PROCESSING AND PRODUCING HEAD-FINAL STRUCTURES
STUDIES IN THEORETICAL PSYCHOLINGUISTICS VOLUME 38 Managing Editors Lyn Frazier, Dept. of Linguistics, University of Massachusetts at Amherst Thomas Roeper, Dept. of Linguistics, University of Massachusetts at Amherst KennethWexler, Dept. of Brain and Cognitive Science, MIT, Cambridge, Mass.
Editorial Board Robert Berwick, Artificial Intelligence Laboratory, MIT, Cambridge, Mass. Matthew Crocker, Saarland University, Germany Janet Dean Fodor, City University of New York, New York Angela Friederici, Max Planck Institute of Human Cognitive and Brain Sciences, Germany Merrill Garrett, University of Arizona, Tucson Lila Gleitman, School of Education, University of Pennsylvania Chris Kennedy, Northwestern University, Illinois Manfred Krifka, Humboldt University, Berlin, Germany Howard Lasnik, University of Connecticut at Storrs Yukio Otsu, Keio University, Tokyo Andrew Radford, University of Essex, U.K.
For further volumes: http://www.springer.com/series/6555
PROCESSING AND PRODUCING HEAD-FINAL STRUCTURES Edited by HIROKO YAMASHITA Rochester Institute of Technology, Rochester, NY, USA YUKI HIROSE University of Tokyo, Tokyo, Japan and JEROME L. PACKARD University of Illinois at Urbana-Champaign, IL, USA
13
Editors Hiroko Yamashita Department of Modern Languages and Cultures Rochester Institute of Technology 92 Lomb Memorial Drive Rochester, NY, 14623 USA
[email protected]
Yuki Hirose Department of Language and Information Sciences The University of Tokyo 3-8-1 Komaba, Meguro-ku Tokyo, 153-8902 Japan
[email protected]
Jerome L. Packard Department of East Languages and Cultures University of Illinois at Urbana-Champaign 2090A Foreign Language Building 707 S. Mathews Urbana, IL, 61801 USA
[email protected]
ISSN 1873-0043 ISBN 978-90-481-9212-0 e-ISBN 978-90-481-9213-7 DOI 10.1007/978-90-481-9213-7 Springer Dordrecht Heidelberg London New York Library of Congress Control Number: 2010937565 # Springer ScienceþBusiness Media B.V. 2011 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, microfilming, recording or otherwise, without written permission from the Publisher, with the exception of any material supplied specifically for the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work. Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)
Acknowledgements
We would like to thank all contributors to the current volume. This book is the fruit of the International Conference on Processing Head-final Structures held at the Rochester Institute of Technology in September, 2007. Building on the seminal Symposium of Japanese Sentence Processing held at Duke University in 1991 and the Workshop of Japanese Sentence Processing held at the Ohio State University in 1999, the International Conference on Processing Headfinal Structures was the first of its kind, comprehensively dealing with the processing and production of head-final structures in multiple languages. The conference was made possible by grants from the National Science Foundation (BCS-0642367), Center for Language Sciences at University of Rochester, University of Richmond, 21st Century Center of Excellence Program at the University of Tokyo, Center for Evolutionary Cognitive Science at the University of Tokyo, College of Liberal Arts at Rochester Institute of Technology and the Department of Modern Languages and Cultures at Rochester Institute of Technology. The editing of this book was funded by the National Science Foundation (BCS-0642367), and the Office of the Provost for Research at the Rochester Institute of Technology. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation or other funding sources.
v
Head-Direction and its Effect on Comprehension and Production
‘‘What about head-final structures?’’ This question, which began to be raised in the late 1980s, has expanded the depth and breadth of human sentence processing research. Earlier work on sentence comprehension in the 1970s revealed the remarkable efficiency of the human language parser. In daily life, listeners and readers receive linguistic input in linear fashion over the time course of a sentence. As information comes in, we encounter many potential lexical and syntactic ambiguities that cause a sentence fragment to have more than one possible interpretation. In English, for example, a fragment such as The lawyer investigated. . . could begin a simplex sentence with a direct object, such as The lawyer investigated the case, or it may be the beginning of a sentence with a relative clause, The lawyer investigated for tax evasion was interviewed on the morning show. Likewise, the multiple selectional restriction options on the verb knew lead the sentence fragment John knew the girl. . . to the two possible completions John knew the girl or John knew the girl was from New York. Such possibilities raise the question of how humans can so effortlessly process language in spite of such great potential for ambiguity. As a means of navigating such ambiguity, researchers have proposed that the parser tends to follow certain principles. For example, the parser may hypothesize the simplest possible structure consistent with the input data – an approach known as the Minimal Attachment Principle (Frazier, 1978) – or it may attempt to associate adjacent words as clause-mates, referred to as the Local Association Principle (Frazier & Fodor, 1978). Researchers have also found that information in the input stream tends to be exploited early on for comprehension purposes. Of particular prominence are phrasal heads – especially verbs – which provide critical information about syntactic constituency (e.g., Boland, Tanenhaus, & Garnsey, 1990; Trueswell, Tanenhaus, & Kello, 1993). Verbs do the important work of specifying selectional restrictions and assigning thematic roles to complements. In a head-initial language like English, the verb appears early, not only marking the onset of the VP but also providing selectional restriction and probabilistic information on the type vii
viii
Head-Direction and its Effect on Comprehension and Production
of complements the verb may take. For example, Garnsey, Perlmutter, Meyers, and Lotocky (1997) found that English readers utilize the statistical preference of the verb, expecting (1b) as the likely continuation when they read the verb argued but expecting (2a) when they read discovered.
(1)
(2)
The divorce lawyer argued the issue. . . a. The divorce lawyer argued the issue. b. The divorce lawyer argued the issue was irrelevant to the case. The scuba diver discovered the wreck. . . a. The scuba diver discovered the wreck. b. The scuba diver discovered the wreck was caused by a collision.
This example, in which the parser exploits information contained in the information-laden verb head, involves the notion of head-driven parsing (Abney, 1989; Pritchett, 1988, 1992). Head-driven parsing posits that the parsing of an incoming phrase critically depends on the appearance of the phrasal head, because the phrasal head provides key information about syntactic constituency. According to this theory, the parser waits until the phrasal head appears in the input stream before it assigns a structural interpretation to the phrase. As the examples in (1) and (2) demonstrate, information contained in the verb head is especially important for head-initial languages like English, making the head-driven parsing theory a natural one for head-initial structures. But it also raises an obvious question: if sentence processing is head-driven, then how are structures processed when the head follows a constituent rather than occurring in phrase-initial position (e.g., Inoue, 1991; Inoue & Fodor, 1995; Mazuka & Lust, 1990; Yamashita, 1994)? Languages such as Japanese, Korean, Hindi and Basque are predominantly head-final, with the information contained in the head naturally not available to the parser until the head appears in the input stream. In addition to the early unavailability of head information in those languages, there is also often no clear marking of the beginning of embedded clauses–so-called ‘‘left-edge marking.’’ In such cases the parser has less information about, for example, the degree of embedding within a sentence, until all phrase-final heads are processed. Often these languages have flexible word-order and phonologically null pronouns, both contributing more ambiguity to the parse (e.g., Inoue, 1991; Mazuka & Lust, 1990). To illustrate, the two arguments in Japanese marked by nominative and dative case markers in (3) may be the beginning of a simplex clause as in (4), part of a relative clause as in (5), or with a phonologically null pronoun (indicated
Head-Direction and its Effect on Comprehension and Production
ix
here as e) they may be two arguments that belong respectively to matrix and subordinate clauses, as in (6). John-ga Mary-ni. . . John-NOM Mary-DAT (4) John-ga Mary-ni pen-o ageta. John-NOM Mary-DAT pen-ACC gave. ‘John gave Mary an apple.’ (5) [[John-ga Mary-ni ageta] pen]-ga kowareta. John-NOM Mary-DAT gave pen-NOM broke ‘The pen that John gave to Mary broke.’ (6) John-ga [ e Mary-ni pen-o ageta]-to itta. John-NOM Mary-DAT pen-ACC gave-comp said ‘John said that he gave a pen to Mary.’
(3)
The question of interest is: at what point in the processing of sentences like (4)–(6) does the parser commit itself to a simplex clause analysis or posit embedded structures? With information from the syntactic head available only at clause-final position, the possibility suggested by head-driven parsing is that the parser waits to posit syntactic structure until the information from the head becomes available. As an alternative to head-driven parsing, it was proposed that the parser is more opportunistic, using all available information to build structure even before the information from the head is available. This fully incremental parsing approach (e.g., Crocker, 1994; Gorrell, 1995; Lombardo & Sturt, 2002; Stabler, 1994; Sturt & Crocker, 1996) entails that every incoming word incrementally contributes its information to the structure of the phrase under construction. The multiplicity of possible analyses inherent in such an approach entails that the parser frequently revises its initial hypotheses, potentially resulting in increased processing difficulty. In 1995 Inoue and Fodor introduced ‘‘Information–paced Processing,’’ in which the ease of revising incorrect initial parses depends upon the reliability of the existing analysis. Thus even though the delayed availability of information from the head in head-final languages might lead to more parsing possibilities than in headinitial languages, it was suggested that head-final languages may have more tolerance for adjusting initial analyses. Those questions involving head-final parsing led to a flurry of processing studies in languages with head-final structures (e.g., Berwick & Fong, 1995; Konieczny, Hemforth, Scheepers, & Strube, 1997; Miyamoto, 1999; Nakayama, 1995; Sakamoto, 1991; Vasishth, 2003; Yamashita, 1994). As the field continued to develop based on the syntax-based approach followed in the 1970s, sentence processing researchers examined factors such
x
Head-Direction and its Effect on Comprehension and Production
as discourse (Altmann & Steedman, 1988; Crain & Steedman, 1985), stochastic information (Merlo & Stevenson, 2002) and prosody (Beach, 1991; Kjelgaard & Speer, 1999). New experimental procedures were devised, with eye-tracking, event-related potential (ERP), brain imaging and most recently corpus studies, coming onto the scene and used alongside the reliable standard self-paced moving-window technique. Head-final sentence processing research developed concomitantly, with some studies focusing on areas common to head-initial and head-final structures. They include the studies on the immediate utilization of case and other morphological markings or verb information for predicting (a) forthcoming structure(s) (e.g., Kamide & Mitchell, 1999; Yamashita, 1995; Yoshida, 2006), the role of thematic roles at the different stages of processing (Hirose, 2002; Hirose & Inoue, 1998), individual differences in comprehension (Jincho, Namiki, & Mazuka, 2008) and preference on ambiguous relative clause interpretation (relative clause attachment) (e.g., Gibson, Perlmutter, Gonzalez, & Hickok, 1996; Hemforth, Konieczny, Scheepers, & Strube, 1998; Miyamoto, Gibson, Pearlmutter, Aikawa, & Miyagawa, 1999). Some researchers used head-final structures to pose questions not easily addressed in head-initial languages, such as detecting or processing relative clauses, where the head is at the final position of the construction (e.g., Hsiao & Gibson, 2003; Miyamoto & Nakamura, 2003; Ueno & Garnsey, 2007). And some investigated the processing that is unique to languages with head-final structures such as effects of scrambling in processing (Kim, 1999; Mazuka, Itoh, & Kondo, 2002; Miyamoto & Takahashi, 2002; Yamashita, 1997). These approaches were critical in advancing the field of psycholinguistics, as they drew commonalities among different types of languages or highlighted the effects of language-specific linguistic phenomena on human language processing. Studies on head-final languages are no longer uncommon in the field of sentence processing but at the same time the increasing number of topics and variety of examined structures signaled the need for a comprehensive examination of head-final sentence processing research. The International Conference on Processing Head-final Structures held at the Rochester Institute of Technology in September 2007 served that purpose. Our conference was intended to bring together cutting-edge research on head-final sentence processing and production, to generate comparisons with head-initial languages such as English and to have a hand in setting the research agenda over the next several years. The chapters in this book find their provenance in the presentations at that conference, offering experimental data and theoretical discussions from head-final structures in languages as diverse as Basque, Chinese, Hindi, Japanese, Korean and German. The chapters in Part I focus on the comprehension of head-final structures and its incremental nature. Although the details remain a matter of debate, results on head-final processing suggest that the parser is at work incrementally employing all available sources of information – such as case, agreement, or thematic roles – well before the head is encountered. Evidence for incremental
Head-Direction and its Effect on Comprehension and Production
xi
processing is presented in Part I and also appears in various chapters throughout the book. The chapters in Part II investigate the role of prosodic information in the processing of head-final structures. The Part II chapters include studies investigating the listener’s parsing of ambiguous structures with varied prosody and how prosody underlies possible differences in the ways syntactic structures are encoded and decoded by speakers and listeners. Part III addresses the relatively understudied area of language production and its relevance to head-direction. Head-initial and head-final structures would seem to be subject to different production constraints. For example, in a head-initial language with fairly rigid word-order such as English, when a subject and verb are uttered, what follows is restricted to the set of possible complements of the verb. In contrast, in head-final languages the verb does not lead the production of the VP because it is uttered clause-finally after all arguments. Furthermore, many head-final languages allow flexible word order, which would allow a greater number of possibilities in beginning a phrase. The chapters in Part III address differences and similarities in producing head-final and head-initial languages by adult native speakers and also the effects of an L2 with opposite head direction in bilingual production. The chapters in Part IV present a relatively new approach in psycholinguistics, which is the use of corpora as a hypothesis-testing tool. The examination of information obtained from language corpora – such as the frequency of words in actual use or the frequency of a verb’s complements – has become an integral part of psycholinguistic research as a means of obtaining statistical information to supplement experimental results. Recently, researchers have begun to investigate processing or production hypotheses using corpora, complementing experimental studies by offering new evidence or investigating areas not easily tested in the laboratory. These chapters offer prime examples of the use of corpora to investigate sentence processing and production. The chapters in Part V examine the case of Chinese, focusing on the processing of relative clauses. Along with German subordinate structures, Chinese relative clause constructions are ‘‘mixed-headed’’ structures, since the clause VP is head-initial while the NP is head-final. There has been a debate over why object relative clauses are generally found to be more difficult to process than the subject relative clauses. Two possible accounts – structural complexity and linear distance – are not easily tested in English due to confounds present in the relative clause structure. The ‘‘mixed-headed’’ nature of relative clauses in Chinese ameliorates this confound, allowing the effects of linear versus structural distance to be teased apart. In addition, the processing of Chinese relative clauses entails the parser’s commitment to the presence and type of empty categories. The chapters in Part V advance the discussion of Chinese relative clause processing by examining syntactic, semantic (animacy) and preceding cue information, and by reporting results obtained from experimentation using ERP.
xii
Head-Direction and its Effect on Comprehension and Production
The chapters in Part VI conclude the volume with theoretical treatments on the relationship among head direction, sentence processing and Universal Grammar. Consensus seems to favor an incremental universal core parsing architecture, in which processing constraints arise conditioned by values set on underlying parameters such as head direction. The discussion includes whether the directionality parameter is orthogonal to the occurrence of backward- versus forward-looking parsing, and to what extent locality versus antilocality effects covary with the value of the head direction parameter – notably in languages with mixed headedness. This critically includes how an open directionality parameter is to be reconciled with the fixed temporal order of human speech – surely one of the fundamental issues in natural language processing. And so the challenge to the field that was presented with the simple question ‘‘what about head-final structures’’ is substantively addressed in this collection of experimental and theoretical works. While the exact nature of head-final processing is yet to be fully explicated, the papers in this volume collectively imply that the processing of these structures critically utilizes complement information – relegating the head to a more peripheral role – contrary to what head-driven parsing approaches based mostly on data from head-initial languages would suggest. After considering the papers in this volume, the reader may surmise that the role of the syntactic head in parsing is not as critical as previously thought but we still cannot assert with complete confidence what the relative role of heads and complements is once the variable of head direction is held constant. It is hoped that this volume has made a small contribution toward shedding light on this important question and that the work presented herein will serve to encourage further research on this fascinating topic. Rochester Institute of Technology University of Tokyo University of Illinois at Urbana-Champaign
Hiroko Yamashita Yuki Hirose Jerome L. Packard
References Abney, S. (1989). A computational model of human parsing. Journal of Psycholinguistic Research, 18, 129–144. Altmann, G., & Steedman, M. (1988). Interaction with context during human sentence processing. Cognition, 30, 247–264. Beach, C. M. (1991). The interpretation of prosodic patterns at point of syntactic ambiguity: Evidence for a cue trading relations. Journal of Memory and Language, 30, 644–663. Berwick, R. C., & Fong, S. (1995). Madama butterfly redux: Parsing English and Japanese with a principles-and-parameters approach. In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing (pp. 177–208). Hillsdale, NJ: Lawrence Erlbaum.
Head-Direction and its Effect on Comprehension and Production
xiii
Boland, J. E., Tanenhaus, M. K., & Garnsey, S. M. (1990). Evidence for the immediate use of verb control information in sentence processing. Journal of Memory and Language, 29, 413–432. Crain, S., & Steedman, M. J. (1985). On not being led up the garden path: the use of context by the psychological parser. In D. Dowty, L. Kartunnen & A. Zwicky (Eds.), Natural language parsing: Psychological, computational, and theoretical perspectives (pp. 320–358). Cambridge, MA: Cambridge University Press. Crocker, M. W. (1994). On the nature of the principle-based sentence processor. In C. Clifton, L. Frazier & K. Rayner (Eds.), Perspectives on sentence processing (pp. 245–266). Hillsdale, NJ: Erlbaum. Frazier, L. (1978). On comprehending sentences: Syntactic parsing strategies. Doctoral dissertation, University of Connecticut, Storrs. Frazier, L., & Fodor, J. D. (1978). The sausage machine: A new two-stage parsing model. Cognition, 6, 1–34. Garnsey, S. M., Pearlmutter, N. P., Myers, E., & Lotocky, M. (1997). The contributions of verb bias and plausibility to the comprehension of temporarily ambiguous sentences. Journal of Memory and Language, 37(1), 58–93. Gibson, E., Pearlmutter, N. J., Canseco-Gonzalez, E., & Hickok, G. (1996). Recency preference in the human sentence processing mechanism. Cognition, 59, 23–59. Gorrell, P. (1995). Syntax and parsing. Cambridge, MA: Cambridge University Press. Hemforth, B., Konieczny, L., Scheepers, C., & Strube, G. (1998). Syntactic ambiguity resolution in German. In D. Hillert (Ed.), Sentence processing: A crosslinguistic perspective (pp. 293–312). Syntax and Semantics, 31, San Diego, CA: Academic Press. Hirose, Y. (2002). Resolution of reanalysis ambiguity in Japanese relative clauses: early use of thematic compatibility information and incremental processing. In M. Nakayama (Ed.), Sentence processing in East Asian languages (pp. 31–52). Stanford, CA: CSLI Publications. Hirose, Y., & Inoue, A. (1998). Ambiguity of reanalysis in parsing complex sentences in Japanese. In D. Hillert (Ed.), Sentence processing: A crosslinguistic perspective (pp. 71–93). Syntax and Semantics, 31, San Diego, CA: Academic Press. Hsiao, F., & Gibson, E. (2003). Processing relative clauses in Chinese. Cognition, 90, 3–27. Inoue, A. (1991). A Comparative study of parsing in English and Japanese. Doctoral dissertation, University of Connecticut, Storrs. Inoue, A., & Fodor, J. D. (1995). Information-paced parsing of Japanese. In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing (pp. 9–63). Hillsdale, NJ: Erlbaum. Jincho, N., Namiki, H., & Mazuka, R. (2008). Effects of verbal working memory and cumulative linguistic knowledge on reading comprehension. Japanese Psychological Research, 50(1), 12–23. Kamide, Y., & Mitchell, D. C. (1999). Incremental pre-head attachment in Japanese parsing. Language and Cognitive Processes, 14(5/6), 631–662. Kim, Y. (1999). The effects of case marking information on Korean sentence processing. Language and Cognitive Processes, 14(5/6), 687–714. Kjelgaard, D., & Speer, S. (1999). Prosodic facilitation and interference in the resolution of temporary syntactic closure ambiguity. Journal of Memory and Language, 40, 153–194. Konieczny, L., Hemforth, B., Scheepers, C., & Strube, G. (1997). The role of lexical heads in parsing: Evidence from German. Language and Cognitive Processes, 12(2/3), 307–348. Lombardo, V., & Sturt, P. (2002). Incrementality and lexicalism. In P. Merlo & S. Stevenson (Eds.), The lexical basis of sentence processing (pp. 137–156). Amsterdam: John Benjamins. Mazuka, R., Itoh, K., & Kondo, T. (2002). Cost of scrambling in Japanese sentence processing. In M. Nakayama (Ed.), Papers from international East Asian psycholinguistics workshop. Stanford, CA: CSLI. Mazuka, R., & Lust, B. (1990). On parameter-setting and parsing: Predictions for acquisition. In L. Frazier & J. de Villiers (Eds.), Language processing and acquisition (pp. 163–206). Dordrecht: Kluwer.
xiv
Head-Direction and its Effect on Comprehension and Production
Merlo, P., & Stevenson, S. (Eds.) (2002). The lexical basis of sentence processing: Formal, computational and experimental issues. Amsterdam: John Benjamins. Miyamoto, E. T. (1999). Relative clause processing in Brazilian Portuguese and Japanese. Doctoral dissertation, Massachusetts Institute of Technology. Miyamoto, E. T., Gibson, E., Pearlmutter, N. J., Aikawa, T., & Miyagawa, S. (1999). A U-shaped relative clause attachment preference in Japanese. Language and Cognitive Processes, 14, 663–686. Miyamoto, E. T., & Nakamura, M. (2003). Subject/object asymmetries in the processing of relative clauses in Japanese. In G. Garding & M. Tsujimura (Eds.), Proceedings of the 22nd WCCFL (pp. 342–355). Somerville, MA: Cascadilla Press. Miyamoto, E. T., & Takahashi, S. (2002). Sources of difficulty in processing scrambling in Japanese. In M. Nakayama (Ed.), Sentence processing in East Asian languages (pp. 167–188). Stanford, CA: CSLI. Nakayama, M. (1995). Scrambling and probe recognition. In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing (pp. 257–273). Hillsdale, NY: Erlbaum. Pritchett, B. L. (1988). Garden path phenomena and the grammatical basis of language processing. Language, 64(3), 539–576. Pritchett, B. L. (1992). Grammatical competence and parsing performance. Chicago and London: The University of Chicago Press. Sakamoto, T. (1991). Processing empty subjects in Japanese: Implications for the transparency hypothesis. Doctoral dissertation, City University of New York. Stabler, E. P. (1994). The finite connectivity of linguistic structure. In C. Clifton, L. Frazier & K. Rayner (Eds.), Perspectives on sentence processing (pp. 303–336). Hillsdale, NJ: Erlbaum. Sturt, P., & Crocker, M. (1996). Monotonic syntactic processing: a cross-linguistic study of attachment and reanalysis. Language and Cognitive Processes, 11, 449–494. Trueswell, J. C., Tanenhaus, M. K., & Kello, C. (1993). Verb-specific constraints in sentence processing: Separating effects of lexical preference from garden-paths. Journal of Experimental Psychology: Learning, Memory and Cognition, 19(3), 528–553. Ueno, M., & Garnsey, S. M. (2007). Gap-filling vs. filling gaps: Event-related brain indices of subject and object relative clauses in Japanese. In N. H. McGloin & J. Mori (Eds.), Proceedings of the 15th Japanese/Korean Linguistics Conference (pp. 288–301). Stanford, CA: CSLI. Vasishth, S. (2003). Working memory in sentence comprehension: Processing Hindi center embeddings. New York: Garland Press. Yamashita, H. (1994). Processing of Japanese and Korean. Doctoral dissertation, The Ohio State University. Yamashita, H. (1995). Verb argument information used in a prodrop language: An experimental study in Japanese. Journal of Psycholinguistic Research, 24, 333–347. Yamashita, H. (1997). The effects of word-order and case marking information on the processing of Japanese. Journal of Psycholinguistic Research, 26, 163–188. Yoshida, M. (2006). Classifier mismatch effects and relative clause prediction in Japanese. Doctoral dissertation, University of Maryland.
Contents
Part I 1
Incremental Processing and Head-final Structures
Rich Agreement in Basque: Evidence for Pre-verbal Structure Building . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Leticia Pablos
2
The Processing of Japanese Control Sentences . . . . . . . . . . . . . . . . . . Jeffrey D. Witzel and Naoko O. Witzel
3
Individual Differences in Sentence Processing: Effects of Verbal Working Memory and Cumulative Linguistic Knowledge . . . . . . . . . . . . . . . . . . Nobuyuki Jincho and Reiko Mazuka
Part II 4
5
6
7
23
49
Prosody and Processing
Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension – ERP Evidence from German Ambiguous DPs . . . . . Petra Augurzky and Matthias Schlesewsky Production-Perception Asymmetry in Wh-scope Marking . . . . . . . . . . Yuki Hirose and Yoshihisa Kitagawa
Part III
3
69
93
Production of Head-final Structures
The Production of Head-Initial and Head-Final Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mikihiro N. Tanaka, Holly P. Branigan, and Martin J. Pickering
113
Incremental Sentence Production: Observations from Elicited Speech Errors in Japanese . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Noriko Iwasaki
131
xv
xvi
8
Contents
The Status of Dative Constructions in Korean, English and in the Korean-English Bilingual Mind . . . . . . . . . . . . . . . . . . . . . . Jeong-Ah Shin and Kiel Christianson
Part IV 9
10
12
13
14
Corpus-based Approach to Processing and Production
Subject Preference, Head Animacy and Lexical Cues: A Corpus Study of Relative Clauses in Chinese . . . . . . . . . . . . . . . . . . Fuyun Wu, Elsi Kaiser, and Elaine Andersen
173
Why Speakers Produce Scrambled Sentences: An Analysis of a Spoken Language Corpus in Japanese . . . . . . . . . . . . . . . . . . . . Tadahisa Kondo and Hiroko Yamashita
195
Part V 11
153
Processing Relative Clauses in Chinese
Filler-Gap Processing in Mandarin Relative Clauses: Evidence from Event-Related Potentials. . . . . . . . . . . . . . . . . . . . . . . Jerome L. Packard, Zheng Ye, and Xiaolin Zhou
219
Animacy and the Resolution of Temporary Ambiguity in Relative Clause Comprehension in Mandarin. . . . . . . . . . . . . . . . . . . . . . . . . . Yowyu Lin and Susan M. Garnsey
241
Garden Path and the Comprehension of Head-Final Relative Clauses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chien-Jer Charles Lin and Thomas G. Bever
277
Use Your Headedness: An Exercise in Psycholinguistic Exploitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shukhan Ng and Janet Dean Fodor
299
Part VI
Head-Direction and Processing Theory
15
On Being Both Head-Initial and Head-Final . . . . . . . . . . . . . . . . . . . Markus Bader
325
16
Integration and Prediction in Head-Final Structures . . . . . . . . . . . . . Shravan Vasishth
349
17
Directionality in the Architecture of the Language Faculty: Integrating with Real Time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Barbara Lust
369
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
395
List of Contributors
Elaine Andersen University of Southern California, Los Angeles, CA, USA Petra Augurzky University of Tu¨bingen, Tu¨bingen, Germany Markus Bader University of Konstanz, 78457 Konstanz, Germany Thomas G. Bever University of Arizona, Tucson, AZ, USA Holly P. Branigan University of Edinburgh, Edinburgh, Scotland, UK Kiel Christianson University of Illinois at Urbana-Champaign, Urbana-Champaign, IL, USA Janet Dean Fodor The Graduate Center, City University of New York, New York, NY, USA Susan M. Garnsey University of Illinois at Urbana-Champaign, Urbana-Champaign, IL, USA Yuki Hirose The University of Tokyo, Tokyo, Japan Noriko Iwasaki School of Oriental and African Studies, University of London, London, UK Nobuyuki Jincho Laboratory for Language Development, RIKEN Brain Science Institute, Saitama, Japan Elsi Kaiser University of Southern California, Los Angeles, CA, USA Yoshihisa Kitagawa Indiana University, Bloomington, IN, USA Tadahisa Kondo Human and Information Science Laboratory, NTT Communication Science Laboratories, NTT Corporation, Atsugi, Kanagawa, 243-0198 Japan Chien-Jer Charles Lin Department of East Asian Languages and Cultures, Indiana University, Bloomington, IN, USA Yowyu Lin College of Foreign Languages and Literature, National Cheng Chi University, Taipei City, Taiwan, China
xvii
xviii
List of Contributors
Barbara Lust Department of Human Development, College of Human Ecology, Cornell University, Ithaca, NY, USA Reiko Mazuka Brain Science Institute, RIKEN, Tokyo, Japan; Duke University, Durham, NC, USA Shukhan Ng University of Texas, San Antonio, TX, USA Leticia Pablos Leiden University Centre for Linguistics, Leiden University, Leiden, The Netherlands Jerome L. Packard University of Illinois at Urbana-Champaign, Urbana-Champaign, IL, USA Martin J. Pickering University of Edinburgh, Edinburgh, Scotland, UK Matthias Schlesewsky University of Mainz, Mainz, Germany Jeong-Ah Shin Seoul National University, Seoul, Korea Mikihiro N. Tanaka Showa University, Yamanashi-ken, Japan Shravan Vasishth Universita¨t Potsdam, Potsdam, Germany Jeffery D. Witzel University of Texas at Arlington, Arlington, Texas, USA; The University of Arizona, Tucson, AZ, USA Naoko O. Witzel University of Texas at Arlington, Arlington, TX, USA; The University of Arizona, Tucson, AZ, USA Fuyun Wu Institute of Linguistics Studies, Shanghai International Studies University, Shanghai, China Hiroko Yamashita Rochester Institute of Technology, Rochester, NY, USA Zheng Ye Peking University, Beijing, China Xiaolin Zhou Peking University, Beijing, China
Part I
Incremental Processing and Head-final Structures
Chapter 1
Rich Agreement in Basque: Evidence for Pre-verbal Structure Building Leticia Pablos
Head-final languages pose an interesting question for human sentence parsing. In previous discussions in the literature about processing head-final languages, two main alternatives for how the parsing of these languages could proceed have been suggested. The first alternative is the head-driven parsing model (Abney, 1989; Mulders, 2003; Prittchet, 1992), which assumes that the parser needs to wait to encounter the verbal head to interpret all the relevant information. The second is the fully incremental parsing model (Crocker, 1994; Gorrell, 1995; Lombardo & Sturt, 2002; Stabler, 1994; Sturt & Crocker, 1996), which assumes that arguments are interpreted incrementally and incorporated into the structure of the sentence through the use of language-specific grammatical information (such as word order, case marking, etc.). This second approach presupposes that the parser can build a fully connected syntactic structure even before the verbal head becomes available. Many studies on the processing of head-initial-languages have focused on how the argument structure of the verb can guide parsing by providing direct cues about the number of complements and their probability of occurrence (Boland, Tanenhaus, Garnsey, & Carlson, 1995; Garnsey, Pearlmutter, Myers, & Lotocky, 1997; MacDonald, Pearlmutter, & Seidenberg, 1994; Trueswell, Tanenhaus, & Kello, 1993). Taking this evidence as a starting point, in this chapter I would like to explore whether there are additional cues that the parser can use to project structure in an anticipatory fashion, before the verb becomes available in the input. In particular, I will discuss whether agreement information can assist pre-head structure building in head-final languages and whether the parser uses an indirect cue provided by an agreement mismatch to make inferences about the upcoming structure. Throughout the chapter I will discuss a study by Pablos (2006) that used the agreement information provided by auxiliaries in Basque with two goals in L. Pablos (*) Leiden University Centre for Linguistics, Leiden University, Postbus 9515, 2300 RA, Leiden, The Netherlands e-mail:
[email protected]
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_1, Ó Springer ScienceþBusiness Media B.V. 2011
3
4
L. Pablos
mind. One of the goals was to examine whether agreement creates a positive expectation for certain arguments in the upcoming input when the auxiliary is fronted and the verb’s information (participle) is not available until the end of the sentence. The other was to test whether an indirect cue provided by the morphological mismatch between the auxiliary and the immediately following noun phrase (NP) is used by the parser in order to make inferences about the presence of a clause boundary and avoid a garden-path. The present study on Basque auxiliaries shows that these agreement mismatches may provide cues to the parser for detecting clause boundaries and predicting an upcoming clause before the specific lexical-semantics of the verb becomes available. This provides evidence that the parser can use more than bottom-up projection of lexical information on a head and that it can make inferences about the upcoming material based on indirect cues. In what follows, I first present evidence from head-final languages for the claim that the parser builds structure incrementally and immediately. I then present the study that uses agreement information in the auxiliaries in Basque and show its relation to existing literature on agreement in comprehension and to studies that use word order and case marking as clause boundary inducers. Finally, the experimental results of the Basque auxiliary agreement study are discussed in the context of the issues of head-finality.
1.1 Evidence from Head-Final Languages in Favor of Incremental Structure Building Previous studies on head-final languages have demonstrated an incremental attachment of phrases in the structure before the verbal head is accessed, mainly based on evidence from Japanese and German (for Japanese, see Kamide, 2006). Mazuka and Itoh (1995) showed in an eye-tracking study that the three initial NPs that had identical nominative case marking were read slower than those NPs that had three different case markings (NOM-ACC-DAT). They considered this as evidence to show that structure is built prior to the verb being processed. A second piece of evidence against head-driven parsers in Japanese comes from a sentence completion study by Inoue and Fodor (1995) for which they used fragments with the NOM-DAT-ACC case NP sequence. They reported that Japanese speakers consistently completed the fragments as if the three NPs were co-arguments of the same verb and they all belonged to the same clause. Finally, the findings of Bader and Lasser’s (1994) self-paced reading study on German also provide evidence that there is association of NPs before reaching the verbal head. In German, embedded clauses are introduced by a complementizer and verbs appear in clause-final position. The experimental items they used were embedded sentences with a head-final configuration, where the pronoun sie was ambiguous between a nominative and an accusative and
1 Rich Agreement in Basque: Evidence for Pre-verbal Structure Building
5
could either be taken for the subject of the main clause or for the object of the embedded clause. The sentence was disambiguated at the clause-final auxiliary, which was passive or active depending on the condition. Their results provided evidence in favor of incremental parsing models since participants initially took the ambiguous pronoun sie to be the subject of the main clause in both sentences, as shown by reading times that were significantly longer for the passive auxiliary than for the active auxiliary. Other arguments in favor of incremental structuring of sentences include Japanese studies on the processing of dative marked NP phrases (Kamide & Mitchell, 1999); scrambled wh-phrases (Miyamoto & Takahasi, 2000, 2003) and backward anaphora cases (Aoshima, Yoshida, & Phillips, 2009). Pre-verbal filler-gap dependencies provide one further piece of evidence. Numerous studies have shown preverbal dependency completion effects in long distance dependencies involving wh-gaps (e.g., V2 constructions in German, Clahsen & Featherston, 1999; long-distance scrambled wh-phrases in Japanese: Aoshima, Phillips, & Weinberg, 2004; Nakano, Felser, & Clahsen, 2002). Complementing previous studies, the Basque experiment in this chapter provides new evidence supporting incremental structure building in head-final languages. I will show that Basque speakers use case information and a morphological mismatch between the auxiliary and the NP that follows to: (i) attach every word incrementally and without delay and (ii) formulate predictions about the upcoming input.
1.2 Basque: An Ergative Rich-Agreement Language Basque is a head-final language with SOV order in sentences with canonical word order, where the case array is subject-dative-object-attributive-verb (Hualde & Ortiz de Urbina, 2003, p. 448f.). Additionally, Basque is an ergative-absolutive language, with morphological ergativity manifested in its case system. This means that the intransitive subject and the transitive object both receive absolutive case, as shown in (1) and (2) respectively. The absolutive case is the unmarked case of the case system in Basque and often has zero morphology. On the other hand, the transitive subject bears morphologically marked ergative case as shown in (3)1 (Dixon, 1994; Etxepare, 2003, p. 363f.). (1)
1
Gizona etorri Man-ABSSG come ‘The man has come.’
da. aux
Both the absolutive and the ergative can have other syntactic functions and appear in different syntactic contexts. I am only stating the syntactic functions that are relevant for the present discussion. In order to learn further about the other contexts the reader is referred to Hualde & Ortiz de Urbina (2003, pp. 179f., 180f., 364f.).
6
(2)
(3)
(4)
L. Pablos
Eklipsea ikusi du. aux Eclipse-ABSSG see ‘S/he has seen the eclipse.’ Gizonek eklipsea ikusi dute. Man-ERGPL eclipse-ABSSG see aux ‘The men have seen the eclipse.’ Basque Case Morphology for Absolutive and Ergative cases for stems ending in vowels Indefinite
Singular
Plural
ABS
-Ø
-a
-ak
ERG
-k
-ak
-ek
(Saltarelli, 1988) As illustrated in (4), the endings -a (singular), -ak (plural) are the definite articles for absolutive case, whereas the endings –ak (singular), -ek (plural) are the definite articles for ergative case. Thus, the absolutive plural and ergative singular forms are identical (-ak ending) in Basque, as shown in (5) and (6) respectively. (5)
(6)
Gizonak etorri Man-ABSPL come ‘The men have come.’ Gizonak eklipsea Man-ERGSG eclipse-ABSSG ‘The man has seen the eclipse.’
dira. aux ikusi see
du. aux
Verbal auxiliaries in Basque agree with all arguments in a clause, providing morphological information about agreement, number, tense and mood (Hualde & Ortiz de Urbina, 2003, p. 205f.). In (7) and (8), the verb erosi ‘‘buy’’ is preceded by empty pronouns and followed by the verbal auxiliary. (7)
(8)
Erosi d-it-u-t. proiproj buy ABSj. tense – number –have-ERGi ‘I have bought them.’ Erosi d-izk-i-o-t. proiprok proj buy ABSj. tense – number –have-DATK – ERGi ‘I have bought them for him/her.’
Thus, auxiliaries in Basque provide the reader with exact information about the number of internal arguments of the verb even when (optional) arguments have been dropped.
1 Rich Agreement in Basque: Evidence for Pre-verbal Structure Building
7
As illustrated in (7) and (8), in addition to rich agreement information, Basque has another property that makes it very suitable for testing theories about syntactic structure prediction: it is a pro-drop language. Due to the rich verbal morphology, the object(s) and the subject can be elided with all relevant information preserved in the morphology of the auxiliary. Hence, this can generate a great deal of ambiguity for the parser since, being a head-final language, the verbal head is delayed and the exact number of arguments that the verb takes cannot normally be recognized prior to the auxiliary. The verbal cluster in Basque generally consists of the participle, which is a non-finite form and the auxiliary, which is the inflected form. Critically, under negation, the auxiliary that contains information on the number of arguments splits from its position following the verbal head, as in (9) and is fronted, appearing before the verbal head as shown in (10). (9)
(10)
Nik liburua erosi dut. I-ERG book-ABSSG buy aux ‘I have bought the book.’ Nik ez dut liburua erosi. I-ERG Neg aux book-ABSSG buy ‘I haven’t bought the book.’
In the context of negation, therefore, Basque auxiliaries allow the anticipation of the upcoming argument structure of a clause final verb and eliminate the temporary ambiguity that is created by empty pronouns and by verbs that can have more than one argument structure. This is the configuration that the Basque experiment covered in this chapter examines.
1.2.1 Brief Overview of Basque Psycholinguistic Research The on-line study of the comprehension and production of Basque is a recent addendum to the psycholinguistic and neurolinguistic literature. Much of this literature has focused on the comparison of monolingual (L1) and bilingual (L2) speakers of Basque, where some of the studies looked at the comprehension of case marking and phrase structure violations using Event-related Potentials (ERP) in L1 and L2 speakers (Diaz, Sebastia´n-Galles, Erdozia, Mueller, & Laka, 2006; Zawiszewski, 2007) while others focused on attachment preferences for relative clauses in Basque-Spanish bilinguals (Gutierrez Ziardegi, 2006). A third set of studies examined semantic priming effects in morpheme recognition by proficient Basque-Spanish bilinguals (Perea, Dun˜abeitia, & Carreiras, 2008) and a final set of experiments looked at the production of determiner phrases (DP) and other linguistic phenomena in Basque-Spanish bilingual speakers (Santesteban, 2006; Santesteban & Costa, 2006).
8
L. Pablos
Very few studies have investigated language comprehension in pure monolingual speakers of Basque. Erdozia (2006) used behavioral and ERP methods to examine comprehension of canonical and non-canonical sentences in Basque with ambiguous case markers. Carreiras, Dun˜abeitia, Vergara, Zieza, and Laka (2007) used the same methods to investigate object versus subject preference in relative clauses in Basque. Finally, Dun˜abeitia, Perea, and Carreiras (2007a, 2007b) used a masked priming lexical decision paradigm to look for morphological processing effects in compounds and in transposed-letter contexts in Basque.
1.2.2 Agreement in Sentence Comprehension Most studies on agreement in language comprehension have mainly focused on the processing of subject-verb agreement errors through behavioral data (Nicol, Forster, & Veres, 1997; Pearlmutter, Garnsey, & Bock, 1999; Thornton & McDonald, 2003) and through ERP data related to the study of agreement violations in languages like English (Coulson, King, & Kutas, 1998; Hagoort, Brown, & Groothusen, 1993; Kaan, 2002; Osterhout & Mobley, 1995), where the agreement information is relatively scarce and in languages with richer agreement systems such as Spanish (Barber & Carreiras, 2005), Basque (Diaz et al. 2006; Pablos & Saddy, 2009; Zawiszewski, 2007) and Hindi (Nevins, Dillon, Malhotra, & Phillips, 2007). Studies on agreement violations in Basque have tested both subject and object verb agreement violations in bilingual and monolingual speakers. Diaz et al. (2006) and Zawiszewski (2007) examined ERP responses to agreement violations like those reflected in (11b) for subject-verb agreement and (12b) for object-verb agreement at the auxiliaries, comparing them with their grammatical counterparts in (11a) and (12a). (11)
a. Keparen arrebek egunkaria saskian ekarri dute Kepa-GEN sister-PL newspaper basket bring they-have-it kioskotik. news-stand-the-from ‘Kepa’s sisters have brought the newspaper from the newsstand in a basket.’ b. Keparen arrebek egunkaria saskian ekarri *du Kepa-GEN sister-PL newspaper basket bring she-have-it kioskotik. newsstand-the-from ‘Kepa’s sisters *has brought the newspaper from the newsstand in a basket.’
(12)
a. Zuk You
ni hondartzara eramaten nauzu batzuetan. me beach-the-to take-PROG me-have-youSG sometimes
1 Rich Agreement in Basque: Evidence for Pre-verbal Structure Building
9
‘Sometimes you take me to the beach.’ b. Zuk ni hondartzara eramaten *duzu batzuetan. You me beach-the-to take-PROG it-have-youSG sometimes ‘Sometimes you *take me to the beach.’ In these studies the agreement information in the auxiliaries acts as a final check on the grammaticality of the structure. Participants should detect the ungrammaticality of the sentence when the morphology at the auxiliary does not match the agreement they expected considering the unambiguous case array of previous NPs in the sentence and the lexical information provided by the participle. The agreement violations at the auxiliary position in (11) and (12) generated an anterior negativity and a P600 in the case of subject-verb agreement violations and a rightward posterior negativity and posterior negativity followed by a positivity, for non-natives and natives respectively, in object-verb agreement violations. The effects associated with the agreement violations in these experiments could therefore be considered an effect of failure to encounter the predicted agreement information at the auxiliary. Moreover, these results also suggest that the morphological information of the auxiliary can indeed signal the type and number of arguments. In connection to this, the experiment discussed in this chapter uses fronted auxiliaries to examine whether the agreement information provided by the auxiliary can create expectations on the valence of the verb and the structure of the sentence. This study therefore may be considered the first in which agreement is examined in Basque as a means of guiding the analysis of the sentence and not as a final check on the grammaticality of the sentence.
1.2.3 Case Markers and Word Order as Inducers of Clause Boundaries The logic of the present experiment follows a long tradition in proposing that information about argument structure appears in many forms that might not involve the head of the predicate. In light of the evidence coming from previous studies that have claimed that case markers and word order can be inducers of clause boundaries in Japanese, this experiment investigates whether the combination of agreement and case marking can predict the presence of a clause and if the cue provided by this combination generates predictions about syntactic structure in the parser. Within the literature on the use of case markers in Japanese, Inoue (1991) was among the first to discuss the use of case markers to induce the correct structure of a sentence. He investigated the processing of the verb or the following word to test how reanalysis occurred but he did not focus on whether the case marked NPs would be considered co-arguments before the verb was
10
L. Pablos
processed. Yamashita (1997), on the other hand, reported that Japanese readers are sensitive to the types of case markers that appear in NP sequences by showing disruption in reaction times at the pre-verbal NPs, when there were two nominative marked NPs together (in the sequence NP-Nom NP-Dat NPNom). Miyamoto (2002, 2003) extended Yamashita’s work and showed that the slowdown found in a particular sequence of NPs (nominatives or accusatives) was a reflection of a general case driven strategy to assign clause boundaries. More specifically, he argued that in Japanese a second nominative or accusative NP can mark the left-edge of a clause, because it is not common to have two NPs with nominative case and grammatically prohibited to have two NPs with accusative case within a single clause. The results of these studies can be taken as direct evidence for incremental structure building models since they show that the case marking of NPs is used to link the NPs together and derive the kind of structure being processed (e.g., monoclausal vs. biclausal) before the verbal head is encountered. In what follows, the basic manipulation of the Basque study is introduced to differentiate between two issues regarding case marking expectations. One is the case order expected within sentences (canonical vs. non-canonical) and the other is the case order that can cause the parser to postulate a clause boundary when combined with auxiliary agreement information (this agreement information containing both argument structure and number information). On the one hand, the target sentence in (13a) shows an ERG-DAT-ABS case marking combination, which, as pointed out in Section 1.2, is the canonical word order for Basque (Ortiz de Urbina, 2003, p. 448f.). Based on previous discussion on case markers in Section 1.1 and 1.2.3, it is predicted that in this specific example, the ABS-marked NP will be considered a co-argument of the preceding ERG-marked and DAT-marked NPs. Thus, the sequence ERGDAT should strongly predict a following ABS-marked NP. This expectation to take the ABS-marked NP to be part of the main clause is used in (13a) to lead the parser down the garden-path, since it is predicted that the parser will not consider this ABS-marked NP as part of an embedded clause until the embedded verb and auxiliary are encountered. It is predicted that the parser will have to reanalyze the structure processed so far at the embedded verb and consider the ABS-marked NP as part of the embedded clause. This condition is contrasted with that in (13b), where the agreement at the auxiliary clearly marks a monotransitive structure to which the DAT-marked argument cannot belong. Therefore, this should be an unambiguous indication for the parser to posit an embedded clause before the embedded verb and auxiliary are encountered. In (13b) the embedded sentence null subject does not generate any ambiguity in the parsing of the sentence. The fact that there is an agreement mismatch between the auxiliary and following dative-marked NP should be enough to determine the existence of an embedded sentence. In contrast, in (13a), where the auxiliary and the following NP match and the ABS-SG-marked NP can be considered a co-argument of the previous ERG and DAT-marked NPs, the fact that there is a null subject does indeed contribute to the ambiguity.
1 Rich Agreement in Basque: Evidence for Pre-verbal Structure Building
(13)
11
a. Match-null subject Maitanek ez dio ikasleari [ txostena Mi-ERG neg aux.ditran student-DAT [proi document-ABS idatzi duela] esan. write aux.trans-Comp] say. ‘Maitane hasn’t told the student that she has written the document.’ b. Mismatch-null subject Maitanek ez du [ ikasleari txostena Mi-ERG neg aux.trans [proi student-DAT document-ABS idatzi diola] esan. write aux.ditrans-Comp] say. ‘Maitane hasn’t said that she has written the document to the student.’
On the other hand, the control sentences in (14) show a word order combination where the NP zuzendariak ‘‘director’’ unambiguously marks a clause boundary. Recall, as noted in Section 2.2, that the ambiguous case marking –ak in Basque can correspond to an absolutive plural and an ergative singular. Nonetheless, none of these case markings are compatible with the main clause interpretation of the NP zuzendariak ‘‘director’’. The ERG singular interpretation for the main clause is impossible because a simple clause cannot contain two ERG-marked NPs in Basque. And the number agreement marked in the fronted auxiliaries du/dio (singular) disallows the ABS plural interpretation for the NP zuzendariak ‘‘director’’ within the main clause. This said, this case ambiguity cannot be completely discarded and is only resolved at the ABSSG-marked NP txostena ‘‘document’’ in (14ab) when this argument makes the ABS-PL interpretation of the NP zuzendariak ‘‘director’’ in the embedded clause impossible2. This case ambiguity, however, does not affect the predictions for our experimental sentences about the inference of clause boundaries from an agreement mismatch cue3. The resulting word order in the control conditions is either ERG-DAT-ERG or ERG-ERG-DAT, since the option to have the NP director belong to the same clause as the NPs unambiguously marked ERG-DAT in (14a) and ERG in (14b) becomes unavailable. (14)
2
a. Match-overt subject [zuzendariak txostena Maitanek ez dio ikasleari M-ERG neg aux.ditrSG student-DAT [director-ERGSG document-ABS
Overall, it is not clear if the parser has a preference to interpret initially the ERG case over the ABS case in this kind of context when both case markings can have a subject role. 3 For an experiment that examines a purely ambiguous context with these two cases outside an agreement mismatch context, the reader is referred to Erdozia (2006).
12
L. Pablos
idatzi duela] esan. write aux.trans-Comp] say. ‘Maitane hasn’t told the student that the director has written the document.’ b. Mismatch-overt subject Maitanek ez du [zuzendariak ikasleari txostena M-ERG neg aux.transSG [director-ERGSG student-DAT document-ABS idatzi diola] esan. write aux.ditrans-Comp] say. ‘Maitane hasn’t said that the director has written the document to the student.’
In sum, in the items of the experiment, clause boundaries are predicted based on the following factors: (1) the case information combination between the agreement information in the auxiliary and the NP following the auxiliary (match/mismatch); (2) the expectation of the parser for a particular case marked NP considering the case array of the previous NPs (ERG-DAT) and (3) the number agreement marked in the fronted auxiliary (SG). Therefore, in the examples in (13) and (14), a clause boundary is posited to be more clearly predicted using agreement information at the auxiliary than when case marking or word order information alone is used. In addition to case marking and word order information, the agreement information at the auxiliary is crucial in determining whether there is an embedded clause or not.
1.2.4 Experiment on Basque Auxiliary Fronting In this study, two contexts were tested to induce different expectations in the parser by means of a morphological match/mismatch. In the conditions with a mismatch between the auxiliary and the NP, the prediction was that the parser would be able to infer a clause boundary and avoid a garden-path. In contrast, it was predicted that the matching conditions would initially be analyzed as mono-clausal and would lead to reanalysis later in the sentence upon encountering the embedded auxiliary. In order to maintain the ambiguity in the argument structure of the embedded verb, optionally ditransitive verbs were selected4. The complete set of experimental conditions is shown in (15).
4
These verbs are the so-called ‘‘trivalent ditransitives’’, which can increase their valency with the addition of a dative-marked argument fulfilling the role of recipient or beneficiary. They all have therefore a transitive bivalent use. The dative argument of these verbs is doubly marked by case on the NP and dative affixes on the auxiliary (Etxepare, 2003, p. 411f.).
1 Rich Agreement in Basque: Evidence for Pre-verbal Structure Building
13
(15) a. Match-null subject Maitaneki ez dio ikasleari [proi ikastaroko txostena Mi-ERG neg aux.ditran student-DAT [proi workshop’s document-ABS oso presaka idatzi duela] esan. very fast write aux.trans-Comp] say. ‘Maitane hasn’t told the student that she/he has written the workshop’s document in a hurry.’ b. Match-overt subject Maitanek ez dio ikasleari [zuzendariak ikastaroko M-ERG neg aux.ditr student-DAT [director-ERG workshop’s txostena oso presaka idatzi duela] esan. document-ABS very fast write aux.trans-Comp] say. ‘ Maitane hasn’t told the student that the director has written the workshop’s document in a hurry.’ c. Mismatch-null subject Maitaneki ez du [proi ikasleari ikastaroko txostena Mi-ERG neg aux.trans [proi student-DAT workshop’s document-ABS oso presaka idatzi diola] esan. very fast write aux.ditrans-Comp] say. ‘Maitane hasn’t said that she has written the workshop’s document to the student in a hurry.’ d. Mismatch-overt subject Maitanek ez du [zuzendariak ikasleari ikastaroko M-ERG neg aux.trans [director-ERG student-DAT workshop txostena oso presaka idatzi diola] esan. document-ABS very fast write aux.ditrans-Comp] say. ‘Maitane hasn’t said that the director has written the workshop’s document to the student in a hurry.’
The rationale for the experiment is as follows. If the ditransitive auxiliary in the matching conditions in (15a) and (15b) prevents a clause boundary from being inferred due to lack of agreement mismatch between the auxiliary and the NP that follows it, then we would expect a garden-path effect at the region of the embedded verb and auxiliary in the matching null subject condition in (15a) when compared to the same region in the matching overt subject condition in (15b), because at that point the parser needs to reanalyze (15a) but not (15b), as a bi-clausal sentence. (15b) is already analyzed as bi-clausal with the appearance of the ERG-marked zuzendariak, because a simple clause in Basque cannot contain two ERG-marked NPs. By the same token, if the transitive auxiliary in the mismatch condition in (15c) causes a clause boundary to be inferred due to the agreement mismatch between the transitive auxiliary and the DAT-marked NP that follows it, then we would expect to see no garden-path effect for (15c) at the embedded auxiliary
14
L. Pablos
region (diola) when (15c) and (15d) are compared. In (15d), a clause boundary is already inferred with the appearance of the ergative-marked zuzendariak, because a simple clause in Basque cannot contain two ergative-marked NPs. As a result, it is predicted that there will be no bi-clausal reanalysis at the embedded auxiliary region (diola) in (15d). When (15c) and (15d) are compared, therefore, there should be no garden-path effect if the agreement mismatch caused a clause boundary to be inferred in (15c). The evidence for the prediction of an embedded clause in (15c) would be the lack of a garden-path effect at the embedded auxiliary region (diola) when (15c) and (15d) are compared. Conversely, if no clause boundary is inferred in the mismatch condition in (15c), we would expect a garden-path effect at the embedded auxiliary region (diola) when (15c) and (15d) are compared. A set of 24 items of four conditions each was used in a word-by-word selfpaced reading experiment, along with 72 filler sentences; and a comprehension question followed each experimental sentence and filler (see appendix). Fortyone native speakers of Basque (students at the University of Deusto and the University of the Basque Country) participated in the experiment. They were all from Basque speaking families and were bilingual speakers of Basque and Spanish, differing in their level of proficiency in Spanish. Figure 1.1 shows the residual reading times for all regions in the matching sentences (15a) and (15b) and Fig. 1.2 shows the residual reading times for all regions in the mismatching sentences (15c) and (15d).
Mean Residual Reading Time (ms)
200 Match - null subject Match - overt subject
150 100 50 0
Emb verb Emb aux
–50 –100 –150 –200
0
1
2
3
4
5
6
7
8
9
10
11
12
13
Fig. 1.1 Mean residual reading times per region for the overt subject and null subject matching conditions
Maitanek1 ez2 NP-ERG1 neg2 oso8 presaka9 mod8 adv9
dio3 ikasleari4 zuzendariak5 ikastaroko6 txostena7 aux.ditran3 NP-DAT4 NP-ERG5 mod6 NP-ABS7 idatzi10 duela11 esan12. V10 aux.trans11 V12
1 Rich Agreement in Basque: Evidence for Pre-verbal Structure Building
15
Mean Residual Reading Time (ms)
200 Mismatch - overt subject Mismatch -null subject
NP Mismatch
150 100 50 0
Emb verb
–50 –100
Emb aux
–150 –200 0
1
2
3
4
5
6
7
8
9
10
11
12
13
Fig. 1.2 Mean residual reading times in milliseconds per region for the mismatching null and overt subject conditions
Maitanek1 ez2 NP-ERG1 neg2 oso8 presaka9 mod8 adv9
du3 zuzendariak4 ikasleari5 ikastaroko6 txostena7 aux.trans3 NP-ERG4 NP-DAT5 mod6 NP-ABS7 idatzi10 diola11 esan12. V10 aux.ditrans-Comp11 V12
The results indicate a garden-path effect for the matching null subject condition (15a), i.e., a slower RT in the matching null subject condition (15a) at the embedded verb and the verbal auxiliary (regions 10 idatzi and 11 duela) when compared with the same regions in (15b), the matching overt subject condition (see Fig. 1.1). A 2 2 ANOVA was performed on regions 10 and 11. For region 10, the ANOVA yielded a main effect for the embedded subject presence ðF1ð1; 40Þ ¼ 11:45; p50:01; F2ð1; 23Þ ¼ 4:93; p5:0Þ, such that the average residual reading time was significantly slower for the null subject condition (15a) when compared with the overt subject condition (15b). The main effect of matching was non-significant F1(1,40) = 0.397, p < .05; (F2(1,23) = 1.825, p < .05). No significant interaction effect was observed between the two factors ðF1ð1; 40Þ ¼ 0:149; p4:05; ðF2ð1; 23Þ ¼ 0:804; p50:05Þ. The post-hoc pairwise t-test, however, shows a significant difference in the means of the match/null vs. match/overt conditions ðF1ð1; 40Þ ¼ 9:008; p50:01; (F2(1,23) = 4.451, p < 0.05). For region 11, the ANOVA failed to yield a main effect for the embedded subject presence ðF1ð1; 40Þ ¼ 3:56; p40:06; ðF2ð1; 23Þ ¼ 1:17 p4:05Þ. A significant interaction was found by subject ðF1ð1; 40Þ ¼ 4:928; p50:04Þ,
16
L. Pablos
however it does not reach significance when the analysis is performed by items ðF2ð1; 23Þ ¼ 1:54; p4:05Þ. A post-hoc analysis at this region, by means of pairwise t-test, shows a significant effect in the match/null vs. match/overt comparison by items ðF2ð1; 23Þ ¼ 4:509; p50:05Þ as well as subjects F1(1,40) = 5:693; p50:02. The mismatch/null vs. mismatch/overt comparison (see Fig. 1.2) was not significant for region 11 (F1(1,40) = 0.045, p > 0.05; F 2 ð1; 23Þ ¼ 0:505; p > 0.05), or region 12 (F1(1,40) = 0.045, p > 0.05; F2(1,23) = 0.505, p > 0.05). Overall, these results show a garden-path effect in the (15a) and (15b) matching conditions that is absent in the mismatch conditions (15c) and (15d)5. This garden-path effect in the matching conditions reflects the fact that the parser has to reanalyze previous commitments (Frazier & Rayner, 1982; Sturt, Pickering, & Crocker, 1999). In (15a), the linear sequence of words and their case array matches the predictions established by the fronted ditransitive auxiliary, meaning that the parser presumably considers the dative and absolutive marked NPs as part of a single ditransitive clause when the first verbal head (idatzi ‘‘to write’’) is encountered (recall the discussion in Section 1.2.3 about case marking expectations in Basque). This is ostensibly due to the contrast between null and overt subjects. The overt subject in (15b) disambiguates the structure whereas in (15a) the null subject retains the ambiguity up to the embedded auxiliary verb position. The sentence in (16) shows the monoclausal structure that could be generated when the embedded verb is encountered and incorporated with previous arguments if no null subject is postulated. (16) Maitanek ez dio ikasleari ikastaroko txostena Maitane-ERG neg aux-ditrans student-DAT workshop’s document-ABS oso presaka idatzi. very fast write-ditrans. ‘Maitane has not written the document for the student in a hurry.’
It is not until the embedded monotransitive auxiliary (duela) is processed that the parser unambiguously determines that the embedded verb (idatzi ‘‘to write’’) is used monotransitively and that the structure it has generated so far is bi-clausal. The parser then has to infer that the previous dative and accusative arguments are not co-arguments of a single verb and it is forced to reanalyze the structure into the new structure shown by the brackets in (17). In this new structure the ABS-marked NP txostena ‘‘document’’ becomes part of the embedded clause and a null subject has to be postulated. This reanalysis is 5
Note that in the matching conditions the NP-dative following the main auxiliary (region 5 ikasleari) also showed a weak significant effect. This result is not expected since the sequence of words is the same up to this region in both conditions (i.e., Maitanek ez dio ikasleari) and it is taken to be caused by experimental conditions, not by the experimental design.
1 Rich Agreement in Basque: Evidence for Pre-verbal Structure Building
17
considered to be the reason why there is a slowdown at the embedded auxiliary in the matching null subject condition in (15a). (17) Maitanek ez dio Maitane-ERG neg aux.ditrans { [null subject insertion= pro]
ikasleari student-DAT ikastaroko txostena workshop’s document-ABS oso presaka idatzi du-ela}....[VERBAL HEAD EXPECTED] very fast write-monotrans aux.monotrasn-Comp ‘Maitane has not [EXPECTED VERB] the student that [NULL SUBJECT INSERTION = she/he] has written the document in a hurry.’
In sum, the evidence at the embedded verbal cluster in conditions (15a) and (15b) seems to indicate the parser’s attempt to attach each new word incrementally to the syntactic structure being built and to use any available local information (e.g., rich agreement at the auxiliaries) to connect and interpret as much input as possible, without needing to wait for the verbal head for confirmation. If the parser did not attempt to connect the structure previously processed, there should be no reason to expect a RT difference at these regions in the sentence. The possibility of being led down the garden-path only exists if the parser is trying to attach each word in the sentence incrementally. Results from the Basque experiment in Pablos (2006) showed that this is the case and that the garden-path effect cannot be explained unless an incremental parser is assumed (Crocker, 1994; Gorrell, 1995; Lombardo & Sturt, 2002; Stabler, 1994; Sturt & Crocker, 1996). Furthermore, the fact that Basque speakers do not show a garden-path at the embedded verb and auxiliary in sentences (15c) and (15d) suggests that they are using the agreement mismatch cue and that they are building an embedded clause as a consequence of this mismatch. This evidence also cannot be explained under models that lack anticipatory structure building (Abney, 1989; Mulders, 2003; Pritchett, 1992), since they would predict that the parser would not use the agreement mismatch as a cue for structure building and would still fall into a garden-path at the embedded clause in the mismatching conditions. Since this is not the case, the most reasonable assumption is that there is an incremental parser.
1.3 Conclusions In this chapter, I have discussed the different parsing models that have been proposed for the processing of head-final languages such as Basque, outlining the evidence in the literature that supports incremental models for Japanese and German, two other languages with head-final configurations.
18
L. Pablos
In particular, I have shown how the combination of rich agreement and case marking can be used to anticipate the upcoming syntactic structure and guide syntactic parsing. With this purpose in mind, I have examined how the match/mismatch between the agreement information encoded in Basque auxiliaries and the case information in the adjacent NP provides the parser with cues that help predict upcoming structure and avoid possible misanalyses of the sentence. In sum, results from this experiment on agreement in Basque support incremental processing models in head-final languages. These results show that the parser does not wait for the confirmation provided by heads to combine pre-verbal arguments,but rather uses indirect mismatch cues provided by the combination of agreement and case to postulate a clause boundary. In addition to providing evidence for the incremental nature of the parser in Basque, I have also shown how the information prior to the verb can be used to predict upcoming structure to avoid a garden-path. The data from fronted auxiliaries in Basque adds an additional dimension to the existing incremental literature, by making use of agreement in a different way from previous agreement comprehension studies and by predicting clause boundaries based on a combination of two types of information. I have shown how the agreement information at the auxiliary becomes crucial for determining clause boundaries and for predicting upcoming sentence structure. Critically, the prediction of a clause boundary is done based on a much more indirect cue than in previous studies since several factors have to be considered: the case information combination between the auxiliary and the adjacent NP, the case array of previous NPs and the specific agreement marked in the fronted auxiliary. All in all, it seems clear that the evidence from different head-final languages discussed in this chapter cannot be explained under head-driven models. These effects are best accounted for by assuming an incremental parser that tries to integrate every new word into the existing structure and does not wait for the confirmation provided by heads. Acknowledgements This work is part of a doctoral research grant awarded to the author by the Department of Research and Education of the Basque Government and from the Basque Government Research Project GIC07/144-IT-210-07. I would like to thank Nina Kazanina, Jon Ortiz de Urbina, Colin Phillips, J. Douglas Saddy, Juan Uriagereka, Amy Weinberg and Masaya Yoshida for their helpful comments at different stages of this work. All errors remain mine.
References Abney, S. (1989). A computational model of human parsing. Journal of Psycholinguistic Research, 18, 129–144. Aoshima, S., Phillips, C., & Weinberg, A. (2004). Processing long-distance dependencies in a head-final language. Journal of Memory and Language, 51, 23–54.
1 Rich Agreement in Basque: Evidence for Pre-verbal Structure Building
19
Aoshima, S., Yoshida, M., & Phillips, C. (2009). Incremental processing of coreference and binding in Japanese. Syntax, 12, 93–134. Bader, M. & Lasser, I. (1994). German verb-final clauses and sentence processing: evidence for immediate attachment. In C. Clifton, L. Frazier & K. Rayner (Eds.) Perspectives on Sentence Processing (225–242). Hillsdale, NJ: Erlbaum. Barber, H. & Carreiras, M. (2005). Grammatical gender and number agreement in Spanish: An ERP comparison. Journal of Cognitive Neuroscience, 17, 137–153. Boland, J. E., Tanenhaus, M. K., Garnsey, S. M., & Carlson, G. N. (1995). Verb argument structure in parsing and interpretation: Evidence from wh-questions. Journal of Memory and Language, 34, 774–806. Carreiras, M., Dun˜abeitia, J. A., Vergara, M., Zieza, I., & Laka, I. (2007). Object relative clause preference in Basque. AMLaP Conference, Turku, Finland. Clahsen, H., & Featherston, S. (1999). Antecedent priming at trace positions: evidence from German scrambling. Journal of Psycholinguistic Research, 28, 415–437. Coulson, S., King, J. W., & Kutas, M. (1998). Expect the unexpected: event-related brain response to morphosyntactic violations. Language and Cognitive Processes, 13, 21–58. Crocker, M. W. (1994). On the nature of the principle-based sentence processor. In C. Clifton, L. Frazier & K. Rayner (Eds.) Perspectives on Sentence Processing (245–266). Hillsdale, NJ: Erlbaum. Diaz, B., Sebastia´n-Galles, N., Erdozia, K., Mueller, J. L., & Laka, I. (2006). Individual Differences in the Syntactic Processing/Learning of Second Languages: An Erp Study. Granada, Spain: Congreso Espan˜ol de Psicofisiologı´ a. Dixon, R. M. W. (1994). Ergativity. Cambridge: Cambridge University Press. Dun˜abeitia, J. A., Perea, M., & Carreiras, M. (2007a). The role of frequency of constituents in compound words: evidence from Basque and Spanish. Psychonomic Bulletin and Review, 14, 1171–1176. Dun˜abeitia, J. A., Perea, M., & Carreiras, M. (2007b). Do transposed-letter similarity effects occur at a morpheme level? Evidence for morpho-orthographic decomposition. Cognition, 105 (3), 691–703. Erdozia, K. (2006). Processing Ambiguous Word Orders in Basque, An ERP Investigation. PhD Dissertation, University of the Basque Country. Etxepare, R. (2003). Valency and argument structure in the Basque verb. In J. I. Hualde & J. Ortiz de Urbina (Eds.) A Grammar of Basque (363–426). Berlin: Mouton de Gruyter. Frazier, L., & Rayner, K. (1982). Making and correcting errors during sentence comprehension: eye-movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14, 178–210. Garnsey, S. M., Pearlmutter, N. J., Myers, E., & Lotocky, M. A. (1997). The contributions of verb bias and plausibility to the comprehension of temporarily ambiguous sentences. Journal of Memory and Language, 37, 58–93. Gorrell, P. (1995). Syntax and Parsing. Cambridge: Cambridge University Press. Gutierrez Ziardegi, E. (2006). Ambiguity resolution in Basque-Spanish bilinguals. In B. Fernandez & I. Laka (Eds.) Andolin Gogoan. Essays in Honour of Professor Eguzkitza (417–434). Zarautz: EHU-ko Argitalpen Zerbitzua, Itxaropena. Hagoort, P., Brown, C. M., & Groothusen, J. (1993). The syntactic positive shift SPS as an ERP measure of syntactic processing. Language and Cognitive Processes, 8, 439–484. Hualde, J. I., & Ortiz de Urbina, J. (2003). A Grammar of Basque. Berlin: Mouton de Gruyter. Inoue, A. (1991). A Comparative Study of Parsing in English and Japanese. PhD Dissertation, University of Connecticut. Inoue, A., & Fodor, J. D. (1995). Information-paced parsing of Japanese. In R. Mazuka & N. Nagai (Eds.) Japanese Sentence Processing (9–63). Hillsdale, NJ: Erlbaum. Kaan, E. (2002). Investigating the effects of distance and number interference in processing subject-verb dependencies: An ERP study. Journal of Psycholinguistic Research, 31, 165–193.
20
L. Pablos
Kamide, Y. (2006). Incrementality in Japanese Sentence Processing. In M. Nakayama & S. Shirai (Eds.) Handbook of East Asian Psycholinguistics 2: Japanese (249–256). Cambridge: Cambridge University Press. Kamide, Y., & Mitchell, D. C. (1999). Incremental Pre-head attachment in Japanese parsing. Language and Cognitive Processes, 14, 631–662. Lombardo, V., & Sturt, P. (2002). Incrementality and lexicalism. In P. Merlo & S. Stevenson (Eds.) The Lexical Basis of Sentence Processing (137–156). Amsterdam: John Benjamins. MacDonald, M. C., Pearlmutter, N. J., & Seidenberg, M. S. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101, 676–703. Mazuka, R., & Itoh, K. (1995). Can Japanese speakers be lead down the garden-path? In R. Mazuka & N. Nagai (Eds.) Japanese Sentence Processing (295–329). Hillsdale, NJ: Erlbaum. Miyamoto, E. T. (2002). Case markers as clause boundary inducers in Japanese. Journal of Psycholinguistic Research, 31, 307–347. Miyamoto, E. T. (2003). Reanalysis and clause boundaries in Japanese as a constraint-driven process. Language and Speech, 461, 23–52. Miyamoto, E. T., & Takahashi, S. (2000). The processing of wh-phrases and interrogative complementizers in Japanese. In N. Akatuka & S. Strauss (Eds.) Japanese Korean Linguistics, 10, (62–75). Stanford, CA: CSLI. Miyamoto, E. T., & Takahashi, S. (2003). Typing mismatch effects in the processing of whphrases in Japanese. (Submitted for publication). Mulders, I. (2003). Transparent Parsing: Head-Driven Processing of Verb-Final Structures. PhD Dissertation, Utrecht Universiteit: LOT. Nakano, Y., Felser, C., & Clahsen, H. (2002). Antecedent priming at trace positions in Japanese long-distance scrambling. Journal of Psycholinguistic Research, 31, 531–571. Nevins, A., Dillon, B., Malhotra, S., & Phillips, C. (2007). The role of feature-number and feature-type in processing Hindi verb agreement violations. Brain Research, 1164, 81–94. Nicol, J. L., Forster, K. I., & Veres, C. (1997). Subject-verb agreement processes in comprehension. Journal of Memory and Language, 36, 569–587. Ortiz de Urbina, J. (2003). Word order. In J. I. Hualde & J. O. de Urbina (Eds.) A Grammar of Basque (448–459). Berlin: Mouton de Gruyter. Osterhout, L., & Mobley, L. A. (1995). Event-related brain potentials elicited by failure to agree. Journal of Memory and Language, 34, 739–773. Pablos, L. (2006). Pre-Verbal Structure Building in Romance Languages and Basque, PhD Dissertation. University of Maryland. Pablos, L., & Saddy, J. D. (2009). NPIs and complementizer agreement in Basque. In K. Alter, M. Horne, M. Lindgren, M. Roll & J. von Koss Torkildsen (Eds.) Papers from Brain Talk: Discourse with and in the Brain. The 1st Birgit Rausing Language Program Conference in Linguistics. Lund: Lund University, Media Tryck. Pearlmutter, N. J., Garnsey, S. M., & Bock, K. (1999). Agreement processes in comprehension. Journal of Memory and Language, 41, 427–456. Perea, M., Dun˜abeitia, J. A., & Carreiras, M. (2008). Masked associative/semantic priming effects across languages with highly proficient bilinguals. Journal of Memory and Language, 58, 916–930. Pritchett, B. (1992). Grammatical Competence and Parsing Performance. Chicago: The University of Chicago Press. Saltarelli, M. (1988). Basque. London: Croom Helm. Santesteban, M. (2006). Lexical Representation and Selection on Bilingual Speech Production. PhD Dissertation, University of Barcelona. Santesteban, M., & Costa, A. (2006). Does L1 syntax affect L2 processing? A study with highly proficient early bilinguals. In B. Fernandez & I. Laka (eds.) Andolin Gogoan. Essays in honour of Professor Eguzkitza (817–834). Zarautz: EHU-ko Argitalpen Zerbitzua, Itxaropena.
1 Rich Agreement in Basque: Evidence for Pre-verbal Structure Building
21
Stabler, E. P. (1994). The finite connectivity of linguistic structure. In C. Clifton, L. Frazier & K. Rayner (Eds.) Perspectives on Sentence Processing (303–336). Hillsdale, NJ: Erlbaum. Sturt, P., & Crocker, M. M. (1996). Monotonic syntactic processing: A cross-linguistic study of attachment and reanalysis. Language and Cognitive Processes, 11, 449–494. Sturt, P., Pickering, M., & Crocker, M. M. (1999). Structural change and reanalysis difficulty in language comprehension. Journal of Memory and Language, 40, 136–150. Thornton, R., & McDonald, M. C. (2003). Plausibility and grammatical agreement. Journal of Memory and Language, 48, 740–759. Trueswell, J. C., Tanenhaus, M. M., & Kello, C. (1993). Verb-specific constraints in sentence processing: separating effects of lexical preference from garden-paths. Journal of Experimental Psychology: Learning, Memory and Cognition, 19, 528–553. Yamashita, H. (1997). The effects of word-order and case marking information on the processing of Japanese. Journal of Psycholinguistic Research, 26, 163–188. Zawiszewski, A. (2007). Basque Bilinguals Processing Syntax: Case and Agreement as Revealed by ERPs. PhD Dissertation, University of the Basque Country.
Chapter 2
The Processing of Japanese Control Sentences Jeffrey D. Witzel and Naoko O. Witzel
2.1 Introduction Japanese is a head-final language that allows relatively free movement of constituents (through scrambling) and permits null arguments in virtually any structural position (for a review of null arguments in Japanese, see Zushi, 2003; for more on the structural characteristics of Japanese, see Shibatani, 1990; Tsujimura, 1996, 1999). These three aspects of Japanese – its head-finality, free word order and abundant empty elements – pose challenges for models of sentence processing (Berwick & Fong, 1995; Inoue & Fodor, 1995; Mazuka & Nagai, 1995; Nakayama, 1999). Indeed, these characteristics force sentence processing models to explain how the parser,1 or structural computation mechanism, interprets displaced constituents and empty structural elements when access to phrasal heads is delayed. The present study examines the processing of Japanese Obligatory Control sentences – sentences that highlight the ‘‘problem’’ of interpreting empty structural elements in a head-final language. Specifically, this study investigates the processing of empty embedded clause subjects in these sentences. Because Japanese is a strongly head-final language, the Control information necessary to specify the referent of this empty element is often not available until the end of the sentence, at the Control verb. Under incremental sentence processing frameworks (see e.g., Inoue & Fodor, 1995; Kamide, 2006; Miyamoto, 2002), available information is used to generate the most complete structural/interpretive representation of the sentence as it unfolds from beginning to end. We hypothesized that if incrementality applies to the gap-filling operation involved J.D. Witzel (*) Department of Linguistics and TESOL, College of Liberal Arts, University of Texas at Arlington, Arlington, Texas 76019, USA e-mail:
[email protected] 1
The term parser refers to the mechanism that generates structural interpretation, including the association of constituents as well as the assignment of coreference, during online sentence comprehension.
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_2, Ó Springer ScienceþBusiness Media B.V. 2011
23
24
J.D. Witzel and N.O. Witzel
in Japanese Obligatory Control sentences, then a referent should be provisionally assigned to the empty embedded clause subject in these sentences.
2.1.1 Control and Head Directionality In order to understand the phenomenon of interest, consider the following English sentences (example (1) is drawn from Haegeman (1994, p. 279)): (1) Johni promised Maryj [PROi to behave himself / *herself / *oneself.] (2) Johni persuaded Maryj [PROj to behave *himself / herself / *oneself.] Each sentence consists of a main, or matrix, clause and an embedded nonfinite clause (in brackets). According to generative syntactic analyses, the embedded clause includes the empty external argument PRO, which is present to satisfy the requirements of the Extended Projection Principle (EPP) and the Theta Criterion (Chomsky, 1981, 1982; Chomsky & Lasnik, 1977). In both sentences, antecedents are assigned to PRO through Obligatory Control. That is, PRO is referentially dependent on another noun phrase (NP) in the sentence. In the Subject Control (SC) sentence (1), the referential properties of PRO are ‘‘controlled’’ by the matrix subject (John). In the Object Control (OC) sentence (2), on the other hand, the matrix object (Mary) controls the empty embedded clause subject. These referential restrictions on PRO are illustrated by the grammaticality/ungrammaticality of the reflexives in the embedded clause of each example sentence. Specifically, this reflexive must match the number, person and gender of its binder PRO, which in turn receives these referential properties from its controller in the main clause. Finally, as shown in these examples, whether a sentence is SC or OC depends largely on the matrix verb – for example, promise in the SC sentence (1) and persuade in the OC sentence (2). That is, these verbs contain information about Control type (presumably in their lexical entries) that determines the referent for the empty embedded clause subject (for more on Control, see Davies & Dubinsky, 2004). With respect to the processing of these sentences, the important thing to notice is that in a head-initial language like English, Control information is often available prior to the appearance of the embedded clause. Therefore, it is reasonable to assume that during the online processing of Control sentences in such languages, Control information in the matrix verb can be accessed to establish the antecedent of the empty embedded clause subject as soon as this clause is encountered (Betancort, Carreiras, & Acun˜a-Farin˜a, 2006; Boland, Tanenhaus, & Garnsey, 1990; Crain & Fodor, 1985; but see also Bever & McElree, 1988; McElree & Bever 1989; Nicol & Swinney, 1989). The same assumption cannot be made with regard to the processing of Japanese Control sentences. Indeed, because Japanese is a strongly head-final
2 The Processing of Japanese Control Sentences
25
language, Control information provided by the verb is not available until late in the sentence. Consider the following Japanese SC sentences, (3a) and (3b), and OC sentences, (4a) and (4b), along with their glosses and translations (these sentences are based on Sakamoto’s (1996, p. 24) examples (33a) and (33b); embedded clauses are bracketed; empty subjects are indicated with ‘‘e’’2): (3) a. Jirooi-ga senshuu Kazumij-ni [ei/j* Tookyoo iki]-o Jirooi-NOM last week Kazumij-DAT [ei/j* Tokyo going]-ACC wazato kakushiteita. purposely concealed. b. Jirooi-ga senshuu Kazumij-ni [ei/j* Tookyoo iku koto]-o Jirooi-NOM last week Kazumij-DAT [ei/j* Tokyo-to go fact]-ACC wazato kakushiteita. purposely concealed. ‘Last week, Jiroo purposely concealed his traveling to Tokyo from Kazumi.’ (4) a. Jirooi-ga senshuu Kazumij-ni [ei*/j Tookyoo iki]-o Jirooi-NOM last week Kazumij-DAT [ei*/j Tokyo going]-ACC wazato tanonda. purposely requested. b. Jirooi-ga senshuu Kazumij-ni [ei*/j Tookyoo iku koto]-o Jirooi-NOM last week Kazumij-DAT [ei*/j Tokyo-to go fact]-ACC wazato tanonda. purposely requested. ‘Last week, Jiroo purposely requested of Kazumi that she go to Tokyo.’ The (a) and (b) versions of these sentences have essentially the same meaning but differ in terms of their embedded clauses – sentences (3a) and (4a) have gerundive embedded clauses, while (3b) and (4b) involve embedded clauses that are nominalized with koto. Again, in the SC sentences (3a) and (3b), the antecedent of the empty embedded clause subject is the matrix subject (Jiroo). In the OC sentences (4a) and (4b), the antecedent of the empty subject is the object (or dative-marked NP) of the main clause (Kazumi). As in their English counterparts, the matrix verbs in these Japanese Control sentences – kakushiteita (‘‘concealed’’) in (3) and tanonda (‘‘requested’’) in (4) – contain the information necessary to specify the referent of the empty embedded clause subject.
2 The precise nature of this empty element – that is, whether it is pro, PRO, or a trace-like element – will remain unspecified in this paper (see Sakamoto (1996) and references therein for more on the syntax of Japanese Control).
26
J.D. Witzel and N.O. Witzel
However, unlike in English, information about this antecedent is not available until after the embedded clause. Given the structure of Japanese Control sentences, the safest way for the parser to analyze the empty embedded clause subject would be to delay interpretation until information in the Control verb becomes available. At that point, a full interpretation of the sentence could be established, complete with an empty embedded clause subject and the correct filler for that subject. A purely bottom-up, head-driven parser (a la Pritchett, 1991) would presumably delay the assignment of referents in this way. A relevant question then is whether this type of language processor is viable for and, more importantly, evinced in Japanese sentence comprehension. There is good reason to think that it is not.
2.1.2 Incrementality and the Processing of Japanese Control Sentences Theoretical arguments against head-driven parsing in Japanese tend to focus on the processing costs that such a mechanism would entail for even the most straightforward sentences (see e.g., Inoue & Fodor, 1995; Kamide, 2006; Miyamoto, 2002). Indeed, such a parser would leave the relationships among sentence constituents undefined often until the end of the sentence. Seemingly excessive processing resources would be required in order for these constituents to be held in memory until they could receive complete structural interpretation. These arguments are bolstered by experimental evidence demonstrating that rich structural/interpretive decisions are made before phrasal heads are encountered (see e.g., Aoshima, Phillips, & Weinberg, 2003, 2004; Kamide & Mitchell, 1999; Kamide, Altmann, & Haywood, 2003; Miyamoto, 2002). Such findings suggest a pre-head, incremental structure-building mechanism for Japanese. Although a complete review of this evidence is beyond the scope of this paper, two sets of findings related to incremental parsing in Japanese are germane to the present study: (a) results demonstrating the importance of case marking for establishing clause boundaries and (b) those showing the time course of gap filling during the processing of sentences involving displaced constituents. With regard to the first of these sets of findings, Miyamoto (2002) showed that during online Japanese sentence comprehension, relationships among constituents are determined before all of the relevant information from phrasal heads becomes available. Specifically, in three self-paced reading experiments, Miyamoto demonstrated that clause boundaries are established when case marking prevents a single clause interpretation – or, when case marking indicates that NPs cannot be coarguments of the same predicate. For example, Miyamoto’s Experiment 2 tested the following sentence types:
2 The Processing of Japanese Control Sentences
27
(5) a. Obasan-ga yoboyobono toshiyori-o guuzen-ni koosaten-de woman-NOM feeble old man-ACC by chance intersection-at mita takushii-ni isoide noseta. saw taxi-DAT hurry put. ‘The woman hurriedly put the feeble old man in the taxi that she saw at the intersection by chance.’ b. Obasan-ga yoboyobono toshiyori-o gakusei-ga koosaten-de woman-NOM feeble old man-ACC the student intersection-at mita takushii-ni isoide noseta. saw taxi-DAT hurry put. ‘The woman hurriedly put the feeble old man in the taxi that the student saw at the intersection.’ Sentence (5b) begins with the constituents NP1-nom NP2-acc NP3-nom. Because no Japanese verb takes three arguments case marked in this way, at the incompatible constituent NP3-nom, it is clear that the sentence involves more than one predicate/clause. In sentence (5a), however, the relative clause head takushii-ni (taxi-dat) is the first point at which a single clause analysis is revealed to be incorrect. Although reading times (RTs) were longer at NP3-nom in sentences like (5b) relative to the same region in sentences like (5a), RTs at the matrix verb were faster in (5b) than in (5a). This pattern of results suggests that while processing difficulty occurs when a clause boundary is induced by an incompatible argument, establishing this multi-clause interpretation early in the sentence seems to obviate the need for reanalysis at the matrix clause verb. These results therefore indicate the pre-head generation of clause boundaries based on local information provided by NP case markers (for comparable results, see Kamide & Mitchell, 1999; Kamide et al., 2003). Furthermore, again in a series of self-paced reading experiments, Aoshima and colleagues (2004; see also Aoshima et al., 2003) provided evidence for the pre-head assignment of displaced constituents to potential gap sites in the online processing of scrambled Japanese sentences. The following scrambled (6a) and canonically-ordered (6b) sentence types were among those tested: (6) a. Dono-shain-ni senmu-wa (__) shachoo-ga kaigi-de which employee-DAT managing director-TOP president-NOM meeting-at kachoo-ni shookyuu-o yakusokushita-to iimashita-ka? assistant director-DAT raise-ACC promised-DeclC told-Q ‘To which employee did the managing director tell that the president promised a raise to the assistant director at the meeting?’ b. Dono-shain-ga senmu-ni (__) shachoo-ga kaigi-de which employee-NOM managing director-TOP president-NOM meeting-at kachoo-ni shookyuu-o yakusokushita-to iimashita-ka?
28
J.D. Witzel and N.O. Witzel
assistant director-DAT raise-ACC promised-DeclC told-Q ‘Which employee told the managing director that the president promised a raise to the assistant director at the meeting?’ The scrambled sentence (6a) involves a sentence-initial, dative-marked wh-constituent that is displaced from its canonical position (indicated with parentheses just after senmu-wa). The canonically-ordered sentence (6b) has the same wh-constituent in sentence-initial position but, in this instance, it is marked with nominative case. Based on this case marking, it is clear that this wh-constituent is in its canonical position. It was predicted that the (ultimately erroneous) association of the scrambled wh-constituent with a gap position in the embedded clause would yield a filled-gap effect (Crain & Fodor, 1985; Stowe, 1986) at the dative-marked NP in this clause, kachoo-ni. Consistent with this prediction, longer RTs were obtained at this NP in scrambled sentences than in canonical sentences. The authors interpret these results in support of a Full Constraint-Driven model of processing filler-gap dependencies. According to this model, a filler initiates a search for the first gap position that resolves its thematic dependencies. Given that this gap position can be identified without reference to information provided by verbal heads, this filler-gap parsing operation is consistent with the incremental processing of Japanese sentences. In light of this evidence for incremental parsing in Japanese, it is at least plausible that the structural computation mechanism would attempt to generate the most complete representation possible for Japanese Control sentences prior to accessing Control information in the matrix verb. The relevant questions, then, are as follows: (a) Is an antecedent provisionally assigned to the empty embedded clause subject in Japanese Control sentences before Control information becomes available at the matrix verb? (b) If so, which NP – the matrix subject or object – is assigned as this antecedent? A positive answer to the first of these questions would provide further support for incremental parsing in Japanese, while a clear answer to the second would shed light on the types of information that contribute to this incremental processing. In fact, several studies to date have examined these questions regarding the processing of Japanese Control sentences (Ninose, Oda, Sakaki, Sakamoto, & Gyoba, 1998; Oda, Ninose, Sakaki, Gyoba, & Sakamoto, 1997; Sakamoto, 1995, 1996; see Sakamoto, 2002, 2006, for review). Although all of these studies have indicated a bias in the interpretation of the empty subject in these sentences, the nature of this bias has proven to be task-dependent. Specifically, experiments using end-of-sentence antecedent retrieval tasks have revealed an OC bias, whereas experiments using end-of-sentence antecedent recognition tasks have yielded results consistent with an SC bias. The first of these studies (reported in Sakamoto, 1995, 1996) tested gerundive and koto-clause SC/OC sentences (again, see example sentences (3a), (3b), (4a), and (4b)) as well as scrambled versions of the gerundive sentences with an
2 The Processing of Japanese Control Sentences
29
end-of-sentence antecedent retrieval task. The scrambled sentences were generated by moving the dative-marked NP from its canonical position to sentence-initial position (e.g., Kazumi-ni Jiroo-ga. . ..). Each sentence in these experiments (test and filler) expressed the idea that a specific person went to Tokyo (as in the examples above). After listening to each sentence, participants had to say who went to Tokyo as quickly and accurately as possible. For the Control sentences, this effectively meant that participants had to provide the antecedent of the empty embedded clause subject. For all three Control sentence types, subjects were faster and more accurate when producing the antecedent of the empty subject for OC sentences than for SC sentences. These results were taken to indicate an OC bias in the interpretation of Japanese Control sentences, regardless of whether the object occurred in canonical or noncanonical position. Subsequent experiments by Oda et al. (1997) and Ninose et al. (1998) investigated the processing of Japanese Control sentences using an end-ofsentence antecedent recognition task. The first of these experiments (Oda et al., 1997) tested canonical gerundive Control sentences, while the second (Ninose et al., 1998) examined scrambled versions of these sentences. In both experiments, subjects listened to Control sentences in one ear and, after the offset of each, heard the name of a potential referent for the empty subject (either the matrix subject or object) in the other ear. This target was presented at one of six probe points after the sentence: 0, 300, 600, 900, 1200 and 1500 ms. Again, the embedded clauses in these sentences expressed the idea that a specific person went to Tokyo. Therefore, when presented with a name, subjects had to indicate (as quickly and accurately as possible) whether this person went to Tokyo with a yes-no, push-button response. In a departure from Sakamoto’s (1995, 1996) previous Japanese Control studies, these experiments did not include any filler items. Rather, subjects were exposed only to 12 items of each Control type (in contrast to 24 of each type in Sakamoto’s earlier experiments), with each item presented 12 times during the experiment – once at each probe point with each yes/no response. The results showed that yes responses were made more quickly and accurately on SC sentences than on OC sentences at the 0 ms probe point. No other probe point revealed significant response time differences. Taken together, these experiments seem to reveal an SC bias in the interpretation of canonical and scrambled Japanese Control sentences – again, exactly the opposite of the bias reported in the earlier antecedent retrieval experiments. However, it is important to ask whether any of these experiments are capable of revealing an interpretive bias in the online processing of Japanese Control sentences. First, all of these experiments rely on the interpretation of data from end-of-sentence tasks. Most studies investigating antecedent reactivation, however, use intrasentential probe points (see e.g., McElree & Bever, 1989; Nicol & Swinney, 1989). The reason for this is clear: Only through the use of intrasentential probes can one determine the time course of antecedent reactivation. End-of-sentence tasks are only able to provide (at best) a rough indication of
30
J.D. Witzel and N.O. Witzel
this time course and, thus, are inadequate for the examination of real-time sentence processing. Furthermore, each of these experiments required subjects to search for the same information in the sentences of interest (i.e., who went to Tokyo). This aspect of the experimental design is particularly striking in the antecedent recognition experiments (Ninose et al., 1998; Oda et al., 1997), in which each sentence was presented 12 times with no filler sentences to distract subjects from the sentence types of interest. Such information searches are undesirable requirements in sentence processing experiments, which are ideally designed to preclude the use of strategic processing.
2.2 The Experiment: Rationale, Sample Items and Predictions 2.2.1 Rationale As discussed above, in online Japanese sentence comprehension, relationships among constituents and interpretations of displaced elements are provisionally determined before relevant phrasal heads are encountered. Experimental evidence to date also suggests that this incrementality applies to the gap filling operation in Japanese Control sentences (Ninose et al., 1998; Oda et al., 1997; Sakamoto, 1995, 1996; reviewed in Sakamoto, 2002, 2006). That is, experimental studies of Japanese Control have indicated (a) that an antecedent is assigned to the empty embedded clause subject before Control information becomes available and (b) that this process is guided by a bias for positing a specific main clause NP as this antecedent. However, because these experiments have yielded conflicting findings as to the nature of this bias (i.e., some experiments indicate an SC bias, while others suggest an OC bias), finer-grained tests of the online processing of Japanese Control sentences are necessary. Therefore, the present experiment examined these sentences using a self-paced, region-byregion reading task – a paradigm that is considered sensitive to online structure building.
2.2.2 Sample Test Sentences Table 2.1 provides examples of the four sentence types tested in this experiment: Subject Control sentences with empty embedded clause subjects (SC/empty), Subject Control sentences with overt embedded clause subjects (SC/overt), Object Control sentences with empty embedded clause subjects (OC/empty) and Object Control sentences with overt embedded clause subjects (OC/overt). The self-paced reading task used in this experiment (described in detail below) required participants to read each sentence one region at a time. For the sake of clarity, the presentation regions for these example sentences are divided into separate cells and numbered 1–8. Above each example sentence, a schematic
NPj-DAT
Region 2
Region 3 EC* constituent
EC verb + koto ACC
Table 2.1 Example test sentences Region 4 Region 5
NPj-DAT
reflexivei-NOM
EC constituent
EC verb +koto ACC
NPj-DAT
EC constituent
EC verb +koto ACC
NPj-DAT
reflexivei-NOM
ECconstituent
EC verb +koto ACC
Daisuke-ga Kyoko-ni karejishin-ga kuruma-o koonyuu suru koto-o Daisuke-NOM Kyoko-DAT he-himself-NOM car-ACC purchase fact-ACC ‘It seems that Daisuke showed off to Kyoko in the office kitchen that he would purchase a car’ * EC stands for Embedded Clause
b. NPi-NOM
Daisuke-ga Kyoko-ni kuruma-o koonyuu suru koto-o Daisuke-NOM Kyoko-DAT car-ACC purchase fact-ACC ‘It seems that Daisuke showed off to Kyoko in the office kitchen that he would purchase a car’
(8) a. NPi-NOM
Daisuke-ga Kyoko-ni karejishin-ga kuruma-o koonyuu suru koto-o Daisuke-NOM Kyoko-DAT he-himself-NOM car-ACC purchase fact-ACC ‘It seems that Daisuke showed off to Kyoko in the office kitchen that he would purchase a car’
b. NPi-NOM
Daisuke-ga Kyoko-ni kuruma-o koonyuu suru koto-o Daisuke-NOM Kyoko-DAT car-ACC purchase fact-ACC ‘It seems that Daisuke showed off to Kyoko in the office kitchen that he would purchase a car’
(7) a. NPi-NOM
Region 1
kyuutoo-shitsu de office kitchen in
adverbial
kyuutoo-shitsu de office kitchen in
adverbial
kyuutoo-shitsu de office kitchen in
adverbial
kyuutoo-shitsu de office kitchen in
adverbial
Region 6
evidential + desu yoo desu seems
evidential + desu yoo desu seems
Region 8
evidential + desu saisokushita yoo desu urged seems
OC verb
evidential + desu saisokushita yoo desu urged seems
OC verb
jimanshita showed off
SC verb
jimanshita showed off
SC verb
Region 7
2 The Processing of Japanese Control Sentences 31
32
J.D. Witzel and N.O. Witzel
representation indicates the structural content of each region. A gloss and translation are also provided below each example. In each sentence, region 1 was the matrix subject (NP-nom), and region 2 was the matrix object (NP-dat). As shown in the example sentences, these NPs contrasted with each other in terms of gender (e.g., Daisuke is a male name; Kyoko is a female name). In SC/empty and OC/empty sentences, region 3 was simply skipped. In SC/overt and OC/overt sentences, on the other hand, this region provided a gender-marked reflexive (either karejishin (‘‘he-himself’’) or kanojojishin (‘‘she-herself’’)) as the overt subject of the embedded clause. This reflexive specified either the matrix subject or object – again, one being male and the other being female – as the referent of the embedded clause subject. Regions 4 and 5 contained the (remaining) constituents of the embedded clause. Region 4 was either an argument or adjunct of this clause, while region 5 was made up of the embedded clause verb, the nominalizer – koto and the accusative case marker –o. Regions 6, 7 and 8 contained the remaining constituents of the matrix clause. Region 6 was an adverbial phrase; region 7 contained the Control verb – an SC verb for SC sentences or an OC verb for OC sentences; and region 8 consisted of an evidential and the copula desu.
2.2.3 Predictions The predictions for these sentences hinge on the presence/absence of an overt subject in region 3. Indeed, this overt subject has several important structural/ interpretive implications. First, this element yields the series NP1-nom NP2-dat NP3-nom at the beginning of SC/overt and OC/overt sentences. Because no Japanese verb takes three arguments case marked in this way, these NPs cannot be coarguments of the same predicate. Thus, consistent with Miyamoto (2002), the NP in region 3 should allow the parser to begin building a multi-clause interpretation of the sentence. In sentences that lack this overt subject (SC/ empty and OC/empty sentences), however, the presence of an embedded clause was not indicated until the last region of this clause, region 5. It is also clear at this point that an argument is missing from the subject position of the embedded clause in these sentences. In SC/overt and OC/overt sentences, on the other hand, region 5 should merely provide the right boundary for a previously-initiated embedded clause. It was predicted that the structural reanalysis required for empty subject sentences in region 5 would incur measurable processing costs. SC/empty and OC/empty sentences were therefore predicted to have longer RTs than SC/overt and OC/overt sentences in this region. RT differences were also predicted at the Control verb, in region 7. Recall that the NP in region 3 of overt subject sentences was a reflexive element that unambiguously specified either the matrix subject or object as its referent. Thus, in SC/overt and OC/overt sentences, the referential properties of the embedded clause subject could be established well before (and without reference to) the
2 The Processing of Japanese Control Sentences
33
Control verb. The SC/empty and OC/empty sentences, however, required information from the Control verb in order to establish the referent of the embedded clause subject. If there is a bias for assigning a specific NP as the antecedent of the empty embedded clause subject in these sentences, Control verbs that are inconsistent with this bias should initiate reanalysis and, thus, should incur processing costs. Specifically, if there is an SC bias, OC verbs in OC/empty sentences should have longer RTs than in OC/overt sentences. If, on the other hand, there is an OC bias, then SC verbs in SC/empty sentences should have longer RTs than in SC/overt sentences. Alternatively, if the assignment of antecedents to empty embedded clause subjects is delayed until Control information becomes available, RTs for Control verbs in both OC/empty and SC/empty sentences should be longer than in their overt subject counterparts.
2.3 Methods 2.3.1 Participants Sixty native speakers of Japanese participated in the experiment. The participants were undergraduate students, graduate students and staff members at the University of Tokyo and the University of Arizona.
2.3.2 Materials Forty-eight sets of experimental sentences were created. The sentences in each set were similar to those in Table 2.1.3 These items were constructed from 24 SC verbs and 24 OC verbs drawn from those used in Sakamoto (1995, 1996). These were verbs that had been consistently analyzed as belonging to their respective Control types in the norming tasks and experiments described in these earlier studies.4 Before creating these items, eight Japanese native speakers were asked to compose sentences using each of the verbs in a frame very similar to that which was planned for the experimental sentences (NP-ga/NP-ni/clause-o/ Control Verb). These sentences were used (a) to verify that each verb could be used in this frame, (b) to identify any collocational uses of these verbs that should be avoided and (c) to provide natural base sentences from which to generate the more tightly constrained experimental items. 3
The complete set of experimental sentences is available on the publisher’s website (or will be provided upon request to the first author). 4 One verb in this experiment did not come from those used in Sakamoto’s earlier work – settoku shita ‘persuaded’. Although not used in Sakamoto’s studies, this verb shows up regularly in the syntax literature related to Japanese Control.
34
J.D. Witzel and N.O. Witzel
The following factors were controlled in the experimental item sets: As illustrated in the example sentences, regions 1 and 2 contained NPs that contrasted in terms of gender. The gender of these NPs was indicated either with names (as in the example sentences) or with gender-specific NPs such as hahaoya ‘‘mother’’ and musuko ‘‘son’’. Thirty-six items had names for these NPs, while 12 involved gender-specific NPs. The position of these masculine and feminine NPs was balanced such that each occurred with equal frequency in subject and object position. This also ensured that the two gender-marked reflexives, karejishin (‘‘he-himself’’) and kanojojishin (‘‘she-herself’’), in region 3 of overt subject sentences occurred with equal frequency across test sentences and across sentences of the two Control types. The embedded clause regions 4 and 5 were constructed such that these clauses varied in terms of their semantic content across items, thus precluding the use of strategic searches for information during the reading task. Region 6 consisted of a matrix clause adverbial. Time, location and manner adverbials were used in this region, with each type roughly equally represented across the item sets. Region 7 consisted of the Control verb. Each experimental item set involved two Control verbs, one SC verb and one OC verb, with each verb used in two different item sets. Finally, region 8 consisted of one of four evidential elements, rashii, mitai, yoo and soo (all of which mean something roughly equivalent to ‘‘seems’’), followed by the copula desu. The four evidentials were balanced across the item sets, such that each occurred in 12 sets. From the 48 sets of experimental sentences, eight counterbalanced lists of items were compiled for the reading task. These lists were generated as follows: First, four lists of items were created, each of which included 12 items from each sentence condition, and one instance of each Control verb. Two versions of each of these lists were then created. In one version, half of the experimental items (i.e., 24 items) were followed by comprehension questions; while the second version included comprehension questions for the other half of the experimental items. In this way, there was a comprehension question for each condition in the experimental sentence sets across the eight item lists. As in previous experiments on Japanese Control, the comprehension questions for these sentences required participants to identify the subject of the embedded clause. However, in contrast to these earlier studies, the questions varied depending on the content of the embedded clause (i.e., the question was not always about who went to Tokyo). Each list also included 80 filler sentences, 32 of which were followed by comprehension questions. Items were presented in pseudorandom order such that (a) an experimental item was never immediately followed by another experimental item from the same condition and (b) there were never more than three comprehension questions in succession. Based on the experimental item sets, an offline ‘‘naturalness’’ judgment questionnaire was also created. Four versions of this questionnaire were constructed, each of which included 12 empty subject sentences from each Control condition (SC and OC), for a total of 24 experimental sentences. Each list also contained 24 of the filler items in the experiment. These items were presented in
2 The Processing of Japanese Control Sentences
35
pseudorandom order such that there were no consecutive items from the same experimental condition. Next to each item, there was a 5-point (Likert-style) naturalness scale.
2.3.3 Procedures The experiment employed a region-by-region, self-paced, non-cumulative moving-window reading task (Just, Carpenter, & Woolley, 1982) presented on laptop computers using DMDX software. For each item, the subject performed as follows: First, the subject pressed the space bar to begin the item. The initial display was a string of asterisks, with each asterisk representing a single character in the sentence. (Each sentence was of course presented in standard Japanese script – that is, in a combination of kanji and kana.) The subject then pressed the right shift key on the keyboard to reveal the first region of the sentence. After reading this region, the subject again pressed the right shift key. The first region would then disappear and the next region would be presented. The subject would continue through the remaining regions of the sentence in the same manner. Fifty-six items (24 experimental items and 32 fillers) were followed by comprehension questions. Under each comprehension question, there were two answer choices – one on the left side of the screen and another on the right. If the correct answer was on the left, the subject pressed the left shift key. If the correct answer was on the right, the subject pressed the right shift key. Before beginning this task, subjects read a set of instructions and were given 10 practice items. After finishing the reading task, subjects completed the off-line (paper-and-pencil) naturalness questionnaire. On this questionnaire, the subject rated each sentence on a five-point naturalness scale – ‘‘1’’ for an unnatural sentence and ‘‘5’’ for a natural sentence. The experimental session lasted around 30 min, with 15–20 min for the reading task and about 10 min for the questionnaire.
2.4 Results The data from participants (a) with error rates (ERs) of 30% or greater on the complete set of comprehension questions (i.e., on the questions for filler and experimental items) or (b) with ERs of 30% or greater on the comprehension questions for the Control sentences (i.e., on the questions for experimental items only) were eliminated from the analysis. The data sets of 12 participants were excluded based on these cut-off scores. Therefore, analyses were conducted on the data from 48 participants (with six participants for each of the eight counterbalanced lists). After these subjects were removed, the data for Control sentence items (a) with total ERs greater than 33.33% or (b) with ERs greater than 33.33% in any of the four experimental conditions (SC/empty,
36
J.D. Witzel and N.O. Witzel
SC/overt, OC/empty, OC/overt) were removed from the analysis. The data from 14 items were excluded based on these criteria. Therefore, analyses were conducted on the data from 34 Control sentence items.5
2.4.1 Reading Time (RT) Analyses The RT data for items on which comprehension questions were answered incorrectly were eliminated from the analysis. The RT data for items on which a participant spent 4 s or longer reading any region were also excluded (accounting for 0.17% of the remaining data points). Outlier RTs were then trimmed as follows: RTs that were two standard deviations above or below a subject’s mean RT for a given region were replaced with the value two standard deviations above or below the subject’s mean RT for that region. 4.38% of the data were trimmed in this way. Table 2.2 provides the mean RTs for each region of the experimental sentences. These means are also illustrated in Fig. 2.1. Table 2.3 provides the mean combined RTs for the experimental sentences. These combined RTs were calculated by adding the RTs of regions 4–8 for each item after outlier data points had been trimmed. In the statistical analyses of the RT data, both subjects and items were treated as random factors. The by-subjects analyses (F1) consisted of an 8 (22) ANOVA for each region, with List as a non-repeated factor (to remove the variability associated with the counterbalancing procedures), Control type (SC, OC) and embedded clause type (empty, overt) as repeated measures and RT as the dependent variable (DV). The by-items analyses (F2) consisted of a 22 ANOVA for each region, with Control type (SC, OC) and embedded clause type (empty, overt) as repeated measures and RT as the DV. The same by-subjects (F1) and by-items (F2) analyses were conducted on the combined RTs. Significant interactions of the repeated measures in these analyses were followed up with tests of the simple effect of embedded clause type for each Control type. Because region 3 occurred only in overt subject sentences (SC/ overt and OC/overt), embedded clause type (empty, overt) was not a relevant factor in the analyses for this region. Therefore, the by-subjects analysis (F1) for region 3 consisted of an 8(2) ANOVA with List as a non-repeated factor, Control type (SC, OC) as a repeated measure and RT as the DV. The by-items analysis (F2) consisted of a one-way ANOVA with Control type (SC, OC) as a repeated measure and RT as the DV. 5
The data from participants and items were excluded in these ways in order to conform to accepted standards in the self-paced reading literature. It is important to note, however, that the pattern of results for all participants and items was virtually the same as the pattern reported in the Results section for the subset of participants and items that met the inclusion criteria.
700
661
overt
Kyoko-ni
empty
NPj-dat
NPi-nom
Daisuke-ga
OC
709
720
726
685
overt
Kyoko-ni
738
Daisuke-ga
685
NPj-dat
NPi-nom
empty
SC
Region 2
Region 1
856
702
737
EC constituent kuruma-o
(_)/reflexivei-nom
667
(_)/karejishin-ga
820
kuruma-o
(_)/karejishin-ga 706
EC constituent
(_)/reflexivei-nom
625
692
koonyuusuru koto-o
EC verb + koto acc
618
677
koonyuusuru koto-o
EC verb + koto acc
530
552
kyuutoo-shitsu de
adverbial
506
542
kyuutoo-shitsu de
adverbial
Table 2.2 Mean RT (in milliseconds) per region of experimental sentences Region 3 Region 4 Region 5 Region 6
evidential + desu
492
553
yoo desu
evidential + desu
Region 8
514
511
527
541
saisokushita yoo desu
OC verb
476
547
jimanshita
SC verb
Region 7
2 The Processing of Japanese Control Sentences 37
38
J.D. Witzel and N.O. Witzel 900 850 800
Mean RT
750 SC/empty
700
SC/overt
650
OC/empty
600
OC/overt
550 500 450 400 Region Region Region Region Region Region Region Region 1 2 3 4 5 6 7 8
Fig. 2.1 Mean RT (in milliseconds) per region of experimental sentences
Table 2.3 Mean combined RTs (in milliseconds) for experimental sentences Control type Embedded clause type Mean combined RT Subject control (SC)
Empty subject (empty) Overt subject (overt)
3025 2759
Object control (OCT)
Empty subject (empty) Overt Subject (overt)
3033 2898
For region 1, the main effects of Control type and embedded clause type and the interaction of these factors were not significant by subjects or by items (Control type: both F’s<1; embedded clause type: F1ð1; 40Þ ¼ 1:96; p40:05; F251; Control type x embedded clause type: F1ð1; 40Þ ¼ 2:22; p40:05; F251. Similar results were obtained for region 2. Again, the main effects of Control type and embedded clause type and the interaction of these factors were not significant by subjects or by items (all F’s<1). For region 3, the difference between SC/overt and OC/overt sentences was not significant by subjects or by items ðF1ð1; 40Þ ¼ 1:30; p40:05; F2ð1; 33Þ ¼ 1:97; p40:05Þ. The lack of statistically significant results for regions 1–3 indicates essentially no processing differences in the early regions of these Control sentences. For region 4, the main effect of Control type was significant in the bysubjects analysis ðF1ð1; 40Þ ¼ 4:78; p50:05Þ and approached significance in the by-items analysis ðF2ð1; 33Þ ¼ 3:92; p ¼ 0:056Þ. The main effect of embedded clause type was significant by items ðF2ð1; 33Þ ¼ 4:82; p50:05Þ but not by subjects ðF1ð1; 40Þ ¼ 2:33; p40:05Þ. The interaction of Control type and embedded clause type was not significant either by subjects or by
2 The Processing of Japanese Control Sentences
39
items (both F’s<1). The main effect of Control type indicates that SC sentences were read faster than OC sentences in this region. This result does not lend itself to clear interpretation. The main effect of embedded clause type suggests faster RTs for overt subject sentences (SC/overt, OC/overt) than for empty subject sentences (SC/empty, OC/empty) in this region. As discussed above, region 3 of SC/overt and OC/overt sentences both indicates the presence of an embedded clause and specifies the subject of this clause. Therefore, this element places certain interpretive restrictions on region 4 of these overt subject sentences. The reduced potential for ambiguity in these sentences may have facilitated the processing of this region. However, because this main effect of embedded clause type was not statistically reliable by subjects, this interpretation is somewhat speculative. For region 5, the main effect of Control type was not significant in the bysubjects or by-items analysis ðF151; F2ð1; 33Þ ¼ 1:33; p40:05Þ. However, the main effect of embedded clause type was significant in both analyses ðF1ð1; 40Þ ¼ 21:56; p50:001; F2ð1; 33Þ ¼ 19:38; p50:001Þ. The interaction of Control type and embedded clause type was not significant either by subjects or items (both F’s<1). The main effect of embedded clause type indicates longer RTs for empty subject sentences (SC/empty and OC/empty) than for overt subject sentences (SC/overt and OC/overt) in this region. As detailed above, in empty subject sentences, region 5 is the first point at which it is clear that the sentence involves an embedded clause. In order to create the most complete representation of these sentences at this point, the reader must establish both boundaries of the embedded clause, recognize that this clause is missing an argument in subject position and posit a referent for this empty element. The inflated RTs in this region for empty subject sentences can be taken to reflect the processing difficulty associated with (some or all of) these structure-building/ interpretive operations. For region 6, the main effect of Control type was not significant by subjects or by items ðF1ð1; 40Þ ¼ 2:13; p40:05; F2ð1; 33Þ ¼ 2:67; p40:05Þ. However, the main effect of embedded clause type was significant in both analyses ðF1ð1; 40Þ ¼ 11:30; p50:005; F2ð1; 33Þ ¼ 5:50; p50:05Þ. The interaction of Control type and embedded clause type was not significant in either analysis (both F’s<1). The significant main effect of embedded clause type indicates longer RTs for empty subject sentences (SC/empty and OC/empty) than for overt subject sentences (SC/overt and OC/overt) in this region. This effect might be attributed to the spill over of processing difficulty for empty subject sentences in region 5. For region 7, the main effect of Control type was not significant by subjects or items (both F’s<1). However, the main effect of embedded clause type was significant in both analyses ðF1ð1; 40Þ ¼ 14:14; p50:005; F2ð1; 33Þ ¼ 18:50; p50:001Þ. The interaction of Control type and embedded clause type was also significant by subjects and items ðF1ð1; 40Þ ¼ 26:77; p50:001; F2ð1; 33Þ ¼ 19:30; p50:001Þ. For OC sentences, the simple effect of embedded
40
J.D. Witzel and N.O. Witzel
clause type was not significant by subjects or items (both F’s<1). For SC sentences, on the other hand, the simple effect of embedded clause type was significant in both the by-subjects and by-items analyses ðF1ð1; 40Þ ¼ 29:52; p50:001; F2ð1; 33Þ ¼ 34:79; p50:001Þ. Indeed, the mean RT for SC/empty sentences was 71 ms slower than the mean RT for SC/overt sentences in this region. These results indicate processing difficulty when the SC verb is encountered in SC/empty sentences – or in SC sentences in which the embedded clause subject is not overtly specified before the Control verb is encountered. In OC sentences, on the other hand, processing difficulty was not evidenced at the Control verb, regardless of whether the subject of the embedded clause was overtly specified. This pattern of results is consistent with an OC bias in the provisional assignment of antecedents to empty embedded clause subjects in Japanese Control sentences. For region 8, the main effect of Control type was not significant in the bysubjects or by-items analysis ðF1ð1; 40Þ ¼ 1:16; p40:05; F251Þ. However, the main effect of embedded clause type was significant by both subjects and items ðF1ð1; 40Þ ¼ 8:49; p50:01; F2ð1; 33Þ ¼ 4:44; p50:05Þ. The interaction of Control type and embedded clause type was significant by subjects ðF1ð1; 40Þ ¼ 4:41; p50:05Þ and approached significance by items ðF2ð1; 33Þ ¼ 3:07; p ¼ 0:089Þ. For OC sentences, the simple effect of embedded clause type was not significant either by subjects or by items (both F’s<1); for SC sentences, this simple effect was significant in both analyses ðF1ð1; 40Þ20:36; p50:001; F2ð1; 33Þ ¼ 15:93; p50:001Þ. Therefore, in this region, SC/empty sentences again had longer RTs than SC/overt sentences, whereas OC/empty and OC/overt sentences showed no such difference. Given the similarity in the pattern of results for regions 7 and 8, the inflated reading times for SC/empty sentences in this last region of the sentence can be attributed to the spill over of processing difficulty from the region of the Control verb. For the combined RTs (again, the combined RTs of regions 4–8), the main effects of Control type and embedded clause type as well as the interaction of Control type and embedded clause type were significant by subject and by items (Control type: (F1ð1; 40Þ ¼ 4:89; p50:05; F2ð1; 33Þ ¼ 5:28; p50:05 embedded clause type: F1ð1; 40Þ ¼ 26:77; p50:001; F2ð1; 33Þ ¼ 31:79; p50:001; Control type x embedded clause type: F1ð1; 40Þ ¼ 7:31; p50:05; F2ð1; 33Þ ¼ 8:27; p50:01). For both SC and OC sentences, the simple effect of embedded clause type was significant by subjects and by items (SC: F1ð1; 40Þ ¼ 25:47; p50:001; F2ð1; 33Þ ¼ 36:29; p50:001; OC: F1ð1; 40Þ ¼ 13:06; p50:005; F2ð1; 33Þ ¼ 7:60; p50:01). The results for these combined RTs reflect the overall pattern of results in the region-by-region analyses: Longer RTs were obtained for empty subject sentences (SC/empty and OC/empty sentences) than for overt subject sentences (SC/overt and OC/overt sentences), and this difference was more pronounced for SC sentences than for OC sentences.
2 The Processing of Japanese Control Sentences
41
2.4.2 Error Rate (ER) Analyses The mean ER on the complete set of comprehension questions was 10.5%. For the Control item comprehension questions, the mean ER was 12.4%. The mean ERs for the four experimental sentence types (SC/empty, SC/overt, OC/ empty and OC/overt) are presented in Table 2.4. Control type
Table 2.4 Mean ERs (in percentages) for experimental sentences Embedded clause type Mean ER
Subject control (SC)
Empty subject (empty) Overt subject (overt)
17.64% 6.70%
Object control (OC)
Empty subject (empty) Overt subject (overt)
12.67% 13.02%
Interestingly, the ER results for these sentences pattern along with the RT results for regions 7 and 8. Specifically, although there was no significant effect of Control type (both F’s<1), both the main effect of embedded clause type and the interaction of Control type and embedded clause type were significant by subjects and by items (embedded clause type: F1ð1; 40Þ ¼ 6:51; p50:05; F2ð1; 32Þ ¼ 6:27; p50:05; Control type x embedded clause type: F1ð1; 40Þ ¼ 6:19; p50:05; F2ð1; 32Þ ¼ 7:53; p50:05). For SC sentences, the simple effect of embedded clause type was significant by subjects and by items ðF1ð1; 40Þ ¼ 11:88; p50:005; F2ð1; 32Þ ¼ 19:60; p50:001Þ; for OC sentences, the simple effect of embedded clause type was not significant in either analysis (both F’s<1). These results indicate that questions on SC/empty sentences were answered less accurately than questions on SC/overt sentences. However, there was no such accuracy disparity between the questions on OC/empty sentences and those on OC/overt sentences.
2.4.3 Naturalness Ratings The offline questionnaire revealed no significant difference in the participants’ naturalness ratings for SC and OC sentences (mean SC rating = 3.90, mean OC rating = 3.91; both F’s<1). This indicates that the experimental sentences were comparable in terms of naturalness. Therefore, the observed disparities between SC and OC sentences in the reading task cannot be attributed to differences in the acceptability of these sentence types.
2.5 Discussion 2.5.1 Discussion of the Experimental Results The results of this experiment provide support for incrementality in the processing of Japanese sentences generally and in the filler-gap parsing operation
42
J.D. Witzel and N.O. Witzel
involved in Japanese Control sentences in particular. First, sentences with empty embedded clause subjects (SC/empty and OC/empty sentences) had inflated RTs in the final region of the embedded clause, region 5, relative to their overt subject counterparts (SC/overt and OC/overt sentences). We interpret this RT disparity to indicate processing difficulty when the parser was forced to revise a single-clause analysis for the empty subject sentences. Again, in these sentences, the presence of an embedded clause was not clearly indicated until this clause-final region. In the overt subject sentences, however, case information provided by the embedded clause subject could be used to establish a multi-clause interpretation before the entire embedded clause became available. In these sentences, because region 5 simply provided the right boundary for a previously-initiated embedded clause structure, reanalysis was not necessary at this point. Consistent with Miyamoto (2002), these results suggest that in the online processing of these sentences, case information was used to establish clause boundaries before information from relevant phrasal heads became available. It is important to note, however, that a full reanalysis of empty subject sentences at the final region of the embedded clause would involve more than generating clause boundaries. Rather, in order to create the most complete representation of these sentences at this point, the reader would also have to posit an empty subject for the embedded clause (i.e., recognize that this clause was missing an argument in subject position) and assign a referent to this element. That is, the incremental processing of the empty embedded clause subject would require its antecedent to be assigned without delay and without reference to Control information provided by the main clause verb. The results obtained at the Control verb, in region 7, are consistent with such incremental processing. Specifically, processing difficulty was shown at this verb only for SC/empty sentences relative to SC/overt sentences. We interpret this result to indicate that during the comprehension of Japanese Control sentences, the main clause object is initially assigned as the antecedent of the empty embedded clause subject – or, in other words, that there is an OC bias in the provisional assignment of an antecedent to the empty embedded clause subject. When this provisional assignment is then checked against information in the Control verb, disconfirmation (with an SC verb) necessitates structural reanalysis and incurs processing costs. As discussed in the Introduction, this experiment builds on earlier studies examining the comprehension of Japanese Control sentences (Ninose et al., 1998; Oda et al., 1997; Sakamoto, 1995, 1996; see Sakamoto, 2002, 2006, for review). These previous experiments also investigated whether there is bias for positing a specific NP as the antecedent of the empty embedded clause subject in these sentences. As detailed above, however, this work yielded conflicting findings – experiments that required subjects to produce the antecedent of the empty subject revealed an OC bias (Sakamoto, 1995, 1996), while those that required subjects to respond (yes/no) to potential antecedents revealed an SC bias (Ninose et al., 1998; Oda et al., 1997). It is important to reiterate that all of
2 The Processing of Japanese Control Sentences
43
these experiments involved end-of-sentence tasks that are arguably incapable of indicating online processing biases. The present experiment, on the other hand, employed a reading task that is considered sensitive to real-time structural computation. Therefore, we would argue that this experiment provides the only clear indication of the online processes involved in the comprehension of Japanese Control.
2.5.2 Early Disconfirmation of a Processing Bias Leads to Facilitation? The most important result in this experiment is the indication of processing difficulty at the Control verb for SC/empty sentences relative to SC/overt sentences. However, the apparent ease of processing in this region for SC/ overt sentences relative to OC/overt sentences is also striking. Indeed, SC/ overt sentences were read most quickly in this region. Again, the processing difficulty for SC/empty sentences at this verb can be attributed to the costs of disconfirming the provisional, OC-bias-driven assignment of an antecedent. If this is correct, then one might speculate that because this provisional assignment is disconfirmed early in SC/overt sentences, the processing of these sentences is in some way facilitated. The obvious objection to this interpretation is that it relies on the comparison of RTs for two different verbs – one SC verb and one OC verb. Indeed, because these verbs were not matched in terms of frequency or length, comparisons between them cannot be considered reliable.
2.5.3 Why an OC bias? The current experimental outcome revealed an OC bias in the provisional assignment of antecedents to empty embedded clause subjects in Japanese Control sentences. However, it does not dictate the principle that underlies this bias. One possibility is that in Japanese, OC sentences simply occur more frequently than SC sentences. Indeed, SC sentences occur infrequently in English largely because there are few transitive SC verbs (e.g., promise). But the same does not appear to be true of Japanese (Sakamoto, 1996). In the present study alone, 24 such verbs were identified and used in comprehensible Japanese sentences. Despite this fact, a thorough examination of the relative frequency of SC and OC sentences is necessary in order to assess the possibility of a frequency account for the apparent OC bias in the processing of Japanese Control. Alternative explanations for this bias might appeal to the semantics and/or syntax of Japanese Control sentences or to the linear position of constituents in these sentences. For example, a semantic account of this OC bias might be based
44
J.D. Witzel and N.O. Witzel
on Nishigauchi’s (1984) suggestion that thematic Goals are preferred controllers. Consider sentences (9) and (10) (Nishigauchi’s (1984) (4a) and (4b)): (9) Bill bought for Susani a large flashy car [PROi to drive]. (10) Johni received from Susan a book [PROi to read]. In both sentences, the controller of the embedded clause subject is the thematic Goal of the sentence (Susan in (9) and John in (10)). This is true despite the difference in the grammatical roles of these NPs. Nishigauchi accounts for this and other Control phenomena with reference to the following hierarchy of thematic relations: Goal > Location, Source. Essentially, this hierarchy states that in the functional structure that specifies a Goal argument, this thematic relation is preferentially selected as a controller over the otherwise preferred Location and Source thematic roles. To the extent that one can classify the dative-marked NPs in the present study as thematic Goals, the OC bias might be explained in terms of this hierarchy (for more on this interpretation, see Sakamoto, 2002). One might also propose that this OC bias results from a preference for oneto-one matchings between subjects/external arguments and predicates. That is, once an NP has been established as the subject of a given predicate, there may be a preference for maintaining an exclusive relationship between these elements. In the SC/empty and OC/empty sentences in this experiment, the need for an embedded clause subject becomes clear at the rightmost boundary of this clause. It is also clear at this point that the preceding nominative-marked element is the subject of the main clause. If there were a preference for a subject to be associated with a single predicate, then the only other available element – the dative-marked NP – would be selected as the controller of the embedded clause subject. Another explanation for this OC bias might appeal to the structural relations between constituents in these sentences. In particular, it may be that the empty embedded clause subject in these sentences prefers to take the nearest NP (in terms of structural distance) as its controller. Such an account would seemingly match well with Least Effort axioms common in the Minimalist approach to syntactic analysis (Chomsky, 1995). Alternatively, it may be that this OC bias is not based on the abstract semantic or syntactic structure of these sentences at all. Rather, it is entirely possible that this bias results from a preference for positing the nearest element in the linear word string as the antecedent of the empty embedded clause subject. A precedent for this explanation can be found in the early work of Frazier, Clifton, and Randall (1983) and Clifton and Frazier (1986) suggesting that fillers are assigned to PRO gaps in Control sentences through the Most Recent Filler (or MRF) strategy. As its name suggests, the MRF strategy dictates that the nearest lexical filler is assigned to PRO in these sentences (for evidence partially consistent with this strategy in
2 The Processing of Japanese Control Sentences
45
the processing of Spanish Control sentence, see Betancort, Meseguer, & Carreiras, 2004; for more on the influence of locality/recency in sentence processing, see Gibson, 1998 and references therein). It is difficult to conceive of an experiment (or set of experiments) that would confirm one and only one of the above explanations for the OC bias in the processing of Japanese Control. However, by scrambling the main clause dative-marked NP over the main clause nominative-marked NP (to the sentence-initial position), it would be possible to test between the linear recency account for the OC bias and the family of semantically-/structurally-based explanations entertained above (i.e., the thematic hierarchy account, the subject-predicate exclusivity account, and the structural distance account). Scrambling this dative-marked NP would result in a sentence structure roughly as follows (where t indicates the trace/copy of the scrambled NP-dat and e represents the empty embedded clause subject that receives referential specification from Control information in the main clause verb): (11) [NP-DATi NP-NOM ti [e embedded clause] Control verb] If the linear recency account for the OC bias in canonically-ordered Japanese Control sentences is correct, then the opposite bias – an SC bias – should be obtained in their scrambled counterparts. Such a result would eliminate all of the semantically-/structurally-based accounts presented above from contention. If, on the other hand, the OC bias were retained in the processing of these scrambled Control sentences, the linear recency account would not be a viable explanation for this bias (and any of the semantically-/structurally-based accounts for this bias would still be possible).
2.5.4 Why Any Bias? In addition to questions concerning the OC direction of this bias, it is also important to consider why there should be any bias at all. We have argued that this bias is consistent with the incremental processing of Japanese Control sentences. That is, the provisional assignment of antecedents to the empty embedded clause subjects allows for the most complete structural representation of the sentence to be generated before information in the verb becomes available. It may be, however, that we can put a finer point on the mechanism that underlies this incremental parsing decision. As discussed above, with respect to the processing of filler-gap dependencies, Aoshima et al. (2004) posited a Full Constraint Satisfaction model. According to this model, a displaced filler initiates a search for the first gap position that resolves its thematic dependencies. In light of the results of the present study, it might be possible to expand this model. Specifically, the results of the present experiment suggest that when a gap is detected, a referent/filler is immediately assigned to this element. The Full Constraint Satisfaction model might therefore
46
J.D. Witzel and N.O. Witzel
be restated as follows: In the processing of filler-gap operations, satisfy the dependencies of identified fillers and gaps as soon as possible. Clearly, the dependencies introduced by these elements are different: A displaced filler requires thematic specification, whereas a gap requires referential specification. However, the same underlying mechanism might account for the satisfaction of these dependencies during online sentence processing. Acknowledgements This research was supported in part by an East Asia and Pacific Summer Institute grant from the National Science Foundation and the Japanese Society for the Promotion of Science. We would like to thank Yuki Hirose for her help and encouragement throughout this project. We would also like to thank Ken Forster, Merrill Garrett, Janet Nicol, Andy Barss, Mariko Karatsu, Simin Karimi and the participants at the First International Conference on the Processing of Head-Final Structures for their comments and suggestions. All mistakes are our own. Corresponding author: Jeffrey Witzel (
[email protected]).
References Aoshima, S., Phillips, C., & Weinberg, A. (2003). Processing of Japanese wh-scrambling constructions. In W. McClure (Ed.), Japanese/Korean linguistics (Vol. 12, pp. 179–191). Stanford, CA: CSLI Publications. Aoshima, S., Phillips, C., & Weinberg, A. (2004). Processing filler-gap dependencies in a headfinal language. Journal of Memory and Language, 51, 23–54. Berwick, R. C., & Fong, S. (1995). Madama Butterfly redux: Parsing English and Japanese with a principles and parameters approach. In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing (pp. 177–208). Hillsdale, NJ: Lawrence Erlbaum Associates. Betancort, M., Carreiras, M., & Acun˜a-Farin˜a, C. (2006). Processing controlled PROs in Spanish. Cognition, 100, 217–282. Betancort, M., Meseguer, E., & Carreiras, M. (2004). The empty category PRO: Processing what cannot be seen. In M. Carreiras & C. Clifton (Eds.), The online study of sentence comprehension: Eyetracking, ERPs and beyond (pp. 95–118). New York, NY: Psychology Press. Bever, T. G., & McElree, B. (1988). Empty categories access their antecedents during comprehension. Linguistic Inquiry, 19(1), 35–43. Boland, J. E., Tanenhaus, M. K., & Garnsey, S. M. (1990). Evidence for the immediate use of verb control information in sentence processing. Journal of Memory and Language, 29, 413–432. Chomsky, N. (1981). Lectures on government and binding. Dordrecht: Foris. Chomsky, N. (1982). Some concepts and consequences of the theory of government and binding. Cambridge: MIT Press. Chomsky, N. (1995). The minimalist program. Cambridge: MIT Press. Chomsky, N., & Lasnik, H. (1977). Filters and control. Linguistic Inquiry, 8(3), 425–504. Clifton, C., & Frazier, L. (1986). The use of syntactic information in filling gaps. Journal of Psycholinguistic Research, 15(3), 209–224. Crain, S., & Fodor, J.D. (1985). How can grammars help parsers? In D.R. Dowty, L. Karttunen & A. Zwicky (Eds.), Natural language parsing: Psychological, computational, and theoretical perspectives (pp. 95–128). Cambridge: Cambridge University Press. Davies, W. D., & Dubinsky, S. (2004). The grammar of raising and control: A course in syntactic argumentation. Oxford: Blackwell Publishers. Frazier, L., Clifton, C., & Randall, J. (1983). Filling gaps: Decision principles and structure in sentence comprehension. Cognition, 13, 187–222.
2 The Processing of Japanese Control Sentences
47
Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68, 1–76. Haegeman, L. (1994). Introduction to government and binding theory. Oxford: Blackwell. Inoue, A., & Fodor, J. D. (1995). Information-paced parsing in Japanese. In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing (pp. 9–63). Hillsdale, NJ: Lawrence Erlbaum Associates. Just, M. A., Carpenter, P. A., & Woolley, J. (1982). Paradigms and processes in reading comprehension. Journal of Experimental Psychology: General, 11, 228–238. Kamide, Y. (2006). Incrementality in Japanese sentence processing. In M. Nakayama, R. Mazuka & Y. Shirai (Eds.), The handbook of East Asian psycholinguistics: Japanese (pp. 249–256). Cambridge: Cambridge University Press. Kamide, Y., Altmann, G. T. M., & Haywood, S. L. (2003). The time-course of prediction in incremental sentence processing: Evidence from anticipatory eye movements. Journal of Memory and Language, 49, 133–156. Kamide, Y., & Mitchell, D. C. (1999). Incremental pre-head attachment in Japanese parsing. Language and Cognitive Processes, 14, 631–662. Mazuka, R., & Nagai, N. (1995). Japanese sentence processing: An interdisciplinary approach. In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing (pp. 1–8). Hillsdale, NJ: Lawrence Erlbaum Associates. McElree, B., & Bever, T. G. (1989). The psychological reality of linguistically defined gaps. Journal of Psycholinguistic Research, 18(1), 21–35. Miyamoto, E. T. (2002). Case markers as clause boundary inducers in Japanese. Journal of Psycholinguistic Research, 31(4), 307–347. Nakayama, M. (1999). Sentence processing. In N. Tsujimura (Ed.), Handbook of Japanese linguistics (pp. 398–424). Oxford: Blackwell. Nicol, J., & Swinney, D. (1989). The role of structure in coreference assignment during sentence comprehension. Journal of Psycholinguistic Research, 18(1), 5–19. Ninose, Y., Oda, J., Sakaki, Y., Sakamoto, T., & Gyoba, J. (1998). On the real-time processing of empty subjects in Japanese using a dichotic-listening method (2). Cognitive Science, 5, 82–88. Nishigauchi, T. (1984). Control and the thematic domain. Language, 60(2), 215–250. Oda, J., Ninose, Y., Sakaki, Y., Gyoba, J., & Sakamoto, T. (1997). On the real-time processing of empty subjects in Japanese using a dichotic-listening method. Cognitive Science, 4, 58–63. Pritchett, B. L. (1991). Head position and parsing ambiguity. Journal of Psycholinguistic Research, 20, 251–270. Sakamoto, T. (1995). Transparency between parser and grammar: On the processing of empty subjects in Japanese. In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing (pp. 275–294). Hillsdale, NJ: Lawrence Erlbaum Associates. Sakamoto, T. (1996). Processing empty subjects in Japanese: Implications for the transparency hypothesis. Fukuoka: Kyushu University Press. Sakamoto, T. (2002). Processing filler-gap constructions in Japanese: The case of empty subject sentences. In M. Nakayama (Ed.), Sentence processing in East Asian languages (pp. 189–221). Stanford, CA: CSLI Publications. Sakamoto, T. (2006). Processing empty categories in Japanese. In M. Nakayama, R. Mazuka & Y. Shirai (Eds.), The handbook of East Asian psycholinguistics: Japanese (pp. 270–276). Cambridge: Cambridge University Press. Shibatani, M. (1990). The languages of Japan. Cambridge: Cambridge University Press. Stowe, L. A. (1986). Parsing WH-constructions: Evidence for on-line gap location. Language and Cognitive Processes, 1, 227–245. Tsujimura, N. (1996). An introduction to Japanese linguistics. Oxford: Blackwell. Tsujimura, N. (1999). Handbook of Japanese linguistics. Oxford: Blackwell. Zushi, M. (2003). Null arguments: The case of Japanese and Romance. Lingua, 113, 559–604.
Chapter 3
Individual Differences in Sentence Processing: Effects of Verbal Working Memory and Cumulative Linguistic Knowledge Nobuyuki Jincho and Reiko Mazuka
3.1 Introduction Previous psycholinguistic research has revealed that many linguistic factors such as structural complexity and temporal ambiguity can affect processing demands during sentence comprehension. However, not many studies have investigated individual differences in sentence processing. At the word level and the discourse level, many studies have demonstrated that various cognitive factors could modulate comprehension performance (e.g., Amano & Kondo, 1999; Stanovich & West, 1989 for the word level, Kintsch & van Dijk, 1978 for the discourse level). Thus, it is likely that some cognitive factors can also explain individual differences in sentence level processing. The present study discusses whether two cognitive factors, verbal working memory (VWM) and cumulative linguistic knowledge (CLK, discussed in detail below), contribute to sentence processing.
3.1.1 Two Factor Model of Individual Differences in Reading Comprehension at Text and Word Levels Previous literature has primarily addressed VWM as a predictor of individual differences in text reading comprehension (Daneman & Carpenter, 1980). We will call this the VWM model of Individual differences. Working memory is a mental resource that is used for both mental processing and the storage of intermediate and final products (Baddeley, 1986; Just & Carpenter, 1992). During text reading, VWM is critical not only for a buffer of preceding context but also for integration of context and new information (Kintsch & van Dijk, 1978). Readers with high VWM capacity, which is typically measured by N. Jincho (*) Laboratory for Language Development, RIKEN Brain Science Institute, 2-1 Hirosawa, Wako-shi, Saitama, Japan e-mail:
[email protected]
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_3, Ó Springer ScienceþBusiness Media B.V. 2011
49
50
N. Jincho and R. Mazuka
the Reading Span test (RST), can comprehend a text better than low VWM capacity readers (Daneman & Carpenter, 1980; Daneman & Merikle, 1996). Although VWM capacity may be one factor contributing to general reading comprehension, an additional language-related cognitive factor may contribute to individual differences. Jincho, Namiki, and Mazuka (2008) proposed that an additional factor, Cumulative Linguistic Knowledge (hereafter CLK), also makes an independent contribution to reading comprehension. CLK consists of various types of knowledge that are acquired in language related experience (e.g., reading) throughout one’s life. We call this the Two Factor Model of Individual Differences. In their study, vocabulary size, orthographic knowledge and general world knowledge, all of which were assumed to be CLK measures, were highly correlated with one another but none of them significantly correlated with the RST score (i.e., VWM capacity). In addition, both VWM capacity and scores in CLK measures could predict performance in a general reading comprehension task independently (see also Dixon, LeFevre, & Twilley, 1988). Metaphorically, individual differences in VWM are like the speed of a CPU and the size of RAM of a computer, while CLK is like the quality and the size of an encyclopedic database installed on the hard disk. Both of these would contribute to how well a text can be processed but the two are independent; CPU and RAM can be faster/larger independent of the quality of the database and vice versa. In word level processing, previous studies of individual differences among skilled adult readers showed that one’s CLK was a good predictor of word level processing (Amano & Kondo, 1999; Stanovich & West, 1989). Amano and Kondo (1999) found that people with better orthographic knowledge of kanji words named words faster than those with less orthographic knowledge. Stanovich and West (1989) demonstrated that readers with more print exposure showed better performance in orthographic processing (e.g., recognition of spelling of words) and phonological processing (e.g., pseudoword naming).
3.1.2 VWM Model of Individual Differences in Sentence Processing Individual differences in VWM capacity are found to be a good predictor of one’s on-line sentence processing (Just & Carpenter, 1992; King & Just, 1991; MacDonald, Just, & Carpenter, 1992; Miyake, Just, & Carpenter, 1994; Waters & Caplan, 1996). King and Just (1991) found that readers with low VWM capacity took longer to read object-relative clauses than subject-relative clauses, although high VWM capacity readers did not. MacDonald et al. (1992) found that high VWM capacity readers showed a longer reading time for disambiguating temporal syntactic ambiguities than low VWM capacity readers. However, low capacity readers’ performance in probe questions was lower than high capacity readers. In Miyake et al. (1994), readers with only low VWM capacity showed
3 Individual Differences in Sentence Processing
51
significantly longer reading times for resolving lexical ambiguity. These studies suggest that high VWM capacity readers can hold multiple interpretations for a longer time than low capacity readers. In other words, VWM capacity may determine whether one has to select one interpretation or affords to keep multiple interpretations (Capacity hypothesis). An alternative interpretation for the results in MacDonald et al. (1992) was proposed by MacDonald, Pearlmutter, and Seidenberg (1994). According to them, the differences between high and low capacity readers were due to on-line utilization of plausibility information to resolve syntactic ambiguity rather than to capacities for holding multiple syntactic structures (Verb Plausibility hypothesis).
3.1.3 Implications of VWM Model for Head-Final Languages In head-final languages, furthermore, VWM capacity may become more important than in head-initial languages. For instance, a verb typically comes at the end of a clause in head-final languages. It has been claimed that the lexical information from a verb is critical for on-line comprehension of a sentence (MacDonald et al., 1994; Trueswell, Tanenhaus, & Garnsey, 1994). Thus, people may delay parsing syntactic structure until a head appears (Head-driven model, Pritchett, 1991). This processing model, however, places great demands on VWM since the increasing number of arguments must consume VWM resources without any chunking. Thus, the size of one’s VWM capacity may determine whether one can correctly interpret a sentence even if there are no temporal syntactic ambiguities in the sentence. Alternatively, people process each word incrementally even in head-final languages (Incremental model, Inoue & Fodor, 1995). Although this strategy bares the risk that the initial parsing decision is found to be incorrect later in the sentence, the demand on VWM is reduced by chunking arguments. Therefore, the size of VWM capacity may not be so critical in reading sentences with a simple structure. It becomes important only when temporal syntactic ambiguity allows multiple interpretations of the syntactic structure. Previous studies reporting a significant relationship between VWM and syntactic ambiguity resolution examined mainly head-initial languages and raised ambiguities generated by verbs. This may be one of the reasons why multiple accounts (i.e., Capacity hypothesis and Verb Plausibility hypothesis, see above) have been proposed for the role of VWM. In head-final languages, however, syntactic ambiguity may occur before a verb is presented. Thus, examining the effect of VWM in head-final languages enables us to determine which of the two accounts is more valid. If the capacity of VWM is critical for maintenance of multiple syntactic structures as argued in the Capacity hypothesis, capacity differences should be crucial for the resolution of syntactic ambiguities not generated by verbs. In contrast, if VWM is related to utilization of plausibility information of verbs as argued in the Verb Plausibility hypothesis, the VWM capacity may not have critical importance for syntactic ambiguity in head-final languages.
52
N. Jincho and R. Mazuka
3.1.4 Two Factor Model of Individual Differences in Sentence Processing As argued in the previous sections, VWM capacity contributes to individual differences in sentence processing. It is likely that other factors such as CLK also contribute. In the present paper, we examine whether the Two Factor model of individual differences (VWM + CLK; Jincho et al., 2008) predicts individual differences in sentence processing better than VWM alone and how VWM and CLK interact in processing various types of sentences. In the VWM model, the limitation in VWM capacity should predict processing difficulty in general. The Two Factor model, in contrast, assumes that VWM and CLK are independent factors and they both contribute to sentence processing. As described above, Stanovich and West (1989) suggested that people with high CLK read words faster and more accurately than people with low CLK, thus, they should have less difficulty in sentence comprehension than low CLK readers in general. It is not clear, however, how the two factors may interact in sentence processing. On the one hand, a high capacity in either of these factors may mediate processing difficulty; for example, one’s CLK may influence how much of one’s VWM resource needs to be allocated for processing a sentence, resulting in low CLK readers requiring more VWM. Alternatively, the resources CLK employs may be independent from one’s VWM capacity. This assumption predicts only main effects of individual differences in VWM and CLK. As regards the processing of head-final structures, the Two Factor model may be able to hypothesize about the specific roles of VWM and CLK. As argued in the previous section, head-final structure languages have a verb at clause-final position. Thus, VWM may contribute to how many words can be stored without any commitment or to how many structures one can keep in ambiguous situations. In addition to VWM contribution, CLK may have an independent role. In head-initial languages, the verb’s semantic information comes in an earlier position of a clause and constrains the following contents (Altmann & Kamide, 1999; Kamide, Altmann, & Haywood, 2003), which might reduce the demand on VWM. However, in head-final languages, we cannot receive any assistance from verb meanings. Thus, readers or listeners have to predict the following contents from earlier parts of a clause such as from nouns and case markers (e.g., Kamide et al., 2003). It is not clear what types of cognitive factors are utilized for predicting the subsequent contents of sentences. One’s CLK may contribute to how precise a prediction one can make for a sentence structure without verb information. More concrete predictions are made below.
3.1.5 Purpose of the Present Study The goal of the present study is to test whether the Two Factor model can explain individual differences in sentence processing better than the VWM
3 Individual Differences in Sentence Processing
53
model. More specifically, we test how the two factors, VWM and CLK interact in sentence processing by examining different types of sentences. In Experiment 1, we use ditransitive verb sentences in which high/low-frequency words are embedded in the direct object position. As readers try to integrate a new word into the meaning of the sentence as a whole, the processing may be influenced by whether the word is a high-frequency word or low-frequency word (Rayner & Duffy, 1986). The VWM model of individual differences predicts that low VWM readers should take longer to read sentences containing low-frequency words than high VWM capacity readers (Perfetti, 1988). In the Two Factor model, two predictions are possible regarding the effects of lexical frequency. First, if low-frequency words require more resources than high-frequency words, as assumed in the VWM model, the lexical demands of low-frequency words should interactively increase for low VWM/ low CLK readers. Alternatively, it is possible that lexical frequency effects will be found only on a CLK scale, independent of VWM. This is because although the lexical frequency may be part of CLK, it does not make a sentence ambiguous, thus it may not place increased demand on VWM. In Experiments 2 and 3, transitive sentences with canonical and scrambled order are used. Previous literature has pointed out multiple sources of difficulty in the comprehension of scrambled sentences. One difficulty comes from filler-gap dependency. According to Mazuka, Itoh, and Kondo (2002; Miyamoto & Takahashi, 2002), scrambled word order of transitive sentences contains a gap following the subject noun phrase. The object noun phrase acts as a filler for this gap. Successful retrieval of the filler at the gap position requires enough working memory capacity. Low VWM capacity readers may show a larger cost of gap-filling (i.e., increase in reading time at subject noun phrase in scrambling sentences compared to object noun phrase in canonical sentences) than high VWM capacity readers do. Another difficulty is temporal ambiguity within a sentence structure. Facing an object noun, readers may have multiple interpretations for the sentence. Other than the scrambling sentence interpretation (e.g., Mary-o John-ga mita. ‘John saw Mary.’), for instance, a verb may follow the object noun and it is found to be a null subject sentence (e.g., Mary-o mita. ‘(someone) saw Mary’). Otherwise, a verb and a subject noun phrase may follow the object noun and the object may be a part of a relative clause (e.g., Mary-o mita John-ga . . . ‘John, who saw Mary, . . .’). According to previous literature (MacDonald et al., 1992), we predict that readers take longer to read object noun phrases of scrambled sentences than subject noun phrases of canonical sentences and that low VWM readers will show longer reading times on object noun phrases of scrambled sentences than high VWM readers. In addition, structural frequency may also determine processing demands reflected in reading time in scrambled sentences (MacDonald et al, 1994). We assume that canonical subject-object order is much more frequent than scrambled object-subject order for any type of transitive verb (Yamashita, 2002). Thus, we predict low CLK readers to have more difficulties with rare
54
N. Jincho and R. Mazuka
processing (e.g., filling-a-gap) while reading scrambled sentences than high CLK readers since they do not have enough experience. Thus, it is predicted that low CLK readers will show larger structural frequency effects than high CLK readers, similar to the lexical frequency effects in Experiment 1. In summary, multiple linguistic factors can explain the difficulty of two types of sentences: lexical and structural frequency, syntactic ambiguity and fillergap-dependency. Thus, these sentences tell us which types of processing are mediated by one’s VWM and CLK. Furthermore in Experiment 3, we compare sentences with canonical and scrambled word order where two adverbial or prenominal modifiers precede each of the NPs. Previous studies have demonstrated that long distance dependency places an increased demand on VWM (King & Just, 1991; Miyamoto & Takahashi, 2002; Ueno & Kluender, 2003). Thus, this is a structure in which individual differences in VWM are likely to have an effect. If people read words without constructing syntactic structure as the Head-driven model claims, low VWM readers should have greater difficulty in reading canonical sentences than high VWM readers. If readers incrementally construct syntactic structure as the Incremental model claims, individual differences in VWM capacity should not be related to processing canonical sentences.
3.2 Experiment 1 (Lexical Frequency) 3.2.1 Method 3.2.1.1 Participants Forty native speakers of Japanese (mean age = 21.1, SD = 2.4) participated in Experiment 1. To assess individual differences in VWM and CLK, we administered a Japanese version of RST (Osaka, 1998) and a Hyakurakan kanji word reading test (Amano & Kondo, 1999). In the RST, participants read sentences aloud and remembered an underlined word in the sentence. One’s VWM capacity was defined by the number of successfully recalled underlined words. In the Hyakurakan test, participants wrote pronunciations for 100 kanji words. One’s CLK was defined by the number of correct items. Considering performance in both tests, they were divided into four groups (mean scores in the RST and the Hyakurakan test are shown in Table 3.2). 3.2.1.2 Materials Experimental sentences were ditransitive verb sentences in which the frequency of accusative nouns was manipulated (see Table 3.1). High-frequency words appear at least 50,000 times in a newspaper corpus (Amano & Kondo, 2001), while low-frequency words appear only once. Other phrases and the length in letters and morae of accusative phrases were matched between the two conditions.
3 Individual Differences in Sentence Processing
Condition
Table 3.1 Example set of stimulus sentences in Experiment 1 NP-NOM Adv NP-DAT NP-ACC
55
Verb
High-frequency
Ace-ga totsuzen kantoku-ni mondai-o Ace-NOM suddenly director-DAT problem-ACC ‘The ace pitcher suddenly appealed a problem to the director.’
uttaeta. appealed
Low-frequency
Ace-ga totsuzen kantoku-ni haitsu-o uttaeta. Ace-NOM suddenly director-DAT pain in the back-ACC appealed ‘The ace pitcher suddenly appealed pain in his back to the director.’
3.2.1.3 Procedure Participants read 84 sentences in a non-cumulative self-paced reading task. Sixteen of them were critical sentences in which high- or low-frequency words were embedded in the direct object position. Another 20 sentences are analyzed in Experiment 3. The remaining 48 sentences were used as filler sentences. After reading each sentence, participants answered a yes/no comprehension question about thematic roles of the noun phrase (e.g., ‘‘Did the director appeal for anything?’’ for the example sentence in Table 3.1).
3.2.2 Results and Discussion We compared the four groups’ performance in high- and low-frequency conditions. In off-line comprehension scores (see Table 3.2), a 2 (VWM: high, low) 2 (CLK: high, low) 2 (Frequency: high, low) ANOVA revealed that lowfrequency sentences led lower accuracy than high-frequency sentences. In Table 3.2 Mean scores in Reading Span Test and Hyakurakan test, and accuracy in comprehension question in Experiments 1 and 3 Experiment 1 Experiment 3 High Low Group RST Hyakurakan frequency frequency Canonical Scrambled HighVWM/ 4.0 (0.6) 74.9 (4.0) 0.95 (0.09) 0.88 (0.20) 0.95 (0.05) 0.74 (0.14) HighCLK (n = 12) HighVWM/ 3.9 (0.6) 55.3 (8.4) 0.97 (0.06) 0.86 (0.14) 0.93 (0.07) 0.76 (0.11) LowCLK (n = 7) LowVWM/ 2.2 (0.3) 76.4 (4.4) 0.96 (0.08) 0.89 (0.14) 0.90 (0.12) 0.73 (0.08) HighCLK (n = 11) LowVWM/ 2.2 (0.3) 59.5 (5.3) 0.98 (0.05) 0.86 (0.19) 0.89 (0.11) 0.72 (0.13) LowCLK (n = 10) Parentheses show SDs. VWM: verbal working memory, CLK: cumulative linguistic knowledge, RST: reading span test.
56
N. Jincho and R. Mazuka
addition, critically, an interaction between Frequency and CLK was significant, F1 ð1; 33Þ ¼ 12:41; p50:01; F2 ð1; 30Þ ¼ 4:46; p50:05. Lower performance in low-frequency sentences was significant in low CLK readers but not in high CLK readers. On-line reading times for high- and low-frequency words among the four groups revealed a similar pattern of results. We calculated participants’ residual reading times as an on-line reading time measure in which effects of visual- and phonological-length were reduced, using the following regression equation (for details on this procedure, see also Mazuka, Itoh, & Kondo, 1997). RTðestimatedÞ ¼ a ðthe number of moraÞ þ b ðthe number of characterÞ þ intercept
RTðresidualÞ ¼ RTðobservedÞ RTðestimatedÞ
200
Residual reading time (ms)
Residual reading time (ms)
High VWM
First, the coefficients of two predictors (a and b) and the intercept were estimated. Then, we were able to calculate estimated RT for each phrase. Finally, residual RT was obtained by subtracting estimated RT from observed RT. We assumed that longer residual reading time reflected high processing demands in reading the phrase. A three-factor ANOVA showed that longer reading times of accusative nouns were observed for the low-frequency condition (75 ms) than for the high-frequency condition (–73 ms), F1 (1,33) = 24.62, p < 0.001, F2 (1,30) = 22.23, p < 0.001. Moreover, we found a significant interaction between Frequency and CLK, F1 ð1; 33Þ ¼ 4:47; p50:05; F2 ð1; 30Þ ¼ 15:49; p50:001. As Fig. 3.1 shows, all groups seemed to have greater difficulties with low-frequency 150 100 50 0 –50 –100 NP–NOM
low frequency
100 50 0 –50 –100
Adv NP–DAT NP–ACC Verb
200
NP–NOM
Residual reading time (ms)
Residual reading time (ms)
high frequency
150
–150
–150
Low VWM
200
150 100 50 0 –50 –100
Adv NP–DAT NP–ACC Verb
200 150 100 50 0 –50 –100 –150
–150 NP–NOM
Adv
NP–DAT NP–ACC Verb
Low CLK
NP–NOM Adv NP–DAT NP–ACC Verb
High CLK
Fig. 3.1 Mean residual reading time on each phrase in Experiment 1
3 Individual Differences in Sentence Processing
57
words than with high-frequency words. However, reliable effects of frequency were not obtained in high CLK readers (right two graphs) while they were obtained in low CLK readers (left two graphs). Differences in VWM did not influence off-line comprehension or on-line reading time performance. These results suggest that the frequency effect was mediated only by readers’ CLK and if readers have high linguistic knowledge, the demand of lexical processing is reduced.
3.3 Experiment 2 (Adjacent Scrambling Sentences) 3.3.1 Method 3.3.1.1 Participants Thirty-seven Japanese speakers (mean age = 20.4, SD = 2.0) participated. As in Experiment 1, they were divided into four groups in terms of performance on the RST and the Hyakurakan test (Table 3.4). 3.3.1.2 Materials Twenty-four transitive verb sentences were tested in Experiment 2 (See Table 3.3). In a half of them, the nominative noun preceded the accusative noun (Canonical condition). In the other half, the order of nominative and accusative nouns was reversed (Scrambled condition).
Condition
Table 3.3 Example set of stimulus sentences in Experiment 2 NP-LOC NP-NOM/ACC NP-ACC/NOM
Verb
Canonical
Kaisha-de Yamamoto-ga Hiroko-o office-at name-NOM name-ACC ‘At the office, Yamamoto teased Hiroko.’
hiyakashita. teased
Scrambled
Kaisha-de Yamamoto-o Hiroko-ga office-at name-ACC name-NOM ‘At the office, Hiroko teased Yamamoto.’
hiyakashita. teased
3.3.1.3 Procedure Participants read 24 sentences in a similar manner as in Experiment 1. Half of them were canonical order sentences and the other half were scrambled order sentences. They also read 96 filler sentences in the same session.
3.3.2 Results and Discussion We conducted separate 2 (VWM: high, low) 2 (CLK: high, low) 2 (Word order: canonical, scrambled) ANOVAs on off-line comprehension performance
58
N. Jincho and R. Mazuka
Table 3.4 Mean scores in Reading Span Test and Hyakurakan test, and accuracy in comprehension questions in Experiment 2 Group RST Hyakurakan Canonical Scrambled HighVWM/highCLK 4.1 (0.6) 76.6 (6.8) 0.88 (0.17) 0.72 (0.28) (n = 8) HighVWM/lowCLK 4.0 (0.6) 59.9 (3.3) 0.88 (0.18) 0.73 (0.25) (n = 8) LowVWM/highCLK 2.3 (0.3) 74.9 (4.7) 0.93 (0.12) 0.63 (0.10) (n = 11) LowVWM/lowCLK 2.4 (0.4) 61.9 (3.0) 0.93 (0.08) 0.70 (0.03) (n = 10) Parentheses show SDs. VWM: verbal working memory, CLK: cumulative linguistic knowledge, RST: reading span test.
and residual reading time of each phrase. In comprehension questions, all four groups showed almost perfect accuracy for canonical sentences (Table 3.4). However, their scores for scrambled sentences were lower than for canonical sentences, F1 ð1; 36Þ ¼ 26:20; p50:001; F2 ð1; 46Þ ¼ 40:41; p50:001. Other main effects and interactions involving VWM capacity or CLK did not reach significance. In on-line reading time data (Fig. 4.2), however, an interaction between CLK and word order was found as well as a main effect of word order in the third phrase, F1 ð1; 36Þ ¼ 3:53; p50:07; F2 ð1; 23Þ ¼ 9:40; p50:01 for Residual reading time (ms)
Residual reading time (ms)
High VWM
250 200 150 100 50 0 –50 –100 –150 –200 NP– NOM/ACC
NP– ACC/NOM
Verb
250 200 150 100 50 0 –50 –100 –150 –200
Residual reading time (ms)
Low VWM
Residual reading time (ms)
NP–loc
NP–loc
NP– NOM/ACC
NP– ACC/NOM
250 200 150 100 50 0 –50 –100 –150 –200
Verb
canonical scrambled
NP–loc
NP– NOM/ACC
NP– ACC/NOM
Verb
NP–loc
NP– NOM/ACC
NP– ACC/NOM
Verb
250 200 150 100 50 0 –50 –100 –150 –200
Low CLK Fig. 3.2 Mean residual reading time on each phrase in Experiment 2
High CLK
3 Individual Differences in Sentence Processing
59
the interaction between word order and CLK; F1 (1, 36) = 12.10, p < 0.01; F2 ð1; 23Þ ¼ 15:97; p50:01 for word order effect. Post-hoc comparisons revealed that only the low CLK group (left two graphs) showed longer reading time for scrambled sentences than for canonical sentences. The high CLK group (right two graphs) did not show significant increase in scrambled sentences compared to canonical sentences. These results support the Two Factor model. The interaction observed between scrambling and CLK could indicate that CLK mediates the size of the frequency effect: The more experience readers have, the smaller the frequency effect becomes. Contradictory to our prediction, we did not find any effect of VWM capacity. Thus, the difficulty of scrambled word order might come from the structural frequency. However, absence of VWM effect might be due to the fact that subject and object nouns were adjacent. This might reduce demands for keeping multiple interpretations and retrieving the filler item at the gap position. In Experiment 3, we further investigate the contribution of VWM capacity in scrambling by using scrambled sentences in which subject and object words were not adjacent.
3.4 Experiment 3 (Distant Scrambling Sentences) 3.4.1 Method 3.4.1.1 Participants and Procedure Data for the sentences (20 sentences) in Experiment 3 were collected in the same session as Experiment 1.
3.4.1.2 Materials We manipulated the word order of nominative and accusative nouns in transitive verb sentences as in Experiment 2 (Table 3.5). In the canonical condition, nominative nouns preceded accusative nouns. In the scrambled condition, accusative nouns appeared earlier than nominative nouns. In addition, the sentences had three phrases between nominative and accusative nouns.
3.4.2 Results and Discussion We conducted 2 (VWM) 2 (CLK) 2 (Word order) ANOVAs for comprehension performance and reading time of each phrase separately. In off-line comprehension performance (Table 3.2), the main effect of Word order was highly reliable, F1 ð1; 33Þ ¼ 58:2; p50:001; F2 ð1; 38Þ ¼ 10:2; p50:01. Effects of VWM and CLK were not significant.
ginko-no uraguchi-de keikan-ga hitogomi-no aida-kara shinchona bank-GEN back entrance-LOC policeman-NOM crowd-GEN among careful ‘At the back entrance of the bank, a policeman watched a careful criminal among the crowd.’
ginko-no uraguchi-de keikan-o hitogomi-no aida-kara shinchona bank-GEN back entrance-LOC policeman-ACC crowd-GEN among careful ‘At the back entrance of the bank, a careful criminal watched a policeman among the crowd.’
Scrambled
Table 3.5 Example set of stimulus sentences in Experiment 3 NP-NOM/ACC I1 I2 I3
Canonical
B2
B1
Condition
hannin-ga criminal-NOM
hannin-o criminal-ACC
NP-ACC/NOM
kanshishita. watched
kanshishita. watched
Verb
60 N. Jincho and R. Mazuka
3 Individual Differences in Sentence Processing
61
On reading time data (Fig. 3.3), however, VWM and CLK interactively affected word order. In the fourth phrase (I1 in Fig. 3.3), the main effects of VWM and word order were significant, F1 (1, 33) = 10.03, p < 0.05, F2 (1, 35) = 5.75, p50:01 for VWM; F1 ð1; 33Þ ¼ 6:59; p50:05 but F2 (1, 35) = 0.69, p > 0.10 for word order. In addition, an interaction between VWM and word order was also significant, F1 ð1; 33Þ ¼ 10:55; p50:01; F2 ð1; 35Þ ¼ 8:81; p50:01. Low VWM capacity readers (bottom two graphs) showed longer reading times for scrambled sentences than for canonical order sentences, p < 0.01. However, high VWM readers did not show significantly increased times for scrambled sentences. When participants read this region, multiple interpretations of sentence structures were assumed to be held. Thus, the obtained interaction suggests that the maintenance costs of those structures might be larger for low VWM readers than for high VWM readers. In the fifth and sixth phrases (I2 and I3 in Fig. 3.3), word order effects were reliable F1 ð1; 33Þ ¼ 12:06; p50:01; F2 ð1; 33Þ ¼ 21:151; p50:001 for fifth phrase; F1 ð1; 33Þ ¼ 12:66; p50:01; F2 ð1; 37Þ ¼ 19:09; p5:001 for sixth phrase. Increased reading times were found for all participants, which suggests that even high VWM readers have to spend their VWM capacity keeping multiple interpretations for a very long time. In the seventh phrase, where the accusative noun could lead to a scrambled transitive verb sentence, only word order effect was reliable, F1 ð1; 33Þ ¼ 8:98; p50:01; F2 ð1; 37Þ ¼ 57:43; p50:001. In the last phrase (Verb in Fig. 3.3), reading time in scrambled sentences was significantly longer than in canonical sentences, F1 (1, 33) = 11.68, p < 0.01, F2 (1, 35) = 12.58, p < 0.01. 250 Residual reading time (ms)
High VWM
Residual reading time (ms)
250 200 150 100 50 0 –50 –100 B1
100 50 0 –50 –100
B2 NP–NOM/ACC I1 I2 I3 NP–ACC/NOMVerb
250
250
200
200
Residual reading time (ms)
Low VWM
150
–150
–150
Residual reading time (ms)
canonical scrambled
200
150 100 50 0 –50 –100 –150
B1
B2 NP–NOM/ACC I1 I2 I3 NP–ACC/NOMVerb
B1
B2 NP–NOM/ACC I1 I2 I3 NP–ACC/NOMVerb
150 100 50 0 –50 –100 –150
B1
B2 NP–NOM/ACC I1 I2 I3 NP–ACC/NOMVerb
Low CLK
High CLK
Fig. 3.3 Mean residual reading time on each phrase in Experiment 3
62
N. Jincho and R. Mazuka
Although other main effects and interactions were not significant, the high VWM/ high CLK group (top right graph) showed smaller differences between canonical and scrambled sentences than the other three groups did. Therefore, we conducted paired comparisons between canonical and scrambled sentences in each group. High VWM/ high CLK group readers did not have longer reading times for scrambled sentences compared to canonical sentences, t1 ð7Þ ¼ 0:79; p40:10; t2 ð37Þ ¼ 1:11; p40:10. However, the other three groups did show longer reading times for scrambled sentences than for canonical sentences, t1 ð9Þ ¼ 1:62; p ¼ 0:14; t2 ð37Þ ¼ 2:40; p50:05 for low VWM/ low CLK group; t1 ð10Þ ¼ 2:31; p50:05; t2 ð37Þ ¼ 2:40; p50:05 for low VWM/ high CLK group; t1 (7) = 1.97, p = 0.09, t2 (36) = 2.34, p < 0.05 for high VWM/ low CLK group. At the preceding nominative phrase position, only a main effect of scrambling was significant and all groups showed longer reading times in scrambled sentences. Thus, the longer reading time at the final verb of scrambled sentences compared to canonical sentences can be interpreted as a spill-over effect. Along this interpretation, high VWM/ high CLK readers had less difficulty with the gap-filling process than the other readers. It should be noted that the Head-driven model might expect the effects of VWM capacity in canonical sentences. However, the results did not show any difference among the four groups in canonical sentences. This suggests that readers do not just rehearse each word until facing a verb but they process each word incrementally.
3.5 General Discussion In the present study, we compared the VWM and Two Factor models of individual differences in sentence processing. Results clearly favored the Two Factor model: individual difference in CLK is a predictor of sentence level processing and its contribution is independent of VWM capacity. If we take the VWM model, we can only explain individual differences in processing distant scrambling sentences (Experiment 3). The Two Factor model can capture individual differences in processing both types of constructions in the three experiments. High CLK readers took a shorter time processing sentences containing lowfrequency words (Experiment 1) and scrambled word order (Experiments 2 and 3) than low CLK readers. This could mean that CLK and frequency are language-experience factors and readers’ processing costs for sentence level processing are reduced as they accumulate the linguistic knowledge by language-related experience. Interestingly, CLK mediated lexical frequency effect and structural frequency effect on reading time and comprehension accuracy. These results are consistent with MacDonald et al. (1994) who claimed that lexical- and structural-frequency are treated in the same fashion. Miyamoto and Nakamura (2005) also demonstrated that readers expected the null subject
3 Individual Differences in Sentence Processing
63
structure more than the scrambled word order structure when they read the object nouns of the scrambling sentences. According to their studies, increase in reading time for scrambling sentences in our study may reflect reader’s surprise about unexpected subject nouns. We assumed that one’s CLK was the factor that was influenced by one’s language related activity. Thus, high CLK readers might attenuate the surprise effect due to their experience. In this study, effects of VWM capacity were observed only in distant scrambling sentences. Two types of demands on VWM were assumed in scrambling sentences: temporal syntactic ambiguity (MacDonald et al., 1992) and filler-gap dependency (Miyamoto & Takahashi, 2002). The costs for keeping multiple interpretations of syntactic structures would be expected to appear before the object nouns were presented, while the filling-a-gap process should be observed in reading object nouns. The results suggest that both of these processes interacted with individual differences in readers’ VWM capacity. However, we found no VWM effects in the adjacent scrambling sentences in Experiment 2. This suggests that if the gap and its filler are close enough, even low VWM readers can process scrambling sentences. We also examined whether the processes affected by CLK and VWM share the same cognitive resources. We found an interaction between CLK and VWM in reading the scrambling sentences of Experiment 3: high VWM/ high CLK readers showed shorter reading time in scrambled sentences than the other three groups. Thus, it seems that these cognitive factors share the same resources. When we comprehend a sentence, the processing demand is affected by one’s CLK. The cost in this stage consumes resources in one’s VWM. Then, we store the intermediate and final products of comprehension using the remaining VWM. In Experiments 1 and 2, the effect of CLK but not VWM capacity, was found. These results were also consistent with the resource sharing hypothesis. Even low VWM capacity readers do not have difficulties unless the CLK process exhausts the resources in one’s VWM capacity. Finally, we discussed the possibility of a specific demand on processing headfinal structure. The Head-driven model assumes readers store all arguments preceding a verb into VWM. The Incremental model assumes that each argument is processed one by one. Our results favor the incremental model because individual differences in VWM relate to reading time in scrambling sentences but not in canonical sentences. In addition, contribution of VWM capacity is likely to be related to maintenance of multiple interpretations for the sentence structure (i.e., Capacity hypothesis, Just & Carpenter, 1992). Since our results were obtained in the region preceding verbs, the Verb Plausibility hypothesis of VWM (MacDonald et al., 1994) cannot explain our results. Individual differences in CLK were also found to have significant contribution to lexical and structural frequency effects as discussed above. Other studies (e.g., Kamide et al., 2003) suggested that readers incrementally process sentences in head-final languages. The current study showed that individual differences in cognitive factors such as one’s VWM capacity and CLK are critical for utilization of information preceding verbs.
64
N. Jincho and R. Mazuka
References Altmann, G., & Kamide, Y. (1999). Incremental interpretation at verbs: restricting the domain of subsequent reference. Cognition, 73(3), 247–264. Amano, S., & Kondo, T. (1999). NTT database series: Lexical properties of Japanese (I). Tokyo: Sanseido. Amano, S., & Kondo, T. (2001). NTT database series: Lexical properties of Japanese (VII). Tokyo: Sanseido. Baddeley, A. D. (1986). Working memory. New York: Oxford University Press. Daneman, M., & Carpenter, P. A. (1980). Individual differences in working memory and reading. Journal of Verbal Learning and Verbal Behavior, 19, 450–466. Daneman, M., & Merikle, P. M. (1996). Working memory and language comprehension: A meta-analysis. Psychonomic Bulletin and Review, 3(4), 422–433. Dixon, P., LeFevre, J., & Twilley, L. C. (1988). Word knowledge and working memory as predictors of reading skill. Journal of Educational Psychology, 80, 465–472. Inoue, A., & Fodor, J. D. (1995). Information-paced parsing of Japanese. In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing (pp. 9–63). Hillsdale, NJ: Lawrence Erlbaum Associates. Jincho, N., Namiki, H., & Mazuka, R. (2008). Effects of verbal working memory and cumulative linguistic knowledge on reading comprehension 1. Japanese Psychological Research, 50(1), 12–23. Just, M. A., & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in working memory. Psychological Review, 99(1), 122–149. Kamide, Y., Altmann, G. T. M., & Haywood, S. L. (2003). Prediction and thematic information in incremental sentence processing: Evidence from anticipatory eye movements. Journal of Memory and Language, 49, 133–156. King, J., & Just, M. A. (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language, 30, 580–602. Kintsch, W., & van Dijk, T. A. (1978). Toward a model of text comprehension and production. Psychological Review, 85, 363–394. MacDonald, M. C., Just, M. A., & Carpenter, P. A. (1992). Working memory constraints on the processing of syntactic ambiguity. Cognitive Psychology, 24, 56–98. MacDonald, M. C., Pearlmutter, N. J., & Seidenberg, M. S. (1994). Lexical nature of syntactic ambiguity resolution. Psychological Review, 101, 676–703. Mazuka, R., Itoh, K., & Kondo, T. (1997). Processing down the garden path in Japanese: Processing of sentences with lexical homonyms. Journal of Psycholinguistic Research, 26, 207–228. Mazuka, R., Itoh, K., & Kondo, T. (2002). Costs of scrambling in Japanese sentence processing. In M. Nakayama (Ed.), Sentence processing in East Asian languages. Stanford, CA: CSLI Publication. 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. Miyamoto, E. T., & Nakamura, M. (2005). Unscrambling some misconceptions: A comment on Koizumu and Tamaoka (2004). Gengo Kenkyu, 128, 113–130. Miyamoto, E. T., & Takahashi, S. (2002). Sources of difficulty in processing scrambling in Japanese. In M. Nakayama (Ed.), Sentence processing in East Asian languages. Stanford, CA: CSLI Publication. Osaka, M. (1998). Reading and working memory. In N. Osaka (Ed.), Reading: Information processing of brain and mind. Tokyo: Asakura shoten. Perfetti, C. A. (1988). Verbal efficiency in reading ability. In G. E. MacKinnon, T. G. Waller & M. Danamen (Eds.), Reading research: Advances in theory and practice (pp. 109–143). New York: Academic Press.
3 Individual Differences in Sentence Processing
65
Pritchett, B. L. (1991). Head position and parsing ambiguity. Journal of Psycholinguistic Research, 20, 251–270. Rayner, K., & Duffy, S. A. (1986). Lexical complexity and fixation times in reading: effects of word frequency, verb complexity, and lexical ambiguity. Memory and Cognition, 14, 191–201. Stanovich, K. E., & West, R. F. (1989). Exposure to print and orthographic processing. Reading Research Quarterly, 24, 402–433. Trueswell, J. C., Tanenhaus, M. K., & Garnsey, S. M. (1994). Semantic influences on parsing: Use of thematic role information in syntactic ambiguity resolution. Journal of Memory and Language, 33, 285–318. Ueno, M., & Kluender, R. (2003). Event-related brain indices of Japanese scrambling. Brain and Language, 86, 243–271. Waters, G. S., & Caplan, D. (1996). Processing resource capacity and the comprehension of garden path sentences. Memory and Cognition, 24(3), 342–355. Yamashita, H. (2002). Scrambled sentences in Japanese: Linguistic properties and motivations for production. TEXT, 22, 597–633.
Part II
Prosody and Processing
Chapter 4
Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension – ERP Evidence from German Ambiguous DPs Petra Augurzky and Matthias Schlesewsky
4.1 Introduction A central idea within the field of sentence comprehension has been the assumption that parsing proceeds in an incremental fashion. Thus, every newly incoming item is thought to be immediately integrated and assigned an interpretation. Incremental processing has been claimed to be operative even in head-final languages, where relations between nominal elements have to be established in the absence of the sentence’s main information carrier, i.e., before the clause-final verb has been encountered (e.g., Gorrell, 1995). One of the first languages that has been experimentally examined with respect to head-final processing is German (e.g., Bader & Lasser, 1994). Besides having main-clause subject-verb-object (SVO) surface order, German also has subordinate clause SOV orders and has sometimes been considered to be underlyingly verb-final (e.g., Bach, 1962; Bierwisch, 1963). In addition to being inherently verb-final, German has the property of allowing a flexible ordering of its sentential arguments, comparable to languages such as Turkish, Russian, or Japanese. Consider the following comparison between German and English, the latter being a language with a rigid word order. (1)
English (rigid word order) a. He knew that Patrick ignored Paula. b. He knew that the man saw the spy.
(2)
German (flexible word order) a. Er wusste, dass Patrick Paula ignorierte. He knew that Patrick-AMB Paula-AMB ignored ‘He knew that Patrick ignored Paula.’
P. Augurzky (*) Department of Linguistics, University of Tu¨bingen, Wilhelmstr. 19, 72074 Tu¨bingen, Germany e-mail:
[email protected]
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_4, Ó Springer ScienceþBusiness Media B.V. 2011
69
70
P. Augurzky and M. Schlesewsky
b. Er wusste, dass der Mann den Spion sah. He knew that the-NOM man the-ACC spy saw. ‘He knew that the man saw the spy.’ Er wusste, dass den Mann der Spion sah. He knew that the-ACC man the-NOM spy saw. ‘He knew that the spy saw the man.’ In the English examples in (1), the first encountered nominal element has to be interpreted as the subject of the sentence, i.e., as the performer of the action described by the sentential event. By contrast, the examples in (2) demonstrate that, contrary to languages like English, word order information is not always informative with respect to the integration of currently processed elements in German: The first nominal element can either be interpreted as the sentential subject or as the object. If morphological information is ambiguous, as in (2a), both an SOV as well as an OSV reading is therefore, in principle, available. In (2b), case information on the determiner of the first DP distinguishes between the two readings: An initial nominative case marker unambiguously reflects an SOV word order, whereas an initial accusative reflects an OSV word order. Behavioral studies such as self-paced reading experiments or off-line tasks revealed that there is a general preference for the SOV interpretation, which has been termed the ‘‘subject preference’’ (Crocker, 1994; Schriefers, Friederici, & Ku¨hn, 1995, for German data; see also de Vincenzi, 1991, for Italian data, or Frazier & Flores d’Arcais, 1989, for evidence in Dutch). In the past few years, research on head-final processing has benefited from the emergence of neurolinguistic techniques, such as event-related potentials or ERPs. By isolating language-related activity from on-going brain impulses, ERPs reflect a precise on-line record of the brain’s reaction to linguistic information as the sentence unfolds over time (for an overview upon this technique, see Coles & Rugg, 1995, or Osterhout, 1994). In addition to providing a high temporal resolution, ERPs also add a further dimension to behavioral measures by systematically correlating with different linguistic aspects. For example, the above mentioned subject preference has been found to elicit different ERP components depending on a range of factors, such as animacy of the DPs (Frisch & Schlesewsky, 2001; Philipp, BornkesselSchlesewsky, Bisang, & Schlesewsky, 2008), their realization as a noun or as a pronoun (Schlesewsky, 2003), or the specific kind of case the object bears (e.g., Bornkessel, McElree, Schlesewsky, & Friederici, 2004; see Bornkessel & Schlesewsky, 2006, for a review). This diversity suggests that distinct underlying mechanisms may have been subsumed under the notion of a ‘‘subject preference’’. The first ERP component that has been associated with the processing of object-initial orders can be observed when unambiguous case information is already available at the first encountered argument in the verb-final clause, such
4 Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension
71
as in (2b) above. In examples like these, a so-called ‘‘scrambling negativity’’ at the position of the accusative case-marked argument reflects increased processing costs as opposed to its nominative-initial counterpart (see Ro¨sler et al., 1998; Bornkessel, Schlesewsky, & Friederici, 2002; Schlesewsky, Bornkessel, & Frisch, 2003). Besides German, the scrambling negativity has also been observed to occur in other verb-final languages like Turkish (Demiral, Schlesewsky, & Bornkessel-Schesewsky, 2008) and Japanese (Wolff, Schlesewsky, Hirotani, & Bornkessel-Schlesewsky, 2008, see also Ueno & Kluender, 2003, for the observation of a scrambling-related negativity). As this component has been shown to be absent under certain circumstances (i.e., in pro-drop constructions in Turkish, see Demiral et al., 2008, or for initial dative objects in German, see Bornkessel et al., 2002), it has been suggested that the effect does not reflect a processing disadvantage for initial objects per se but rather a preference for structures involving one argument over a two-argument structure. More precisely, based on economy considerations, the processing system is thought to adhere to a one-argument reading until conflicting accusative case information forces it to abandon this preference. Thus, the scrambling negativity is associated with the costliness of a two-argument structure rather than with a general object-initiality in the examples above (see Bornkessel & Schlesewsky, 2006, for discussion). A second ERP component that has been associated with the subject preference is the N400. This component is usually found on the second argument when an initially assigned subject reading is no longer possible and has therefore been interpreted as signaling reanalysis towards the dispreferred word order. This effect has been observed, for example, in Japanese (see the sentences in (3), taken from Wolff, Schlesewsky, & Bornkessel-Schlesewsky, 2007). (3)
a. 二週間前 判事は 大臣を two weeks ago judge-TOP minister-ACC ‘Two weeks ago, the judge invited the minister.’
招きました invited
b. 二週間前 判事は 大臣が two weeks ago judge-TOP minister-NOM ‘Two weeks ago, the minister invited the judge.’
招きました invited
In the examples in (3), the initial argument is ambiguous between a subject and an object reading, as the topic marker ‘‘-wa’’ is compatible with both alternatives. According to the subject preference, this element is incrementally interpreted as the subject. When subsequently processing a nominative case-marked second argument, the initially preferred reading is no longer available, yielding to a reanalysis-related negativity (for a discussion on the N400 as a reanalysis component, see Haupt et al., 2008). Based upon the Japanese data, this effect was interpreted as reflecting the parser’s preference
72
P. Augurzky and M. Schlesewsky
for structures involving only one participant of an intransitive event over a transitive reading involving two participants.1 Comparable to the scrambling negativity, this effect is considered to be the result of economy considerations. However, unlike the former component, the N400 reflects relational processing between the two arguments rather than phrase-structural processes. Note that a very similar process is found in German (Bornkessel, Fiebach, & Friederici, 2004): (4)
Ja¨ger gesehen hat. a. . . ., welcher Ga¨rtner [. . .] den . . ., which-NOM gardener [. . .] the-ACC hunter seen has . . ., which gardener saw the hunter b. . . ., welchen Ga¨rtner [. . .] der Ja¨ger gesehen hat. . . ., which-ACC gardener [. . .] the-NOM hunter seen has . . ., which gardener the hunter saw
When comparing ERPs on the second case marker in (4), an N400 was observed for the subject-initial word order (4a). Following the above mentioned rationale, the parser prefers to interpret the initial argument as signaling an intransitive event. When accusative case information rules out an intransitive reading, a transitive event has to be established already at the position of the initial argument in (4b). When the parser subsequently arrives at the second argument, no additional processing cost arises for this condition. By contrast, an initial nominative as in (4a) is compatible with both a transitive and an intransitive reading. As the establishment of a transitive interpretation is associated with increased processing cost, an N400 is observed at the position of the second accusative case-marked argument. In sum, the presence of distinct ERP components in head-final constructions indicates that what has been subsumed under the notion of a ‘‘subject preference’’ might be related to distinct processing mechanisms. First, the presence of a scrambling negativity signals increased processing cost related to the phrasestructural level. By contrast, the N400 has been found to co-occur with processing difficulties related to the interpretive level, namely the costliness of the establishment of a transitive event. Both mechanisms can be seen as being instances of an economy-based parsing strategy that ensures minimal dependencies in incremental processing (i.e., the Minimality principle, Bornkessel & Schlesewsky, 2006, p. 790). As the above mentioned examples have illustrated, case information can be used for effectively determining argument interpretation in head-final 1
Note that in this study, sentences were presented auditorily. As nominative case-marked second arguments are also compatible with a reading in which the second DP is the subject of a subordinate clause, the effect only showed up when the prosodic manipulation ruled out a subordinate clause interpretation.
4 Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension
73
languages and constructions. However, even in morphologically elaborate languages like German, case information can sometimes be ambiguous, as the example in (2a) has demonstrated. Besides case, prosodic phrasing has been found to guide ambiguity processing in spoken language (e.g., Price et al., 1991). From a neurolinguistic perspective, very little is known about the influence of prosodic phrasing on incremental processing when verbal information is not yet available. As initial data in Japanese suggests, prosodic phrasing may affect phrase-structural minimality, analogous to case. When sentences comparable to those in (3) were presented with unambiguously case marked (accusative or nominative) first arguments, the presence or absence of a scrambling negativity crucially depended on the prosodic realization of the sentences. As Japanese is a pro-drop language, an initial accusative is compatible with a one-argument structure if this interpretation is permitted by prosodic phrasing. Indeed, a scrambling negativity for an initial object was only found when a prosodic phrase boundary was following the first argument, a realization incompatible with a pro-drop reading (see Wolff et al., 2008, for details). To summarize, it has been observed that prosodic phrasing might incrementally affect minimality in head-final languages like Japanese. Analogous to case information, prosodic phrasing affected the occurrence of the scrambling negativity, which has been shown to systematically correlate with the processing of structures involving two (or more) arguments. Given this background, the question arises whether prosodic phrasing also affects the above introduced preference for a minimal-event reading. More precisely, the present experiment was designed to investigate whether prosody is able to constrain reanalysis processes toward a transitive reading, similar to case. In order to examine the interplay between prosodic phrasing and case information in German verb-final constructions, we constructed DP1-DP2-V ambiguities, in which the second DP can either modify the first DP, leading to an intransitive reading (5a), or be an argument of the clause-final verb, leading to a transitive reading (5b): (5) a. . . ., dass der
Arzt der Patientin / des Patienten schla¨ft. . . ., that the-NOM doctor the-AMB patient / the-GEN patient sleeps-INTR . . ., that the doctor of the patient sleeps b. . . ., dass der Arzt der Patientin / dem Patienten hilft. . . ., that the-NOM doctor the-AMB patient / the-DAT patient helps-TR . . ., that the doctor helps the patient
In the examples in (5), sentences are initially ambiguous until they are disambiguated by transitivity information on the clause-final verb, or they are
74
P. Augurzky and M. Schlesewsky
locally disambiguated by case information on the second determiner. Whereas a genitive case maker immediately signals an intransitive reading, a dative case marker signals a transitive reading. From a prosodic perspective, it has been shown that verbal arguments are not to be separated by a prosodic phrase boundary from their verbal heads, thus making a DP1-#-DP2-V realization (where the ‘‘#’’ diacritic stands for a prosodic phrase boundary) more appropriate for the two-participant reading. Extending the argument rule toward nominal arguments, a DP1-DP2-#-V phrasing might be more appropriate for the one-participant reading (see Gussenhoven, 1992; see also Watson, Breene, & Gibson, 2006). We were therefore interested in the question whether an early boundary separating the DPs would be able to introduce a two-participant reading, analogous to overt case marking. Prosodic as well as case-related effects should be reflected in an N400 on the second determiner. Note that besides the above mentioned strategy of a minimality preference, constructions like those in (5) may also be constrained by the so-called ‘‘argument preference’’. Largely based upon data from behavioral experiments on PP attachment, it has been claimed that when the parser faces optionality, it preferably interprets incoming elements as arguments and not as adjuncts (see Abney, 1989; Clifton, Speer, & Abney, 1991; Schu¨tze & Gibson, 1999). Importantly for the present purposes, previous ERP experiments on sentences with ambiguous PPs such as in (6) revealed increased processing costs for the clause-final integration of facultative elements (6b) as indexed by an N400 (e.g., Juranek, 2005). (6)
a. Gestern wollte der Berater die Grafik auf der Leinwand abbilden. Yesterday wanted the advisor the drawing on the canvas displayDITRANS ‘Yesterday, the advisor wanted to display the drawing on the canvas.’ b. Gestern wollte der Berater die Grafik auf der Leinwand u¨berpru¨fen. Yesterday wanted the advisor the drawing on the canvas check-TRANS ‘Yesterday, the advisor wanted to check the drawing on the canvas.’
Therefore, our experiment additionally examined whether a similar argument preference is found when processing the clause-final verb in German DP1-DP2-V constructions and whether this effect is modulated by prosody.
4.2 ERP Study In the present ERP study, we explored whether prosodic phrasing incrementally affects the processing of ambiguously case-marked DPs in German verb-final constructions. Moreover, we examined whether the parser’s choice is guided by a preference for minimal events. As argued above, we expected that an
4 Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension
75
intransitive reading is maintained as long as possible, i.e., until contradicting case or prosodic information is encountered. Finally, we were interested in the processes that guide the clause-final integration of the DPs with the logical structure of the verb. As shown in previous experiments (e.g., Juranek, 2005), increased integrational effort might be expected at the clause-final verb for arguments as opposed to adjuncts. Table 4.1 summarizes all experimental conditions. Sentences were either disambiguated locally by means of case information or clause-finally via verbal transitivity information. In the local disambiguation condition, the second determiner was either dative case marked and biased toward a two-argument reading, or it was a genitive and thus biased toward a one-argument reading. In the first case, the second DP had to be interpreted as a dative argument of the clause-final transitive verb (conditions A and B). By contrast, if the determiner of DP2 was genitive case marked (conditions C and D), the second DP had to be interpreted as being in a possessive relation with DP1. In this case, the whole DP complex was interpreted as one verbal argument. Both interpretations are temporarily available for conditions E to H, Table 4.1 Experimental stimuli used in the ERP study. The ‘‘#’’ diacritic is indicative of a prosodic phrase boundary Condition Sentence . . ., dass der Optiker # dem Sa¨nger ha¨ufig hilft. . . ., that the optician # the-DAT singer often helps-TRANS . . . that the optician often helps the singer B LTL . . ., dass der Optiker dem Sa¨nger # ha¨ufig hilft. . . ., that the optician the-DAT singer # often helps-TRANS . . . that the optician often helps the singer C LIE . . ., dass der Optiker # des Sa¨ngers ha¨ufig schnarcht. . . ., that the optician # the-GEN singer often snores-INTRANS . . ., that the optician of the singer often snores D LIL . . ., dass der Optiker des Sa¨ngers # ha¨ufig schnarcht. . . ., that the optician the-GEN singer # often snores-INTRANS . . ., that the optician of the singer often snores E NTE . . ., dass der Optiker # der Sa¨ngerin ha¨ufig hilft. . . ., that the optician # the-AMB singer often helps-TRANS . . ., that the optician often helps the singer F NTL . . ., dass der Optiker der Sa¨ngerin # ha¨ufig hilft. . . ., that the optician the-AMB singer often helps-TRANS . . ., that the optician often helps the singer G NIE . . ., dass der Optiker # der Sa¨ngerin ha¨ufig schnarcht. . . ., that the optician # the-AMB singer often snores-INTRANS . . ., that the optician of the singer often snores H NIL . . ., dass der Optiker der Sa¨ngerin # ha¨ufig schnarcht. . . ., that the optician the-AMB singer # often snores-INTRAN . . ., that the optician of the singer often snores Conditions are coded with respect to DISAMBIGUATION (first letter: L = local; N = non-local), TRANSITIVITY (second letter: T = transitive; I = intransitive) and PROSODY (third letter: E = early boundary; L = Late boundary). The disambiguating region is indicated by underlining.
A
LTE
76
P. Augurzky and M. Schlesewsky
where disambiguation is realized on the clause-final verb, which is either transitive (conditions E and F) or intransitive (conditions G and H). A female speaker of German was instructed to realize prosodic phrase boundaries either at the region between the two DPs (early boundary condition, conditions A, C, E and G), or at the region following the second DP (conditions B, D, F and H). Whereas the early boundary condition should be associated with a transitive reading, the late boundary condition should bias toward an intransitive reading. Both prosodic variants were recorded for transitive and intransitive sentences, which were either disambiguated locally on the determiner, or clause-finally on the verb. The session was recorded in an acoustically shielded chamber using a DAT recorder. The material was digitized (44.1 kHz, 16 bit sampling rate/ mono) using Cool Edit software (Syntrillum). As there were 40 experimental items in each condition, a set of 320 sentence pairs was recorded. Our hypotheses were as follows: first, if there is an initial preference for an intransitive-event reading, increased processing costs in terms of an N400 are expected at the position of the second determiner if this reading has to be revised. This process might either be caused by overt case information, or, in the absence of morphological information, by a prosodic phrase boundary. Thus, if an early phrase boundary separates the two DPs, an N400 component on the second determiner is expected for dative case and for case-ambiguous sentences but not for genitive case. For the late boundary condition, an effect is only expected for dative determiners. Second, if there is a clause-final integration difficulty for adjuncts per se, we expect an N400 component on the verb for intransitive as opposed to transitive conditions. If integration processes are modulated by prosodic phrasing, the N400 might be even more pronounced when prosody supports an adjunct reading. Finally, we expect an increased integrational effort for sentences that are disambiguated clause-finally as opposed to sentences that have already been disambiguated at the determiner.
4.2.1 Acoustic Analyses Before entering the ERP study, experimental items were analyzed with respect to their acoustic properties in order to ensure that our speaker produced perceptually reliable prosodic contrasts. Sentences were analyzed separately with respect to their durational and F0 properties. As pre-final lengthening (= durational increase of the element preceding a prosodic phrase boundary) and a rise in F0 of the final part of the element preceding a prosodic phrase boundary are considered as being indicative of a prosodic phrase boundary (e.g., Vaissie`re, 1983; Cruttenden, 1986; Fe´ry, 1993), we examined the regions around the potential boundary positions. For the durational analysis, complement clauses were divided into six constituents: complementizer, D1, N1 (including any following pause), D2, N2 (including any following pause),
4 Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension
77
adverb, final verb. All sentences were triggered in Praat (http://www.fon.hum. uva.nl/praat) by using a special script. Afterwards, average durational values over all items were calculated for each fragment. For statistical purposes, one-factorial ANOVAs with the factor PROSODY (early boundary vs. late boundary) were carried out for each sentential region. For F0 analyses, both nouns were divided into seven intervals relative to their complete length, resulting in eight measurement points of mean F0 values. Onefactorial ANOVAs were then carried out for F0 values on each of the eight measurement points. The results of the durational analyses are given in Table 4.2. Descriptively, both prosodic variants were clearly distinguishable with respect to durational cues. As expected, the length of the element preceding the boundary (i.e., N1 for the early boundary condition and N2 for the late boundary condition) was increased. As in the case of the second noun, this increase might as well affect elements following the boundary. These descriptive observations were confirmed by statistical analyses (see Table 4.2). Comparable to the durational results, F0 values of the elements preceding the respective boundary regions were significantly increased. Table 4.3 presents
Table 4.2 Mean durational values (in ms) of the single constituents within the complement clause. Durational values of the nouns include any following silent phase Region Early boundary Late boundary F p comp D1 N1 D2 N2 Adverb Verb
77.9 97.1 722.4 132.7 542.6 378.7 687.0
75.6 83.1 476.3 136.6 768.1 415.5 665.1
2.43 39.94 920.83 1.21 831.12 39.94 13.44
> 0.1 < 0.001 < 0.001 > 0.2 < 0.001 < 0.001 < 0.01
Table 4.3 Mean F0 values (in Hz) for each of the single measurement points of N1 and N2 N1 N2 Prosody Significance Prosody Significance E L F p E L F p Onset 177.7 173.0 0.60 > 0.5 173.4 170.9 .02 > 0.8 p1 177.0 181.8 4.20 < 0.05 180.0 177.2 3.7 < 0.07 p2 189.3 188.2 0.29 > 0.5 181.5 196.5 10.23 < 0.01 p3 208.0 193.5 5.53 < 0.05 192.6 222.0 49.04 < 0.001 p4 225.8 199.8 9.87 < 0.01 211.5 229.6 12.38 < 0.01 p5 239.9 190.7 42.45 < 0.001 216.9 242.7 15.95 < 0.001 p6 241.8 192.4 131.79 < 0.001 222.3 245.6 54.06 < 0.001 Offset 208.6 170.9 83.01 < 0.001 209.7 235.5 29.05 < 0.001 Conditions are coded with respect to PROSODY (E = Early boundary; L = Late boundary)
78
P. Augurzky and M. Schlesewsky
mean F0 values for the respective conditions, as well as the results of the onefactorial analysis of variance. Whereas F0 values at the end of N1 were increased for the early boundary condition (consistent results from point 3 until the end of the constituent), F0 at the final part of N2 was increased for the late boundary condition (consistent results from point 2). In sum, our speaker produced reliable acoustic cues indicative of a prosodic phrase boundary that was either realized at an early position (separating DP1 from DP2) or at a later position (separating the two DPs from the following material). Boundaries were characterized by a rise in F0 and a durational increase of the element preceding the boundary.
4.2.2 Method 4.2.2.1 Participants Twenty-four German volunteers (12 male, 12 female) participated in the ERP study. They were paid for their participation and had normal or corrected-tonormal vision. Participants’ mean age was 25.1 years (range 21–29 years). 4.2.2.2 Materials A 222 factorial design was used: DISAMBIGUATION (local vs. non-local) vs. TRANSITIVITY (transitive vs. intransitive) vs. PROSODY (early vs. late boundary). Nouns and clause-final verbs were matched for frequency employing values from the Wortschatz data base (http://www.wortschatz.uni-leipzig.de). A leadin matrix sequence, e.g., Hans hat erza¨hlt. . . (‘‘Hans has told. . .’’) preceded the verb-final complement clause. Complement clauses were always introduced by the complementizer dass (‘‘that’’), followed by two DPs. Then a VP-adverbial like manchmal (‘‘sometimes’’) followed and then the clause-final verb, which was either transitive or intransitive. Sentences were either produced with an early prosodic phrase boundary separating the two DPs or with a late boundary following. A set of 40 sentences per condition was constructed. The total number of critical items was 320. 4.2.2.3 Procedure Volunteers were seated in a dimly-lit, soundproof booth in front of a 17’’ computer screen. Sentences were presented auditorily via loudspeakers and appeared in a pseudo-randomized fashion. The experimental session was divided into eight blocks, with breaks between each block. Each block consisted of 40 experimental trials. At the beginning of each trial, a fixation cross appeared in the center of the screen for the time the auditory stimulus was presented. Then a 1000 ms pause
4 Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension
79
followed and a question mark appeared on the screen, signaling that participants now had to judge the acceptability of the sentence by pressing one of two buttons on a handheld answer box. Sentences were expected to vary in their acceptability as items with cooperating as well with non-cooperating prosody were presented. The keys for acceptable and unacceptable answers were counterbalanced across participants. The above timeout for making the acceptability judgment was 2000 ms. Following the judgment, a blank screen appeared for 1000 ms and a single probe word was shown on the screen. Participants were then requested to decide whether this word had (in 50%) or had not been (50%) included in the preceding sentence by pressing a button on the answer box. Again, the above timeout for the decision was 2000 ms and answer buttons were counterbalanced. A short practice session preceded the experimental session.
4.2.2.4 ERP Recordings The EEG was continuously recorded from 26 AgAgCl-type scalp electrodes (the labeling is adopted by the nomenclature of the American Encephalographic Society (Sharbrough et al., 1991)): AFz (Ground), F7, F3, Fz, F4, F8, FC5, FC1, FCz, FC2, FC6, Cz, CP5, CP1, CPz, CP2, CP6, P7, P3, Pz, P4, P8, POz, O1, O2. In order to control for eye movement artifacts, the EOG (electrooculogram) was recorded bipolar. The horizontal EOG (EOGH) was recorded from two electrodes placed on the outer canthus of each eye, while the vertical EOG (EOGV) was recorded from two electrodes placed above and below the right eye. The activity over the right mastoid was recorded and re-referenced to linked mastoids off-line. Electrode impedances were kept below 4 kO. Signals were amplified and recorded continuously with a frequency of 250 Hz. For the analyses, raw EEG data were filtered off-line with a 0.3-20.0 Hz band pass filter in order to exclude slow signal drifts. For illustrative purposes, EEGs were filtered off-line using a 7 Hz low-pass filter for the figures below.
4.2.2.5 Data Analysis For the behavioral data, repeated measures analyses of variance with the within-subject factors DISAMBIGUATION (2), TRANSITIVITY (2) and PROSODY were calculated. For the acceptability judgment task, mean acceptability ratings and reaction times were statistically analyzed. For the probe detection task, mean error rates and reaction times were computed and then analyzed. Incorrectly answered trials were excluded from reaction time and EEG analyses. Before entering the ERP analysis, data were automatically and manually screened in order to exclude trials with EOG and muscular artifacts, as well as amplifier-saturation artifacts. Data per participant and per condition were aggregated from the onset of the critical element (D2 and the clause-final verb) to 1000 ms post onset. Afterwards, grand averages were calculated over all participants. The averages were aligned to a 200 ms pre-stimulus baseline.
80
P. Augurzky and M. Schlesewsky
EEG data were analyzed statistically by carrying out a repeated-measures ANOVA for mean amplitude values within time windows per condition, which were chosen based on the visual inspection of the data. Four regions of interest (ROIs) were introduced into the analysis of variance: left anterior (Roi 1: F3, F7, FC1, FC5), right anterior (Roi 2: F4, F8, FC2, FC6), left posterior (Roi 3: CP1, CP5, P3, O1) and right posterior (Roi 4: CP2, CP6, P4, O2). Statistical analyses were carried out in a hierarchical manner and only significant interactions (p < 0.05) were resolved. Corrected p-values (Huynh & Feldt, 1970) were chosen when the analysis involved more than one degree of freedom in the numerator. 4.2.2.6 Results Behavioral Data Mean acceptability judgment ratings and corresponding reaction times as well as mean error rates and reaction times for the probe detection task are given in Table 4.4.
Condition
Table 4.4 Behavioral results in the ERP experiment. Acceptability judgment Probe detection Acceptability Reaction times Correct responses (%) (ms) (%)
Reaction times (ms)
LTE 84.1 822.8 94.3 1128.3 LTL 71.4 828.5 89.4 1144.8 LIE 74.2 846.7 92.4 1133.6 LIL 81.0 792.6 93.9 1139.4 NTE 84.8 834.8 91.8 1134.3 NTL 66.0 823.8 90.0 1165.8 NIE 67.4 844.1 89.6 1178.8 NIL 77.1 807.3 92.2 1161.6 Conditions are coded with respect to DISAMBIGUATION (first letter: L = local; N = non-local), TRANSITIVITY (second letter: T = transitive; I = intransitive) and PROSODY (third letter: E = early boundary; L = Late boundary).
Acceptability Judgment Task Repeated-measures ANOVAs on acceptability ratings with the factors DISAMBIGUATION, TRANSITIVITY and PROSODY revealed a main effect of DISAMBIGUATION ðF1 ¼ 9:65; p50:01; F2 ¼ 40:77; p50:001Þ, indicating an advantage for sentences that were disambiguated locally. Moreover, an interaction of TRANSITIVITY PROSODY was observed ðF1 ¼ 7:98; p50:01; F2 ¼ 218:55; p50:001Þ. Transitive constructions with an early boundary were judged as being more acceptable than with a late prosodic boundary (F1 = 17.22; p < 0.001; F2 = 207.28: p < 0.001). Moreover, intransitive constructions were judged as being more
4 Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension
81
acceptable when accompanied by a late boundary (F1 = 5.11; p < 0.05; F2 = 46.34; p < 0.001). Finally, a three-way interaction between DISAMBIGUATION, TRANSITIVITY PROSODY was found ðF1 ¼ 6:56; p50:05; F2 ¼ 7:56; p50:01Þ. When sentences with an early boundary were disambiguated locally, transitive constructions were judged more acceptable ðF1 ¼ 5:60; p50:05; F2 ¼ 20:30; p < 0.001). By contrast, locally-disambiguated targets with a late boundary showed an advantage for intransitive constructions ðF1 ¼ 4:07; p50:06; F2 ¼ 17:62; p < 0.001). When sentences with an early boundary were disambiguated clause-finally, again an advantage for transitive constructions was visible ðF1 ¼ 13:62; p50:01; F2 ¼ 59:93; p50:001Þ. Finally, ambiguous sentences with a late boundary showed an advantage for an intransitive construction ðF1 ¼ 4:87; p50:05; F2 ¼ 13:67; p50:01Þ. The analysis of reaction times showed a main effect of PROSODY ðF1 ¼ 10:05; p50:01; F2 ¼ 5:81; p50:05Þ, indicating that participants’ decisions were faster for sentences with a late boundary. Probe Detection Task The analysis of error rates in the probe detection task revealed a main effect of DISAMBIGUATION, which was significant in the subject analysis (F1 = 8.00; p< 0.05 F2 = 2.25; p > 0.1). Generally, participants made more errors when sentences were disambiguated at the clause-final region. Moreover, an interaction between TRANSITIVITY PROSODY ðF1 ¼ 22:15; p50:001; F2 ¼ 4:97; p50:05Þ indicates that error rates were increased for misleading prosody. Participants made more errors when transitive constructions were accompanied by a late prosodic phrase boundary than by an early boundary ðF1 ¼ 15:04; p50:001; F2= 3.4; p = 0.06). Moreover, errors were increased when intransitive constructions were accompanied by an early prosodic boundary ðF1 ¼ 7:40; p< 0.01; F2 = 1.61; p>0.1). The reaction times for the probe detection showed a main effect of DISAMBIGUATION ðF1 ¼ 47:03; p50:001; F2 ¼ 5:81; p50:05Þ – generally, participants made less errors when sentences were disambiguated locally. In addition, an effect of PROSODY was observed ðF1 ¼ 5:08; p50:05; F2 = 5.81; p < 0.05). Moreover, an interaction between TRANSITIVITY PROSODY was found ðF1 ¼ 14:9; p50:001; F2 ¼ 9:3; p50:01Þ. Whereas decisions were faster for transitive constructions when accompanied by an early prosodic boundary than when accompanied by a late prosodic boundary ðF1 ¼ 12:86; p < 0.01; F2 = 6.55; p < 0.05), no such prosodic effect was observed for the intransitive constructions ðF1 ¼ :75; p40:3; F2 ¼ 2:70; p40:1Þ. ERP Data D2 Figures 4.1 and 4.2 show grand averages on the determiner of the second DP. While Fig. 4.1 shows conditions with an early phrase boundary, Fig. 4.2 shows conditions with a late phrase boundary. For each prosodic variant, case-
82
P. Augurzky and M. Schlesewsky
F7
F3
FZ
F4
F8
FC2
FC1 CZ
CP5
CP1
CP2
CP6
PZ P3
P4
O1 Early boundary ambiguous dative genitive
O2
−5 µV s 0.5 5
Fig. 4.1 Grand average ERPs from the onset of D2 in the early boundary condition with ambiguous (dotted), dative (dashed) and genitive (solid) case marking. Negativity is plotted upwards
ambiguous conditions were subsumed under one label, as they do not differ physically on the determiner. As is descriptively apparent, conditions differ only in the early boundary condition: A broadly-distributed negativity is observed for both dative casemarked determiners and ambiguous determiners as opposed to genitives. No comparable effect was observed in the late-boundary condition. The descriptive observations were confirmed by statistical analyses, which were carried out in a 330–450 ms time window. The repeated measures ANOVA with the factors DISAMBIGUATION, TRANSITIVITY, PROSODY and ROI revealed a significant effect of DISAMBIGUATION ðF ð1; 23Þ ¼ 12:03; p50:01Þ, as well as a significant effect of PROSODY ðF ð1; 23Þ ¼ 21:08; p50:001Þ. Moreover, an interaction between ROI and PROSODY was observed ðFð1; 23Þ ¼ 24:51; p50:001Þ, indicating a frontal distribution of the prosodic effect (Roi 1: F(1,23) = 28.91; p < 0.01; Roi 2: F(1,23) = 31.11; p < 0.001); Roi 3: Fð1; 23Þ ¼ 1:22; p40:05; Roi 4: Fð1; 23Þ ¼ 2:15; p40:05). Finally, a three-way interaction between DISAMBIGUATION, TRANSITIVITY and PROSODY was found ðFð1; 23Þ ¼ 4:84; p50:05Þ. In the early boundary condition, an effect of DISAMBIGUATION was observed ðFð1; 23Þ ¼ 24:04; p < 0.001), as well as an interaction between DISAMBIGUATION and TRANSITIVITY
4 Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension
F7
F3
FZ
83
F8
F4
FC2
FC1 CZ
CP5
CP1
CP2
CP6
PZ P3
P4
O1
O2
−5
µV
Late boundary ambiguous dative genitive
s 0.5 5
Fig. 4.2 Grand average ERPs from the onset of D2 in the late boundary condition with ambiguous (dotted), dative (dashed) and genitive (solid) case marking. Negativity is plotted upwards
ðFð1; 23Þ ¼ 4:20; p ¼ 0:05Þ. Resolving the interaction towards TRANSITIVITY showed an effect of DISAMBIGUATION for the intransitive condition ðFð1; 23Þ ¼ 25:30; p50:001Þ but no effect for the transitive condition ðFð1; 23Þ ¼ 10:45; p50:001Þ. No significant effects were observed for the late boundary condition (all Fs < 1.3). In sum, effects on the second determiner were only observed for the early boundary condition. An N400-like negativity for ambiguously case-marked determiners and dative case-marked determiners as opposed to genitives reflects increased processing cost for these conditions. Clause-Final Verb Figure 4.3 shows the comparison between transitive and intransitive clausefinal verbs, whereas Fig. 4.4 shows the comparison between sentences that were disambiguated locally vs. non-local disambiguation. Finally, Fig. 4.5 shows ERPs for the interaction between verbal transitivity and prosodic phrasing. Descriptively, conditions differed in terms of a negative deflection with a maximum at around 400–600 ms. Statistical analyses were carried out in an 450–600 ms time window.
84
P. Augurzky and M. Schlesewsky
F7
F3
FZ
F4
F8
FC2
FC1 CZ
CP5
CP1
CP2
CP6
PZ P3
P4
O1
O2
−5 µV
Transitivity transitive intransitive
s 0.5
1.0
5
Fig. 4.3 Grand average ERPs from the onset of the clause-final verb for transitive (solid) vs. intransitive (dotted) verbs. Negativity is plotted upwards
The repeated measures ANOVA with the factors DISAMBIGUATION, TRANSIPROSODY and ROI revealed a main effect of TRANSITIVITY (F(1,23) = 19.30; p50:001Þ, as well as an interaction between TRANSITIVITY and ROI (F (1,23) = 8.23; p50:001Þ. Resolving the interaction towards ROI showed that the effect was more significant in posterior regions Fð1; 23Þ ¼ 8:38; p50:01; Roi 2: Fð1; 23Þ ¼ 8:70; p50:01; Roi 3: Fð1; 23Þ ¼ 22:70; p50:001; Roi 4: Fð1; 23Þ ¼ 26:45; p50:001). Moreover, a more frontally-distributed effect of PROSODY was observed ðFð1; 23Þ ¼ 20:83; p50:001Þ, as indexed by an interaction between PROSODY and ROI: Fð1; 23Þ ¼ 6:83; p50:01; Roi 1: F (1,23) = 18.81; p < 0.001; Roi 2: Fð1; 23Þ ¼ 26:23; p50:001; Roi 3: Fð1; 23Þ ¼ 3:70; p ¼ 0:06; Roi 4: Fð1; 23Þ ¼ 10:23; p50:01. Finally, an interaction between DISAMBIGUATION and ROI was observed ðFð1; 23Þ ¼ 12:31; p50:01Þ, indicating a posterior effect (Roi 1: Fð1; 23Þ ¼ 0:08; p40:9; Roi 2: Fð1; 23Þ ¼ 0:02; p40:8; Roi 3: F = 5.92; p < 0.05; Roi 4: Fð1; 23Þ ¼ 11:65; p50:01). Finally, a marginally significant interaction between ROI, TRANSITIVITY and PROSODY was found (F(1,23) = 2.64; p 50:07Þ. However, resolving the interaction did not lead to any significant interactions within the single Rois (all F values < 1.5). To sum up, an N400 effect reflected integration difficulties at the clause-final verb. A posterior effect for TRANSITIVITY shows increased processing cost
TIVITY,
4 Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension
F7
F3
FZ
F4
85
F8
FC2
FC1 CZ
CP5
CP1
CP2
CP6
PZ P3
P4
O1
O2
−5
Ambiguity late early
µV s 0.5
1.0
5
Fig. 4.4 Grand average ERPs from the onset of the clause-final verb for late (dotted) vs. early (solid) disambiguation. Negativity is plotted upwards
associated with the clause-final processing of facultative elements as opposed to verbal arguments. Moreover, a processing advantage for sentences that were already disambiguated on the determiner was mirrored in a posterior N400 for the late disambiguation condition. As our behavioral results indicated that there is a strong mismatch effect for transitive conditions with a late boundary and intransitive conditions with an early boundary, we carried out an additional analysis with the factors DISAMBIGUATION, MATCH (Match vs. Mismatch), TRANSITIVITY and ROI. Besides the above-reported effects for the factors DISAMBIGUATION and TRANSITIVITY, we also found an interaction between MATCH and TRANSITIVITY ðFð1; 23Þ ¼ 20:83; p50:001Þ. Whereas a TRANSITIVITY effect was observed for the matching conditions ðFð1; 23Þ ¼ 37:36; p50:001Þ, no such effect was obtained for the mismatches ðFð1; 23Þ ¼ 0:30; p50:5Þ An interaction between ROI, MATCH and TRANSITIVITY ðFð1; 23Þ ¼ 6:82; p50:01Þ revealed that the mismatch effect was more significant in frontal regions (Roi 1: Interaction MATCH and TRANSITIVITY: Fð1; 23Þ ¼ 18:81; p50:001; effect of TRANSITIVITY for the matching condition: Fð1; 23Þ ¼ 24:69; p50:001; for the mismatch condition Fð1; 23Þ ¼ 1:71; p ¼ 0:2; Roi 2: Interaction MATCH and TRANSITIVITY: Fð1; 23Þ ¼ 26:26; p50:001; effect of TRANSITIVITY for the matching condition:
86
P. Augurzky and M. Schlesewsky
F7
F3
FZ
F4
F8
FC2
FC1 CZ
CP5
CP1
CP2
CP6
PZ P3
Conditions O1 early boundary, transitive early boundary, intransitive late boundary, transitive late boundary, intransitive
P4
O2
−5
µV 0.5
s 1.0
5
Fig. 4.5 Grand average ERPs from the onset of the clause-final verb for transitive vs. intransitive verbs with an early or a late boundary. Negativity is plotted upwards
Fð1; 23Þ ¼ 25:24; p50:001; for the mismatch condition F(1,23) = 3.28; p > 0.06; Roi 3: Interaction MATCH and TRANSITIVITY: Fð1; 23Þ ¼ 3:70; p50:06; Roi 4: Interaction MATCH and TRANSITIVITY: Fð1; 23Þ ¼ 10:24; p ¼ 0:01; effect of TRANSITIVITY for the matching condition: Fð1; 23Þ ¼ 44:20; p50:001; for the mismatch condition Fð1; 23Þ ¼ 4:41; p50:05) To summarize, ERP effects differed between early and later processing regions. In the absence of any verbal information, i.e., on the position of the second determiner, argument processing was guided by a preference for a one-participant reading, which was reflected in an N400 when either prosodic or case information of the second DP biased toward a transitive reading. This effect was restricted on the early boundary condition. By contrast, in accordance with previous experiments on the argument-adjunct distinction, increased integrational effort was observed on the clause-final verb for adjuncts as opposed to arguments. This effect was even more pronounced when prosodic phrasing additionally biased toward an adjunct interpretation.
4.2.3 Discussion The present ERP study investigated the interplay between prosodic phrasing and case information in the processing of head-final structures in German. To
4 Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension
87
this end, we aimed to explore whether prosodic phrasing is able to trigger reanalysis processes on nominal elements if an initially assigned intransitive structure can no longer be maintained. Analogous to former studies in German (Bornkessel et al., 2002) and in Japanese (Wolff et al., 2007), we expected locally increased processing costs for a transitive event interpretation to be reflected in an N400 component. Moreover, we investigated whether the often reported ‘‘argument preference’’ was observed at the cause-final region and whether this effect was modulated by the prosodic manipulation. For examining these issues, we presented DP1-DP2-V ambiguities, which can be either interpreted as a possessive construction (intransitive event) or as a two-participant construction (transitive event). Sentences were disambiguated locally on the second DP by case information, or were disambiguated clausefinally, by means of verbal transitivity information. Based upon former approaches on the prosody of verbal argument structure (e.g., Gussenhoven, 1992; Watson et al., 2006), we hypothesized that these constructions might be constrained by prosodic phrasing. Whereas a prosodic phrase boundary separating the two DPs should be associated with a twoparticipant reading, a prosodic phrase boundary following DP2 should rather bias toward an intransitive possessive reading. Overwhelmingly, our behavioral data supported this prediction: First, acceptability rates were drastically decreased when a two-argument construction was accompanied by a late prosodic phrase boundary, as well as when a one-argument construction was accompanied by an early prosodic phrase boundary. At the same time, correct performance for the probe detection task was decreased when prosodic information was not in accord with verbal transitivity. Thus, we concluded that prosodic phrasing might lead to the perception of a mismatch between transitivity and prosody. With respect to the question of incrementality in head-final structures, we first analyzed ERPs from the onset of the second determiner. Based upon current data in Japanese (Wolff et al., 2007), we expected to observe increased processing cost on the second determiner if either case information or prosodic information is not in accord with the initially kept intransitive reading. Considering the early boundary condition, such an effect was indeed observed: When a prosodic phrase boundary preceded the second DP, an N400 was found either for unambiguously case-marked datives, or for case-ambiguous determiners. By contrast, no such effect was observed for the genitive casemarked determiner. These results are clearly in accord with previous data in German that suggest that the establishment of a (non-minimal) two-participant reading is reflected in an N400 component (e.g., Bornkessel et al., 2002). Note that a further implication of the obtained pattern is that unambiguous case information is somehow a more reliable cue for the processor than prosodic phrasing – the prosodic effect was restricted to those instances where case information was ambiguous. This effect is probably not surprising, as prosodic phrasing can induce a bias toward one of the two readings, as reflected by the pattern of behavioral results. A mismatch between prosodic phrasing and
88
P. Augurzky and M. Schlesewsky
transitivity thus leads to the perception of a decrease in acceptability. However, contrary to case information, prosodic phrasing is not able to lead to the perception of ungrammaticality: (7)
a.
?
b.
*
dass der Optiker dem Sa¨nger # ha¨ufig hilft. that the optician the-DAT singer # often helps-TRANS ‘that the optician often helps the singer’ dass der Optiker dem Sa¨nger ha¨ufig schnarcht that the optician the-DAT singer often snores, ‘that the optician often snores the singer’
We would therefore like to suggest that the prosodic influence is restricted to case-ambiguous constructions in German.2 Against our hypotheses, we did not observe any effect on the determiner for the late-boundary condition. However, if overt case information is a valid cue for introducing transitive events, we would have expected an N400 for the datives in the late-boundary condition as well. Without any further evidence, we can only speculate about the absence of an effect at this position. One possible answer could be that the clear intonational break preceding the critical region was responsible for the more pronounced effect in the early-boundary condition. By contrast, the absence of a prosodic break might have forced the effect to occur with a slight delay. Indeed, there seems to be an indication of a negativity for the dative condition in Fig. 4.2 around 500 ms post onset. However, additional data and a comparison with visually presented ERP experiments might be needed in order to explain the weakness of the effect. In general, the observed pattern on the determiner is compatible with former behavioral data on German DP1-DP2-V ambiguities (see Konieczny et al., 1997). In an eye-tracking experiment, targets comparable to ours were presented visually. Analogous to the present results, an early advantage for the transitive reading was observed (in the eye-tracking experiment, the analysis region included the first noun and the determiner). The authors of this study suggested that the principle of Head Attachment is responsible for the effect. According to this principle, attachment to an already processed head is preferred over an attachment to upcoming heads. As one potential attachment site, namely the nominal head of DP1, has already been processed when arriving at the second determiner, the second DP is preferably attached to the nominal head. Our data are clearly compatible with such an explanation, though we prefer a transitivity-related approach based upon the qualitative aspect of the obtained ERP effect. It has been repeatedly found that attachment-related processing difficulties are reflected by the occurrence of late positivities rather 2
Note that this might also explain the fact that there is no processing difficulty for unambiguous genitives on the determiner.
4 Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension
89
than negative ERP deflections (see, for example, Carreiras, Salillas, & Barber, 2004; Kaan & Swaab, 2003). Thus, the finding of an N400 component seems to be better compatible with a minimal event interpretation. However, as former ERP studies on attachment ambiguities exclusively deal with the processing of optional elements, a direct comparison to the Konieczny et al. data might be problematic. Further studies examining the interplay between word order and the current manipulation might be needed in order to address these questions. Finally, in accordance with previous studies on argument-adjunct ambiguities (e.g., Clifton et al., 1991; Juranek, 2005; Schu¨tze & Gibson, 1999;), we observed increased processing cost for the clause-final integration of adjuncts as opposed to verbal arguments. This effect was reflected in an N400 for intransitive over transitive verbs, which is in accord with previous ERP studies on the argument-adjunct distinction. As expected, an interaction between prosody and verbal transitivity was observed. Whereas transitive constructions accompanied by an early prosodic phrase boundary were easiest to process, intransitive constructions accompanied by a late prosodic phrase boundary were most difficult, as indexed by an increased N400 component. Both ‘‘mismatch conditions’’, i.e., transitive constructions with a late boundary, as well as intransitive constructions with an early boundary, occupy a position between those extremes. Therefore it seems to be the case that the processor is sensitive to the amount of information characterizing an element as an argument or an adjunct. The more explicitly an adjunct is characterized as such (i.e., by prosodic and verbal information together), the more difficult it is to integrate. In sum, the present results can be taken as further evidence for an incremental parsing strategy according to which ambiguous nominal elements are preferably interpreted as the single participants of an intransitive event, a mechanism that has been sometimes related to the so-called ‘‘subject preference’’. The present findings also motivate the dissociation of processing routines related to the hierarchizing of potential arguments on the one hand and verbal integration on the other hand – contrary to the finding of a local advantage for intransitive events, a general advantage for arguments over adjuncts was observed at the clause-final verb.
4.3 Conclusion The present ERP study demonstrated that prosodic phrasing information plays a crucial role in head-final sentence comprehension. As our findings indicate, prosody might incrementally affect parsing decisions when neither configurational information nor case information offers reliable cues with respect to the processing of ambiguous elements. Moreover, our data suggest that, analogous to languages like Japanese, German exhibits a local tendency to interpret ambiguous DPs preferably as participants of an intransitive event, as long as
90
P. Augurzky and M. Schlesewsky
no contradictory information is encountered. Finally, we replicated the wellknown finding of an argument preference, which has been often reported for clause-final verbal integration.
References Abney, S. (1989). A computational model of human parsing. Journal of Psycholinguistic Research, 18, 129–144. Bach, E. (1962). The order of elements in a transformational grammar of German. Language, 38, 263–269. Bader, M., & Lasser, I. (1994). German verb-final clauses and sentence processing: Evidence for immediate attachment. In C. Clifton Jr., L. Frazier & K. Rayner (Eds.), Perspectives on sentence processing (pp. 225–242). Hillsdale, NJ: Lawrence Earlbaum. Bornkessel, I., Fiebach, C. J., & Friederici, A. D. (2004). On the cost of syntactic ambiguity in human language comprehension: An individual differences approach. Cognitive Brain Research, 21, 11–21. Bornkessel, I., McElree, B., Schlesewsky, M., & Friederici, A. D. (2004). Multi-dimensional contributions to garden path strength: Dissociating phrase structure from case marking. Journal of Memory and Language, 51, 495–522. Bornkessel, I., Schlesewsky, M., & Friederici, A. D. (2002). Grammar overrides frequency: Evidence from the online processing of flexible word order. Cognition, 85, B21–B30. Bornkessel, I., & Schlesewsky, M. (2006). The extended argument dependency model: A neurocognitive approach to sentence comprehension across languages. Psychological Review, 113, 787–821. Bierwisch, M. (1963). Grammatik des deutschen Verbs. Berlin: Akademie Verlag. Carreiras, M., Salillas, E., & Barber, H. (2004). Event-related potentials elicited during parsing of ambiguous relative clauses in Spanish. Cognitive Brain Research, 20, 98–105. Clifton, C., Speer, S., & Abney, S. (1991). Parsing arguments: Phrase structure and argument structure as determinants of initial parsing decisions. Journal of Memory and Language, 30, 251–271. Coles, M. G. H. & Rugg, M. D. (1995): Event-related brain potentials: An introduction. In M.D. Rugg & M. G. H. Coles (Eds.), Electrophysiology of mind: Event-related brain potentials and cognition (pp. 1–26). New York: Oxford University Press. Crocker, M. W. (1994). On the nature of the principle-based sentence processor. In C. Clifton Jr., L. Frazier & K. Rayner (Eds.), Perspectives on sentence processing (pp. 245–266). Hillsdale, NJ: Erlbaum. Cruttenden, A. (1986). Intonation. Cambridge: Cambridge University Press. Demiral, S¸. B., Schlesewsky, M., & Bornkessel-Schlesewsky, I. (2008). On the universality of language comprehension strategies: Evidence from Turkish. Cognition, 106, 484–500. de Vincenzi, M. (1991). Syntactic parsing strategies in Italian. Dordrecht, The Netherlands: Kluwer. Fe´ry, C. (1993). German intonational patterns. Tu¨bingen: Niemeyer. Frazier, L., & Flores d’Arcais, G. B. (1989). Filler-driven parsing: A study of gap filling in Dutch. Journal of Memory and Language, 28, 331–344. Frisch, S., & Schlesewsky, M. (2001). The N400 indicates problems of thematic hierarchizing. Neuroreport, 12, 3391–3394. Gorrell, P. (1995). Syntax and parsing. Cambridge, England: Cambridge University Press. Gorrell, P. (1996). Parsing Theory and phrase-order variation in German V2 clauses. Journal of Psycholinguistic Research, 25(1), 135–156. Gussenhoven, C. (1992). Sentence accents and argument structure. In I. M. Roca (Ed.), Thematic structure: Its role in grammar. Berlin: Foris.
4 Prosodic Phrasing and Transitivity in Head-Final Sentence Comprehension
91
Haupt, F. S., Schlesewsky, M., Roehm, D., Friederici, A. D., & Bornkessel-Schlesewsky, I. (2008). The status of subject-object reanalyses in the language comprehension architecture. Journal of Memory and Language, 59(1), 54–96. Huynh, H., & Feldt, L. S. (1970). Conditions under which the mean square ratios in repeated measurement designs have exact F distributions. Journal of the American Statistic Association, 65, 1582–1589. Kaan, E., & Swaab, T. Y. (2003). Repair, revision, and complexity in syntactic analysis: An electrophysiological differentiation. Journal of Cognitive Neuroscience, 15(1), 98–110. Konieczny, L., Hemforth, B., Scheepers, C., & Strube, G. (1997). The role of lexical heads in parsing: Evidence from German. Language and Cognitive Processes, 12(2/3), 307–348. Osterhout, L. (1994). Event-related brain potentials as tools for comprehending language comprehension. In C. Clifton, L. Frazier & K. Rayner (Eds.), Perspectives on sentence processing (pp. 15–44). Hillsdale, NJ: Lawrence Erlbaum. Philipp, M., Bornkessel-Schlesewsky, I., Bisang, W., & Schlesewsky, M. (2008). The role of animacy in the real time comprehension of Mandarin Chinese: Evidence from auditory event-related brain potentials. Brain and Language. Price, P., Ostendorf, M., Shattuck-Hufnagel, S., & Fong, C. (1991). The use of prosody in syntactic disambiguation. Journal of the Acoustical Society of America, 90(6), 2956–2970. Ro¨sler, F., Pechmann, T., Streb, J., Ro¨der, B., & Hennighausen, E. (1998). Parsing of sentences in a language with varying word order: Word-by-word variations of processing demands are revealed by event-related brain potentials. Journal of Memory and Language, 38, 150–176. Schlesewsky, M., Bornkessel, I., & Frisch, S. (2003). The neurophysiological basis of word order variations in German. Brain and Language, 86, 116–128. Schriefers, H., Friederici, A. D., & Ku¨hn, K. (1995). The processing of locally ambiguous relative clauses in German. Journal of Memory and Language, 34, 499–520. Schu¨tze, C. T., & Gibson, E. (1999). Argumenthood and English prepositional phrase attachment. Journal of Memory and Language, 40, 409–431. Ueno, M., & Kluender, R. (2003) Event-related brain indices of Japanese scrambling. Brain and Language, 86, 243–271. Vaissie`re, J. (1983). Language-dependent prosodic features. In A. Cutler & D. R. Ladd (Eds.), Prosody: Models and Measurements (pp. 53–66). Berlin: Springer. Watson, D., Breen, M., & Gibson, E. (2006). The role of syntactic obligatoriness in the production of intonational phrase boundaries. Journal of Experimental psychology: Learning, Memory, and Cognition, 32, 1045–1056. Wolff, S., Bornkessel-Schlesewsky, I., Hirotani, M., & Schlesewsky, M. (2008). The neural mechanisms of word order processing revisited: Electrophysiological evidence from Japanese. Brain and Language, 107, 133–157. Wolff, S., Schlesewsky, M., & Bornkessel-Schlesewsky, I. (2007). The interaction of universal and language-specific properties in the neurocognition of language comprehension: Evidence from the processing of word order permutations in Japanese. Journal of Cognitive Neuroscience, Supplement, 288.
Chapter 5
Production-Perception Asymmetry in Wh-scope Marking1 Yuki Hirose and Yoshihisa Kitagawa
5.1 Introduction: Computing a Wh-scope Domain in a Head-Final Language A complex interrogative clause containing a Wh-phrase in its subordinate clause such as (1) potentially allows two Wh-scope readings: (1) hokenzyo-wa [kanzya’tati-ga na’ni-o ta’beta-ka] tasika’metandesu-ka? health.dept.-TOP patients-NOM what-ACC ate-COMP-Q confirmed COMP-Q
a. Subordinate wh-scope reading: ‘Did the health department confirm [what1 the patients had eaten t1]?’ b. Matrix Wh-scope reading: ‘What1 was such that the health department confirmed [whether the patients had eaten it1]?’ Unlike head-initial languages such as English, the COMP in Japanese is located clause-finally, realizable as a question particle such as ka (COMP-Q), which is ambiguous between ‘‘Wh-COMP’’ and ‘‘whether/Yes-No-Q-COMP’’ in its function. The Wh-phrase, in this case na’ni (‘‘what’’), is located in situ within the subordinate clause. Whether this Wh-phrase takes subordinate scope or matrix scope depends on whether it is interpretively associated with a Wh-COMP Y. Hirose (*) Department of Language and Information Sciences, The University of Tokyo, 3-8-1 Komaba, Meguro-ku, Tokyo, 153-8902 Japan e-mail:
[email protected] 1
We are grateful to the participants of WPSI 3 and the Head-final Processing Conference for their invaluable comments. This material is based upon work supported by the National Science Foundation (NSF) under Grant No. 0650415 and Japan Society for the Promotion of Science (JSPS) under grant No.19720092. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation or Japan Society for the Promotion of Science.
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_5, Ó Springer ScienceþBusiness Media B.V. 2011
93
94
Y. Hirose and Y. Kitagawa
appearing in the subordinate or matrix clause. As will become crucial in our analysis below, in the left-to-right on-line processing, the subordinate in-situ Wh-phrase is processed prior to either of the candidate COMP-Qs with which the Wh-phrase needs to be licensed. The matrix Wh-scope reading tends to be more difficult to obtain, or marginally acceptable and hence rarely encountered, while the subordinate Wh-scope reading is easily obtained for most native speakers. The marginal status of the matrix Wh-scope reading has often been taken to be an instance of Subjacency effects in the syntax literature since Nishigauchi (1990) and Watanabe (1992). It has been claimed, in other words, that a Wh-in-situ within a Wh-island cannot be associated with the matrix Wh-COMP. It is also well-known in the literature, however, that the alleged unacceptability is subtle and does not necessarily hold across speakers. In this paper, we argue against the view that the matrix Wh-scope interpretation of Wh-questions such as (1) above should be regarded as ungrammatical. Based on the production and the comprehension study, we will point out that there arises discrepancy between the way the speakers encode the scope information in the prosodic structure and the way the listeners identify the syntactic structure through the prosodic cues. We will further argue that this mismatch comes from the processing strategy under incrementality pressure and that it results in apparently more error-inducing processing when the Wh-item takes matrix scope.
5.2 Prosody and Processing of Wh-scope Ambiguity 5.2.1 Relationship Between Wh-scope and Prosody It has been discussed in the recent literature that Wh-questions in Japanese are accompanied by an appropriate prosodic (pitch) contour that marks the Wh-focus with an elevated pitch, followed by a pitch reduction domain that continues up to the end of the Wh-scope domain (Fujisaki & Kawai, 1988; Kori, 1989; Maekawa, 1991). The domain of the Wh-scope can therefore be disambiguated by this pitch reduction domain, in particular, by where it ends. Deguchi & Kitagawa (2002) and Ishihara (2003) pointed out that, at least for some speakers of Tokyo Japanese, the Wh-scope domain correlates with the domain of focus prosody (henceforth FPD). The two alternative Wh-scope readings in (2) below (= (1)) contrast in the end position of FPD wherein the Wh-focus receives pitch prominence (indicated by bold-face capitals), followed by post-focal pitch-range reduction (henceforth PFR; indicated by underlines). The end of FPD is marked by post-COMP F0 rise (whose high tones are indicated by capitals). This rise occurs at the matrix verb (e.g., tasika’meta ‘‘confirmed’’) in (2a) for subordinate Wh-scope, while no such rise takes place
5 Production-Perception Asymmetry in Wh-scope Marking
95
for matrix Wh-scope and hence PFR continues through the sentence-final matrix Wh-COMP in (2b). We will refer to the former prosodic pattern as ‘‘Local’’ FPD and the latter pattern as ‘‘Global’’ FPD.2 (2) a. Subordinate Wh-scope (Local FPD): hokenzyo-wa [kanzya’tati-ga NA’ni-o ta’beta-ka] taSIKA’metandesu-ka? health.dept.TOP patients-NOM what-ACC ate-COMP-Q confirmed COMP-Q Wh-prominence PFR Post-COMP Rise ‘Did the health department confirm [what1 the patients had eaten t1]?’
b. Matrix Wh-scope (Global FPD): hokenzyo-wa [kanzya’tati-ga NA’ni-o ta’beta-ka] tasika’metandesu-ka? health.dept.TOP patients-NOM what-ACC ate-COMP-Q confirmed COMP-Q Wh-prominence PFR ‘What1 is such that the health department confirmed [whether the patients had eaten it1]?’
For those speakers who can detect such a prosody-scope correlation, the matrix Wh-reading of (1) is a legitimate interpretation, which they can obtain when the sentence is accompanied by an appropriate prosodic pattern and an appropriate pragmatic context.3 Kitagawa and Fodor (2003, 2006) further argued that the apparent overwhelming preference for the subordinate Whscope reading for sentences such as (1) is caused by at least three independent factors. One of them is a processing constraint, which we will discuss in some detail in the next subsection. Another factor is pragmatic in nature. They pointed out that the satisfaction of the presuppositions involved in the extraction of Whscope out of a Wh-island generally requires the envisioning of somewhat unusually elaborate pragmatic context, which is typically disregarded in the syntactic tests carried out in a null discourse context. They also argued, based upon some experimental results, that Local FPD is a default prosodic pattern automatically computed and imposed on the input material during the on-line processing of a Wh-in-situ construction as in (1) in silent reading – and presumably also in an instantaneous production (Kitagawa, 2005). Hirotani (2005) defined the two patterns of prosody corresponding to the two Wh-scope readings in terms of the prosodic phrasing. The difference in the prosodic patterns depends on the presence or absence of a Major prosodic 2
FPD here corresponds to Ishihara’s FI (Focus Intonation). The terms Local FPD, Global FPD and PFR replace the terms used in Deguchi & Kitagawa (2002) such as Short EPD, Long EPD and Eradication, which were somewhat misleading. 3 Kitagawa (2005) offers an analysis in which the specific pairing of Wh-scope and FPD (‘‘Matrix scope – Global FPD’’ and ‘‘Subordinate scope – Local FPD’’) is grammatically encoded in terms of the formal features associated with Wh-phrases and COMPs in the Numeration.
96
Y. Hirose and Y. Kitagawa
phrase (MaP) boundary immediately following the subordinate COMP, in contrast to Deguchi & Kitagawa and Ishihara above, who defined the two prosodic patterns in terms of the patterns of surface pitch contours. Hirotani also argued that there is no strict one-to-one relationship between the prosodic phrasing and the Wh-scope; the subordinate scope reading requires a MaP boundary preceding the matrix verb whereas the matrix scope reading only optionally accompanies a MaP boundary in that position. The difficulty of the matrix Wh-reading could be attributed to the lack of the strict relation between its syntactic structure and the prosodic structure. What is common among these analyses on the relationship between prosody and scope interpretation is that the matrix Wh-scope reading is not ruled out outright by grammar. Yet, there is a strong preference for the subordinate Whscope reading. All the analyses above maintain that it is the end of the Wh-scope domain that is important for prosody to mark.
5.2.2 Processing Account for the Preference Towards the Subordinate Wh-scope The preference in sentence comprehension for the subordinate Wh-scope reading can in fact be naturally predicted by a more generalized version of a processing principle such as the Minimal Chain Principle (de Vincenzi, 1991) or the Active Filler Strategy (Frazier, 1987). In Wh-questions in English, with more than one possible gap position (for Wh-traces), the parser always prefers to associate the fronted Wh-phrase to the first encountered gap as demonstrated by ‘‘filled gap effects’’ (Crain & Fodor, 1985; Frazier & Clifton, 1989; Stowe, 1986). In Japanese, cases of ‘‘in-situ equivalent of the filled-gap effect’’ (Miyamoto, 2008) called typing mismatch effect (Miyamoto & Takahashi, 2002) have been reported in literature (Aoshima, Phillips, & Weinberg, 2004; Miyamoto & Takahashi, 2002; Yoshida, 2004; Ono, Yoshida, Aoshima, & Phillips, 2006). Based upon their experimental results, Miyamoto & Takahashi demonstrated a locality effect between the Whin-situ and its associated Wh-COMP while Wh-interrogative constructions in this language do not involve overt Wh-movement. In both (3a) and (3b), the entire subordinate clause that has the Wh-phrase in it is scrambled to the sentence-initial position. The trace of the fronted clause is in the object position of the matrix clause. Their results showed that people took longer to read the subordinate verb plus the COMP when the COMP is not interrogative (as in (3b)) than when it is (as in (3a)), indicating that the parser attempted to establish the binding relationship between the Wh-phrase and the first-encountered COMP. dono-pasokon-o tukatteiru-ka]1 [kakarichoo-ga (3) a. [senmu-ga director-NOM wh-computer-ACC is using-COMP-Q supervisor-NOM t1itta-no]? said-COMP-Q ‘‘Did the supervisor say what type of computer the director was using?’’
5 Production-Perception Asymmetry in Wh-scope Marking
97
(3) b. [senmu-ga dono-pasokon-o tukatteiru-to]1 [kakarichoo-ga director-NOM which-computer-ACC is using-COMP supervisor-NOM t1 itta-no]? said-COMP-Q ‘What type of computer did the supervisor say that the director was using?’ In the processing of the example in (1) above, when the Wh-phrase na’ni-o (‘‘what-ACC’’) in the subordinate clause is encountered, the parser attempts to find a Wh-COMP that can license the Wh-phrase as early in the input as possible. The Wh-COMP at the end of the subordinate clause is the first one encountered and therefore the parser should create the dependency relationship between this COMP and the subordinate Wh-in-situ. This presumably induces the preference for the subordinate Wh-scope reading. We assume that such a processing principle for locality in Wh-COMP dependency operates in the online processing of Japanese sentences and we will further investigate how this can also relate to the puzzling mismatch between the production and comprehension data we will encounter and discuss below.
5.3 Production Study4 The purpose of this experiment is to examine if the speakers indeed establish prosody-scope correlation when they utter scopally ambiguous Wh-questions and if they do, in what exact phonetic respect the two prosodic patterns are distinct from each other. The previous studies predicted that the crucial difference is to be found in the post-COMP F0 rise (Deguchi & Kitagawa, 2002; Ishihara, 2002, 2003), or the presence or absence of the Major phrase boundary preceding the matrix item (Hirotani, 2005).5 In this study, we measure the F0 peak values of post-COMP items – those of the matrix verbs in our examples, as well as several other items. We also measure durational cues (silence interval and segment length) that are relevant to marking the beginning of FPD in a preWh position and the end of the subordinate Wh-scope domain (= Local FPD) in a post-COMP position.
5.3.1 Materials Thirteen scopally-ambiguous Wh-interrogative sentences including and similar to (1) (repeated below) were created as target sentences. Each target sentence was embedded in a dialogue-like context including an answer to the target question, 4
The presented data are part of a more comprehensive study to appear in Kitagawa and Hirose (in prep). 5 In fact, other than the post-COMP F0 rise, what exact phonetic properties the presence of MaP boundaries amount to is not entirely clear.
98
Y. Hirose and Y. Kitagawa
so that the subject would assign some specific Wh-scope to the target sentence induced by the context. The example of the dialogue for the subordinate Whscope reading is shown in (4) and the matrix Wh-scope reading in (5) below. (1) Target sentence hokenzyo-wa [kanzya’tati-ga na’ni-o ta’beta-ka] tasika’metandesu ka? health.dept.TOP patients-NOM what-ACC ate-COMP-Q confirmed -COMP-Q a. Subordinate Wh-scope reading: ‘Did the health department. confirm [what1 the patients had eaten t1]?’ b. Matrix Wh-scope reading: ‘What1 was such that the health department confirmed [whether the patients had eaten it1]?’ (4) Dialogue for subordinate Wh-scope reading
(English translation) Journalist: It’s been 4 h since you started interviewing the patients of the food poisoning. We need to know whether you finally identified the cause. Spokesman: We are not ready to announce the name of the item yet. Journalist: You don’t have to tell us what it is. We simply would like to know if you have already identify the cause.<
> Spokesman: Oh, yes. That they have. (5) Dialogue for Matrix Wh-scope reading
(English translation)
5 Production-Perception Asymmetry in Wh-scope Marking
99
Journalist: I heard that you have already identified the food item that caused the mass food poisoning. Tell us what it is.<> Spokesman: It is bean sprouts. The 13 pairs of dialogues were divided into two lists. In addition to the dialogues with the target experimental sentences with Wh-scope ambiguity, 13 dialogues with various types of questions without a scope ambiguity were created as filler items and were inserted between experimental dialogues. The same filler dialogues were used in the two sets. Both sets (i.e., both versions of each experimental sentence) were to be read by all subjects.
5.3.2 Subjects (Speakers) Six native Tokyo Japanese speakers from Tokyo, Kanagawa, Saitama and Chiba participated in the production experiment. All subjects, referred to as YA(male), RA(female), SZ(female), FN(male), YI(female) and YN(male), were undergraduate and graduate students of The University of Tokyo and were unaware of the purpose of the study.
5.3.3 Procedure First, in the screening session, the subjects were shown all experimental dialogues in pairs so that they could compare each item and familiarize themselves with the two contrasting Wh-scope readings. The subjects were told to compare the two dialogues and to make sure if the target sentence (indicated by an underline) made sense in each context. It was intended to screen out the subjects who disapproved of the matrix Wh-reading in case such speakers were included. With respect to our subjects, however, all six of them agreed that the target Whquestions made sense in both versions of the context. After this screening session, the subject took a short break and then was given one of the lists for the recording session. In the recording session, they were instructed to skim through each dialogue in the list again before they started to read the sentences aloud. They were told to start only after they understood the whole dialogue. The target sentences were always indicated by an underline. When they made a speech error, they were asked to repeat from the point prior to the target (indicated by the experimenter). After they had read through all the sets of dialogues, they took a short break and repeated the same set of dialogues for the second time. They were asked to come back at least 1 day later to read the other list. This time, the subjects did not go through the preview session and went straight to the recording session (since they had checked all the experimental sentences in both versions on the first day). The procedure for the recording
100
Y. Hirose and Y. Kitagawa
session for the second set of items was the same as that for the first set. Only the second recordings for all dialogues were used for the analyses. The F0 peak values were measured at four positions of the recorded target sentences: the item immediately preceding the Wh-phrase, the Wh-phrase itself, the subordinate verb and the matrix verb – in the case of example (1), kanzya’tati-ga (‘‘patients-NOM’’), na’ni-o (‘‘what-ACC’’), ta’beta-ka (‘‘ate-COMP’’) and tasika’metandesu-ka (‘‘confirmed-COMP’’), respectively. The durational measures included the silent interval duration (if any) immediately preceding the Wh-phrase and immediately following the subordinate COMP (corresponding to the beginning and the end of the subordinate Wh-scope domain). The final-segment duration was also measured for these two items (the last vowel a in kanzya’tati-ga and ta’beta-ka, respectively, in the case of (1)). In all versions of the prosody-scope correlation of Wh-interrogatives discussed by Deguchi & Kitagawa (2002), Ishihara (2002, 2003) and Hirotani (2005) 6, the important information about the Wh-scope domain is considered to be encoded in the prosodic cue to signal the end of the focus domain. The purpose of our production study is to examine if this is true across the speakers and to investigate further if there is any other type or location of cues that is also encoded for the scopal distinction they detect.
5.3.4 Results 5.3.4.1 F0 Analyses The maximum F0 for the matrix verb for each subject is shown in Table 5.1 below with statistical details and for the Wh-phrase it is shown in Table 5.2. As is clear in Table 5.1, the F0 peak for the matrix verb was significantly higher for the subordinate Wh-scope reading for all six speakers, as expected. In addition to this, as shown in Table 5.2, the difference in F0 height was also observed at the Wh-phrase itself: it was significantly higher when the matrix Wh-scope was intended (except that the marginal difference in the opposite direction was observed for SZ). No significant difference was observed in the highest F0 at other positions for any speaker.
6
According to Hirotani (2005), the scopal difference should be, at least partially, reflected in the presence or absence of a Major prosodic boundary immediately preceding the matrix verb. In our study we do not examine the phonological status of the observed difference to answer the question as to whether the prosodic difference is a phenomenon at the surface pitch contour or whether it is directly related to the prosodic structure. See Kubozono (2007) for relevant discussion.
5 Production-Perception Asymmetry in Wh-scope Marking
101
Table 5.1 Mean peak F0 values (in Hz) of the matrix verb for the subordinate and the matrix Wh-scope readings. The t-values and the levels of significance are indicated for the comparison of F0 values between the two scope readings Subject Intended Wh scope Peak F0 t (df=12) p YA(M) RA(F) SZ(F) FN(M) YI(F) YN(M)
Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate
141 169 202 295 276 344 210 254 232 269 113 168
2.37
p< 0.05
6.49
p< 0.0001
4.80
p< 0.0005
2.53
p< 0.05
4.31
p< 0.001
9.89
p< 0.0001
Table 5.2 Mean peak F0 values (in Hz) of the Wh-phrase for the subordinate and the matrix Wh scope readings. The t-values and the levels of significance are indicated for the comparison of F0 values between the two scope readings Subject Intended Wh scope Peak F0 t (df=12) p YA(M) RA(F) SZ(F) FN(M) YI(F) YN(M)
Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate
233 201 335 291 358 366 296 252 350 308 207 186
3.59
p< 0.001
4.51
p< 0.001
2.15
p= 0.052
3.84
p< 0.005
6.92
p< 0.0001
4.74
p< 0.0001
5.3.4.2 Durational Analyses Pause Duration Silence intervals immediately preceding the Wh-domain (right before the Whphrase) and following the subordinate COMP were measured. Some of the silence intervals were noticeable pauses and some of them may have been due to the closure preceding the stop consonant at the beginning of the following item without involving any prosodically-meaningful break between phrases. Since the two Wh-scope readings arose from a single Wh-question in each pair of the stimuli, any difference in the silence duration that may arise between the two scope readings would presumably reflect the scopal difference realized in prosody.
102
Y. Hirose and Y. Kitagawa
Table 5.3 Mean duration of silence intervals in milliseconds at the two positions of the target sentences. (Standard deviations are presented in parentheses.) Subject Intended Wh scope Pre-Wh position Post-COMP position YA(M) RA(F) SZ(F) FN(M) YI(F) YN(M)
Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate
217 (133) 103 (102) 14 (0) 16 (6) 160 (160) 146 (171) 228 (110) 112 (72) 162 (178) 115 (184) 29 (68) 9 (2)
132 (130) 84 (48) 34 (1) 35 (20) 31 (16) 61 (66) 77 (91) 91 (66) 23 (10) 43 (49) 25 (24) 25 (14)
The average duration in milliseconds shown in Table 5.3 may not be very informative of the distribution of pauses. As can be seen from the large values of standard deviation in the duration measurement, some utterances were accompanied by a substantial pause while many did not exhibit any silence interval between phrases. As far as the present analysis is concerned, the only significant difference in duration that was detected between the two scope readings was at the pre-Wh position for one speaker YA and FN in which the silence duration was longer for the Matrix Wh-scope reading (t(12)=2.97, p< 0.05, t(12)=3.61, p< 0.005, respectively). Final Segment Duration The durations of the final segment (the final vowel) of the pre-Wh phrase and that of the subordinate COMP (-a as in –ka) were measured for each utterance. The mean duration values of the six speakers are shown in Table 5.4. Table 5.4 Mean duration of the phrase final segments in milliseconds at the pre-Wh phrase and the subordinate COMP Subject Intended Wh scope Pre-Wh phrase Subordinate COMP YA(M) RA(F) SZ(F) FN(M) YI(F) YN(M)
Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate Matrix Subordinate
135 (39) 92 (38) 82 (29) 80 (29) 120 (40) 80 (37) 131 (29) 109 (32) 82 (18) 80 (42) 81 (41) 70 (41)
98 (40) 96 (35) 59 (13) 54 (13) 56 (9) 81 (47) 62 (11) 77 (41) 46 (27) 50 (44) 46 (27) 47 (33)
5 Production-Perception Asymmetry in Wh-scope Marking
103
For one speaker (YA), the final segment of the pre-Wh phrase was significantly longer for the matrix Wh-scope reading (t(12)=3.59, p< 0.005). For another speaker (SZ), the final segment of the subordinate COMP in the subordinate Wh-scope reading was longer almost significantly (t(12)=1.89, p= 0.064). No other difference in duration was found in any other speaker. 5.3.4.3 Summary of the Results In summary, the prosody-scope correlation (reflecting the Post-COMP F0 rise) discussed by Deguchi & Kitagawa and Ishihara was replicated. The prosodic contrast between the two types of Wh-scope readings consistently observed across speakers was the peak F0 height on the matrix verb. The overall direction in the pattern does not at least contradict with Hirotani’s analysis, in which a Major Phrase boundary (= reset of pitch range, optionally accompanying a pause) is predicted to occur obligatorily preceding the matrix verb for the subordinate Wh-scope reading but only optionally for the matrix Wh-reading. (In our experiment, on the other hand, the F0 peak of the matrix verb was low fairly consistently across speakers when they reported matrix Wh-scope.) One interesting new finding is that the F0 contrast between the two readings was present at the Wh-phrase itself: the F0 peak of the Wh-phrase was higher in the matrix Wh-scope reading than in the subordinate reading. Since the post focal domain tends to be longer for the matrix Wh-scope reading, the contrast in the Wh-peak could be considered to simply reflect the speakers’ preparation for starting a long focus prosody domain with a higher pitch. For some speakers, other types of prosodic contrast reflecting the two distinct scopal readings were also observed. Those included silence interval preceding the Wh-phrase, duration of the final segment of the pre-Wh phrase (longer for matrix Wh-scope reading) and duration of the final segment of the subordinate COMP (longer for subordinate Wh-scope). However, durational contrasts are not consistent relative to the F0 cues and only limited to some speakers. In the subsequent analysis, we will mainly focus on the F0 contrasts, which were more robust and consistent across speakers.
5.4 Comprehension Study A comprehension study was also conducted using the subset of items recorded by two native Tokyo speakers. The aim of the study was to examine how well the speakers’ intention (regarding the Wh-scope interpretation) is conveyed to listeners based upon the prosodic patterns they encoded into their utterances.
5.4.1 Materials The same 13 pairs of items with the Wh-scope ambiguity in two versions were recorded. Since two items were excluded due to presentation error, 11 pairs of
104
Y. Hirose and Y. Kitagawa
sentences were used in the comprehension study. An additional 44 filler items of various types of interrogative sentences were prepared, half of which were ungrammatical. The target items were divided into two lists in the same way as the production study. The filler items were the same across the lists. Recordings by two Tokyo speakers were used who were distinct from the subjects of the production study.7 The speakers included YK, one of the authors (who was of course fully aware of the purpose of the study) and one graduate student SY, who was not told the purpose of the study. The F0 peak data (converted into semitones, to make the inter-speaker comparison easier) and the durational data of these recordings are presented in Tables 5.5–5.8 below. Table 5.5 Mean peak F0 values (in semitones) of the Wh-phrase and the matrix verb for the subordinate and the matrix Wh-scope readings Wh-phrase Matrix verb Speaker Intended Wh scope Peak F0 t (df=10) p Peak F0 t (df=10) p YK YS
Matrix Subord. Matrix Subord.
22.09 16.14 18.70 18.01
7.72
p< 0.0001 5.19 15.85 p> 0.5 8.03 18.53
0.64
5.03
p< 0.0001
20.01
p< 0.0001
Table 5.6 Mean duration of silence intervals in milliseconds at the Pre-Wh and the postCOMP positions of the target sentences. (Standard deviations are presented in parentheses.) Speaker Intended Wh- scope Pre-Wh position Post-COMP position YK YS
Matrix Subordinate Matrix Subordinate
68 (67) 9 (30) 30 (81) 0 (10)
0 (0) 52 (53) 14 (45) 45 (85)
Table 5.7 Mean duration of the phrase final segments in milliseconds at the pre-Wh phrase and the subordinate COMP. (Standard deviations are presented in parentheses.) Speaker Intended Wh scope Pre-Wh phrase Subordinate COMP YK YS
7
Matrix Subordinate Matrix Subordinate
127 (46) 41 (19) 75 (37) 85 (39)
74 (26) 89 (37) 41 (13) 101 (35)
This comprehension study was in fact conducted before the production study reported above. Since the items used in the comprehension study were not exactly matched with the ones used in the production study (besides one of the informants was not naı¨ ve about the purpose of the experiment), we did not include the data recorded by these two speakers in the results of the production study.
5 Production-Perception Asymmetry in Wh-scope Marking
105
Table 5.8 Percentage of the acceptance with matrix Wh-scope Speaker Intended Wh- scope Pre-Wh position YK YS
Matrix Subordinate Matrix Subordinate
30 9 12 2
As shown in Table 5.5, the highest F0 on the Wh-phrase was significantly higher and that on the matrix verb was significantly lower for the matrix reading for YK, contrary to YS’s recordings in which the F0 difference was only found for the matrix verb. For YK, the average pause duration was significantly longer following the pre-Wh phrase (t(10)=3.16, p < 0.05) and shorter following the subordinate COMP (t(10)=4.09, p < 0.005) when the matrix scope was intended. Likewise, the final segment duration of the pre-wh phrase was significantly longer (t(10)=4.09, p < 0.005) and that of the subordinate COMP was shorter (t(10)=5.27, p < 0.0005) for the matrix Wh-reading. As is obvious in the comparison between the data from YK and YS, the difference between the two Wh-scope readings is more emphasized for YK both in the F0 and the durational cues.
5.4.2 Subjects (Listeners) Twenty-eight native Tokyo speakers participated in the experiment. The subjects were divided into two groups for each ‘‘speaker block’’ (each subject only heard the target and filler utterances by one speaker, either YK or YS). Within the speaker block, the subjects were further divided into two subgroups and each group was assigned one of the two counterbalanced lists of materials.
5.4.3 Procedure The utterances were played in sequence. Each item in the written form was also projected onto the screen simultaneously so that they could check how some of the words and proper names are written in kanji in case they were not confident if they caught each word correctly. The subjects were instructed to focus on the auditory stimuli when the utterances were first played out. The subjects were then asked to make a forced-choice judgment on a possible answer (either corresponding to the matrix Wh-scope readings or the subordinate Wh-scope readings, respectively) to the question. In addition to the two kinds of possible answers, there was also an option to reject the question sentence as ungrammatical. One example of the answer set is shown below (corresponding to the example target sentence (1), repeated below):
106
Y. Hirose and Y. Kitagawa
(The example target question sentence repeated from (1)) hokenzyo-wa [kanzya’tati-gana’ni-o ta’beta-ka ] tasika’metandesu-ka? health.dept.TOP patients-NOM what-ACC ate-COMP-Q confirmed COMP-Q (The example answer set) a. The question itself was ungrammatical b. ‘‘Yes, they have already confirmed it’’ (Subordinate Wh-scope reading) c. ‘‘Bean sprouts.’’ (Matrix Wh-scope reading) The order between the two possible answers to the question (‘‘b’’and ‘‘c’’ above) was alternated across items.
5.4.4 Results 5.4.4.1 Comprehension Data The percentage of the responses in which the subjects chose an answer consistent with the matrix Wh-scope reading is shown in the Table 5.8 below. The rejection rate for the target sentences is presented in Table 5.9. Table 5.9 Percentage of the rejection of the target question sentences Speaker Intended Wh- scope Pre-Wh position YK YS
Matrix Subordinate Matrix Subordinate
19 0 7 0
For the speaker block YK, the percentage of matrix Wh-scope interpretation was significantly higher when the speaker had intended the matrix-Wh-scope reading (t1(13)=2.58, p < 0.05, t2(10)=2.81, p < 0.05,). The rejection rate of the target sentences was 19% for intended matrix Wh-scope reading, compared to 0% for intended subordinate Wh-scope reading. For the speaker block YS, the percentage of matrix Wh-scope interpretation was significantly higher when the speaker had intended the matrix-Wh-scope reading in the subject analysis, and the difference was nearly significant in the item analysis (t1(13)=2.88, p < 0.05, t2(10)=2.18, p= 0.054). The target sentences were rejected 7% of the times when the matrix Wh-scope reading was originally intended. No rejection occurred when the subordinate Wh-scope reading was intended. On the whole, the results suggest that the speakers’ intention reflected in the prosodic pattern influenced the listeners’ interpretation when the responses to the two versions of the sentences were compared to each other in relative terms. The overall proportion of the matrix Wh-scope reading, however, was strikingly low. The outcome falls in line with Hirotani (2005) except that her experimental results showed a higher acceptance rate (about 50%) of the matrix interpretation.
5 Production-Perception Asymmetry in Wh-scope Marking
107
Finally, the rejection rate of the question sentences themselves was notably higher when they were accompanied by the prosody intended for the matrix scope reading. It may indicate that the Global FPD, the prosodic pattern for the matrix Wh-scope, indeed is more marked than the Local FPD, as has been argued by Kitagawa and Fodor (2003, 2006).
5.4.4.2 Factors Contributing to the Listeners’ Decision We suspect that the poor comprehension performance for the matrix Wh-scope intended sentences (reported above) may be due to listeners not focusing on the supposedly most useful and consistently present prosodic cue, the F0 information on the matrix verb. We now would like to investigate if the listeners actually rely on some other information in processing the utterances intended for a matrix scope reading. Based on the results of our production experiment, the F0 peak on the Wh-phrase was also considered as a possible cue although it may not be consistently available across speakers. A linear regression analysis was conducted, where the percentage of correct responses (i.e., matrix Wh-scope reading) in the comprehension study was considered as the dependent variable. The F0 peak values of the matrix verb and those of the Wh-phrase were considered as independent variables. The 22 matrix-scope intended utterances made by the two speakers were analyzed together. The results are summarized in Table 5.10. Although the coverage of the data explained by the resulting function was limited, as suggested by the very low R2 value, our present data at least indicate that the F0 peak of the matrix verb is not the most important cue used by the listeners in any obvious way. Instead, the F0 peak of the Wh-phrase had a greater impact on the listeners’ judgment when such a cue was present, although the size of contribution was small in our results. Recall that the average F0 peak of the Wh-phrase did not differ between the two Wh-scope readings in the production data from YS and that the listeners’ comprehension performance for the matrix Wh-scope reading was noticeably lower when the material recorded by YS was used compared to when the utterances by YK were used (See Table 5.8). It is noteworthy, in other words, that the poor comprehension rate of the matrix Wh-scope intended sentences may be related to the lack of the F0 cue on the Wh-phrase itself. Table 5.10 Summary of Linear Regression Analysis predicting the % correct response to the Matrix Wh-scope intended utterances from the two speakers F df p R2 0.252 3.19 2,21 0.064 Variables included in the model F0 peak of Wh-phrase F0 peak of matrix verb
b 0.426 –.132
t 1.90 –.590
p 0.073 0.565
108
Y. Hirose and Y. Kitagawa
5.5 Discussion and Conclusion The results of our production study were consistent with the existing claim in the literature that prosody-scope correspondence is established for Wh-interrogatives in Japanese. When we combine these results with the outcome of our comprehension study, however, they cast doubt on the standard assumption that speakers’ encoding and the listeners’ decoding of the Wh-scope information are mediated by the same prosodic cues. In particular, the hypothesis that the F0 peak of the post-COMP item (in our examples, the matrix verb) marks the difference in the Wh-scope domains for both speakers and listeners becomes hardly sustainable. For production, the language users’ appeal to the F0 peak of the post-COMP item makes perfect sense since, for speakers, the most straightforward way to encode the subordinate Wh-scope in prosody would be to signal the end of its focus prosodic domain with the Post-COMP rise. The matrix Wh-scope, on the other hand, can be encoded by the absence of such Post-COMP Rise. Our comprehension study revealed, however, that this prosodic cue was not necessarily utilized by listeners to distinguish the two Wh-scope readings in the way the speakers had intended. At least for a considerable proportion of the cases, the listeners relied more on other cues, most notably, the relative F0 height of the Wh-item itself. This result implies that the marking of the subordinate Whscope domain by the F0 peak of the post-COMP item should not be regarded as the main factor determining the Wh-scope in sentence processing as proposed by Hirotani (2005), while it may be regarded as a production phenomenon induced by grammatical derivation as proposed by Deguchi & Kitagawa (2002) and Ishihara (2003). One remaining question is why such a mismatch between speakers and listeners arises. While the full answer to this question must await further research, we can identify some lead when we consider the listeners’ general interests in regard to the incremental processing. In on-line real time processing, upon encountering the Wh-item, listeners need to learn at which COMP the Wh-COMP dependency is meant to be resolved. For the Wh-in-situ in a subordinate clause, the first candidate COMP item would be the subordinate COMP, but there is no way for listeners to know if there is another COMP to be encountered in the subsequent input, which turns out to be the relevant COMP. The processing preference would be to resolve the Wh-COMP dependency at the earliest possible point (Miyamoto & Takahashi, 2002), which, as we pointed out above, correctly predicts the preference towards the subordinate Wh-scope reading. For an incremental parser, therefore, the decision as to whether to resolve the dependency at the subordinate COMP or to take a chance for the matrix COMP (that will hopefully be present) needs to be made when (or before) the subordinate COMP is encountered. In this regard, the most informative prosodic cue encoded by the speaker at the first matrix item following COMP, i.e., the post-COMP F0 rise, is not
5 Production-Perception Asymmetry in Wh-scope Marking
109
necessarily the most useful in on-line processing. The listener is under pressure, in other words, to make the decision about the Wh-domain in comprehension by resorting to a cue that appears as early as possible and inform them if the Wh-phrase should NOT be associated with the first encountered COMP. So for the sake of timely decision, it would be ideal if the Wh-phrase itself contained some information that could provide a signal to the parser if it should override the default local dependency analysis. This, we believe, is a highly likely source of the mismatch between speakers’ and listeners’ cues. The mismatch in the type of the prosodic cues between speakers and the listeners has also been reported by Hirose (2006). In her experiment, when the left- versus right-branching ambiguity exists in recursive NPs, the listeners showed tendency to disregard the F0 contour information (with respect to the occurrence of downstep), which the speakers systematically relate to the syntactic structure. Instead, the listeners relied more on the durational information, which occurred in a less consistent manner. This mismatch may arise from the fact that the reliability of the F0 contour information is limited to the cases involving lexical accents. As discussed by Kitagawa & Fodor (2003, 2006), the strong preference towards the subordinate Wh-scope reading in a Wh-island construction in Japanese seems to be at least partly due to the direct effect of the processing preference towards the local Wh-COMP dependency. Based upon the experimental results and theoretical conjectures presented above, we now consider that the discrepancy between speakers’ and listeners’ strategies in encoding/decoding the Whscope information in prosody induced by this processing constraint provides an additional processing cause of the subordinate Wh-scope preference and hence the controversy over the Subjacency effect in Japanese Wh-sentences.
References Aoshima, S., Phillips, C., & Weinberg, A. (2004). Processing filler-gapdependencies in a headfinal language. Journal of Memory and Language, 51, 23–54. Crain, S., & Fodor, J. D. (1985). How can grammars help parsers? In D. R. Dowty, L. Karttunen & A. M. Zwicky (Eds.), Natural language parsing: Psycholinguistic, computational, and theoretical perspectives (pp. 94–128). Cambridge: Cambridge University Press. Deguchi, M., & Kitagawa, Y. (2002). Prosody and Wh-questions. Proceedings of the Thirtysecond Annual Meeting of the North-Eastern Linguistic Society, GLSA, University of Massachusetts at Amherst, 73–92. de Vincenzi, M. (1991). Syntactic parsing strategies in Italian. Dordrecht, The Netherlands: Kluwer. Frazier, L. (1987). Sentence processing: A tutorial review. In M. Coltheart (Ed.), The psychology of reading (pp. 559–586). Hove, England: Erlbaum. Frazier, L., & Clifton, C. Jr. (1989). Successive cyclicity in the grammar and the parser. Language and cognitive processes, 4, 93–126. Fujisaki, H., & Kawai, H. (1988). Realization of linguistic information in the voice fundamental frequency contour of the spoken Japanese. Proceedings of the 1998 IEEE International Conference on Acoustics, Speech, and Signal Processing, 1, 663–666.
110
Y. Hirose and Y. Kitagawa
Hirose, Y. (2006). Missed Cues: Speaker-Hearer Mismatch and Variability. Paper presented at the 19th annual CUNY Conference on Human Sentence Processing, City University of New York, New York. Hirotani, M. (2005). Prosody and LF: Processing of Japanese Wh-questions. Doctoral dissertation, University of Massachusetts, Amherst. Ishihara, S. (2002). Invisible but audible Wh-scope marking: Wh-constructions and deaccenting in Japanese. Proceedings of the Twenty-first West Coast Conference on Formal Linguistics, Cascadilla Press, 180–193. Ishihara, S. (2003). Intonation and interface conditions. Doctoral dissertation, Massachusetts Institute of Technology, Cambridge, MA. Kitagawa, Y. (2005). Prosody, syntax and pragmatics of Wh-questions in Japanese. English linguistics, 22.2, 302–346. Kitagawa, Y., & Fodor, J. D. (2003). Default prosody explains neglected syntactic analyses of Japanese. Japanese/Korean linguistics, 12, CSLI Publication, 267–279. Kitagawa, Y., & Fodor, J. D. (2006). Prosodic influences on syntactic judgments. In G. Fanselow et al. (Eds.), Gradience in grammar: Generative perspectives (pp. 336–358). Oxford: Oxford University Press. Kitagawa, Y., & Hirose, Y. (in prep.). Production and perception of subjacency sentences in Japanese. Unpublished manuscript. Kori, S. (1989). Kyocho-to Intoneshon (Emphasis and Intonation). In M. Sugito (Ed.), Koza Nihongo-to Nihongo-kyoiku 2: Nihongo-no Onsei On’in (Vol. 1, pp. 316–342). Tokyo: Meiji-Shoin. Kubozono, H. (2007). Focus and intonation in Japanese: Does focus trigger pitch reset? In S. Ishihara, S. Jannedy & A. Schwarz (Eds.), Working Papers of the SFB632, Vol. 9. Interdisciplinary Studies on Information Structure (ISIS), (pp. 1–27). Potsdam: Universita¨tsverlag Potsdam. Maekawa, K. (1991). Perception of intonation characteristics of WH and non-WH questions in Tokyo Japanese. Proceedings of the XXIInd International Congress of Phonetic Science, 4, 202–205. Miyamoto, E. T. (2008). Processing sentences in Japanese. In S. Miyagawa & M. Saito (Eds.), The Oxford handbook of Japanese linguistics (pp. 217–249). Oxford: Oxford University Press. Miyamoto, E. T., & Takahashi, S. (2002). The processing of Wh-phrases and interrogative complementizers in Japanese. In N. M. Akatsuka & S. Strauss (Eds.), Japanese/Korean linguistics, 10, CSLI Publications 62–75. Nishigauchi, T. (1990). Quantification in the theory of grammar. Dordrecht, The Netherlands: Kluwer. Ono, H., Yoshida, M., Aoshima, S., & Phillips, C. (2006). Real-time computation of Japanese exclamatives and the strength of locality biases in sentence comprehension. Cognitive Studies: Bulletin of the Japanese Cognitive Science Society, 13, 261–287. Stowe, L. A. (1986). Parsing WH-constructions: evidence for on-line gap location. Language and Cognitive Processes, 1, 227–245. Watanabe, A. (1992). Subjacency and S-structure movement of WH-in-situ. Journal of East Asian Linguistics, 1.3, 255–291. Yoshida, M. (2004). When negative statements are easier: Processing polarity items in Japanese. Paper presented at the 17th annual CUNY Conference on Human Sentence Processing, College Park, MD.
Part III
Production of Head-final Structures
Chapter 6
The Production of Head-Initial and Head-Final Languages Mikihiro N. Tanaka, Holly P. Branigan, and Martin J. Pickering
6.1 Introduction Speakers can sometimes express the same message in many ways. For instance, an event in which a pirate knocks over a swing can be described in English using the active as ‘‘The pirate knocked over the swing’’ or using the passive as ‘‘The swing was knocked over by the pirate’’. In Japanese, the same message can be expressed in at least four ways, either as an active (1-a, 1-c) or a passive (1-b, 1-d) and with either the subject before the object (SOV; 1-a, 1-b) or the object before the subject (OSV; 1-c, 1-d): (1) a. SOV-Active 海賊がブランコを倒した。 Kaizoku-ga buranko-o taoshita. Pirate-NOM Swing-ACC knock over-PAST ‘The Pirate knocked over the swing.’ b. SOV-passive ブランコが海賊によって倒された。 buranko -ga Kaizoku -niyotte taosareta. Swing -NOM Pirate-Oblique knock over-PAST-Passive ‘The swing was knocked over by the pirate.’ c. OSV-Active ブランコを海賊が倒した。 Buranko-o Kaizoku-ga taoshita. Swing-ACC Pirate-NOM knock over-PAST ‘The Pirate knocked over the swing.’ M.N. Tanaka (*) Faculty of Arts and Sciences, Showa University, 4562 Kamiyoshida Fujiyoshida-shi, Yamanashi-ken, 403-0005, Japan e-mail: [email protected]
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_6, Ó Springer ScienceþBusiness Media B.V. 2011
113
114
M.N. Tanaka et al.
d. OSV-passive 海賊によってブランコが倒された。 Kaizoku -niyotte Buranko -ga taosareta. Pirate -Oblique Swing -NOM knock over-PAST-Passive ‘The swing was knocked over by the pirate.’ All four sentences have the same denotational meaning but clearly differ in emphasis. Speakers must choose one sentence over the others. How do they do this? Most studies that investigate these issues have been carried out in English, a largely head-initial language. They suggest that, for instance, English speakers tend to place conceptually more accessible (e.g., animate or concrete) entities earlier than less accessible ones. Depending on which entity is more accessible, such tendencies can lead to a preference to produce either passives or actives (e.g., Bock & Warren, 1985). In addition, many studies suggest that people tend to repeat the syntactic structure that they have previously used (structural priming: Bock, 1986; Pickering & Branigan, 1998; see Pickering & Ferreira, 2008). Other studies have suggested that speakers also tend to place shorter phrases earlier than longer phrases (e.g., Arnold, Wasow, Losongco, & Grinstrom, 2000; Stallings, MacDonald, & O’Seaghdha, 1998; Hawkins, 1994). However, there has been very little work on how speakers choose syntactic structure in head-final languages such as Japanese and even less comparing sentence production in head-initial and head-final languages. It is important to test both types of language, in part because some grammatical alternations occur in one type of language but not the other. In addition, the central role of the verb in production may mean that processing is fundamentally affected by the position of the verb. Hence there may be important differences in the production system between head-initial and head-final languages. Therefore, in this chapter, we examine three factors – the conceptual accessibility of referents, structural priming and the length of noun phrases – which have been hypothesized to affect production, and compare their influence on production in both types of language. We then consider how such results might inform theories of language production and in particular whether common mechanisms can be postulated for head-initial and head-final languages.
6.2 Models of Language Production Most current models of language production distinguish three different stages: conceptualization, determining what message is expressed; formulation, realizing this message as a series of linguistic representations; and articulation, converting these representations into motor movements (e.g., Levelt, 1989). Formulation is
6 The Production of Head-Initial and Head-Final Languages
115
generally divided into two different processing levels: grammatical encoding, during which the syntactic component (lemma) of appropriate lexical entries is retrieved and used to generate syntactic structure; and (morpho)phonological encoding, during which the morphophonological component of those entries is retrieved and used to generate morphological and phonological representations. Although the details of grammatical encoding are controversial (e.g., Hartsuiker, Kolk, & Huiskamp, 1999; Pickering, Branigan, & McLean, 2002), most models of production (e.g., Bock & Levelt, 1994; Garrett, 1975; Levelt, 1989) assume that it involves two different types of processing: functional processing, which maps from the conceptual representation to an unordered set of lemmas that are tagged for grammatical function (e.g., subject, direct object); and positional processing, which determines constituent structure (including word order). The processing tasks at each level are thought to be different. At the functional level, conceptual information is transformed into an unordered representation that incorporates information about grammatical function. In contrast, the positional level is a constituent-structure representation that is specified for linear order. Thus, processing at the two levels could be influenced by different types of information. Since functional processing involves the transformation of a conceptual representation into a sentence, such processing might be affected by conceptual information (e.g., animacy or imageability; Bock & Warren, 1985; McDonald, Bock, & Kelly, 1993). In contrast, positional processing deals with constituent assembly (determination of word order) and might plausibly be influenced by lexical or phonological information (e.g., Bock, 1986, 1987; Kelly, Bock, & Keil, 1986; McDonald et al., 1993). However, some researchers have claimed that conceptual representations might influence positional processing as well (Branigan & Feleki, 1999; Kempen & Harbusch, 2004; Prat-Sala & Branigan, 2000). More importantly, many models assume that language production is highly incremental (Bock, 1982; Kempen & Hoenkamp, 1987; Levelt, 1989), so that as soon as minimal information is available, it will be passed down to the next level. In this sense, the processor does not need to wait until all the information relevant for production of the entire utterance is available but can deal with fragments of information simultaneously. Incremental processing is theoretically attractive because it can explain how speakers generally manage to produce language fluently and efficiently without frequent pausing to plan the next part of their utterance. Evidence that speakers’ production processes are sensitive to the accessibility of different kinds of information (such as the accessibility of particular concepts or particular syntactic structures, as noted above) is taken as support for this assumption. Such effects are easily explained if information is processed as and when it becomes available; it is in contrast difficult to explain why accessibility affects sentence production if the processor does not proceed incrementally but rather waits until all relevant information has been retrieved before initiating further processing of any of that information.
116
M.N. Tanaka et al.
6.3 Factors Affecting Syntactic Choice in Head-Initial and Head-Final Languages Research on language production has investigated a number of factors that affect choice of syntactic structure. In this section, we consider three factors whose influence has been explored in both head-initial and head-final languages and examine the extent to which they exert comparable effects in both language types. As noted above, many researchers have suggested that language production is processed incrementally (e.g., Kempen & Hoenkamp, 1987; Levelt, 1989). That is, the processor deals with information as soon as it becomes available and at the same time that different aspects of processing take place. In this sense speakers do not need to wait until all the information is retrieved but can generate an utterance as soon as minimal input is available. In keeping with an incremental model of production, language production is hypothesized to be influenced by the relative accessibility of information. With respect to syntactic processing, relevant dimensions that might be important are the relative accessibility of concepts associated with particular entities and the relative accessibility of syntactic information.
6.3.1 Conceptual Accessibility Many researchers have suggested that conceptual representation influences grammatical encoding during production. Some evidence comes from an experimental study by Bock and Warren (1985). Bock and Warren proposed that the choice of grammatical function assignment is influenced by what they term conceptual accessibility. They claimed that conceptual accessibility influences grammatical function assignment in two ways. Firstly, the ease of word retrieval from the mental lexicon influences function assignment, so lemmas that are retrieved faster will be assigned a grammatical function before lemmas that are retrieved less quickly. Secondly, grammatical functions are assigned according to Keenan and Comrie’s (1977) Noun Phrase (NP) accessibility hierarchy, such that the subject function is assigned first, direct object second, then indirect object and oblique object. Thus lemmas that are retrieved more quickly tend to be assigned the subject function and lemmas that are retrieved less quickly tend to be assigned an object function. Such a process will partly determine the grammatical structure of the sentence (e.g., active or passive). To test this, Bock and Warren (1985) conducted a sentence recall task, asking participants to memorize sentences and recall them in a randomized order. They used imageability as an index of conceptual accessibility and demonstrated that participants tended to recall active and passive sentences in a way that allowed more concrete (hence more imageable) entities to appear as sentence-initial subjects. For example, they were more likely to recall passive
6 The Production of Head-Initial and Head-Final Languages
117
sentences with more imageable agents and less imageable patients (e.g., The shock was administered by the doctor) as active sentences (e.g., The doctor administered the shock) than to recall passive sentences with less imageable agents and more imageable patients (e.g., The doctor was administered by the shock) as actives (e.g., The shock administered the doctor). Similarly, they were more likely to recall active sentences with less imageable agents and more imageable patients as passive sentences than to recall active sentences with more imageable agents and less imageable patients as passives. In addition to this, Bock and Warren (1985) tested for effects of conceptual accessibility on noun-phrase conjuncts (e.g., The pirate and the swing were gone vs. The swing and the pirate were gone). They used this structure because the nouns that make up the conjuncts in NP conjunctions have the same grammatical function and so word order could be changed freely without any associated variations in grammatical function assignment. Bock and Warren failed to find any tendency for participants to produce more imageable nouns earlier than less imageable ones in NP conjunctions. Thus they argued that conceptual accessibility influences the choice of grammatical function assignment but not word order. Further supporting evidence for this conclusion was reported by McDonald et al. (1993), who found a tendency for animate nouns to precede inanimate nouns in the transitive sentences but not in the conjuncts. When both animate nouns and inanimate nouns appeared in the same subject position of conjunct phrases, the effect of animacy disappeared. This finding was consistent with the hypothesis by Bock and Warren (1985) where conceptual accessibility influences the choice of grammatical function assignment. However, since Bock and Warren (1985) and McDonald et al. (1993) investigated this issue in English, which has a relatively rigid order, it is hard to make a clear dissociation between the effects of grammatical function assignment and word order determination. In fact, further studies using different head-initial languages showed that conceptual accessibility can also influence the choice of word order. For instance, Branigan and Feleki (1999) tested the effect of animacy in Greek using sentence recall and found that speakers were more likely to recall sentences in a form that allowed the conceptually more accessible entity to precede the less accessible entity, irrespective of grammatical function. Thus, their participants tended to recall SVO sentences (e.g., Sta dimokratika politevmata, to sindagma sevete ton politi: ‘‘In democratic regimes, the lawNOM respects the citizen-ACC’’) as OVS sentences (e.g., Sta dimokratika politevmata, ton politi sevete to sindagma: ‘‘In democratic regimes, the citizenACC respects the law-NOM’’) more when the subject was inanimate and the object was animate, than when the subject was animate and the object was inanimate. Equally, their participants were likely to recall OVS orders (e.g., Sta dimokratika politevmata, to sindagma sevete o politis: ‘‘In democratic regimes, the law-ACC respects the citizen-NOM’’) as SVO (e.g., Sta dimokratika politevmata, o politis sevete to sindagma: ‘‘In democratic regimes, the citizen-NOM respects the law-ACC’’) more when the animate noun was the subject than when the inanimate noun was the subject.
118
M.N. Tanaka et al.
Prat-Sala and Branigan’s (2000) picture-description experiment in Spanish showed that Spanish speakers were more likely to describe target pictures using sentences that had an animate subject (e.g., La Mujer fue atropellada por el tren: ‘‘The woman was run over by the train’’) than sentences that had an inanimate subject (e.g., El tren atropello’ a la mujer: ‘‘The train ran over the woman’’). More importantly, participants also tended to produce more dislocated sentences, in which the patient appears sentence-initially preceding the subject, when the patient of the dislocated sentence was animate (e.g., A la mujer la atropello el tren: ‘‘The woman, the train ran her over’’) than when it was inanimate. Thus, despite the claim by Bock and Warren (1985) that conceptual accessibility is associated with variations in grammatical function assignment but not word order, this study provides further evidence that conceptual accessibility may influence the choice of word order in a head-initial language. Tanaka et al. (submitted) conducted two experiments in Japanese that investigated whether conceptual accessibility influences the choice of grammatical functions and word order in a head-final language, specifically Japanese. As noted above, Japanese allows two types of sentences that have the same denotational meaning and the same grammatical function assignment but that have different word orders, depending on the relative ordering of the subject and direct object (Subject-Object vs. Object-Subject order). In addition, it allows both actives and passives, as does English. Tanaka et al. (submitted) used a sentence recall task (similar to Bock & Warren, 1985), asking participants to memorize transitive sentences and recall them in a randomized order. They manipulated the animacy of the nouns as a measure of conceptual accessibility to see (1) if animacy affected the choice of word order (SOV in (1-a) or OSV in (1-c)) in Japanese and (2) if it affected the choice of both grammatical function assignment (active in (1-b) or passive in (1-d)) and word order (SOV or OSV). Thus if accessible elements tend to appear first in head-final languages, as in head-initial languages, we would expect that in the first experiment speakers would tend to produce more inverted sentences (e.g., OSV recalled as SOV) when the original sentence had an inanimate NP first. In addition to this, participants might produce reversed voice forms in the second experiment (e.g., active recalled as passive) whenever the animate NP was not the subject in the original sentence. In Tanaka et al.’s (submitted) first experiment, Japanese participants inverted the order of subject and object significantly more often in sentences presented in OSV order when the result was to place an animate entity in first position in the sentence than when the result was to place an inanimate entity in first position (26 vs. 17%); hence OSV sentences involving an animate subject (e.g., Buranko-o (acc) kaizoku-ga (nom) taoshita: ‘‘The pirate knocked over the swing’’), the pirate (nom) knocked over.) were misrecalled as SOV sentences (e.g., Kaizoku-ga (nom) Buranko-o (acc) taoshita: ‘‘The pirate knocked over the swing’’) more often than OSV sentences involving inanimate subjects. The results strongly suggest that conceptual accessibility affects word order in Japanese, in keeping with Branigan and Feleki’s (1999) findings. This is
6 The Production of Head-Initial and Head-Final Languages
119
inconsistent with Bock and Warren’s (1985) proposal, based on their finding that conceptual accessibility did not affect word order in NP conjunctions in English, that conceptual accessibility affects the assignment of grammatical functions but not of word order. In Tanaka et al.’s (submitted) second experiment, participants were again more likely to misrecall sentences in an order that allowed animate entities than inanimate entities to appear first (21 vs. 12%). In particular, they tended to recall OSV-active sentences with inanimate objects (e.g., The swing (acc), the pirate (nom) knocked over.) as SOV-active sentences (e.g., The pirate (nom) the swing (acc) knocked over.) significantly more often than to recall OSV sentences with animate objects (e.g., The pirate (acc), the swing (nom) knocked over.) as SOV-active sentences (e.g., The swing (nom) the pirate (acc) knocked over). Equally, they tended to recall OSV-passives with inanimate objects (e.g., By the swing (obl) the pirate (nom) was knocked over.) as SOV-passives (e.g., The pirate(nom) by the swing (obl) was knocked over.) significantly more often than to recall OSV-passives with animate objects (e.g., By the pirate (obl) the swing (nom) was knocked over.) as SOV-passives (e.g., The swing (nom) by the pirate (obl) was knocked over). But at the same time, participants were more likely to misrecall conceptually accessible nouns than conceptually inaccessible nouns as subjects (30 vs. 12%). When participants were presented with a sentence in which the conceptually accessible noun (here, animate entity) did not appear as the subject (e.g., The swing (nom) the pirate (acc) knocked over. or ‘‘The swing (nom) by the pirate (obl) was knocked over.’’), they tended to recall that sentence in the alternative voice, hence making the animate entity the subject, irrespective of word order (SOV/OSV). So, SOV-actives were significantly more likely to be recalled as SOV-passives and SOV-passives were significantly more likely to be recalled as SOV-actives, when this resulted in the animate entity appearing as the subject. But more interestingly, OSV-actives were significantly more likely to be recalled as SOV-passives and OSV-passives were significantly more likely to be recalled as OSV-actives, when this resulted in the animate entity appearing as the subject than when it resulted in the inanimate entity appearing as the subject; note that recalling OSV-actives as SOV-passives with animate subjects and OSV-passives as OSV-actives with animate subjects, meant that the animate entity did not appear sentence-initially. Hence, we conclude that the speakers’ tendency to recall animate entities as subjects is independent of the tendency to recall them in sentence-initial position. We can therefore conclude that conceptual accessibility (specifically animacy) influences both the choice of grammatical function assignment and choice of word order in Japanese under some circumstances and thus that conceptual accessibility affects grammatical function assignment and word order in head-final languages. This is in keeping with the finding that it affects word order in head-initial languages (though there is no unequivocal evidence that it independently affects grammatical function assignment in headinitial languages).
120
M.N. Tanaka et al.
In contrast, Tanaka et al. (submitted) found no significant effect of animacy on conjunct order in Japanese, with comparable numbers of misrecalls placing the animate noun first and the inanimate noun first (18 vs. 12%), in keeping with Bock and Warren’s (1985) and McDonald et al.’s (1993) findings. Therefore, it seems that conceptual accessibility does not strongly influence the choice of order of conjuncts in NP conjunctions in head-initial or head-final languages. Tanaka et al. argued that these differences occurred because the conjuncts involved two NPs with the same grammatical relation, whereas transitives involved two NPs with different grammatical relations. In sum, determinants of conceptual accessibility such as animacy influence the choice of grammatical functions and word order in at least some head-initial and head-final languages but do not influence the choice of conjunct order in NP conjunctions. Such evidence indicates that conceptual representations influence grammatical encoding (in particular grammatical function assignment and word order) in both head-initial and head-final languages.
6.3.2 Structural Priming In the section above, we considered how the relative accessibility of different referents might affect choice of syntactic structure in language production. But structure choice also seems to be affected by the relative accessibility of the syntactic structures themselves. Evidence for this comes from structural (or syntactic) priming effects (see Pickering & Ferreira, 2008). Many researchers have suggested that in a normal conversation, people tend to repeat themselves. There is considerable evidence that such repetition has a largely syntactic cause. Many researchers have observed that structural repetition occurs in natural dialogue. For example, Schenkein (1980) observed a repetition of sentences between speakers and listeners (when speakers said ‘‘But can you go to sleep tonight?’’ listeners said ‘‘How am I going to sleep tonight?’’). Weiner and Labov (1983) pointed out that in a corpus there was an increased amount of passive production after having produced passives (see also Estival, 1985; Tannen, 1989). These observations are also confirmed by several experimental studies. The first experimental study that investigated the syntactic priming effect was by Levelt and Kelter (1982). They asked Dutch shopkeepers either Hoe laat gaat uw winkel dicht?: ‘‘What time does your shop close?’’ or Om hoe laat gaat uw winkel dicht?: ‘‘At what time does your shop close?’’ and demonstrated that their participants were more likely to answer the questions with prepositional phrases (e.g., Om vijf uur: ‘‘at five o’clock’’) when the experimenter used a prepositional phrase than otherwise. If the question was without prepositions, participants also answered them without prepositions vijf uur: ‘‘five o’clock’’). Thus they claimed that such results may be an example of structural repetition.
6 The Production of Head-Initial and Head-Final Languages
121
The strongest evidence for structural priming comes from Bock (1986). Bock conducted a series of experiments in which participants had to recall sentences and describe pictures. She showed that participants were more likely to describe the target pictures with Prepositional object (PO) (e.g., The doctor gave the injection to the patient) after PO prime sentences (e.g., A rock star sold some cocaine to an undercover agent) than after Double object (DO) prime sentences (e.g., A rock star sold an undercover agent some cocaine). They also produced more active target descriptions (e.g., Lightning is striking the church) after active prime sentences (e.g., One of the fans punched the referee) than after passive prime sentences (e.g., The referee was punched by one of the fans). Bock argued that this effect was due to priming of the processes responsible for creating syntactic structures in sentence formulation. A further study by Bock and Loebell (1990) helped to exclude other explanations for such priming effects. For instance, locative sentences (e.g., The 747 was landing by the control tower) primed passive sentences even though the type of event was different. In addition, while Susan brought a book to Stella and Susan brought a book to study are similar in terms of the subject noun phrase, metrical structure and position of closed-class words, they differ in terms of constituent structure. Importantly, Bock and Loebell found that the former example primed another PO sentence, while the latter did not. This suggests that the priming effect relies upon the repetition of syntactic processes and/or representations (see also Bock, 1989). Many studies have reported similar effects for other constructions (Chang, Bock, & Goldberg, 2003; Ferreira, 2003; Griffin & Weinstein-Tull, 2003), using various methodologies (e.g., Branigan, Pickering, & Cleland, 2000; Pickering & Branigan, 1998; Potter & Lombardi, 1998; Smith & Wheeldon, 2001). Importantly, structural priming occurs between two different headinitial languages (Spanish-English; Hartsuiker, Pickering, & Veltkamp, 2004; Korean-English; Shin & Christianson, 2009). Thus there is good evidence that choice of syntactic structure is influenced by the relative accessibility of syntactic alternatives. However, does this also happen in a head-final language? Some of the priming studies on languages with head-final constructions such as Dutch (Hartsuiker & Westenberg, 2000) and Japanese (Tanaka, Pickering, & Branigan, 2009) suggest that it does. Hartsuiker and Westenberg (2000) used a sentence completion task to examine Dutch auxiliary verbs and past participles to investigate word order priming for both written and spoken language production in verb-final (hence, headfinal) clauses. The word orders in the Dutch sentences (2-a) and (2-b) were different but the sentences expressed the same meaning: (2) a. De man belde de politie omdat zijn portemonnee was gestolen. ‘The man called the police, because his wallet was stolen.’ b. De man belde de politie omdat zijn portemonnee gestolen was. ‘The man called the police, because his wallet stolen was.’
122
M.N. Tanaka et al.
Hartsuiker and Westenberg’s (2000) participants showed a word order priming effect, in that their target completions were influenced by the word order of the prime sentences. Specifically, they found priming for the order of the main verb and auxiliary in the verb cluster that appeared at the end of the clause. Their results therefore showed not only priming in a head-final construction but also priming for the elements within the head itself when it appeared in a final position (see also Hartsuiker & Kolk, 1998). However, Dutch is not a uniformly head-final language. Tanaka et al. (2009) used a picture-matching/-description task to demonstrate that structural priming also occurs in a uniformly head-final language, Japanese. In Tanaka et al.’s experiments, participants saw on a computer a prime picture (such as a pirate knocking over a swing), followed by a prime sentence such as (1a-d) and had to indicate if this sentence described the prime picture correctly (Branigan, Pickering, McLean, & Cleland., 2007). After participants made this judgment, they were asked to describe verbally a target picture that depicted a transitive action. Tanaka et al. (2009) conducted three experiments using this method. Two experiments in Japanese aimed to investigate at what stage voice priming (active/ passive) and word order priming (SOV/OSV) take place. A further experiment was carried out in English to compare the voice priming effect in Japanese with voice priming in English. Tanaka et al. controlled for the animacy of the subject and the object (either animate subject and inanimate object, or inanimate subject and animate object). If structural priming occurs in Japanese, speakers should show a tendency to re-use the voice (active or passive) or word order (SOV or OSV) of the prime sentence that they have previously comprehended, in their own subsequent description of the target picture. Moreover, we would expect to find comparable effects in both languages. Tanaka et al. (2009) first manipulated voice in a similar way to Bock et al.’s (1992) experiment in English. When the prime sentences were actives such as Gonin no hito-ga (nom) booto-o (acc) hakonda: ‘‘Five people carried the boat’’ , Japanese speakers were more likely to describe target pictures as actives (as in (1-b)) than after passive primes (as in (1-d)) and they were more likely to describe them as passives (e.g., Booto-ga (nom) Gonin no hito-ni (obl) hakobareta: ‘‘The boat was carried by five people’’) ) after passives (as in 1-d) than after actives (as in (1-b)). Overall, there was a 12% priming effect for choice of voice in Japanese. Tanaka et al. also conducted the same priming manipulation using English transitives (as in Bock, Loebell, & Morey, 1992) and found the same pattern of results as in Japanese passive constructions, with a similar magnitude of priming (13%). Using sentences such as (1-a) and (1-c), Tanaka et al.’s final experiment (2008) found that Japanese speakers were significantly more likely to describe target pictures using SOV orders (e.g., Gonin no hito-ga (nom) booto-o (acc) hakonda: ‘‘Five people carried the boat’’) after SOV primes (as in (1-a)) than after OSV primes (as in (1-c)) and more likely to describe them using OSV orders (e.g., booto-o (acc)Gonin no hito-ga (nom) hakonda: ‘‘Five people carried
6 The Production of Head-Initial and Head-Final Languages
123
the boat’’) after OSV primes (as in (1-c)) than after SOV primes (as in (1-a)), with an overall priming effect for choice of word order of 4%. These results suggest that priming for word order also occurs in Japanese sentence production. They are compatible with the findings of Hartsuiker and Westenberg (2000), who showed that their target completions were influenced by the word order of the prime sentences in Dutch (see also Hartsuiker et al., 1999). They are also compatible with the finding by Pickering et al. (2002), who found that people tended to repeat the PO form when they had just produced a PO form (The racing driver showed the extremely dirty and badly torn overall to the mechanic) with the constituents in the same order but not when they had just produced heavy-shifted PO (The racing driver showed to the mechanic the extremely dirty and badly torn overall). Thus this suggests that there is no tendency to repeat constituents without also repeating their orders. In addition, in all three experiments Tanaka et al. (2009) did not find any animacy effect – there was no particular tendency to place a particular animate entity to a particular position (e.g., subject or object). Their experiments therefore demonstrated structural priming in both a head-initial and a head-final language, using the same experimental task and materials. Furthermore, a similar pattern of effects was found for aspects of structure that were comparable across languages (i.e., priming for active and passive structures but no effect of animacy). In sum, then, there is strong evidence that structural priming exerts a comparable influence on syntactic choice in language production in both head-initial and head-final languages.
6.3.3 Length Effects on Word Order We have seen that at least two factors, conceptual accessibility and structural priming, influence choice of syntactic structure in comparable ways in the production of both head-initial and head-final languages. But there is other evidence that seems to suggest that there are also differences in structural preferences between head-initial and head-final languages. There is good evidence for a ‘‘short-before-long’’ preference in English (e.g., Arnold et al., 2000; Hawkins, 1994; Stallings et al., 1998). All other things being equal, English speakers tend to place shorter phrases before longer phrases. For instance, in English the heavy noun phrase in (2c) ‘‘a song that was written by a famous guitar player from Texas’’ tends to be ‘‘shifted’’ after the prepositional phrase ‘‘to friends’’ (Yamashita & Chang, 2001). However, such shifting is highly awkward when it does not involve a heavy noun phrase, as when the canonical order (3a) is ‘‘shifted’’ to (3b). (3) a. Bill sang a song with friends. b. Bill sang with friends a song. c. Bill sang with friends a song that was written by a famous guitar player from Texas.
124
M.N. Tanaka et al.
Such a ‘‘short-before-long’’ preference may reflect an influence of accessibility on syntactic choice in production, on the assumption that shorter phrases are easier to produce than longer phrases (Arnold et al., 2000; Hawkins, 1994; Stallings et al., 1998). Short phrases require the speaker to access fewer words (and typically have fewer alternative orderings among which to select). In fact, speakers tend to pause for longer and be more disfluent before producing longer than shorter utterances or constructions (Clark & Wasow, 1998; Ferreira, 1996). As we have seen, incrementality is hypothesized to be a fundamental principle of language production, so if the ‘‘short-before-long’’ preference reflects differences in accessibility, we might expect it to hold for all languages, irrespective of head-position. But research has shown that this preference does not hold for Japanese. Yamashita and Chang (2001) showed that Japanese speakers tended to correctly recall sentences involving a shorter phrase as in (4-a) but were more likely to misrecall sentences such as (4-b) by placing the longer phrase earlier. This is (4) the opposite pattern to that found in English (e.g., Hawkins, 1994). (4)
a. Keezi-ga hannin-o oikaketa. Detective-NOM suspect-ACC chased ‘The detective chased the suspect.’ b. Keezi-ga se-ga takakute gassiri sita hanninn-o oikaketa. Detective-NOM height tall big-boned suspect-ACC chased ‘The detective chased the suspect who was tall and big-boned.’
This difference in preference for Japanese, a head-final language, versus English, a head-initial language, might reflect a fundamental difference in the mechanisms that underlie the production of head-initial and head-final languages. Yamashita and Chang (2001) claimed that there are two possible explanations of this difference. Firstly, it could be related to the location of the verb (in other words, it depends on either head-initialness or head-finalness of the language). In head-initial languages such as English, word order is relatively rigid, thus syntactic constraints are important in the creation of grammatical sentences. In addition to this, in English all arguments take place after the verb and there is evidence that verbs strongly influence the linear order of phrases in English (Stallings et al., 1998). Thus syntactic constraints could be reduced, resulting in a ‘‘short-before-long’’ preference. However, in head-final languages such as Japanese, all noun phrases precede the verb. Thus it is possible that there are few syntactic constraints and people have relative freedom in how they can construct their syntactic structure and as a result there is a ‘‘long-before-short’’ preference in Japanese. Secondly, Yamashita and Chang (2001) suggested that it is possible that there are cross-linguistic differences in whether people prefer to prioritize lexical
6 The Production of Head-Initial and Head-Final Languages
125
accessibility and complex meaning/form in determining word order. For instance, as we have seen above, since long noun phrases contain rich meaning and complex form and accessibility of meaning and form can have different influences in production, Japanese speakers may place enriched material earlier because all noun phrases precede the verb and there may be few syntactic constraints, while English speakers produce easily accessed words earlier. Such cross-linguistic differences in the prioritization of lexical accessibility versus meaning in determining word order might give rise to differences across languages in structural preferences and this would suggest that some aspects of the production system may not be universal. Thus, this evidence would appear to suggest that not all aspects of language production occur in the same way for head-initial and head-final languages, in contrast to the evidence reviewed earlier.
6.4 Discussion We have considered three different phenomena that have been investigated from the perspective of language production processes and have discussed how studies exploring these phenomena in head-initial and head-final languages have both demonstrated some similarities and some differences across languages. We now consider why this might be the case. Probably the most straightforward of the three phenomena is structural priming. As far as is known, structural priming has similar effects in head-initial and head-final constructions and languages. In all cases, exposure to a particular structure increases the propensity to subsequently use that structure. There is no evidence that the position of the head affects this tendency (though clearly further work is necessary). The other two phenomena are more interesting. In both head-initial and head-final languages, reference to more conceptually accessible entities tends to occur before reference to less conceptually accessible entities. In contrast, the language types do appear to have different preferences with respect to noun phrase length: Whereas head-initial languages prefer short noun phrases before long noun phrases, head-final languages prefer the opposite. Why might this be the case? Yamashita and Chang (2001) essentially proposed an accessibility account of the effects of noun phrase length and argued that two competing types of factor affect accessibility. Long phrases are less accessible than short phrases because their syntactic complexity makes them harder to produce (cf. Ferreira, 1996) and because there tend to be more alternative ways of expressing the same message; but they are more accessible because they make reference to more entities and therefore have stronger links with semantic representations (cf. Bock et al., 1992). In head-initial languages, the former (more syntactic) factors are more important because they occur after syntactic information associated
126
M.N. Tanaka et al.
with the head is accessed. But in head-final languages, the latter (more semantic) factor is more important, presumably because syntactic factors have not had time to have much effect. However, this explanation faces the problem that conceptual accessibility behaves the same way in both types of language. If semantic factors were prominent in head-final languages, we might expect conceptual accessibility to have a much stronger role in head-initial than head-final languages and there is no clear evidence that this is the case. An alternative possibility is that the long-before-short preference in Japanese is not the result of conceptual accessibility. In general, accessibility tends to affect early stages in processing, before the processor makes detailed decisions about the structure of noun phrases. Indeed, it is by no means clear that long phrases are conceptually more accessible than short ones – this would only follow if most entities in the phrase are accessed at roughly the same time (rather than incrementally). Instead, we propose that conceptual accessibility plays a strong role in both types of language in promoting (for example) animate entities before inanimate ones. But we suggest that the long-before-short preference follows from a different source, namely a preference to construct short dependencies where possible. As Hawkins (1994) pointed out, dependencies are shorter (on average) in a head-final construction when the heavy noun phrase follows the light noun phrase but are shorter in a head-initial construction when the light noun phrase follows the heavy noun phrase. Hawkins explained this preference in terms of ease of comprehension but subsequent evidence suggests that the motivation for this preference comes at least partly from ease of production (Wasow, 1997). Note that this proposal works against an extreme incremental view of production, in that the processor must in some sense realize the expected location of the upcoming verb in a head-initial construction.
6.5 Conclusion Although there has been considerable research comparing language comprehension in head-initial and head-final languages, there has been much less research on language production in these different language types. Current models of language production have tended to concentrate on evidence from head-initial languages. Nevertheless, such models are assumed to be based on universal processing principles (e.g., incremental processing) and we might therefore expect to find that the same factors should exert comparable effects on the production of both language types, despite their superficial differences. We have shown that recent evidence suggests that some factors influence production of both language types in consistent ways; in particular, both headinitial and head-final languages show evidence for comparable influences of conceptual accessibility and structural priming on the choice of syntactic
6 The Production of Head-Initial and Head-Final Languages
127
structure, with speakers of both language types tending to choose structures that allow accessible entities to appear earlier or as subjects, or structures that have been made more accessible through prior use. Both of these findings are consistent with a model of production in which processing is incremental, such that speakers’ decisions are driven by considerations of information accessibility. Other evidence, specifically a preference for short noun phrases to precede long noun phrases in head-initial languages but vice versa for head-final languages, appears at first sight to suggest that head-initial and head-final languages may be processed differently in some respects. However, we have suggested that these differences in the realization of surface syntactic structure may be explained in terms of a common underlying processing principle, namely a preference to establish shorter dependencies over longer dependencies. Hence very different linguistic representations can emerge from common underlying processing mechanisms. We therefore suggest that the mechanisms that underlie syntactic choice in production may be largely the same in both head-initial and head-final languages; however, those principles may manifest themselves differently in surface syntax in the two language types. On the basis of currently available evidence, it therefore seems plausible that despite the apparently substantial differences between head-initial and head-final languages, a single model of language production could account for the production of both language types. Acknowledgement We thank Claudine Raffray for her help. Martin Pickering acknowledges support of ESRC Grant RSE-062-23-0376.
References Arnold, J., Wasow, T., Losongco, A., & Grinstrom, R. (2000). Heaviness vs. newness: The effects of structural complexity and discourse status on constituent ordering. Language, 76, 28–55. Bock, J. K. (1982). Toward a cognitive psychology of syntax: Information processing contributions to sentence formulation. Psychological Review, 89, 1–47. Bock, J. K. (1986). Syntactic persistence in language production. Cognitive Psychology, 18, 355–387. Bock, J. K. (1987). An effect of the accessibility of word-forms on sentence structures. Journal of Memory and Language, 26, 119–137. Bock, J. K. (1989). Closed-class immanence in sentence production. Cognition, 31, 163–186. Bock, J. K., & Levelt, W. (1994). Language production: Grammatical encoding. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 945–984). San Diego, CA: Academic Press. Bock, J. K., & Loebell, H. (1990). Framing sentences. Cognition, 35, 1–39. Bock, J. K., Loebell, H., & Morey, R. (1992). From conceptual roles to structural relations: Bridging the syntactic cleft. Psychological Review, 99, 150–171. Bock, J. K., & Warren. R. (1985). Conceptual accessibility and syntactic structure in sentence formulation. Cognition, 21, 47–67.
128
M.N. Tanaka et al.
Branigan, H. P., & Feleki, E. (1999). Conceptual accessibility and serial order in Greek speech production. Proceedings of the 21st Cognitive Science Society Conference, Vancouver. Branigan, H. P., Pickering, M. J., & Cleland, A. A. (2000). Syntactic coordination in dialogue. Cognition, 75, B13–B25. Branigan, H. P., Pickering, M. J., McLean, J. F., & Cleland, A. A. (2007). Syntactic alignment and participant role in dialogue. Cognition, 104, 163–197. Chang, F., Bock, K., & Goldberg, A. E. (2003). Can thematic roles leave traces of their places? Cognition, 90, 29–49. Clark, H. H., & Wasow, T. (1998). Repeating words in spontaneous speech. Cognitive Psychology, 37, 201–242. Estival, D. (1985). Syntactic priming of the passive in English. Text, 5, 7–21. Ferreira, V. S. (1996). Is it better to give than to donate? Syntactic flexibility in language production. Journal of Memory and Language, 35, 724–755. Ferreira, V. S. (2003). The persistence of optional complementizer production: Why saying ‘‘that’’ is not saying ‘‘that’’ at all. Journal of Memory and Language, 48, 379–398. Garrett, M. F. (1975). The analysis of sentence production. In G. Bower (Ed.), The psychology of learning and motivation: Advances in research and theory (Vol. 9, pp. 133–177). New York: Academic Press. Griffin, Z. M., & Weinstein-Tull, J. (2003). Conceptual structure modulates structural priming in the production of complex sentences. Journal of Memory and Language, 49, 537–555. Hartsuiker, R. J., & Kolk, H. H. J. (1998). Syntactic persistence in Dutch. Language and Speech, 41, 143–184. Hartsuiker, R. J., Kolk, H. H. K., & Huiskamp, P. (1999). Priming word order in sentence production. Quarterly Journal of Experimental Psychology: Human Experimental Psychology, 52, 129–147. Hartsuiker, R. J., Pickering, M. J., & Veltkamp, E. (2004). Is syntax separate or shared between languages? Cross-linguistic syntactic priming in Spanish/English bilinguals. Psychological Science, 15, 409–414. Hartsuiker, R. J., & Westenberg, C. (2000). Word order priming in written and spoken sentence production. Cognition, 75, B27–B39. Hawkins, J. (1994). A performance theory of order and constituency. Cambridge: Cambridge University Press. Keenan, E. L., & Comrie, B. (1977). Noun Phrase Accessibility and Universal Grammar. Linguistic Inquiry, 8, 33– 99. Kelly, M. H., Bock, J. K., & Keil, F. C. (1986). Prototypicality in a linguistic context: Effects on sentence structure. Journal of Memory and Language, 25, 59–74. Kempen, G., & Harbusch, K. (2004). A Corpus study into word order variation in German subordinate clauses: Animacy affects linearization independently of grammatical function assignment. In T. Pechmann (Ed.), Language production. Berlin: Mouton Dr Gruyter. Kempen, G., & Hoenkamp, E. (1987). An incremental procedural grammar for sentence formulation. Cognitive Science, 11, 201–288. Levelt, W. J. M. (1989). Speaking: From intention to articulation. Cambridge: MIT Press. Levelt, W., & Kelter, S. (1982). Surface form and memory in question answering. Cognitive Psychology, 14, 78–106. McDonald, J., Bock, K., & Kelly, M. H. (1993). Word and world order: Semantic, phonological, and metrical determinants of serial position. Cognitive Psychology, 25, 188–230. Pickering, M. J., & Branigan, H. P. (1998). The representation of verbs: Evidence from syntactic priming in written production. Journal of Memory and Language, 39, 633–655. Pickering, M. J., Branigan, H. P., & McLean, J. F. (2002). Constituent structure is formulated in one stage. Journal of Memory and Language, 46, 586–605. Pickering, M. J., & Ferreira, V. S. (2008). Structural priming: A critical review. Psychological Bulletin, 134, 427–459.
6 The Production of Head-Initial and Head-Final Languages
129
Potter, M. C., & Lombardi, L. (1998). Syntactic priming in immediate recall of sentences. Journal of Memory and Language, 38, 265–282. Prat-Sala, M., & Branigan, H. P. (2000). Discourse constraints on syntactic processing in language production: A cross-linguistic study in English and Spanish. Journal of Memory & Language, 42, 168–182. Schenkein, J. (1980). A taxonomy for repeating action sequences in natural conversation. In B. Butterworth (Ed.), Language production (Vol. 1, pp. 21–47). London: Academic Press. Shin, J. A., & Christianson, K. (2009). Syntactic processing in Korean–English bilingual production: Evidence from cross-linguistic structural priming. Cognition, 112, 175–180. Smith, M. C., & Wheeldon, L. R. (2001). Syntactic priming in spoken sentence production: An online study. Cognition, 78, 123–164. Stallings, L. M., MacDonald, M. C., & O’Seaghdha, P. G. (1998). Phrasal ordering constraints in sentence production: Phrase length and verb disposition in heavy-NP shift. Journal of Memory and Language, 39, 392–417. Tanaka, M., Branigan, H. P., & Pickering, M. J. (submitted). Conceptual influences on word order and voice in sentence production: Evidence from Japanese. Tanaka, M., Pickering, M. J., & Branigan, H. P. (2009). Structural priming of word order and voice in Japanese and English. Unpublished manuscript. Tannen, D. (1989). Talking voices: Repetition, dialogue and imagery in conversational discourse. Cambridge: Cambridge University Press. Wasow, T. (1997). Remarks on grammatical weight. Language Variation and Change, 9, 81–105. Weiner, E. J., & Labov, W. (1983). Constraints on the agentless passive. Journal of Linguistics, 19, 29–58. Yamashita, H., & Chang, F. (2001). ‘‘Long before short’’ preference in the production of a head-final language. Cognition, 81, B45–B55.
Chapter 7
Incremental Sentence Production: Observations from Elicited Speech Errors in Japanese Noriko Iwasaki
7.1 Introduction1 Sentence production is assumed to proceed from left to right in a piecemeal fashion, and a speaker starts their sentences without planning their full details. (e.g., Yngve, 1960). Such a piecemeal manner of processing, namely, the incremental nature of sentence generation, is broadly agreed upon, and it is considered to be one of the key properties that make speaking efficient (Ferreira, 1996; Ferreira, 2002; Garrett, 1975; Griffin & Bock, 2000; Kempen & Hoenkamp, 1987; Vigliocco & Hartsuiker, 2002). Speakers of all languages speak their languages fluently and effortlessly, but the structures of languages greatly differ from each other – perhaps most notably in the direction of branching. In right-branching languages such as English, elements that are crucial in the generation of phrases and sentences (e.g., a verb in a verb phrase and noun in a noun phrase) occur early on (e.g., Ken gave a pen to Miki), but in left-branching languages such as Turkish and Japanese, they occur later – at the end of phrases or sentences (e.g., Ken-ga Miki-ni pen-o age-ta ‘‘Ken gave Miki a pen’’). Verbs that occur late in leftbranching languages presumably play the central role in sentence generation. In a series of experiments using written sentence completion tasks, Forster (1966, 1968) found that this directionality had consequences on speakers’
N. Iwasaki (*) Department of Linguistics, School of Oriental and African Studies, University of London, London WC1H 0XG, UK e-mail: [email protected] 1
I would like to express sincere gratitude to the book editors for organizing the inspiring and informative conference and for giving me helpful feedback on this chapter, to the anonymous reviewers who also provided me with constructive feedback and to Tammy Gales who carefully proofread and edited the earlier versions of this chapter.
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_7, Ó Springer ScienceþBusiness Media B.V. 2011
131
132
N. Iwasaki
performances in sentence completion tasks.2 Across the different languages that Forster studied, sentences missing the first half were more difficult to complete than sentences missing the last half (Forster, 1967, 1968); however, for speakers of left-branching languages, the difficulty was considerably reduced. These results confirm that the ending parts of sentences in left-branching languages provide more information concerning initial operations that are necessary to generate sentences, as compared to right-branching languages.
7.1.1 How Far Ahead Do Speakers Plan? A question may arise as to whether speakers, especially those of left-branching languages, indeed initiate their utterances without selecting to-be-uttered verbs. Previous research has provided empirical evidence that speakers initiate their utterances without selecting verbs. This appears to be the case even in a rightbranching language. Speakers of English were found to perform some (but not all) verb selection when initiating their utterances (Lindsley, 1975, 1976). In Lindsley’s (1975) study, English speakers produced simple sentences to describe pictures. When neither subject NP or verb was held constant, it took longer for participants to start producing subject-verb utterances such as The man is greeting than only naming the actor (e.g., The man), indicating that at least some processing of the verb occurred. But, when the subject was held constant, it did not take as long to start producing subject-verb utterances as to produce verbs alone (e.g., greeting), suggesting that English speakers initiated their utterances before completing the processing of the verb. In these English sentences, the verbs occur early in the utterance, immediately following the subjects; yet, not all verb selection is performed prior to initiating the utterances. In head-final structures, then, it is plausible that even less processing of verb selection may be performed prior to initiating utterances. In German, verb position is variable depending on linguistic contexts. Schriefers, Teruel, and Meinshausen (1998) used picture-word interference experiments with two types of lead-in fragments presented auditorily (one eliciting verb-initial utterances, und auf dem na¨chsten Bild. . . ‘‘and on the next picture’’ requiring Verb-Subject-Object order, the other verb-final utterances, auf dem na¨chsten Bild sieht man wie. . ., ‘‘on the next picture one sees how,’’ requiring Subject-Object-Verb order). They provided evidence that the verb is 2 Forster (1966, 1967, 1968) used sentences selected from novels and magazines and deleted either the first or the last half of the sentences to create the experimental materials. His subjects were asked to complete the sentences using a specified number of words. An example English sentence given by Forster (1966) was ‘‘Slowly he sat down and thought about his lost dog,’’ from which a left deletion item (__ __ __ __ __ thought about his lost dog) and a right deletion item (Slowly he sat down and __ __ __ __ __) were generated. A Japanese example given by Forster (1968) was Kookooyakyuu o kenbutu site iru to zinsei ga tanosiku naru (high school baseball ACC observe is when life NOM become enjoyable ‘When observing high school baseball games, my life becomes enjoyable’).
7 Incremental Sentence Production
133
not automatically and obligatorily part of the grammatical advanced planning unit. In a picture-word interference paradigm, the participants’ task is to name pictures while distracter words are also presented to them at about the same time as the pictures. Semantically related distracter words are known to delay the naming of the target words. In Schriefers et al.’s experiments, distracters were verbs that were semantically related to the verbs to be used in the target utterances. The semantically related distracters were expected to delay the utterance onset latencies – if verb selection is required for initiating the utterances. Semantic interference was found only for the condition in which participants produced the verbs in utterance-initial position, and it was not observed when the verb occurred in the utterance-final position, indicating that verb selection was not performed prior to the onset of the utterance. The picture-word interference paradigm has also been used to study the activation of grammatical class information of target verbs. In naming action pictures in Italian, verb distracters delayed verb naming when participants were producing inflected verb forms (Vigliocco, Vinson, & Siri, 2005), suggesting that verb distracters compete with the target verb for the verb slot when speakers are engaged in phrasal integration to produce sentences. However, in similar tasks in Japanese, verb distracters did not delay verb naming regardless of whether Japanese speakers were producing citation forms (e.g., sakebu ‘‘shout’’) or sentences (e.g., tomodati ga sakende iru ‘‘My friend is shouting’’) as compared to the noun distracters (Iwasaki, Vinson, Vigliocco, Watanabe, & Arciuri, 2008). Hence, verb distracters did not compete with the verb targets in Japanese, suggesting that the grammatical category information of the distracters was not immediately activated upon viewing the action pictures for verb naming. This lack of urgency for the activation of verb category information may be due to two properties of the Japanese language: head-final structure and the absence of subject-verb agreement.
7.1.2 Mapping Conceptual Structure to Syntactic Structure Despite evidence that speakers do not necessarily select the verb prior to initiating their utterances, the centrality of the role of the verb in formulating sentences is unquestionable. After all, it is the verb that dictates which slots (e.g., subject, direct object, prepositional object) need to be filled and by what kinds of thematic roles (e.g., agent, experiencer, theme, goal) (Pinker, 1994). It is puzzling, to say the least, how speakers of head-final languages initiate their utterances and possibly proceed without selecting the verbs. How do they map their concepts to syntactic structures? Before discussing mapping from concepts to structures, it is necessary to make the framework adopted in the current study explicit. Like many other authors such as Kempen and Hoenkamp (1987), Levelt (1989), Vigliocco and Hartsuiker (2002), I adopt a model of staged sentence production processes proposed by Garrett (1975), with two levels of processing in the formulator – though the precise nature of the two stages differs among authors. Between the
134
N. Iwasaki
conceptualizer (where a message to be articulated is formulated) and articulator (where plans for articulation are computed) is the formulator, which consists of two stages of encoding processes: the Functional and Positional levels. In the Functional stage, grammatical functions such as subject and object are assigned but unordered; in the Positional stage, the constituents are ordered and their morphological and phonological make-up is determined. The current study concerns the mapping from the conceptualizer to the Functional-level stage (see also Tanaka, Branigan, Holly, & Pickering, this volume, for a summary of sentence production models). Kempen and Hoenkamp (1987) assumed incremental processes such as in Fig. 7.1b between the conceptualizer, formulator, and articulator. In the nonincremental serial processing, as shown in Fig. 7.1a, only when the processor completes processing the semantic structure is the output of conceptualization forwarded to the formulator where syntactic structure is computed. It then needs to wait until the syntactic structure processing is complete to forward the information to the articulator. In contrast, in the incremental processes in 1b, as soon as a conceptual fragment has been computed, it is passed over to the formulator, which subsequently attempts to translate it into an utterance fragment that is then articulated. Note that because the grammar of a particular language constrains how certain elements are ordered, the order in which conceptual fragments become available is not always the order in which their utterance fragment counterparts are articulated. For example, if a speaker describes the event of him/her drinking coffee at home in English, and if conceptual fragments become available in order of the ‘‘speaker,’’ ‘‘home,’’ and ‘‘coffee,’’ the utterance fragment that corresponds to the second conceptual fragment, that of ‘‘home,’’ needs to be articulated after the utterance fragment for ‘‘coffee’’ is articulated in order to produce the grammatically well-formed C F A
Conceptualizing Formulating Articulating
a. Non-incremental processing: C semantic structure
syntactic
F
structure
phonetic
A
structure
b. Incremental processing: C cf1
Fig. 7.1 Kempen and Hoenkamp’s illustration of incremental processing (cf = conceptual fragment, uf = utterance fragment)
cf2
cf3
F A uf1
uf2 uf3
7 Incremental Sentence Production Fig. 7.2 Incremental processing in a flexible word order language
C cf1
135 cf2
cf3
F A uf1
uf2 uf3
sentence I drink coffee at home. Hence, in Fig. 7.1b, the second conceptual fragment that becomes available (cf2) is phonologically realized and articulated as the third utterance fragment (uf3) in the resulting utterance. However, in a language with a very flexible word order, something akin to the processes depicted in Fig. 7.2 may be possible; conceptual fragments can be realized in the utterances in the same or similar order in which they become available. Though the word order of Japanese is not completely free, it is much more flexible than many languages whose production processes have been studied.3 Hence, processes similar to what is illustrated in Fig. 7.2 may be at work. For example, if a speaker is to express the event of him/her drinking coffee (the example event given above) in Japanese, the utterance fragments watasi ‘‘I’’, koohii ‘‘coffee’’, and uti ‘‘home’’ that correspond to the three conceptual fragments, ‘‘home,’’ ‘‘speaker (first person),’’ and ‘‘coffee’’ can be expressed in various orders: e.g., watasi-ga uti-de koohii-o, uti-de koohii-o watasi-ga, koohii-o watasi-ga uchi-de. The verb also plays an indispensable role in the incremental procedural grammar proposed by Kempen and Hoenkamp (1987). In their model, sentences are formulated by a team of syntactic procedures working in parallel rather than by a central builder overlooking the entire process. Once a conceptual structure to be expressed, such as shown in (1), is fed to the formulator, which maps the conceptual structure to the syntactic structure, one of the first actions to be taken is lexicalization. This process – the retrieval of lexical entries called lemmas, which contain a list of procedure calls – is very important, especially that of verbs because procedure calls contained in verbs play an important role in generating sentences. (1) bake (actor: Tony) (product: cake) In most cases, even if the verb selection is not complete, because the model is equipped with procedures that build clauses and phrases, the formulation of a syntactic structure can be initiated once a conceptual structure becomes available. A categorical procedure for a sentence initiates the production processes by calling up a functional procedure for the subject, which in turn calls for a NP procedure. In the meantime, the verb lemma may be retrieved, and its 3
The word order of Japanese, especially when it is spoken, is more flexible than what is often assumed. It is not always verb-final in that some constituents can occur after the predicates (Ono, 2006; Ono & Suzuki, 1992).
136
N. Iwasaki
associated procedures become available. The eventual selection of the verb ensures accuracy. However, these processes may not account for all cases. A conceptual structure may not always be fully available when initiating an utterance. Taking the message in (1) as an example, depending on the kind of ‘‘cake,’’ the speaker may become indecisive as to which kind of cake-making event is more appropriate to express – ‘‘making’’ may be a more appropriate event to be expressed than ‘‘baking’’ if the cake is better characterized as a ‘‘steamed’’ kind. In such cases, a speaker may have a fuzzy conceptual structure such as in (2). (2) make/bake/steam (actor: Tony) (product: cake) A speaker formulating SVO sentences such as in Dutch and English needs to select the verb before articulating the Object; thus the delay of the verb retrieval may lead to disfluency. On the contrary, speakers of a head-final language such as Japanese can wait to finalize the verb selection until later. In the meantime, the NP that corresponds to the available concept, ‘‘cake’’ in this example, is likely to be expressed. But a Japanese NP needs to be assigned a case (accusative, dative, or nominative). The question is: How do Japanese speakers select case particles without the verb selection? The above example may have led the reader to speculate that even the fuzzy conceptual structure is sufficient to call upon for a Direct Object procedure, and thus, the most likely structural case to be assigned is accusative. However, a direct object (or what can be considered a direct object) in Japanese can have an accusative, dative, or nominative case particle, or a postposition to ‘‘with’’ as shown in (3).4 The lexical entry of the verb, or the information of the verb lemma, appears necessary to assign the correct case particle. (3)
a.
b.
c.
d.
4
Miki-ga pai-o Miki-NOM pie-ACC ‘Miki made a pie.’ Miki-ga pai-ni Miki-NOM pie-DAT ‘Miki touched the pie.’ Miki-ga pai-ga Miki-NOM pie-NOM ‘Miki likes a pie.’ Miki-ga Ken-to Miki-NOM Ken-with ‘Miki married Ken.’
tukutta. made sawatta. touched sukida. likes kekkon-sita. married
The following abbreviations are used. NOM: nominative, ACC: accusative, DAT: dative.
7 Incremental Sentence Production
137
In most circumstances, the verb must be selected early enough, and the verb lemma should activate the required case particles as suggested by Terao (1987). However, for incremental sentence production to proceed, some secondary procedures must be at work. In this chapter, a procedure to be called for in the absence of the yet-to-be-retrieved verb lemma will be proposed. The procedure utilizes the thematic roles of NPs in mapping concepts to structures. Data that are indicative of the existence of such secondary procedures are also presented.
7.1.3 Thematic Roles in Sentence Comprehension and Production Processes As Jackendoff (1987) noted, thematic relations or thematic roles of NPs play an important role alongside syntactic structure, and hence it is foreseeable that they play an important role in sentence processing. It should be noted, however, that thematic roles cannot directly correspond to grammatical roles or to morphological case across languages (Frawley, 1992). Both the number of thematic roles and precise characterization of the roles are controversial. According to Frawley (1992), literature lists up to 25 thematic roles. But rather than listing discrete thematic roles, Dowty (1991) proposed a theory that considers only two proto-roles, namely, cluster-concepts of protoagent and proto-patient. He suggested that such proto-roles may help explain various linguistic phenomena such as language acquisition and typological observations. The fact that the most widely studied thematic roles in psycholinguistics are also Agent (doer of action) and Patient/Theme (entity on which action is performed/entity that undergoes changes/movement) may reflect the importance of these two proto-roles in (transitive) actions. In discussing thematic roles in the current study, I adopt three broad categories clustering around their proto-roles, similarly to what Dowty proposed: Patient (including Theme), Agent, and Goal/Location. Use of thematic roles in comprehension has been extensively studied though their results are rather mixed. In the processing of reduced relative clauses, for example, some researchers (e.g., Clifton et al., 2003; Ferreira & Clifton, 1986) found that the animacy of the first noun in such a sequence as ‘‘the defendant examined. . .,’’ which increases the likelihood of the noun being the Agent (rather than Patient in a reduced relative clause such as ‘‘the defendant examined by the attorney. . .,’’) did not affect the initial syntactic ambiguity resolution. However, Trueswell, Tanenhaus, and Garnsey (1994) found that animacy led to a difficulty in processing reduced relative clauses, which they attributed to a good fit between the noun and the Agent role (the main clause interpretation). Other research has also suggested that English speakers use their knowledge of Agent and Patient along with world knowledge/contexts in the comprehension or interpretation of sentences or phrases (e.g., Altmann, 1999; Ferretti, Gagne, & McRae, 2003).
138
N. Iwasaki
The most relevant comprehension work to the current study may be Kamide, Altmann, and Haywood’s (2003a) work. In their earlier work utilizing eyetracking experiments with the ‘‘visual-world’’ paradigm, Altmann and Kamide (1999) found that in processing English, selectional information of the verb can be used to anticipate an upcoming Theme. Further, using the same paradigm, Kamide, Scheepers, and Altmann (2003b) demonstrated that case-marking information (i.e., nominative vs. accusative) and verbs’ semantic constraints are integrated rapidly to predict the most plausible forthcoming NP in processing German sentences, and Kamide et al. (2003b) showed that in processing Japanese, syntactic and semantic constraints are extracted from pre-verbal arguments and further forthcoming arguments can be anticipated in the absence of the verb. Specifically, while experimental participants were viewing a picture containing a waitress, a customer, and a hamburger, the auditory presentation of such a sequence as ‘‘waitress-NOM customer-DAT’’ in the sentence (4a) enabled the processor to anticipate the third argument, the object in the scene that can be transferred. The participants looked more at the hamburger when hearing the continuing sequence tanosigeni ‘‘merrily’’ than in cases in which they heard such sequences as ‘‘waitress-NOM customer-ACC’’ in the sentence (4b). (4)
a. Weitoresu-ga kyaku-ni tanosigeni hambaagaa-o hakobu. waitress-NOM customer-DAT merrily hamburger-ACC bring ‘The waitress merrily brings a hamburger to the customer.’ b. Weitoresu-ga kyaku-o tanosigeni karakau. waitress-NOM customer-ACC merrily tease ‘The waitress merrily teases the customer.’
The thematic role information extracted from the sequence ‘‘waitress-NOM customer-DAT’’, namely, Agent followed by Goal, contributed to the anticipation of a Theme role NP with the absence of the verb. The crucial source of the information was the dative case particle ni. Kamide et al. presented their findings as a ‘‘hallmark of an incremental processor,’’ which utilizes different sources of information to establish the fullest possible interpretation of the input. There appears to be little research on the use of thematic roles in sentence production, though more research can be found on other semantic features such as imageability (Bock & Warren, 1985), given/new information (Ferreira & Yoshita, 2003), and animacy (Bock, Loebell, & Morey, 1992; McDonald, Bock, & Kelly, 1993). Both McDonald et al. (1993) and Bock et al. (1992) found the effect of animacy in producing sentences in English. MacDonald et al. found that animate nouns tended to appear as subjects in recall tasks, and Bock et al. found that English speakers tended to produce inanimate-subject (active) sentences in picture descriptions after repeating an inanimate-subject (active or passive) sentence in a syntactic priming experiment. Thus, in both McDonald
7 Incremental Sentence Production
139
et al.’s and Bock et al.’s studies, the effects of animacy were observed in functional assignments – in the Functional-level processing. Interestingly, Tanaka, Branigan, and Pickering (submitted) found that in Japanese sentence production in recall tasks, animacy affected both functional assignments (at the Functional-level stage) and word order. There is much less research that investigates the use of thematic roles in mapping concepts to syntactic structures. Bock and Loebell (1990) examined whether speakers use thematic role information in syntactic processes in sentence production. They used a syntactic priming paradigm, in which participants first heard and repeated a priming sentence (of manipulated syntactic form) and then described a picture in each trial. In their first experiment, using priming sentences such as prepositional dative sentences (e.g., The wealthy widow gave an old Mercedes to the church) and prepositional locatives (e.g., The wealthy widow drove a Mercedes to the church), they found that both types of sentences primed prepositional dative sentences for describing pictures regardless of the thematic roles of the prepositional phrase (i.e., Beneficiary or Location). However, both of these roles can be construed as Goal (e.g., Jackendoff, 1983). Thus they conducted the second experiment using passive sentences (e.g., The construction worker was hit by the bulldozer) and locative sentences (e.g., The construction worker was digging by the bulldozer) and found that both sentences with prepositional phrases, regardless of their thematic roles (i.e., Agent or Location), primed passive sentences. Further, in order to rule out the possibility that the identical metrical patterns containing the same closed-class items, namely, to in the first experiment and by in the second experiment, caused priming, they conducted the third experiment using prepositional dative sentences (e.g., Susan brought a book to Stella) and sentences containing infinitives (e.g., Susan brought a book to study) and showed that only prepositional datives primed prepositional dative sentences. They thus concluded that thematic roles had no consequence on the priming of the hierarchical constituent patterns – processes of event-to-sentence mapping at the Functional level. Chang, Bock, and Goldberg (2003), however, noted that in Bock and Loebell’s (1990) study the order of thematic roles and syntax in the priming sentences covaried, making it difficult to assess the independent contribution of thematic roles of NPs to the ordering of these NPs. Hence, Chang et al. conducted experiments to examine the independent effect of thematic roles. They used ‘‘spray-load’’ locative alternation sentences in which the order of inanimate constituents varied within the same syntactic structure NP [V NP [P NP]PP]VP as in The man sprayed wax on the car (theme NP-location NP) vs. The man sprayed the car with wax (location NP-theme NP). Adopting Potter and Lombardi’s (1998) rapid serial visual presentation (RSVP) repetition paradigm, in which English speakers silently read stimulus sentences and repeated them after a distraction task, they found that speakers tended to use the order of thematic roles found in the priming sentence in the preceding trial to recall and
140
N. Iwasaki
repeat the target sentence. They thus supported the evidence that thematic role information is used in mapping concepts to sentence structures. Yamashita, Hirose, and Chang (2003) provided some evidence that the thematic roles of NPs may have an effect on the production of dative sentences in Japanese as well, though the contribution of thematic roles and grammatical roles is difficult to distinguish in their study. Like English, Japanese has two dative alternatives, but unlike English, which has the double-object V-NP-NP and the prepositional V-NP-PP structures, both alternatives in Japanese have a NP-NP-V structure. The difference in the two alternatives in Japanese is in the assignment of thematic roles. One alternative has the sequence ‘‘Theme-o Beneficiary-ni V’’ and the other has the sequence ‘‘Beneficiary-ni Theme-o V’’. Yamashita et al. found that Japanese speakers tended to change the target ‘‘Theme-o Beneficiary-ni V’’ sentences to ‘‘Beneficiary-ni Theme-o V’’ sentences after having produced ‘‘Beneficiary-ni Theme-o V’’ sentences – even when both NPs were inanimate (a condition for which a possible effect of animacy was absent). Hence, the ordering of the thematic roles may have affected sentence production – although the grammatical roles covaried with the thematic roles (Theme NPs were direct object, Beneficiary NPs indirect object). In F. Ferreira’s (1994) study, the prominence of thematic roles (i.e., Agent and Experiencer are more prominent than Patient or Theme) was found to be related to functional assignments – i.e., the more prominent thematic roles are assigned the functional assignment higher in the hierarchy, namely the sentential subject. In F. Ferreira’s study, participants were presented with two nouns and a verb and instructed to produce a sentence. When the verb requires Theme-Experiencer (e.g., The rock [Theme] alarmed Tom [Experiencer].) in the active voice, the participants produced more passives than with the other verbs (e.g., Agent-Theme verbs). Though Agent/Experiencer is strongly associated with animacy, in F. Ferreira’s experiment in which both NPs were animate, the effects of the thematic roles were observed. In Yamashita et al.’s (2003) and Chang et al.’s (2003) studies, the recent mapping of the thematic roles to serial positions was used in subsequent utterances, but it is not clear what it implies for the potential use of thematic roles in incremental sentence production. Perhaps, F. Ferreira’s (1994) findings that the prominent thematic role tends to be mapped to the sentential subject are more pertinent to concept-to-structure mapping in incremental sentence production. It is foreseeable that Agent NP, which is the most prominent thematic role in the event, may become available and get assigned the sentential subject in the absence of the yet-to-be-retrieved verb lemma. There are also other indications that thematic roles are essential in incremental sentence production. In an incremental sentence production model for dative construction that V. Ferreira (1996) proposed, the binding of message level representation (thematic roles of referents, such as Goal and Theme) to NPs is indispensable in implementing selectional restriction. In fact, Iwasaki’s (2006) findings concerning native speakers’ naturally-occurring case particle errors on object NPs suggest that these thematic roles are utilized in concept-to-
7 Incremental Sentence Production
141
structure mapping. Though accusative particle o is the most common particle of the direct object in Japanese, some object(-like) NPs take other particles as shown in Examples (3a-d) above. Iwasaki examined the errors of the accusative o when the verb required the dative ni and those of ni when the verb required o and found that o errors were often observed when the events being expressed were high in transitivity according to Hopper and Thompson (1980) (e.g., the action has indeed occurred; Agent volitionally acted; the referent of the direct object is affected), while ni errors often occurred when the transitivity of the events being expressed were low (and the NP could be construed as Goal with verbs such as ooen-suru ‘‘support, cheer’’, sukuu ‘‘rescue’’). In other words, erroneous o occurred when the NP had more features of the prototypical Patient role in a highly volitional event as in (5a), and erroneous ni occurred when the NP lacked prototypical features of Patient and instead had some features similar to Goal as shown in (5b). (Note that erroneous particles will be indicated in capital letters below.) (5)
a.
b.
Keikaini monogoto-O taioo suru. cope-with-NONPAST nimbly things-ACC ‘. . .nimbly cope with things’ Zyoosi-NI sukuu koto-ga dekinakatta. boss-DAT rescue nominalize-NOM can-NEG-PAST ‘I could not save my boss.’
In sum, studies on comprehension show that thematic roles play a role in predicting forthcoming elements, and in particular, Kamide et al.’s (2003a) work demonstrated that in processing Japanese sentences, thematic role information extracted from the case particles can be used to predict the thematic role of the forthcoming NP. It is not surprising, then, if the Japanese production system also actively utilizes conceptual information such as thematic roles to assign functional roles and to select case particles. Indeed, evidence has been provided by Iwasaki (2006) to support such a mechanism. The current study provides more evidence that it is indicative of such mapping.
7.1.4 Thematic Roles and Case Particles in Japanese A secondary mechanism that enables incremental processing in Japanese proposed here is the use of common correspondences between case particles and thematic roles, though such correspondence is by no means very reliable. To facilitate the discussion, the most common case particles that concern thematic relations are shown in (6): the nominative ga, the accusative o, and the dative/ locative ni.
142
(6)
N. Iwasaki
Maya-ga Ken-ni Ken-DAT Maya-NOM ‘Maya gave Ken a pen.’
pen-o pen-ACC
ageta. gave
The most problematic particle in terms of finding the commonly associated thematic role is the nominative ga, which can be associated with a variety of thematic roles, including the Agent/Experiencer and Patient/Theme roles. Moreover, it has been suggested by researchers that ga may be a default case particle (e.g., Fukui & Takano, 1995; Inoue, 1997). Japanese speakers’ speech errors also suggest that ga may be used as a default particle in sentence production (Iwasaki, 2007; see also Terao, 1995). Such a default selection of ga may conceal the possible effects of the thematic roles if there are any (e.g., the effect of the prominent thematic role, Agent, on the subject NP suggested by Ferreira, 1994). Though ni is also used for numerous thematic roles, there seem to be prototypical central roles from which other usages may be derived – Location/Goal (Kabata, 2000). Hence, the Japanese sentence production processor may map Locative/Goal to ni in the absence of the verb lemma, or in cases in which the predicate is not fully planned. The particle that has the most reliable association with a thematic role is the accusative o and the associated thematic role is Patient, although there is another commonly recognized thematic role associated with o – Passage/Passing Point (Inoue, 1998) in such sentences as kooen-o aruku ‘‘(s/he) walks through the park.’’ But this latter thematic role can also be construed as Patient, since its property is ‘‘Stationary relative to movement of another participant’’ (Dowty, 1991), which is one of the contributing properties for the Patient protorole. Thus, the mapping from Patient to the accusative o is plausible, and the current study examines speech errors among native speakers of Japanese to see if the errors reflect such a mechanism. Before discussing the current study, however, I address the question of which grammatical encoding stage (the Functional-level or the Positionallevel) may be responsible for assigning case and selecting case particles. In the original model proposed by Garrett (1975), closed-class elements (i.e., grammatical elements that do not carry substantial meanings such as inflectional morphemes and prepositions) are assigned in the structural frames constructed at the Positional level; however, Kempen and Hoenkamp (1987), who largely adopted Garrett’s model, diverged from Garrett’s in this respect and allotted the task of inserting closed-class elements to the Functional-level. One possibility is that the stages during which closed-class elements are inserted depend on the closed-class elements: functional words that are more meaningful (e.g., the English preposition from) may be inserted at the Functional-level stage and those that are structural and have no obvious semantic content (e.g., the third-person subject-verb agreement –s) may be inserted at the Positional-level. Taking such a position, I assume that case particles that
7 Incremental Sentence Production
143
are (arguably) structurally assigned, such as the nominative ga and genitive no, are assigned in the syntactic frame at the Positional level while case particles with more substantive meanings are selected and inserted at the Functional level. Added to the complication, however, is the fuzziness of the syntactic and semantic statuses of the case particles. Though linguists generally agree that the nominative ga (and the genitive no) are structural and that they do not have obvious semantic content, it is not evident whether case particles such as the accusative o and the dative ni are structural or are meaningful (see, for example, Inoue, 1998, for the status of o, and Sadakane & Koizumi, 1995, for the status of ni). In the current study, I consider both o and ni as potentially meaningful elements in the sense that they are associated with thematic roles, which are unarguably features related to intended concepts, without classifying particles o and ni into a number of homonymous particles based on their functions or associated thematic roles.
7.2 Thematic Roles and Japanese Speakers’ Speech Errors The current study examines the relationship between the thematic roles of NPs and speech errors (slips of the tongue) of case particles produced by native speakers of Japanese, focusing on the errors (overuse) of the accusative particle o. As summarized in the introduction, previous research on German indicates that the retrieval of the syntactic information of verbs is not an obligatory part of advanced planning when speakers are producing verb-final utterances in German (Schriefers et al., 1998), and the syntactic information of the verb does not appear to be immediately activated upon viewing the action to name in a sentence in Japanese (Iwasaki et al., 2008). When speaking Japanese, it is not always necessary to finalize verb selection when initiating sentences in Japanese. The question is how Japanese speakers produce sentences without the syntactic information of the verb available.
7.2.1 Hypothesis and Prediction It is hypothesized that speakers utilize the thematic roles of NPs and provisionally select case particles that commonly occur with the respective thematic roles in the absence of the verb lemma’s syntactic information and that such selection will occasionally result in errors. It is important to note here that the sequences ‘‘noun + case particle’’ are the important basic unit of sentence production in Japanese, as suggested by the analysis of repairs by Iwasaki (2000). In Iwasaki’s data, the majority of repairs were repairs of ‘‘noun + case particle’’ sequences (71.2% of repairs). When an NP possesses the Patient role in an event to be expressed, then the likelihood that the NP is to be marked with the accusative o is high. Likewise,
144
N. Iwasaki
when an NP to be expressed possesses the Goal/Location role, then the likelihood that the NP should be marked with locative ni is high. Thus, selecting o and ni on the basis of the thematic roles should help speakers to produce sentences fluently and accurately (most of the time). But such mechanisms are expected to result in speech errors when the Patient NP is realized as the subject in passive sentences or in intransitive sentences that require Patient role subjects, or when the Locative NP occurs with verbs referring to activities (rather than existence) since the particle de is required in such cases. In the current study, error occurrences of case particles on Patient NPs in subject positions are primarily examined. In addition, occurrences of erroneous ni with events with activities are examined.
7.2.2 Method The current study reanalyzes a subset of data collected by Iwasaki (2000). 7.2.2.1 Participants Sixty-three Japanese college students in Kobe, Japan, participated. 7.2.2.2 Materials and Procedures Errors were elicited through picture description tasks. Native speakers of Japanese described 120 pictures presented on a computer using the SuperLab program. The pictures were drawn to elicit various events that contain a variety of thematic roles (Agent, Patient, Location) with various types of verbs in both canonical and non-canonical orders. There were 60 line-drawings depicting the events, and each of them was first presented without any color for 700 ms followed by the same picture with coloring on one of the participants/entities in the event. The ‘‘no color followed by colored picture’’ sequence of the same drawing was presented twice in different blocks, with coloring on a different participant/entity in the event. For instance, a drawing of a cat eating a fish was presented for 700 ms followed by the same drawing with a cat colored in pink in one of the two blocks and in the other block, the same drawing appeared with the fish colored in pink (Figs. 7.3 and 7.4).
Fig. 7.3 Cat eating fish with the cat colored
7 Incremental Sentence Production
145
Fig. 7.4 Cat eating fish with the fish colored
Responses to 64 pictures depicting events of three verb types were analyzed, which consisted of five unaccusative verbs that required the Patient role NPs as subjects listed in (7), seven unergative verbs that required the Agent role NPs as subjects listed in (8)5 and transitive verbs, which allowed the passive constructions (11 with animate Agents and 9 with inanimate Agents) listed in (9). The English translations illustrate the events depicted in the pictures. (7)
(8)
(9)
5
tuku/tootyaku-suru ‘a train arriving at the station’ todoku ‘a letter getting delivered to the mail box’ sizumu/tinbotu-suru ‘a ship sinking in the ocean’ otiru/tuiraku-suru ‘a plane falling into the ocean’ ahureru ‘water flooding a bathtub’ aruku/sampo-suru ‘a woman walking in the park’ hasiru ‘a boy running in the hallway’ suwaru ‘an old woman getting seated on a bench’ naku ‘a baby crying on a chair’ odoru/dansu-suru ‘a woman dancing on a stage’ neru ‘a girl sleeping in a classroom’ oyogu ‘a man swimming in a pool’ a. Transitive verbs with animate Patient utu ‘a hunter shooting a bird’ taberu ‘a cat eating a fish’ sikaru ‘a mother scolding a child’ tataku ‘a boy hitting a girl’ keru ‘a boy kicking a dog’ tukamaeru ‘a cat catching a mouse’ okosu ‘a mother waking a child’ nerau ‘a man aiming at a bird’ daku ‘a mother holding a baby’
The data for two verbs that require Patient role NPs were excluded due to the observed difficulty that participants had in describing the pictures. As a result, the numbers of the two types of intransitive verbs were not the same.
146
N. Iwasaki
tasukeru ‘a fireman rescuing a girl’ tukamareru/taiho suru ‘a police officer arresting a thief’ b. Transitive verbs with inanimate Patient nomu ‘a baby drinking milk’ taberu ‘a man eating a hamburger’ kaku ‘a girl writing a letter’ yomu ‘a man reading a book’ miru ‘a man watching TV’ arau ‘a girl washing dishes’ hiku ‘a girl playing the piano’ sagasu ‘a man searching for his key’ kiku ‘a girl listening to music’ All intransitive events were depicted with Goal/Location to elicit two NPs in the sentences. Specifically, six of seven unergative verbs occurred with Location requiring the particle de for the location of the activity (e.g., a girl dancing on a stage), and one of the unergative verbs, suwaru ‘‘sit’’ and four of the five unaccusative verbs occurred with Goal NPs requiring the particle ni (e.g., a woman getting seated on a bench, a train arriving at the station, a plane falling into the ocean). Error occurrences of case particles with two types of the subject NPs, both requiring the nominative ga, were compared: (i) Agent role NPs (in the case of intransitive verbs with Agent role subjects), (ii) Patient NPs in passive sentences and in unaccusative sentences. In addition, error occurrences of case particles with Location NPs were examined. If the errors were not random, and if the thematic roles are used for concept-structure mapping, then o errors on Patient NPs and ni errors on Location NPs are expected – due to the common correspondence. Participants were told that the objective of the research was to examine speaking patterns in fast speech, and they were instructed to describe the pictures as fast as possible. They were also instructed to explicitly mention the participants, entities, and the place depicted in the pictures. Clauses that did not contain intended target verbs or target NPs were discarded; consequently, there was a total of 2596 pertinent clauses.
7.2.3 Results The error occurrence in this task was very infrequent; there was a total of only 48 errors (confirming that Japanese speakers speak very accurately in a situation in which they are required to speak fast). However, error patterns observed in the study corroborate the predictions.
7 Incremental Sentence Production
147
7.2.3.1 O Errors There were indeed errors of the accusative o for Patient role NPs in subject positions, both with unaccusative verbs such as todoku ‘‘gets delivered’’ and tuiraku-suru ‘‘fall’’ (3 in 386 pertinent clauses) as shown in (10–11) and passive construction (3 in 418 pertinent clauses) as shown in (12–13). No other case particle errors occurred with the subject NPs of these verbs. (10) Tegami-O... tegami-ga posuto-ni todoita. reached letter-ACC letter-NOM post-LOC ‘A letter, a letter was delivered to the mailbox’. (11) Hikooki-O... umi-no-naka-ni tuitaku-sita. airplane-ACC ocean-GEN-inside-LOC fell ‘The plane fell into the ocean.’ (12) Onnanoko-O okaasan-ni dakkos-arete iru. girl-ACC mother-by hold-PASSIVE is ‘A girl is being held by her mother.’ (13) Tori-O. . . tori-wa Taro-ni raifuru-de uta-reta. bird-ACC bird-TOP Taro-by rifle-INST shoot-PASSIVE was ‘A bird, a bird was shot with a rifle by Taro.’ In these examples, the Patient NPs such as tegami ‘‘letter’’ and onnanoko ‘‘girl’’ are erroneously marked with the accusative o when they should be marked with ga. No such errors occurred within 572 clauses using unergative verbs such as aruku ‘‘walk’’ and odoru ‘‘dance.’’ Though the application of statistical tests with such a small number of errors may be suspect, Fisher’s Exact tests showed a significant difference between unaccusative and unergative (X2=4.46) and between passives and unergatives (X2=4.12) with p < 0.05. (The more conservative tests with Yates’ correction do not yield statistical significance in either case ðX2 ¼ 2:317; p ¼ 0:128; X2 ¼ 2:085; p ¼ 0:149Þ).
7.2.3.2 NI Errors As expected, ni was also a common error for Location NPs that require de, especially for the sentence-initial NPs: 15 errors with the sentence-initial NP and four with the second NP. An example of a ni error is shown in Example (14), in which the location of activity, puuru ‘‘swimming pool’’ is marked with ni. In contrast, the occurrence of de errors in cases in which ni was the target for Goal NPs for unaccusative verbs was very rare: there was only one error of de occurring with intransitive verbs in this data set.
148
N. Iwasaki
(14) Puuru, puuru-NI otokonohito-wa oyoide ite, swim-was pool, pool-LOC man-TOP etto, sono hito-wa agatta. uh, that person-TOP went-out ‘A man was swimming in the pool, pool and got out.’ Albeit inconclusively (due to the small number of errors), these errors support the hypothesis that the processor refers to the thematic roles in the preverbal message in order to incrementally produce sentences.
7.3 Discussions and Conclusions If speakers select the verb or predicate promptly, then the syntactic information of the selected verb should guide the selection of the case particles for the NPs. This may be the primary sentence production mechanism. But, speakers do not need to wait until the verb selection to start producing sentences. The current study proposes and supports the secondary, incremental mechanism that enables Japanese speakers (or speakers of other languages that utilize case particles in a similar manner) to produce sentences in a piece-meal fashion. Though the occurrence of errors was very infrequent, errors were observed in the patterns expected by the hypothesized procedure that maps the thematic roles to NPs. The accusative o error occurred with Patient NPs that occurred in the subject positions, and ni errors occurred with Location. The data then support the case particle selection mechanism based on the mapping between the thematic role of the NP and the most strongly associated case particle, which enables Japanese speakers to produce sentences fluently (and accurately, most of the time) without finalizing the verb selection. In most cases, the verb/ predicate that is eventually selected and articulated is likely to require the case particle that is earlier selected for the NP based on the thematic roles (e.g., Patient in a transitive event that involves some kind of effect on the affected entity). In some cases, speakers may monitor their own selection of the case particle and utilize the information to finalize the verb selection in order to achieve accuracy. In the remainder of cases, the types of errors that we observed in the current study may occur. However, there is an alternative explanation that needs to be considered before concluding that o errors in the contexts examined were due to the thematic roles of the NPs. The errors on the subject NPs in passive sentences may have been due to the blending of two sentence plans: a scrambled Patient-o Agent-ga V sentence and a passive Patient-ga Agent-ni V sentence. The error could be construed as the result of either the competition of two simultaneously activated plans or a change of plan from the scrambled alternative to the passive alternative. While this explanation cannot be completely ruled out, this is not a
7 Incremental Sentence Production
149
very plausible explanation. This is because Japanese speakers predominantly used passives for animate Patient (a logistic regression with animacy as a predictor revealed the significant effect of animacy at p < 0.0001). Yet, two of the three errors of the accusative o on subject NPs occurred with an animate Patient: events of a bird shot by a hunter and a baby held by a mother. These two events were indeed predominantly described by passive sentences, 36 out of 43 analyzable descriptions for the former event and 16 out of 18 for the latter, respectively. Hence, there is some indication that mapping from thematic roles to case particle may be at work. Because thematic roles are conceptual features of intended messages, this proposed mapping mechanism is likely to be the Functional-level processing. The mechanism proposed here may be merely one of many processing mechanisms that enable incremental sentence production. And the current study has a limitation in that their production task diverges from natural speaking (especially because the first NP to be mentioned was specified by a colored picture) and that the data set was very small. More research is called for to shed further light on incremental sentence production. It would be particularly important, even necessary, to study a language with variable verb positions (such as German and Russian) in order to show that similar morphological errors of case based on thematic roles occur only when the argument precedes the verb, not when it follows the verb.6 Head-initial and head-final languages may share many processing properties that enable incremental processing as suggested by Tanaka et al. (this volume). Yet, each language (especially a head-final language) may also be equipped with additional procedures. Human language processors are likely to take advantage of all available resources maximally.
References Altmann, G. T. M. (1999). Thematic role assignment in context. Journal of Memory and Language, 41, 124–145. Altmann, G. T. M., & Kamide, Y. (1999). Incremental interpretation at the verb: Restricting the domain of subsequent reference. Cognition, 73, 247–264. Bock, K., & Loebell, H. (1990). Framing sentences. Cognition, 35, 1–39. Bock, K., Loebell, H., & Morey, R. (1992). From conceptual roles to structural relations: Bridging the syntactic cleft. Psychological Review, 99(1), 150–171. Bock, K., & Warren, R. K. (1985). Conceptual accessibility and syntactic structure in sentence formulation. Cognition, 21, 47–67. Chang, F., Bock, K., & Goldberg, A. E. (2003). Can thematic roles leave traces of their places? Cognition, 90, 29–49. Clifton, C. J., Traxler, M. J., Mohamed, M. T., Williams, R. S., Morris, R. K., & Rayner, K. (2003). The use of thematic role information in parsing: syntactic processing autonomy revisited. Journal of Memory and Language, 49, 317–334. Dowty, D. (1991). Thematic proto-roles and argument selection. Language, 67(3), 547–617. 6
I am grateful to an anonymous reviewer for raising the need for such a study.
150
N. Iwasaki
Ferreira, F. (1994). Choice of passive voice is affected by verb type and animacy. Journal of Memory and Language, 33, 715–736. Ferreira, F. (2002). How incremental is language production? Evidence from the production of utterances requiring the computation of arithmetic sums. Journal of Memory and Language, 46, 57–84. Ferreira, F., & Clifton, C. (1986). The independence of syntactic processing. Journal of Memory and Language, 25, 348–368. Ferreira, V. S. (1996). Is it better to give than to donate? Syntactic flexibility in language production. Journal of Memory and Language, 35, 724–755. Ferreira, V. S., & Yoshita, H. (2003). Given-new ordering effects on the production of scrambled sentences in Japanese. Journal of Psycholinguistic Research, 32(6), 669–692. Ferretti, T. R., Gagne, C., L., & McRae, K. (2003). Thematic role focusing by participle inflections: Evidence from conceptual combination. Journal of Experimental Psychology: Learning, Memory and Cognition, 29(1), 118–127. Forster, K. I. (1966). Left-to-right processes in the construction of sentences. Journal of Verbal Learning and Verbal Behavior, 5, 285–291. Forster, K. I. (1967). Sentence completion latencies as a function of constituent structure. Journal of Verbal Learning and Verbal Behavior, 6, 878–883. Forster, K. I. (1968). Sentence completion in left- and right-branching languages. Journal of Verbal Learning and Verbal Behavior, 7, 296–299. Frawley, W. (1992). Linguistic Semantics. Hillsdale, NJ: Lawrence Erlbaum. Fukui, N., & Takano, Y. (1995). Symmetry in syntax: Merge and demerge. Journal of East Asian Linguistics, 7, 17–86. Garrett, M. F. (1975). The analysis of sentence production. In G. H. Bower (Ed.), The psychology of learning and motivation: Advances in research and theory (Vol. 9, pp. 133–177). New York: Academic Press. Griffin, Z. M., & Bock, K. (2000). What the eyes say about speaking. Psychological Science, 11(4), 274–279. Hopper, P. J., & Thompson, S. A. (1980). Transitivity in grammar and discourse. Language, 56(2), 251–299. Inoue, K. (1997). Case marking vs. Case checking in Japanese generative grammar: An alternative proposal. Paper presented at the 40th Anniversary of Generativism FCCL (Web Journal of Formal, Computational & Cognitive Linguistics). Inoue, K. (1998). On the Japanese particle o. In M. Janse (Ed.), Productivity and creativity: Studies in general and descriptive linguistics in honor of E.M. Uhlenbeck (pp. 449–471). Berlin: Mouton de Gruyter. Iwasaki, N. (2000). Speaking Japanese: L1 and L2 grammatical encoding of case particles and adjectives/adjectival nouns. Unpublished doctoral dissertation, University of Arizona, Tucson, AZ. Iwasaki, N. (2006). Transitivity in Japanese sentence production: Speech errors of the dative NI and the accusative O. Journal of Japanese Linguistics, 22, 43–57. Iwasaki, N. (2007). Case particle errors in Japanese: Is the nominative ga a default case marker in sentence production? In C. T. Schutze & V. S. Ferreira (Eds.), The state of the art in speech error research: Proceedings of the LSA institute workshop (Vol. MIT working papers in linguistics, pp. 205–219). Cambridge, MA: MIT Press. Iwasaki, N., Vinson, D., P., Vigliocco, G., Watanabe, M., & Arciuri, J. (2008). Naming action in Japanese: Effects of semantic similarity and grammatical class. Language and Cognitive Processes, 23(6), 889–930. Jackendoff, R. (1983). Semantics and cognition. Cambridge, MA; MIT Press. Jackendoff, R. (1987). The status of thematic relations in linguistic theory. Linguistic Inquiry, 18(3), 369–411. Kabata, K. (2000). Speakers’ judgments about the lexical complexity of ni. LACUS Forum, 26, 467–478.
7 Incremental Sentence Production
151
Kamide, Y., Altmann, G. T. M., & Haywood, S. L. (2003a). The time-course of prediction in incremental sentence processing: Evidence from anticipatory eye movements. Journal of Memory and Language, 49, 133–156. Kamide, Y., Scheepers, C., & Altmann, G. T. M. (2003b). Integration of syntactic and semantic information in predictive processing: Cross-linguistic evidence from German and English. Journal of Psycholinguistic Research, 32(1), 38–55. Kempen, G. H., & Hoenkamp, E. (1987). An incremental procedural grammar for sentence formulation. Cognitive Science, 11, 201–258. Levelt, W. J. M. (1989). Speaking: From intention to articulation. Cambridge, MA: MIT Press. Lindsley, J. R. (1975). Producing simple utterances: How far ahead do we plan? Cognitive Psychology, 7, 1–19. Lindsley, J. R. (1976). Producing simple utterances: details of the planning process. Journal of Psycholinguistic Research, 5(4), 331–354. McDonald, J. L., Bock, K., & Kelly, M. H. (1993). Word and word order: Semantic, phonological, and metrical determinants of serial position. Cognitive Psychology, 25, 188–230. Ono, T. (2006). An emotively motivated post-predicate constituent order in a ‘strict predicate final’ language: Emotion and grammar meet in Japanese everyday talk. In S. Suzuki (Ed.), Emotive communication in Japanese (pp. 139–152). Amsterdam: John Benjamins. Ono, T., & Suzuki, R. (1992). Word order variability in Japanese conversation: Motivations and grammaticalization. Text, 12(3), 429–445. Pinker, S. (1994). The language instinct: How the mind creates language. New York: HarperPrennial. Potter, M. C., & Lombardi, L. (1998). Syntactic priming in immediate recall of sentences. Journal of Memory and Language, 38(3), 265–282. Sadakane, K., & Koizumi, M. (1995). On the nature of the ‘‘dative’’ particle ni in Japanese. Linguistics, 33, 5–33. Schriefers, H., Teruel, E., & Meinshausen, R. M. (1998). Producing simple sentences: results from picture-word interference experiments. Journal of Memory and Language, 39, 609–632. Tanaka, M., Branigan, H. P., & Pickering, M. J. (Submitted). Conceptual influence on word order and voice in Japanese sentence production. Journal of Memory and Language. Tanaka, M., Branigan, H. P. & Pickering, M. J. (this volume). The production of head-initial and head-final languages. Terao, Y. (1987). Nihongo ni okeru jyoshi no iiayamari ni tsuite no ichi kousatsu. Bulletin of Tokoha Gakuen Junior College, 18, 141–154. Terao, Y. (1995). Bunsanshutsu-katei ni okeru toogobumon-kenkyuu no tenboo: joshi no hatsuwa deeta o shiryou to shite. [Outlook of research on syntactic aspects of sentence production: Evidence from case particle production data] Bulletin of Tokoha Gakuen Junior College, 26, 245–255. Trueswell, J. C., Tanenhaus, M. K., & Garnsey, S. M. (1994). Semantic influence on parsing: Use of thematic role information in syntactic ambiguity resolution. Journal of Memory and Language, 33, 285–318. Vigliocco, G., & Hartsuiker, R. J. (2002). The interplay of meaning, sound, and syntax in sentence production. Psychological Bulletin, 128(3), 442–472. Vigliocco, G., Vinson, D. P., & Siri, S. (2005). Semantic similarity and grammatical class in naming actions. Cognition, 94(3), B91–B100. Yamashita, H., Hirose, Y., & Chang, F. (2003). The nature of grammatical encoding: Syntactic-independent formation of a constituent structure. Paper presented at the 20th Meeting of Japanese Cognitive Science Society, Tokyo. Yngve, V. H. (1960). A model and an hypothesis for language structure. Proceedings of the American Philosophical Society, 104(5), 444–466.
Chapter 8
The Status of Dative Constructions in Korean, English and in the Korean-English Bilingual Mind Jeong-Ah Shin and Kiel Christianson
8.1 Introduction In the research described here, we examine a controversial issue in Korean formal syntax: the status of dative structures. By way of adjudication between two conflicting accounts of these structures, we present the results of two psycholinguistic experiments that support one of the two accounts. One side of this debate proposes a structural analysis of the Korean datives as equivalent in important ways to the English structure (Urushibara, 1991). According to this view, Korean datives have two alternate structures, postpositional and double-object structures, similar to the English ‘‘dative alternation’’ between prepositional and double-object structures (Larson, 1988). The other side argues that Korean datives have only one double-object structure with two morphological variants (Oh, 2006). By examining this issue through the lens of cross-linguistic structural priming in language production, we hope to accomplish three things. First, we want to determine how, if at all, abstract syntactic knowledge from one language can affect production in the other language. Second, we wish to shed light on the relationship between syntactic rules in both languages in the minds of Korean-English bilinguals. Third and most globally, we wish to demonstrate how the results of these experiments might inform the conflicting formal accounts of Korean dative construction morphosyntax. In so doing, we hope that this study will serve as a useful example of how psycholinguistic evidence can be of interest to formal linguists and how formal theories can guide psychological research.
J.-A. Shin (*) College of Humanities, Seoul National University, 599 Gwanak-ro Gwanak-gu, Seoul, 151-742 Korea e-mail: [email protected]
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_8, Ó Springer ScienceþBusiness Media B.V. 2011
153
154
J.-A. Shin and K. Christianson
8.1.1 The question of Korean dative structures English contains a syntactic alternation termed the ‘‘dative alternation,’’ which is illustrated in (1). The alternation arises from the fact that the thematic role and case relations can be expressed with either a preposition (to) assigning oblique (dative) case to the indirect object, or with the verb (or a light verb) assigning dative case and thematic role directly to the indirect object (Beck & Johnson, 2004; Larson, 1988). (1)
a. Mary gave the book to John. b. Mary gave John the book.
In terms of a constituent that groups together basic parts of speech such as nouns (N), verbs (V), or prepositions (P) and phrasal categories headed by those parts such as noun phrases (NP), verb phrases (VP) and prepositional phrases (PP), the verb phrase of (1a) is analyzed as [V NP PP] while that of (1b) is [V NP NP]. Like English, Korean has dative alternations, which are potentially analogous to their English counterparts. However, the status of dative structures in Korean is controversial. The examples of the canonical Korean dative constructions are in (2). (2)
a. Mary-ka Mary-NOM b. Mary-ka Mary-NOM
John-eykey John-to/DAT John-ul John-ACC
chayk-ul book-ACC chayk-ul book-ACC
cwu-ess-ta1 gave-PAST-DECL cwu-ess-ta gave-PAST-DECL
The sentence in (2a) is the -eykey-ul construction in which the recipient (or goal) and the theme are licensed by particles/markers, -eykey and -ul, respectively and the sentence in (2b) is the -ul-ul construction, in which the recipient and the theme can alternately be licensed by an accusative marker -ul (Choe, 1988; O’Grady, 1991; Song, 1993). Whereas the -eykey-ul construction is used in every dative sentence, the double accusative -ul-ul construction (lul after vowels) is used in restricted contexts such as with benefactive verbs like cwu‘‘give’’ and kaluchi- ‘‘teach’’ and with verbs with special verbal morphology, cwu- that can productively form the benefactive verbal construction as in (3) (Jung & Miyagawa, 2004; Whong-Barr & Schwartz, 2002).
1
Yale Romanization is used to transliterate the Korean examples.NOM: Nominative case, ACC: Accusative case, DAT: Dative case, L: Linker, PAST: Past tense, DECL: Declarative marker
8 The Status of Dative Constructions
(3) a. Mary-ka John-eykey kulim-ul kuly-e Mary-NOM John-for/DAT picture-ACC draw-L b. Mary-ka John-ul kulim-ul kuly-e Mary-NOM John-ACC picture-ACC draw-L
155
cwu-ess-ta gave-PAST-DECL cwu-ess-ta gave-PAST-DECL
At issue is whether the -eykey particle is a postposition or a dative case marker in Korean linguistics (Hong, 1991; O’Grady, 1991). This controversial issue is directly related to a debate about the nature of dative structures in Korean. One side of this debate argues that a structural analysis of the Korean postpositional-object dative is analogous to that of the English prepositionalobject structure like [NP PP V] (Urushibara, 1991). The other side of the debate proposes that the -eykey particle is a dative case marker and the maximally projected phrase is an NP. In this case, the -eykey-ul construction is analyzed as structurally equivalent to the English double-object dative like [NP NP V] (Oh, 2006). According to the view that holds the -eykey-ul construction is a postpositional-object construction, the -eykey particle is a postposition with evidence of four criteria such as Case drop, Case stacking, Plural copying and Conjunction (for a detailed description, see Urushibara, 1991) and with evidence of a passive test (see Jung & Miyagawa, 2004, for details). Also, Whong-Barr and Schwartz (2002) argued that the -ul-ul construction is likened to the English double-object construction, quoting Song’s (1993) argument that both dative-shift in English and the Korean -ul-ul construction ‘‘promote a noncore or nonaccusative NP into the position of a core or accusative NP’’ (p. 41). On the other hand, according to the other view that holds the -eykey-ul construction as a double-object construction, Oh (2006) argued that -eykey is a dative case marker, with the similarity of the c-command asymmetry between two languages as evidence. She argued that the -eykey-marked recipient (or goal) argument asymmetrically c-commands the -ul-marked theme argument as in the English double-object construction (Larson, 1988) in the domain of anaphor binding, quantifier-pronoun binding, weak crossover, superiority, the each. . .the other construction with a reciprocal reading and negative polarity items (see Oh, 2006 for details). In addition, she argued that the Korean dative alternations are associated with only one structure irrespective of the morphology (-eykey-ul vs. -ul-ul) and the only variation in Korean ditransitives is wordorder scrambling as in (4).2 (4) Mary-ka chayk-ul John-eykey cwu-ess-ta Mary-NOM book-ACC John-to/DAT gave-PAST-DECL 2
The scrambled -ul-ul construction is theoretically possible (Cho & Choe, 2001) but it will not be covered in this paper because many native speakers of Korean consider it unacceptable.
156
J.-A. Shin and K. Christianson
In terms of this scrambled sentence, both accounts hold that the underlying structure of the scrambled construction remains the same as that of the canonical sentence in (2a) (Baek & Lee, 2004; Takano, 1998). In addition, both accounts hold that the abstract structural relations are identical in both English and Korean, although Korean is a head-final language while English is a headinitial language. In this paper, we examine the relationship between the Korean dative construction in the minds of native Korean speakers who are also bilingual English speakers. The basic idea is that if the Korean -eykey-ul construction does indeed bear structural similarities to the English prepositional dative structure (Urushibara, 1991), then we might expect to observe cross-linguistic structural priming (Desmet & Declercq, 2006; Loebell & Bock, 2003) of the structure in one language by its counterpart in the other language. Before summarizing the results from two cross-linguistic structural priming experiments (which are reported in more detail in Shin, 2008; Shin & Christianson, 2009), we discuss the use of psycholinguistic methods and results to address theoretical linguistic questions.
8.1.2 Mutually Informative Linguistic and Psycholinguistic Studies The relationship between linguistic theory and experimental psycholinguistics has, for the past 50 or so years, resembled many human relationships. At times the two disciplines have been very close, while at others the relationship has been considerably cooler. As explained by F. Ferreira (2005), one potential source of disconnect between the two fields is the view of some linguists and psycholinguists that the goals of those working in their respective disciplines are different: While linguists seek to elucidate knowledge structures underlying language competence, psycholinguists seek to understand the cognitive mechanisms that allow us to comprehend and produce language input and output in real time and real life. Obviously, this view of the two disciplines is faulty: If one wants to know how certain knowledge is deployed, one must first have some hypothesis of what that knowledge is. And if one wants to lay bare the structure of knowledge, one must also be concerned with positing knowledge structures that can interface in a psychologically plausible manner (Halle, Bresnan, & Miller, 1978). It is not always clear, however, that linguistics and psycholinguistics are mutually and reciprocally informative. In a classic example, Chomsky (2000) points out that the discovery that auditory clicks are perceptually displaced to major phrase boundaries (Abrams & Bever, 1969) had more to say about the click detection paradigm than it did about syntactic theory. The reason, basically, is that we have independently motivated and substantiated theories about phrase structure; we do not have any theories about click detection.
8 The Status of Dative Constructions
157
All interactions between formal linguistics and psycholinguistics have not been so one-sided, however. Examination of the contraction of want to to wanna in spoken English revealed that the contraction is disallowed in sentences in which a theoretically postulated null element called a trace, resulting from the displacement of an element, was hypothesized to remain in the phrase structure (Pullum, 1997; but cf. Hudson, 2006, for a comprehensive discussion). This finding not only supported the ‘‘psychological reality’’ of traces but also lent considerable support to the syntactic theory that postulated them (Chomsky, 1981, 1986). A more recent example of psycholinguistic research verifying predictions made by one formal syntactic account of a language’s grammar – and perhaps even adjudicating between conflicting accounts of that grammar – can be found in the lone study of sentence comprehension in the Algonquian language of Odawa (Ottawa) (Christianson, 2002). In that study, it was observed that processing and comprehension of sentences containing the so-called Direct verb form (Valentine, 2001) was fastest and easiest in the SVO order (Odawa permits all six possible orders of major constituents); however, for sentences containing the so-called Inverse verb form, the OSV order was processed fastest and most accurately. Importantly, sentences in OSV order, while grammatical, are exceedingly rare in actual production (Christianson, 2002; Christianson & Ferreira, 2005). Under Bruening’s (2001) account of Algonquian syntax, however, OSV is taken to be the unmarked, canonical word order for Inverse sentences (which are also far less common than Direct sentences). The psycholinguistic results are unexplained by any other account of Odawa syntax and as such, are taken as strong support for Bruening’s account. Likewise, the debate about the nature of the dative structures in Korean can be addressed by psycholinguistic data (Shin, 2008; Shin & Christianson, 2009). We exploited the existence of the dative alternation in English and Korean in the context of sentence production by Korean-English bilingual speakers using a reliable, well-attested psycholinguistic method, namely, structural priming. After an introduction to the structural priming paradigm and a summary of the method and the results, we discuss how the psycholinguistic data lend support for one of the competing views about the nature of the dative structures in Korean. We also discuss what the data suggest about the organization of syntactic structures in the minds of bilinguals.
8.1.3 Structural priming in production Dative alternation structures have been proven extremely useful in the psycholinguistic examination of the role in sentence production and comprehension by native speakers of English (Bock, 1986) and bilingual speakers of GermanEnglish (Loebell & Bock, 2003), Spanish-English (Hartsuiker, Pickering, & Veltkamp, 2004; Meijer & Fox Tree, 2003), Dutch-English (Salamoura &
158
J.-A. Shin and K. Christianson
Williams, 2006; Schoonbaert, Hartsuiker, & Pickering, 2007) and Greek-English (Salamoura & Williams, 2007), as well as by non-native English speakers (McDonough, 2006). In particular, they have been used in investigating the existence and role of abstract syntactic structure in sentence production. Most notable among these investigations is the large body of research yielding structural, or syntactic, priming effects. Structural priming refers to the phenomenon whereby speakers tend to repeat structural patterns (Bock, 1986). In the same way that people often repeat words as a usual strategy for fluent communication (Tannen, 1987), they repeat the same syntactic structure in utterances with different words, independently of meaning (Bock & Loebell, 1990; Bock, Loebell, & Morey, 1992). If, for instance, a speaker hears or speaks a double-object dative construction like John gave Mary a gift is previously spoken or heard, then s/he is more likely to have the same structure in a subsequent utterance like Sarah sent her friend a letter instead of the alternate prepositional-object dative structure like Sarah sent a letter to her friend. Structural priming has been investigated in contexts such as naturalistic corpus data (Gries, 2005), experimental settings (Bock, 1986) and dialogue (Branigan, Pickering, & Cleland, 2000) and with various syntactic structures such as transitives (e.g., Bock, 1986), datives (e.g., Bock & Loebell, 1990), phrasal-verbs (Konopka & Bock, 2005), optional complementizer that constructions (Ferreira, 2003), noun-phrase structures (e.g., Cleland & Pickering, 2003) and relative clause constructions (e.g., Corley & Scheepers, 2002). Moreover, structural priming has been tested in several priming paradigms involving picture description tasks (Bock, 1986), immediate recall tasks with rapid serial visual presentation (RSVP; Potter & Lombardi, 1998), sentence fragment completion tasks (Desmet & Declercq, 2006), sentence recall tasks (Fox Tree & Meijer, 1999) and confederate-scripting tasks (Branigan et al., 2000). Structural priming has been widely exploited to help untangle the mechanism of human language production,3 especially grammatical encoding processes (Bock & Levelt, 1994; Levelt, 1989). Grammatical encoding is a level of syntactic processing linking a message-level representation (i.e., conceptual level) to phonological encoding. Structural priming effects are interpreted as evidence for the existence of the grammatical encoding process, because structural priming occurred irrespective of meaning and sound (see Pickering & Ferreira, 2008, for a detailed review). Also, it has been argued that grammatical encoding comprises two stages: functional-level and positional-level processing. In functional-level processing, grammatical relations or thematic roles are assigned. In positional-level processing, serial order is determined. As evidence for these two stages, priming of word order was proposed and found (Hartsuiker, Kolk, & Huiskamp, 1999; Hartsuiker & Westenberg, 2000). Hartsuiker and colleagues 3
The production mechanism is assumed to have three main components: the message, the grammatical component and the phonological component.
8 The Status of Dative Constructions
159
argued that word order priming effects are interpreted as evidence for positional-level processing because the separate linearization process can cause priming when the sentences have identical grammatical relations but different surface orderings. In addition, structural priming has been interpreted as evidence for the psychological reality of phrase structure rules in the human mind (Pickering & Ferreira, 2008). Because structural priming is lexically independent, it shows the existence of abstract syntactic frames like VP ! V NP PP. Bock and Loebell (1990), for example, found that a verb phrase like in The girl is handing a paintbrush to the boy was primed by a same kind of verb phrase in Susan brought a book to Stella but not by a different kind of verb phrase in Susan brought a book to study, even though the infinitive is phonologically similar to the prepositional phrases. This shows that there is an abstract syntactic rule in the human mind independent of words and phonological components. In this way, structural priming serves as another example of how psycholinguistics and formal linguistics complement one another. Structural priming has also been found across languages in bilingual production, i.e., cross-linguistic priming (Bernolet, Hartsuiker, & Pickering, 2007; Desmet & Declercq, 2006; Hartsuiker et al., 2004; Loebell & Bock, 2003; Meijer & Fox Tree, 2003; Salamoura & Williams, 2006, 2007; Schoonbaert et al., 2007). As within one language, syntactic repetition occurs without the speaker’s intention of repeating the sentence structure and without specific lexical, thematic, or pragmatic support across languages (Loebell & Bock, 2003). It is also robust in several structural priming paradigms that have been exploited in monolingual structural priming studies such as picture description tasks (Loebell & Bock, 2003), sentence fragment completion tasks (Desmet & Declercq, 2006; Salamoura & Williams, 2006), sentence recall tasks (Meijer & Fox Tree, 2003) and confederate scripting tasks (Hartsuiker et al., 2004; Schoonbaert et al., 2007). The only difference from the original paradigms in monolingual structural priming studies is that prime sentences are provided in one language and target sentences are produced in the other language. Cross-linguistic priming results have been interpreted as shared syntax or syntactic processing between two languages in the bilingual mind. Loebell and Bock (2003), for example, observed structural priming across German and English. German and English dative constructions have the same structural configuration. The examples of double-object dative counterparts are presented in (5) and those of prepositional dative equivalents are in (6). (5) (6)
a. The b. Der a. The b. Der
boy Junge boy Junge
sent schickte sent schickte
his pen pal a letter. seinen Brieffreund einen Brief. a letter to his pen pal. einen Brief an seinen Brieffreund.
160
J.-A. Shin and K. Christianson
The results showed that sentence (6b) was primed by The girl bought a newspaper for the blind woman, not by The girl bought the blind woman a newspaper. It was argued that the syntax and processing for these structures were shared in the minds of the German English bilinguals. Like the dative example described above, when two languages contain structures that are syntactically analogous, cross-linguistic priming occurs. In other words, cross-linguistic priming effects are evidence for identical structures across two languages. Likewise, cross-linguistic priming results have showed that structural priming cannot occur across languages when the structures are not equivalent. For example, in Loebell and Bock’s (2003) experiment, German and English passives did not prime one another, because German passives are structurally different from English passives (e.g., Die Bo¨den werden ta¨glich von dem Hausmeister gereinigt [The floors are daily by the janitor cleaned]). Also, in Meijer and Fox Tree’s (2003) Experiment 3, the negation structure in English did not prime the use of the single negation structure in Spanish (e.g., Nunca quedaba gente a ese hotel cuando el otro existı´a [Never there-were people in this hotel when the other existed]) instead of double negation (e.g., No quedaba nunca gente a ese hotel cuando el otro existı´a [No there-were never people in this hotel when the other existed]). Meijer and Fox Tree argued that the two structures, single negation and double negation, are not truly interchangeable structures across these two languages. Cross-linguistic priming is thus a useful tool for investigating the bilingual mind, focusing on syntax and syntactic representation in production, constituting evidence for shared syntactic representation in the bilingual lexicon (Hartsuiker et al., 2004; Schoonbaert et al., 2007), as well as for a cognitive mechanism for usual bilingual language performance and processing (Loebell & Bock, 2003). In addition, it is a promising technique for testing whether structures in one language are analogous to ones in another language in that cross-linguistic priming has only been observed in the same structural relationship across languages (Loebell & Bock, 2003). Using this well-established methodology, the two experiments in this paper investigated whether cross-linguistic priming can occur in the dative constructions in English and Korean during bilingual production. Korean-to-English priming was used in Experiment 1 (see also Shin & Christianson, 2009) and English-to-Korean priming in Experiment 2 (see also Shin, 2008). These experiments examined the nature of syntactic processing and relationship between syntax in both languages in the Korean-English bilingual mind.
8.1.4 Predictions for Dative Structures Across Languages in Korean-English Bilinguals We have three conditions in the Korean datives: the double-accusative -ul-ul, canonical -eykey-ul and scrambled -ul-eykey constructions. The English datives have two conditions: the double-object and prepositional constructions. With
8 The Status of Dative Constructions
161
these conditions, we make several initial predictions. The first prediction concerns the relevance of the priming data reviewed here to the competing accounts of Korean dative constructions. If the -eykey-ul construction can prime the prepositional dative in English but not the double-object dative in English (or/ and vice versa), then the -eykey-ul construction is the postpositional dative, equivalent to the English prepositional dative and the Korean datives have two different structural alternations: double-object and postpositional datives. Accordingly, the -ul-ul construction can then be assumed to be structurally different from the -eykey-ul construction. In addition, with respect to predictions about the organization of Korean and English grammars in the minds of bilinguals, we make the following predictions. First, if the Korean dative structures can prime English dative production (or/and vice versa) regardless of scrambled word order, then the shared syntax in the bilingual mind is at the functional level of grammatical encoding. Second, if the scrambled word order in Korean can affect English double-object production (or/and vice versa), then positional level processing can be also involved in bilingual language processing.
8.1.5 Experiments Demonstrating Dative Structures in the KoreanEnglish Bilingual Mind There were several motivations for conducting the experiments that we describe next. First, we sought to determine which syntactic account of Korean dative structures is supported by the psycholinguistic empirical data. We were also interested in testing whether bilingual syntactic processing is shared across head-initial and head-final languages. We also wanted to address how the syntactic processing of two languages by bilinguals are cognitively integrated (for a more full report, see Shin & Christianson, 2009). We focus here on the first of these aims but also summarize results related to shared syntactic processing by Korean-English bilinguals. The experiments followed the general procedure of Meijer and Fox Tree’s (2003) sentence recall paradigm, which is believed to be sensitive to structural priming effects (Fox Tree & Meijer, 1999). The only change was that this experiment used auditory sentence stimuli instead of the written format used in the original paradigm, because Korean double-accusative datives are not frequently used in the written register. Aural presentation is, however, quite typical of other structural priming paradigms (e.g., Bock, 1986; Loebell & Bock, 2003). Experiment 1 had three Korean prime conditions (-ul-ul vs. -eykey-ul vs. -ul-eykey construction) and two English target conditions (Prepositional-object vs. Double-object construction). Experiment 2 used the same stimuli as Experiment 1 but in the reverse order: two English prime conditions and three Korean target conditions. In each trial, the target sentence was first provided (in English in Experiment 1 and in Korean in Experiment 2) through the headset and then followed by the
162
J.-A. Shin and K. Christianson
prime sentence in the other language (i.e., Korean in Experiment 1 and in English in Experiment 2). After hearing the semantically unrelated prime sentences and performing a word probe distraction task, the participants were asked to recall the previous target sentence in the language originally presented. The pivotal aspect of the procedure is that the structure of the in-between prime sentence in one language can affect the recall production in the other language. As previous research on cross-linguistic priming showed, the similar structure between two languages causes structural priming effects. The structure of the prime sentence in one language can prime the analogous structure in the other language, leading to switching to the alternate structure if the target structure is not the same as that of the prime sentence. For example, if the Korean -eykey-ul construction corresponds to the English prepositional construction, participants are likely to produce the English prepositional construction after the Korean -eykey-ul construction during recall even though the original target sentence that should have been recalled is the English double-object construction. Also, if the Korean -eykey-ul construction corresponds to the English double-object construction, participants are likely to produce the English double-object construction after the Korean -eykey-ul construction during recall even though the target sentence, which should have been recalled, is the English prepositional construction.
8.1.6 Method The participants were Korean-English bilinguals (48 in Experiment 1 and 36 in Experiment 2) recruited from the University of Illinois at Urbana-Champaign community. They were all unbalanced bilinguals who spoke Korean as their first language (L1) and English as their second language (L2). The English proficiency test, which was a cloze test conducted before the experiment, revealed that participants were fairly fluent speakers of English (Mean = 64%, SD = 0.15). Through a language history questionnaire filled out at the end of the experimental session, a detailed profile of the participants was obtained. They had had formal instruction in English for at least six years. They included 22 males and 62 females. The mean age was 28 (range 18–40). Participants received $7 for their participation. Thirty-six English dative target sentences were created on the basis of the materials of Loebell and Bock (2003). Thirty-six Korean dative target sentences were created by native speakers of Korean. These verbs allow -ul-ul and -eykey-ul constructions. The Korean prime and English target sentence in each trial were not semantically related. In addition to these 36 experimental trials, 80 filler trials were created as well as five practice trials. None of the filler or the practice sentences had either an -ul-ul or an -eykey-ul construction. The experiment involved a 2 (English type: prepositional-object vs. double-object) 3 (Korean type: -ul-ul vs. -eykey-ul vs. -ul-eykey construction) design. Both argument order
8 The Status of Dative Constructions
163
and structure factors were manipulated completely within subjects and within items, counterbalanced across experimental lists. Examples of items in each English and Korean condition are provided in (7–11). (7) The lawyer sent the contract to his client. (8) The lawyer sent his client the contract. (9) Halmeni-ka sonnye-ul chayk-ul ilk-e grandma-NOM granddaughter-ul book-ul read-L (10) Halmeni-ka sonnye-eykey chayk-ul ilk-e grandma-NOM granddaughter-eykeybook-ul read-L (11) Halmeni-ka chayk-ul sonnye-eykey ilk-e grandma-NOM book-ul granddaughter-eykeyread-L
cwu-ess-ta gave-PAST-DECL cwu-ess-ta gave-PAST-DECL cwu-ess-ta gave-PAST-DECL
Participants first took the L2 proficiency cloze test (approximately 5–10 min). They were then seated in front of a computer screen and read the instructions. They received one of the six lists. After the five practice trials were administered, the sentence recall experiment was administered in a pseudorandom order of 116 trials, with the constraint that no more than two experimental trials appeared on consecutive trials. After the participants completed the sentence recall experiment, they filled out the language background questionnaire and self-ratings of their English proficiency for 5 min. The experiment took between 40 and 45 min in total. Experimental target descriptions were scored on the basis of transcriptions, which two native speakers of English in Experiment 1 and of Korean in Experiment 2 made while listening to the recordings of participants’ production. The transcriptions were first scored as recalled and forgotten. If the event and number of arguments of target sentences were correctly recalled in a wellformed sentence including three arguments (i.e., AGENT, THEME, and RECIPIENT), then they were coded as recalled. Even when sentences did not contain up to two noncritical words, which did not influence the major syntactic constituents of the sentence (Fox Tree & Meijer, 1999), they were also considered as recalled. All recalled targets were coded into the complementary categories of Prepositional-Object (PO) and Double-Object (DO) structures in Experiment 1 and Canonical -eykey-ul construction, Scrambled -ul-eykey construction and -ul-ul condition in Experiment 2. The dependent variables were the proportions of PO productions out of the sum of PO and DO productions in Experiment 1 and the proportions of Canonical -eykey-ul constructions out of the sum of Canonical -eykey-ul constructions, Scrambled -ul-eykey constructions and -ul-ul constructions in Experiment 2. Other target recalls such as incomplete sentences were not included in these calculations. Six participants in Experiment 1 were replaced because they could not remember at least half the targets. The data were analyzed and the proportions
164
J.-A. Shin and K. Christianson
of productions in each condition for each subject (across items) and for each item (across subjects) were entered to yield analyses by participants and items, respectively, as random effects. They were subsequently arcsine-transformed and entered to yield repeated measures ANOVAs and planned pair-wise comparison analyses (Smith, 1976). All effects reported here were treated as significant when the probabilities were less than 0.05 (p < 0.05), unless otherwise indicated.
8.1.7 Results The overall accuracy of the target recalls was 77% in Experiment 1 and 90% in Experiment 2. Because the participants were unbalanced bilinguals speaking Korean as their dominant language, they recalled Korean targets in Experiment 2 better than English targets in Experiment 1. Among recalled trials, the participants generally recalled and produced the original target sentences correctly, except that they produced the canonical -eykey-ul constructions instead of the original scrambled -ul-eykey constructions in Experiment 2. When they did not correctly recall the original target sentences, they switched to other alternations during recall, more frequently producing the English prepositional-object dative in Experiment 1 and the Korean canonical -eykey-ul construction in Experiment 2. Experiment 1 showed that there was a main effect of the Korean prime on the English PO production (F1 (2, 94) = 3.91, F2(2, 70) = 3.89). The Korean -eykey-ul construction primed the English prepositional-object datives more than the Korean -ul-ul construction did. Planned comparisons showed that the participants produced the alternate prepositional-object dative in English significantly more often after the Korean -eykey-ul construction prime (33.8%, with canonical and scrambled word order pooled) than after the Korean -ul-ul dative prime (24.5%, t1(47) = 2.83; t2 (35) = 3.18), demonstrating 9.3% structural priming across languages. Even when the Korean primes had the same structure but different argument order from the English targets (i.e., the canonical -eykey-ul construction in Korean), the argumentorder independent structural effect (10.7% priming) was found. That is, the Korean canonical -eykey-ul construction primed the English prepositionalobject dative alternate production more than the Korean double-accusative -ul-ul datives did ðt1 ð47Þ ¼ 2:72; t2 ð35Þ ¼ 3:01Þ. Importantly, no unique thematic argument order difference was found between the scrambled -ul-eykey construction primes and the canonical -eykey-ul construction primes in the English target productions. In Experiment 2, the results revealed no main effect of English primes on Korean production because the participants mainly switched the original sentences to the alternate canonical -eykey-ul constructions as default production regardless of English prime sentences. Among these alternate Korean
8 The Status of Dative Constructions
165
canonical -eykey-ul construction recalls, the participants were more likely to switch to the alternate canonical -eykey-ul constructions from the original scrambled -ul-eykey construction than the original double-accusative-ul-ul constructions independently of the English prime conditions (t1 (35) = 3.71; t2 (35) = 2.82, p<0.01). Performance on an English cloze test showed that participants’ L2 proficiency (Mean = 64%) positively correlated with accurate English recall. In terms of priming effects, when the participants were divided into two groups according to the English cloze test results, the low proficiency group with below 60.88%4 accuracy showed no structural effect in priming while the high proficiency group showed the priming effect of the Korean canonical and scrambled -eykey-ul constructions on the English prepositional-object datives and a partial priming effect in the opposite direction.5
8.2 Discussion The results of the two experiments reported here are informative on several levels. First, an overall cross-linguistic priming effect of 9.3% was observed in Experiment 1 in Korean -eykey-ul constructions and English prepositionalobject datives compared to Korean -eykey-ul constructions and English double-object datives. The results suggest that the Korean -eykey-ul construction is analogous to the English prepositional-object dative but not the English double-object dative. This finding lends strong support to the syntactic account that posits two alternate dative structures in Korean (Urushibara, 1991; WhongBarr & Schwartz, 2002). Second, the experiments revealed that structural priming across languages occurred from L1 Korean to L2 English in Experiment 1, not from L2 to L1 in Experiment 2, showing asymmetric processing in the organization of unbalanced bilingual processing. The two-way directions of bilingual syntactic processing had different impacts on each other. L1 processing had a great influence on L2 processing and production when both languages shared the same syntactic structure and argument order and even when only syntactic structure was shared independently of argument order, while L2 processing was not strong enough to affect overall L1 processing. Third, structural priming was found independently of thematic argument order. The results of Experiment 1 showed that the participants were more likely to produce English prepositional dative sentences influenced by Korean eykey-ul construction primes regardless of argument order. The results revealed 4
This cut-off value was the overall mean of both experiments. The mean of the high and low proficiency group was 73.62% (SD = 8.99) and 49.14% (SD = 8.98), respectively. 5 Although this finding is of considerable interest with respect to foreign language acquisition (Cummins, 1979, 1980; McDonough, 2006; Segalowitz, 2003), further discussion is beyond the scope of this paper. See Shin (2008) and Shin and Christianson (2009) for details.
166
J.-A. Shin and K. Christianson
no unique argument order priming across languages. This suggests that shared syntactic processing is mainly due to abstract structure at the functional level of grammatical encoding in Bock and Levelt’s (1994) production model, not at the positional level. Furthermore, it appears that these shared syntactic representations are present only in Korean-dominant Korean-English bilinguals who are relatively more proficient in English. The lack of argument order effect is consistent with the syntactic view of scrambling (Saito, 1992, 2003), which argues that scrambling involves semantically vacuous movements (Yamashita, 1997). The results of this study showed that the Korean scrambled -ul-eykey constructions primed the English prepositional-object datives as much as the Korean -eykey-ul constructions did in Experiment 1 and Korean scrambled -ul-eykey constructions were susceptible to switching to canonical -eykey-ul constructions when primed by English prepositional-object datives, while the Korean double-accusative -ul-ul constructions were not in Experiment 2. Thus, it seems that the word order information fades quickly and mainly functional information continues to influence recall. Finally, the structural priming observed was also independent of the head position. The prepositional datives in English and postpositional datives in Korean are similar in terms of structural and functional relations but different in the position of heads, including the verb. This result stands as intriguing evidence for what sort of representations are shared in the bilingual mind. The order of constituents in the phrase, that is, the head position, does not appear to be a crucial element of shared bilingual syntactic representations. Rather, what appears to be shared is abstract structure at the functional level of representation, independent of the surface realization of argument order. In summary, our study showed structural priming in dative constructions across two different languages and in so doing speaks to several issues in formal linguistics and psycholinguistics. Most important for the focus of this paper, the results provide evidence for the account that holds two alternative dative structures in Korean, which are analogous to English double-object and prepositional-object dative structures (Urushibara, 1991; Whong-Barr & Schwartz, 2002). As such, it has been a case study in how psycholinguistic methods can be used to test predictions generated by competing syntactic accounts and in turn provide data that serve to adjudicate between those accounts.
References Abrams, K., & Bever, T. G. (1969). Syntactic structure modifies attention during speech perception and recognition. Quarterly Journal of Experimental Psychology, 21, 280–290. Baek, J. Y.-K., & Lee, J.-H. (2004). Double object constructions in Korean: Asymmetry between theme and goal. Language Research, 40, 669–679. Beck, S., & Johnson, K. (2004). Remarks and replies: Double objects again. Linguistic Inquiry, 35, 97–124.
8 The Status of Dative Constructions
167
Bernolet, S., Hartsuiker, R. J., & Pickering, M. J. (2007). Shared syntactic representations in bilinguals: Evidence for the role of word-order repetition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 33, 931–949. Bock, J. K. (1986). Syntactic persistence in language production. Cognitive Psychology, 18, 355–387. Bock, J. K., & Levelt, W. (1994). Language production. Grammatical encoding. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 945–984). San Diego, CA: Academic Press. Bock, J. K., & Loebell, H. (1990). Framing sentences. Cognition, 35, 1–39. Bock, J. K., Loebell, H., & Morey, R. (1992). From conceptual roles to structural relations: Bridging the syntactic cleft. Psychological Review, 99, 150–171. Branigan, H. P., Pickering, M. J., & Cleland, A. A. (2000). Syntactic co-ordination in dialogue. Cognition, 75, B13–B25. Bruening, B. (2001). The morphosyntax of the Algonquian conjunct verb. New York: Garland Publishing Inc. Cho, S.-Y., & Choe, J.-J. (2001). Scrambling in Korean: A marker-based approach. Language and Information, 5, 73–85. Choe, H. (1988). Restructuring parameters and complex predicates: A transformational approach. Unpublished doctoral dissertation, Massachusetts Institute of Technology, Cambridge. Chomsky, N. (1981). Lectures on government and binding. Dordrecht: Foris. Chomsky, N. (1986). Knowledge of language. New York, NY: Praeger. Chomsky, N. (2000). New horizons in the study of language and mind. Cambridge: Cambridge University Press. Christianson, K. (2002). Sentence processing in a ‘‘nonconfigurational’’ language. Unpublished doctoral dissertation, Michigan State University, Lansing, MI. Christianson, K., & Ferreira, F. (2005). Conceptual accessibility and sentence production in a free word order language (Odawa). Cognition, 98, 105–135. Cleland, A. A., & Pickering, M. J. (2003). The use of lexical and syntactic information in language production: Evidence from the priming of noun-phrase structure. Journal of Memory and Language, 49, 214–230. Corley, M., & Scheepers, C. (2002). Syntactic priming in English sentence production: Categorical and latency evidence from an Internet-based study. Psychonomic Bulletin & Review, 9, 126–131. Cummins, J. (1979). Linguistics interdependence and the educational development of bilingual children. Review of Educational Research, 49, 222–251. Cummins, J. (1980). The cross-lingual dimensions of language proficiency: Implications for bilingual education and the optimal age issue. TESOL Quarterly, 14, 175–187. Desmet, T., & Declercq, M. (2006). Cross-lingustic priming of syntactic hierarchical configuration information. Journal of Memory and Language, 54, 610–632. Ferreira, F. (2005). Psycholinguistics, formal grammars, and cognitive science. The Linguistic Review, 22, 365–380. Ferreira, V. S. (2003). The persistence of optional complementizer production: Why saying ‘‘that’’ is not saying ‘‘that’’ at all. Journal of Memory and Language, 48, 379–398. Fox Tree, J. E., & Meijer, P. J. A. (1999). Building syntactic structure in speaking. Journal of Psycholinguistic Research, 28, 71–92. Gries, S. (2005). Syntactic priming: A corpus-based approach. Journal of Psycholinguistic Research, 34, 365–399. Halle, M., Bresnan, J., & Miller, G. A. (1978). Linguistic Theory and Psychological Reality. Cambridge, MA: MIT Press. Hartsuiker, R. J., Kolk, H. H. J., & Huiskamp, P. (1999). Priming word order in sentence production. The Quarterly Journal of Experimental Psychology, 52A, 129–147.
168
J.-A. Shin and K. Christianson
Hartsuiker, R. J., Pickering, M. J., & Veltkamp, E. (2004). Is syntax separate or shared between languages?: Cross-linguistic syntactic priming in Spanish-English bilinguals. Psychological Science, 15, 409–414. Hartsuiker, R. J., & Westenberg, C. (2000). Word order priming in written and spoken sentence production. Cognition, 75, B27–B39. Hong, K. S. (1991). Argument selection and case marking in Korean. Unpublished doctoral dissertation, Stanford University, Stanford, CA. Hudson, R. (2006). Wanna revisited. Language, 82, 604–627. Jung, Y.-J., & Miyagawa, S. (2004). Decomposing ditransitive verbs. Paper presented at the 6th Seoul international conference on generative grammar: Minimalist views on interfaces, Seoul. Konopka, A., & Bock, J. K. (2005, April). Helping syntax out: What do words do? Paper presented at the 18th Annual CUNY Conference on Human Sentence Processing, Tucson, AZ. Larson, R. K. (1988). On the double object construction. Linguistic Inquiry, 19, 335–391. Levelt, W. J. M. (1989). Speaking: From intention to articulation. Cambridge, MA: MIT Press. Loebell, H., & Bock, J. K. (2003). Structural priming across languages. Linguistics, 41, 791–824. McDonough, K. (2006). Interaction and syntactic priming: English L2 speakers’ production of dative constructions. Studies in Second Language Acquisition, 28, 179–207. Meijer, P. J. A., & Fox Tree, J. E. (2003). Building syntactic structures in speaking: A bilingual exploration. Experimental Psychology, 50, 184–195. O’Grady, W. (1991). Categories and case: The sentence structure of Korean. Amsterdam: Benjamins. Oh, E. (2006). Second language acquisition of English double object construction by Korean speakers. Unpublished doctoral dissertation, University of Southern California, Los Angeles, CA. Pickering, M. J., & Ferreira, V. S. (2008). Structural priming: A critical review. Psychological Bulletin, 134, 427–459. Potter, M. C., & Lombardi, L. (1998). Syntactic priming in immediate recall of sentences. Journal of Memory and Language, 38, 265–282. Pullum, G. (1997). The morpholexical nature of English to-contraction. Language, 73, 79–102. Saito, M. (1992). Long distance scrambling in Japanese. Journal of East Asian Linguistics, 1, 69–118. Saito, M. (2003). A derivational approach to the interpretation of scrambling chains. Lingua, 113, 481–518. Salamoura, A., & Williams, J. N. (2006). Lexical activation of cross-language syntactic priming. Bilingualism: Language and Cognition, 9, 299–307. Salamoura, A., & Williams, J. N. (2007). Processing verb argument structure across languages: Evidence for shared representations in the bilingual lexicon. Applied Psycholinguistics, 28, 627–660. Schoonbaert, S., Hartsuiker, R. J., & Pickering, M. J. (2007). The representation of lexical and syntactic information in bilinguals: Evidence from syntactic priming. Journal of Memory and Language, 56, 153–171. Segalowitz, N. (2003). Automaticity and second languages. In C. J. Doughty & M. H. Long (Eds.), The handbook of second language acquisition (pp. 382–408). Malden, MA: Blackwell. Shin, J. A. (2008). Structural priming in bilingual language processing and second language learning. Unpublished doctoral dissertation, University of Illinois at Urbana-Champaign, Urbana, IL. Shin, J.-A., & Christianson, K. (2009). Syntactic processing in Korean-English bilingual production: Evidence from cross-linguistic priming. Cognition, 112, 175–180.
8 The Status of Dative Constructions
169
Smith, J. E. K. (1976). Data transformations in analysis of variance. Journal of Verbal Learning and Verbal Behavior, 15, 339–346. Song, N. S. (1993). Thematic relations and transitivity in English, Japanese, and Korean. Honolulu: Center for Korean Studies, University of Hawai’i. Takano, Y. (1998). Object shift and scrambling. Natural Languages & Linguistic Theory, 16, 817–889. Tannen, D. (1987). Repetition in conversation: Toward a poetics of talk. Language, 63, 574–605. Urushibara, S. (1991). Ey/Eykey: A postposition or a case marker? In S. Kuno, I.-H. Lee, J. Whitman, J. Maling, Y.-S. Kang & Y.-J. Kim (Eds.), Harvard studies in Korean linguistics (Vol. 4, pp. 421–432). Seoul: Hanshin Publishing Company. Valentine, S. (2001). Nishnaabemwin reference grammar. Toronto: University of Toronto Press. Whong-Barr, M., & Schwartz, B. D. (2002). Morphological and syntactic transfer in child L2 acquisition of the English dative alternation. Studies in second language acquisition, 24, 579–616. Yamashita, H. (1997). The effects of word-order and case marking information on the processing of Japanese. Journal of Psycholinguistic Research, 26, 163–188.
Part IV
Corpus-based Approach to Processing and Production
Chapter 9
Subject Preference, Head Animacy and Lexical Cues: A Corpus Study of Relative Clauses in Chinese Fuyun Wu, Elsi Kaiser, and Elaine Andersen
9.1 Introduction Over the past two decades, there has been an increasing amount of research on the type of information that influences sentence processing. The sources of information that have been argued to be used during real-time processing include syntactic hierarchical structure, lexical subcategorization, thematic roles, pragmatic contexts and statistical probability. Current models of language processing differ in their claims and assumptions regarding how and when different constraints interact with each other. But most of these models have been developed based on data from head-initial languages like English, with some also taking into consideration evidence from head-final languages like Japanese. We focus on the use and processing of relative clauses (RC) in Mandarin, a mixed-order language that is typologically different from both English and Japanese (Huang, 1982); it therefore raises interesting questions for models of language processing. Relative clauses (RC) have played a prominent role in cross-linguistic research, because there exist two major types of RC construction in the world’s languages: head-initial (or postnominal) and head-final (or prenominal). When the head noun (underlined in (1)) precedes the RC, the RC is head-initial, as is typical in English (ex. (1a)). When the head noun follows the RC, the RC is head-final, as is strictly the case in Japanese (ex. (1b)). Mandarin is very unusual in combining both Subject-Verb-Object word order and the N-final property (Dryer, 1992, p. 86), see ex. (1c). (1)
Typologically different RCs a. English: head-initial The dog chased the cati [that ti has stripes]
F. Wu (*) Institute of Linguistics Studies, Shanghai International Studies University, Shanghai, 200089 China e-mail: [email protected]
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_9, Ó Springer ScienceþBusiness Media B.V. 2011
173
174
F. Wu et al.
b. Japanese: head-final The dog [stripes has ti ] cati chased c. Chinese: mixed The dog chased [ti has stripes DE] cati. Following standard notation, we abbreviate RC types as SS, SO, OS and OO. The first letter in each pair indicates the role of the head noun in the main clause. The second letter indicates the role of the head noun within the RC (i.e., gap position). Thus, SS stands for subject-modifying, subject-gapped RC, SO stands for subject-modifying, object-gapped RC, OS for object-modifying, subject-gapped RC, and OO for object-modifying, object-gapped RC. Existing research in languages with head-initial RCs has shown that subjectgapped RCs are easier to process than object-gapped RCs. This has been observed in SVO English (e.g., King & Just, 1991; King & Kutas, 1995; Traxler, Morris, & Seely, 2002) and in SOV Dutch (Frazier, 1987; Mak, Vonk, & Schriefers, 2002); it is also the case for OV languages with head-final RCs such as Korean (Kwon, Polinsky, & Kluender, 2004) and Japanese (Miyamoto & Nakamura, 2003; Ueno & Garnsey, 2008). In head-final Mandarin Chinese, however, the picture is less clear. Some experimental studies (e.g., Hsiao & Gibson, 2003; Lin & Garnsey, this volume; Packard, Ye, & Zhou, this volume) have suggested that object-gapped RCs are easier to process than subject-gapped RCs, while others (e.g., Lin, 2006; Lin & Bever, 2006) have found the reverse. In addition, corpus studies (e.g., Hsiao, 2003; Kuo & Vasishth, 2006; Pu, 2007) have suggested that subject-gapped RCs occur more frequently than object-gapped RCs. Clearly more research is needed to clarify this issue. Another line of research explores whether there are additional cues to RC structure that are available before the head noun that could potentially facilitate processing. It appears that mismatching classifiers (CL) could provide one such cue for RC structure-building. For instance, in the grammatical sentence fragment nei-ben laoshi tuijian de shu (‘‘that-CL(modifying book) [teacher recommend DE] book,’’ meaning ‘‘the book [that the teacher recommended]’’, the pre-RC classifier ben is incongruent with the immediately adjacent embedded noun laoshi ‘‘teacher’’, yet it is consistent with the head noun shu ‘‘book’’. Therefore, the local classifier mismatch potentially serves as a predictor for the RC that could render it easier to process than a comparable structure with a classifier that matches the adjacent noun (e.g., nei-wei laoshi tuijian de shu ‘‘[that-CL(modifying teacher) teacher recommend DE] book’’, still meaning ‘‘the book [that the teacher recommended]’’). In the matching-classifier structure, the parser might initially assume a monoclausal structure (rather than a RC construction) and analyze ‘‘teacher’’ as the matrix subject, until the relativizer DE is encountered. Research on Japanese has shown that classifier mismatch facilitates RC processing (Yoshida, Aoshima, & Phillips, 2004) but evidence for Mandarin is mixed (see Hsu, 2006; Hsu, Phillips, &
9 Subject Preference, Head Animacy and Lexical Cues
175
Yoshida, 2005; Wu, Haskell, & Andersen, 2006). It is worth noting that the studies of Mandarin used stimuli in which the classifier (e.g., ben) was in a pre-RC position and therefore dislocated from the head noun (e.g., shu ‘‘book’’) but it is unclear how frequent this displaced structure is in actual usage. A useful way to probe these unsettled issues is to examine a large corpus of production data. Recent research has suggested that distributional information plays an important role during incremental language comprehension, especially in the context of ambiguity resolution (e.g., MacDonald, 1994; MacDonald & Christiansen, 2002; MacDonald, Pearlmutter, & Seidenberg, 1994; Trueswell 1996). These findings suggest a correlation between ease of processing and frequency of occurrence. Similarly, Hawkins’ performance-grammar correspondence hypothesis argued that ‘‘conventionalized syntactic structures are in proportion to their degree of preference in performance, as evidenced by patterns of selection in corpora and by ease of processing in psycholinguistic experiments’’ (Hawkins, 2004, p. 3). Given this connection between frequency and ease of processing, a detailed corpus study of RCs in Chinese could contribute significantly to our understanding of what factors make RCs easy or hard to process. The current study expands on prior research and aims to shed light on the issues of preferences and processing in Mandarin by examining the distribution of different types of RCs. Our corpus study focuses on three main issues: (1) the distribution and frequency of subject-gapped and object-gapped RCs in Chinese; (2) the animacy properties of the head noun in RCs; and (3) the relative frequency of RCs that contain a classifier dislocated from, versus adjacent to, its head noun. In each case, we will ask what might be responsible for any frequency differences we identify. In the next section we review two prominent theories of RC processing and prior evidence for and against the claim that Chinese RCs demonstrate a subject preference. Section 9.3 presents findings from the current corpus study, focusing on the distribution of RC types and animacy properties of head nouns. In Section 9.4, we turn to corpus findings on classifier positioning and propose two principles, Early Occurrence Strategy and Semantic Clash Avoidance, to account for the asymmetries. Section 9.5 concludes the paper and addresses the implications of our findings for the design of future sentence processing experiments.
9.2 Chinese RCs: Subject Preference? Before reviewing relevant prior empirical work, we briefly summarize two theories of sentence processing that lead to different predictions about RC processing in Mandarin.
9.2.1 Two Theoretical Accounts Theories proposed to explain the consistent subject-object processing asymmetry found in English make different predictions when applied to Mandarin Chinese. One approach is represented by Gibson (1998, 2000)’s resource-based Dependency
176
F. Wu et al.
Locality Theory (DLT). DLT suggests that parsing is constrained by the limits of working-memory capacity. Roughly speaking, DLT states that the shorter the linear distance between the gap and the head, the easier the structure should be to process. Thus subject-gapped RCs should be easier to process in English but object-gapped RCs should be easier in Mandarin, due to the shorter linear distance between the gap and the head than their respective counterparts (see ex. 2). (2)
Different predictions by the DLT: English: subject-gapped RC > object-gapped RC a. The cat [that t chased the dog] has stripes. (subject-gapped RC) b. The cat [that the dog chased t] has stripes. (object-gapped RC) Chinese: object-gapped RC > subject-gapped RC a. [t chase the dog DE ] cat has stripes. (subject-gapped RC) b. [The dog chased t DE] cat has stripes. (object-gapped RC)
An alternative perspective, represented by Keenan and Comrie’s (1977) Noun Phrase Accessibility Hierarchy, proposes an implicational universal for relativization in terms of the ranking of grammatical relations: (3) Accessibility Hierarchy: Subject (SU)4Direct Obj (DO)4Indirect Obj (IO)4Oblique (OBL)4Genitive (GEN)
By this account all languages must have at least one relativization strategy that will relativize Subject (i.e., to produce subject-gapped RCs). This relativizing strategy then continues to apply down the Accessibility Hierarchy to ‘‘lower’’ grammatical relations on the scale from Object to Indirect Object, etc. Accordingly, there should be a processing advantage across all languages that would mean a preference for subject-gapped RCs in Chinese as well as in English. It is worth noting that Keenan and Comrie’s typological universal is systematically captured by Hawkins’ (2004, pp. 175–178) Minimize Domains (MiD) Hypothesis. Simplifying somewhat, the MiD states that parsing prefers to minimize the size of the domain (i.e., the smallest set of nodes) that a dependency relation spans. Assuming that subjects generally occupy the highest argument position in a tree c-commanding other NP positions, whereas a direct object requires the co-occurrence of a c-commanding subject and a tight bond with the verb (Tomlin, 1986), Hawkins’ MiD predicts that subject-gapped RCs are universally easier than object-gapped RCs.
9.2.2 Inconsistency in Experimental Studies Existing sentence processing work on Chinese regarding the ease of processing subject-gapped and object-gapped RCs has led to inconsistent results. Hsiao
9 Subject Preference, Head Animacy and Lexical Cues
177
and Gibson’s (2003) self-paced reading study appeared to support Gibson’s DLT theory: using singly- and doubly-embedded RCs (ex. 4), they found that object-gapped RCs in Mandarin are processed faster than subject-gapped RCs, despite the temporary ambiguity they present. (4)
doubly embedded RCs: a. Subject-gapped RCs embedded in subject-gapped RCs (nested dependency): [[ e1 yaoqing[e2 goujie faguande] fuhao2 de] guanyuan1 xinhuaibugui. [gap1] invite [gap2] conspire judge DE tycoon1 DE official2 have bad intentions ‘The official who invited the tycoon who conspired with the judge has bad intentions.’ b. Object-gapped RCs embedded in object-gapped RCs (serial dependency) [[ fuhao yaoqing e2 de] faguan2 goujie e1 de] guanyuan1 xinhuaibugui tycoon invite [gap2] DE judge2 conspire [gap1]DE official1 have bad intentions ‘The official who the judge who the tycoon invited conspired with has bad intentions.’
However, it is worth noting that it is not clear how often people produce or encounter complex doubly-embedded sentences in real life. If we follow the assumption that frequency of occurrence influences processing ease, then the processing difficulties found in doubly-embedded subject-gapped RCs might be partially due to the infrequency of this construction. Kuo and Vasishth (2006) further noted that Hsiao and Gibson’s argument for the DLT hypothesis was largely built upon what they termed ‘‘the Gap Assumption’’; that is, as soon as a sentence-initial verb is encountered, the parser immediately ‘‘realize[s] that an RC is being processed, because there is no subject for the verb’’ (Hsiao & Gibson, 2003, p. 6). However, since some Chinese words can be both nouns and verbs1 and since Chinese as a pro-drop language commonly allows null pronominals, the presence of a sentence-initial verb is not an unambiguous signal for a RC. In their discussion of Hsiao and Gibson, Lin and Bever (2006) also noted that doubly-embedded subject-gapped RCs involve nested dependencies of fillers and gaps, whereas doubly-embedded object-gapped RCs involve serial dependencies. They suggest that, as serial dependencies are less complex than nested dependencies, the reported faster processing of object-gapped RCs could be due to the easiness of the dependencies involved. In sum, Hsiao and Gibson’s 1 For instance, zhuguan can be a noun (‘‘person in charge’’) or a verb (‘‘manage’’). Its lexical ambiguity cannot be resolved until the next word is encountered.
(1) zhuguan jingji de fu shizhang. manage economy DE vice mayor ‘the mayor who manages economy’
(2) zhuguan jintian mei lai. person-in-charge today not come ‘The manager didnot come today.’
178
F. Wu et al.
(2003) finding that object-gapped RCs are easier than subject-gapped RCs has resulted in considerable discussion among psycholinguists working on Chinese. Research by Lin (2006) argued against the object advantage observed by Hsiao and Gibson (2003). Lin (2006) found that subject gaps are easier to process than object gaps in Mandarin possessor RCs (PRC). The crucial manipulation involved the distance between the possessee (the gap) and the head noun in three experimental conditions (ex.5): shortest in PRCs with canonical SVO order, longer in PRCs with the SOV BA-construction (a construction in which the NP following BA is usually the patient) and longest in PRCs with the OSV BEI-construction (a passive construction in which the NP following the passive marker BEI is the agent). In contrast to what Gibson’s linear distance theory predicts, Lin found that passives were read faster and rated more natural and more grammatical than the canonical and BA sentences. Given that the gap position in passives occupies the highest subject position, whereas the gap position in the canonical sentences occupies the object position (see Lin, 2006, p. 178 for tree diagrams), Lin concluded that subject-gapped RCs are easier to process, supporting Hawkin’s MiD and Keenan and Comrie’s typological universal hierarchy (5)
Sample PRC stimuli from Lin (2006) a. [Bad guys kidnap t wife] DE chairman decide call police. [PRC w/ canonical order] b. [Bad guys BA t wife kidnap] DE chairman decide call police. [PRC w/ BA] c. [ t wife BEI bad guys kidnap] DE chairman decide call police. [PRC w/ BEI] ‘The chairman whose wife some bad guys kidnapped decided to call the police.’
Aspects of Lin’s stimuli may however complicate the conclusions that can be drawn from his results. In particular, the gap position in Lin’s stimuli is normally realized with an overt resumptive pronoun (RP)2 (e.g., ta DE laopo ‘‘his wife’’). While a covert/null possessee is possible only for constructions involving inalienable possession, half of Lin’s stimuli can only be categorized as non-inalienable ownership and therefore should contain an overt RP. This may have been a factor in why Lin’s participants rated the canonical sentences as fairly unnatural (average rating of 4.48, on a scale where 6=most unnatural) and tended to find them ungrammatical (only 18.06% of sentences were rated as ‘‘grammatical’’). Overall, Lin’s approach of manipulating linear distance between the gap and the filler to test Gibson’s DLT is ingenious, yet his actual implementation seems to have resulted in non-acceptable sentences, or his data might be consistently explained by the lack of overt RPs. If another experiment could be done, using
2
This intuition is also shared by Audrey Li (p.c.).
9 Subject Preference, Head Animacy and Lexical Cues
179
overt RPs in the gap positions and if the result of BEI > canonical > BA3 were still obtained, Lin’s subject-preference argument would be more convincing.
9.2.3 Previous Corpus Results Existing corpus studies have suggested that subject-gapped RCs are more frequent than object-gapped RCs in Chinese. For example, Hsiao (2003)’s analysis of the Chinese Treebank 3.0 corpus found more subject-gapped RCs (n=507, 57.5%) than object-gapped RCs (n=375, 42.5%). Similarly, Kuo and Vasishth’s (2006) corpus study of the Taiwan-based Sinica Corpus 3.0 (5 million words) found that subject-gapped RCs were more frequent than object-gapped RCs (119 tokens vs. 46 tokens). In fact, their search yielded 639 sentences with the form of [t2 V N1] DE N2, among which only 119 instances were real subject-gapped RCs (the remaining 520 instances were complementations4 (e.g., ‘‘increase company DE competitiveness’’, ‘‘hire employee DE cost’’)). In contrast, they found only 117 sentences of the form [N1 V t2] DE N2, among which were 46 real object-gapped RCs and 71 complementations (e.g., ‘‘tea-leave import DE quantity’’). This indicates that the most frequently occurring sentences that superficially match the subjectrelative pattern are not real subject-gapped RCs but rather complementations that do not contain gaps. Kuo and Vasishth noted that this goes against Hsiao and Gibson’s Gap Assumption, since the verb encountered first is more likely to be part of a prepositional complementation, rather than a subject-gapped RC. In another recent corpus study, Pu (2007) examined RCs in spoken and written descriptions of a short silent movie as well as modern short stories. In total, Pu identified 271 RCs, with the following frequency ranking (in descending order): SS > OS > OO > SO. Overall, the corpus contained more subject-gapped RCs (SS and OS) than object-gapped RCs (SO and OO) (74% vs. 26%), suggesting a strong preference for subject-gapped RCs. In addition, subject-gapped RCs modified the subject more frequently than the object (SS > OS), whereas object-gapped RCs modified the object more frequently than the subject (OO > SO). 3
However, it is worth noting that if the subject preference is correct, the RTs should presumably be BEI > BA > Canonical, since the gap in the BA condition occupies the VPinternal subject position (according to Lin’s tree diagram), which should be preferred over the object position in the Canonical condition. Thus, there seems to be an inconsistency between these predictions and the reading/rating data. 4 Kuo and Vasishth called these sentences adjunct relatives. However, since there is no gap in these sentences, they are not relative clauses in the strict sense. We refer to them as complementations, because, based on their English translation, they are more like a complementation structure where the second NP is the complement of the preceding NP connected by a preposition (e.g., ‘‘the competitiveness of the company to increase’’, ‘‘the cost of hiring employees’’). See similar coding criteria in Chang, Jurafsky & Manning (2009) for their five classes of the homograph DE in Chinese-English machine translation.
180
F. Wu et al.
To explain this asymmetry between subject-gapped and object-gapped RCs, Pu (2007) investigated factors such as animacy and givenness, building on existing work (e.g., Givon, 1983; Fox & Thompson, 1990). Since her claims regarding animacy are relevant for our research, we summarize them in some detail. It has been noted that human referents tend to be realized in subject position (e.g., Croft, 1990; Givon, 1983). Thus, as Pu noted, it follows that (i) subject heads (in subject-modifying RCs) tend to refer to human referents but object heads (in object-modifying RCs) tend to refer to nonhuman referents; and (ii) subject-gapped RCs tend to occur with human heads but object-gapped RCs overwhelmingly modify nonhuman heads (on animacy in RCs, see also Gennari & MacDonald, 2008; Mak et al., 2002; Traxler et al., 2002). To explain why OS RCs are less frequent than SS RCs, Pu argued that in addition to a preference for human referents to be realized in subject position, there is also a preference against realizing human referents in object position, which is ‘‘a less preferred syntactic role for coding human referents’’ (Pu, 2007, p. 46). Since a subject-gapped RC is likely to have a human head (e.g., 70% of Pu’s subject-gapped RC have human heads), it is therefore more likely to be subject-modifying (SS) than object-modifying (OS). For object-gapped RCs (OO > SO), Pu argued that because the object position is dispreferred for human referents, object-gapped RCs are likely to have non-human heads (94.4% of Pu’s object-gapped RCs have non-human heads). Given that the subject position is associated with human referents, it is not surprising that RCs with non-human heads tend to occur more in object position (OO) than in subject position (SO). Pu’s corpus study made an important contribution by highlighting the role of humanness in explaining the different RC types in Chinese, based on a corpus consisting primarily of narratives. One of the aims of our corpus study is to address the question of replicability across genres by testing whether the patterns replicate for a somewhat larger corpus in a different genre, namely the news reports in the Chinese Treebank corpus.
9.3 Current Corpus Study: RC Types and Animacy To shed light on the factors that influence the frequency and ease of processing of RCs in Mandarin Chinese, we first consider the distribution of different RC types and the effects of head animacy5 to see (i) whether the corpus patterns can inform the debate regarding ease of processing of object-gapped vs. subjectgapped RCs and (ii) whether the findings of Pu (2007) can be replicated for a slightly bigger corpus in a different genre. The corpus analysis was conducted using the first 1,000 (of 1,151) files of the third released version of the Chinese Treebank 5.0 (CTB) (Palmer, Chiou, Xue, & Xia, 2005). The corpus contains 507,222 words, equivalent to 824,983 Hanzi. It 5
For more on animacy in Chinese, see MacWhinney (1989).
9 Subject Preference, Head Animacy and Lexical Cues
181
is compiled from three different sources: Xinhua6 News from 1994–1998 (698 articles), Hong Kong news from Information Services Department of HKSAR in 1997(55 articles) and Taiwan-based Sinoranma magazine in 1996–1998 and 2000–2001 (132 articles).7 We extracted 1,218 RCs from the corpus and manually coded them for: head type; gap type; RC-internal verb type; head noun animacy; and classifier position. According to the grammatical function of the relativized head noun, these RCs were further coded into six types: subject-modifying; object-modifying; object of preposition; predicate nominal; existential; and possessive. The two most common types of RC in the corpus are subject-modifying (n=521, 42.8% of the total) and object-modifying (n=297, 24.4%). The remaining four types of RCs together constitute a third of the total corpus, with 20.2% for object-ofpreposition, 8.5% for predicate nominal, 2.6% for possessive and 1.6% for existential. Further analyses will focus only on subject-modifying and objectmodifying RCs.
9.3.1 Frequency Ranking of RCs As shown in Table 9.1, the frequency of occurrence of the different RC types is SS4OS4SO4OO: Table 9.1 and Fig. 9.1 clearly show that subject-gapped RCs are significantly more frequent than object-gapped RCs (676/818, 86.6%, and 142/818, 17.4% respectively) ðw2 ð1; N ¼ 818Þ ¼ 348:6; p50:0001Þ. This pattern holds for both subject-modifying and object-modifying RCs: Out of the 521 subject-modifying RCs, 425 (81.6%) are subject-gapped and only 96 (18.4%) are object-gapped. The chi-square difference is significant (w2 (1, N = 521) = 207:8; p50:001Þ. Out of the 297 object-modifying RCs, 251 (84.5%) are subject-gapped and only 46 (15.5%) object-gapped. The chi-square difference is also significant ðw2 ð1; N ¼ 297Þ ¼ 141:5; p50:001Þ. Following the criteria in Li and Thompson (1981), the different types of RCs were further coded by verb types within RCs (transitives, intransitives and Table 9.1 Token counts of the four major types of RCs Subject-gapped Object-gapped Total
6
Subject-modifying Object-modifying
425 251
96 46
521 297
Total
676
142
818
Xinhua is the official news agency of the People’s Republic of China. The remaining 151 files are mostly from Taiwan Sinoranma magazine, the style of which does not quite match that of the files analyzed.
7
182
F. Wu et al. 500 Number of occurrences
Fig. 9.1 Distribution of four major types of RCs
400 300
Subject-modifying Object-modifying
200 100 0 Subject-gap
Object-gap
ditransitives). As is most common in the existing psycholinguistics literature, we focus on RCs with transitive action verbs (i.e., leaving out possessive verbs like you or yongyou ‘‘have’’, prepositional verbs like guanyu ‘‘about’’ and other nonaction verbs). This results in 347 RCs with transitive action verbs, with the frequencies of the different RC types shown in Table 9.2. By examining RCs with transitive action verbs, we find that the ranking hierarchy of the four types of RCs in descending order of frequency changes from SS > OS > SO > OO (all verb types) to SS > SO > OS > OO (only transitives). Both orders differ from Pu’s (2007) result, namely, SS > OS > OO > SO. This may be due to genre differences, an issue that we return to later. The frequency ranking of SS > SO > OS > OO shows that subject-modifying RCs (SS and SO) are significantly more frequent than object-modifying RCs (OS and OO; w2 ð1; N ¼ 347Þ ¼ 35:5; p50:0001Þ. Furthermore, subjectgapped RCs are significantly more frequent than object-gapped RCs for both subject-modifying RCs and object-modifying RCs, which is consistent with Keenan and Comrie’s (1977)’s Noun Phrase Accessibility Hierarchy. As discussed in the next section, the distinct animacy patterns within the four RC types are also relevant for understanding the frequency ranking.
Table 9.2 Token counts of four types of RCs with transitive action verbs RC types with SubjectObjectTransitives gapped gapped Total w2 Analysis Subject-modifying Object-modifying
137 74
92 44
229 118
w2 ð1; N ¼ 229Þ ¼ 8:8; p50:005 w2 ð1; N ¼ 118Þ ¼ 7:627; p50:01
Total
211
136
347
w2 ð1; N ¼ 347Þ ¼ 16:2; p50:0001
9.3.2 Interim Discussion The rather consistent corpus finding that subject-gapped RCs are more frequent than object-gapped RCs seems to go against those experimental findings
9 Subject Preference, Head Animacy and Lexical Cues
183
that found object-gapped RCs easier to process – unless, of course, one assumes that there is no correlation between corpus frequency and processing ease. Indeed, this is the position that Hsiao and Gibson (2003) (see also Hsiao, 2003, p. 106) assumed in arguing for Gibson’s resource-based DLT account. However, those experimental results that found that subject-gapped RCs were easier to process (Kuo & Vasishth, 2006; Lin, 2006) are consistent with the corpus findings and the notion of a correlation between corpus frequency and processing ease. In the remainder of this section, we explore whether and how animacy interacts with structural preferences. Before turning to our corpus analysis, it is worth noting that a number of prior processing experiments used sentences with two human referents (e.g., Hsiao & Gibson, 2003; Lin, 2006) but corpus analyses (Hsiao, 2003; Kuo & Vasishth, 2006) seem to suggest that RCs with two animate NPs are not very frequent. The infrequency of RCs with two animate NPs has also been observed in languages with head-initial RCs. In a corpus study of Dutch RCs in newspaper texts, Mak et al. (2002) found 206 SS RCs as opposed to 80 SO RCs. Over half of the SS RCs (119 out of 206) had an animate head NP and an inanimate NP inside the RC (i.e., embedded NP). In contrast, SO RCs almost exclusively involved an inanimate head NP and an animate embedded NP. Surprisingly, Mak et al. found no object-gapped RCs containing both an animate head NP and an animate embedded NP. The greater frequency of RCs with two entities differing in animacy was taken to support an idea proposed by Zubin (1979) regarding the tendency to use an animate head noun as the subject of the RC (see also Pu, 2007). In sum, there appears to be a correlation between being animate, being the head of the RC and being the grammatical subject. In the next subsection, we report our corpus results regarding the animacy of the head noun in the RCs and show how the relations between grammatical position and animacy of the head noun can help explain the frequency of RC types obtained from our corpus.
9.3.3 Head NP Animacy and its Distribution Our corpus results concerning the animacy of the head noun in RCs show a correlation between head NP animacy and RC type. The head nouns of the 364 transitive RCs were coded into two animacy categories: Animate (Human and Humanized) and Inanimate (Nonhuman/Animate and Inanimate). As summarized in Table 9.3 and Fig. 9.2, Head NP animacy shows a distinct distribution pattern: 99 of 137 SS RCs (72.3%) have animate heads, whereas only 38 (or 27.7%) have inanimate heads. The chi-square difference is significant ðw2 ð1; N ¼ 137Þ ¼ 27:2; p50:001Þ. In contrast, the 74 OS RCs show no clear bias in favor of animate heads (51.4% versus 48.7%). However, of the 143 animate heads in our corpus, the vast majority (95.8%, 137/143)
184
F. Wu et al. Table 9.3 Token count of head noun animacy in different RCs SS SO OS OO Total Animate head Inanimate head Total
99 38
4 88
38 36
2 42
143 204
137
92
74
44
347
are in subject-gapped RCs (SS and OS); only six animate heads are in objectgapped RCs (SO and OO). The chi-square difference is significant ðw2 ð1; N ¼ 143Þ ¼ 120; p50:001Þ. Thus, we conclude that (i) SS RCs tend to have animate heads and (ii) animate heads generally tend to occur in subjectgapped RCs. These findings are consistent with what Pu (2007) found in her smaller corpus. On the other hand, object-gapped RCs (SO and OO) show an opposite pattern, for both us and Pu: In our SO RCs, only 4.4% (4/92) of the Head NPs are Animate, whereas 95.7% (88/92) are Inanimate (similar to Pu’s finding of 75% inanimate heads in SO RCs). The chi-square difference is significant ðw2 ð1; N ¼ 92Þ ¼ 76:7; p50:0001Þ. For OO RCs, we found only 4.6% (2/44) of Head NPs to be Animate and 95.5% (42/44) Inanimate. (Similarly, Pu found zero animate Heads for OO RCs). Again the chi-square difference is significant ðw2 ð1; N ¼ 44Þ ¼ 36:4; p50:0001Þ. Moreover, if we look at the 204 inanimate heads in our corpus, we see that 130 (63.7%) occur in object-gapped RCs (compared to 53% for Pu). Thus for object-gapped RCs, the head NP is more likely to be inanimate than animate ðw2 ð1; N ¼ 204Þ ¼ 15:4; p50:0001Þ. Our findings differ from Pu’s, however, in the distribution of nonhuman/ inanimate heads. Pu’s corpus showed a preference for non-human heads to occur in object-modifying RCs (72%), whereas we find that 126/204 (62%) of all inanimate heads occur in subject-modifying RCs. We turn to this issue below.
Number of occurrences
120
Fig. 9.2 Head-noun animacy distribution across four RCs
100 80 Animate head
60 Inanimate head
40 20 0 SS
SO
OS
OO
9 Subject Preference, Head Animacy and Lexical Cues
185
In sum, although our ranking hierarchy (SS>SO>OS>OO for transitives) differs from that of Pu (2007) (SS>OS>OO>SO), the animacy patterns in our study are very similar to her results, except for the distribution of inanimate heads. Both corpus studies found that Head NPs in SS RCs are predominantly Animate, whereas Head NPs in SO and OO RCs are predominantly Inanimate and OS RCs show no overwhelming bias for Animate or Inanimate heads. The overall finding that subject-gapped RCs are more frequent than object-gapped RCs in both our and Pu’s corpus fits with Keenan and Comrie (1977)’s NP Accessibility Hierarchy and suggests that Chinese follows this common cross-linguistic relativization preference. Furthermore, within the class of subject-gapped RCs, our findings fit Pu’s results: SS > OS. We agree with her interpretation that this stems from a preference for human/animate entities to be subjects and a dispreference for them to appear in object-position (see also Croft, 1990; Miao, 1981 in Chinese). Given that a subject-gapped RC is likely to have a human head, it is therefore more likely to be subject modifying (SS) than object-modifying (OS). However, Pu also found that object-gapped RCs modify objects more frequently than subjects (OO > SO). She attributed this to a bias for object-gapped RCs to have non-human heads and therefore to occur more in object position than subject position, since subject-position is for human referents. In contrast, we did not find a preference for inanimate heads to be restricted to object-modifying RCs: 62% of the inanimate heads in our corpus are in subject-modifying RCs. This is also why, in absolute frequency terms, our corpus contains more SO RCs than OO RCs: the overall frequency of SO RCs seems to be boosted by the number of inanimate heads in SO RCs. The reason for this difference between our and Pu’s results is not clear but may be related to genre differences. For example, Xinhua news reports tend to objectively cover events or happenings, which could bias toward inanimate heads in subject position. On the whole, despite the different genres, the general patterns that we obtained are similar to those of Pu (2007). The results of our study thus confirm the importance of taking animacy patterns into account when looking at the frequency of different RC types.
9.4 Current Corpus Study: Classifier Position We now turn to another factor that has been argued to influence ease of processing, namely classifier position. We analyzed the distribution pattern of classifiers in Mandarin RCs, in order to help clarify the findings of prior experimental research on the potential function of classifier mismatch as an early predictor of RCs. Existing experimental research on this topic has used structures that involve displacement of the classifier from its head noun. We
186
F. Wu et al.
investigate how often such dislocation occurs in natural language production and address the underlying reasons for the observed distribution. Section 9.4.1 discusses the structure of DPs in Chinese and the two positions that classifiers can occupy. Section 9.4.2 summarizes and evaluates psycholinguistic evidence regarding the possibility that classifier information provides a signal for an upcoming RC. Section 9.4.3 presents our findings regarding the distribution of numeral + classifier sequences in the Treebank corpus, as well as a proposal to explain the characteristics of the observed distributional patterns and experimental evidence supporting the proposal.
9.4.1 Classifier Positions in Chinese DP Structure Chinese DPs are standardly head-final. Thus DPs can contain several phrasal heads that occur before the head noun, including demonstratives (Dem), numerals (NumP), classifiers (ClP), adjectival phrases (AP), possessive phrases (PossP), and RCs (Li, 1998). A classifier is obligatory when a noun contains a numeral or a demonstrative.8 There exists a hierarchical relation between demonstrative, numeral and classifier, with the head noun occupying the final position (Li, 1998; Tang, 1990). Crucially, RCs can occur either before or after the fixed string of [demonstrative + numeral + classifier]. For instance, for an English sentence such as ‘‘the three books [that I bought]’’, there are two Chinese translation equivalents: a version in which the numeral+classifier sequence precedes the RC: zhe san ben [wo mai de] shu (‘‘the three CL [I buy DE] book’’) and a version in which the numeral+classifier sequence follows the RC: [wo mai de] zhe san ben shu (‘‘[I buy DE] the three CL book’’).
9.4.2 Classifier as a Cue for RC? The role of classifiers in ambiguity resolution in Mandarin RC processing provides an interesting empirical issue for psycholinguistic research because classifiers require agreement with the nouns they modify, a striking property in a language with very limited morphological agreement. Chinese has a considerable number of classifiers that interact with nouns. That is, every noun has at least one classifier that can modify it and one classifier can be applied to a set of several nouns. Given that some classifiers can only modify a restricted set of nouns based on semantic congruity, then in a sentence where the RC occurs after the classifier, a mismatching pre-RC classifier (i.e., a classifier that is not congruent with the subsequent embedded NP inside the RC and only matches 8
The wording ‘‘obligatory’’ means the syntactic phrasal head of ClP must be projected, though its phonetic content can be null given special discourse or pragmatic contexts (e.g., zhe-ø ren ‘‘this-ø person’’).
9 Subject Preference, Head Animacy and Lexical Cues
187
the head NP) could be a potential cue for early prediction of RC. To see how this works, let us build on the example in Section 9.1, zhixiang [nei-ben laoshi tuijian de shu] ‘‘Point to the book [that the teacher recommended]’’ (lit. Point-to that-CL [teacher recommend DE] book). In this sentence, the pre-RC classifier ben matches the head noun shu ‘‘book’’ but mismatches the noun that immediately follows it, namely the RC subject laoshi ‘‘teacher’’. The incongruity between the classifier ben and the immediately following noun might signal to the comprehender that an RC is upcoming. A few researchers have experimentally investigated this possibility (Hsu 2006; Hsu et al., 2005, Wu et al., 2006), with mixed results. The early work of Hsu et al. (2005) suggested that mismatching classifiers are a relatively weak cue. In an on-line self-paced reading study, they found a long-lasting slowdown at the embedded noun position in the mismatching condition in comparison with the same region in the matching condition, suggesting that mismatching classifier-noun sequences create greater processing difficulty than matching ones. Furthermore, the reading time at the head noun position was surprisingly higher in the mismatching condition than in the matching condition, suggesting that the mismatching classifier was not used for RC-structure prediction. Wu et al. (2006) questioned the plausibility of expecting participants to utilize the classifier cue when the target sentences were presented in isolation, since a RC is only necessary in a context where comprehenders must single out one of several referents. To test whether classifiers could be facilitatory in a more supportive discourse context, Wu et al. had Chinese speakers view pictures and listen to four-sentence trials while monitoring their eye movements. On each trial, the first three sentences created a context, followed by the instruction to point to a target picture. The visual display was designed so that a RC was necessary in the targets, to indicate which of the two pictures (e.g., two windows) was being referred to. In the critical instructions in the classifier mismatch condition (e.g., ‘‘Point to the window that the ball broke’’), a mismatch between the classifier (e.g., shanCL-window) and the following noun (e.g., ‘‘ball’’) provided an early cue to the RC. In the classifier match condition (e.g., ‘‘Point to the vase that the ball broke’’), the classifier (e.g., zhiCL-ball/vase) matched the following noun (e.g., ‘‘ball’’). Participants showed a small but reliable increase in early looks to the target picture following a mismatch relative to the classifier-match condition, suggesting that an appropriate discourse context facilitates the use of classifier mismatch cues. Following Wu et al. (2006), Hsu (2006) also investigated the role of discourse context by comparing RC processing in two-referent versus one-referent contexts. In a self-paced reading study, she found that mismatching classifiers could indeed help predict upcoming RCs. However, a close examination of her data suggests a possible confound that may have contributed to her results: her stimuli contained a manner adjunct phrase between a preceding classifier mismatch and a following head noun that might have provided an extra cue for an upcoming RC.
188
F. Wu et al.
It is worth noting that all of these studies used sentences in which the head noun’s classifier was dislocated from the head noun due to the intervening RC. As discussed earlier (Section 9.4.1), classifiers that modify the head of the RC can either precede the RC (pre-RC) or follow the RC (post-RC). We would like to suggest that in order for a classifier cue to be maximally useful during realtime processing, the classifier needs to be located at the correct location in the sentence to help avoid ambiguity (i.e., pre-RC) and such a structure also needs to be sufficiently frequent in naturally-occurring language. It is therefore worth asking whether pre-RC classifiers occur as often as post-RC classifiers in the corpus. Since existing research suggests that a mismatch between the classifier and the immediately following noun causes a local parsing disruption (e.g., neiliang daxuesheng baoyang de motuoche ‘‘that-CL-vehicle college-student maintain DE motocycle’’ in Hsu et al., 2005), it could be that constructions with pre-RC classifiers are dispreferred or less frequent. In order to test whether this is indeed the case, we examine the distribution of classifiers in pre- and post-RC position in the corpus.
9.4.3 Distribution of Classifier and Factors in Noun Coding Table 9.4 below shows, for each of the four RC types, how many contained a head noun classifier in pre-RC or post-RC position (only RCs with transitive action verbs are included in this analysis). As demonstrated in Table 9.4, most sentences with RCs contain no classifiers.9 If we compare the pre-RC and post-RC rows of Table 9.4, we see a mirror asymmetry for object-gapped and subject-gapped RCs: For object-gapped RCs (SO and OO), classifiers are almost exclusively post-RC (only one occurrence of a pre-RC classifier). Interestingly, this asymmetry is reversed for the subjectgapped RCs (SS and OS). For OS RCs, pre-RC classifiers occur more frequently than post-RC classifiers, accounting for 24.3% (18/74) of the total, with only two instances (2.7%, 2/74) of post-RC classifiers. For SS RCs, the
Table 9.4 Token counts of head noun classifiers in RCs with transitive verbs SS OS SO OO Total
9
Pre-RC Post-RC No classifier
22 15 100
18 2 54
0 18 74
1 12 31
41 47 259
Total
137
74
92
44
347
This may not be surprising since classifiers are necessitated by the presence of a numeral or a demonstrative. Its use is subject to special discourse contexts such as previous mention or referential events that specify numbers.
9 Subject Preference, Head Animacy and Lexical Cues
189
preference for pre-RC classifiers is less dramatic (16% of the total 137 tokens compared to 10.9% for post-RC classifiers).
9.4.4 Early Occurrence Strategy and Semantic Clash Avoidance Principle We suggest that two principles might be posited to account for the asymmetrical distribution of classifiers: (i) an Early Occurrence Strategy and (ii) Semantic Clash Avoidance. The Early Occurrence Strategy states that classifiers prefer to occur as early as they can. This strategy fits in with the growing body of research on anticipatory and expectation-based processing. Classifiers provide information about the upcoming head noun and thus would make it possible for the language comprehension system to start to build expectations about that noun. As discussed below, the Early Occurrence Strategy can explain the pre-RC preference for subject-gapped RCs. The Semantic Clash Avoidance principle states that classifier positioning should avoid disrupting lexical access to the following noun, which can help explain why object-gapped RCs prefer post-RC classifiers. To better see the two principles at work, it is helpful to take a close look at the structure of the four types of RCs (Table 9.5). For subject-gapped RCs (SS and OS), the gap (denoted with t) occurs at the left edge of the RC (marked with [RC . . .]). The early presence of the numeral + classifier phrase at the left edge provides useful information about the upcoming RC, since the classifier phrase necessarily implies an upcoming noun, possibly activating lexical access to nouns compatible with the classifier. When the next immediately available word is a verb, the parser has to delay assigning the classifier to its noun, while keeping open other possible options for ending the sentence. As more character strings unfold, the probability of positing a RC structure increases (see e.g., Levy, 200810), until the RC-marker DE becomes available, confirming that the RC structure is the only way to complete the sentence. In this way, the Early Occurrence Strategy can help with anticipatory processing.
RC SS OS OO SO
Table 9.5 Structures with pre-RC and post-RC classifiers across four types of RCs Pre-RC Post-RC ðNum þ CLÞ½RC ti V O Si V O S V ðNum þ CLÞ ½RC ti V O Oi S V ðNum þ CLÞ ½RC S V ti Oi ðNum þ CLÞ½RC S V ti Si V O
½RC ti V OðNum þ CLÞ Si V O S V½RC ti V O ðNum þ CLÞ Oi S V ½RC S V ti ðNum þ CLÞOPi ½RC S V ti ðNum þ CLÞ Si V O
10 Levy’s incremental probabilistic approach predicts that the more dependents seen in an utterance, the more information is available to the comprehender to expect the governor (e.g., the head noun’s identity and location).
190
F. Wu et al.
Results from Kuo and Vasishth (2006) provide some preliminary evidence suggesting that the Early Occurrence Principle may be on the right track. Kuo and Vasishth demonstrated that reading times at the matrix verb region were much faster when subject-gapped RCs (SS) contained a sentence-initial determiner+classifier sequence than when they did not. This suggests a late processing advantage facilitated by the presence of the pre-RC classifier. Having seen how the Early Occurrence Strategy works in subject-gapped RCs, let us now turn to object-gapped RCs (OO and SO). The Early Occurrence Strategy predicts that classifiers in object-gapped RCs will also prefer to occur as early as possible, i.e., pre-RC. Clearly, though, this does not fit the corpus patterns (Table 9.4), as object-gapped RCs mostly have post-RC classifiers. We suggest that the infrequency of pre-RC classifiers in object-gapped RCs stems from a crucial difference between subject-gapped and object-gapped RCs: In contrast to subject-gapped RCs, in object-gapped RCs, a pre-RC classifier is adjacent to the ‘‘wrong’’ noun (the subject of the RC, rather than its actual target, the head noun). Given that the presence of a classifier implies that a noun of a certain semantic class is coming up, then in a situation where the immediately following noun (subject of RC) does not have the expected semantic class, the resulting mismatch may interfere with lexical access. In other words, if the classifier leads to increased activation of nouns compatible with the classifier, forcing the parser to deal with a semantic clash when encountering an incompatible noun may result in increased processing load. What about a situation where an object-gapped RC has a pre-RC classifier that matches the noun inside the RC (e.g., nei-wei daxuesheng baoyang de motuoche ‘‘that-CL-human college-student maintain DE motorcycle’’)? No semantic clash occurs in such a structure and thus lexical access should not be disrupted. Note, however, that such a sentence will also fail to provide an early cue to an upcoming RC structure. In other words, no classifier mismatch is present to indicate to the parser that the noun adjacent to the classifier is not the object of the main clause. Eye-tracking data from Wu et al. (2006) (discussed in Section 9.4.2) can be seen as providing tentative evidence for the Semantic Clash Avoidance principle. Wu et al. looked at object-gapped RCs and found that when a pre-RC classifier mismatches the subsequent noun (embedded noun, RC-subject), the proportion of fixations to the picture of the embedded noun is lower than in conditions where the pre-RC classifier matches the embedded noun. This is what the Semantic Clash would predict: classifiers that mismatch the embedded noun cause problems with lexical access as compared to the classifier-match condition, where no such lexical disruption occurs. Interestingly, however, the absence of a semantic clash also means that a possible early cue to the presence of an RC is absent. This raises the question of whether classifier placement is different in RCs where the classifier matches the embedded noun as well as the head noun vs. RCs where the classifier only matches the head noun. We leave this question for future research.
9 Subject Preference, Head Animacy and Lexical Cues
191
In sum, although the Early Occurrence Strategy and Semantic Clash Avoidance principle are still speculative and need to be rigorously tested in future work, they appear to provide a promising way of capturing the corpus patterns and receive some support from comprehension and production research.
9.5 Conclusions This corpus study aims to shed light on three unresolved issues in Mandarin Chinese processing: (i) whether subject-gapped RCs are easier to process than object-gapped RCs, (ii) whether head NP animacy information contributes to RC processing and (iii) whether pre-RC classifier mismatch is a reliable predictor of an upcoming RC. In light of the debate concerning the ease of processing subject-gapped vs. object-gapped RCs in Chinese, we analyzed the frequency of different RC types in a large corpus of naturally-occurring text. The results show that SS RCs consistently occur more frequently than other types (OS, SO, OO). If frequency and ease of processing are related (e.g., MacDonald & Christiansen, 2002), this finding suggests that the subject-gap preference found in other languages also exists in Chinese. Our frequency count of RCs with transitive action verbs demonstrates a frequency ranking of SS > SO > OS > OO. We explain this pattern by building on Keenan and Comrie (1977)’s Accessibility Hierarchy, as well as the general tendency for animate entities to be realized in subject position and inanimate entities to be realized as objects. In looking at classifier position, we asked whether a mismatch between a preRC classifier and the following embedded noun can provide a cue to an upcoming RC and how frequently such cues occur in naturally-occurring text. The corpus analysis suggests that pre-RC classifiers are more frequent in subject-gapped than in object-gapped RCs, an asymmetry that we explain by means of the Early Occurrence Strategy and the Semantic Clash Avoidance principle. This corpus study lays the foundation for future research on Chinese RC processing. The results presented here further our understanding not only of the overall frequency of different RC types in Chinese but also of the crucial effect of animacy as well as the asymmetries in classifier position in different RC types. Thus they are relevant for research investigating the nature of the relationship between processing ease and frequency and also highlight factors (e.g., animacy) that should be taken into account when designing future experiments.
References Chang, P.-C., Jurafsky, D., & Manning, C. (2009). Disambiguating ‘‘DE’’ for Chinese-English machine translation. EACL 2009 Fourth Workshop on Statistical Machine Translation. Croft, W. (1990). Typology and universals. Cambridge textbooks in linguistics. Cambridge, England: Cambridge University Press.
192
F. Wu et al.
Dryer, M. S. (1992). The Greenbergian word order correlations. Language, 68, 81–138. Frazier, L. (1987). Syntactic processing: Evidence from Dutch. Natural Language and Linguistics Theory, 5, 519–559. Fox, B. A., & Thompson, S. A. (1990). A discourse explanation of the grammar of relative clauses in English conversation. Language, 66, 297–316. Gennari, S. & MacDonald M.C. (2008). Semantic indeterminacy in object relative clauses. Journal of Memory and Language, 58, 161–187. Gibson, E. (1998). Linguistic complexity: The locality of syntactic dependencies. Cognition, 68, 1–76. Gibson, E. (2000). The dependency locality theory: A distance-based theory of linguistic complexity. In A. Marantz, Y. Miyashita & W. O’Neil (Eds.), Image, language, brain (pp. 95–126). Cambridge, MA: MIT Press. Givon, T. (1983). Topic continuity and word order pragmatics in Ute. In T. Givon (Ed.), Topic continuity in discourse: Quantitative cross-language studies (pp. 343–363). Amsterdam: Benjamins. Gordon, P. C., Hendrick, R., & Levine, W. H. (2002). Memory load interference in syntactic processing. Psychological Science, 13, 425–430. Hawkins, J. (2004). Efficiency and complexity in grammar. Oxford: Oxford University Press. Hsiao, F. (2003). The syntax and processing of relative clauses in Mandarin Chinese. Doctoral dissertation, MIT. Hsiao, F., & Gibson, E. (2003). Processing relative clauses in Chinese. Cognition, 90, 3–77. Hsu, C.-C. N. (2006). Issues in head-final relative clauses in Chinese – Derivation, processing and acquisition. Doctoral dissertation, University of Delaware. Hsu, C.-C. N., Phillips, C., & Yoshida, S. (2005). Cues for head-final relative clauses in Chinese. Poster presented at the 18th Annual CUNY Sentence Processing Conference. Tucson, AZ. Huang, C.-T. J. (1982). Logical relations in Chinese and the theory of grammar. Doctoral dissertation, MIT. Keenan, E. L., & Comrie, B. (1977). Noun phrase accessibility and universal grammar. Linguistic Inquiry, 8, 63–99. King, J., & Just, A. (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language, 30, 580–602. King, J. & Kutas, M. (1995). Who did what and when? Using word- and clause-level ERPs to monitor working memory usage in reading. Journal of Cognitive Neuroscience, 7(3), 376–395. Kuo, K., & Vasishth, S. (2006). Processing relative clauses: Evidence from Chinese. Unpublished manuscript, University of Potsdam. Kwon, N., Polinsky, M., & Kluender, R. (2004). Processing of relative clause sentences in Korean. Poster presented at the 10th Annual Conference on Architectures and Mechanisms for Language Processing (AMLaP 2004). Universite de Province, Aix-en-Provence, France. Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106, 1126–1177. Li, A. Y.-H. (1998). Argument determiner phrases and number phrases. Linguistic Inquiry, 29(4), 693–702. Li, C. N., & Thompson, S. A. (1981). Mandarin Chinese: A functional reference grammar. Berkeley, CA: University of California Press. Lin, C.-J. C. (2006). Grammar and parsing: A typological investigation of relative-clause processing. Ph.D. thesis, University of Arizona. Lin, C.-J. C., & Bever, T. (2006). Subject preference in the processing of relative clauses in Chinese. In D. Baumer, D. Montero & M. Scanlon (Eds.), Proceedings of the 25th West Coast Conference on Formal Linguistics (pp. 254–260). Somerville, MA: Cascadilla Proceedings Project. Lin, Y.-Y., & Garnsey, S. (this volume). Plausibility and the resolution of temporary ambiguity in relative clause comprehension in Mandarin. In H. Yamashita, Y. Hirose & J. Packard (Eds), Processing and producing head-final structures. Dordrecht: Springer.
9 Subject Preference, Head Animacy and Lexical Cues
193
Mak, W. M., Vonk, W., & Schriefers, H. (2002). The influence of animacy on relative clause processing. Journal of Memory and Language, 47, 50–68. MacDonald, M. C. (1994). Probabilistic constraints and syntactic ambiguity resolution. Language and Cognitive Processes, 12, 121–136. MacDonald, M. C., & Christiansen, M. H. (2002). Reassessing working memory: A comment on Just & Carpenter (1992) and Waters & Caplan (1996). Psychological Review, 109, 35–54. MacDonald, M. C., Pearlmutter, N., & Seidenberg, M. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101, 676–703. MacWhinney, B. (1989) Competition and connectionism. In B. MacWhinney & E. Bates (Eds.), The crosslinguistic study of sentence processing, (pp. 422–457). New York: Cambridge University Press. Miao, X. (1981). Word order and semantic strategies in Chinese sentence comprehension. International Journal of Psycholinguistics, 8, 23–33. Miyamoto, E. T., & Nakamura, M. (2003). Subject/object asymmetries in the processing of relative clauses in Japanese. In G. Garding & M. Tsujimura (Eds.), WCCFL 22 Proceedings (pp. 342–355). Somerville, MA: Cascadilla Press. Packard, J., Ye, Z., & Zhou, X. (this volume). Filler-gap processing in Mandarin relative clauses: evidence from event-related potentials. In H. Yamashita, Y. Hirose & J. Packard (Eds), Processing and producing head-final structures. Dordrecht: Springer. Palmer, M., Chiou, F.-D., Xue, N., & Xia, F. (2005). The Penn Chinese Treebank corpus. Catalog number LDC2005T0. Linguistic Data Consortium, Philadelphia. Pu, M.-M. (2007). The distribution of relative clauses in Chinese discourse. Discourse Processes, 43(1), 25–53. Tang, C.-C. J. (1990). Chinese phrase structure and the extended X’-theory. Doctoral dissertation, Cornell University, Ithaca, NY. Tomlin, R. S. (1986). Basic Word Order: Functional Principles. London: Groom Helm. Traxler, M, J., Morris, R. K., & Seely, R. E. (2002). Processing subject and object relative clauses: Evidence from eye movements. Journal of Memory and Language, 47, 69–90. Trueswell, J. C. (1996). The role of lexical frequency in syntactic ambiguity resolution. Journal of Memory and Language, 35(4), 566–585. Ueno, M., & Garnsey, S. (2008). An ERP study of the processing of subject and object relative clauses in Japanese. Language & Cognitive Processes, 23, 646–688. Wu, F.-Y., Haskell, T., & Andersen, E. (2006). The interaction of lexical, syntactic, and discourse factors in on-line Chinese parsing: Evidence from eye-tracking. Poster presented at the 19th Annual CUNY Conference on Human Sentence Processing, CUNY Graduate School and University Center. Yoshida, S., Aoshima, M., & Phillips, C. (2004). Relative clause prediction in Japanese. Poster presented at the 17th Annual CUNY Sentence Processing, Baltimore, MD. Zubin, D. (1979). Discourse function of morphology: the focus system in German. Discourse and Syntax, 469–504.
Chapter 10
Why Speakers Produce Scrambled Sentences: An Analysis of a Spoken Language Corpus in Japanese Tadahisa Kondo and Hiroko Yamashita
10.1 Introduction How do speakers choose words and put them into sentences in order to express their ideas? Is the process the same regardless of the type of the language? A number of studies of language production in English have reported some factors that influence the speaker’s choice in his syntactic output in utterances: animacy of nouns (Ferreira, 1994; McDonald, Bock, & Kelly, 1993; Tanaka, Branigan, & Pickering, 2005; this volume), discourse information (Arnold, Wasow, Losongco, & Grinstorm, 2000; Bock, 1977; Bock & Irwin, 1980), relative length of phrases (Arnold et al., 2000; Stallings, MacDonald, & O’Seaghdha, 1998), imageability (Bock & Warren, 1985) and prototypicality (Kelly, Bock, & Keil, 1986), among others. They reported that when there is a choice in syntactic structure as a syntactic output, such as the choice between the two ditransitive structures below, speakers tend to choose the structure that enables them to say those words with such influencing factors ahead of those that do not have such properties. For example, Bock and Irwin (1980) found that speakers of English tend to choose prepositional object structure in (1a) more than the double object structure in (1b) when the horse is ‘‘given’’ information that appeared in the previous discourse and (1b) is chosen over (1a) when the cowboy represents given information. (1)
a. The rancher sold the horse to the cowboy. b. The rancher sold the cowboy the horse.
Notice that English is a head-initial language with fairly rigid word-order. Examining speaker choices in speaking of a language that has different T. Kondo (*) Human and Information Science Laboratory, NTT Communication Science Laboratories, NTT Corporation, 3-1, Morinosato-Wakamiya, Atsugi, Kanagawa, 243-0198 Japan e-mail: [email protected]
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_10, Ó Springer ScienceþBusiness Media B.V. 2011
195
196
T. Kondo and H. Yamashita
linguistic properties, such as a head-final language with relatively free wordorder gives us opportunities to gain further understanding of the process of language production. In this paper, we report some findings of factors influencing word-order in the production of Japanese. We offer the first analysis of word-order choices in Japanese that is based on production in a natural setting: in a spoken language corpus (the Corpus of Spoken Japanese: hereafter CSJ). The results demonstrate the two factors that influence speakers’ choice of sentences in Japanese: length (weight) and referentiality. The results complement the existing experimental and corpus studies and at the same time present issues to be further explored by both experimental and corpus settings.
10.1.1 Flexible Word-Order and its Determinant in Japanese Japanese is a strictly head-final language that allows a great degree of flexibility in word-order. Except for the clause-final position of predicates, arguments and other phrases trade places while keeping the semantic content relatively the same. Thus in a transitive sentence, the order of the accusative-marked object may follow or precede the nominative-marked subject, as shown in (2a,b). (2)
a. John-ga John-NOM b. Mary-wo Mary-ACC ‘John saw Mary.’
Mary-o Mary-ACC John-ga John-NOM
mita. saw mita. saw
In a ditransitive sentence, the nominative-marked subject, accusativemarked object and dative-marked indirect object may trade places as follows. While the order in (3a) is ‘‘canonical,’’ i.e., the most intuitive or commonly used order,1 all three arguments may change their places as in (3b-f). (3)
1
a. Otokonoko-ga boy-NOM b. Otokonoko-ga boy-NOM c. Onnanoko-ni girl-DAT
onnanoko-ni girl-DAT pizza-wo pizza-ACC otokonoko-ga boy-NOM
pizza-wo pizza-ACC onnanoko-ni girl-DAT pizza-wo pizza-ACC
ageta. gave ageta. gave ageta. gave
The notion of ‘‘canonical order’’ may be defined in several ways. It may be defined as the most widely assumed order, the most frequently observed order in natural use, or as an order at the pre-syntactic derivation stage (‘base’ order; Miyagawa & Tsujioka, 2004). In the current study, canonical order is defined as the first. See the discussion of the frequency of word-order and the strength of keeping the canonical order (canonicality) in the later sections.
10
Why Speakers Produce Scrambled Sentences
d. Pizza-wo otokonoko-ga onnanoko-ni pizza-ACC boy-NOM girl-DAT e. Pizza-wo onnanoko-ni otokonoko-ga pizza-ACC girl-DAT boy-NOM f. Onnanoko-ni pizza-wo otokonoko-ga girl-DAT pizza-ACC boy-NOM ‘The boy gave the girl (some) pizza.’
197
ageta. gave ageta. gave ageta. gave
In addition to the flexibility of words among arguments, virtually any argument may be phonologically ‘‘null,’’ when its content is retrievable from the context. This further adds flexibility in the structure of overtly spelled sentences in Japanese. With all these possible choices of sentences, with only a slight difference in meaning from others, how speakers of a language such as Japanese choose one structure over others has been a question (e.g., Dryer, 1980; Hawkins, 1994; Yamashita, 2002; Yamashita & Chang, 2001; Yamashita & Kondo, 2008). In the field of language production, some experimental studies on Japanese report that Japanese speakers choose a syntactic structure that enables them to say more accessible words ahead of those that are not in a sentence, as in English. For example, in the same paradigm as Bock and Irwin (1980), Ferreira and Yoshita (2003) found the ‘‘given’’ information ahead of the ‘‘new’’ information in Japanese. They reported that when the direct object was given information, speakers chose a dative structure whose direct object was spoken ahead of the indirect object by scrambling but the indirect object tended to be spoken ahead of the direct object when the indirect object represented the given information. Tanaka et al. (2005, this volume) reported that Japanese speakers tend to choose the sentence that positions the animate argument ahead of the inanimate argument by scrambling or by passivization. These studies reported that the factors that have been found to influence syntactic choice in English (Bock, 1977; Bock & Irwin, 1980; Ferreira, 1994; McDonald et al., 1993) also influence the choices speakers make in Japanese. In contrast, some studies on Japanese and other head-final languages report a tendency that is different from English. It is known that when choosing between two phrases of different length in English, positioning a short phrase ahead of a long one is preferred (‘‘short-before-long’’ preference), sometimes known as ‘‘heavy NP shift.’’ Thus, the sentence that places the long indirect object later in the sentence, as in (4a), is preferred over the one that places it ahead of a short direct object, as in (4b) (e.g., Arnold et al, 2000; Hawkins, 1994; Kimball, 1973; Stallings et al., 1998; Wasow, 1997). (4)
a. John showed the toy to the boy who came into the room crying. b. ??John showed the boy who came into the room crying the toy.
Such preference of order is reversed in Japanese: Yamashita & Chang (2001) reported that when solicited to produce sentences with different phrase lengths,
198
T. Kondo and H. Yamashita
Japanese speakers tend to produce sentences in the ‘‘long-before-short’’ order, as reported in the written corpus analysis in Hawkins (1994). When one of the argument phrases, for example, onnanoko ‘‘the girl’’ or pizza ‘‘pizza’’ in (3) is modified by a long modifier such as a relative clause timata-de uwasa-ni natteiru ‘‘everyone is talking about,’’ speakers showed a tendency to place such a long argument ahead of the short phrase, even at the expense of changing the canonical order. Experimental studies are beneficial in the sense that conditions can be controlled better than in observational studies. At the same time, the factors to be examined tend to be more limited than with naturally obtained data. Corpusbased study enables researchers to gain insight on multiple factors observed in the naturally occurring utterances and examine them in the same set of data. In a small corpus study based on written languages, Yamashita (2002) reported cases of word-order changes in Japanese. Yamashita examined five types of sentences. Out of 2,635 sentences found in written language (a mix of short magazine interviews, essays and articles), 19 cases of scrambled sentences were found. In the qualitative analysis, Yamashita observed that long phrases such as complex NPs and sentential complements tended to be scrambled ahead of other short phrases. The study also found that scrambled phrases tended to include words representing ‘‘old’’ information, more specifically ‘‘referential’’ phrases (sonna ‘‘such’’, sono ‘‘its’’) that referred directly to the preceding context. Due to the small number of sentences examined manually, unfortunately, no statistical analyses were conducted in Yamashita (2002). A large spoken corpus offers a unique opportunity in the study of production by enabling us to investigate multiple factors in a natural setting and observe the phenomena in a large number of sentences. The current study attempts to examine the effects of possible influential factors on word-order choices in the largest corpus of spoken Japanese: The Corpus of Spoken Japanese (CSJ) (National Institute for Japanese Language & National Institute of Information and Communications Technology, 2004). Two factors that were found in the literature, the length of phrases and the referentiality of words in natural utterances, are examined. Taking advantage of the wide range of structures and large number of utterances in a corpus, the frequency of scrambling by the sentence type will also be reported.
10.1.2 Target Sentence Types and Factors Analyzed in the Study Yamashita (2002) examined several structures found in written articles, essays and transcribed (and also edited) interviews in magazines. She examined transitive sentences that included nominative (NOM)/topic (TOP) and accusative (ACC) phrases, ditransitive sentences that included dative (DAT) and ACC phrases and ditransitive sentences that included sentential complements with NOM/TOP phrases, inflectional phrases and/or DAT phrases. Four of 19 cases
10
Why Speakers Produce Scrambled Sentences
199
of scrambling were found in transitive sentences between NOM/TOP and ACC phrases, five cases were found between DAT and ACC phrases in dative sentences and the remainder were found between NOM/TOP and a sentential complement, or between DAT and a sentential complement. Sentential complements, however, may not be appropriate for examining the length effect in scrambling because sentential complements are usually longer than ordinary NP-phrases. Therefore the current study examined the following three types of structures in which the ACC phrase’s position could be changed: transitive sentences (5a: TR sentences) with ‘‘NP-(NOM) NP(ACC)’’ and the two structures that contain ni-marked NP, ditransitive sentences (5b: DTR sentences) and transitive sentences with a locative (LOC) phrase (5c: TRL sentences). These two structures share the same markers on the surface, ‘‘NP-ga NP-ni NP-wo V’’ but they were separately examined due to the difference in the nature of ni-marked NP. While the dative ni-marked argument in DTR has a role as a receiver of the object, the ni-marked phrase in TRL has a role as the goal or resulting position of the ACC-marked object (Sadakane & Koizumi, 1995). (5) a. Transitive sentence (TR): (ex. taberu ‘eat,’ yomu ‘read’) John-ga/wa pizza-wo tabeta. John-NOM /TOP pizza-ACC ate ‘John ate some pizza’ b. Ditransitive sentence (DTR): (ex. ageru ‘give,’ kau ‘buy,’ watasu ‘hand’) John-ga/wa Mary-ni pizza-wo John-NOM/TOP Mary-DAT pizza-ACC ‘John gave some pizza to Mary’ c. Transitive sentence with ni-marked LOC-phrase (TRL): (ex. oku ‘put,’ tasu ‘add’) John-ga/wa teeburu-ni pizza-wo John-NOM/TOP table-LOC pizza-ACC ‘John put the pizza on the table.’
ageta. gave
oita. put
The DTR and TRL included sentences without the wa-marked topic or NOM-marked subject for analysis but they were analyzed separately from sentences with NOM/TOP markers. This is because NOM/TOP phrases are frequently deleted in Japanese, especially in spoken utterances. Passive and causative structures were excluded from this study. The first factor examined is the phrase length. The number of Bunsetsu (a content word followed by a case marker or a post position when applicable) was used as the measure of length in this study because it is the minimal semantic unit and most intuitively the smallest unit for native speakers. An example of the number of Bunsetsu is shown in (6).
200
(6)
T. Kondo and H. Yamashita
Kooiuhuuna kotae-wo hutuu-no this way answer-ACC ordinary ‘The answer like this, humans write’ [Kooiuhuuna kotae-wo] : [hutuu-no ningen-ga] number of Bunsetsu = 2 : 2
ningen-ga human-NOM
kaku write
The second factor considered is ‘‘referentiality.’’ In the small-scaled corpus study in Japanese, Yamashita (2002) observed that phrases that contain referential words such as kore ‘‘this’’ or sonna ‘‘such,’’ referring to the immediately preceding context, tend to be scrambled ahead of other phrases in Japanese. These referential phrases represent ‘‘given/old’’ information (Bock, 1977; Bock & Irwin, 1980; Ferreira & Yoshita, 2003). Furthermore, these phrases are ‘‘salient’’ because they have discourse prominency (Levelt, 1989). The study tagged the phrases that contain referential phrases as below: (7)
a. Deictic NPs: kore ‘this one’, sore ‘that one (near the listener),’ are ‘that one (away from both speaker and listener)’ b. Deictic Determiners: kono ‘this ’ sono ‘that (near the listener)’, ano ‘that (away from both speaker and listener)’ c. Locatives: koko/kotira ‘here (this place),’ soko/sotira ‘there (near the listener),’ asoko/atira ‘there (away from both speaker and listener)’ d. like : kono yoona ‘like this,’ sono yoona ‘like that (referring to information recently mentioned),’ ano yoona ‘like that (referring to information mentioned while ago or information provided by someone else)’
10.2 Method 10.2.1 Corpus The CSJ (National Institute for Japanese Language and National Institute of Information and Communications Technology, 2004) contains 7.5 million words, 660 h of spontaneous speech in total. This study used only the data from conference presentations, which contain 3.3 out of 7.5 million words and whose length is 275 h out of 660 h. The CSJ provided morphological information for all the data. Every morpheme is extracted from the transcription of utterances and several tags are attached, such as part-of-speech and mora length. However, some syntactic information necessary for this study is not provided, such as syntactic structure and syntactic role. Only a portion of the data (the ‘‘Core’’ data) in the CSJ is parsed (about 0.2 million words from 3.3 million words). Therefore, some
10
Why Speakers Produce Scrambled Sentences
201
Table 10.1 Size of corpus (CSJ) and actual analyzed data in this study Conference presentations Core Non-core (by hand) CSJ total Total Analyzed Words (million) Total (hours) Presentation Sentences Clauses Male / female Range of age
7.5
3.3
0.2
–
660 –
275 987 79,010 317,625 814/173 63.5–67.5
15 55 6665 18119 37/18 63.3-67.3
– – 15,000 65,000 – –
– – –
additional data that are not in the ‘‘Core’’ data needed to be parsed by the authors (see ‘‘Non-Core’’ column in Table 10.1) because the ‘‘Core’’ data were too small to obtain enough data to conduct statistical analyses on relevant factors.
10.2.2 Procedure As the first step of the analysis, the target utterances were extracted from the conference presentation data in the CSJ. The target utterances were extracted from all of the ‘‘Core’’ data and randomly selected items from the ‘‘non-Core’’ data. In total, more than 20,000 sentences were extracted for this study. The sentences consisted of several clauses connected with conjunctions such as ‘‘and’’ or ‘‘then.’’ Clauses were the unit of analysis. There were more than 80,000 clauses. Some clauses contained embedded clauses, such as the relative clauses shown in (8). Such clauses were analyzed recursively, thus an embedded clause and a main clause containing the relative clause modifier were treated as two different sentences. For simplicity, we refer to these analysis units (clauses) as ‘‘sentences’’ in this study. (8) Yoko-ni Hanako-ga ageta hon-wo Taro-ni miseta. Yoko-DAT Hanako-NOM gave book-ACC Taro-DAT showed ‘(Yoko/someone) showed Taro a book that Hanako presented to Yoko.’ [analyzed unit 1] Yoko-ni Hanako-ga ageta [analyzed unit 2] (Yoko-ni Hanako-ga ageta) hon-wo Taro-ni miseta. Then the sentences that matched the target sentence types TR, DTR and TRL were selected and counted. As a result, 1,107 sentences (including embedded sentences) were selected. Among them 706 sentences were TR, 120 sentences were DTR and 281 sentences were TRL. Then in all those cases, the number of Bunsetsu, morae and morphemes in each phrase were counted for
202
T. Kondo and H. Yamashita
length effect analysis. Additionally, the referential phrases in relevant arguments were identified and marked for the analysis of referential phrase effects.
10.3 Results 10.3.1 Frequency of Scrambling The frequencies of scrambled structures for the three types of sentence are shown in Table 10.2. Table 10.2a shows the number of canonical and scrambled order in TR. Out of 706 TR sentences, 463 had a NOM-marked subject, of which 8.2% (38/463) were the scrambled order (ACC-NOM). Two hundred and forty-three TR sentences had a topic, of which 3.3% (8/243) were scrambled. When both TR with a subject and with a topic were combined, 6.5% (46/706) of sentences were scrambled.2,3 Table 10.2b shows the frequency of scrambling and canonical sentences in DTR. The frequency of scrambling is 37.6% (35/93) between ACC and DAT phrases in sentences without a NOM-marked subject or topic and the frequency Table 10.2 Frequency of scrambling in the three types of sentence With NOM subject with TOP total a. TR (NOM-ACC) Canonical Scrambled Scrambling%
425 38 8.2
235 8 3.3
660 46 6.5
Without NOM subject/TOP
With NOM subject/TOP
total
58 35 37.6
24 3 11.1
82 38 31.7
b. DTR (DAT-ACC) Canonical Scrambled Scrambling% c. TRL (LOC-ACC) Canonical 134 9 143 Scrambled 135 3 138 Scrambling% 50.2 25.0 49.1 Note: In all cases, canonical indicates the order in which the ACC-phrase follows the other phrase; scrambled indicates that the ACC-phrase precedes the other phrase. 2
The analysis excluded cases in which sentences contained both a subject and a topic in one sentence. 3 In TR with a TOP phrase and in DTR or TRL with a topic or subject, the number of scrambled sentences was too small to investigate the effects of phrase length and referential phrases. Therefore, these cases were excluded in the subsequent analysis.
10
Why Speakers Produce Scrambled Sentences
203
of scrambling was 11.1% (3 / 27) when the sentences contained the overt subject or topic. Combined, the frequency of scrambling was 31.7%. What is of particular interest is the high frequency of the third sentence type, TRL. As shown in Table 10.2c, the frequency of scrambling in TRL was around half (50.2% without NOM/TOP phrase, 49.1% in total). These results show that the frequency of scrambling varied depending on the sentence type. The scrambling frequency was low in TR, suggesting that the tendency to keep canonical order, or canonicality, is strong. On the other hand, a higher frequency of scrambling was found between the DAT and the ACC in DTR: around 30%. Therefore, it is not always the case that the frequency of scrambling is low in Japanese. These results mean that TR has a stronger canonicality than DTR. Furthermore, in TRL, the frequency of the ACC phrase preceding the LOC phrase was nearly half of the time. This shows that the canonicality of TRL is quite low. In this type of sentence, phrase length and referential phrases may be more important factors, questioning the notion of ‘‘canonical’’ order in this sentence type.
10.3.2 Phrase Length and Word-order 10.3.2.1 Number of Bunsetsu and Frequency of Scrambling Table 10.3a shows the number of scrambled sentences in TR according to the number of Bunsetsu in the ACC-marked object. Table 10.3b shows the number of scrambled sentences according to the number of Bunsetsu in ACC with DAT and Table 10.3c shows ACC with LOC. These results show that the postposed phrases (NOM, DAT and LOC phrases) were only one Bunsetsu long in almost all scrambled sentences, though there were a small number of exceptions: in particular, scrambled sentences were observed even when the LOC phrases were longer than one Bunsetsu in TRL. Here it appears that for scrambling to occur, the postposed phrases (the phrase over which a phrase is scrambled) must be quite short. In the next section, differences in the number of Bunsetsu between ACC and other phrases are examined.
10.3.2.2 Differences in Phrase Length and the Frequency of Scrambling This section examines the difference between the length of the scrambled phrase and the postposed phrase that was in the preceding position in the canonical sentence. The relative frequency distributions of the differences in length in canonical and scrambled sentences are shown in Fig. 10.1a for TR, Fig. 10.1b for DTR and Fig. 10.1c for TRL. The difference here is the number of Bunsetsu in ACC-phrases in comparison to the other phrases, NOM, DAT and LOC, respectively. The positive values in the differences indicate that the ACC-phrase
204
T. Kondo and H. Yamashita Table 10.3 Number of scrambled sentence (a) Number of Bunsetsu in NOM
in ACC
1
2
3
4
>=5
1 2 3 4 5 >=6 total
10 6 6 4 2 5 33
0 0 0 0 0 1 1
0 0 2 2 0 0 4
0 0 0 0 0 0 0
0 0 0 0 0 0 0
total 10 6 8 6 2 6 38
(b) Number of Bunsetsu in DAT in ACC
1
2
3
4
>=5
1 2 3 4 5 >=6 total
15 9 4 1 1 1 31
0 2 0 0 0 0 2
0 1 0 0 0 0 1
0 0 0 0 0 0 0
1 0 0 0 0 0 1
total 16 12 4 1 1 1 35
(c) Number of Bunsetsu in LOC in ACC
1
2
3
4
>=5
1 2 3 4 5 >=6 total
46 19 11 6 5 7 94
6 8 2 3 2 3 24
5 2 0 4 1 1 13
0 0 0 0 0 1 1
0 1 0 1 1 0 3
total 57 30 13 14 9 8 135
is longer than the other phrase and the negative values indicate that the ACCphrase is shorter than the other phrase. The overall distribution patterns of the three sentence types are similar. In all three sentence types, the distribution patterns of the differences in the length of Bunsetsu for the scrambling sentences are shifted to the right in comparison to the canonical sentences, indicating that the length of ACC-phrases are longer in the scrambling sentences than in the canonical sentences. The results from ANOVA confirmed the significant differences in length between canonical and scrambled sentences [TR: Fð1; 461Þ ¼ 22:27; p50:001; DTR: F(1,91) = 12.36, p50:001; TRL: Fð1; 267Þ ¼ 41:62; p50:001]. These results again confirm length effects in scrambling and the tendency of long before short in scrambled sentences in Japanese. Among the three structures, there were some differences observed in terms of distribution. In TR, no scrambled sentence was observed when the ACC-phrase was shorter than the NOM-phrase (the difference is less than zero). At the same
Why Speakers Produce Scrambled Sentences
Fig. 10.1 Relative frequency of canonical and scrambled sentences in terms of differences in the number of Bunsetsu in each type of sentence. The difference here is the number of Bunsetsu in ACC-phrases in comparison to NOM phrases in TR(a), DAT phrases in DTR(b) and LOC phrases in TRL(c)
205
a.TR 0.5 relatitve freq.
10
Canonical
0.4
Scrambled
0.3 0.2 0.1 0.0 –6
–4
–2
0
2
4
6
8
length difference
b. DTR 0.5
Canonical
relatitve freq.
Scrambled 0.4 0.3 0.2 0.1 0.0 –6
–4
–2
0
2
4
6
8
length difference
relatitve freq.
c. TRL 0.5
Canonical Scrambled
0.4 0.3 0.2 0.1 0.0 –6
–4
–2
0
2
4
6
8
length difference
time, there are many canonically ordered sentences even though the short NOM-phrases are placed before the long ACC-phrases (the difference is more than zero), for example, there are about 20% canonical sentences in cases where the ACC-phrase is one Bunsetsu longer than the NOM-phrase (the difference is
206
T. Kondo and H. Yamashita
one) as shown in the canonically ordered sentence of Fig. 10.1a. These results demonstrate that canonical order is strongly preferred and that the influence of length on scrambling was limited in TR. Likewise, no scrambled sentence was observed when the ACC-phrase was shorter than the DAT-phrase in DTR. In contrast to TR and DTR, there are several scrambled sentences when the ACCphrase is shorter than the DAT-phrase (the difference is less than zero) in TRL. At the same time, there exists a smaller portion of canonically ordered sentences in DTR or TRL than in TR when the ACC-phrase is longer than the DATphrase or LOC-phrase (the difference is more than zero). These results portray the difference in the tendency to prefer canonical order, canonicality of each sentence type. Namely, canonicality is ranked in the order of TR, DTR and TRL.
10.3.3 Referential Phrases and Word-order This section summarizes the effect of words with referential expressions on word-order. Table 10.4 shows the observed number of scrambled and canonical sentences divided by phrases that contain a referential phrase, such as kore ‘‘this,’’ or sono hanashi, ‘‘that story.’’4 There is no scrambled sentence with a referential phrase in the NOM or DAT phrase in the TR sentence type shown in Table 10.4a and the DTR type in Table 10.4b. This means that when there is a referential phrase in the NOM or DAT phrase, the sentences are always in canonical order in TR and DTR sentences, keeping those referential phrases ahead of the ACC phrases. At the same time, there is no canonical sentence with a referential phrase in the ACC phrase in Table 10.4b and only one canonical sentence with that in Table 10.4c, meaning that the referential phrase in the ACC phrase always or almost always causes scrambling in DTR and TRL sentences. From these results, the effects of referential phrases on word-order were clear, namely, word order in DTR seems to be influenced by which phrase contains the referential phrase. In other words, the other factors were effective only when no referential phrase appeared. As for TR, the other factors seem to be effective only when no referential phrase occurs in the NOM phrase and as for TRL, the other factors seem to be effective only when there is no referential phrase in the ACC phrase. The chi-square analysis of binary divided cross table by scrambling (canonical vs. scrambled) and referential phrases (ACC vs. the others excluding ‘‘Both’’) shows significant bias by the referential phrase [TR: chi2 ¼ 57:50; p50:001; DTR: chi2 ¼ 20:19; p50:001; TRL: chi2 = 37.38, p < 0.001]. 4
A very small number of sentences contained referential phrases in both relevant arguments: only five sentences in the transitive, two sentences in the ditransitive and 10 in the transitive with locative, as shown in the ‘‘Both’’ column in Table 4. Those 17 sentences were excluded in the following analyses.
10
Why Speakers Produce Scrambled Sentences
207
Table 10.4 Number of sentences according to word order and referential phrases None Both NOM/DAT/LOC ACC total a. TR (NOM-ACC) canonical scrambled total
331 22 353
5 0 5
67 0 67
22 16 38
425 38 463
b. DTR (DAT-ACC) canonical scrambled total
41 22 63
1 1 2
16 0 16
0 12 12
58 35 93
c. TRL (LOC-ACC) canonical 81 3 49 1 134 scrambled 71 7 21 36 135 total 152 10 70 37 269 Note: canonical: ACC-phrase is placed later; scrambled: ACC-phrase is placed earlier. None: no referential phrase exists; Both: referential phrases exist in both phrases; NOM/DAT/LOC: a referential phrase exists in the NOM/DAT/LOC-phrase; ACC: a referential phrase exists in the ACC-phrase.
The results might show the stronger effect of referential phrases than phrase length but the interaction should be examined. In the next section, the interaction between these two factors is examined.
10.3.4 Interaction between Length and Referentiality The logistic regression analysis was conducted for the statistical analysis of the probability of scrambling with interaction effects between length difference and referential phrases. The logistic model is shown in equation (1). log
p 1p
¼ b0 þ b1 x1 þ b2 x2 þ b3 x1 x2
(1)
p: probability of scrambling x1: length difference in number of Bunsetsu x2: referential phrase exists whether in {1: ACC, 0: the others excluding ‘‘Both’’} x1 x2: interaction
Table 10.5 shows the estimated parameters from the results of logistic regression. Length difference (b1) has a significant contribution to the probability of scrambling in all sentence types. This is consistent with the result from ANOVA in the previous section. However, in all sentence types the interaction (b2) is non-significant. As for the existence of a referential expression (b3), there
208
T. Kondo and H. Yamashita
Table 10.5 Estimated coefficients for equation (1) from logistic regression analysis Coef. S.E. Wald Z p a.TR: R2 ¼ 0:274; p50:001 3.195 0.259 12.32 < 0.001 b0: interception 0.311 0.084 3.68 < 0.001 b1: length difference b2: reference 2.612 0.466 5.61 < 0.001 0.112 0.247 0.45 = 0.651 b3: length*reference b. DTR: R2 ¼ 0:461; p50:001 0.949 b0: interception 0.618 b1: length difference 10.283 b2: reference b3: length*reference 0.618
0.272 0.229 32.824 10.699
3.49 2.70 0.31 0.06
< 0.001 < 0.01 = 0.754 = 0.954
c. TRL: R2 ¼ 0:365; p50:001 0.442 b0: interception 0.438 b1: length difference 3.864 b2: reference b3: length*reference 0.051
0.148 0.092 1.029 0.755
2.99 4.75 3.76 0.07
< 0.01 < 0.001 < 0.001 = 0.946
is a difference among the three types of sentence. In TR and TRL, the existence of a referential phrase in ACC has a significant contribution to the scrambling probability. In DTR, the existence of a referential phrase does not have a significant contribution, presumably because of a small number of items for the category. Figure 10.2 shows the observed scrambling probabilities (plots) and their fitting curves (dotted line) according to the length difference by referential phrases. The interaction was not statistically significant. However these figures show the different tendencies of the probability of scrambling according to the existence of referential phrases. The scrambling probability is always low in TR and always 1.0 when the ACC phrase contains a referential phrase in DTR and TRL, namely, always scrambled, as already shown in the previous section. To confirm this phenomenon statistically, additional data are needed. This is because there is only a small amount of data when the data are divided by two factors, especially for cases where the referential phrase existed in the ACC phrase when the length difference was minus. To sum up, we observed both the effects of phrase length and referential phrases on word-order in TR, DTR and TRL sentence types. Both factors increased the speakers’ tendency to change the canonical order to the scrambled order. Further analysis of length effects revealed that the scrambling is most likely to occur when the scrambled phrase is placed ahead of a canonically ordered phrase that is short – mostly when it is one Bunsetsu. The existence of referential phrases affected word-order significantly in TR and TRL. The analysis also found that the frequency of scrambling varied among the sentence types. In the next section, we will discuss these points in detail.
10
Why Speakers Produce Scrambled Sentences
209
(a) TR 1.0
0.0 –10
–5
0
5
10
(b) DTR Scrambling Rate
1.0 none/ No. of NOM/ obs. DAT
0.0
ACC
1000
–10
–5
0
5
10
100 10
(c) TRL
1
1.0
0.0 –10 –5 0 5 10 Length Difference (in bunsetsu)
Fig. 10.2 Probability of scrambling of ACC-phrases
The fitting curves were drawn by applying the coefficients in Table 10.5a for TR, 5b for DTR and 5c for TRL to Eq. (1). The size of circles is logarithmically proportionate to the number of the items.
10.4 General Discussion 10.4.1 Frequency of Scrambling and Sentence Types The current analyses of the spoken corpus found that the frequency of scrambled sentences varied by sentence type. While scrambling was observed infrequently in all TR sentences (6.5%), the frequency of scrambling was higher in DTR (31.7%) and in TRL (49.1%). Of particular interest is the low frequency in TR of the ACC-marked object scrambled in front of the NOMmarked subject or the TOP-marked topic. Here we attempt to account for such differences among structures by examining word-order as a result of the production process. Most theories of production commonly assume three levels in its architecture: the message level, at which the message to be conveyed is formulated; the functional level, at which grammatical functions (subject, object, indirect object, etc.) are assigned; and the positional level, at which the phonological content of the words are retrieved and assembled in the order of utterance (Bock & Levelt,
210
T. Kondo and H. Yamashita
1994; Levelt, 1989). It is considered that the elements of an utterance are processed incrementally, such that each level processes a portion of a sentence without waiting for the whole sentence to be assembled. The fact that a phrase with given information is produced ahead of those without given information as a scrambling (Ferreira & Yoshita, 2003) suggests that the concept of a phrase (lemma) that receives the grammatical function at the functional level quicker tends to be produced early in the sentence in a language like Japanese. The low frequency of scrambling in TR, which is the ordering between the subject/topic and the ACC-marked object, is presumably due to the prominent functional role of the subject and topic in transitive sentences in Japanese. In a discourse in which multiple sentences are uttered, the subject of a transitive sentence is commonly realized as a phonologically null pronoun if its semantic content is retrievable from the context. When a subject is overtly realized, it is particularly motivated to occur, such as to bear exhaustive (‘‘X and only X’’) reading meaning (Kuno, 1973). The role of overt Topic phrases also plays a special function in a discourse, such as contrastive use (Kuno, 1973) or establishing a topic from the existing (given) information among the speakers (Tsujimura, 2007). Phrases containing such prominent functions are salient, thus grammatical assignment to them is made more quickly than to other phrases, resulting in their being uttered earlier than other phrases. Overriding such a preference requires a special cause, either length or saliency, in the other phrase. The low frequency of scrambled sentences in TR, that is, the word-order between the NOM-marked subject/TOP-marked topic and the ACC-marked object, must reflect such an informational balance between two phrases. At the same time, what is of particular interest is the motivation to still produce the postposed subject or topic after the fronted ACC-marked object, instead of completely deleting it from the sentence. The examples are as follows. (9)
a. Hanashite-no ito-wo kikite-ga sikkari sooteesuru speaker-GEN intention-ACC listener-NOM well expect koto-wo sezu thing-ACC do without ‘(without doing things like) the listener very well expects what the speaker intends’ b. Kooiu gainen-wo eezyento-ga atukau (tame-ni-wa) this kind of concept-ACC agent-NOM handle (in order to-TOP) ‘(in order) for an agent to handle this kind of concept’
What linguistic factor motivated kikite ‘‘the listener’’ and eezyento ‘‘the agent’’ to be overtly present in the sentence above needs further examination; such investigation must involve a comprehensive approach of the discourse prominency of each word in the sentence, as well as the nature of all words and phrases, including their animacy, saliency and familiarity.
10
Why Speakers Produce Scrambled Sentences
211
The frequency of scrambling was higher in DTR and TRL than in TR. This is presumably because the phrases that scrambled phrases precede do not have a prominent semantic or discourse function as NOM-marked subject or topic. In DTR, the two phrases observed are the DAT-marked indirect object and the ACC-marked direct object. Although there seems to be in general a preference for canonical order, in which the DAT-phrase precedes the ACC-phrase (but see Matsuoka, 2003; Miyagawa & Tsujioka, 2004), such a preference appears to be overridden by factors such as phrase length or the existence of a referential phrase. The scrambled sentences were most frequent in the TRL structure, in which the LOC phrase was preceded by the ACC-marked object almost half the time, presumably suggesting that canonicality may be defined mainly among arguments that are strongly connected to the meaning of the event described by the predicate.
10.4.2 Head-Direction and Phrase Length The current study found that both phrase length and referentiality affect wordorder. Long phrases tend to be scrambled ahead of a short phrase, often with a single Bunsetsu; the phrase containing the referential phrase tended to be scrambled ahead of a phrase that did not contain them. Coupled with the findings of the tendency to produce a phrase representing given information ahead of those with new information by Ferreira & Yoshita (2003), Japanese raises a question about the process of language production. When all relevant phrases are equally short, for example, a single word, Japanese speakers place a given or salient phrase ahead of those that are not by either preserving the canonical order or by scrambling (Ferreira & Yoshita, 2003; Tanaka et al., 2005). On the other hand, the tendency for a long phrase to be scrambled ahead of a short phrase creates a reverse preference in a language such as English. While existing theories account for the length preference (Hawkins, 1990, 1994; Kimball, 1973; Stallings et al., 1998; Wasow, 1997; Yamashita & Chang, 2001), there has not been a theory that accounts for the accessibility and length effects comprehensively. While it is beyond the scope of the current study, we attempt to suggest a possible account of these seemingly paradoxical results. It is difficult to imagine that within one language, a different mechanism operates in producing different types of sentences: those that contain information that is discourse prominent and thus accessible and those that contain long phrases. The preference of scrambling referential phrases and long phrases, ahead of those that do not have such attributes, must come from the same set of constraints driven by the characteristics of Japanese. Likewise, the preference of shifting a long phrase to a position after the short one and the preference of positioning an accessible phrase earlier than those that are not must derive from the same set of principles in English.
212
T. Kondo and H. Yamashita
The differences in the two types of language, Japanese and English, are the direction of syntactic heads, the flexibility of word-order and the optionality of null pronouns. The grammar of English does not allow a great degree of flexibility in word-order within a structure and even the elements understood by the preceding discourse must be spelled out in a determined slot with appropriate case assigned. A typical exchange in word-order occurs in a ditransitive structure, in whose semantic content syntactic choice (NP-NP or NP-PP) is controlled by the verb. Within a long phrase, word-order is not flexible; speakers simply cannot just start with any words and conclude it grammatically. The priority in the language production process, then, is accuracy of phrases – their internal structure, assigned case and the order in which they can appear. English is a language in which the speakers must thoroughly plan and prepare phrases before they speak. When such preparation is asked of speakers, phrases that are accessible or short, whose words are more quickly selected, may tend to be produced earlier than those that are longer or hard to retrieve. In contrast, Japanese scrambling occurs before the verb and within a sentence and clause the order or appearance of phrases is flexible. The role of the verb in Japanese is to ‘‘summarize’’ what was spoken, rather than giving a ‘‘road map’’ within the unfolding VP as in English. As suggested in Iwasaki (2007, this volume), Japanese speakers start out without the concrete selection of a specific verb. After speaking only the relevant and necessary arguments, a verb is spoken, sometimes resulting in the wrong selection of the verb with the wrong subcategorization. Coupled with the flexibility of word-order and the optionality of overt arguments, the preparedness of the structure of each phrase does not need to be a priority. Instead, with a fragment of a phrase, what comes to mind, a speaker may begin speaking. Once the speaker starts to speak, there are many ways to continue and complete the phrase. Semantically prominent phrases, such as referential phrases or given information may be prepared and assigned a grammatical role at the functional level earlier than those that are not; at the same time, a long phrase contains many words. The likelihood of the words in the long phrase being prepared for speaking is more than that of a short phrase and almost any word in a phrase may start the phrase without violating the grammar. One predicts that speakers benefit from producing those long phrases before short phrases because of the ease of ‘‘simply starting talking’’ of the long phrase. Some studies have reported indirect evidence of such an effect. In a new typing-based production task, Chang, Yamashita, & Hirose (2008) found that people took less effort to change and produce word-order that resulted in long-before-short order than word order that did not. Another way to confirm this may be to measure the amount of disfluency discourse markers in the speech. In Arnold et al. (2000), when the speakers were having a difficulty producing a sentence and making many disfluency discourse markers, they tended to produce the more accessible phrase ahead of the less accessible one. We would predict that when speakers are having difficulty in production, more long-before-short order would be observed, so that their difficulty would be relieved. In this way, the paradox
10
Why Speakers Produce Scrambled Sentences
213
of long-before-short and short-before-long may be accounted for by the nature of the constraints that speakers must observe in each language.5
10.4.3 Summary The current study found the effects of length and referentiality in the wordorder of Japanese found in a large spoken corpus. Such findings had been only reported by experimental setting or by a small-scale written corpus. The current corpus-based study leads to opportunities for further exploration in the study of human language production. One is that with corpus study we can observe the linguistic factors that override the prominency of the subject or the topic in the TR type sentences. When such prominent words are preceded by a scrambled phrase, what causes scrambling must be studied through the analysis of the internal structures of such sentences and the preceding utterances that lead to such utterances. Corpora of spoken language enable researchers to analyze both aspects in naturally-occurring sentences. Large corpora enable us to examine multiple factors influencing each other and in some cases enable us to conduct statistical analyses. In some cases, out of tens of thousands of sentences, the phenomenon sought after may not occur frequently enough to conduct statistical analyses. However, that is indicative in itself (Jaeger, 2007). Another strength of corpus analysis is to question or make suggestions about experimental studies (Roland, Dick, & Elman, 2007). Close examination of naturally-occurring utterances brings new insights to researchers and corpus studies together with experimental studies may effectively advance the investigation of human language production. Acknowledgement Part of this chapter was presented at the joint workshop of the Mental Architecture for Processing and Learning of Language (MAPLL) and the Thought and Language (TL) Workshop by the Institute of Electronics, Information, and Communication Engineers (IEICE) in August 2008. We thank the reviewers of this book, the participants of those workshops, Yuki Hirose, Florian Jaeger and Franklin Chang for their comments.
5
Another possibility is suggested by a computational model of English and Japanese sentence production. The model proposed in Chang (2009) learns the syntactic representations for each language from message-sentence pairs and it can exhibit heavy NP shift in the appropriate direction for each language. In English, it depends more on statistical structural regularities in the post-verbal positions (light phrases are more frequent than heavy phrases in these positions). Japanese has few structural cues at the position where sentence structures are chosen and therefore the model depends more on meaning in these positions and heavy elements have an enriched meaning representation that biases them to go earlier.
214
T. Kondo and H. Yamashita
References Arnold, J., Wasow, T., Losongco, A., & Grinstorm, R. (2000). Heaviness vs. newness: The effects of structural complexity and discourse status on constituent ordering. Language, 76, 28–55. Bock, J. K. (1977). The effect of a pragatic presupposition on syntactic structure in question answering. Journal of Verbal Learning and Verbal Behavior, 16, 723–732. Bock, J. K., & Irwin, D. E. (1980). Syntactic effects of information availability in sentence production. Journal of Verbal Learning and Verbal Behavior, 19, 467–484. Bock, J. K., & Levelt, W. J. M. (1994). Language production: Grammatical encoding. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics. Hillsdale, NJ: Erlbaum. Bock, J. K., & Warren, R. (1985). Conceptual accessibility and syntactic structure in sentence formulation. Cognition, 21, 47–67. Chang, F. (2009). Learning to order words: A connectionist model of heavy NP shift and accessibility effects in Japanese and English. Journal of Memory and Language 61, 374–397. Chang, F., Yamashita, H., & Hirose, Y. (2008). Typing speed as a reflection of incremental sentence planning: Application of a new task to Heavy NP shift in Japanese. Poster presented at the 21st CUNY Sentence Processing Conference, Chapel Hill, NC. Dryer, M. (1980). The positional tendencies of sentential noun phrases in universal grammar. Canadian Journal of Linguistics, 25, 123–195. Ferreira, F. (1994). Choice of passive voice is affected by verb type and animacy. Journal of Memory and Language, 33, 715–736. Ferreira, V. S., & Yoshita, H. (2003). Given-New ordering Effects on the production of scrambled sentences in Japanese. Journal of Psycholinguistic Research, 32(6), 669–692. Hawkins, J. (1990). A parsing theory of word order universals. Linguistic Inquiry, 21, 223–261. Hawkins, J. (1994). A performance theory of order and constituency. Cambridge: Cambridge University Press. Iwasaki, N. (2007). Case particle errors in Japanese: Is the nominative ga a default case marker in sentence production? In C. Schutze & V. Ferreira (Eds.), The state of the art in speech error research: Proceedings of the LSA institute workshop. MIT Working Papers in Linguistics, 53, 205–219. Cambridge, MA: MIT Press. Iwasaki, N. (this volume). Incremental sentence production in Japanese: Observations from elicited speech errors in a picture description task. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. Jaeger, F. (2007). Corpus-based research on language processing: Working with naturally unbalanced data. Paper presented at the International Conference on Processing Headfinal Structures, September, Rochester, NY. Kelly, M. H., Bock, J. K., & Keil, F. C. (1986). Prototypicality in a linguistic context: Effects on sentence structure. In M. H. Kelly, K. Bock & F. C. Keil (Eds.), Journal of Memory and Language, 25, 59–74. Kimball, J. (1973). Seven principles of surface structure parsing in natural language. Cognition, 2, 15–47. Kuno, S. (1973). The Structure of the Japanese Language. Cambridge, MA: MIT Press. Levelt, W. J. M. (1989). Speaking: From intention to articulation. Cambridge, MA: MIT Press. Matsuoka, M. (2003). Two types of ditransitive construction in Japanese. Journal of East Asian Linguistics, 12, 171–203. McDonald, J., Bock, K., & Kelly, M. H. (1993). Word and world order: Semantic, phonological, and metrical determinants of serial position. Cognitive Psychology, 25, 188–230. Miyagawa, S., & Tsujioka, T. (2004). Argument structure and ditransitive verbs in Japanese. Journal of East Asian Linguistics, 13, 1–38. National Institute for Japanese Language & National Institute of Information and Communications Technology. (2004). The Corpus of Spontaneous Japanese.
10
Why Speakers Produce Scrambled Sentences
215
Roland, D., Dick, F., & Elman, J. (2007). Frequency of basic English grammatical structures: A corpus analysis. Journal of Memory and Language, 57, 348–379. Sadakane, K., & Koizumi, K. (1995). On the nature of the ‘‘dative’’ particle in Japanese. Linguistics, 33, 5–33. Stallings, L. M., MacDonald, M. C., & O’Seaghdha, P. G. (1998). Phrasal ordering constraints in sentence production: Phrase length and verb disposition in heavy-NP shift. Journal of Memory and Language, 39(3), 392–417. Tanaka, M., Branigan, H., & Pickering, M. (this volume). The production of head-initial and head-final languages. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. Tanaka, M., Branigan, H. P., & Pickering, M. J. (2005, March). The role of animacy in Japanese sentence production. Paper Presented at the CUNY Sentence Processing Conference, Tucson, AZ, USA. Tsujimura, N. (2007). An introduction to Japanese linguistics. Oxford: Blackwell. Wasow, T. (1997). Remarks on grammatical weight. Language Variation and Change, 9, 81–105. Yamashita, H. (2002). Scrambled sentences in Japanese: Linguistic properties and motivations for production. TEXT, 22(2), 597–633. Yamashita, H., & Chang, F. (2001). ‘‘Long before short’’ preference in the production of a head-final language. Cognition, 81(2), B45–B55. Yamashita, H., & Kondo, T. (2008). Effects of phrase length and referenciality in the Wordorder in Japanese production: A corpus analysis. IEICE Technical Report, 108(184), 125–130.
Part V
Processing Relative Clauses in Chinese
Chapter 11
Filler-Gap Processing in Mandarin Relative Clauses: Evidence from Event-Related Potentials Jerome L. Packard, Zheng Ye, and Xiaolin Zhou
11.1 Introduction One of the goals of sentence processing research is to determine how utterances are parsed and comprehended in real time. Most models of sentence processing posit that syntactic structure is built incrementally based on the information contained in each incoming word (e.g., Kamide, Altmann, & Haywood, 2003). According to these models, each incoming word is processed in two ways: it is both integrated with previous input and it also plays a role in generating a prediction of input to follow. The lexical properties of verbs have been found to play a large role in the prediction and integration of incoming sentence constituents. Such properties include a verb’s syntactic requirements, which may be termed its subcategorization or selectional restrictions. Verb selectional restrictions figure prominently in sentence processing research, because parser expectations are affected by how strongly information has been predicted by those selectional restrictions (e.g., Altmann & Kamide, 1999; Boland, Tanenhaus, Garnsey, & Carlson, 1995; Kamide et al., 2003).1 In an SVO language like Chinese, the integration of a verb’s internal arguments is facilitated when those arguments match expectations predicted by the selectional restrictions specified in the verb’s lexical entry (Li, Shu, Liu, & Li, 2006). As an example, the selectional restrictions on the Mandarin verb xihuan ‘‘to like’’ includethatitselectsforanobligatorynominalcomplementsisterthatcanserveeither as a simple NP direct object, as in the object NP Xiaolan in (1), or as a complex NP in the form of an embedded sentence such as diao yu cast-fish ‘‘catch fish’’ in (2). J.L. Packard (*) Department of East Asian Languages and Culture, University of Illinois at Urbana-Champaign, 707 S. Mathews, 61801, Urbana-Champaign, IL, USA e-mail: [email protected] 1
The context affecting an incoming word’s interpretation can include other knowledge such as discourse pragmatic information (e.g., Tanenhaus, Carlson & Trueswell, 1989) in addition to the lexico-semantic and syntactic contextual information that we will focus on here.
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_11, Ó Springer ScienceþBusiness Media B.V. 2011
219
220
J.L. Packard et al.
(1)
Zhangsan xihuan Xiaolan Zhangsan like Xiaolan ‘Zhangsan likes Xiaolan.’
(2)
Zhangsan xihuan diao yu Zhangsan like cast fish ‘Zhangsan likes fishing.’
Given the parser expectation of a following NP for xihuan, if instead Zhangsan xihuan is followed by the copula shi ‘‘is’’ or the classifier for books ben, these words are difficult to integrate into the phrase under construction (and result in ungrammatical strings) because they run counter to the expectations generated by the selectional restrictions on the verb xihuan. As another example, the lexical listing of the Mandarin ditransitive verb gei ‘‘give’’ includes the requirement that one of its two internal arguments be a lexical item that semantically can fill the role of Goal and another be a lexical item that can fill the role of Theme. So in the infelicitous sentence ?Ta gei wo shuizai he give me water-disaster (?‘‘he gave me a flood’’), the second NP shuizai ‘‘flood’’ would be difficult to integrate because it does not fulfill parser expectations generated by the selectional restrictions on the verb gei involving Theme. Evidence for integration difficulty has been found in sentence processing studies using self-paced reading, eye-tracking and event-related potential (ERP) research methods. In self-paced reading studies, integration difficulty is measured by the time it takes to read the word being integrated (e.g., Hsiao & Gibson, 2003; Just & Carpenter, 1992; King & Just, 1991). In eye-tracking studies, increased integration difficulty may be inferred by gaze duration and other measures (e.g., Rayner, 1998). In ERP studies, integration difficulty has been associated with the P600 – a characteristic positive wave form that occurs approximately 600 ms after word onset. The P600 originally was associated with garden-path sentences and a variety of different syntactic violations and was thought to reflect the detection of syntactic anomaly (Hagoort, Brown, & Groothusen, 1993; Neville, Nicol, Barss, Forster, & Garrett, 1991; Osterhout & Holcomb, 1992). More recently, P600 ERP effects have been observed in processing sentences that are not necessarily ungrammatical or atypical but that require more cognitive effort to process. For example, a P600 component has been found for syntactic diagnosis and reanalysis processes (Friederici, Hahne, & Mecklinger, 1996; Friederici, Mecklinger, Spencer, Steinhauer, & Donchin, 2001) and for the detection of ambiguity (Frisch, Schlesewsky, Saddy, & Alpermann, 2002). P600 effects have also been found for processing long-distance syntactic dependencies in well-formed sentences. For example, a P600 has been observed for the integration of the verb containing the gap that completes a filler-gap dependency
11
Filler-Gap Processing in Mandarin Relative Clauses
221
(Felser, Clahsen, & Mu¨nte, 2003; Fiebach, Schlesewsky, & Friederici, 2002; Kaan, Harris, Gibson, & Holcomb, 2000; Phillips, Kazanina, & Abada, 2005). To illustrate, Kaan et al. (2000) found a larger P600 reflecting integration difficulty for processing the verb in a ‘‘wh-’’ complement (such as the verb imitated in (3).) when compared with the same verb in a ‘‘whether’’ complement (as in (4)). (3) Emily wondered whoi the performer in the concert had imitated Øi (. . . for the audience’s amusement) (4) Emily wondered whether the performer in the concert had imitated (. . . a pop star for the audience’s amusement) The authors argue that the observed P600 effect – the larger P600 in the whcomplement condition – reflects increased cost for processing the verb in the whcomplement, because integrating the verb includes integrating its anaphoric gap with the antecedent filler who, which is held in memory over the length of the dependency. A similar result was obtained by Phillips et al. (2005), who found a larger P600 component on the verb that marks the completion of a wh-dependency (i.e., the verb recognize in (5)) than the P600 that occurs on the same verb (recognize in (6)) in a sentence with no dependency. The authors argue that the P600 effect – the larger P600 in the dependency condition – indicates increased processing cost, because integrating the verb in the dependency condition also entails integrating the verb gap with its antecedent filler stored in memory. (5) The detective hoped that the lieutenant knew [which accomplice]i the shrewd witness would recognize Øi in the lineup. (6) The detective hoped that the lieutenant knew that the shrewd witness would recognize the accomplice in the lineup. Relative clauses contain the same types of filler-gap dependencies seen in the wh- complement sentences (in (3)–(6)) above. Research has shown that in English, object-gap clauses (as in (7)), which have a longer filler-gap dependency, are harder to process than subject-gap clauses (as in (8)), which have a shorter filler-gap dependency (e.g., Gibson, Desmet, Watson, Grodner, & Ko, 2005; King & Just, 1991; Traxler, Morris, & Seely, 2002; Traxler, Williams, Blozis, & Morris, 2005; Wanner & Maratsos, 1978).2 The distance between the filler (‘‘reporteri’’) and the gap (marked by Øi) in the 2
Several proposals have been offered to account for processing difficulty in object-gap relatives, including accessibility of the relativized NP (Keenan & Comrie, 1977), differences in perspective between the relative clause and matrix sentence (MacWhinney, 1982; MacWhinney & Pleh, 1988), canonical vs. non-canonical word order in the relative clause (MacDonald & Christiansen, 2002), preference for the role of agent to be expressed by the clauseinitial NP (Diessel & Thomasello, 2005) and distance between filler and gap, either in linear
222
J.L. Packard et al.
(7) the reporteri who the senator attacked Øi (8) the reporteri who Øi attacked the senator object-gap and subject-gap clauses in (7) and (8) respectively is greater in the object-gap clauses in English. Further, ERP effects reflecting the increased processing demands for the more difficult object-gap relatives have been found both for the reading (King & Kutas, 1995) and auditory processing (Muller, King, & Kutas, 1997) of English relative clauses. The ERP effect found in both the King and Kutas and Muller et al. studies was an increased anterior negativity for the object-gap clauses beginning after the relative pronoun who and continuing through the processing of the following main clause verb. Notably, the King and Kutas and Muller et al. studies did not search for and did not observe a P600 effect at the point of filler-gap integration (i.e., the relative clause verb), even though such an effect is predicted to occur there if the P600 indexes syntactic integration difficulty, as demonstrated by Kaan et al. (2000) and Phillips et al. (2005). But because the relative clause verb occurs at different places in the subject-gap and object-gap clauses in English, to look for a P600 effect at the relative clause verb in the King and Kutas and Muller et al. studies would have required comparing the ERP profiles of words that occur in different positions within the clause, namely, clause initial position in the subject-gap clauses and clause final position in the object-gap clauses.3 In Mandarin, however, because relative clauses are right-headed, the point of filler-gap integration is predicted to occur at a constant position within the clause for the different dependency lengths, namely, at the position of the filler. As seen in 9, in a Mandarin relative clause the modifier (zhui gou de in (9)) occurs to the left of the head (the noun nanhair in (9)) and so the verb containing the gap (zhui in (9)) occurs before the filler (nanhair in (9)). (9) Øi zhui gou de nanhairi chase dog DE boy ‘the boy who chased the dog’ Because filler-gap integration is predicted to occur at the head, if there is a differential sentence processing cost for subject-gap versus object-gap relatives terms, measured by the number of words or constituents that intervene between the filler and the gap (King & Just, 1991; Kluender & Kutas, 1993; Gibson, 1998; Hsiao & Gibson, 2003), or in structural terms, measured by differences between subject-gap and object-gap relatives in the depth of embedding of the filler versus the gap (Lin and Bever, 2006; O’Grady, Lee, & Choo, 2003). 3 King and Kutas did directly compare the ERP profiles of the relative clause verbs in subjectgap and object-gap clauses but the area they examined was the 200–500 ms interval post-verbonset, which is earlier than would have been necessary to observe a P600 effect.
11
Filler-Gap Processing in Mandarin Relative Clauses
223
associated with integrating the filler with the gap over a longer distance, in Mandarin it would be the cost of holding the antecedent gap-plus-verb (i.e., Øi zhui in (9)) in memory until the filler (e.g., the relative clause head nanhair in (9)) is reached. It is expected that in Mandarin, as in English, the head of the relative clause functions as the filler that would satisfy the relative clause verb selectional restriction, because the relative clause head is an argument (the subject, in (9)) of the relative clause verb. In Mandarin, however, there is an alternative possibility. Mandarin routinely allows a null element to serve as the head of a relative clause (see, e.g., Lin & Garnsey, this volume; Simpson & Wu, 1999), as seen in comparing (9) and (10). In (10), the null status of the head allows it to be interpreted as a lexically null, unspecified NP, usually translated as ‘‘the one’’ in English. This suggests (10)
Øi zhui gou de Øi chase dog DE Ø ‘the one that chased the dog’
that in Mandarin, the selectional restrictions of the relative clause verb could be satisfied with the appearance of the relative marker de, because when de appears in the input stream, the parser receives information that (a) an embedded complex NP is being constructed and (b) that complex NP is capable of satisfying the selectional restrictions of the relative clause verb, because the position of the gap at the antecedent verb tells the parser what the thematic specifications of that NP are. Note that the selectional restriction satisfied by relative marker de in this instance may be considered the form class categorial or ‘‘c-selection’’ selectional restriction and not the word-specific semantic or ‘‘s-selection’’ selectional restriction, as discussed in Grimshaw (1979) and Pesetsky (1993). This is because the marker de contains no lexical semantic information and so its appearance only informs the parser of the presence of an NP that will satisfy the structural NP subcategorization requirements of the verb. The lexically filled, full-NP relative clause head, on the other hand, satisfies the word-specific ‘‘s-selection’’ selectional restriction, because it contains the lexical information needed to satisfy the verb’s semantic selectional restrictions. One way to investigate whether it is the relative clause head or the relative marker de that satisfies the selectional restrictions of the relative clause verb in on-line sentence processing would be to set up a ‘‘difficult/easy’’ integration contrast within the Mandarin relative clause that, following previous research (Kaan et al., 2000; Phillips et al., 2005), is predicted to elicit a P600 ERP component effect and then observe whether the P600 effect occurs on the relative clause head or the relative marker de. Research in Mandarin has found that subject-gap relative clauses (as in 11) are more difficult to process
224
J.L. Packard et al.
than object-gap relative clauses (as in 12; Hsiao, 2003; Hsiao & Gibson, 2003, Lin & Garnsey, 2006, 2007; cf. Lin & Bever, 20064). (11)
Øi xihuan Mali de goui Ø like Mary DE dog ‘the dog that likes Mary’
(12)
Mali xihuan Øi de gou Mary like Ø DE dog ‘the dog that Mary likes’
One of the theories (see footnote 3) that has been proposed to explain this contrast is that filler-gap integration is more difficult in the subject-gap relatives because the linear distance between the integrating verb (xihuan in 11 and 12) and the integrated constituent (gou in 11 and 12) is greater in the subject-gap relatives (Gibson, 1998; Hsiao & Gibson, 2003). If this is so, then a P600 effect – a larger P600 for the subject-gap relative clauses – is predicted. Furthermore, if the predicted P600 is in fact observed, its location would indicate where fillergap integration occurs in Chinese relative clauses. Our study had the following objectives. First, we wanted to see whether in Mandarin the neurophysiological cost of integrating an anaphoric argument with a verb-plus-gap antecedent would elicit a P600 ERP component. If filler-gap integration is more costly in subject-gap relative clauses in Mandarin, then we would expect to see a larger P600 associated with the processing of subject-gap relatives when compared with object-gap relatives. Our second objective was to determine where filler-gap integration – and by extension relative clause verb selectional restriction satisfaction – occurs in Mandarin, as indicated by the precise location of the predicted P600 effect. If the P600 effect occurs on the relative clause head, this would suggest that filler-gap integration occurs on the head and therefore relative clause verb selectional restrictions are satisfied by the head. If on the other hand the P600 component occurs on the relative marker de, it would imply that filler-gap integration occurs on de and therefore relative clause verb selectional restrictions are satisfied with the appearance of de.
11.2 Method 11.2.1 Participants Twenty students (mean age 22.2 years; 10 females) from universities in Beijing participated in the experiment. All were native speakers of Mandarin Chinese 4
See also the corpus study by Wu, Kaiser and Andersen, this volume, who found subject-gap relatives to be more frequent, thus potentially easier to process, than object-gap relatives.
11
Filler-Gap Processing in Mandarin Relative Clauses
225
and had normal or corrected-to-normal vision. All but one participant was right-handed according to performance on a Handedness Questionnaire (Li, 1983). The data from the left-handed participant were subsequently removed from the analysis, yielding data from a final total of 19 participants. All participants were paid for their participation.
11.2.2 Stimuli Three hundred and eighty draft stimulus sentences were constructed consisting of 95 sentences with each of the following four different sentence types: subject-gap relative construction as the subject of the matrix sentence (SS), object-gap relative construction as the subject of the matrix sentence (SO), subject-gap relative construction as the object of the matrix sentence (OS) and object-gap relative clause as the object of the matrix sentence (OO; please see Table 11.1 for example sentences). These initial 380 sentences were randomized across sentence types and then rated by 18 native Mandarin speakers for acceptability on a scale of 1–5, with 1 being the least acceptable and 5 being the most acceptable. The 60 most highlyrated sentences within each sentence type were selected, yielding a total of 240 stimulus sentences. The mean acceptability rating for the selected stimulus sentences was 4.21 for type SS, 3.91 for type SO, 4.27 for type OS and 4.23 for type OO. An additional 240 sentences of different lengths and syntactic structures were constructed to be used as filler sentences, yielding a total of 480 experimental sentences.
11.2.3 Procedure Participants were tested in a soundproof, electrically-shielded booth, seated in a comfortable chair with a distance of approximately 100 cm between their
Sentence types
Table 11.1 Stimulus sentence examples Example sentences
SS
完成 论文 的 学生 获得了 学位 complete thesis de student obtained degree ‘‘The student that completed her thesis obtained her degree’’
SO
教授 指导 的 学生 发表了文章 professor advise de student published article ‘‘The student that the professor advised published the article’’
OS
老师 表扬了 完成 作业 的 学生 teacher praised complete assignment de student ‘‘The teacher praised the student that completed the assignment’’
OO
公司 录用了 老师 推荐 的 学生 company hired teacher recommend de student ‘‘The company hired the student that the teacher recommended’’
226
J.L. Packard et al.
forehead and the computer screen. For each experimental item, participants were instructed to silently read the experimental sentence and then to push one of two buttons located side-by-side on a keypad to respond yes or no to a following probe sentence, based on the content of the experimental sentence. The probe sentence response task was included to ensure that participants attentively read the experimental sentences. The keypad rested in the participant’s lap held by both hands, with the button used to indicate a ‘‘yes’’ response located on the right side of the keypad pressed by the right thumb for half of the participants and located on the left side of the keypad pressed by the left thumb for the other half. The experimental sentences were displayed using Presentation software in 24-point-font simplified Chinese characters word-by-word in the center of the screen in Rapid Serial Visual Presentation (RSVP) format. Each experimental item began with a 64-point-font fixation dot in the center of the screen lasting 500 ms, followed by successive presentation of each sentence word within a 540 ms time frame. Each frame consisted of the appearance of the word for 210 ms followed by 330 ms of blank screen. The offset of the last sentence word frame was followed by 1000 ms of blank screen, followed by projection of the entire probe sentence. The probe sentence remained on the screen for 7000 ms or until the participant’s yes/no response, whichever occurred first. This was followed by 500 ms of blank screen before the appearance of the fixation point signaling the beginning of the next item. The experiment began with instructions and a short practice session, followed by presentation of 10 blocks of 48 sentences each, with the sentences pseudorandomized within each block by filler and stimulus sentence type, such that a participant never saw more than two filler or two experimental sentences consecutively. Participants were given a short break after each block. The EEG recording portion of the experiment lasted about 1 h, with the entire experiment lasting about 2 h.
11.2.3.1 EEG Recording Continuous EEGs were recorded from 30 electrodes (FP1, FP2, F7, F3, Fz, F4, F8, FT7, FC3, FCz, FC4, FT8, T7, C3, Cz, C4, T8, TP7, CP3, CPz, CP4, TP8, P7, P3, Pz, P4, P8, O1, Oz and O2) attached to an elastic cap (Electro Cap International) worn by the participant. EEGs were then amplified by a SynAmp amplifier with a band pass from 0.05 to 70 Hz digitized at 250 Hz, yielding one voltage measurement every 4 ms. The vertical electrooculogram (VEOG) was monitored from electrodes located above and below the left eye and the horizontal electrooculogram (HEOG) from electrodes located on the cap. The AFz electrode on the cap served as ground. Recordings were referenced to the bilateral mastoids. Electrode impedances were kept below 5 kOhm. Electrode P7 was found to be defective in several of the trials and so was removed from the analysis, along with the corresponding right-side electrode P8.
11
Filler-Gap Processing in Mandarin Relative Clauses
227
11.2.4 Data Analysis 11.2.4.1 Behavioral Data Response latency means (in milliseconds) and accuracy rates (%) were compiled for the 19 participants’ true/false responses to the probe sentences for both the filler and the stimulus sentences.
11.2.4.2 EEG Data The continuous EEG files were first processed by subject to adjust for blink artifacts. This was done by using the VEOG channel to identify all blinks and then constructing an averaged blink artifact value in the remaining electrode channels based on the degree of deviation that occurred in those channels as a response to the VEOG channel reading. The averaged blink artifact for each participant was then subtracted from all channels in spans defined as blinks. The continuous EEG files were then epoched (i.e., averaged) over sentence type to isolate 4000-ms ERP sweeps corresponding to all SS, OS, SO and OO stimulus sentences. Each 4000-ms sentence epoch consisted of (1) the last 300 ms of the blank screen that started the first word frame, (2) the first word of the sentence (210 ms), (3) the remaining five word frames (540 ms each) and (4) 790 ms of the 1,000 ms blank screen that preceded the probe question. This procedure yielded a total of 4525 (1119 SS, 1133 SO, 1138 OS and 1135 OO) of a possible 4560 (19 participants 4 sentence types 60 tokens per sentence type) useable epoched sentence tokens. The epoched sentence tokens were next processed to remove sentences that contained artifacts below –75 or above 75 mV. This procedure resulted in the removal of an additional 257 (44 SS, 45 SO, 61 OS and 107 OO) sentence tokens. The tokens were then baseline-adjusted using a 100 ms prestimulus average baseline voltage measure. This was done by computing the average voltage for each electrode over the first 100 ms of the epoch and subtracting it from each of the remaining voltage measurements for each respective electrode within the remaining 3900 ms of the epoch. An additional 79 tokens (17 OS and 62 OO) were subsequently discarded because the sentences were found to have been displayed either with incorrect characters or with clauses that did not contain a bona fide anaphoric gap. For each of the 19 participants an average ERP trace was then computed for their remaining tokens of each of the four sentence types, yielding 76 average ERP traces. It was the voltage values taken from those 76 traces that were entered into the statistical analysis to test for relative clause ERP effects. P600 component measurements for the particle de and the relative clause head were based on average amplitudes calculated within the four 50 ms intervals within a 200-ms span centered at 600 ms post-stimulus onset. The first interval was the first 50 of the 100 ms preceding the 600-ms midpoint (= 500–550), the second was for the second 50 of the 100 ms preceding the
228
J.L. Packard et al.
midpoint (= 550–600), the third was for the first 50 of the 100 ms following the midpoint (= 600–650) and the fourth was for the second 50 of the 100 ms following the midpoint (= 650–700). For statistical analyses, six regions of interest (ROIs) were used in the ANOVAs, with three electrodes representing each ROI: left anterior (F3, FC3, C3), anterior midline (Fz, FCz, Cz), right anterior (F4, FC4, C4), left posterior (CP3, P3, O1), posterior midline (CPz, Pz, Oz) and right posterior (CP4, P4, O2). Separate ANOVAs were conducted for relative clauses located in matrix subject and matrix object positions and for the particle de and the relative clause head for each of the four 50-ms intervals surrounding the 600-ms midpoint. The separate one-way within-subjects repeated measures ANOVAs were each run with ROI as the independent variable and with the voltage difference between subject- and object-gap relative clauses as the dependent variable.
11.3 Results 11.3.1 Behavioral Data 11.3.1.1 Probe Sentence Response Accuracy Rate In their responses to the probe sentences, participants correctly responded to 92.8% of the probes following filler sentences and 95.2% of the probes following the experimental sentences. The participants were significantly more accurate on the experimental sentences ðX1 ¼ 24:2; p 5 0:01Þ, indicating that they paid sufficiently close attention to the experimental sentences. For the experimental sentences, accuracy was significantly lower for probes following objectgap (94.55%) than subject-gap (95.88%) relative clauses ðX1 ¼ 4:39; p 5 0:05Þ. Accuracy was also significantly lower for probes following sentences with the relative clause located in matrix subject position (94.55%) than for probes with the relative clause located in matrix object position (95.87 %; X1 = 4.34; p < 0.05). Controlling for matrix position, for probes following sentences with the relative clause located in matrix object position there was no significant response accuracy difference for sentences containing object-gap (95.61%) versus subject-gap (96.13%) clauses ðX1 ¼ 0:399; p 5 0:05Þ. Following sentences with the relative clause located in matrix subject position, probe responses to sentences containing object-gap clauses were significantly less accurate (93.49%) than to sentences containing subject-gap (95.62%) clauses ðX1 ¼ 4:96; p 5 0:05Þ. Controlling for gap position, following sentences containing a subject-gap relative clause, there was no significant response accuracy difference for sentences following a relative clause in matrix subject (95.62%) versus matrix object (96.13%) position ðX1 ¼ 0:0375; p 4 0:05Þ. Following sentences containing an object-gap relative clause, probe responses following
11
Filler-Gap Processing in Mandarin Relative Clauses
229
sentences with a relative clause located in matrix subject position were significantly less accurate (93.49%) than when they followed sentences with a relative clause located in matrix object (95.61%) position ðX1 ¼ 4:93; p 5 0:05Þ. 11.3.1.2 Probe Sentence Response Latency The mean response latency to the probes following the experimental sentences was 2456 ms, with significant main effects found for gap site and matrix position and a significant effect found for the gap-by-matrix interaction. In all of the following statistical comparisons of response latency, the number of characters in the probe sentence was used as a covariate to control for probe sentence length, due to the existence of a significant probe sentence length effect, in which subjects responded significantly slower to probe sentences containing more characters ðF 1; 18 ¼ 163:39; p 5 0:001Þ. The significant main effect found for gap site indicated that responses to probe sentences following sentences with object-gap relative clauses were significantly slower (2496 ms) than responses to probes following subject-gap relative clauses (2415 ms; F 1; 18 ¼ 8:00; p 5 0:01). The significant main effect for matrix position found that responses to probes following sentences with the relative clause located in matrix subject position were significantly slower (2527 ms) than responses to probes following sentences with relative clauses located in matrix object position (2385 ms; F 1; 18 ¼ 40:39; p 5 0:01). The significant interaction between gap site and matrix position ðF 1; 18 ¼ 29:56; p 5 0:0001Þ reveals that all of the object-gap slowness effect occurs with clauses located in matrix subject position, i.e., object-gap relatives were significantly slower (2629 ms) than subject-gap relatives (2425 ms) in matrix subject position only (F 1; 18 ¼ 33:75; p 5 0:0001 with the Tukey-Kramer multiple comparison adjustment). Responses to probes following subjectgap relatives in matrix object position were numerically slower than responses to probes following object-gap relatives (2405 vs. 2365 ms respectively), with the difference marginally significant before the Tukey-Kramer adjustment ðF 1; 18 ¼ 3:41; p ¼ 0:065Þ but not after the adjustment (adjusted p > 0.10). Reaction to probes following sentences with relative clauses located in matrix subject position were significantly slower (2626 ms) than when following sentences with the clause located in matrix object position (2365 ms) only when the clauses contained an object gap (F 1,18 = 69.46, adjusted p < 0.0001). There was no significant difference in response latency to probes following clauses located in matrix subject (2425 ms) versus matrix object (2405 ms) position when the clauses contained a subject gap (F 1,18 = 0.42, adjusted p > 0.10).
11.3.2 ERP Data Visual inspection of the overlaid subject- and object-gap relative clause ERP traces revealed a greater positive voltage for the subject-gap relatives in the
230
J.L. Packard et al.
P600 region of the particle de for the relative clauses located in matrix subject position and in the P600 region of the relative clause head for clauses located in matrix object position. The positivity following the particle de for clauses in matrix subject position (see Fig. 11.1) began at around 500 ms and ended at around 600 ms, with the peak positivity located at around 550 ms. The positivity following the head for clauses in matrix object position (see Fig. 11.2) began at around 480 ms and ended at around 560 ms, with the peak positivity located at around 525 ms. To test the significance of the greater positive voltage in the subject-gap relatives, separate one-way within-subjects ANOVAs were conducted for relative clauses located in matrix subject and matrix object positions for each of the four 50-ms intervals centered on the 600-ms point following the particle de and following the relative clause head. The separate ANOVAs were each run with the voltage difference between subject- and object-gap relative clauses as the dependent variable and with ROI as the independent variable. The ANOVA run on the first 50-ms pre-de interval (500–550 ms post-deonset) for relative clauses in matrix subject position found no overall significant
Fig. 11.1 ERP trace of P600 on ‘‘de’’ on subject- and object-gap relative clauses in matrix subject position
11
Filler-Gap Processing in Mandarin Relative Clauses
231
Fig. 11.2 ERP trace of P600 on relative clause head on subject- and object-gap clauses in matrix object position
difference for ROI ðFð5; 318Þ ¼ 0:99; p 4 0:05Þ, indicating that the voltage difference in that time interval between subject- and object-gap relatives did not differ among all six defined ROIs. However, the voltage difference within each of the six ROIs was significantly greater than zero (left-anterior t 318 ¼ 2:74; p 5 0:01, left-posterior t 318 ¼ 2:83; p 5 0:01, midline-anterior t 318 ¼ 3:63; p 5 0:01, midline-posterior t 318 ¼ 3:01; p 5 0:01, right-anterior t 318 ¼ 2:87; p 5 0:01, right-posterior t 318 ¼ 2:74; p 5 0:01), demonstrating a significant late positivity in the 500–550 ms window following de onset and that the positivity was distributed broadly and equally over all six ROIs. The ANOVA on the second pre-de interval (550–600 ms post-de-onset) also found no overall significant difference for ROI ðFð5; 318Þ ¼ 1:70; p 4 0:05Þ and found the voltage difference between subject- and object-gap relatives within three of the six ROIs to be significantly greater than zero (left-anterior t 318 ¼ 2:46; p 5 0:01, left-posterior t 318 ¼ 2:34; p 5 0:01, midline-anterior t 318 ¼ 2:45; p 5 0:01), demonstrating a late positivity on the left and at the anterior midline.
232
J.L. Packard et al.
The ANOVA run on the first post-de interval (600–650 ms post-de-onset) for the relative clause in matrix subject position found an overall significant difference for ROI ðFð5; 318Þ ¼ 2:61; p 5 0:05Þ but no difference between subjectand object-gap relatives within any of the six ROIs (all p > 0.05). Post hoc comparisons showed that the significant main effect for ROI was due to greater positivity for the subject-gap clauses in the left-anterior over the midline-posterior ROI (t 318 = 2.23, Tukey-adjusted p < 0.05), the left-anterior over the right-anterior ROI (t 318 = 2.28, adjusted p < 0.05) and the leftanterior over the right-posterior ROIs (t 318 = 2.90, adjusted p < 0.01). The ANOVA run on the second post-de interval (650–700 ms post-de-onset) for relative clauses in matrix subject position found an overall significant difference for ROI ðFð5; 318Þ ¼ 3:22; p 5 0:01Þ and found that the average difference between subject- and object-gap relatives within five of the six ROIs (all except right-anterior) was significantly greater than zero (leftanterior t 318 ¼ 2:20; p 5 0:01, left-posterior t 318 ¼ 2:69; p 5 0:01, midlineanterior t 318 ¼ 2:80; p 5 0:01, midline-posterior t 318 ¼ 3:08; p 5 0:01, right-posterior t 318 ¼ 2:46; p 5 0:01, demonstrating a late positivity over all ROIs except right-anterior. For all of the pre- and post-head intervals for relative clauses located in matrix subject position, there were no significant main effects for ROI (all p > 0.05) and no significant differences within any of the six ROIs (all p > 0.05). Therefore, the results for relative clauses located in matrix subject position generally indicate a late positivity at the particle de and not at the relative clause head. For the relative clauses located in matrix object position, there were no significant main effects for ROI (all p > 0.05) and no differences between subjectand object-gap relatives within any of the six ROIs in any of the pre- and post-de intervals (all p > 0.05). The ANOVA run on the first pre-head interval found an overall significant effect for ROI ðFð5; 318Þ ¼ 10:57; p 5 0:01Þ and found that the difference between subject- and object-gap relatives within the midline-anterior ROI was significantly greater than zero ðt 318 ¼ 2:34; p 5 0:05Þ and that the difference between subject- and object-gap relatives within the left-anterior and right-anterior ROIs was marginally greater than zero (t 318 = 1.69 and 1.70 respectively, both p < 0.10). For the remaining pre- and post-head intervals for relative clauses located in matrix object position, there were no significant main effects for ROI (all p > 0.05) and no differences between subject- and object-gap relatives within any of the six ROIs (all p > 0.05).
11.4 Discussion The ERP results indicate a significantly greater positivity for the subject-gap over object-gap relative clauses in the P600 region of the relative marker de for clauses located in matrix subject position and in the P600 region of the relative
11
Filler-Gap Processing in Mandarin Relative Clauses
233
clause head for clauses located in matrix object position. These P600 findings may be interpreted as an index of syntactic integration cost (Kaan et al., 2000; Phillips et al., 2005) and suggest that filler-gap integration is more costly for subject-gap than object-gap relative clauses in Mandarin. The findings further suggest that the increased integration cost for subject-gap relatives occurs when integrating the relative marker de for clauses located in matrix subject position and when integrating the relative clause head for clauses in matrix object position. The different locus of the P600 effect in matrix subject versus matrix object relative clauses may be explained as follows. When the relative clause modifies the matrix subject, because the clause occurs in sentence-initial position it is relatively isolated from the effects of matrix sentence processing. In this condition, the relative marker de is able to play the role of ‘‘filler’’ in the gap-filler relation, because de represents a complex NP that can satisfy the thematic requirements of the relative clause verb. This possibility is made all the more likely in Mandarin by the optional but common occurrence of relative clauses with null heads, which allow the head to be interpreted structurally as a full NP but with content that is lexically and semantically unspecified (see Simpson & Wu, 1999). Because relative clauses with null heads are common in Mandarin, we would expect the parser to begin constructing a complex NP when de follows a gapped verb. The marker de, therefore, is posited to play a major role in signaling the appearance of an upcoming complex NP in Mandarin. Because the incoming de represents an NP that fits the thematic requirements of the relative clause verb gap, when de is integrated it satisfies the selectional restriction of the relative clause verb, resulting in the P600 effect observed on de. Note that the selectional restriction satisfied by de in this instance may be considered the form class categorial or ‘‘c-selection’’ selectional restriction and not the word-specific semantic or ‘‘s-selection’’ selectional restriction, as discussed in Grimshaw (1979) and Pesetsky (1993). So when the parser encounters de it knows that there is an incoming NP that can serve as a filler but it does not yet know the lexical identity of that NP. The lexical identity of the NP is revealed with the appearance of the head. In this situation, it is not that the head of the RC is integrated with the gap before the head appears, it is that de is integrated as a substitute for the head, because when the parser reaches de it receives all the information that the head would provide – except for its precise lexical identity. When the relative clause occurs in matrix object position, the parser – following the appearance of the matrix subject and verb – is expecting the matrix sentence object and fulfills that expectation (i.e., satisfies the selectional restriction of the matrix sentence verb) by integrating the marker de. No P600 integration effect is observed on de in that case, however, because integrating the matrix object with the matrix verb does not involve the subject-gap/objectgap processing contrast. We suggest a P600 integration effect occurs instead on the relative clause head because the head satisfies the selectional restrictions of the relative clause verb (both the ‘‘c-selection’’ and ‘‘s-selection’’ selectional
234
J.L. Packard et al.
restriction). In other words, when the relative clause is located in matrix object position, the marker de satisfies the selectional restrictions of the matrix verb and cannot also (i.e., simultaneously) satisfy the selectional restrictions of the relative clause verb. Satisfaction of the relative clause verb selectional restrictions occurs with the appearance of the relative clause head, resulting in the P600 effect observed there. The ERP results indicating greater processing difficulty for subject-gap relative clauses appear to conflict with some of our behavioral data, which seem to suggest that object-gap relatives are more difficult to process (see also Lin & Bever, 2006). Recall that probe responses to sentences with object-gap relatives were less accurate (when the relative clauses were located in matrix subject position) and also that response latency was slower for probe responses following the sentences with object-gap relative clauses (also when the relative clauses were located in matrix subject position). A solution to this apparent conflict emerges when we take a closer look at those behavioral results. First, the behavioral results consist of responses to probe sentences that follow the experimental sentences and so completely lack temporal resolution regarding the locus of processing difficulty within the experimental sentences. The speed and error rate of probe responses tells us little about where processing difficulty actually occurs: all it suggests is that probe sentences are more difficult to answer when the preceding stimulus sentence contains a certain type of clause located in a given sentence position. Second, recall that in sentences with relative clauses located in matrix object position, it was the probe responses to subject-gap relatives that were actually marginally slower than those to objectgap relatives before the Tukey correction. In essence, what the behavioral data tell us is that there is a sizable interaction between the location of the relative clause in the matrix sentence and the position of the relative clause gap, indicating that probes following sentences with relative clauses containing an object gap located in matrix subject position (SO) are more difficult. The fact that this effect occurs only when the relative clauses are located in matrix subject position suggests that the entire SO sentence is more difficult to process for reasons that may be orthogonal to the processing of the relative clause per se. One possible explanation for this finding is that the SO sentences are more difficult to process due to a ‘‘thematic mismatch’’ effect, because in the SO condition the subjects of the matrix and embedded sentence verbs refer to different nouns. In other words, in an SO sentence the single relative clause head NP has dual thematic identities – that NP is both the subject of the matrix verb and the object of the relative clause verb. This recalls the ‘‘parallel function’’ hypothesis of Sheldon (1974), which states that a relative clause is easier to process if the head NP plays the same thematic role in the relative clause and the matrix sentence. Following this hypothesis, the locus of processing difficulty in an SO sentence would be predicted to occur at the matrix verb, because it is at that point in the input stream that the parser is confronted with the conflicting thematic assignments between the relative clause and matrix sentence verbs: the
11
Filler-Gap Processing in Mandarin Relative Clauses
235
Fig. 11.3 N400 on matrix sentence verb in SO sentences
matrix verb is where the second, different, thematic role is assigned to the single NP head. When we compare the matrix sentence verb ERP traces of the SO and SS sentences, we do indeed see an N400 amplitude on the SO matrix verb that is larger than the N400 amplitude on the SS matrix verb (see Fig. 11.3). The mean amplitude of the object-gap clauses over all electrodes in the 100 ms interval that occurs 300–400 ms after the onset of the matrix verb is –2.65 (SE = 0.34), while the mean amplitude of the subject-gap clauses in the same interval is –1.86 (SE = 0.38). The difference between the two is significant ðF 1; 19 ¼ 12:13; p 5 0:01Þ, demonstrating an N400 that is distributed broadly over all six ROIs. Although the N400 effect is usually associated with such phenomena as anomalous words (Kutas & Hillyard, 1980) and cloze probability of content words (Kutas & Hillyard, 1984), other investigators have proposed a thematic source for the N400 effect, resulting from conflicts in the assignment of thematic roles (Frisch & Schlesewsky, 2001, 2005; Mueller, Hirotani, & Friederici, 2007). So in the SO sentences we may be witnessing a thematic assignment conflict, with the matrix verb thematic role assignment conflicting with the thematic role assigned by the
236
J.L. Packard et al.
relative clause verb. In further support of this hypothesis, an examination of the results of a Mandarin relative clause processing study by Hsiao (2003) clearly show that for both groups tested (the Toronto group and the Taiwan group), the participants responded to the matrix verb in the SO condition substantially more slowly than to the matrix verb in the SS condition (Hsiao, 2003, Figs. 11.4 and 12.5, pp. 92–93), although no analysis was conducted to test the statistical significance of that starkly unambiguous numerical result. Now, a thematic assignment mismatch between the matrix and relative clause verbs also occurs in OS sentences, i.e., when a relative clause in object position contains a subject gap but the probes following the OS sentences were not harder to process than probes following the OO sentences, as the ‘‘thematic mismatch’’ hypothesis would predict. A possible reason for this is that it is offset by a temporary ambiguity ‘‘garden-path’’ effect in the OO sentences. In the OO sentences, the noun that occurs just after the matrix verb is the subject of the relative clause but its position causes it to be temporarily misanalyzed as the direct object of the matrix verb, resulting in the temporary ambiguity that yields the garden-path effect. This garden-path effect for the OO sentences increases
Fig. 11.4 Garden-path P600 on relative clause verb in OO sentences
11
Filler-Gap Processing in Mandarin Relative Clauses
237
their difficulty, potentially offsetting the ‘‘thematic mismatch’’ effect, making them approximately equal in difficulty to the OS sentences. If there is a garden-path effect occurring on the relative clause verb in the OO sentences, we should be able to observe it in the ERP data, since gardenpathing is traditionally associated with a P600 effect (Osterhout & Holcomb, 1992, 1993; Osterhout, Holcomb, & Swinney, 1994). When we examine the ERP trace at the relative clause verb of the OO sentences, we do in fact see a larger P600 amplitude suggestive of a garden path effect (see Fig. 11.4). The mean amplitude of the OO clauses over all electrodes in the interval that occurs 540–620 ms after the onset of the relative clause verb is 0.92 (SE = 0.27), while the mean amplitude of the OS clauses in the same interval is 0.58 (SE = 0.25). The difference between the two is marginally significant (F1,19 = 3.77, p ¼ 0:07Þ, suggesting a P600 that is distributed broadly over all six ROIs. It is therefore not unreasonable to suggest that a garden-path effect in the OO sentences increases their difficulty, making them equal in difficulty to the thematically-mismatched OS sentences in the context of probe sentence responses.
11.5 Conclusion The results of this study provide evidence that subject-gap relative clauses are more difficult to process than object-gap relative clauses in Mandarin, because filler-gap integration resulted in a larger P600 amplitude in the subject-gap relative clauses than in the object-gap relative clauses. Our results also suggest that the precise locus of filler-gap integration in a Mandarin relative clause depends on the location of that clause within the matrix sentence: filler-gap integration occurs on the grammatical marker de when the relative clause is the subject of the matrix sentence and it occurs on the relative clause head when the clause is the object of the matrix sentence. In linguistic theory, the notion of verb selectional restrictions is primarily representational in nature. This means that if a verb ‘‘V’’ has selectional restrictions X and Y associated with it in its lexical entry, when ‘‘V’’ is represented in sentence ‘‘S’’ together with arguments AX and AY that match its selectional restrictions, V’s selectional restrictions are said to be satisfied by virtue of the fact that AX and AY are present in the representation of S. In online sentence comprehension, on the other hand, a verb’s selectional restriction requirements can be thought of as being satisfied in real time when arguments that satisfy the requirements are actually parsed in the incoming speech stream. The present research suggests that it is possible to determine precisely at what point in the speech stream a verb’s selectional restrictions are satisfied and that a verb can satisfy its selectional restrictions either ‘‘lexically’’ with overt lexical arguments (‘‘s-selection’’), or ‘‘structurally’’ with a grammatical marker (such as de) that represents a structure-satisfying NP (‘‘cselection’’).
238
J.L. Packard et al.
Acknowledgments The authors would like to thank Xiaoming Jiang and Longjun Yu for help in running subjects, Gary Huang for help in Presentation programming, Hengqing Chu for help in locating subjects and Susan Garnsey, Gary Dell, Kara Federmeier, Kay Bock and Kiel Christiansen for helpful feedback and advice. We especially thank James Yoon for his input on c-selectional and s-selectional restrictions. Our appreciation also goes to the Research Board at the University of Illinois at Urbana-Champaign for funding this research.
References Altmann, G., & Kamide, Y. (1999). Incremental interpretation at verbs: Restricting the domain of subsequent reference. Cognition, 73, 247–264. Boland, J., Tanenhaus, M., Garnsey, S., & Carlson, G. (1995). Verb argument structure in parsing and interpretation: Evidence from wh-questions. Journal of Memory and Language, 34, 774–806. Diessel, H., & Thomasello, M. (2005). A new look at the acquisition of relative clauses. Language, 81.4, 882–906. Felser, H., Clahsen, H., & Mu¨nte T. (2003). Storage and integration in the processing of fillergap dependencies: An ERP study of topicalization and wh-movement in German. Brain and Language, 87, 345–354. Fiebach, C., Schlesewsky, M., & Friederici, A. (2002). Separating syntactic memory costs and syntactic integration costs during parsing: the processing of German WH-questions. Journal of Memory and Language, 47, 250–272. Friederici, A., Hahne, A., & Mecklinger, A. (1996). The temporal structure of syntactic parsing: early vs. late effects elicited by syntactic anomalies. Journal of Experimental Psychology. Learning, Memory, and Cognition, 22, 1219–1248. Friederici, A., Mecklinger, A., Spencer, K. M., Steinhauer, K., & Donchin, E. (2001). Syntactic parsing preferences and their on-line revisions: A spatio-temporal analysis of event-related brain potentials. Cognitive Brain Research, 11, 305–323. Frisch, S., & Schlesewsky, M. (2001). The N400 reflects problems of thematic hierarchizing. Neuroreport, 12, 3391–3394. Frisch, S., & Schlesewsky, M. (2005). The resolution of case conflicts from a neurophysiological perspective. Cognitive Brain Research, 25, 484–498. Frisch, S., Schlesewsky, M., Saddy, D., & Alpermann, A. (2002). The P600 as an indicator of syntactic ambiguity. Cognition, 85, B83–B92. Gibson, E. (1998). Linguistic complexity: locality of syntactic dependencies. Cognition, 69, 1–76. Gibson, E., Desmet, T., Watson, D., Grodner, D., & Ko, K. (2005). Reading relative clauses in English. Cognitive Linguistics, 16–2, 313–353. Grimshaw, J. (1979). Complement selection and the lexicon. Linguistic Inquiry, 10(2), 279–326. Hagoort, P., Brown, C., & Groothusen, J. (1993). The syntactic positive shift (SPS) as an ERP measure of syntactic processing. Language and Cognitive Processes, 8, 439–483. Hsiao, F. (2003). The syntax and processing of relative clauses in Mandarin Chinese. Massachusetts Institute of Technology Ph.D. dissertation. Hsiao, F., & Gibson, E. (2003). Processing relative clauses in Chinese. Cognition, 90, 3–27. Just, M. A., & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in working memory. Psychological Review, 98, 122–149. Kaan E., Harris A., Gibson E., & Holcomb, P. (2000). The P600 as an index of syntactic integration difficulty. Language and Cognitive Processes, 15, 159–201. Kamide, Y., Altmann, G., & Haywood, S. (2003). The time-course of prediction in incremental sentence processing: Evidence from anticipatory eye movements. Journal of Memory and Language, 49(1), 133–156
11
Filler-Gap Processing in Mandarin Relative Clauses
239
Keenan, E. L., & Comrie, B. (1977). Noun phrase accessibility and universal grammar. Linguistic Inquiry, 8, 63–99. King, J., & Just, M. A. (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language, 30, 580–602. King, J., & Kutas, M. (1995). Who did what and when? Using word and clause-related ERPs to monitor working memory usage in reading. Journal of Cognitive Neuroscience, 7, 378–397. Kluender, R., & Kutas, M. (1993). Bridging the gap: Evidence from ERPs on the processing of unbounded dependencies. Journal of Cognitive Neuroscience, 5, 196–214. Kutas, M., & Hillyard, S. (1980). Reading senseless sentences: brain potentials reflect semantic incongruity. Science, 207, 203–205. Kutas, M., & Hilyard, S. A. (1984). Brain potentials during reading reflect word expectancy and semantic association. Nature, 307, 161–163. Li, T. (1983). Handedness questionnaire (Chinese version). Acta Psychologica, 15, 268–276. Li, X, Shu, H., Liu, Y., & Li, P. (2006). Mental representation of verb meaning: Behavioral and electrophysiological evidence. Journal of Cognitive Neuroscience, 18, 1774–1787. Lin, C.-J., & Bever, T. (2006). Chinese is no exception: Universal subject preference of relative clause processing. Paper presented at the 19th Annual CUNY Conference on Human Sentence Processing, New York, NY, March 2006. Lin, Y., & Garnsey, S. (2006). Relative clause comprehension in Mandarin. Poster presented at the 46th Annual Meeting of the Psychonomics Society, Toronto, Canada. Lin, Y., & Garnsey, S. (2007). Plausibility and the resolution of temporary ambiguity in relative clause comprehension in Mandarin. Poster presented at the 20th Annual CUNY Conference on Human Sentence Processing, University of California, San Diego, CA. Lin, Y., & Garnsey, S. (this volume). Animacy and the resolution of temporary ambiguity in relative clause comprehension in Mandarin. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. MacDonald, M., & Christiansen, M. (2002). Reassessing working memory: Comment on Just and Carpenter (1992) and Waters and Caplan (1999). Psychological Review, 109, 35–54. MacWhinney, B. (1982). Basic syntactic processes. In S. Kuczaj (Ed.), Language acquisition, vol. 1: Syntax and semantics. Hillsdale, NJ: Lawrence Erlbaum Associates. MacWhinney, B., & Pleh, C. (1988). The processing of restrictive relative clauses in Hungarian. Cognition, 29, 95–141. Mueller, J., Hirotani, M., & Friederici A. (2007). ERP evidence for different strategies in the processing of case markers in native speakers and non-native learners. BMC Neuroscience, 8, 18. Published online 2007 March 2. doi: 10.1186/1471-2202-8-18. Muller, H., King, J., & Kutas, M. (1997). Event-related potentials elicited by spoken relative clauses. Cognitive Brain Research, 5, 193–203. Neville, H., Nicol, J., Barss, A., Forster, K., & Garrett, M. (1991). Syntactically based sentence processing classes: evidence from event-related brain potentials. Journal of Cognitive Neuroscience, 3, 151–165. O’Grady, W., Lee, M., & Choo, M. (2003). A subject-object asymmetry in the acquisition of relative clauses in Korean as a second language. Studies in Second Language Acquisition, 25, 433–448. Osterhout, L., & Holcomb, P. (1992). Event-related brain potentials elicited by syntactic anomaly. Journal of Memory and Language, 31, 785–806. Osterhout, L., & Holcomb, P. (1993). Event-related potentials and syntactic anomaly: Evidence of anomaly detection during the perception of continuous speech. Language and Cognitive Processes, 8, 413–438. Osterhout, L., Holcomb, P., & Swinney, D. (1994). Brain potentials elicited by garden-path sentences: Evidence of the application of verb information during parsing. Journal of Experimental Psychology: Learning Memory and Cognition, 20(4), 786–803.
240
J.L. Packard et al.
Pesetsky, D. (1993). Topic. . .comment. Natural Language & Linguistic Theory, 11, 557–558. Phillips, C., Kazanina, N., & Abada, S. (2005). ERP effects of the processing of syntactic longdistance dependencies. Cognitive Brain Research, 22, 407–428. Rayner, K. (1998). Eye movements in reading and information processing: 20 years of research. Psychological Bulletin, 124, 372–422. Sheldon, A. (1974). On the role of parallel function in the acquisition of relative clauses in English. Journal of Verbal Learning and Verbal Behavior, 13, 272–281. Simpson, A., & Wu, Z. (1999). The syntax and interpretation of sentence-final DE. Proceedings of NACCL, 10, 257–274. Tanenhaus, M., Carlson, G., & Trueswell, J. (1989). The role of thematic structures in interpretation and parsing. Language and Cognitive Processes, 4, 211–234. Traxler, M., Morris, R., & Seely, R. (2002). Processing subject and object relative clauses: evidence from eye movements. Journal of Memory and Language, 47, 69–90. Traxler, M., Williams, R., Blozis, S., & Morris, R. (2005). Working memory, animacy, and verb class in the processing of relative clauses. Journal of Memory and Language, 53(2), 204–224. Wanner, E., & Maratsos, M. (1978). An ATN approach to comprehension. In M. Halle, J. Bresnan & G. Miller (Eds.), Lingusistic theory and psychological reality. Cambridge, MA: MIT Press. Wu, F., Kaiser, E., & Andersen, E. (this volume). Subject preference, head animacy, and lexical cues: A corpus study of relative clauses in Chinese. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer.
Chapter 12
Animacy and the Resolution of Temporary Ambiguity in Relative Clause Comprehension in Mandarin Yowyu Lin and Susan M. Garnsey
Language comprehension does not always proceed completely smoothly. One source of difficulty is structural ambiguity, which is quite prevalent in language. Even in the absence of ambiguity, though, some structures can be harder to process than others. Relative clause constructions have proven useful in shedding light on how comprehenders deal with both of these sources of difficulty. The two kinds of relative clause construction in English can be illustrated by the following examples (1a) and (1b) (1) a. Object relative clause The reporter who the senator attacked __t__ admitted the error. b. Subject relative clause The reporter who __t__ attacked the senator admitted the error. Relative clauses are one type of wh-construction where relativizers such as ‘‘who’’ or ‘‘which’’ appear in designated positions, leaving a gap or trace (marked as t in the sentences above) in their original position. These two types of relative clauses are distinguished by the role that the head noun (‘‘the reporter’’) plays in the relative clause. In (1b), the reporter serves as the doer of the action ‘‘attacked’’ in the relative clause, while in (1a), in contrast, the reporter is the object of the action ‘‘attacked’’. Based on the head noun’s function within the relative clauses, (1b) is called a subject relative clause while (1a) is called an object relative clause. The main clauses and the subject role of the relative clause head noun in the main clause are identical in the two sentences. Traditionally, relativizers such as ‘‘who’’ have been called ‘‘fillers’’ since they can be mapped to fill gaps that are posited in the canonical position for noun phrases with their function. The relationship between filler and gap can be Y. Lin (*) English Department, College of Foreign Languages and Literature, National Cheng Chi University, Taipei City 11605, Taiwan, China e-mail: [email protected]
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_12, Ó Springer ScienceþBusiness Media B.V. 2011
241
242
Y. Lin and S.M. Garnsey
characterized either in terms of Linear Distance or Structural Distance. The linear distance of a filler-gap dependency is the number of words between the filler and the gap, as illustrated below in (2). (2)
a. Object relative clause The reporter who the senator attacked __t__ admitted the error. filler gap b. Subject relative clause The reporter who __t__ attacked the senator admitted the error. filler gap
Example (2) shows that the distance between filler and gap is shorter in subject relative clauses (2b) than in object relative clauses (2a) in English. The shorter linear distance between filler and gap in subject relative clauses has been cited as one of the reasons why subject relative clauses are easier to process in English, which has been shown in many previous studies (e.g., Just & Carpenter, 1992; King & Just, 1991; Traxler, Morris, & Seely, 2002). Another way to characterize distance between filler and gap is in terms of hierarchical structures (O’Grady, 1997). The more nodes there are between filler and gap (i.e., the deeper the embedding), the longer the structural distance and possibly the more difficult the construction.
(3) a. Object relative clause
b. Subject relative clause
NP
NP
NP The reporter
S’
NP who NP the senator
NP The reporter S
S’
NP who VP
V attacked
S
NP __t__ NP __t__
VP
V attacked the senator
NP
As illustrated in (3), fewer nodes intervene between filler and gap in a subject relative (3b) than in an object relative (3a), which may also contribute to the comparative ease English speakers have with subject relatives. Note that the two kinds of distance are confounded in English relative clauses. Other factors have also been proposed to account for processing asymmetry in English relative clauses. Another account attributes the processing asymmetry to the relative frequency of different word orders (Bever, 1970;
12
Animacy and the Resolution of Temporary Ambiguity
243
MacDonald & Christiansen, 2002; Reali & Christiansen, 2007; Tabor, Juliano, & Tanenhaus, 1997). According to this hypothesis, the word order that is most common in a language will be the easiest for people to process. Consider the English relative clauses below: (4)
a. Subject relative clause The reporter who attacked the senator admitted the error. S V O b. Object relative clause The reporter who the senator attacked admitted the error. O S V
Ignoring the relativizer ‘‘who’’, subject relative clauses have SVO word order, which is canonical in English, while object relative clauses have the less frequent OSV word order. Thus, subject relatives should be easier for English speakers to understand, which they are. Note that all of the factors described so far are a consequence of word order, since that is what determines Linear and Structural Distance. Since canonical word order varies across languages, these kinds of accounts would make different predictions for different languages. In contrast, there are other accounts whose predictions do not differ across languages. One is the noun phrase Accessibility Hierarchy proposed by Keenan and Comrie (1977). It aims to provide a single generalization across human languages. The accessibility to relative clause formation of certain noun phrases is given as follows (5): (5) Subject > Direct Object > Indirect Object > Oblique Object > GEN > Object of Comparison Being higher on the hierarchy means that there are more languages that can relativize on this grammatical function. For example, all languages can relativize on subjects but fewer languages can relativize on both subjects and direct objects. If a language can relativize on indirect objects, it can also relativize on direct objects and subjects. While English allows relativization of all grammatical functions (Fox, 1987), some other languages such as Mandarin can relativize only on some of them. The Accessibility Hierarchy puts subject highest in the hierarchy for all languages, so this account predicts that subject relative clauses should be universally easier than object relative clauses, regardless of other properties shown in different languages. Another explanation (MacWhinney & Pleh, 1988) hypothesizes that people are oriented to the subject of a clause by default because the subject is what the clause is about and shifting perspective to another functional role in the sentence will be harder than maintaining the subject’s perspective. Subject relative clauses modifying the main clause subject will be easiest since they require no shifting of perspective. In object relative clauses, however, people have to shift
244
Y. Lin and S.M. Garnsey
their perspective between the subject of the main clause (‘‘reporter’’) and the subject of the relative clause (‘‘senator’’) and then back to the main clause subject. In contrast, for relative clauses modifying main clause objects, both subject and object relatives require a single shift. For object relatives modifying main clause objects, there must be a single shift from the object of the main clause to the subject of relative clause. For subject relatives modifying main clause objects, a single shift is also required to shift from object of the main clause to subject of the relative clause. To summarize, this account predicts that subject relatives modifying main clause subjects should be easiest, subject and object relatives modifying main clause objects should both be about the same amount harder and object relatives modifying main clause subjects should be hardest. This ranking of difficulty should be consistent across languages, according to the Perspective Shift account. The accounts described so far do not stress the importance of the processing demands of the moment-by-moment integration of incoming words. Other recent work has emphasized the representations that comprehenders construct based on the moment-by-moment integration of different kinds of information (e.g., Altmann & Steedman, 1988; Gibson & Pearlmutter, 1998). Gibson (1998, 2000) proposed the Dependency Locality Theory, which has two major sources of processing cost, called integration cost and memory cost, both of which are related to locality and working memory demands. Integration cost refers to the effort involved in linking a new incoming word into the existing structure. There is also integration cost at the discourse level, where one energy unit is consumed for each new discourse referent. For structural integration, it is cost-free to integrate a VP with an NP when there are no intervening items between them. If there are intervening words, one energy unit will be consumed for each intervening noun and verb until integration can take place. In sum, Dependency Locality Theory provides an account for why object relative clauses are harder in English in terms of the total amount of energy consumed. Each of the accounts described above tries to explain an asymmetry in the difficulty of subject and object relative clauses in English. It is impossible to tease them apart in English because they all predict object relative clauses to be harder. However, some of these accounts make opposite predictions about relative clauses in other languages. One of the controversial issues in current psycholinguistic research is whether object relative clauses are universally more difficult than subject relative clauses. So far, most research in other languages has found object relative clauses to be harder than subject relative clauses, just as in English (Dutch: Mak, Vonk, & Schriefers, 2002; French: Cohen & Mehler, 1996; German: Schriefers, Friederici, & Kuhn, 1995; Japanese: Miyamoto & Nakamura, 2003; Ueno & Garnsey, 2008; Korean: Kwon, Polinsky, & Kluender, 2006). However, research on Mandarin relative clauses has yielded a different pattern of results. Mandarin is different from Indo-European languages in several important ways: it has no case marking, inflection, or agreement and allows pro-drop, meaning that arguments that would be pronominalized in English can sometimes be omitted completely in Mandarin. Another important
12
Animacy and the Resolution of Temporary Ambiguity
245
difference is that the head noun occurs after the relative clause in Mandarin, rather than before as it does in English. Hsiao and Gibson (2003) predicted that Mandarin speakers should find subject relatives harder than object relatives and found reading time evidence supporting that prediction. However, their findings have been disputed and Lin (2006; Lin & Bever, this volume) and Kuo and Vasishth (2007) have both found Mandarin object relative clauses to be more difficult, consistent with findings from other languages. In the following section, the construction of relative clauses in Mandarin will be introduced and discussed in more detail.
12.1 Mandarin Relative Clauses Relative clauses in Mandarin use the word ‘‘DE’’, which functions as a relativizer but also has several other functions. In relative clauses, DE functions like the relative pronoun in English and it is used with both animate and inanimate head nouns. Although English and Mandarin both have default SVO basic word order, in Mandarin relative clauses the head noun occurs at the end of the clause, unlike English, where the head occurs at the beginning of the relative clause. Mandarin relative clauses are thus said to be ‘‘head-final’’. The following examples illustrate the construction of Mandarin relative clauses: (6) a. Mandarin object relative clause 人們
people
[伯爵 批評 _t_ definitely not believe [count criticize _t_ S V 完全
不
相信
的
公主]。
DE princess] O
(People definitely do not believe [the princess who(m) the count criticized].) b. Mandarin subject relative clause 人們
people
完全
不
相信
[_t_ definitely not believe [_t_
批評
伯爵
的
公主] 。
criticize Count DE princess] V O S
(People definitely do not believe [the princess who criticized the count].) DE serves as the relativizer in the relative clause and a trace (marked above as t) is posited at the position where the head noun would be if it were not moved to the end of the relative clause. For Mandarin object relative clauses, the trace position is between the relative clause verb and DE, while for subject relative clauses, the trace is at the beginning of the clause. As illustrated in (6), an important difference between English and Mandarin is that in Mandarin it is object relative clauses that have SVO word order, which is the canonical word order in Mandarin. Subject relative clauses, on the other hand, begin with a verb and have a VOS word order, which is non-canonical and thus less frequent.
246
Y. Lin and S.M. Garnsey
Therefore Mandarin and English differ in which kind of relative clause has default word order. Given Mandarin’s default SVO order, it is not typical for clauses to begin with a verb. However, Mandarin is also a pro-drop language, which means that subject nouns are often dropped, so sentences beginning with a verb are not so uncommon. In fact, one of the most common expressions that Mandarin speakers use to greet each other is: ‘‘Eaten yet?’’ as (7) below: (7)
吃
飽
了
嗎
?
Eat full asp Q (Have you eaten yet?) Since there is no case marking or inflection on the verb in Mandarin, the example in (7) begins with an ordinary verb. Sentences containing ‘‘before’’ or ‘‘after’’ clauses in Mandarin can also begin with verbs. For example, ‘‘eat’’ comes before ‘‘before’’ and starts the clause with a verb in the Mandarin translation of ‘‘Before getting full’’. Thus, when a Mandarin sentence begins with a verb, it does not necessarily imply that there is a relative clause, though that is certainly one of the possibilities. It is not until DE appears that a relative clause becomes close to a certainty.
12.1.1 Predictions About Mandarin Relative Clauses The theoretical accounts described above provided accounts for the asymmetry in the difficulty of subject and object relative clauses in English. Researchers such as Just and Carpenter (1992) and Keenan and Comrie (1977) have argued that object relative clauses are universally more difficult than subject relative clauses, for reasons outlined earlier. So far, several cross-linguistic studies have provided evidence consistent with this claim (Dutch: Mak et al., 2002; French: Cohen & Mehler, 1996; German: Schriefers et al., 1995; Japanese: Miyamoto & Nakamura, 2003; Ueno & Garnsey, 2008; Korean: Kwon et al., 2006) but results so far have been inconsistent for Mandarin relative clauses. In the following sections, the predictions of the different accounts of relative clause processing difficulty for Mandarin relative clauses will be described first and then the findings of recent studies on Mandarin relative clauses will be reviewed. According to the Accessibility Hierarchy account (Keenan & Comrie, 1977), subject relatives should be easier than object relatives in all languages. The Perspective Shift account (MacWhinney & Pleh, 1988) also predicts that Mandarin subject relatives should be easier than object relatives. Mandarin object relative clauses begin with a noun and since first nouns are most likely to be subjects, comprehenders should take its perspective. However, when DE appears, people realize that the first noun is part of an object relative clause and
12
Animacy and the Resolution of Temporary Ambiguity
247
that another upcoming noun will be the main clause subject. Therefore they will have to shift their perspective. Since subject relative clauses begin with a verb, the predictions of the Perspective Shift account are less clear. If people are still oriented to subjects by default even when they appear to be missing, then subject relative clauses in Mandarin require no Perspective Shift since the subject perspective that is assumed at the beginning of the sentence matches with the role that the head noun plays in the relative clause. Since object relative clauses require a shift while subject relative clauses require none, the Perspective Shift account predicts that subject relative clauses should be easier than object relative clauses in Mandarin. The other accounts of relative clause processing asymmetry make different predictions for Mandarin than they do for English. The Word Order Frequency account (MacDonald & Christiansen, 2002) predicts that it is object relatives that should be easier in Mandarin because they have the canonical and thus more frequent SVO word order. Gibson’s Dependency Locality Theory (Gibson, 1998, 2000) also predicts that object relatives should be easier in Mandarin. Object relatives incur no integration cost because there are no words intervening between the trace and the relative marker DE, while subject relatives consume two energy units since there are two intervening words. (8)
a. Mandarin object relative clause councilman
interrogate
t
DE
official
Discourse integration
1
1
0
0
1
Structural integration
0
0
0
0
0
Total
1
1
0
0
1
b. Mandarin subject relative clause t
interrogate
councilman
DE
official
Discourse integration
0
1
1
0
1
Structural integration
0
0
0
2
0
Total
0
1
1
2
1
This analysis follows from Hsiao and Gibson’s (2003) argument that comprehenders would treat (8b) as a subject relative clause from the start, since no subject noun precedes the verb. A trace is therefore posited at the beginning of the subject relative and the distance between that trace and its head noun filler increases the integration cost at the head noun. This analysis could be disputed, since sentences beginning with verbs could turn out to be other kinds of constructions, as described earlier. However, Grodner, Gibson, and Tunstall (2002) argued that Dependency Locality Theory predicts Mandarin subject relative
248
Y. Lin and S.M. Garnsey
clauses should be harder than object relative clauses even without this assumption because Mandarin subject relative clauses require some reanalysis when the initial analysis turns out to be incorrect, which will be more costly. (Storage costs are also predicted to be greater for Mandarin subject relatives but we will not work through those predictions in detail here.) Finally, distance accounts also make predictions about Mandarin relative clause processing asymmetry but their accounts are also based on the assumption that sentences beginning with verbs are treated as subject relative clauses from the start. Example (9) below shows that there is a shorter linear distance between gap and filler in object relative clauses than subject relative clauses in Mandarin, leading to the prediction that object relative clauses should be easier to process than subject relative clauses.
A˘
Recall that for English, Linear and Structural Distance both predicted that subject relatives should be easier. Interestingly, in Mandarin the linear and Structural Distance accounts make opposite predictions. As example (10) below shows, Structural Distance is greater in object relative clauses in Mandarin, which should make them harder. (10) a. Object Relative Clause
b. Subject Relative Clause
NP CP OP
NP
reporter C’
IP
reporter
OP C’ DE
Senator VP attack
CP
IP DE __t__
__t__
VP attack senator
For English, all of the accounts agree in predicting that object relative clauses should be more difficult. It seems likely that each of these accounts captures some aspect of the reasons for differences in processing difficulty between subject and object relatives, so since they all agree, differences should be quite robust in English and they are. However, in a language like Mandarin, where some factors point in one direction and others in the opposite direction, processing asymmetries might be expected to be smaller and that may account
12
Animacy and the Resolution of Temporary Ambiguity
249
for the inconsistent results obtained in studies so far, which are described in the next section. In summary, the Accessibility Hierarchy, Perspective Shift and Structural Distance accounts all predict that object relatives should be harder in Mandarin, while the Word Order Frequency, Dependency Locality Theory and Linear Distance accounts all predict that it is subject relatives that should be harder in Mandarin.
12.1.2 Previous Studies of Mandarin Relative Clause Processing There have been several studies of Mandarin relative clause processing, which have yielded inconsistent results. Hsiao and Gibson’s (2003) study was the first on Mandarin relative clauses and employed a self-paced reading time paradigm. Their stimuli included both singly-embedded and doubly-embedded relative clauses and reading time differences were robust only in the doubly-embedded versions. The doubly-embedded sentences were quite difficult to understand, so it has been argued that the results for those may not reflect normal sentence processing procedures (e.g., Christianson, Hollingworth, Halliwell, & Ferreira, 2001). If comprehenders stop trying to figure things out at some point, results for doubly-embedded relative clauses may not be the best indicator of the relative difficulty of subject and object relative clauses more generally. On the other hand, if the relative clause processing asymmetry is smaller in Mandarin than it is in English, it may be observable only when it is exacerbated by some degree of added difficulty in the sentences. For sentences with singly-embedded relative clauses, the only position at which there was a reliable difference in Hsiao and Gibson’s study was the second word in the sentence. In subject relative clauses, this word was a noun that followed a sentence-initial verb. Longer reading times on this noun might reflect the atypicality of the sentence beginning with a verb, spilling over onto the next word. There is also more ambiguity about how the sentence will continue when the first word is a verb, which could also increase reading times. Thus longer times for subject relatives at the second word may not have been due specifically to relative clause processing. Lin (2006; Lin & Bever, this volume) criticized Hsiao and Gibson’s stimuli because some verbs were used in multiple items and some constructions involved negation in the relative clause region, which could complicate matters. Lin modified the stimuli to remove these objections and also added conditions with relative clauses modifying the main clause object, in addition to the conditions used by Hsiao and Gibson where the relative modified the main clause subject. However, Lin’s stimuli were not without their own problems. Some verbs were still used more than once and a verb in one sentence set did not sound natural. For the sentences with relative clauses modifying the main clause subject, Lin found no reliable differences in reading times at any word. For the sentences with relative clauses modifying the main clause object, there were reliable differences
250
Y. Lin and S.M. Garnsey
only on the final two words in the relative clauses, i.e., DE and the head noun. At these two positions, it was object relative clauses that were read more slowly. Kuo and Vasishth (2007) also used Hsiao and Gibson’s materials and added new materials using either determiners or ‘‘bei’’, a passivization marker. Like Lin, they did not replicate Hsiao and Gibson’s results even for the sentences taken from their study but instead found Mandarin object relatives to be harder. Yang, Johnson, and Gordon (2008) conducted two reading time studies of singly-embedded relative clauses using self-paced reading and eye-tracking and found mixed results. At some sentence positions it was subject relatives that were read more slowly while at others it was object relatives. The direction of the difference on the head noun itself differed depending on whether the relative clause modified the main clause subject (subject relative slower) or main clause object (object relative slower). The results in the eye-tracking study were even less clear. Several other recent studies of Mandarin relative clause processing have either found subject relatives to be harder or found no difference. Hsu and Chen (2007) found subject relatives to be harder but only when sentences were made difficult by increasing the linear distance between gap and filler. When that distance was short, there were no reliable differences. Chen, Ning, Bi and Dunlap (2008) also found subject relatives to be harder but only for readers who tested low on a working memory span measure and thus presumably had more difficulty processing the sentences. Wu and Gibson (2008) found subject relative clauses to be harder when they were embedded in discourse contexts that made relative clause structures especially likely in the target sentences. Interestingly, the size of the effect was much larger in Wu and Gibson study than in any of the other previous studies with singly-embedded relative clauses. They argued that this was because using context to make relative clauses highly likely removed obscuring effects due to ambiguity about whether a sentence had a relative clause or some other structure. Lin and Bever (this volume) also argue that context and other kinds of cues that remove or mitigate temporary ambiguity about whether the sentence contains a relative clause can influence the relative difficulty of subject and object relatives. (See also Hsu & Chen, 2007; Hsu, Hurewitz, & Phillips, 2006; Hsu, Phillips, & Yoshida, 2005; Wu, Haskell, & Andersen, 2006; Yoshida, Aoshima, & Phillips, 2004 for studies investigating how certain sentence-internal cues and/or discourse context can help signal the likely presence of a relative clause in Mandarin or Japanese and how and when such information influences relative clause processing.) A few studies have used event-related brain potentials (ERPs) to investigate relative clause processing in head-final languages, including Japanese (Ueno & Garnsey, 2008), Korean (Kwon et al., 2006), Mandarin (Packard, Ye, & Zhou, this volume) and Basque (Carreiras, Dun˜abeitia, Vergara, Zieza, & Laka, 2007). In these studies, it was hypothesized that whichever kind of relative clause was less expected and/or more difficult to process should lead to larger P600s in the ERP waveforms, since P600 is associated with revision and with integration
12
Animacy and the Resolution of Temporary Ambiguity
251
difficulty (e.g., Kaan, Harris, Gibson, & Holcomb, 2000). Packard et al. (this volume) found larger P600s for subject relatives than for object relatives, though the word on which these effects emerged differed depending on whether the relative modified the matrix clause subject or object. (They argued that the effect emerged at different positions because of differences in when the selectional restrictions of the relative clause verb could be satisfied.) Most importantly, it was subject relatives that consistently elicited larger P600s, leading Packard and colleagues to argue that subject relatives are more difficult in Mandarin. Basque is another language that has head-final relative clauses but it differs from Mandarin in that its default word order is also head-final (SOV). Carreiras et al. (2007) took advantage of ambiguous case and number morphology to construct Basque sentences with relative clauses that were fully ambiguous as to whether they were subject or object relatives until the main clause verb disambiguated them. In two reading time studies, subject relatives were read more slowly than object relatives at the disambiguating words and in an ERP study, subject relatives elicited more P600-like positivity at the disambiguation, suggesting that readers preferred the object relative interpretation up until that point. Japanese and Korean are both like Basque in being generally head-final but without Basque’s morphology. ERP studies of Japanese (Ueno & Garnsey, 2008) and Korean (Kwon et al., 2006) have found evidence for a preference for subject relatives, in contrast with the results for Mandarin and Basque. Another approach to investigating relative clauses in Mandarin has been to examine text corpora to determine which kinds of relatives occur more frequently (Hsiao & Gibson, 2003; Pu, 2007; Wu, Kaiser, & Andersen, this volume), on the assumption that structures that are easier to process should occur more often (Hawkins, 2004). Probably because different text corpora were used in the different studies, there are some inconsistencies in the results. Hsiao and Gibson (2003) found more object relatives than subject relatives in their corpus study but Pu (2007) and Wu et al. (this volume) both found the reverse, i.e., more subject relatives than object relatives. Beyond these major findings about the overall frequency of subject and object relatives, there were also interesting differences depending on the animacy of the head noun and on whether the relative clause modified the matrix subject or matrix object, which we will return to in the introduction of Experiment 2 and in the discussion of our results. There were also some minor differences between Pu’s and Wu et al.’s results, which they attributed to differences in the genre of the texts included in the corpora. For now, however, the most important point is that both Pu and Wu et al. argued that the greater frequency of subject relatives in Mandarin suggests that they should be easier to process than object relatives. The previous experimental studies of Mandarin have used relatively small numbers of stimuli. Given that differences between subject and object relatives seem likely to be smaller in Mandarin than in English because different relevant factors conflict for Mandarin, it seems important to use a larger set of stimuli in order to have a better chance to detect small effects.
252
Y. Lin and S.M. Garnsey
12.1.3 Topicalization and Dropping the Head Noun in Mandarin Relative Clauses In Mandarin, it is possible to topicalize the main clause object by moving it to the beginning of the sentence. An example of this construction in English would be: ‘‘That guy, I really hate.’’ Such topicalizations are relatively rare in English but they are considerably more common in Mandarin (Liu, 2005). Examples of object and subject relatives modifying a main clause object are shown below in (11) and topicalized versions of them are shown in (12). (11)
a. Mandarin object relative clause modifying main clause object 1
記者 已經 開始 詳細 報導 [議員 質詢 的 官員] 。 Reporter already begin detail report [councilman interrogate DE official] (Reporters have already begun to report in detail about [the official who the councilman interrogated].) b. Mandarin subject relative clause modifying main clause object
記者 已經 開始 詳細 報導 [質詢 議員 的 官員] 。 Reporter already begin detail report [interrogate councilman DE official] (Reporters have already begun to report in detail about [the official who interrogated the councilman].) (12)
a. Mandarin object relative clause modifying topicalized main clause object
[議員
質詢 的 官員] 記者 已經 開始 詳細 報導。 [Councilman interrogate DE official] reporter already begin detail report (About [the official who the councilman interrogated], reporters have already begun to report in detail.) b. Mandarin subject relative clause modifying topicalized main clause object
[質詢
議員 的 官員] 記者 已經 開始 詳細 報導 [Interrogate councilman DE official] reporter already begin detail report (About [the official who interrogated the councilman], reporters have already begun to report in detail.)
Notice that in the topicalized examples in (12), because the relative clause is head-final and the main clause is subject-first, the head noun ‘‘official’’ immediately precedes the main clause subject noun ‘‘reporter’’. The same is not true for the English translation, since in English the head noun precedes the relative clause. This is a point we will return to shortly but first one more property of Mandarin relative clauses must be introduced. A final interesting property of Mandarin relative clauses is that the head noun can be omitted, especially when it is recoverable from context (Chu & Chi, 1999). For example, the head noun underlined in example (13) can be omitted without causing any obvious difficulties in comprehension. DE remains in its normal position when the head noun following it is dropped. 1
Although the word ‘‘report’’ in the translation can be used as either a noun or a verb in English, the equivalent word used in the Mandarin sentences can only be a verb. This was true across the materials used in our studies.
12
Animacy and the Resolution of Temporary Ambiguity
253
(13) 叫
[賣 菜 的 人] 明天 再 來。 Ask [sell vegetable DE person] tomorrow again come (Ask the person who sells vegetable to come again tomorrow.)
Now we return to the point raised earlier about the adjacency of the relative clause head noun and the main clause subject noun in topicalized Mandarin sentences (see example 12). When a main clause object is modified by a relative clause and is topicalized, then when the head noun of the relative is dropped, the main clause subject noun immediately follows DE (see example 14 below, where the examples in 12 above are repeated with the head noun omitted.) That leads to temporary ambiguity. On one possible interpretation, ‘‘reporter’’ is the head of the relative, as in ‘‘The reporter who interrogated the councilman. . .’’ On another possible interpretation (i.e., the one that ultimately turns out to be correct in our materials), the relative clause head is unexpressed and ‘‘reporter’’ is the subject of the subsequent main clause, as in ‘‘About the interrogation of the councilman (by someone unspecified), the reporter . . .’’ When readers pursue the first interpretation, they are forced to revise it when the obligatorily transitive Mandarin verb ‘‘report’’ is not followed by an object. Note that it is the use of obligatorily transitive Mandarin verbs like ‘‘report’’2 that makes it impossible to treat the main clause subject as also being the relative clause head noun. Since ‘‘report’’ requires a direct object and the main clause role of the only available noun is subject, the relative clause must have a missing head noun that serves as the main clause object. It was the possibility of creating this temporary ambiguity that was our main reason for using sentences with topicalized main clause objects, so that properties of the main clause subject nouns could be manipulated to determine how the temporary ambiguity is resolved. (14) a. Mandarin object relative clause modifying topicalized main clause object, with head noun omitted 質詢 的 ____] 記者 已經 開始 詳細 報導。 [議員 [Councilman interrogate DE ____] reporter already begin detailed report (About the person who the councilman interrogated, reporters have already begun to report in detail.) b. Mandarin subject relative clause modifying topicalized main clause object, with head noun omitted 質詢 的 ____] 記者 已經 開始 詳細 報導。 [質詢 [Interrogate councilman DE ____] reporter already begin detailed report (About the person who interrogated the councilman, reporters have already begun to report in detail.)
It is important to note that topicalizing a main-clause-object-modifying relative does not change its Linear or Structural Distance properties, nor in fact does it change the direction of the predictions of any of the accounts of 2
Although the verb usage of ‘‘report’’ is only optionally transitive in English, the equivalent verb in Mandarin is obligatorily transitive. This was true across the materials used in our studies.
254
Y. Lin and S.M. Garnsey
relative clause processing difficulty. For example, the tables in (15) below show that Gibson’s Dependency Locality Theory predicts that object relatives should be easier in Mandarin even when they are topicalized. (15) a. Topicalized main-clause-object-modifying object relative clause in Mandarin councilman interrogate t1 DE official reporter begin report t2 Discourse integration
1
1
0
1
1
1
1
Structural integration
0
0
0
0
0
0
3
Memory units
2
1
2
1
2
1
1
Total
3
2
2
2
3
2
5
b. Topicalized main-clause-object-modifying subject relative clause in Mandarin t1 interrogate councilman DE official reporter begin report t2 Discourse integration
1
1
0
1
1
1
1
Structural integration
0
0
2
0
0
0
3
Memory units
4
3
3
1
2
1
1
Total
5
4
5
2
3
2
5
12.2 Experiment 1 The first goal of Experiment 1 was to re-examine asymmetries in processing difficulty for Mandarin relative clauses and the second goal was to investigate ambiguity resolution in relative clauses with dropped heads. The experimental stimuli contained singly-embedded subject and object relative clauses, modifying a topicalized main clause object noun.
12.2.1 Using Topicalization and Head Dropping in the Experiment Our examples include topicalized object-modifying subject and object relative clauses in Mandarin with and without head nouns, as illustrated below in (16): (16) a. Mandarin object relative clause modifying topicalized main clause object with head noun present 質詢 的 官員] 記者 已經 開始 詳細 報導。 [議員 [Councilman interrogate DE official] reporter already begin detail report (About [the official who the councilman interrogated], reporters have already begun to report in detail.) b. Mandarin object relative clause modifying topicalized main clause object with head noun omitted [議員 質詢 的 ____] 記者 已經 開始 詳細 報導。 [Councilman interrogate DE ____] reporter already begin detail report (About [the person who the councilman interrogated], reporters have already begun to report in detail.)
12
Animacy and the Resolution of Temporary Ambiguity
255
c. Mandarin subject relative clause modifying topicalized main clause object with head noun present 議員 的 官員] 記者 已經 開始 詳細 報導。 [質詢 [Interrogate councilman DE official] reporter already begin detail report (About [the official who interrogated the councilman], reporters have already begun to report in detail.) d. Mandarin subject relative clause modifying topicalized main clause object with head noun omitted [質詢 議員 的 ____] 記者 已經 開始 詳細 報導。 [Interrogate councilman DE ____] reporter already begin detail report (About [the person who interrogated the councilman], reporters have already begun to report in detail.)
In speech, sentences with topicalized object phrases such as these would probably have some prosodic marking of the clause boundary following the relative clause, realized on and after the head noun when it is present, or on DE when there is no head noun. (We know of no systematic observations about this, however.) In written Mandarin, though, it is optional whether to put a comma after the topicalized relative clause.
12.2.2 Materials, Design and Procedure Two experiments were conducted using a word-by-word self-paced reading paradigm with each word appearing centrally on the screen. In each experiment, there were 80 sets of sentences with four versions in each set, namely, subject and object relative clauses, each with and without a head noun present, as illustrated in example (16) above. Most words consisted of two characters but a few had three or four characters. In addition, there were 60 fillers, which were created using other functions of DE in Mandarin to try to distract participants from the relative clause manipulations. For example, DE is used with adjectives such as ‘‘beautiful (‘‘piao liang de’’, 漂亮的)’’ and can also be used to mean ‘‘during (‘‘de shi hou’’, 的時候)’’. Four lists were created out of 80 sets of stimuli using a Latin Square design. Each participant saw only one list, which contained 140 sentences, including the 80 experimental stimuli and 60 fillers. The two characters of words such as ‘‘老師 (‘‘lao shi’’, teacher)’’ were presented together. The period in Mandarin ‘‘。’’ was presented separately at the end of each sentence. Subjects controlled their own reading pace by pressing the space bar. The session began with instructions and three practice trials. Yes/no comprehension questions were asked at the end of approximately one quarter of the experimental stimuli. The whole experiment took less than 30 min. Across the 80 sets of stimuli, each relative clause was transitive and thus included both a subject and an object, and nouns filling both roles were present in the (a) and (c) versions of each sentence set but whichever of those nouns was the head of the relative was dropped in the (b) and (d) versions. As described
256
Y. Lin and S.M. Garnsey
above, after topicalizing the object of the main clause, the head noun of the relative clause and the main clause subject noun were adjacent. In Experiment 1, these two nouns were always both animate so that when the head noun was missing, the matrix subject noun could plausibly be interpreted initially as the head noun in conditions with the head noun omitted. In addition, in order for both the subject relatives and object relatives to be felicitous, the other noun involved in the relative clause (e.g., ‘‘councilman’’ in example 16 above) also had to be animate, since the relative clause verbs were chosen to require both their subjects and objects to be animate. Making all three of these nouns animate makes our sentences somewhat atypical for sentences containing relative clauses, since corpus studies have found that it is rare for both nouns in a relative clause to be animate (Mak, et al., 2002; Pu, 2007; Wu et al., this volume). It seems likely that this would contribute to making our sentences generally harder to understand, which is a point we will return to later. Since the materials were constructed so that the same animate head noun could be felicitous in both subject and object relatives, a plausibility norming study was conducted to insure that the head noun was approximately equally plausible in both roles, given the verb and the other noun in the relative. Fortytwo native Mandarin speakers who did not participate in the reading time studies were asked to rate the plausibility of the following types of sentences. (17) The councilman interrogates the official. (18) The official interrogates the councilman. Sentences like (17) and (18) were both rated as highly acceptable, though sentences like (17), in which the head noun plays the subject role, as it does in our subject relatives, were rated as slightly more plausible (5.74 vs 5.46, on a 7-point scale, with 1 meaning unacceptable and 7 meaning very natural). Thus, plausibility slightly favored the subject relatives in our materials, so any effect of this small difference in plausibility should facilitate subject relative clause processing. Care was taken in creating the stimuli to avoid an unlikely but possible problem with the two-noun sequences in the sentences with head noun present. Noun-noun compounds are fairly common in Mandarin, so to prevent readers from treating the two sequentially presented nouns as a compound, rather than as two distinct noun phrases, nouns that could not plausibly combine to form compounds were used, according to the first author’s intuition. In addition, across all of the stimuli, there were compound nouns in only two sentences (one experimental item and one distracter). In both cases, the two components of the compound were presented together. Finally, interpreting the two-word sequences as compounds was unlikely because four-character words are rare, even among compounds. Thus the stimuli and the experimental setting discouraged readers from trying to combine nouns presented sequentially into compounds.
12
Animacy and the Resolution of Temporary Ambiguity
257
The experiment was conducted in Taiwan at National Taiwan Normal University where most of the 48 subjects were recruited. All were native speakers of Mandarin and most were either college students or held college degrees (approximate mean age 21). Although most spoke English as a foreign language, they all used Mandarin as their primary language in daily communication. At the end of the experiment, subjects were paid a small sum for participation.
12.2.3 Results The yes/no comprehension questions were answered correctly 88% of the time across subjects. An error was discovered in one of the test sentences, so that item was omitted from all analyses. No subjects were dropped from the analyses. The experimental items varied in both total number of words and in the position within the sentence of certain critical words. For example, in 44 of the 80 experimental items there was an adverb, quantifier, or auxiliary verb of some sort intervening between the main clause subject noun and the main clause verb (e.g., ‘‘already’’ in example (16) above), while in the other 36 items there was not. Two different approaches to handling this kind of variation were taken in the data analyses. In one approach, the reading time for a word that was both the word immediately following the main clause subject noun and the main clause verb contributed to means calculated for both of those sentence positions. This resulted in equal numbers of trials contributing to means for each position. The logic was that this word conveyed important information both because it was a non-noun immediately following the main clause subject and also because it was a verb. In the other approach, if the word immediately following the main clause subject noun was a verb, it contributed only to the means for the verb position and not to the means for the word immediately following the subject noun, with the result that different numbers of trials contributed to the means at different sentence positions. The results turned out to be nearly identical for both approaches, so only the former analyses are reported here. Figure 12.1 shows the mean reading times at each sentence position for object and subject relative clauses that contained head nouns. We focus first on just these two conditions because one important goal was to determine whether it is subject or object relative clauses that are more difficult to process in Mandarin. Figure 12.1 clearly shows that subject relatives were read more slowly than object relatives at almost every word position. This difference was reliable at the relative clause head noun (‘‘official’’: ðF1 ð1; 47Þ ¼ 11:2; p 5 0:01; F2 ð1; 78Þ ¼ 11:2; p 5 0:01Þ. This word immediately follows DE, which is the first unambiguous cue that the sentence begins with a relative clause, so this is the first position at which a difference between subject and object relative clauses can be attributed specifically to relative clause processing. The next word, the main clause subject noun (‘‘reporter’’), was
258
Y. Lin and S.M. Garnsey
Fig. 12.1 Reading times for sentences with head noun present
also read more slowly in subject relative clauses (F1(1,47) = 13.9, p<0.01; F2 ð1; 78Þ ¼ 16:8; p50:01Þ. It is at this position that it first becomes apparent that the relative clause head noun is not also the main clause subject, since there is another noun to play that role, and therefore the sentence has some kind of topicalization, which probably explains why reading times at this word were the slowest in the sentence. At the first two words in the sentence, it was not yet clear that there was a relative clause but these words were nonetheless read more slowly in subject relatives than in object relatives (word 1: F1(1,47) = 14.0, p<0.05, F2 ð1; 78Þ ¼ 22:0; p50:01; word 2: F1(1,47) = 14.6, p<0.05, F2(1,78) = 19.7, p50:01). This may have been in part because readers interpreted the sentenceinitial verb in subject relatives as signaling the beginning of a subject relative but it could also have been more simply that it is less typical for a sentence to begin with a VN sequence than a NV sequence. Another likely factor is that there is more ambiguity about possible continuations following a VN sequence. Subject relative sentences were read reliably more slowly than object relative sentences by both subjects and items at all of the rest of the words except for the word following the two-noun sequence (‘‘already’’ in Fig. 12.1, which will be labeled ‘‘subject noun plus one’’ or ‘‘SN+1’’ here), which was reliable only by subjects. (DE: F1 ð1; 47Þ ¼ 4:5; p50:05; F2 ð1; 78Þ ¼ 4:8; p50:05; SN+1 (‘‘already’’): F1 ð1; 47Þ ¼ 13:9; p 5 0:01; F2 ð1; 78Þ ¼ 1:1; p 50:05; verb (‘‘begin’’): F1 ð1; 47Þ ¼ 6:3; p 5 0:05; F2 ð1; 78Þ ¼ 3:0; p 5 0:01; verb + 1 (‘‘detail’’): F1 ð1; 47Þ ¼ 7:6; p50:01; F2 ð1; 49Þ ¼ 5:6; p50:05; last word (‘‘report’’):
12
Animacy and the Resolution of Temporary Ambiguity
259
F1 ð1; 47Þ ¼ 18:0; p50:05; F2 ð1; 78Þ ¼ 16:2; p50:01; sentence-final period: F1 ð1; 47Þ ¼ 4:6; p50:05; F2 ð1; 78Þ ¼ 6:2; p50:05). Thus, subject relative clauses were read more slowly than object relative clauses throughout the sentence, including at the first two words before it was clear there was a relative clause. In the stimulus norming study, simple sentences in which the head nouns (e.g., ‘‘official’’) played the subject role (see example 17 above) were rated as slightly more plausible than those in which the same head nouns played the object role (see example 18 above). Thus, it is unlikely that the subject relatives were read more slowly because readers found them generally less plausible. Figure 12.2 shows the conditions with omitted head nouns superimposed on the head-present conditions that were shown alone in Fig. 12.1. The solid lines represent the conditions with head noun present and the dotted lines the conditions with head noun absent. Since the head noun was omitted in the conditions plotted with the dotted lines, there is a break in those lines in Fig. 12.2. The sentence stimuli in the head-present and head-absent conditions were identical across the first three words and an omnibus ANOVA across all conditions at those positions showed no effects of head-presence (word l: F1 ð1; 47Þ ¼ 2:7; p40:1; F2 ð1; 78Þ ¼ 2:1; p40:1; word 2: Fs < 1; DE: Fs < 1) and no interaction between relative clause type and head presence (words l and 2: all Fs < 1; DE: F1 ð1; 47Þ ¼ 2:2; p40:1; F2 ð1; 78Þ ¼ 3:3; p40:05). Just as in the analysis described above that included only the head-present conditions, in the omnibus ANOVA subject relatives were read more slowly
Fig. 12.2 Reading times for all sentence types in Experiment 1
260
Y. Lin and S.M. Garnsey
overall than object relatives at all of the first three words (word 1: F1 ð1; 47Þ ¼ 22:3; p50:01; F2 ð1; 78Þ ¼ 50:8; p40:01; word 2: F1(1,47) = 31.7, p50:01; F2 ð1; 78Þ ¼ 37:7; p50:01; DE: F1(1,48) = 17.1, p < 0.01; F2 (1,78) = 13.5, p < 0.01). Sentences in the head-present and head-absent conditions first became different at the word following DE, which was the relative clause head noun in the head-present conditions (‘‘official’’) and the main clause subject noun in the head-absent conditions (‘‘reporter’’). A direct comparison of the results for these two different words will be postponed until after comparisons of the rest of the words that were identical across conditions. At the main clause subject noun (‘‘reporter’’), subject relatives continued to be read more slowly overall than object relatives ðF1 ð1; 47Þ ¼ 14:1; p50:01; F2 ð1; 78Þ ¼ 17:1; p50:01Þ. In the head-present conditions, this noun was the second in a two-noun sequence, while in the head-absent conditions it immediately followed DE rather than another noun, and this is reflected in the fact that this word was read more slowly overall in head-present sentences than in head-absent sentences ðF1 ð1; 47Þ ¼ 11:8; p50:01; F2 ð1; 78Þ ¼ 44:4; p50:01Þ. In addition, there was an interaction between head-presence and relative clause type at this word, such that differences between subject and object relatives were larger when the head was present than when it was absent (F1(1,47) = 6.6, p<0.05; F2 ð1; 78Þ ¼ 5:2; p50:05Þ. Starting at the next word position and continuing throughout the rest of the sentence, differences between head-present and head-absent sentences reversed, with head-absent sentences read much more slowly than head-present sentences (SN+l (‘‘already’’): F1 ð1; 48Þ ¼ 93:5; p50:01; F2 ð1; 78Þ ¼ 166:4; p50:01; verb (‘‘begin’’):F1 ð1; 47Þ ¼ 57:1; p50:01; F2 ð1; 78Þ ¼ 241:7; p50:01; verb+l (‘‘detail’’): F1 ð1; 47Þ ¼ 86:8; p50:01; F2 ð1; 49Þ ¼ 116:3; p50:01; last word (‘‘report’’): F1 ð1; 47Þ ¼ 54:3; p50:01; F2 ð1; 78Þ ¼ 186:4; p50:01; sentencefinal period: F1 ð1; 47Þ ¼ 75:6; p50:01; F2 ð1; 78Þ ¼ 245:9; p50:01). Across these same later word positions, differences between subject and object relatives disappeared at the word following the main clause subject noun in all conditions and then that overall difference re-emerged and became reliable again only at the last word ðF1 ð1; 47Þ ¼ 17:4; p50:01; F2 ð1; 78Þ ¼ 10:7; p50:01Þ and the sentence-final period ðF1 ð1; 48Þ ¼ 8:6; p50:01; F2 ð1; 78Þ ¼ 11:3; p50:01Þ. (All Fs at both SN+l and verb <1; verb+l: F1(1,47) = 3.7, p<0.05; F2 ð1; 49Þ = 2.5, p < 0.1). Although it can be seen in Fig. 12.2 that the pattern of subject relatives being slower than object relatives re-emerged earlier in head-present conditions than in head-absent conditions, there were no interactions at any position after the main clause subject (all Fs<2), except for the word after the verb in the subject analysis only (verb+1 (‘‘detail’’): F1 ð1; 47Þ ¼ 4:1; p50:05). One additional analysis was conducted to directly compare the words immediately following DE, which were different words in the head-present and headabsent conditions, i.e., the relative head noun in the head-present conditions (‘‘official’’) but the main clause subject noun (‘‘reporter’’) in the head-absent conditions. We hypothesized that the main clause subject would initially be
12
Animacy and the Resolution of Temporary Ambiguity
261
mistaken as the relative clause head noun in the head-absent conditions and the reading times at that word were quite similar to those at the head noun in the head-present conditions (see Fig. 12.2). However, the main clause subject noun in the head-absent conditions was read more slowly overall than the relative clause head noun in the head-present condition (F1(1,47) = 12.3, p<0.01; F2 ð1; 78Þ ¼ 12:5; p50:01Þ. Since different words are being compared, the overall reading time difference could be partly due to differences between the words, such as familiarity and/or the length and complexity of the characters. We do not have frequency of occurrence or familiarity information about the words and characters used in the stimuli but the number of strokes per word was counted as a measure of length and complexity. The relative clause head nouns had on average 18.5 strokes while the main clause subject nouns were slightly more complex with an average count of 20.6 strokes, which may at least partially explain why the latter were read more slowly. This word was also read more slowly overall in subject relatives than object relatives in the subject analysis ðF1 ð1; 47Þ ¼ 5:5; p50:05Þ but that difference did not reach reliability in the item analysis ðF2 ð1; 79Þ ¼ 3:5; p ¼ 0:06Þ. There was no interaction between head-presence and relative clause type in this comparison (Fs<2).
12.3 Discussion The results for the head-present conditions replicated Hsiao and Gibson’s original finding that subject relative clauses are harder to process than object relative clauses in Mandarin and do so for a different kind of sentence with the relative clauses modifying topicalized main clause objects. Furthermore, this difference was observed across most of the words in the sentence, in contrast with the more restricted differences found in most previous studies, probably because the topicalized structures were relatively difficult to process, thus amplifying differences, as Hsiao and Gibson’s doubly embedded sentences did. In the head-absent conditions, reading times slowed down dramatically starting at the word after the main clause subject noun (‘‘already’’). The word in this position was always an adverb or a quantifier or a verb such as ‘‘start’’ or ‘‘decide’’, which was followed later in the sentence by another verb, as in ‘‘decided to buy’’. Thus, the word in this position signaled that there would not be two sequential nouns but it did not unambiguously signal that there was no head noun for the relative clause, since the main clause subject noun could still be interpreted as both the relative head noun and the main clause subject. Reading times may have slowed down so much at this position because readers were considering both possible interpretations in parallel, i.e., that the one noun was both relative head noun and main clause subject, or that the relative was missing its head and the noun was only the main clause subject. For the latter possibility, it seems likely that the unusualness of both topicalization and an
262
Y. Lin and S.M. Garnsey
omitted head noun in isolated sentences without any contextual support also contributed to the slow reading times starting at this position. Of the various proposals to explain relative clause processing, the Word Order Frequency account, the Linear Distance account and Gibson’s Dependency Locality Theory can all account for our results. The Accessibility Hierarchy, Perspective Shift and Structural Distance accounts all predict that object relatives should be harder in Mandarin and thus fail to account for our results. Because Dependency Locality Theory makes detailed predictions about the changing degree of difficulty at each word position in the sentences, it is tempting to derive similar predictions about word-by-word reading times. Such predictions do not fare very well, however. In particular, Dependency Locality Theory predicts higher processing cost at DE in subject relatives than in object relatives (see 15 above) but we found no reliable differences at this word. In addition, Dependency Locality Theory predicts the same processing costs for subject and object relatives across the main clause words but we found subject relatives to be harder than object relatives across those word positions. It is perhaps unfair to derive such predictions, however, since the relationship between processing cost and word-by-word reading times is likely to be complex, especially at high-frequency function words like DE, which tend to be read quickly. An important property of the sentences used in Experiment 1 was that the main clause subject nouns were always animate and thus plausible potential head nouns for the relative clauses, so that when the relative head noun was missing it would be possible to mistake that noun as the head noun. That was changed in Experiment 2.
12.4 Experiment 2 In Experiment 1, we obtained evidence against the idea that object relative clauses are universally harder than subject relative clauses by showing that it is subject relatives that are harder to process in Mandarin sentences where the relative clauses modify topicalized main clause object nouns. The second experiment had two goals. The first was simply to replicate the finding in the first experiment that subject relatives are harder and the second was to investigate the role of a semantic cue, animacy, in resolving temporary ambiguity in Mandarin relative clauses. In example (16b), repeated below in (19a), the animate main clause subject noun ‘‘reporter’’ immediately follows ‘‘DE’’ because the relative head noun has been dropped and it is plausible as a possible head noun for the relative clause, since reporters can both interrogate and be interrogated. The extreme difficulty that readers had in the head-absent conditions in Experiment 1 when it turned out that ‘‘reporter’’ was not the relative clause head suggests that it was initially misinterpreted as the head.
12
Animacy and the Resolution of Temporary Ambiguity
263
(19) a. Mandarin object relative with head noun omitted and animate main clause subject [議員 質詢 的 ____] 記者 已經 開始 詳細 報導。 [Councilman interrogate DE ____] reporter already begin detail report (About [the person who the councilman interrogated], reporters have already begun to report in detail.) b. Mandarin object relative with head noun omitted and inanimate main clause subject [議員 質詢 的 ____] 報紙 已經 開始 詳細 報導。 [Councilman interrogate DE ____] newspaper already begin detail report (About [the person who the councilman interrogated], newspapers have already begun to report in detail.)
Compare (19a) to (19b), where the inanimate noun ‘‘newspaper’’ is not plausible as the relative clause head because an inanimate newspaper can neither interrogate nor be interrogated. In Experiment 2, we investigated whether making the main clause subject noun inanimate would prevent mistaking it as the relative clause head noun even though it immediately follows ‘‘DE’’ in the head-omitted conditions and thus alleviate the difficulty participants experienced in those conditions in Experiment 1. In both English and Mandarin, word order is a strong cue about the sentential roles of noun phrases but English also has subject-verb number agreement cues that Mandarin does not. There is evidence that Mandarin speakers rely instead on the animacy of nouns to help determine their sentential roles. In offline sentence interpretation studies testing the Competition Model developed by Bates and MacWhinney (1979) to explain cross-linguistic processing differences, Mandarin speakers have been found to treat animacy as equally important as word order, in contrast with English speakers, who tend not to rely on animacy even when word order is ambiguous and agreement cues are absent (Lin, 2005; Su, 2001). Thus, we might expect Mandarin speakers to make rapid use of the animacy of the main clause subject noun in our sentences to avoid interpreting it as the head noun when the head is omitted and so to have substantially less difficulty than in Experiment 1 with the head-absent conditions. There has been considerable controversy about just how quickly semantic and pragmatic information come into play during sentence comprehension (e.g., Clifton et al., 2003; Ferreira & Clifton, 1986; Trueswell, Tanenhaus, & Garnsey, 1994). Rather than reviewing all of the evidence in this area, we focus here on previous studies specifically investigating the impact of semantic and pragmatic factors on relative clause comprehension in different languages. In both German and Dutch, the roles of nouns in sentences are indicated by case inflections in addition to word order and agreement. Case marking is sometimes ambiguous but word order and subject-verb agreement can often compensate. However, verbs are clause-final in relative clauses in both languages, so word order cues are disrupted and agreement cues become useful only at the verb at the end of the relative clause. Schriefers et al. (1995) took advantage of these properties to investigate the role of plausibility in relative clause comprehension in German. Across three studies, they found consistently
264
Y. Lin and S.M. Garnsey
longer reading times on the disambiguating relative-clause-final auxiliary verb when it disambiguated toward an object relative, even when preceding pragmatic bias favored the object relative interpretation. The authors concluded that plausibility does not overrule a general preference for subject relatives in German. However, Mak, Vonk, and Schriefers (2002) argued that the absence of plausibility effects in Schriefers et al.’s studies was due to the fact that all of the relative clause head nouns were animate. In corpus studies of Dutch and German newspaper text, Mak et al. found object relatives to be common only when the head noun was inanimate. Animate nouns are more likely to play subject/agent roles and thus to be more felicitous in subject relatives, while inanimate nouns are more likely to play object/patient roles and thus be more felicitous in object relatives. In two reading time studies in Dutch, Mak et al. found object relatives to be as easy to understand as subject relatives when the object relatives had inanimate heads. In a similar series of studies, Traxler, Morris and Seely (2002) also found English object relatives to be as easy to read as subject relatives when their heads were inanimate. As in Schriefers et al.’s German studies, a plausibility manipulation in English did not eliminate the difficulty of object relatives when head nouns were animate. Interestingly, Mandarin corpus studies conducted by Wu et al. (this volume) and Pu (2007) found similar effects of the animacy of the head noun, though they differed from each other in some of the details. In both studies, subject relatives were found to be more common when the head noun was animate and object relatives were more common when the head noun was inanimate, just as in Dutch (Mak et al., 2002). In addition, both studies found relative clauses modifying the matrix subject to be considerably more common than those modifying the matrix object and Wu et al. also found the preference for animate heads of subject relatives and for inanimate heads of object relatives to be especially strong when the relative modified the matrix subject. Thus, not surprisingly, it seems that the semantic and pragmatic factors that determine which kinds of relative clauses are likely to be used in Dutch also apply in Mandarin. In the experimental studies of German, Dutch and English described above, the effects of animacy were observed at or after the end of the relative clause. Given that English relative clauses are head-initial and typically have a relative pronoun immediately following the head, the animacy of the head noun could start to influence expectations about the kind of relative clause to follow quite early. In an event-related brain potential (ERP) study of English, Weckerly and Kutas (1999) found effects of animacy beginning at the first noun, suggesting that good readers expected sentences to begin with animate nouns, but that when they began with inanimate nouns, object relative clauses were a better continuation than subject relatives. In Experiment 1, because the relative clauses were constructed to require animate heads, animate main clause subject nouns were used so that it would be possible to mistake them for the relative clause head nouns when those were absent. One consequence of making both nouns animate is that it made them
12
Animacy and the Resolution of Temporary Ambiguity
265
similar to one another. The observation that similarity across the nouns in sentences with multiple embeddings can increase difficulty can be dated back to Miller and Chomsky’s (1963) observations about multiply center-embedded relative clauses, illustrated in (20) below. (20) The salmon that the man that the dog chased smoked fell. Lewis and colleagues (1996; Lewis, Vasishth, & Van Dyke, 2006; Van Dyke & Lewis, 2003) promoted the notion of ‘‘similarity-based interference’’ to explain why sentences containing multiple words having similar properties can be more difficult to process than sentences whose words are less similar to one another. They argued that when multiple elements must be held in working memory during sentence processing in order to integrate them with later elements, the more similar those elements are, the more potentially confusable they are, making it more difficult to be sure to retrieve and integrate the right ones at the right times. Evidence supporting this general notion has been found in a number of recent studies (Fedorenko, Gibson, & Rohde, 2006; Gordon, Hendrick, & Johnson, 2001, 2004; Gordon, Hendrick, Johnson, & Lee, 2006; Gordon, Hendrick, & Levine, 2002; Van Dyke & Lewis, 2003; Van Dyke & McElree, 2006; Warren & Gibson, 2002). For example, Gordon and colleagues (Gordon et al., 2001) used different noun phrase types, such as proper names and job descriptors (e.g., ‘‘Ethan’’ vs ‘‘the fireman’’), to examine processing difficulty for English relative clauses. In addition to confirming that object relative clauses are the harder ones in English, they also found that reading times were slower when the two nouns in the sentence were of the same type. Warren and Gibson (2002) found that some doubly-embedded relative clauses only become acceptable when some of the nouns are pronouns and others full nouns, as illustrated in (21) below. (21) The reporter who everyone that I met trusts said the president won’t resign yet. (Bever, 1974) Given the results of previous studies showing effects of animacy in Dutch and English relative clause processing (Mak et al., 2002; Traxler et al., 2002; Weckerly & Kutas, 1999), studies showing effects of animacy in other kinds of Mandarin sentences (Lin, 2005; Su, 2001), and studies showing similarity-based interference effects in English (Fedorenko et al., 2006; Gordon et al., 2001, 2002, 2004, 2006; Van Dyke & Lewis, 2003; Van Dyke & McElree, 2006; Warren & Gibson, 2002), in Experiment 2 we used inanimate nouns as the main clause subject nouns in our sentences. For the sentences with head noun present, the goals were to determine whether making the two nouns less similar (i.e., animate relative clause head noun and inanimate main clause subject noun) would make the sentences easier to understand, as well as to replicate
266
Y. Lin and S.M. Garnsey
the overall finding that subject relatives were more difficult in Experiment 1. For the sentences with head noun absent, the goal was to determine whether the implausibility of the inanimate main clause subject as the relative clause head noun would prevent people from pursuing that interpretation and thus make the sentences without head nouns easier than they were in Experiment 1.
12.5 Materials, Design and Procedure The design and procedures for the second experiment were identical to the first. Forty participants from the same population of college students at National Taiwan Normal University as in Experiment 1 (approximate mean age 21) were paid a small sum for their participation. The only difference from Experiment 1 was that the main clause subject nouns were inanimate, as illustrated in (22) below. (This required a few other small changes to the main clause words in a few items to make the sentences completely felicitous.) In the stimulus norming study described above for Experiment 1, sentences like those in (22) below were included and they were rated as much less acceptable (1.65) than the versions with two animate nouns that were used in Experiment 1 (5.74 and 5.46). Thus the inanimate nouns used in Experiment 2 were indeed very poor potential relative clause heads in the head-absent conditions. (22) The councilman interrogates the newspaper. It is important to note that the sentences still contained two animate nouns in the relative clause itself, which is atypical for relative clauses, according to corpus studies (Mak et al., 2002; Pu, 2007; Wu et al., this volume) and which seems likely to make the sentences somewhat harder to understand. We will return to this point in the discussion. There were again 80 sets of test sentences and 60 sets of fillers. Care was again taken to avoid possible compound interpretations of the two-noun sequence of the relative clause head noun followed by the main clause subject noun. The fillers were the same as in the first experiment. The paradigm was again selfpaced, word-by-word reading. (23) a. Mandarin object relative clause modifying topicalized main clause object with head noun present 質詢 的 官員] 報紙 已經 開始 詳細 報導。 [議員 [Councilman interrogate DE official] newspaper already begin detail report (About [the official who the councilman interrogated], newspapers have already begun to report in detail.) b. Mandarin object relative clause modifying topicalized main clause object with head noun omitted
12
Animacy and the Resolution of Temporary Ambiguity
267
[議員 質詢 的 ____] 報紙 已經 開始 詳細 報導。 [Councilman interrogate DE ____] newspaper already begin detail report (About [the person who the councilman interrogated], newspapers have already begun to report in detail.) c. Mandarin subject relative clause modifying topicalized main clause object with head noun present [質詢 議員 的 官員] 報紙 已經 開始 詳細 報導。 [Councilman interrogate DE official] newspaper already begin detail report (About [the official who interrogated the councilman], newspapers have already begun to report in detail.) d. Mandarin subject relative clause modifying topicalized main clause object with head noun omitted 議員 的 ____] 報紙 已經 開始 詳細 報導。 [質詢 [Interrogate councilman DE ____] newspaper already begin detail report (About [the person who interrogated the councilman], newspapers have already begun to report in detail.)
12.6 Results The reading times in Experiment 2 are shown in Fig. 12.3 below, with solid lines representing conditions with head noun present and dashed lines representing conditions with head noun absent. A comparison of Fig. 12.3 with Fig. 12.2 reveals three important differences. First, it appears that subject relatives were
Fig. 12.3 Reading times for all sentence types in Experiment 2
268
Y. Lin and S.M. Garnsey
again read more slowly than object relatives but at fewer word positions. Second, readers appeared to have less difficulty in the head-absent conditions in Experiment 2 than they did in Experiment 1. And third, there were no differences between subject and object relatives at the end of the sentence in Experiment 2, while there were in Experiment 1. Before further comparing the results across the two studies, however, we first present the analyses for Experiment 2 alone. An omnibus ANOVA revealed that subject relative clauses were reliably read more slowly overall than object relative clauses at just the first word ðF1 ð1; 39Þ ¼ 16:6; p50:01; F2 ð1; 79Þ ¼ 89:7; p50:01Þ and the main clause subject noun (‘‘newspaper’’: F1(1,39) = 8.9, p<0.01; F2(1,79) = 12.8, p50:01; all Fs < 2 at other word positions). Since it was not yet clear at the first word that there was a relative clause, it is only the results at the main clause subject noun that unambiguously show subject relatives to be harder than object relatives. As in Experiment 1, the difference at the first word could be due to an expectation that the sentence will continue with a subject relative but could also reflect the fact that it is somewhat atypical for Mandarin sentences to begin with a verb and/or the fact that there is more ambiguity about possible continuations for a sentence that begins with a verb. There were no main effects of head noun presence until the main clause subject noun, which was where the sentences in the head-present and headabsent conditions first became different (all Fs < 2 for words preceding the main clause subject), except for a difference at the second word that was reliable by subjects and marginal by items (F1(1,39) = 4.5, p<0.05; F2 ð1; 79Þ ¼ 2:5; p50:1Þ. Since the head-present and head-absent sentences were still identical at the second word, this effect is inexplicable. The main clause subject noun itself was read more slowly in the head-present than the head-absent conditions, reliably by items and marginally by subjects ðF1 ð1; 39Þ ¼ 3:8; p50:1; F2 ð1; 79Þ ¼ 13:1; p50:01Þ. As in Experiment 1, the difference between head-present and head-absent conditions reversed at the word following the main clause subject noun, with the head-absent conditions read more slowly throughout the rest of the sentence (all Fs > 10, all p<0.01). As in Experiment 1, there was an interaction between relative clause type and head noun presence such that the difference between subject and object relatives was larger in the head-present conditions than the head-absent conditions at the main clause subject noun, reliable by items and marginal by subjects ðF1 ð1; 39Þ ¼ 2:7; p50:1; F2 ð1; 79Þ ¼ 4:8; p50:05Þ. Unlike Experiment 1, there was a similar interaction at the first word in the sentence ðF1 ð1; 39Þ ¼ 11:2; p50:01; F2 ð1; 79Þ ¼ 8:6; p50:01Þ, which is inexplicable, since the head-present and head-absent conditions were identical at that point. The reading times at the words immediately following DE, which differed in head-present and head-absent sentences, were directly compared and unlike Experiment 1, there were no reliable differences in this comparison (all Fs < 2.5, all p>0.1), even though the main clause subject noun
12
Animacy and the Resolution of Temporary Ambiguity
269
(‘‘newspaper’’) was again a bit more complex (22.8 strokes on average) than the relative clause head noun (‘‘official’’: 18.5 strokes on average).
12.7 Discussion The second study confirmed the finding in Experiment 1 that subject relative clauses were more difficult overall than object relative clauses in Mandarin but the differences were more restricted, presumably because the sentences were easier because they contained inanimate main clause subject nouns that 1) could not be mistaken for the head nouns when those were missing and 2) were less confusable with the relative clause head nouns when those were present. Before further discussion of the results, an analysis combining the two studies will be reported.
12.8 Results Across Both Studies To directly evaluate the effect of the animacy of the main clause subject noun, an analysis was conducted on the combined results of the two studies, starting from the relative clause head noun.3 Since the sentences did not yet differ across studies at the relative clause head noun itself, the only effect was that head nouns were read more slowly overall in subject relatives than in object relatives (640 vs 566 msec; F1 ð1; 86Þ ¼ 11:3; p50:01; F2 ð1; 157Þ ¼ 10:3; P50:01Þ, as was found in each study separately. This effect of relative clause type continued at the main clause subject noun itself, which was read more slowly following subject relatives than following object relatives (755 vs 658 ms; F1 ð1; 86Þ ¼ 20:8; p50:01; F2 ð1; 157Þ ¼ 24:3; p50:01). The overall difference between subject and object relatives then disappeared at the next two word positions but re-emerged on the last word of the sentence (757 vs 689 ms; F1 ð1; 86Þ ¼ 12:7; p50:01; F2 ð1; 157Þ ¼ 9:3; p50:01) and the sentence-final period (885 vs 803 ms; F1 ð1; 86Þ ¼ 7:8; p50:01; F2 ð1; 157Þ ¼ 11:0; p50:01). In addition, the main effect of head presence that was observed at the main clause subject noun in each study separately also carried over into the analysis combining studies. This word was read more slowly when it was the second noun in a two-noun sequence (because it immediately followed the head noun in the head-present conditions) than when it immediately followed DE 3
Animacy was treated as a between-subjects factor and the item-based analysis across both studies was done in two different ways, with animacy treated as a between-items factor in one and as a within-items factor in the other. It was not obvious which was the more appropriate approach because the items were very similar but not identical across studies except for the animacy manipulation, because a few items had to be modified in other ways to remain felicitous in Experiment 2. The patterns of reliability were identical in both item-based analyses, so only the between-items version is reported here.
270
Y. Lin and S.M. Garnsey
(because the head noun was absent) (773 vs 639 ms; F1(1,86) = 14.3; p50:01; F2 ð1; 157Þ ¼ 55:2; p50:01). At all subsequent positions, however, once it became clear that the main clause subject noun was not the relative clause head noun in the head-absent conditions, the direction of this effect reversed, with head-absent conditions read more slowly than head-present ones throughout the rest of the sentence (all Fs > 90, all p<0.01). Finally, the interaction between relative clause type and head presence that was observed at the main clause subject noun (and not at any subsequent positions) separately in each study also carried over into the combined analysis ðF1 ð1; 86Þ ¼ 8:7; p50:01; F2 ð1; 157Þ ¼ 9:8; p50:01Þ, reflecting the fact that the difference between subject and object relatives was larger when the head noun was present than when it was absent (129 ms difference vs 63 ms difference) Most important in this analysis were a number of effects of animacy starting at the main clause subject noun, which was the noun whose animacy was manipulated. Animate main clause subject nouns themselves were read more slowly overall than inanimates (755 vs 657 ms), producing a reliable main effect of animacy in the item-based analysis only ðF1 52; F2 ð1; 157Þ ¼ 19:7; p40:01Þ. There was also an interaction between animacy and head-noun-presence starting at the main clause subject noun and persisting throughout the rest of the sentence, which arose because the difficulty due to an absent relative clause head noun was substantially mitigated when the main clause subject noun was inanimate and thus could not be mistaken as the head. At the main clause subject noun itself, this effect was reliable only by items ðF1 52; F2 ð1; 157Þ ¼ 7:5; p50:01Þ but throughout the rest of the sentence it was reliable in both analyses (all Fs>20, all p<0.01). Interactions between animacy and relative clause type emerged starting at the next-to-last word in the sentence, where they were marginal by both subjects and items ðF1 ð1; 86Þ ¼ 3:24; p50:1; F2 ð1; 99Þ ¼ 3:0; p50:1Þ and then became reliable at the last word and the period (all Fs > 9, all p<0.01, except for the subject-based analysis at the sentence-final period, where F1 ð1; 86Þ ¼ 5:8; p50:05. Finally, the three-way interaction among animacy, relative clause type and head-presence did not reach reliability at any position, though it was marginal for the last three words in the sentence (all Fs > 2.5, all p<0.1), reflecting the fact that differences between subject and object relatives re-emerged at these positions but only when the main clause subject noun was animate and the head noun was present. Simple effects tests supported this interpretation of the marginal three-way interaction, since the two-way interaction of animacy and relative clause type was reliable at these three words only when the head noun was present (all Fs > 7 and p<0.01 for the last two words, and F1 ð1; 86Þ ¼ 5:0; p50:05; F2 ð1; 157Þ ¼ 2:9; 50:1 for the third-to-last word). There were no similar two-way interactions between animacy and relative clause type at any of these word positions when the head noun was absent (all Fs < 2, all p>0.1).
12
Animacy and the Resolution of Temporary Ambiguity
271
12.9 General Discussion In both experiments, object relatives were read more quickly than subject relatives at multiple word positions. This was especially true when the relative clause head noun was present but was also found to a lesser extent when the head noun was absent. The finding that object relatives were easier than subject relatives is consistent with the predictions of the Word Order Frequency, Linear Distance and Dependency Locality Theory accounts for asymmetry in relative clause processing difficulty and inconsistent with the predictions of the Accessibility Hierarchy, Perspective Shift and Structural Distance accounts. It appears that Mandarin object relative clauses are easier to process than subject relatives, probably because their canonical SVO word order 1) is more frequent, 2) makes the linear distance between filler and gap shorter and 3) thereby decreases the storage and integration costs incurred during processing. Differences in our studies were more robust than in previous studies of Mandarin relative clauses, except for Wu and Gibson’s (2008) study, which also found large differences. We believe the size of the effects in our studies was probably due to using relatively difficult sentences with relative clauses modifying topicalized main clause objects, while Wu and Gibson (2008) argued that their differences were large because all temporary ambiguity was removed by using contexts that made relative clauses virtually certain. In addition to contributing to the growing body of evidence about relative clause processing in different languages, our studies also examined the role of animacy in disambiguating sentences whose relative clause heads were omitted, as is allowed in Mandarin. Using relative clauses that modified topicalized main clause objects and including conditions where the relative clause head nouns were omitted allowed us to create temporary ambiguity about whether or not the noun immediately following the relativizer DE was the head noun of the relative clause. Topicalized main clause objects were used so that when the relative clause head noun was omitted, the main clause subject noun was the word immediately following DE and thus could be mistaken as the relative head noun. The relative clauses were constructed to require animate head nouns, so that when the noun immediately following DE in the head-absent conditions was inanimate (as in Experiment 2), readers might be able to use that cue to avoid being garden-pathed into thinking that the main clause subject noun was the relative head noun. Readers had much less difficulty in the head-absent conditions in Experiment 2 than in Experiment 1, showing that they were indeed able to make rapid use of the animacy cue to rule out the possibility that the main clause subject noun was the relative head noun in the head-absent conditions. Although using topicalized relative clauses and omitting their heads allowed us to investigate the usefulness of animacy cues in Mandarin sentence comprehension, using such sentences without any discourse context naturally raises concerns about the generalizability of the results to relative clause processing in
272
Y. Lin and S.M. Garnsey
other more typical kinds of sentences. Topicalization and head-dropping are probably felicitous only in discourse contexts that promote them. However, in a series of more recent studies using non-topicalized sentences containing relative clauses whose heads were always present, we have consistently found subject relatives to be read more slowly than object relatives (Lin & Garnsey, 2009). Thus, we do not believe that the general pattern of results reported here is limited to the particular kinds of sentences we used. However, in both the studies reported here and in our more recent studies, the two nouns in the relative clause were usually both animate, which is atypical for relative clauses according to corpus studies (Mak et al., 2002; Pu, 2007; Wu et al., this volume), so it remains to be seen whether our results will generalize to sentences with relative clauses involving at least one inanimate noun. A puzzle remains about how to resolve the apparent discrepancy between results like ours showing that Mandarin subject relatives are harder to understand with the higher frequency of occurrence of subject relatives observed in analyses of Mandarin corpora (Wu et al., this volume; Pu, 2007). Given that communication goals and most components of the language production process should be highly similar across languages, it is not surprising that similar frequencies of occurrence for subject and object relatives are observed across languages. However, it is rather surprising that Mandarin comprehenders apparently end up having to deal more often with a structure that is harder for them to understand. As described earlier in our introduction, in the case of Mandarin the several factors that have been invoked to explain differences in the processing difficulty of subject and object relative clauses make different predictions. Perhaps the relative weights of these various factors differ for production and comprehension, leading to the discrepancy. Additional evidence that there are conflicting pressures in Mandarin comes from the relatively small effect sizes and mixed pattern of results across studies using Mandarin sentences that are not made more complex by double embeddings or topicalization. Perhaps which kind of relative clause is more difficult to understand is simply not as consistent in Mandarin as it is in English, where the various relevant factors all point in the same direction. Thus, there may be more room in Mandarin for various other factors to influence relative clause processing, resulting in a mixed pattern of results across studies. Finally, our results provide new evidence supporting the role of similaritybased interference during sentence processing. In Experiment 1, where the relative clause head noun and the main clause subject noun were similar because both were animate, the difficulty of subject relatives persisted throughout the sentence but in Experiment 2 where the nouns were less similar because one was animate and the other inanimate, the difficulty of subject relatives did not persist. Because the sentences were not fully disambiguated until an obligatorily transitive verb appeared at the end of the sentence, readers could not fully integrate earlier elements of the sentence until then and thus suffered additional processing cost when the elements they were forced to continue holding onto were more similar. This is the first instance we know of where a semantic feature
12
Animacy and the Resolution of Temporary Ambiguity
273
such as animacy, rather than the pronoun/noun/name status of nouns or syntactic functions of nouns, has been observed to contribute to similaritybased interference.
References Altmann, G., & Steedman, M. (1988). Interaction with context during human sentence processing. Cognition, 38, 419–439. Bates, E., & MacWhinney, B. (1979). A functionalist approach to the acquisition of grammar. In E. Ochs & B. Schieffelin (Eds.), Developmental pragmatics (pp. 167–209). New York: Academic Press. Bever, T. G. (1970). Cognitive basis for linguistic structures. In J. R. Hayes (Ed.), Cognition and the development of language (pp. 279–362). New York: Wiley. Bever, T. G. (1974). The ascent of the specious, or there’s a lot we don’t know about mirrors. In D. Cohen (Ed.), Explaining linguistic phenomena (pp. 173–200). Washington, D.C.: Hemisphere. Carreiras, M., Dun˜abeitia, J. A., Vergara, M., Zieza, I., & Laka, I. (2007). Object relative clause preference in Basque: Subject relative clauses are not universally easier to process. Presented at the Architectures and Mechanisms of Language Conference, Turku, Finland. Chen, B., Ning, A., Bi, H., & Dunlap, S. (2008). Chinese subject-relative clauses are more difficult to process than the object-relative clauses. Acta Psychologica, 129, 61–65. Christianson, K., Hollingworth, A., Halliwell, J. F., & Ferreira, F. (2001). Thematic roles assigned along the garden path linger. Cognitive Psychology, 42, 368–407. Chu, C., & Chi, T. J. (1999). A cognitive-functional grammar of Mandarin Chinese. Taipei: The Crane Publishing Co. Clifton, C., Traxler, M. J., Mohamed, M. T., Williams, R. S., Morris, R. K., & Rayner, K. (2003). The use of thematic role information in parsing: Syntactic processing autonomy revisited. Journal of Memory & Language, 49, 317–334. Cohen, L., & Mehler, J. (1996) Click monitoring revisted: An on-line study of sentence comprehension. Memory & Cognition, 24, 94–102. Fedorenko, E., Gibson, E., & Rohde, D. (2006). The nature of working memory capacity in sentence comprehension: Evidence against domain-specific working memory resources. Journal of Memory & Language, 54, 541–553. Ferreira, F., & Clifton, C. Jr. (1986). The independence of syntactic processing. Journal of Memory & Language, 25, 348–368 Fox, B. (1987). The noun phrase accessibility hierarchy reinterpreted: Subject primacy or the absolutive hypothesis. Language, 63, 856–870. Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68, 1–76. Gibson, E. (2000). The dependency locality theory: a distance-based theory of linguistic complexity. In Y. Miyashita, A. Marantz & W. O’Neil (Eds.), Image, language, brain (pp. 95–126). Cambridge, MA: MIT Press. Gibson, E., & Pearlmutter, N. (1998). Constraints in sentence comprehension. Trends in Cognitive Science, 2, 262–268. Gordon, P. C., Hendrick, R., & Johnson, M. (2001). Memory interference during sentence comprehension. Journal of Experimental Psychology: Learning, Memory, & Cognition, 27, 1411–1423. Gordon, P. C., Hendrick, R., & Johnson, M. (2004). Effects of noun phrase type on sentence complexity. Language & Cognitive Processes, 51, 97–114. Gordon, P. C., Hendrick, R., Johnson, M., & Lee, Y. (2006). Similarity-based interference during language comprehension: Evidence from eye tracking during reading. Journal of Experimental Psychology: Learning, Memory & Cognition, 32, 1304–1321.
274
Y. Lin and S.M. Garnsey
Gordon, P. C., Hendrick, R., & Levine, W. H. (2002). Memory-load interference in syntactic processing. Psychological Science, 13, 425–430. Grodner, D., Gibson, E., & Tunstall, R. (2002). Syntactic complexity in ambiguity resolution. Journal of Memory and Language, 46, 267–295. Hawkins, J. (2004). Efficiency and complexity in grammar. Oxford: Oxford University Press. Hsiao, F., & Gibson, E. (2003). Processing relative clauses in Chinese. Cognition, 90, 3–27. Hsu, C.-C. N., & Chen, J.-Y. (2007). A new look at the subject-object asymmetry: The effects of linear distance and structural distance on the processing of head-final relative clauses in Chinese. Presented at the Conference on Interdisciplinary Approaches to Relative Clauses, Cambridge. Hsu, C.-C. N., Hurewitz, F., & Phillips, C. (2006). Contextual and syntactic cues for processing head-final relative clauses in Chinese. Presented at the 19th Annual CUNY Conference on Human Sentence Processing, New York, NY. Hsu, C.-C. N., Phillips, C., & Yoshida, M. (2005). Cues for head-final relative clauses in Chinese. Presented at the 18th Annual CUNY Conference on Human Sentence Processing, Tucson, AZ. Just, M. A., & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in working memory. Psychological Review, 98, 122–149. Kaan, E., Harris, A., Gibson, E., & Holcomb, P. (2000). The P600 as an index of syntactic integration difficulty. Language & Cognitive Processes, 15, 159–201. Keenan, E. L. & Comrie, B. (1997). Noun phrase accessibility and universal grammar. Linguistic Inquiry, 8, 63–99. King, J., & Just, M. A. (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory & Language, 30, 580–602. Kuo, K., & Vasishth, S. (2007). Processing Chinese relative clauses: Evidence for the universal subject preference. Unpublished manuscript. Kwon, N., Polinsky, M., & Kluender, R. (2006). Subject preference in Korean. In D. Baumer, D. Montero & M. Scanlon (Eds.), Proceedings of the 25th West Coast Conference on Formal Linguistics (pp. 1–14). Somerville, MA: Cascadilla Proceedings Project. Lewis, R. L. (1996). Interference in short-term memory: The magical number two (or three) in sentence processing. The Journal of Psycholinguistic Research, 25, 93–115. Lewis, R. L., Vasishth, S., & Van Dyke, J. A. (2006). Computational principles of working memory in sentence comprehension. Trends in Cognitive Sciences, 10, 447–454. Lin, Y. (2005). Word order, animacy, and agreement cues in sentence processing by L1 Mandarin EFL learners. Presented at the Midwestern Conference on Culture, Language, and Cognition, Northwestern University, IL. Lin, C.-J. (2006). Grammar and parsing: A typological investigation of relative clause processing. Ph.D dissertation, University of Arizona. Lin, C.-J., & Bever, T. G. (this volume). Garden path and the comprehension of head-final relative clauses. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. Lin, Y., & Garnsey, S. M. (2009). The contributions of classifiers and pronouns to Mandarin relative clause comprehension. Presented at the 22nd Annual CUNY Conference on Human Sentence Processing, Davis, CA. Liu, A. K.-L. (2005). The structure of relative clauses in Jianshi Squliq Atayal. Concentric: Studies in Linguistics, 31, 89–110. MacDonald, M. C., & Christiansen, M. H. (2002). Reassessing working memory: A comment on Just & Carpenter (1992) and Waters & Caplan (1996). Psychological Review, 109, 35–54. MacWhinney, B., & Pleh, C. (1988). The processing of restrictive relative clauses in Hungarian. Cognition, 29, 95–141. Mak, W., Vonk, W., & Schriefers, H. (2002). The influence of animacy on relative clause processing. Journal of Memory & Language, 47, 50–68.
12
Animacy and the Resolution of Temporary Ambiguity
275
Miller, G. A., & Chomsky, N. (1963). Finitary models of language users. In R. D. Luce, R. R. Bush & E. Galanter (Eds.), Handbook of mathematical psychology (Vol. II, pp. 419–491). New York: Wiley. Miyamoto, E., & Nakamura, M. (2003). Subject/object asymmetries in the processing of relative clauses in Japanese. Presented at the 22nd West Coast Conference on Formal Linguistics. Somerville, MA: Cascadilla Press. O’Grady, W. (1997). Syntactic development: The acquisition of English. Chicago, IL: University of Chicago Press. Packard, J., Ye, Z., & Zhou, X. (this volume). Filler-gap processing in Mandarin relative clauses: Evidence from Event-related potentials. In H. Yamashita, Y. Hirose, & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. Pu, M. M. (2007). The distribution of relative clauses in Chinese discourse. Discourse Processes, 43, 25–53. Reali, F., & Christiansen, M. H. (2007). Processing of relative clauses is made easier by frequency of occurrence. Journal of Memory & Language, 53, 1–23 Schriefers, H., Friederici, A. D., & Kuhn, K. (1995). The processing of locally ambiguous relative clauses in German. Journal of Memory & Language, 34, 499–520. Su, I. R. (2001). Context effects on sentence processing: A study based on the competition model. Applied Psycholinguistics, 22, 167–189. Tabor, W., Juliano, C., & Tanenhaus, M. K. (1997). Parsing in a dynamical system: an attractor-based account of the interaction of lexical and structural constraints in sentence processing. Language & Cognitive Processes, 12, 211–272. Traxler, M. J., Morris, R. K., & Seely, R. E. (2002). Processing subject and object relative clauses: evidence from eye movements. Journal of Memory & Language, 47, 69–90. Trueswell, J. C., Tanenhaus, M. K., & Garnsey, S. M. (1994). Semantic influences on parsing: Use of thematic role information in syntactic ambiguity resolution. Journal of Memory and Language, 33, 285–318. Ueno, M., & Garnsey, S. (2008). An ERP study of the processing of subject and object relative clauses in Japanese. Language & Cognitive Processes, 23, 646–688. Van Dyke, J. A., & Lewis, R. (2003). Distinguishing effects of structure and decay on attachment and repair: A cue-based parsing account of recovery from misanalyzed ambiguities. Journal of Memory & Language, 49, 285–316. Van Dyke, J. A., & McElree, B. (2006). Retrieval interference in sentence comprehension. Journal of Memory & Language, 55, 157–166. Warren, T., & Gibson, E. (2002). The influence of referential processing on sentence complexity. Cognition, 85, 79–112. Weckerly, J., & Kutas, M. (1999). An electrophysiological analysis of animacy effects in the processing of object relative sentences. Psychophysiology, 36, 559–570. Wu, F.-Y., Haskell, T., & Andersen, E. (2006). The interaction of lexical, syntactic, and discourse factors in on-line Chinese parsing: Evidence from eye-tracking. Presented at the 19th Annual CUNY Conference on Human Sentence Processing, New York, NY. Wu, F., Kaiser, E., & Andersen, E. (this volume). Subject preference, head animacy, and lexical cues: A corpus study of relative clauses in Chinese. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. Wu, H.-H. I., & Gibson, E. (2008). Processing Chinese relative clauses in context. Poster presented at the 21st annual CUNY Conference on Human Sentence Processing, Chapel Hill, NC. Yang, C.-L., Johnson, M., & Gordon, P. C. (2008). The effect of contrasting linear ordering of filler-gap dependency in the processing of sentence integration: Evidence from the processing of relative clauses in Chinese. Presented at the CUNY Conference on Human Sentence Processing, Chapel Hill, NC. Yoshida, M., Aoshima, S., & Phillips, C. (2004). Relative clause prediction in Japanese. Presented at the 17th Annual CUNY Conference on Human Sentence Processing, College Park, MD.
Chapter 13
Garden Path and the Comprehension of Head-Final Relative Clauses Chien-Jer Charles Lin and Thomas G. Bever
13.1 Headedness and the Comprehension of Relative Clauses Constructing dependent relations between linguistic units has been a central issue in sentence processing research. In order to establish a dependency, the parser needs to detect the dependent elements and formulate syntactic and semantic relations between these elements. The head plays an important role in sentence comprehension as it specifies subcategorizing and adjunctive relations with other elements within the same phrase and gets integrated with linguistic units outside of the phrase. Two main proposals in sentence processing have been made about the integration process involving the head. One takes an incremental position to argument integrations, positing that arguments are assigned their thematic roles as soon as they come into presence in a sentence (Bader & Lasser, 1994; Miyamoto, 2002). The other argues for a head-driven approach to thematic assignments (Abney, 1989; Pritchett, 1991, 1992), maintaining that thematic roles are not assigned until the head is reached. In this chapter, we focus on the comprehension of head-final relative clauses. A critical issue in the processing of these relative clauses is when the parser realizes it has encountered a relative clause and whether reanalysis after the temporary ambiguity successfully takes place. Incremental processing and head-driven parsing each have distinct predictions about the comprehension of head-final relative clauses. According to incremental processing, pre-head sequences are analyzed based on top-down heuristics (e.g., Canonical Form Constraints of Bever, 1970 and Townsend & Bever, 2001) and processing strategies such as minimal attachment (Frazier, 1987a). Thus a head-final relative clause creates a garden path for the parser since prior to the head noun, a relative clause is preferably analyzed as a main clause. Based on the head-driven approach, however, the temporary ambiguity prior to the head C.-J.C. Lin (*) Department of East Asian Languages and Cultures, Indiana University, Bloomington, IN, 47405-7005 USA e-mail: [email protected]
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_13, Ó Springer ScienceþBusiness Media B.V. 2011
277
278
C.-J.C. Lin and T.G. Bever
makes processing laborious since multiple possibilities need to be maintained before the head is reached. In this chapter, the incremental garden-path analysis is supported by the processing of object-extracted relative clauses that modify the objects of the main clauses in Chinese (creating a superficial NVN sequence), while the processing of Chinese relative clauses at sentence-initial positions (i.e., those modifying the subjects of the main clauses) suggests multiple parses have been retained. We further discuss how such a garden path should be dealt with so that an unbiased comparison between the processing of head-final relative clauses and that of head-initial relative clauses can be made. In the following sections, we first introduce the structural properties of head-final relative clauses and demonstrate the existence of garden path in Section 13.2. Section 13.3 then reviews different approaches to resolve this garden path and evaluates the degrees of success of each approach. Section 13.4 discusses the significance of the garden-path effect in studying head-final relative clauses.
13.2 Structural Properties of Head-Final Relative Clauses and the Issue of Garden Path Head-final relative clauses present a word order mirroring that of headinitial relative clauses (1–2). Such a typological difference produces distinct filler-gap relations. In a head-initial language like English, the head (i.e., the filler) precedes the relativized gap. When the filler is encountered, the parser initiates a gap-searching process, actively postulating a gap at the earliest possible position in the upcoming string of words (i.e., the Active Filler Strategy, Frazier & d’Arcais, 1989). With head-final relative clauses, however, the head and the relativizer are both linearized to the right of the relative clause. As a result, the gap precedes the filler. Without overt marking, detecting the gap prior to the head noun becomes a challenge for the parser. (1)
(2)
head-initial relative clause head noun – relativizer – relative clause filler >>>>>>>>>>>>> gap e.g., the guyFILLER that you bumped into __GAP yesterday [English] head-final relative clause relative clause – relativizer – head noun gap >>>>>>>>>>>>>> filler e.g., ni zuotian pengjian __GAP de nage renFILLER [Chinese] you yesterday bumped into __GAP relativizer that guyFILLER ‘the guy that you bumped into yesterday’
13
Garden Path and the Comprehension of Head-Final Relative Clauses
279
A crucial difference in the comprehension of head-initial and head-final relative clauses lies in the relative ordering between the filler and the gap. When the filler precedes the gap as in a head-initial relative clause, the left edge of a relative clause is unambiguously marked by the relativizer. When the filler follows the gap as in a head-final relative clause, however, the parser is additionally challenged by the structural ambiguity prior to the head noun. Since the left edge of the relative clause is not overtly indicated, both a mainclause analysis and a relative-clause analysis are initially possible. Before presenting processing evidence for this potential misanalysis, let us first consider how relative clauses appear within sentences. In the following, we take Chinese relative clauses as an example. In (3), relative clauses with subject and object extractions appear at sentence-initial positions to modify the subjects of the matrix clauses; in (4), they appear in medial positions to modify the objects. (3)
(4)
a. Subject-extracted relative clause modifying the matrix subject dashang shangren de daitu jiandaole matrix jizhe 1 reporter hurt businessman REL gangster saw [V N de HN ] ‘The gangster who hurt the businessman saw the reporter.’ b. Object-extracted relative clause modifying the matrix subject shangren dashang de daitu jiandaole jizhe businessman hurt REL gangster saw reporter [N V de HN ] ‘The gangster who the businessman hurt saw the reporter.’ a. Subject-extracted relative clause modifying the matrix object jizhe jiandaole dashang shangren de daitu reporter saw hurt businessman REL gangster [V N de HN ] ‘The reporter saw the gangster who hurt the businessman.’ b. Object-extracted relative clause modifying the matrix object jizhe jiandaole shangren dashang de daitu reporter saw businessman hurt REL gangster [N V de HN ] ‘The reporter saw the gangster who the businessman hurt.’
The sentences in (3) begin with relative clauses. A subject relative clause as in (3a) contains a sentence-initial gap, which can be mistaken as a null subject in a 1
REL stands for ‘‘relativizer.’’ HN stands for ‘‘head noun.’’ CL stands for ‘‘classifier.’’ Relative clauses involving subject extractions are abbreviated as SRCs; those involving object extractions as ORCs.
280
C.-J.C. Lin and T.G. Bever
main clause (i.e., a subject that is pro-dropped). In Chinese, this sequence can be analyzed as a sentence containing a pro (as in 5a), a PRO (as in 5b), or correctly as a gapped relative clause (as in 5c). Both (5a-b) are main-clause misanalyses; only (5c) is the correct subordinate relative-clause analysis. (5)
a.
b.
c.
pro dashang shangren hurt businessman ‘(Somebody) hurt the businessman.’ PRO dashang shangren hui dailai gengduo mafan hurt businessman will bring more trouble ‘Hurting the businessmen will bring more trouble.’ ti dashang shangren de daitui jiandaole jizhe GAP hurt businessman REL gangster saw reporter ‘The gangster who hurt the businessman saw the reporter.’
Object relative clauses like (3b) are also ambiguous prior to the relativizers. These fragments can be misread as part of a main clause (as in 6a) because of its apparent NV order (conveniently mapped onto an agent-action-(patient) template). The correct gap analysis of a relative clause (in 6b) requires postulating more layers of syntactic nodes and is more consuming to construct.2 (6)
a.
b.
shangren dashang yige daitu businessman hurt one gangster ‘The businessman hurt a gangster.’ shangren dashang ti de daitui jiandaole jizhe businessman hurt GAP REL gangster saw reporter ‘The gangster who the businessman hurt saw the reporter.’
The discussion so far shows that both subject and object relative clauses can be misread as main clauses before the relativizer and the head nouns. Object relative clauses are possibly more susceptible to the main-clause misanalysis because of the initial NV sequence, following the canonical main-clause order – NVN – in Mandarin Chinese. Similar issues of temporary ambiguity in Japanese have been discussed by Hirose (2006). When relative clauses modify the object of a sentence, they appear in the medial position of a sentence. Due to the lack of overt marking on the left edge, the leftmost element of a relative clause can be mistaken as part of the ongoing 2
In a sentence completion task, Hsu et al. (2005) found NV fragments predominantly solicited main-clause completions. See also Chapter 14 (Ng & Fodor) of this volume on the interpretation of empty categories in Chinese in relation to the left-edge ambiguity in headfinal structures and its interplay with the existence of a plausible referent in the contexts.
13
Garden Path and the Comprehension of Head-Final Relative Clauses
281
main clause. In the case of a subject relative clause, a PRO analysis in (7a) is still possible, in contrast to a relative-clause parse (like 7b). (7)
a.
b.
jizhe xihuan PRO dashang shangren reporter like hurt businessm ‘The reporter enjoys hurting businessmen.’ jizhe xihuan ti dashang shangren de daitui reporter like hurt businessman REL gangster ‘The reporter likes the gangster who hurt the businessman.’
An object relative clause is also susceptible to misanalysis. The embedded subject can be mistaken as the object of the matrix clause, following an apparent NVN order as in 8a, thus calling for subsequent reanalysis at the embedded verb or the relativizer regions. Similar problems of garden path can be found in Japanese and Korean relative clauses. (8)
a.
b.
jizhe jiandaole shangren . . . reporter saw businessman . . . ‘The reporter saw the businessman . . ..’ jizhe jiandaole shangren dashang ti de daitui reporter saw businessman hurt REL gangster ‘The reporter saw the gangster who the businessman hurt.’
The Chinese examples presented above show that in isolated sentence comprehension, garden-path readings are likely to take place. Crucially, since there is no overt marking on the left edge of a relative clause, the parser is constantly challenged by the uncertainty of main-clause versus subordinate-clause analyses.3 In the following, we compare two self-paced reading experiments of Chinese relative clauses. In one experiment (hereafter cited as Experiment 1), participants were given specific instructions about the existence and the position of a relative clause in each experimental sentence. The purpose of Experiment 1 was to reduce the possibility of garden path by indicating exactly where in the 3
In this chapter, we focus on garden path associated with canonical relative clauses, namely simple relative clauses that involve subject or object extractions. Lin (2006, Experiment 5) investigated ‘‘movement-induced’’ garden path in head-final relative clauses, where topicalization inside a subject relative clause produces an apparent NVN sequence and an unrecoverable garden-path effect: nyuyanyuan conglai bu juede houhui [nanyoui fangqiguo haoji wei ti] de boyfriend abandoned many CL REL actress always not feel regretful N V N ‘The actress who has abandoned many boyfriends never felt regretful.’
282
C.-J.C. Lin and T.G. Bever
sentence the relative clause is located so that participants knew where to expect an upcoming relative clause in a sentence. In another experiment (Experiment 2 below), the sentences containing relative clauses were presented along with filler sentences of various syntactic types; no specific instructions about the existence and position of relative clauses were given. In Experiment 2, therefore, there is the possibility of garden path described in (5–8), while in experiment 1, the possibility of misreading is greatly reduced. Comparing the reading patterns in these two experiments will shed light on whether the comprehension of headfinal relative clauses without specific cues would involve main-clause gardenpath misanalyses.4 These two experiments each recruited as participants 48 college students in Taiwan, who were native speakers of Mandarin Chinese. Two factors were investigated in a factorial design—whether the relative clauses modified the subject or the object positions of the main clauses and whether the relative clauses involved subject or object extractions. Figures 13.1a, b present the average reading times of the relative-clause regions in these two experiments. In Fig. 13.1a, the relative clauses modified the subjects of the main clauses (as exemplified by 3a-b); in Fig. 13.1b, they modified the objects (exemplified by 4a-b). Comparison of these two experiments demonstrated a reading-time increase on the relativizer and the head-noun regions of the object relative clauses in the object-modifying condition when the sentences were read without special instructions about where the relative clauses were located (i.e., in Experiment 2). It is important to note that the increase of reading time was significant only at the relativizer ðFð1; 23Þ ¼ 17:85; p 5 0:001Þ and the head noun ðFð1; 23Þ ¼ 4:13; p 5 0:05Þ of the object-modifying object-extracted relative clause, not at the pre-head relative-clause regions and not on the subjectextracted relative clauses. This reading-time increase demonstrated that object relative clauses that modified the objects of the main clauses are especially susceptible to garden-path readings. When the participants were provided with information about where the relative clause was located as in Experiment 1, the garden-path reading (with increased reading time) was not observed.5 The results suggest that head-final relative clauses are susceptible to gardenpath readings due to the lack of markings on the left edge. According to the selfpaced reading data, such a garden path was not obvious when the relative clauses appeared at the subject-modifying position (e.g., 5–6). However, when 4
Experiment 2 cited in this article was first reported in Lin and Bever (2006). Even though the reading time on the head noun in the object-modifying object relatives was long (over 1400 ms), the comprehension accuracies were over 95% and did not differ across conditions (Experiment 2: SRC-S modification 97%, ORC-S modification 95%, SRC-O modification 97%, ORC-O modification 96%), suggesting that after reanalysis, the parser was able to reach similar comprehension accuracies regardless of the garden path. Note also the possibility that the long reading time on this position may be due to the end-of-sentence wrap-up effect, thus enlarging the differences.
5
13
Garden Path and the Comprehension of Head-Final Relative Clauses
283
a. subject-modifying relative clauses 1600 1400 1200 1000 800 600 400 V/N
N/V Experiment 1-src Experiment 2-src
DE
HN
Experiment 1-orc Experiment 2-orc
b. object-modifying relative clauses 1600 1400 1200 1000 800 600 400 V/N
N/V Experiment 1-src Experiment 2-src
DE
HN
Experiment 1-orc Experiment 2-orc
Fig. 13.1 Reading times (ms) of the relative clauses modifying (a) the subject and (b) the object positions of the main clauses in Experiments 1 & 2
the relative clauses appeared at the object-modifying position and when they involved object extractions, the superficial NVN sequence engendered a mainclause illusion and induced a garden-path parse. When the participants were expecting a relative clause (as instructed in Experiment 1), however, this gardenpath effect was not observed. This garden-path effect is further supported by an ERP study by Packard, Ye, and Zhou (this volume), who examined the processing of subject and object
284
C.-J.C. Lin and T.G. Bever
extractions in relative clauses that modified either the subjects or the objects of the main clauses. The relative clauses in their study appeared without contextual or instructional motivations (like Experiment 2 of our study). They found greater P600s in object-modifying object-gap relative clauses and attributed them to the cost of revising a garden-path analysis. This comparison has two important implications. First, head-final relative clauses that appear at sentence-initial positions are not severely garden-pathed. That is, multiple parses may be maintained for superficial NV and VN sequences. However, when the superficial sequence appears to be argumentcomplete (e.g., NVN) as in the object-modifying object relative clauses, the parser would be misled to a main-clause analysis. Second, this comparison demonstrated that overt instructions about the existence of relative clauses are able to reduce the garden-path effect on head-final relative clauses. This latter methodological implication is further discussed in Section 13.3.4.
13.3 Attempts to Avoid the Garden Path If one attempts to compare the processing of head-final relative clauses with that of head-initial relative clauses, the issue of garden path discussed in Section 13.2 has to be considered. Much previous sentence processing research has focused on the processing differences between subject and object extractions in head-initial languages (Brazilian Portuguese: Gouvea, 2003; Dutch: Frazier, 1987b; Mak, Vonk, & Schriefers, 2002; English: Ford, 1983; Gibson, Desmet, Grodner, Watson, & Ko, 2005; King & Just, 1991; King & Kutas, 1995; Traxler, Morris, & Seely, 2002; French: Cohen & Mehler, 1996; Frauenfelder, Segui, & Mehler, 1980; Holmes & O’Regan, 1981; German: Mecklinger, Schriefers, Steinhauer, & Friederici, 1995; Schriefers, Friederici, & Kuhn, 1995). So far, there is little dispute that subject extractions are easier to comprehend in headinitial relative clauses. In English, for example, (9) is easier than (10) according to self-paced reading studies (King & Just, 1991), eye-movement studies (Traxler et al., 2002) and ERP studies (King & Kutas, 1995). (9) (10)
The gangster who hurt the businessman saw the reporter. The gangster who the businessman hurt saw the reporter.
There are various plausible explanations for this advantage on subject extractions. A locality-based working-memory account (Gibson, 1998), a structural account (Lin, 2006; O’Grady, 1997) and a canonical templatic account (Bever, 1970) are all compatible with this subject advantage in head-initial languages (see Hsiao & Gibson, 2003; Lin, 2006 for reviews of these accounts). Head-final languages, on the other hand, distinguish among the predictions of these theories, making it possible to test the validity of these accounts.
13
Garden Path and the Comprehension of Head-Final Relative Clauses
285
Much research has recently been conducted regarding subject versus object extractions in head-final relative clauses (Chinese: Hsiao & Gibson, 2003; Lin & Bever, 2006, 2007; Lin & Garnsey, this volume; Packard et al., this volume; Qiao & Forster, 2008; Wu & Gibson, 2008; Korean: Kwon, Polinsky, & Kluender, 2004, 2006; Japanese: Ishizuka, 2005; Ishizuka, Nakatani, & Gibson, 2006; Miyamoto & Nakamura, 2003; Ueno & Garnsey, 2008). Given that the issue is about extractions out of subject and object positions, the additional problem induced by garden path in head-final relative clauses need to be controlled for. As discussed in Section 13.2, there is potentially more garden path on object relative clauses than subject relative clauses in object-modifying relative clauses. Therefore, research investigating extraction effects of headfinal relative clauses need to consider the issue of garden path (especially when they appear in the medial positions of the main clauses) so that a fair comparison between extraction effects in head-initial and head-final relative clauses can be made.6 In this section, we review previous attempts to tackle the garden-path issue in head-final relative-clause processing. In particular, we examine whether and how garden paths could be successfully avoided. We restrict this review to sentence comprehension in Chinese, Japanese, and Korean. Japanese and Korean are both strictly head-final. Examples in (11) and (12) illustrate relative clauses in Japanese and Korean. The basic word orders in Japanese and Korean are SOV. In both languages, the pre-head nominal entities are case-marked. The primary difference between these two languages is the existence of the suffix -n on the embedded verb functioning as a relative-clause marker in Korean. (11)
6
a. Japanese SRC (Miyamoto & Nakamura, 2003, p. 343) [__ tosiyorino obaasan-o basutei-made miokutta] onnanoko elderly woman-ACC bus stop-to accompanied girl ‘the girl that accompanied the elderly woman to the bus stop’
Note that the study of Lin and Garnsey (this volume), which used topicalized objectmodifying relative clauses with the head nouns missing, was additionally complicated by a garden-path effect. When the head nouns of these topicalized relative clauses were omitted, the subjects of the matrix clauses, which now follow these object-modifying relative clauses, can be taken as the missing head nouns. It is unlikely that the participants came to the correct parse since the only cues were matrix clauses with objects missing (which is not uncommon in Chinese). In their Experiment 2, where inanimate matrix subjects were used to help disambiguate, it is still unclear whether participants have successfully constructed a relative clause with head nouns omitted. Lin (2006, Experiment 5), for example, showed that animacy is not sufficient to indicate the existence of a relative clause. It is also noteworthy that relative clauses with missing head nouns are best motivated by contexts that supplied the missing heads. In Lin and Garnsey’s study, however, no context was provided.
286
(12)
C.-J.C. Lin and T.G. Bever
b. Japanese ORC (Miyamoto & Nakamura, 2003, p. 343) [tosiyorino obaasan-ga __ basutei-made miokutta] onnanoko elderly woman-NOM bus stop-to accompanied girl ‘the girl that the elderly woman accompanied to the bus stop’ a. Korean SRC (Kwon, p.c.) [__ naitun pwuin-ul bes cenkecang-kkaci tonghayngha]-n sonye elderly woman-ACC bus stop-to accompany-REL girl ‘the girl that accompanied the elderly woman to the bus stop’ b. Korean ORC (Kwon, p.c.) [naitu-n pwuin-i __ bes cenkecang-kkaci tonghayngha]-n sonye elderly woman-NOM bus stop-to accompany-REL girl ‘the girl that the elderly woman accompanied to the bus stop’
In Chinese, the head positions are mixed. In the verbal domain, Chinese is head-initial. The object follows the verb, resulting in the same canonical word order SVO as in English. In the nominal domain, however, it is head-final, with the nominal modifiers appearing prior to the nominal heads. In all three languages, the challenge posed by a head-final relative clause is that no overt marking indicates its left edge. In order to adopt a relative-clause parse in left-to-right incremental processing, the relative clause needs to be motivated before the head appears. So far, researchers have adopted both covert and overt cues to indicate the existence of a head-final relative clause. Overt cues include grammatical markings inside the relative clauses and special instructions about the existence of relative clauses as part of the experimental design. Covert cues include referential contexts that motivate relative clauses and semantic clashes that indicate syntactic discontinuity. We evaluate each of these in turn.7
13.3.1 Markings Inside the Relative Clauses In Chinese, object relative clauses can be optionally marked by the preverbal marker suo as in (13).8 Suo does not appear in main clauses; therefore, when it appears, it indicates the existence of an object relative clause. In an off-line sentence completion experiment, Hsu, Phillips, and Yoshida (2005) used sentence fragments containing suo (e.g., laoshi xiang xuesheng suo . . . – ‘teacher to students SUO . . .’) and was able to induce more completions of relative clauses than when they used sentence fragments without suo. 7
See research by Kang and Speer (2003) and Kang, Speer, and Nakayama (2004) on how prosody facilitates the identification of clausal boundaries in head-final languages, such as Korean and Japanese. 8 See Chiu (1995) and Ting (2003) for relevant syntactic analyses of suo.
13
(13)
Garden Path and the Comprehension of Head-Final Relative Clauses
287
shangren (suo) dashang de daitu jiandaole jizhe businessman (SUO) hurt REL gangster saw reporter ‘The gangster who the businessman hurt saw the reporter.’
In an on-line self-paced reading experiment, Hsu et al. (ibid.) compared the processing of object relative clauses with and without suo. They found that the parser slowed down upon reaching suo due to its general lower frequency of occurrence. When the parser reached the relativizer and the head noun, however, the reading time was shorter in sentences with suo than in those without it. The results suggested that suo did successfully indicate the existence of an object relative clause in the sentence. Without it, an object relative clause was more difficult to process due to the temporary ambiguity. The sentence-completion and self-paced reading results showed that suo can indeed indicate an object relative clause before the head-noun regions. Nevertheless, there are critical insufficiencies about using suo markings to indicate Chinese relative clauses. First, suo appears before the embedded verb inside an object relative clause. The embedded subject is not marked and can still be incorrectly treated as part of a main clause (i.e., the kinds of garden path described in (6) and (8)). Even though these sentences are disambiguated relatively early, the possibility of a garden path is still not removed. Second, suo only marks relative clauses with object extractions in Chinese. Since there is no equivalent marker in subject-extracted relative clauses, it is not possible to compare subject and object relative clauses by using suo.
13.3.2 Classifier-Noun Mismatch In East-Asian languages such as Chinese, Japanese, and Korean, nominal classifiers appear before count nouns under constraints on semantic consistency. Different classifiers selectively co-occur with nouns that share semantic properties in terms of shape and other physical dimensions. In Chinese, for instance, the classifier tiao denotes nominal entities that are thin, long and flexible in shape. Nouns such as shengzi ‘‘rope’’, she ‘‘ snake’’, malu ‘‘road’’ but not gunzi ‘‘rod’’ or xiang ‘‘elephant,’’ co-occur with the classifier tiao as in (14). (14)
a. yi tiao she one CL snake ‘a snake’ b. *yi tiao xiang one CL elephant ‘an elephant’ c. yi tiao [xiang caibian de] she one CL elephant step-flat REL snake ‘a snake that an elephant stepped on (and became flat)’
288
C.-J.C. Lin and T.G. Bever
The semantic agreement between the classifier and the noun phrase following it has therefore been used to indicate potential phrasal boundaries. In (14c), for instance, the fact that tiao and xiang cannot compose one constituent implies that xiang should be part of a different phrase. Thus a relative clause boundary may be postulated between tiao and xiang. Yoshida, Aoshima, & Phillips (2004), Hsu, et al. (2005), and Hsu, Hurewitz, & Phillips (2006) used this classifier-noun mismatch to signal a relative clause boundary. Yoshida et al. (2004) found that, in Japanese, the reading time at the disambiguating region (i.e., the embedded verb yonda in 15) was shorter in the classifier-noun mismatch condition. Classifier-noun mismatch was picked up as an indicator of an embedded relative clause in Japanese. (15)
[san-satu-no [sensee-ga yonda] hon]. . . three-Classifier(book)-GEN teacher-NOM read book ‘three books that the teacher read’ (Yoshida et al., 2004)
In Chinese, however, Hsu et al. (2005) did not find similar patterns when using classifier-noun mismatch as the only cue for relative clauses. Instead, they found longer reading times on the mismatching condition (16) than the matching condition (17) starting from the mismatching noun phrase till the head noun region. (16)
(17)
yonggong-de xuesheng zaoshang laoshi caixiang na-si-pian teacher guess that-four-CL(article) diligent student morning jingchang yu tongxue taolun de wenzhang . . . often with classmate discuss REL article ‘The teachers thinks that the four articles that the diligent students often discussed with a classmate in the morning . . .’ (Hsu et al., 2005) laoshi caixiang na-si-wei yonggong-de xuesheng zaoshang teacher guess that-four-CL(human) diligent student morning jingchang yu tongxue taolun de wenzhang . . . often with classmate discuss REL article ‘The teachers thinks that the four articles that the diligent students often discussed with a classmate in the morning . . .’ (Hsu et al., 2005)
The study by Hsu et al. (2005) suggested that classifier-noun mismatch alone was not adequately suggestive of a relative clause parse in Chinese. Comparing the Chinese and Japanese materials in (15) and (16), it seems plausible that when the mismatching noun phrase follows the classifier in Chinese, the parser is still inclined to treat the classifier and the adjacent noun phrase as one classifier
13
Garden Path and the Comprehension of Head-Final Relative Clauses
289
phrase. Such a misconstrued classifier-NP constituent is taken as syntactically well-formed but semantically incongruent. In such cases, the classifier-noun mismatch induced a surprisal effect but not a correct syntactic reanalysis; therefore, an embedded relative clause was not successfully postulated. With incongruity as such, it is likely that the parser never arrived at the correct parse (Ferreira, 2003; Lin, 2008).9 In Japanese, however, the genitive marker no makes a potential phrasal boundary salient, thus increasing the likelihood of a relative-clause analysis. In a follow-up experiment, Hsu et al. (2006) tested the disambiguating power of classifier-noun mismatch with the help of a felicitous context (which provided a narrowed set of referents consistent with the mismatched classifier and the head noun). Sentences with classifier-noun mismatches were then indeed read faster on the head-noun region. With appropriately designed contexts, the parser can then treat classifier-noun mismatch as an indicator of a relative clause boundary. This leads to our discussion on the role of context in the processing of head-final relative clauses.
13.3.3 Context That Induces Relative Clauses Crain and Steedman (1985, p. 342) showed that an appropriate referential context facilitates a relative-clause parse in the target sentence following the context. They contrasted two kinds of contexts, one with and one without contrasting referents. When they presented a context without contrasting referents as in (18), a complement clause like (20) was preferred. When the context provided referents that need to be further specified as in (19), then a reduced relative clause like (21) became the preferred parse. (18)
(19)
(20) (21)
9
Context that induces a complement clause A psychologist was counseling a married couple. One member of the pair was fighting with him but the other one was nice to him. Context that induces a relative clause A psychologist was counseling two married couples. One of the couples was fighting with him but the other one was nice to him. Complement target sentence The psychologist told the wife that he was having trouble with her husband. Relative target sentence The psychologist told the wife that he was having trouble with to leave her husband.
Hsu et al. surmised that the longer reading times on the true head nouns in the mismatching condition may be due to the larger integration cost between a distant classifier and the real matching head noun in (16).
290
C.-J.C. Lin and T.G. Bever
Relative clauses, as discussed, serve particular pragmatic functions within context. They increase ‘‘referential coherence’’ in texts and tend to appear ‘‘when the speaker assumes that the referent’s identity is accessible to the hearer – but not easily accessible (Givo´n, 1993, p. 108).’’10 The propositional content inside a relative clause usually refers to the knowledge about the referent that is already familiar to the hearer.11 These referential relative clauses are the so-called restrictive relative clauses. In English, restrictive and non-restrictive relative clauses are distinguished by the existence of a comma or pause after the head nouns of non-restrictive relatives but not after those of restrictive relatives. Chinese relative clauses are not overtly marked regarding restrictiveness. While most Chinese relative clauses are restrictive, whether non-restrictive relative clauses exist in Chinese remains controversial. See J. Lin (2003) for a review of this controversy. A reasonable way to motivate a relative clause is therefore to provide a context such as the sort constructed by Crain and Steedman (1985). Since a relative clause can help select a referent from a set of referents previously mentioned, an appropriate context constructs a situation in which a small set of referents compete to be selected as the topic of the text that follows. As mentioned in 13.3.2, Hsu et al. (2006) used contexts in addition to classifiernoun mismatches to induce relative clauses. Following Crain and Steedman (1985), they contrasted two kinds of contexts. The first context provided two referents of the same kind (as illustrated by 22) followed by a target relative clause selecting one of the referents. The second context provided two different kinds of referents (as illustrated by 23). Self-paced reading results showed that when the relative clause served to select between referents of the same kind as 10
Givo´n (1993, p. 108) summarized the referential properties of Restrictive Relative Clauses (RRCs) in terms of ‘‘referent tracking strategies’’ as the following: a. The speaker assumes that a certain state or event is known, familiar or mentally accessible to the hearer. b. The proposition corresponding to that familiar state/event is thus pragmatically presupposed. c. The referent to be identified is a participant – subject, direct object, indirect object, etc. – in the state or event coded in the proposition. d. The familiar proposition thus helps guide the hearer toward identifying the referent in his/ her mentally stored knowledge. It grounds the referent in the hearer’s knowledge-base. e. The proposition used for such grounding is coded syntactically as a RRC. 11 The information status inside sentences with relative clauses also matters. Gibson et al. (2005) compared relative clauses that appeared early (e.g., modifying the subject of the matrix clause) and those that appeared late (e.g., modifying the object of the matrix clause) in a sentence. They showed that relative clauses are read with greater ease when they appear early in the sentence. The propositional content inside the relative clause serves as grounding information within the sentence, referring to knowledge already present in text or new information that is not likely to be challenged (i.e., Givo´n’s pragmatic definition of a restrictive relative clause). As relative clauses serve to ground the referent as old information in the sentence, they tend to appear early, e.g., in the subject position.
13
Garden Path and the Comprehension of Head-Final Relative Clauses
291
in (22), a sentence in which the left edge of the relative clause was indicated by a classifier-noun mismatch was read with greater ease. That is, the classifier-noun mismatch works with a felicitous context to foreshadow a relative clause. In (23), where the context did not provide a need for further specification, the grounding information provided by a relative clause was redundant and the classifier-noun mismatch was insufficient to indicate a relative-clause parse. (22)
(23)
Context with two same kinds of referents: The college student upstairs has two motorcycles. He does not maintain one of the motorcycles, but he maintains the other motorcycle very carefully. Target sentence: This semester, the motorcycle that the college student maintained carefully was stolen. Context with different referents: The college student upstairs has one motorcycle and one computer. He does not carefully maintain his computer, but he maintains his motorcycle very carefully. Target sentence: This semester, the motorcycle that the college student maintained carefully was stolen.
Ishizuka et al. (2006) and Wu and Gibson (2008) also adopted referential contexts prior to their target relative clauses (as in 24). The goal was to motivate a relative clause to avoid potential garden-path effects.12 (24)
A reporter interviewed a writer on a TV program. Then the writer interviewed another reporter for his new novel. Taro: ‘Which reporter stands as a candidate for the election?’ Hanako: ‘It seems to be the reporter who {the writer interviewed / interviewed the writer}.’ (Ishizuka et al., 2006)
To evaluate whether a garden-path effect was successfully avoided, one needs to compare the reading patterns in the experiments with and without contexts. With contexts, Ishizuka et al. (2006) and Wu and Gibson (2008) both reported easier comprehension on object extractions. Ishizuka et al.’s (2006) findings in Japanese contrasted with findings by Ishizuka (2005), Miyamoto and Nakamura (2003), and Ueno and Garnsey (2008), who presented relative clauses without contexts and found comprehension advantages for subject extractions. Wu and Gibson’s results in Chinese also contrasted with the subject-extraction advantage reported by Lin and Bever (2006).13 It was suggested 12 Hsu and Chen (2007) adopted similar contexts and are subject to similar problems as those discussed below. 13 See Lin and Bever (2006) for critiques on the object-extraction advantage reported by Hsiao & Gibson (2003). Lin and Bever argued that the reported advantage resulted from
292
C.-J.C. Lin and T.G. Bever
that when relative clauses are presented in isolation, the advantage for a subject extraction may be due to the garden-path effect on object extractions. When relative clauses are presented in context, an advantage for object relative clauses is actually obtained. Several crucial aspects about these experiments need to be considered to evaluate the significance of the results. While a context potentially motivates the occurrence of a relative clause, it also brings forth irrelevant effects. In both studies, the context was about three referents (two were of the same kind and one was of a different kind). The head of the target relative clause was one of the two referents of the same kind. The format of the context was the following: (25)
(26)
Context of Ishizuka et al. (2006) and Wu and Gibson (2008): a. A verbed B, and b. B verbed another A. Target sentence of Wu and Gibson (2008): a. Subject relative clause: [__ verbed B] relativizer A (meaning ‘the A that verbed B’) b. Object relative clause: [B verbed __] relativizer A (meaning ‘the A that B verbed’)
When the target relative clause (in the form of (26a-b)) was read after the context in (25), (26b) was found to be comprehended faster than (26a). It is crucial to note that thematic priming effect may exist between the context and the target sentence to facilitate the comprehension of the object relative but not the subject relative. That is, the context itself highlights the canonical thematic order (NVN mapping onto Agent-Action-Patient in Chinese and N-nominative N-accusative V mapping onto Agent-Patient-Verb in Japanese), which appears in the object relative but not in the subject relative. The fact that (26b) maps directly onto the context (25b) maintaining an NVN order in Chinese, while (26a) does not, potentially makes (26b) easier than (26a). This processing asymmetry is not related to subject versus object extractions per se. Lin’s (2009b) self-paced reading experiments manipulated the thematic orders in the contexts and found that when (25b) was passivized, with the thematic roles repositioned as Patient-Agent-Verb, the advantage for the object relative could no longer be obtained.
nested dependencies being more difficult than serial dependencies. In a self-paced reading experiment focusing on doubly-embedded relative clauses (Lin & Bever, 2007), it was demonstrated that the effect of dependency types (nested dependencies being harder than serial dependencies) actually contributed to the advantage of object extractions reported by Hsiao and Gibson.
13
Garden Path and the Comprehension of Head-Final Relative Clauses
293
To sum up, an appropriately constructed referential context does facilitate the processing of relative clauses. However, it is crucial to attend to the potential thematic priming effects introduced by the context.
13.3.4 Overt Instructions on the Existence of a Relative Clause Even though providing a context is a natural way to induce relative clauses, the fact that there is no context that ensures the appearance of a relative clause and that the context itself may affect the comprehension of the target sentences calls for a more controlled experimental methodology. One way to ensure the participants know they are reading head-final relative clauses is to simply tell them they are doing so in the experimental instructions. Lin and Bever (2007) explicitly instructed the participants that all the experimental sentences contained relative clauses. In addition, we specified the exact position of the relative clause inside the sentence – whether it appeared early in the sentence, modifying the subject position, or late modifying the object of the sentence.14 The goal was to ensure that there was no ambiguity in the reading of these head-final relative clauses by letting the participants know exactly where to expect relative clauses in the experimental sentences. As discussed earlier in Section 13.2 (summarized in Fig. 13.1), reading times collected in this way (Experiment 1) were reduced on the relativizer and head-noun regions (but not at the pre-relativizer regions). We concluded that with instructions like those of Lin and Bever (2007), the parser was indeed expecting the coming of a relative clause and thus no garden-path effect or reanalysis took place on the relativizer and head-noun regions. This methodology makes it possible to evaluate the processing differences between subject and object extractions in head-final relative clauses. Comparing subject and object extractions in such a way, we found an advantage for subject extractions at the relativizer and the head-noun regions of doubly embedded (nested) relative clauses. No advantage was found on object extractions. This finding was also consistent with the greater number of subject relatives found in the Chinese Treebank corpus (Wu, Kaiser, & Andersen, this volume; Lin, 2009a).
14 The participants were provided with examples of relative clauses in the instructions. All the sentences with subject-modifying relative clauses were presented in one block. All sentences with object-modifying relative clauses were presented in a separate block. Sentences with single layers of relative clauses and those with double layers of relative clauses were tested separately.
294
C.-J.C. Lin and T.G. Bever
13.4 Implications In this chapter, we demonstrated the garden-path effect in the processing of head-final relative clauses and evaluated the different methods that have been adopted to deal with it. Our discussions on this garden-path effect have the following implications. In terms of headedness and sentence processing, it was demonstrated that a head-final relative clause is susceptible to garden-path readings due to the lack of markings on the left edge. The parser constructs syntactic structures incrementally before the head is reached, preferably following the canonical word orders in the language (Bever, 1970). The comparison between subject-modifying relative clauses and object-modifying relative clauses reveals a garden-path analysis when the sequence follows an apparent canonical NVN order. Unless otherwise motivated, a relative-clause analysis is only postulated as a last resort – when the disambiguating regions, including the relativizer and the head noun, are reached. This garden-path effect was found on object-extracted relative clauses modifying the object of the main clause due to the superficial NVN main-clause illusion. However, relative clauses that modify subjects did not show the same gardenpath effects, suggesting the possibility that when a sentence-initial NV sequence and a VN sequence are encountered, multiple parses (e.g., main clauses and relative clauses) may have been maintained. A seemingly argument-complete sequence produces the greatest garden-path effect. In terms of a general processing theory for relative clauses in both headinitial and head-final languages, one needs to consider both the similar and different factors that are involved. A head-initial relative clause is unambiguously marked on its left edge with a relativizer (with the exception of a reduced relative clause like the horse (which was) raced past the barn); therefore, the cost of processing is primarily associated with filler-gap integrations. In head-final relative clauses, however, the gap precedes the filler, creating temporary ambiguity. Hence, in addition to costs on filler-gap integrations, there is also a cost on ambiguity resolution. Research focusing on issues of locality, filler-gap integration, and relativization effects in head-final relative clauses should therefore attend to the issue of temporary ambiguity and this garden-path effect. Last but not least, contexts that are used to motivate relativization should be used with caution. Even though a context provides a natural condition for the comprehension of relative clauses, they also serve as a primary source of influence on the processing of the target sentences. As relative clauses serve to increase referential coherence by picking the referents from a set of plausible candidates, an infelicitous context (such as a context that provides two distinct kinds of referents) does not motivate a relative clause. The thematic ordering inside the context also affects the comprehension of the target sentences. These properties of the contexts should all be controlled for to better understand the filler-gap effects and to ultimately theorize filler-gap integrations based on the processing of both head-initial and head-final relative clauses.
13
Garden Path and the Comprehension of Head-Final Relative Clauses
295
Acknowledgments We are grateful to participants at the International Conference on Processing Head-Final Structures at Rochester Institute of Technology, for their comments and discussions and especially to the organizers/editors and reviewers for their valuable suggestions on previous versions of this chapter. We also thank Natalie Hsu for useful discussions. Research hereby presented has been supported by research grants from the National Science Council of Taiwan (NSC 95-2411-H-003-056 & NSC 96-2411-H-003-035). Research assistance from Paul Chang, Li-Hsin Ning and Larry Hong-Lin Li is gratefully acknowledged.
References Abney, S. P. (1989). A computational model of human parsing. Journal of Psycholinguistic Research, 18, 51–60. Bader, M., & Lasser, I. (1994). German verb-final clauses and sentence processing: Evidence for immediate attachment. Hillsdale, NJ and England: Lawrence Erlbaum Associates. Bever, T. G. (1970). Cognitive basis for linguistic structures. In J. R. Hayes (Ed.), Cognition and the development of language (pp. 279–362). New York: Wiley. Chiu, B. (1995). An object clitic projection in Mandarin Chinese. Journal of East Asian Linguistics, 4, 77–117. Cohen, L., & Mehler, J. (1996). Click monitoring revisited: An on-line study of sentence comprehension Memory and Cognition, 24, 94–102. Crain, S., & Steedman, M. (1985). On not being led up the garden path: The use of context by the psychological parser. In D. R. Dowty, L. Karttunen & A. M. Zwicky (Eds.), Natural language parsing: Psychological, computational, and theoretical perspectives (pp. 320–358). New York: Cambridge University Press. Ferreira, F. (2003). The misinterpretation of noncanonical sentences. Cognitive Psychology, 47, 164–203. Ford, M. (1983). A method for obtaining measures of local parsing complexity throughout sentences. Journal of Verbal Learning and Verbal Behavior, 22, 203–218. Frauenfelder, U., Segui, J., & Mehler, J. (1980). Monitoring around the relative clause. Journal of Verbal Learning and Verbal Behavior, 19, 328–337. Frazier, L. (1987a). Sentence processing: A tutorial review. In M. Coltheart (Ed.), Attention and performance XII: The psychology of reading. Hillsdale, NJ: Lawrence Erlbaum Associates. Frazier, L. (1987b). Syntactic processing: Evidence from Dutch. Natural Language and Linguistic Theory, 5, 519–559. Frazier, L., & d’Arcais, F. G. (1989). Filler-driven parsing: A study of gap-filling in Dutch. Journal of Memory and Language, 28, 331–344. Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68, 1–76. Gibson, E., Desmet, T., Grodner, D., Watson, D., & Ko, K. (2005). Reading relative clauses in English. Cognitive Linguistics, 16, 313–353. Givo´n, T. (1993). English grammar: A function-based introduction (Vol. 2). Amsterdam, The Netherlands and Philadelphia, PA.: John Benjamins. Gouvea, A. C. (2003). Processing syntactic complexity: Cross-linguistic differences and ERP evidence. College Park, MD: University of Maryland. Hirose, Y. (2006). Processing relative clauses in Japanese: Coping with multiple ambiguities. In M. Nakayama, R. Mazuka & Y. Shirai (Eds.), The handbook of East Asian psycholinguistics, Volume II: Japanese (pp. 264–269). Cambridge: Cambridge University Press. Holmes, V. M., & O’Regan, J. K. (1981). Eye fixation patterns during the reading of relativeclause sentences. Journal of Verbal Learning and Verbal Behavior, 20, 417–430. Hsiao, F., & Gibson, E. (2003). Processing relative clauses in Chinese. Cognition, 90, 3–27.
296
C.-J.C. Lin and T.G. Bever
Hsu, C.-C. N., & Chen, J.-Y. (2007). A new look at the subject-object asymmetry: The effects of linear distance and structural distance on the processing of head-final relative clauses in Chinese. Paper presented at the Interdisciplinary Approaches to Relative Clauses (REL07), Cambridge University, Cambridge, UK. Hsu, C.-C. N., Hurewitz, F., & Phillips, C. (2006). Contextual and syntactic cues for processing head-final relative clauses in Chinese. Paper presented in the 19th Annual CUNY Conference on Human Sentence Processing, New York: City University of New York. Hsu, C,-C. N., Phillips, C., & Yoshida, M. (2005). Cues for head-final relative clauses in Chinese. Poster presented at the 18th Annual CUNY Conference on Human Sentence Processing. Tucson, AZ: University of Arizona. Ishizuka, T. (2005). Processing relative clauses in Japanese. UCLA working papers in Psycholinguistics, 2, 135–157. Ishizuka, T., Nakatani, K., & Gibson, E. (2006). Processing Japanese relative clauses in context. Paper presented at the 19th Annual CUNY Conference on Human Sentence Processing, CUNY Graduate Center, New York. Kang, S., & Speer, S. R. (2003). Prosodic disambiguation of syntactic clause boundaries in Korean. In G. Garding & M. Tsujimura (Eds.), Proceedings of the 22th West Coast Conference on Formal Linguistics. Somerville, MA: Cascadilla. Kang, S., Speer, S. R., & Nakayama. M. (2004). Effects of prosodic boundaries on ambiguous syntactic clause boundaries in Japanese. Paper presented at the 8th International Conference on Spoken Language Processing Proceedings, Jeju, Korea. King, J., & Just, M. A. (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language, 30, 580–602. King, J. W., & Kutas, M. (1995). Who did what and when? Using word- and clause-level ERPs to monitor working memory usage in reading. Journal of Cognitive Neuroscience, 7, 376–395. Kwon, N., Polinsky, M., & Kluender, R. (2004). Processing of relative clause sentences in Korean. Poster presented at Architectures and Mechanisms for Language Processing (AMLaP). Kwon, N., Polinsky, M., & Kluender, R. (2006). Subject preference in Korean. In D. Baumer, D. Montero & M. Scanlon (Eds.), Proceedings of the 25th West Coast Conference on Formal Linguistics (pp. 1–14). Somerville, MA: Cascadilla Proceedings Project. Lin, J-W. (2003). On restrictive and non-restrictive relative clauses in Mandarin Chinese. Tsinghua Journal of Chinese Studies, New Series, 33, 199–240. Lin, C.-J. C. (2006). Grammar and parsing: A typological investigation of relative-clause processing. Doctoral dissertation, The University of Arizona, Tucson, AZ. Lin, C.-J. C. (2008). Distinguishing linguistic and processing explanations of grammar. Manuscript. National Taiwan Normal University, Taipei. Lin, C.-J. C. (2009a). Chinese relative clauses in corpus: Processing considerations. Paper presented at the 2009 International Conference on Applied Linguistics & Language Teaching (ALLT), Taipei, April 16–18, 2009. Lin, C.-J. C. (2009b). Thematic patterns and comprehending Chinese relative clauses in context. Talk given at Southern Taiwan Psycholinguistic Circle, National Cheng-Kung University, Tainan, June 5, 2009. Lin, C.-J. C., & Bever, T. G. (2006). Subject preference in the processing of relative clauses in Chinese. In D. Baumer, D. Montero & M. Scanlon (Eds.), Proceedings of the 25th West Coast Conference on Formal Linguistics (pp. 254–260). Somerville, MA: Cascadilla Proceedings Project. Lin, C.-J. C., & Bever, T. G. (2007). Processing doubly-embedded head-final relative clauses. Poster presented at Interdisciplinary Approaches to Relative Clauses (REL07), Cambridge University, Cambridge, UK. Lin, Y., & Garnsey, S. M. (this volume). Animacy and the resolution of temporary ambiguity in relative clause comprehension in Mandarin. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Berlin: Springer.
13
Garden Path and the Comprehension of Head-Final Relative Clauses
297
Mak, W. M., Vonk, W., & Schriefers, H. (2002). The influence of animacy on relative clause processing. Journal of Memory and Language, 47, 50–68. Mecklinger, A., Schriefers, H., Steinhauer, K., & Friederici, A. D. (1995). Processing relative clauses varying on syntactic and semantic dimensions: An analysis with event-related potentials. Memory and Cognition, 23, 477–494. Miyamoto, E. T. (2002). Case markers as clause boundary inducers in Japanese. Journal of Psycholinguistic Research, 31(4), 307–347. Miyamoto, E. T., & Nakamura, M. (2003). Subject/object asymmetries in the processing of relative clauses in Japanese. In G. Garding & M. Tsujimura (Eds.), Proceedings of the 22nd west coast conference on formal linguistics (pp. 342–355). Somerville, MA: Cascadilla Press. Ng, S., & Fodor, J. D. (this volume). Use your headedness: An exercise in psycholinguistic exploitation. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Berlin: Springer. O’Grady, W. (1997). Syntactic development. Chicago, IL: The University of Chicago Press. Packard, J., Ye, Z., & Zhou, X. (this volume). Filler-gap processing in Mandarin relative clauses: Evidence from event-related potentials. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. Pritchett, B. L. (1991). Head position and parsing ambiguity. Journal of Psycholinguistic Research, 20, 251–270. Pritchett, B. L. (1992). Grammatical competence and parsing performance. Chicago, IL: The University of Chicago Press. Qiao, X., & Forster, K. I. (2008). Object relatives ARE easier than subject relatives in Chinese. Poster presented at the 14th Annual Conference on Architectures and Mechanisms for Language Processing (AMLaP). Schriefers, H., Friederici, A. D., & Kuhn, K. (1995). The processing of locally ambiguous relative clauses in German. Journal of Memory and Language, 34, 499–520. Ting, J. (2003). The nature of the particle suo in Mandarin Chinese. Journal of East Asian Linguistics, 12, 121–139. Townsend, D. J., & Bever, T. G. (2001). Sentence comprehension: The integration of habits and rules. Cambridge, MA: MIT Press. Traxler, M. J., Morris, R. K., & Seely, R. E. (2002). Processing subject and object relative clauses: Evidence from eye movements. Journal of Memory and Language, 47, 69–90. Ueno, M., & Garnsey, S. M. 2008. An ERP study of the processing of subject and object relative clauses in Japanese. Language and Cognitive Processes, 23, 646–688. Wu, H. I., & Gibson, E. 2008. Processing Chinese relative clauses in context. Poster presented at the 21th Annual CUNY Conference on Human Sentence Processing, University of North Carolina, Chapel Hill. Wu, F., Kaiser, E., & Andersen, E. (this volume). Subject preference, head animacy, and lexical cues: A corpus study of relative clauses in Chinese. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. Yoshida, M., Aoshima, S., & Phillips, C. (2004). Relative clause prediction in Japanese. Paper presented at the 17th Annual CUNY Conference on Human Sentence Processing, College Park, MD.
Chapter 14
Use Your Headedness: An Exercise in Psycholinguistic Exploitation1 Shukhan Ng and Janet Dean Fodor
14.1 Introduction For the parsing mechanism, differences in the position of the head in a syntactic construction affect the flow of information through a sentence, creating different temporary ambiguities and fluctuations in working memory load. Headedness differences thereby enrich the range of issues that psycholinguistics is able to explore. They provide, most obviously, the opportunity to investigate the specific contributions of heads in syntactic processing: do heads deliver more information than modifiers and complements do; are they the engine that drives phrasal integration; and so forth. But also, headedness differences between languages can create novel structures and word orders that permit other important issues to be studied, issues that could not be addressed otherwise. The work we report here takes advantage of the mixed headedness of Mandarin Chinese, in order to investigate how the human parser trades off effort between phrase structure building and empty category interpretation.2 Can there be competition between these? If so, which one wins? Chinese is mixed-headed in a somewhat unusual fashion. It is subject-initial and verb-initial, so it has canonical SVO word order as English and many languages do. But it has strictly head-final noun phrases, which is characteristic of consistently head-final languages such as Japanese and Hindi. Nominal modifiers typically precede the head noun, with de, a nominal modifier marker, between the two. S. Ng (*) Department of Biology, University of Texas, San Antonio, TX 78249-1644, USA e-mail: [email protected] 1
We are very grateful to Eva Ferna´ndez and Marcel den Dikken for their helpful advice on this project. We have also benefited from the assistance of Li Ma and Chi-Chen Bredeche, and from discussion at a Psycholinguistics Supper meeting at CUNY Graduate Center and at the International Conference on Processing Head-final Structures at the Rochester Institute of Technology. 2 Other topics have been investigated in the growing number of experimental studies of parsing Chinese. We cannot summarize all such research here but see especially Chapters 9, 11, 12 and 13 of this volume and references therein.
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_14, Ó Springer ScienceþBusiness Media B.V. 2011
299
300
S. Ng and J.D. Fodor
This is the case for all types of nominal modifiers: adjectives, complex adjectival phrases, genitives and relative clauses. For example, the English noun phrase the senator that Jim admired is expressed in Chinese with the word order: Jim admired ei de senatori. We do not discuss here the headedness of other constituents such as prepositional or adverbial phrases, where the facts are more complex. Head-finality of NPs combines with other facts in Chinese that are typical of head-final languages, to create an array of on-line ambiguities. Chinese has no relative pronouns, no complementizers (unless de counts as a complementizer), and few left-edge clause boundary markers. Hence, the English sentence That Mary won surprised Tim is expressed as ‘‘Mary won surprised Tim’’ in Chinese, with no indication on-line that the initial clause is a subject complement rather than the main clause. The same is true for clausal objects; English has both John knows Peter lied and John knows that Peter lied, where the complementizer in the latter version could cue the clausal object structure, but Chinese has only the former, which is temporarily ambiguous on-line. Another construction that appears in the input without warning is a relative clause, which could easily be taken as a main clause until the marker de is encountered. Moreover, even when de is processed, it is not a reliable indicator of the presence of a relative clause. Since de is a nominal modifier marker for many types of modifiers, it is often unclear how many of the phrases that precede de should be taken as comprising the modifier. Experiment 2 discussed below takes advantage of this modifier left-edge ambiguity.3 These various facts conspire to create considerable structural ambiguity in on-line processing of Chinese. Among other things, a clause that has been identified in the input string may subsequently turn out to be a main clause, a complement clause, a relative clause, or even possibly an adverbial clause. Experiment 1 takes advantage of this clause-type ambiguity. These phrase structure characteristics of Chinese interact in interesting ways with the distribution of Chinese empty categories. Chinese is a null subject and null object language: the subject or the object (or indeed any argument or adjunct) can be omitted (i.e., can be pro) if its interpretation is recoverable from the discourse context. In addition, Chinese has the usual array of empty categories in other clause types. It has traces of movement in relative clauses and topicalization constructions, and it can have PRO (or pro) subjects in complement clauses.4 As a result, even when an EC has been detected in the 3
Lin and Bever (this volume) indicate several potential cues for the left edge of a relative clause in Chinese, noting that they are not always present. 4 There has been considerable discussion concerning the type of empty subject in a complement clause in Chinese. (See among others Huang 1989; Xu 1986.) Because Chinese has no overt tense marking, it is not easy to distinguish a finite clause from a non-finite clause (Hu, Pan, & Xu, 2001), and hence pro from PRO. While it is important to some studies to determine whether an empty category behaves as PRO or pro, that distinction is not crucial to our experiments because it has little or no impact on how the how the parser interprets the empty category or what expectations that gives rise to for the rest of the sentence. For simplicity here, and without prejudging the linguistic issues, we treat the distribution of PRO as analogous to English or Italian.
14
Use Your Headedness: An Exercise in Psycholinguistic Exploitation
301
word string, Chinese exhibits additional on-line ambiguities concerning its status: is it pro, PRO or trace? These ambiguities of EC-type are largely aligned with the phrase structure ambiguities described above. A pro can be in a main clause; a PRO can be in a complement clause; a trace can be in a relative clause. Thus, the parser might in principle make its ambiguity-resolution choices on either basis: it might choose to interpret an empty category as PRO and conclude that the clause is a complement clause; or it might choose to analyze the clause as a complement clause and conclude that its empty category must be PRO. Similar bi-directional inference could occur for pro and main clauses, and for trace and relative clauses: the parser might infer the structure from the EC, or the EC from the structure, or perhaps it might make an integral judgment based on both. It is known that the human parsing mechanism has its preferences in ambiguity resolution. For phrase structure ambiguities these preferences have been widely studied. However, to the best of our knowledge there has been no investigation of preferences for one type of EC over another, nor of how structural and EC preferences might interact. Of particular interest are constructions where the parser’s principles would favor one analysis of an ambiguous construction because it optimizes the phrase structure but would favor a different analysis on the basis of optimizing EC interpretation. For instance, Minimal Attachment (a structural economy principle) favors main clause analyses over subordinate clause analyses, but this may clash with the needs of EC interpretation. If a clause with an empty subject occurs in a discourse context that offers no suitable discourse antecedent for a pro, that might tip the balance towards preferring a complement clause analysis with a PRO subject instead, since a PRO could be controlled by a noun phrase later in the sentence. (This is so in Chinese, as in English examples like To lose the race would upset Sally). The experiments reported below explore what happens when phrase structure principles and EC-interpretation principles are thus pitted against each other. Does one set of interests invariably win, or is there a trade-off between them?
14.2 Experiment 1 Experiment 1 employed a written sentence completion task to establish the preferred analysis of an initial clause with an empty subject. Each fragment to be completed consisted of a single verb, such as zhengli (‘‘tidy’’), and participants were instructed to add words to create a complete sentence, in any way they wished. How the sentence is completed reveals what analysis the parser imposed on the presented fragment. Three different ways in which such a fragment can be completed are illustrated in (1).
302
(1)
S. Ng and J.D. Fodor
Fragment: Zhengli. . . a. e Zhengli fangjian. tidy room ‘He/she/they tidied the room.’ b. [e Zhengli tushu] (dui Lifang) hen zhongyao. tidy book (to Lifang) very important ‘Tidying the books is important (to Lifang).’ c. [ei Zhengli bangongshi] de mishui huanying Lili. tidy office DE secretary welcome Lili ‘The secretary who tidied the office welcomed Lili.’
(1a) shows a completion as a main clause; hence the empty subject is pro, whose referent should be supplied by the preceding discourse. (1b) shows a completion as a clausal subject, with empty subject PRO. The PRO can acquire its referent by association with a noun phrase (such as Lifang in (1b)) in the following matrix clause, an instance of non-obligatory control (Hornstein, 1999; Landau, 2001). Alternatively, in (1b) if Lifang is not present, PRO may be locally assigned a generic referent (referring to people in general, not excluding the speaker; see Egerland, 2003). In the possible continuation (1c), the empty subject is construed as a trace in a relative clause that modifies the subject of the main clause. The trace is necessarily bound by the head noun mishu (‘‘secretary’’) that immediately follows the relative clause. The experiment was run in two versions: without any context in Experiment 1A, and with a preceding context in Experiment 1B. The method was the same in both; only the materials differed, in ways described below.
14.2.1 Experiment 1A (Without Context) 14.2.1.1 Materials In the without-context condition, four sentence forms were tested, as illustrated in (2). (2)
a. V:
Zhengli. . . tidy
b. SV:
Mali zhengli. . . Mary tidy
c. AdvV:
Changchang zhengli. . . often tidy
14
Use Your Headedness: An Exercise in Psycholinguistic Exploitation
d. QV:
303
Daduoshu ren zhengli. . . most people tidy
Subjectless fragments like (2a) were compared with similar fragments with an overt subject as in (2b). The latter are ambiguous in similar ways to the subjectless fragments: they could be continued as a main clause, or as a clausal subject, or as a relative clause. Comparison of the ways in which participants continue the (2a) fragments versus the (2b) fragments can reveal whether and how the presence of an empty category affects structure assignment. The two other sentence forms included in this experiment served to control for the fact that the fragments (2a/b) differed not only with respect to the presence of an empty category but also in their length and informativeness. In the absence of such a control it might be suggested that participants would tend to complete the shorter fragment (2a) with a more complex structure in order to compensate for it being shorter and/or less informative than (2b). The possible role of these factors can be checked by reference to (2c) and (2d): AdvV and QV. This pair also contrasts the presence versus absence of an empty subject, but the overt subject in (2d) is a quantified phrase (daduoshu ren ‘‘most people’’) and the empty subject version (2c) contains an adverb (changchang ‘‘often’’) that carries very similar information. A difference in the completion patterns for (2a,b) can safely be attributed to the empty category, rather than to information load, if a comparable difference is found for (2c,d). A possible effect of fragment length can also be checked since the length in characters progressively increases (from 2–6) from version (2a) through (2d).5 The materials included 24 target fragments each with a V version (empty subject) and an SV version (overt subject), plus another 24 target fragments, each with an AdvV (empty subject with Adverb) version and a QV (quantified overt subject) version. All verbs were transitive. Two counterbalanced questionnaires were created, each with 48 target fragments intermingled with 24 filler fragments of varied construction and length. No participant saw the same verb twice.6 14.2.1.2 Participants and Procedure Twenty-six native speakers of Mandarin Chinese (living in New York but with primary and secondary education in mainland China, 20 female, age 18–23, mean age 20) participated in the study in return for course credit or modest monetary compensation. They were instructed to continue the fragments in writing in any way they wished, to make a complete and well-formed sentence. A practice session preceded the experiment. 5
The individuation of words is not a simple matter in Chinese, so length in terms of characters is a more practical – and possibly more relevant – measure here. 6 Due to an oversight, the verb zhengli ‘‘tidy’’ was used twice in both lists.
304
S. Ng and J.D. Fodor
14.2.1.3 Predictions A Chinese reader encountering a sentence beginning with zhengli as in (2a) faces uncertainty both with respect to its structure and with respect to the interpretation of its empty category. As observed above, these ambiguities are not independent of each other and the preferences for one interpretation or the other may clash. The materials for this experiment were designed (on the basis of pilot data) to create a potential conflict between optimization of the structural factor and optimization of the EC-interpretation factor. The question of interest is whether the structural preferences or the EC-interpretation preferences will dominate the parser’s analysis of the fragment. If the parser analyzes a fragment like (2a) as the start of a main clause, then the EC is pro and it needs a discourse antecedent to provide its referent. (The empty subject of a main clause cannot usually have a generic referent.) No discourse antecedent is present in this without-context version of the experiment, so in this respect the main clause analysis is not optimal, even though it would be preferred on structural grounds. A reader might of course conjure up a discourse referent to serve the purpose, but this mental labor could be expected to carry a processing cost (Crain & Steedman, 1985) and thus would also be less than optimal. If, instead, the parser analyzes the fragment as the beginning of a clausal subject, its EC would be PRO. In this case there are two opportunities for interpreting the empty category in a satisfactory way. As noted above, PRO could be associated with a suitable referential phrase (a ‘‘controller’’) in a following main clause, or it could be assigned a generic reference with no need for establishing a dependency with any other element. Thus, if the parser does not always give priority to optimizing structure building, it might well favor a clausal subject analysis with PRO. Finally, if the fragment is parsed as the start of a relative clause, the structure building cost (in terms of number of syntactic nodes to be created) will be greater than for a main clause or even a clausal subject, since the relative clause has an additional NP node above the clause node. (Note that if the parser opts for the relative clause analysis, this extra node has to be inserted right away, at the beginning of the clause.) On the other hand, possibly offsetting this structural penalty, on the relative clause analysis the continuation of the sentence will necessarily supply a head noun for the relative clause to modify, which will provide the interpretation for the EC. Our predictions are thus as follows. On grounds of structural economy alone, the main clause analysis should be the most preferred, followed by the clausal subject analysis, with the relative clause analysis least preferred. This would be expected for the fragments with an overt subject. It would also apply to fragments with an empty subject if structural economy is the sole or primary factor driving ambiguity resolution. On the other hand, if EC-interpretation is the more powerful factor in ambiguity resolution, the pattern of completions for the fragments with empty subjects will differ from those with overt subjects. For the fragments containing an EC, the main clause analysis will be dispreferred for lack of any discourse context that could supply an antecedent. The
14
Use Your Headedness: An Exercise in Psycholinguistic Exploitation
305
parser will instead opt for either the clausal subject analysis or the relative clause analysis, despite their greater structure-building costs. Between these two subordinate clause types, the RC structure is more costly than the clausal subject analysis with respect to structural economy, which might sway the choice. It could be argued that the relative clause is also less preferred in respect of EC processing because it violates the Minimal Chain Principle (‘‘Avoid postulating unnecessary chain members. . .’’; de Vincenzi, 1989), which advocates not starting to build a syntactic chain unless it is necessary to do so. The association of a trace with the head noun of a relative clause construction constitutes a syntactic chain (Simpson, 2002), while the relation between a PRO in a clausal subject and a (non-obligatory) controller in the main clause is a semantic/pragmatic association, not a syntactic chain (Landau, 2001). Thus, since it is arguably nonoptimal on both EC and structural grounds, it seems reasonable to predict that the relative clause structure will be the least preferred of the three. In short: the goal of minimizing structural complexity competes with the ease of empty category interpretation in the fragments with no overt subject. The proportions of different completions provided by the participants can reveal which processing factor was given higher priority during the processing of the fragments.
14.2.1.4 Results Figure 14.1 shows the results for Experiment 1A. It is clear from the figure that the fragments with overt subjects (SV and QV) were treated similarly, as were the fragments with empty subjects (V and AdvV). Main clause continuations predominated overall, with the primary competitor being clausal subject continuations; the relative clause construction accounted for a negligible proportion of the data (4%), as did other continuations (such as taking the subjectless fragment as a noun or as the start of an adjunct clause, 7%).
Continuations (%)
100 80 Main Clause
60
Clausal Subject Relative Clause
40
Others
20 0 V
SV
AdvV
QV
Fragment type
Fig. 14.1 Results of the without-context sentence completion task
306
S. Ng and J.D. Fodor
The percentage of main clause continuations was used as a measure of the preference for simpler phrase structure. Analyses of variance performed on percentages of main clause continuations included the factors Subject Type (empty versus overt) and Informativeness (imbalanced versus balanced). The Subject Type manipulation was hugely reliable (F1 (1,24) = 160.86, p < 0.001; F2 ð1; 44Þ ¼ 251:10; p50:001Þ; fragments with overt subjects were completed with main clause continuations at an average rate of 90%, while fragments with empty subjects were completed with main clause continuations at an average rate of 38%. By contrast, the Informativeness manipulation showed no main effect ðF1 ð1; 24Þ ¼ 3:14; p40:05; F2 ð1; 44Þ51Þ, and the interaction of the two factors was also not significant (F1 (1,24) = 3.43, p > 0.05; F2 (1,44) = 1.24, p40:25Þ. The data also show no evidence of an effect of fragment length in characters. Although the materials design did not allow for a precise comparison, it is clear that there is no consistent tendency to provide structurally richer continuations either for the shorter fragments (V and SV) or for the longer ones (AdvV and QV). It can be concluded that it is the empty category, not other factors, that influences the resolution of the ambiguity in the target fragments. Fragments with an overt subject were overwhelmingly assigned main clause continuations, quite predictably since a main clause analysis satisfies Minimal Attachment and for these fragments requires no EC-interpretation. For fragments with an empty subject, where structural concerns and EC concerns were pitted against each other, the two influences were fairly evenly balanced in their effects, but the EC factor prevailed by a small margin. For the V fragment, there were more continuations as a clausal subject than as a main clause. Out of the total of main clause plus clausal subject continuations, only 44% were main clauses, which is reliably less than chance ðt1 ð25Þ ¼ 2:17; p50:02; t2 ð23Þ ¼ 2:98; p50:005Þ. As predicted, the relative clause construction, lacking merit in either respect, was rarely adopted.
14.2.2 Experiment 1B (With Preceding Context) The with-context version of the sentence completion study was designed to confirm whether the absence of a readily available discourse referent for the empty category was the primary reason for the relatively high proportion of bi-clausal structures produced for fragments with an empty subject in Experiment 1A.
14.2.2.1 Materials, Procedure and Participants All the sentence fragments in Experiment 1B consisted of two clauses. The second clause was one of the V fragments (like (2a), with no overt subject) from Experiment 1A. The first clause was an adverbial clause that either did or
14
Use Your Headedness: An Exercise in Psycholinguistic Exploitation
307
did not provide a plausible antecedent for the empty category, as in (3a) and (3b) respectively. (3)
a. Plausible antecedent: Ritian zhaodao xin-de zhengju yihou, tuifan. . . Ritian found new evidence after reject ‘After Ritian found new evidence, e rejected. . .’ b. No plausible antecedent: Shiyan shibai zhihou, tuifan. . . experiment fail after reject ‘After the experiment failed, e rejected. . .’
The first clause of (3a) mentions Ritian, thus satisfying the EC’s need for a discourse referent on the main clause analysis. The first clause of (3b) is not helpful in this respect. Although it does offer some contextualization, it does not provide a referent for the EC any more than the without-context materials of Experiment 1A did. Twenty-four test items were created in two versions: with or without a potential referent in the first clause. They were intermingled with 24 filler sentences as in Experiment 1A, and presented for completion in counterbalanced lists. Twenty-six native speakers of Mandarin Chinese drawn from the same population as for Experiment 1A completed the task.7 The procedure was the same as for Experiment 1A. 14.2.2.2 Predictions If the lack of an antecedent for pro in Experiment 1A was indeed the sole or primary cause of the significant decrease in main clause continuations for fragments with empty subjects compared with fragments with overt subjects, then in the present experiment we should find that the items like (3a), with an empty subject preceded by a good antecedent, will behave like the items with overt subjects in the previous experiment. The items like (3b), with an empty subject but no suitable antecedent in the preceding clause, should behave like the items with empty subjects in the previous experiment. 14.2.2.3 Results Figure 14.2 shows the percentage of continuations of each kind for the two fragment types in Experiment 1B. 7
Some participants in Experiment 1B (evenly distributed across the groups) had participated previously in Experiment 1A, with an interval of at least 3 weeks between the two sessions. The materials are sufficiently different that it is unlikely that there was any carryover in their responses between the two experiments.
308
S. Ng and J.D. Fodor
Continuations (%)
100 80 Main Clause
60
Clausal Subject Relative Clause
40
Others
20 0
Plausible antecedent
No plausible antecedent
Fragment type
Fig. 14.2 Results of the with-context sentence completion task
Once again, RC continuations (9%) and other continuations (4%) were negligible. The main contenders were main clause and clausal subject continuations. Analyses were performed on percent main clause continuations, as before. As is evident in Fig. 14.2, fragments with a preceding plausible antecedent for the EC were almost always continued as a main clause (93%), while only half of the fragments without a plausible antecedent had main clause completions (49%). This difference was highly significant in a one-way analysis of variance with Antecedent Type as the single factor ðF1 ð1; 24Þ ¼ 96:94; p < 0.001; F2 (1,22) = 59.75, p < 0.001). This pattern of results in Experiment 1B, where the target fragments differed only in their preambles, is very similar to that observed with the SV and V materials in Experiment 1A. When there is an overt subject (Experiment 1A), or when there is an empty subject with a plausible antecedent (Experiment 1B), Minimal Attachment applies with no hindrance from EC-interpretation, so main clause continuations prevail. When there is no discourse context for an EC (Experiment 1A), or when there is a context but it provides no plausible antecedent (Experiment 1B), clausal subject continuations greatly increase, from 2 to 31% in 1B. The results of Experiment 1B are thus as predicted by the hypothesis that it was the failure of EC-interpretation that was responsible for the overriding of Minimal Attachment in Experiment 1A. A salient difference between the two experiments concerns the balance between main clause and clausal subject continuations for the items whose EC lacks a good antecedent (V in Experiment 1A, and the ‘‘no plausible antecedent’’ condition in Experiment 1B). In Experiment 1A, the usual preponderance of main clause continuations was reversed for the V fragments, as observed above. Main clause continuations in Experiment 1B, though much reduced for the fragments without a plausible antecedent, were still the preferred response. There were more continuations as a main clause than as a clausal subject; out of the total of main clause plus clausal subject
14
Use Your Headedness: An Exercise in Psycholinguistic Exploitation
309
continuations, 62% were main clauses, which is reliably different from chance ðt1 ð25Þ ¼ 1:95; p50:05; t2 ð23Þ ¼ 1:96; p50:05Þ. This difference between the two studies suggests that the participants in Experiment 1B may have found it easier than those in Experiment 1A to conjure up a situation in which there was a likely referent for pro, even though no specific individual was mentioned in the 1B context. More generally we may speculate that any contextualization may facilitate interpretation of a discourse-dependent item such as a phonologically empty pronoun, though an explicit antecedent has a greater effect.
14.2.3 General Discussion of Experiment 1 Together, the sentence completion results for these constructions reveal a powerful effect of the need to assign an antecedent to pro. Even in the clearly artificial situation of the experiment, under conditions in which the participants knew they were being presented with a sequence of entirely unrelated sentences out of context, they evidently found it difficult to switch out the grammatical requirement that a pro should have an identifiable referent in the preceding discourse. Nevertheless, they did do so sometimes, when the alternative was to adopt a structurally more complex analysis of the word string. Thus, EC-interpretation is indeed a powerful factor but it is not all-powerful: it did not completely eliminate all main-clause continuations. Instead, we see something more like an accommodation between these two factors. It seems that the parser is able to strike a balance between satisfying structural economy and optimizing EC-interpretation. This is a novel finding, which challenges the hegemony of structure-building concerns and possibly even points toward a non-modular processing regime – depending on whether EC-interpretation is regarded as properly a syntactic or a semantic matter. Resolving issues of modularity has never been an easy matter and we will not aim to do so here. Nevertheless we can move closer to an understanding of how these different influences on ambiguity resolution interact, by shifting now to a self-paced reading paradigm that provides some time-course information.
14.3 Experiment 2 The findings of Experiment 1 have shown that when facing a structural ambiguity, the processor is sometimes willing to sacrifice Minimal Attachment in order to secure a suitable referent for the empty category. Sentence completion methodology is an excellent tool for an initial exploration of a previously unstudied phenomenon. However, it is possible that the completion task unduly encourages the comprehension routines to
310
S. Ng and J.D. Fodor
complete their work on the presented fragment before sending it to the production routines for continuation. It might be argued that this alters the normal balance of structural and interpretive operations. In Experiment 2, therefore, we employed a self-paced reading paradigm, in which the parsing system is free to make its decisions at whatever moment in the parse it elects to. Continuing to explore the interplay of structural and EC-interpretation pressures, our specific aim in this study was to replicate, with a different set of constructions, the finding that EC concerns may sometimes override structural economy concerns. Experiment 1 showed that under pressure for EC-interpretation, a subordinate clause can sometimes be preferred over a matrix clause, contrary to the standard findings for structures in which ECs are not involved. In Experiment 2 we seek to find out whether, under pressure for EC-interpretation, a relative clause can sometimes be preferred over a complement clause, contrary to the usual findings. It is this experiment which most directly exploits the special features of Chinese sentence structure, particularly its mixed-headedness. The fact that its noun phrases are strictly head-final creates a temporary ambiguity between relative clauses and clausal objects, since (unlike in English) a relative clause may directly follow a higher verb, as a clausal object does. With no left-edge clause marker to differentiate them, the parser may not know the nature of a post-verbal clause until several words later. The fact that the verb phrase in Chinese is head-initial (unlike a strictly head-final language such as Japanese) means that the possible interpretation of an empty subject in a post-verbal subordinate clause can be contextually manipulated by the choice of matrix verb. As will be clear below, this is essential to the design of Experiment 2. The comparison between clausal objects and relative clauses is well selected for purposes of our investigations. In Experiment 1 we compared main clause versus various kinds of subordinate clause analyses. However, since every sentence must contain a main clause but need not contain a subordinate clause, the strong preference for a main clause analysis even in the face of EC-interpretation problems might be ascribed to its necessarily greater frequency. We know of no data on the relative frequencies of clausal objects and relative clauses in Chinese, but they are surely better matched than are main and subordinate clauses.
14.3.1 Materials The test sentences for this experiment are illustrated in (4). In each case there is a main clause followed by a subordinate clause. For the processor, the decision to be made on-line, when the subordinate clause begins, is whether to parse it as a clausal object or as a relative clause.
14
(4)
Use Your Headedness: An Exercise in Psycholinguistic Exploitation
311
a. Coref+CO: Coreferential verb + clausal object [ei tuichu zhegebisai de zongjuesai]. Jimi jujue Jim refuse withdraw this-tournament DE final ‘Jim refused to withdraw from this tournament’s final.’ b. Non-coref+CO: Non-coreferential verb + clausal object Jimi fengci [ej tuichu zhegebisai de zongjuesai]. Jim ridicule withdraw this-tournament DE final ‘Jim ridiculed (him/her) for withdrawing from this tournament’s final.’ c. Coref+RC: Coreferential verb + relative clause Jim jujue [ei tuichu zhegebisai] de yundongyuani. Jim reject8 withdraw this-tournament DE athlete ‘Jim rejected the athletes who withdrew from this tournament.’ d. Non-coref+RC: Non-coreferential verb + relative clause Jim fengci [ei tuichu zhegebisai] de yundongyuani. Jim ridicule withdraw this-tournament DE athlete ‘Jim ridiculed the athletes who withdrew from this tournament.’
Two factors are varied in these sentences. One is the type of matrix verb, which we have designated Coreferential vs. Non-coreferential and which we will explain below. The other is the type of subordinate clause (clausal object vs. relative clause). The clause type is established only by the last word of each sentence: the noun that is encountered at the end of the subordinate clause. In (4a,b) the noun zongjuesai ‘‘final’’ (as in ‘‘final round of a contest’’) forces the clausal object reading because it makes sense only as the object of the verb tuichu ‘‘withdraw’’ (a person can withdraw from a final); the relative clause analysis would be nonsensical (‘‘the final that withdrew from the tournament’’). In (4c,d) the noun yundongyuan ‘‘athlete’’ forces the relative clause analysis because it makes sense only as the subject of tuichu (‘‘the athlete that withdrew from the tournament’’, not ‘‘someone withdrew from the tournament’s athlete’’9). In general, for all target items in the experiment, the final nouns in the (a,b) versions were best suited as the object (theme) of the subordinate verb, thus implying the clausal object analysis, while the final nouns in the (c,d) versions were better suited as the subject (agent) of the subordinate verb, thus implying the relative clause analysis. The clause-type factor is crossed with the matrix verb type factor. The verb jujue in (4a,c) is fully compatible with a clausal object whose empty category is coreferential with the matrix subject (i.e., the person who does the refusing is the 8
We have translated the verb jujue into English as both ‘‘refuse’’ (in 4a) and ‘‘reject’’ in (4c), since no English verb captures its different shades of meaning in Chinese. 9 This English translation is not nonsensical if it means emotional withdrawal, but tuichu in Chinese does not have that sense. In general, it is important not to be misled by the inadequacy of the closest English translations of the Chinese lexical items.
312
S. Ng and J.D. Fodor
person who does the withdrawing). The clausal object analysis is also structurally simpler than the relative clause analysis, since the latter demands an extra NP node to unite the relative clause and the head noun, as noted above. Hence it can be anticipated that the parser will readily accept the clausal object analysis for the jujue sentences. In that case, the beginning of the subordinate clause will be easy to process in both (4a) and (4c). However, the final word of (4a), which forces the clausal object analysis, will be easier to process (other things being equal) than the final word of (4c), which forces the relative clause analysis. By contrast, the verb fengci ‘‘ridicule’’ in (4b,d) does not naturally tolerate coreference of the empty category with the matrix subject (it is unlikely that the person who does the withdrawing is the same as the person who does the ridiculing), so a different interpretation has to be assigned when the processor detects the empty category in the subordinate clause following fengci. One possibility is a clausal object analysis whose empty category refers to a salient individual in the preceding discourse, but since in this experiment there was no preceding discourse, it can be inferred from Experiment 1 that this analysis is not highly valued. It may therefore be open to competition from a relative clause analysis, in which the empty category is a trace bound by the head noun that will occur later in the sentence.10 A relative clause is dispreferred by Minimal Attachment, as we have noted. But the relative clause structure has the advantage that it needs no prior antecedent for the EC. So it is fully acceptable on-line, even when the matrix verb offers no prior referent for a PRO in versions (4b/d). Thus, which of these two analyses (clausal object versus relative clause) is adopted on-line by the processor for the fengci sentences (4b/d) will reveal how the parser prioritizes structural economy relative to felicitous EC-interpretation. The matrix verbs used in Experiment 2 were required to differ in their bias for or against coreference between the matrix subject and the empty subject of the subordinate clause. Two groups of verbs were sought that would contrast in this respect as clearly as possible. It was also required that the verbs in the two coreference-bias groups should not differ with respect to their preference for a clausal object or a non-clausal noun phrase object. For a matrix verb that more frequently has a clausal object in normal use, the clausal object analysis might be preferred in Experiment 2 regardless of the preferred interpretation of the empty category. Similarly, a verb that usually has a nominal object might lead to a nominal object analysis in Experiment 2 regardless of the extra cost of building a relative clause structure within the noun phrase. Two norming studies were conducted to check these properties of Chinese verbs and guide the selection of appropriate materials for Experiment 2. The first was a written 10
Another logical possibility is a clausal object analysis with an arb (arbitrary, generic) referent for the empty category but this is rare in Chinese in this syntactic context (Huang, 1994). There are also some three-clause structures that might be constructed here but complexity considerations suggest they would not be adopted by the parser as long as there is a simpler alternative. We will discount these possibilities in the discussion that follows.
14
Use Your Headedness: An Exercise in Psycholinguistic Exploitation
313
sentence completion study, the results of which were used to pair up matrix verbs with similar complement preferences. The second was a pencil-and-paper judgment study, which checked the (non-)coreferentiality bias of the matrix verbs to be used in the target sentences of the main study. Further details can be found in Ng (2009). One of the important target regions for measuring processing load is the last word of each sentence. As shown in many previous studies, a sentencefinal word may be subject to a ‘‘wrap-up’’ effect, in which the parser reviews and consolidates the structure and meaning of the whole sentence. As a result, reading times on the final word (or region) may reflect not only the processing load pertinent to that region but also the accumulated processing difficulty of other parts of the sentence. A standard way to mitigate this wrap-up effect at a target region of interest is to add a phrase after it, so that wrap-up effects will occur at a later point and will not interfere with measurements in the target region. This, however, is not easy to achieve in Chinese. Chinese allows only a few phrase types except for the object to occur after the verb. Thus, common tactics used for English, such as adding a sentence-final adverb after the object, are not generally feasible in Chinese. Instead, a better way to ameliorate this problem is to display the final period (in Chinese orthography the symbol ) separately from the last word. This avoids overtly signaling the end of the sentence until after the final noun position. Thus, in all sentences, the final region contained only the period. Reading times for the preceding region should therefore be relatively free of wrap-up effects, and reading times for the final region can be compared since it was fully matched across all sentence versions. One further property of the experimental materials needs to be addressed. The nearest English translations of the target sentences are reminiscent of the well-known contrast between subject-control constructions with verbs like ‘‘try’’, and object-control constructions with verbs like ‘‘persuade’’, discussed by Chomsky (1981 and elsewhere) and in many works since. In the English examples, the persuade-type construction has an object noun phrase in the matrix clause preceding the subordinate clause, as in (5b), contrasting with the lack of a matrix object in subject control examples such as (5a). (5)
a. Suei tried PROi to open the window. b. Jim persuaded Suei PROi to open the window.
The controller of PRO in (5a) is the only NP in the matrix clause, the subject. The controller of PRO in (5b) is the closest NP in the matrix clause, the object. In the Chinese sentences used in Experiment 2 there is no noun phrase following the matrix verb in either case. This is what permits the close match between the word strings in the Coref and Non-coref sentence types in the experiment, as for instance in (4a,b) above. However, the comparison with English raises the
314
S. Ng and J.D. Fodor
question of whether there might be a phonologically empty object in the main clause of the Non-coref example (4b), which controls the PRO in the subordinate clause. On that analysis the Non-coref examples would have two empty categories in a row between the matrix verb and the subordinate verb, and this might be expected to affect the processing load (see Boland, Tanenhaus, & Garnsey, 1990; Crain & Fodor, 1985). However, linguistic evidence concerning these constructions in Chinese, presented in Ng (2009), shows that for both the Coref and the Non-coref items there are two possible structures (with more or less identical meanings): one with and one without an empty matrix object. There are no experimental data that show which of these structures is preferred on-line, but there is no obvious reason to think that the verb types differ in this respect. Hence, it may be concluded that the possible presence of a matrix gap is not a factor that differentiates the Coref and Non-coref target items in the experiment.
14.3.2 Procedures The self-paced reading study was designed as an investigation of whether online empty category interpretation could influence the usually observed preference for a clausal object analysis over a relative clause analysis. Verb Type (coreferential vs. non-coreferential bias) was crossed with Clause Type (clausal object vs. relative clause) in a counterbalanced 2 2 Latin Square Design. There were four materials lists, each containing three implicit practice sentences followed by 96 items: 32 target sentences, 16 target sentences from another experiment serving as filler items here, and 48 additional fillers of varied construction. These 96 sentences were grouped into blocks. Each list had eight blocks, each block consisting of 12 sentences: four target sentences and eight fillers. The order of sentences within blocks was randomized, as was the order of blocks, for each participant individually. Sentences were presented sequentially on a computer screen. For all sentence types, the presentation regions were as shown in (6). (6)
Presentation regions:
Matrix subject / matrix verb / subordinate verb / NP / DE / Noun / PERIOD 1 2 3 4 5 6 7 A comprehension question following each sentence was presented with two possible answers: one on the left of the screen and the other on the right. The position of the correct answer was counterbalanced within blocks and within sentence types. The comprehension questions did not directly interrogate the local ambiguities of interest to this study, in order to avoid drawing too much attention to them. Thus, the accuracy rate and reaction times for question
14
Use Your Headedness: An Exercise in Psycholinguistic Exploitation
315
answering shed no light on the processing of the target regions and were intended only to provide a criterion for excluding any inattentive participants.
14.3.3 Predictions It can be anticipated that the sentences with a Coref verb will favor the clausal object analysis on encounter with the EC at the beginning of the subordinate clause. In that case it is predicted that processing will be untroubled at the verb of the subordinate clause, and that reading will be faster for the final word of sentences like (4a) than for the final word of sentences like (4c), other things being equal. By contrast, in the sentences with Non-coref verbs, a slowdown in processing is expected at the beginning of the subordinate clause, where the parser is faced with two unattractive alternatives: either conjure up a referent for the empty subject of a clausal object, or adopt the structurally more complex relative clause analysis which does not need a prior antecedent for its EC. There are no previous data on this topic on which to base a firm prediction, but the results of Experiment 1 encourage the speculation that EC-considerations may sometimes prevail in this conflict, leading to adoption of the relative clause analysis. If this is so, then a final noun that is suitable as a binder for the EC subject of the subordinate verb, as in (4d), will be easier to process than a word that is not sensible as the head noun for this relative clause to modify, as in (4b). It may be predicted, then, that reading will be faster in this experiment for the final word of the (d) version than for the final word of the (b) version. These predictions concern unmatched lexical items, so they cannot be made directly. But the four-way comparison among versions (a,b,c,d) controls for lexical differences. If a significant interaction is observed between the factors Verb Type and Clause Type, it can be inferred that there is a tendency for the relative clause analysis to be adopted on-line for the Non-coref sentences. This outcome would indicate that structural simplicity can be outweighed by optimal empty category interpretation. A significant effect of clause type that does not interact with matrix-verb type would show that structural simplicity is paramount.
14.3.4 Results Fifty-eight native speakers of Mandarin Chinese drawn from the same population as for Experiments 1A and 1B completed the task. Accuracy rates for the comprehension questions were averaged by sentence type and participant. Two participants were excluded due to an error rate of 15% on the comprehension questions, leaving 14 participants for each of the four materials lists. Their average error rate for the comprehension questions was 4.7%. Reading times for each region were trimmed by discarding times that were more than three standard deviations from the mean for that individual
316
S. Ng and J.D. Fodor
participant for that sentence region. This procedure removed 283 data points. Together with four display errors, the total number of data points discarded was 287 (2.3% of the total). An analysis of variance was then performed on the reading times for each region to identify the effects of matrix verb type and subordinate clause type. Mean reading times for all sentence regions of the four sentence types are shown in Fig. 14.3. Significant differences between sentence types were observed in three regions only: regions 2, 4 and 7. For reasons of space, data for only these and related regions will be discussed here.
Reading time (msec)
700
600
500
400 refuse/ridicule withdraw tournament
DE
final/athlete PERIOD
Coref+CO
Non-coref+CO
Coref+RC
Non-coref+RC
Fig. 14.3 Mean reaction times (in milliseconds) to all regions in Experiment 2
Region 2 (e.g., refuse/ridicule; 2–3 characters, mean 2.13): This region showed a main effect of verb type ðF1 ð1; 52Þ ¼ 18:40; p50:01; F2 ð1; 28Þ ¼ 14:16; p < 0.01): the Non-coref verbs were significantly harder to read than the Coref verbs. The two classes of verbs were distinct lexical items, so they cannot properly be compared. Thus this difference in region 2 has no implications for the issues addressed in the experiment. Region 3 (e.g., withdraw; 1–3 characters, mean 1.88): No main effects were observed. This region, with a verb immediately following the matrix verb, is where the parser can detect the beginning of a subordinate clause with an empty subject. The parser is expected to start building the subordinate clause at this point, and the experimental hypothesis is that in doing so it will incur a greater processing cost if the clausal object analysis is unavailable or non-optimal, due to lack of a referent for its PRO subject. Yet the region 3 reading times show no reliable increase in difficulty for the Non-coref items. The predicted slowdown shows up only in the following region; this is a standard phenomenon in selfpaced reading; see discussion below.
14
Use Your Headedness: An Exercise in Psycholinguistic Exploitation
317
Region 4 (e.g., tournament; 2–6 characters, mean: 4.19): There was a significant main effect of the matrix verb type ðF1 ð1; 52Þ ¼ 9:66; p50:01; F2 (1,28) = 5.11, p < 0.04) in this region, with mean reading time for Non-coref items longer by 40 ms than for Coref items. Thus by this point in the parse, the predicted differential difficulty in building the subordinate clause has appeared. It may reflect either the cost of building the more complex relative clause structure or else the cost of interpreting a clausal object construction that lacks a fully acceptable antecedent for its empty category. At this region the data do not differentiate these two possibilities, although by the final region the data support the hypothesis that a relative clause construction was often built on-line. Region 6 (e.g., final/athlete; 2–4 characters, mean: 3.06): There were no reliable main effects or interaction, though the reading times for the Noncoref+CO and Coref+RC sentences were numerically somewhat higher than those for Coref+CO and Non-coref+RC sentences. Slower reading for the Non-coref+CO and Coref+RC items was predicted for this region due to a clash between the type of clause analysis previously adopted on-line and the disambiguation by the noun in region 6 toward either the CO or the RC analysis. If a Coref sentence was preferentially parsed on-line as containing a clausal object, as suggested above, then a noun requiring an RC analysis would be an unexpected disambiguation calling for structural reanalysis. Of particular interest is the converse situation: if a Non-coref sentence was parsed as a relative clause on-line, despite its violation of Minimal Attachment, then a final noun requiring a CO analysis would be an unexpected disambiguation (a gardenpath) and would require reanalysis. That should be reflected in reading time, yielding an interaction between clause type and verb type. No such interaction was observed at region 6 (F1, F2 < 1). Once again, the predicted difference emerged only one region later in this self-paced reading task. Region 7 (e.g., PERIOD; 1 symbol): In this region containing just the period (Chinese symbol ), there were no main effects but a significant interaction of clause type and verb type ðF1 ð1; 52Þ ¼ 10:48; p50:01; F2 ð1; 28Þ ¼ 5:50; p < 0.03). Reading times were higher for Non-coref+CO and Coref+RC sentences than for Coref+CO and Non-coref+RC sentences. If effects in region 7 can be regarded as spillover from region 6, this result is fully in keeping with the prediction that a Coref matrix verb promotes a CO analysis on-line and a Non-coref matrix verb promotes an RC analysis on-line, causing surprise and reanalysis when the noun in region 6 conflicted with the on-line analysis. Incidentally, it may be noted that the reading times for the Non-coref+CO and Coref+RC items are strikingly similar. This would have to be just a coincidence if region 7 is a pure spillover site (i.e., if region 7 is where the processing of region 6 is registered), because the lexical items in region 6 are unmatched. However, if it is only the syntactic integration cost that spills over to the next region, the similarity in reading times in region 7 could carry the interesting suggestion that reanalysis from a relative clause to a clausal object is no less (and no more) arduous than reanalysis from a clausal object to a relative clause.
318
S. Ng and J.D. Fodor
While it is plausible to ascribe the interaction at the period in region 7 to spillover of integration processing from region 6, it is also true that a sentencefinal period could trigger end-of-sentence wrap-up processing, in which the overall structure and meaning of the sentence are reviewed and evaluated. However, this possibility poses no problem to the interpretation of the data proposed here. Note that at earlier regions the sentence types that show difficulty are the two Non-coref sentence types, but this difficulty is not evident at all in the pattern of data for the end-of-sentence region. At the end of the sentence, the sentence types group quite differently, with difficulty evident for the Noncoref+CO and Coref+RC versions relative to the others. Thus it appears that the cause of the interaction at the period is not wrap-up processing but is specifically the disambiguation in region 6. At the very least, it is clear that prior patterns of difficulty encountered at Regions 2 and 4 are substantially canceled out at Region 7 by garden-path effects revealing that the wrong analysis was computed on-line. Thus the conclusion still follows: it is not always the case that a relative clause analysis is less preferred than a clausal subject analysis in on-line sentence processing. The presence of an empty category in need of interpretation evidently can override the usual preference for structural simplicity. A methodological note: The self-paced reading paradigm is commonly used to provide insight into incremental processing on-line. But we have seen here a partial asynchrony between the processing events that would be triggered by the currently presented word, and the reading slowdown that those events could be expected to cause. This varies in interesting ways with the nature of the parsing difficulty. The lexical effect observed in Region 2 was local in the sense that it was evident immediately in the reading times for the lexical items that differed, whereas syntactic processing effects showed up in the region following the triggering word. For example, the subordinate verb in Region 3 was the indicator that an empty category was present in the subordinate clause, yet the importantly different implications of that empty category for the sentences with Coref verbs and sentences with Non-coref verbs did not show up until region 4. Also, the sentence-final noun in region 6 was what indicated whether the subordinate clause was a clausal object or a relative clause, yet the very different effects of that noun, depending on whether it confirmed or disconfirmed the parser’s on-line analysis, did not show up until the period in region 7. We see no problem here for our interpretation of the data, for three reasons. One is the closeness of fit between the syntactic effects that were predicted and those that were observed, albeit one region later. The second is that there is no alternative explanation, based on the words in the regions where the effects were observed (regions 4 and 7), since the four sentence versions were all identical at those regions. Finally, the brief delay that we observed in the responsiveness of participants’ reading times to the triggering stimulus has often been noted in the self-paced reading task (e.g., Mitchell, 1984, 2004; Vasishth, this volume; Yamashita, Stowe, & Nakayama, 1993). What needs to be pinned down in future research is whether the delay is specific to syntactic integration
14
Use Your Headedness: An Exercise in Psycholinguistic Exploitation
319
phenomena or is also characteristic of semantic or pragmatic processing and/or other aspects of the stimuli.
14.4 General Discussion Although the data we have presented here pertain to certain specific constructions in Chinese, they suggest two general conclusions, one of which is methodological and the other more substantive. The methodological conclusion is that there is much to be learned by exploiting the interesting structural configurations that occur in a mixed-headed language when a head-final structure meets a head-initial structure, or vice versa. The early position of the main verb in the Experiment 2 materials made it possible to bias the processing of the EC in the clause that followed, while the late position of the noun that disambiguates a relative clause from a clausal object made it possible to find out which clause type is preferred in case of ambiguity. See also Bader (this volume) for processing issues that arise in German where mixed headedness creates branching direction changes in the tree structure. Other novel configurations in mixed branching languages can be expected to open up more topics for future investigation. The substantive conclusion is that empty category interpretation can exert a considerable influence on processing load and structural disambiguation. In Experiment 1 we saw that a subordinate clause is not always dispreferred relative to a structurally simpler main clause if the latter has an uninterpreted empty category. Indeed, the processing penalty for a non-optimal EC appears to be of about the same order of magnitude as the processing penalty for added structural complexity. The conclusion is further supported by the results of the on-line reading study, which compared clausal objects and relative clauses and showed that a relative clause is not always dispreferred compared to a structurally simpler clausal object if the latter has an uninterpreted empty category. Of course more research is needed to establish how generally this conclusion holds. To date, filler-first dependencies such as wh-movement constructions have received the most attention, in part because their long-distance span creates special challenges for the parsing mechanism. Gap-first constructions have attracted less psycholinguistic research. In English, gap-first dependencies can be created by rightward movement (e.g., Luke put e on the table all the food he could find in the fridge) or by backwards anaphora (e.g., After e practicing the cello, Sue took a siesta), and occur in some control contexts (e.g., e To clear myself of the charges is important to me). Though they are found in many languages, until recently they have been little studied experimentally. They fall together in one important respect with filler-first constructions where the filler is not recognizable as such until the gap is encountered, which have been extensively studied (see Fodor, 1995 for references). Examples in English are passives, subject-raising, and control constructions such as I hope e to clear
320
S. Ng and J.D. Fodor
myself of the charges. The processing of such filler-gap dependencies must be gap-initiated, even though the filler occurs first. We may consider formulating a comprehensive parsing strategy for gapinitiated dependencies, albeit very tentatively at this stage. Based on the data here, one might speculate that when a gap is encountered, the parser will first look leftward for the filler in the already-parsed segment of the sentence or a preceding part of the discourse. In the sentence completion study, when a plausible referent was available in the preceding clause, the parser used it for gap interpretation, resulting in a substantial amount of main clause continuations. In the reading study, the matrix subject was highly preferred as the filler for the gap in the subordinate clause, provided that the semantic and syntactic properties of the matrix verb did not discourage such a co-indexation. If the search for a prior filler does not yield any good candidates, we may suppose that the parser then initiates a search for the filler in the incoming words, while it is simultaneously assigning structure to those words.11 This is where the balance between optimizing phrase structure and optimizing interpretation of the empty category comes into play.
References Bader, M. (this volume). On being both head-initial and head-final. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. Boland, J. E., Tanenhaus, M. K., & Garnsey, S. M. (1990). Evidence for the immediate use of verb control information in sentence processing. Journal of Memory and Language, 29, 413–432. Chomsky, N. (1981). Lectures on Government and Binding. Dordrecht and Holland: Foris Publications. Crain, S., & Fodor, J. D. (1985). How can grammars help parsers? In D. R. Dowty, L. Karttunen & A. M. Zwicky (Eds.), Natural language parsing: Psycholinguistic, computational, and theoretical perspectives (pp. 94–128). Cambridge: Cambridge University Press. Crain, S., & Steedman, M. (1985). On not being led up the garden path: The use of context by the psychological parser. In D. R. Dowty, L. Karttunen & A. M. Zwicky (Eds.), Natural language parsing: Psycholinguistic, computational, and theoretical perspectives (pp. 320–358). Cambridge: Cambridge University Press. de Vincenzi, M. (1989). Syntactic parsing strategies in a null subject language. Unpublished doctoral dissertation, University of Massachusetts, Amherst. Egerland, V. (2003). Impersonal pronouns in Scandinavian and Romance. Working Papers in Scandinavian Syntax (WPSS), 71, 75–102. Fodor, J. D. (1995). Comprehending sentence structure. In L. R. Gleitman & M. Liberman (Eds.), Invitation to cognitive science (Vol. 1, 2nd ed., pp. 209–246). Cambridge, MA: MIT Press. 11
It is not clear from our data here whether the parser actively continues searching for a filler if an easier alternative is available in the form of a locally assigned generic (‘‘arb’’) interpretation, which does not require associating the EC with any other element. Our latest research (Ng, 2008a, 2008b) indicates that the parser does indeed seek out a filler.
14
Use Your Headedness: An Exercise in Psycholinguistic Exploitation
321
Hornstein, N. 1999. Movement and control. Linguistic Inquiry, 30, 69–96. Hu, J., Pan, H., & Xu, L. (2001). Is there a finite vs. nonfinite distinction in Chinese? Linguistics, 39, 1117–1148. Huang, Y. (1994). The syntax and pragmatics of anaphora. Cambridge: Cambridge University Press. Huang, C.-T. J. (1989). Pro-drop in Chinese: A generalized control theory. In O. Jaeggli & K. J. Safir (Eds.), The null subject parameter (pp. 185–214). Dordrecht: Kluwer Academic Publishers. Landau, I. (2001). Control and extraposition: The case of Super-Equi. Natural Language and Linguistic Theory, 19, 109–152. Lin, C.-J., & Bever, T. G. (this volume). Garden path and the comprehension of head-final relative clauses. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. Mitchell, D. C. (1984). An evaluation of subject-paced reading tasks and other methods of investigating immediate processes in reading. In D. E. Kieras & M. Adam Just (Eds.), New Methods in Reading Comprehension Research. Hillsdale, NJ: Erlbaum. Mitchell, D. C. (2004). On-line methods in language processing: Introduction and historical review. In M. Carreiras & C. Clifton Jr. (Eds.), The on-line study of sentence comprehension: Eye-tracking, ERPs and beyond (pp. 15–32). New York: Psychology Press. Ng, S. (2008a). Gap-driven processing of filler-gap dependencies in Chinese. Poster presented at the 21st Annual CUNY Conference on Human Sentence Processing, University of North Carolina, Chapel Hill, NC, March 2008. Ng, S. (2008b). On-line interpretation of empty categories in Chinese. Paper presented at the 20th North America Conference on Chinese Linguistics (NACCL-20), Ohio State University, April 2008. Ng, S. (2009). Processing Chinese empty categories. Unpublished Ph.D. dissertation, Graduate Center, City University of New York. Simpson, A. (2002). On the status of modifying de and the structure of the Chinese DP. In S.-W. Tang & C.-S. Luther Liu (Eds.), On the formal way to Chinese languages (74–101). Stanford, CA: CSLI Publications. Vasishth, S. (this volume). Integration and prediction in head-final structures. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. Xu, L. (1986). Free empty category. Linguistic Inquiry 17: 75-93. Yamashita, H., Stowe, L. A., & Nakayama, M. (1993). Processing of Japanese relative constructions. In P. M. Clancy (Ed.), Japanese/Korean linguistics 2 (pp. 248–264). Stanford, CA: CSLI Publications.
Part VI
Head-Direction and Processing Theory
Chapter 15
On Being Both Head-Initial and Head-Final1 Markus Bader
15.1 Introduction Starting with the seminal work of Greenberg (1963), the question of which word orders are preferred by the languages of the world has occupied a central place in typological investigations. One tendency has become known under the name of the head parameter: Languages prefer consistent orders between heads and their complements – either head-initial or head-final. In various works, J. Hawkins argued that the head-parameter is an adaptation of the grammar to the needs of the human parsing mechanism (e.g., Hawkins, 1990, 1994): For purposes of parsing efficiency, it is optimal to order heads and complements in a consistent way. However, there are also certain well-known exceptions to the consistent ordering between heads and complements. One of these exceptions, which has been pointed out by Hawkins himself, concerns the relationship between complementizer position and verb-object order. As summarized in (1) (from Hawkins, 1990, p. 225), the complementizer is consistently clause-initial in VO languages whereas OV languages occur both with clause-initial and clause-final complementizers (see also Dryer, 1992). (1)
a. VO languages: S[Comp S] only b. OV languages: S[Comp S] or S[S Comp]
1 This work was supported by the Deutsche Forschungsgemeinschaft (SFB 471, Project D2). Thanks are due to the organizers and the audience of the conference on processing head-final structures, as well as to Lyn Frazier, Jana Ha¨ussler and Tanja Schmid for discussion and helpful comments.
M. Bader (*) Department of Linguistics, University of Konstanz, 78457 Konstanz, Germany e-mail: [email protected]
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_15, Ó Springer ScienceþBusiness Media B.V. 2011
325
326
M. Bader
One of those inconsistent languages that have a head-initial CP together with a head-final VP is German, the language that is the focus of the current chapter. In German, the property of mixed headedness interacts in an interesting way with the syntax of infinitival complement clauses: They are not introduced by an overt complementizer and they can occur on either side of the verb. Thus, German allows both right and left branching structures, as schematically shown in (2). (2)
a. [Comp ... [XP ... V] ... V] b. [Comp ... V [XP ... V]]
While neither left- nor right-branching per se seems to be problematic for the Human Sentence Processing Mechanism (HSPM) (cf. Miller & Chomsky, 1963), the co-existence of head-initial and head-final phrases as found in German raises certain questions that will be explored in the present paper. The first question focuses on the verb-final nature of German and asks for the time course of building up the head-final VP. This question is thus relevant for verb-final languages more generally. I will review experimental evidence arguing in favor of a strictly incremental HSPM that integrates each word into an ongoing syntactic structure as soon as it is encountered, whether heads are initial or final in their respective projections. This implies that there is no fundamental difference between the processing of head-initial and head-final languages (cf. Inoue and Fodor (1995) for a similar view, and Mazuka and Lust (1990) for an alternative view). A second question to be explored in this paper concerns the grammar-parser relationship. In particular, does the grammar show special adaptations to parsing problems that result from the particular mixture of head-initial and head-final properties found in German? I will argue that German grammar (and related West-Germanic OV-languages) can only be understood in the context of requirements imposed by the HSPM. This is in particular true of so-called clause-union phenomena (for an overview, cf. Wurmbrand, 2006). Certain left-branching structures have been hypothesized to allow an alternative syntactic analysis in which a clause-like complement is pruned in favor of a compound-like verb cluster. I will review recent experimental evidence in support of the assumption that the grammar provides the means for verb-cluster formation. The organization of this paper is as follows. The next section introduces two basic approaches to human parsing – head-driven parsing and strictly incremental parsing – and discusses how they apply to the processing of head final structures. Section 15.3 presents a new argument in favor of strictly incremental parsing. Certain consequences of strictly incremental parsing for the grammar are the topic of Section 15.4. A short summary of this chapter is provided in Section 15.5.
15
On Being Both Head-Initial and Head-Final
327
15.2 Head Position, Branching Direction and Mode of Parsing With regard to the parsing algorithm used by the HSPM, the best-known distinction is probably the one between top-down and bottom-up parsing. Within discussions of the HSPM a somewhat different distinction has played a more prominent role: the distinction between head-driven parsing and strictly incremental parsing. Head-driven parsing, which is defined in (3), was proposed in reaction to the fact that, starting in the late seventies, grammatical theory was attributing more and more weight to lexical information. (3)
Head-Driven Parsing The parser constructs a phrase XP at the point where it encounters the head of XP.
Head-driven parsing is assumed by many parsing models that assume a strongly lexicalized representation of grammatical knowledge (cf. MacDonald, Pearlmutter, & Seidenberg, 1994; Pritchett, 1988; 1992; Trueswell & Tanenhaus, 1994; Vosse & Kempen, 2000). By making the postulation of a phrasal projection dependent on the presence of the phrasal head in the input string, a head-driven parsing scheme implies a tight coupling between the time course of structure building and the positions of heads. For purposes of illustration, (4) shows a strictly head-initial and a strictly head-final clause structure.
(4)
a. Head-Initial Clause Structure:
b. Head-Final Clause Structure:
CP C
CP IP
NP
IP I’
I
NP VP
V
C I’
VP XP
XP
...
...
I V
An HSPM that builds up a phrase at the point where the head of the phrase is encountered in the input can build the head-initial tree in (4a) by and large in an incremental way. The only exception is the subject NP in the specifier of IP that has to wait for the head I before it can be integrated into the ongoing structure. Building up a head-final structure as in (4b), in contrast, involves heavy delays. No phrase can be integrated into any higher structure before the end of the clause.
328
M. Bader
With respect to the time course of structure building, head-driven parsers contrast with parsers that work in a strictly incremental way. A strictly incremental parser can be defined as a parser adhering to the Left-to-Right Constraint of Frazier and Rayner (1988). This constraint is shown in (5). (5)
Left-to-Right Constraint Each item is incorporated into a constituent structure representation of a sentence (essentially) as the item is encountered.
Because it cannot (always) rely on lexical information associated with heads, the actions of a strictly incremental parser must at least sometimes be driven by item-independent structural knowledge about the shape of syntactic phrases. In this sense, a strictly incremental parser might be called structuredriven. I will avoid the term ‘‘structure-driven parsing’’ in the following, however, because it is often used with the further implication that itemspecific lexical information beyond syntactic category labels is not available to the HSPM for the initial construction of a syntactic structure. For present purposes, this stronger claim is not necessary. The important point is only that the HSPM can draw on item-independent structural knowledge in order to postulate phrasal projections even if their head has not yet been encountered in the input. It is immaterial at this point how much of the information associated with heads is used in those cases where the head is available from the beginning, as in head-initial phrases. With the distinction between head-driven and strictly incremental parsing in the background, let us now turn to the issue of mixed headedness. Combining the upper part (CP) of the head-initial structure in (4a) with the lower part (IP and VP) of the head-final structure in (4b) results in the structure that is shown in (6) (6)
Mixed-Headed Clausal Structure (Version 1): CP IP
C NP
I’ VP
XP
I V
...
Mixing head-initial- and head-final phrases as in (6) is potentially harmful for the HSPM, in particular when XP in (6) is a complex constituent not introduced by any overt marker. As already indicated in the introduction, this
15
On Being Both Head-Initial and Head-Final
329
is the situation that results when an infinitival complement is embedded to the left of the verb. An example sentence illustrating this point is shown in (7).2 (7)
. . . dass Peter letzte Woche [CP PRO ein Buch zu lesen] versucht hat. that Peter last week a book to read tried has ‘that last week Peter tried to read a book.’
Assuming syntactic structures as in (6) and a parsing mechanism that works strictly incrementally, sentences like (7) pose two problems. The first problem is the Complexity Problem: (7) is an instance of center embedding. The second problem is the Ambiguity Problem: (7) necessarily involves reanalysis because the parser will overlook the clausal boundary on first-pass parsing. As shown by the example in (8), the initial part of sentence (7) (dass Peter letzte Woche ein Buch ‘‘that Peter last week a book’’) is easily compatible with a continuation in which the NP ein Buch is not embedded in a further clausal constituent. In fact, sentences as in (8) are much more frequent than sentences as in (7). (8)
. . . dass Peter letzte Woche ein Buch las. that Peter last week a book read ‘that Peter read a book last week.’
It is important to note at this point that the problem of overlooking a clausal boundary on first-pass parsing is in no way peculiar to the property of mixedheadedness. A well-known example from English illustrating the same kind of parsing problem is shown in (9). (9)
I believe [CP the author was lying].
Sentences like (9) are English garden-path sentences (e.g., Ferreira & Henderson, 1990; Trueswell, Tanenhaus, & Kello, 1993). On first-pass parsing, the HSPM analyzes the NP the author as object of believe. Only when the auxiliary was is encountered will the HSPM realize that this NP is the subject of an embedded clause instead. On the one hand, this shows that inserting a clausal node on second-pass parsing is indeed a costly operation, although the costs are only quite moderate in examples of this sort. On the other hand, this raises the question of why mixed-headed structures as in (6) are singled out as a special topic of research in the present paper. 2
The exact syntactic category of infinitival complements is a matter of dispute, with answers ranging from VP to CP. For the following argument, the label of the constituent under consideration is immaterial; for reasons of concreteness, I will assume a CP.
330
M. Bader
With regard to this question, I would like to propose that in qualitative terms, mixed-headedness (at least of the sort considered here) does not pose any special problems at all. This is a claim that I am obliged to make since I will argue in the next section that the HSPM works alike in all languages as far as the time course of structure building is concerned. In quantitative terms, however, the situation is somewhat different. In English, overlooking a clause boundary is a restricted phenomenon. Furthermore, the grammar of English provides a means to avoid the potential garden-path effect in sentences like (9), namely by inserting a complementizer in front of the embedded clause. In German, in contrast, the problem is more widespread, at least when we take into account all categories along the extended projection of V (that is, VP and all functional projections above). In fact, it occurs in every embedded sentence containing more than a single verb.3 In addition to control verb examples as in (7), this includes all embedded sentences with a complex tense form, a modal verb, or a raising verb, to list the most prominent possibilities. In main clauses, the problem will occur whenever a clause contains three or more verbs. Thus, mixed-headedness should lead to permanent parsing problems of the sort discussed above if the following two assumptions are correct: the assumption of strictly incremental parsing and the assumption of syntactic structures as in (6). In the next section, I present a new argument for the assumption that the HSPM does indeed work in a strictly incremental way. With regard to the two problems under consideration, help must therefore come from the grammar. How this might work will be discussed after the experimental evidence in favor of strictly incremental parsing has been presented.
15.3 A New Argument for Strictly Incremental Parsing Although strictly incremental parsing and head-driven parsing make opposing claims when it comes to the processing of head-final structures, finding firm experimental evidence for distinguishing between these two hypotheses has turned out to be an intricate task. In my earlier work, I already attempted to provide experimental evidence in favor of strictly incremental parsing (see Bader, 1996; Bader & Lasser, 1994) with the help of embedded infinitival clauses similar to the ones discussed above. However, as pointed out in later work (e.g., Kamide & Mitchell, 1999; Vosse & Kempen, 2000),
3
German does not have a class of verbs that are defective in the same sense as English modal verbs and which therefore could be argued to be directly inserted under I .
15
On Being Both Head-Initial and Head-Final
331
the conclusions were much less compelling than initially thought because a structural ambiguity was confounded with a case ambiguity that makes the results open to explanations in terms of structural and/or lexical frequencies. I therefore present a new argument for strictly incremental parsing that is based on a lexical category ambiguity found in German. Lexical-category ambiguities have been extensively studied for English but have not received much attention in the literature on verb-final languages so far.
15.3.1 The Verb-Adjective Ambiguity The local syntactic ambiguity that forms the basis of the upcoming argument takes advantage of the fact that German exhibits a lexical-category ambiguity between finite main verbs and deverbal adjectives. Considered abstractly, this gives rise to an ambiguity between a structure like (10a) and a structure like (10b). (10)
a.
b.
VP XP
VP
V° w
AP XP
… A° w
(10) involves a phrase XP, which can be attached to VP either directly or indirectly via an intermediate AP. Depending on the particular structure, the lexically ambiguous word w is either a verb or an adjective. An illustrative sentence pair containing this kind of ambiguity is given in (11) (from Bader, 1996). (11)
a.
b.
. . ., dass er Maria daru¨ber informierte abzureisen. that he Maria there-about informed to-leave ‘that he informed Maria about his leaving’ . . ., dass er Maria daru¨ber informierte Leute vorgestellt hat. that he Maria there-about informed people introduced has ‘that he introduced people thatwere informed about it to Maria’
In (11a), informierte is the finite main verb of the sentence. In (11b), in contrast, informierte is the head of an adjective phrase that modifies the immediately following noun. In both cases, the preceding word daru¨ber is a complement of informierte. Disambiguation in these examples is achieved by the material following informierte. The complete syntactic structures of the two sentences in (11) are shown in (12).
332
(12)
M. Bader
a. Structure of main-verb reading (cf.(11a)): CP C
VP
dass
NP
V’ NP
er
V’ PP
Maria
V’
darüber
CP
V informierte
abzureisen
b. Structure of adjectival reading (cf. (11b)): CP C dass
VP NP er
V’ NP
V’
Maria
NP AdjP
PP darüber
V N
A
vorgestellt hat
Leute
informierte
The verb-adjective ambiguity introduced in this section is the German counterpart of the ambiguity between past-participle and main-verb in English, as exemplified by the famous example in (13) (from Bever, 1970). (13)
a. The horse raced past the barn fell. b. The horse raced past the barn and fell.
Over the years, experimental evidence has accumulated that indicates that the resolution of the main-verb – past participle ambiguity is governed at least partially by the ratio between main verb usages and past participle usages of the individual verbs that give rise to this kind of ambiguity (e.g., MacDonald, 1994). When we next consider in detail how the HSPM might handle the parallel lexical category ambiguity in German sentences like (11), the role of probabilistic lexical information will play a crucial role for deciding between headdriven and strictly incremental parsing.
15.3.2 V-A-Ambiguity: Processing by a Non-Incremental, Head-Driven Parser When a head-driven parsing scheme is applied to head-final structures, phrases often have to wait for some while before they can be integrated into a higherlevel structure. A head-driven parser must therefore contain a memory buffer
15
On Being Both Head-Initial and Head-Final
333
for holding phrases that cannot immediately be attached to their mother phrase. We consider next how this memory buffer unfolds over time when processing sentences with a local verb-adjective ambiguity. Since we are dealing with verb-final clauses, all constituents of VP will have to stay in the buffer until the verb is encountered. After processing the words that precede the ambiguity under consideration in sentence (11), the memory buffer will look as in (14). (14)
Memory buffer after processing of dass er Maria [CP dass], [DP er], [DP Maria]
(14) contains three elements. A CP node that has been projected from the complementizer dass and two DP nodes. These three nodes must stay unconnected within the memory buffer because the parser has not yet seen the head of the intermediate projection VP that links the CP with the various DPs. Although the local ambiguity starts with the word daru¨ber, on encountering this word in the input string nothing else can be done than putting it into the memory buffer, as shown in (15). (15)
Memory buffer after processing of dass er Maria + daru¨ber [CP dass], [DP er], [DP Maria], [PP daru¨ber]
It is only when the word informierte is encountered at the next step that the local ambiguity inherent in these sentences can be detected by the parser. Because informierte is lexically ambiguous between adjective and verb, the parser can either project an AP or a VP; the PP daru¨ber that was put last onto the memory buffer will become the complement of the newly projected phrase, whether AP or VP. The two possible extensions of the memory buffer that result from processing the ambiguous word are shown in (16a) and (16b). (16)
Memory buffer after processing of dass er Maria daru¨ber + informierte a. [CP dass], [DP er], [DP Maria], [VP daru¨ber informierte] b. [CP dass], [DP er], [DP Maria], [AP daru¨ber informierte]
Assuming that the parser cannot pursue alternative structures in parallel, it now has to make a decision between the two structures in (16). Most current head-driven models assume that probabilistic lexical information plays a crucial role in decisions of this kind. In the simplest case, this would mean that the proportion with which the parser chooses the verb reading approximates the proportion of verb usages of the ambiguous words
334
M. Bader
under consideration. For our example sentences, this means that the ratio between pursuing (16a) and pursuing (16b) approximates the frequency ratio between the adjective and verb reading of informierte as found in actual language use. With respect to reading sentences that contain a local ambiguity between verb and adjective, frequency-based head-driven parsing thus predicts gardenpath effects for both kinds of disambiguation. The strength of the ensuing garden-path effect should correlate with lexically based ratios between adjective and verb usage.
15.3.3 V-A-Ambiguity: Processing by a Strictly Incremental Parser Let us consider next how a strictly incremental parser handles the syntactic ambiguity under consideration. (17) shows the partial tree that the HPSM will have built after processing of the clause initial complementizer dass (‘‘that’’). (17)
Partial tree after processing of dass: CP C
VP
dass
(17) does not only contain a CP node reflecting the presence of a complementizer but also a VP as a sister to C. As discussed above, this VP can already be postulated at this early point because it is a necessary node for any sentence. The following two NPs can then straightforwardly be integrated into the already existing VP. This will result in the partial structure shown in (18) (V a stands for a verbal category that is unspecified as to its X-bar level). (18)
Partial tree after processing of dass er Maria: CP C dass
VP NP er
V’ NP
Vα
Maria
The next word to integrate is the word daru¨ber. The two ways to integrate this item are shown in (19).
15
(19)
On Being Both Head-Initial and Head-Final
335
Partial trees after processing of dass er Maria daru¨ber:
a.
b.
CP C
VP
dass NP er
C dass
V’ NP
Maria
CP
NP er
V’ PP
VP
V
V’ V’
NP Maria
NP N
AdjP
darüber PP
V
A
darüber
In the left-hand tree (19a), the PP daru¨ber has been attached directly to the VP that was already set up when processing the preceding part of the clause. In the right-hand tree (19b), this PP has been analyzed as the complement of an adjective that itself modifies a following noun. Note that both adjective and noun are hypothesized nodes in (19b). The input thus far contained neither of the two. Any economy-based parsing principle, like Minimal Attachment (Frazier, 1979) and its successors, will predict that the HSPM computes tree (19a) on first-pass parsing. This tree is clearly less complex than its competitor in (19b). Furthermore, this tree embodies only a single hypothesis about the remainder of the sentence, namely that a finite verb will follow. Since every clause is in need of a finite verb, the parser can be sure that the remaining input will fulfill this hypothesis. Structure (19b), in contrast, incorporates two additional hypotheses, both of which are risky: that the remaining input contains an adjective and a noun in addition to a finite verb. In fact, (19b) is not the only structure in which the PP daru¨ber appears on the left edge of a phrase intervening between it and the VP. A further instance of this structure is provided in (20). (20)
. . .dass er Maria [CP daru¨ber zu sprechen] gezwungen hat. that he Maria there-about to talk forced has ‘. . . that he forced Maria to talk about it.’
In (20), the PP daru¨ber is part of an embedded infinitival clause of the sort discussed at the beginning of this chapter. Postulating a structure as in (20) on encountering the PP daru¨ber would imply to hypothesize an upcoming infinitival verb. Thus, any structure other than the one in (19a) involves making risky predictions about how the sentence continues. By sticking to the structure in (19a), the HSPM adopts the most economical structure for the current input and as a side effect also the most probable one.
336
M. Bader
The next word to integrate is the category-ambiguous word informierte, which is the focus of the current discussion. A decision in favor of structure (19a) does not yet settle the issue of whether informierte should be analyzed as a verb or as an adjective. As illustrated by the example in (21), adjectival informierte can be used without a PP complement and thus as part of an NP that starts directly after the PP daru¨ber. (21)
. . .dass er Maria daru¨ber [NP informierte Leute] kontaktieren sah. that he Maria there-about informed people contact saw ‘. . . that he saw Maria contact informed people regarding this.’
The structural ambiguity that still exists after the HSPM has opted in favor of structure (19a) actually recapitulates the situation in (19). The word informierte can be analyzed as a verb that forms the head of the already present VP. Alternatively, informierte might be taken as an adjective that modifies a noun that has not yet been seen in the input, analogously to the tree in (19b). Again, the more economical structure is the one in which the current word is directly attached to VP, which means analyzing informierte as a verb and thus as the head of VP. For a strictly incremental HSPM, we thus arrive at the following predictions. When the HSPM processes sentences of the kind under consideration, it should prefer the verb reading on its first pass through the sentence. The sole reason for this preference is the fact that the verb reading – that is, the reading where the ambiguous word is the main verb of its clause – is structurally simpler than the competing adjective reading. This implies in particular that the frequency properties associated with the two lexical categorizations of the ambiguous word should not have any effect on first-pass parsing. The verb reading should always be preferred. When the remainder of the sentence supports this reading, there is nothing else to do and sentences disambiguated toward the verb reading should therefore not show any effects related to lexical frequency. Because of the general preference for the verb reading, disambiguation toward the adjective reading should always trigger reanalysis. It is at this point where lexical frequencies might come into play even under the hypothesis of strictly incremental parsing. In particular, the strength of the garden-path effect that is predicted for disambiguation towards the adjective reading might be expected to correlate with lexical frequencies. After all, when the HSPM has detected that the part of the sentence that follows the lexically ambiguous word informierte cannot be integrated into the verb structure built thus far, a crucial part of the reanalysis processes that must now be invoked is the re-categorization of informierte. Initially, this word was analyzed as a verb but now it is needed as an adjective. Recognizing that this kind of change is allowed can plausibly be assumed to be easier the more often such a word occurs as an adjective in language use.
15
On Being Both Head-Initial and Head-Final
337
In summary, with regard to the role of lexical frequencies, the hypothesis of strictly incremental parsing predicts a different outcome than the hypothesis of head-driven parsing. Sentences disambiguated towards the verb reading should be unaffected by lexical frequencies whereas sentences disambiguated towards the adjective reading might well show frequency effects. The rationale for this prediction went as follows. The verb reading is predicted to be always preferred on first-pass parsing, independent of the ratio between verb and adjective occurrences of the lexically ambiguous words. Disambiguation towards this reading will thus always fit into the structure computed on first-pass parsing. Conversely, disambiguation towards the adjectival reading of the ambiguous words will always cause the HSPM to revise the structure initially built. This revision might be sensitive to effects of lexical frequency.
15.3.4 Testing the Predictions Bader (in prep.) ran a series of self-paced reading experiments in order to test the predictions of head-driven and strictly incremental parsing. A stimulus sentence of one of the experiments is shown in (22). (22)
a. Ich habe gesehen, dass Jonas sehr sorgfa¨ltig kopiert(e), I have seen that Jonas very carefully xerox(es|ed) was Birgit fu¨r den Vortrag braucht. what B. for the talk needs ‘I have seen that Jonas (xeroxes|xeroxed) very carefully what Birgit needed for the talk.’ b. Ich habe gesehen, dass Jonas (einige) sehr sorgfa¨ltig kopierte I have seen that Jonas some very carefully xeroxed Unterlagen heute Morgen bereitwillig verteilt hat. documents today morning promptly distributed has ‘I have seen that Jonas promptly distributed (some) documents this morning which were xeroxed very carefully.’
Investigating the local ambiguity under consideration experimentally is faced by two practical problems. First, as a comparison between (22a) and (22b) shows, kopierte is followed by a comma when it is a clause-final finite main verb as in (22a) but not when it is an adjective as in (22b). Since this might affect the processing of these sentences in ways that are unrelated to the issues at hand, Bader (in prep.) also ran experiments in which the ambiguous word was followed by a comma in both conditions. This was achieved by making the adjective phrase in sentences like (22b) more complex by adding a further adjective. As it turned out, comparable results were
338
M. Bader
obtained whether punctuation was matched across conditions or not. For reasons of space, in the following I will only discuss experimental results for sentences as in (22). A second practical problem concerns the use of control sentences. Normally, one uses locally unambiguous sentences as a baseline for measuring whether locally ambiguous sentences lead to garden-path effects. For the construction under consideration, it is not possible to construct corresponding sentences without any local ambiguity. I will therefore only talk of control sentences in the following. Control sentences in the verb condition use the same verb as in the locally ambiguous condition but in the present tense (kopiert) instead of the past tense (kopierte). For inflectional reasons, present tense verbs are not homophonous with attributive adjectives. Control sentences in the adjective condition had a determiner (einige ‘‘some’’ in (22b)) in front of the otherwise ambiguous NP. This determiner introduces a strong bias towards reading the following ambiguous word as adjective. Bader (in prep.) constructed 28 sentences with each sentence appearing in four versions according to the two factors Structure (Verb versus Adjective) and Ambiguity (Ambiguous versus Control). Each sentence contained a different lexically ambiguous word, which could be categorized either as a finite verb or as an adjective. For each item, a random selection of 100 corpus instances from the newspaper corpus of the Institut fu¨r Deutsche Sprache (IDS, Mannheim, Germany) was analyzed in order to obtain a measure of lexical preference. For the 28 items selected for the experiment, the percentages of verb instances ranged from 27% to 100%, with a mean value of 68%. Participants read sentences using a word-by-word non-cumulative selfpaced reading procedure. Each participant saw only one version of each sentence. The 28 experimental sentences read by each participant were interspersed among a list of about 100 filler sentences. From the raw reading times we computed residual reading times, which correct for effects of word-length. Residual reading times are a measure of how much reading times on a particular word of a given length w (measured in numbers of letters) deviate from the average reading times for words of length w. They can thus be both positive (longer reading times than expected for words of the given length) and negative (shorter reading times than expected for words of the given length). For reasons of space, I only discuss results for the complete disambiguating region here. This region always contained exactly six words following the lexically ambiguous word. Figure 15.1 shows the residual reading times that were obtained for the disambiguating region. Figure 15.1 reveals a clear difference between verb and adjective sentences. In the verb condition, control and ambiguous sentences do not differ with regard to their reading times. In the adjective condition, in contrast, ambiguous sentences needed much more time in the disambiguating region than control sentences. Thus, we see a garden-path effect only when the local ambiguity was disambiguated in favor of the adjective reading. This is exactly the pattern predicted by the hypothesis of strictly incremental parsing: Because of its
15
On Being Both Head-Initial and Head-Final
Fig. 15.1 Residual reading times (in ms) for the disambiguating region of sentences as in (22); see main text for details
339
500
Verb
400
Adjective
300 200 100 0 Control
Ambiguous
structural simplicity, the verb reading is always chosen on first-pass parsing. When the disambiguating region is compatible with this choice, it is processed in the same way as it is in control sentences. When the disambiguating region contradicts the initial choice, in contrast, revision processes become necessary, which are reflected in the higher reading times for ambiguous adjective sentences. According to the hypothesis of head-driven parsing, a garden-path effect should have become visible for both ambiguous verb and ambiguous adjective sentences but this is not what was found. This conclusion is strengthened by the further finding of a substantial and significant correlation between residual reading times and lexical corpus frequencies for ambiguous adjective sentences (r = 0.52, p < 0.01). For ambiguous verb sentences, in contrast, there was no correlation at all (r = 0.06, not significant). The control conditions also did not show significant correlations with the corpus frequencies. This again confirms the predictions derived from the assumption of strictly incremental parsing. Because the verb reading is always chosen on first-pass parsing for purely structural reason, lexical frequencies do not play any role for processing of verb sentences. For adjective sentences, in contrast, such frequencies were expected to have an effect on the revision processes that have to be invoked in order to transform the verb structure computed on first-pass parsing into the adjective structure required by the disambiguating region. As part of this revision, the lexically ambiguous word has to be re-categorized from verb to adjective. For each ambiguous word, the ease of this re-categorization seems to depend on the frequency ratio between verb and adjective usage, as reflected by the correlation between reading times and lexical frequencies. In summary, the results of the self-paced reading experiment that I have reviewed here confirm the predictions of a strictly incremental parser: A general preference for the verb reading and a garden-path effect for the adjective reading; the strength of this garden-path effect correlates with lexical frequencies. These results are unexpected under a head-driven perspective because head-driven parsing had predicted garden-path effects for verb and adjective sentences and correlations with lexical frequencies for sentences in both conditions.
340
M. Bader
15.4 Some Consequences for the Grammar-Parser Relationship Accepting the conclusion that the HSPM works in a strictly incremental manner, let us now explore some consequences for German grammar. To recapitulate shortly, consider again sentence (7) that is repeated below as (23). (23)
. . . dass Peter letzte Woche [ein Buch zu lesen] versucht hat. that Peter last week a book to read tried has ‘that last week Peter tried to read a book.’
Under the assumption that the infinitival complement, which intervenes between the clause-initial complementizer and the clause-final verb, is a clausal constituent itself, sentence (23) poses two problems: The Complexity Problem – (23) is an instance of center embedding – and the Ambiguity Problem – (23) necessarily induces a garden-path effect because the clausal boundary will be overlooked on first-pass parsing. These two problems owe their existence to the juxtaposition of two assumptions: strictly incremental parsing and syntactic structures in which the infinitival complement forms a clause-like constituent. The correctness of the former assumption has been argued for in the preceding section. What is at issue now is the correctness of the second assumption, which is illustrated by the syntactic tree in (24a) (repeated from above). (24)
a. Bi-Clausal Structure
b. Mono-Clausal Structure
CP C
CP IP
X P1
C I’
VP CP
X P1 I
V
VP
…
V’ X P2 …
V Vinf
V
N P 2 ...Vinf
In the syntactic literature on German, there is a broad consensus that the grammar does indeed provide bi-clausal structures as in (24a) for sentences as in (23). However, it has also been proposed that these sentences can be assigned an alternative, mono-clausal structure, which is shown in (24b). In (24b), infinitival verb and matrix verb have fused to form a so-called verb cluster. The structure in (24b) is almost identical to the structure of a simple sentence with a single verb. The only difference is that (24b) contains a verb cluster where a simple sentence would contain just a single verb. Verb-cluster formation has been proposed in
15
On Being Both Head-Initial and Head-Final
341
many syntactic frameworks (e.g., Haider, 1993; Johnson, 1986; Steedman, 1983) but it is nevertheless a controversial assumption (for an overview, cf. Wurmbrand, 2006). In the rest of this chapter, experimental evidence for syntactic structures like (24b) will be reviewed. As already pointed out at the beginning, German grammar provides several positions for infinitival complements: The two most important positions are the position directly in front of the clause-final verb (intraposed infinitivals) and the position directly behind the clause-final verb (extraposed infinitivals). When an infinitival complement occurs extraposed, syntactic tests show that it forms a clausal constituent. When infinitival complements occur intraposed, some syntactic tests again indicate the presence of a clausal constituent; other tests, however, argue against such a constituent (cf. von Stechow & Sternefeld, 1988). For reasons of space, only a single test for mono-clausality will be discussed here – the so-called long distance passive. Consider first sentence (25). This is a simple active sentence with an intraposed infinitival clause. If the grammar allows both structures in (24), than this sentence is compatible with both of them. (25)
. . . dass Maria den Kuchen zu essen versucht hat. that Maria the-ACC cake to eat tried has ‘. . . that Maria tried to eat the cake.’
The effect of passivizing a sentence like (25) with its infinitival complement in intraposed position is shown in (26). The interesting point is that the accusative object of the active clause can appear either with accusative case or with nominative case after passivization. This is exactly what is expected if sentence (25) can be assigned either a mono- or a bi-clausal structure. (26)
Kuchen zu essen] versucht wurde. a. . . . dass [CP den that the-ACC cake to eat tried was ‘. . . that it was tried to eat the cake’ b. ?. . . dass der Kuchen [V zu essen versucht wurde ]. that the-NOM cake to eat tried was ‘. . . that it was tried to eat the cake’
Under passivization, the former subject gets suppressed and the former object becomes the subject. The former subject Maria has been suppressed in both (26a) and (26b). In the bi-clausal structure, the infinitival clause is the object when the matrix verb is in the active voice. Thus, it becomes the subject when the matrix verb is put into the passive voice, as in (26a). Note that the
342
M. Bader
infinitival clause itself remains completely unaffected because clauses do not inflect in German and thus the promotion from object clause to subject clause is not overtly visible. In the mono-clausal structure, the NP den Kuchen (‘‘the-ACC cake’’) is the object of the verb cluster as a whole when the cluster is in the active voice. Thus, this NP becomes the new subject under passivization. The accusative NP den Kuchen accordingly appears as the nominative NP der Kuchen in (26b). Because passivizing the matrix verb versuchen (‘‘to try’’) affects the object of the embedded infinitival verb zu essen (‘‘to eat’’), the construction exemplified by (26b) is called long-distance passive. As indicated in (26b), the grammatical status of the long-distance passive is under dispute. Bader and Schmid (2009) investigated how the HSPM processes sentences with intra- and extraposed infinitival complements, with special emphasis given to the phenomenon of long-distance passive. These experiments had two main aims: First, providing experimental evidence for the hypothesis that the grammar provides mono-clausal structures for intraposed infinitival complements and second, clarifying the grammatical status of sentences with long-distance passive. In order to pursue the second aim, the experimental procedure that was used in these experiments was the method of speeded grammaticality judgments. In the experiments reported in Bader and Schmid (2009), this method worked as follows. Participants read sentences, which were displayed on a computer screen in a word-by-word fashion. The presentation time for each word was 225 ms plus an additional 25 ms per character, with no interval between words. Immediately after the last word, three question marks appeared on the screen, indicating to participants that they now had to judge the grammaticality of the sentence by pressing either the right shift key on a regular computer keyboard for judging a sentence as grammatical or the left shift key for judging a sentence as ungrammatical. If no judgement was given within 2000 ms, the trial was aborted automatically. The method of speeded grammaticality judgment has been used successfully to study the mental processes involved in syntactic ambiguity resolution (e.g., Ferreira & Henderson, 1991; Warner & Glass, 1987). In the current context, it is particularly suited because it is sensitive to parsing complexity and to degrees of grammaticality. With regard to parsing complexity, Bader and Ha¨ussler (2009) showed that self-paced reading and speeded grammaticality judgments deliver identical results for sentences in which a local syntactic ambiguity is disambiguated by a verb in clause-final position. With regard to degrees of grammaticality, Bader and Ha¨ussler (2010) showed that speeded grammaticality judgments, which require binary judgments from participants and magnitude estimation, a particular way of rating the grammaticality of sentences in a gradient way (see Bard, Robertson, & Sorace, 1996), lead to equivalent results. Let us now first consider bi-clausal passive clauses as in (26a). Such clauses are of particular relevance in the present context because they provide yet another argument for strictly incremental parsing and against head-driven
15
On Being Both Head-Initial and Head-Final
343
parsing. Because strictly incremental parsing and head-driven parsing make opposing claims concerning the processing cost of sentences like (26a), it is important to have some baseline against which sentences like (26a) can be compared. One of the baselines used by Bader and Schmid (2009) is illustrated by the sentence in (27). (27)
Kuchen zu essen]i wurde tiversucht. [CP Den the-ACC cake to eat was tried ‘To eat the cake was tried.’
(27) is the main clause-counterpart to (26a) in which the infinitival clause has been topicalized to the clause-initial specifier of CP position.4 Because of the structure of main clauses in German, which always have the finite verb in second position (C ), the existence of an embedded infinitival clause becomes obvious in (27) immediately on encountering the infinitival verb zu essen (‘‘to eat’’). (27) therefore does not pose the kind of ambiguity problem caused by intraposed infinitival clauses. Working under the assumption of strictly incremental parsing, Bader and Schmid (2009) proposed the Clause-Union Preference Hypothesis, which is shown in (28). (28)
Clause-Union Preference Hypothesis On first -pass parsing, the HSPM always assigns a mono-clausal structure to a sentence containing an intraposed infinitival complement.
The Clause-Union Preference Hypothesis predicts that sentence (26a) with its intraposed infinitival clause should cause a garden-path effect. For sentence (27) with the infinitival clause topicalized to SpecCP, in contrast, no gardenpath effect should occur. Thus, sentences like (26a) should be more difficult to process than sentences like (27). A different prediction follows from the hypothesis of head-driven parsing. The content of the memory buffer after processing sentence (26a) up to the accusative NP is shown in (29a). As before, all constituents seen so far must remain unconnected within the buffer.
4
Bader and Schmid (2009) also looked at main clauses in which the infinitival clause appeared in the same intraposed position as in (26a). For intraposed infinitival clauses, it did not matter whether the matrix clause was a main clause or an embedded clause itself. For reasons of space, main clauses with intraposition are therefore omitted from the current discussion.
344
(29)
M. Bader
a. Memory buffer after processing of dass den Kuchen [CP dass], [DP den Kuchen] b. Memory buffer after processing of dass den Kuchen zu essen [CP dass], [VP [DP den Kuchen] zu essen]
When the following infinitival verb is processed, a VP can be projected and the NP can be integrated into the VP. The memory buffer now looks as in (29b). When the remaining part of sentence (26a) is encountered, namely the passive verb versucht wurde (‘‘tried was’’), all elements in the buffer can be integrated into a completely connected clause structure. For sentence (27), the crucial processing steps look almost the same, as shown in (30). (30)
a. Memory buffer after processing of Den Kuchen [DP Den Kuchen] b. Memory buffer after processing of Den Kuchen zu essen [VP [DP Den Kuchen] zu essen]
Under a head-driven parsing scheme, processing cost is therefore unaffected by the position of the infinitival clause. Because of center-embedding, an intraposed infinitival clause might still be at a disadvantage. However, topicalization is also associated with additional costs, outweighing the cost of centerembedding: Topicalization involves syntactic movement and results in a sentence with a marked focus structure. All in all, the processing cost of infinitival clauses according to a head-driven parsing schema should not differ depending on whether the clause is in topicalized or in intraposed position. In Experiment 3 of Bader and Schmid (2009), it was found that sentences with a topicalized infinitival clause are easier to process than sentences with an intraposed infinitival clause. This supports the Clause-Union Preference Hypothesis in (28) and thus the assumption of strictly incremental parsing from which this hypothesis was derived. At the same time, the predictions of head-driven parsing are refuted again. Sentences like (26b), which instantiate the long-distance passive, were also part of the experiments conducted by Bader and Schmid (2009). As pointed out above, the grammatical status of such sentences is controversial. In our experiments, such sentences were judged as grammatical about 70% of the time when they contained an inanimate subject, as in the example above. This indicates that the long-distance passive is indeed a grammatical option of German but one of somewhat degraded grammaticality. Bader and Schmid (2009) proposed that the degraded grammaticality of the long-distance passive is in fact a processing effect, as captured by the VerbCluster Complexity Hypothesis in (31).
15
(31)
On Being Both Head-Initial and Head-Final
345
Verb-Cluster Complexity Hypothesis The argument-structure operations involved in verb-cluster formation are costly for the HSPM.
According to the Verb-Cluster Complexity Hypothesis, fusing several verbs into a single verb cluster is costly because it involves relatively complex operations like argument-structure merging. This hypothesis thus complements the Clause-Union Preference Hypothesis by making explicit the drawback of verbcluster formation. The Clause-Union Preference Hypothesis captures the advantages of verb-cluster formation, which are brought about by turning a problematic bi-clausal into a less problematic mono-clausal structure. The Verb-Cluster Complexity Hypothesis, in contrast, captures the price to pay for verb-cluster formation – a price that can even offset the advantages of mono-clausal structures, as witnessed by the degraded grammaticality of the long-distance passive.
15.5 Conclusion I have discussed two issues in the processing of German, a language exhibiting mixed headedness – CPs are head-initial but VPs are head-final. The first issue concerned the time course of building up a head-final VP. With regard to this issue, which is of general relevance in the context of processing verb-final languages, I have discussed experimental evidence from a syntactic ambiguity based on a lexical-category ambiguity between verb and adjective. The experimental evidence from the verb-adjective ambiguity that I have reviewed supports the assumption of a strictly incremental HSPM that works alike for all languages. The second issue that I have discussed concerns the grammar-parser relationship. Does the grammar help the parser by providing easy-to-process structures? With regard to the domain of infinitival complementation in German, I have argued that the answer is yes. Due to the mixed-headedness of German, intraposed infinitival complement clauses would be particularly harmful for the HSPM if they formed a clausal constituent of their own, as they do in the bi-clausal structure. The mono-clausal structure does not pose problems of the same kind and can thus be understood as being motivated by considerations of efficient parsing. As a final note, let me point out that the adaptation of the grammar to the parser’s need that was hypothesized in the preceding section is of a much more modest nature than what is sometimes proposed in discussions of this topic. What is at stake is not whether the grammar – in the sense of Universal Grammar (UG) – has been shaped by pressure from the HSPM (for critical discussion, cf. Fodor, 1981). Instead, the claim is only that the grammar – in the sense of the grammar of an individual language – provides structures that make the parser’s life easier. These structures are assumed to fall within the space of possible structures made available by UG.
346
M. Bader
References Bader, M. (1996). Sprachverstehen: Syntax und Prosodie beim Lesen. Opladen: Westdeutscher Verlag. Bader, M. (in prep.). Lexical-category ambiguity and the processing of head-final clauses. Manuscript in preparation. Bader, M., & Ha¨ussler, J. (2009). Resolving number ambiguities during language comprehension. Journal of Memory and Language, 61(3), 352–373. Bader, M., & Ha¨ussler, J. (2010). Toward a Model of Grammaticality Judgments. Journal of Linguistics, 46(2), 273–330. Bader, M., & Lasser, I. (1994). German verb-final clauses and sentence processing: Evidence for immediate attachment. In C. Clifton, Jr., L. Frazier & K. Rayner (Eds.), Perspectives on sentence processing (pp. 225–242). Hillsdale, NJ: Lawrence Erlbaum. Bader, M., & Schmid, T. (2009). Minimality in verb-cluster formation. Lingua, 119(10), 1458–1481. Bard, E. G., Robertson, D., & Sorace, A. (1996). Magnitude estimation of linguistic acceptability. Language, 72(1), 32–68. Bever, T. G. (1970). The cognitive basis for linguistic structure. In J. R. Hayes (Ed.), Cognition and the development of language (pp. 279–362). New York: Wiley and Sons. Dryer, M. S. (1992). The Greenbergian word-order correlations. Language, 68(1), 81–138. Ferreira, F., & Henderson, J. M. (1990). Use of verb information in syntactic parsing: evidence from eye movements and word-by-word self-paced reading. Journal of Experimental Psychology: Learning, Memory and Cognition, 16(4), 555–568. Ferreira, F., & Henderson, J. M. (1991). Recovery from misanalysis of garden-path sentences. Journal of Memory and Language, 30(6), 725–745. Fodor, J. D. (1981). Does performance shape competence? Philosophical Transactions of the Royal Society London, B 295, 285–295. Frazier, L. (1979). On comprehending sentences: Syntactic parsing strategies. Unpublished doctoral dissertation, University of Connecticuts, CT. Frazier, L., & Rayner, K. (1988). Parameterizing the language processing system: Left- vs. right-branching within and across languages. In J. A. Hawkins (Ed.), Explaining language universals (pp. 247–279). Oxford: Blackwell. Greenberg, J. H. (1963). Some universals of grammar, with particular reference to the order of meaningful elements. In J. Greenberg (Ed.), Universals of language (pp. 73–113). Cambridge, MA: MIT Press. Haider, H. (1993). Deutsche Syntax – generativ. Vorstudien zu einer projektiven Theorie der Grammatik. Tu¨bingen: Narr. Hawkins, J. A. (1990). A parsing theory of word order universals. Linguistic Inquiry, 21(2), 223–261. Hawkins, J. A. (1994). A performance theory of order and constituency. Cambridge, MA: Cambridge University Press. Inoue, A., & Fodor, J. D. (1995). Information-paced parsing of Japanese. In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing (pp. 9–63). Hillsdale, NJ: Lawrence Erlbaum. Johnson, M. (1986). A GPSG account of VP structure in German. Linguistics, 24(5), 871–882. Kamide, Y., & Mitchell, D. C. (1999). Incremental pre-head attachment in Japanese parsing. Language and Cognitive Processes, 14(5), 631–662. MacDonald, M. C. (1994). Probabilistic constraints and syntactic ambiguity resolution. Language and Cognitive Processes, 9(2), 157–201. MacDonald, M. C., Pearlmutter, N., & Seidenberg, M. S. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101(4), 676–703. Mazuka, R., & Lust, B. (1990). On parameter setting and parsing: predictions for crosslinguistic differences in adult and child processing. In L. Frazier & J. de Villier (Eds.),
15
On Being Both Head-Initial and Head-Final
347
Language processing and language acquisition (pp. 163–205). Dordrecht, Boston & London: Kluwer. Miller, G. A., & Chomsky, N. (1963). Finitary models of language users. In R. D. Luce, R. R. Bush & E. Galanter (Eds.), Handbook of mathematical psychology (Vol. II, pp. 419–490). New York, London & Sidney: Wiley. Pritchett, B. L. (1988). Garden path phenomena and the grammatical basis of language processing. Language, 64(3), 539–576. Pritchett, B. L. (1992). Grammatical competence and parsing performance. Chicago and London: The University of Chicago Press. Steedman, M. (1983). On the generality of the nested-depedency constraint and the reason for an exception in Dutch. Linguistics, 21(1), 35–66. Trueswell, J. C., & Tanenhaus, M. K. (1994). Toward a lexicalist framework of constraintbased syntactic ambiguity resolution. In C. Clifton, Jr., L. Frazier & K. Rayner (Eds.), Perspectives on sentence processing (pp. 155–179). Hillsdale, NJ: Lawrence Erlbaum. Trueswell, J. C., Tanenhaus, M. K., & Kello, C. (1993). Verb-specific constraints in sentence processing: Separating effects of lexical preference from garden-paths. Journal of Experimental Psychology: Learning, Memory, and Cognition, 19(3), 528–553. von Stechow, A., & Sternefeld, W. (1988). Bausteine syntaktischen Wissens. Opladen: Westdeutscher Verlag. Vosse, T., & Kempen, G. (2000). Syntactic structure assembly in human parsing: a computational model based on competitive inhibition and a lexicalist grammar. Cognition, 75(2), 105–143. Warner, J., & Glass, A. L. (1987). Context and distance-to-disambiguation effects in ambiguity resolution: Evidence from grammaticalicity judgements of garden path sentences. Journal of Memory and Language, 26(6), 714–738. Wurmbrand, S. (2006). Verb clusters, verb raising, and restructuring. In M. Everaert & H. van Riemsdijk (Eds.), The Blackwell companion to syntax (Vol. 5, pp. 229–343). Oxford: Blackwell.
Chapter 16
Integration and Prediction in Head-Final Structures Shravan Vasishth
16.1 Introduction What determines incremental online processing difficulty? This is a central question in sentence comprehension research. In this paper I will briefly review the history of two phenomena – locality and interference – that together constitute a crucial piece of the incrementality puzzle. The main goal of this paper is to lay out some open research issues regarding these two phenomena. Incrementality becomes particularly interesting in head-final structures. If a verb occurs after its arguments rather than before, the human sentence parsing mechanism (presumably) faces greater demands than in non-head final structures: it must hold the arguments in memory as well as predictively build syntactic structure until the verb is processed. Indeed, the existence of headfinal languages like Japanese has occasionally led researchers (e.g., Pritchett, 1992) to propose a less incremental, head-driven parsing strategy where decisions about structure-building are postponed until the head is encountered. A great deal of the research on incrementality presupposes a universal parsing mechanism that applies equally to head-final and non-head final languages. However, there is no reason for this to be necessarily true. One alternative is that ‘‘the’’ human parsing mechanism has fundamentally different properties depending on the language being parsed. A third, intermediate position (which could turn out to be the correct one) is also possible: a univerally applicable core parsing architecture exists but the processing constraints are conditioned by underlying properties (such as head-finality) of a language. Such conditioning through word order constraints (which derive from grammar) could result in quite different parsing events in head-final languages compared to non-head final ones. It is in this logical space of possibilities that locality and similarity-based interference become relevant. S. Vasishth (*) Department of Linguistik, Exzellenzbereich Kognitionswissenschaft, Universita¨t Potsdam, D-14476 Potsdam, Germany e-mail: [email protected]
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_16, Ó Springer ScienceþBusiness Media B.V. 2011
349
350
S. Vasishth
My goal is to discuss some of the recent theories that have been proposed to explain locality and interference effects, and to summarize the cross-linguistic empirical base of these theories. In doing so, I hope to lay out the state of the art on the issue, and outline some of the important open empirical and theoretical issues, especially in relation to the processing of head-final languages. Although the facts I present are well known, the gaps in the theoretical debate on locality and interference effects may not be. Throughout this paper I use grayed-out boxes to provide additional discussions and more detailed definitions of concepts discussed in the text. These can be skipped by the reader without loss of flow; they are included merely to allow readers new to the area to obtain a quick overview of the theoretical claims mentioned in the text.
16.2 Locality in Sentence Comprehension Locality is the claim that the distance – however quantified – between a dependent and a head determines integration difficulty at the head. An example is the self-paced reading study by Grodner and Gibson (2005), which showed monotonically increasing reading time at the verb supervised as a function of the distance between the subject ‘nurse’ and the verb: (1) a. The nurse supervised the administrator while . . . b. The nurse from the clinic supervised the administrator while . . . c. The nurse who was from the clinic supervised the administrator while . . . Chomsky (1965, pp. 13–14) was perhaps the first to propose that the reduced acceptability of sentences containing a ‘‘nesting of a long and complex element’’ arises from ‘‘decay of memory.’’ In related work, Just and Carpenter (1980, 1992) directly address dependency resolution in sentence comprehension in terms of memory retrieval (similar early approaches are the production-system based models of Anderson, Kline, and Lewis, 1977). Just and Carpenter developed a model of integration that involved activation decay (as a side-effect of capacity limitations) as a key determinant of processing difficulty. For example, under the rubric of distance effects, they describe the constraints on dependency resolution as follows (1992, 133): The greater the distance between the two constituents to be related, the larger the probability of error and the longer the duration of the integration process.
The explanation for the distance effect in terms of activation decay was taken a great deal further in the Syntactic Prediction Locality Theory or SPLT (see Gibson, 1998, 9 for a historical overview of the connection between decay and
16
Integration and Prediction in Head-Final Structures
351
Fig. 16.1 A schematic illustration of DLT’s predictions for multiply embedded structures. Integration costs are labeled along the arcs that define the argument-head dependencies, computed by counting the number of intervening discourse referents. Another component of the theory is storage cost; these costs are presented under each verb for illustration. The storage costs are computed by counting the number of heads predicted at each point
distance) and, more recently, the Dependency Locality Theory or DLT (Gibson, 2000). The DLT proposes (among other things) that the cognitive cost of assembling a dependent with a head is partly a function of the number of new intervening discourse referents that are introduced between the dependent and the head; see Fig. 16.1 for an example. In effect, the DLT discretizes the concept of activation decay in the DLT complexity metric (Gibson, 2000, p. 103). The predictions of SPLT and DLT find quite good empirical support from online experiments involving English (e.g., Gibson & Thomas, 1999; Grodner & Gibson, 2005; Warren & Gibson, 2005) and also Chinese (Hsiao & Gibson, 2003) (see the papers on Chinese in this volume). At least one offline study involving Japanese is also consistent with the SPLT’s (the precursor of DLT) predictions (Babyonlyshev & Gibson, 1999). As mentioned above, locality cost is characterized by the DLT in terms of the number of discourse referents intervening between the dependent and the head. One may ask: what is so special about the number of new discourse referents? Why not count the number of intervening syntactic nodes, words, letters, syllables, etc.? The rationale within the DLT is that building discourse referents is computationally costly; independent evidence for this idea comes from studies showing that the accessibility of the intervening discourse referent (as defined by the accessibility hierarchy) can modulate retrieval difficulty (Warren, 2001; Warren & Gibson, 2005). In direct opposition to the locality hypothesis, Lewis (1996, p. 15) proposed that increased difficulty in resolving a long-distance dependency could at least in certain cases be attributable to syntactic interference. His proposal was that dependency resolution would become difficult if there are multiple intervening potential filler sites that correspond to the gap. This is how he explains the existence of the wh-island constraint (Ross, 1967): *Who does Phineas know a boy who hates the man who saw _? The sentence is perceived to be ungrammatical because of an upper bound on the number of similar filler sites for the wh-gap. In later work, Lewis and colleagues generalized the interference idea beyond
352
S. Vasishth
structure-based interference and suggest that the similarity of any kind of feature (not just syntactic) can make processing more difficult. For example, Lee, Nakayama, and Lewis (2005) found evidence for interference due to phonological similarity of case markers in Korean (cf. Vasishth 2003, which failed to find evidence for this idea in Hindi case-marking). The interference idea now exists in four distinct variants: the original conception by Lewis (1996), Van Dyke’s retrieval interference model based on Search of Associative Memory or SAM (Van Dyke, 2002) (discussed below), the cue-based retrieval model’s interference theory (Lewis & Vasishth, 2005) (which subsumes but goes beyond Lewis’ (1996) proposals), and Gordon and colleagues’ idea of interference due to feature-similarity of noun phrases (Gordon, Hendrick, & Johnson, 2001, 2004; Gordon, Hendrick, Johnson, & Lee, 2006; Gordon, Hendrick, & Levine, 2002). Are locality and interference two alternative explanations or do both the factors operate independently? It is plausible to assume that they are two independent factors, but I will return to this question in the next section. At this juncture I discuss some well-known but still open empirical problems with the locality hypothesis. A major issue is that locality does not seem to have much empirical support when we look beyond head-final structures in English (indeed, Jaeger, Fedorenko, Hofmeister, and Gibson (2008) have recently presented evidence that the locality constraint may not apply even in English, the language that has the most-attested instances of locality effects).
16.2.1 Counterexamples to Locality: Antilocality Konieczny (2000) presented an important counterexample to the locality hypothesis. In a self-paced reading study involving German center-embedded relative clauses, he showed that increasing argument-head distance, analogous to example 1 above, resulted in faster reading time at the verb, not slower, as predicted by locality. Konieczny’s explanation for the result was that the strength of prediction for the upcoming verb increases if more intervening material is present between the dependent and the head (he calls this the anticipation hypothesis). Konieczny’s finding would not necessarily have been so damaging to the locality idea: the effect he found could have been a consequence of confounding factors such as spillover from the region(s) preceding the verb, or due to differences in word position that result from the locality manipulation. Here, it is worth briefly considering these possibilities. Mitchell and Green (1978, p. 632) appear to have been the first to discuss the issue of spillover in print; they do so in the context of sentence-final wrap-up effects: One further possibility is that the pauses occurred because certain syntactic processes lagged behind the initial presentation of the material. For example, the subjects might have left some or all of the parsing to be carried out at the end of the clause.
16
Integration and Prediction in Head-Final Structures
353
Similarly, in later work Mitchell (1984, p. 76) writes: In most immediate processing tasks the end of one response measure is immediately followed by the beginning of another, together with a new portion of text. In this situation any uncompleted processing will spill over from one response measure to the next. In others words, certain aspects of processing will be postponed and join a queue or buffer so that they can be dealt with later. . . Here, the response measure will be influenced not only by the problems in the current display but also by any backlog or processing that may have built up in the buffer.
Spillover is a legitimate concern in Konieczny’s study, where the regions preceding the verb were not identical, and therefore differential amounts of spillover from the preverbal regions could well have been responsible for the antilocality effect. Another possible confound in Konieczny’s finding was that the location of the critical word was different in the local versus non-local condition. It could be argued as one reads a sentence from the left to right, reading time becomes faster and faster; this might be responsible for the facilitation Konieczny observed at the verb (note, however, that if this were a systematic speedup in reading regardless of language, one might ask why a speedup was not observed in the Grodner and Gibson 2005 study). The evidence for systematic speedups as a function of word position are based, however, on a misreading of the literature. The key finding was not, as is often claimed, that reading time steadily increases as one proceeds through the sentence but that average reading time is faster in longer sentences. The history of this idea is discussed elsewhere (Vasishth, 2003, pp. 170–185) so I will not repeat it here; rather, I simply note that although the alleged speedup-by-word-position is invoked as fact, as far as I know there exists no empirical demonstration in the literature that speedup generally occurs as one proceeds through a sentence, and there is at least one demonstration (Vasishth, 2003) that shows an absence of such a tendency. In particular, there is currently no evidence that the Konieczny result is a consequence of the alleged speedup-by-word-position. Such a demonstration would also have to explain why the locality manipulation of Grodner and Gibson (2005) is not subject to a word-position effect (i.e., why their results showed a locality effect). Following Konieczny’s work, Vasishth (2003) and Vasishth and Lewis (2006) presented further evidence from Hindi that increasing argument-head distance resulted in faster reading times at the verb; the latter work showed that such a speedup was seen even after spillover was taken into account as a covariate in the data analysis. They proposed an explanation for both locality and antilocality effects based on very general assumptions about forgetting (decay) and reactivation effects in working memory as derived from the ACT-R cognitive architecture (Anderson et al., 2004). Under this view, which is spelled out in a series of articles (Lewis & Vasishth, 2005; Lewis, Vasishth, & Van Dyke, 2006; Vasishth, Bruessow, Lewis, & Drenhaus, 2008; Vasishth & Lewis, 2006), instead of defining constraints on retrieval in terms of the number of intervening new discourse referents (DLT’s proposal), the cognitive costs of
354
S. Vasishth
dependency resolution are derived from an independently motivated theory of working memory retrieval: Dependents are retrieved through a content-based retrieval process that relies on cues expressed as feature-value specifications, and probability of correct retrieval and retrieval difficulty are dependent on the dynamic interaction of constraints on processes such as decay, reactivation, interference, partial-cue matching and stochastic noise. Locality effects arise if the dependent decays over time, but antilocality effects can arise if a to-beretrieved element is reactivated prior to its retrieval at the head (for details, see the articles cited above). Differences Between DLT and the ACT-R Model One might ask whether the DLT and ACT-R based model are identical theories. I point out some differences here. First, the ACT-R based model relies on a generally applicable theory of working memory (Anderson et al., 2004); by contrast, DLT instantiates a theory of decay specific to sentence processing – the number of discourse referents are counted to quantify distance. The decay equation in ACT-R would apply equally to (for example) list recall tasks, whereas the DLT instantiation of decay does not have a correlate in working memory tasks outside of sentence comprehension. The second difference is that the ACT-R theory of decay subsumes a reactivation component that is missing in the DLT. The ACT-R decay equation says that an item will decay in memory unless it is reused (reactivated), and the more often it is reactivated, the greater the boost in activation. The reactivation through reuse could account for the antilocality effects that serve as a counterexample to the DLT’s predictions (however, not all antilocality effects can be so explained, as discussed below). Independent evidence for reactivation effects comes from Hofmeister (2009). Hofmeister presents evidence from self-paced reading suggesting that in clefting constructions such as It was John who bought a book, the noun is retrieved faster at the verb when it is clefted versus nonclefted. He argues that the clefting boosts activation of the noun, resulting in faster retrieval at the verb. The third difference between the DLT and the ACT-R model is that the latter includes an interference theory; by contrast, DLT has no explanation for interference effects.
16.2.2 Surprisal: A New Explanation for Antilocality Effects An interesting theoretical development in the locality debate was a paper by Levy (2008). He proposed that antilocality effects could be explained by assuming that the material intervening between the dependent and head could serve to sharpen the expectation for the upcoming verb. This sharpened expectation emerges from the elimination of alternative possible parses. The expectation
16
Integration and Prediction in Head-Final Structures
355
hypothesis – which Levy argues is related to surprisal (Hale, 2001) – was argued to be the explanation for the antilocality effects seen in German (Koniechzny, 2000) and Hindi (Vasishth & Lewis, 2006). Surprisal Theory (Quoted from Boston, Hale, Patil, Kliegl, & Vasishth, 2008) The idea of surprisal is to model processing difficulty as a logarithmic function of the probability mass eliminated by the most recently added word. This number is a measure of the information value of the word just seen as rated by the grammar’s probability model; it is nonnegative and unbounded. More formally, define the prefix probability of an initial substring to be the total probability of all grammatical analyses that derive w ¼ w1 wn as a left-prefix (definition 1). Where the grammar G and prefix string w (but not w’s length, n) are understood, this quantity is abbreviated by the forward probability symbol, an . prefix-probabilityðw; GÞ ¼
X
ProbðdÞ ¼ an
(16:1)
d2DðG;wvÞ
(Note: In this definition, G is a probabilistic grammar; the only restriction on G is that it provide a set of derivations, D that assign a probability to particular strings. When DðG; uÞ ¼ ; we say that G does not derive the string u. The expression DðG; wvÞ denotes the set of derivations on G that derive w as the initial part of larger string, the rest of which is v. (See Charniak, 1993; Jurafsky & Martin, 2000; or Manning & Schu¨tze, 2000 for more details on probabilistic grammars.) Then the surprisal of the nth word is the log-ratio of the prefix probability before seeing the word, compared to the prefix probability after seeing it (definition 2). surprisalðnÞ ¼ log2
an1 an
(16:2)
As the logarithm of a probability, this quantity is measured in bits. Consider some consequences of this definition. Using a law of logarithms, one could rewrite definition 2 as log2 ðan1 Þ log2 ðan Þ But on a well-defined probabilistic grammar, the prefix probabilities a are always less than one and strictly nonincreasing from left to right. This implies that the two logarithms are to be subtracted in the opposite order. For instance, if a given word brings the prefix probability down from 0:6 to 0:01, the surprise value is 4:09 bits.
356
S. Vasishth
Intuitively, surprisal increases when a parser is required to build some low-probability structure. The key insight is that the relevant structure’s size need not be fixed in advance as with Markov models. An interesting prediction of Levy’s view is that English would also show antilocality effects. Indeed, Levy quotes a self-paced reading study by Jaeger and colleagues (also see the further studies in Jaeger et al., 2008) which, contrary to Grodner and Gibson’s study, confirmed this prediction. One novel aspect of Jaeger and colleagues’ study was that they tried to bring spillover under experimental control (as opposed to statistical control, as Vasishth and Lewis (2006) did). They manipulated dependent-head distance while holding the pre-verbal region constant; i.e., they tried to ensure that differential spillover from the preverbal region would be minimized (this assumes that spillover only occurs from the n 1th word to the nth word; although this is an assumption that does not hold in general, the design is a significant improvement on earlier work). Of course, the expectation account fails to explain the Grodner and Gibson locality effects in English; one possibility is that the Grodner and Gibson locality effect could be due to spillover. Thus, the story comes full circle. We began with the assumption that the locality hypothesis holds and that any demonstrations of antilocality could have been due to confounding factors such as spillover; with the Jaeger et al work, it appears that locality effects found in the literature may be due to confounding factors such as spillover and positional differences. Jaeger et al. have made significant advances in bringing the positional confound under experimental control, although they are still of course forced to control for spillover statistically. A central problem is that we have no theory of spillover. The current assumptions, discussed by Vasishth and Lewis (2006) and Jaeger et al. (2008), are simply that spillover occurs and that different stimulus items and participants may display different amounts of spillover; it is therefore taken into account as a covariate in the statistical data analysis. A better approach would be to flesh out Mitchell and Green’s original formulations into a process model of spillover; I believe such a model will help us address its role in reading experiments better than the current methodology. Once such a theory exists, its predictions could be taken into account in a more systematic manner in (implemented) models of parsing.
16.2.3 Is the Locality Effect an Illusion? Given the recent English findings of Jaeger and colleagues, one might be inclined to reject locality altogether as a constraint. There are at least two problems with dismissing the locality effect. First, Van Dyke and Lewis (2003) demonstrated the existence of a locality effect while experimentally controlling for the spillover confound. They conducted a self-paced study
16
Integration and Prediction in Head-Final Structures
357
involving sentences such as 2. One factor was ambiguity (presence/absence of the sentential complement that), and another was distance between an argument (here, the noun student) and verb (was standing). (2)
a. b. c. d.
The assistant forgot that the student was standing in the hallway. The assistant forgot the student was standing in the hallway. The assistant forgot that the student who knew that the exam was important was standing in the hallway. The assistant forgot the student who knew that the exam was important was standing in the hallway.
The ambiguity manipulation ensures that reanalysis takes place at was standing – the NP student must be reanalyzed as the subject of a sentential complement rather than the object of forgot. The distance manipulation ensures that the reattachment of the NP as subject of was standing is affected by locality. The reanalysis requires an integration between the verb and the argument, which is either near or distant from the verb. Consequently, if a significantly greater reanalysis cost is observed in the intervening-items conditions 2 c,d than in the non-intervening-items conditions 2 a,b, this would be a locality effect, and it would be independent of spillover confounds because the comparison is no longer a direct one between conditions with differing regions preceding the critical verb. The interaction was in fact observed in the Van Dyke and Lewis study, suggesting that locality can affect processing. Second, a puzzling asymmetry regarding the locality effect was found by Vasishth, Suckow, Lewis, and Kern (2010) between English and German in a series of experiments. They carried out self-paced and eyetracking studies investigating double center-embedding structures in English and German and found that, as predicted by the DLT and its earlier variants (Gibson & Thomas, 1999), omitting the middle verb in a double center embedding can result in facilitation in processing (compare the ungrammatical 3b with the missing middle verb with its grammatical but harder to process counterpart 3a). Vasishth et al. (2010) generally found faster reading times at the final verb and the region following it in the ungrammatical condition. (3)
a. b.
The apartment that the maid who the service had sent over was cleaning every week was well decorated. *The apartment that the maid who the service had sent over was well decorated.
DLT’s explanation for the facilitation in the ungrammatical English structure depends on locality assumptions: the long-distance argument-head dependency associated with the middle verb involves the greatest integration cost
358
S. Vasishth
compared to the other verbs, resulting in the parser forgetting the prediction for the VP corresponding to the second verb. This counterintintuitive result is an important piece in the locality puzzle. Interestingly, however, contrary to the DLT’s prediction German does not present the pattern seen in English: omitting the middle verb renders processing more difficult. German examples analogous to the English ones above are shown in 4.
(4)
a.
b.
Der Anwohner, den der Wanderer, den der Pfo¨rtner suchte, The resident that the hiker that the doorman searched-for sto¨rte, verarztete den Verletzten. disturbed tended-to the injured-person ‘The resident that the hiker that the doorman was looking-for disturbed tended to the injured person.’ *Der Anwohner, den der Wanderer, den der Pfo¨rtner suchte, The resident that the hiker that the doorman searched-for verarztete den Verletzten. tended-to the injured-person ‘The resident that the hiker that the doorman was looking-for tended to the injured person.’
Here, one might object that the obligatory presence of commas in German relative clauses might provide cues for remembering the serial position of the verb, resulting in facilitation of the grammatical structure, whereas in English the fact that commas did not delimit the relative clauses might be the reason that English speakers cannot process the grammatical structures quite as easily. However, a further English experiment was carried out that included commas in the English sentences; this study showed that English speakers persist in forgetting the middle verb (see Vasishth et al., 2010 for details). Based on these studies, Vasishth and colleagues hypothesized that German speakers are able to maintain predictions of upcoming verbs much better than speakers of a non-head-final language like English. In other words, the absence of the locality effect in German seems to have its source in some other factor: The high frequency of verb-final structures in German could result in German speakers being more practised than English in maintaining the prediction of the upcoming verb (see Engelmann & Vasishth, 2009 for a connectionist model that correctly predicts the English and German asymmetry). If this language-based difference turns out to be valid, a languageindependent account of locality/antilocality effects – in particular, one that applies to both head-final and non-head-final languages – is unlikely to furnish a complete explanation.
16
Integration and Prediction in Head-Final Structures
359
16.2.4 Working Memory Structures and Locality: An Unexplored Puzzle A further open question relating to the locality/antilocality and interference literature is that the working memory mechanisms that have been proposed by models like the DLT and the cue-based retrieval model are at odds with other related work in the working memory and language processing literature. Using speed-accuracy tradeoff and eyetracking studies, McElree and colleagues (McElree, 2000, 2006; McElree, Foraker, & Dyer, 2003; Foraker & McElree, 2007) have argued that increasing dependent-head distance does not affect its accessibility (the speed of the dependency resolution process), but does affect the availability (the probability that the dependent would be retrieved). If McElree and colleagues are right, locality and interference effects (which are reflected in longer reading times) are due to availability, not accessibility. However, none of the current sentence comprehension models draw out the consequences of this presumed difference between availability and accessibility. In fact, the working memory theory of ACT-R, although similar to McElree and colleagues’ in assuming a content-addressable architecture, directly contradicts the availability-accessibility distinction because it assumes a close relationship between probability of retrieval and retrieval latency: they are inversely related. If an item has a lower retrieval probability than another item, its activation will on average will be lower; since retrieval latency depends on activation, on average retrieval latency will be higher. In essence, the account by McElree and colleagues is a fourth theory of dependency resolution, and it predicts neither locality nor antilocality effects. Their account would ascribe locality effects in reading studies to reduced availability as distance is increased (McElree et al., 2003). It is difficult to test their claims relative to the other theories using reading-time and eyetracking studies because the key differences (availability versus accessibility) are not directly measurable in such studies (cf. Foraker & McElree, 2007). However, even in the theoretical literature there are no hypotheses about what the locality and antilocality effects index (availability or accessibility). I believe this is a critical gap in the locality debate.
16.2.5 Locality: A Summing Up To sum up the discussion so far, both locality and antilocality effects have been attested in English, and mainly antilocality effects have been seen in German and Hindi (cf. (Sommerfeld, Vasishth, Logacˇev, Baumann, & Drenhaus, 2007) and (Drenhaus, Vasishth, Wittich, & Patil, 2007) for evidence for locality effects in German). The candidate theories explaining locality effects are the DLT and the cue-based retrieval (ACT-R based) model, and at least two competing explanations for antilocality effects have been offered: the ACT-R based
360
S. Vasishth
model and surprisal. Unfortunately, the empirical data do not match any of the theories: DLT uniformly fails to explain antilocality effects, the ACT-R model fails to explain the antilocality effects Jaeger and colleagues found in English, and all three theories are unable to explain the asymmetry of the missing verb effect (Vasishth et al., 2010), which seems to be grounded in structural frequency differences. It is important to note here that the predictions of Levy’s expectation theory are not so easy to determine, and although Levy relates the idea to surprisal, it is not clear to me that surprisal and expectation theory predict the same things. The predictions of surprisal as a scalable computational theory of sentence processing difficulty have been investigated recently in several recent articles (Boston, Hale, Patil, Kliegl, Vasishth, 2008; Boston, Hale, Kliegl, & Vasishth, 2008; Demberg & Keller, 2008); all these articles rely on treebank-based estimates of surprisal. The predictions of expectation theory are derived quite differently. For example, Jaeger et al. (2008) base their predictions of expectation theory on introspection, not on a computation derived from a probabilistic context-free grammar. There are of course good reasons not to compute surprisal from a treebank corpus: the theoretically interesting structures may occur so rarely in the corpus that a meaningful calculation may not be possible. However, in its present form expectation theory as conceived by Levy does not provide any objective way to derive a prediction (cf. DLT and other computational models).
16.2.6 Solving the Locality Puzzle What can be done to disentangle the predictions of these competing theories? The most obvious point is that we need to develop non-messy experimental designs that manipulate locality without introducing the word position and spillover confounds; until we can resolve these confounds, we don’t really know what the evidence is. Note that bringing position and/or spillover under statistical control is a start but insufficient; these factors need to be brought under experimental control. Second, the debate has centered around only a few languages, a limited range of syntactic structures, and only two methodologies, self-paced reading and eyetracking. Missing is more evidence from a variety of languages, involving different methods, including speed-accuracy tradeoff studies and event-related potentials. It is here that a closer study of head-final languages other than German and Hindi can be informative. Third, in the comparison between DLT, the ACT-R based model and surprisal one detail needs to be noted. The first two models are ‘‘backward-looking’’ theories: they define the constraints on retrieving previously processed or encoded material in working memory. By contrast, surprisal is a ‘‘forward-looking’’ metric: it defines the processing cost of predicting upcoming material (Demberg & Keller, 2008; Levy, 2008). Theoretically, these are two orthogonal classes of explanation;
16
Integration and Prediction in Head-Final Structures
361
a priori, they are not competing explanations, but rather explain different aspects of processing difficulty. This observation is also discussed in Sommerfeld et al. (2007) and Jaeger et al. (2008). One implication is that the relative contributions of these theories can be empirically evaluated; see Boston, Hale, Vasishth, and Kliegl (in press) for one such attempt; this work showed that retrieval cost and surprisal may be independent contributors of processing difficulty. We turn next to another theoretical position that relates closely to the locality issue – similarity-based interference.
16.3 Interference in Sentence Comprehension Interference is the claim that memory traces with similar feature specifications can render integration more difficult due to the reduced ability to distinguish between the target of integration and the similar non-target items. Extensive evidence exists for similarity-based interference in sentence comprehension: for example, (see Gordon et al., 2001, 2002, 2004, 2006; Lee et al., 2005; Lewis, 1996; Van Dyke, 2007; Van Dyke & Lewis, 2003; Van Dyke & McElree, 2006). Interestingly, however, the differences between competing explanations for interference effects have not yet been discussed in the literature, and several misunderstandings about the differences between these interference-accounts persist in published work. Consider the interference theory proposed by Gordon and colleagues. They assume that the featural properties of noun phrases (NPs) cause interference when the target NP is to be retrieved, e.g., at a verb. For example, consider example 5. (5)
The banker [that praised the barber/Sophie] climbed the mountain just outside of town.
Here, the subject relative clause has a grammatical object that is either a definite description, the barber or a proper name, Sophie. The claim is that type identity of the object with respect to the subject, the banker, renders it more difficult to integrate with the verb. Following up on their previous work, Gordon and colleagues conducted a reading study using eyetracking that provided evidence consistent with this claim: longer reading times (gaze durations, rightbounded reading time and re-reading time) were found in the relative clause (RC) region (marked in square brackets); no effect of similarity was found at the matrix verb ‘climbed’. As they put it (Gordon et al., 2006, 1309): The finding that interference . . . occurs in close proximity to the embedded and matrix verbs is consistent with the idea that the similarity-based interference occurs at the time of memory retrieval, as has also been indicated by work manipulating memory load during self-paced reading.
362
S. Vasishth
The emphasis is mine. The above quote raises the question: what is the precise relationship between the feature specifications of NPs and the retrieval process at the verb? Gordon and colleagues assume that featural similarities among the NPs somehow come into play at the moment of retrieval. But it is not clear how or why this happens. In the interference theory proposed by Gordon and colleagues, type similarity of NPs has been invoked as a source of interference during retrieval, but what does type-similarity have to do with retrieval? I argue below that Gordon and colleagues’ type-similarity assumption distinguishes their theory from other interference accounts. An alternative interference theory to the proposal by Gordon and colleagues has been offered in Lewis and Vasishth (2005); Lewis et al. (2006); Van Dyke (2007); Van Dyke and Lewis (2003); Van Dyke and McElree (2006). The proposal, which has been referred to as the cue-based retrieval theory of interference, is that the retrieval event at the verb triggers interference due to the subcategorization requirements of the verb. For example, a verb such as kissed requires a human-referring agent; consequently, integrating the verb with its subject involves a search for a noun phrase that refers to a human entity and is consistent with the agent role. Under this view, the retrieval cues set by the verb preparatory to completing the integration could be responsible for interference effects.
Van Dyke’s Retrieval Interference Theory In Van Dyke’s model, retrieval cues are combined into a retrieval probe, which determines the strength of association between each probe cue and memory traces of items in memory. The probability of retrieving an item Ii given probe cues (Q1 ; ; Qm ) is a function of the strength of association S between each probe cue Qj and the features of the memory trace, denoted by SðQj ; Ii Þwj , where wj is a weighting factor denoting the relative salience of the different cues. Equation (3) formalizes this: m Q
PðIi j Q1 Qm Þ ¼
SðQj ; Ii Þwj
j¼1 N Q m P
(16:3) SðQj ; Ik Þ
wj
k¼1 j¼1
This equation quantifies the probability of retrieval of an item Ii given cues (Q1 ; ; Qm ) as the proportion of the total strength of association for the item Ii (computed as the product of the strengths of association of the probe cues with the item), with respect to the sum of all such strengths of associations for all items Ik. Van Dyke’s interference theory is the only one among the candidate theories that defines interference explicitly in terms of retrieval probability (cf. the discussion about availability versus accessibility above).
16
Integration and Prediction in Head-Final Structures
363
A self-paced reading study described in (Van Dyke & McElree, 2006) illustrates the key idea. In this study, participants had to memorize triplets of nouns such as table, sink, truck (see 6), and then read target sentences that had a highinterference or low-interference manipulation. The high-interference target sentences had a critical verb, e.g., fixed, that could plausibly take as subject any of the three nouns in the memorized triplet, as well as the NP in the target sentence, boat, 6b. The low-interference target sentences had a critical verb, e.g., sailed, that could plausibly take as its subject none of the three nouns in the memorized triplet, but it could have as subject the NP present in the target sentence, boat, 6b. (6)
a. b.
c.
Memorized set: TABLE SINK TRUCK High interference It was the boat that the guy who lived by the sea fixed in two sunny days. Low interference It was the boat that the guy who lived by the sea sailed in two sunny days.
Compared to baseline conditions that required no list memorization, Van Dyke and McElree found longer reading times at the verb fixed compared to sailed. To summarize, two distinct explanations have been offered for interference effects. The feature-similarity theory of Gordon and colleagues is about the features common to NPs but not necessarily relevant to the retrieval process per se, while the cue-based retrieval theory is about the features directly triggering the retrieval process. Clearly, the two theories explain qualitatively different empirical phenomena: the feature-similarity based theory cannot explain the Van Dyke and McElree results, while the cue-based retrieval theory cannot explain the interference effect found by Gordon and colleagues. Apart from the orthogonality in empirical coverage, there is a further important difference between the featureoverlap and cue-based retrieval accounts. Interference due to feature-similarity entails that the ease with which an item is encoded and maintained in memory depends on its featural similarities with previously encountered items (Lewis et al., 2006). There are different ways in which such feature-similarity could affect processing adversely: The features representing the new item could partially overwrite features of the old items, rendering the older items harder to maintain in memory; or the new item could compete with old items for common features, making it harder to encode the new item; or the new item could entirely displace older items from memory (Jonides et al., 2008; Lewis et al., 2006) (cf. Oberauer & Kliegl, 2006). When we consider that the feature-overlap theory implies the existence of interference effects even before retrieval occurs, a natural question arises: can
364
S. Vasishth
the interference effect observed by Gordon and colleagues plausibly be attributed to the encoding stage (during the processing of the second NP) rather than at the retrieval stage (during processing at the verb)? Gordon and colleagues’ findings could equally be explained as interference during encoding: they analyze reading times over the entire relative clause region, which contains both the second NP and the verb, making it impossible to determine whether the effect was occurring at the NP or the verb. Indeed, in earlier work (2004) they considered the possibility that locus of interference effects may occur even earlier than during the retrieval event; however, they concluded that the 2004 experimental manipulations ‘‘do not allow us to ascertain definitively the locus of interference within working memory’’ (2004, p. 112). This difficulty in isolating the locus of interference also holds for their 2006 experiment discussed above. Thus, an important open question is disentangling the claims of the two classes of interference theory. Finally, Logacˇev and Vasishth (2010) show that all existing interference theories are currently unable to explain the fact that a match between two noun phrases along two dimensions can facilitate rather than hinder processing – what Logacˇev and Vasishth refer to as similarity-based facilitation. All existing theories of interference incorrectly predict greater rather than reduced difficulty. Logacˇev and Vasishth propose an explanation for this puzzle based on an assumption about the nature of memory representations: conflicting bindings (Hommel, 1998). They show that, under this representational assumption, not only can the published similarity-based interference facts be explained but also the similarity-based facilitation findings.
16.4 Concluding Remarks I have summarized some of the recent work on locality and interference effects, especially as they relate to head-final structures. Regarding interference theories, there appear to be several distinct theories which have – at least in principle – empirically distinguishable impact on processing difficulty. These differences have apparently not been noted in the literature. Regarding locality, it is clear that current explanations – I have discussed only three in some detail, DLT, the ACT-R-based model, and surprisal – need to be teased apart empirically. As discussed above, surprisal is a ‘‘forwardlooking’’ metric and the other two theories mainly refer to retrieval of already processing material; they propose ‘‘backward looking’’ metrics. In principle, they are orthogonal explanations and therefore their impact should be independent (Boston, Hale, Vasishth, & Kliegl, in press). One related problem is that we do not yet understand whether locality effects are confounded with position-effects and spillover. Another open issue is that we do not understand
16
Integration and Prediction in Head-Final Structures
365
what the implications for theories of locality (and interference) are of McElree and colleagues’ claims about retrieval accuracy versus latency. Thus, many important research questions remain open in the locality and interference debates. All these issues are of great relevance to research on head-final languages. Head-final languages like German and Hindi have played an important role in discovering the limits of locality theories. However, missing in the literature is a broader study of head-final languages in the context of phenomena like locality and interference. Acknowledgements I am grateful to Ted Gibson, Florian Jaeger, Roger Levy, Richard L. Lewis, Don Mitchell and Julie Van Dyke for discussions about locality and interference theories over the last few years. I also benefitted from many comments from, among others, Mike Tanenhaus, Florian Jaeger and Jeff Runner. I also thank my students Esther Sommerfeld, Pavel Logacˇev, Umesh Patil and Titus von der Malsburg for extended discussions regarding the issues presented here. Marisa Ferrara Boston contributed valuable comments to an earlier draft of this paper.
References Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S., Lebiere, C., & Qin, Y. (2004). An integrated theory of the mind. Psychological Review, 111(4), 1036–1060. Anderson, J. R., Kline, P., & Lewis, C. (1977). A production system model for language processing. In M. Just & P. Carpenter (Eds.), Cognitive processes in comprehension. Hillsdale, NJ: Lawrence Erlbaum Associates. Babyonyshev, M., & Gibson, E. (1999). The complexity of nested structures in Japanese. Language, 75(3), 423–450. Boston, M. F., Hale, J. T., Kliegl, R., & Vasishth, S. (2008). Surprising parser actions and reading difficulty. Proceedings of ACL-08: HLT Short Papers (pp. 5–8). Columbus, OH: The Ohio State University. Boston, M. F., Hale, J. T., Patil, U., Kliegl, R., & Vasishth, S. (2008). Parsing costs as predictors of reading difficulty: An evaluation using the Potsdam Sentence Corpus. Journal of Eye Movement Research, 2(1), 1–12. Boston, M. F., Hale, J. T., Vasishth, S., & Kliegl, R. (in press). Parallelism and syntactic processes in reading difficulty. Language and Cognitive Processes. Charniak, E. (1993). Statistical language learning. Cambridge, MA: MIT Press. Chomsky, N. (1965). Aspects of the theory of syntax. Cambridge, MA: MIT Press. Demberg, V., & Keller, F. (2008). Data from eye-tracking corpora as evidence for theories of syntactic processing complexity. Cognition, 109(2), 193–210. Drenhaus, H., Vasishth, S., Wittich, K., & Patil, U. (2007). Locality and working memory capacity: An ERP study of German. Proceedings of the AMLaP conference, Turku, Finland. Engelmann, F., & Vasishth, S. (2009). Processing grammatical and ungrammatical center embeddings in English and German: A computational model. Proceedings of international conference on cognitive modeling, Submitted to ICCM 2009. Foraker, S., & McElree, B. (2007). The role of prominence in pronoun resolution: Active versus passive representations. Journal of Memory and Language, 56(3), 357–383. Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68, 1–76.
366
S. Vasishth
Gibson, E. (2000). Dependency locality theory: A distance-based theory of linguistic complexity. In A. Marantz, Y. Miyashita & W. O’Neil (Eds.), Image, language, brain: Papers from the first mind articulation project symposium. Cambridge, MA: MIT Press. Gibson, E., & Thomas, J. (1999). Memory limitations and structural forgetting: The perception of complex ungrammatical sentences as grammatical. Language and Cognitive Processes, 14(3), 225–248. Gordon, P. C., Hendrick, R., & Johnson, M. (2001). Memory interference during language processing. Journal of Experimental Psychology: Learning, Memory and Cognition, 27(6), 1411–1423. Gordon, P. C., Hendrick, R., & Johnson, M. (2004). Effects of noun phrase type on sentence complexity. Journal of Memory and Language, 51, 97–104. Gordon, P. C., Hendrick, R., Johnson, M., & Lee, Y. (2006). Similarity-based interference during language comprehension: Evidence from eye tracking during reading. Journal of Experimental Psychology: Learning Memory and Cognition, 32(6), 1304–1321. Gordon, P. C., Hendrick, R., & Levine, W. H. (2002). Memory-load interference in syntactic processing. Psychological Science, 425–430. Grodner, D., & Gibson, E. (2005). Consequences of the serial nature of linguistic input. Cognitive Science, 29, 261–290. Hale, J. T. (2001). A probabilistic Earley parser as a psycholinguistic model. Proceedings of the Second Meeting of the North American Chapter of the Association for Computational Linguistics, Pittsburgh, PA. Hofmeister, P. (2009). Encoding effects on memory retrieval in language comprehension. Proceedings of cuny conference. Davis, CA: University of Davis. Hommel, B. (1998). Event files: Evidence for automatic integration of stimulus response episodes. Visual Cognition, 5, 183–216. Hsiao, F. P.-F., & Gibson, E. (2003). Processing relative clauses in Chinese. Cognition, 90, 3–27. Jaeger, F. T., Fedorenko, E., Hofmeister, P., & Gibson, E. (2008). Expectation based syntactic processing: Antilocality outside of head-final languages. Cuny sentence processing conference, NC. Jonides, J., Lewis, R. L., Nee, D. E., Lustig, C. A., Berman, M. G., et al. (2008). The mind and brain of short-term memory. The Annual Review of Psychology, 59, 193–224. Jurafsky, D., & Martin, J. H. (2000). Speech and language processing: An introduction to natural language processing, computational linguistics and speech recognition. Upper Saddle River, NJ: Prentice-Hall. Just, M., & Carpenter, P. (1980). A theory of reading: From eye fixations to comprehension. Psychological Review, 87(4), 329–354. Just,M., & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in working memory. Psychological Review, 99(1), 122–149. Konieczny, L. (2000). Locality and parsing complexity. Journal of Psycholinguistic Research, 29(6), 627–645. Lee, S.-H., Nakayama,M., & Lewis, R. L. (2005). Difficulty of processing Japanese and Korean center-embedding constructions. In M. Minami, H. Kobayashi, M. Nakayama & H. Sirai (Eds.), Studies in language science (Vol. 4, pp. 99–118). Tokyo: Kurosio Publishers. Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106, 1126–1177. Lewis, R. L. (1996). Interference in short-term memory: The magical number two (or three) in sentence processing. Journal of Psycholinguistic Research, 25(1), 93–115. Lewis, R. L., & Vasishth, S. (2005, May). An activation-based model of sentence processing as skilled memory retrieval. Cognitive Science, 29, 1–45. Lewis, R. L., Vasishth, S., & Van Dyke, J. (2006). Computational principles of working memory in sentence comprehension. Trends in Cognitive Sciences, 10(10), 447–454.
16
Integration and Prediction in Head-Final Structures
367
Logacˇev, P., & Vasishth, S. (2010). Morphological ambiguity and working memory. In P. de Swart & M. Lamers (Eds.), Case, word order, and prominence: Psycholinguistic and theoretical approaches to argument structure. Berlin: Springer. Manning, C. D., & Schu¨tze, H. (2000). Foundations of statistical natural language processing. Cambridge, MA: MIT Press. McElree, B. (2000). Sentence comprehension is mediated by content-addressable memory structures. Journal of Psycholinguistic Research, 29(2), 111–123. McElree, B. (2006). Accessing recent events. In B. H. Ross (Ed.), The psychology of learning and motivation (Vol. 46). San Diego, CA: Academic Press. McElree, B., Foraker, S., & Dyer, L. (2003). Memory structures that subserve sentence comprehension. Journal of Memory and Language, 48, 67–91. Mitchell, D. C. (1984). An evaluation of subject-paced reading tasks and other methods of investigating immediate processes in reading. In D. E. Kieras & M. Just (Eds.), New methods in reading comprehension research. Hillsdale, NJ: Erlbaum. Mitchell, D. C., & Green, D.W. (1978). The effects of context and content on immediate processing in reading. Quarterly Journal of experimental Psychology. Oberauer, K., & Kliegl, R. (2006). A formal model of capacity limits in working memory. Journal of Memory and Language. Pritchett, B. L. (1992). Grammatical competence and parsing performance. Chicago: University of Chicago Press. Ross, J. R. (1967). Constraints on variables in syntax. Unpublished doctoral dissertation, Massachusetts Institute of Technology, Cambridge, MA. Sommerfeld, E., Vasishth, S., Logacˇev, P., Baumann, M., & Drenhaus, H. (2007). A twophase model of integration processes in sentence parsing: Locality and antilocality effects in German. Proceedings of the CUNY sentence processing conference, La Jolla, CA. Van Dyke, J. (2002). Parsing as working memory retrieval: Interference, decay, and priming effects in long distance attachment. Unpublished doctoral dissertation, University of Pittsburgh, PA. Van Dyke, J. (2007). Interference effects from grammatically unavailable constituents during sentence processing. Journal of Experimental Psychology Learning Memory Cognition, 33(2), 407–30. Van Dyke, J., & Lewis, R. L. (2003). Distinguishing effects of structure and decay on attachment and repair: A cue-based parsing account of recovery from misanalyzed ambiguities. Journal of Memory and Language, 49, 285–316. Van Dyke, J., & McElree, B. (2006). Retrieval interference in sentence comprehension. Journal of Memory and Language, 55, 157–166. Vasishth, S. (2003). Working memory in sentence comprehension: Processing Hindi center embeddings. New York: Garland Press. (Published in the Garland series Outstanding Dissertations in Linguistics, edited by Laurence Horn) Vasishth, S., Bruessow, S., Lewis, R. L., & Drenhaus, H. (2008). Processing polarity: How the ungrammatical intrudes on the grammatical. Cognitive Science, 32(4), 533–567. Vasishth, S., & Lewis, R. L. (2006). Argument-head distance and processing complexity: Explaining both locality and antilocality effects. Language, 82(4), 767–794. Vasishth, S., Suckow, K., Lewis, R., & Kern, S. (2010). Short-term forgetting in sentence comprehension: Crosslinguistic evidence from head-final structures. Language and Cognitive Processes, 25(4), 533–567. Warren, T. C. (2001). Understanding the role of referential processing in sentence complexity. Unpublished doctoral dissertation, Massachusetts Institute of Technology, Cambridge, MA. Warren, T. C., & Gibson, E. (2005). Effects of NP type in reading cleft sentences in English. Language and Cognitive Processes, 20(6), 751–767.
Chapter 17
Directionality in the Architecture of the Language Faculty: Integrating with Real Time Barbara Lust1
17.1 Introduction The fundamental question we are concerned with is: What is the nature of, or role of, directionality in natural language, and more specifically in the Language Faculty? We assume it is the Language Faculty that makes language knowledge and language acquisition possible. By directionality, we refer to what is generally referred to as ‘‘right’’ or ‘‘left’’ in spatial terms, or ‘‘before’’ and ‘‘after’’ in temporal terms.2 Thus ‘‘directionality’’ is that which determines linear order of elements, i.e., whether an item precedes or follows another. While we know that directionality is fundamental to biological and physical reality (McManus, 2002), our question is: What is its position and role in cognitive reality, in linguistic representation, or more specifically, in the human competence for linguistic representation. Questions regarding the representation of linear order in natural language knowledge have long been fundamental both in typological linguistics and psycholinguistics, e.g., classic papers by Greenberg (1963) and Lashley (1948/ 1951) respectively, as well as in linguistic theory. Issues of language processing arise intrinsically when we consider directionality, since language processing – whether instantiated in terms of language production or language comprehension – inherently involves the instantiation of language in real time and at some level the human language processing
B. Lust (*) Department of Human Development, College of Human Ecology, Cornell University, MVR, HD, G57, Ithaca, NY, 24853 USA e-mail: [email protected] 1 We thank James Gair, Michael Wagner, Suzanne Flynn, Claire Foley and Jerry Packard for discussion of the issues raised in this paper. 2 For purposes of simplification we will use the terms, ‘‘right’’ and ‘‘left’’ here, although these spatial terms to some degree beg the issues we will raise.
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7_17, Ó Springer ScienceþBusiness Media B.V. 2011
369
370
B. Lust
mechanism must work from left to right or before to after.3 Issues of processing arise also because in any realistic model of language acquisition, the child must map between whatever is programmed in the Language Faculty to experience the speech stream in real time. In this paper, we will first review our basic assumptions regarding Universal Grammar (UG) with a view to explicating the foundations of the issues regarding the positioning of directionality in UG (Section 17.2). We will assume UG to provide a leading theory of the Language Faculty, one which can provide precise hypotheses regarding its architecture. We will articulate two currently opposing positions regarding the formulation of directionality in UG and show that they critically involve the issue of the relation of UG to real time (Section 17.3). We will then consider deriving empirical predictions from these diverse interpretations of UG (Section 17.4) and briefly consult typological (1.4.1), language acquisition (1.4.2 and 1.4.3) and language processing data (1.4.4) in this regard. Our results (1.5) call into question current assumptions regarding the nature of UG and its fundamental architecture as well as its relation to human cognition in general.
17.2 Assumptions Regarding UG We will assume a working definition of UG as in its initial and now classic formulation, as in 1: (1) UG a. ‘In a highly idealized picture of language acquisition, UG is taken to be a characterization of the child’s pre-linguistic initial state’ (Chomsky, 1981, p. 7) b. ‘. . .UG is a theory of the initial state S0 of the relevant component of the language faculty’ (Chomsky, 1993, p. 1) c. ‘. . .S0 is the ‘initial state’, prior to any language learning. . .’ (Chomsky, 1975, p. 119) d. ‘Linguistic theory, the theory of UG, . . ., is an innate property of the human mind. In principle, we should be able to account for it in terms of human biology’ (Chomsky, 1975, p. 34) e. ‘What many linguists call ‘universal grammar’ may be regarded as a theory of innate mechanisms, an underlying biological matrix that provides a framework within which the growth of language proceeds’ (Chomsky, 1980, p. 187); ‘. . .an abstract partial specification of the genetic program that enables the child to interpret certain events as linguistic experience and to construct a system of rules and principles on the basis of this experience’ (Chomsky, 1980, p. 187) 3
Although any model of language processing must operate on the left to right process of speech (whether oral or visual as in sign language) in real time, we know that knowledge regarding language need not. The productivity of anticipatory speech errors suggests this.
17
Directionality in the Architecture of the Language Faculty
371
In general, then, we assume the definition of UG to be at one and the same time both (i) a theory of the fundamental principles and parameters that underlie all possible natural languages; and (ii) a theory of what is biologically programmed in the Initial State of the Language Faculty (cf. Lust, 2006). In this way, UG is both (i) a theory concerning abstract linguistic representation and (ii) a theory concerning the physical reality of biological programming that makes language acquisition and language knowledge empirically possible. We may assume that, given this bidimensional definition, if UG is valid, it must have empirical consequences both for the proper representation of variation observed in the grammars of the world’s languages and for their acquisition. If so, it must have consequences for language processing. The variation in directionality across languages points up this challenge. Given our assumed definition of UG, not only facts of syntax but also facts of both language typology regarding directionality and facts of language acquisition should follow empirically from the formal articulation of UG. The fundamental problem of learnability of natural language directionality should be rendered tractable. Pursuing these additional dimensions requires both precise formalization of the grammatical proposal, including empirical predictions and precision with regard to the interpretation of empirical data.
17.3 The Grammatical Representation of Directionality Attempts to represent directionality in natural language, e.g., 2, have varied with regard to their fundamental assumptions. Generalizations such as those in 2a and 2b are based on observed facts regarding the world’s languages; while generalizations such as those attempted in proposals like 2c, d and e are based on theories of grammar and in fact not merely specific language grammar but Universal Grammar. We will focus here on the proposals in 2c, d and e in order to raise the issues regarding the Language Faculty that we are concerned with. In these theories of Universal Grammar (UG), alternative views have appeared. They differ precisely with regard to how and where real time is integrated in linguistic representation. (2)
Proposed Principles of Directionality: Selected Examples a. Word Order Universals: ‘General Framework of Harmonic Relations’ (Greenberg, 1963, p. 101)
For example, languages that are SOV are often postpositional, modifiers precede the noun, adverbial adjuncts precede main verb. This concept is ‘‘very obviously connected with the psychological concept of generalization’’ (Greenberg, 1963, p. 97; cf. Croft, 1990, p. 54 f. for explication). b. Cross-Categorical Harmony (CCH) (Hawkins, 1985)
372
B. Lust
‘‘Underlying the relative sizes of the word order co-occurrence types there emerges a distributional generalization which I will call the CCH’’ (1983, 1983, pp. 133F, 133F). ‘‘It asserts that a quantifiable preference exists for the position of the head relative to its modifiers within one phrasal category (i.e., the NP, the VP, the adjective phrase, or the adposition phrase) to generalize to the others’’ (1985, p. 571). c. Head Direction Parameter (Chomsky, 1988) In phase structure rules, the head may be left or right terminal in the phrases, according to language. (See also: Baker, 2001, 2008; Fukui & Saito (1997)/Saito & Fukui, 1998; Fukui & Takano, 1998, 2006.) d. Fractionation of parameters (e.g., Travis, 1989, 1991) Head Direction may differentiate by; Theta-role Assignment Direction; Case Marking Direction or other grammatical factor. e. Antisymmetry proposal (Kayne, 1994)4 UG does not make both right headed and left headed phrasal orders available equivalently; one order is unmarked. There is no directionality parameter.
17.3.1 The Classic Program in UG: Head Direction Parameter (e.g., 2c,d) In the now classic UG based proposal represented in 2c and d, specific language grammars are free to vary as ‘‘head initial’’ (e.g., English, Arabic, Indonesian) or ‘‘head final’’ (e.g., Japanese, Korean, Tamil, Malayalam); linear order is determined by parameter setting through experience.5 Both orders are available equivalently since both parameter settings are equally possible grammatically (e.g., Chomsky, 1988, p. 67f.; Smith, 1999 for discussion; Baker, 2001). UG here is viewed as ‘‘prior to experience’’ and in this way we would assume ‘‘prior to time’’ (e.g., 1 above). Thus time is irrelevant to the grammatical representation provided by UG. Directionality is unspecified in UG and purely arbitrary in biologically programmed grammatical competence for natural language. Mapping from grammar to specific language, which necessarily takes place 4 Another antisymmetry proposal is possible, e.g., Haider (2000). For Haider and Kayne, opposite orders, OV or VO respectively, are proposed to be unmarked. We will concentrate here on the Kayne version of antisymmetry, because of the precision with which it specifies its integration with real time. 5 The definition of ‘‘head’’ (e.g., functional or lexical) is critical to this theory. We leave aside this issue here and pursue it elsewhere (Lust, in prep.) Also, the issue of the ‘‘branching direction’’ of the language is frequently confounded with the issue of ‘‘head direction’’; we leave this correlation aside here also.
17
Directionality in the Architecture of the Language Faculty
373
through experience in real time, will necessarily result in establishment of one order or another.6 But the choice will be arbitrary grammatically. Thus, either right headed or left headed languages are equally possible. In this view, what UG provides is universal linguistic principles such as ‘‘structure dependence’’, assumed to provide knowledge of hierarchical relations defined configurationally and operations over these. UG also provides the systematicity leading to the generalizations noted in 2a and b above, i.e., to the widespread structural harmonies and constrained syntax that specific language grammars reveal, whether head initial or head final. Deviations, i.e., disharmonic orders, must involve explicit learning. Such disharmonic orders arise in the mapping from Universal Grammar to specific language both in language acquisition and in typological variations. Asymmetries characterize right and left headed languages. For example, the behavior of Wh-interrogative phenomena, or the behavior of subject anaphors (e.g., Kayne, 1994, pp. 50–54 for early articulation of several of these) differ across them. On the classic view of UG, such asymmetries between right and left headed languages must arise not at the level of UG but at the level of specific languages that are expressed in real time. They may be computed as ‘‘deductive consequences’’ of parameter setting. In this classical view, the issues that persist include how precisely to define the parameter of variation, e.g., what definitions of ‘‘headedness’’ are critical to 2c, how can languages with ‘‘mixed’’ headedness or ‘‘disharmonic orders’’ be accounted for, which aspects of grammar are affected (as in 2d), or what aspects of language experience can ‘‘trigger’’ or determine such parameter setting for the child (e.g., Mazuka, 1996; Nespor et al., 1996). Systematic application of the model to head final languages is necessary (e.g., Fukui 1995 for Japanese).
17.3.2 The Antisymmetric Program in UG (e.g., 2e) In a contrasting view of UG, inspired by the revolutionary work of Kayne 1994, analysts of directionality make the strong proposal that UG itself is essentially antisymmetric, challenging the parametric view. On this view, a specifier-head-complement (S-H-C) order such as in (SVO) word orders is proposed to be universal.7 6
Because directionality is necessarily binary and order is arbitrary in fundamental grammatical competence for natural language, in fact on this view there may be no need to specify a parameter to refer to directionality at all within UG itself. In other words, a parameter may become a second order phenomenon in this case, describing the results of UG->Specific Language Grammar->Language mapping, not its cause. 7 We present only the general idea here, leaving aside details. The asymmetric proposal refers to the order of subconstituents of a phrase. Both the early parameter setting proposal and the asymmetric proposal involve a form of phrase structure grammar (X-bar theory) including an assumption that all phrases are headed; although for Kayne the properties of X-bar theory are not themselves a primitive component of UG but derive from the asymmetry proposal (1994, 3). (See Chomsky, 1995, pp. 334–340 for integration of the asymmetric proposal with Minimalist assumptions of ‘‘bare phrase structure’’.)
374
B. Lust
An order such as specifier-complement-head (S-C-H), as in (SOV) order, reflects a derived phrase structure. To be specific, in the latter case, constituent movement has occurred, raising the complement to a specifier position to the left of the head.8 Generalizations such as in 2a and b above reflect either the unmarked order or movement operations that permute this order systematically. This ‘‘antisymmetry’’ program, summarized in 3, helps to focus critical issues regarding the nature of UG with regard to directionality and real time. (3) An Antisymmetry Proposal (Kayne, 1994) <x,y> is interpreted as ‘x precedes y’ rather than as ‘x follows y’.(36) ‘. . .UG does not allow languages a choice between S-H-C and C-H-S.’(36)9 ‘. . .UG does not make both orders available. . ..’(38) To be more precise, the foundations for this proposal include 4a-c: (4) Fundamental Assumptions of ‘antisymmetry’: a. Hierarchical structure and linear order are wedded, such that if x asymmetrically c-commands y, x precedes y. A relation of precedence is configurationally determined. At a basic level, configuration determines linearity.10 b. Selection of the particular linear relation ‘precedes’ as the only one given by UG, as opposed to the opposite relation of ‘subsequence’, results from a representation that is fundamentally related to the antisymmetric nature of time in keeping with the following mechanism. c. Mechanism of ‘antisymmetry’: In line with the first assumption above (4a), a basic phrase structure configuration is assumed to map to a set of terminal elements that must be linearly ordered (‘at least in the PF wing of the grammar’ (4)). In this configuration, an abstract node A is posited that is adjoined to the root node (which dominates every node in the structure (except itself)); this abstract node ‘A’ asymmetrically c-commands every other node. This abstract root node is then mapped by ‘d’ into the abstract beginning terminal ‘a’ and the full set of terminals (37). Here, ‘d’ refers to the non terminal to terminal dominance relation in this phrase structure (5).
8
In this notation, the Verb (V) in Subject Verb Object (SVO) or Subject Object Verb (SOV) sentences is considered to be a head (H) taking the object (O) as complement (C) of the Verb Phrase. In both cases the sentence subject is treated as a ‘‘specifier’’ (S). Thus the SVO notation is frequently annotated as ‘S-H-C’ and the SOV notation as ‘‘S-C-H’’. 9 Other orders, e.g., S-C-H etc., are ruled out by a postulate that spec and complement must always be on opposite sides of the head. 10 More specifically, asymmetric c-command is mediated by dominance relations but the ‘‘asymmetric c-command relation is significantly similar to the dominance relation. . .both are locally linear’’(Kayne, 1994, p. 36) (i.e., in the sense that the dominance relation ‘‘becomes linear if one restricts oneself to the local environment of a given node’’)(4).
17
Directionality in the Architecture of the Language Faculty
375
Kayne is explicit on the relation of this representation to real time in the mapping proposed in 4c. ‘‘Let us think of the string of terminals as being associated with a string of time slots. . ..what is paired with each time slot is not simply the corresponding terminal, but the substring of terminals ending with that terminal (i.e., the substring produced up to that time).’’(37) ‘‘In other words, a string of terminals ‘‘abcdz’’ (with ‘‘a’’ and ‘‘z’’ abstract) is mapped to a set of substrings). . .a, ab, abc, abcd, abcdz . . .. An asymmetry between a and z has now appeared: a precedes every terminal in every substring, but z does not follow every terminal in every substring (since z figures in only one substring.’’(37). The underlying logic for mapping to the initial terminal ‘‘a’’, rather than to the final terminal, ‘‘z’’ in 4c is the following: ‘‘If the abstract root node for asymmetric c-command needs to be mapped by d to a corresponding abstract ‘‘root node’’ for terminals, and if that root node for terminals must be in some fixed relation to every terminal in every substring, then that abstract terminal must be ‘‘a’’ and the fixed relation must be ‘‘precedes’’.’’(37) This representation of the relation between terminals and non-terminals, these assumptions and this mechanism, underlie 5. (5)
Linear Correspondence Axiom I(LCA): (Kayne, 1994, p. 6) ‘d(A) is a linear ordering of T ’11
From the above it can be seen that in the antisymmetry proposal the ‘‘. . . S-H-C property of UG, as well as the fact that UG does not make both orders available, is ultimately related to the asymmetry of time.’’(38) Without the assumption of time determining both the ‘‘beginning terminal element’’ and the temporally achieved cumulative set of substrings, e.g., ab, abc, abcd, etc., the LCA (5, above) may hold but it may hold equally as well given ‘‘subsequence’’ rather than ‘‘precedence’’ of the phrasal subconstituents. In sum, two essential components underlie the antisymmetry proposal. These include the proposal (1) for a critical link between hierarchical structure (c-command) and linearity of terminal elements as in (4a) and (2) for asymmetry of configuration privileging the ordering relation of precedence as in (4b), where an integration of the formal representation of directionality in UG with regard to real time language processing links these two as in (4c). A priori, each of these two components is independent of the other. Indeed the first component, linking hierarchical structure and linear order, converges with a long and wide history of results in earlier interpretations of UG (e.g., Reinhart, 1983, 1986 and many others). It is the second component that is most unique. This antisymmetry proposal now joins a history of attempts to capture linear variation in natural language; 2 exemplifies only a selected few of these (Lust, in Here A consists in all sets of ordered pairs, <Xj, Yj> of non terminals ‘‘such that for eachj, Xj asymmetrically c-commands Yj’’, T is a set of terminals, d is a dominance mapping from nonterminals to terminals (5) and d(X) is the set of terminals that X dominates. (Kayne, 1994).
11
376
B. Lust
prep for review). At the same time, it revolutionizes and sharpens our focus on directionality because of its formal precision and its strong proposal (cf. Cinque, 1996; Kayne, 2000). As Kayne suggested, this antisymmetric view of UG is ‘‘not logically necessary’’. An opposite view is equally possible, e.g., one interpreting <x,y> as meaning that x follows y and thus producing the CH or SOV order as basic. (36) In fact, such a reverse proposal has been made (e.g., Haider, 2000). Alternatively, if time is not relevant at the level of UG, <x,y> may be equivalently interpreted as x precedes y or x follows y in real time in natural languages. Two distinct hypotheses regarding the architecture of the Language Faculty, modeled by UG, are now seen as possible. Hypothesis 1: Linear order (and the directionality that determines it), is irrelevant to UG itself. On the definition of UG as in 1c, e.g., as ‘‘prior to experience’’, time is irrelevant. There is no directionality in UG. There is no underlying markedness in organization of specific language grammars as SVO or SOV or as head initial or head final. <x,y> is not specified at all for linearity. UG determines that linear order is linked to configuration, i.e., it provides a principle of ‘‘Structure Dependence’’ but makes no reference to order or directionality in the interpretation of the LCA (5 above). Hypothesis 2: Directionality is determined by UG; there is a unique setting consistent with ‘‘precedence’’ universally across languages. This is critically determined by formulating the architecture of UG with reference to ‘‘precedence’’ of real time speech processing (4c above). Antisymmetries that arise in specific language grammars are explained by the organization of UG. In general, SOV languages are in some way marked with regard to SVO languages. On the first hypothesis, linear order and the directionality that determines it, arises at the level of mapping from UG to Specific Language Grammar (SLG) and to Language, i.e., it arises at the point where UG interacts with experience and real time in a specific language. Because there is no directionality programming in UG, this mapping is initially arbitrarily leftward or rightward. In natural language grammars, either forward or backward orders are equivalent. Antisymmetries in language organization that arise, e.g., a potentially universal initial ‘‘spec’’ position, arise at the level of the mapping to language and to experience in real time. Harmonic universals and/or tendencies in languages arise because the language faculty provides essential structure dependence leading to systematicity in grammar construction. While the second hypothesis is consistent with the full ‘‘antisymmetry’’ proposal (in 4) as a theory of UG, the first leaves open a possible form of directionality parameter in conjunction with the basic first component of the antisymmetry proposal, i.e., that which links linear order to underlying configuration of sentence structure. On the first hypothesis, the full antisymmetry proposal is resituated with regard to the architecture of the language faculty; it is viewed as a model of the integration of UG with real time processing of natural language, rather than as a model of UG itself.
17
Directionality in the Architecture of the Language Faculty
377
The contrast between these two hypotheses now provokes us to consider the fundamental nature of the Language Faculty, represented by a theory of UG. The choice between the two equally logical possibilities for formulation of directionality in UG, i.e., the choice between a unidirectional ‘‘antisymmetry’’ proposal and a parameterized bidirectional proposal, derives specifically from the definition of a formal theory of UG with regard to how it integrates with the empirical irreversible asymmetric order of real time.12 The contrast between the two hypotheses raises the issue of how and where this integration exists. A leading question now, given 4 and 5 above, is twofold: in the language faculty is there a fundamental link between configuration and linearity and does this link require precedence?
17.4 Deriving Theoretical and Empirical Predictions We must now ask: how can the varied proposals for a theory of UG be tested? How can empirical evidence be derived to differentiate between them and/or to distinguish them? Clearly, the linguistic investigation of the head final or left branching languages, those that ostensibly do not evidence an unmarked SVO, or S-H-C order (about half the languages of the world), would appear to provide a critical challenge to the asymmetric proposal (e.g., Japanese, Korean, Malayalam, Tamil, Sinhala, etc.). Their linguistic analysis is even more critical now. Linguistic inquiry in the antisymmetric framework currently pursues this challenge through a theory of constrained movements and assumed underlying functional structure guiding these movements (e.g., Jayaseelan, in press; Kayne, 2003; and many others) in order to deal with the apparent head final nature of these languages. These proposals lead us to seek a principled explanation for widespread and coherent triggering of movements in these head final languages as opposed to the head initial languages. At the same time, asymmetric syntactic and semantic operations exist across head initial and head final languages and must be explained in either framework. Given these considerations and our assumed definition of UG in 1.2, any data where Universal Grammar maps to language in real time would appear to provide critical evidence regarding the fundamental issues here. Typological variation, language acquisition and language processing across head final and head initial languages must be consulted now. We will briefly sketch developments along each of these dimensions in order to adumbrate the relevant research areas and their potential research types.
12
More precisely, we recognize that UG is ‘‘of course not a grammar but a system of conditions on the range of possible grammars for possible human languages’’(Chomsky, 1980, p. 189). Thus we are inquiring about the nature of the representation of the linguistic ‘‘system of conditions’’ provided by the language faculty.
378
B. Lust
17.4.1 Typology The ‘‘harmonic regularities’’ wherein various structures within a language are so frequently found to cohere, confirming much of Greenberg’s (1963) original ‘‘word order universals’’, would appear to adumbrate the regularities of an underlying system, i.e., a grammar; e.g., the overwhelming tendency for VO order to correlate with N-relative clause order and for OV to correlate with relative-N order (e.g., Gair, 1983). At the same time the ‘‘disharmonic’’ orders within and across language impugn the regularities of the grammar underlying the language(s). Cross linguistic typological work continues to reveal the scope of these harmonies and disharmonies and to raise issues regarding their interpretation. (See for example, Dryer, 1991, 1992, 1998, to appear and Haspelmath, Dryer, Gil, & Comrie (Eds.) 2005, World Atlas of Language Structures. Haspelmath, 2002) For example, what does the critical case of Chinese imply regarding the nature of Universal Grammar? Chinese Chinese represents a critical case with regard to the definition of ‘‘word order universals’’ and explanatory ‘‘parameters’’ of variation. (We focus here on Mandarin). It provides the cell with only one language in Dryer’s 1992 crosslinguistic frequency tables based on a survey of 625, challenging cross-categorical harmony in directionality.13 Although Chinese is SVO in basic main clause word order, ‘‘the fact that modifiers of the noun, the verb and the adjective precede their heads would predict in Greenberg’s framework that Chinese is an SOV language. . .’’ (Huang, 1982, p. 29). Chinese appears to accord with what Greenberg’s Universal 7 predicted for SOV languages. For example, order of ‘‘adverbial modifiers of the verb,’’ or adverbial subordinate clauses, are left of the main clause in Chinese. In addition, relative clauses and complements precede their head noun in Chinese as in many SOV languages. Interpreting Typological Data In general, testing hypotheses regarding the role of grammatical representations in typological data involves considering distributional frequencies of language types, as well as of within-language cross-structural harmonies and disharmonies. Languages like Chinese, which offend ‘‘harmonic orders,’’ are sometimes considered ‘‘more marked or less preferred’’ in typological terms (Hawkins, 1985, p. 576). At the same time, from the point of view of defining language 13 In Dryer’s more recent typological studies (pc and Haspelmath et al., 2005), which extend the database, more languages have been added to the cell wherein RELN order occurs with VO order. These include three Chinese languages (Mandarin, Hakka and Cantonese), Bai (Tibeto-Burman) and Amis (a Sinicized Austronesian language of Taiwan). (Dryer, pc., and Haspelmath et al., 2005) Dryer’s database distinguishes genera and language (Dryer, 1989).
17
Directionality in the Architecture of the Language Faculty
379
competence, or the Language Faculty, such frequency data are to some degree irrelevant. Mandarin Chinese, the official language of the People’s Republic of China, the most populous as well as the third largest country in the world, accounts for a significant proportion of the world population and is thus clearly in this sense not ‘‘less natural’’ or ‘‘more marked’’ with regard to grammatical competence. Viewed in this way, the ‘‘tendencies’’ for coherence, or harmonies, shown in cross-linguistic typological studies may indirectly reflect properties of UG but they also reflect the mapping between UG, Specific Language Grammar and Language; and thus they are subject to numerous factors that intervene in this mapping, including the wide array of pragmatic factors that may intervene often in language-specific ways (e.g., Tomlin, 1986). Consequently, typological data analyzed in terms of frequencies with which languages occur, although providing important empirical evidence with regard to hypotheses involving UG, will not easily allow straightforward interpretation with regard to the fundamental issues regarding the nature of the Language Faculty that we raise here. The basic obvious fact that languages are free to vary in directionality of basic structures and to reflect grammars coherent in directionality in either direction appears to connote no markedness in one direction or another at the level of UG. At the same time, various types of coherences may be susceptible to either asymmetric (e.g., Cinque on Greenberg’s Universal 20, 2005) or parametric proposals. Constraints against disharmonies that do not occur may also be formulated and explained grammatically (e.g., Biberauer, Holmberg, & Roberts, 2008). Explanation of various disharmonies within and across languages is now under active investigation.
17.4.2 The Learnability Challenge The facts of surface language typology alone would suggest that all basic orders of <S,V,O> are possible as basic orders for a language and therefore would appear to need to be learned. However, in many ways, this learning appears a priori to be an essentially intractable feat. How would children distinguish the syntactic, pragmatic, semantic, stylistic, or psychological factors and their interactions in their original decisions about the basic order of the language they were acquiring when any order is a priori possible? Even more critically, within all languages, surface orders vary; reflecting language-specific interactions with varying syntactic movements and with various language-specific sensitivities to pragmatic factors, e.g., ‘‘interrogative’’ speech acts, or more specific pragmatic factors related to focus or topicalization for example. Each observed order may reflect optional or obligatory syntactic movements. All of these interactions render any observed ‘‘word order’’ indeterminate in itself with regard to ‘‘basic grammatical order.’’ How would children determine the basic order of a language given order variation within a language and across languages, such as produced by various
380
B. Lust
movement types (like ‘‘raising’’ or ‘‘scrambling’’ or Wh- movement) and their linked interactions? If knowledge of basic order must be derived from Primary Linguistic Data (PLD), then given such movements, how could the learners determine what the fundamental word order for a language is and what is derived? The learners have to ‘‘undo’’ the movement to determine the basic order in a form of ‘‘reverse engineering’’. But how can they do this if they do not know what has moved and what has not? Since different constituents can move and multiple movements are possible, how could the learners determine which constituent had moved at any particular time? For example, given VO/OV variation in a language, how would children know whether the verb or the object, or both, had moved in order to determine the underlying order? Critically, this problem is compounded since, on the surface, elements appear to move in either direction, as in right or left dislocation for example. This means that an observed order, [X Y], is not only indeterminate as to whether it is the result of movement or not but if moved, it a priori could result as seen below from 6a or b (cf. Barbier, 1994 on this issue, with particular reference to Dutch). On these assumptions, the indeterminacy in any particular observed linear string is thus complete. (6)
Opacity of Observed word orders a. -Y X —> X Y
b. Y X- —> X Y
Obviously, when movements may be covert as well as overt, the problem would appear to be compounded exponentially. These learnability problems would appear to characterize either classic parametric or asymmetric proposals.14 Given such essential indeterminacy, how can critical data be available to the child for determining the basic order of their language? If some movement is optional, how is the child to know which movement variation is allowable and which is not in their language? On the other hand, given the presence of obligatory movements in a language as well, how is the child to determine that these are not also optional movements? Where constituent deletion is involved, or where multiple movements may occur, of various types, e.g., head movement or ‘‘remnant movement’’, possibly intricately linked and ordered, in different ways in different languages, the problem is further complicated.
14
The details of approach would differ here, e.g., only leftward movement is allowed in the asymmetric proposal.
17
Directionality in the Architecture of the Language Faculty
381
While specific movements involved in topicalization or focus, for example, would presumably be linked to various pragmatic and semantic factors, how would the child come to learn which such factors correlated with which basic and derived orders in their language, especially if they had not already settled on a basic word order? Since the interaction of movement with pragmatic factors is language-specific to a large degree, the child cannot assume a particular correlation of order with such a factor, or even whether order should be correlated with such a factor, in order to begin to solve the problem of establishing a basic order in extra-linguistic terms. The child could not execute or understand a specific pragmatically triggered order variation without the assumption of an alternative order to which the pragmatically motivated one could and should be contrasted. Similarly, semantic interpretation does not correlate in a one-to-one manner with surface order or with the existence of movement, which can be overt but ‘‘semantically vacuous’’ or covert and semantically significant. No obvious form of semantic bootstrapping can solve this learnability problem. A parametric approach to UG must provide the child with an appropriate parameter setting in order to render this learnability problem tractable; an asymmetric approach must begin with an assumption of no rightward movement followed by potential overt and covert leftward movements on an initial assumed order defined over particular forms of assumed representations. To our knowledge, no systematic analysis has fully confronted this learnability problem on either approach.
17.4.3 Language Acquisition Since every normal child everywhere solves this learnability problem and in fact does so remarkably early, we may look to the nature of the Language Faculty as it is instantiated in language acquisition. We seek evidence from language acquisition with regard to the nature of the Language Faculty that initiates the language acquisition process and that presumably guides and constrains the course of language acquisition. We can begin to conjoin this line of evidence with analysis of the linguistic theory that provides precise hypotheses regarding the design of the Language Faculty and with results of research on real time language processing. Through this evidence, we must evaluate both aspects of the leading questions regarding UG, viz., are configuration and linearity integrated in such a way that linearity reflects configuration and does a precedence relation in linear order characterize this relation uniquely? Configuration and Linearity Numerous studies of child language acquisition have now evidenced an early and continuous structure dependence in early grammatical knowledge. (See Lust, 2006 for general review.) This structure dependence involves an early and
382
B. Lust
continuous sensitivity to phrase structure as well as to functional headedness of this phrase structure (e.g., Gerken, Remez, & Landau, 1990; Kedar, Casasola, & Lust, 2006; Lust, 1994). Young children at early periods of language acquisition integrate this configurational sensitivity with movement, e.g., with verb raising, as in Dutch and German and link these associations to the syntax of their language, e.g., finiteness (Barbier, 1995a, 1995b, 1996, 2000; Boser, 1997; Boser et al., 1991, 1992; Boser, Santelmann, Barbier, & Lust, 1995; Lust, Sun˜er, & Whitman, 1994; Lust & Wakayama, 1981).15 Young children link linearity to configuration, both configuration of individual structures and of general direction of headedness across languages. This has been documented both within and across languages when computational processes such as anaphora have been studied (e.g., the early collections in Lust 1986, 1987 and subsequent research by several labs). For example, precedence effects in anaphora in child language have been shown to be modulated by structure as well as by the grammar of the language being acquired. Precedence Effects Children acquiring English were observed to show early linear precedence effects in anaphora production as well as interpretation; and these were argued by some to reflect a universal linear precedence effect (e.g., Bauer, 1995; Carden, 1986; O’Grady, Suzuki-Wei, & Cho, 1986). As attested by several studies including ours (e.g., Lust, 1981, 1986; Lust, Loveland, & Kornet, 1980; Lust, Solan, Flynn, Cross, & Schuetz, 1986; Eisele & Lust, 1996), children acquiring English productively assign a pronoun to be coreferential with an antecedent that precedes the pronoun (forward anaphora) but less productively with an antecedent that follows the pronoun (backward anaphora) in complex sentences containing adverbial subordinate clauses (i.e., they demonstrate a ‘‘precedence effect’’ in this domain), even when pronouns are free and optionally coreferential in this domain. Correspondingly, they produce sentences with forward anaphora such as 7a or b, or 8a – significantly more successfully than those with backward anaphora such as 7c or 8b – and make more anaphora errors (e.g., pronoun reversals) when producing the backward forms: (7)
15
Lexical Pronouns Forward a. Jenna drank some juice while she was having lunch b. While Jenna was having lunch, she drank some juice Backward c. When she threw the block, Jenna began to dance
Elsewhere we offer systematic comparative experimental study of the acquisition of several head initial and head final languages including the critical Chinese; we assess both relative markedness of directionality and potentially universal forms of structure dependence across these. (Lust, in prep.)
17
Directionality in the Architecture of the Language Faculty
(8)
Null Proforms a. Jenna drank some juice when 0 having lunch b. When 0 having lunch, Jenna drank some juice
383
These precedence facts led some, (e.g., O’Grady et al., 1986) to propose a ‘‘Forward Preference Hypothesis’’ or ‘‘forward directionality constraint’’ in child language (See also Carden, 1986; Lust, 1981) as in 9:16 (9)
‘. . .the known developmental facts seem to support the view that any directionality preference for anaphora is the same for all languages (i.e., it is forward). . ..This in turn would seem to suggest analyses of anaphora in which linear order is taken to be a critical factor and backward patterns of coreference are predicted to be more marked in all languages’ (O’Grady et al., 1986, p. 420).
However, further study within English structures such as 10 compared to those such as 11–13 revealed that the backward anaphora ‘‘constraint’’ in children’s grammars was not structure independent. On structures like 10 for example, which involve forward anaphora (where reconstruction effects are implied), anaphora was significantly blocked, when compared to either 11 or 12 with backward anaphora, even by children as young as 3 (e.g., Eisele & Lust, 1996; Lust & Clifford, 1986; Lust, Eisele, & Mazuka, 1992). At the same time, children as young as 3 significantly block anaphora in backward cases like 13, where the structure (and Binding Principle C) rule this out. (See also Crain & McKee, 1985; Lidz, 2007.) Related results have also been found in Italian (Guasti & Chierchia, 1999/2000). (10) (11) (12) (13)
On the top of Ernie’s head, he rubbed the donut On the top of his head, Ernie rubbed the donut When he pushed the pillow, Big Bird threw the candybar He pushed the pillow, when Big Bird threw the candybar
In fact, Principle C effects, which crucially involve computation over configurational structure, replicate across languages in early acquisition, e.g., in Hindi (Gair, Lust, Bhatia, Sharma, & Khare, 1998; Lust, Bhatia, Gair, Sharma, & Khare, 1995; Somashekar, 1995; Somashekar et al., 1997). They do so even when forms of grammatical reconstruction are involved (e.g., cases 10 or 11). In addition, when acquisition of head final languages was studied, the precedence effects seen in early English were found to be either muted or negated, i.e., backward anaphora is freely available in left branching structures in these languages early in language acquisition (e.g., Japanese, Sinhala; see Lust, in 16 Bauer 1995: In fact, it has been proposed that there is consistent directionality and asymmetry in language change as well as in language acquisition. (Bichakjian, 1987, 1988a, 1988b; Lust in prep for discussion.)
384
B. Lust
prep., 1983, Gair, Lust, Sumangala, & Rodrigo, 1998; Lust, deAbrew, & Sharma, 1982; Lust & Lee, 1989; Lust & Mangione, 1983; Lust, Mangione, & Chien, 1984a, b; Lust, Wakayama, Hiraide, Snyder, & Bergmann, 1983). In fact, in Chinese – a left branching language with relative clause-Noun order and adverbial subordinate clause-main clause order – even when backward pronominal anaphora is ruled out by the adult language, children allow this backward anaphora, where it is compatible with the left-branching configuration (Lust, Chien, Chiang, & Eisele, 1996). (See also Kazanina & Phillips, 2001 for related results in Russian acquisition.) In summary, evidence from first Language Acquisition suggests: 1. 2.
3.
4.
Children show a very early sensitivity to sentence configuration. They link linearity (i.e., linear precedence effects in anaphora) to this configuration. There is no universal ‘forward directionality of anaphora’ constraint on grammatical competence. Backward anaphora is as unmarked as forward anaphora when the configuration of the language supports this. In each language, children modulate anaphora linearity both in terms of sentence configuration and in terms of specific language grammar (right or left branching/head initial or head final). They do so even when the adult grammar does not model this. For example, where Chinese adult grammar demonstrates a forward directionality constraint on lexical pronominal anaphora, children first assume that a left-branching-backward anaphora correlation is supported.
In these ways, not only does configuration trump linear order early in language acquisition but it does so even when surface facts appear inconsistent with configuration, e.g., Chinese pronominal anaphora. To date, then, there is significant evidence from the study of language acquisition that favors the first component of a UG proposal that critically links configuration and linear order and makes linear order to some degree dependent on configuration. The child’s early knowledge of language, across both right-headed and left-headed languages, appears to reflect this form of structure dependence. These results support the first component of the LCA (4 and 5 above); they are consistent with structure dependence in a parametric view of directionality in UG as well.
17.4.4 Language Acquisition and Language Processing Converge The results from language acquisition that evidence early integration of linear order and configuration, with configuration capable of trumping linearity, cohere in general with results from several studies of natural language processing in the adult (e.g., Bornkessel, Schlesewesky, & Friederici, 2002; Friederici, 2002; Friederici & Gorrell, 1988; Vigliocco & Nicol, 1998). For example, in early work,
17
Directionality in the Architecture of the Language Faculty
385
Vigliocco and Nicol (1998) expressly tested hierarchical relations and word order factors in language production and argued that their experimental results supported a stage in which syntactic structure was built prior to linear assignments of word positions. Friederici and Gorrell (1988) argued on the basis of crosslinguistic evidence from aphasic subjects that a ‘‘structural prominence hypothesis’’ explained patients’ thematic reversal errors rather than a linear or directionality strategy. These results would appear to cohere with a recent neurocognitive model of auditory sentence comprehension, supported by fMRI and ERP data, by Friederici (2002) that revealed an early (100–300 ms) initial syntactic parse in the left anterior temporal region that is based on phrase structure and word category information and with Friederici, Bahlmann, Heim, Schubotz and Anwander (2006). Bornkessel et al’s (2002) ERP studies of flexible word order parsing in German would appear to be consistent in their argument that grammar guides word order processing in these cases. The language acquisition results above also fundamentally converge with natural language processing studies when we consider the infant’s processing of the speech stream in early periods of language acquisition. In general, very young children show early and continuous sensitivity to linear order in the speech stream as well as to clausal and phrasal structure. (See Lust, 2006, p. 186f. for general review.) They generalize an abstract knowledge of word order (Gertner, Fisher, & Eisengart, 2006); and, as is well-known, show very few word order errors from the time they first begin producing sentences. At the same time, they show sensitivity to clausal structure by 4–6 months of age (e.g., Hirsh-Pasek et al., 1987; Johnson & Seidl, 2008). Thus the foundations both for constituent structure and for linearity appear to be set at once at earliest periods of language acquisition. Current studies are testing more precise hypotheses regarding the nature of this mapping of structure and linearity in the speech stream by the infant. It had been speculated earlier that infants may determine essential directionality properties of their language even before they produce their first words (Mazuka, 1996). Now more recent research has begun to test this hypothesis empirically. A recent cross-linguistic comparative study of 8-month-old Japanese and Italian infants, for example, demonstrated ‘‘opposite order preferences’’ in an artificial grammar experiment using a head turn preference procedure. The infants preferred opposite sets of test items depending on whether these instantiated frequent initial (Italian) or frequent final (Japanese) regularities. The authors suggest ‘‘infants might use the relative order of functors and content words at utterance boundaries to create one of the first rudimentary representations of word order already before they build their lexicon’’ (Gervain, Nespor, Mazuka, Horie, & Mehler, 2008, p. 23). These results imply that once again, configuration and linearity are tightly integrated in the earliest phases of language knowledge and they begin to suggest a mechanism for this integration in early language processing. Recent and current studies have begun to investigate precise properties of the speech stream that may be relevant to the infant’s language processing so that
386
B. Lust
directionality may be established in a language (e.g., Mazuka, 1996; Nespor, Guastic, & Christophe, 1996). The issues here include the questions of which structural factors are consulted by the infant, (e.g., clausal functional category heads and clausal domains, or lexical category heads and smaller phrasal domains) and how each of these may correlate with prosodic aspects of the speech stream. Specific phonetic and/or phonological properties involving prosody may cohere with cross-language typological order differences and with different directions of headedness across languages (e.g., Wagner, 2005). Comparative developmental studies of the child’s processing of head-final languages are now also critical. Mazuka’s seminal (1990, 1998) comparative studies with 4–8 year old children (English and Japanese) lead the way in this endeavor, suggesting early systematic differences in processing strategies across head-initial and head-final languages at early periods of language knowledge and they motivate continued research in this paradigm. They can now be merged with studies of adult processing of head-final languages, such as reflected in this volume, as well as with adult processing studies of head-initial languages and together they can be brought to bear on current theories of the Language Faculty. Study of second language acquisition provides another avenue of discovery (e.g. Flynn, 1989; Flynn & Lust, 2002).
17.5 Conclusions Even the brief sketch above suggests that each line of evidence – language acquisition and language processing – is advancing now, producing ever more subtle results regarding the nature of the language faculty and its role in language acquisition and use. However, to date, while the results are building more and more support for the fundamental integration of configuration and linear order along all lines of research, and for the result that to some degree configuration determines linear order, they leave open the issue of the second leading question. Namely, is precedence and left-headedness predetermined by the Language Faculty as the foundation upon which all languages are founded, acquired, developed and known? What are needed now are clear predictions from each of the two leading hypotheses regarding the nature of UG that would differentiate them. Any particular language structure can be linguistically represented in terms of either approach (e.g., 2c,d versus 2e). At the same time, if we assume the definition of UG as in 1 above, then we would expect that any full model of UG would generate empirical predictions that bear on both language acquisition and language processing, as well as language variation. In the strongest model, predictions across these dimensions will cohere. They will address the learnability challenge addressed in Section 17.4.2. We know of no comprehensive attempt to formulate such predictions, especially those that may apply across all these relevant dimensions, required by a theory of UG defined as in 1.
17
Directionality in the Architecture of the Language Faculty
387
To some degree, studies of language processing stand as the central core that may link these dimensions of evidence. For example, as Mazuka suggested, in some ways, ‘‘When we examine the mechanisms of language acquisition, it becomes apparent that the language acquisition problem is the language processing problem’’ (1998, p. ix). The relation between grammatical knowledge and language processing comes to the fore here. The child must process the speech stream in order to establish the particular grammar of their language. In particular, the directional value of the grammar being acquired must be derived from speech stream processing in real time. Yet, the child cannot process the linguistic value of the real time speech stream without some prior grammatical knowledge (cf. Fodor, 1998; Valian, 1993). We look to theories of the Language Faculty as well as empirical studies of child and adult language processing to resolve this paradox, which bears on the learnabilty problem we raised in Section 17.4.2 above. Earlier cross-linguistic studies of language processing have raised issues concerning the relation between grammatical variation in directionality (e.g., head final versus head initial, SOV versus SVO, or left-branching versus right-branching languages, however characterized) and the nature of language processing across language types. A series of early studies raised these issues through a focused contrastive study of the left-branching or head-final languages, long absent from early studies of language processing (e.g., Frazier & Rayner, 1988; Inoue, 1991; Inoue & Fodor, 1995; Mazuka & Lust, 1987, 1990; Mazuka & Nagai, 1995; Mazuka & Itoh, 1995); and these are now complemented and significantly extended by the rich set of studies of processing of head-final structures in this volume. The issues regarding the relation between grammar types that vary in directionality and their processing organization that were raised in the early studies persist and are confronted in papers collected in this volume: e.g., are there principled differences between the organization of processing strategies in head initial and head final languages, such as in parameterized ‘‘top-down’’ or ‘‘bottom-up’’ processing models, which consult parametric grammatical differences across languages, or are there universal strategies that apply incrementally on-line, without consulting such principled grammatical variation, or does some combination of both hold? Are ‘‘incremental’’ processors able to accomplish equally efficient processing of both head initial and head final languages without depending on grammatical headedness at all? More specific questions now arise: are functional and lexical heads differentiated in their role or non-role in on-line processing? If grammatical headedness is not critical in on-line processing, then in what way does grammar inform such processing? If in head final languages, language processing proceeds as it must from left to right and material must be held before being integrated in a higher level structure such as a headed phrase, then what form does the representation of this material take and what aspects of grammar may inform it? For Bader (this volume) for example, the grammar ‘‘provides structures that make the parser’s life easier’’
388
B. Lust
in right headed structures but it is specific language grammar, not Universal Grammar that does so.17 Only systematic comparative cross-linguistic research that is precisely focused on the head final languages, such as reflected in this volume, in systematic comparison to head initial languages and in conjunction with precise linguistic theory, such as exemplified in Section 17.3, can fully resolve the fundamental issue of the nature of the Language Faculty that we have raised here. Disharmonic languages provide a further challenge (Bayer, 1996). In sum, issues regarding directionality in natural language knowledge as they are currently formulated, i.e., in terms of symmetry vs. asymmetry, or parametric versus non-parametric treatments, force us to reconsider the foundations of a theory of UG as a model of the Language Faculty. Fundamentally, given all lines of evidence, the most basic question still unresolved in the field involves the degree to which UG provides an abstract specification ‘‘prior to experience’’, i.e., prior to integration of language knowledge in real time; this would leave directionality arbitrary for the language faculty. This issue has important consequences for cognitive science as well as for linguistic theory. To the degree that the Language Faculty may be formulated in terms of real time (as in hypothesis 2 above), then this linguistic module can be rendered more directly coherent with other cognitive modules or processes relevant to the processing of time. If not (hypothesis 1), this leaves the linguistic module at least partially distinct in nature from these other cognitive components, potentially in part at least comparable to other cognitive representations such as those involved in mathematical knowledge, where time is irrelevant. We have suggested that sharpened empirical predictions are necessary now in the framework of a full definition of the theory of UG. We can assume that sharpened theoretical predictions, combined with advances in the study of each of the relevant dimensions of empirical evidence – language typology, language acquisition and language processing – each systematically contrasting head final with head initial languages, will now allow us to approach a solution to this fundamental question.
References Bader, M. (this volume). On being both head-initial and head-final. In H. Yamashita, Y. Hirose & J. Packard (Eds.), Processing and producing head-final structures. Dordrecht: Springer. Baker, M. (2008). The macroparameter in a microparametric world. In T. Biberauer (Ed.), The limits of syntactic variation (pp. 351–373) Amsterdam: Benjamins. 17 Note that to the degree that adult language processing may refer only to specific language grammar, the importance of the study of the child’s language processing is accentuated; i.e., at early periods before the specific language grammar is fully represented, one may hypothesize that UG may be more directly reflected. Note too that in order to more directly evaluate directionality factors relevant to UG in adult language processing, systematic comparisons across head final and head initial languages are necessary.
17
Directionality in the Architecture of the Language Faculty
389
Baker, M. (2001). The atoms of language. New York: Basic Books. Barbier, I. (1995a). Configuration and movement: Studies of the first language acquisition of dutch word-order. Unpublished doctoral dissertation, Cornell University, Ithaca, NY. Barbier, I. (1995b). The acquisition of embedded clauses with finite verbs in nonfinal position in Dutch. In E. Clark (Ed.), Proceedings of the 26th Annual Child Language Research Forum, 26, 147–158. Barbier, I. (1996). The head-direction of Dutch VPs: Evidence from L1 acquisition. In C. Koster & F. Wijnen (Eds.), Proceedings from the groningen assembly on language acquisition. Groningen: Center for Language and Cognition. Barbier, I. (2000). An experimental study of scrambling and object shift in the acquisition of Dutch. In C. Hamann & S. Powers (Eds.), The Acquisition of Scrambling and Cliticization. Dordrecht: Kluwer. Bauer, B. L. M. (1995) The emergence and Development of SVO Patterning in Latin and French: Diachronic and psycholinguistic perspectives. New York: Oxford. Bayer, J. (1996). Final Complementizers in Hybrid Languages. Paper presented at the Conference on Final Heads. Katholieke Universiteit Brabant, Tilburg, November 24–25, 1995. Biberauer, T., Holmberg, A., & Roberts, I. (2008). Structure and linearization in disharmonic word orders. In C. Chang & H. Haynie (Eds.), Proceedings of the 26th west coast conference on formal linguistics (pp. 96–104). Somerville, MA: Cascadilla Proceedings Project. Bichakjian, B. H. (1987). The Evolution of Word Order: A paedomorphic explanation. In A Giacalone Ramat, O. Carruba and G. Bernini (Eds.), Papers from the 7th international conference on historical linguistics (pp. 87–107). Amsterdam: Benjamins. Bichakjian, B. H. (1988a). Evolution in language. Ann Arbor, MI: Koroma. Bichakjian, B. H. (1988b). J’ai tombe´ pour Je suis tombe´: L’aboutissement d’une longue e´volution. In R. Landheer (Ed.), Aspects de linguistique francaise (pp. 31–48). Amsterdam: Rodopi. Bornkessel, I., Schlesewsky, M., & Friederici, A. (2002). Grammar overrides frequency: evidence from the online processing of flexible word order. Cognition, 85, B21–B30. Boser, K. (1997). The acquisition of word order knowledge in early child German: Interactions between synax and pragmatics. Unpublished doctoral dissertation, Cornell University, Ithaca, NY. Boser, K., Lust, B., Santelmann, L., & Whitman, J. (1991). The theoretical significance of auxiliaries in early child German. 16th Annual Boston University Conference on Language Development, Boston, MA. Boser, K., Lust, B., Santelmann, L., & Whitman, J. (1992). The syntax of CP and V-2 in early German child grammar: The strong continuity hypothesis. Proceedings of the North Eastern Linguistics Association, University of Massachusetts: Amherst, 22, 51–66. Boser, K., Santelmann, L., Barbier, I., & Lust, B. (1995). Grammatical mapping from UG to language specific grammars: Variation in the acquisition of German, Dutch and Swedish. In D. MacLaughlin & S. McEwen (Eds.), Proceedings of the 19th Annual Boston University Conference on Language Development, 1, 130–142. Carden, G. (1986). Blocked forward coreference. In B. Lust (Ed.), Studies in Acquisition of Anaphora (pp. 319–357). Dordrecht: Reidel. Chomsky, N. (1975). Reflections on language. New York: Pantheon. Chomsky, N. (1980). Rules and representations. New York: Columbia University Press. Chomsky, N. (1981). Lectures on government and binding. Dordrecht and Holland: Foris. Chomsky, N. (1988). Language and problems of knowledge (pp. 67–74). Cambridge, MA: MIT Press. Chomsky, N. (1993). Language and thought. Wakefield, RI and London: Moyer Bell. Chomsky, N. (1995). The minimalist program. Cambridge, MA: MIT Press. Cinque, G. (1996). The antisymmetric programme: Theoretical and typological implications. Journal of Linguistics, 32, 447–464.
390
B. Lust
Cinque, G. (2005). Deriving Greenberg’s universal 20 and its exceptions. Linguistic Inquiry, 36 (3), 315–332. Crain, S., & McKee, C. (1985). Acquisition of structural restrictions on anaphora. In S. Berman, J.-W. Choe & J. McDounough (Eds.), Proceedings of the 16th Annual Meeting of the North Eastern Linguistics Society (NELS) (pp. 94–110). Montreal, QC: McGill University. Croft, W. (1990). Typology and universals. Cambridge: Cambridge University Press. Dryer, M. (to appear). The branching direction theory of word order correlations revisited. In S. Scalise, E. Magni & A. Bisetto (Eds.), Universals of Language Today. Berlin: Springer. http://linguistics.buffalo.edu/people/faculty/dryer/dryer/DryerBDTrevisited.pdf Dryer, M. (1989). Large linguistic areas and language sampling. Studies in Language, 13, 257–292. Dryer, M. (1991). SVO languages and the OV:VO typology. Journal of Linguistics, 27, 443–482. Dryer, M. (1992). The Greenbergian word order correlations. Language, 68(1), 81–138. Dryer, M. (1998). Aspects of word order in the languages of Europe. In A. Siewierska (Ed.), Constituent Order in the Languages of Europe (pp. 283–319). European Science Foundation Language Typology series. Berlin: Mouton de Gruyter. Eisele, J., & Lust, B. (1996). Knowledge about pronouns: A developmental study using a truth value judgment task. Child Development, 67, 3086–3100. Flynn, S. (1989). Spanish, Japanese and Chinese speakers’ acquisition of English relative clauses: New evidence for the head-direction parameter. In K. Hyltenstam & L. Obler (Eds.), Bilingualism across the lifespan (pp. 116–131). Cambridge: Cambridge University Press. Flynn, S., & Lust, B. (2002). A minimalist approach to L2 solves a dilemma of UG. In V. Cook (Ed.), Portraits of the L2 User (pp. 93–120). Clevedon: Multilingual Matters, Ltd. Fodor, J. D. (1998). Unambiguous triggers. Linguistic Inquiry, 29, 1–36. Frazier, L., & Rayner, K. (1988). Parameterizing the language processing system: Left- versus Right-branching within and across languages. Explaining language universals (pp. 247–279). New York: Basil Blackwell. Friederici, A. (2002). Towards a neural basis of auditory sentence processing. Trends in Cognitive Sciences, 6, 78–84. Friederici, A., Bahlmann, J., Heim, S., Schubotz, R., & Anwander, A. (2006). The brain differentiates human and non-human grammars: Functional localization and structural connectivity. PNAS, 103, 2458–2463. Friederici, A., & Gorrell, P. (1998). Structural prominence and agrammatic theta-role assignment: A reconsideration of linear strategies. Brain and Language, 65, 253–275. Fukui, N. (1995). The principles and parameters approach: A comparative syntax of English and Japanese. In M. Shibatani & T. Byron (Eds.), Approaches to language typology (pp. 327–372). New York: Oxford University Press. Fukui, N., & Saito, M. (1997). Order in phrase structure and movement. Linguistic Inquiry, 29, 439–474. Fukui, N., & Takano, Y. (1998). Symmetry in syntax: Merge and demerge. Journal of East Asian Linguistics, 7, 27–86. Fukui, N., & Takano, Y. (2006). Theoretical comparative syntax: Studies on macroparameters (pp. 132–178). New York: Routledge. Gair, J. (1983). Ambiguity is no relative matter, or how do you treat your doubtful relatives. In F. Agard & G. Kelley, (Eds.), Essays in honor of Charles F. Hockett (pp. 128–146). Amsterdam: Brill. Gair, J., Lust, B., Bhatia, T., Sharma, V., & Khare, J. (1998). A parameter-setting paradox: Children’s acquisition of Hindi anaphora in ‘‘Jab’’ clauses. In J. Gair (Ed.), Studies in South Asian linguistics (pp. 286–304). New York: Oxford University Press.
17
Directionality in the Architecture of the Language Faculty
391
Gair, J., Lust, B., Sumangala, L., & Rodrigo, M. (1998). Acquisition of null subjects and control in some Sinhala adverbial clauses. In J. Gair (Ed.), Studies in South Asian linguistics (pp. 271–285). New York: Oxford University Press. Gertner, Y., Fisher, C., & Eisengart, J. (2006). Learning words and rules. Abstract knowledge of word order in early sentence comprehension. Psychological Science, 17(8), 684–691. Gerken, L., Remez, R., & Landau, B. (1990). Function morphemes in young children’s speech perception and production. Developmental Psychology, 26, 204–216. Gervain, J., Nespor, M., Mazuka, R., Horie, R., & Mehler, J. (2008). Bootstrapping word order in prelexical infants: a Japanese-Italian cross-linguistic study. Cognitive Psychology, 57, 56–74. Greenberg, J. (1963). Some universals of grammar with particular reference to the order of meaningful elements. In J. Greenberg (Ed.), Universals of language (pp. 58–90). Cambridge, MA: MIT Press. Guasti, M., & Chierchia, G. (1999/2000). Backward versus forward anaphora: reconstruction in child grammar. Language Acquisition, 8(2), 129–170. Haider, H. (2000). OV is more basic than VO. In P. Svenonius (Ed.), The derivation of VO and OV (pp. 45–67). Amsterdam: John Benjamins. Haspelmath, M. (2002). On understanding word order asymmetries (comments on John A. Hawkins, ‘Symmetries and asymmetries, their grammar, typology and parsing’). Theoretical Linguistics, 28(2), 159–170. Haspelmath, M., Dryer, M., Gil, D., & Comrie, B. (2005). World atlas of language structures. Oxford: Oxford University Press. [Also: http://wals.info] Hawkins, J. (1983). Word order universals. New York: Academic Press. Hawkins, J. (1985). Complementary methods in universal grammar: A reply to Coopmans. LG, 61, 569–587. Hawkins, J. (Ed.). (1988). Explaining language universals. Oxford: Basil Blackwell. Hirsh-Pasek, K., Kemler-Nelson, D., Jusczyk, P., Wright-Cassidy, K., Druss, B., & Kennedy, L. (1987). Clauses are perceptual units for young children. Cognition, 26, 269–286. Huang, C.-T. (1982). Logical relations in Chinese and the theory of grammar. Doctoral dissertation, Massachusetts Institute of Technology, Cambridge, MA. Inoue, A. (1991). A comparative study of Parsing in English and Japanese. Doctoral dissertation, University of Connecticut Storrs. Inoue, A., & Fodor, J. D. (1995). Information-paced parsing of Japanese. In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing (pp. 9–63). Hillsdale, NJ: Erlbaum. Jayaseelan, K. A. (in press). Remnant movement and word order. Syntax. Johnson, E., & Seidl, A. (2008). Clause segmentation by 6-month old infants: A Crosslinguistic perspective. Infancy, 13, 440–455. Kayne, R. (2000). On the left edge in UG. A reply to a reply to McClosky. Syntax, 3, 44–51. Kayne, R. (1994). The antisymmetry of syntax. Cambridge, MA: MIT Press. Kayne, R. (2003). Antisymmetry and Japanese. English Linguistics, 20, 1–40. Kazanina, N., & Phillips, C. (2001). Coreference in child Russian: Distinguishing syntactic and discourse constraints. In A. H.-J. Do, L. Dominguez & A. Johansen (Eds.), Proceedings of the 25th annual Boston University Conference for Language Development (pp. 413–424). Somerville, MA: Cascadilla Press. Kedar, Y., Casasola, M., & Lust, B. (2006). Getting there faster: 18- and 24-month old infants’ use of function words to determine reference. Child Development, 77(2), 325–338. Lashley, K. (1948/1951/1960). The problem of serial order in behavior. In F. Beach, D. Hebb, C. Morgan & H. Nissen (Eds.), The Neurosychology of Lashley. Selected Papers of K.S. Lashley (pp. 507–529). New York: McGraw Hill. Reprinted in B. Lust & C. Foley (Eds.), First Language Acquisition: The Essential Readings (pp.316–334). New York: Blackwell Publishing.
392
B. Lust
Lidz, J. (2007). The abstract nature of syntactic representations: Consequences for a theory of learning. In E. Hoff & M. Shatz (Eds.), Blackwell handbook of language development (pp. 277–303). New York: Blackwell. Lust, B. (in prep.). Universal grammar and the initial state. Cross-linguistic studies of directionality. Cambridge, MA: MIT Press. Lust, B. (1981). Constraint on anaphora in early child language: A prediction for a universal. In S. Tavakolian (Ed.), Linguistic theory and first language acquisition. Cambridge, MA: Massachusetts Institute of Technology. Lust, B. (1983). On the notion ‘‘Principal Branching Direction:’’ A parameter of universal grammar. In Y. Otsu, H. VanRiemsdijk, K. Inoue & K. Kawasaki (Eds.), Studies in generative grammar and language acquisition (pp. 137–151). Tokyo: Monbusho. Lust, B. (Ed.). (1986). Studies in the acquisition of anaphora. Vol. I. Defining the constraints. Dordrecht: Reidel Press. Lust, B. (Ed.). (1987). Studies in the acquisition of anaphora. Vol. 2. Applying the constraints. Dordrecht: Reidel Press. Lust, B. (1994). Functional projection of CP and phrase structure parameterization: an argument for the strong continuity hypothesis. In B. Lust, J. Whitman & M. Sun˜er, (Eds.), Cross Linguistic Perspectives: Vol. I. Heads, Projections, and Learnability (pp. 85–118). Hillsdale, NJ: Lawrence Erlbaum. Lust, B. (2006). Child language. Acquisition and growth. Cambridge: Cambridge University Press. Lust, B., Bhatia, T., Gair, J., Sharma, V., & Khare, J. (1995). Children’s acquisition of Hindi anaphora: A parameter-setting paradox. In V. Ghambir (Ed.), Teaching and acquisition of South Asian languages (pp. 172–189). Philadelphia, PA: Penn Press. Lust, B., Chien, Y.-C., Chiang, C.-P., & Eisele, J. (1996). Chinese pronominals in universal grammar: A study of linear precedence and command in Chinese and English children’s first language acquisition. Journal of East Asian Linguistics, 5, 1–47. Lust, B., & Clifford, T. (1986). The 3-D study: Effects of depth, distance and directionality on children’s acquisition of anaphora. In B. Lust (Ed.), Studies in the acquisition of anaphora: Vol. I. Defining the constraints (pp. 203–244). New York: Reidel Press. Lust, B., deAbrew, K., & Sharma, V. (1982). A cross-linguistic study of the acquisition of complex syntax in Hindi and Sinhalese: An argument for a universal. The 44th South Asian Languages Roundtable. Syracuse University. Syracuse, NY. Lust, B., Eisele, J., & Mazuka, R. (1992). The binding theory module: evidence from first language acquisition for principle C. Language 68(2), 333–358. Lust, B., & Lee, K.-O. (1989). On the acquisition of Korean pronominal anaphora: The roles of configuration and linearity. Presented at the International Circle of Korean Linguistics, University of Toronto. July 1988. In E. Baek (Ed.), Papers from the 6th International Conference on Korean Linguistics (pp. 479–493). Toronto, ON: University of Toronto. Lust, B., Loveland, K., & Kornet, R. (1980). The development of anaphora in first language: Syntactic and pragmatic constraints. Linguistic Analysis, 6(2), 217–249. Lust, B., & Mangione, L. (1983). The principle branching direction parameter constraint in first language acquisition of anaphora. In P. Sells & C. Jones (Eds.), Proceedings of the 13th Annual Meeting of the North Eastern Linguistic Society (pp. 145–160). Amherst: University of Mass. Lust, B., Mangione, L., & Chien, Y.-C. (1984a). The determination of empty categories in first language acquisition of Chinese. In W. Harbert (Ed.), Cornell Working Papers in Linguistics (Vol. 6, pp. 151–165). Ithaca, NY: Cornell University. Lust, B., Mangione, L., & Chien, Y.-C. (1984b). First language acquisition of Mandarin Chinese: Constraints on free and bound null anaphora. In S. Hattori & K. Inoue (Eds.), Proceedings of the 13th International Congress of Linguistics (pp. 1127–1130). Tokyo: Japan.
17
Directionality in the Architecture of the Language Faculty
393
Lust, B., Solan, L., Flynn, S., Cross, C., & Schuetz, E. (1986). A comparison of null and pronoun anaphora in first language acquisition. In B. Lust (Ed.), Studies in the acquisition of anaphora: Vol. I. Defining the constraints (pp. 245–278). New York: Reidel Press. Lust, B., Sun˜er, M., & Whitman, J. (Eds.). (1994). Syntactic theory and first language acquisition: cross-linguistic perspectives. Vol. I. heads. Projections and learnability. Hillsdale, NJ: Lawrence Erlbaum Assoc. Lust, B., & Wakayama, T. K. (1981). Word order in Japanese first language acquisition. In P. Dale & D. Ingram (Eds.), Child language: An international perspective (pp. 73–90). Baltimore, MD: University Park Press Publishers. Lust, B., Wakayama, T., Hiraide, H., Snyder, W., & Bergmann, M. (1983). Comparative studies on the first language acquisition of Japanese and English – language universal and language-specific constraints. In S. Hattori & K. Inoue (Eds.), Proceedings of the 13th International Congress of Linguists (pp. 1131–1135). Tokyo: Japan. Mazuka, R. (1990). Japanese and English children’s processing of complex sentences: An experimental comparison. Unpublished doctoral dissertation, Cornell University, Ithaca, NY. Mazuka, R. (1996). Can a grammatical parameter be set before the first word? Prosodic contributions to early setting of a grammatical parameter. In J. L. Morgan & K. Demuth (Eds.), Signal to syntax: Bootstrapping from speech to grammar in early acquisition (pp. 313–330). Hillsdale, NJ: Lawrence Erlbaum Associates. Mazuka, R. (1998). The development of language processing strategies: A cross-linguistic study between Japanese and English. Hillsdale, NJ: Lawrence Erlbaum Associates. Mazuka, R., & Itoh, K. (1995). Can Japanese speakers be led down the garden path? In R. Mazuka & N. Nagai (Eds.), Japanese sentence processing (pp. 295–331). Hillsdale, NJ: Lawrence Erlbaum. Mazuka, R., & Lust, B. (1987). Why is Japanese not difficult to process?: A proposal to integrate parameter-setting in UG and parsing. In J. Blevins & J. Carter (Eds.), Proceedings for NELS 18 (pp. 333–356). Amherst: University of Massachusetts. Mazuka, R., & Lust, B. (1990). On parameter setting and parsing: Predictions for crosslinguistic differences in adult and child processing. In L. Frazier & J. de Villiers (Eds.), Language acquisition (pp. 147–162). Amsterdam: Kluwer. Mazuka, R., & Nagai, N. (Eds.). (1995). Japanese sentence processing. Hillsdale, NJ: Lawrence Erlbaum. McManus, C. (2002). Right hand, left hand. The origin of asymmetry in brains, bodies, atoms and cultures. Cambridge, MA: Harvard UP. Nespor, M., Guastic, M. T., & Christophe, A. (1996). Selecting word order: The rhythmic activation principle. In U. Kleinhenz (Ed.), Interfaces in phonology: studia grammatica (Vol. 41, pp. 1–26). Berlin, Germany: Academie Verlag GmbH. O’Grady, W., Suzuki-Wei, Y., & Cho, S. W. (1986). Directionality preferences in the interpretation of anaphora: Data from Korean and Japanese. Journal of Child Language, 13(2), 409–420. Reinhart, T. (1983). Anaphora and semantic interpretation. London and Carberra: Croom Helm. Reinhart, T. (1986). Center and periphery in the grammar of anaphora. In B. Lust (Ed.), Studies in the acquisition of anaphora (pp. 123–150). Dordrecht: D. Reidel. Saito, M., & Fukui, N. (1998). Order in phrase structure and movement. Linguistic Inquiry, 29(3), 439–474. Smith, N. (1999). Chomsky. Ideas and ideals. Cambridge: Cambridge UP. Somashekar, S. (1995). Indian children’s acquisition of pronominals in Hindi ‘jab’ clauses: Experimental study of comprehension. Unpublished Master’s thesis, Cornell University, Ithaca, NY.
394
B. Lust
Somashekar, S., Lust, B., Gair, J., Bhatia, T., Sharma, V., & Khare, J. (1997) Principles of pronominal interpretation in Hindi ‘jab’ clauses. Experimental test of children’s comprehension. In S. Somashekar, K. Yamakoshi, M. Blume & C. Foley (Eds.), Papers on language acquisition: Cornell university working papers in linguistics (pp. 65–87). Ithaca, NY: Cornell University. Tomlin, R. (1986). Basic word order. Functional principles. London and Carberra: Croom Helm Linguistics Series. Travis, L. (1989). Parameters of phrase structure. In A. S. Kroch & M. Baltin (Eds.), Alternative conceptions of phrase structure (pp. 263–279). Chicago, IL: University of Chicago Press. Travis, L. (1991). Parameters of phrase structure and verb-second phenomena. In R. Freidin (Ed.), Principles and parameters in comparative syntax (pp. 339–364). MA: MIT Press. Valian, V. (1993). Parser failure and grammar change. Cognition, 46, 195–202. Vigliocco, G., & Nicol, J. (1998). Separating hierarchical relations and word order in language production: is proximity concord syntactic or linear? Cognition, 68, B13–B29. Wagner, M. (2005). Asymmetries in prosodic domain formation. MIT Working Papers in Linguistics, 49, 329–367. http://ling.auf.net/lingBuzz/000429
Index
A Abada, S., 221 Abney, S., 3, 17, 74 Abney, S. P., 227 Abrams, K., 156 Absolutive, 5–6, 11, 16 Accessibility hierarchy, 116, 176, 182, 185, 191, 243, 246, 249, 262, 351 Accusative (case particle), 4, 10, 16, 32, 54, 56–57, 59, 61, 70–73, 136, 138, 141–143, 145–149, 154–155, 160–161, 164–166, 196, 198, 292, 341–343 Acquisition, 137, 165, 369–371, 373, 377, 381–388 Active, 5, 96, 113–119, 121–123, 138, 140, 278, 341–342, 379 Active filler strategy, 96, 278 Active gap strategy, 320 ACT-R, 353–354, 359–360, 364 Acun˜a-Farin˜a, C., 24 Adjunct, 32, 74–76, 86, 89, 179, 187, 277, 300, 305, 371 Adjunct relative clauses, 179 Agent, 117, 121, 133, 137–142, 144–146, 148, 163, 178, 210, 221, 264, 280, 292, 311, 362 Agreement, 3–18, 133, 142, 186, 244, 263, 288 Agreement mismatch, 3–4, 10–11, 13–14, 17 Alpermann, A., 220 Altmann, G., 52, 219, 244 Altmann, G. T. M., 26, 52, 137–138 Amano, S., 49–50 Ambiguity detection, 220 lexical, 51, 177 local, 333–334, 337–338 structural, 241, 279, 300, 309, 331, 336
syntactic, 51, 54, 63, 137, 331, 334, 342, 345 temporal, 49, 53 temporary, 7, 177, 236, 241–273, 277, 280, 287, 294, 310 Wh-scope, 94–97, 99, 103 Anaphora, 5, 319, 382–384 Anaphoric gap, 221, 227 Andersen, E., 173–191, 250–251 Anderson, J. R., 173–191, 250–251 Animacy, 70, 115, 117–120, 122–123, 137–140, 149, 173–191, 195, 210, 241–373 Animate, 114, 117–120, 122–123, 126, 138–140, 145–146, 149, 183–185, 197, 245, 256, 262–266, 269–272, 285, 344 Antecedent, 24–26, 28–30, 33, 40, 42–45, 221, 223–224, 301, 304, 307–309, 312, 315, 317, 382 Anticipation, 7, 138, 352 Anticipatory processing, 189 Anticipatory Structure Building, 17 Antilocality, 352–356, 358–360 Antisymmetry, 372, 374–377 Anwander, A., 385 Aoshima, M., 174 Aoshima, S., 5, 26–27, 45, 96, 250, 288 Arciuri, J., 133 Argument preference, 74, 87, 90 structure, 3, 7, 9–10, 12, 71, 73, 87, 345 Arnold, J., 114, 123–124, 195, 197, 212 Articulation, 114, 134, 371, 373 Auditory sentence processing, 161, 385 Augurzky, P., 69–90 Auxiliaries, 3–4, 6–9, 11, 17–18
H. Yamashita et al. (eds.), Processing and Producing Head-final Structures, Studies in Theoretical Psycholinguistics 38, DOI 10.1007/978-90-481-9213-7, Ó Springer ScienceþBusiness Media B.V. 2011
395
396 B Bach, E., 69 Backward Anaphora, 5, 382–384 Baddeley, A. D., 49 Bader, M., 4, 69, 277, 319, 325–345, 388 Baek, J. Y.-K., 156 Bahlmann, J., 385 Baker, M., 372 Barber, H., 8, 89 Barbier, I., 380, 382 Bard, E. G., 342 Barss, A., 220 Basque, 3–18, 250–251 Bates, E., 263 Bauer, B. L. M., 382–383 Baumann, M., 359 Bayer, J., 388 Beck, S., 154 Beneficiary, 12, 139–140 Bergmann, M., 384 Bernolet, S., 159 Berwick, R. C., 23 Betancort, M., 24, 45 Bever, T., 174, 177, 222, 224, 234 Bever, T. G., 24, 29, 156, 242, 245, 249–250, 265, 277–295, 300, 332 Bhatia, T., 383 Biberauer, T., 379 Bichakjian, B. H., 383 Bierwisch, M., 69 Bi, H., 250 Bilingual mind, 153–166 Bilingual production, 159–160 Bisang, W., 70 Blozis, S., 221 Bock, K., 8, 115, 117, 121, 131, 138–139, 195 Bock, J. K., 114–122, 125, 156–162, 166, 195, 197, 200, 203 Boland, J., 219 Boland, J. E., 3, 24, 314 Bornkessel, I., 70–72, 87, 384–385 Bornkessel-Schlesewsky, I., 70–71 Boser, K., 382 Boston, M. F., 355, 360–391 Bottom-up parsing, 327 Branching Direction, 319, 327–330, 372 See also Left branching; Right branching Branigan, H. P., 113–127, 134, 139, 158, 195 Bresnan, J., 156 Brown, C., 220 Brown, C. M., 8 Bruessow, S., 353
Index C Canonical, 5, 8, 10, 27–29, 45, 53–55, 57–63, 123, 144, 154, 156–157, 160, 163–166, 178–179, 196, 198, 202–208, 211, 221, 241, 243, 245, 247, 271, 277, 280–281, 284, 286, 292, 294, 299 Canonicality, 196, 203, 206, 211 Canonical/non-canonical, 8, 144 Canonical order, 57, 61, 123, 144, 178, 196, 198, 203, 206, 208, 211 Canonical word order, 5, 10, 157, 221, 243, 245, 286, 294 Caplan, D., 50 Carden, G., 382–383 Carlson, G., 219 Carlson, G. N., 3 Carpenter, P., 350 Carpenter, P. A., 35, 49–50, 63, 220, 242, 246, 350 Carreiras, M., 7–8, 24, 45, 89, 250–251 Casasola, M., 382 Case ambiguity, 11, 331 Case marking, 3–4, 7, 9–12, 16, 18, 26, 28, 74, 82–83, 138, 244, 246, 263, 352, 372 Case marking expectations, 10, 16 Case particles, 136–138, 140–144, 146–149 Categorial selectional restriction, 219, 223, 233–234 Center embedding, 329, 340, 344, 357 Chang, F., 121, 123–125, 139–140, 197, 211–213 Chang, P.-C., 179 Charniak, E., 355 Chen, B., 250 Chen, J.-Y., 250, 291 Chiang, C.-P., 384 Chien, Y.-C., 277–295, 384 Chierchia, G., 383 Child, 145, 370, 373, 379–387 Chinese, 173–191, 219, 224, 226, 278–282, 285–288, 290–293, 299–301, 303–304, 307, 310–315, 317, 319, 351, 378–379, 382, 384 Chinese characters, 226 Chiou, F.-D., 180 Chi, T. J., 252 Chiu, B., 286 Choe, H., 154 Choe, J.-J., 155 Chomsky, N., 24, 44, 156–157, 313, 326, 350, 370, 372–373, 377 Choo, M., 22 Cho, S. W., 382
Index Cho, S.-Y., 155 Christiansen, K., 238 Christiansen, M., 221 Christiansen, M. H., 175, 191, 243, 247 Christianson, K., 121, 153–166, 249 Christophe, A., 386 Chu, C., 238, 252 Chu, H., 238 Cinque, G., 376, 379 Clahsen, H., 5, 221 Clark, H. H., 124 Classifier match, 187, 190 Classifier mismatch, 174, 185, 187, 190–191 Classifier position, 175, 181, 185–191 Clause boundary, 4, 10–14, 18, 27, 255, 288–289, 300, 330 Clause union, 326, 343–345 Cleland, A. A., 121–122, 158 Clifford, T., 383 Clifton, C., 44, 74, 89, 137, 263 Clifton, C. J., 137 Clifton, C. Jr, 96, 263 Closed-class elements, 142 Cloze probability, 235 Cloze test, 162–163, 165 Cognition, 370 Cognitive Science, 388 Cohen, L., 244, 246, 284 Coles, M. G. H., 70 Comrie, B., 128, 185, 191, 221, 243, 246, 378 Conceptual accessibility, 114, 116–120, 123, 126 Conceptualization, 114, 134 Conceptual level, 158 Conceptual structure, 133–137 Confederate scripting task, 158–159 Context, 4–5, 7–8, 12, 49, 95, 97–99, 132, 137, 148, 154, 157–158, 173, 175, 186–188, 197–198, 200, 210, 237, 250, 252, 262, 271–272, 284–286, 289–294, 300–309, 319, 326, 342, 352, 360 Control, 11, 23–46, 79, 121, 229, 302–303, 313, 319, 330, 338–339, 356, 360 Control verb, 23, 26, 32–34, 40, 42–43, 45, 330 Coreference, 312, 383 Corley, M., 158 Corpus, 54, 120, 158, 173–191, 195–213, 224, 251, 256, 264, 266, 272, 293, 338–339, 360 Corpus analysis/studies, 174, 179–180, 183, 185, 191, 196, 198, 213, 256, 264, 266, 272
397 Costa, A., 7 Coulson, S., 8 Crain, S., 24, 28, 96, 289–290, 304, 314, 383 Crocker, M. M., 3, 16–17 Crocker, M. W., 3, 17, 70 Croft, W., 180, 185, 371 Cross, C., 382 Cross-categorical harmony, 371, 378 Cross-linguistic, 124–125, 153, 156, 159–160, 162, 165, 173, 185, 246, 263, 350, 379, 385, 387–388 Cross-linguistic priming, 159–160, 162, 165 Cruttenden, A., 76 C-selection selectional restriction, 223, 233 Cummins, J., 165 Cumulative linguistic knowledge, 49–63
D D’Arcais, F. G., 70, 278 Dative (case particle), 74–76, 82–83, 138, 154–155 Dative sentences, 139–140, 154, 165, 199 Dative structure, 153–158, 160–161, 165–166, 197 Davies, W. D., 24 DeAbrew, K., 384 Decay, 350–351, 353–354 Declercq, M., 156, 158–159 Default case particle, 142 Deguchi, M., 94–97, 100, 103, 108 Dell, G., 238 Demberg, V., 360 Demiral, S. B., 71 Dependency Locality Theory (DLT), 176–178, 183, 244, 247, 249, 254, 262, 271, 351, 354, 357, 359–360, 364 De relative clause marker, 227–228, 230, 232–233, 237, 245–255, 257–260, 262–263, 269–271, 311 Desmet, T., 156, 158–159, 221, 284 De Urbina, J. O., 5–6, 10 De Vincenzi, M., 70, 96, 305 Diaz, B., 7–8 Dick, F., 213 Diessel, H., 221 Dillon, B., 8 Directionality, 24–26, 131, 369–388 Discourse context, 95, 187–188, 250, 271–272, 300–301, 304, 308 Discourse referent, 244, 304, 306–307, 351, 353–354 Disharmony, 373, 378–379, 388
398 Ditransitive, 12–13, 16, 53–54, 155, 182, 195–196, 198–199, 206, 212, 220 Dixon, P., 50 Dixon, R. M. W., 5 Donchin, E., 220 Double-object dative, 155, 158–159, 161, 165 Dowty, D., 137, 142 Drenhaus, H., 353, 359 Dryer, M., 197, 378 Dryer, M. S., 173, 325 Dubinsky, S., 24 Duffy, S. A., 53 Dun˜abeitia, J. A., 7–8, 250 Dunlap, S., 250 Dutch, 70, 120–123, 136, 157, 174, 183, 244, 246, 263–265, 284, 380, 382 Dyer, L., 359
E EEG, 79–80, 226–228 Egerland, V., 302 Eisele, J., 382–384 Eisengart, J., 385 Elman, J., 213 Embedded noun, 174, 187, 190–191 Empty category, 299–301, 303–307, 309, 311–312, 314–315, 317–320 Empty subject, 25, 28–30, 32, 34, 38–42, 300–308, 310, 312, 315–316 Encoding interference, 364 Engelmann, F., 358, 382 Erdozia, K., 7–8, 11 Ergative, 5–17 Estival, D., 120 Etxepare, R., 5, 12 Event-related brain potentials (ERPs), 7–8, 69–90, 219–238, 250–251, 264, 283–284 Event-related potentials (ERPs), 7–8, 69–90, 219–238, 250–251, 264, 283–284, 360 Eye-tracking, 4, 88, 190, 220, 250
F Fe´ry, C., 76 Featherston, S., 5 Federmeier, K., 238 Fedorenko, E., 265, 352 Feldt, L. S., 80 Feleki, E., 115, 117 Felser, C., 5 Felser, H., 221
Index Ferreira, F., 131, 137, 142, 156–157, 195, 197, 249, 263, 289, 329, 342 Ferreira, V. S., 114, 120–121, 124–125, 138, 140, 158–159, 197, 200, 210–211 Ferretti, T. R., 137 Fiebach, C., 221 Fiebach, C. J., 72 Filler, 14, 26, 29–30, 34–35, 44, 46, 53, 55, 57, 59, 63, 96, 99, 104–105, 162, 177–178, 221–228, 233, 241–242, 247–248, 250, 255, 266, 271, 278–279, 282, 294, 303, 307, 314, 319–320, 338, 351 Filler-gap integration, 222, 224, 233, 237, 294 Finiteness, 382 Fisher, C., 385 Flores d’Arcais, G. B., 70 Flynn, S., 382, 386 Focus prosody (FPD), 94–95, 97, 103, 107 Fodor, J. D., 4, 23–24, 26, 28, 51, 95–96, 107, 109, 299–320, 326, 345, 387 Fong, S., 23 Foraker, S., 359 Ford, M., 284 Formal linguistics, 157, 159, 166 Formulation, 114, 121, 135, 356, 370, 377 Forster, K., 220 Forster, K. I., 8, 131–132, 285 Forward Anaphora, 382–384 Fox, B., 243 Fox, B. A., 180 Fox Tree, J. E., 157–159, 161, 163 Frauenfelder, U., 284 Frawley, W., 137 Frazier, L., 16, 44, 70, 96, 174, 277–278, 284, 328, 335, 387 Frequency, 34, 43, 53–59, 62–63, 78–79, 175, 177, 179–183, 185, 190–191, 196, 198, 202–203, 205, 208–211, 242, 247, 249, 251, 261–262, 271–272, 287, 310, 334, 336–337, 339, 358, 360, 378–379 Frequency of scrambling, 198, 202–203, 208–211 Friederici, A., 220–221, 235 Friederici, A. D., 70–72, 244, 284, 384–385 Frisch, S., 70–71, 220, 235 Fujisaki, H., 94 Fukui, N., 142, 372–373 Full Constraint Satisfaction, 45 Functional level, 115, 134, 142–143, 149, 158, 161, 166, 209–210, 212 Functional-level processing, 139, 149, 158 Functional processing, 115
Index G Gagne, C. L., 137 Gair, J., 378, 383–384 Gap, 5, 23, 26–28, 30, 41, 45–46, 53–54, 59, 62–63, 96, 174, 176–179, 181–182, 189, 191, 219–238, 241–242, 248, 250, 271, 278–280, 284, 294, 314, 319–320, 351, 359 Garden path effect, 13–17, 236–237, 278, 281, 283–285, 291–294, 318, 330, 334, 336, 338–340, 343 sentences, 220, 329 Garnsey, S., 3, 8, 24, 51, 137, 174, 219, 223–224, 238, 241–273, 285, 291, 314 Garnsey, S. M., 3, 8, 24, 51, 137, 241–273, 285, 291, 314 Garrett, M., 220 Garrett, M. F., 115, 131, 133, 142 Gaze duration, 220, 261 Generic referent, 302, 304, 312 Gennari, S., 180 Gerken, L., 382 German, 4–5, 7, 17, 26, 69–90, 132, 138, 143, 149, 157, 159–160, 244, 246, 263–264, 284, 319, 325–326, 330–332, 338, 340–345, 352, 355, 357–360, 382, 385 Gertner, Y., 385 Gervain, J., 385 Gibson, E., 45, 74, 89, 174–175, 177–178, 183, 220–222, 224, 244–245, 247, 249–251, 265, 271, 284–285, 290–292, 350–353, 356–357 Gil, D., 378 Givo´n, T., 180, 290 Glass, A. L., 342 Goal, 3–4, 44, 52, 133, 137–142, 144, 146–147, 154–156, 199, 219–220, 254, 257, 262, 265–266, 272, 291, 293, 305, 349–350 Goldberg, A. E., 121, 139 Gordon, P. C., 250, 265, 352, 361–364 Gorrell, P., 3, 17, 69, 384–385 Gouvea, A. C., 284 Grammar-parser relationship, 326, 340, 345 Grammatical class, 133 Grammatical encoding, 115–116, 120, 142, 158, 161, 166 Grammatical functions, 115–120, 134, 181, 209–210, 243 Greenberg, J., 369, 371 Greenberg, J. H., 325 Green, D. W., 352
399 Gries, S., 158 Griffin, Z. M., 121, 131 Grimshaw, J., 223, 233 Grinstorm, R., 195 Grodner, D., 221, 247, 284, 350–351, 353, 356 Groothusen, J., 8, 220 Guastic, M. T., 386 Guasti, M., 383 Gussenhoven, C., 74, 87 Gutierrez Ziardegi, E., 7 Gyoba, J., 28
H Ha¨ussler, J., 325, 342 Haegeman, L., 24 Hagoort, P., 8, 220 Hahne, A., 220 Haider, H., 341, 372, 376 Hale, J. T., 355, 360–361, 364 Halle, M., 156 Halliwell, J. F., 249 Harbusch, K., 115 Harris, A., 221, 251 Hartsuiker, R. J., 115, 121–123, 131, 133, 157–160 Haskell, T., 175, 250 Haspelmath, M., 378 Haupt, F. S., 71 Hawkins, J., 114, 123–124, 175–176, 197–198, 211, 251, 371, 378 Hawkins, J. A., 325 Haywood, S., 219 Haywood, S. L., 26, 52 Head direction parameter, 372–373 driven parser/parsing, 3–4, 26, 277, 326–328, 330, 332–334, 337, 339, 343–344, 349 noun, 173–175, 178, 181, 183–190, 241, 245, 247, 250–272, 277–280, 282, 285, 287, 289–290, 293–294, 299, 302, 304–305, 312, 315, 378 parameter, 325 Headedness, 277–278, 294, 299–320, 326, 328–330, 345, 373, 382, 386–387 processing, dependencies, 5, 28, 45, 219–238, 320 Head-initial structure, 319, 328 Heim, S., 385 Henderson, J. M., 329, 342
400 Hendrick, R., 265, 352 Hillyard, S., 235 Hindi, 8, 299, 352–353, 355, 359–360, 383 Hiraide, H., 384 Hirose, Y., 93–109, 140, 212, 280 Hirotani, M., 71, 95–97, 100, 106, 108, 235 Hirsh-Pasek, K., 385 Hoenkamp, E., 115–116, 131, 133–135, 142 Hofmeister, P., 352, 354 Holcomb, P., 220–221, 237, 251 Hollingworth, A., 249 Holmberg, A., 379 Holmes, V. M., 284 Hommel, B., 364 Hong, K. S., 155, 181 Hopper, P. J., 141 Horie, R., 385 Hornstein, N., 302 Hsiao, F., 174, 176–179, 183, 220, 222, 224, 236, 245, 247, 249–251, 284–285, 291 Hsiao, F. P.-F., 351 Hsu, C.-C. N., 174, 187–188, 250, 280, 286–291 Hualde, J. I., 5–6 Huang, C.-T., 378 Huang, C.-T. J., 173, 300 Huang, G., 173, 300, 312, 378 Huang, Y., 312 Hudson, R., 157 Huiskamp, P., 115, 158 Hu, J., 300 Hurewitz, F., 250, 288 Huynh, H., 80 Hyakurakan, 54–55, 57–58
I Inanimate, 117–120, 122, 126, 138–140, 145–146, 183–185, 191, 197, 245, 263–265, 269–272, 285, 344 Incrementality, 23, 26–30, 41, 87, 94, 124, 349 Incremental model, 17, 51, 54, 63, 116 Incremental parsing/processing, 3–18, 23–46, 49–63, 89, 326–328, 330–340, 342–344 Individual differences, 49–63 Infinitival clauses, 330, 335, 341–344 Inoue, A., 4, 9, 23, 26, 51, 326, 387 Inoue, K., 142–143 Integration, 49, 70, 74–76, 84, 86, 89–90, 133, 219–224, 233, 237, 244, 247, 250, 254, 271, 277, 289, 294, 299, 317–318, 349–365, 372–373, 375–377, 384–388
Index Integration cost, 233, 244, 247, 271, 289, 317, 351, 357 Internal argument, 6, 219–220 Intransitive, 5, 72–76, 78, 80–81, 83–87, 89, 144–147, 181 Irwin, D. E., 195, 197, 200 Ishihara, S., 94, 96–97, 100, 103, 108 Ishizuka, T., 285, 291–292 Italian, 70, 133, 300, 383, 385 Itoh, K., 4, 53, 56, 387 Iwasaki, N., 131–149, 212
J Jackendoff, R., 137, 139 Jaeger, F., 213 Jaeger, F. T., 352, 356, 360–361 Japanese, 4–5, 9–10, 23–46, 54, 57, 69, 71, 73, 87, 89, 93–94, 96–97, 99, 108–109, 113–114, 118–126, 131–149, 173–174, 195–213, 244, 246, 250–251, 280, 285–289, 291–292, 299, 310 Jayaseelan, K. A., 377 Jiang, X., 238 Jincho, N., 49–63 Johnson, E., 385 Johnson, K., 154 Johnson, M., 250, 265, 341, 352 Jonides, J., 363 Juliano, C., 243 Jung, Y.-J., 154–155 Jurafsky, D., 179, 355 Just, A., 174 Just, M., 350 Just, M. A., 35, 49–50, 54, 63, 220–222, 242, 246, 284
K Kaan, E., 8, 89, 221–223, 233, 251 Kabata, K., 142 Kaiser, E., 173–191, 251, 293 Kamide, Y., 4–5, 23, 26–27, 52, 63, 138, 141, 219, 330 Kang, S., 286 Kawai, H., 94 Kayne, R., 372–377 Kazanina, N., 221, 384 Kedar, Y., 382 Keenan, E. L., 116, 176, 182, 185, 191, 221, 243, 246 Keil, F. C., 115, 195 Keller, F., 360
Index Kello, C., 3, 329 Kelly, M. H., 115, 138, 195 Kelter, S., 120 Kempen, G., 115–116 Kempen, G. H., 131, 133–135, 142, 327, 330 Kern, S., 357 Khare, J., 383 Kimball, J., 197, 211 King, J., 50, 54, 174, 220–222, 242, 284 King, J. W., 8, 284 Kintsch, W., 49 Kitagawa, Y., 93–109 Kliegl, R., 355, 360–361, 363–364 Kline, P., 350 Kluender, R., 54, 71, 174, 222, 244, 285 Koizumi, K., 143, 199 Ko, K., 221, 284 Kolk, H. H. J., 122, 199 Kolk, H. H. K., 115 Kondo, T., 49–50, 53–54, 56, 195–213 Konieczny, L., 88–89, 352–353 Konopka, A., 158 Korean, 121, 153–166, 174, 244, 246, 250–251, 281, 285–287, 352, 372, 377 Kori, S., 94 Kornet, R., 382 Ku¨hn, K., 70 Kubozono, H., 100 Kuhn, K., 244, 284 Kuno, S., 210 Kuo, K., 174, 177, 179, 183, 190, 245, 250 Kutas, M., 8, 174, 222, 235, 264–265, 284 Kwon, N., 174, 244, 246, 250–251, 285–286
L Labov, W., 120 Laka, I., 7–8, 250 Landau, B., 382 Landau, I., 302, 305 Language Faculty, 369–388 Language production, 114–116, 120–121, 123–127, 153, 158, 186, 195–197, 211–213, 272, 369, 385 Language typology, 371, 379, 388 Larson, R. K., 153–155 Lashley, K., 369 Lasnik, H., 24 Lasser, I., 69, 277, 330 Learnability, 371, 379–381, 386 Least Effort, 44 Lee, J.-H., 156 Lee, K.-O., 384
401 Lee, M., 222 Lee, S.-H., 352, 361 Lee, Y., 265 LeFevre, J., 50 Left branching, 131–132, 326, 377, 383–384, 387 Left-headed, 384, 386 Lemma, 115–116, 135–137, 140, 142–143, 210 Levelt, W., 115, 120, 158 Levelt, W. J. M., 114–116, 158, 200, 209–210 Levine, W. H., 265, 352 Levy, R., 189, 354–356, 360 Lewis, C., 350 Lewis, R., 265, 357 Lewis, R. L., 265, 351–353, 355–357, 361–363 Lexical access, 125, 189–190 Lexical category ambiguity, 331–332, 345 Lexical frequency, 53–62, 336–337 Lexically null head, 223 Li, A. Y.-H., 186 Li, C. N., 181 Lidz, J., 383 Lin, C.-J., 222, 224, 234, 245, 249–250, 300 Lin, C.-J. C., 174, 177–178, 183, 277–294 Lindsley, J. R., 132 Linear Correspondence Axiom, 375 Linear distance, 176, 178, 224, 242, 248–250, 262, 271 Linearization process, 159 Linear Order, 115, 124, 369, 372, 374–376, 381, 383–386 Lin, J-W., 290 Lin, Y., 223–224, 241–273, 285 Lin, Y.-Y., 174 Li, P., 219 Li, T., 225 Liu, A. K.-L., 252 Liu, Y., 219 Li, X., 219 Locality, 45, 96–97, 176, 244, 247, 249, 254, 262, 271, 284, 294, 349–361 Locative/location, 34, 44, 100, 121, 124, 126, 137, 139, 141–142, 144, 146–148, 188–189, 199–200, 206, 224, 234, 237, 353 Loebell, H., 121–122, 138–139, 156–162 Logacˇev, P., 359, 364 Lombardi, L., 121, 158 Lombardo, V., 3, 17 Long-distance dependency, 351 Long-before-short order, 198, 212 Long-before-short preference, 124, 126
402 Losongco, A., 114, 195 Lotocky, M. A., 3 Loveland, K., 382 Lust, B., 326, 369–388
M MacDonald, M., 221 MacDonald, M. C., 3, 50–51, 53, 62–63, 114, 138, 175, 180, 191, 195, 243, 247, 327, 332 MacWhinney, B., 180, 221, 243, 246, 263 Maekawa, K., 94 Mak, W., 244, 246, 256, 264–266, 272 Mak, W. M., 174, 180, 183, 284 Malayalam, 372, 377 Malhotra, S., 8 Mandarin, 173, 175, 177, 185–186, 219–238, 241–273, 378 Mandarin Chinese, 174–175, 180, 191, 224, 280, 282, 299, 303, 307, 379 Mangione, L., 384 Manning, C., 179 Manning, C. D., 355 Maratsos, M., 221 Martin, J. H., 355 Matrix object, 24, 32, 228–234, 251, 264, 279, 313–314 Matrix subject, 24–25, 28–29, 32, 174, 228–230, 232–234, 251, 256, 264, 279, 311–312, 314, 320 Matsuoka, M., 211 Mazuka, R., 4, 23, 49–63, 326, 373, 383, 385, 387 McDonald, J., 115, 117, 120, 195, 197 McDonald, J. L., 138 McDonald, M. C., 8 McDonough, K., 158, 165 McElree, B., 24, 29, 70, 265, 359, 361–363, 365 McKee, C., 383 McLean, J. F., 115, 122 McManus, C., 369 McRae, K., 137 Mecklinger, A., 220, 284 Mehler, J., 244, 246, 284, 385 Meijer, P. J. A., 157–161, 163 Meinshausen, R. M., 132 Merikle, P. M., 50 Meseguer, E., 45 Message-level representation, 158 Miao, X., 185 Miller, G. A., 156, 265, 326
Index Minimal attachment, 277, 301, 306, 308–309, 312, 317, 335 Minimalist (program), 44, 373 Minimality, 72–74 Minimize Domain Hypothesis, 176 Mitchell, D. C., 5, 26–27, 318, 330, 352–353, 356 Mixed headedness, 299, 310, 319, 326, 328, 330, 345, 373 Mixed-order language, 173 Miyagawa, S., 154–155, 196, 211 Miyake, A., 50 Miyamoto, E. T., 5, 10, 23, 26, 42, 53–54, 62–63, 96, 108, 174, 244, 246, 277, 285–286, 291 Mobley, L. A., 8 Morey, R., 122, 138, 158 Morphology, 5–7, 9, 154–155, 251 Morris, R., 221 Morris, R. K., 174, 242, 264, 284 Most Recent Filler strategy, 44 Moving window, 35 Mu¨nte T., 221 Mueller, J., 235 Mueller, J. L., 7 Mulders, I., 3, 17 Muller, H., 222 Myers, E., 3
N N400, 71–72, 74, 76, 83–89, 235 Nakamura, M., 62, 174, 244, 246, 285–286, 291 Nakano, Y., 5 Nakatani, K., 285 Nakayama, M., 23, 286, 318, 352 Namiki, H., 50 Negation, 7, 160, 249 Nespor, M., 373, 385–386 Neville, H., 220 Nevins, A., 8 Ng, S., 280, 299–320 Nicol, J., 24, 29, 220, 384–385 Nicol, J. L., 8 Ning, A., 250 Ninose, Y., 28–30, 42 Nishigauchi, T., 44, 94 Nominative (case particle), 136 Noncanonical, 29 Non-obligatory control, 302, 305 Noun phrase accessibility hierarchy, 176, 182, 243
Index NP conjunction, 117, 119–120 Null head, 233 Null subject, 10–11, 13–17, 53, 62, 279, 300
O Oberauer, K., 363 Object control (OC), 24–25, 28–45, 313 bias, 24, 28–30, 40, 42–46 Object-gap, 182, 221–222, 228–229, 231, 234–235 Object-gap relative clause (RC), 174–185, 188–191, 224–225, 228–230, 232–234, 237, 284 Object-modifying relative clauses, 283, 285, 293–294 Object relative, 50, 233, 241–266, 268–272, 280–282, 284–287, 292 Obligatory control, 23–24, 302, 305 Oda, J., 28–30, 42 O’Grady, W., 383 Oh, E., 153, 155 Ono, H., 96 Ono, T., 135 O’Regan, J. K., 284 Ortiz de Urbina, J., 5–6, 10 Osaka, M., 54 O’Seaghdha, P. G., 114, 195 Osterhout, L., 8, 70, 220, 237 OSV, 70, 113–114, 118–119, 122–123, 157, 178, 243 Overt subject, 11–16, 32–34, 36, 38–42, 203, 303–304
P P600, 9, 220–224, 227, 230–234, 236–237, 250–251, 284 Pablos, L., 3–18 Packard, J., 174, 250–251, 283, 285, 369 Packard, J. L., 219–237 Palmer, M., 180 Pan, H., 300 Parallel function hypothesis, 234 Parameter, 207, 325, 371–373, 376–378, 381, 387 Parsing model, 3, 5, 17, 327 Passive, 5, 113–114, 116–123, 138–140, 144–149, 160, 178, 199, 319, 341–342, 344–345 Patient, 73, 93, 95, 98–100, 106, 117–118, 121, 137, 140–149, 178, 264, 280, 292, 385 Patil, U., 355, 359–360
403 Pearlmutter, N., 175, 244, 327 Pearlmutter, N. J., 3, 8, 51 language, 3, 24, 51–52, 93, 114, 117–119, 124–127, 173, 195, 278, 284, 386 Perea, M., 7–8 Perfetti, C. A., 53 Perspective shift, 244, 246–247, 249, 262, 271 Pesetsky, D., 223, 233 Philipp, M., 70 Phillips, C., 5, 8, 26, 96, 174, 221–223, 233, 250, 286, 288, 384 Phonological encoding, 115, 158 Phrase-length, 125, 197, 199, 202–208, 211–213 Pickering, M., 16, 195 Pickering, M. J., 113–127, 134, 139, 157–159, 195 Picture description task, 144, 158–159 Picture-matching/-description task, 122 Picture-word interference, 132–133 Pinker, S., 133 Plausibility, 51, 63, 187, 256, 263–264, 266 Pleh, C., 221, 243, 246 Polinsky, M., 174, 244, 285 Positional level, 115, 134, 142–143, 158–159, 161, 166, 209 Positional-level processing, 158–159 Positional processing, 115 Post-COMP F0 rise, 94, 97, 103, 108 Post-focal reduction (PFR), 94–95 Postnominal relative clauses, 173 Postposition, 136, 153, 155, 161, 166, 371 Potter, M. C., 121, 139, 158 Prat-Sala, M., 115, 118 Precedence, 374–377, 381–384, 386 Prediction, 5, 7, 9, 11–12, 14, 16, 18, 28, 30, 32–33, 52–53, 59, 87, 143–144, 146, 157, 160–161, 166, 175–176, 179, 187, 219, 243–249, 253, 262, 271–272, 277, 284, 304–305, 307, 315–319, 335–339, 343–344, 349–365, 370–371, 377, 386, 388 driven model, 18, 51, 54, 62–63, 333 Prenominal relative clauses, 173 Prepositional-object dative, 158, 164–166 Price, P., 73 Prime, 121–123, 139, 159–162, 164–166 Principle C, 383 Pritchett, B., 17 Pritchett, B. L., 26, 51, 277, 327, 349 PRO, 24–25, 44, 280–281, 300–302, 304–305, 312–314, 316, 329 Processing bias, 43
404 Processing ease, 177, 183, 191 Pro-drop, 7, 71, 73, 177, 244, 246, 280 Production Japanese, 141 language, 114–116, 120–121, 123–127, 153, 158, 186, 195–197, 211–213, 272, 369, 385 sentence, 114–115, 123, 131–149, 157–158, 213 Pronouns, 4–7, 40, 70, 76, 86, 88, 155, 178, 210, 212, 222, 245, 264–265, 300, 309, 382 Prosodic phrasing, 69–90, 95–96 Prosody, 67–109, 286, 386 Prosody-scope correlation, 95, 97, 100, 103 Psycholinguistics, 7–8, 49, 137, 153, 156–157, 159, 161, 166, 175, 182, 186, 244, 299–320, 369 Psychological reality, 157, 159 Pullum, G., 157 Pu, M. M., 174, 179–180, 183–185, 251, 256, 264, 266, 272
Q Qiao, X., 285
R Randall, J., 44 Rapid Serial Visual Presentation (RSVP), 139, 158, 226 Rayner, K., 16, 53, 220, 328, 387 Reading Span Test, 50, 55, 58 Reali, F., 243 Reanalysis, 9, 12, 14, 16, 27, 32–33, 42, 71, 73, 87, 220, 248, 277, 281–282, 289, 293, 317, 329, 336, 357 Recency, 45 Referent, 23–26, 29, 32, 39, 42, 44–45, 114, 120, 140–141, 180, 183, 185, 187–188, 196, 198, 200, 202–203, 206–208, 211–213, 244, 280, 286, 289–294, 302, 304, 306–307, 309, 311–317, 320, 351, 353–354, 382 Referential phrase, 198, 200, 202–203, 206–208, 211–212, 304 Region of interest (ROI), 313 Reinhart, T., 375 Relational minimality, 72–74 Relative clause (RC) garden-path, 220, 236–237, 271, 278, 281–285, 291–294, 318
Index head-final structures, 299 sentence comprehension, 237, 263, 271, 277, 281, 285 Relative marker de, 223, 232–233, 247 Relativizer, 174, 241, 243, 245, 271, 278–282, 287, 292–294 Remez, R., 382 Retrieval interference, 352, 362 Right branching, 109, 131–132, 326, 387 Right-headed, 222, 372–373, 384, 388 Ro¨sler, F., 71 Roberts, I., 379 Robertson, D., 342 Rodrigo, M., 384 Rohde, D., 265 Roland, D., 213 Ross, J. R., 351 RSVP, 139, 158, 226 Rugg, M. D., 70 Russian, 69, 149, 384
S Sadakane, K., 143, 199 Saddy, D., 220 Saddy, J. D., 8 Saito, M., 166, 372 Sakaki, Y., 28 Sakamoto, T., 25, 28, 30, 33, 42–44 Salamoura, A., 157–159 Salillas, E., 89 Saltarelli, M., 6 Santelmann, L., 382 Santesteban, M., 7 Scheepers, C., 138, 158 Schenkein, J., 120 Schlesewsky, M., 69–90, 220–221, 235 Schmid, T., 325, 342–344 Schoonbaert, S., 158–160 Schriefers, H., 70, 132–133, 143, 174, 244, 246, 263–264, 284 Schu¨tze, C. T., 74, 89 Schu¨tze, H., 355 Schubotz, R., 385 Schuetz, E., 382 Schwartz, B. D., 154–155, 165–166 Scrambled sentences/scrambling, 28–29, 53–54, 57–59, 61–63, 156, 195–213 Scrambled word order, 53–54, 59, 63, 161, 164 Scrambling, 23, 45, 53, 57–59, 62–63, 71–73, 155, 166, 197–199, 202–204, 206–213, 380
Index language, 3–5, 7, 17–18, 23, 51–52, 69, 73, 93–94, 113–127, 149, 156, 161, 163, 196–197, 250, 284, 286, 294, 299–300, 310, 326, 349–350, 358, 360, 365, 386–387 Scrambling negativity, 71–73 Sebastia´n-Galles, N., 7 Seely, R., 221 Seely, R. E., 174, 242, 264, 284 Segalowitz, N., 165 Segui, J., 284 Seidenberg, M., 175 Seidenberg, M. S., 3, 51, 327 Seidl, A., 385 Selectional restriction, 140, 219–220, 223–224, 233–234, 237, 251 Self-paced reading, 4, 14, 26–27, 30, 36, 55, 70, 177, 187, 220, 249, 250, 255, 281–282, 284, 287, 290, 292, 309–310, 314, 316–318, 337–339, 342, 350, 352, 354, 356, 360–361, 363 Self-paced reading task, 30, 55, 317–318 Sells, P., 253 Semantic interference, 133 Semantic selectional restriction, 223 Sentence completion, 4, 121, 131–132, 280, 286–287, 301, 305–306, 308–309, 313, 320 Sentence comprehension, 8–9, 23, 26, 30, 49, 52, 69–89, 96, 137–141, 157, 237, 263, 271, 277, 281, 285, 349–352, 354, 359, 361–364 Sentence fragment completion task, 158–159 Sentence processing, 23, 30, 45–46, 49–63, 108, 137, 173, 175–176, 219–220, 222–223, 233, 249, 265, 272, 277, 284, 294, 318, 326, 354, 360 Sentence production, 114–115, 123, 131–149, 157 Sentence recall task, 116, 118, 158–159 Shared bilingual processing, 165 Shared syntax, 159, 161 Sharma, V., 383–384 Sheldon, A., 234 Shibatani, M., 23 Shin, J. A., 121, 153–166 Short-before-long order, 123–124, 197, 213 Short-before-long preference, 123–124, 197 Shu, H., 219 Similarity-based interference, 265, 349, 361, 364 Simpson, A., 223, 233, 305 Sinhala, 377, 383
405 Siri, S., 133 Smith, J. E. K., 164 Smith, M. C., 129 Smith, N., 372 Snyder, W., 384 Solan, L., 382 Somashekar, S., 383 Sommerfeld, E., 359 Song, N. S., 123, 154 Sorace, A., 342 SOV, 5, 69–70, 113, 118–119, 122–123, 174, 178, 251, 285, 371, 374, 376, 378, 387 Speaker-listener asymmetry, 94, 106, 108 Speech errors, 99, 131–149, 370 Speeded grammaticality judgments, 342 Speer, S., 74 Speer, S. R., 286 Spencer, K. M., 220 Spillover, 317–318, 352–353, 356–357, 360, 364 S-selection selectional restriction, 223, 233 Stabler, E. P., 3, 17 Stallings, L. M., 114, 123–124, 195, 197, 211 Stanovich, K. E., 49–50, 52 Steedman, M., 244, 289–290, 304, 341 Steinhauer, K., 220, 284 Sternefeld, W., 341 Stowe, L. A., 28, 96, 318 Structural ambiguity, 241, 279, 300, 309, 331, 336 Structural distance, 44–45, 242–243, 248–249, 253, 262, 271 Structural frequency, 53–54, 59, 62–63, 360 Structural integration, 244, 247, 254 Structural priming, 114, 120–123, 125–126, 153, 156–162, 164–166 Structural (syntactic) priming, 114, 120–123, 125–126, 138–139, 153, 156–162, 164–166 Sturt, P., 3, 16–17 Subcategorization, 173, 212, 219, 223, 362 Subjacency, 94, 109 Subject advantage, 284 Subject control (SC), 24–25, 29–36, 38–45, 313 bias, 28–30, 33, 42, 45 Subject-gap, 174–185, 188–191, 221–225, 228–230, 232–237 Subject-gap relative clause, 223–224, 228–229, 234, 237 Subject-modifying relative clauses, 283, 293–294
406 Subject relative, 50, 241–262, 264, 266–269, 271–272, 279, 281, 285, 292–293, 305, 308, 361 Suckow, K., 357 Su, I. R., 176, 263, 265 Sumangala, L., 384 Sun˜er, M., 382 Surprisal, 289, 354–356, 360–361, 364 Suzuki, R., 135 Suzuki, S., 135 Suzuki-Wei, Y., 382 SVO, 69, 117, 136, 157, 174, 178, 219, 243, 245, 247, 271, 286, 299, 373–374, 376–378, 387 Swaab, T. Y, 89 Swinney, D., 24, 29, 237 Syntactic ambiguity, 51, 54, 63, 137, 331, 334, 342, 345 Syntactic integration cost, 233, 317 Syntactic priming, 120, 138–139, 158 Syntactic processing, 116, 158–161, 165–166, 299, 318
T Tabor, W., 243 Takahashi, S., 54, 63, 96, 108 Takano, Y., 142, 156, 372 Tamil, 372, 377 Tanaka, M., 134, 139, 149, 195, 197, 211 Tanaka, M. N., 113–127 Tanenhaus, M., 219 Tanenhaus, M. K., 3, 24, 51, 137, 243, 263, 314, 327, 329 Tang, C.-C. J., 186 Tannen, D., 120, 158 Target, 10, 29, 81, 88, 97–100, 102, 104–106, 118, 121–123, 133, 140, 146–147, 159, 161–164, 187, 190, 198–201, 250, 289–294, 303, 306, 308, 311, 313–315, 361, 363 Temporary structural ambiguity, 241, 310 Terao, Y., 137, 142 Teruel, E., 132 Thematic mismatch, 234, 236–237 Thematic roles, 44, 55, 133, 137–144, 146, 148–149, 151, 154, 158, 173, 234–235, 277, 292 Theme, 133, 137–140, 142, 154–155, 163, 220, 311 Theta role, 372 Theta role satisfaction, 24 Thomasello, M., 221
Index Thomas, J., 351, 357 Thompson, S. A., 141, 180–181 Thornton, R., 8 Time (real), 30, 43, 108, 156, 173, 219, 237, 369–388 Ting, J., 286 Tomlin, R., 379 Tomlin, R. S., 176 Top-down parsing, 277, 327, 387 Topicalization, 252–255, 258, 261, 272, 281, 300, 344, 379, 381 Townsend, D. J., 277 Trace, 25, 45, 96, 157, 227, 229–231, 235, 237, 241, 245, 247, 300–302, 305, 312, 361–362 Transitive, 5, 12–13, 43, 53, 57, 59, 61, 72–76, 78, 80–81, 83–89, 117–118, 122, 137, 145–146, 148, 182–183, 188, 191, 196, 198–199, 206, 210, 253, 255, 272, 303 Transitivity, 69–89, 141 Travis, L., 372 Traxler, M., 221 Traxler, M. J., 174, 180, 242, 264–265, 284 Trueswell, J., 219 Trueswell, J. C., 3, 51, 137, 175, 263, 327, 329 Tsujimura, M., 23 Tsujimura, N., 23, 210 Tsujioka, T., 196, 211 Tunstall, R., 247 Twilley, L. C., 50 Typology (language), 371, 379, 388
U Ueno, M., 54, 71, 174, 244, 246, 250–251, 285, 291 Unaccusative verbs, 145–147 Unergative verbs, 145–147 Universal Grammar, 345, 370–371, 373, 377–378, 388 Urushibara, S., 153, 155–156, 165–166
V Vaissie`re, J., 76 Valency, 12 Valentine, S., 157 Valian, V., 387 Van Dijk, T. A., 49 Van Dyke, J., 265, 353, 357, 361–363 Van Dyke, J. A., 265, 352 Vasishth, S., 174, 177, 179, 183, 190, 245, 250, 265, 318, 349–365
Index Veltkamp, E., 121, 157 Verbal Cluster, 7, 17 Verbal Head, 3–4, 7, 10, 16–17, 28, 74 Verbal working memory, 49–63 Verb cluster, 122, 326, 340, 342, 344–345 Verb-final processing, 345 Verb internal argument, 6, 219 Verb selectional restriction, 219, 223–224, 234, 237 Verb selectional restriction satisfaction, 224 Verb subcategorization, 173, 212, 219, 223, 362 Veres, C., 8 Vergara, M., 8, 250 Vigliocco, G., 131, 133, 384–385 Vinson, D. P., 133 Voice priming, 122 Vonk, W., 174, 244, 264, 284 Von Stechow, A., 341 Vosse, T., 327, 330
W Wagner, M., 369, 386 Wakayama, T., 384 Wakayama, T. K., 382 Wanner, E., 221 Warner, J., 342 Warren, R., 114–118, 195 Warren, R. K., 138 Warren, T., 265 Warren, T. C., 351 Wasow, T., 114, 124, 195, 197, 211 Watanabe, A., 94 Watanabe, M., 133 Waters, G. S., 50 Watson, D., 74, 87, 221, 284 Weckerly, J., 264–265 Weinberg, A., 5, 26, 96 Weiner, E. J., 120 Weinstein-Tull, J., 121 Westenberg, C., 121, 123, 158 West, R. F., 49–50, 52 Wh-comp dependency, 97, 108–109 Wh- complement, 221 Wh-construction, 241 Wheeldon, L. R., 121 Wh-focus prosody, 94 Whitman, J., 382 Whong-Barr, M., 154–155, 165–166 Wh-scope ambiguity, 94–96 Williams, J. N., 158–159
407 Williams, R., 221 Wittich, K., 359 Witzel, J. D., 23–46 Witzel, N. O., 23–46 Wolff, S., 71, 73, 87 Woolley, J., 35 Word-order, 195–196, 198, 203, 206–210, 212 Word order, 3–5, 9–12, 23, 53–54, 57–59, 61–63, 69–72, 89, 115, 117–125, 135, 139, 157–158, 161, 164, 166, 173, 195–196, 198, 203–210, 212, 221, 242–245, 247, 249, 251, 263, 271, 278, 285–286, 294, 299–300, 325, 349, 371–373, 378–381, 385 Word order priming, 121–122, 159 Working memory, 49–63, 244, 250, 265, 284, 299, 353–354, 359–360, 364 Wu, F., 173–191 Wu, H.-H. I., 285, 291–293 Wu, H. I., 250–251, 256, 264, 266, 271–272 Wurmbrand, S., 326, 341 Wu, F.-Y., 175 Wu, Z., 223–224, 233
X Xia, F., 180 Xue, N., 180 Xu, L., 300
Y Yamashita, H., 10, 53, 123–125, 140, 166, 195–213, 318 Yang, C.-L., 250 Ye, Z., 174, 219–238 Yngve, V. H., 131 Yoon, J., 238 Yoshida, M., 5, 96, 174–175, 250, 286, 288 Yoshida, S., 5, 96, 174–175, 250, 286, 288 Yoshita, H., 138, 197, 200, 210–211 Yu, L., 238
Z Zawiszewski, A., 7–8 Zhou, X., 174, 250, 283 Zieza, I., 8, 250 Zubin, D., 183 Zushi, M., 23