sensitivity to the ambiguity present in French clitic/causative constructions would be correlated with reading speed in the second language. That is, whether "slow" L2 readers would be less sensitive to structural ambiguity than "fast" readers. Quite clearly, the answer to this question was negative. First, it can be noted that both groups of L2 readers were significantly slower than native readers. Despite this, no interactions between reading group (native, fast L2, and slow L2) and ambiguity resolution were observed at the point in the sentence where disambiguation occurred. Both slow and fast readers showed immediate sensitivity to structural ambiguity in the L2, in like fashion to native readers. That is, all groups took significantly longer to read the second verb in causative/clitic sentences such as (la) than in unambiguous constructions. Thus, in this study reading speed per se was not indicative of
On-Line Syntactic Processing
221
differences in syntactic processing, either between L2 readers themselves or as compared to native speakers. ~ A similar pattern of results was reported by Frenck-Mestre and Pynte (1997, Experiment 2), in an eye-movement study involving proficient English-French and French-English bilinguals. These bilingual subjects displayed slower overall reading times in their second as compared to native language, and as compared to native speakers. However, this factor was not correlated with differences in syntactic processing as compared to native processing. The results of this study showed that not only are L2 readers sensitive to syntactic ambiguities in their second language, but moreover they can use lexical subcategorisation information to resolve certain structural ambiguities. Consider the example provided in (2), below. (2a) Whenever the dog obeyed the little girl she showed her approval (2b) Whenever the dog obeyed the little girl showed her approval (2c) Whenever the dog barked the little girl showed her approval The first of the above examples, (2a), poses less difficulty than the second, (2b), for native speakers (Rayner, Carlson, & Frazier, 1983). This is due, according to one theoretical framework, to the parser systematically adopting an initial direct object analysis of the second NP, only to find this inappropriate in the case of (2b) upon the reading of the second verb (Frazier, 1987). Reanalysis is thus necessary in (2b), which is costly. In addition, in this framework (2b) and (2c) should both cause the parser to initially adopt a direct object NP analysis, although this error would be quickly recognized in (2c) due to the lexical properties of the subordinate verb coming into play. As such, (2c) is also easier to parse than (2b) (Frenck-Mestre & Pynte, 1995; Mitchell, 1989; Mitchell & Holmes, 1995; Trueswell, Tanenhaus, & Kello, 1993). The same result was obtained for proficient French-English and English-French bilinguals reading in their second language (Frenck-Mestre & Pynte, 1997, Experiment 2). Sentences such as (2c) were processed faster, causing fewer re-readings and fewer regressive saccades, than sentences such as (2b). Otherwise stated, these L2 readers were "led up the garden path" in the case of sentences like (2b) just as native readers are, whereas in the case of (2c) the subcategorisation information present in the subordinate verb helped these L2 readers to quickly resolve this syntactic ambiguity, just as it has been shown in native readers. 2 Another recent on-line study of L2 syntactic processing examined the same ambiguity as illustrated in (2b) and (2c), but in less-proficient bilinguals. Juffs and Harrington (1996) looked at Chinese ESL students' processing of this structural ambiguity by means of a word-by-word self-paced reading experiment. The authors themselves concluded that their ESL readers were sensitive to the syntactic ambiguity present in these structures, and able to resolve the ambiguity faster for
222
Cheryl Frenck-Mestre
sentences in which the subordinate verb was intransitive (2c) than optionally transitive (2b). Indeed, these L2 readers did experience apparent difficulty with the matrix verb in sentences such as (2b) as revealed by an abrupt increase in reading times at the matrix verb for these sentences. Moreover, this abrupt increase was not observed at the matrix verb in sentences such as (2c), but one constituent earlier (i.e., at the NP following the intransitive subordinate verb, thus indicating that the subcategorisation information provided by the subordinate verb was used to block/re-analyse a direct-object analysis of the NP). However, some caution may be warranted prior to concluding that these ESL readers had lesser difficulty with the structural ambiguity in the intransitive than optionally transitive cases. In fact, this direct comparison was not made in the study. The inspection of means does not lead one to believe, moreover, that at the matrix verb itself any significant interaction would obtain in the ESL group. As such, the conclusions drawn by Juffs and Harrington that ESL readers experienced the same processing difficulty as native speakers with "garden path" sentences may need to be nuanced. Yet another eye-movement experiment, reported in Frenck-Mestre and Pynte (1997, Experiment 1) revealed quite impressive use of subcategorisation information in the resolving of structural ambiguities by second-language readers. The ambiguity studied therein involved the attachment of a prepositional phrase. Consider, first, monolingual studies of this ambiguity, illustrated in (3). (3a) The spy saw the cop with the binoculars but the cop didn't see him (3b) The spy saw the cop with the revolver but the cop didn't see him In an early eye-movement study with monolinguals, Rayner, Carlson, and Frazier (1983) demonstrated that readers systematically prefer (3a)to (3b). That is, readers had greater difficulty (reflected by increased reading times) when the sentence biased "low" attachment of the prepositional phrase, to the preceding noun phrase ("with the revolver" is associated with "the cop"), than when it biased toward "high" attachment to the preceding verb phrase ("with the binoculars" thus, modifies the verb "to see"). In one theoretical framework, this preference to attach "high" is attributed to an autonomous parser which obeys heuristic principles and prefers the least costly structure in terms of syntactic nodes (Frazier, 1987). The parser is assumed to operate independently of other levels of analysis--lexical, pragmatic, and/or referential--as attested by the fact that even when semantic/pragmatic factors favor "low" attachment for this structure, as in (3b) (cops are known to carry revolvers) the initial analysis of the sentence, as revealed by first pass reading times, still reflected a "high attachment preference." The conclusions from this seminal study were brought into question, however, by subsequent monolingual studies in which the influence of thematic constraints and referential context were clearly demonstrated (Altmann & Steedman, 1988;
On-Line Syntactic Processing
223
Taraban & McClelland, 1988). Parsing preferences were not found in these on-line (self-paced reading) studies to be systematic, but to vary as a function of these "extra-syntactic" factors. Low attachment was shown to be preferred when either thematic constraints or the referential context disambiguated toward this structure. In addition, the lexical constraints of the initial verb phrase (VP) have been shown to immediately influence the processing of this ambiguity in the native language of readers (Frenck-Mestre & Pynte, 1997; see also Boland & Boehm-Jernigan, 1998 for similar work on prepositional phrase (PP) attachment, but see Ferreira & Henderson, 1990, for counter-arguments). The same is true of proficient EnglishFrench bilinguals reading in their second language (Frenck-Mestre & Pynte, 1997). Consider the examples provided in (4). (4a) I1 rate le train de nuit et d6cide alors de chercher un hotel (He missed the train of night and decided to look for a hotel) (4b) I1 rate le train de peu et d6cide alors de chercher un hotel (He missed the train by little and decided to look for a hotel) (4c) I1 avertit la police du quartier puis se f61icite de son action (He warned the police of the district and congratulated himself for it) (4d) II avertit la police du crime puis se f61icite de son action (He warned the police of the crime and congratulated himself for it) In (4), the bias towards <
> or <> attachment of the ambiguous PP (highlighted in the examples) is provided by the lexical constraints of the verb preceding it. In (4a) and (4b), the-matrix verb is monotransitive, that is, it generally selects for a single direct object complement. For this class of verb, then, the immediately following NP can satisfy the thematic grid. In contrast, the matrix verb in (4c) and (4d) is ditransitive, that is, it generally selects for two complements, only one of which is provided by the post verbal NP. For this class of verb a second complement is expected, which can be satisfied by the PP. This type of subcategorisation information influences the immediate decisions made by the parser (Frenck-Mestre &Pynte, 1997). For monotransitive verbs, the immediate preference of native French readers--as shown both off-line (via sentence completion data) and on-line (via the recording of eye movements during reading)-is to attach low, to the preceding NP. That is, at the disambiguating element (the noun of the PP, highlighted in the examples) readers experience less difficulty with (4a) than with (4b). The exact opposite is observed for sentences containing ditransitive verbs. In this instance, first pass reading times are longer at the disambiguating element for low than high attachment of the PP, that is for (4c) than (4d).
224
Cheryl Frenck-Mestre
Fluent bilingual readers show the same sensitivity to lexical constraints when resolving the syntactic ambiguity illustrated in (4). English native speakers reading in French prefer low attachment of the PP for monotransitive verbs (4a as compared to sentence 4b), but high attachment following ditransitive verbs (4d as compared to sentence 4c). Moreover, this effect is observed on first pass reading times, as it is in native speakers. Indeed, the pattern of processing did not differ as a function of "type of reader," that is whether first or second language. What can be concluded from the above set of studies? Quite clearly, fluent bilinguals are not only sensitive to structural ambiguities when reading in their L2, but moreover are sensitive to fine levels of analysis. Although we did highlight some important differences across these three studies, it is more likely, in our opinion, that the difference in sensitivity to syntactic ambiguity in the second language across these three studies is related to L2 proficiency rather than simply reading speed. Slower second language reading is not necessarily indicative of different processing as compared to native speakers, as shown by Hoover and Dwivedi (.1998) as well as by Frenck-Mestre and Pynte (1997). However, level of proficiency indeed seems to be an important factor, as can be seen in the Juffs and Harrington (1996) study. Less-proficient bilinguals may, first, be unable to immediately process syntactic ambiguity in their second language. Second, as we will outline in the next section, when a structural ambiguity can be resolved differently across languages, non-proficient bilinguals may be susceptible to effects of forward transfer. Hence, they may process the ambiguity differently from native speakers.
The Role of Experience (Years of Exposure) and Effects of Transfer in L2 Parsing In the previous section we reviewed three on-line studies where the native language of the bilingual readers was generally not considered to inhibit second language syntactic processing. In those studies, either the structure was absent from the native language of subjects (cf. Hoover & Dwivedi, 1998), or it was both identical across the bilinguals' two languages and resolved in similar manner in the two (Frenck-Mestre & Pynte, 1997; Juffs & Harrington, 1996). In this section, we will take a look at second language syntactic processing for ambiguous structures that are apt to be influenced by the reader's native language. This question can be addressed most easily through the study of structures which exist in both the native and second language but which are processed differently in the two. Just such a case is provided by the ambiguity illustrated in (5), below. (5a) Arnold watched the wife of the doctor who was leaving the medical center (5b) La police cherche left& du mddecin qui est soup~onn6 de trafic de drogues (5c) Alguien dispar6 contra la criada del actor que estaba en el balc6n
On-Line Syntactic Processing
225
(5d) Patricia conosceva il ragazzo de direttore che era svenuto alia festa (5e) De gangsters schoten op de zoon van die actrice die op het balkon zat The sentences presented in (5a) through (5e) all present the same structural ambiguity. That is, the subject of the relative clause can be either the head of the complex noun phrase (i.e., in 5a, "the wife of the doctor") or the second noun phrase alone (i.e., in 5a, "the doctor"). To date, and despite numerous on-line studies performed in a wide range of languages, from English to Japanese, no single model can account for the resolution of this ambiguity (for reviews, see Cuetos, Mitchell, & Corley, 1996; Fernandez, 2000; Fodor, 1998; Frazier & Clifton, 1996; FrenckMestre & Pynte, 2000a, 2000b; Gibson, Pearlmutter, Canseco-Gonzalez, & Hickok, 1996). Factors known to influence the resolution of this ambiguity are, among others, the respective "weight" of the complex NP and the subsequent relative clause (Fern~indez, 2000; Fodor, 1998; Pynte& Colonna, 2000), the frequency of the two noun phrases (Frenck-Mestre &Pynte, 2000a; Pynte & Colonna, 2000), the preposition in the complex NP (Cuetos et al., 1996; De Vincenzi & Job, 1993; Frenck-Mestre & Pynte, 2000b; Gilboy, Sopena, Clifton, & Frazier, 1995), segmentation of the sentence (Gilboy & Sopena, 1996, but see Carreiras & Clifton, 1999), whether one considers "on-line" or "off-line" measures of noun preference (De Vincenzi & Job, 1995), and indeed the language that the ambiguity is presented in (cuetos et al., 1996; Fern~indez, 2000; Frenck-Mestre, 1999; Frenck-Mestre & Pynte, 2000b; Gibson et al., 1996). It is this latter factor that is of particular interest for the present discussion. For purposes of the present paper, we will consider the "simplest" case, that is where the preposition in the complex NP is "of" or its equivalent in French (and Spanish, reviewed later) the printed frequency of the two nouns is either roughly equivalent or neutralized, and the sentence contains no breaks. Consider the examples provided in (6). (6a) Charles photographie lafille de Marc qui est plus d~daigneu(se/x) que jamais (Charles photographs the daughter of Mark who is more disdainful [fem/masc] than ever) (6b) Aline t616phone auxfilles de la gardienne qui reviennent/revient de Paris (Aline calls the daughters of the nanny who are/is returning from Paris) (6c) Aline t616phone ~ la gardienne desfilles qui revient/reviennent de Paris (Aline calls the nanny of the girls who is/are returning from Paris) In (6a), the structure is disambiguated by gender agreement between one of the nouns in the complex NP and the adjective in the relative clause (RC). In (6b) and (6c), the same structure is disambiguated by number agreement between either the
226
Cheryl Frenck-Mestre
first or second noun phrase (NPI or NP2) and the verb of the RC. We have shown that in French, the preferred interpretation of these sentences is to attach the RC to the head of the complex NP, as revealed by the first-pass reading times of native French speakers (Frenck-Mestre, 1999; Frenck-Mestre & Pynte, 2000b; see also Zagar, Pynte, & Rativeau, 1997, but also the debate between Baccino, De Vincenzi, & Job, 2000 and Frenck-Mestre & Pynte, 2000b). This preference to attach "high" can be accounted for in terms of the "construar' hypothesis, forwarded by Frazier and Clifton (1996), or the "predicate proximity vs. recency" model proposed by Gibson et al. (1996), or alternatively in the framework of the "linguistic tuning" hypothesis, advocated by Mitchell, and colleagues (Cuetos et al., 1996; Mitchell, 1994; Mitchell, Brysbaert, Grondelaers, & Swanep0el, 2000). These models also offer accounts for why the same structure, presented in English, is apparently resolved differently by native English speakers (for further discussion, see Cuetos, Mitchell, & Corley 1996; Frenck-Mestre & Pynte, 2000a). That is, English native speakers have been shown to either prefer "low" attachment or not to display a definite preference for either host (Carreiras & Clifton, 1993; 1999; Cuetos & Mitchell, 1988; Fernandez, 2000). Of interest here, however, is which, if any of these models can account for experienced-based differences, that is, for changes in parsing preferences in the second language of bilingual readers as these readers become more proficient in their L2. Indeed, given the reported differences across French and English as concems the resolution of this ambiguity, it is of interest to examine how English-French bilinguals resolve it when reading in French. Furthermore, it is important to determinewhether these bilinguals' parsing preferences will be modified by their experience with their second language. Indeed, in a beginning stage, it would not be surprising to find that English-dominant bilinguals are influenced by their native language when processing L2 sentences. 3 What difference, however, will experience with the L2 make? The answer to this question may lie in the type of structure that is processed, as we will outline shortly. Consider first, however, the results from a series of recent studies completed in our centre, with beginning and more advanced English-French bilinguals. In a first study (reported in Frenck-Mestre, 1999), we asked relatively nonproficient English-French "late" bilinguals (n = 16) to participate in a reading experiment wherein we recorded their eye-movements during the reading of individually presented French sentences (sentences were presented on a single line). These bilinguals had a mean of three years of formal learning of French in a classroom setting outside of France, and nine months of immersion in French, in France. They rated their abilities as concerns written and oral expression and comprehension at a level of roughly 5 on a ten-point scale of proficiency. The sentences of interest (n = 20) were of the structure illustrated in (6b) and (6c), where the verb of the subordinate clause agreed in number with either NPI or NP2. These sentences were interspersed with three times as many filler sentences of varying
On-Line Syntactic Processing
227
syntactic structures (some ambiguous), such that subjects were not prone to engage in strategic processing. Subjects were not informed of the ambiguous nature of the experimental sentences. Native French speakers were included in the study as a control group. The results are presented in Figure 1. As can be seen, the two groups showed a quite different pattern of results as concerns the immediate resolution of this ambiguity. Native French speakers showed a definite preference for "high" attachment of the RC. That is, first pass (as well as total) reading times were significantly faster at the disambiguating subordinate verb when it agreed with the head of the complex NP than when it agreed with NP2. The opposite was observed in-the group of English-dominant bilinguals. These beginning bilingual readers, who had been living in France for a mean of nine months and had a mean of three years of formal classroom learning in French, showed a trend toward "low" attachment, to the second noun phrase. That is, their first pass (and total) reading times tended to be faster when the subordinate verb agreed in number with the second than with the first noun of the complex NP (Note that in both the native French and the beginning bilingual group, the effect of attachment was independent of whether the subordinate verb was singular or plural). As we have argued elsewhere (cf. Frenck-Mestre, 1999), the pattern obtained in the English-dominant group is to be attributed most likely to the influence of their native language on second language processing, rather than to a general strategy of attaching new elements to the most recently processed constituent. First Pass Gaze Durations (ms) across all regions
800
"-0-
N1 Attachment
-{~-
N2 Attachment
"~ 700
E o
600
u) a. 500 ,m
~
4OO
300
NP1
NP2
Verb
VP+I
Native French Speakers
NP1
NP2
Verb
VP+I
Beginning B i l i n g u a l s
Figure 1. Mean first pass reading times as a function of type of reader and type of relative clause attachment ("high" to NP 1 or "low" to NP2) and sentence region.
228
Cheryl Frenck-Mestre
In another study (e.g., Frenck-Mestre, 1998), we examined the performance of proficient English-French "late" bilinguals for the same ambiguous structure in a new eye-movement experiment. The same set of sentences and same presentation were employed as in the previous experiment (see the examples provided in 6b and 6c). The group had roughly three years of formal learning of French outside of the country, at least two years of study in a French University in a mainstream curriculum, and had been in France for a mean of five years. They rated themselves at a level of 7 or better on a ten-point scale of proficiency in the second language, for written and oral comprehension and production skills. The results of these "proficient" bilinguals are compared to the performance of the same group of native French speakers examined in the first study, in Figure 2. As can be seen in Figure 2, the proficient bilinguals' syntactic processing of the ambiguity in French was highly similar to that of native French speakers. In fact, the statistical comparisons of these two groups did not reveal any significant interactions with type of reader. The proficient bilinguals were not slower in their first pass reading times than native French speakers (p > . 10), nor did they demonstrate a different pattern of ambiguity resolution than native speakers.
First
Pass Gaze Durations
- [~-
(ms) across
NI
Attachment
N2
Attachment
all regions
700
w
600
E :o 500
4,w
a 400 G) N
~1 300 200
i
i
i
i
i
,
i
l
NP1
NP2
Verb
VP+I
NPI
NP2
Verb
VP+I
Native
French Speakers
Proficient Bilinguals
Figure 2. Mean first pass reading times as a function of type of reader and type of relative clause ("high" to NP 1 or "low" to NP2) and sentence region.
On-Line Syntactic Processing
229
A comparison was made of the three groups--native French, beginning, and advanced English-French bilinguals--as concems the first pass gaze duration at the disambiguating subordinate verb. This comparison is presented in Figure 3. As can be readily seen, the native and advanced groups showed a highly similar pattern of ambiguity resolution for the NPI-Prep-NP2-RC structure, which differed from the group of beginning (English-dominant) bilinguals. This was bom out in statistical analyses, showing a significant interaction between type of reader and attachment preference (FI(1,36) = 4.89, p <.01 ).
First Pass Gaze at the Disambiguating VP
I
A
-0-
N1 A t t a c h m e n t
-O-
N2 Attachment I
500
E o
,,m 4~
Im
= a
400
N
300
I
i
French
Beginning
I
Advanced
R e a d i n g t i m e s of t h e d i s a m b i g u a t i n g s e g m e n t as a f u n c t i o n of r e a d e
Figure 3. Mean first pass reading times as a function of type of attachment of the ambiguous relative clause (N 1 vs. N2) and type of reader.
Post-hoc comparisons revealed that the beginning bilinguals differed from the advanced and native groups as concerns the time spent reading the subordinate verb in the case of "high" (i.e., N1 attachment). The beginning bilinguals were significantly slower to process the subordinate verb than advanced bilinguals and native speakers when the verb agreed in number with the first noun phrase. Otherwise stated, these English-dominant bilinguals demonstrated specific difficulty with "high" attachment. Our data show a clear evolution of processing strategies, from performance closely tied to the native language of late bilinguals, to that closely resembling the
230
Cheryl Frenck-Mestre
performance of native speakers of the language. It is important to underline that the difference between the two groups of bilinguals was experience with the second language. The advanced group had lived for several years in the country of their second language, in environments where they used the language as a daily means of communication. They had been "exposed" to their second language, and its patterns, for several years. Prior to examining which model of processing can best account for these data, we will consider the results of another study of the same syntactic ambiguity. Our data can be compared to those reported recently by Fernandez (1998), for Spanish-English and Japanese-English bilinguals, in off-line studies of the same, relative clause attachment ambiguity. It is clear from the questionnaire data provided in this study that the deciding factor behind ambiguity resolution in the second language was level of expertise in this language. Whether the bilinguals learned their second language early or late was not predictive of the amount of forward transfer from the bilinguals' native language. Self-rated proficiency, however, was a good predictor. Those Spanish speakers who rated themselves as more proficient in English than in Spanish processed this ambiguity akin to native English speakers, whereas those who rated themselves as more proficient in Spanish tended to process this ambiguity as would monolingual Spanish speakers. In like manner, the Japanese bilinguals, despite having learned English fairly, rated themselves in general as "Japanese-dominant" and showed clear effects of transfer from their native language. In our study, the number of years of experience with and/or immersion in the second language (French) was the deciding factor: less proficient bilinguals showed clear effects of forward transfer, whereas the proficient group was well on its way to processing the ambiguity in question as would a monolingual. 4 These data clearly show that any processing model of L2 syntactic analysis must take into account the evolution of processing with experience. The most readily adapted framework for this would seem to be the "linguistic tuning" hypothesis (Cuetos et al., 1996; Mitchell, 1994). The model incorporates both cross-linguistic differences in ambiguity resolution and changes over time in processing as a result of exposure, due to the very nature of the model. Indeed, it is assumed that the particular analysis adopted by the adult reader will be the product of his/her exposure to a given language and its statistical properties as concerns the frequency of structures. Changes in syntactic processing are predicted as a result of exposure, as clearly stated in Cuetos et al. (1996, p. 175), "The model predicts that parsing preferences will change if, during some period prior to testing, the reader or listener has been exposed to an unusual preponderance of one ambiguity resolution rather than another." Indeed, our data (as well as Fern~indez') show that English-French bilinguals will adopt a "French" strategy of "high attachment" for the RCattachment ambiguity atter having been exposed to French for numerous years. 5
On-Line Syntactic Processing
231
The linguistic tuning hypothesis is not alone, however, in being able to account for cross-linguistic variation as concems the ambiguity under question, nor, indeed, for the evolution of processing we observed in our bilingual subjects as concerns this structure. The construal hypothesis, forwarded by Frazier and Clifton (1996), stipulates that 'non-primary' relationships such as relative clauses are 'construed' rather than being governed by strictly syntactic principles. Attachment would be the result, in this case, of "late" discourse level processes. This leaves open the possibility, first, for cross-linguistic variation. As concerns bilingual processing, the construal hypothesis could also account for the change we observed in English native speakers' processing of French sentences with increased exposure to their second language. Given the apparently strong preference for "high" attachment of the RC following a genetive relationship (e.g., "the daughters of the nanny") in French (and Spanish), it is perhaps not surprising that native English speakers would adhere to this processing strategy. This is all the more plausible given that there is apparently no clear-cut preference in English for the resolution of this structure. Hence, when faced with this structure in their second language, and provided substantial past experience with a certain type of resolution thereof, the English speaker might well adopt the most frequent form of ambiguity resolution in their L2. Note, conversely, that French and Spanish speakers of English should thus show a reluctance to drop their LI strategies for this particular ambiguity, as there is no "dominant" strategy in their L2 (English). This type of asymmetry in forward transfer is open for investigation.
Conclusions
In summary, the present review of a handful of on-line studies of second language syntactic processing has shown, first, that second language syntactic analysis need not be considered as a separate case from native-language parsing. Proficient bilinguals show immediate sensitivity to structural ambiguities in like manner to native readers, quite independently of reading speed. These proficient bilinguals are also able to take advantage of fine-levels of analysis to guide (or filter) the parsing of structurally ambiguous sentences in the second language. Second, in the instance that the bilingual's two languages differ as concerns the processing of a given syntactic structure, changes in parsing strategies will occur over time as the bilingual gains experience with the second language. A decrease in forward transfer is observed for more proficient bilinguals along with increased "native-like" processing.
232
Cheryl Frenck-Mestre Notes
~Note, there were differences between the fast and slow L2 readers at the final segment of sentences, whereby the latter group was significantly slower for structurally ambiguous sentences. Hoover and Dwivedi (1998) attributed this to the slower readers' lesser ability to process these structures. 2 Note, that there were also effects of inappropriate forward transfer, from the native to second language as concerns lexical constraints (see Frenck-Mestre & Pynte, 1997, for further details). 3 Since the end of the nineteenth century, linguists have provided us with examples of "transfer" whenever two languages come into contact (Sweet, 1899/1972). The sentences produced by non-native speakers often bear the mark of their first language (Durguno~;lu & Hancin, 1992; Koda, 1993; Odlin, 1989; MacWhinney, 1992). For example, word order in second language sentences is frequently affected by native language constraints (Meisel, Clahsen & Pienemann, 1981). Nonetheless, the results of various studies have underlined cross-linguistic commonalities in second language acquisition as concerns syntactic development, which support the hypothesis of linguistic universals in second language development (Muysken, 1984; Zobl, 1986), rather than a "cognitive-based" approach to second language learning. At the intersection of these two theoretical viewpoints Flynn (1984), Flynn and Espinal (1985) highlight the important role played by the native language of adult learners in the acquisition of second language syntax while underlining the universal nature of acquisition. 4 In a recent ERP study of early and late bilinguals (Weber-Fox & Neville, 1996), it was found that "late" bilinguals do not show the same sensitivity as early bilinguals or monolinguals to phrase structure violations. It would be of interest to see whether eye-movement recordings would replicate these findings, or, rather, if the reported effect is linked to the inability to "repair" for this type of sentence. 5 It is worth noting that, though attractive, the linguistic tuning hypothesis has experienced what would appear to be a serious set-back in recent studies involving Dutch (Mitchell & Brysbaert, 1998). That is, when on-line RC attachment is examined for materials modeled closely on corpora-based sentences, the on-line preferences are quite the opposite that which would be predicted by the statistical frequency of structures in written Dutch.
References
Altmann, G., & Steedman, M. (1988). Interaction with context during human sentence processing. Cognition, 30, 191-238. Baccino, T., De Vincenzi, M., & Job, R. (2000). Cross-linguistic studies of the late closure strategy: French and Italian. In M. De Vincenzi & V. Lombardo (Eds.),
On-Line Syntactic Processing
233
Cross-linguistic perspectives on language processing (pp. 89-118). Boston, MA: Kluwer Academic. Boland, J. E., & Boehm-Jernigan, H. (1998). Lexical constraints and prepositional phrase attachment. Journal of Memory and Language, 39, 684-719. Carreiras, M., & Clifton, C. (1993). Relative clause interpretation preferences in Spanish and English. Language and Speech, 36, 353-372. Carreiras, M., & Clifton, C. (1999). Another word on parsing relative clauses: Eye-tracking evidence from Spanish and English. Memory & Cognition, 27, 826833. Cuetos, F., & Mitchell, D. C. (1988). Cross-linguistic differences in parsing: Restrictions on the late-closure strategy in Spanish. Cognition, 30, 73-105. Cuetos, F., Mitchell, D. C., & Corley, M. (1996). Parsing in different languages. In M. Carreiras, J. Garcia-Albea, & N. Sebastien-Galles (Eds.), Language processing in Spanish (pp. 145-187). Mahwah, NJ: Erlbaum. De Vincenzi, M., & Job, R. (1993). Some observations on the universality of the late closure strategy. Journal of Psycholinguistic Research, 22, 189-206. De Vincenzi, M., & Job, R. (1995). An investigation of late closure: The role of syntax, thematic structure and pragmatics in initial and final interpretations. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1303-1321. Durguno~lu, A. Y., & Hancin, B. J. (1992). An overview of cross-language transfer in bilingual reading. In R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 391-413). Amsterdam: Elsevier. Fern~indez, F. (1998). Language dependency in parsing: Evidence from monolingual and bilingual processing. Psychologica Belgica, 38, 197-230. Fern~indez, F. (2000). Bilingual sentence processing: Relative clause attachment in English and Spanish. Unpublished Doctoral Dissertation, City University of New York. Ferrand, L., & Grainger, J. (1993). The time course of orthographic and phonological code activation in the early phases of visual word recognition. Bulletin of the Psychonomic Society, 31, 119-122. Ferreira, F., & Henderson, J. M. (1990). Use of verb information in syntactic processing: Evidence from eye movements and word-by-word self-paced reading. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 555568. Flynn, S. (1984). A universal in L2 acquisition based on a PBD typology. In F. Eckman, L. Bell, & D. Nelson (Eds.), Universals of second language acquisition (pp. 75-87). Rowley, Mass: Newbury House. Flynn, S., & Espinal, I. (1985). Head initial/head final parameter in adult Chinese L2 acquisition of English. Second Language Research, 1, 93-117. Fodor, J. (1998). Learning to parse? Journal of Psycholinguistic Research, 27, 285-319.
234
Cheryl Frenck-Mestre
Frazier, L. (1987). Sentence processing: A tutorial review. In M. Coltheart (Ed.), Attention and performance XII: The psychology of reading (pp. 559-586). Hillsdale, NJ: Erlbaum. Frazier, L., & Clifton, C. (1996). Construal. Cambridge, MA: MIT Press. Frenck-Mestre, C. (1993). Use of orthographic redundancies and word identification speed in bilinguals. Journal of Psycholinguistic Research. 22, 397410. Frenck-Mestre, C. (1998, November). Overcoming parameters: An on-line look at bilingual sentence processing. Paper presented at the 39 th annual meeting of the Psychonomic Society, Dallas, TX. Frenck-Mestre, C. (1999). Examining second language reading: An on-line look. In A. Sorace, C. Heycock, & R. Shillcock (Eds.), Language acquisition: Knowledge representation and processing (pp. 474-478). Amsterdam: North-Holland. Frenck-Mestre, C., &Pynte, J. (1995). Lexical influences on parsing strategies: Evidence from eye movements. In J. M. Findlay, R. W. Kentridge, & R. Walker (Eds.), Eye movement research: Mechanisms, processes and applications (pp. 433444). Amsterdam: Elsevier. Frenck-Mestre, C., & Pynte, J. (1997). Syntactic ambiguity resolution while reading in second and native languages. The Quarterly Journal of Experimental Psychology, 50,4, 119-148. Frenck-Mestre, C. & Pynte, J. (2000a). Romancing syntactic ambiguity: Why the French and the Italians don't see eye to eye. In A. Kennedy, R. Radach, D. Heller, & J. Pynte (Eds.), Reading as a perceptual process (pp. 549-564). Amsterdam: Elsevier. Frenck-Mestre, C., &Pynte, J. (2000b). Resolving syntactic ambiguities: Crosslinguistic differences? In M. De Vincenzi & V. Lombardo (Eds.), Cross-linguistic perspectives on language processing (pp. 119-148). Boston, MA: Kluwer Academic. Frost, R., Katz, L., & Bentin, S. (1987). Strategies for visual word recognition and orthographical depth: A multilingual comparison. Journal of Experimental Psychology: Human Perception and Performance, 13, 104-115. Gibson, E., Pearlmutter, N., Canseco-Gonzales, E., & Hickok, G. (1996). Crosslinguistic attachment preferences: Evidence from English and Spanish. Cognition, 59, 23-59. Gilboy, E., & Sopena, J. M. (1996) Segmentation effects in processing complex NPs with relative clauses. In M. Carreiras, J. Garcia-Albea, & N. Sebastien-Galles (Eds.), Language processing in Spanish (pp. 191-206). Mahwah, NJ: Erlbaum. Gilboy, E., Sopena, J. M., Clifton, C., & Frazier, L. (1995). Argument structure and association preferences in Spanish and English complex NPs. Cognition, 54, 131-167. Hoover, M. L., & Dwivedi, V. D. (1998). Syntactic processing by skilled bilinguals. Language Learning, 48, 1-29.
On-Line Syntactic Processing
235
Juffs, A., & Harrington, M. (1996). Garden path sentences and error data in second language sentence processing. Language Learning, 46, 283-326. Koda, K. (1993). Transferred L1 strategies and L2 syntactic structure in L2 sentence comprehension. The Modern Language Journal 77, 490-499. MacWhinney, B. (1992). Transfer and competition in second language learning. in R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 371-390). Amsterdam: Elsevier. Meisel, J., Clahsen, H., & Pienemann, M. (1981). On determining developmental stages in natural second language acquisition. Studies in Second Language Acquisition, 3, 109-135. Mitchell, D. C. (1989). Verb-guidance and other lexical effects in parsing. Language and Cognitive Processes, 4, 123-154. Mitchell, D. C. (1994). Sentence parsing. In M. A. Gernsbacher (Ed.), Handbook ofpsycholinguistics (pp. 375-409). New York: Academic Press. Mitchell, D. C., & Brysbaert, M. (1998). Challenges to recent theories of crosslinguistic variation in parsing: Evidence from Dutch. Syntax and Semantics, 31, 313-335. Mitchell, D. C., Brysbaert, M., Grondelaers, S., & Swanepoel, P. (2000). Modifier attachment in Dutch: Testing aspects of construal theory. In A. Kennedy, R. Radach, D. Heller, & J. Pynte (Eds.), Reading as a perceptual process (pp. 493516). Amsterdam: Elsevier Mitchell, D. C., & Holmes, V. M. (1995). The role of specific information about the verb in parsing sentences with local structural ambiguities. Journal of Memory and Language, 24, 542-549. Muysken, P. (1984). The Spanish that Quechua speakers learn. In R. Anderson (Ed.), Second languages: A cross-linguistic perspective. Rowley, Mass: Newbury House. Odlin, T. (1989). Language transfer: Cross linguistic influence in language learning. Cambridge: Cambridge University Press. Paap, K., McDonald, J. E., Schvaneveldt, R. W., & Noel, R. W. (1987). Frequency and pronounceability in visually presented naming and lexical decision tasks. In M. Coltheart (Ed.), Attention and performance XII: The psychology of reading (pp. 221-244), Hillsdale, NJ: Erlbaum. Pynte, J., & Colonna, S. (2000). Decoupling syntactic parsing from visual inspection. The case of relative clause attachment in French. In A. Kennedy, R. Radach, D. Heller, & J. Pynte (Eds.), Reading as a perceptual process (pp. 529548). Amsterdam: Elsevier. Rayner, K., Carlson, M., & Frazier, L. (1983). The interaction of syntax and semantics during sentence processing: Eye movements in the analysis of semantically biased sentences. Journal of Verbal Learning and Verbal Behavior, 22, 358-374.
236
Cheryl Frenck-Mestre
Rayner, K., & Pollatsek, S. (1987). Eye movements in reading: A tutorial review. In M. Coltheart (Ed.), Attention and performance XII: The psychology of reading (pp. 327-362). Hillsdale, NJ: Erlbaum. Rayner, K., Sereno, S. C., Morris, R. K., Schmauder, A. R., & Clifton, C. (1989). Eye movements and on-line language comprehension processes. Language and Cognitive Processes, 4, 21-49. Segalowitz, N. (1986) Skilled reading in the second language. In J. Vaid (Ed.),
Language processing in bilinguals: Psycholinguistic and neuropsychological perspectives (pp. 3-19). Hillsdale, NJ: Erlbaum. Segalowitz, N., & H6bert, M. (1990). Phonological recoding in the first and second language reading of skilled bilinguals. Language Learning, 40, 503-538. Sweet, H. (1899/1972). The practical study of languages. Language and language learning series. London: Oxford University Press. Taraban, R., & McClelland, J. L. (1988). Constituent attachment and thematic role assignment in sentence processing: Influences of content-based expectations. Journal of Memory and Language, 27, 1-36. Trueswell, J. C., Tanenhaus, M. K., & Kello, C. (1993). Verb-specific constraints in sentence processing: Separating the effects of lexical preference from garden paths. Journal of Experimental Psychology: Learning, Memory, and Cognition, 19, 528-553. Weber-Fox, C. M., & Neville, H. (1996). Maturational constraints on functional specializations for language processing: ERP and behavioral evidence in bilingual speakers. Journal of Cognitive Neuroscience, 8, 231-256. Zagar, D., Pynte, J., & Rativeau, S. (1997). Evidence for early closure attachment on first-pass reading times in French. Quarterly Journal of Experimental Psychology, 50A, 421-438. Zobl, H. (1986). Word order typology, lexical government and the prediction of multiple, graded effects in L2 word order. Language Learning, 36, 159-183.
Bilingual Sentence Processing- R.R. Heredia and J. Altarriba (Editors) 9 2002 ElsevierScience B.V. All rights reserved.
10
Cross-Linguistic Aspects of Anaphor Resolution
Dieter Hillert University of California, San Diego
Abstract This chapter discusses how speakers of different languages tend to compute, during real-time sentence comprehension, a variety of different anaphoric relationships. The results are interpreted according to the Universal Grammar paradigm. It is concluded that experimental approaches to the examination of the Universal Grammar should consider more and more a linguistic profile that takes into account different degrees of nonnative language competence as well as a broad range of typologically different languages.
Universal and Typological Structures Among philosophers and linguists it is quite non-controversial that context must be evoked to account for presuppositional phenomena (e.g., Austin, 1962; Quine, 1960; Searle, 1958; Strawson, 1956). The nature of this intuition has been formerly expressed in natural language (e.g., Ariel, 1985, 1988, 1990; Chomsky, 1982; Gazdar, 1979; Reinhart, 1984). The present paper will concentrate on the system of accessing noun phrase antecedents by considering syntactic, semantic, and pragmatic properties of the sentence structure. In linguistics, a noun phrase is defined by phrase structure (PS) rules. Each PS rule consists of a head that is related to its phrasal counterpart. For instance, the three heads in "the beautiful language" consists of a determiner (the), an adjective (beautiful) and a noun (language). Simplified for illustrative purposes, the relevant PS rule is NP --~ Det Adj N in which the heads "the" (Det), "beautiful" (Adj), and "language" (N) are attached to a noun phrase (NP). The degree of accessibility varies between various NPs that are partly expressed in their binding conditions across different languages. According to Chomsky (1982) an anaphor is bound in its governing category, a pronominal is free in its governing category, and a referential expression is free. With other words, reflexives and reciprocals must be linked to an antecedent within a minimal domain (usually the clause they are in). In contrast, pronouns can be either independently referential (deictic) or referentially dependent on a linguistic antecedent (anaphor). In considering Ariel's (1990) "accessibility theory", different context types guide the degree of accessibility of a word. For instance, in considering Chomsky's binding conditions, reflexives and reciprocals are highly accessible because they cannot have an independent reference and they must be
238
Cross-Linguistic Aspects of Anaphor Resolution
linked to an antecedent within a minimal domain. On the other hand, pronouns are lower accessible than reflexives and reciprocals because they can be deitic or anaphoric. Again, different context variants such as world knowledge, physical situation, and immediate discourse contribute to the degree of accessibility. In this chapter, we focus on the accessibility and use of world knowledge to interpret an anaphoric relationship across different languages. A fundamental issue is how typological diverse patterns are computed in bilingual speakers. In general, one might separate between three acquisition accounts on the relationship between the universal grammar (UG) and the language-specific grammar(s). It is a welldocumented research area in cognitive (neuro)science that the acquisition of (a first) language is based on a genetic disposition for language. Children acquire language during a certain period of time according to certain stages and patterns of acquisition, independent of typological differences. The biolinguistic disposition as well as the universal characteristics of acquisition (and disorders) strongly supports Chomsky's idea of a universal grammar (UG). As defined in the present context, UG refers to all natural languages, which includes dead languages, newly developed languages, sign languages, and all symbolic systems people can make use of. But what is the function of the UG? Are UG principles and parameters modified according to principles and parameters of the language in question or does the UG remain invariant and does it constrain the acquisition language-particular structures? Without discussing controversial issues in detail, we support a view that supports a non-modified and steady UG (Flynn & Martohardjono, 1994). We consider UG as an underlying computational device (deep structure level) that provides adult speakers with the option to learn new languages. However, this does not exclude the possibility of generalization among languages. A transfer usually takes place on the surface structure level when grammatical functions of language A are used to compensate for non-acquired grammatical functions in language B. Thus, from the present viewpoint it is always necessary to control interfering factors from typological characteristics of those non-target languages that are relatively frequently used by the bilingual or the polyglot speaker. Certainly, from adult foreign language acquisition we should learn to what extent the UG is contaminateable or modifiable. We claim that the UG per se remains unaffected by the acquisition of new structures because surface level structures are modified, but not the underlying structures of the UG. However, cross-linguistic research may expand our knowledge of those principles and parameters that belong to a single language and those that belong to the UG.
The Structure of Coreference
Contemporary process or sentence-level models. word recognition models word recognition models about the degree to what
models of language comprehension are either discourseIn turn, sentence-level models can be divided between (lexical access models) and parsing models. Typically, are divided with respect to the controversial discussion extent linguistic context influences the speed of lexical
Dieter Hillert
239
access (see for example Hillert & Swinney, 2001; Lucas, 1999; Swinney, 1979). Similarly, several accounts have focused on how perceivers parse sentences in realtime (e.g., Hillert, 1998; 2002; MacDonald, Perlmutter, & Seidenberg, 1994; Pynte, 1998; Pynte & Prieur, 1998). The empirical evidence seems to be strongly dependent on the specific empirical approach selected, as discussed below. One approach to the cognitive analysis of language comprehension is divided between three computational levels: surface level, propositional level, and situative level (e.g., Gernsbacher, 1990; Kintsch, 1988, 1998; van Dijk & Kintsch, 1983). The surface level refers to the specific syntactic information and lexical information used in sentence or discourse context. A listener memorizes surface information only a few seconds (e.g., Kintsch, Welsch, Schmalhofer, & Zimny, 1990), unless pragmatic aspects play a particular role in a given discourse context. The common view is that listeners will construct a "situation model" when referential functions are dominant, as in a scientific discourse, but not when a poetic function is dominant, as in literature. Which information will be reconstructed in a situation model dependents upon the specific characteristics of the discourse itself, as well as from the listener's cognitive strategies. In the present context, we will primarily focus on sentence-level processing aspects by considering universal cognitive strategies as well as language-particular strategies that influences the cognitive strategies used by non-native speakers. A central aspect of a language comprehension model is the phenomenon of anaphora resolution. An anaphor co-references an expression with its antecedent, that is, an expression usually refers back to a preceding expression to share the same referent. However, in "backward anaphora" or "crossing coreference sentences" (sometimes called "Bach-Peters sentences") the first pronoun acts like its full NP antecedent, as shown in (la); the second NP, however, does not include this property (see 1b). This condition lets us assume that there is an asymmetry between both pronouns (cf. Bach, 1970). (1 a) The lawyer who loves her kissed his spouse twice (1 b) The lawyer who loves his spouse kissed her twice In general, an anaphor implies a repeated or paraphrased discourse referent that is a major source for text coherence. For instance, some well-known,anaphora are Cicero's (In Catilinam)"...non feram, non sinam, non patiar ..." (I shall not bear it, I shall not endure it, I shall not allow it.), Gerdrude Stein's "A rose is a rose is a rose", or Martin Luther King's "I have a dream speech". The notion of coreference does not imply that this term is primitive in grammar. We assume, however, that it can be inferred from the semantic-pragmatic representations of some sentences. Referential indices will be used pragmatically, that is, they do not refer to a certain grammatical theory, but we consider referential indices as disambiguating surface sentence structures. Accordingly, the term "antecedent" has no theoretical implications. If a pronoun is co-referential to an NP (referent), the latter will be called the antecedent of the former. An anaphor, however, might be also expressed by an explicit or implicit clitic. A clitic can be regarded as hybrid between an affix
240
Cross-Linguistic Aspects of Anaphor Resolution
and a content word. While clitics are syntactically independent, they are usually phonologically and/or semantic-pragmatically bounded to a preceding or a following (content) word. Depending on language-particular characteristics, clitics might stay on their own at a prescribed syntactic position, if they are overtly expressed, or may function as a suffix. In Spanish, for example, the subject pronoun can be dropped (2a), and the object pronoun might precede the verb (2b), or might be attached to the ending of the verb (2c). (2a) pro explica la situaci6n (Explain the situation; S/he explains the situation) (2b) La explica la situaci6n (Her explain the situation; S/he explains her the situation) (2c) Explicala la situaci6n (Explain-her the situation; S/he explains her the situation) Pronouns are non-anaphoric because they must be free in a local domain. For instance, in the French phrase (3a) the reflexive pronoun "se" (her-/himself) must refer to the entity denoted by "elle" (she); in (3b), however, the object pronoun "le" (masculine) cannot refer to the head NP, but refers to an external entity. (3a) (She (3b) (She
Elle se blesse pendant la s6ance d'entraTnement her-/himself hurt during the workout; She hurts herself during the workout) Elle le blesse pendant la s6ance d'entraTnement him hurt during the workout; She hurts him during the workout)
The above examples illustrate that an anaphoric relationship can be expressed in various ways depending on the typological characteristics of a single language. In Non-Romance languages like German in which the word order is relatively "relaxed," the pronoun can be placed in different syntactic positions. While (4a) shows the non-canonical IOPro-V-S-DO (indirect object pronoun-verb-subjectdirect object) sentence structure, (4b) shows a DO-V-S-IOPro sentence structure. (4a) Ihm gab sie das Kochbuch, weil er Frtihlingsrollen zubereiten wollte (Him gave she the book; She gave him the book) (4b) Das Kochbuch gab sie ihm, weil er Frtihlingsrollen zubereiten wollte (The book gave she him; The book, she gave him) In particular, when we consider long-distance dependencies, the position of the pronoun may influence how fast listeners are able to access the antecedent. In (5a), the pronoun that refers to the first lexical element of the sentence is part of a subordinate clause (subclause) and in (5b) it is part of sub-subclause. Thus, it is not necessarily the time span between pronoun and antecedent that manipulates antecedent access, but the syntactic environment of the anaphor.
Dieter Hillert
241
(5a) Paul is going to hold his breath for several minutes until he turns blue (5b) John said yesterday to Carmen that Mary thought that he loved Linda Moreover, different nouns might refer to the same concept, and in this case a content word functions like a clitic (6) "Audrey," "the actress," "Breakfast at Tiffany's chief protagonist," "this beautiful woman." All these nouns might, for instance, refer to the same concept of [Audrey]
(6) [Audrey]
Audrey the actress this beautiful woman Breakfast at Tiffany's chief protagonist
However, in the present context we are not interested in accounting for the meaning of what the antecedent is referring to. In the case of code switching (e.g., non-native language processing), culture-specific presuppositions may have a direct influence on the accuracy and speed of antecedent processing. Thus, the variability of conceptual presuppositions must be carefully controlled in psycholinguistic experiments on referential assignments. Because we are interested in discussing some universal linguistic and cognitive principles that guide access and resolution of anaphoric relationships, we focus in the present context on the properties of the language system as a cognitive module. One of the initial questions is how listeners access the antecedent when they recognize an anaphoric relationship during realtime sentence processing. For example, do we store a potential antecedent in a passive format that requires re-activation of the antecedent? To what extent do syntactic properties guide or facilitate access to the antecedent? Does semantic information interfere with the access mechanism? What kind of syntactic and/or semantic/conceptual information is used to resolve anaphoric ambiguities?
Real-Time Comprehension of Anaphora There are two general accounts of normal sentence processing in the literature: (a) the constraint satisfaction account and (b) the context-independent account. The constraint satisfaction account claims that language processing is a highly interactive operation in which the selection of a certain parsing strategy occurs as lexical information is encountered (MacDonald, Pearlmutter, & Seidenberg, 1994). Thus, access to words is facilitated by sentence context (Tabossi, 1988). The context-independent account of sentence processing, in contrast, claims that initial processes are modular. Thus, parsing is an autonomous process that takes place independent of lexical entries (J. D. Fodor, 1995), and words are initially accessed without contextual penetration (e.g., J. A. Fodor, 1983; Seidenberg, Tanenhaus,
242
Cross-Linguistic Aspects of Anaphor Resolution
Leiman, & Bienkowski, 1982; Tanenhaus, Leiman, & Seidenberg, 1979; Wilson & Sperber, 1986). In the following section, the divergent findings and models of coreference assignment will be discussed with respect to a coherent picture of the language comprehension process. This analysis will be taken as evidence for a model that considers the sentence as well as the discourse level and catches different temporal stages that operate simultaneously during comprehension of a continuous stream of speech. The major questions are, at which level of representation do we (re)access an antecedent and in which format do we store temporarily coreferenced information? Nicol (1988) found evidence for the distinction between anaphoric and nonanaphoric pronouns with English monolinguals. She used the cross-modal lexical priming (CMLP) technique. In this technique, participants simultaneously perform a passive and an active task. First, they pay attention to spoken sentences presented one at a time (passive task); second, during sentence presentation, a probe (word or non-word) flashes on the computer screen for about 300 milliseconds (ms) and participants perform a word/nonword lexical decision task'(active task). The probes are displayed at the point at which the examiner intends to measure lexical priming. For example, in sentences such as (7a), one may wish to test whether or not the antecedent "doctor" (depicted by subscript "i") can be reactivated at the offset of the anaphor "himself' (depicted by the subscript "ii"). Notice that in (7a), "himself' can only refer to the "doctor." Faster reaction time to a word that previously appeared in the sentence (e.g., "doctor") rather than to a word that is not related to the sentence context (a control word) would indicate that access to the relevant antecedent was facilitated by previous activation patterns (see for example Hillert, 1999). In sentences in which the reflexive is bounded to the subject NP of the subclause (e.g., 7a), Nicol found only re-activation of the antecedent "doctor" at the offset of the anaphor "himself," but no priming for both subject NP "boxer" and object NP "skier" of the main clause. In contrast, in sentences in which the pronoun within the subclause refers to the subject NP of the main clause, Nicol found significant priming not only for the subject NP "boxer" but also for the object NP "skier" (e.g., 7b). However, no priming was found for the subject NP "doctor" of the subclause. This finding is not surprising because both NPs of the main clause are possible candidates for being the antecedent of the pronoun. (7a) The boxer told the skier that the doctor0) for the team would blame himself(ii) for the recent injury (7b) The boxer(i) told the skier that the doctor for the team would blame him(i0 for the recent injury The data show several important aspects about how listeners recognize potential antecedents. First, if a link is compiled between a reflexive pronoun and its NP antecedent within a minimal domain, only the relevant antecedent within this domain will be activated. Therefore, this effect on reflexives was not due to a residual effect. Otherwise, Nicol should have found priming for "doctor" in object
Dieter Hillert
243
pronoun sentences. These results are in line with the syntactic binding constraints, that is, pronouns must be free in a local domain (no priming for "doctor") and reflexives must be anaphoric linked (priming for "doctor"). Similarly, Swinney, Ford, & Bresnan (1989) using the CMLP paradigm examined sentences such as (8a) by presenting potential referents at the probe positions "1-3. In this case "boxer" is the only possible antecedent of the pronoun "him." The direct object "doctor" is linked to a syntactic trace position at the offset of the infinitive verb "to see." (8a) The boxer~ visited the doctorj thatj the swimmer at the competition had ^! advised ^2 himi ^3 to s e e j about the injury" (8b) The boxer visited the doctori that the swimmer at the competition had advised m.i While no priming was found at the control probe position ^1, significant priming was measured at the verb probe position "2 for the object but not for the potential antecedent "boxer." Again, at the pronoun probe position A3 activation of both NPs of the main clause, "boxer" and "doctor", were found. That priming was found for "boxer" at the offset of the pronoun, but no priming for "swimmer" at any probe position confirms the account that personal pronouns must be locally flee. Moreover, that no priming was found for the antecedent "boxer" at the first and second probe positions points to the issue that the antecedent has been re-activated at the relevant third probe position. However, the question arises why the obviously irrelevant NP "doctor" was activated at the second and third probe position? Swinney et al. (1989) assume that the parser may "erroneously" anticipate a whtrace at the vicinity of the verb (probe point ^2), as shown in (8b). (Wh-trace analysis: X visited Y that Z had advised whom?) This analysis goes in line with the findings on wh-traces, suggesting that the correct antecedent is immediately linked to the trace position and alternative referents are immediately excluded. It should be emphasized at this point that it is unkown whether the antecedent (re)activation is triggered by the pronoun itself or simply by the implicit argument structure of the verb. A reliable control study might use sentences in which the relevant pronoun is located in a pre-verbal and in a post-verbal position (Hillert, 2001). Significant priming difference between a pre- and post-verbal position would inform us that access to a verb triggers a process that attempts to fill the arguments of a verb. Another question is whether coreference assignment is more syntactic or conceptual in nature. Clifton and Ferreira (1994) measured the reading time required for the segment containing the relevant pronoun. For example, in (9a) both NPs of the first clause (John, Mary) belong to different constituents, but not in (9b).
(9a) [John] [pushed [Mary to the head of the line]], but suddenly / they discovered / that all their money was missing
244
Cross-Linguistic Aspects of Anaphor Resolution
(9b) [John and Mary] [pushed toward the head of the line], but suddenly / they discovered / that all their money was missing Clifton and Ferreira (1994) reported negative evidence, that is, they did not measure reading time difference of the relevant segments in both sentences. They concluded therefore that pronoun coreference captures referential identity rather than syntactic representations. Referential identity refers to discourse representations or the mental model (situation model) the perceiver constructs during text comprehension. However, still one must take into account that this interpretation is based on reading time data that do not necessarily reflect initial access processes during spoken language comprehension. Alonso and Carreiras (1994) posed a similar question. They manipulated the superficial distance (S-structure) and the conceptual distance (D-structure) of the antecedent and pronoun in Spanish sentences. For example, the pronoun "elias" feminine-plural (fem-pl) must refer to a female antecedent. In (11) the pronoun must refer to antecedents mentioned in the first sentence, respectively to "Maria" and "Eva". In this case the female antecedents are superficially close. (10a) Maria y Eva compiten en un campeonato de atletismo en Barcelona Jaime compite en un campeonato de Judo en Madrid Elias han pasado a la final (Mary and Eve compete in athletic trials in Barcelona James competes in a Judo tournament in Madrid They (fem) have qualified for the final) (10b) Maria y Eva compiten en campeonatos de atletismo en Barcelona y Sevilla respectivamente Jaime compite en un campeonato de Judo en Madrid Elias han pasado a la final (Mary and Eve compete in athletic trials in Barcelona and Seville respectively James competes in a Judo tournament in Madrid They (fem) have qualified for the final) In (11) both referents are superficially distant because they appear in different sentences. However, in (10a) and (1 l a) the antecedents are conceptually close (Maria and Eva compete in Barcelona), but in (10b) and (1 l b) they are conceptually distant (Maria competes in Barcelona and Eva in Seville). (11 a) Maria y Jaime compiten en un campeonato de atletismo en Barcelona Eva compite en un campeonato de Judo en Madrid Elias han pasado a la final
Dieter Hillert
245
(1 l b) Maria y Jaime compiten en campeonatos de atletismo en Barcelona y Sevilla respectivamente Eva compite en un campeonato de Judo en Madrid Elias han pasado a la final If participants require less time to read sentences in which both female referents are presented in the same sentence (10), it would support a surface structure assignment strategy. Otherwise, if participants read sentences faster in which both antecedents are located in the same physical space (see 11), a conceptual assignment strategy would be plausible. The reading data indicate that pronoun coreference is assigned by a superficial strategy rather than by a conceptual one. However, as suggested by the authors, participants might have simply focused on the gender marker and their responses might reflect a training effect. Accordingly, Alonso and Carreiras (1994) conducted a follow-up study that appears to be less contaminated by task-specific strategies. After participants read the first two sentences, the relevant pronoun (ellas; fem-pl) was immediately displayed. They are asked to produce a meaningful continuation of the text they read before. In this experiment, Alonso and Carreiras indeed found an interaction between superficial and conceptual distance. Participants were faster to complete the phrase when the referents were superficially or conceptually close. The authors conclude therefore that both strategies may play a role in coreference assignment of overt, morphologically realized pronouns. Still it seems to be important to keep in mind that the experimental techniques used might not be sensitive enough to clarify whether both the morpho-syntactic (superficial) or the conceptual strategy are applied at the same time or during different processing stages.
Points of Discussion
We restricted the discussion of coreference assignment to a certain set of crosslinguistic phenomena. Different typological structures illustrate that the process of coreference assignment may involve different cognitive strategies. While in English coreference assignment relies strongly on the syntactic position of the antecedent, other languages, as we briefly exemplified, rely to different degrees on morphologically realized properties of the anaphoric relationship. We predict that certain coreference assignment strategies are dominant in certain languages. For example, it is reasonable to assume that native German speakers use case and gender information to (re)activate an antecedent; in Romance languages we predict, however, that an anaphoric relationship will be recognized by gender information. This does not imply that a conceptual strategy will not be used in these languages. Taking into account data on lexical access during real-time sentence processing, one may favor the prediction of a modular process model (Fodor, 1995; Hillert, 2002). In the first step, morphological and/or syntactic information may be sufficient to (re)activate the correct antecedent; if this strategy fails, a conceptual strategy will be applied that relies strongly on the plausibility of the correct antecedent. However, as
246
Cross-Linguistic Aspects of Anaphor Resolution
we discussed before, the range of evidence that supports this view is small. One reason is that there is only a small number of studies that examined coreference assignment in real-time. Another reason is that most experimental evidence relies on English data that do not allow conclusions beyond certain typological characteristics. In general, one may state that a conceptual strategy might be used across all languages. For example, if a listener must recognize novel NP referents as in (13), s/he is forced into use semantic-pragmatic information to select the correct antecedent. (13) The doctor asked the nurse to let enter the next x-ray in about five minutes Alternatively, listeners of highly inflected languages may rely strongly on morphological information to assign coreference. Thus, there are multiple morphological and syntactic strategies possible that depend on the typological characteristics of a particular language. Accordingly, we assume that in the case of bilingual processing, multiple process strategies are possible, depending on the type of native language and depending on the degree of non-native language competence. To exemplify a single case, let us predict the process strategies of native English speakers with non-native language competence in Spanish (ES speaker). We predict that the ES speaker is significantly delayed in accessing an antecedent in Spanish, if the coreference assignment relies solely on morphological characteristics such as gender information. In contrast to a native Spanish speaker, s/he would show significantly delayed lexical priming effects taking place several hundred milliseconds after encountering the anaphor. The ES speaker would rely to a certain degree, depending on the degree of Spanish language competence, on a conceptual strategy, or a syntactic strategy transferred from her/his native English language competence. Inversely, a native Spanish speaker with non-native language competence in English (SE) speaker wouldn't be able to use a morphological strategy to assign coreference in English, but would rely on a conceptual strategy. Because the SE speaker would need to acquire a syntactic strategy, that native English speakers use to assign anaphoric relationships, again a significant delay of antecedent (re)activation may be predicted.
Conclusion
The present discussion on anaphor resolution should be regarded as a starting point for further discussion and experimental research. Further research on bilingual anaphor processing requires explicit knowledge of how native speakers of typologically different languages assign coreference during real-time sentence processing. It is obviously that any universal model of anaphor resolution must take into account language-particular properties. Thus, to explore in-depth language processing as a cognitive and neuronal event, we must consider a linguistic profile that includes non-native language competence.
Dieter Hillert
247
References
Alonso, M.A., & Carreiras, M. (1994). Syntactic and semantic cues in plural anaphora resolution. Paper presented at the 7 th Conference of the European Society for Cognitive Psychology, Lisbon, Portugal. Ariel, M. (1985). Givenness marking. Unpublished doctoral dissertation, TelAviv University. Ariel, M. (1988). Referring and accessibility. Journal of Linguistics, 24, 65-87. Ariel, M. (1990). Accessing noun-phrase antecedents. London, New York: Routledge. Austin, J. L. (1962). How to do things with words. Cambridge, MA: Harvard. Bach, E. (1970). Problominalization. Linguistic Inquiry, 1, 121-122. Chomsky, N. (1982). Some concepts and consequences of the theory of government and binding. Cambridge, MA: MIT Press. Clifton, C., & Ferreira, F. (1994). Discourse structure and anaphora: Some experimental results. In M. Coltheart (Ed.), Attention and performance XII: The psychology of reading (pp. 23-37). Hillsdale, NJ: Erlbaum. Flynn, S., & Martohardjono, G. (1994). Mapping from the initial state to the final state: The separation of universal principles and language-specific properties. In B. Lust, M. Suner, & J. Whitman (Eds.), Heads, projections, and learnability (pp. 319-335). Hillsdale, NJ: Lawrence Erlbaum. Fodor, J. A. (1983). The modularity of mind An essay on faculty psychology. Cambridge: MIT Press. Fodor, J. D. (1995). Comprehending sentence structure. In L. R. Gleitman & M. Liberman (Eds.), An invitation to cognitive science, Language: Vol. 1. (pp. 209245). Cambridge, MA: MIT Press. Gazdar, G. J. M. (1979). A solution to the projection problem. In C. K. Oh & D. Dinneen (Eds.), Syntax and semantics: Vol. 11. Presupposition (pp. 57-90). New York: Academic Press. Gernsbacher, M.A. (1990). Language comprehension as structure building. Hillsdale, NJ: Erlbaum. Hillert, D., & Swinney, D. (2001). The processing of fixed expressions during sentence comprehension. In A. Cienki (Ed.), Conceptual structure, discourse, and language (pp. 107-121). Stanford, CA: CSLI. Hillert, D. (1998). Verb processing in English and German: The interface of parsing and lexical access during real-time sentence comprehension. In D. Hillert (Ed.), Syntax and semantics: Vol. 31. Sentence processing: A crosslinguistic perspective (pp. 247-263). San Diego, CA: Academic Press. Hillert, D. (1999). On processing lexical concepts in aphasia and in Alzheimer's disease. Brain and Language, 69, 95-118. Hillert, D. (2002). Biolinguistic inquiries. New York, NY: Academic Press. Hillert, D. (2001). Coreference assignment in Spanish. Unpublished manuscript, University of California, San Diego. Kintsch, W. (1988). The role of knowledge in discourse processing. A construction-integration model. Psychological Review, 95, 163-182.
248
Cross-Linguistic Aspects of Anaphor Resolution
Kintsch, W. (1998). Comprehension. A paradigm for cognition. New York, NY: Cambridge University Press. Kintsch, W., Welsch, D., Schmalhofer, F., & Zimny, S. (1990). Sentence memory: A theoretical analysis. Journal of Memory and Language, 29, 133-159. Lucas, M. (1999). Context effects in lexical access. A meta-analysis. Memory & Cognition, 27, 385-398. MacDonald, M. C., Pearlmutter, N. J., & Seidenberg, M. S. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101, 676-703. Nicol, J. (1988). Coreference processing during sentence comprehension. Unpublished doctoral dissertation, MIT. Nicol, J., & Swinney, D. (1989). The role of structure in coreference assignment during sentence comprehension. Journal of Psycholinguistic Research, 18, 5-19. Pynte, J. (1998). The time course of attachment decisions: Evidence from French. In D. Hillert (Ed.), Syntax and semantics: If"ol. 31. Sentence processing: A crosslinguistic perspective (pp. 227-245). San Diego, CA: Academic Press. Pynte, J., & Prieur, B. (1998). Prosodic breaks and attachment decisions in sentence parsing. Language and Cognitive Processes, 11, 165-191. Quine, W. V. (1960). Word and object. Cambridge, MA: MIT Press. Reinhart, T. (1984). Center and periphery in the grammar of anaphora. In B. Lust (Ed.), Studies in the acquisition ofanaphora (pp. 123-150). Dordrecht: Reidel. Searle, J. R. (1958). Proper names. MindLXVII, 266, 166-173. Seidenberg, M. S., Tanenhaus, M. K., Leiman, J. M., & Bienkowski, M. (1982). Automatic access of the meaning of ambiguous words in context" Some limitations of knowledge-based processing. Cognitive Psychology, 14, 489-537. Strawson, P. F. (1956). On referring. In F. Zabeeh, E. D. Klemke, & A. Jacobson (Eds.), Readings in semantics (pp. 159-192). Urbana: University of Urbana Press. Swinney, D. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18, 645-660. Swinney, D., Ford, M., & Bresnan, J. (1989). On the temporal course of gapfilling and antecedent assignment during sentence comprehension. In B. Grosz,, R. Kaplan, M. Macken, & I. Sag (Eds.), Language structure and processing (pp. 1224). Stanford, CA: CSLI. Tabossi, P. (1988). Accessing lexical ambiguity in different types of contexts. Journal of Memory and Language, 27, 324-340. Tanenhaus, M. K., Leiman, J. M., & Seidenberg, M. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18, 427-440. Van Dijk, T., & Kintsch, W. (1983). Strategies of discourse comprehension. New York, NY: Acdemic Press. Wilson, D., & Sperber, D. (1986). Pragmatics and modularity. In A. M. Farley, P. T. Farley, & K. E. McCullough (Eds.), CLS 22 Part 2: Papers from the parasession on pragmatics and grammatical theory (pp. 67-84). Chicago: Chicago Linguistic Society.
Part V: Figurative Language Processing
This Page Intentionally Left Blank
Bilingual Sentence Processing- R.R. Heredia and J. Altarriba (Editors) 9 2002 Elsevier Science B.V. All rights reserved.
11
Understanding Phrasal Verbs in Monolinguals and Bilinguals
Teenie Matlock University of California, Santa Cruz Roberto R. Heredia Texas A&M International University
Abstract
This study investigates how English monolinguals and bilinguals access lexical forms such as call on, which can be interpreted as either a phrasal verb, as in John called on the student, or as a verb-preposition combination, as in John called on the third floor. The results of an on-line sentence comprehension experiment indicate that both monolinguals and early bilinguals access phrasal verbs more quickly than verb-preposition combinations. The results also suggest that early bilinguals access both phrasal verbs and verb-preposition combinations as quickly as monolinguals do, whereas late bilinguals (even with high proficiency in English) do not. Everyday language is replete with phrasal verbs such as turn on, run off, or p u t up. There are an estimated 3,000 two- or three-part phrases such as these in English alone (McArthur & Atkins, 1974). These multi-word conventionalized forms-which on the surface look like a verb and a preposition -- feature a verb and a particle. They are often highly polysemous, and open to multiple interpretations. Some of the meaning uses of such a phrase can differ substantially from others, as demonstrated with turn on in sentence (1 a) below. (1 a) M a r y turned on C o l o m b o One possible interpretation of turned on in sentence (la) pertains to electronic operability. If the noun phrase Colombo is interpreted as the television drama starring Peter Falk, then turned on refers to pressing a button on a television remote control. However, if Colombo refers to a real person (or even the main character on Colombo), then turned on implies that Mary sexually aroused that person. Yet another reading of turned on is possible in sentence (1 a) if Colombo is interpreted as a person. Namely, at one point, Mary and Colombo were aligned in some way (for instance, politically), but later, Mary went against him or betrayed him. These are just a few meanings of turn on, .and each of these is figurative. To make matters even more complicated, we have to contend with the non-phrasal
252
Teenie Matlock and Roberto R. Heredia
verb, literal meaning of t u r n on in sentence (la) above, in which t u r n refers to rotation of a physical object, and o n to a proximal spatial relationship. This interpretation of t u r n o n is achieved for the sentence in (1 a) when one imagines that Mary is first driving down a street called F i r s t S t r e e t , and then changes direction by going down a street called C o l o m b o S t r e e t . In this case, t u r n on is not a phrasal verb, but what we will call a verb+preposition combination. Arriving at the appropriate interpretation of English two-word phrases such as t u r n on, which nearly always have a phrasal verb interpretation and a verb+preposition combination interpretation, would seem problematic. However, correct interpretation and appropriate use of these forms is rarely a problem for the native English speaker, even when there is little or no linguistic context. For instance, when hearing the command, T u r n it on, a listener standing next to a computer may automatically infer that it refers to the computer. In an entirely different situation, for instance, after hitting a ball into center field and running to first base, the same person may interpret T u r n it on (yelled by the first base coach) to mean keep running to second base, or run faster. In this case, it refers to running speed, not an electronic device. Depending on context and situation, including speakers' intentions and shared knowledge, native English speakers are readily able to access the correct meaning of a form such as t u r n on, even when there is little or no explicit linguistic information available. This ability is not limited to phrasal verb understanding. Language speakers resolve all sorts of ambiguity this way (e.g., Clark, 1996; Gibbs, 1999). And of course it is also possible that native English speakers are keenly aware of which meaning of phrasal verbs are dominant and which are secondary. For instance, for many of these ambiguous forms, the figurative, phrasal verb interpretation of a form such as t u r n on may be the default, and not the literal verb+preposition combination interpretation. Or it might be the case, that the default is the simultaneous activation of both meanings. The situation is not as straightforward for the second language (L2) learner or bilingual, especially when English proficiency is limited. For individuals whose first language (L1) is not English, even those with high proficiency in the language, understanding highly polysemous forms such as t u r n o n can pose problems. For instance, there is often confusion as to whether a word such as on, as in t u r n on, is functioning as a particle in a phrasal verb or a preposition in a verb+preposition combination. Similarly, there is also confusion as to the relationship between the figurative interpretation of a form such as t u r n on and the literal meaning of the verb without the particle ( t u r n by itself) (Matlock, Heredia, & McLaughlin, 1996; Sjoholm, 1995). In addition, bilingual individuals (or second language learners) who do not attempt to use phrasal verbs frequently or who use them inappropriately are viewed as being less proficient in English than those who use them frequently and correctly (Cornell, 1985). Thus, it appears that phrasal verb comprehension is not as natural for the bilingual, and part of the reason may relate to the absence of this form in L1. Because phrasal verbs are pervasive in English, but relatively uncommon in other languages, it might be that the phrasal verb form is not the default case for bilinguals. But it may also relate to not having a default interpretation for certain situations (e.g., in certain grammatical contexts),
Understanding Phrasal Verbs
253
or not knowing how to move fluidly from one meaning interpretation to another. The goal of this chapter is to examine how bilingual and monolingual English speakers understand and use phrasal verbs, such as turn on, which have an ambiguous verb-preposition interpretation. The article begins with a linguistic survey of phrasal verbs, and is followed by discussion of empirical research, including a report of results from two experiments. The article closes with discussion of the implications for currently popular processing theories and suggestions for further research.
Phrasal Verbs
Phrasal verbs and verb-prepositions share a common surface form (e.g., turn on in turn on a TV s h o w versus turn on a street), but semantically and syntactically, they behave differently. Phrasal verbs, for instance, adhere to certain syntactic constraints that do not apply to verb-preposition combinations. For example, consider the utterances with the frequently used ate up in sentences (2a) and (2b). (2a) B o b ate up the l a s a g n a (2b) Bob ate the lasagna up In sentences (2a) and (2b), we see phrasal verb uses of ate up. In (2a), the noun phrase the l a s a g n a is a direct object. It immediately follows ate up, and in (2b), it is situated between the verb ate and the particle up. In sentence (3a), we see a verbpreposition combination with ate up. In this case, the noun phrase the hill is an indirect object, not a direct object as the l a s a g n a was in sentence (2a). As such, it appears right after the preposition up. The sentence shown in (3b) is semantically infelicitous, because the hill is an indirect object and not a direct object. As such, it cannot occur between ate and up. (Throughout this paper, "??" specifies a grammatically anomalous or semantically odd sentence). (3a) B o b ate up the hill (3b) ?? B o b ate the hill up Another difference concerns phrasal verbs and anaphoric referents such as it or them. In such cases, a pronoun refers to an entity that has already been specified or which is known to both the speaker and listener. For instance, in sentence (4a), suppose it refers to lasagna, and that both the speaker and listener know it refers to lasagna. For transitive phrasal verbs (i.e., phrasal verbs that take a direct object, such as eat up, cut up, drop off), the pronoun must appear between the verb and the particle, as shown in sentence (4a). The pronoun cannot appear after the verb particle, as shown in (4b). Other examples with the transitive phrasal verb write d o w n are shown in (4c) and (4d). (4a) B o b ate it up (it = lasagna) (4b) ?? B o b ate up it (it - lasagna)
254
Teenie Matlock and Roberto R. Heredia
(4c) Carlos wrote them down (them = phone numbers) (4d) ?? Carlos wrote down them (them = phone numbers) In addition, phrasal verb constructions usually do not welcome adverbs between the verb and particle. For example, the sentence in (5a) sounds strange because the adverb "quickly" is situated between the verb (ate) and the particle (up). For verbpreposition combinations, however, this is usually not a problem, especially in informal English, and when there is sufficient contextual information (see sentence
5b). (5a) ?? Bob ate quickly up the lasagna (5b) Bob ate quickly up the street at Romano 's, and then rushed off to work This is also true of reflexives. For instance, Bob wiped himself off is correct, but ? ?Bob wiped off himself is not acceptable (at least in most dialects and situations). There are additional syntactic differences between phrasal verbs and verb+preposition combinations. Some differences relate to syntactic flexibility of the verb phrase. Namely, there is a greater degree of flexibility with verb+preposition combinations than there is with phrasal verbs. As shown in sentence (6a), a syntactic variant of sentence (3a), the prepositional phrase, which includes the preposition up and the indirect object the hill can occur in sentenceinitial position. This is not possible with the phrasal verb counterpart, as shown in sentence (6b), a syntactic variant of sentence (2a), in part because ate and up form a semantic unit. ~This pattern is also observed in idioms (see Gibbs, Nayak, & Cutting, 1989). (6a) Up the hill is where Bob ate (6b) ?? Up the lasagna is what Bob ate Thus, even though on the surface, phrasal verbs and verb-preposition combinations are identical in form, they differ in many respects when it comes to their syntactic behavior (see also Cornell, 1985). Phrasal verbs and verb-prepositions also differ with respect to meaning. Each word in a verb+preposition combination has its own meaning, and each meaning is transparent, or clear. For instance, the word up in sentence (3a) above refers to vertical position or ascension. It specifies where Bob ate. To some extent, its meaning is independent of the verb's meaning. The meaning of up in sentence (2a) is less transparent. It does not refer to vertical orientation or ascension, and it does not have a meaning that is independent of the verb. However, this is not to say that up in this case has no meaning on its own. As a matter of fact, Lindner (1983) convincingly argues that in cases such as eat up, the particle up means completion. This type of alternation is not unique to phrasal verbs. It occurs with sentence pairs such as John walked down the street and Down the street John walked (for further transformations, see Perlmutter & Soames, 1979).
Understanding Phrasal Verbs
255
Together with the verb eat, up highlights the end of eating. This is evidenced by sentences (7a) and (7b). Notice however, that sentence (7b) sounds somewhat unusual because the word up highlights the completion of eating. Thus, in terms of semantics, phrasal verbs also contrast with their verb-preposition combination counterparts. (7a) Bob ate the lasagna until he was full (7b) ?? Bob ate up the lasagna until he was full
Phrasal Verbs and the Lexicon
Within both linguistics and psycholinguistics there is a general agreement that phrasal verbs are lexicalized or stored in the lexicon similar to the way that words or idioms are stored (e.g., Jackendoff, 1995). When hearing a phrasal verb such as turn on, the listener retrieves information about argument structure, including whether the phrase takes a direct object, and if so, where a pronoun can be placed, for instance, not after the particle, but after the preposition, as shown in sentences (4a) and (4b). Sometimes this gets quite involved, especially when it comes to separability. Some phrasal verbs are separable, eat up, as in sentences (2a) and (2b). In such cases, the object can appear between the verb and particle, or after the particle. Other phrasal verbs are not separable. In these cases, the object must appear after the particle, as with call on in The teacher called on the student and not between the verb and the particle (??The teacher called the student on). In addition, not all phrasal verbs are transitive. Some do not take an object, such as speak up, as in He spoke up. Though there is consensus that phrasal verbs or idioms are lexicalized, and that their understanding involves connecting grammatical information and semantic information, there is little agreement on exactly how these multi-word phrases are stored, accessed, and retrieved. Three psycholinguistic theories attempt to explain the manner in which phrasal verbs and idioms are retrieved from semantic memory. According to the lexical representation hypothesis (Swinney & Cutler, 1979), idioms are stored as individual items and are retrieved in the same way as words. In this view, the retrieval of an idiom is initiated as soon as the first word of an idiomatic expression is encountered. The computation of the literal phrase and the computation of the idiomatic phrase run in parallel for a given expression (e.g., Swinney & Cutler, 1979). Note that researchers in this domain make a theoretical distinction between literal and idiomatic or nonliteral meaning. Thus, according to this distinction, the idiomatic expression in sentence (8a) can be understood in terms of its idiomatic interpretation (to die) or the literal interpretation (to knock over a pail with one's foot). Thus, according to this view, both meanings of the idiomatic expression become active simultaneously. (8a) My grandmother kicked the bucket
256
Teenie Matlock and Roberto R. Heredia
A second hypothesis, the literal processing model, or the idiom list hypothesis (Bobrow & Bell, 1973), holds that idioms or idiomatic expressions are stored and accessed from a special list which is not part of the regular lexicon (Swinney & Cutler, 1979). According to this hypothesis, during idiom comprehension the listener first attempts to construct a literal interpretation of the string of words. If this process fails, an idiom's mode is instantiated and the idiom meaning is recovered from the idiomatic list. An alternative view is offered by the direct access hypothesis (Gibbs, 1980, 1986). In this view, idiom comprehension does not occur in parallel with literal comprehension. Instead, the idiomatic interpretation precedes the literal interpretation. A literal interpretation results only if the idiomatic interpretation fails because of incongruent contextual cues. Thus, when hearing kick, of kick the bucket, the idiomatic interpretation is accessed first. If it turns out that it is inappropriate (e.g., context refers to a painter knocking over a pail), then the idiomatic meaning is abandoned and the literal one is activated. However, in order for this hypothesis to be true, one would have to distinguish between high and low familiar idiomatic expressions, in which highly familiar idioms would behave according to this hypothesis (cf. Cacciari & Tabossi, 1988; Titone & Connine, 1994). Although both the lexical representation and the direct access hypotheses reject the idea that processing literal language is faster than processing figurative language, and that processing idiomatic meaning takes less time than accessing a non-idiomatic meaning, they have different underlying assumptions. However, both approaches have been criticized for not employing sensitive enough measures for online studies (e.g., Cacciari & Tabossi, 1988). For instance, in both Swinney and Cutler (1979) and Gibbs (1986) participants made a decision about an ambiguous (literal versus idiomatic) target sentence such as He kept it under his hat only after processing it. Nonetheless, studies in either area have revealed that literal interpretations of multi-word expressions are no more difficult to process than figurative interpretations (but see Titone & Connine, 1994).
Bilingualism and Figurative Language Processing Although much of the work in bilingualism has looked at how the bilingual lexicon is organized (see Heredia, 1997; Kroll & Stewart, 1994), some studies on bilingualism and figurative language suggest interesting relationships between fluency in the second language and the comprehension of figurative language. For example, Johnson and Rosano (1993) gave second language learners metaphors such as My shirt is a mirror and asked them to understand what the metaphor meant and to provide as many interpretations of the metaphorical expression as possible. Results showed that metaphor fluency (the number of interpretations provided) was positively correlated withmeasures of second language communicative proficiency and general language proficiency. More specifically, processing capacity and relevant knowledge were the major factors determining complexity level in the bilinguals'
Understanding Phrasal Verbs
257
interpretations. Other studies, which examined experienced versus non-experienced second language learners (bilinguals learning a second language) and metaphor interpretation, show experienced (L2) learners performing at monolingual levels (Johnson, 1989; Nelson, 1992). Moreover, other studies investigating the comprehension of idiomatic expression by advanced L2 learners have found that idioms that are identical or literal translations across English and Spanish (e.g., to be all ears vs. ser todo oido) provide the least difficulty. Idioms that are similar across the two languages are the second most difficult and idioms that are different across the two languages are the most difficult to comprehend. Overall, these studies suggest that language knowledge or language proficiency is related to the comprehension ofnonliteral language. The more experience in the L2, the easier it becomes to use and comprehend nonliteral language. In relation to figurative language processing, the data reviewed here suggests that bilinguals at early stages of L2 learning must establish direct connections between literal and nonliteral meanings of the figurative part of speech. So, during the comprehension of an idiomatic expression, the beginning bilingual engages in the following stages: (a) translate the given idiomatic expression literally into L1, (b) access the literal meaning of the idiomatic expression and attempt to make sense of it, and (c) access the figurative meaning if there is one. However, at more advanced stages of L2 learning, the bilingual speaker may come to behave in the same manner as a native monolingual speaker (c.f. Heredia, 1997; Heredia & Altarriba, 2001).
Bilingualism and Phrasal Verb Understanding Research has shown that phrasal verbs present problems for bilinguals or second language learners. One problem relates to avoidance. While communicating in English, proficient second language learners avoid using phrasal verbs. Instead, they rely on a single word or a concatenation of words. This is true when the first language does not contain lexicalized phrases such as phrasal verbs, or perhaps even when the first language is genetically unrelated to English. For instance, Dagut and Laufer (1985) conducted a study in which participants were given three tasks. One was a multiple-choice cloze task, in which participants had to fill in a missing blank with either a phrasal verb, a one-word verb, or one of two distractors. Another was a memorization task, in which participants first memorized lists of words or phrasal verbs, and ten minutes later filled in blanks (with either form). A final task required Hebrew participants to translate into English the missing word from their own language. The results of the Dagut and Laufer study showed that Hebrew speakers learning English as a second language preferred using one word (e.g., disappoint, confuse) over a semantically equivalent phrasal verb (e.g., let down, mix up). The authors attributed this avoidance to the absence of phrasal verbs in Hebrew. The type of avoidance seen in Hebrew ESL learners is also seen in situations where the first language is genetically related to English, and even in situations in which the first language has phrasal verbs, such as Dutch. Hulstijn and Marchena (1989)
258
Teenie Matlock and Roberto R. Heredia
conducted the same studies in their research on English phrasal verbs with Dutch ESL speakers, and found similar results to those of Dagut and Laufer. This was found to be the case even in situations in which an English speaker would prefer using a phrasal verb over a one-word verb, and the tendency was more pronounced in intermediate Dutch ESL learners than in advanced Dutch ESL learners. Hulstijn and Marchena attributed the problem to a perceived difference related to the overall semantic characteristics of phrasal verbs and their one-word counterparts. In particular, they argued that Dutch speakers avoid phrasal verbs (e.g., let down) because these forms are perceived as having a more specialized, idiomatic meaning than one-word phrases (e.g., disappoint). Drawing from Kellerman (1977), Hulstijn and Marchena also claimed that Dutch ESL learners avoided using phrasal verbs because they feared making an error, or because they perceived the second language form as being too similar to a form in the first language. In sum, regardless of the first language, research suggests that ESL learners avoid using a phrasal verb when an equivalent one-word expression is available. Sjoholm (1995) found similar results in a study comparing Swedish ESL learners and Finnish ESL learners on their understanding of English phrasal verbs. To see how these ESL learners compared, Sjoholm had participants do a multiple-choice task. In this task, participants were given English sentences without verb phrases (e.g., The dentist had to one of this front teeth), and were asked to fill in the blanks. In making their decisions, participants chose from one of the following items: a phrasal verb (e.g., take out), a synonymous one-word verb (remove), or two incorrect distractor verbs. Many of the phrasal verbs offered in the list were intended to bear a resemblance to Swedish phrasal verbs. The results showed that both the Finnish and Swedish participants performed well overall, but the Finns made significantly more errors than the Swedes. In addition, Sjoholm found that overall, Swedes chose a phrasal verb over a one-word Synonym more often than the Finns did, suggesting that Swedes had a stronger preference for phrasal verbs due to their familiarity with phrasal verbs in their first language. It is important to note that although phrasal verbs exist in Swedish and other Germanic languages, English surpasses other Germanic languages with respect to productivity, frequency of use, and semantic range (see Bolinger, 1971). Apparently, Finns did not choose phrasal verbs as often because Finnish shares fewer properties with English than Swedish does. Swedish and English are both Germanic languages, while Finnish is a Finno-Ugric language. Consequently, Finnish ESL learners were less familiar with phrasal verbs than Swedes were. The only Finns who preferred to use phrasal verbs were the most advanced ESL learners in that group, suggesting the influence ofthe first language gets diluted over a long period of time. To avoid making errors by using this form, they simply avoided using it whenever they could (See Sjoholm, 1995). Moreover, avoidance oflexicalized multi-word forms by ESL learners and bilinguals is not unique to phrasal verbs. It has also been seen with idioms. As reported by Irujo (1993), Spanish dominant Spanish-English bilinguals are unable to produce an English idiom when asked to provide one except for cases where a virtually identical idiom exists in their first language (See also Yorio, 1989). Even considering the aforementioned work on avoidance of phrasal verbs by ESL learners, little psycholinguistic work has been devoted to phrasal verb
Understanding Phrasal Verbs
259
comprehension in bilinguals or even monolinguals. Based on the research we surveyed, no studies appeared to specifically address the understanding and use of lexically and syntactically ambiguous forms such as eat up, shown in sentences (2a) and (2b). The goal of the current study is to investigate how bilinguals and monolinguals use these ambiguous forms. Of particular interest is the likelihood of generating a phrasal verb interpretation for a form such as eat up. To this end, in Experiment 1, we investigate whether monolinguals and bilinguals are equally likely to produce a phrasal verb when asked to complete a sentence with an ambiguous form such as eat up or turn on (e.g., John ate up ). Perhaps monolinguals will be more likely to supply a phrase such as the pizza, which implies a phrasal verb interpretation, than they will a phrase such as the hill, which implies a verb+preposition interpretation. If so, it will suggest that monolinguals tend to activate the figurative, phrasal meaning first. In addition, monolinguals might be more likely to supply a phrasal verb interpretation than bilinguals will. This would suggest that retrieving a phrasal verb reading is more natural for a monolingual than it is for a bilingual. Also of interest is whether a bilingual will be more likely to use a verb+prepositional form than a phrasal verb form. If so, further evidence may be provided to support the idea that bilinguals retrieve the literal meaning first. Extending the same line of thinking to an on-line reading time study, in Experiment 2, we wanted to compare monolinguals and bilinguals on how quickly they process sentences with phrasal verbs (e.g., look up, in Teri looked up the pilot's address on the computer), and sentences with verb+preposition combinations (e.g., look up, in Teri looked up her friend's chimney with a flashlight).
Experiment 1 Method
Participants. Seventy-six participants, including 50 monolinguals and 26 bilinguals volunteered for the experiment. The bilingual participants represented 11 different native language backgrounds. Their language backgrounds are presented in Table 1. None of the bilingual participants reported having had exposure to English before the age of 7. Stimuli. The stimuli consisted of 27 partially completed sentences. Each contained a subject noun phrase that referred to a person, an ambiguous phrasal verb/verb + prepositional form, and a blank for a direct or indirect object, for . For consistency, in all cases, the past tense was used. instance, John ate up In addition to these primary stimuli, there were 16 filler sentences. All sentences were randomly ordered. Stimuli are included in the Appendix. Procedure. The experiment was conducted on the World Wide Web (WWW) (for a review of experiments on the WWW, please see, Birnbaum, 2000). Participants accessed the experiment by clicking on a hyperlink from a web site that listed psychology experiments on the web, or they heard about the experiment
260
Teenie Matlock and Roberto R. Heredia
through word of mouth. First, participants were asked to complete a survey that contained questions about language background, including when they began learning English. After completing the survey, participants were instructed to read each partially completed sentence, and to complete each one by typing in a word or a string of words. The partially completed sentences on the list were randomly ordered so that no participant would see the same order. Participants were instructed to work quickly, and were told they could not use a dictionary or consult native English speakers.
Table 1. Number of L1 Bilinguals in Experiment 1 Language
Number
Chinese Hebrew Indonesian Italian Korean Portuguese Russian Spanish Tagalog
4 1 1 1
2 11 1 4 1
R e s u l t s and D i s c u s s i o n
Data from 5 participants were omitted from the analysis because they contained over 30 percent errors. For instance, each completed sentence that was nonsensical or ungrammatical, such as John ate up a radio sale or John ate up he's hungry, counted as an error. These responses were assumed to have been contributed by English language learners with limited proficiency or people who did not take the experiment seriously. Correct responses included well-formed completed sentences. Completed sentences such as John ate up the pizza were coded as phrasal verb responses, and given a value of 2. Completed sentences such as John ate up the street were coded as verb + preposition responses, and given a value of 1. A single score, the phrasal verb ratio, was computed for each participant. This score was obtained by dividing the sum of phrasal verb and verb + preposition forms by the total number of responses (excluding errors). Scores on this metric, range from 1 (indicating all verb + preposition responses) to 2 (all phrasal verb responses). Seven participants did the experiment under typical laboratory conditions (4 monolingual and 3 bilingual). Their data were consistent with those
Understanding Phrasal Verbs
261
obtained from participants in the web survey. The performance of the monolinguals was compared to that of the bilingual group. We took into account the possibility that a participant would respond at random. If so, the phrasal verb ratio would be 1.5. A one-sample t-test was conducted for each group to determine if the group mean differed from random behavior. The results showed that the mean for the monolingual group (M = 1.76), differed significantly from 1.5, t(49) - 16.81, p < .001, and so did the mean for the bilingual group (M = 1.66), t ( 2 0 ) - 4.79, p <. 001. Comparison of the two groups indicated that monolinguals were significantly more likely to produce a sentence with a phrasal verb than bilinguals were, t(69)= 2.98, p < .01. In general, the results show that when given an ambiguous form such as eat up, bilinguals are less likely to initially produce a phrasal verb interpretation than monolinguals. Although there was a clear difference in the production of phrasal verb interpretations by monolinguals and bilinguals, the bilinguals, as a group, produced more phrasal verb interpretations than verb+prepositional forms. These results suggest that bilinguals in this experiment were proficient enough in English to do well on the task. In relation to whether or not bilinguals were able to directly produce the phrasal verb interpretation, our study did not address this particular issue. However, the hypothesis that bilinguals in this study were relying on the literal interpretation (i.e., verb+prepositional form) and then generating the figurative (phrasal interpretation) cannot be rejected. That is, it is possible that bilinguals were activating the literal meaning of a form such as eat up before producing the figurative meaning. This explanation is consistent with our proposed hypothesis that during nonliteral or figurative language comprehension, bilinguals rely on the literal interpretation before accessing the figurative meaning (cf. Irujo, 1993). In any event, phrasal verbs appear to be more naturally produced by monolinguals than by bilinguals.
Experiment 2 The goal of this experiment was to explore the understanding of ambiguous forms such as eat up in an on-line reading task. In part, we wanted to see if monolinguals and bilinguals would read a sentence with a phrasal verb more quickly than they would a sentence with an identical verb+preposition, such as go over, for instance, P a u l went over the bridge with his bicycle (verb+preposition) and Paul went over the exam with his students (phrasal verb). Such a result would suggest a difference in how these ambiguous forms are processed. Of interest was whether bilinguals would comprehend sentences with phrasal verbs as quickly as monolinguals. Such a result would suggest a similar processing pattern for activating lexical and syntactic information for these ambiguous forms. On the other hand, a difference would suggest a different way of activating lexical and syntactic information by the two groups.
262
Teenie Matlock and Roberto R. Heredia
Method
Participants. Fifty-seven undergraduate students from the University of California, Santa Cruz, participated in the experiment to meet a requirement in a psychology course. Of these, 22 were monolingual, and 35 were bilingual. Bilinguals came from a wide range of first language backgrounds, as shown in Table 2. Over 25 percent of the bilingual participants were Spanish-English bilinguals (15).
Table 2. Number of L1 of Bilingual Participants in Experiment 2 Language
Early Bilingual
Late Bilingual
Amharic Chinese Farsi French German Portuguese Romanian Russian Spanish Tagalog Vietnamese
1 5 0 1 0 0 0 1 12 2 0
0 4 1 0 1 1 1 1 3 0 1
Stimuli and design. The stimuli included 80 pairs of English sentences. The first sentence of each pair was the target sentence. Each target sentence contained an ambiguous two-word verb form that could be interpreted as a phrasal verb or a verb+preposition, for instance go over. The two-word verb forms were taken from the Longman Dictionary o f Phrasal Verbs (1983). The second sentence in each pair was either a true or a false paraphrase of the first sentence (half the paraphrases were true and half were false). The purpose of the paraphrase was to make sure that participants would pay attention to the target sentence. For instance, a sentence such as Paul went over the exam with his students was paired with a (false) sentence paraphrase such as Paul had to take the exam. In addition, for every sentence pair with a phrasal verb target sentence there was a sentence pair with a verb+preposition target sentence, for instance, as Paul went over the exam with his students and Paul went over the bridge with his bicycle. This formed the verb type factor. All corresponding target sentences (e.g., sentence with phrasal verb go over and with verb+preposition go over) were matched for word length. All target sentences
Understanding Phrasal Verbs
263
contained a subject noun phrase that referred to a human (e.g., Paul). All sentences appeared in the simple past tense. The prepositions and particles used were limited to in, out, off, over, and into. Finally, half the target sentences featured the direct object between the verb and the particle, as with ragingfire in Herb put the raging fire out in only seconds. The other half of target sentences featured the direct object after the particle, as with the exam in Paul went over the exam with his students. Stimuli are included in the Appendix. Procedure. Each participant completed a questionnaire that included queries relating to first language and first exposure to English. After filling out the questionnaire, participants sat in front of an Apple Macintosh Quadra 650 computer that was running PsyScope (Cohen, MacWhinney, Flatt, & Provost, 1993). Participants were instructed to read the first sentence of the pair (target sentence), and to press a button to see the next sentence, which they would read and decide as to whether it was an appropriate paraphrase. Pressing a green "yes" button indicated an affirmative response. Pressing a red "no", button indicated a negative response. After pressing one of the decision keys, the participant was presented with the prompt, "Ready?" which marked the beginning of a new trial. All sentence pairs were randomized so that no participant would see the same order. Before beginning the actual experiment, each participant was exposed to two practice trials.
Results and Discussion
Reaction time measured the time taken to read and comprehend the target sentence. Response times for the second sentence were not analyzed because as mentioned, they merely served to ensure the participant would pay attention. We were only interested in the time taken to read a sentence with a phrasal verb or a verb+preposition. All incorrect responses and reaction times below 1 second or above 15 seconds were excluded because these clearly indicated that the participant did not understand the sentence or was not paying attention. Two participants had over 30 percent errors (incorrect response to paraphrase), and their data were not included in the final analysis. Overall error rates (in percentages) are presented in Table 3. Mean reading times in milliseconds are shown in Table 4. A paired t-test comparing the monolinguals on verb type revealed a significant effect for verb type, t(21) = 4.43, p < .001, indicating that reading times were quicker for phrasal verb sentences than for verb+preposition sentences. A t-test was also conducted to see how all bilinguals (both groups combined) compared to monolinguals. Only a marginally significant difference was obtained, t(34) = 2.05, p < .06, suggesting that monolinguals' reading times were overall faster than bilinguals'. Pilot data indicated that further analysis would benefit by separating the bilinguals into an early bilingual group and a late bilingual group. Bilinguals who began speaking English before 12 were placed into the early bilingual group and those who began speaking English after 12 were placed into the late bilingual
264
Teenie Matlock and Roberto R. Heredia
group. Such divisions are common in bilingual research (see Asher & Garcia, 1969; Extra & Verhoeven, 1994; McLaughlin, 1994; Vaid & Genesee, 1980).
Table 3. Overall Error Rates in Experiment 2 Group English Monolingual (n = 22) Early Bilingual (n = 22) Late Bilingual (n = 13) Total Bilingual (n = 35)
Percent Errors 3.3 5.8 7.5 6.5
Table 4. Mean Reading Times: Group x Verb Type in Experiment 2 Verb Type
Group
Phrasal Verb
Verb Phrase
English Monolingual (n = 22) Early Bilingual (n = 22) Late Bilingual (n = 13) Total Bilingual (n = 35)
4312 4140 6882 5159
4635 4354 7053 5356
To determine whether there was a difference in reading times between the two bilingual groups, we conducted a 2 (bilingual type: early vs. late) x 2 (verb type: phrasal verb vs. verb phrase) ANOVA. The results indicated a main effect of group, F(1,33) = 37.87, p < .05, with faster reading times for the early bilingual group. No effect was observed for verb type. Two paired t-tests were also run within each bilingual group to see if there was an effect for verb type. For the early bilingual group, a significant difference was obtained, t(21) = 2.18, p < .001, with faster reading times for phrasal verb sentences. No difference was observed for the late bilingual group. Finally, a 2 (language group: monolingual vs. early bilingual) x 2 (verb type: phrasal verb vs. verb phrase) ANOVA was conducted to compare the reading times
Understanding Phrasal Verbs
265
of the monolingual group to those of the early bilingual group. No significant differences were noted, suggesting that the early bilinguals were equally as quick in comprehending the target sentences as the monolinguals were. Taken together, the results for monolinguals and early bilinguals indicate that phrasal verbs such as eat up and go over are initially activated more quickly than their seemingly identical two-part verb phrases with a verb+preposition interpretation. The results also show that this is not the case for late bilinguals. This disparity suggests a difference in how these ambiguous forms are processed. There are two plausible explanations for this. It is not unreasonable to assume that for monolinguals and early bilinguals, the figurative, highly conventionalized meaning is always activated before the literal meaning. If the phrasal verb meaning that is activated coheres to the current context, then the literal meaning is not entertained. An equally plausible explanation relates to analyzability. It is possible that the literal reading takes longer to process because each word meaning in the phrase (e.g., the meaning of go and the meaning of over) must be activated. For the late bilingual group, it is possible that the literal meaning is always activated first, regardless of context. To be on the "safe side," bilinguals initially activate the literal meaning.
General Discussion
Overall, the results from the sentence completion task in Experiment 1 and the on-line reading task in Experiment 2 suggest differences in the processing of ambiguous phrases such as eat up and go over. In Experiment 1, both monolinguals and bilinguals generated a phrasal verb interpretation more than they did a verb+preposition interpretation. Similarly, in Experiment 2, both monolinguals and early bilinguals read sentences with phrasal verbs more quickly than they read sentences with an ambiguous verb+preposition. However, late bilinguals in Experiment 2 did not read the phrasal verb sentences any faster than they read the verb+preposition sentences. How might these results be interpreted? Phrasal verbs are highly conventionalized linguistic forms. Based on our results (especially for Experiment 2) it is possible that for monolinguals and early bilinguals, the figurative meaning, which is assigned to the entire phrase, is activated first. Thus, when processing the phrasal verb, go over in context, as in Paul went over the exam with his students, the reader activates the meaning of go over as a unit, not the meaning of go and over individually. This interpretation of the data is consistent with the direct access hypothesis for processing idioms (Gibbs 1980), which is based on similar results. When reading idioms, such as kick the bucket, for instance, people process the figurative meaning (to die) more quickly than the literal meaning (to knock over a pail). Even when given a literal context, it appears that the idiomatic interpretation is processed first (Gibbs, 1980, 1986, 1995). Another possible explanation is that at the onset monolinguals and early bilinguals start activating both the literal and figurative interpretations of an ambiguous form such as go over, and then settle on
266
Teenie Matlock and Roberto R. Heredia
the meaning that is congruent with the context. If the appropriate meaning is a phrasal verb meaning, less time is required because the meaning of the entire phrasal verb is activated much like a single word is activated. If the appropriate meaning is a verb+preposition meaning, more time is required because the meaning for each word (go, over) must be activated Of course, based on our data alone we cannot make conclusions for either of these explanations. Though the data from our experiments show a trend, namely, that phrasal verb activation occurs more readily or more quickly than verb+preposition activation, there are methodological issues that need to be considered, especially for Experiment 2. One concern relates to the type of measurement we used. In Experiment 2, we measured time taken to read the entire sentence. As Ferreira and Anes (1994) have pointed out, reading tasks of this type fall at the low end of acceptability for on-line processing. Our psycholinguistic technique (time taken to read the entire sentence) may not have been sensitive enough to capture what was going on during the actual processing of the ambiguous form. A more sensitive measure, for instance, one that employs both the spoken and visual modalities would have been more preferable as compared to the measure we used. For instance, the cross-modal lexical priming paradigm (Swinney, 1979; Tabossi, 1996) used for idiom understanding by Cacciari and Tabossi (1988) and Titone and Connine (1994) might be successfully adapted for the understanding of phrasal verbs and verb+prepositions and bilingualism. This technique, for example, has been shown to be successful in measuring the activation of multiple meanings and differences between literal and figurative meaning, and has been used successfully in addressing bilingual sentence processing issues (see Heredia & Stewart this volume). Future research on the on-line processing of phrasal verbs should utilize psycholinguistic techniques with a greater degree of sensitivity to get at processing differences that arise in various stages of second language learning in bilinguals. For instance, in Experiment 2 we observed no difference in reading times for sentences with a phrasal verb or a verb+preposition form for late bilinguals, but we observed differences for early bilinguals. Perhaps, there are processing differences between these two phrases, but our task was not able to detect such differences simply because our measurement was simply at the offset of the entire sentence. It might be the case that there would have been a difference in processing times if activation had been measured at different places along the course of the entire sentence. It is possible that the phrasal verb interpretation remains activated slightly later in the stream of words for late bilinguals than for early bilinguals. In addition, for in-depth research in this area, we need to consider a more careful analysis of the components of the ambiguous phrasal verb/verb+preposition form. Work on idioms has begun to examine the subparts of an idiom, and this technique should also be carried over to work on phrasal verbs, for as Jackendoff (1995) argues, phrasal verbs are small idioms. The type of work we are thinking about is seen in Hamblin and Gibbs (1999), which focuses on the semantic contribution of the verb, and how it influences the understanding of an idiom. For example, the action expressed by kick (quick action, involves a human) in the
Understanding Phrasal Verbs
267
idiom kick the bucket helps understand the meaning of that idiom. The same attention should be paid to the semantic contribution to the verb and to the particle in phrasal verbs. As Linder (1983), Morgan (1997), and other linguists have shown, even though phrasal verbs have a meaning that is associated with the entire phrase, the subparts, in particular the particle, contribute semantic information to the overall meaning. One form of evidence is based on semantic regularities observed in language. For instance, as mentioned in the introduction, the particle up often contributes information about completion, as in drive up and dry up. As an example, consider the difference between drive up and drive. The former accentuates the aspect of completion. Hence, because it emphasizes completion of the act, one does not use the present progressive -- which emphasizes the on-going nature of an event -- with this form (e.g., ?? He has been driving up), but there is no problem with using the simple past tense (e.g., He just drove up). More attention should also be given to the degree to which a phrasal verb is idiomatic. Some phrasal verbs are highly idiomatic, such as get on (e.g., He got into drugs at the age of 13, She really got into it with another faculty member), while others are barely idiomatic at all, such as rush off(e.g., He rushed off to work). In addition, highly polysemous phrasal verbs have some meanings that are very idiomatic and others that are less so, for example, hang up (e.g., John hung up the laundry, Maria is hung up at work). This range of idiomaticity may be another reason second language learners and bilinguals avoid using phrasal verbs or attempt to understand a literal, non-phrasal verb meaning first. It is simply easier to be "safe" and go for a one-word form, which is likely to have less variability (see Sjoholm, 1995). Only careful experimentation that targets a range of idiomaticity will tell. Finally, psychological studies on phrasal verbs have not paid attention to the polysemous nature of the subparts of the phrasal verb. Across the board, phrasal verbs recruit as their verb subpart words that are highly polysemous in English. One rarely hears or uses phrasal verbs that contain words with only one or two meanings, such as administer up or facilitate out. Phrasal verbs, which are Germanic in origin, most often refer to basic, often bodily activities, for instance, turn on, go over, and dry up (Bolinger, 1971). For a better understanding of the semantic interpretation of these forms and to other conventionalized multi-forms, closer attention needs to be paid to the polysemous nature of the subparts, and how that might influence on-line understanding. For instance, how might a phrasal verb with a verb subpart with 30 possible meanings differ from one with a verb subpart with two meanings? The polysemous nature of both subparts is important because it can drive the meaning of the multi-word form as a whole (e.g., Matlock, 1998). Moreover, interpretation of an ambiguous form such as turn on also depends on the range of polysemy available with the verb+preposition meaning. Representational issues abound with polysemous word forms like these, and there is no clear-cut way to divide things up in many cases (See Gibbs & Matlock, 1999, 2001). Nonetheless, one experimental approach to get at the difficulty ofpolysemy might be to employ the methods used in Frisson, Sandra, Brissard, and Cuyckens (1996). In their study, Dutch speakers were asked to learn the literal meaning of
268
Teenie Matlock and Roberto R. Heredia
several polysemous spatial expressions in an unfamiliar artificial language. Later, when participants were given the same words in non-literal, non-spatial contexts, they were able to correctly identify their new semantic extensions, suggesting that learning one meaning of a polysemous word has an influence on learning another meaning. With phrasal verbs, it might be the case that knowing a literal meaning of a polysemous verb might facilitate the understanding of a phrasal verb containing the same verb, especially if the phrasal verb is not particularly idiomatic. It might also be the case that knowing the meaning of the verb might also hinder the understanding of a phrasal verb with the same verb, for instance, if the phrasal verb was highly idiomatic, in which case, the meaning of the verb and the phrasal verb might be completely different. Taken together, the results suggest processing differences between bilinguals and monolinguals in the comprehension of phrases such as turn on, which have phrasal verb and verb+preposition interpretations. The results are in line with avoidance research and with idiom research. Further work should serve to clarify and solidify our preliminary findings here.
Author Notes
Teenie Matlock is now a Postdoctoral Researcher in the Psychology Department at Stanford University.
References
Asher, J., & Garcia, R. (1969). The optimal age to learn a foreign language. Modern Language Journal 8, 334-341. Birnbaum, M. H. (Ed.). (2000). Psychological experiments on the internet. San Diego, CA: Academic Press. Bolinger, D. (1971). The phrasal verb in English. Cambridge, MA: Harvard University Press. Bobrow, S., & Bell, S. (1973). On catching on to idiomatic expressions. Memory & Cognition, 1, 343-346. Cacciari, C., & Tabossi, P. (1988). The comprehension of idioms. Journal of Memory and Language, 27, 668-683. Clark, H. H. (1996). Using language. Cambridge: Cambridge University Press. Cohen, J. D., MacWhinney, B., Flatt, M., & Provost, J. (1993). PsyScope: A new graphic interactive environment for designing psychology experiments. Behavioral Research Methods, Instruments, and Computers, 25, 257-271. Cornell, A. (1985). Realistic goals in teaching and learning phrasal verbs. IRAL, 23,269-280. Dagut, M., & Laufer, B. (1985). Avoidance of phrasal verbs--A case for contrastive analysis. Studies in Second Language Acquisition, 7, 73-79. Extra, G., & Verhoeven, L. (1994). Cross-linguistic perspectives on bilingual
Understanding Phrasal Verbs
269
development. In G. Extra & L. Verhoeven (Eds.), The cross-linguistic study of bilingual development (pp. 7-23). Amsterdam: Royal Academy of Arts and Sciences. Ferreira, F., & Anes, M. (1994). Why study spoken language? In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 33-56). San Diego, CA: Academic Press. Frisson, S., Sandra, D., Brisard, F., & Cuyckens, H. (1996). From one meaning to the next: The effects of polysemous relationships in lexical learning. In M. Ptitz & R. Dirven (Eds.), The construal of space in language and thought (pp. 613-645). Berlin: Mouton de Gruyter. Gibbs, R. W. (1980). Spilling the beans on understanding and memory for idioms in conversation. Memory & Cognition, 8, 149-156. Gibbs, R. W. (1986). Skating on thin ice: Literal meaning and understanding idioms in conversation. Discourse Processes, 9, 17-30. Gibbs, R. W. (1995). Idiomaticity and human cognition. In M. Everaert, E-J. van der Linden, A. Schenk, & R. Schreuder (Eds.), Idioms." Structural and psychological perspectives (pp. 97-116). Hillsdale, NJ: Erlbaum. Gibbs, R. W. (1999). Intentions in the experience of meaning. Cambridge: Cambridge University Press. Gibbs, R.W., & Matlock, T. (1999). Psycholinguistics and mental representations. Cognitive Linguistics, 10, 263-269. Gibbs, R. W., & Matlock, T. (2001). Psycholinguistic perspectives on polysemy. In H. Cuykens, D. Tuggy, & S. Rice (Eds.), Polysemy in cognitive linguistics (pp. 213239). Amsterdam: John Benjamins. Gibbs, R.W., Jr., Nayak, N. P., & Cutting, C. (1989). How to kick the bucket and not decompose: Analyzability and idiom processing. Journal of Memory and Language, 28, 576-593. Hamblin, J. L., & Gibbs, R.W. (1999). Why you can't kick the bucket as you slowly die. Verbs in idiom comprehension. Journal of Psycholinguistic Research, 28, 25-39 Heredia, R. R. (1997). Bilingual memory and hierarchical models: A case for language dominance. Current Directions in Psychological Science, 6, 34-39. Heredia, R. R., & Altarriba, J. (2001). Bilingual language mixing: Why do bilinguals code-switch? Current Directions in Psychological Science, 1O, 164-168. Hulstijn, J. H., & Marchena, E. (1989). Avoidance: Grammatical or semantic causes? Studies in Second Language Acquisition, 11, 241-255. Irujo, S. (1993). Steering clear: Avoidance in the production of idioms. International Review of Applied Linguistics, 31,205-219. Jackendoff, R. (1995). The boundaries of the lexicon. In M. Everaert, E. van der Linden, A. Schenk, & R. Schreuder (Eds.), Idioms: Structural and psychological perspectives (pp. 133-165). Hillsdale, NJ: Erlbaum. Johnson, J. (1989). Factors related to cross-language transfer and metaphor interpretation in bilingual children. Applied Psycholinguistics, 10, 157-177. Johnson, J., & Rosano, T. (1993). Relation of cognitive style to metaphor interpretation and second language proficiency. Applied Psycholinguistics, 14, 159175. Kellerman, E. (1977). Towards a characterization of the strategy of transfer in
270
Teenie Matlock and Roberto R. Heredia
second language learning. Interlanguage Studies Bulletin, 2, 58-145. Kroll, J. F., & Stewart, E. (1994). Category interference in translation and picture naming: Evidence for asymmetric connections between bilingual memory representations. Journal of Memory and Language, 33, 149-174. Lindner, S. (1983). A lexico-semantic analysis of English verb particle constructions. Bloomington, Indiana: Indiana Linguistics Club. Longman Dictionary of Phrasal Verbs. (1983). London: Longman. McLaughlin, B. (1994). Retrospective on the study of early bilingualism. In G. Extra & L. Verhoeven (Eds.), The cross-linguistic study of bilingual development (pp. 27-38). Amsterdam, Holland: Royal Academy of Arts and Sciences. Matlock, T. (1998). What is missing in research on idioms. American Journal
of Psychology, 111,643-648. McArthur, T., & Atkins, B. (1974). Dictionary of English phrasal verbs and
their idioms. London, UK: Collins. Nelson, E. M. (1992). Memory for metaphor by nonfluent bilinguals. Journal of
Psycholinguistic Research, 21, 111-125. Perlmutter, D. M., & Soames, S. (1979). Syntactic argumentation and the structure of English. Berkeley, CA: University of California Press. Swinney, D., & Cutler, A. (1979). The access and processing of idiomatic expressions. Journal of Learning and Verbal Behavior, 18, 523-534. Tabossi, P. (1996). Cross-modal semantic priming. Language and Cognitive Processes, 11, 569-576. Titone, D. A., & Connine, C. M. (1994). Comprehension of idiomatic expressions: Effects of predictability and literality. Journal of Experimental Psychology: Learning Memory and Cognition, 20, 1126-1138. Vaid, J., & Genesee, F. (1980). Neuropsychological approaches to bilingualism: A critical review. Canadian Journal of Psychology, 34, 417-444. Yorio, C. A. (1989). Idiomaticity as an indicator of second language proficiency. In K. Hyltenstam & L. K. Obler (Eds.), Bilingualism across the lifespan (pp. 55-72). Cambridge: Cambridge University Press.
Appendix Stimuli used in Experiment 1
1. Bret spoke o n 5. Bill banked on 9. They slept i n 13. Rita went over 17. Carl turned on
2. Herb called on __ 3. Mark went into __ 4. Mary ran o v e r 6. Phil ran into 7. Tina went over 8. Mina ran out 10. Jeri went out 11. John ate up __ 12. Neil came up__ 14. Kari worked out m 15. Pete slept over 1 6 . Teri wrote u p _
Stimuli used in Experiment 2
Eighty total stimuli" 40 Phrasal Verbs (PV) and 40 Verb Phrases (VP). For each, 20 with Noun Phrases (NP) in final position (following particle or preposition) and 20 with NP in middle
Understanding Phrasal Verbs
2 71
position (between verb and particle or preposition). Half the stimuli were followed by true paraphrases and half by false paraphrases. Sentences in lower case are paraphrases of the sentences in uppercase letters. PV/NP FINAL/False Paraphrase (Sentence in Lowercase) JANE WROTE OUT THE LYRICS IN BLACK INK Jane used a pencil to write TERI LOOKED UP THE PILOT'S ADDRESS ON THE COMPUTER Teri glanced at the pilot PHIL RAN INTO AN OLD FRIEND FROM COLLEGE YESTERDAY Phil jogged with his old friend TINA SPOKE ON THE RECENT INTERNET INITIATIVE YESTERDAY Tina communicated over email yesterday BILL WENT OFF WHEN HIS MOM WAS LATE FOR THE FIFTH TIME Bill was pleased with his mom's tardiness HUGO CALLED UP THE HILLYARD FAMILY ON THE PAY PHONE Hugo used the Hillyard's phone KERI CALLED ON THE STUDENT WHO HAD A PHONE IN CLASS Keri borrowed the phone PAUL WENT OVER THE EXAM WITH HIS STUDENTS Paul had to take the exam THEY SPED UP THE MUSIC SO SOMEBODY WOULD DANCE They started dancing faster and faster THEY DROVE OUT THE TOURISTS BY FORECASTING BAD WEATHER They gave the tourists a ride PV/NP Middle/False Paraphrase (Sentence in Lowercase) EARL SHOWED THE OTHER BOYS UP BY DOING 100 PUSHUPS Earl watched the boys exercise together ALAN TYPED HIS RESUME UP THIS MORNING ON AN IBM PC Alan has never used a computer MARK RIPPED THE VALUABLE PAINTING OFF SO HE COULD SELL IT Mark stole the inexpensive painting KIRK CLOSED THE OLD THEATER UP 30 MINUTES AGO Kirk just opened the theater RONA SHUT THE SHOP DOWN THE MOMENT SHE BECAME ILL Rona was always sick JANE TOOK THE NOTES DOWN QUICKLY ON HER PAD AS SHE LISTENED Jane distributed the notes THEY SAW ME OFF AT THE DENVER AIRPORT LAST WEDNESDAY They noticed me as I arrived BILL FOLLOWED THE PROJECT THROUGH TO COMPLETION Bill quit the project MATT ROLLED THE MAP UP AND PUT IT AWAY Matt hung up a map TODD CHEERED HIS FOOTBALL TEAM ON ALL NIGHT LONG Todd left the game early
2 72
Teenie Matlock and Roberto R. Heredia
PV/NP Final/True Paraphrase (Sentence in Lowercase) LENA BANKED ON FIFTY DOLLARS IN TIPS LAST NIGHT Lena planned to get $50.00 SEAN SLEPT IN SEVERAL HOURS LATER THAN HE HAD PLANNED Sean did not get up early MARY RAN OVER A FOX WHILE DRIVING DOWN HIGHWAY ONE Mary hit a fox on the road MATT RAN DOWN THE MUGGER WHO STOLE HIS MOTHER'S PURSE Matt caught the thief JILL RAN OFF THE DISRUPTIVE STUDENTS IN HER CLASS YESTERDAY Jill forced the boys to leave ALAN WORKED OFF THE TEN EXTRA POUNDS HE HAD GAINED Alan lost weight BART WORKED OUT IN HIS RED T-SHIRT EVERY OTHER DAY Bart exercised regularly PAUL STRIPPED DOWN THE MOMENT HE GOT HOME FROM SCHOOL Paul undressed right away OPAL STOOD OUT BECAUSE OF HER BRIGHT RED CURLY HAIR Opal was very noticeable because of her hair GLEN DREW UP THE FLOOR PLAN USING THE BLACKBOARD Glen sketched the floor plan in chalk PV/NP Middle/True Paraphrase (Sentence in Lowercase) JESS BROUGHT HIS SALARY INCREASE UP AT THE MEETING Jess mentioned a raise JOEL LEFT HIS OLD FRIEND OUT DURING THE PARTY Joel excluded his friend at the party BERT PULLED THE BANK ROBBERY OFF LAST WEEK Bert successfully stole money BARB TOOK THE WHOLE WEEK OFF AND STAYED HOME TO REST Barb did not go to work HERB PUT THE RAGING FIRE OUT IN ONLY SECONDS The fire was extinguished TAMI TURNED HER BAG OVER TO THE POLICE The bag was given to the police JOEL PUT HIS GOOD FRIEND UP FOR THE REST OF THE NIGHT The friend stayed with Joel LEON PULLED THE MAN OVER FOR DRUNK DRIVING AGAIN Leon made the man stop the car RITA LED EVERYBODY ON EVEN THOUGH SHE WAS MARRIED Rita was flirtatious THEY TOOK THE GOVERNMENT EMBASSY OVER THE SECOND HE DIED They started controlling the embassy 9VP/NP Final/False Paraphrase (Sentence in Lowercase) BART WORKED OUT OF HIS HOME ON MONDAYS Bart exercised in his home gym
Understanding Phrasal Verbs
2 73
MATT RAN DOWN THE HILL BEHIND MARIA'S HOUSE Matt chased Maria on the hill LENA BANKED ON FIFTH AVENUE AT THE BANK OF AMERICA ALL YEAR Lena managed a bank SEAN SLEPT IN SEVERAL DIFFERENT HOTELS ON HIS VACATION Sean stayed in youth hostels MARY RAN OVER THE BRIDGE WHILE JOGGING YESTERDAY Mary damaged the bridge JILL RAN OFF THE EDGE OF THE ROAD IN HER CAR YESTERDAY Jill crashed a party recently ALAN WORKED OFF THE MAIN BOULEVARD IN SANTA CRUZ Alan lost his job PAUL STRIPPED DOWN THE HILL AT BEEFY BOYS EVERY NIGHT Paul was a fast-food cook OPAL STOOD OUT IN THE RAIN AT THE BUS STOP FOR AN HOUR Opal waited indoors GLEN DREW UP ON THE BLACKBOARD WITH LARGE CHALK Glen collected more chalk VP/NP Middle/False Paraphrase (Sentence in Lowercase) JESS BROUGHT THE NEW BICYCLE UP THE WALKWAY Jess mentioned the tires JOEL LEFT HIS NEW STEREO OUT IN HIS TOYOTA LAST NIGHT Joel installed the stereo in his living room BERT PULLED THE STAINED CLOTH OFF THE TABLE Bert spilled his drink on the cloth BARB TOOK THE DOLL COLLECTION OFF THE DISPLAY TABLE Barb left the dolls on the table HERB PUT THE GRAY POODLE OUT IN THE BACKYARD The dog was just a puppy TAMI TURNED HER BAG OVER ON THE TABLE The bag was taken off the table JOEL PUT HIS GOOD FRIEND UP THE HALL IN A SMALL ROOM The friend was petite LEON PULLED THE MAN OVER THE HUGE HOLE IN THE ROAD Leon was joking with the man RITA LED EVERYBODY ON A TOUR OF THE HOLY CITY Rita was by herself THEY TOOK THE EMBASSY OFFICIALS OVER THE BAY BRIDGE They murdered the embassy officials VP/NP Final/True Paraphrase (Sentence in Lowercase) TINA SPOKE ON THE STAGE AT OUR HIGH SCHOOL ABOUT AAA Tina talked to a group of students BILL WENT OFF THE ROAD INTO A GULCH WHILE DRIVING DRUNK Bill had a driving accident HUGO CALLED UP THE HILL IN AN OLD PHONE BOOTH Hugo used a public phone
2 74
Teenie Matlock and Roberto R. Heredia
KERI CALLED ON THE PATIO USING A CORDLESS PHONE Keri used the telephone TERI LOOKED UP HER FRIEND'S CHIMNEY WITH A FLASHLIGHT Teri shined light into the chimney PHIL RAN INTO AN OLD HOUSE TO HIDE FROM THE POLICE Phil entered the house quickly PAUL WENT OVER THE BRIDGE WITH HIS STUDENTS Paul crossed the bridge JANE WROTE OUT ON THE PATIO IN THE SUN Jane was outside while writing THEY SPED UP THE STEEP HILL IN THE NEW PORSCHE They drove fast THEY DROVE OUT THE GARAGE DOOR AND HEADED FOR THE BEACH They went to the beach by car VP/NP Middle/True Paraphrase (Sentence in Lowercase) EARL SHOWED THE OTHER BOYS UP THE STAIRS TO THE OFFICE Earl guided the boys upstairs ALAN TYPED HIS RESUME UP THE STREET IN MARY'S OFFICE Alan was in Mary's office MARK RIPPED THE BUTTONS OFF HIS SHIRT AND THREW THEM Mark removed the buttons KIRK CLOSED THE OLD THEATER UP THE STREET PERMANENTLY Kirk went out of business RONA SHUT THE DOOR DOWN THE HALLWAY VERY LOUDLY Rona slammed the door JANE TOOK THE NOTES DOWN THE STAIRS TO HER MOM Jane delivered the notes THEY SAW ME OFF ON THE SIDE ASSISTING WITH COSTUMES They noticed I was helping BILL FOLLOWED THE GUIDE THROUGH THE ANCIENT TUNNEL Bill did not lead MATT ROLLED THE BALL UP THE SIDEWALK Matt kept the ball moving TODD CHEERED HIS FOOTBALL TEAM ON THE BALCONY T o d d was seated on the balcony
Bilingual SentenceProcessing- R.R. Herediaand J. Altarriba(Editors) 9 2002 ElsevierScienceB.V. All rights reserved.
12
What Native and Non-Native Speakers' Images for Idioms Tell Us About Figurative Language
Heather Bortfeld Brown University
Abstract
In a series of experiments, native and non-native English speakers were asked to form mental images of familiar American English idioms. Even when given explicit instructions to base their images on the phrases' literal meanings, both groups of speakers reported images that reflected fusions of the phrases' literal and figurative meanings. In a subsequent rating task, non-native speakers were judged to have produced more natural sounding sentences using the target idioms after the imaging task than they did before the task. In a second experiment, non-native English speakers were asked to form images of unfamiliar American English idioms both before and after being told the phrases' figurative meanings. Non-native speakers' images did not noticeably change between the first and the second imaging session, indicating that their images were not dependent on explicit knowledge of the phrases' figurative meanings. Emergent properties of metaphorical language are discussed. Introduction
How is it that we understand not only literal language (when a phrase means what is explicitly stated), but also figurative language (when a phrase means something different from what is explicitly stated)? Various accounts have been offered to explain how people understand figurative language. Some researchers have argued that conceptual metaphors guide people's comprehension of figurative language (e.g., Gibbs, 1994). Others propose that conceptual schemas are constructed on-line and made available during comprehension (Glucksberg, Brown, & McGlone, 1993). These proposals are based largely on data collected from speakers comprehending idioms from their native language. Yet even as native speakers, people have variable knowledge of their own language's idioms. Some phrases are highly overlearned while others have never been heard before. Some studies have examined the effects of differences in idiom familiarity (e.g., B lasko & Connine, 1993), but determining which processes are at work, given different knowledge states, is difficult. Few studies have examined how people process completely unfamiliar idioms (however see Keysar & Bly, 1995), but this is exactly the situation one finds oneself in when learning another language. Such a situation
276
Heather Bortfeld
renders the adult learner almost child-like in his or her ignorance of the figurative meanings intended by these novel turns of phrase. Anyone who has spent time leaming an additional language as an adult is familiar with this phenomenon. Yet, unlike children, adult learners are fully developed language comprehenders. They understand that their job is to determine what the novel combinations of words mean when used figuratively in the target language.
Why Study Idioms? Idioms are remarkably frequent in everyday language and, indeed, entire dictionaries have been written about these phrases. For example, Boatner, Gates, and Makkai's original (1975) Dictionary of American Idioms presented over 4000 expressions that occur in English alone. The latest edition of this same dictionary (Makkai, Boatner, & Gates, 1995) contains over 8000. Increasingly, language scholars are acknowledging the wide range of idiomatic phrases that people use daily. Of course, language learners become aware of this as soon as they try to join a colloquial conversation being held in their new tongue. Since idioms are abundant in spoken language (any spoken language), they must be accounted for by any theory of language comprehension.
Idiom Processing Gibbs and his colleagues have actively focused on the "decomposition" of idiomatic phrases into smaller meaningful units. Based on this work, they have come to characterize idiomatic phrases as variably decomposable (Gibbs & Nayak, 1989). This characterization is consistent with more recent linguistic analyses of idiomaticity (e.g., Nunberg, Sag, & Wasow, 1994) that criticize the traditional view (e.g., that idioms are noncompositional). Gibbs' analysis of idiom structure has led him to examine the relationship between metaphors and the phrases' surface content. He proposes that idiomatic meaning is motivated to varying degrees by speakers' implicit knowledge of the metaphors the phrases refer to. Gibbs argues that these conceptual structures arise from the connections we make everyday between domains of meaning. Other researchers have argued that these conceptual structures are instead constructed anew and made accessible in working memory during idiom comprehension (e.g., Glucksberg & Keysar, 1990; Glucksberg, Brown, & McGlone, 1993; Glucksberg, Keysar, & McGlone, 1992). Cultural knowledge in idiom processing: But cultural knowledge and experience play an important role in the lexicalization of concepts in a given language as well. Cultural differences are often what make idiomatic phrases appear arbitrary to cultural outsiders. Of course, even members of a particular language group may find historically-based phrases in their language opaque in origin. Yet as native speakers of the language, these individuals come to access the intended meanings of the seemingly arbitrary phrases automatically, such that they quickly and easily understand them. But of interest here are those culturally-colored phrases that require aiaalysis to be understood yet that are not completely opaque (e.g., based on some elusive historical event or fact). These variably "analyzable" phrases, which
Native and Non-Native Speakers' Images for Idioms
277
differ from language to language, will help us understand more about how idiom processing works. Perhaps the kernels of meaning upon which many idioms are based are general across languages and it is the stretching and warping that takes place when those concepts are lexicalized that is culturally motivated (Kittay, 1997). If so, can people apply conceptual knowledge established in their first language to make sense of figurative speech of this sort in another? Where recognizing a familiar idiom's figurative meaning won't necessarily require that we access the phrase's underlying metaphorical reference, trying to understand a novel idiom (e.g., one from another language) should prompt us to do the necessary mapping between the phrase's surface content and the conceptual structure that motivates its figurative meaning.
Mentally Imaging Idioms Linguists and other scholars have documented the historical origins of many idioms. Native speakers tend not to have access to this historical information (e.g., why kick the bucket means 'to die'). However, the same speakers demonstrate tacit knowledge of the metaphorical basis for many other idioms. When asked to form a mental image for a phrase like spill the beans, people can do so without difficulty and can answer a variety of questions about that image (Gibbs & O'Brien, 1990). Although their reports are somewhat varied, native speakers generally report that the container for the beans is something like a large jar or bowl about the size of a person's head, that the beans were supposed to remain in the container but were accidentally spilled, that the beans go all over the place once spilled, and that they are not easy to get back into the container. How does this knowledge come about? There is no linguistic explanation for this tacit knowledge about certain idioms. The studies reported here attempt to determine whether this kind of knowledge crosses linguistic boundaries to help non-native speakers make sense of novel idioms from other languages. Gibbs and O'Brien (1990) found that native English speakers' images for subsets of idiomatic phrases with similar figurative meanings were highly consistent, despite large differences in their surface forms (e.g., keep it under your hat and hold your tongue). Responses to follow-up questions about the causes and effects of different events within their images were also highly consistent. These researchers concluded that the image schemas they observed across speakers' images reflect knowledge that is the product of conceptual metaphors (e.g., THE MIND IS A CONTAINER and IDEAS ARE ENTITIES) that motivate the phrases' figurative meanings. However, in trying to replicate Gibbs and O'Brien's (1990) finding with a group of native Italian speakers, Cacciari and Glucksberg (1995) found that the majority of the images these speakers generated for Italian idioms reflected only the literal actions and events denoted by an idiom's surface structure. Cacciari and Glucksberg concluded that such literal interpretations demonstrate that people's images do not reflect figurative meanings, "let alone the conceptual metaphors that presumably underlie those meanings" (p. 290). Given these contradictory findings, perhaps there is some confusion over what aspects of people's mental images reflect what aspects of the phrases' meanings.
278
Heather Bortfeld
Indeed, in other work (e.g., Gibbs, 1994) Gibbs has argued that the distinction between literal and figurative is not at all categorical and should instead be conceptualized as falling along a continuum. It is not possible in either of these two imaging studies to determine whether participants construed experimental instructions to be asking for an image of a phrase's literal or figurative meaning. If no further instruction was given and no clarification was sought, then perhaps the difference between the two studies' results is due to different interpretations of the task by the different participant populations. Given that figurative interpretations of familiar idioms are made just as quickly as literal ones (Gibbs, 1980, 1986; McElree & Nordlie, 1999), even if native speakers were explicitly asked to produce an image based on an idiom's literal meaning, they would likely experience interference from the phrase's figurative meaning while forming that image. The findings reported by Gibbs and O'Brien (1990) may reflect difficulties that native speakers have in separating familiar idioms' literal and figurative meanings. Non-native speakers who are unfamiliar with their new language's idioms, on the other hand, shouldn't have automatic access to any pre-stored figurative meanings for the phrases and thus should find it easier to produce strictly literal images for them. If non-native speakers instead produce the kind of hybrid images that the native speakers in Gibbs and O'Brien's study produced, this would indicate that they engage in a process of mapping the phrases' surface forms to conceptual structures similar to those held by the native speakers. While the on-line aspect of this mapping process is consistent with the view that Glucksberg and his colleagues hold on idiom processing, the similarity between the outcomes of native and non-native speakers' mapping would likewise support the view held by Gibbs and his colleagues. Experiment 1 was designed to test this hypothesis.
Experiment 1: Native and Non-Native English Speakers' Mental Images for American English Idioms Method Native English and native Latvian speakers learning English as a foreign language were asked to form mental images based on the literal meanings of a group of American English idiomatic phrases.
Participants Native speakers: Twenty-five undergraduates (17 women and 8 men) at the State University of New York at Stony Brook received course credit in exchange for participation in this experiment. All were native speakers of English and none reported the ability to speak another language. Non-native speakers: Twenty-five undergraduate students (20 women and 5 men) at the University of Latvia received course credit in exchange for participation in this experiment. All were in their first or second year of the English philology
Native and Non-Native Speakers' Images for Idioms
279
concentration. All were native speakers of Latvian, spoke Russian ~ to varying degrees, and had been studying English for five years or less. Materials
Stimuli were the 25 idioms used by Gibbs and O'Brien (1990). The phrases represent 5 concepts (e.g., anger, revelation, secretiveness, insanity, and control), with 5 different idioms expressing each specific concept (e.g., anger = blow your stack, flip your lid, lose your cool hit the ceiling, and foam at the mouth). All idioms were "abnormally analyzable," as defined by Gibbs and Nayak (1989) in their taxonomy of idiom types. These five groups of idioms are listed in the Appendix. The six probe questions from Gibbs and O'Brien's (1990) original study were also used. As noted by those researchers, the probe questions were designed to reveal participants' beliefs about various aspects of their mental images. Probe questions were of two types: 1) questions requiring yes/no responses (e.g., Intentionality--"Was the action done intentionally?"; Reversibility--"Is it difficult to reverse the action?"), and 2) questions eliciting open-ended responses (e.g., Causation--" What caused the action to happen?"; Manner--"How was the action performed?"; Consequence--" What happens as a result of the action?"; Negative Consequence--"What would happen if the action didn't happen?"). During the imaging task, each probe question was altered to match the specific mental image that participants described for each idiom. For example, if a subject reported that her image for blow your stack was someone with fire and smoke shooting out of the top of her head, they would be asked "What causes the fire and smoke to shoot out of the top of this person's head?" for the Causation question. All probe questions were adjusted in this way according to participants' images. PrOcedure Native speakers: Native speaking participants were first given a sheet of paper with a list of the 25 idioms on it and were asked to generate a sentence for each, using the phrase figuratively. They were then asked to form mental images based on the literal meanings of the 25 different idiomatic phrases. Idioms were presented to each participant sequentially in a random order. Once a mental image was generated, native speakers were asked the series of detailed probe questions (outlined above) in an invariant order (e.g., Causation, Intentionality, Manner, Consequence, Negative Consequence, and Reversibility). Each imaging session was audiotaped and later transcribed verbatim for subsequent coding. Following the imaging task, native speakers were given a second sheet of paper listing the 25 idioms and again were asked to generate sentences using the phrases figuratively. Idioms for which native speakers were unable to provide a mental image were skipped and re-presented at the end of the interview. If a native speaker was still unable to provide a mental image for that phrase, the phrase was skipped entirely. The interview was conducted according to a script, and very little discussion of the images took place, aside from simple clarification questions (e.g., "What do you see coming out of the person's head?" when someone described "stuff' exploding out of
280
Heather Bortfeld
the top of a person's head for blow your stack). Each interview took less than one hour to complete. Non-native speakers: The protocol for the non-native speakers was similar to that followed for the native speakers. Upon scheduling an imaging session, they were given a list of the 25 idioms (in English) accompanied by the phrases' figurative paraphrases (also in English). Paraphrases for each idiom were based on Boatner, Gates, and Makkai's (1995) definition for the phrase. Non-native speakers were told to "become familiar with" the figurative meanings of the idiomatic phrases, as they would be asked several questions about those phrases during the imaging session. Sessions always took place two weeks after scheduling. As was the case with native speakers, non-native speakers were asked to generate written sentences using the idioms figuratively both prior to and following the imaging session. At the beginning of the session, they were told that they would be presented with the idiomatic phrases they had been given two weeks earlier and that they were to form a mental image based on the literal meaning of that phrase. Non-native speakers were asked to describe the first image that came to mind, and to answer the probe questions in as detailed and accurate a manner as possible. Each interview took less than one hour to complete.
Coding The interviews were transcribed and imported into Sequence | a software package designed specifically for coding linguistic corpora. Two research assistants naive to the experimental hypothesis being tested were trained to code the images and answers to probe questions following the procedure outlined by Gibbs and O'Brien (1990). The native speakers' images and answers to probe questions were coded first, in order to establish a baseline for comparing the non-native speakers' images and answers. Images were first identified as reflecting either the literal or figurative meaning of an idiom, or some hybrid of the two meanings. In order to be considered entirely literal, an image had to describe only the literal contents of the phrase, with no reference to the underlying emotion or desire described by the figurative meaning on the phrase. Conversely, those images that were considered entirely figurative had to describe only the emotion or desire that the figurative meaning of the phrase entailed with no reference to the surface structure of the phrase. Hybrid images had to include references to both the concrete actions described by a literal reading of the phrase, in addition to references to an underlying (figuratively accurate) emotion or desire that motivated those actions. Examples of the three types of images for the idiom go offyour rocker from the insanity concept group are presented in Table 1. For example, when a native speaker reported that his image for flip your lid was of the top of a person's head being forcefully blown up, this was scored as a hybrid image in which "some force causes the top of a container to release pressure violently." Reliability was coded with satisfactory agreement between coders (96.5% agreement, Cohen's Kappa = .95). Coders discussed any discrepancies until agreement was reached.
Native and Non-Native Speakers' Images for Idioms
281
Table 1. Sample Responses by Native Speakers to Questions Regarding Their Image for the Phrase "Go Off your Rocker"
Literal General Image: Causation: Intentionality: Manner: Consequence: Negative Consequence: Reversibility:
Getting off of my rocking chair to get food Hunger Yes, it was intentional Casually standing up I get food More hunger It's not difficult to reverse, just sit back down
Figurative General Image: Causation: Intentionality: Manner: Consequence: Negative Consequence: Reversibility:
The old man loses his mind He's getting too old No Absent mindedly (effect of old age) He gets put away He'd be normal Difficult too, because it happens sooner or later
Hybrid General Image: Causation: Intentionality: Manner: Consequence: Negative Consequence: Reversibility:
A crazy woman rocks and rocks in her rocking chair until she loses her grip on the chair and flies off Loss of balance due to her agitation No Surprisingly Breaks her hip Healthy hip No, it's hard to stop once it's happening
Collapsing across concept groups, a total of 70% of the native English speakers' images qualified as hybrids of phrases' figurative and literal meanings, 28% were based entirely on a literal reading of the phrases, and 2% clearly referred to only the phrases' figurative meanings. Since the majority of images for phrases represented a hybrid of the two meanings, the hybrid images in each concept group were then coded for their general characteristics. In order to do this, each coder read through all of the participants' hybrid images and responses to the corresponding probe questions and independently developed five schemas intended to capture the most common answers provided for each conceptual group. Once coders had established
282
Heather Bortfeld
independent schemas for the images' general characteristics, these were discussed until a single representative schema for each concept was agreed upon. Table 2 outlines the image schema derived from native English speakers' most frequent images and responses to probe questions for anger idioms. A schema was established for each of the five concept groups.
Table 2. Schema Derivedfrom Native English Speakers' Most Frequent Responsesfor Anger Idioms
Anger General Image: Causation: Intentionality: Manner: Consequence: Negative Consequence: Reversibility:
Pressure forces some container to release its contents violently upwards and outward Stress, anger, and/or frustration Action is not intentional Forcefully, violently, uncontrollably Pressure is released or relieved Container remains intact Action is difficult if not impossible to reverse
Once schemas were established, coders independently scored the native and non-native speakers' image descriptions and probe question responses for all images (including those that had been classified as entirely literal or figurative) according to whether or not they matched the schemas. The coding program allowed images and responses to be grouped according to concept with the variable labels hidden, so that coders would not be biased by such information (e.g., whether the image had been considered literal, figurative, or a hybrid of the two; which participant provided that image; which idiom was being described). Coders were able to view the image schema key (developed earlier) corresponding to a particular concept while coding any given image and probe question answers. Images and probe responses were coded independently of one another. This prevented coders from "reading into" or projecting a response onto a probe question based on supplementary information available in responses to other questions for the same idiom. Images and responses were coded as either matching or not matching the corresponding schema. In order to be considered a match, a general image had to include both the literal and figurative aspects of the image schema that had been identified for that image's concept group. For example, if a participant reported that she saw "a lion tamer cracking a whip to discipline a lion" for the idiom crack the whip, this was coded as a match with the general schema for control. That is, this general image includes a concrete action (cracking the whip) for a purpose consistent with the figurative meaning of the phrase (that order is imposed in some way by someone else). A general image was considered a non-match if only a literal
Native and Non-Native Speakers' Imagesfor Idioms
283
action was included (e.g., "a man jumping too high and hitting his head on the ceiling" for hit the ceiling) with no reference to the underlying figurative meaning of the phrase. Conversely, if the general image included only a reference to the figurative meaning (e.g., "to sneak around with a secret" for behind one's back) with no reference to the literal contents of the phrase, this was also considered a non-match. In order for responses to the open-ended probe questions to be coded as matches, they had to include at least one of the characteristics listed in the corresponding image schema. For example, if a person answered the Causation question for the phrase blow your stack by saying "something upset them," this was counted as a match since the image schema for that concept's Causation probe includes "stress, anger, and/or frustration." Responses were coded as a non-match if they did not include at least one aspect of that probe's listing for the corresponding concept. For example, when a participant reported that "someone wanted to roll a ball to someone else" for the phrase keep the ball rolling, this was coded as a nonmatch because the Causation probe for control had to include "exertion of authority or control over others." A test of reliability showed satisfactory agreement between coders for native speakers' images (92.4% agreement, Cohen's Kappa = .91) and for non-native speakers' images (94% agreement, Cohen's Kappa = .92). Again, all coding discrepancies were discussed until agreement was reached. Compiling codes across participants and across idioms within each of the five conceptual groups provided a baseline summary of the most frequent image characteristics for each conceptual group for the two groups of speakers. Results
Where relevant, two analyses are reported; tl is the analysis by-subjects, and t: is the analysis by-items. The criterion for significance was set at p < .05. Overall, both groups of speakers produced images and responses to probe questions that matched the hybrid image schemas developed for each conceptual group of idioms. Averaging across the different idiom groups reveals that 74% of native speakers' mental images matched the schemas for hybrid images (ranging from 61% to 89% for the five concepts). This proportion is significantly different from chance (t1(24) = 1.93, p < .05; t2(24) = 1.72, p < .05) even with a conservative null hypothesis of .50 that assumes participants could only give two possible responses. An average of 72% of non-native speakers' images matched the native speakers' most common schemas for hybrid images (ranging from 58% to 87%), also significantly different from chance (t~(24) = 1.73, p < .05; t2(24) = 1.72, p < .05). In addition to their general image schemas, participants' answers to the probe questions illustrate participants' understanding of the details characterizing their images. Data from probe questions were of two types: Some questions required yes/no responses and others elicited open-ended responses that were coded in the same manner as participants' images. Across the five idiom groups, native-speakers gave responses that matched the schemas 80% of the time to the Causation questions, 86% to the Intentionality questions, 79% to the Manner questions, 82% to the Consequence questions, 76% to the Negative Consequence questions, and
284
Heather Bortfeld
75% of the time to the Reversibility questions. The proportions for all the probes were greater than chance (t1(24) = 2.03, 2.78, 1.98, 2.23, 1.81, and 1.79, respectively, p < .05 for each comparison). Non-native speakers gave responses that matched the schemas 73% of the time to the Causation questions, 82% to the Intentionality questions, 78% to the Manner questions, 81% to the Consequence questions, 75% to the Negative Consequence questions, and 72% of the time to the Reversibility questions. The proportions for all but the reversibility probes were greater than chance (t1(24) = 1.74, 2.19, 1.88, 2.25, and 1.78 respectively, p < .05 for each comparison). Since each percentage represents a single item from the imaging questionnaire, no item analyses were carried out for this set of data. Discussion
Native and non-native speakers both formed images for idioms that reflect a fusing of the phrases' figurative and literal meanings. For example, the idioms go off your rocker and lose your marbles both mean "to go crazy" when used figuratively. Both native and non-native speakers described images for these idioms that involved a complete loss of stability (most often physical) on the part of the individuals imagined, yet involved rocking chairs in the former instance and marbles in the second. If they were interpreting the phrases literally, as they had been instructed to do, nothing in the lexical makeup of these phrases should have constrained their image production to such a degree. Instead, both groups appeared to have difficulty forming images separating the phrases' literal from their figurative meanings. These results support and extend Gibbs and O'Brien's (1990) findings regarding native speakers' images for idioms. First, native speakers more often produced images for idioms that hybridized the phrases' figurative and literal meanings than reflected just their literal or figurative meanings. Non-native speakers did the same. In fact, in certain instances, non-native speakers' images were more consistent with the established schemas than a subset of the native speakers' images. This was because a subset of the native English speakers interpreted certain idioms in terms of more contemporary, slang meanings of the terms (e.g., blow your stack to mean 'spend all your money' rather than 'to get extremely angry'). The non-native speakers, on the other hand, had only just learned the phrases' canonical, idiomatic meanings and did not have enough cultural exposure to English to be aware of the phrases' more recently developed colloquial meanings. Despite this, the overall similarity between both groups' images and their responses to subsequent probe questions is noteworthy. More surprising is the fact that both groups consistently hybridized the phrases' literal and figurative meanings while imaging them.
Experiment 2: Naturalness of Native and Non-Native English Speakers' Use of Idioms Before and After Imaging
If native speakers' images of idioms reflect mappings between surface structures and underlying conceptual structures, then non-native speakers stand to gain
Native and Non-Native Speakers' Images for Idioms
285
linguistic knowledge in their new language by doing this mapping themselves. Of course, the rich linguistic knowledge that native speakers have to characterize each of these domains is what distinguishes them as native speakers. Nevertheless, adult learners who successfully acquire another language do so in large part because they adapt to the way speakers of the new language lexicalize familiar concepts. One way to promote this is for adult learners to analyze phrases from the new language in terms of those underlying structures. This is a process that the mental imaging task appears to promote as well. If this is the case, then the sentences that the native and non-native speakers produced both prior to and following the mental imaging task in Experiment 1 provide an ideal opportunity toexamine whether the task had any influence on the non-native speakers' linguistic abilities. Experiment 2 tests the prediction that nonnative speakers will use idiomatic phrases from their new language more naturally following participation in a task that prompts them to analyze how the phrases' surface content maps to corresponding conceptual structures. Method Native English speakers rated the sentences that were collected before and after the imaging task (completed in Experiment 1) for naturalness. Participants Thirty undergraduate psychology students (16 women and 14 men) at the State University of New York at Stony Brook received course credit for participating in the study. All identified themselves as native speakers of English and none reported the ability to speak another language. Materials The sentences generated by both native and non-native speakers before and after the imaging task were compiled into groups reflecting the five concepts described in Experiment 1. One idiom was randomly selected from each set in the five concept groups. Before-and-after sentences produced by each of the 50 participants in Experiment 1 (25 native speakers and the 25 non-native speakers) were then transferred into a booklet, resulting in 250 pairs of sentences. The sentences were blocked by concept. Within each block, sentences were randomized such that pairs of sentences produced by native and non-native speakers were mixed. Within each pair, sentences produced prior to and following the mental imaging task were also randomized. The following instructions preceded each pair of sentences: (1) Circle the sentence that sounds more natural to you and (2) Rate how natural that sentence sounds. The second question was followed by a 7-point scale with 1 anchored as "not very natural" and 7 anchored as "very natural."
286
Heather Bortfeld
Procedure
Participants were instructed to compare each pair of sentences, circle the more natural sounding one, and then rate that sentence for how natural it sounded. Participants completed the questionnaire individually. Results
Before and after choices: The naturalness of native speakers' expressions was not affected by their participation in the mental imaging task; the sentences they produced following the mental imaging task were judged to sound more natural than those produced before the task 52.9% of the time, a difference that is not greater than chance (tj(29) = 1.57, p > .05; t2(123) = .52, p > .05). In contrast, the sentences that non-native speakers produced after participating in the mental imaging task were judged to sound more natural than those they produced before the task 62.8% of the time, a difference that is greater than chance (tj(29) = 6.44, p < .001; t2(123) = 3.21, p < .003). Naturalness ratings: Although the non-native speakers used English idioms in sentences more naturally following the mental imaging task, the sentences they produced were still rated as less natural than those produced by native speakers (4.01 vs. 5.47; t~(29) = 6.50, p < .001; t2(248) = 7.76, p < .001).
Discussion
Findings from Experiment 2 indicate that analysis of an idiom's surface form (as is promoted by the mental imaging task) can help people recognize an idiom's conceptual underpinnings. Despite the fact that non-native speakers were aware of the figurative meanings of the idioms prior to participating in the mental imagery task, the sentences they produced based on that knowledge were not as natural sounding as the sentences they produced following the imaging task. Although the increase in naturalness in post-imaging sentences could be a product of the increased time non-native speakers were actually aware of the canonical definitions of the phrases, it seems likely that analyzing the phrases helped non-native speakers better understand the mapping between surface forms and conceptual structures.
Experiment 3: Influence of Prior Knowledge on Native and Non-Native Speakers' Mental Images for Idioms
It is still unclear whether non-native speakers' knowledge of the canonical definitions of target idioms prior to participation in the imaging task resulted in their production of hybrid images or whether their hybrid images emerged from the analysis promoted by the task. Three issues that need to be resolved include, 1) whether or not one's prior knowledge of the figurative meaning of an idiomatic phrase influences the mental image one forms for that phrase's literal meaning; 2) whether the similarities seen between native and non-native speakers' images for idioms in Experiment 1 disappear when non-native speakers' native language is not
Native and Non-Native Speakers' Images for Idioms
287
from the same language family as that of the target language; and 3) whether analyzing idioms' surface structure helps non-native speakers determine their figurative meanings. Experiment 3 was designed to address these issues.
Method As in Experiment l, Experiment 3 compares the mental images generated by native and non-native speakers for the same set of 25 English idioms. In the present study, however, participants performed the imaging task twice, both before and after being told the target phrases' canonical figurative meanings. To eliminate any confounds that may have been introduced due to genetic similarities between English and Latvian (both Indo-European languages), non-native speakers in the present study were native speakers of Mandarin (a Sino-Tibetan language). As in Experiment 1, participants generated sample sentences using the target idioms both prior to and following the first imaging session. This served the dual purpose of measuring the participants' familiarity with the idioms and whether engaging in the imaging task helped participants determine the figurative meanings of the phrases. Finally, a questionnaire was substituted for the interview format used in Experiment 1 to control for any influence the interactive nature of the interview may have had on participants' responses.
Participants Native speakers." Twenty-five undergraduate students (12 women and 13 men) of the State University of New York at Stony Brook received monetary compensation for their participation in the experiment. All identified themselves as native speakers of English and none reported the ability to speak another language. Non-native speakers." Participants were twenty-five graduate students (14 women and 11 men) at the State University of New York at Stony Brook who received monetary compensation for their participation in the experiment. All were native Mandarin Chinese speakers from Mainland China. Although it is common for people from China to speak any number of SinoTibetan dialects z depending on what region they are from, formal education takes place in Mandarin and all forms of media in the country are transmitted in Mandarin. As graduate students, the non-native speakers in Experiment 3 all completed several years of higher education in China and thus were all fluent in Mandarin. Given that any of the Chinese dialects the students might speak in addition to Mandarin are also unrelated to Indo-European languages, their multilingualism did not present itself as a confound in the present study. Participants were recruited from English as a Second Language (ESL) courses offered by Stony Brook's Department of Applied Linguistics to new non-English speaking graduate students. All had Test of English as a Foreign Language (TOEFL) scores between 500 and 600 (scores can range from 200 to 800). TOEFL is a comprehension and writing test that does not involve spoken English. TOEFL scores were used as a rough indicator of English knowledge because the test is used as a standard measure of fluency among foreign students. Graduate students with scores under 600 are typically asked to attend an ESL course in order to improve their English skills prior
288
Heather Bortfeld
to serving as teaching assistants. However, even a score above 600 does not guarantee fluency in spoken English nor familiarity with idiomatic English. Materials
Stimuli were the same 25 idiomatic phrases used in Experiment 1 (see Appendix). Five versions of the questionnaire were prepared, each with the target idioms presented in a different random order. Questionnaires were identical for native and non-native speakers, except that the non-native speakers' questionnaires consisted of the English idiomatic phrases and their literal translations in Mandarin. The literal translations were provided to discourage participants from using Mandarin-English dictionaries to look up individual words. Written instructions and probe questions were also translated. Translations: Although the difficulties inherent in translating experimental stimuli from one language to another are well-documented (see Au, 1983, 1984), literal translations of the idioms were necessary, given the non-native speakers' relatively limited English vocabularies. Every effort was made to ensure that the translations from English to Mandarin were accurate and reflected the appropriate meanings of both the instructions and experimental stimuli. Participants were encouraged to write in Mandarin if they preferred to do so, and all did. Translations were completed by two fully bilingual Mandarin-English speakers and were judged for accuracy by 5 native Mandarin speakers who had resided in the United States for 10 years or more. Any ambiguities were corrected, based on these 5 speakers' recommendations. Of the test idioms, only two out of the 25 literal translations needed to be discussed and the ambiguities in those translations were noted by only 2 of the 5 reviewers. Procedure The experimental protocol for each of the two sessions included a pre-test, the mental image questionnaire, and a post-test identical to the pre-test. Participants were tested individually. Data from the native speakers' pre-tests e~tablished a baseline for how familiar they were with the test idioms and data from the nonnative speakers provided information about how much idiomatic English they knew prior to participating in the study. For the pre-test, participants were told to provide a paraphrase of the figurative meaning for each phrase they identified. Participants then completed the mental image questionnaire. Following completion of the questionnaire, participants were given a post-test and again were told to simply provide figurative definitions for as many of the phrases as they could. Following completion of the first session, all participants were given a sheet of paper with the canonical figurative definitions of the 25 idiomatic phrases. They were told to check their understanding of the idioms' meanings against the canonical figurative definitions provided on the sheet of paper prior to returning for the second session. At this time, a second session was scheduled to take place approximately two weeks after the first session. The second session followed the same format as the first, except that no post-test was administered.
Native and Non-Native Speakers' Images for Idioms
289
Coding Native English speakers' responses to the first and second mental image questionnaires were transcribed and imported into Sequence | (in the same manner described in Experiment 1). Native Mandarin speakers' responses were translated by two bilingual Mandarin-English speakers who were nal've to the experimental hypotheses being tested. No significant semantic differences were noted between the two sets of translations, ensuring that participants' responses were accurately represented by the translations. These responses were then transcribed and imported into the coding program. Coding was completed by the two research assistants who coded participant responses for Experiment 1. The image schemas established in that experiment were used as a comparison measure. As in the first experiment, images were coded for whether they were based on an idiom's figurative or literal meaning, or instead represented a hybrid of the two. The definitions that each participant provided preceding and following the first mental imaging session were coded for whether participants knew the canonical figurative meanings of each phrase or not. A test of reliability indicated satisfactory agreement between coders across the different material (91.6% agreement, Cohen's Kappa = .90).
Results The percentage of images reported by native and non-native speakers that were coded as literal, figurative, and hybridized are shown in Table 3. Again, even when explicitly instructed to base their images on the literal meanings of the phrases, both groups of speakers reported hybrid images the majority of the time. Collapsing across the different idiom groups for the first imaging session, an average of 69% of native English speakers gave hybrid images for the different English idioms (ranging from 56.1% to 76.1%). For the second session this average was 67% (ranging from 58.7% to 74.8%). Collapsed across the different idiom groups, an average of 47% of the native Mandarin speakers' images during the first session were hybrids (ranging from 27.7% to 57.4%). For the second session this average was 54% (ranging from 27.1% to 73.4%). These data indicate that it is more difficult for native speakers to separate the literal from the figurative meanings of idiomatic phrases from their own language than it is for non-native speakers. However, the data also indicate that non-native speakers fused the literal and figurative meanings of the idioms almost 50% of the time both prior to and following being made explicitly aware of the phrases' canonical meanings. This was the case even when non-native speakers had no prior knowledge of the majority of the phrases' figurative meanings. Data from the pre-tests indicate that non-native speakers were familiar with just 15% of the English idioms prior to the imaging task. Focusing on results from the hybrid images, the data from the first and second imaging sessions indicate that both groups of participants gave responses that were consistent with the concept schemas established during Experiment 1. Collapsing across the different idiom groups, an average of 76% of native English speakers gave responses that matched the schemas for the different English idioms (ranging
290
Heather Bortfeld
from 69.5% to 80.1%) during the first imaging session, a proportion that is significantly different from chance (t(30) = 1.81, p < .05). An average of 73% of the non-native speakers' responses during the first session matched the schemas for the different English idioms (ranging from 66.5% to 78.5%). This result is marginally different from chance (t(31) = 1.62, p < .10). These results indicate that non-native speakers' analyses of the phrases during the imaging task gave them an indication of the phrases' figurative meanings, at least to some degree.
Table 3. Image Types Reported by Participants During First and Second Mental Imaging Sessions English Background: Session: Image Type
I
H
Non-Native I III
Anger
Hybrid: Literal: Figurative:
56.1 40.0 3.9
65.2 29.0 5.8
54.2 32.9 12.9
65.8 12.3 21.9
Control
Hybrid: Literal: Figurative:
76.1 22.6 1.3
74.8 21.9 3.2
57.4 37.4 5.2
73.4 18.8 7.8
Secrecy
Hybrid: Literal: Figurative:
75.9 23.2 0.6
74.2 21.3 4.5
56.8 26.5 16.8
60.0 14.2 25.8
Insanity
Hybrid: Literal: Figurative:
72.3 25.8 1.9
64.5 30.3 5.2
37.0 51.9 11.0
41.3 36.1 22.6
Revelation
Hybrid: Literal: Figurative:
65.8 32.3 1.9
58.7 36.1 5.2
27.7 66.5 5.8
27.1 44.5 28.4
Concept Group
Native
Collapsing across idiom groups for the second session reveals that an average of 77% of native English speakers' hybrid images for the English idioms matched the baseline schemas (ranging from 69.5% to 80.8%), a proportion that is significantly different from chance (t(31) = 1.95, p < .05). An average of 76% of non-native English speakers' hybrid images for the different English idioms collapsed across the different idiom groups described similar general images (ranging from 73% to 78.5%), also significantly different from chance (t(31) = 1.93, p < .05). These data indicate that the Mandarin speakers were even more likely to fuse literal and
Native and Non-Native Speakers' Images for Idioms
291
figurative meanings of the phrases once they were made aware of the canonical figurative definitions of the phrases. Overall, both groups of participants reported highly consistent responses to the different probe questions about their images for the English idioms (see Table 3). This was particularly so for images produced during the second imaging session. The probe question data indicate that when participants' images were hybrids, their intuitions about the characteristics of their images tended to match the schemas developed based on native speakers' most common responses. Finally, data from pre- and post-tests conducted during the first imaging session support the hypothesis that analysis of idioms' surface structure helps non-native speakers deduce their figurative meanings. These data are presented in Table 4.
Table 4. Types of Definitions Provided at Pre- and Post-Test During Imaging Session I Concept Group"
Natives
Non-natives
Test."
Pre-
Post-
Pre-
Post-
Anger
Correct: Incorrect: Slang:
74.8 7.1 11.0
91.0 6.5 2.6
9.7 31.0 10.3 58.1 . . . .
Control
Correct: Incorrect: Slang:
75.5 5.2 16.1
87.7 1.9 10.3
5.2 49.0 22.6 36.8 . . . .
Secrecy
Correct: Incorrect: Slang:
91.6 2.6 3.2
97.4 0.6 1.9
45.2 82.6 12.9 13.5 . . . .
Insanity
Correct: Incorrect: Slang:
76.8 5.2 15.5
90.3 1.9 7.7
5.8 41.0 11.7 51.3 . . . .
Revelation
Correct: Incorrect: Slang:
81.3 6.5 8.4
96.1 1.9 1.9
12.3 45.2 16.1 45.2 . . . .
Where non-native speakers gave correct responses only 15% of the time on the pre-tests, they correctly defined the test idioms 49.7% following completion of the imaging task. This result makes a compelling argument for the analyzability of at least a subset of idiomatic phrases. It is notable that at no time prior to the post-test were these non-native speakers given information about the phrases' figurative meanings. The increase in correct figurative definitions at post-test can only have been a product of analysis of the phrases' surface structure and how this mapped to conceptual knowledge (an analysis promoted by the imaging task). In contrast,
292
Heather Bortfeld
changes between native speakers' pre- and post-test responses were almost entirely due to a reduction in slang interpretations of certain phrases at post-test.
Discussion
Results from Experiment 3 are consistent with the initial findings regarding speakers' mental images for idioms. Non-native speakers produced lots of hybrid images and these hybrids matched schemas that were developed based on native speakers' images, as did their answers to a standard set of probe questions asked regarding those images. Following a two week period, during which non-native speakers had access to the test idioms' canonical figurative meanings, even more of their responses matched native speaker-based image schemas. Pre- and post-tests from the first imaging session indicate that speakers are able to infer figurative meaning based on an analysis of the surface structure of idiomatic phrases, without additional supporting context. That these phrases can be understood by non-native speakers with no instruction regarding their canonical meanings is compelling evidence that at least a subset of idiomatic phrases are analyzable. Although the images that non-native speakers produced prior to being told the phrases' figurative meanings were more often based on literal interpretations of the phrases than those produced by native speakers, the increase in accurate definitions of the idioms' figurative meanings at post-test indicates that the non-native speakers were engaged in this process of analysis. The similarities between the native and non-native speakers' images in both Experiment 1 and 3 indicate that this process of analysis guides and constrains idiom comprehension more than knowledge of the phrases' canonical figurative meanings itself. Again, there is no explanation for such regularity in people's knowledge about figurative phrases in traditional linguistic theories of idiomaticity. Data from the probe questions indicate that both native and non-native speakers are able to access very specific information about the causes and effects of actions in their images for idioms. Given the striking similarities between native and non-native speakers' images and their answers to probe questions about those images, these findings indicate that specific conceptual structures underlie at least some subset of languages' idiomatic phrases. Another issue that these data address is that of the universality of these conceptual structures. Figurative language is fascinating precisely because it is so prevalent and in such diverse forms across languages. It is important to ask how people who speak languages that are quite literally a world apart can relatively easily map figurative tums of phrase from one language onto corresponding turns of phrase in another. The fact that figurative phrases originating in one language can be understood and appreciated by speakers of another attests to the importance of this question. The conceptual structure or structures tapped by a given idiomatic phrase must be shared by the two languages in order for it to be analyzed. This is indicated by the cross-linguistic data reported here. Our ability to appreciate literature that has been translated from other languages addresses the universality question as well. Translators do not translate idioms
Native and Non-Native Speakers' Images for Idioms
293
word-for-word from the source to the target language. Instead, they try to locate an idiom in the target language that best represents the sense intended by the idiom used in the source language. In doing so, good translators maintain the source language's original mapping from words to conceptual structure. The art of translation depends on the translator's ability to recognize the concept or group of concepts being tapped by a given idiom in the source language and find the closest lexicalization in the target language. Truly talented translators work to preserve original structural representation (e.g., turns of phrase) so that people can at least in part experience different manners of lexicalizing a familiar concept. It is this opportunity--the opportunity to experience new lexicalizations of old concepts--that makes the experience of reading well-translated foreign literature distinct from the experience of reading literature that originated in one's own language.
Conclusion
These studies provide evidence that our use and understanding of idioms is guided by conceptual structures that are not created or constrained by our a priori knowledge of the phrases' canonical figurative meanings (cf., Keysar & Bly, 1995, 1999). Analyzability itself can be conceptualized as ranging along a continuum (Bortfeld & McGlone, 2001). The data presented here suggest that the continuum of analyzability may well reflect a more fundamental continuum ranging between universal conceptual knowledge and culture-specific knowledge. The idioms used in the current studies can all be located roughly in the middle of the analyzability continuum; they are all somewhat metaphorical. Further work needs to be done examining how both native and non-native speakers process idioms from the rest of the analyzability continuum. Ultimately, it seems that the aptness of the conceptual mapping rather than simply one's familiarity with the phrase is what determines an idiom's analyzability. After all, language--both one's first and any subsequently acquired--is initially encountered without the benefit of familiarity. For figurative speech in our own language, we must initially map words to conceptual structures in whatever way is meaningful. Over time, that mapping becomes well established. The sense that an unfamiliar figurative phrase is apt is something that most people who have learned a second language can appreciate, whether that learning took place during child- or adulthood. Certain phrases in a new language seem appropriate and, quite simply, make sense. It is not the case that the phrases are familiar; rather it is that they are apt, regardless of the particular surface structure the source language has given them. By looking more closely at what motivates this apparent aptness, we stand to learn much about the structures guiding our comprehension of nonliteral language. Thus far, this discussion has been framed in terms of second language learners dealing with figurative speech in their new language. However, the research reported here can likewise be evaluated in terms of the benefits of multilingualism. Perhaps what distinguishes good language learners is how willing they are to consider an unfamiliar turn of phrase as representative of a familiar concept. All the non-native speakers who took part in these experiments were able to form images for various idiomatic phrases. But some people seemed more comfortable or adept
294
Heather Bortfeld
at mapping words to meaning than others. Perhaps individuals who have been raised in a bilingual environment are particularly flexible in their approach to this mapping process. Just as bilingual speakers have less difficulty than monolinguals in appreciating that two different words can both be symbols for the same conceptual object, they may likewise more readily appreciate that distinct turns of phrase map to the same conceptual structure. Perhaps bilingual speakers simply map words to underlying meaning more flexibly than monolingual speakers. Such a finding would add further support to claims about the universality of conceptual structures. Our understanding of language processing, in general, and idiom comprehension, in particular, stands to gain from the extension of the questions addressed here to bilingual populations. Such research would also benefit adult second language learners by enriching our understanding of the source of at least some of their difficulties. Idioms are fascinating precisely because they are so prevalent and in such diverse forms across languages. Language processing research must continue examining how idioms can or cannot cross linguistic boundaries. Only then will we understand how it is that people who speak languages that are, quite literally, a world apart can relatively easily map the surface forms of novel phrases to their appropriate underlying concepts.
Notes
Russian is a Slavic language, relatively closely related to Latvian in that both stem from the Balto-Slavic branch of Indo-European languages. 2 These dialects tend to be entirely different languages, although for political and cultural reasons they are typically not referred to as such.
References
Au, T. (1983). Chinese and English counterfactuals: The Sapir-Whorf hypothesis revisited. Cognition, 15, 155-187. Au, T. (1984). Counterfactuals: In reply to Alfred Bloom. Cognition, 17, 289302. Blasko, D., & Connine, C. (1993). Effects of familiarity and aptness on metaphor processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 19, 295-308. Boatner, M., Gates, J., & Makkai, A. (1975). ,4 dictionary of,4merican idioms. Woodbury, NY: Barron's Educational Series, Inc. Bortfeld, H., & McGlone, M. (2001). The continuum of metaphor processing. Metaphor and Symbol 16, 75-86. Cacciari, C., & Glucksberg, S. (1995). Understanding idioms: Do visual images reflect figurative meanings? European Journal of Cognitive Psychology, 7, 283305. Gibbs, R. (1980). Spilling the beans on understanding and memory for idioms in conversation. Memory & Cognition, 8, 449-456.
Native and Non-Native Speakers' Imagesfor Idioms
295
Gibbs, R. (1986). Skating on thin ice: Literal meaning and understanding idioms in conversation. Discourse Processes, 9, 17-30. Gibbs, R. (1994). The poetics of mind. New York, NY: Cambridge University Press. Gibbs, R., & Nayak, N. (1989). Psycholinguistic studies on the syntactic behavior of idioms. Cognitive Psychology, 21, 100-138. Gibbs, R., & O'Brien, J. (1990). Idioms and mental imagery: The metaphorical motivation for idiomatic meaning. Cognition, 36, 35-38. Glucksberg, S., Brown, M., & McGlone, M. (1993). Conceptual metaphors are not accessed during idiom comprehension. Memory & Cognition, 21, 711-719. Glucksberg, S., & Keysar, B. (1990). Understanding metaphorical comparisons: Beyond similarity. Psychological Review, 97, 3-18. Glucksberg, S., Keysar, B., & McGlone, M. (1992). Metaphor understanding and accessing conceptual schema: Reply to Gibbs. Psychological Review, 99, 578581. Keysar, B., & Bly, B. (1995). Intuitions of the transparency of idioms: Can one keep a secret by spilling the beans? Journal of Memory and Language, 34, 89-109. Keysar, B., & Bly, B. (1999). Swimmers against the current: Do idioms reflect conceptual structure? Journal of Pragmatics, 31, 1559-1578. Kittay, E. (1997). Of "men" and metaphors: Shakespeare, embodiment, and filing cabinets. In T. Ward, S. Smith, & J. Vaid (Eds.), Creative thought." An investigation of conceptual structures and processes (pp. 375-402). Washington, DC: American Philosophical Association. Makkai, A., Boatner, M., & Gates, J. (1995). A dictionary of American idioms. Hauppauge, NY: Barron's Educational Series, Inc. McElree, B., & Nordlie, J. (1999). Literal and figurative interpretations are computed in equal time. Psychonomic Bulletin and Review, 6, 486-494. Nunberg, G., Sag, I., & Wasow, T. (1994). Idioms. Language, 70, 491-538.
Appendix
Idioms used for Mental Imaging Task (adapted from Gibbs & O'Brien, 1990) Anger
Control
Secretiveness
blow your stack hit the ceiling lose your cool foam at the mouth flip your lid
crack the whip lay down the law call the shots wear the pants keep the ball rolling
keep it under your hat button your lips hold your tongue behind one's back keep in the dark
Insanity
Revelation
go off your rocker lose your marbles go to pieces lose your grip bounce off the walls
spill the beans let the cat out of the bag blow the whistle blow the lid off loose lips
This Page Intentionally Left Blank
Part VI: Language Skill Development in Bilingual
Children
This Page Intentionally Left Blank
Bilingual Sentence Processing- R.R. Heredia and J. Altarriba (Editors) 9 2002 Elsevier Science B.V. All rights reserved.
13
The Role of Formal Definitions in Reading Comprehension of Bilingual Students
Aydin Y. Durguno~lu University of Minnesota, Duluth Zehra F. Peynircioglu American University Montserrat Mir Illinois State University
Abstract In upper elementary grades, once word recognition is efficient, reading comprehension is more affected by processes such as vocabulary and background knowledge. One facet of vocabulary knowledge that is closely related to comprehension is the quality of formal definitions, as this knowledge reflects the student's awareness of decontextualized language. Spanish-English fourth-grade students who have just been transitioned to allEnglish classrooms gave definitions to words from Spanish and English expository passages before reading the texts and answering questions about them. The quality of formal definitions was correlated across the two languages, indicating cross-language transfer of this metalinguistic awareness. In addition, formal definition quality was related to reading comprehension both within and across languages. In the first few years of schooling, one of the biggest accomplishments of young children is learning to read proficiently. Later, in the upper elementary grades, it is assumed that this skill is in place and can be used as a tool for knowledge acquisition. For most children, once the decoding process is efficient, reading comprehension is affected by higher order processes such as vocabulary and background knowledge rather than by word recognition proficiency. There has been considerable research showing that vocabulary knowledge is a variable that affects reading comprehension, especially with expository texts (Anderson & Freebody, 1981; Stahl, Jacobson, Davis, & Davis, 1989). Vocabulary knowledge is multifaceted. For example, the word "lemon" has many different aspects of its meaning represented in the mind of an individual. These can include among others, categorical membership (fruit, food, plant, acidic), function (used in dishes), characteristics (yellow, hard, small, sour), relationships with other concepts (similar to lime), or some personal information (my mother used to make lemonade from fresh lemons). Vocabulary knowledge also includes concepts
300
Aydin Y. Durguno~lu, Zehra F. Peynircio~lu and Montserrat Mir
represented at different levels of specificity and sophistication. For example the representation for the word "emperor" will be considerably richer and deeper for a historian than for a non-historian. Another dimension of vocabulary knowledge is how this existing knowledge is expressed, that is, how concepts are formally defined for an audience. Formal definitions require the writer/speaker to activate the existing vocabulary knowledge and to express it in a decontextualized manner that is without the personal links to which the reader/listener may not have access. A study by Gutierrez-Clellen and DeCurtis (1999) provided evidence that the quality of formal definitions indicates specialized linguistic knowledge. They compared the quality of formal definitions of Spanish-speaking children with or without language impairments. Impairment was defined by low Spanish Assessment of Basic Education scores and the clinical judgments and parent reports of language delays. Both groups of children were given 10 familiar words to define. Children without language impairment provided formal definitions 50% of the time, whereas this number was 19% for children with language impairment. Children without language impairment used superordinate categories (e.g., animal) rather than the generic term (e.g., "una cosa," a thing) more often. They also provided more specific details about the object. Why is formal definition quality related to reading achievement? As Perfetti (1991 ) summarized, print is decontextualized language as it does not have the rich nonlinguistic information (such as gestures, immediate feedback) that is part of oral communication. Snow and colleagues (Snow, Tabors, Nicholson, & Kurland, 1995) suggested that oral decontextualized language skills account for increasing variance in reading success as reading becomes a task of comprehension rather than decoding. There have been studies indicating that the decontextualized language of classrooms is a challenge for students as they move from home to school. For example, in schools, writing tasks require the students to provide information to a distant audience that may not be present in the same location at the same time. Likewise, texts (especially expository texts) that may not have personally relevant contexts, require children to have an understanding of decontextualized language. Formal definitions can be considered an index of a student's awareness of decontextualized language. In fact, several studies have shown that quality of formal definition sentences is correlated with reading achievement. Snow, Cancini, Gonz~lez, and Shriberg (1989) collected definitions from 2 nd to 5th grade children at a private international school and related the quality of the definitions to their reading scores on the Califomia Achievement Test. The quality of formal definitions, but not the quality of informal definitions increased with grade. In addition, quality of formal definitions, but not informal, colloquial definitions, was related to reading achievement of 4 th and 5th grade students. However, formal definition skills are not necessarily predicted by oral communication proficiency. Children with similar oral communication levels may have varying decontextualized language skills and hence varying literacy proficiencies (Dickinson & Snow, 1987). Formal definitions require children to provide the description of a concept in a "culturally prescribed" sentence (Snow et al., 1989, p. 234) because the definitions usually have an "X is a Y that Z" format, with Y representing the superordinate category, and Z specifying other characteristics such as
Formal Definitions in Reading Comprehension
301
function and features, among others (Gutierrez-Clellen & DeCurtis, 1999). For example, the following definitions follow such a prescribed format: "a submarine is a ship that goes under the water," "an author is a person who wrote a book." In contrast, an informal, personal definition describes a concept by giving personally relevant information that may or may not be shared by the listener/reader. For instance, compare the following two definitions for the word "ocean" taken from the responses of the students in our study: "Ocean is like when [it] is summer, you could go to ocean" versus "ocean is water that surrounds the earth." In the first example, the definition is given in the context of personal experiences, whereas in the second definition, the description is more complex, and does not necessarily assume shared personal experiences, but provides decontextualized information. The ability to provide high quality formal definitions is not only semantic, but also metalinguistic knowledge, because it involves judging how to present the information to a distant audience, when the possibility of common personal interactions is slim. This is especially true for tasks in which non-linguistic clues are not available, for example, tasks requiring written responses. If formal definition skill involves a metalinguistic awareness, then can it transfer across languages of a bilingual reader? Given the data summarized above, the goal of the current study was to explore the link between formal definition quality and reading comprehension across the two languages of bilingual students. The following questions were addressed: 1) How are formal definitions related to reading comprehension? 2) How are formal definitions related across languages? 3) How does formal definition proficiency develop? We conducted a study with fourth grade, Spanish-speaking children who had exited from Spanish-English bilingual education programs to all-English classrooms. This ensured us a sample who had formal schooling in both Spanish and English. Our outcome variables were English and Spanish reading comprehension. We also assessed quality of formal definitions, syntactic knowledge, oral communicative proficiency (vocabulary and story retelling), and word recognition proficiency in both languages and observed which of those variables best predicted success in English and Spanish reading comprehension.
Method Participants Participants were 26 fourth-grade students (8 boys and 18 girls) from an elementary school in a suburb of Chicago. All students had Spanish-speaking parents (18 from Mexico, and the rest from E1 Salvador, Puerto Rico, and Colombia), with blue-collar jobs either in the service industry or in the factories around town. When asked about how much Spanish they knew, the majority of the children (n -- 19) said "a lot." Except for one child who reported using only English, all children used Spanish (solely or in
302
Aydin Y. Durguno~lu, Zehra F. Peynircio~lu and Montserrat Mir
combination with English) when they spoke to their parents. However, they were more likely to use English with friends and teachers. Except for the five who had attended kindergarten in Mexico, all children had been attending school in the USA and they had been in bilingual education classrooms. Currently they were in all-English classrooms. In these classrooms, although several teachers understood and spoke Spanish, they used only English. In their self-reports, the majority of the children (n 16 for English and n = 16 for Spanish) rated their reading proficiency in both Spanish and English as "good" or "very good." Materials and Procedure
The students' performance in Spanish and in English was assessed using the tasks described below. All tasks were given in November. Outcome Measures
Reading comprehension. The major goal was to determine variables that affect reading comprehension in Spanish and in English. The students read two expository passages in Spanish and two expository passages in English and answered multiple choice and short answer questions about the passages. Expository passages were selected because students usually find them more difficult as these texts use more decontextualized language. In addition, this type of reading is what the students need to do most of the time to acquire content knowledge in and out of school. The Spanish passages were about forestkeepers and vitamin C and the English passages were about Jules Verne and the origin of chocolate (see Appendix). The passages were taken from Spanish reading books, translated and modified. The reading comprehension tests (see Appendix) were developed by researchers and included both multiple-choice and shortanswer questions. The tests were scored by giving one point to each correct answer on the multiple-choice and three points for each short-answer question. Partial credit was given to correct, but incomplete answers on the short answer part. Students completed the reading comprehension tasks in an intact classroom. They read the passages once and answered the questions at their own pace without looking back at the text. They completed the two Spanish texts in one day, and the two English texts on another day (maximum score was 13 for each language task across the two passages). Predictor Measures
English word recognition tasks. Two different tasks were used to determine the level of English word recognition and awareness of English orthography. The students were tested individually and all responses were tape-recorded and later transcribed. The first task was the word identification subtest of the Woodcock test. In this open-ended, 106-item test, children attempted to read progressively more difficult words and continued until missing six consecutive words. The second task was used to determine how children read exception words in English, such as "ocean" and "island" (Adams & Huggins, 1985) which cannot be pronounced using spelling-to-sound correspondences. Twenty exception words were included in this task, with one point given for each correct pronunciation. The two word recognition measures were highly correlated (r =
Formal Definitions in Reading Comprehension
303
84), hence in the following analyses, the two tasks were combined into a single word recognition measure. The score was the total number correct in both tasks with maximum possible 126.
Table 1. Vocabulary Items and the Means and Standard Deviations (SD) of Definition Scores
Language/Text/Word Mean
SD
2.9 3.1 5.3 5.1 4.0
4.0 2.8 3.1 2.7 2.9
4.6 0.1 7.2 4.5 4.2
2.6 0.7 3.0 2.8 2.0
6.2 0.6 3.9 4.6 4.4
3.3 1.1 2.2 3.4 1.8
1.2 0.7 3.7 2.1 4.5
2.5 1.7 2.3 2.7 2.2
SPANISH Forestkeepers forestkeeper fire tree insect lightning Vitamin C vitamin scurvy fish lemon boat ENGLISH Jules Verne author science fiction story submarine ocean Chocolate Aztecs emperor chocolate cocoa milk
English syntax. In this task, the students listened to a sentence read by the experimenter and corrected the syntactic error in the sentence. To reduce the memory load, each sentence was also presented visually as typed on an index card. The 20 sentences included errors of different types, such as tense, inflection, word order, and pluralization (cf. Johnson & Newport, 1989). The responses were tape-recorded and transcribed. The students were told that each sentence had an error in it, and they needed to correct that error. Each correct response was given 3 points. If the students corrected the error, but made a mistake in another part of the sentence, they got 2
304
Aydin Y. Durguno~lu, Zehra F. Peynircio~lu and Montserrat Mir
points. I f they identified the error, but did not c o r r e c t it a c c u r a t e l y , t h e y got one point. M a x i m u m p o s s i b l e score w a s 60.
Table 2. Scoring Framework for Definitions
I- Syntax, formal structure: X is a Y that is ...... 0 = no syntax 1 = "X is a Y" present 2 = "X is a Y that is ..." (incomplete/reduced clause) 3 = "X is a Y that is ..." (full clause) II- Superordinate term, Y 0 = none given 1 = correct but very general, e.g. "thing" 2 = adjective + object, or more specific category 3 = best superordinate III- Relative clause: that .... 0 = none given 1 = incorrect 2 = correct but limited 3 = properly defines correct subset IV- Number of distinct descriptive features given (excluding the ones already in the relative clause) V- Number of examples, enabling conditions VI- Use, purpose, context 0 = none 1 = vague or limited 2 = correct VII- Comparison (like z, but/except) 0 = none 1 = partial 2 = example given VIII- Synonym 0 = none 1 - vague or limited 2 = correct Examples: (each digit by the definition refers to the score from each subscale above): CHOCOLATE chocolate is samting [something] that you eat and it's sweet: 31310000 = 8 chocolate is junk food 12010000 = 4 LEMON un limon[lim6n] as[es] agrio 00010000 = 1 [a lemon is sour] una fruta verde y amarga 0 3 0 2 0 0 0 - 5 [a fruit green and sour] es una fruta que es color verde o amarilla 33200000 = 8 [it is a fruit that is green or yellow]
Formal Definitions in Reading Comprehension
305
English oral proficiency. To assess oral proficiency, two subsections of the LAS test (Duncan & De Avila, 1987) were used. In the oral vocabulary section, the experimenters pointed to 20 common objects in a picture and the students named those objects. In the story retelling subsection, the students listened to a short story on the tape and then retold the story. The retelling was tape-recorded, transcribed, and scored according to the five levels described in the test manual. Spanish word recognition. In this test, students read 20 common Spanish words with varying degrees of morphological complexity (e.g., "leer," "charlar," "trabajoso," and "deshojado"). The responses were recorded and later transcribed. The number of words correctly read was the score on this test. Spanish syntax. This test was parallel to the English version, although they were not translations of each other. Students were given 20 Spanish sentences and asked to correct the syntactic problem in each sentence. The sentences had different types of syntactical errors, such as in subject-verb agreement, inflections, or word order. The testing and scoring procedures were identical to those for the English syntax test, with a maximum score of 60. Spanish oral proficiency. Two subsections from the Spanish LAS test (Duncan & De Avila, 1987) were used. As was the case with the English LAS test, the subsections used were: oral vocabulary, in which the students named the 20 common objects in a picture; and, story retelling, in which the students listened to a short story on the tape and then retold the story. The retelling was tape-recorded and later transcribed with the maximum score of 5. Spanish and English formal definitions. Before reading the two passages in each language, the students wrote formal definitions (Snow et al., 1989) to 10 words, 5 from each passage. The to-be-defined words included several common words (e.g., milk), as well as less frequent words that indicated background knowledge about the contents of each passage (e.g., Aztecs). Table 1 gives the list of vocabulary items for each of the reading passages. The quality of the formal definitions was scored using a framework based on the one developed by Snow et al. (1989). The framework is summarized in Table 2, along with some examples. It must be noted that both the format used in the definitions, as well as the scientific content were evaluated in the scoring procedure. However, the accuracy of spelling or syntax was not taken into consideration while evaluating the definitions. If the children used the "X is a Y that is Z" format, and also provided the right superordinate category and details about features, their scores were higher. Some examples of definitions and their scores are given at the bottom of Table 2. The means for each of the vocabulary words are in Table 1. As would be expected, common words such as tree, lemon, fish, and author were defined more accurately than low frequency, content-specific words such as science fiction, scurvy, and emperor.
Results and Discussion
Table 3 presents the means and standard deviations from each of the tasks. The children showed high levels of performance on the vocabulary knowledge for common
306
Aydin Y. Durguno~,lu, Zehra F. Peynircio~lu and Montserrat Mir
objects, in both Spanish and English, 89% and 99%, respectively. On the story retelling part of the oral proficiency test, English retellings had a score of 77% and Spanish retellings 81%. Children showed some room for improvement on the syntax measures, 75% and 64% for Spanish and English tests, respectively. Word recognition levels were 58% for English and 83% for Spanish.
Table 3. Means and Standard Deviations (SD) for Each Task Task
Mean (SD)
Maximum
English word recognition
72:73 (12.2)
126
English syntax
38.40 (14.6)
60
English story retelling
3.84 (0.6)
English vocabulary
19.08 (1.9)
20
Spanish word recognition
16.50(3.9)
20
Spanish syntax
44.89 (11.3)
60
Spanish story retelling
4.04 (0.5)
Spanish vocabulary
17.88 (1.4)
English formal definitions
32.60 (7.9)
Spanish formal definitions
43.73(15.1)
English Reading Comprehension
5.20 (2.5)
13
Spanish Reading Comprehension
5.81 (2.1)
13
20
The correlations among predictor measures are presented in Table 4. First, looking at within language correlations, there were no significant correlations among the English word recognition, oral proficiency, and syntax measures at the p < .05 level. For Spanish, there were several within-language correlations: Knowledge of Spanish syntax was correlated with Spanish word recognition (r = .41). Spanish syntax was also correlated with vocabulary knowledge (r = .60). Finally, Spanish syntax was correlated with Spanish word recognition (r = .56). Because the students had Spanish as their first language and also had received more instruction in it, the Spanish language and literacy variables seem to be more integrated with each other.
Formal Definitions in Reading Comprehension
307
Table 4. The Correlations Among English and Spanish Word Recognition, Syntax and Vocabulary Measures Tasks 1) E-word recognition
( 1) 1.0
(2) .33
(3)
(4)
(5)
(6)
(7)
(8)
91
.07
.11
.36"
.05
-.07
(9) .04
(10) .06
.2 3
-.05
-.02
-.25
.44*
-.05
-.17
-.22
1. 0
-.37"
-.24
-.17
-.14
.03
.56*
-.16
1.0
.27
.20
-.06
.48*
.16
.13
1.0
.37"
.29
.17
.21
.69*
1.0
.41"
.03
.56*
.37"
1.0
.13
.60*
.26
1.0
.14
.25
1.0
.23
7
2) E-syntax 3) E-story retelling 4) E-vocabulary 5) E-formal definitions 6) S-word recognition 7) S-syntax
8) S-story retelling 9 ) S-vocabulary 10) S-formal definitions
1.0
1.0
Spanish *p < .05, ap < .10
E = English, S =
Across languages, English and Spanish syntax were correlated (r = .44). Because the syntax test required the analytic strategy of pinpointing and correcting a syntactic error, children who analyzed sentences well in one language tended to do it well in another language too. Another example of such a language-independent analytic skill was found with formal definitions. The quality of formal definitions was highly correlated across the languages (r = .69). The next step was to analyze the correlations of reading comprehension scores with the predictor measures (Table 5). For both languages, the quality of formal definitions was correlated with reading comprehension (r = .67 and .46 for English and Spanish, respectively). Of course the vocabulary words were from the passages, so this correlation is hardly surprising. However, it still provides evidence that vocabulary knowledge is an important component of reading comprehension. It is likely that the formal definitions activated some background knowledge, facilitating the comprehension of passages. However, there was also a significant correlation between English reading comprehension and the quality of Spanish definitions (r = .46).
308
Aydin Y. Durguno~lu, Zehra F. Peynircio~lu and Montserrat Mir
Similarly, Spanish reading comprehension showed a marginally significant correlation with the quality of English definitions (r = .34). None of the other predictor measures correlated with reading comprehension (except for Spanish vocabulary with Spanish reading comprehension). The oral proficiency and word recognition measures did not correlate with the reading comprehension tests.
Table 5. The Correlations Among English and Spanish Reading Comprehension and Syntax, Vocabulary and Word Recognition Measures Tasks 1) E-Reading Comprehension 2) S-Reading Comprehension 3) E-word recognition 4) E-syntax 5) E-story retelling 6) E-vocabulary 7) S-word recognition 8) S-syntax 9) S-story retelling 10) S-vocabulary 11) E-formal definitions 12) S-formal definitions
E-Reading Comprehension 1.0 -.01 -.09 -.30 .21 .17 .08 -. 10 .14 .67* .46*
S-Reading Comprehension .13 1.0 .12 -.01 -.26 .23 .30 .33 .21 .45* .34a .46*
Spanish *p<. 05, ap < .10
E = English, S =
These patterns were further analyzed using multiple regression analyses to predict Spanish and English reading comprehension performance. For English reading comprehension, the predictor variables (English word recognition, English formal definitions, English vocabulary, English syntax and English story retelling) when entered into the equation explained 48% of the variance. However, in the final equation, only English formal definitions was significant (beta = .648, t = 3.57). When only English formal definitions were entered as a predictor, it explained 45% of the variance (beta = .672, t = 4.35). For Spanish reading comprehension, the predictor variables (Spanish word recognition, Spanish formal definitions, Spanish vocabulary, Spanish syntax, and Spanish story retelling) together explained 3 5% of the variance. Again replicating the pattern above, only Spanish formal definitions was significant (beta = .379, t = 1.74). When only Spanish formal definitions was entered into the equation, it explained 21% of the variance (beta = .462, t = 2.55)
Formal Definitions in Reading Comprehension
309
Because the quality of formal definitions was such an integral part of reading comprehension, we examined their development. Results showed that quality of formal definitions were correlated across languages although the sets of words were completely different. Moreover, formal definition quality in one language was correlated with reading comprehension in the other language (which also involved different topics and words). Therefore, the big impact of formal definitions on comprehension could not have been only due to their overlap with the text passage. To analyze formal definition performance, two different multiple regression analyses were conducted. In each equation, the predictors were: other-language formal definition skill, same-language word recognition proficiency, and same-language vocabulary (see Table 6). When Spanish formal definitions are considered, quality of English formal definitions, Spanish word recognition, and Spanish vocabulary together explained 49% of the variance. However, English formal definition quality was the only significant predictor. Similarly, for English formal definitions, the predictors explained 53% of the variance with Spanish formal definitions being the only significant variable. Here we see a close connection between the quality of definitions although the word sets are completely different in the two languages.
Table 6. Multiple Regression Analyses on Formal Definition Scores /~
Spanish formal definitions English definitions Spanish word recognition Spanish vocabulary Total Variance explained English formal definitions Spanish definitions English word recognition English vocabulary Total variance explained
R2
.644* .108 .034 49% .678* .144 .173 53%
*p <.05
A qualitative analysis of the definitions indicated that children who had good formal definitions in one language had similarly good definitions in the other language as well. (The spelling errors were ignored when scoring the quality of definitions, hence they are left as is in the following samples.) In the samples in Table 7, both the Spanish and English definitions of Child #22 and Child #32 can be compared with the definitions of Child #20 and Child #31. The first two children provide more categorical information and critical details. In addition they tend to use a more formal format.
310
Aydin Y. Durguno~lu, Zehra F. Peynircio~lu and Montserrat Mir
Table 7. Samples of Formal Definitions
Child #22 cocoa: Is a fruit from a tree in the beach author: is a person who write a book submarine: is a ship that goos under the water ocean: is a big body of water ~.rbol: un ~irbol tiene un tronco de madera y unas hojas (a tree has a wood trunk and some leaves) lim6n: sale de un b.rbol y es muy agrio (it comes out of a tree and is very sour) insectos: son unos animales que viven en la tierra (they are animals that live in the ground) barco: es algo que va en el agua (it is something that goes in the water) Child #32 cocoa: is like choclate milk author: is a person that writes a book submarine: is a boat that goes under water ocean: is blue water ~irbol: un ~irbol es como una planta grande o chica (a tree is like a big or small plant) lim6n: es algo agrio pero se le echa al caldo (it is something sour but it ispoured into the broth) insectos: un insecto es un animal con seis pies [although technically incorrect] (an insect is an
animal with six legs) barco: un barco es como un carr0 pero no est~i tan chico y no tiene llanta (a boat is like a car but
is not so small and it does not have wheels) Child #20 cocoa: it like has water inside author: an author is whom they write stories submarine: is an boat that goes in the water ocean: an ocean is water b.rbol: es un palo que tiene muchas hojas (It is a stick that has many leaves) lim6n: un lim6n es agrio (a lemon is sour) insectos: un insecto tiene muchos pies (an insect has many legs) barco: un barco viaja por el agua (a boat travelsby water) Child #31 cocoa: is something to drink author: author is something they do a book submarine: submarine is something thorsting [interesting??] ocean: is a part of lend ~.rbol: b.rbol es un pedazo de madera que tiene hojas (tree is a piece of wood that has leaves) lim6n: es algo pa comer (it is something to eat) insectos: es un animal (it is an animal) barco: es algo para viajar (it is something to travel)
Formal Definitions in Reading Comprehension
311
Conclusions
In this study, reading comprehension performance of Spanish-speaking children was investigated as a function of different variables in their two languages. The quality of formal definitions was the best predictor for reading comprehension in both languages, even when word recognition, syntax, and oral proficiency variables were taken into consideration. Because formal definitions were such an important part of reading comprehension, the next step was to examine their development. These analyses indicated that quality of formal definitions was correlated across the two languages even when the word sets were different. Multiple regression analyses also highlighted this interrelatedness. When analyzing quality of formal definitions, once the definition score in another language was entered into the equation, most of the variance was explained. These results suggest that the ability to give decontextualized definitions using a formal syntactic structure is a metalinguistic skill that can be transferred across languages. There was very little cross-language transfer at the basic word recognition levels. This is not surprising given that these children were in upper-elementary grades and had relatively well-established word recognition proficiencies. It can be assumed that at this stage of literacy development word recognition processes can become modular and impervious to the influence of the basic processes in the other language. This is parallel to a finding with monolingual readers indicating that when word recognition becomes proficient, semantic context exerts less influence (Stanovich & West, 1983). However, analyzing syntax and correcting errors was a skill that was related across the two languages, even when the sentences had errors reflecting the different syntactic structures of Spanish and English. Therefore, analyzing and critically evaluating the structure of a sentence seems to be a metalinguistic skill that also transfers across languages. Another cross-linguistic effect was found with definitions and reading comprehension. Quality of Spanish definitions were related to English reading comprehension and vice versa. Of course, quality of definitions depended on the ability to define words in a formal, decontextualized manner. This can be considered a metalinguistic ability that is relatively language-independent and transfers across languages, just like syntactic analysis and phonological awareness (Durguno~,lu & Hancin, 1992; Durguno~lu, Nagy, & Hancin-Bhatt, 1993). Overall, once bilingual readers acquire a specific metacognitive/metaliguistic awareness in one language, this insight seems to be used in the other language as well. For the future, identifying more of these insights that need to be "learned only once" and applying them in classrooms can help (bi)literacy instruction become more efficient and effective.
Author Notes
This research was supported by a Small Grant from the Spencer Foundation to A. Y. Durguno~,lu. We thank Sofia Arifio-Marti, Patricia MacGregor, and Maria-Jos6 Gonz~.lez for their assistance in testing subjects and scoring the data.
312
Aydin Y. Durguno~,lu, Zehra F. Peynircio~,lu and Montserrat Mir References
Adams, M. J., & Huggins, A. W. F. (1985). The growth of children's sight vocabulary: A quick test with educational and theoretical implications. Reading Research Quarterly, 20, 262-281. Anderson, R. C., & Freebody, P. (1981). Vocabulary knowledge. In J. Guthrie (Ed.), Comprehension and teaching: Research reviews (pp. 77-110). Newark, DE: International Reading Association. Dickinson, D., & Snow, C. E. (1987). Interrelationships among prereading and oral language skills in kindergartners from two social classes. Early Childhood Research Quarterly, 2, 1-25. Duncan, S., & De Avila, E. A. (1987). Language assessment scales (LAS). Monterey, CA: CTB/McGraw-Hill. Durguno~lu, A. Y., & Hancin, B. J. (1992). An overview of cross-language transfer in bilingual reading. In R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 391411). Amsterdam: Elsevier. Durguno~lu, A. Y., Nagy, W. E., & Hancin-Bhatt, B. J. (1993). Cross-language transfer of phonological awareness. Journal of Educational Psychology, 85, 453-465. Gutierrez-Clellen, V. F., & DeCurtis, L. (1999). Word definition skills in Spanishspeaking children with language impairment. Communication Disorders Quarterly, 21, 23-35. Johnson, J. S., & Newport, E. L. (1989). Critical period effects in second language learning: The influence of maturational state on the acquisition of English as a second language. Cognitive Psychology, 21, 60-99. Perfetti, C. A. (1991). Representations and awareness in the acquisition of reading competence. In L. Rieben & C. A. Perfetti (Eds.), Learning to read (pp. 33-46). Hillsdale, NJ: Erlbaum. Snow, C., Cancini, H., Gonz~ilez, P., & Shriberg, E. (1989). Giving formal definitions: An oral language correlate of school literacy. In D. Bloome (Ed), Classrooms and literacy (pp. 233-249). Norwood, NJ: Ablex. Snow, C. E., Tabors, P. O., Nicholson, P. A., & Kurland, B. E (1995). SHELL: Oral language and early literacy skills in kindergarten and first-grade children. Journal of Research in Childhood Education, 10, 37-48. Stahl, S. A., Jacobson, M. G., Davis, C. E., & Davis, R. L. (1989). Prior knowledge and difficult vocabulary in the comprehension of unfamiliar text. Reading Research Quarterly, 24, 27-43. Stanovich, K. E., & West, R. F. (1983). On priming by a sentence context. Journal of Experimental Psychology: General 112, 1-36.
Appendix The reading passages and the comprehension questions Los guardabosques y los ~irboles Los guardabosques saben mucho acerca de los ~irboles. Su trabajo consiste en proteger los
Formal Definitions in Reading Comprehension
313
~irboles y a los seres vivos que viven en ellos. Los guardabosques tienen muchas responsabilidades. Ellos deciden que ~irboles se puede talar. Adem~.s, siembran arbolitos nuevos y tratan de impedir que los insectos darien los ~.rboles. Una de las mayores tareas de los guardabosques es combatir los incendios que se producen en los bosques. Algunos incendios se producen a causa de los rayos, pero la mayor parte son provocados por la gente. Cuando hay un incendio se emplean grandes m~.quinas para detener el avance del fuego. Los guardabosques tambi6n utilizan estas m~quinas para despejar el terreno al acercarse un incendio. AI no poder propagarse a trav6s del terreno despejado, el fuego se apaga. Los guardabosques pueden ensefiarnos muchas cosas sobre los ~.rboles y sobre como cuidarlos. 1) LCual es la causa principal de los incendios? a) los rayos b) la gente c) los animales d) las m~,quinas 2) Se utiliza una mS.quina grande para: a) talar los ~.rboles que est~in viejos b) plantar ~.rboles nuevos c) despejar el terreno para parar el fuego d) educar a los nifios acerca de los bosques 3) Si el terreno est,. despejado, el fuego: a) se extiende m~is b) se eleva c) se hace m~.s peligroso d) no se propaga 4) Escribe una o dos frases para contestar la siguiente pregunta. [Escribe en espafiol] /,Cu~iles son dos de los trabajos de un guardabosques? La vitamina C En tiempos pasados, cuando los barcos navegaban a vela, los viajes duraban varias semanas e incluso meses. Naturalmente, estos barcos tenian que estar bien almacenados con comida. Generalmente, habia una gran cantidad de legumbres secas, carne salada, pescado seco, jam6n ahumado, quesos y pan. Todos estos alimentos se conservaban bien por largo tiempo. No obstante, durante estos prolongados viajes, muchos de los marineros contraian una enfermedad misteriosa, llamada escorbuto. Esta enfermedad causaba la caida de los dientes y del pelo y frecuentemente graves trastornos intestinales queen algunos casos llegaban a causar la muerte. Los m6dicos estudiaron esta enfermedad durante muchos afios. Finalmente descubrieron que el escorbuto se relacionaba con la dieta alimenticia de a bordo. M~.s adelante se confirm6 este diagn6stico cuando los doctores observaron que la enfermedad desaparecia cuando los marineros desembarcaban en areas donde habia muchas naranjas y limones. En 1795, la marina brit~.nica comenz6 a suministrar limones a los marifieros durante las navegaciones prolongadas. Ninguno de estos marineros se enferm6 de escorbuto. La conclusi6n era 16gica: los limones y las naranjas contenian una sustancia que faltaba en la dieta de los marineros y que preservaba contra aquella enfermedad. A finales del siglo pasado, esta beneficiosa sustancia se identific6 como la vitamina C.
314
Aydin Y. Durguno~lu, Zehra F. Peynircio~,lu and Montserrat Mir
1) El escorbuto es una enfermedad que: a) ocurria en los viajes por tierra b) ocurria solo en las montafias c) ocurria en los largos viajes por mar d) ocurria solo en el invierno 2) El escorbuto causa: a) la caida de los dientes y el pelo b) problemas intestinales c) la muerte d) todo 1o mencionado (a, b, and c) 3)/,Qu6 alimento previene el escorbuto? a) el pescado seco b) la leche y el queso c) las naranjas y los limones d) el pan 4) E1 escorbuto esta causado por: a) el movimiento ondulante de los barcos b) la falta de vitamina C c) la suciedad en el barco d) la came salada 5. Escribe una o dos frases para contester la siguiente pregunta: [Escribe en espafiol] /~C6mo descubrieron los m6dicos los alimentos que previenen el escorbuto? A strange drink called xoco-latl About four hundred years ago, Hernando Cortez, an adventurer from Spain, sailed to the coast of Mexico. He traveled inland, to the capital of the Aztec empire. At the capital, Cortez met the emperor named Montezuma. Montezuma welcomed the Spaniards. He offered them food and gifts of gold and jewels. The Spaniards noticed that Montezuma was enjoying a strange drink. It was foamy, spicy, fragrant and dark in color. Aztecs called the drink xoco-lati. Today we call it chocolate. To make chocolate, the Aztecs picked the seeds of the cocoa tree and roasted these seeds. The next step was to grind the seeds. Then they added water and some spices, such as vanilla and pepper, and made a thick paste. Finally, they beat this mixture until it was foamy. The Aztecs ate this spicy chocolate-vanilla foam with a jeweled spoon or drank it from a golden cup. Montezuma said that he finished fifty pitchers of xoco-latl in a day. He drank it with his meals, and between meals from morning to night. Spaniards took many bags of cocoa seeds with them when they returned home. They used cocoa to make the Aztec drink in Spain. They also started to add sugar to the drink to make it sweeter. In the next century, many people in Europe were making and enjoying chocolate drinks. Some people added milk to make the chocolate creamy. They also began to make solid chocolate. Today chocolate is more than a popular drink. It is a favorite flavor in ice cream, cookies and cakes. Many people eat chocolate bars for quick energy. 1) The Aztecs: a) learned from the Spanish how to make chocolate b) taught the Spanish how to make chocolate
Formal Definitions in Reading Comprehension
315
c) sailed the oceans d) lived in Brazil 2) Cortez is: a) an adventurer from Spain b) an emperor of the Aztecs c) a special drink d) a kind of spice 3) Montezuma is: a) an adventurer from Spain b) an emperor of the Aztecs c) a special drink d) a kind of spice 4) Write one or two sentences to answer the following question. How did the Aztecs make xoco-latl? Jules Verne Over 100 years ago, a very famous author named Jules Verne lived in France. Jules Verne had an incredible imagination. He wrote many stories. They were the type of stories we call science fiction, a mixture of science and fantasy. For example, in one of Verne's books, the characters traveled to the hot core in the center of the earth. Another book was about people living underwater in a submarine called the Nautilus. In one adventure, the captain of the Nautilus navigated the submarine under the ice at the South Pole. In 1958, 90 years after Jules Verne wrote the story about the Nautilus, an American submarine sailed under the ice of the North Pole. The submarine was also named Nautilus. It is possible that Jules Verne's book was more than a good story. Jules Verne had a vision of the future. Perhaps his novel made people curious about the bottom of the oceans. As a result, some people tried to construct vehicles to explore underwater. In this way, part of Verne's vision of the future came true. 1) Jules Verne is the name of: a) an author b) an adventurer c) a captain d) a book 2) Which of the following is an example of a science fiction book? a) A book about George Washington's childhood in the year 1740 b) A book about people living on the moon in the year 2250 c) A book about how to draw pictures d) A book about how paper is made 3) Nautilus is the name of: a) a captain b) the ice around the North Pole c) a submarine d) an ocean
316 4) a) b) c) d)
Aydin Y. Durguno~lu, Zehra F. Peynircio~lu and Montserrat Mir
Jules Verne: built a submarine sailed the sea around the North Pole imagined and wrote about a submarine traveled to the center of the earth
5. Write one or two sentences to answer the following question. Why does the text say that Verne had a "vision of the future?"
Bilingual Sentence Processing- R.R. Heredia and J. Altarriba (Editors) 9 2002 Elsevier Science B.V. All rights reserved.
14
Syntactic Structure, Grammatical Accuracy, and Content in Second-Language Writing: An Analysis of Skill Learning and On-Line Processing
Wendy S. Francis University of Texas at El Paso Laura F. Romo California State University, Northridge Rochel Gelman Rutgers University
Abstract This chapter examines second-language sentence production both as a skill and as an on-line process by measuring relationships among linguistic and content characteristics of language samples both between and within participants. Written language samples were obtained from high school students enrolled in an intermediate-level English as a Second Language (ESL) class with embedded science instruction. Across participants, the skills of producing sentences with correct English grammar and accurate science content were positively associated with the skills of producing complex syntactic structures and more elaborate science content. Within participants, the on-line production of correct grammar was compromised when complex science content was expressed. The pattern of results obtained highlight the importance of analyzing bilingual language acquisition data in terms of both skill learning and on-line performance.
Introduction The primary purpose of this chapter is to examine second-language sentence production both as a skill and as an online process. This is accomplished by measuring relationships among linguistic and content characteristics of written language samples both between and within participants. In particular, we wanted to find out the strengths and directions of the correlations among syntactic complexity, grammatical accuracy, content complexity, and content accuracy, and the implications of these associations for the cognitive processes underlying sentence production in bilinguals.
318
Wendy S. Francis, Laura F. Romo, and Rochel Gelman
Written language production is a complex skill. It involves the recruitment and coordination of several component subskills. A competent speaker of a language must retrieve linguistic structures, grammatical rules, vocabulary, and other features of language to produce novel sentences. For late learners of a second language, proficiency in these skills and their coordination takes time to develop, and the learner must be able to retain these skills across periods of time in which they are not using the language. The long-term leaming of a skill and the quality of performance during leaming have long been distinguished in the skill learning literature (e.g., Estes, 1955; Tolman, 1932). The primary basis for this distinction is that manipulations that reduce error during leaming tend to increase error in a delayed retention test, and vice-versa. The factors leading to optimum performance during learning and those leading to optimum long-term retention are not the same. Achievement of long-term learning and transfer in second language leaming is likely based on the same factors that lead to good skill retention and transfer more generally. If so, then language learning in an immersion context is ideally suited to long-term retention because of the unpredictably spaced repetitions of various syntactic structures and vocabulary, the variation of semantic context for syntactic structures, and the variation of syntactic context for vocabulary. Speaking in conversation and writing for communicative purposes also requires retrieving vocabulary and linguistic structures from memory. As with other skills, under these learning conditions, although retention and transfer should be high, errors in second language performance are to be expected until a high level of proficiency is reached. Based on the distinction between performance during learning and long-term retention, some predictions can be made about the relationships among various characteristics of second language production. First, long-term learning and retention of second language skills ought to be a function of practice. Immersion in a language environment exposes the learner to many components of language simultaneously, thereby giving the learner practice with the various language subskills. This simultaneous practice implies that the learning of language components should be positively correlated across time and across individuals. Evidence of positive associations among various aspects of language is found in the literature. For example, vocabulary size is correlated with grammar skills (Harrington, 1992) and with the syntactic complexity of utterances (Kemper, Kynette, Rash, O'Brien, & Sprott, 1989). The second set of predictions can be made about performance during learning, or online processing, which will depend heavily on both the level of learning that has been achieved and on the cognitive resources available for each component of language processing. In capacity models of attention (e.g., Kahneman, 1973; Wickens, 1984), there exists a limited pool of cognitive processing resources that can be focused on one task or allocated across multiple tasks. The more attention that is required to perform one task, the less cognitive resources will be left to allocate to other tasks. Therefore, if a person tries to perform two tasks that require high levels of attention, one or the other will suffer. The deficit in performance under such conditions of divided attention can be manifested either as an increase in
Second Language Writing
319
the time necessary to carry out the task or as a decrease in the quality or accuracy of performance. Under certain circumstances, some pairs of tasks can be performed simultaneously with little or no detectable decrement in performance relative to performing the tasks individually. The first circumstance, and least controversial, is that one or both of the tasks can become more automatic with practice. As each task is better learned, it requires less attention and leaves more resources available to allocate to the other task. The second circumstance is when the two tasks rely on separate pools of attentional resources. Under a multiple-resource theory of attention (Wickens, 1984), the pool of cognitive resources is subdivided into pools that are used for different modalities of processing. Two tasks that draw upon the same partition will interfere with each other, but two tasks that draw upon separate partitions will not suffer when performed simultaneously. In healthy adult native speakers of a language, production of correct grammar and complex syntactic structures is thought to require little attention, because these skills are so well learned that they have become automatic (Bock, 1982; Levelt, 1989). In a not-yet-fluent second language, all aspects of second language production ought to require more attention than they do in the person's own native language and more attention than they do for native speakers of the language being learned. Under conditions of divided attention, healthy adult native speakers of a language do not appear to suffer cognitive consequences in terms of grammatical accuracy. For example, in one study in which the primary language task was describing a picture, having to perform a secondary tone-discrimination task did not compromise any aspect of language production relative to describing the picture alone (Murray, Holland, & Beeson, 1998). In the same study, mildly aphasic individuals, for whom language presumably requires more attention, exhibited profound decrements in language production performance when attention was divided. Specifically, dividing attention decreased the rate of sentence production, decreased the syntactic complexity (in terms of embedded clauses) of the utterances, increased the proportion of sentences with grammatical errors, increased the number of word-finding errors, and decreased the number of correct information units produced. Thus, it appears that all of these language performance deficits are potential consequences of attentional overload to examine in second language production. However, there is little published research on the attentional demands of producing sentences in a developing second-language. A related approach is to think about the potential consequences of attentional overload as a function of working memory. Both comprehension and production of language require maintenance of information in working memory. The importance of working memory for language comprehension has been demonstrated in studies in which individual differences in working memory capacity are associated with measures of sentence comprehension. Individuals with lower working memory span showed lesser comprehension of sentences with complex embedded clause structures (King & Just, 1991). Although the causal relationship has not been definitively established, it appears that working memory limitations of the lowerspan individuals reduce their comprehension. Working memory also appears to be important for language production. Among older native speakers, working memory
320
Wendy S. Francis, Laura F. Romo, and Rochel Gelman
capacity as measured by digit span is positively correlated with the mean number of clauses per utterance in speaking and writing, even controlling for the age of participants (Kemper et al., 1989). That is, participants with greater short-term memory capacity produced more syntactically complex utterances with more subordinate/embedded clauses. Working memory has been linked with some aspects of second language processing, although it has not been studied as much as with first language processing. In studies of adult leamers of English as a Second Language, a measure of working memory based on reading span was positively correlated with both the grammar and reading portions of the Test of English as a Foreign Language (TOEFL) and with other tests of vocabulary and grammar (Harrington, 1992). These studies of how individual differences in working memory are associated with specific language skills and characteristics may help us to make predictions about second-language sentence production. We know that the capacity of verbal working memory in a person's not-yet-fluent second language is lower than that of a native speaker of the language and increases with proficiency (e.g., Chincotta, HyOn~, & Underwood, 1997). Therefore, if verbal working memory capacity is a limiting factor in the quality of first-language processing, then second-language processing ought to exhibit the same performance deficits as that of native speakers who have low working memory capacity. A further implication would be that these features will improve with increased proficiency, because of the increase in working memory capacity. Generally, a second-language leamer's reduced working memory capacity will make them more susceptible to attentional overload. A third set of predictions can be derived from the models that have been developed to explain the complex series of cognitive steps necessary for on-line language production in healthy adult native speakers. These models are in agreement that the idea or message to be expressed is conceptualized before the syntax of an utterance is planned (Bock, 1982; Garrett, 1975; Grabowski, 1996; Kellog, 1996; Levelt, 1989). The temporal sequence of these processes suggests that characteristics of the idea or message will influence characteristics of the syntax. Once the idea or propositional structure is formulated, an appropriate syntactic flame must be selected. Thus, the idea or propositional structure constrains the possible syntactic structures for the sentence, and one of these structures is planned and produced. In the present study, students' complex ideas were hypothesized to constrain their choices of syntactic structures to those containing embedded clauses or prepositional phrases. Some of the models include assumptions about the level of attention required for each stage of sentence production. Those researchers who address attention or working memory in their models are in agreement that attention is required for conceptualization of the idea or message (Bock, 1982; Grabowski, 1996; Kellog, 1996; Levelt, 1989). However, there is some disagreement about the role of attention in syntactic processing. In their models of speech production, Bock and Levelt claim that at least for healthy adult native speakers, syntactic processes are automatic. In contrast, in their models of writing, Grabowski and Kellogg claim that attention is required for translating the message into a syntactic flame. Again, these
Second Language Writing
321
models are meant to explain the production performance of adult native speakers, who have a high level of expertise in sentence production in their language. In relatively novice second-language leamers, the level of attention required for the various production steps may be greater. Clearly, conceptualization of the message, which requires attention in monolinguals, will require attention in second-language learners, but probably not more attention than in monolinguals, because it is thought to be a non-linguistic set of processes. However, planning of the syntactic frame of the sentence and application of morphosyntactic markings may require more processing resources for second-language leamers, because of their relative inexperience or lack of expertise. The attentional resources that second-language learners need to produce correct grammar may not always be available to them when the idea they intend to express is complex. Therefore, in the present study, a trade-off would be expected between idea complexity and correct grammar at the sentence level.
The Present Study Written language can provide a rich database for studying complex structures, because writing tends to have longer sentences and sentences with more complex syntactic structure than spoken language (e.g., Kemper et al., 1989). Writing is also the modality in which late second-language leamers initiate the use of new syntactic structures (Weissberg, 2000). In the present study, writing samples were obtained from late learners of English, in order to explore the relationships among features of sentence production. The participants were high school students of English as a Second Language (ESL). The content of the writing samples was constrained by the curriculum of their ESL class, which was based on science content. Students were to write sentences about what they had leamed in their science lessons. Evaluations of the program and examinations of changes over time using the microdevelopmental approach appear elsewhere (Gelman, Meck, Romo, Meck, Francis, & Fritz, 1995; Gelman, Romo, & Francis, in press). The analyses here focused on linguistic characteristics of sentence production, specifically correct grammar, use of embedded clauses, use of prepositional phrases, and also included the total number of sentences produced. In addition to the relationships among these language variables, we were interested in how linguistic characteristics of the sentences would be related to the content expressed. Constraining the topic to science made it possible to evaluate both the accuracy of the content in each statement and the complexity of the ideas expressed. These aspects of sentence production could be studied both as characteristics of students and as characteristics of sentences. Each of the language and content variables was measured for each sentence and summarized to characterize the sentence production of each student. The summary measures obtained for each student were used to examine the sentence characteristics as skills. The individual measures for each Sentence were used to examine how sentence characteristics go together or trade off in on-line processing. We expected that the difficult aspects of
322
Wendy S. Francis, Laura F. Romo, and Rochel Gelman
second-language sentence production, such as the use of correct grammar and complex syntax, would be well acquired by the same students. However, it is not clear whether those features would appear in the same sentences. Those aspects of sentence production that require attention and draw upon the same resource pool should tend not to appear in the same sentences. In contrast, aspects of sentence production that draw upon separate resources or do not require attention should be independent or positively associated at the sentence level. Another prediction was that focusing on communication of complex information would compromise linguistic characteristics of sentences produced. Anecdotally, many second-language speakers (and even many first language speakers) report that their pronunciation and grammar are not as accurate when discussing complicated topics as they are in everyday conversation. In contrast to potential trade-offs among different language components during on-line processing, some aspects of sentence production may have a positive relationship. For example, certain syntactic structures appear to lend themselves well to particular types of semantic content, such as how adverbial clauses (adjuncts and participials) can be used to convey cause, conditions, purpose, manner, time, location, and comparison. Therefore, we also looked at how semantic content was related to the language variables at both the student and sentence level.
Method
Participants The study participants were 9th- and 10th-grade students (aged 14-17) at a large public high school in the San Fernando Valley of the Los Angeles metropolitan area. These students were in the second level of a 4-level ESL program. Their class was an experimental section that included a science curriculum covering a range of related topics in physical and biological science. A detailed description of the program and curriculum are given elsewhere (Gelman et al., 1995; Meck & Gelman, 1993). A total of 30 students were enrolled for at least part of the term. Of the 30 students, 8 either started the semester late, moved before the semester was over, or did not attend regularly. The notebooks of the remaining 22 students who were present throughout the semester were considered for analysis. Two of these students were excluded (as explained in the next section), leaving 20 participants (11 male, 9 female) for the analyses of skill learning and on-line processing. The native languages of the participants were Spanish (11), Korean (6), Farsi (1), Hebrew (1), and Tagalog (1).
Data Collection Procedure The ESL teacher taught 10 science themes over the course of a 20-week semester. At the end of each science unit, the class was asked to generate as a group
Second Language Writing
323
a list of 10 to 20 science terms relevant to the theme under study. After the class agreed upon the list, students were given approximately 15 minutes to independently construct concept maps in their notebooks. In this exercise, students drew lines to link terms that they thought to be related to each other. Subsequently, they were given approximately 15 minutes more to independently write up to 10 sentences to describe the relationships among the terms they had linked in their concept maps. Thus, the maps were used as a mechanism both to constrain the content and to get students to think deeply about the set of terms. Both of these activities were graded on a credit/no credit basis.
Characteristics of the Database
A sentence was defined as a numbered word sequence that contained at least one verb. If two complete sentences were combined, they were separated. The database had to be carefully checked to eliminate cases in which students copied sentences from other students' notebooks or from the sentences that other students were asked to write on the chalkboard to explain to the class. There were several such cases, but in each, it was clear from examination of differential sentence quality which student copied and which student composed the sentences. The 10-sentence exercise was performed for 10 units, and the 2nd through 9th units were analyzed. The original database, after excluding the students with low attendance contained 1271 sentences. Of these sentences, 37 sentences, all produced by the same student, were judged to be inappropriate, in that they stated simply which terms in the map could be connected, without describing the relationship. Two independent coders evaluated whether or not each of the remaining 1234 sentences was comprehensible enough to allow for coding of its content and structure. Classifications matched in 98% of the cases (Cohen's ~c = .89). Consensus was reached on all discrepancies, resulting in 81 sentences being deemed as incomprehensible. One student accounted for 38 of the incomprehensible sentences. This student and the student who produced 37 inappropriate sentences were excluded from analysis because they were outliers and because they each produced less than 25 codeable sentences. The remaining 20 students produced a total of 1160 sentences, of which 42 (3.6%) were incomprehensible. Thus, the body of codeable data contained 1118 sentences, with a mean of 55.9 sentences (SD = 14.2) produced by each student.
Variables Grammatical Errors
For each sentence, two coders, both native speakers of English, each wrote independently a "correct model" version of each sentence (i.e., how the sentence could have been written to be grammatically correct in English). In correcting the
324
Wendy S. Francis, Laura F. Romo, and Rochel Gelman
sentences, coders used a minimal change approach, making as few changes as were necessary to make the sentence grammatically correct. For each participant the percentage of sentences containing at least one grammatical error was computed. Averaging across participants, the mean percentage of sentences containing grammatical errors was 70.1% (SD = 11.9%). Further classification of the grammatical errors revealed that the 4 most prevalent error types were subject-verb agreement, determiner errors, auxiliary verb errors, and preposition errors (see Gelman et al., in press). These four error types were examined for their potential as variables in a student-level and in a sentencelevel associative analysis. In order to be a good candidate for a sentence-level analysis, the error would have to occur in at least 20% of the sentences. Determiner errors, preposition errors, and auxiliary verb errors did not occur in enough sentences to provide a good sentence-level analysis. The error type that had the potential to occur across the largest proportion of sentences was subject-verb agreement errors. That is, on average, 94% of sentences produced by any student required subject-verb agreement marking, and only 12% of these sentences required multiple instances of subject-verb agreement marking. On average, a student made subject-verb agreement errors in 38.4% (SD = 13.1%) of their sentences, allowing for a powerful analysis of the associations between subject-verb agreement errors and other language and content variables in the study.
Syntactic Structure Variables The sentence structure analysis focused on two syntactic structures: embedded clauses and prepositional phrases. Embedded clauses are a widely accepted indicator of sentence complexity. Prepositional phrases were also examined as a secondary indicator of more sophisticated sentence structure. These structures allow one to increase the number of propositions expressed in a sentence through the recursive property of language. Previous findings in the literature support the interpretation of prepositional phrase usage as a measure of complexity. For example, the lag time taken prior to producing a memorized sentence is increased both when it contains a prepositional phrase and when it contains an embedded clause, relative to a more simple structure, even when the length of the sentences in words and syllables is held constant (Ferreira, 1991). Two coders, both with formal training in linguistics, coded the clauses and prepositional phrases of each sentence and checked each other's work to ensure accuracy. The average proportions of sentences containing embedded clauses and prepositional phrases are shown in Table 1. Across participants, the mean proportion of sentences with embedded clauses was 22.7% (SD = 13.2%). On average, students produced prepositional phrases in 48.2% of their sentences (SD = 16.7%). Several sentences contained more than one prepositional phrase. The validity of both embedded clause usage and prepositional phrase usage as indicators of proficiency in creating sentence structures is supported by the significant increase in the percentage of sentences with each of these structures over the duration of the course (Gelman et al., in press).
Second Language Writing
325
Table 1. Descriptive Statistics on Linguistic and Content Variables
Mean
SD
55.9
14.2
% of sentences containing grammatical errors
70.1
11.9
% of sentences containing subject-verb agreement errors
38.4
13.1
% of sentences containing embedded clauses
22.7
13.2
% of sentences containing prepositional phrases
41.9
14.3
29.4
13.9
3.0
0.4
Variable Total number of comprehensible sentences Grammatical accuracy measures
Grammatical complexity measures
Idea complexity measure % of sentences expressing a complex idea Content accuracy measure Average concept accuracy score
Number of Sentences Produced Because the students were given a limited amount of time (approximately 15 minutes) to complete the sentence assignment, the total number of sentences produced is a measure of the speed of processing. The main reason to include this variable was to find out whether students who produced more sentences had done so at the expense of grammatical accuracy or syntactic complexity.
Complexity of Ideas Expressed Based on content, sentences were classified as expressing simple or complex ideas. Sentences classified as expressing simple ideas included those that described properties, category membership, sources, mechanics, or functions of entities. Sentences classified as expressing complex ideas included those that explained conditions and circumstances for how and why entities, substances, or processes were related to one another. Thus, description of science terms was considered to be simple and explanation of science terms was considered to be complex. Table 2 shows several examples of simple and complex ideas that students produced. Two independent raters coded the complexity of each sentence. Agreement on classifications into the two categories was 95% (Cohen's • = .88). Consensus was reached on all discrepancies. As shown in Table 1, the average proportion of sentences that expressed complex ideas was 29.4% (SD - 13.9). It is important to note that this classification was based on the content rather than the structure of the
326
Wendy S. Francis, Laura F. Romo, and Rochel Gelman
sentence. Although, as will be explained later, idea complexity was associated with syntactic complexity, complex clause structures were not necessary for the expression of complex ideas, as shown by the examples in Table 2.
Table 2. Examples of Simple and Complex Ideas Idea Complexity Category
Examples of Student Sentences a
Simple Ideas
Water and dirt don't have the same temperature. Atmosphere is made up of different elements and compounds. The heat comes from the sunlight. Our body needs a lot of carbohydrates. We inhale oxygen. Water absorbs energy from sunlight.
Complex Ideas
The balloon has gas when it has air. Plants need water, carbon dioxide, and sunlight for photosynthesis. When we add heat energy to the water, it causes the water to change in temperature. An egg can't float in hot water.
aSpelling, punctuation, and grammatical errors have been corrected.
Content Accuracy Content accuracy was coded on a 5-point anchored scale from 0 (no correct science content) to 4 (completely correct science content). Scale anchor points were 0 = no display of understanding; 1 = mostly unclear and inaccurate; 2 = about half correct and half incorrect; 3 = mostly clear and accurate; 4 = clearly accurate display of understanding. For each student, an average accuracy score was computed by dividing the total score by the total number of sentences. On average, students' sentences were scored as "mostly clear and accurate", a score of approximately 3 (M = 3.03, SD = .44).
Second Language Writing
32 7
Relationships Among the Linguistic and Content Variables Relationships among the linguistic and content variables were examined at two levels of analysis, one that emphasized skill levels, and one that emphasized on-line processing. The skill analysis was performed at the student level. This student-level analysis would reveal the features that go together to characterize individual students' ideas and language skills. For example, this analysis will show whether students who express more complex ideas tend to use more prepositional phrases or make fewer grammatical errors. The on-line processing analysis was performed at the level of the individual sentence. This sentence-by-sentence analysis would reveal the characteristics that co-occur in individual sentences. For example, this analysis will show whether complex ideas are expressed using complex sentence structures and whether complex structures are more likely to contain grammatical errors. Thus, it can reveal cognitive trade-offs among different aspects of written language production.
Skill Learning: Student-Level Associations Seven variables were used in the skill-level analysis. The linguistic variables for the student-level analysis were the proportion of sentences containing at least one grammatical error, the proportion of sentences containing at least one subject-verb agreement error, the proportion of sentences containing at least one embedded clause, and the proportion of sentences containing at least one prepositional phrase. The total number of sentences produced was also included as a variable. The content variables were the proportion of sentences that expressed complex ideas and the average content accuracy score. Table 3 shows the correlations among the seven variables at the student level. The proportion of sentences containing grammatical errors was not significantly associated with any of the content or syntactic structure variables, although this variable appears to be negatively associated with all of the content and sentence structure variables. The small sample size limits the strength of the conclusions to be drawn, but the negative direction of these non-significant correlations suggests that students who express more complex or accurate ideas or who use more embedded clauses and prepositional phrases are not compromised on producing grammatically correct sentences. That is, there were no trade-offs among the skills of producing grammatically correct sentences, using complex sentence structures, and learning about science. An analysis focusing specifically on subject-verb agreement errors more definitively demonstrated the relationship between grammatical errors and complexity. The rate of subject-verb agreement errors was significantly negatively correlated with the proportion of complex ideas, science content accuracy, and the use of both embedded clauses and prepositional phrases. Thus, competence in producing correct verb morphology was associated with competence in producing more complex syntactic structure and more complex content.
328
Wendy S. Francis, Laura F. Romo, and Rochel Gelman
Table 3. Correlations Among the Main Content and Language Variables Across Students (N =20)
% of sentences expressing a complex idea Average content accuracy
Content accuracy
Embedded clause
.739**
.955**
.756**
% of sentences containing an embedded clause % of sentences containing a prepositional phrase % of sentences containing any grammatical error % of sentences with a subjectverb agreement error
Prep. phrase
Total Sentences
Grammar error-any
S-V Agr. error
.433 M
-.327
-.521'
.001
.346
-.306
-.562**
.278
.347
-.341
-.501'
.087
-.280
-.523*
.198
N/A a
-. 547*
-.248
Total number of sentences produced
*p <.05; **p <.01; ***p<.001; Mp<.10 Note" All p-values are based on a two-tailed test against the hypothesis that P = 0. a This correlation of .612 is not meaningful, because all sentences with subject-verb agreement errors by definition contain grammatical errors.
Students who wrote about more complex ideas also wrote more accurate ideas, which shows that those with more complex ideas had fewer misconceptions about the material, as explained by Gelman et al. (in press). Students who wrote more complex and/or accurate ideas also used complex sentence structures that contained
Second Language Writing
329
more embedded clauses and more prepositional phrases. That is, students who knew more science had more linguistic tools to express their knowledge. The correlations of the total number of sentences produced by each participant with the linguistic and content variables were examined to determine whether the students who produced more sentences did so at the expense of accuracy and/or complexity. Only one variable, the overall proportion of sentences with grammatical errors, was significantly associated with the total number of sentences produced, such that students who produced more sentences had more accurate grammar. Thus, it appears that the students who were more competent at producing correct grammar were able to produce more sentences in the time allotted, suggesting that those students had developed a greater degree of automaticity in sentence production. Importantly, none of the syntactic or idea complexity variables turned out to be negatively associated with the total number of sentences produced. Therefore, it is clear that there was no trade-off between students' productivity and the complexity of their sentence structures or content. In addition, students who worked faster did not sacrifice the accuracy of the scientific content expressed. From the student-level analysis alone, it does not appear that there are any cognitive trade-offs among the skills underlying the various language and content variables. The following sentence-level analysis will examine whether trade-offs occur in on-line production of sentences. On-Line Performance: Sentence Level Associations
In language production, several different cognitive processes must be executed simultaneously. As explained in the introduction, the attentional requirements of processing a newly learned second language are greater than those of processing one's native language. To the extent that the different components of language production and thought tap the same cognitive resources, trade-offs among the content and language features within sentences were to be expected. Such trade-offs imply that the mental effort used to produce one feature compromises the production of another feature and that the features exhibiting the trade-off require attention. The linguistic variables examined at the sentence level were whether the sentence contained a grammatical error, whether the sentence contained a subjectverb agreement error, whether the sentence contained an embedded clause, and whether the sentence contained a prepositional phrase. The content variables examined at the sentence level were the idea complexity classification, and the content accuracy score for the sentence. To examine sentence-level associations, it is not sufficient to merely sort all sentences bivariately and calculate a single measure of association for the entire set of sentences. When the marginal distributions of the variables vary across participants, such a measure will not necessarily show the same degree or even direction of association (Wickens, 1993). In this data set, it is clear that the marginal distributions (i.e., proportions of sentences with embedded clauses, grammatical errors, etc.) vary across students. Therefore, a statistical analysis that takes betweensubject variability into account was employed. For each student, a contingency table
330
Wendy S. Francis, Laura F. Romo, and Rochel Gelman
was constructed for the bivariate classification of each sentence on each pair of variables. Kendall's tau-b (Xb), a measure of ordinal association was calculated for each participant for each pair of variables. As suggested by Wickens (1993), the association measures obtained for each participant were subjected to a single sample t test against the hypothesis that the population mean sentence-level association is zero. The values of the observed mean Xb across participants for each pair of variables are shown in Table 4. It should be noted that measures of ordinal association are not on the same scale as the correlation coefficients obtained from continuously-distributed variables, such as those in the subject-level analysis. Therefore, the values should not be compared to those of Table 3.
Table 4. Mean Within-Student Sentence-Level Associations Among Content and Linguistic Variables: Values of Kendall's tau-b
Idea complexity classification Content accuracy Sentence contains embedded clause Sentence contains prepositional phrase
Content accuracy
Embedded clause
Prep. phrase
.122"**
.787***
.135'*
.148'**
.021 -.017
Grammar error-any .134'**
-.039
Subj-verb agr. error .082*
-.060
.111 **
.042
.062ra
-. 11 O*
Sentence contains any grammatical error Sentence has subjectverb agreement error * p < .05; ** p < .01; ***p < .001, rap < .10 Note: All p-values are based on a two-tailed single-sample t test against the hypothesis that the mean association in the population is zero.
At the sentence level, the rate of grammatical errors was positively associated with content complexity and the use of embedded clauses. That is, sentences that expressed complex ideas were more likely to contain grammatical errors (M = 79.5%) than were those that expressed simple ideas (M = 65.7%). Likewise, sentences with embedded clauses (M = 78.2%) were more likely to contain grammatical errors (M = 78.2%) than were sentences with simple structures (M =
Second Language Writing
331
66.8%). For sentences with prepositional phrases, the pattern was similar, in that sentences with prepositional phrases had more grammatical errors (M = 73.2%) than sentences without prepositional phrases (M = 68.0%), but the association was only marginally significant. These patterns clearly indicate a trade-off in sentence processing resources between the syntactic complexity and grammatical accuracy of the output. Looking specifically at subject-verb agreement errors, a similar pattern emerged. Sentences with complex ideas were more likely to contain subject-verb agreement errors (M = 42.9%) than were sentences with simple ideas ( M - 35.3%). The association between embedded clause usage and subject verb agreement errors was not significant. The direction of the association between prepositional phrase usage and subject-verb agreement errors was unexpected--sentences with prepositional phrases were less likely to contain subject-verb agreement errors (M = 32.6%) than were sentences without prepositional phrases (M = 43.5%). At the sentence level, the usage of embedded clauses and the usage of prepositional phrases were not associated, meaning that these structures did not necessarily appear in the same sentences. Having an embedded clause in a sentence did not increase or decrease the probability of having a prepositional phrase in the sentence. Content complexity was positively associated with both syntactic variables. Of the complex ideas, on average, 72.0% were expressed using sentences with embedded clauses. In contrast, less than 1% of sentences with simple ideas contained embedded clauses. A similar, but less dramatic pattern was observed in the usage of prepositional phrases; on average, 54.4% of the complex ideas were expressed using sentences with prepositional phrases, as compared to a lower rate of 39.2% for simple ideas. Content accuracy and grammatical accuracy were not associated, suggesting that they draw upon different resource pools. That is, the accuracy of the idea was not associated with the accuracy of the grammar used to express it. The accuracy of ideas was significantly associated with their complexity; the complex ideas were rated as more accurate than the simple ideas. More accurate ideas were also more likely to contain embedded clauses than were less accurate ideas.
Discussion Sentence Production Skill Analysis The analysis of summary measures of language skill across students indicated that the different components of language learning tend to be learned well by the same students. Specifically, the production of more complex sentence structure, in terms of either embedded clauses or prepositional phrases, was positively associated with correct usage of subject-verb agreement marking. Perhaps it is not surprising that the students who have a better mastery of verb morphology would produce more sentences with complex clause structures, because sentences with embedded clauses contain multiple verbs. The analysis of the relationship between syntactic
332
Wendy S. Francis, Laura F. Romo, and Rochel Gelman
complexity and overall grammatical accuracy was less conclusive, but suggests the same pattern, that students who produced more complex sentence structures produced fewer grammatical errors. This pattern of association suggests that these language subskills develop together in early stages of second-language learning. Across students, there was considerable variability in the total number of sentences produced in the time frame allowed, which prompted us to examine the relationship between the quantity of sentences and their quality. One possibility was that the students who produced more sentences would have done so at the expense of complexity or accuracy, or that the students who wrote more complex ideas or used more complex syntactic structures would not have time to write as many sentences. Considering all of the language and content complexity and accuracy variables, there was no indication of such a trade-off between sentence quantity and sentence quality across participants. If anything, the students who produced more sentences produced better quality sentences, in that they made fewer grammatical errors. The number of sentences produced was not reliably associated with any of the content or language complexity variables (although the directions of the associations were positive), which suggests that the students wrote sentences at a level of complexity commensurate with their language skills. The content variables, idea complexity and content accuracy, were highly correlated with each other, perhaps as indicators Of a greater understanding of the scientific material. These variables were also correlated with the use of complex syntactic structures. Furthermore, the proportion of sentences with complex ideas increased over the duration of the course (Gelman et al., in press). One explanation is that the complexity of students' ideas increased, because the students' capacity to remember the complex ideas may have increased. The influence of memory on complexity could have taken place either during learning, where over time they ~ became better able to store complex ideas in long-term memory, or during output, where they had to maintain the idea in working memory to be able to produce it. This suggests that one needs greater working memory capacity to produce complex ideas and to hold a complex idea in mind while producing a sentence to express it.
Online Processing Analysis At the sentence level, the majority of sentences with complex ideas were expressed using complex clause structures, whereas very few of the sentences with simple ideas had complex clause structures. Similarly, sentences with complex ideas were more likely to contain prepositional phrases than were those with simple ideas. In fact, a secondary analysis revealed that all of the sentences with complex ideas contained at least one embedded clause or prepositional phrase. This finding suggests that students generated sentence structures that were useful, or perhaps necessary, to explain their complex ideas. Scientific explanation typically includes purposes, causes, necessary and sufficient conditions, time and location circumstances, consequences, and comparison. Any of these components of explanation can be expressed using an adverbial clause or adverbial prepositional phrase. Another way to express causality is through causal verbs, such as "cause" or
Second Language Writing
333
"make", many of which can take an embedded noun clause as a complement. For increased precision in scientific explanation, it also helps to specify the identity of entities involved or give additional information about the entities involved. These types of information can be given using adjectival clauses (relative clauses) and adjectival prepositional phrases. With few exceptions, the students did not express their ideas with complex clause structures unless they were expressing a complex idea. This extreme positive association at the sentence level made the corresponding positive student-level correlation inevitable. Grammatical errors were more likely to occur in sentences with complex ideas. This trade-off in performance suggests that both production of correct grammar and production of complex ideas require attention and draw upon the same pool of cognitive resources. Grammatical errors were also more likely to occur in sentences with complex syntactic structure, indicating that production of correct grammar and production of complex syntax both draw attentional resources from the same pool. That sentences with complex ideas would suffer subject-verb agreement errors is not surprising given that the verb appears to be particularly sensitive to processing demands of the sentence. For example, Ferreira (1991) showed that in producing memorized sentences with complex objects, participants paused just before the main verb. In reading comprehension, King and Just (1991) found that reading times for the main verb in sentences with relative clauses were affected more than reading times for other words in the sentences by the syntactic complexity of that relative clause (subject-relative or object-relative). However, in our study, subject-verb agreement errors were reliably positively associated only with idea complexity, and not with the measures of syntactic complexity. Because grammatical accuracy is compromised both in sentences with complex ideas and sentences with embedded clauses or prepositional phrases, and because of the high association between idea complexity and use of embedded clauses, it is not immediately obvious which type of complexity draws the cognitive resources away from grammatical accuracy. To answer this question, an additional analysis examined the proportion of sentences with grammatical errors as a function of both idea complexity and syntactic structure for sentences. As shown in Table 5, the answer is not simple, because all complex ideas were expressed using either an embedded clause or a prepositional phrase, and only 6 of the simple ideas were expressed using embedded clauses, leaving two cells of the table insufficient for analysis. Of the sentences expressing simple ideas, the error rates for sentences with one prepositional phrase (64%) were no higher than for those with no prepositional phrase (65%), suggesting that sentences with prepositional phrases were no more taxing than those without. Of the sentences expressing complex ideas, the grarrunatical error rate for sentences containing an embedded clause with no prepositional phrase (72%) was no higher than for sentences containing a prepositional phrase but no embedded clause (75%). This suggests that the embedded clauses were not more taxing on attentional resources than prepositional phrases. (The error rate for complex idea sentences containing both embedded clauses and prepositional phrases is higher, but these are longer sentences with more
334
Wendy S. Francis, Laura F. Romo, and Rochel Gelman
opportunities for errors.) Taken together, these two findings suggest that syntactic complexity in and of itself does not compete with grammatical accuracy for attention. In contrast, considering only sentences that contained prepositional phrases, complex ideas were more likely to have a grammatical error (75%) than were simple ideas (64%). This pattern suggests that idea complexity, rather than clause complexity, competes with grammar for cognitive resources.
Table 5. Proportion of Sentences Containing Grammatical Errors as a Function of Idea Complexity and Syntactic Structure
Syntactic structure
Simple ideas
C o m p l e xideas b
No prepositional phrase and no embedded clause
65%
One prepositional phrase but no embedded clause
64%
75%
One embedded clause but no prepositional phrase
...a
72%
a Only 6 such sentences were produced. b No such sentences were produced.
A Comparison Between Skill Learning and On-Line Performance The pattern of associations among the language and content variables was different in the student-level analysis than in the sentence-level analysis, which shows that relationships among the sentence production skills were different from relationships among features of sentences in on-line production. The dissociations between these two levels of analysis are particularly informative. The strongest dissociations were observed among the correlations of grammatical accuracy with the complexity variables. At the student level, students who wrote more complex sentences in terms of content or clause structure made fewer grammatical errors. However, at the sentence level, the sentences with more complex content or clause structure were more likely to contain grammatical errors. Thus, the direction of the relationship was different for the two levels of analysis, meaning that complex sentences tended to have grammatical errors, but that didn't mean that students who attempted more complex sentences made more errors overall. Despite the fact that these students attempted more complex sentences, it seems that these students were less likely to make errors than students who wrote less complex sentences. Features that indicate syntactic complexity, embedded clauses and prepositional phrases, are positive indicators of more sophisticated language production skills. Therefore, the skills ought to be correlated across second-language learners as they
Second Language Writing
335
were in this study. The substantial increase in the use of such structures over the course of language acquisition indicates that sentences containing such structures are more difficult to produce. If one of the structures is to be produced, the attentional resources for producing another complex structure may not remain available. Given these attentional demands, if one of the structures is present in a sentence, it would be reasonable to expect that the other would not be present. In fact, in the present study, there was little if any indication that the uses of these two syntactic structures were associated either positively or negatively at the sentence level. Other indicators of complexity, say different types of embedded clauses, m!ght in fact be expected to have a negative or disjunctive relationship at the sentence level in the early stages of second language acquisition because of the increased processing demands of producing two rather than only one of the structures. Ih the sample analyzed here, instances of sentences with multiple embedded clauses were infrequent and therefore could not be analyzed in this way. This study was our initial attempt at implementing a new approach to the study of second language acquisition, and as such, it is limited in scope. We did not have access to comparable data for native English speakers. Therefore, we do not know whether these patterns are characteristic of English-speaking students in regular science classes. Perhaps a more interesting comparison would have been sentences produced by the same students in their native languages, but such data were not collected. Thus, although some of the findings run contrary to what the literature suggests for adult native speakers, we do not have primary data for native speakers. A second limitation of this study is that the only second language examined was English. Our sample did include students with different native languages, and the two larger language groups, Spanish and Korean, both showed the same patterns of association among the critical variables, despite the fact that they differed in their propensity to make some types of errors. It remains to be seen whether the patterns of association among the content and linguistic variables generalize to other second languages. We speculate that similar patterns may exist among science content sophistication, clause structures, and correct grammar across different second languages, although the specific error types involved may differ.
Conclusions
As a complex behavior, language production requires the successful simultaneous performance of several cognitive sub-tasks. In second-language sentence production, some of the components that do not appear to require attention in adult native speakers may require attention until sufficient experience with the language is logged. In learning to perform multiple attention-demanding tasks simultaneously, it is not uncommon to observe degraded performance relative to performing each task in isolation. So long as the essential components of second language production require attention, language production will be reduced in complexity relative to the native language and contain syntactic imperfections. As automaticity develops in the various component subskills of language production,
336
Wendy S. Francis, Laura F. Romo, and Rochel Gelman
trade-offs among sentence features in on-line production ought to be diminished, increasing speed and complexity and decreasing errors. In particular, as the production of correct syntactic morphology, such as subject-verb agreement becomes more automatic, sentence production should become more fluent (i.e., faster), and more resources should be available for translating complex ideas into written or spoken language. When studying characteristics of complex behavior, such as language, it is useful to examine relationships among critical characteristics both as indicators of students' language skill development and at the level of the individual sentence. Studying characteristics of second-language sentence production both at the level of the student and at the level of the sentence provided different kinds of information. Some seemingly conflicting sets of relationships emerged, but they can be understood by considering the different properties of skill acquisition and on-line performance. This distinction may also help us to interpret how attentional demands of bilingual sentence processing change with acquisition or vary according to proficiency.
Author Notes
This research was supported by grants from the University of California Linguistic Minority Research Institute, the U.S. Office of Education, and the National Science Foundation to Rochel Gelman; and by a Fernald Child Study Center Fellowship and University Elementary School Seed Grant to Laura Romo. We wish to thank George Meck, the ESL instructor who taught the curriculum and collected the classroom data.
References
Bock, J. K. (1982). Toward a cognitive psychology of syntax" Information processing contributions to sentence formulation. Psychological Review, 89, 1-47. Chincotta, D., Hy6n~i, J., & Underwood, G. (1997). Eye fixations, speech rate and bilingual digit span" Numeral reading indexes fluency not word length. Acta Psychologica, 97, 253-275. Estes, W. K. (1955). Statistical theory of distributional phenomena in learning. Psychological Review, 62, 369-377. Ferreira, F. (1991). Effects of length and syntactic complexity on initiation times for prepared utterances. Journal of Memory and Language, 30, 210-233. Garrett, M. F. (1975). The analysis of sentence production. In G. Bower (Ed.), The psychology of learning and motivation, Vol. 9 (pp. 133-177). New York: Academic Press. Gelman, R., Meck, G., Romo, L., Meck, B., Francis, W., & Fritz, C. O. (1995). Integrating science concepts into intermediate English as a second language (ESL) instruction. In R. F. Macias & R. G. Garcia-Ramos (Eds.), Changing schools for
Second Language Writing
337
changing students: An anthology of research on language minorities, schools, & society (pp. 181-203). Santa Barbara, CA: University of California Linguistic Minority Research Institute. Gelman, R., Romo, L. F., & Francis, W. S. (in press). Notebooks as windows on learning: The case of a science-into-ESL program. In N. Granott & J. Parziale (Eds.), Microdevelopment. Cambridge, England: Cambridge University Press. Grabowski, J. (1996). Writing and speaking: Common grounds and differences toward a regulation theory of written language production. In C. M. Levy & S. Ransdell (Eds.), The science of writing: Theories, methods, individual differences, and applications (pp. 73-91). Mahwah, NJ: Erlbaum. Harrington, M. (1992). Working memory capacity as a constraint on L2 development. In R. J. Harris (Ed.), Cognitive processing in bilinguals (pp. 123135). Elsevier Science Publishers. Kahneman, D. (1973). Attention and effort. Englewood Cliffs, NJ: Prentice Hall. Kellog, R. T. (1996). A model of working memory in writing. In C. M. Levy & S. Ransdell (Eds.), The science of writing: Theories, methods, individual differences, and applications (pp. 57-71). Mahwah, NJ: Erlbaum. Kemper, S., Kynette, D., Rash, S., O'Brien, K., & Sprott, R. (1989). Life-span changes to adults' language: Effects of memory and genre. Applied Psycholinguistics, 10, 49-66. King, J., & Just, M. A. (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language, 30, 580-602. Levelt, W. J. M. (1989). Speaking: From intention to articulation. Cambridge, MA: MIT Press. Meck, G. H., & Gelman, R. (1993). Interrelated science concepts in English as a Second Language: A science content text for intermediate level ESL instruction. First version of a curriculum designed to embed science instruction in ESL classes for 9th grade students in a Los Angeles School. UCLA. Murray, L. L., Holland, A. L., & Beeson, P. M. (1998). Spoken language of individuals with mild fluent aphasia under focused and divided-attention conditions. Journal of Speech, Language, and Hearing Research, 41, 213-227. Tolman, E. C. (1932). Purposive behavior of animals and men. New York: Century. Weisberg, B. (2000). Developmental relationships in the acquisition of English syntax: Writing vs. speech. Learning and Instruction, 10, 37-53. Wickens, C. D. (1984). Processing resources and attention. In R. Parasuraman" & D. R. Davies (Eds.), Varieties of attention (pp. 63-102). New York: Academic Press. Wickens, T. D. (1993). Analysis of contingency tables with between-subjects variability. Psychological Bulletin, 113, 191-204.
This Page Intentionally Left Blank
Bilingual Sentence Processing - R.R. Heredia and J. Altarriba (Editors) 9 2002 Elsevier Science B.V. All rights reserved.
15
Code-Switching in Preschool Bilingual Children
Zehra F. Peynircio~lu American University Aydln Y. Durguno~lu University of Minnesota, Duluth
Abstract Code-switching refers to the alternating use of two or more languages, either within a sentence (intrasentential) or between sentences (intersentential). Because code-switching is governed by grammatical rules, both language systems are presumed to be active while producing mixed sentences. Codeswitching occurs in discourse of fluent and nonfluent bilinguals alike, although intrasentential switches are often thought to be illustrative of the level of bilingualism or comfort with another language (e.g., Poplack, 1980). This chapter focuses on code-switching in preschool children, with a special emphasis on the linguistic context, comparisons between intra- and intersentential switching, the effect of level of bilingualism, differences in production and comprehension of sentences, and experimental demands. To this end, we present new data from 36 Spanish-English bilingual children between the ages of 3 and 5, and tie the results to the theoretical framework of Kroll and her colleagues (e.g., Kroll & Stewart, 1994; Sholl, Sankaranarayanan, & Kroll, 1995) on conceptual and lexical links between words in sentence processing.
Overview Code-switching, or alternating between two or more languages in a given sentence or between sentences, is a common but intriguing phenomenon in bilingual or multilingual sentence processing. It is quite prevalent, especially in cultures where the use of multiple languages are commonplace and taken for granted, such as amongst the Hispanic populations in the U.S., the later generations of guest workers in industrialized European countries, and even students in schools where intensive foreign languages are taught. Interestingly, it is almost exclusively limited to spoken language, and in the rare written cases, it is often limited to conversation-like correspondence. Start of serious research on code-switching can be dated to the early 1950s as part of research on bilingualism (e.g., Weinreich, 1953), and the effort has proliferated since the 1960s. One of the problems with integrating so much research, however, has been the many different approaches to studying code-switching. Within the linguistic,
340
Zehra F. Peynircio~lu and Aydm E Durguno~lu
sociological, cognitive, and educational settings, similar questions have been explored from vastly different perspectives. For instance, whereas some social scientists have been interested in the group dynamics or identity that lead to code-switching or codeswitching as expression of position of power (e.g., Pandharipande, 1990; Scotton, 1988), others have focused on detailed analyses of lexical issues in code-switching (e.g., Sridhar & Sridhar, 1980; Yoon, 1993). In addition, not everyone agrees on exactly what code-switching should refer to. In the literature, sometimes detailed distinctions have been drawn between such terminology as code-mixing, codechanging, borrowing, deliberate vs. accidental code-mixing, etc. For instance, borrowing has sometimes been excluded from code-mixing or intrasentential codeswitching because theoretically monolinguals can also intersperse a single target word from another language without having any competence in that language (e.g., Pfaff, 1979). Similarly, situational switching (e.g., Blom & Gumperz, 1972; Lipski, 1978), or using different languages in different settings or when switching topics, has also sometimes been excluded because it does not capture the linguistic constraints highlighted when switching occurs in the same mindset. In this chapter, we define code-switching as any alternation between English and Spanish, including borrowings because we know that all of our participants in the experiment we report had at least a working knowledge of both languages. Thus, in this chapter, intrasentential code-switching refers to any switch from the starting language in a sentence (e.g., "swimming muy muy faster" or "she put her swimsuit para ira la playa"), and intersentential code-switching refers to switching languages between sentences, (e.g., from Spanish to English-," Ella esta llorando. He took her balloon" or from English to Spanish--"To go home. Para tomar jugo de naranja"). We concentrate on such code-switching from the perspective of sentence processing. First we briefly review results relevant to comprehension and production of code-switched speech in adults as well as results obtained in a few studies with children. Next we present data from a detailed study with bilingual preschool children, which shed some light on the prevalance of code-switching in such children under different circumstances as well as touch on developmental factors. We then compare our results with those from relevant studies and also speculate on the implications of these results on the cognitive representations of two or more languages and interactions between them.
Effects of Fluency, Context, and Age on Inter- and Intrasententiai Code-Switching
Code-switching occurs naturally in bilinguals. One immediate question that arises is whether a critical level of fluency needs to be reached in one's second language (L2) or in both languages for young children before code-switching can be engaged in. Or as an extension, whether code-switching behavior increases with increasing degree of balance between the two languages of a bilingual. From a purely cognitive, sentenceprocessing perspective (as opposed to say social or cultural demands, cf. Fishman, 1965), in nonbalanced bilinguals, when one language is definitely more dominant than the other, one may expect lots of code switches from L2 to L1 as people are speaking in L2 and cannot find words or phrases to express themselves or are uncomfortable
Code-Switching in Bilingual Children
341
with the syntactic structure of the sentence they have started. There should be fewer switches from L 1 to L2. In addition, code-switching has been shown to bridge not only lexical gaps but structural gaps, as well (e.g., Gawlitzek & Tracy, 1996), which would mean that more "need" might arise for switching to L 1 from L2 than vice versa. On the other hand, because nonbalanced bilinguals predominantly use L1 as their base language in natural conversational settings, most of their switches should be from L 1 to L2 (cf. Poplack, 1980). It has also been found to be easier for them to switch from L 1 to L2 (e.g., Meuter & Allport, 1999). In balanced bilinguals, the direction of code switches should be more unpredictable from a cognitive perspective, although the context or social factors might have an effect. Thus, ease of finding the words or phrases, or compensating for structural difficulty is only one factor in determining whether code switching occurs, and probably less of a factor with more fluent bilinguals (cf. Clyne, 1967). Comfort with and dominant use of one language are factors which create an artifactual preponderance of L1 to L2 switches for nonbalanced bilinguals. Whether the balanced or the nonbalanced bilinguals engage in more code switching then is not really a fair question and the answer depends on the reason and the direction of switching. It also depends on the content (e.g., whether it is of single words or interjections or whole phrases) and context (e.g., the topic being addressed and the linguistic setting) of switching. Whether the switch occurs intersententially or intrasententially, however, has been found to be highly indicative of bilingual fluency. Usually, balanced bilinguals have been found to engage in many more intrasentential switching than nonbalanced bilinguals, and nonbalanced bilinguals have been found to prefer intersentential switching (e.g., Muysken, Kook, & Vedder, 1996; Nortier, 1990; Poplack, 1980). The prediction then is that as the degree ofbilingualism increases, so should the differences between intra and intersentential switching preferences in favor of intrasentential switching. The reasoning is that balanced bilinguals feel equally at home with the syntactic constraints of both languages and can manipulate them both with equal ease, whereas nonbalanced biliguals feel more comfortable keeping to a single language within a sentence so as not to violate syntactic structures while determining compatibility. The underlying assumption is that at least in balanced bilinguals, both L1 and L2 are activated automatically regardless of which language is being spoken (cf., Macnamara, 1967). Code-switching has also been shown to be highly dependent on context (e.g., Heredia & Altarriba, 2001; Verhoeven & Boeschoten, 1986). For instance, Verhoeven and Boeschoten (1986) have found that for Turkish children living in the Netherlands, most code-switching occurred when describing life in general, followed by when describing a picture of a market place, and least code-switching occurred when describing a teacher's visit to their homes. Social and cultural demands aside, and only from the perspective of sentence processing, availability of and comfort with linguistic elements in L 1 and L2 in different topics or contexts is certainly a factor that affects code-switching frequencies. The context created by who is involved in the conversation is also important. For instance, if a person is addressed in a code-switched manner, he/she should be more likely to respond in like manner (cf. Martin-Jones, 1995). Finally, the context created by the type of discourse (e.g., formal interview versus
342
Zehra F. Peynirciofflu and Aydm Y. Durguno~lu
friendly joshing) also affects code-switching tendencies (cf. McClure, 1981). Children who are exposed to more than one language can differentiate between those languages at least by age two (Nicoladis & Genesee, 1997). Most code-switching studies, however, have been done with adult bilinguals, and, with a few exceptions, many of the studies with children have involved detailed case-studies, observational studies, or experiments with only a handful of participants. We know that just like adults, bilingual children also code-switch readily, both intersententially and intrasententially. Interestingly, although intrasentential switching in adults increases with degree ofbilingualism and is less frequent than intersentential switching for those not very fluent bilinguals, it is more common than intersentential switching in young children, and there is usually no relationship between code-switching in general and children's ages (cf. Redlinger & Park, 1980). If it is assumed that with increasing age comes increasing linguistic competence, and with increasing linguistic competence in two languages comes increasing intrasentential code-switching, then the finding that age and intrasentential code-switching are not related might seem to be at odds. But perhaps increasing linguistic competence in L2 to be able to cope with syntactic constraints in both languages is less important in intrasentential code-switching propensity than the degree of balance per se between L1 and L2, however impoverished both languages may be at the moment, especially for young children. For compound bilingual children, whose L 1 and L2 are developing together and at equal rates, the degree of balance is very high, and indeed code-switching is explained using the Unitary Language System Hypothesis (e.g., Genesee, 1989), which holds that the two languages for such children are really undifferentiated. Syntactic constraints start appearing later (cf. Meisel, 1989), but this does not pose a problem for compound bilingual children in terms of code-switching because they also develop together. Thus, if it is the degree of balance between the two languages that affects intrasentential code-switching, then the finding of no difference between rates of such code-switching and age is no longer surprising. The question remains, however, about whether young children with distinct L 1 and L2 systems and of varying degrees ofbilingualism show differences in code-switching behavior, especially in intrasentential code-switching behavior.
The Present Experiment We tested young children with varying degrees of bilingualism in a controlled experiment. We created three linguistic context situations such that the participants were more encouraged to speak in L1 in one, in L2 in another, and in a mixed fashion in the third. All participants participated in all situations. Given that more fluent or balanced bilinguals engage in more code-switching, at least intrasententially, but such switches have usually been observed mainly from L1 to L2 for all bilinguals (cf., Meuter & Allport, 1999; Poplack, 1980), we thought it would be interesting to create situations in which the context or presentation language was L2. In that case, from the perspective of ease 0fsentential processing, less balanced bilinguals might be expected to code-switch (L2 to L 1) more often than when the context/presentation language was
Code-Switching in Bilingual Children
343
L 1. That is, the gap between intrasentential code-switching tendencies might narrow between the more and less balanced bilinguals when the context/presentation language was L2. Likewise when the context/presentation language was itself mixed. In addition, all of the stimuli were rotated through each of the three presentation language conditions (L 1, L2, and Mixed) across participants so that the content of the story or the picture would not be a confound in determining language choice. This was important given the findings that code-switching in children was highly dependent on the topic or context (e.g., Verhoeven & Boeschoten, 1986). In the present experiment, because all children participated in all conditions and different topic contents were provided and held constant, it would be possible to observe code-switching behavior only as a function of changing language context and as a within-participant variable. For each presentation language condition, we presented two types of stimuli, a picture that participants were asked to describe, and a short story the participants were asked to retell and answer questions about. We assumed that whereas the story condition would look at both comprehension of L 1, L2, and Mixed presentations and production, the picture condition would encourage production without being dependent on correct comprehension. Drawing such a distinction between comprehension and production was important because in his studies, Kolers (1966, 1974) had shown that comprehension of mixed-language presentation was as good as comprehension of the average of L1 and L2 unilingual presentations (cf. Blair & Harris, 1981), whereas production of mixed-language took longer. Thus, he had concluded that the possible detrimental effects of code-switching in production was irrelevant in comprehension. In the present experiment, we added another variable, that of the level of bilingualism, while exploring such possible differences in code-switching between comprehension and production situations in young children.
Method Participants The participants were a total of 36 preschool children from two Spanish-English bilingual jump-start programs in the metropolitan Washington, D.C. area. They were all between the ages of 3 and 5 (with a mean age of 4.31); 31 were more fluent in Spanish than in English, 4 were more fluent in English than in Spanish, and 1 was equally fluent in the two languages. They were all rated by their teachers in terms of fluency in both languages separately on a scale of 1-10. The difference between the scores for the two languages constituted the primary measure of our bilingualism balance. We considered a child with a score of, for instance, 6 on the Spanish scale and 4 on the English scale (a difference of 2) to have the same degree of imbalance as a child with a score of, for instance, 9 on the Spanish scale and 7 on the English scale (again a difference of 2). The degree of imbalance ranged from 0 to 8 (mean of 2.67). These measures of participants' degrees of bilingualism were corroborated through several weeks of classroom observation of each of the participants before the experiment was conducted.
344
Zehra F. Peynircioglu and Aydm Y. Durgunoglu
Materials For the pictorial presentation, three busy scenes, one depicting the town, one the home, and one the beach, were chosen from children's books (see Appendix). For the verbal presentation, three short paragraphs, one about going to the zoo, one about going swimming, and one about going to the doctor were written by the experimenters (see Appendix). All paragraphs were written in English, in Spanish, and in two mixedlanguage versions; the two mixed versions differed from each other in that the words written in Spanish and English were completely interchanged. A tape recorder was used to record participants' speech. The tapes were transcribed later by two independent judges who were Spanish-English bilinguals. Design All participants were shown all the pictures and read all the stories, in two sessions. Half (18 participants) were shown the pictures in the first session and read the stories in the next session, and the other half did the reverse. The presentation order of both the pictures and the stories was the same for all participants. Within each half of the participants, however, the language of presentation was varied across three groups of participants such that all stories and pictures were presented in all languages (Spanish, English, Mixed) equally often. Thus, a third (6 participants) were shown the first picture and read the first story in English, the second picture and the second story in Spanish, and the third picture and the third story in mixed-language (with 3 participants getting one mixed version and the other 3 getting the complementary mixed version). For another third of the participants the presentation order was Spanish, Mixed, English, and for the final third, Mixed, English, Spanish. Procedure All participants' parents had signed parental consent forms presented in either English or in Spanish, and were told in detail about the experiment. The teachers had also rated each participant's English and Spanish competence on a scale from 1-10, and the experimenters had observed each participant in free play and classroom situations to make their own independent assessments. The experiment was conducted in a separate room during the participants' regular classroom sessions. Each participant was tested individually in two separate sessions, neither of which lasted more than 15 minutes. They were also given breaks in between stories or pictures. In the pictures condition, the participants were shown each picture and asked to tell what was going on in the picture. They could either describe the things in it or tell a story about it. The experimenters took care to speak only in English, only in Spanish, or in a code-switching fashion depending on the presentation language condition of the picture. Thus, for instance, if the first picture was in the English condition, the experimenter gave the instructions and interacted with the child only in English. The children were of course free to speak in any language they wished, although the experimenter would answer in only the language of the condition. Because the children
Code-Switching in Bilingual Children
345
were used to being addressed in both Spanish and English as well as in a mixed fashion during their days at the jumpstart program (by their teachers, friends, as well as the teachers' aides, two of whom also collected the data and did the classroom observations), they did not think it unusualto have the experimenter talk to them in the language of the condition and switch for the next condition. If the child was shy and did not speak, the experimenter would cue by pointing randomly and asking what he/she saw there. After each picture, a mini-break was given and the language switched for the next picture. In the stories condition, the procedure wa~s very similar except that the children were read the stories, then asked to relate what had happened in the story, again using only the language of the condition. Ifa child did not speak spontaneously, then a set of four questions were asked, such as "what animals did the children see at the zoo?" in the zoo story. Again, after each story, a mini-break was given and the language switched for the next story. Half of the participants had the pictures first and then the stories in the next session, and the other half had the reverse.
Results and Conclusions
In both pictures and stories, because we had both English-dominant and Spanishdominant participants, we report the results as a function of first language (L 1) and second language (L2), which in all but one case also coincides with the participants' dominant and weaker languages, respectively, rather than as a function of English and Spanish. For our only balanced bilingual participant, L1 was Spanish. Also, the "Mixed" condition refers to intrasentential code-switching (e.g., "she put her swimsuit para ir a la playa" or "swimming muy muy faster"). We should note that the number of intrasentential switches within a given sentence were not relevant; as long as there was any switch, that sentence was considered to belong to the Mixed condition. The direction of the switch was determined from the starting language. Thus, "este apples" would be a Spanish-English switch whereas "swimming muy muy faster" would be an English-Spanish switch. We report the intersentential code-switching results in comparison with intrasentential code-switching results separately, but when comparing the percentage of times participants spoke in L 1, L2, or Mixed conditions, the Mixed condition necessarily refers to intrasentential code-switching. There were no effects of order of presentation language or the version of Mixed presentation participants received in stories (i.e., which words were presented in Spanish and which words in English), and thus all data were collapsed across these variables. We first analyzed the results using a 2 x 3 x 3 ANOVA (pictures/stories x presentation language condition--L 1, L2, M x response language--L 1, L2, M). Overall, there was no main effect of the type of task (pictures vs stories) or of presentation language (L 1, L2, or M), both ps >. 10. There was a main effect of response language, however, F(2, 34) - 121.41, p < .01; L 1 was the preferred language. There was also an interaction between task and response language, F(2, 34) - 4.03, p < .05; more L1 r6sponses were given in the story condition than in the picture condition. Finally, and perhaps most importantly for present purposes, there was an interaction between
346
Zehra F. Peynircio~tu and Aydm E Durguno~lu
presentation language and response language, F(4, 32) = 19.66, p < .01. Participants changed their language depending on the situation and responded in the language that the materials were presented in. We then analyzed the pictures and stories separately. By and large, there were a great many more responses for the pictures overall (mean of 84.3) than for the stories (mean of 20.4). Thus, there were more data with the pictures, and the overall ANOVA had indicated that more L2 responses were given with these pictures. Looking at the participants' responses with the two types of materials in more detail, however, we saw that both the numbers of responses and code-switching proportions were positively correlated, rs = 0.44 and 0.46, respectively, both ps < .01. Thus, children who were more likely to speak or code-switch in the pictures condition were also more likely to speak or code-switch in the stories condition. It is just that they did both more frequently in the pictures condition when there was a lot more information that could be conveyed as well as when there was no comprehension involved but only production. We should also note that although the correlation between total number of responses and age was significant, r = 0.45, p < .01, that between mixed-language use percentages and age was not r = .05, p > . 10. Thus, older children did tend to speak more often and give more responses than younger children, but younger children were just as likely to switch codes intrasententially as the older children (cf. Redlinger & Park, 1980). In addition, the correlation between mixed-language use percentages and degree ofbilingualism was significant, r - 0.32, p < .05. Thus, although age per se did not influence intrasentential code-switching tendencies, consistent with previous research (e.g., Poplack, 1980), the more proficient bilinguals were more likely to engage in code-switching, at least intrasententially. Pictures
Table 1 summarizes the results of the pictures condition. As can be seen, collapsed across language of presentation condition, participants spoke most often in their L1, then in L2, and finally in a Mixed fashion (ts(35)= 2.39 and 5.19 between L1 and L2 and between L2 and Mixed, respectively, ps < .02). Thus, the results obtained with the overall ANOVA were mirrored with pictures alone, as well. In further analyses, however, we found that even though the number of times the participants spoke did not differ as a function of presentation language (M = 26.9 times when presentation language was L 1, 29.3 times when it was L2, and 28.0 times when it was Mixed), F(2, 105) = 0.58, p > .10, the percentage of responses given in each language did. Specifically, although participants gave more L 1 responses than L2 responses when the presentation language was L1, t(35)= 7.52, p < .01, when the presentation language was L2, their language of responding also changed and they gave more responses in L2 than in L 1, t(35) = 4.06, p < .01. Mixed-language responding was the lowest in both presentation-language conditions; it was significantly lower than the less preferred L 1 when the presentation language was L2, ts(35) = 3.36, p < .01, although the difference between the less preferred language and Mixed language conditions did not reach significance when the presentation language was L1, t(35)= 1.43, p > .10. Again, the
Code-Switching in Bilingual Children
347
results obtained with the overall ANOVA were mirrored with pictures alone. Interestingly, when the presentation language was itself Mixed, again L1 responses were the most frequent, followed by L2 responses, and then Mixed-language responses (ts(35) = 3.90 and 3.92 between L1 and L2 and between L2 and Mixed Language, respectively, ps < .01).
Table 1. Mean Percentages of Responses in Different Languages as a Function of Presentation Language for Picture Stimuli Presentation in
% of Responses in L1 L2 Mixed
L1
L2
Mixed
Overall
65.9 21.3 12.8
31.6 62.0 6.4
57.8 33.4 8.8
51.8 38.9 9.3
In addition, proportions of Mixed responses, or intrasentential code-switching, also changed as a function of language of presentation, F(2, 105) = 3.59, p < .03. Contrasts revealed that the main difference in intrasentential code-switching rates or proportion of Mixed responses was between L 1 and L2 presentation language conditions, t(3 5) = 2.66, p < .01; somewhat surprisingly, differences in intrasentential code-switching rates were not different between L 1 and Mixed and L2 and Mixed presentation conditions. That is, participants did not tend to code-switch more when the experimenter codeswitched and set the context. Indeed, Mixed presentation acted pretty much like L 1 presentation in determining language preference in responding and intrasentential codeswitching rates, implying perhaps that participants found it easier to function within L 1 as their base structure when they were confronted with code-switches. Given that our language-context change was successful in influencing the participants' language preferences in responding, we next looked at whether the degree of balance between the participants' languages affected the likelihood of their responding patterns in the different presentation language conditions. We divided the participants into two groups as a function of bilingual fluency, or balance. Those whose rating differences between the two languages were less than 3 were classified as "more" balanced and those whose rating differences were greater than 3 were classified as "less" balanced. There were 8 participants whose rating differences were exactly 3, and we assigned 4 of them to the more balanced and 4 of them to the less balanced group on the basis of complementary data from our classroom observations. Thus, there were 18 participants in the more balanced and 18 participants in the less balanced group. Of particular interest were any differences in mixed responding or
348
Zehra F. Peynircioglu and Aydm Y. Durgunoglu
intrasentential code-switching rates in the different conditions. The number of times the participants spoke was not influenced by degree of balance in bilingualism in any of the presentation language conditions (all 12s >. 10). However, even though the patterns of responding, both in general and as a function of presentation language, were the same for both the more and the less balanced group, overall, participants in the less balanced group tended to use L 1 more often than those in the more balanced group (54.7% tO 48.9%) and engage in intrasentential codeswitching less often than those in the more balanced group (7.7% to 11.6%), although these differences were not statistically significant. Interestingly, in both L2 and L1 presentation language conditions, it was the more balanced bilinguals who engaged in more intrasentential code-switching than the less balanced bilinguals (7.6% to 2.6% in L2 and 17.1% to 5.9% in L1). Thus, more balanced bilinguals engaged in more intrasentential code-switching, and it was easier for both groups to code-switch when the presentation context was in L1 than L2. In fact, the gap in the proportion of intrasentential code-switches did not close between the more and less balanced bilinguals as we had predicted when the general linguistic context was L2. That is, although less proficient in L2, the less balanced bilinguals preferred to speak in L2 at least as much as the more balanced bilinguals (62.8% vs 58.5%). Perhaps lexical difficulties that arose when speaking in L2 were not strong enough to overcome structural difficulties that arose when code-switching intrasententially. Age within each degree of balance group did not make a difference. Similar patterns were obtained for the younger more balanced bilingual children as the older more balanced bilingual children as well as for the younger and older less balanced bilingual children (all ps >. 10). Thus, as long as degree ofbilingualism was the same, syntactic competency that was assumed to have increased with age did not play a role in intrasentential code-switching rates. The results of the intersentential switches and in comparison with intrasentential switches are shown in Table 2 and as a function of degree of bilingualism. As can be seen, there were 6 possible types ofintersentential switches: L l-L2: "Ella est~i llorando. He t o o k h e r balloon"; L2-LI: "To go home. Para tomar jugo de naranja"; L2-M: "Swimming, swimming. Under de agua"; M-L2: "That's pap~i, nifio y mamfi, that's papfi y grama. He put the door on"; L l-M: "Un gato. Gato y clock"; M-L 1: "Perro con bunny rabbit y gato. Est~i dormiendo." Following a Mixed sentence by another Mixed sentence was not considered to be an intersentential switch even if the base languages were different (e.g., "El la de quiere con du con porque esta de swim? He gets in the water que se condo que tiene suefio") because by definition both sentences were uttered in the same language (i.e., Mixed). Overall, all participants code-switched intersententially much more often than intrasententially; in fact, this was true for 35 out of the 36 participants (p < .01 on a sign test). Similar results were obtained when the presentation language was L2 (27 showed this result, 1 showed the opposite results, and there were 8 ties), L 1 (30 showed this result, and there were 6 ties), or Mixed (30 showed this result, and there were 6 ties), all ps < .02 on a sign test. There were no differences as a function of whether the participants were more or less balanced bilinguals or as a function of their age (all ps > .10).
Code-Switching in Bilingual Children
349
Table 2. Inter- and Intrasentential Code-Switching Ratios in Percentages, the Directions of Intrasentential Switches, and the Rate and Directions of lntersentential Switches as a Function of Presentation Language Condition and Degree of Bilingualism for Picture Stimuli
Presentation Language L1
L2
Deg. of Biling. High Low M High Ratio of Intersentential 69 to Intrasentential 31 Switching
Mixed
Overall
Low
M
High Low M
High
Low
M
75
72
71
75
73
78
73
76
73
74
74
25
28
29
25
27
22
27
24
27
26
26
Intrasentential Switches 89 % L1 to L2 11 % L 2 t o L1
77 23
83 17
51 49
35 65
43 57
58 42
73 27
66 34
66 34
62 38
64 36
Intersentential Switching Rate Overall 36 L1 toL2 7.4 L2toL1 6.1 L2toM 4.0 MtoL2 3.7 L1 toM 7.3 MtoL1 7.0
21 4.6 4.3 1.3 1.5 4.9 4.8
34 6.0 5.2 2.7 2.6 6.1 5.9
18 3.1 4.5 2.7 2.2 2.3 3.0
15 2.5 3.1 3.4 3.5 0.9 1.8
17 2.8 3.8 3.1 2.9 1.6 2.4
29 7.3 6.7 2.6 3.0 5.0 4.4
28 7.2 8.3 1.8 1.3 4.4 5.2
29 7.3 7.5 2.2 2.2 4.7 4.8
28 5.9 5.8 3.1 3.0 4.9 4.8
21 4.8 5.2 2.2 2.1 3.4 3.9
25 5.4 5.6 2.7 2.6 4.2 4.4
In addition, when intrasentential code switching occurred, overall, it was easier to switch from L 1 to L2 than vice versa (24 participants switched more often from L 1 to L2, 7 switched more often from L2 to L1, and there were 5 ties), p <. 02) and there was no difference in this tendency as a function of whether they were more or less balanced bilinguals). However, when the presentation language was L2, the proportion of L1 to L2 and L2 to L1 switches were about equal for more balanced bilinguals and indeed, as expected, the proportion of L2 to L l switches dominated for less balanced bilinguals. Thus, even though the rate of intrasentential code-switching was always higher for more balanced bilinguals and regardless of presentation language (although a higher rate was obtained when the presentation language was L 1), as mentioned above, when code-switching did occur, its direction was indeed influenced by the presentation language. More code-switching occurred when the base language was L1, but the
350
Zehra F. Peynircio~lu and Aydm Y. Durguno~lu
direction of the switches changed when the base language was L2, at least for less balanced bilinguals. When we look at the more common intersentential code-switching results, we see that except when the presentation language was L1, both the more and the less balanced participants tended to have similar code-switching rates. In fact, overall about a quarter of their total utterances with the picture stimuli were intersententially codeswitched. The only influence of degree of bilingualism was apparent when the language of presentation was L 1, and the more balanced bilinguals code-switched more than did less balanced bilinguals. Age of participants did not play a role. Also of interest is the direction of the switches occurring between sentences. Because there were not enough data to yield meaningful statistical analyses, none were done. But simply looking at Table 2, it appears that intersentential code-switching occurred regardless of the base language of the preceding sentence. And although certainly not strikingly so, it also appears that the highest percentage of switches occurred between non-intrasententially code-switched sentences, and the lowest between L2 and intrasententially code-switched sentences. Stories
Table 3 summarizes the results of the stories condition. Overall, there was considerably less data because participants did not speak as much in response to the story stimuli. Many of the results were, however, similar to those obtained with pictures, and all of the results obtained with the overall ANOVA were mirrored with just the stories, as well.
Table 3. Mean Percentages of Responses in Different Languages as a Function of Presentation Language for Story Stimuli
Presentation in
% of Responses in L1 L2 Mixed
L1
L2
Mixed
Overall
79.8 7.2 13.0
29.5 60.9 9.6
65.0 19.6 15.4
58.1 29.2 12.7
Participants spoke most often in their L 1, then in L2, and finally in a Mixed fashion (ts(3 5)= 5.59 and 3.17 between L 1 and L2 and between L2 and Mixed, respectively, ps < .01, respectively). When we looked at the same results in the three presentation language conditions separately, again, we found that even though the number of times
Code-Switching in Bilingual Children
351
the participants spoke did not differ as a function of presentation language (a mean of 6.2 times when presentation language was L 1, 6.3 times when it was L2, and 7.6 times when it was Mixed), F(2, 105) = 1.59,.p > . 10, the percentages of responses given in each language did differ as a function of presentation language. Specifically, the same switch in preferred languages occurred as with the picture stimuli and participants gave more L 1 responses than L2 responses when the presentation language was L 1, t(35) = 14.31, p < .01, and they gave more L2 responses than L1 responses when the presentation language was L2, t(35) = 4.22, p < .01. When the presentation language was itself Mixed, again L 1 responses were the most frequent (t(3 5) = 6.83, p < .01, for the difference between L1 and L2), but this time there was no statistical difference between L2 and Mixed Language (t(3 5) = 0.32, p > . 10). Also, unlike with the picture stimuli, proportions of Mixed responses, or intrasentential code-switching, did not change as a function of language of presentation, F(2, 105) = 1.13, p > . 10.
Table 4. Inter- and Intrasentential Code-Switching Ratios in Percentages, the Directions of Intrasentential Switches, and the Rate and Directions of lntersentential Switches as a Function of Presentation Language Condition and Degree of Bilingualism for Story Stimuli
Presentation Language L1
L2
Deg. of Biling. High Low M High Ratio of Intersentential 67 to Intrasentential 33 Switching
Mixed
Overall
Low
M__
High Low M
High
Low
M
50
59
74
69
72
63
71
67
68
63
66
50
41
26
31
28
37
29
33
32
37
34
Intrasentential Switches 95 % L 1 to L2 5 % L2 to L1
66 34
81 19
33 67
17 83
25
74 26
91 9
83 17
67 23
58 42
63
75
Intersentential Switching Rate Overall 29 LltoL2 2.8 L2toL1 3.7 L2toM -Mto L2 2.3 LltoM 11.7 Mto L1 8.2
12 -1.1 1.1 -3.7 6.0
21 1.4 2.4 0.6 1.2 7.7 7.1
25 2.7 5.1 4.6 5.8 3.2 3.9
13 2.7 3.2 2.9 2.8 -1.8
19 2.7 4.2 3.8 4.3 1.6 2.9
29 3.4 3.4 3.0 2.7 10.0 7.0
20 3.7 4.2 -1.6 5.1 5.2
25 3.6 3.8 1.5 2,2 7.6 6.1
28 3.0 4.1 2.5 3.6 8.3 6.4
15 2.1 2.8 1.3 1.5 2.9 4.3
22
37
2.6 3.5
2.0 2.6 5.6 5.4
352
Zehra F. Peynircioglu and Aydm Y. Durgunoglu
Again, the number of times the participants spoke was not influenced by degree of balance in bilingualism in any of the presentation language conditions (all ps > . 10). Even though the pattems of responding were the same for both the more and the less balanced group, however, overall, participants in the less balanced group again tended to use L1 more often than those in the more balanced group (63.8% to 51.6%). Also, they engaged in intrasentential code-switching less often than the participants in the more balanced group (overall: 9.9% to 16.3%; when the presentation language was L 1: 6.6% to 16.9%; when the presentation language was L2: 5.8% to 10.8%). Thus, it was easier for both groups to code-switch when the presentation context was in L 1 than L2; and the gap in the proportion of intrasentential code-switches did not close between the more and less balanced bilinguals when the general linguistic context was L2. Once again, age did not influence any of these results. The results of the intersentential switches and in comparison with intrasentential switches and as a function of degree of bilingualism are shown in Table 4, and they pretty much mirrored those with pictorial stimuli. Most of the time, the differences that were observed were exaggerated even more. The only differences between the patterns of responding with the pictures and with the stories were that with stories, unexpectedly, less balanced bilinguals' intra- and inter-sentential code-switching rates were about the same in the L 1 presentation context, more balanced bilinguals, just like less balanced bilinguals, switched more often from L2 to L1 in the L2 presentation context, and a higher rate of switching seemed to occur between L1 and Mixed sentences rather than non-intrasententially code-switched sentences.
Grammatical Mistakes and Comprehension Mistakes We looked at grammatical mistakes while speaking in L 1, L2, and Mixed Language in response to both pictures and stories. Not surprisingly, in general, there were fewer such mistakes for both types of stimuli with increasing age, r = .37, p < .03 for pictures and r = .29,p < .09 for stories. The number of mistakes (means of 3.9 in L1,4.7 in L2, and .7 in Mixed with pictures and 1.1 in L1, 1.9 in L2, and .3 in Mixed) with pictures did change as a function of response language, F(2, 105)= 8.57 with pictures and F(2, 105) = 6.52 with stories, bothps < .01. This was true for both more balanced bilinguals (F(2, 51) = 3.40, p < .05 with pictures and F(2, 51) - 2.69, p < .08 with stories) and less balanced bilinguals (F(2, 51) = 5.60, p < .01 with pictures and F(2, 51) = 3.71, p < .05) with stories. With pictures, there was no difference between the number of mistakes made while speaking in L1 and in L2, t(35) - . 7 3 , p >. 10 whereas both L1 and L2 mistakes were greater than Mixed Language mistakes (ts(35) = 3.89 and 3.16, respectively, both ps < .01). Of course, for the most part this was probably due to the differences in the sheer numbers of times spoken in each language, but it might also have been reflective of code-switching in order to have an easier time in expressing oneself. With stories, more mistakes were made in both L2 and L1 compared separately with the Mixed Language condition, ts(35) = 1.81 and 3.61, and ps < .01 and .08, respectively. We also looked at comprehension of stories by looking at the number of conceptually wrong responses to questions. In general, although the number of wrong
Code-Switching in Bilingual Children
353
responses tended to decrease with increasing age, the correlation between age and number of wrong answers was not statistically significant, r = .22, p >. 10. There was also no effect of presentation language on the number of wrong responses, F(2, 105) = 1.14, p >. 10. The mean numbers of wrong responses after hearing the story in L 1 was 1.44, in L2 was 1.72, and in a Mixed Language was 1.97. Comprehension as measured by correct responses as a function of presentation language or the language in which the story was told was about the same for all languages. Thus, consistent with Kolers' (1974) observations, comprehension was not affected by code-switching in children, either. When we looked at the above results as a function of degree of balance in bilingualism, as well, we found that there were no differences whatsoever between the more and the less balanced bilinguals.
Conclusions
In this experiment we looked at children's intra- and intersentential code-switching tendencies as a function of age, degree of balance in bilingualism, and the context of language of presentation. By and large, the results were consistent with those found with adult bilinguals. Unlike the results reported in the Redlinger and Park (1980) study with young children, intrasentential code-switching was used less often than intersentential code-switching and the instances of switching (e.g., as a function of presentation language) were influenced by the degree of bilingualism; the more often used intersentential code-switching, however, was relatively independent of degree of bilingualism (cf. Meuter & Allport, 1999; Poplack, 1980). It appears that comfort with sentence processing, both lexically and syntactically, is related to the degree of intrasentential code-switching children will engage in, just as with adults. Switching between sentences on the other hand is easier because judging compatibility is no longer an issue, and thus less balanced bilinguals also do it quite readily. Interestingly, code-switching appeared to be independent of age as well as linguistic proficiency even in an age group (3 to 5) where language development is at a fast pace (cf. Redlinger & Park, 1980). The direction of intrasentential code-switching was another question of interest. Although somewhat counterintuitive in that it meant ostensibly going from the easier to the more difficult, children engaged in more L1 to L2 switches than vice versa. Consistent with those of other studies (e.g., Poplack, 1980), such a finding has usually been attributed to the greater difficulty faced in syntactic structure changes in L2. That is, it has been assumed that L2 to L1 switches are more difficult because the less balanced bilinguals are more uncomfortable with the word order or L2 and hence more uncomfortable with where and how to make the switches correctly. Another way of viewing the present results might be to connect these findings to those reported in translation studies in general. In exploring translation from one language to another with adults, usually, translation from L2 to L1 is faster and more accurate than translation from L1 to L2, and it has been proposed that translation from L2-L1 is lexically mediated whereas translation from L1 to L2 is conceptually mediated (e.g., Kroll & Stewart, 1994; but see Altarriba & Mathis, 1997). Within this framework,
354
Zehra F. Peynircio~lu and Aydm Y. Durgunoglu
Sholl, Sankaranarayanan, and Kroll (1995) showed that picture naming in either L 1 or L2, through priming the concept itself, helped only L 1 to L2 translation and not L2 to L1 translation because the latter was lexically mediated. As an extension, at the sentence level, participants in our study preferred L1 to L2 switches when codeswitching was intrasentential, possibly because speaking about both the pictures and the stories was conceptually rather than lexieally driven. Thus, even though we were not testing translation of equivalents but sentence processing, a larger unit than single words might have been activated at the conceptual level and made it easier to access compatible L2 phrases. Such a finding is also consistent with Meuter and Allport's (1999) findings of a lesser cost of switching in terms of response latencies from L 1 to L2. Thus, the preferred direction of code-switching in sentence processing appears to lend support to previous findings at the lexical level. The direction of intrasentential code-switching also depended on the linguistic context, however. When the presentation language was L2, participants code-switched more often from L2 to L1 than vice versa. Of course, such a result may not be surprising given that participants' base language became L2 and hence more switching would be expected from the base language to the other simply based on sheer usage frequency. But we had predicted that when the presentation language was L2, from the perspective of ease of sentential processing, the less balanced bilinguals might be expected to code-switch (L2 to L l) more often than when the presentation language was L l, and the gap between intrasentential code-switching tendencies might narrow between the more and less balanced bilinguals when the presentation language was L2. These predictions did not hold up. In addition, given that participants still preferred L 1 to L2 switches even when the presentation language was mixed such that the responses could be based on either language, it appears that the reversal of the preferred direction of switching when the presentation language was L2 was worthy of note and influenced by the context itself in addition to the sheer number of sentences started in one or the other language. Overall, the degree of balance in bilingualism had a moderating effect in that, especially with the stories, the more balanced bilinguals preferred the L l-L2 switch more than the less balanced bilinguals. Based just on comfort with and dominant use of L 1 in all situations, we should have expected the opposite--that the less balanced bilinguals would engage in a greater proportion of L1-L2 switches compared to more balanced bilinguals. But it appears that just like in other contextual domains, the linguistic context set by the experimenter also influenced code-switching tendencies and direction, especially for the less balanced bilinguals. Finally, consistent with results found with adult bilinguals (e.g., Blair & Harris, 1981; Kolers, 1966, 1974) comprehension as measured by being able to answer content questions about the stories correctly was not influenced by the language in which the story was presented. Production, however, as measured by both the preferred direction of intrasentential code-switching (L 1 to L2 vs. L2 to L l) and number of grammatical mistakes made while speaking in L1, L2, or Mixed Language, was influenced by presentation language or linguistic context as well as by the degree of balance in children's bilingualism. Future directions in studying code-switching behavior in young children might include conducting separate smaller scale experiments to explore the influences of linguistic proficiency and degree ofbilingualism more systematically, to
Code-Switching in Bilingual Children
355
concentrate on qualitative content analysis, especially within each presentation language context separately, and to investigate similar questions with bilinguals whose languages are more or less related to each other than English and Spanish.
Author Notes
We thank Megan Reilly and Daviana Greenberg for their help in data collection and Lisa Korenman for her help in statistical analyses.
References
Altarriba, J., & Mathis, K. M. (1997). Conceptual and lexical development in second language acquisition. Journal of Memory and Language, 36, 550-568. Blair, D., & Harris, R. J. (1981). A test ofinterlingual interaction in comprehension by bilinguals. Journal of Psycholinguistic Research, 10, 457-467. Blom, J. P., & Gumperz, J. J. (1972). Social meaning in linguistic structures: Codeswitching in Norway. In J. J. Gumperz & D. Hymes (Eds.), Directions in sociolinguistic: The ethnography of communication (pp. 407-434). New York: Holt, Rinehart & Winston. Clyne, M. G. (1967). Transference and trigerring. The Hague: Martinus Nijhoff. Fishman, J. A. (1965). Who speaks what language to whom and when? La Linguistique, 2, 67-88. Gawlitzek-Maiwald, I., & Tracy, R. (1996). Bilingual bootstrapping. Linguistics, 34, 901-926. Genesee, F. (1989). Early bilingual development: One language or two? Journal of Child Language, 16, 161-179. Heredia, R. R., & Altarriba, J. (2001). Bilingual language mixing: Why do bilinguals code-switch? Current Directions in Psychological Science, 10, 164-168. Kolers, P. A. (1966). Reading and talking bilingually. American Journal of Psychology, 79, 357-376. Kolers, P. A. (1974). Remembering trivia. Language and Speech, 17, 324-336. Kroll, J. F., & Stewart, E. (1994). Concept interference in translation and picture naming: Evidence for asymmetric connections between bilingual memory representations. Journal of Memory and Language, 33, 149-174. Lipski, J. M. (1978). Code-switching and the problem of bilingual competence. In M. Paradis (Ed.), Aspects of bilingualism (pp. 250-264). Columbia, SC: Hornbeam Press. Macnamara, J. (1967). The bilingual's linguistic performance--a psychological overview. Journal of Social Issues, 23, 58-77. Martin-Jones, M. (1995). Code-switching in the classroom: Two decades of research. In L. Milroy & P. Muysken (Eds.), One speaker, two languages (pp. 90-112). Cambridge: Cambridge University Press. McClure, E. (1981). Formal and functional aspects of the codeswitched discourse
356
Zehra F. Peynircioglu and Aydm Y. Durgunoglu
of bilingual children. In R. P. Duran (Ed.), Latino language and communicative behavior (pp. 69-94). Norwood: Ablex. Meisel, J. M. (1989). Early differentiation of languages in bilingual children. In K. Hyltenstam & L. Obler (Eds.), Bilingualism across the lifespan: Aspects of acquisition, maturity, and loss (pp. 13-40). Cambridge: Cambridge University Press. Meuter, R. F. I., & Allport, A. (1999). Bilingual language switching in naming: Asymmetrical costs of language selection. Journal of Memory and Language, 40, 2540. Muysken, P., Kook, H., & Vedder, P. (1996). Papiamento/Dutch code-switching in bilingual parent-child reading. Applied Psycholinguistics, 17, 485-505. Nicoladis, E., & Genesee, F. (1997). Language development in preschool bilingual children. Journal of Speech-Language Pathology and Audiology, 21,258-270. Nortier, J. (1990). Moroccan Arabic-Dutch code-switching. Dordrecht: Foris. Pandharipande, R. (1990). Formal and functional constraints on code-mixing. In R. Jacobsen (Ed.), Code-switching as a worldwide phenomenon (pp. 15-31). New York: Peter Lang. Pfaff, C. W. (1979). Constraints on language mixing: Intrasentential code-switching and borrowing in Spanish/English. Language, 55, 291-318. Poplack, S. (1980). Sometimes I'll start a sentence in Spanish Y TERMINO EN ESPAlqOL: Toward a typology of code-switching. Linguistics, 18, 581-618. Redlinger, W., & Park, T .Z. (1980). Language mixing in young bilinguals. Journal of Child Language, 7, 337-352. Scotton, C. M. (1988). Self-enhancing codeswitching as international power. Language & Communication, 8, 199-211. Sholl, A., Sankaranarayanan, A., & Kroll, J. F. (1995). Transfer between picture naming and translation: A test of asymmetries in bilingual memory. Psychological Science, 6, 45-49. Sridhar, S. N., & Sridhar, K. K. (1980). The syntax and psycholinguistics of bilingual code mixing. Studies in the Linguistic Sciences, 1O, 203-215. Verhoeven, L. T., & Boeschoten, H. E. (1986). First language acquisition in a second language submersion environment. Applied Psycholinguistics, 7, 241-255. Weinreich, U. (1953). Languages in contact. The Hague: Mouton. Yoon, K. K. (1993). New perspectives on intrasentential code-switching: A study of Korean-English switching. Applied Psycholinguistics, 13, 433-449.
Code-Switching in Bilingual Children
35 7
Appendix The Three Pictures: The originals were 12 x 18 in. and in color.
9 9............ 9 ~ '~ii ~ ...:. ii..........::::::99 li 9149 ~ 9176149 .......'L~I~YlI~:~,,~I: f~iii~: i, ~ ! ~ i ii~~ ~.... .........~ii~
~~'~.~~'~ 9~
~. . . .
84149149
~!i/!'~....
~~i~
~:~!~i~iiii~i~i~i~!iiii~!~i~i~L~i~i~iiiiiii!i~iiii i
~
~ .....~%
. . . . . .
. ~~"~~!iiiiiiiii'i'iii~i~i'i~i~i'............... i'~i~i~i~i'i'i'~......... '"~'~:'9~ :TIIIII '-i ~
Z~L LIZ ~~....?":'~,; :%~,: . . . . . . . !~i~'~i'i~~ii~ :i:' ,,:~~'.....
~
~iili 9
9
~::
9.... ~
' ? ~!~
~
~i~i;~9149 ~i ~: 9 84149149149i!!
i~'~'~~i ~~f~:::!~:~i~!i~
:!~i. . . . . .
~ ~
:
358
Zehra F. Peynircio~lu and Aydm Y. Durguno~lu
!ii;:i~iii::.... i...........~:.:.:.:,.,.,
.....
.....
iii
. . . . . . . . .
t:~~' / ! ! i : : ~
..... ...S!~!i!!::~,~ii,~,i~i~':::::. . . . . . . . . . . . ~......i.!.~i~......... i [~ii .....".:~':~:%i!iii,~"~ " ~ ~"~.-:,:'~. .:.'.!~!ii~.... ~!i::~:; "i:~,i::ii i ::~:!!::i:-~i~:~,,~,,~,,,~ .................. ~: .....
:: [~..~:,:),-~
~
..... ~iii~,
...~" :~ ........
..~,:~
......... !-ii:~.....
.''.i::~::::~:.... [~.
~~;~
The Three Stories 1. Three children went to the zoo with their teacher. There they saw monkeys, bears, fish, and birds. They liked looking at the animals very much. The children fed the fish and the birds. When they got tired, they sat down and their teacher gave them apple juice and cookies. Then they went back to school. 2. One day it was very hot outside, so Mary asked her sister if they could go to the beach. Her sister said yes, so Mary put on her bathing suit and sunscreen. Then they went to the beach. Mary swam in the waves until she was tired. Then she ate ice cream with her sister. After she finished eating, Mary played in the sand until it was time to go home. 3. John didn't feel well, so his mom took him to see the doctor. The doctor sat him on the table and looked at his throat, but he was not sick. The doctor told John to go home and take a nap because what happened was that he had a cold and he was very tired. John's mom took him home and gave him some orange juice. Then John ate a bowl of soup and took a nap. When he woke up, he felt great!
Authors
Lise Abrams is an Assistant Professor in the Department of Psychology, University of Florida, Gainesville, FL 32611. Email may be sent to [email protected]. Jeanette Altarriba is an Associate Professor of Psychology, Linguistics and Cognitive Science, and Cognitive Area Director in the Department of Psychology, University at Albany, State University of New York, Albany, NY 12222. Email may be sent to [email protected]. Heather Bortfeld is an Assistant Professor (Research) in the Department of Cognitive and Linguistic Sciences, Brown University, Providence, RI 02912. Email may be sent to [email protected]. Aydin Y. Durguno~;lu is a Professor in the Department of Psychology, University of Minnesota, Duluth, MN 55812. Email may be sent to [email protected]. Igor Farkas is a Postdoctoral Fellow supported by a grant from the National Science Foundation in the Department of Psychology, University of Richmond, Richmond, VA 23173. Igor is also a Research Fellow at the Institute of Measurement Science, Slovak Academy of Sciences, Bratislava, Slovakia. Email may be sent to [email protected]. Eva M. Fermindez is an Assistant Professor in the Department of Linguistics and Communication Disorders, Queens College, City University of New York, Flushing, NY 11367. Email may be sent to [email protected]. Wendy S. Francis is an Assistant Professor in the Department of Psychology, University of Texas, E1 Paso, TX 79968. Email may be sent to [email protected]. Cheryl Frenck-Mestre is a Chargee de Recherche au Centre National de Recherche Scientifique (CNRS) - Laboratoire Parole et Langage, Universit6 de Provence, 13621 Aix-en-Provence, Cedex 02, France. Email may be sent to [email protected]. Rochel Gelman is a Professor of Psychology and Cognitive Science at the Rutgers Center for Cognitive Science, Rutgers University-New Brunswick, Piscataway, NJ 08854. Email may be sent to [email protected]. Jennifer L. Gianico is a Graduate Student in the Department of Psychology, University at Albany, State University of New York, Albany, NY 12222. Jennifer is currently working with Jeanette Altarriba.
360
Authors
Roberto R. Heredia is an Assistant Professor in the Department of Psychology and Sociology, Texas A&M International University, Laredo, TX 78041. Email may be sent to [email protected]. Arturo E. Hern~indez is an Assistant Professor in the Department of Psychology, University of California, Santa Barbara, CA 93106. Email may be sent to hernande @psych. ucsb. edu. Dieter Hillert is an Adjunct Professor in the Biolinguistic Research Laboratory, Department of Psychology, University of California at SanDiego, La Jolla, CA 92093, and a Computational Linguist in the Lernout & Hauspie Speech Products R & D Department, San Diego, CA 92123. Email may be sent to [email protected]. Lori E. James is an Assistant Professor in the Department of Psychology, University of Colorado, Colorado Springs, CO 80919. Email may be sent to [email protected]. Ping Li is an Associate Professor in the Department of Psychology, University of Richmond, Richmond, VA 23173. Email may be sent to [email protected]. Don MacKay is a Professor in the Department of Psychology, University of California, Los Angeles, CA 90095. Email may be sent to [email protected]. Brian MacWhinney is a Professor of Psychology in the Department of Psychology, Carnegie Mellon University, Pittsburgh, PA 15213. Email may be sent to [email protected]. Teenie G. Matlock is a Postdoctoral Researcher in the Psychology Department, Stanford University, Stanford, CA 94305. Email may be sent to [email protected]. Montserrat Mir is an Assistant Professor in the Department of Foreign Languages, Illinois State University, Normal, IL 61790. Email may be sent to mmirl @ilstu.edu. Timothy K. Miura is a Graduate Student in the Department of Psychology, University of Illinois, Chicago, IL 60607. Timothy is currently working with Gary E. Raney. Sharon M. Obeidallah is a Graduate Student in the Department of Psychology, University of Illinois, Chicago, IL 60607. Sharon is currently working with Gary E. Raney. Zehra F. Peynircio~lu is an Associate Professor in the Department of Psychology, American University, Washington, DC 20016. Email may be sent to [email protected]. Gary E. Raney is an Associate Professor in the Department of Psychology, University of Illinois, Chicago, IL 60607. Email may be sent to [email protected].
Authors
361
Laura F. Romo is an Assistant Professor in the Department of Child Development and the Department of Chicano Studies, California State University, Northridge, CA 91330. Email may be sent to [email protected]. Greg B. Simpson is a Professor and Chair of the Department of Psychology, University of Kansas, Lawrence, KS, 66045. Email may be sent to [email protected]. Mark T . Stewart is an Assistant Professor in the Department of Psychology, Willamette University, Salem, OR 97301. Email may be sent to [email protected].
This Page Intentionally Left Blank
Author Index
A Aaronson, D. 171, 179 Abrams, L. ix, 89-91, 102-104, 359 Adams, M. J. 302, 312 Albert, M. 75, 85 Alderson, J. C. 169-170, 173-174, 179 Allport, A. 341-342, 353-354, 356 Alonso, M. A. 244-245, 247 Altarriba, J. v, ix, 8, 10, 12-14, 21, 23, 25, 75-76, 81, 83, 89-93, 95, 97-98, 100-101,103-104, 111, 115-116, 121, 128-130, 131-132, 139-141,143, 157, 160, 167-169, 174-175, 179, 210, 213, 257, 269, 341,353,355, 359-360 Altenberg, E. P. 191, 211 Altmann, E. 82 Altmann, G. 222, 232 Amlund, J. T. 176, 182 Amrhein, P. C. 14, 23 Anderson, J. 33, 36-37, 52, 55-56 Anderson, R. 235, 299, 312 Anes, M. D. 7, 19, 22, 24, 266, 269 Ariel, M. 237, 247 " Arifio-Marti, S. 180, 311 Asher, J. 264, 268 Aslin, R, N. 33, 41, 52, 54 Atilano, R. 169, 182 Atkins, B. 251,270 Atkinson, R. 43, 52 Au, T. 288, 294 Auer, P. 178-179 Austin, J. L. 237, 247 /kvila, L. X. 8, 25, 139, 141,161-162
Baccino, T. 226, 232 Bach, E. 239, 247 Baddeley, A. D. 101, 104 Balota, D. A. 117, 119, 131,139, 143, 160, 163,214 Barry, S. 171,179 Bartolome, L. 182 Bates, E. 7-8, 10, 18-19, 24-26, 31, 3637, 52, 54-55, 57, 80, 82, 139, 141, 145, 161-163
Bavelier, D. 98, 104, 128 Beauvillain, C. 16, 24, 123, 127, 129, 131,139, 168, 179 Becker, C. A. 122, 134 Beeson, P. M. 319, 337 Bell, L. 233 Bell, S. 256, 268 Bentin, S. 218, 234 Berge, C. A. H. 10, 27 Berger, S. 122, 132 Berlin, B. 43, 52 Bernhardt, E. B. 169-170, 173-174, 179180 Bertera, J, H. 118, 133 Besner, D. 162 Betancort, M. 195, 211 Bever, T. 63, 83 Bhatia, T. K. 189, 215 Bialystok, E. 180, 214 Bickerton, D. 81-82 Bienkowski, M. 123, 134, 242, 248 Binder, K. 126, 131 Birnbaum, M. H. 259, 268 B irren, J. E. 104 Blackwell, A. 37, 52, 55-56 Blair, D. 343,354 Blank, M. A. 11, 19, 24 Blasko, D. G. 19, 24, 275,294 Blom, J. P. 340, 355 Bloom, P. A. 112, 132, 294 Blumentritt, T. L. 19, 25 Bly, B. 275, 293,295 Boatner, M. 276, 280, 294-295 Bobrow, S. 256, 268 Bock, J. K. 319-320, 336 Boehm-Jernigan, H. 223,233 Boeschoten, H. E. 341,343,356 Boland, J. E. 223,233 Bolinger, D. 258, 267-268 Boiler, F. 85 Borning, L. 141, 150, 162 Bortfeld, H. x, 275,293,294, 359 Bouhuis, D. G. 105 Bouma, H. 105 Bourassa, D. C. 172, 180 Bowen, J. 43, 57 Bower, G. 336
364
Author Index
Bowerman, M. 64, 82 Bowman, R. W. 91, 104 Bradley, D. 194, 214-215 Bresnan, H. J. 53,243,248 Briihl, D. 118, 131 Brisard, F. 267 Brislin, R. W. 56 Brones, I. 168, 180 Brown, C. M. 165, 168-170, 171-172, 173-174, 180, 183 Brown, M. 275-276, 295 Brysbaert, M. 194, 211, 214, 226, 232, 235 Bubka, A. 122, 132 Burgess, C. 65, 82, 112, 123, 128, 134, 138, 163 Burke, D, M. 102, 104 Bybee, J. 34, 53
Cacciari, C. 19, 24, 256, 266, 268, 277, 294 Cairns, H. S. 191, 211 Cancini, H. 300, 312 Canseco-Gonz~.lez, E. 127, 132, 192, 213,225,234 Caplan, D. 22, 28, 102, 104 Caramazza, A. 105, 128, 131, 168, 175, 180 Carey, S. 43, 53 Carlo, M. S. 128, 130 Carlson, M. 221-222, 235 Carpenter, P. 37, 56, 77, 83, 126, 133 Carreiras, M. 194-195, 211, 213,225226, 233-234, 244-245,247 Carroll, S. 39, 53 Casey, A. 172, 180 Casteel, M. A. 112, 134, 138, 163 Chamot, A. 31, 32, 56 Chang, F. R. 171-172, 180 Chang-Rodriguez, E. 145, 162 Chen, H. C. 26, 174, 180 Chincotta, D. 320, 336 Chomsky, N. 31, 53,237-238, 247 Chrysler, S. T. 167, 174, 176-177, 181 Cienki, A. 247 Clahsen, H. 232, 235 Clark, H. H. 252, 268 Cleeremans, A. 82 Clifton, C. 127, 132, 135, 192, 194-196, 211-213, 215, 218, 225-226, 231,233-
234, 236, 243-244, 247 Clyne, M. G. 341,355 Cohen, J. D. 145, 160, 263,268 Colombo, L. 123, 135 Colonna, S. 209, 215, 225,235 Coltheart, M. 234-236, 247 Connine, C. M. 11, 13, 19, 24, 28, 256, 266, 270, 275, 294 Cook, V. 191, 211 Cooper, R. P. 41, 56 Cooreman, A. 36, 54 Corder, S. P. 31, 53 Corley, M. 194-195,212, 214, 217, 225226, 233 Cornell, A. 252, 254, 268 Corrigan, R. 55 Cortese, C. 15, 27, 127, 134 Costa, A. 128, 131, 175, 180 Cottrell, G. W. 27 Cregut, I. 22, 25 Cristoffanini, P. 121, 131 Cuetos, F. 126-127, 131,133, 193-195, 212, 214, 217, 225-226, 230, 233 Cummings, J. 170, 180 Cutler, A. 198, 212, 255-256, 270 Cutting, C. 254, 269 Cuyckens, H. 267
Dagut, M. 257-258, 268 Damasio, A. R. 38, 53 Damasio, H. 38, 53 Dannenburg, L. 7, 24, 140, 161 Davidian, R. 41, 53 Davies, D. R. 337 Davis, C. E. 299, 312 Davis, J. N. 177, 180 Davis, R. L. 299, 312 Day, P. 159, 162 de Almeida, R. G. 194, 215 De Avila, E. A. 305, 312 de Bot, K. 36, 53, 189, 212 de Gelder, B. 16, 26, 139, 143, 162 de Groot, A. M. B. 7, 24, 74, 80, 82-83, 85, 121,128, 131,140, 161, 165-169, 174-175, 180, 182, 188, 211-214 De Houwer, A. 45, 53 De Vincenzi, M. 192, 196, 212, 225-226, 232-234 Dechert, H. 56 DeCurtis, L. 300-301,312
Author Index
Deevy, P. L. 195, 212 del Castillo Pintado, J. 10, 27 Dell, G. S. 159, 162 Delmaar, P. 128, 131 Derry, S. J. 82 Devescovi, A. 8, 10, 24 Dickinson, D. 300, 312 Dietrich, E. 82 Dijkstra, T. 38, 45, 53, 57, 60-61, 78, 82, 123, 128, 131 Dinneen, D. 247 Dirven, R. 269 Dowin, S. 172, 180 Duarte, A. 7, 24 Dubois-Charlier, F. 93, 104 Duffy, S. A. 112-113, 132 Dufon, P. 177, 181 Dufour, R. 189, 212 Dumas, G. 31, 56 Duncan, S. 305, 312 Duran, R. P. 356 Durguno~,lu, A. Y. x, 7-8, 24, 168-169, 171, 173, 175-176, 180, 232-233,299, 311-312, 359 Dussias, P. E. 188-189, 193,212 Dwivedi, V. D. 218-220, 224, 232, 234 Dyer, J. R. 91, 103
Eben-Ezra, S. 139, 163 Edelman, G. 38, 53 Ehrlich, K. 193, 212 Ehrlich, S. F. 113, 117, 133-134 Ellis, R. 39, 53, 54, 56, 188, 212 Ellison, G. 176, 182 Elman, J. 35, 47, 53, 59-62, 80, 82, 84 Emerson, W. A. 112, 135 Epstein, S. 31, 53 Ericsson, K. A. 32, 53 Espinal, I. 232-233 Estes, W. K. 318, 336 Everaert, M. 269 Extra, G. 264, 268-270
Fabbro, F. 80, 82 Farkas, I. ix, 59, 62, 63, 66-67, 69-70, 73, 82, 359 Farley, A. M. 248 Farley, P. T. 248
365
Faulkner, H. J. 172, 180 Favreau, M. 8, 24, 218 Feeman, D. J. 158, 163 Feldman, L. B. 166-167, 182 Fennema-Notestine, C. 13, 25, 138, 142, 161 Fern~indez, E. M. x, 187, 189, 192-196, 198-200, 207, 210-212, 215, 359 Fern~indez, F. 218, 225-226, 230, 233 Ferrand, L. 218, 233 Ferreira, F. 7-8, 19, 22, 24, 223,233, 243-244, 247, 266, 269, 324, 333,336 Ferres, S. 171, 179 Fias, W. 191,214 Fifer, W. P. 41, 56 Fillmore, C. J. 34, 54 Findlay, J. M. 234 Fischler, I. 112, 132 Fishman, J. A. 340, 355 Fitzgerald, J. 170-171, 174, 181 Flatt, M. 145, 160, 263,268 Flege, J. 41-42, 54 Fletcher, C. R. 166-169, 174, 176-177, 181 Fletcher, P. 53, 55-57 Flores d'Arcais, G. B. 160, 163, 214 Flynn, S. 31., 53,232-233,238, 247 Fodor, J. A. 241,247 Fodor, J. D. 192, 193, 196, 200, 211-215, 225, 233, 241,245,247 Ford, M. 243, 248 Forster, K. I. 139, 161 Foss, D. 11, 24, 112, 135, 139, 161 Francis, N. 169, 171, 181 Francis, W. N. 94, 98, 104, 145, 162 Francis, W. S. x, 189, 213,317, 321, 336,359 Frauenfelder, U. H. 8, 25 Frazier, L. 125, 127, 132-133, 135, 192, 194-196, 213, 215, 221-222, 225-226, 231,234-235 Freebody, P. 299, 312 French, R. M. 60-63, 70, 82 Frenck-Mestre, C. x, 188, 191, 194, 213, 217-219, 221-226, 228, 232, 234, 359 Frisson, S. 267 Fritz, C. O. 321,336 Frost, R. 180, 218, 234 G Garcia, G. E. 171,181
366
Author Index
Garcia, R. 264, 268 Garcia-Albea, J. E. 212, 233-234 Garcia-Ramos, R. G. 336 Gardner-Chloros, P. 178, 181 Garrett, M. F. 320, 336 Gass, S. 36, 53, 171,181 Gates, J. 276, 280, 294-295 Gathercole, S. E. 101,104 Gawlitzek-Maiwald, I. 355 Gazdar, G. J. M. 237, 247 Gelman, R. x, 317, 321-322, 324, 328, 332, 336-337, 359 Genesee, F. 264, 270, 342, 355-356 Gentner, D. 39, 53 Gerard, L. 15, 27, 123, 127, 132, 134 Gernsbacher, M. A. 24, 82, 163, 167, 172, 181,214, 235, 239, 247, 269 Gerver, D. 38, 53 Gianico, J. L. ix, 11 l, 360 Gibbs, R. W. 252, 254, 256, 265-267, 269, 275-280, 284, 294-295 Gibson, E. 127, 132, 192, 213,225-226, 234 Gil, M. 127, 135 Gilboy, E. 127, 132, 192, 194, 196, 213, 225, 234 Giv6n, T. 55 Glanzer, M. 7, 24 Gleitman. L. R. 82, 84, 247 Glucksberg, S. 19, 25, 275-278, 294-295 Gomez, L. 169, 182 Gonz~ilez, P. 300, 312 Goodman, K. S. 170, 181 Gordon, B. 80, 85 Gorfein, D. S. 122, 132 Grabowski, J. 320, 337 Graesser, A. C. 165-167, 171,173-174, 176, 181 Grafman, J. 85 Grainger, J. 16, 24, 38, 45, 53, 57, 60, 82, 123, 127-128, 131,139, 168, 179, 218, 233 Granott, N. 337 Gray, W. 82 Green, C. D. 75, 78-79, 82 Green, D. W. 189, 191,213 Greth, D. 188, 214 Griffin, Z. M. 122, 132 Griffith, T. 112, 133 Grondelaers, S. 194, 214, 226, 235 Grosjean, F. 7, 8, 1l- 12, 15-18, 24-27, 45, 54, 60-61, 75, 78, 81-83, 127, 132,
157, 161, 168, 175, 181, 189, 198, 213 Grossberg, S. 37-38, 54 Grosz, B. 248 Guillelmon, D. 8, 11, 25 Gumperz, J. J. 177-178, 181,340, 355 Gupta, P. 37, 54 Gutierrez-Clellen, V. F. 300-301, 312 H
Hahn, M. 84 Hakuta, K. 189, 214 Halle, M. 53 Hamblin, J. L. 266, 269 Hancin, B. 41, 54, 169, 171,175-176, 180, 232-233, 311-312 Hancin-Bhatt, B. 41, 54, 169, 171, 175176, 180, 311-312 Hare, M. L. 159, 162 Harrington, M. 36, 54, 219, 221,224, 235, 318, 320, 337 Harris, R. J. 1, 3, 23, 26-27, 53-54, 131, 160-162, 179, 233,235,312, 337, 343, 354-355 Hart, J. 80, 85 Hatim, B. 178, 181 Hauser, M. 41, 54 Hebb, D. 64, 83 H6bert, M. 218-219, 236 Heilenman, L. K. 8, 10, 27, 37, 55 Heller, D. 214-215, 234-235 Hemforth, B. 194, 215 Henderson, J. M. 7-8, 22, 24, 112, 132, 223,233 Henstra, J. -A. 195, 213 Heredia, R. R. v, ix, x, 7-8, 10, 12-13, 1923, 25, 27, 76, 83, 115, 124, 127, 130, 132, 139-140, 143, 144, 156, 160-161, 210, 213,251-252, 256- 257, 266, 269, 341,355, 360 Hernb.ndez, A. E. ix, 8-10, 13-16, 19-20, 25, 137-139, 141-143, 145, 148, 150, 156-158, 161-162, 360 Herron, C. 39, 57 Herron, D. 18, 25 Hess, T. 104 Heycock, C. 213,234 Hickok, G. 127, 132, 192, 213,225,234 Hillert, D. x, 84, 214-215, 237, 239, 242243,245, 247-248, 360 Hinton, G. E. 63, 83, 138, 159, 161-162 Hirshkowitz, M. 19, 27
Author Index
Hoefnagel-Hohle, M. 50, 57 Hoeks, J. C. J. 140, 161 Hofstfidter, D. 39, 54 Hogaboam, T. W. 123, 132 Holland, A. L. 319, 337 Holmes, V. M. 123, 132, 221,235 Hong, F. G. 177, 181 Hoover, M. L. 218-220, 224, 232, 234 Hopper, P. 34, 53 Horiba, Y. 166, 169, 171,173-174, 181 Horine, M. D. 22, 24 Huggins, A. W. F. 302, 312 Hulstijn, J. H. 257, 269 Humphreys, G. W. 162 Hyltenstam, K. 270, 356 Hymes, D. 355 HyOnfi, J. 320, 336
Igoa, J. M. 194, 213 Ijaz, H. 42, 54 Inhoff, A. W. 118, 120, 131-132 Irujo, S. 258, 261,269 Ito, T. 37, 54
Jackendoff, R. 255, 266, 269 Jacobs, M. 82, 177, 181 Jacobsen, R. 356 Jacobson, A. 248 Jacobson, M. G. 299, 312 Jain, M. 72, 83 James, L. ix, 89, 360 Jim6nez, R. T. 171, 173, 181 Job, R. 225-226, 232-233 Johnson, J. 49, 192, 196, 212, 256, 269, 303, 312 Johnson, M. H. 80, 82 Johnson, P. O. 177, 181 Joshi, A. 84 Juffs, A. 219, 221,224, 235 Juilland, A. 145, 162 Just, M. 37, 40, 56, 77, 83, 126, 133, 319, 333,337 K
Kahneman, D. 318, 337 Kambe, G. 111,130 Kamil, M. L. 169-170, 173-174, 180
367
Kanwisher, N. G. 97, 104 Kaplan, R. 248 Kardash, C. A. 176, 182 Karmiloff-Smith, A. 80, 82 Kassler, M. A. 166, 181 Katz, L. 180, 218, 234 Kaup, B. 112, 135 Kawamoto, A. H. 59, 84, 159, 162 Kay, P. 34, 43, 52, 54 Keatley, C. 16, 26, 139, 143, 162 Kellas, G. 112, 122, 132, 135 Kellerman, E. 258, 269 Kellerman, K. 57 Kello, C. 221,236 Kellog, R. T. 320, 337 Kemper, S. 318, 320-321,337 Kennedy, A. 214-215, 234-235 Kentridge, R. W. 234 Keysar, B. 275-276, 293,295 Kilborn, K. 8, 10, 26, 36-37, 54 Kimball, J. 192, 213 King, J. 319, 333,337 King, K. T. 177, 182 King, M-L. 72, 83 Kintsch, W. 165-168, 171,174, 176-177, 182-183,239, 247-248 Kirsner, K. 72, 74, 83, 121, 131 Kittay, E. 277, 295 Klein, E. C. 212, 215 Klemke, E. D. 248 Kletzien, S. B. 173, 182 Koda, K. 168, 172, 175, 182, 232, 235 Kohnert, K. J. 145, 161-162 Kohonen, Y. 62-64, 68, 70, 83 Kolers, P. 12, 26, 343,353,354-355 Konieczny, L. 194, 215 Kook, H. 341,356 Krashen, S. 39, 54 Kreuz, R. J. 19, 25, 166, 181 Kroll, J. F. 7-8, 20, 23-24, 26, 42, 54, 75, 80-81, 83, 85, 115, 128, 130, 139-141, 144, 150, 157, 160, 162, 165-169, 175, 179-180, 182-183, 188, 211-214, 256, 270, 339, 353-356 Krueger, M. A. 126, 134 Ku(,era, H. 94, 98, 104, 145, 162 Kulhavy, R. W. 176, 182 Kurland, B. E. 300, 312 Kushnir, S. L. 7, 12, 26 Kutas, M. 125, 135, 139, 163 Kynette, D. 318, 337
368
Author Index
Labarca, A. 179 Lachter, J. 63, 83 LaCount, K. L. 112, 114, 116, 134 Ladefoged, P. 68, 83 Langer, J. A. 169, 171, 182 Lantolf, J. 179 Lasnik, H. 161 Laufer, B. 257-258, 268 Lazarte, A. A. 171, 179 Lebiere, C. 37, 52 Lederberg, A. R. 15, 26 Leiman, J. M. 123, 125, 134-135, 242, 248 Leinbach, J. 35, 55, 67, 84 Lesch, M. 120, 133 Levelt, W. J. M. 80, 83, 319-320, 337 Levy, B. A. 172, 180 Levy, C. M. 337 L6wy, N. 15, 27, 29, 60-61, 81, 83 Li, P. ix, 8-10, 12-13, 15-19, 21, 26, 37, 54, 59, 61-63, 66-67, 69-70, 72-73, 75, 78-82, 84, 123, 132, 360 Liberman, M. 247 Lima, S. 55 Lindner, S. 254, 270 Lindsay, P. 38, 54 Lipski, J. M. 340, 355 Liu, H. 7-8, 10, 18-19, 24, 26, 37, 54 Locke, J. L. 41, 54 Lockhart, R. S. 72, 83 Lombardo, V. 232, 234 Lorch, R. F. 143, 160 Love, T. 19, 26 Lovri~, N. 194, 211,214 Lucas, M. M. 125, 132, 239, 248 Lucas, T. 169, 182 Lund, K. 65, 82 Lupker, S. J. 128, 131 Lust, B. 247-248 Lyons, J. 80, 84 M
MacDonald, M. C. 34, 54, 127, 132, 135, 239,241,248 Macias, R. F. 336 MacKay, D. G. ix, 89-92, 97, 99-100, 102-105, 360 Macken, M. 248 Macnamara, J. 7, 12, 26, 75, 84, 341,355
MacWhinney, B. ix, 8, 26, 31, 34-37, 39, 41, 43, 46, 49, 52, 54-57, 62, 67, 69, 81, 84, 145, 160, 232, 235,263,268, 360 Makkai, A. 276, 280, 294-295 Malakoff, M. 189, 214 Malle, B. 55 Mann, V. 42, 45, 51, 53 Marchena, E. 257, 269 Markman, A. 39, 53, 82 Marslen-Wilson, W. D. 16, 27, 139, 163 Martin, C. 122, 132 Martin, J. 43, 57 Martin, N. 159, 162 Martin-Jones, M. 341,355 Martohardjono, G. 31, 53, 212, 215, 238,247 Masson, M. J'. 138, 159, 162 Matessa, M. 36, 55 Mathis, K. M. 140, 160, 167-168, 179, 353,355 Matlock, T. x, 21-23, 127, 251-252, 26"/-270, 360 Matthews, S. 69, 85 McArthur, T. 251,270 McClelland, J. L. 38, 55, 59-61, 80, 8485, 117, 119-120, 133, 138, 161,176, 182, 223,236 McClure, E. 342, 355 McConkie, G. W. 117, 133 McCullough, K. E. 248 McDonald, J. E. 218, 235 McDonald, J. L. 8, 10, 27, 34-37, 47-49, 55 McElree, B. 112, 133,278, 295 McFarlane, D. K. 7, 24 McGlone, M. 275-276, 293-295 McKay, T. 112, 135 McKoon, G. 19, 27, 138, 163 Mclain-Allen, B. 166, 181 McLaughlin, B. 139, 161,252, 264, 270 McNamara, T. P. 138, 162 Meck, B. 331,336 Meck, G. H. 321-322, 336-337 Mehler, J. 198, 212 Meisel, J. 232, 235, 342, 356 Mendelsohn, A. 194, 214 Menn, L. 41, 56 Merriman, W. 36, 56 Messeguer, E. 194-195, 211, 213 Metcalf, K. 122, 132 Meuter, R. F. I. 341-342, 353-356
Author Index
Meyer, D. 122, 134 Miikkulainen, R. 64-65, 79-81, 85 Milech, D. 121, 131 Miller, G. 53 Miller, J. 78, 83 Miller, M. 89-90, 99, 100, 102-105 Millis, A. C. 165, 181 Milroy, L. 179, 181,355 Minkoff, S. 169, 183 Miozzo, M. 128, 131 Mir, M. x, 69, 180, 299, 360 Mitchell, D. C. 126, 127, 131,133, 192, 194-195, 211-212, 214, 217, 221,225226, 230, 232-233,235 Miura, Y. K. ix, 165, 360 Miyake, A. 37, 56, 126, 133 Monsell, S. 102, 105 Moon, C. 41, 56 Morales, C. 15, 26 Morris, R. K. 112-113, 120, 132-135, 218, 236 Morton, J. 43, 56 Moss, H. E. 139, 159, 162 Murray,. L. L. 319, 337 Muysken, P. 179, 181,232,235,341,355 Myers-Scotton, C. 15, 27, 189, 214
Nagy, W. E. 311-312 Nas, G. L. J. 74, 82, 121, 131,140, 161, I66, 168, 174, 180 Nayak, N. 254, 269, 276, 279, 295 Neely, J. H. 138, 143, 163 Nelson, D. 233 Nelson, E. M. 257, 270 Neuman, O. 104 Neville, H. 232, 236 Newell, A. 32, 34, 56 Newport, E. L. 33, 41, 52, 54, 303, 312 Newsome, S. L. 119-120, 133 Ng, M. L. 174, 180 Nguyen, T. 80, 85 Nicholson, P. A. 300, 312 Nicol, J. 19, 27, 188-189, 212, 214, 242, 248 Nicoladis, E. 342, 356 No~l, M. -P, 191,214 Noel, R. W. 218, 235 Noonan, M. 55 Nordlie, J. 278, 295 Norman, D. A. 38, 54, 56
369
Norris, D. 198, 212 Nortier, J. 341,356 Nosofsky, R. M. 101, 105 Nunberg, G. 276, 295
O'Malley, M. 31, 32, 56 O'Regan, J. K. 117, 119-120, 133 O'Seaghdha, P. G. 138, 163 Obeidallah, S. M. ix, 165, 361 Obler, L. 75, 85, 270, 356 O'Brien, J. 277-280, 284, 295 O'Brien, K. 318, 337 Odlin, T. 232, 235 Oh, C. K. 247 Onifer, W. 19, 27, 123-124, 126, 129, 133 Osherson, D. N. 161
Paap, K. R. 119-120, 133,218, 235 Palermo, H. D. 161 Pandharipande, R. 340, 356 Pandit, R. 8, 10, 28 Paradis, M. 79, 85, 355 Parasuraman, R. 337 Parisi, D. 80, 82 Park, T. Z. 342, 346, 353,356 Parziale, J. 337 Patterson, K. 80, 85 Pavlenko, A. 168-169, 175, 182 Pearlmutter, N. 127, 132, 192, 213,225, 234, 239, 241,248 Pearson, P. D. 171, 181 Pedroza, M. J. 90, 104 Pellegrino, J. W. 180 Perfetti, C. A. 123, 132, 158, 163, 312 Perlmutter, D. M. 254, 270 Perlmutter, N. 194, 214 Peterson, R. R. 112, 134, 138, 163 Peynircio~lu, Z. F. x, 8, 299, 339, 361 Pfaff, C. W. 115, 133,340, 356 Pienemann, M. 232, 235 Pimsleur, P. 181 Plaut, D. C. 80, 85, 159, 163 Pl6h, C. 34, 46, 55 Pletsch de Garcia, K. 15 Plunkett, K. 80, 82 Pollatsek, A. 111, 113, 118, 120, 130, 133, 176, 183,218, 236
370
Author Index
Poplack, S. 339, 341-342, 346, 353,356 Potter, M. 97, 104, 166-167, 182 Poulisse, N. 189, 214 Powell, T. 18, 26 Poyner, D. V. 113, 134 Prather, P. 19, 27 Prentice, K. J. 22, 28 Prieur, B. 239, 248 Prinz, W. 104 Pritchard, D. R. 93, 104 Provost, J. 145, 160, 263,268 Pulvermtiller, F. 39, 56 Ptitz, M. 269 Pynte, J. 188, 191,209, 213-215, 218219, 221-226, 232, 234, 235-236, 239, 248
Quine, W. v. 237, 248 Quinlan, P. 84 Quinn, D. 194, 211, 215 Quinn, T. 181 R
Radach, R. 214-215, 234-235 Radvansky, G. A. 167, 171-172, 174, 183 Raney, G. E. ix, 165, 169, 174, 177, 182183,360,. 361 Ransdell, S. 337 Rash, S. 318, 337 Rashotte, C. A. 172, 183 Ratcliff, R. 19, 27, 138, 163 Rativeau, S. 226, 236 Raupach, M. 56 Rayner, K. 8, 23, 75, 81, 111, 113, 115, 117-120, 125-126, 130-131,133-135, 141,160, 163, 168, 173, 176-177, 179, 183, 214, 218, 221-222, 235-236 Reder, L. 39, 56 Redlinger, W. 342, 346, 353,356 Reinhart, T. 237, 248 Requin, J. 134 Rho, S. H. 19, 25 Rice, S. 269 Rieben, L. 312 Ritchie, W. 189, 215 Robinson, P. 55 Roediger, H. L. 7, 24, 168, 173, 180 Romo, L. x, 317, 321,336-337, 361 Ronna, D. F. 180
Rosano, T. 256, 269 Rosenbloom, P. S. 34, 56 Rumelhart, D. E. 38, 55, 59-60, 79, 8485, 138, 161,182
Saffran, E. M. 102, 105 Saffran, G. S. 159, 162 Saffran, J. R. 33, 52 Sag, I. 276, 295 Samuels, S. J. 172, 183 Sandra, D. 267 Sankaranarayanan, A. 168, 183, 339, 354, 356 Santiago-Rivera, A. L. 115, 131 Scarborough, D. L. 15, 27, 123, 127, 132, 134 Schaie, K. W. 104 Schenk, A. 269 Schmalhofer, F. 166, 182, 239, 248 Schmauder, A. R. 113, 134, 218, 236 Schmidt, R. 39, 56 Schreuder, R. 3, 82, 162-163, 180, 182, 269 Schumann, J. 39, 56 Schunn, C. 82 Schustack, M .W. 113, 134 Schuster, S. P. 102, 105 Schvaneveldt, R. W. 122, 134, 218, 235 Schwanenflugel, P. J. 112, 114, 116, 134 Schwartz, M. F. 159, 162 Scotton, C. M. 340, 356 Searle, J. R. 237, 248 Sebasti~.n-Gall6s, N. 212 Seelig, H. 194, 215 Seely, R. E. 112, 135 Segalowitz, N. 8, 24, 27, 218-219, 236 Segui, J. 198, 212 Seidenberg, M. S. 80, 85, 123, 125, 127, 129, 132, 134-135, 159, 163,239, 241, 248 Sejnowski, T. 63, 83 Seleskovitch, D. 38, 57 Seliger, H. 47, 56 Selinker, L. 31, 53, 56, 171,181 Senerth, D. T. 93, 104 Sera, M. D. 10, 27 Sereno, S. C. 218, 236 Shallice, T. 104, 138, 159, 162, 163 Sharkey, N. E. 138, 163 Sharwood Smith, M. 57
Author Index
Shiffrin, R. M. 101, 105 Shillcock, R. 213,234 Shindler, K. L. 120, 132 Shirai, Y. 80-81, 84 Shoben, E. J. 112, 116, 134 Sholl, A. 8, 20, 23, 26, 42, 54, 75, 81, 115, 130, 139, 141,144, 160, 162, 168, 179, 183,339, 354, 356 Shriberg, E. 300, 312 Simon, H. A. 32, 53, 56, 102, 105 Simpson, G. B. v, viii, ix, 112, 122-124, 126, 128-129, 134, 138, 163,361 Small, S. L. 27, 80, 85 Smith, L. 13.36, 56 Smith, M. C. 72, 74, 83, 85, 133,214 Smith, S. 295 Snow, C. E. 40-4I, 50, 56, 82, 300, 303, 312 So, K. F. 166-167, 182 Soames, S. 254, 270 Soares, C. 12, 25, 27 Sokolov, J. L. 34, 57 Sola, J. M. 93, 104 Soltano, E. G. ix, 89-93, 95, 97-98, 100101,103, 111,121,131 Sopena, J. M. 127, 132, 192, 213,225, 234 Sorace, A. 213,234 Speer, S. R. 139, 161 Sperber, D. 242, 248 Spitzer, M. 64, 85 Sprott, R. 318, 337 Sridhar, S. N. 340, 356 Stahl, S. A. 299, 312 Stanovich, K. E. 112, 134-135, 138, 158, 163, 173, 183,311-312 Starr, M. 120, 132 Steedman, M. 222, 232 Stenshoel, E. 193, 212 Stewart, E. 7, 20, 26, 139, 140, 144, 162, 166-168, 182, 256, 266-270, 339, 353, 355 Stewart, M. T. ix, 7, 19-20, 22, 25, 27, 124, 361 Stockwell, R. 43, 57 Stoel-Gammon, C. 41, 56 Stoness, S. C. 84 Strawson, P. F. 237, 248 Suner, M. 247 Swain, M. 31, 39, 53, 56 Swanepoel, P. 226, 235 Swanepoel, S. 194, 214
371
Sweet, H. 232, 236 Swinney, D. A. 7, 19, 26-27, 123-124, 126, 129, 133-134, 239, 243,247-248, 255-256, 266, 270
Tabors, P. O. 300, 312 Tabossi, P. 19, 24, 28, 129, 135,241, 248, 256, 266, 268, 270 Takagi, J. 42, 53 Tanenhaus, M. K. 27, 123, 125, 127, 134135, 221,236, 241,248 Tang, H. 170-172, 183 Taraban, R. 35, 55, 223,236 Teller, M. 188, 214 Ten Brinke, S. 128, 131 Therriault, D. J. 169, 177, 182-183 Thornton, R. 127, 135 Titone, D. A. 11, 13, 19, 22, 24, 28, 256, 266, 270 Tolman, E. C. 318, 337 Tomasello, M. 39, 57 Torgeson, J. K. 172, 183 Tracy, R. 341,355 Traxler, M. J. 112, 135 Tr6vise, A. 47, 57 Trueswell, J. C. 127, 135, 221,236 Tuggy, D. 269 Tyler, L. K. 16, 28, 159, 162 Tzelgov, J. 139, 163 Tzeng, O. J. L. 26
Udell, C. 13, 25, 138, 142, 161 Underwood, G. 320, 336 Urquhart, A. H. 179
Vaid, J. 1, 3, 8, 10, 25, 27-28, 236, 264, 270,295 Vallar, G. 104 Van den Broek, P. W. 169, 182 Van der Linden, E. 269-270 van Dijk, T. A. 166-168, 171, 174, 182183,239, 248 van Hell, J. G. 7, 24 140, 161 van Heuven, W. 38, 57, 60-61, 78, 82, 85, 123, 128, 131 Van Jaarsveld, H. 128, 131
3 72
Author Index
van Montfort, R. 36, 53 Van Petten, C. 125, 135, 139, 163 Vasquez, O. 169, 182 Vedder, P. 341,356 Verhoeven, L. T. 264, 268-270, 341,343, 356 Vinereanu, M. 193,212 Von Eckardt, B. 166-167, 182 Vu, H. 122, 132 W Walker, R. 234 Walter, M. 194, 215 Wanner, E. 82 Ward, T. 295 Wasow, T. 276, 295 Waters, G. 22, 28, 102, 104 Waterson, N. 82 Weber-Fox, C. M. 232, 236 Weeks, P. A. 7, 24 Weinreich, U. 339, 356 Weisberg, B. 337 Well, A. D. 118, 133 Welsch, D. 166, 182, 239, 248 Weltens, B. 1, 3, 82, 161-162, 180, 182 West, R. F. 112, 134-135, 138, 158, 163, 311-312 Whitman, J. 247 Whitney, P. 112, 135 Wickens, T. D. 318, 329-330, 337 Wiley, J. 177, 183 Williams, J. N. 123, 135 Wilson, D. 242, 248 Wingfield, A. 22, 28 Winitz, H. 52 Wulfeck, B. 8, 10, 18, 24, 26
Yates, J. 123-124, 135 Yip, M .C. 13, 16, 19, 21, 26, 69, 85, 123, 132 Yoon, K. K. 115, 135, 340, 356 Yorio, C. A. 258, 270
Zabeeh, F. 248 Zagar, D. 195, 214, 226, 236 Zhang, G. 102, 106 Zimny, D. 239, 248
Zimny, S. 166, 182 Zipf, G. K. 94, 105 Zobl, H. 232, 236 Zola, D. 117, 133 Zwaan, R. A. 165-174, 181,183
Subject Index
A Acquisition, 160, 198 Activation, 38, 60-62, 64, 67, 70, 74, 78, 137, 159, 176, 191 Adaptive Resonance Theory (ART), 3839 Adjectives, 69 Adverbial clause, 322 Adverbs, 69 Agreement errors, 324-325, 327, 329331,333 Ambiguity, 7, 12, 20-21, 59, 91-99, 102103, 105-106, 116, 122-130, 138, 159, 192, 199, 201,252 American English idioms, 275, 278 Analogies, 39 Analyzability, 291-293 Anaphor, 237-248 Anaphora comprehension, 241-245 Animacy, 8-11, 34, 36, 48 Antecedent, 237-246 Arabic, 35, 45, 50, 178 Associative pathways, 59, 66, 74-75, 7778, 8O Associative priming, 16 Asymmetries, 137, 139-160 Attentional activation, 38 Auditory Moving Window (AMW), 7, 22-23 Automaticity, 329 Automatic priming, 138, 160
Backward anaphora, 239 Balanced bilinguals, 349-350, 352-354 Balto-Slavic, 294 Base-language effect, 15 Between-language, 45, 92-100, 152-154 Between-language priming, 137, 139, 143-144 Bilingual Interactive Activation Model (BIA), 60-61, 70, 78, 80 Bilingual lexical access, 17, 20, 25, 137139, 158-160 Bilingual lexical representation, 168
Bilingual lexicon, 7, 59, 62, 67-68, 72, 75-77, 79-80, 256 Bilingual mode, 13-15, 61 Bilingual Model of Lexical Access (BIMOLA)~ 60-62, 70, 78 Bilingual models of lexical organization, 175 Bilingual "performance deficit," 191 Bilingual reading, 166, 170, 174 Bilingualism and figurative language, 256-257 Bilingualism and phrasal verbs, 256259 Borrowers, 17, 22 Bottom-up activati ,38
Canonical, 9, 288, 292 Cantonese, 35, 40, 51, 68-69, 71-72 Causation, 279, 281-284 Ceiling effects, 94 CHILDES, 69 Children, 40-42, 44, 50, 52, 238, 299300, 305-306, 339-358 Chinese, 10, 16-20, 44-45, 49, 55, 68-75, 80, 171,287 Chinese-English, 10, 16, 18-20, 26, 68, 75, 80, 171 Chinese phonetics, 19 Chromacity, 68 Chunking, 33, 37-38 Clitics, 240 Code-switchers, 16-17, 22 Code switching, 11-23, 40, 72, 339-35~ Cognates, 111, 120-121, 174 Cognitive sub-tasks, 335 Cognitive trade-offs, 329 Competition model, 31-57 Competitive priming, 142-143 Complexity of ideas, 325-326 Computational level, 239 Conceptual constraint, 112-116 Conceptual links, 339 Conceptual metaphors, 275,277 Conceptually mediated, 353 Connectionist models, 31-57, 59-85, 159
374
Subject Index
Connectionist principles, 79 Consonant, 17, 67-68, 71, 81 Content accuracy, 317, 325-327, 329, 331 Context, 13-23, I 11-135, 137-160, 169, 208, 222, 238, 252, 292, 318, 339-343, 347-348, 351-355 Context-dependent model, 122 Context effects, 13-23, 114, 138-164 Controlled priming, 138 Coreference, 238-241 Correlations, 306-308, 328-329, 346, 353 Crossing-coreference sentences, 239 Cross-language ambiguity, 123, 128 Cross-language facilitation, 82, 91, 98, 102 Cross-language masked priming, 121 Cross-language priming, 13-16, 73-75, 77, 79, 139 Cross-language transfer, 311 Cross-linguistic studies, 8, 237 Cross-modal, 13-16, 19-21, 124, 146, 152-153 Cross-Modal Lexical Priming (CMLP), 7, 19-21,124, 242-243,266 Cross-modal naming task, 13-16, 146156 Cue reliability, 34-35 Cue strength, 34, 36, 48 Cued shadowing task, 7, 18-19 Cultural knowledge, 276 Culturally motivated, 277
Decontextualized language, 299-300, 302 Deep structure, 238 Difference times, 22 Digit span, 320 Direct access hypothesis, 256, 265 Distributed representations, 159 Dominance effects, 140, 198, 208 Dominant cue, 35 Dominant language, 10, 13, 21, 140, 175, 187, 189, 208 D-structure, 244 Dual-process, 51 Dutch, 10, 47-49, 60, 80, 128, 194, 257, 267 Dutch-English, 10, 47, 128
Early bilinguals, 137 Educational, 340 Embedded clauses, 319, 324-325, 328333 Emergentist, 31 English, 8-22, 34-36, 42-50, 59-61, 6873, 75-76, 80-81, 92-97, 99-100, 103, 113, 121-123, 127-129, 137, 140-142, 144-150, 152-159, 168, 172, 174-176, 178, 188, 192, 193-199, 202-209, 242246, 251-254, 257-264, 267, 275-278, 280-282, 284-291,299-303,305-309, 311,317, 320-321,323,335, 340, 343345, 355 English as a Second Language (ESL), 197, 257-258, 287, 317, 321-322, 337 English-Chinese, 10 English-dominant, 145, 147, 152, 198, 205-208, 345 English-Dutch, 10, 47 English formal definitions, 305-306 English-French, 10, 127 English idioms, 275-294 English oral proficiency, 305-306 English-Spanish, 14, 107, 121, 141,345 English-speaking, 172, 193,202, 335 English syntax, 303, 306 English word recognition, 302, 306 Episodic memory, 200 Error detection, 33, 39-40, 52 European, 339 Event Related Potential (ERP), 125, 139 Exhaustive access, 122-123 Explicit learning, 39 Expository passages, 302 Eye movements, 113, 115, 117 Eyetracking, 113, 115-117, 123, 125-126
Facilitation effect, 91 Figurative interpretation, 256-267, 275, 292 Figurative language, 19, 256-267, 275, 292 Figurative meaning, 266 Finnish, 45, 258 Finno-Ugric, 258 Fixations, 113, 115-121, 126, 141 Flashbulb memory, 35
Subject Index
Floor effects, 94 Fluency, 21, 37-38, 75-77, 169, 340-342 Formal definitions, 299-316 French, 35, 47, 50, 127, 172, 240, 262 Frequency, 21, 34 G Gating task, 16-18 Gaze durations, 125-126 Gender, 10-11, 18, 35, 37, 45, 49-50 German, 10, 35-36, 38-40, 43-45, 49-51, 80, 240, 245, 262 German-English, 10 Germanic languages, 258 Good learners, 41,293 Grammar, 31, 34, 45-46, 50, 237, 317, 325, 329, 330-331 Grammatical accuracy, 317, 319, 325, 330-333 Grammatical errors, 319, 323-329, 333 Grammatical mistakes, 352-353 Grammatical structure, 8-11
375
Immersion context, 318 Implicit learning, 39 Individual differences, 32, 50, 52, 75, 77 Indo-European, 45, 80, 287, 294 Inhibition, 138 Input strategies, 31, 33, 38, 40-41, 49-50 Integrating, 165 Intentionality, 279, 281-284 Interactive Activation (IA), 38, 60-61 Interactive activation model, 176 Interference, 15, 21, 44-45, 59, 70-71, 73-75, 77-79, 116, 129 Interlexical homographs, 127-128 Interresponse Times (IRT), 22 Intersentential, 339-343,345, 348-352 Intra-lexical priming, 139 Intrasentential, 339-353 Italian, 35, 50, 260, 277
Japanese, 45 K
Hebbian learning, 62, 64-68, 74, 78 Hebrew, 47, 257, 260, 322 Hidden-unit, 61-62 Hierarchical model of bilingualism (see Revised hierarchical model), 167 High attachment, 196 High-constraint, 141 High frequency, 127-128, 141 Hindi, 10, 45 Hispanic, 339 Homograph, 20, 111, 120, 123, 127-129 Homographic noncognates, 120, 128-129 Homophones, 13, 20 Hong Kong, 44, 68-69, 72 Hungarian, 35, 40, 43, 45-46 Hyperspace Analogue to Language (HAL), 65
Idiomatic expression, 255-257 Idiomaticity, 276 Idiom processing, 276 Idioms, 251,257, 275-294 Images, 275-294 Imaging task, 279-280, 291,295
Kintsch and van Dijk's model of text comprehension, 166 Korean, 43, 45, 260, 322, 335
Language asymmetries, 137-162 Language context, 343,355 Language-dependent, 188, 207-208 Language dominance, 10, 13, 76, 196, 198, 204 Language-independent, 168, 170, 189, 204, 207-208, 311 Language separation, 70-72 Language specific-grammar, 238 LAS test, 305 Late closure, 192-193, 196, 204 Latvian, 278-279, 287, 294 Learning strategies, 31-32, 40-42, 52 Less Dominant Language (LDL), 140 Levels of proficiency, 77 Lexical access, 27, 114, 123-125,238239 Lexical access in bilinguals, 139 Lexical access models, 238 Lexical ambiguity, 20, 122-129, 138 Lexical categories, 70-73, 77
376
Subject Index
Lexical constraint, 112-116 Lexical context effects, 138 Lexical decision, 11-14, 19, 114, 123125, 127-128, 141 Lexicalization, 276, 293 Lexical-level processing, 172 Lexical links, 339 Lexical priming, 125, 138, 141,147-152, 153-160 Lexical representation hypothesis, 255 Lexical representations, 167 Lexical-to-lexical links, 168 Lexical tones, 71 Lexicon, 32, 41-55, 72, 255 Linguistic cues 15, 31-57 Linguistic interdependence hypothesis, 170-171 L!nguistic threshold hypothesis, 170, 173 Linguistic variables, 327-328 Literacy variables, 306 Literal interpretation, 255-256, 261,265 Literal language, 275 Literal meaning, 266 Literal processing model, 256 Lithuanian, 45 Long-term retention, 318 Low-constraint, 141 M
Mandarin, 287-290 Mandarin-English, 288 Manner, 68 Mechanisms, 31-33, 35, 38, 41, 44, 52 Memory, 7-8, 20, 33, 35, 41, 50, 89-111, 115, 126, 130 Memory for surface, 166 Mental model construction, 33, 40 Mentally imaging idioms, 277 Merging text comprehension, 168 Message-level, 113-114 Metalinguistic awareness, 299, 301, 311 Mixed-language, 12, 61-62, 69-70, 72, 80, 89, 101, 109, 343-344, 346-347, 354 Models of bilingual language, 165, 178 Models of bilingual text comprehension, 165 Models of word processing, 175 Modular, 45, 311 Monolingual, 111-114, 117, 121-123,
127-129, 187, 189, 191,196-198, 202, 204, 251-268 Mono-syllabic, 67 More Dominant Language (MDL), 140 Morphosyntax, 49-50 Multilingual, 45, 293 Multiple meaning, 266 Multi-store theories, 102 Multi-syllabic, 67
Native English speakers, 282, 284-285, 289 Native speakers, 69, 93,275-294, 321 Nativist, 31, 32 Naturalness ratings, 284-286. Neighborhood, 68, 72 Neural network, 33-35, 39, 49-50, 59, 63, 77 Nodes, 59-62, 70, 77-79 Non-cognate translations, 93 Nonliteral language, 251-268, 275-294 Non-native, 275-294 Nouns, 8, 9, 11, 15, 34-35, 37, 43-45, 4950, 61, 70, 72-74, 81, 114, 188, 193, 195-196, 199, 221-223,225-227, 229, 251,253,259, 263,270 Noun phrase, 220, 222, 225-226, 229
Off-line, 123, 187, 193, 199, 201,204, 207, 209 On-line, 123,137-138, 187, 193, 199, 202-201,204, 209, 251,317, 329-332 On-line performance, 334-335 On-line processing, 266, 332-334 On-line reading task, 259 On-line sentence comprehension, 251 On-line tasks, 7-28 Order of acquisition, 139-140 Oral proficiency, 305, 311 Oral vocabulary, 305 Orthography, 39, 44-45, 50
Parafoveal, 117-121,130 Parafoveal preview, 117-118, 120-121 Parafoveal processing, 117, 121 Parallel Distributed Processing (PDP),
Subject Index
59-60, 79 Parsing models, 238 Pattern detection, 33 Performance, 191 Phoneme, 11-13, 18, 61-62, 67-68, 70 Phoneme monitoring lexical decision task, 11- 13 Phonological, 112, 120, 128 Phonological awareness, 311 Phonology, 32, 34, 41-42, 74, 68 Phrasal verbs, 22-23, 251-268 Phrasal verbs and lexicon, 255-256 Phrase-congruent sentences, 101, 103 Pictures, 346-351, 357 Polysemy, 251,267-268 Portuguese, 12, 260, 262 Portuguese-English, 12 Postlexical effects, 138, 160 Pragmatics, 32, 34, 49-50 Predictive association, 33-36, 38, 50-51 Prepositional phrases, 222-223,325, 331333 Preschool children, 339 Preview effects, 116-122 Priming, 7, 13-16, 19-20, 24, 59, 70, 7375, 77, 79, 120-121,123-125, 127-128, 137-145, 147-160 Processing linguistic input, 191 Production, 198, 316-319, 321-322, 329, 332, 334-335, 339-340, 343, 346, 354 Proficiency, 7, 10, 22, 59, 70, 75-76, 78, 170, 173, 189, 198-199, 299-303,320 Pronouns, 35, 69, 72, 220, 235,237-246, 253,255, 271 Propositional level, 239
Qualitative analysis, 309
Rapid Serial Visual Presentation (RSVP), 89, 91, 93-94, 103-104, 115-116 Reading, 111-135, 167-168, 170-171, 174, 176, 198-199, 201-202, 204 Reading comprehension, 198, 299-316 Reading skills, 170 Reading times, 263-266 Reading universal hypothesis, 170 Real-time comprehension, 239, 241 Recognition, 38, 59-61, 75, 77, 81,303
3 77
307-308 Recurrent networks, 47 Relative attachment, 188, 196 Relative clause, 187-211,225,227-230, 304 Repetition Blindness (RB), 89-110 Representation, 64 Representation and text processing, 165184 Re-revised hierarchical model, 140, 156 Resonance, 33, 38-39, 41-42, 44, 52 Response times, 263 Reversibility, 279, 281-282, 284 Revised hierarchical model, 20, 140, 144, 152, 156-157, 159, 167 Romance languages, 245 Romanized script, 44-45 Russian, 35, 260, 262, 294
Saccades, 113, 116-121 Schema, 275, 277, 282-284, 289-292 Second-language acquisition, 139 Second language learning, 31-57 Second language writing, 317-338 Selective attention, 33, 35-37, 39,42 Self-organization, 64 Self-Organizing Map (SOM), 63-81 Self-Organizing Model of Bilingual Processing (SOMBIP), 59-83 Semantic, 102 Semantic Blindness (SB), 90-91, 99-100, 103 Semantic facilitation, 101 Semantic priming, 137-138, 141,157158 Sentence completion task, 260, 265 Sentence constraint, 114, 116, 130 Sentence interpretation task, 8-11, 34, 36, 47 Sentence level associations, 329-332 Sentence-level modes, 238-239, 258 Sentence production, 319, 321-322, 325, 331 Sentential context effects, 137-138 Sentential priming, 125, 138, 142-144, 147-158 Shadowing, 7, 18-19 Simple Recurrent Network (SRN), 60-62, 70 Simulations, 59-83
378
Subject Index
Sino-Tibetan, 287 Situation model, 166-167, 176-177, 239 Situative level, 239 Skill learning, 327 Skilled readers, 172 Slavic language, 294 Social factors, 341 Sociological, 340 Spanish, 7-15, 20-22, 35, 37, 39, 42-44, 49-51, 89-97, 99-100, 115, 121,123, 127-129, 137, 140-150, 152-160, 168, 174-176, 187-188, 192, 194-199, 202209, 240, 244, 246, 257-258, 260, 262, 299-303, 305-309, 311,322, 335, 339340, 343-345, 355 Spanish-dominant, 198, 205-208, 345 Spanish-English, 7, 9-11, 13-15, 20, 8993, 115, 127-128, 137, 140-142, 144, 148, 153, 158, 160, 168, 174, 187-188, 193, 197, 199, 258, 262, 299, 301,305, 339, 343-345 Spanish formal definitions, 305-306 Spanish oral proficiency, 303 Spanish-speaking, 193-194, 202, 300301, 311 Spanish word recognition, 305-306 Speech, 41-49 Split-language lists, 91 Spoken language, 1-28 Spreading activation, 38 S-structure, 244 Statistical learning, 41 Stimulus Onset Asynchrony (SOA), 125, 127-128 Strategies, 160, 171, 187 Study of idioms, 276 Subject-verb agreement, 325, 328, 330331 Superordinate, 304 Surface features, 175 Surface form, 174, 177 Surface level, 175, 239 Surface structure, 238 Swedish, 258 Syllabic, 67, 71, 81 Syllabic templates, 71 Synonym, 304 Syntactic, 102 Syntactic complexity, 317-319, 325-326, 330-335 Syntactic constraints, 342 Syntactic processing, 217, 219, 221
Syntactic structural variables, 324-325 Syntactic structure, 317-318, 334, 34 l, 353 Syntax, 31-35, 45-50, 170, 197, 304, 306308, 31 l, 322, 342
Taiwan, 44 Test of English as a Foreign Language (TOEFL), 287, 320 Text-base level constraints, 175-176 Text-based strategies, 171 Text comprehension in bilinguals, 165184 Text comprehension models, 170, 174 Text-structured based strategies, 171 TRACE Model, 60-61 Transfer, 33-34, 36, 38, 41-43, 45-47, 4950, 52 Translation, 143, 168, 353 Translation equivalents, 92-93, 143, 168, 200 Translation priming, 147 Tuning Hypothesis, 194 Turkish, 34 l Typological structure, 237-238 U
~
Universal grammar, 33,237-238 Universality, 292 Utterance-based learning, 50-51 V Verbs, 8, 9, 61, 69, 72-73, 79, 81,193, 195, 199, 220, 229, 240, 243,251-268, 305, 323-325, 327, 331,336 Verb morphology, 331 Verb-prepositions, 251-268 Vocabulary, 299, 304-305, 307-308 Vowels, 35, 40, 67-68, 71 W Within-language, 45, 90, 92-98, 102, 147-149, 306 Word-based learning, 50-51 Word Co-Occurrence Detector (WCD), 65-67, 69 Word order, 8-11, 34, 36, 48, 240
Subject Index
Working memory, 77, 126, 276, 319-320, 333 World Wide Web (WWW), 259 Writing, 317-338 Written language production, 318
379
This Page Intentionally Left Blank
