ABSTRACT Title of Document: CONSTRAINTS AND MECHANISMS IN LONG-DISTANCE DEPENDENCY FORMATION. Masaya Yoshida, Ph.D. 2006 Directed By: Profesor Colin Philips, Department of Linguistics This thesis aims to reveal the mechanisms and constraints involving in long-distance dependency formation in the static knowledge of language and in real-time sentence procesing. Special atention is paid to the gramar and procesing of island constraints. Several experiments show that in a head-final language like Japanese global constraints like island constraints are applied long before decisive information such as verb heads and relative heads, are encountered. Based on this observation, the thesis argues that there is a powerful predictive mechanism at work behind real time sentence procesing. A model of this predictive mechanism is proposed. This thesis examines the nature of several island constraints, specificaly Complex NP Islands induced by relative clauses, and clausal adjunct islands. It is argued that in the majority of languages, both relative clauses and adjunct clauses are islands, but there is a smal subset of languages (including Japanese, Korean and Malayalam) where extraction out of adjunct clauses sems to be alowed. Applying wel-established syntactic tests to the necesary constructions in Japanese, it is established that dependencies crossing adjunct clauses are indeed created by movement operations, and stil the extraction is alowed from adjuncts. Building on previous findings, the thesis turns to the investigation of the interaction betwen real time sentence procesing and island constraints. Looking specificaly at Japanese, a head-final language this thesis ask how the structure of sentences are built and what constraints are applied to the structure building proces. A series of experiments shows that in Japanese, even before high-information bearing units such as verbs, relative heads or adjunct markers are encountered, the structural skeleton is built, and such pre-computed structures are highly articulated. It is shown that structural constraints on long-distance dependencies are imposed on the pre-built structure. It is further shown that this finding support the incrementality of sentence procesing. CONSTRAINTS AND MECHANISMS IN LONG-DISTANCE DEPENDENCY FORATION By Masaya Yoshida Disertation submited to the Faculty of the Graduate School of the University of Maryland, College Park, in partial fulfilment of the requirements for the degre of Doctor of Philosophy 2006 Advisory Commite: Profesor Colin Philips, Chair Profesor Norbert Hornstein Profesor Robert De Keyser Profesor Jefrey Lidz Profesor Amy Weinberg ? Copyright by Masaya Yoshida 2006 i Acknowledgements At last, I have reached the stage where I can write the ?Acknowledgements? for my thesis. Since the very beginning of my graduate years, I wanted to get to this point so badly (so many times, I thought I could not), and I believed that writing this part of the thesis would be the great relief (and indeed it is!) Now that I am given a chance to write this part and I realize an enormous number of people to who I owe a lot for writing this disertation and surviving through my graduate years. I am most pleased by the fact that I have so many good friends, colleagues, mentors, or whatever you may cal them, who had been there for me whenever I needed them. Here, I would like to expres my deepest gratitude to those who helped me come this far. First and foremost, I would like to thank my advisor Colin Philips. I cannot expres how thankful I am to him in this limited space. I am sure that I could not have reached this point without him. Since I wrote to him the first e-mail mesage ten years ago, he is the one who has always been there for me. For me, he is the best mentor without a doubt, and I am so grateful that I have a chance to work with him. After every meting with him, I had a sense that my work was geting beter and beter. He always showed me how to se my scatered and semingly trivial ideas from various diferent points of view, how to give them shape and what the real questions that I should be asking were. The many discussions with him over years at Maryland, which I greatly enjoyed, form the very foundation of this thesis. I would like to thank the members of my thesis commite, Norbert Hornstein, Amy Weinberg, Jefrey Lidz and Robert DeKeyser. It was a great pleasure that I could atend courses taught by Norbert and I could discuss various topics in ii linguistics with him. Through his courses and several discussions I had with him, I learned a lot about how to construct good syntactic arguments. He is also always acesible and helpful. Whenever I brought smal ideas in syntax and psycholinguistics to him, he came up with briliant new ideas that I could never think about (he is like an alchemist). The syntax part of my thesis has been greatly improved thanks to the discussions with him and his razor sharp comments and criticisms. Having Amy as one of the members of my thesis commite was critical. Whenever I had trouble making progres she was the one who I would go talk to. She was always encouraging, cheerful and showed me that I pointed in the right direction. Thanks to her cheerful mood, I greatly enjoyed writing this thesis. I was realy lucky that Jef kindly acepted to be one of my commite members. Even though I gave him my huge defense draft at the latest moment, he read through al the details of it and gave me important comments. The discussions with him were also realy valuable for writing both psycholinguistics and syntax portions of this disertation. I am also grateful to Robert DeKeyser. He also received the defense draft at the final stage. Even though he has just 10 days or so to read through the draft, and he is not a specialist in the topic explored in this thesis, he read the draft and gave me significant comments. I thank him for his warm encouragements through e-mails too. Even though other profesors in the department could not be on my thesis commite, they played crucial roles at every stage of my thesis writing and completing the Ph.D. requirements. In writing this thesis and developing my skils in syntax, I profited from Howard Lasnik so greatly. Through his courses and discussions with him, I established the basis of my syntax skils. Even though, he iv could not be on my thesis commite, a great part of the syntax chapter would not exist without his help and advice. Juan Uriagereka has contributed a great deal of the development of my syntax skils too. His first syntax course (the very first course in syntax that I took at Maryland) provided me many of the tricks and tools that I used in the syntax chapter of this disertation. I realy enjoyed his course as wel as discussions with him. I also would like to show my gratitude to Jacek Witkos who was the other teacher of my first maryland syntax course. Thanks to his positive and encouraging atitude, the first year of my graduate life was realy enjoyable. For me the course requirements were torturously hard in the first year. However, his encouraging mood always helped me to get through the hardest times. I wil not forget the days we played soccer together. I also benefited from discussion of this material with Tonia Bleam. Her enthusiasm and encouragement made talking to her realy an enjoyable experience. Unfortunately we didn?t have much time to discuss my thesis but I enjoyed just talking to her very much. I am grateful to other profesors, although we couldn?t have many chances to talk about my work. I would like to thank Margaret Antonise, Stephen Crain, Bil Idzardi, Paul Pietroski, David Poeppel, Philip Resnik, Rozz Thornton, Alexander Wiliams and Andrea Zukowski. Al of them ade my graduate life realy enjoyable. Maryland is a unique environment where there are always a lot of Japanese linguists. This high concentration of the density of Japanese population was always an esential part of my life at Maryland. Among the Japanese gang, I owe tremendously huge debts to Takuya Goro and Tomohiro Fuji. Takuya has been one of my oldest linguistics-friends and sempais (senior friends) for me. Discussing things v about linguistics with him always benefited me. Sometimes his super sharp comments kiled my premature ideas, but discussing linguistics with him for years has been realy good training. Aside from linguistics, he was and he is always a good friend and good sempai. I am indeed realy lucky to have him as a friend. I owe him a lot. As wel as Takuya, the presence of Tomohiro Fuji is realy important for me. As a friend, as a roomate and as a teacher, I learned a lot of things from him. I realy enjoyed ever lasting discussion on linguistics at the apartment (sometime we stayed up over night!) very much. He was always generous, patient and acesible for casual chit chating as wel as serious linguistics discussion. I am most grateful having Takuya and Tomo as friends and as (unofficial) mentors. I would like to say thanks to Chizuru Nakao, Hajime Ono, Akira Omaki and Kaori Ozawa too. I greatly enjoyed working with Chizuru and Hajime as wel as casualy chit chating with them. It was realy a valuable experience for me to work with them. While, I did not have the chance to work with Kaori and Akira, talking about linguistics or talking about casual topics over coffe or beer with them was realy a valuable time for me. I am realy fortunate to have them as coauthors as wel as close friends. If Sachiko Aoshima and Mitsue Motomomura and Hirohisa Kiguchi had not been at Maryland, my graduate life and my thesis itself would have been so diferent. The impact of Sachiko?s work on this thesis is too obvious to even mention. The psycholinguistic part of this thesis has been built up on what Sachiko has been doing for years. If she were not there, I could not come up with even the most basic idea of this thesis. Additionaly, I learned a lot of skils in psycholinguistics from her. I am vi realy grateful for her generosity and patience. Her cheerful mood and encouragements made discussing linguistics with her an enjoyable and important experience for me. I would like to show my deep gratitude to Mitsue Motomura. From the first time I met her, Mitsue was realy friendly and helpful. In the first few weks or so in Maryland, she did not hesitate using her time to help me out. Without her generous acommodation, the first part of my graduate life would have been much harder. Discussing linguistics with her was also a great joy. She read many of my badly writen manuscripts, and she was realy patient and generous to give me comments and suggestions. I received a lot of input from her. Hirohisa Kiguchi is another sempai who helped me a lot at various stages in my life at Maryland. Whenever I had a hard time, he was always wiling to help me. Also through showing his papers, he taught me how to do syntax, and how to develop an idea. The input from him is also a crucial ingredient for me. I would also like to show my gratitude to my clasmates, Robert Fiorentino, Utako Minai, and Avi Derow. The first year of the graduate life is realy hard. However, I was lucky enough to have them as my clasmates. The wekly discussion group that we had for the syntax courses helped me very much to survive through the first year. Without their help, and friendship, I do not think I could come to this far. Aside from linguistics, Rob and Utako are realy good friends. Going out to drink with them was one of the most enjoyable times for me through al the graduate years. I am glad to have Scott Fults as my friend. He is also one of my friends whose office I visited from time to time and spent a lot of time discussing linguistics and other many things. Here I would also like to say thanks to his family and friends in vii Michigan who provided me with a warm family-like environment: Daniele Hayot, Frankie Jeanne Hayot Fults, Fran Fults, Sarah Fults, Bil Fults, Sandy Fults, Ruth Anna Fults, Dave Rigo, Mary and A.J. Casper. For friendship and great company over the years, many thanks to Nina Kazanina and Leticia Pablos. They showed me constant friendship and warm encouragement. I realy appreciate the time we shared with them (at the ofice or at the bar). Without their encouragement this disertation would not exist. I would like to show my special thnks also to Andrea Gualmini and Luisa Meroni who also showed me constant and long-lasting friendship. I am indebted to al of my felow students, post-docs, RAs, visitors and people at University of Maryland. Special thanks go to, Diogo Almeida, Tim Hunter, Elen Lau, Philip Monahan, Ivan Ortega Santos, John Sprouse, who were wiling to spend time to chit chat with me and encouraged me. Speicificaly, working with Diogo was realy an enjoyable and valuable experience for me. Dave Booth deserves a special mention. I was realy happy to have him as my friend and as my roomamte. His support and encoragement are realy esential for my life in Maryland. I would like to say thanks also to Ann Adelson, Ilya Arakelyan, He-Don Ahn, Rebeca Bayer, Anita Bowles, Ilhan Cagri, Maria Chait, Pritha Chandra, Ray Cho, Joseph Conroy, Stacey Conroy, Maia Duguine, Brian Dilon, John Drury, Chris Dyer, Annie Gagliardi, Silvia Gennari, Sarah Goodman, Ana Gouvea, Lydia Grevenyova, Maximiliano Guimar?es, Martin Hackl, Ryuichiro Hashimoto, So-One Hwang, Yu Iwasaki, Julian Jenkins, Youngmi Jeong, Chunyuan Jin, Johanes Jurka, Eric Hag, Susannah Hoffman, Masaki Kamiya, Ed Kenschaft, Maki Kishida, Jaiva Larsen, Huan Luo, vii Shiti Malhotra, Yuval Marton, Jobin Mathew, Rebeca McKewon, Jairo Nunes, Miki Obata, Ilknur Oded, Lisa Pearl, Silke Urban Plesch, Ariane Rhone, Josh Riley, Cilene Rodrigues, Graziar Russo-Lasner, Lisa Sanders, Itziar San Martin, Michael Sabotin, Rachel Shorey, Usama Soltan, Clare Stroud, Eri Takahashi, Shin-Ichi Tanigawa, Heather Taylor, Yi-Min Tien, Mat Wagers, Jef Walker, Yadong Wang and Ming Xiang. I also thank Kathi Faukingham and Robert Magee, the great secretaries at department of linguistics. Without their help, my graduate life would have been much much harder. I owe a great debt to people outside of Maryland. I would especialy like to thank Tetsuya Sano, Yukio Otsu, Masaru Kajita and Yasuhiko Kato. Without their help and encouragement studying at Maryland could not have happened. Especialy, I owe great debts to Tetsuya who is my first teacher of linguistics and who cared me al the time. I learned a lot of things from him. I also thank him for his constant friendship. This work has been supported by the help from Nobutaka Tazawa. Whenever I come to Japan, he was the one who was wiling to help me. I thank him for his friendship and help. I would like to show my special thanks to the friends from Hiroshima: Hiromu Sakai, Megumi Yoshimura and Chao Zhang. Their help for running experiments and valuable discussions with them are esential parts of this disertation. I would also like to show my gratitude to those who helped me at various stages of my graduate life. Among al the people who helped me over the years, there are four who deserve special mention: Koji Hoshi, Edson Miyamoto, Asako Uchibori and Akira Watanabe. Whenever I needed help, they did not hesitate spending their time. Without their help, advice, and input, this disertation would have been in much ix diferent shape. My thanks also go to, among others, Sachie Akiba, Pranav Anand, Ken-Ichi Ariji, Angel Jes?s Galego Bartolom?, Cedric Boeckx, ?eljko Bo?kovi?, Suzanne Dikker, Chris Davis, Mika Endo, Sam Epstein, Janet Fodor, Aniela Improta Fran?a, Lyn Frazier, Naoki Fukui, Jon Gajewski, Ted Gibson, Arseni Gosev, Kleanthes Grohmann, Valentine Hacquard, Hiroko Hagiwara, Yurie Hara, Naomi Harada, Kaori Hayashi, Ken Hiraiaw, Tsuneaki Hirooka, Yuki Hirose, Masako Hirotani, Hajime Hoji, Natalie Hsu, Atsu Inoue, Miwa Isobe, K. A. Jayaselan, Kyle Johnson, Kohji Kamada, Ruth Kempson, Jieun Joe Kiaer, Hisatsugu Kitahara, Ivona Kucerova, Shinichiro Ishihara, Hironobu Kasai, Takaomi Kato, Masakazu Kuno, Miriam Lemry, Moti Lieberman, Yoshimi Maeda, Reiko Mazuka, Jason Merchant, Keiko Miyagawa, Satoru Muraoka, Michiko Nakamura, Kentaroo Nakatani, Mineharu Nakayama, Fumikazu Ninuma, Masashi Nomura, Lance Nathan, Andrew Nevins, Reiko Okabe, Koichi Otaki, Bun-Sik Park, Tsutomu Sakamoto, Mamoru Saito, Hiroyuki Shimada, Carmen Dobrovie-Sorin, Patrick Sturt, Koji Sugisaki, Shoichi Takahashi, Jun-Ichi Tanaka, Satoshi Tomioka, Mieko Ueno, Ayumi Ueyama, Naomi Yamada, Hiroko Yamakido, Kyoko Yamakoshi, Hiroko Yamashita, Henny Yeung and Tomoko Yoshino. My thanks go also to people outside of the world of linguistics, for their friendship. Thanks go to Makoto Imaizumi, Yoshito Okuyama, Mika Kato, Miki Tsukada, and Kanako Yano. I would like to show my deep gratitude, especialy to Miki and Kanako for their constant friendship, help and encouragement. Finaly and most of al, I would like to thank my parents, Isao and Fukiko Yoshida, my brother Tomoya Yoshida, and my aunt Setsuko Yokota, for their never- x failing love, support and encouragement. I thank my parents for leting me do whatever I wanted to. This thesis is dedicated to them. xi Table of Contents Acknowledgements..................................................i Table of Contents...................................................xi CHAPTER 1. INTRODUCTION.......................................1 1. Incrementality 1 2. The Predictive Mechanism......................................10 2.1. Evidence for Strong Incrementality.........................13 2.1.1. Evidence for Strong Incrementality 1: Prediction of Relative Clauses.........................................14 2.1.2. Evidence for Strong Incrementality 2: Prediction of Conditional Clauses...............................19 3. Toward the Nature of Predictive Mechanism........................24 CHAPTER 2. SYNTAX OF RELATIVE CLAUSES AND ADJUNCT CLAUSES30 1. Introduction 30 2. The Syntax of Relative Clauses..................................31 2.1. Island Sensitivity.......................................33 2.2. Connectivity/Reconstruction Efects.........................39 2.3. Summary 42 2.4. Isues of Verbal Morphology within Relative Clauses...........43 2.4.1. Nominative Genitive Conversion.....................43 2.4.2. The Predicate Adnominal Form and the C-system.........47 2.4.3. Relative Clauses and the C-system....................50 2.5. Conclusion............................................50 3. The Syntax of Conditional Clauses in Japanese......................51 3.1. The non-islandhood of Japanese Conditionals.................51 3.1.1. Adjuncthood of Conditional Clauses: English............54 3.1.2. Japanese Conditionals as Adjunct Clauses..............58 3.1.3. Summary.......................................63 3.2. Conditional Clauses as Weak Islands........................64 3.2.1. Diagnosing Weak Islands...........................64 3.2.2. Mosi as a eak Island Inducer.......................71 3.2.3. Summary.......................................87 3.3. The Distribution of Mosi.................................87 3.3.1. Japanese Nominal Conditionals......................88 3.3.2. Some Basics of Nominal Conditionals.................89 3.3.3. The Conditional Focus Link.........................91 3.3.3.1. The Syntax of the Clausal Pronoun sore in Japanese....91 3.3.3.1.1. Nakao & Yoshida (2005) on Pronominal Sluicing.....91 3.3.3.2. Summary...................................103 3.3.3.3. Sore in Nominal Conditionals....................103 3.3.3.4. Sore and Mosi................................106 xii 3.3.3.5. A Problem Remains...........................108 3.3.3.6. Toward the Licensing Condition of Mosi...........110 3.3.3.6.1. Rizi (1997).................................110 3.3.3.6.2. Focus Constructions in Japanese: Hiraiwa & Ishihara (2001)...............................117 3.3.3.7. The Analysis of Nominal Conditionals.............136 3.3.3.7.1. Further Properties of Nominal Conditionals.........136 3.3.3.7.2. The Derivation of Nominal Conditionals............145 3.3.3.7.3. The Licensing Condition on Mosi.................148 3.3.4. Conclusion..................................148 3.4. The Question of Mosi and Weak Island Efects...............149 3.4.1. Weak Island Inducers.............................149 3.4.2. Mosi as a Scope Marker...........................152 3.4.2.1. The Basic Observations.........................152 3.4.2.2. Disjunction Scope in English and Focus Particles in Japanese....................................158 3.4.3. Intervention Efects...............................161 4. Approaches to Adjunct (Non)Islands.............................161 4.1. Adjunct Islands and Left/Right Asymmetry..................166 4.1.1. Spanish and English..............................166 4.1.2. Scrambling from Conditionals and the Left/Right Asymmetry ...........................................168 4.2. A Parametric Study on Adjunct Islands.....................188 4.2.1. Languages that Alow Extraction out of Adjuncts........188 4.2.2. Parametric Features...............................193 5. Conclusions ...........................................199 CHAPTER 3. RELATIVE CLAUSE PREDICTION IN JAPANESE..........200 1. Introduction ...........................................200 2. Necesity of extra-lexical knowledge in sentence procesing...........201 3. On the Prediction of Japanese Relative Clauses.....................204 4. The Procesing of Japanese Relative Clauses.......................210 5. Experiment 1 ...........................................219 6. Experiment 2 ...........................................227 7. Relative Clause Islands in Japanese..............................237 8. Procesing Filer-Gap Dependencies in Japanese....................240 9. Filer-gap Dependencies and Relative Clause Islands.................244 10. Experiment 1b ...........................................245 11. Experiment 3 ...........................................261 12. General Discussion............................................. ...........................................272 12.1. Procesing of relative clauses.............................272 12.2. Incremental structure building with a powerful predictive mechanism. ...........................................275 13. Conclusion ...........................................277 CHAPTER 4. CONDITIONALS AND LONGER DEPENDENCY FORMATION .........278 xii 1. Introduction ...........................................278 2. On the Motivations for the Longer Dependency Bias.................279 3. The Source of the Bias for Longer Filer-Gap Dependencies in Japanese..282 3.1. Longer Dependency Bias................................282 3.2. Manipulating the Trigger................................287 3.3. The Japanese Filed-Gap Paradigm and the Motivation for the Longer Dependency Bias......................................289 3.3.1. Aoshima et al., (2004) and Aoshima (2003)............289 3.3.2. Manipulating the Features on the Fronted Phrase........295 3.4. Experiments..........................................297 3.4.1. Experiment 1...................................297 3.4.2. Experiment 2...................................307 3.5. General Discussion.....................................312 4. Conditionals and Long-Distance Dependency Formation..............315 4.1. The Question.........................................315 4.2. Japanese Conditionals from a Real Time Perspective...........316 4.3. Experiments..........................................324 4.3.1. Methodology and Predictions.......................324 4.3.2. Experiment 3: Sentence Fragment Completion Task......327 4.3.3. Experiment 4: Self-Paced Reading Task...............330 4.3.4. Experiment 5: Self-Paced Reading Task: Japanese Filed Gap Paradigm......................................336 4.3.5. Experiment 6: Sentence Fragment Completion Task......347 5. General Discussion: The Silhouete of the Predictive Mechanism.........359 5.1. The Nature of the Predictive Mechanism (or the Things that the Parser should be Able to do)...................................359 5.2. Reconsidering the Procesing of Relative Clause Islands (or the Things that the Parser should be Able to Forese)...................363 CHAPTER 5. ON THE NATURE OF PREDICTIVE MECHANISMS.........366 1. Introduction ...........................................366 2. The Requirements for the Predictive Mechanism....................366 2.1. Subjacency Approaches.................................371 2.2. Morphology-Based Approaches...........................378 2.3. On the Requirements for the Predictive Mechanism: Summary...380 3. Perspectives on Online Sentence Procesing: A Quest for Incrementality..381 3.1. Incrementality........................................381 3.2. Mechanisms for Sentence Procesing.......................386 3.2.1. Purely Bottom-Up Strategy and Their Problems.........387 3.2.2. Purely Top-Down Strategies and Their Problems........390 3.2.3. The Left-Corner Strategy..........................393 3.2.4. SPARSE (Schneider 1999): Incrementality and Flexibility.398 4. How to Block the Longer Dependency?.............................408 4.1. Dealing with the Indirect Cue.............................408 4.2. Enhancing the Predictive Component.......................416 4.3. Dealing with the Direct Cue..............................426 5. Conclusions ...........................................436 xiv CHAPTER 6. GENERAL CONCLUSIONS...........................437 Bibliography.....................................................441 1 CHAPTER 1. INTRODUCTION 1. Incrementality The aim of this thesis is to show that sentence procesing is strongly incremental in the sense that the structural commitments are made and interpretations become available on a word-by-word basis and that strong incrementality is made possible by a powerful predictive mechanism. Throughout this disertation I wil show various lines of supporting evidence for strong incrementality and propose a mechanism that can derive it. In this Introduction I would like to discus isues of incrementality and briefly summarize the themes and organization of this disertation. Over the history of sentence procesing research it has been made clear that sentences are procesed incrementaly, in the sense that constituent structures are constructed and interpretations of constituents become available on a word-by-word basis (Aoshima et al. 2004; Bader and Laser 1994; Frazier 1987; Inoue 1991; Marslen-Wilson 1973; Miyamoto and Takahashi 2002; Miyamoto 2002; Sturt and Crocker 1995; Sturt and Lombardo 2005). Initial support for incrementality came from so-caled garden path sentences (Bever 1970, among many others). It was suggested that garden path efects show that when the parser faces an ambiguous input, it does not delay its decision until crucial disambiguating information becomes available, and thus the parser makes a commitment to an initial analysis that may turn out to be an incorrect analysis. When the parser is garden-pathed, a consciously or experimentaly detectable surprise efect occurs at the point where the initial misanalysis is resolved. Let us take an example from English. 2 (1) The boat floated down the river sank. The string before the main verb sank in (1) can be parsed as a simple independent clause. When the main verb is encountered, a surprise efect takes place, and readers experience dificulty in reading. If the parser waits until a crucial element that can confirm the structure of the sentence, this garden path efect is not expected. The fact that readers experience the surprise efect suggests that the parser commits itself to the initial main clause analysis before the main verb is encountered, and the verb sank forced the parser to reanalyze the structure from the main clause structure to the reduced relative clause structure, resulting in the garden-path efect. However, it has been pointed out that Garden Path sentences do not necesarily support the incrementality of sentence procesing. To acount for the garden path phenomena, many researchers have emphasized the importance of licensing heads. For example, the theory proposed by Abney (1987) or Pritchet (1992), the so-caled head-driven models, explained the garden path phenomena in a way that the sentence procesing is driven to satisfy the requirements of the licensing heads (typicaly verbs) at each step of procesing. Under this approach, an example like (1) creates garden path efect because the materials before the verb sank can satisfy the gramatical requirements that the verb floated such as thematic requirements. As a result the string is understood as a simple sentence initialy. Under strictly head-driven acounts of garden path phenomena, it is asumed that the structure is built when the information of the licensing head becomes available. Thus, 3 under this view, the sentence procesing is not necesarily incremental, and garden path phenomena can be explained by head-driven mechanisms. This means that garden path efect is not necesarily a support for incremental model of sentence procesing. Head-driven strategies imply that in a head-final language such as Japanese the procesor delays structure building until the final word of each constituent. In other words, head-driven theories predict that input materials are not integrated into a syntactic structure before the licensing head becomes available in the input (Mazuka and Lust 1988, 1990; Mazuka 1991; Pritchet 1991b, 1992a, 1992b among others). However, there are various pieces of evidence against this particular view either from head-final languages like Dutch (Frazier 1987) or Japanese (Aoshima et al. 2004; Inoue 1991; Inoue and Fodor 1995; Mazuka and Itoh 1995; Miyamoto 2002) or even from a head-final languages like English (Sturt and Lombardo 2005). Let us take a look at some examples from Japanese. Inoue (1991) cites the following example from Japanese. (2) Brown-ga White-ni ringo-o tabeta inu-o ageta. B-nom W-dat apple-acc ate dog-acc gave ?Daniel gave Paul the dog which ate the apple.? Japanese native speakers experience some dificulty dealing with this sentence when they reach the verb tabeta ?ate?. One of the widely acepted interpretations of this garden path efect is the following. The garden path efect is not expected if the 4 parser delays the structure building until it encounter the verb. When the embedded verb ?ate? becomes available, its argument structure information also becomes available. Thus, the parser recognizes that the verb ?eat? does not take a dative argument. The dificulty indicates that the parser commits an analysis before it encounters the embedded verb. Based on the case-particles that each NP bears, the thre NPs are initialy postulated as coarguments of the same clause. When the embedded verb is encountered, it becomes clear that the initial analysis is failed, resulting in the garden path efect. Inoue?s argument is based on native speakers? intuitive judgments, but some experimental studies support the same point too. Miyamoto (2002) found that the upcoming relative clause structure could be facilitated when a sentence contains a sequence of NPs with the same Case Markers. He examined the following pair of sentences. (3) a. Ofisu-de shokuin-ga kakaricho-o [ RC t 1 ocha-o office-at employeee-nom manager-acc tea-acc dasita] josei-ni teineini shookai-sita. served woman-dat politely introduced ?At the office, the employee politely introduced the manager to the woman who served the tea.? 5 b. Ofisu-de shokuin-ga kakaricho-ni [ RC t 1 ocha-o office-at employeee-nom manager-dat tea-acc dasita] josei-o teineini shookai-sita. served woman-acc politely introduced ?At the office, the employee politely introduced the woman who served the tea to the manager.? Miyamoto makes two important observations. First, when there are two acusative NPs in a sentence as in (3a), the second acusative NP creates a disruption. Second, in (3a), the relative head noun was read more easily than in (3b). Acording to Miyamoto these two observations are not explained by head-driven approaches. Under the head-driven approaches, the disruptive efect of the second acusative NP is not predicted because these NPs come into the input before any of the verbs and thus they should not be procesed. Furthermore, head-driven approaches do not predict the facilitation efect at the relative head position in (3a). In both of the conditions in (3), everything is the same except for the acusative NPs. Thus, the facilitation efect cannot be atributed to any other factors than the presence of the two acusative NPs in (3a). Japanese has the so-caled Double Acusative Constraint (Harada 1973a, 1974; Hiraiwa 2002; Kuroda 1965 among many others). Simply put, the constraint excludes a clause containing two acusative NPs. In (3a) the parser has to insert a clause boundary betwen the two acusative NPs in order to analyze the string gramaticaly because of the Double Acusative Constraint. Once a clause boundary is inserted, an embedded clause structure is created and it becomes clear that the 6 subject position in the embedded clause is empty. Based on this information, the parser can predict the upcoming relative clause structure, resulting in the facilitation of the relative head position. Acording to this acount, the acusative NPs before the embedded verb must be procesed in advance of the embedded verb. Thus this finding is not compatible with head-driven approaches, and supports incremental structure building. Finaly, let us review one more study that supports the strong incrementality. Recently, Aoshima, Yoshida & Philips (2006) show that a hierarchical structural relation such as c-command is established before the verb is encountered. First, they observe that in Japanese a pronoun corefers with a c-comanding antecedent, such as the nominative subject NP in (4a), but if the pronoun is not c-comanded by its antecedent as in (4b), coreference is les aceptable. When the dative NP containing the pronoun is scrambled and precedes the antecedent as in (4c), the coreference is aceptable, even though the pronoun precedes the antecedent. (4c) is aceptable because the pronoun is c-commanded by the antecedent at its original position. (4) a. C-commanded Daigakusei1-ga kare1-no tomodachi-ni college-student-nom he-gen friend-dat okane-o kasita. money-acc lent ?The colege student lent his friend some money.? 7 b. Not C-commanded ? Karei-no tomodachi-ga daigakuseii-ni he-gen friend-nom college-student-dat okane-o kasita. money-acc lent ?His friend lent the colege student some money.? c. Scrambled [Kare1-no tomodachi-ni] 2 daigakusei1-ga t 2 he-gen-friend-dat college-student-nom okane-o kasita. money-acc lent ?The colege student lent his friend some money.? They took advantage of this c-command requirement on pronominal coreference to investigate whether the parser establish c-command relation before the verb is encountered. They tested the following types of sentences. In (5a) and (5b), the dative NP containing the pronoun kare ?he? is scrambled, and the gender of the first nominative subjects is manipulated. In (5a), it is female, oba ?aunt? and in (5b), it is male, oji ?uncle?. On the other hand, in (5c) and (5d), the NP containing the pronoun is the first nominative subject, and thus it is not scrambled. The gender of the second nominative subjects is manipulated in the same way. 8 (5) a. Scrambled Mismatch ? [[Kare-no dono kodomo-ni] 1 oba-ga t 1 obentoo-o he-gen which child-dat aunt-nom lunch-box-acc watasita-ka] 2 titioya-ga t 2 oboeteita? handed-Q father-nom remembered. ?? the father remembered to which of his children the aunt pased a lunch box.? b. Scrambled Match ? [[Kare-no dono kodomo-ni] 1 oji-ga t 1 he-gen which child-dat uncle-nom obentoo-o watasita-ka] 2 titioya-ga t 2 lunch-box-acc handed-Q father-nom oboeteita? remembered. ?? the father remembered to which of his children the uncle pased a lunch box.? c. Unscrambled Mismatch ? [[Kare-no dono kodomo-ga] oba-ni he-gen which child-dat aunt-nom obentoo-o watasita-ka] 2 titioya-ga t 2 lunch-box-acc handed-Q father-nom oboeteita? remembered. 9 ?? the father remembered which of his children handed a lunch box to the aunt.? d. Unscrambled Match ? [[Kare-no dono kodomo-ga] oji-ni obentoo-o he-gen which child-dat aunt-nom lunch-box-acc watasita-ka] 2 titioya-ga t 2 oboeteita? handed-Q father-nom remembered. ?? the father remembered which of his children handed a lunch box to the uncle.? Testing these four types of sentences using self-paced reading task, they found that the first nominative NP is read slower in scrambled mismatch condition than in scrambled match condition. However, the comparison betwen the unscrambled mismatch condition and the unscrambled match condition did not show significant diference in reading time. They base their argument on the conclusions from the previous studies on the procesing of backward anaphora in English that when an anaphor is encountered, the parser initiates an atempt finding its antecedent at the earliest gramaticaly sanctioned position in the sentence, and if the potential antecedent is mismatched in gender with the pronoun there is surprise efect so-caled the Gender Mismatch Efect (van Gompel & Liversedge 2003). They argue that their finding paralels the Gender Mismatch Efect in English. When the pronoun is encountered, the parser starts searching for its antecedent. In the scrambled conditions, the first nominative NPs are potential antecedents for the pronouns in the scrambled NPs. Therefore there was a 10 surprise if the gender of the potential antecedent was mismatched. On the other hand, the first dative NPs in unscrambled conditions are not gramaticaly possible antecedents for the pronouns contained in the nominative subjects. Thus there was no surprise. For our discussion, the most important point in their finding is that the parser calculated c-command relation before the verb becomes available. In other words, even before the verb comes into the input, the parser builds hierarchical syntactic structure. As we can se, the gender mismatch efect is detected at the first nominative NP position in the scrambled conditions that is the position before the verb becomes available. Furthermore, the coreference relation betwen pronouns and their antecedents in Japanese are sensitive to c-command as we have sen. Thus, the gender mismatch efect that they observed indicates that the parser calculated the coreference relation betwen the pronoun and its antecedent (c-command relation) before the verb becomes available. Results of these studies and many others suggest that human sentence procesing is strongly incremental, in which input materials are asembled into a syntactic structure from left-to-right without a delay. 2. The Predictive Mechanism Even though various experimental studies have suggested that human sentence procesing is strongly incremental, there are also various chalenges to the incremental structure building. Lombardo and Sturt (2002) point out an example that creates a potential dificulty to the strong incremental structure building. 11 (6) a. He thinks [ CP [ IP [ NP [ Adj steper] prices] have come about because ?]] b. The example (6) contains an atributive adjective that modifies the subject NP of the embedded clause. When the adjective steper is encountered during online sentence procesing, a strongly incremental parser incorporates the adjective into the current representation without waiting for other lexical items to become available. In order to do so, however, the parser has to build the structure of the NP and IP, which do not have overt heads at the point where the word steper is encountered. Upon encountering the adjective, the parser has to project an NP node to host it. The IP node is to be projected too to host the predicted NP. Thus a strongly incremental parser has to predict an IP node based on the predicted NP node in order to acommodate the adjective into the structure. A chalenge for a strong incremental parser is whether such complicated prediction can be done without a guide from the lexical information. If, on the other hand, the structure building is not strongly incremental and if it is alowed for a certain delay, this example does not create any problem. Facing an example like (6) a delay parser can wait until the NP price becomes available to host the adjective in to the structure. 12 In this respect, Japanese head-final sentences create this type of chalenge almost everywhere. The following simple example can ilustrate this point. (7) a. Brown-ga [ CP [ IP White-ga ringo-o tabeta-to]] B-nom W-nom apple-acc ate-comp itta. said. ?Brown said that White ate an apple.? b. IP NP VP B-nom CP V IP C ita NP VP to W-nom NP V aple-ac tabeta (7) contains a simple multi-clausal structure in which a VP is taking a CP as its complement. When the second nominative subject NP, White-ga is encountered, a strongly incremental parser has to build an embedded IP node, a CP node a VP node and the matrix IP node in order to incorporate two nominative subjects into a structure. In doing so, the parser has to commit to a multiple syntactic predictions. It has to project two IP nodes in order to host Brown-ga and White-ga as subject NPs. It has to project a CP node, because the structure contains White-ga is an embedded clause. It has to project a VP node because there must be a V that takes a CP as a subordinate clause. 13 It is obvious that this type of approach is risky because the structure that the parser builds before encountering lexical verbs can turn out to be incorrect. For example, even if there are two nominative NPs the sentence can be a simple clause with a nominative object in Japanese such as Brown-ga White-ga kiraida ?Brown hates White?. Thus it is easy for a strongly incremental parser to make a mistake. A delay parser, on the other hand, does not need to commit such a complicated multi-step calculation in order to proces a sentence like (7). Furthermore, the delay parser does not need to commit to a particular analysis before it encounters a crucial licensing head. What the delay parser has to do is to wait until lexical heads, such as a verb to come into the input, and build the structure using the information conveyed by the lexical heads. For example the lexical verb with complementizer atached can help the parser to recognize that the there is a embedded clause and the second nominative NP is the subject of the embedded clause and so on. With such an ambiguity problem, for example, is a strongly incremental parser a plausible model for human sentence procesing? Is there any evidence that suggest strong incrementality? 2.1. Evidence for Strong Incrementality In this disertation, I wil show evidence that strongly supports fully incremental models of sentence procesing. Based on these pieces of evidence, I wil investigate what mechanism can capture the strongly incremental nature of sentence procesing. Evidence for strong incrementality comes from the studies on two constructions in Japanese, relative clauses and conditional clauses. 14 The disertation is organized as follows. In Chapter 2, the basic syntactic properties of relative clauses and conditional clauses wil be summarized. In Chapter 3, experimental studies on Japanese head-final relative clauses are summarized. Chapter 4 turns to experimental studies on Japanese conditional clauses. In these two chapters, I wil present evidence for strong incrementality. Finaly in Chapter 5, I wil propose a possible mechanism for a strongly incremental structure building. 2.1.1. Evidence for Strong Incrementality 1: Prediction of Relative Clauses The first evidence for strong incrementality comes from the procesing of head-final relative clauses in Japanese. As is widely recognized, Japanese head-final relative clauses normaly do not have any overt markers corresponding to the relative pronouns in English. Because of this property, it is very dificult to determine whether a string of phrases is part of a relative clause or not. For example, during the online procesing of the example in (8), it is very dificult for the parser to recognize that this sentence contains a relative clause before the verb or the relative head is encountered. There are several reasons for this. One obvious reason is that the same string that constitutes the relative clause in (8) can be an independent simple clause as in (9) because of the lack of unambiguous relative clause markers or because of the availability of empty arguments (Hirose 1999; Inoue 1991; Mazuka 1991; Yamashita et al. 1993; Yamashita 1995 among many others). 15 (8) Brown-ga[ NP [ RC White-ga Blue-ni gap ageta] hon]-o B-nom W-nom B-dat gap gave book-acc yonda. read ?Brown read the book that White gave to Blue ? (9) Brown-ga White-ni gap ageta. B-nom W-dat gave ?Brown gave something to White.? In Chapter 3 of this disertation, I wil show that there are some elements that can potentialy mark the beginning of the relative clause unambiguously. (10) is one of such examples. (10) Brown-ga [ NP 3-satu-no [ NP [ RC White-ga Blue-ni B-nom 3-cl (bok) -gen W-nom B-dat ageta] hon]]-o yonda. gave book-acc read ?Brown read thre books that White gave to Blue.? (10) contains a genitive-marked numeral clasifier that is located before the relative clause. The important property of this type of construction is that the genitive-marked clasifier can only be asociated with the head of relative clause. Normaly genitive clasifiers can be asociated with its adjacent NP. However, in the example above, it can only be asociated with the relative head because it is not semanticaly 16 compatible with its adjacent NP, White-ga. This type of long-distance asociation of clasifier and its host NP, and the semantic incompatibility betwen the clasifier and its adjacent NP only take place in a complex NP environment and typicaly in relative clause environment. Thus, the semantic incompatibility betwen the genitive-marked clasifier and its adjacent NP can unambiguously mark the beginning of the relative clause. During online sentence procesing, if the semantic incompatibility betwen the clasifier and its adjacent NP is encountered, and if the parser can recognize that the semantic incompatibility is possible only in the relative clause environment, it is possible that the parser can predict the upcoming relative clause structure. Thre experiments in Chapter 3 wil show that this is indeed the case. The following two types of sentences were compared. (11) a. Clasifier Mismatch Brown-ga 3-satu-no gakusee-ga katta B-nom [ NP 3-cl (bok) -gen [ NP [ RC student-nom bought] hon-o yonda book]-acc read ?Brown read thre books that the student bought.? b. Clasifier Match Brown-ga [ NP [ RC [ NP 3-nin-no gakusee]-ga katta] B-nom 3-cl (person) -gen student-nom bought hon]-o yonda. book-acc read 17 ?Brown read the book that the thre students bought.? The first example contains a relative clause with semantic incompatibility betwen the clasifier and its adjacent NP. It is caled the clasifier mismatch condition. In this case, as the brackets and the translation show that the clasifier can only be asociated with the relative head. The second example, on the other hand, contains a relative clause with the clasifier that is semanticaly compatible with its adjacent NP. In this case, the clasifier can be asociated with its adjacent NP, and is not necesary to be asociated with the relative head even if the relative head can host the clasifier. This condition is caled the clasifier match condition. A self-paced reading experiment is conducted and it is shown that the embedded relative clause verb is read more easily in the clasifier mismatch condition than that of the clasifier match condition. This result suggests that the parser can predict the upcoming relative clause structure by means of the clasifier mismatch. Because the upcoming relative clause structure is predicted, the relative clause verb, which has a special morphology uniquely licensed in relative clauses and other complex NP environment (se Chapter 2 for details of this point), is read more easily. Furthermore, the result suggests that the relative clause structure is built before the embedded verb or the head of the relative clause is encountered. Thus this finding is fully compatible with strongly incremental parsing models that build the structure before encountering rich information bearing units like the verbs or the head of the relative clause. 18 In chapter 3, we wil also se evidence that the predicted structure of relative clause is rich enough to block the long-distance dependency. This wil be shown through a study on the island efect induced by relative clauses during online procesing. Japanese relative clauses like those in other languages are islands (Saito 1985). Thus, an overt movement like scrambling cannot escape the relative clause island as in the example (12a). (12) a. *Dono-gakusee-ni 1 [Brown-wa [ NP [ RC White-ga t 1 which-student-dat B-top W-nom ageta] hon]-o yonda-no]? gave book-acc read-Q Lit. ?Which student did Brown read the book that White gave the book to?? b. ok Dono-gakusee-ni 1 [Brown-wa[ CP White-ga t 1 hon-o which-student-dat B-top W-nom book-acc ageta-to] itta-no]? gave-comp said-Q ?Which student did Brown say that White gave a book to?? On the other hand, in the experimental studies by Aoshima (2004), it has been shown that when the sentence contains a scrambled phrase as in (12b), the scrambled phrase is preferentialy asociated with the embedded clause. Thus, during the online sentence procesing, when the scrambled phrase dono-gakuse-ni ?which student? is 19 encountered, the parser tries to interpret it in the embedded clause rather than in the matrix clause. Let us cal this preference of the parser, the Longer Dependency Bias (LDB). Given this LDB and the islandhood of relative clauses as wel as the possibility of predicting the upcoming relative clause structure, we can make following two predictions for the behavior of the parser. First, if the predicted relative clause structure does not represent islandhood, the fronted phrase is preferentialy interpreted in the embedded relative clause, resulting in a violation of relative clause island constraints. On the other hand, if the islandhood of the relative clauses is represented, the LDB wil be blocked. The results of the experiment wil be shown in chapter 3 that indicate that LDB is blocked when the upcoming relative clause is predicted by means of the clasifier mismatch. An important observation in this study is the following. It wil be shown that genitive-marked clasifiers do not have any specific gramatical relation to the relative clause structure per se. Thus, the parser makes use of such ?indirect? information to project the upcoming relative clause structure. Taken together, these two pieces of evidence strongly suggest that the parser can build rich representation of the relative clause structure upon encountering the clasifier mismatch before the information from the verb or the head of the relative clause becomes available. 2.1.2. Evidence for Strong Incrementality 2: Prediction of Conditional Clauses In chapter 4, I wil present another piece of evidence for strong incrementality. In this chapter we wil look at Japanese conditional clauses. 20 The detailed descriptive syntactic studies on conditional clauses in chapter 2 wil reveal thre important properties of Japanese conditional clauses. First, Japanese conditional clauses can be introduced by so-caled conditional adverbs. One such conditional adverb is Mosi. An important property of Mosi is that it can be licensed only by conditional clauses. Second, conditional clauses are distinguished from other types of clauses by specific verbal morphology. Typicaly the conditional suffix -ra, which is understood as a complementizer, is atached to the verb. Finaly, conditional clauses are not islands even though they are adjunct clauses. An example of conditional clauses is in (13). (13) Brown-wa [ Cond mosi White-ga Blue-ni present-o B-top mosi W-nom B-dat present-acc ageta-ra] naki-dasu daroo give-cond cry-start will ?Brown wil cry if White gives a present to Blue.? These thre properties of conditional clauses alow us to test strong incrementality too. In chapter 4, I wil show the results of experiment that suggest that conditional verb is read more easily in the sentence containing Mosi compared to the one that does not contain Mosi. The following two types of sentences are compared in self-paced reading experiment. 21 (14) a. Mosi Condition Brown-wa [ cond mosi kawaii onnanoko-ga Blue-ni B-top mosi pretty girl-nom B-dat present-o ageta-ra] naki-dasu daroo. present-acc give-cond cry-start will ?Brown wil cry if the prety girl gives a present to Blue.? b. Adverb Condition Brown-wa [ cond totemo kawaii onnnanoko-ga Blue-ni B-top very pretty girl-nom B-dat present-o ageta-ra] naki-dasu daroo. present-acc give-cond cry-start will ?Brown wil cry if the very prety girl gives a present to Blue.? The Mosi condition contains conditional adverbs. On the other hand Adverb condition contains degre adverbs that do not have any gramatical relation to conditional clause structures. Embedded clauses are conditional clauses in both conditions. In this comparison, the verb with the conditional morpheme, -ra is read more easily in the Mosi condition than in the Adverb condition. This result suggests that the parser sets a prediction of the upcoming conditional structure by means of Mosi. Because Mosi has a direct gramatical relation to conditional clauses, it wil be argued that the parser can make use of such direct cue to predict the upcoming conditional clauses. This finding also is compatible with strong incrementality because the prediction of the upcoming conditional clause is set before the actual conditional verb is 22 encountered. Thus, the parser can project the structure of conditional clauses upon encountering the adverb Mosi. Additionaly, I wil show that the predicted conditional clauses also block the LDB. As we have discussed, Japanese conditional clauses are not islands. Thus an overt movement like scrambling can escape conditional clauses as in (15). (15) a. ok Dono-gakusee-ni 1 Brown-wa [ cond mosi White-ga t 1 which-student-dat B-top mosi W-nom present-o ageta-ra] naki-dasu-no? present-acc give-cond cry-start-Q ?Which student wil Brown cry if White gives a present to?? b. ok Dono-gakusee-ni 1 [Brown-wa [ CP White-ga t 1 which-student-dat B-top W-nom present-o ageta-to] itta-no]? present-acc gave-comp said-Q ?Which student did Brown say that White gave a present to?? Given the results of the studies on relative clauses, we expect that the fronted wh- phrase is preferentialy asociated with the embedded clause in examples like (15a) because of the LDB and non-islandhood of conditional clauses. However, contrary to our expectation, the LDB is blocked when the upcoming conditional clause is predicted by means of Mosi. I wil explain this blocking efect of the LDB in the following way. I wil show two important observations. First, the results of two experiments wil show that 23 the LDB is motivated by wh-feature on the fronted NP. The LDB is observed only when the fronted material is a wh-phrase. I wil argue that the parser tries to asociate wh-phrase with Question particle as soon as possible, resulting in the LDB. Second, the conditional verb, V-ra, cannot host the question particle, -ka. Based on these two observations, I wil argue that the LDB is not sen in the predicted conditional clauses because the morphology of the conditional verb, V-ra, is predicted by Mosi. Because the predicted conditional verbal morphology is not compatible with the motivation for the LDB, it is blocked when the conditional clause is predicted. The finding that the LDB is blocked when the conditional verb is predicted suggests that the parser projects a detailed structure of the conditional clause that can represent the morphology of conditional verbs. Based on the discussion in chapter 2, in which I wil show that conditional verbal morphology is created by V-T-C complex, I wil argue that the parser project the full structure of conditional clauses where V, T and C are al represented to support the conditional morphology. Finaly, I wil argue that the finding that the predicted conditional morphology blocks the LDB, forces us to reconsider the online island efect induced by relative clauses. As we have discused, conditional clauses are not islands in Japanese. However, they block the LDB. If so there can be some factor independent of islandhood, which is relevant to the blocking of the LDB. I wil show that the morphology of the relative clause verbs is also not compatible with the question particle. On the basis of this observation, I wil argue that the LDB is blocked when the relative clause is predicted not because relative clauses are islands but because the morphology of relative clause verbs is predicted. 24 3. Toward the Nature of Predictive Mechanism The findings in chapter 3 and chapter 4 support the strong incrementality of sentence procesing. Experimental results suggest that both relative clause structures and conditional clause structures are built before the verb or the head of the relative clause is encountered. These findings suggest that the parser is equipped with a powerful predictive mechanism that alows for the strong incrementality. Furthermore, this predictive mechanism should be able to project the detailed structure of upcoming relative clauses or conditional clauses by means of indirect cues such as the clasifier mismatch or direct cues such as conditional adverbs like Mosi. In Chapter 5 I wil discuss the isue of the sentence procesing mechanism. Specificaly I wil try to figure out what algorithm can derive the range of predictions that we have sen above. I wil criticaly examine four types of parsing algorithm that have been proposed in the literature, and show that a variant of the so-caled left- corner algorithm is a psychologicaly plausible algorithm for sentence procesing. First, I wil argue that both purely bottom-up algorithms and purely top-down algorithms have problems. It wil be shown that purely bottom-up parsing is not incremental. On the other hand, purely top-down algorithm has incrementality, but it has problem with left-branching structures. I wil show that the so-caled Left-Corner parser is psychologicaly more plausible than these two algorithms (Abney and Johnson 1991; Johnson-Laird 1983; Resnik 1992; Stabler 1994 among others). Asuming some basic phrase structure gramar, a left-corner parser proceses the sentence in the following way. It proceses the leftmost daughter of the right-hand 25 side of a phrase structure rule in a botom-up fashion, and the rest of the rule in a top- down fashion. Let us se how a left-corner parser proceses the sentence ?the man read the book.? (16) The man read the book. When the parser finds the word the, it alows a Det node to be built (the step 1: the bottom up structure projection). Det is the left-most category at the right-hand side of an NP rule like [NP - > Det, N], the NP rule is applied and the NP node is built. The NP rule tels that it can be expanded as Det and N (NP - > Det N). Thus the sibling of the NP, the N node is also projected (the step 2: the top down prediction). The word 26 man comes into the input, and it alows the N node to be built (the step 3), which can be atached as the right-hand member of the NP, and NP is completed by ataching the newly projected N into the predicted N node (the step 4). When the NP is completed, an S rule like [S - > NP VP] tels that NP is the left-most category of the right-hand side of S rule. Thus, the NP alows the S node and its sibling VP-node to be built (the step 5). The word read comes into the input, which can project up to V and VP (the step 6). VP is the right most member of the currently constructed S node. So the newly projected VP is atached to the predicted VP (the step 7). The word the comes into the input and as we have sen it can project an NP node and an N node, the sibling of the NP can be predicted (the step 8 and 9). The word book comes into the input, projecting the N node. The newly projected N node is incorporated into the predicted N node (step 10 and 11). The newly projected NP is the right-most category of the current VP node, thus the newly projected NP can be incorporated into the predicted NP node (step 12) and the sentence structure is completed. In this way, a left-corner parser holds incrementality by means of top-down prediction, and it can resolve the problems of botom-up and top-down algorithms. However, Schneider (1999) points out that even this algorithm has some problems. He argues that because left-corner parsers refer to syntactic categories, they require frequent reanalysis in the procesing of a sentence in head-final languages. For example, in German, a case-marked NP may be licensed by either a verb or a postposition. Thus it is risky to commit a particular analysis. To minimize the amount of reanalysis needed in head-final languages, Schneider proposes a feature-based left-corner parser caled SPARSE. In SPARSE, 27 features are the minimal building blocks. This parser alows for the prediction of features rather than categories. For example, for a German case-marked NP, SPARSE predicts a head with the case feature. Because it does not refer to a particular category, the predicted head is compatible with either a verb or a postposition. In this way, SPARSE has a flexibility in predictive component. Based on this flexibility and incrementality, I wil basicaly adopt SPARSE as the basic structure-building algorithm. SPARASE alows flexible structure building. However, like other algorithms, left-corner parsers in general, including SPARSE, do not provide recursive prediction that is required for an example like (6) where a word to be incorporated in the structure, there must be multiple structural predictions. I wil show that the recursive prediction problem arises in the prediction of relative clauses. To acommodate the recursive prediction, I wil modify SPARSE by enhancing the predictive component. The modification of the parser raises a problem. Enhancing the predictive component to alow for the recursive prediction may make the parser too powerful in a way that the parser alows for an infinite recursive prediction. To resolve this problem, I wil also discuss how we can restrict the power of predictive component. In chapter 5, I wil show how the prediction of relative clauses or conditional clauses can be handled by SPARSE. I wil point out that the original SPARSE algorithm cannot predict the sufficiently rich structure of relative clauses or conditional clauses based on the cues such as the clasifier mismatch or Mosi. To resolve this problem, I wil modify SPARSE so that we can enhance the predictive component of the parser. 28 The key points of my modification rely on the gramatical properties of the clasifiers and Mosi. Both of these items should be able to set a prediction of specific heads. For example, the clasifier for books should be able to predict the upcoming N head with the semantics of book. On the other hand, Mosi should be able to predict the upcoming C which is specified for conditional clause type. The predicted C?s specific morphological form -ra or its variants, which are understood as complementizers (Chapter 2), should also be predicted by means of Mosi because it is the licensor of Mosi (Chapter 2). 1 (17) Based on this asumption, I wil propose that only when the prediction of such specific heads is possible can the parser aces the features of the predicted heads, and only then can these features alow for the top-down prediction. For example, if the N head book is predicted, the information that book does not take a complement clause is also predicted. I wil make this type of top-down prediction using the information from the predicted head possible only when a specific head with specific semantic information is predicted. With this restriction, the parser alows a powerful recursive prediction in a certain limited cases. In a normal situation, the parser can 1 The relation betwen mosi and its licensing morphology is to be understod as a general property of dependent elements and their licensors such as wh-phrase and Q-particle. 29 only predict the upcoming underspecified head as the original SPARSE does, thus the flexible structure building is not sacrificed even though a powerful multi-step prediction is possible. In chapter 5 I wil show how this modification alows for the parser to predict sufficiently rich structure of relative clauses or conditional clauses in which the LDB can be blocked. 30 CHAPTER 2. SYNTAX OF RELATIVE CLAUSES AND ADJUNCT CLAUSES 1. Introduction The aim of this chapter is to figure out the descriptive properties of Japanese Relative Clauses (RCs) and Conditional Clauses. The descriptive syntactic studies help us understand what the representations of these clauses should look like, that the parser construct during online sentence procesing. The study on Conditional Clauses is specificaly important for us because in Japanese Generative Gramar, the syntax of Conditional Clauses have not been studied much, and therefore representation of conditional clauses are not wel understood. Throughout this chapter, we wil try to figure out the internal syntax and external syntax of both RCs and Conditional Clauses by applying various wel-known syntactic tests. In the course of the discussion, it wil become clear that Japanese Adjunct Clauses are not islands. The non-islandhood of adjunct clauses is surprising given the fact that adjunct islands are observed in wide varieties of languages (Stepanov 2001). Thus, in this chapter we wil try to capture the non-islandhood of Japanese adjunct clauses too. Comparing Japanese with various other languages, we wil propose a parameter that regulates the islandhood of adjunct clauses. The organization of the chapter is as follows. In the next section, we wil figure out the descriptive properties of Japanese RCs. In the following section, we wil turn to the syntax of Conditional Clauses. Finaly, we wil turn to the cross- linguistic studies on adjunct clauses. 31 2. The Syntax of Relative Clauses The aim of this section is to briefly review the syntactic properties of relative clauses (RCs) in Japanese. Although there are various types of RCs in Japanese, I wil concentrate on one type of RCs, the so-caled head external restrictive RCs. I wil report thre prominent properties of RCs, namely: (i) island sensitivity; (i) reconstruction efects; and (ii) verbal morphology. I wil show that they are best captured by a CP-analysis of RCs. One of the wel-known properties of Japanese RCs is their apparent insensitivity to island constraints. It has been reported in various places that Japanese relativization can escape islands (Fukui and Takano 2000; Kuno 1973; Murasugi 1991 among others). Taking some examples, Kuno cites the following examples containing, complex NP islands, adjunct islands, and subject islands. They are summarized in (18) to (20). Al of these examples are aceptable even though the relative head is extracted out of island domains. Recently, it has been argued that clausal adjuncts and clausal subjects do not sem to be islands in Japanese (Ishii 1997; Mihara 1994 among others). Therefore, the facts about these two domains are not surprising. However, it is stil surprising that Japanese relativization can escape Complex NPs that normaly show strong island efects (Saito 1985). 32 (18) Complex NP Islands [ NP [ CP __ 2 [ NP [ CP __ 1 kiteiru] huku 1 ]-ga yogoreteiru ] wearing-is suit-nom dirty-is sinsi 2 ] gentleman ?the gentleman who [the suit that he is wearing] is dirty? (19) Adjunct Islands [ NP [ CP [ CP __ 1 sinda no de] minna-ga kanasin-da] died because everyone-nom was-distressed hito 1 ] person Lit. ?a person who, because (he) died, everyone was saddened? (20) Subject Islands [ NP [ CP [ CP watakusi-ga __ 1 au koto/no]-ga I-nom meet comp-nom muzukasii]hito 1 ] difficult person Lit. ?The person whom that I se/met (him) is dificult.? The above examples that demonstrate the island insensitivity of relativization, especialy the examples with Complex NP islands, have created a controversy regarding the treatment of the syntactic derivation and the structure of RCs in Japanese. The above cases of island insensitivity have been one of the important pieces of evidence for the claim that Japanese RCs are diferent from those in English 33 in the important sense that they do not involve movement of the relative head or the relative operator, and that RCs have an IP structure as in (21a) rather than the CP structure such as (21b) (Fukui and Takano 2000; Murasugi 1991). (21) a. [ NP [ IP Quinn-ga __ 1 tabeta] ringo 1 ] Q-nom ate apple b. [ NP [ CP Op 1 [ IP Quinn-ga __ 1 tabeta]] ringo] Op Q-nom ate apple ?The apple that Quinn ate.? In the following subsections, I wil try to show thre arguments against this view, and try to support the movement and CP-analysis of Japanese RCs. 2.1. Island Sensitivity One of the strongest arguments against the movement analysis, and thus non- CP analysis of RCs comes from the apparent island insensitivity of relativization (Kuno 1973; Murasugi 1991). The claim is that if operator movement is involved in relativization, they should show island sensitivity. Because they do not show island sensitivity, operator movement is not motivated. If operator movement is not motivated, then the position that hosts the operator (CP-spec) is not also motivated. Thus, there is no strong reason to asume CP structures inside RCs. However, there are several arguments that this apparent lack of island efects does not necesarily indicate a lack of movement in Japanese relative clause formation. Sakai (1994), incorporating the so-caled major subject analysis of topicalization in Japanese of 34 Kuroda (1986a), suggests that Japanese RCs can involve movement of a null operator (Op) from the major subject position. Acording to this analysis, the derivation of RCs is something like (22a), and that of topicalization is (22b) (22) a. [ CP Op 1 [t 1 [ NP [ CP pro 1 e 2 kiteiru] yoohuku 2 ]-ga Op pro wearing-is suit-nom yogoreteiru] [sinsi 1 ]] dirty-is gentleman ?The gentleman who [the suit that he is wearing] is dirty? b. [ IP (sono) sinsi 1 -ga [ NP [ CP pro 1 e 2 that gentleman-nom pro kiteiru][yoohuku 2 ]]-ga yogoreteiru] wearing-is suit-nom dirty-is ?(that) gentleman is such that the suit that he is wearing is dirty? The upshot of this analysis is the following. In Japanese, major subjects that can be generated outside of RCs are always available (Kuroda 1986a; Watanabe 2003 among others). Thus, if the relative operator is generated as a major subject, and moves from major subject position to the CP-Spec position, the operator movement can circumvent Complex NP Constraint violations because it is a movement that originates outside of the RC (Hoshi 1995, 2004; Sakai 1994). In the same vein, the fact that relativization can escape other types of islands is also explained by major subject analyses. 35 A specific prediction of this major subject analysis is that if the movement from major subject position is somehow not available, then island efects should emerge. Hoshi (2004) points out that the following example in (23a) is one such case. (23a) is derived from (23b). In (23a), the major subject position is occupied by an NP, sono sinsi-ga ?that gentleman-nom?, and the outermost relative head coresponds to the indirect object of the verb okutta ?gave? in the RC. In this example, thus, the relative head is extracted out of an RC and the example creates severe unaceptability. (23) a. *[ NP [ CP sono sinsi 1 -ga kinoo [ NP [ CP pro 1 that gentleman-nom yesterday pro itinen-mae-ni t 2 t 3 okutta]yubiwa 2 ]-ga a-year-ago gave ring-nom nusumareta] okusan 3 ] was-stolen wife ?The wife that that gentleman is such that yesterday the ring which he gave her a year ago was stolen.? b. sono sinsi 1 -ga kinoo [ NP [ CP pro 1 that gentleman-nom yesterday pro itinen-mae-ni t 2 okusan-ni okutta] yubiwa]-ga a-year-ago wife-dat gave ring-nom nusumareta. was-stolen 36 ?That gentleman is such that yesterday the ring which he gave his wife a year ago was stolen.? The same point can be made using another type of island constraint, such as the Coordinate Structure Constraint (Ross 1967) in Japanese (Kato 2005a, 2005b among many others). First of al, Japanese VP-coordination constructions (Takano 2004) do not alow a major subject in the configuration ilustrated in (24). (24) [ IP Auster kyooju-ga [ VP [ VP Quinn-o sikari] A-prof.-nom Q-acc scold [ VP Stillman-o home]]-ta] S-acc praise-past ?Prof. Auster scolded Quinn and praised Stilman.? a. *Sono gakusee 1 -ga, [ IP Auster kyooju-ga that student-nom A-prof.-nom [ VP [ VP Quinn-o sikari][ VP pro 1 home]]-ta]. Q-Acc scold pro praise -past ?That student is such that Prof. Auster scolded Quin and praised him.? b. *Sono gakusee 1 -ga, [ IP Auster kyooju-ga that student-nom A-prof.-nom [ VP [ VP pro 1 sikari][ VP Stillman-o home]]-ta] pro scold S-acc praise-past ?That student is such that Prof. Auster scolded him and praised Stilman.? 37 Furthermore, Japanese VP-coordination constructions are islands as in (25). The Examples in (25) are cases of scrambling (Kato 2005a). Here, I just cite examples of scrambling, but basicaly any type of extraction is sensitive to the Coordinate Structure Constraint. (25) a. *Stillman 1 -o [ IP Auster kyooju-ga [ VP [ VP Quinn-o S-acc A-prof.-nom Q-acc sikari][ VP t 1 home]]-ta] scold praise-past ?Prof. Auster scolded Quinn and praised Stilman.? b. *Quinn 1 -o [ IP Auster kyooju-ga [ VP [ VP t 1 sikari] Q-acc A-prof.-nom scold [ VP Stillman-o home]]-ta] S-acc praise-past ?Prof. Auster scolded Quinn and praised Stilman.? Whatever the explanation of major subject formation might be 2 (Kuno 1973; Kuroda 1986a, 1986b; Mikami 1960; Saito 1982, 1983; Takezawa 1987 among many others), given the unavailability of major subjects and the islandhood of VP- coordination, the straightforward prediction of the major subject analysis of relativization is that Japanese relativization is constrained by the Coordinate Structure 2 For our purpose, just showing the island sensitivity of relativization sufices, so I leave the explanation of major subject formation open. For details, se the references cited. 38 Constraint. This prediction sems to be correct. The following examples show that Japanese relativization is constrained by the Coordinate Structure Constraint. (26) a. *[ NP [ CP Auster kyooju-ga [ VP [ VP Quinn-o sikari] A-prof.-nom Q-acc scold [ VP t 1 home]]-ta] gakusee 1 ] praise-past student Lit. ?The student that Prof. Auster scolded Quinn and praised.? b. *[ NP [ CP Auster kyooju-ga [ VP [ VP t 1 sikari] A-prof.-nom scold [ VP Stillman-o home]]-ta] gakusee 1 ] S-acc praise-past student Lit. ?The student that Prof. Auster scolded and praised Stilman.? These examples suggest that when the major subject becomes unavailable, the relative operator has to move from its thematic position, and thus it creates island efects if the original position is embedded in an island. This, in turn, suggests that there are always two possible derivations for Japanese RCs. One involves the movement of relative operator from the major subject position, and the other involves the movement from its thematic position (Hoshi 2004). Taken together, the data we have sen so far strongly suggest the following two points. First, it sems that there are always two types of possible derivations for Japanese RCs: relative head/operator movement from major subject position; and that from its original position. If one of these derivations becomes unavailable for some 39 reason (e.g., island violation) the other derivation becomes the only possible derivation. Second, there is indeed movement of some sort, and thus, apparent island insensitivity of Japanese RCs is not a strong argument against movement analyses. 2.2. Connectivity/Reconstruction Efects In the previous subsection, we have established that Japanese relativization involves some type of movement (it can either be head raising (Bianchi 2000; Hoshi 2004; Kayne 1994) or null operator movement (Browning 1987; Chomsky 1977; Ishii 1991; Sakai 1994; Watanabe 1992). In this subsection, I would like to point out an additional argument for the movement analysis of Japanese RCs, namely connectivity/reconstruction efects. Ishii (1991) demonstrates that Japanese RCs show connectivity efects when the relative head contains a local anaphor such as kare-zisin ?himself? or kanojo-zisin ?herself?. A relevant example is shown in (27). In this example, the local anaphor that is embedded in the relative head is bound by the subject in the RC. As long as these anaphors in Japanese are to be c-commanded by their antecedents (Katada 1991), this example suggests that the relative head is originated in the RC where the subject can c-command the anaphor. 3 3 Examples of binding reconstruction efects have ben used to suport the head-raising analysis of RCs (Kayne 194). It is problematic to the operator movement analysis because it is normaly asumed that empty operator does not show reconstruction efects (Epstein 1989). 40 (27) [ NP [ CP Quinn 1 -ga t 2 taipu-sita][kare-zisin 1 -no Q-nom type-do himself-gen ronbun] 2 ] paper ?the paper of himself that Quinn typed? Ishii (1991) also observes an apparent counter example to the possibility of connectivity efects within Japanese RCs. He cites the example in (28). In this example, the relative head NP is extracted out of a complex NP, and the coreferential reading of Stilman and the local anaphor is not available. (28) ? *[ NP [ CP Virginia-ga [ NP [ CP Stillman 1 -ga e 3 e 2 miseta V-nom S-nom showed koto-ga aru] hito 3 ]-o sitteiru] [kare-zisin 1 -no fact-nom exist person-acc know himself-gen syasin] 2 ] picture Lit. ?The picture of himself which Virgnina knows the person to whom Stilman has once showed.? Given the fact that the relative head can escape complex NP islands, this example is not expected to be bad. However, Hoshi (2004) points out that this example is not truly problematic. Remember that there are two derivations available for Japanese RCs. One of them involves movement from the major subject position and the 41 starting point of the movement of relative head or relative operator is outside of the complex NP island. Acording to Hoshi, a major subject construction with a local anaphor in the same configuration is indeed unaceptable, as the following example shows. Thus, he concludes that the apparent problematic example in (28) is derived from an underlying structure that involves a major subject. (29) *[kare-zisin 1 -no syasin]-ga [ IP Virginia-ga himself-gen picture-nom H-nom [ NP [ CP Stillman 1 -ga t 3 t 2 miseta koto-ga aru] S-nom show fact-nom exist hito]-o sitteiru.] person-acc know Lit. ?The picture of himself is such that Virginia knows the person to which Stilman showed it.? What I wil argue in the section regarding adjunct clauses in Japanese is that they are not islands for movement. Together with Hoshi?s theory of RCs, the non- islandhood of Japanese adjunct clauses makes a specific prediction, namely that relativization out of adjunct clauses should show binding connectivity efects. This is because, if the relative head or relative operator moves from a non-island domain, movement from the non-major subject position becomes posible. Looking at an example like (30), the prediction sems to be correct. In this example, the coreferential interpretation betwen Stilman and kare-zisin ?himself? in the relative head position is possible without any trouble. 42 (30) [ NP [ CP minna-ga [ Cond Stillman 1 -ga t 2 sutete-simatta-ra] everyone-nom S-nom trash-end-up-with-cond komatte-simau] [kare-jisin 1 -no ronbun] 2 ] have-trouble-end-up-with himself-gen paper Lit. ?The paper of himself that wil make a trouble to everyone if Stilman trash it ? 2.3. Sumary The discussion so far established two important claims about Japanese relativization. First, it involves some type of movement. Specificaly, there are two starting points for the relative head or relative operator: one is from the major subject position; and the other is from the thematic position inside the relative clause. Second, if the derivation involving a major subject is not available, the relative operator has to move out of its original, thematic position inside the relative clause. Thus, in this case, the relativization exhibits island efects if the underlying position of the relative head is inside an island. These two observations basicaly reject the claim that Japanese relativization does not involve movement of any kind. Thus this, in turn, supports the claim that Japanese Relative Clauses and those found in English are basicaly the same type of construction. Only the diference, besides their basic word orders, betwen those two is the availability of a major subject. 43 2.4. Isues of Verbal Morphology within Relative Clauses The third property of Japanese RCs that I would like to discuss concerns the morphology of embedded verbs. Folowing Hiraiwa?s series of studies on the so- caled Nominative-Genitive Conversion, I wil show that embedded verbs in Japanese RCs have special verbal morphology, the so-caled predicate adnominal form. I wil further show that detailed examination of this predicate adnominal form wil provides us with an important clue for the internal phrase structure of RCs in Japanese, namely the existence of CP-layers. 2.4.1. Nominative Genitive Conversion It is wel-known that Japanese nominative Case is optionaly converted to Genitive Case in specific environments, the phenomenon caled Nominative-Genitive Conversion (Fukui 1995; Harada 1971, 1976; Hiraiwa 2000, 2001; Miyagawa 1989, 1993; Ochi 2001; Saito 1982; Sakai 1994; Shibatani 1978; Watanabe 1994, 1996). Nominative-Genitive Conversion is typicaly observed in RCs. In (31), for example, the nominative case on the subject in an RC is converted to genitive case. (31) a. Kinoo Quinn-ga katta hon yesterday Q-nom bought book ?The book that Quinn bought yesterday? b. Kinoo Quinn-no katta hon yesterday Q-gen bought book 44 Traditionaly, it has been argued that genitive case on the subject is licensed in a clause that is headed by a nominal element such as RCs, complement clauses of nouns, or clauses headed by nominalizing complementizers. Backed up by the fact that the genitive case is typicaly licensed by nominals, researchers have argued that the genitive case in Nominative-Genitive Conversion is also licensed by these nominal heads (Fukui 1995; Harada 1971, 1976; Miyagawa 1989, 1993; Ochi 2001; Saito 1982; Sakai 1994; Shibatani 1978). However, recently, problems with this view have been pointed out. Watanabe (1994, 1996) and Hiraiwa (2000, 2001) show that genitive subjects can be licensed in environments with no nominal heads. They cite examples like (32). (32) was first pointed out by Watanabe. (32) contains examples of comparative constructions. Genitive subjects are licensed in comparative constructions without problems, even though they do not utilize any nominal heads in order to mark embedded clauses. This point can be shown by the fact that the marker of the embedded clause yori ?than? does not take the genitive form of the pronoun sono but they select the full DP form sore as summarized in (34). This suggests that yori is not a nominal element. Hiraiwa cites the example in (33) containing an until- clause. Until-clauses also do not contain nominal heads, but license genitive subjects. Hiraiwa cites six more types of examples and confirms that genitive subjects can indeed be licensed in non-nominal environments. 45 (32) a. Quinn-wa [Virginia-ga yonda yori] takusan-no Q-top V-nom read than many-gen hon-o yonda. book-acc read ?Quinn read more books than Virginia did.? b. Quinn-wa [Virginia-no yonda yori] takusan-no Q-top V-gen read than many-gen hon-o yonda. book-acc read (33) a. Quinn-wa [ame-ga yamu made] office-ni ita. Q-top rain-nom stop until office-at be-past ?Quinn was at his office until the rain stopped.? b. Quinn-wa [ame-no yamu made] office-ni ita. Q-top rain-gen stop until office-at be-past (34) a. *sono yori/ *sono made it-gen than it-gen until b. sore yori/ sore made it-gen than it-gen until Hiraiwa concludes that the crucial factor in the licensing of genitive subjects is not a nominal head, rather it is the special verbal morphology that licenses genitive subjects. Acording to Hiraiwa al of the environments in which genitive subjects are legitimate, including RCs, involve the same morphology on the embedded verb, i.e., 46 the so-caled predicate adnominal form. Thus, he concludes that genitive subjects are licensed by a predicate that employs the predicate adnominal form. A note is in order here. In the verbal morphological paradigm in modern Japanese, it is dificult to distinguish the predicate adnominal form and the matrix sentence final form, the so-caled end form, because both of them have the same form, i.e., verbs without overt complementizers. However, the so-caled verbal adjectives and copula in Japanese preserve an explicit distinction betwen the adnominal form and the end form. Thus, we can test whether the environments that license genitive subjects contain the predicate adnominal form or not by using verbal adjectives or copulas. The end form of verbal adjectives is -da and the adnominal form is -na respectively. Thus, we expect that in genitive subject environments, these predicates should bear the -na form rather than the -da form. As we can se in (35), in al the environments where the genitive subject is licensed, these elements inflect with the adnominal form. (35) a. Relative Clauses Quinn-ga suki-na/*-da ongaku Q-nom like-AdNom/*End music ?The music that Quinn likes? b. Comparatives Quinn-no koto-ga simpai-na/*-da yori mo Q-gen thing-nom worried-AdNom/*End than Virginia-ga simpai-da. V-nom worried-be-End 47 ?I am worried about Quinn than about Virginia? c. Made Clause Quinn-wa ijou-na/-*da mad-ni Q-top extraordinary-AdNom/*-End extent-to sinkeisitu da. nervous be-End ?Quinn was extraordinarily nervous? Because the adnominal form is the legitimate form in these environments, we can conclude that the verbs in these environments also bear the adnominal form, and thus, that the genitive subject is licensed by the adnominal morphology of the embedded verb. 2.4.2. The Predicate Adnominal Form and the C-system One of the wel-known constraints on Nominative-Genitive Conversion is the so-caled complementizer blocking efect. Simply put, if the embedded clause is headed by an overt complementizer, the genitive subject is not licensed. Let us se some examples from Hiraiwa (2000). (36) a. [ NP [ RC syoorai daijisin-ga okiru] in-the-future great-earthquate-nom occur kanousei] possibility ?The possibility that a great earthquake wil occur in the future.? 48 b. [ NP [ RC syoorai daijisin-no okiru] in-the-future great-earthquate-gen occur kanousei] possibility ?The possibility that a great earthquake wil occur in the future.? c. [ NP [ CP syoorai daijisin-ga okiru in-the-future great-earthquate-nom occur toiu]kanousei] comp possibility ?The possibility that a great earthquake wil occur in the future.? d. *[ NP [ CP syoorai daijisin-no okiru in-the-future great-earthquate-gen occur toiu] kanousei] comp possibility ?The possibility that a great earthquake wil occur in the future.? The important point of these examples is that, although al of the four sentences have basicaly the same meaning, only the example hosting the overt complementizer disalows the genitive subject. This paradigm suggests that adnominal predicate formation is blocked by the presence of an overt complementizer, and thus adnominal formation has some relation to the complementizer in the embedded clause. As a support for this position, we can cite examples from verbal adjectives again. In the example (37), if the complementizer is present, the verbal adjectives cannot bear the adnominal form, i.e., the -na ending. 49 (37) a. [Quinn-ga gengogaku-ga kirai-na] kanousei Q-nom linguistics-nom hate-AdNom possibility ?The possibility that Quinn does not like linguistics? b. *[Quinn-ga gengogaku-ga kirai-na toiu] Q-nom linguistics-nom hate-AdNom comp kanousei possibility ?The possibility that Quinn does not like linguistics? c. [Quinn-ga gengogaku-ga kirai-da toiu] T-nom linguistics-nom hate-end comp kanousei possibility ?The reason that Quinn does not like linguistics? The discussion so far suggests that there is a tight connection betwen the complementizer and the licensing of the adnominal forms of the embedded predicates. The simplest way to capture the above paradigm regarding the adnominal predicate formation is to asume that there is an empty complementizer position that stands in an agrement-like relation with the embedded verb. Otherwise, it is dificult to explain why the presence of the overt complementizer blocks the adnominal predicate formation. Hiraiwa indeed provides this line of argument. Hiraiwa?s theory of verbal morphology asumes that an adnominal predicate is formed by the amalgamation of v, T and C though the operation Agre (Chomsky 2001 among others), and his theory can give a straightforward explanation for the above paradigm. 50 For our purposes, however, it is more important to show that predicate adnominal form is licensed by a specific type of C, because if it is true, it suggests that Japanese RCs contain a C-layer. 2.4.3. Relative Clauses and the C-system Remembering that RCs are one of the typical environments where nominative-genitive conversion is alowed, it shal now be clear that RCs be the environment where the predicate adnominal form is also alowed. The conclusion we reached in the previous subsection is that the adnominal form is licensed by an empty complementizer. Thus, we can draw the same conclusion for the phrase structure of RCs, i.e., the embedded clause in an RC is headed by the empty complementizer. 2.5. Conclusion In this section, I have reviewed thre arguments for the movement and CP analysis of Japanese RCs. Based on the observations on island sensitivity and connectivity efects, we have concluded that Japanese relativization involves movement of a relative head or relative operator. We have also sen that embedded verbs in Japanese RCs have a predicate adnominal form. Based on the facts about nominative-genitive conversion and complementizer blocking efects on adnominal predicate formation, we concluded that embedded clauses in Japanese RCs are headed by an empty complementizer. Taken together, Japanese RCs, like English RCs, involve movement of a relative operator and a CP structure. The structure of Japanese RCs is thus something like the folowing. 51 (38) [ NP [ CP Op 1 ? t 1 ? V- AdNom ] NP 1 ] 3. The Syntax of Conditional Clauses in Japanese 3.1. The non-islandhood of Japanese Conditionals We wil start our discussion from the curious fact that Japanese conditional clauses do not show strong island efects. The examples in (39) are the relevant cases. Unlike English overt movement, scrambling out of conditional clauses is not degraded. This point can be made clear by comparing examples like (39b) and examples of long-scrambling out of relative clauses or complement clauses. (39) a. Virginia-wa [mosi Quinn-ga sono cake-o tabe- V-top mosi Q-nom that cake-acc eat- ta-ra] naki-dasu daroo. past-cond cry-start will ?Virginia wil start crying if Quinn eats that cake.? b. Sono cake-o 1 Virginia-wa [mosi Quinn-ga t 1 that cake-acc H-top mosi T-nom tabe-ta-ra] naki-dasu daroo. eat-past-cond cry-start will ?Virginia wil start crying if Quinn eats that cake.? As it wil be discussed in later sections, Japanese conditional clauses show the signature properties of so-caled Weak Islands. However, as it wil become clearer, 52 the weak island efects are induced by the Conditional Adverb (CA) mosi. This suggests that conditional clauses per se do not induce island efects. The aim of this subsection is to evaluate one of the possible approaches to the problem of non-islandhood of conditional clauses. The claim that we wil evaluate is that conditional clauses in Japanese are more like arguments than adjuncts. The intuition behind this claim is the following. Overt extraction out of conditional clauses is alowed because conditional clauses are complement clauses, which do not induce island efects under normal circumstances (Chomsky 1981, 1986b; Huang 1982; Ross 1967). The type of analysis that certain adjunct clauses are more like complement clauses has been suggested in previous literature. To the best of my knowledge, it was Mihara (1994) who first acknowledged that adjunct clauses in Japanese alow overt extraction, and he actualy hints at the possibility that adjunct clauses that alow extraction are more like complement clauses. His examples come from scrambling. He shows that long scrambling out of because-clauses in Japanese is alowed. (40) a. Quinn-wa [Stillman-ga bungakubu-ni Q-nom S-nom literature-department-to nyuugaku-sita-node] odoroi-ta enter-did-because get-surprised-past ?Quinn got surprised because Stilman entered the department of literature? 53 b. bungakubu-ni Quinn-wa [Stillman-ga literature-department-to Q-nom S-nom nyuugaku-sita-node] odoroi-ta enter-did-because get-surprised-past ?Quinn got surprised because Stilman entered the department of literature (rather than some other department)? Mihara suggests that the notion of ?degre of subordination? in traditional Japanese gramar (Kuno 1973; Masuoka and Takubo 1992 among others) might be relevant to the adjunct clauses? generosity with regard to extraction. His conclusion is basicaly that adjunct clauses in Japanese are not islands. However, he suggests a correlation betwen extractability and the degre of subordination, i.e., if the degre of subordination is lower, the extraction would be worse. Because in traditional Japanese gramar, however, the term ?degre of subordination? has not been wel defined, we cannot argue anything definitely. Informaly, we can understand degre of subordination as coresponding to the extent to which a clause behaves like a matrix/independent clause or a complement clause. Thus, under this understanding of the term, we can interpret what Mihara suggests as that adjunct clauses that alow extraction are more like complement clauses. In this section, however, I wil show that conditional clauses should be analyzed as adjunct clauses. Thus, contrary to the widely acepted view that clausal adjuncts do not alow overt extraction cross-linguisticaly (Huang 1982; Saito 1985; Stepanov 2001), I wil argue that Japanese conditionals alow the extraction even though they are adjuncts. 54 3.1.1. Adjuncthood of Conditional Clauses: English Basicaly the argument that conditional clauses in English are adjuncts is backed up by four observations. First, there sem to be no verbs that select conditional clauses as their complements. This point is related to the optionality of conditional clauses. Second, unlike arguments, the positioning of conditionals in a sentence is relatively fre. Third, movement of conditional clauses is severely constrained by wh-islands. Fourth, conditionals behave like other adjuncts with respect to constituency tests. I wil review these points one by one. The first indication of the adjuncthood of conditional clauses comes from the fact that they are optional elements and there sems to be no predicate that takes conditionals as its complement. The fact that in the following examples the presence or absence of conditional clauses does not have an impact on aceptability supports this point. These examples also show the adjuncthood of conditional clauses because verbs like leave that do not take complement clause can cooccur with conditional clauses. (41) a. Quinn wil leave [if you get angry]. b. Quinn wil leave. Another major argument for the adjuncthood of conditional clauses is based on the fact that their positioning in a sentence is relatively fre. A conditional clause can appear both sentence-initialy and sentence-finaly without changing the basic interpretation of the sentence. The circumstance is diferent if an argument NP is 55 moved from its original position. For example, if a direct object NP is fronted to the beginning of a sentence, it obligatorily receives a topicalized interpretation. The sentence initial conditionals, however, does not necesarily receive a topicalized interpretation. Thus, this fredom of the positioning in a sentence distinguishes conditional clauses from arguments. (42) a. Quinn wil leave [if you get angry]. b. [If you get angry] Quinn wil leave. (43) a. Quinn wil eat that cake. b. That cake, Quinn wil eat. One of the most powerful arguments for the adjuncthood of conditionals in English comes from extraction from WIs such as wh-islands. It is wel known that argument extraction from wh-islands is more or les alowed but that adjunct extraction creates severe degradation. Collins (1998) shows that English conditionals cannot be extracted out of wh-islands, taking advantage of the fact that they can be clefted. In the examples in (44) an if-conditional is clefted. (44a) shows that the clefting of if-clauses is unbounded, i.e., the if-clause can be extracted out of a complement clause. On the other hand, in (44b) the if-clause is extracted out of a wh- island, and the example is not aceptable. Acording to Collins, the severity of the violation in (44b) can be analyzed as an ECP violation. Given that conditional clauses are obviously not subjects, this extraction patern strongly suggests that they are adjunct clauses. 56 (44) It is if the student fails that the teacher wil fire the TA. a. ? It is if the student fails that the Stilman said that the teacher would fire the TA. b. *It is if the student fails that Stilman wonders why the teacher wil fire the TA. Finaly, let us take a look at an argument based on clasical constituency tests. Constituency tests such as VP-deletion (Jayaselan 1990; Sag 1976) and do so substitution (Lakoff and Ros 1976) show clearly that conditional clauses in sentence final position behave like adjuncts (Bhat and Pancheva 2001). Bhat and Pancheva cite the following examples. In both of the examples in (45), the conditional clauses are interpreted in the elipsis site or do so anaphora. These examples suggest that they are atached to the VP, and thus can be inside the scope of deletion or do so substitution. Together with these examples, the ones in (46) suggest that conditionals are adjuncts rather than arguments, where if-clauses are stranded outside of VP- deletion or do so anaphora. (45) a. I wil leave if you do and Quinn wil leave if you do , too. b. I wil leave if you do and Quinn wil do so too. (46) a. I wil leave if you do and Quinn wil leave if Virginia does. b. I wil leave if you do and Quinn wil do so if Virginia does. 57 So far, I have reviewed a set of argument for the adjuncthood of conditionals in English. One might argue that this isue is too obvious and trivial to spend a lot of space. However, laying out these arguments is crucialy necesary especialy for the analysis of conditionals in Japanese where a systematic argument for their constituency has not been offered in the literature. In the following section, I wil apply some of the constituency tests reviewed above in order to show that conditional clauses in Japanese show adjuncthood in the same way as English conditionals. 58 3.1.2. Japanese Conditionals as Adjunct Clauses Even in Japanese, there sems to be no verb that selects conditional clauses as its complement. Thus, we can basicaly conclude that they are adjuncts rather than arguments. However, syntacticaly showing their adjuncthood is not so easy in Japanese. When we try to test the constituency of sentences in Japanese, we always face the problem of whether Japanese has the same type of constituency tests that we can se in English. For example, we realize that the first two tests in the previous section, optionality on the one hand, and fredom of word order on the other, are not so informative, given that Japanese alows empty elements that can refer to both arguments and adjuncts (Kuno 1973; Kuroda 1965) and given that Japanese alows fre word order. Furthermore, movement of non-wh-phrases is not constrained by wh-islands in Japanese (Boskovic and Takahashi 1998; Saito 1985; Takahashi 1993; Watanabe 1992 among others). Therefore, making use of wh-islands is also not so straightforward. Facing these problems, I can pick only one test from the four that we have sen above for the adjuncthood of conditional clauses, namely VP-constituency tests. Like English, Japanese has a variety of constituency test that can tap on the structure of VP. There are two of them that have been frequently mentioned in the literature. One is VP-fronting (Hoji et al. 1989; Tateishi 1994), and the other is soo su ?do so? substitution (Hinds 1973a, 1973b; Inoue 1976; Nakau 1973; Tateishi 1994). The utility of these two syntactic operations is that both of them are sensitive to the diferences betwen arguments and adjuncts in the same way as English constituency 59 tests. For this reason, by making use of these tests, we can examine the adjuncthood of conditional clauses in Japanese. Let us start our discussion from VP-fronting. Japanese VP-fronting has some prominent gramatical properties. It is alowed if the topic marker -wa, the emphatic marker -mo, or any of the contrastive markers are atached to the fronted VP, and if an auxiliary verb is inserted to support the stranded tense morpheme (Hoji et al. 1989). Even though Japanese makes use of some explicit morphological markers, these properties also hold true in English. The first property is related to the fact that VP-fronting is a phenomena asociated with Topic or Focus (Rochemont 1986) and the second property can be understood as the Japanese counterpart of the constraint on stranded afixes (Lasnik 1981). For example, we can observe the following contrast (Saito 1985; Tateishi 1994). (47) a. Quinn-ga sono hon-o kat-ta koto. Q-nom that book-acc buy-past fact ?The fact that Quinn bought that book.? b. *[ VP Sono hon-o kau] 1 Quinn-ga t 1 ta koto. c. [ VP Sono hon-o kai] 1 -wa Quinn-ga t 1 si-ta koto. (47b) is excluded because there is no contrastive marker on the fronted VP and the tense morpheme is stranded alone. If these two conditions are satisfied, the example becomes aceptable as in (47c). These properties, therefore, alow us to conclude that Japanese VP-fronting is basicaly the same type of operation as in English. 60 VP-fronting in Japanese has various other properties, but for our purposes, the following is the most relevant: if a VP is fronted, the complement must be fronted. Compared to (47c), the example in (48) is not at al aceptable. (48) *[ VP kai] 1 -wa Quinn-ga sono hon-o t 1 si-ta koto. Given this property, it follows that if conditional clauses are arguments, they must be fronted together with the VP. However, as the pair of examples in (49) suggests that it is not the case. The stranding of conditionals does not make the sentence unaceptable. (49) Quinn-wa [(mosi) Stillman-ga kita-ra] nigedasu Q-top mosi S-nom come-cond run-away daroo. will ?Quinn wil run away if Stilman comes.? a. [ VP [(mosi) Stillman-ga kita-ra] nigedasi] 1 -sae mosi S-nom come-cond run-away -even Quinn-wa t 1 suru daroo Q-nom do will ?Quinn would even run away if Stilman comes.? 61 b. ? [ VP nigedasi] 1 -sae Quinn-wa [(mosi) Stillman-ga run-away-even T-top mosi H-nom kita-ra] t 1 suru daroo. come-cond do will For some speakers, (49b) does not sound good. However, this sems to be because of the length and complexity of the conditional clause itself. As an indication of this, if we replace the arguments inside conditionals to empty arguments and omit mosi, the sentence becomes much more aceptable even to these speakers. In the example in (50), a context is inserted before the sentence to support the empty arguments, and this sentence is aceptable to those speakers. (50) Quinn and Virginia are discussing what wil happen to Stilman if he eats a realy spicy hot pepper. Quinn: Naki-sura Stillman 1 -wa [pro 1 pro (hot peper) cry-even J-top tabeta-ra] suru daroo. eat-cond do will ?Stilman wil even cry if he eats the hot pepper.? Another constituency test, the soo su substitution test (Shibatani 1973; Terada 1990 among others), provides a support for the argument above. Just like do so in English, Japanese soo su has been used for testing VP-constituency. Although the status of soo su, whether it is a deep anaphor or a surface anaphor (Hoji 1987), is 62 controversial, the important generalization that has been acepted in the previous studies is that soo su replaces a VP. In other words, when soo su is used, the complement must be replaced together with the verb. The examples in (51) ilustrate this point clearly. (51) Quinn-wa ringo-o tabe-ta. Q-top apple-acc eat-past ?Quinn ate an apple.? a. Stillman-mo soo-si-ta. S-also so-do-past. ?Stilman did so too.? b. *Stillman-mo remon-o soo-si-ta. S-also lemon-acc so-did-past ?Stilman did so a lemon.? Following the same logic as above, we expect that if conditionals are arguments, they must be replaced by soo su together with the other elements in the VP. However, they can be stranded out of the VP. (52) Quinn-wa [mosi Virginia-ga kaetta-ra] kaeru Q-top mosi V-nom go-home-cond go-home deshoo. will ?Quinn wil go home if Virginia goes home.? 63 a. Stillman-mo soo-suru deshoo. S-also so-do will ?Stilman wil do so too.? b. Stillman-mo [mosi Auster-ga kaetta-ra] J-also mosi A-nom go-home-cond soo-suru deshoo. so-do will ?Stilman wil do so if Auster goes home too.? In exactly the same way as the English examples, they strongly indicate that conditional clauses are adjuncts rather than arguments. The above observations tel us some important properties of conditionals in Japanese. One is that Japanese conditional clauses are indeed adjuncts rather than arguments. Otherwise, we cannot capture the diferent behaviors of conditional clauses and arguments with respect to VP-constituency tests. The other is that sentence medial conditional clauses, at least, are generated in a VP-adjoined position. The fact that the conditional can move together with VP, and replaced by soo su together with VP suggests that they are a part of VP. Thus, we can conclude that conditional clauses in Japanese are base generated in a VP-adjoined position. 3.1.3. Sumary In this section, we have reviewed one potential approach to the non- islandhood of conditional clauses, namely that conditionals are complement clauses. Through VP-constituency tests in which conditionals show systematic diferences 64 from complement clauses, we have reached the conclusion that they should be analyzed as adjunct clauses rather than complement clauses. 3.2. Conditional Clauses as Weak Islands The aim of this section is to show that Conditional Clauses in Japanese exhibit signature properties of so-caled Weak Islands (henceforth WIs) (Cinque 1990 among many others). Specificaly, my claim is the following. The Conditional Adverb (CA) mosi is the inducer of the weak-islandhood of conditional clauses and conditional clauses per se are not islands. This means that if mosi is present in a conditional clause, it becomes a WI. On the other hand, if mosi is not present, the conditional clause does not show WI efects. To establish this claim, I wil apply various tests of WIs to conditional clauses. 3.2.1. Diagnosing Weak Islands Referentiality Although there are many properties of WIs that have been reported in the literature, there is one commonly acknowledged feature, namely that extraction out of WIs is dependent on the referentiality of wh-phrases (Cinque 1990; Rizi 1990 among many others). The basic patern in the data is that if a wh-phrase is referential, it is easier to move it out of WIs. On the other hand, if a wh-phrase is not referential, extraction out of WIs becomes worse. Let us se this patern by looking at some examples from English. 65 Rizi, following Cinque (1984), proposes that only elements asigned referential theta roles can be extracted from a WI, and everything else cannot. A referential NP is understood as an NP that refers to specific members of a prestablished set. Based on this notion of referentiality, Rizi draws a distinction betwen arguments with referential theta-roles such as agent, theme, patient, experiencer and so on, and the other elements. The necesity of this distinction can be clearly sen in the following pair of examples. In both of the examples the verb weigh takes a complement NP, apples and 200 lbs respectively. Acording to Rizi, because they are complement NPs, they should not be structuraly diferent, i.e., they are both sisters of the verb weigh, and thus presumably asigned theta roles. (53) a. Quinn weighed apples. a. Quinn weighed 200 lbs. However, as far as movement is concerned, these complements show a clear diference. The question in (54a) is ambiguous but if a wh-phrase is extracted from a wh-island as in (54b), only the agentive reading becomes available. Therefore, the question in (54b) can be properly answered apples, but not 200 lbs. (54) a. What did Quinn weigh t? b. ?What did Quinn wonder how to weigh t? 66 Rizi?s point is that this diference cannot be captured by a theory that simply incorporates a complement/non-complement asymetry (Chomsky 1986b; Huang 1982) because they are both complements. On the other hand, given the referential/non-referential distinction, this diference can be easily captured. NPs such as apples receive referential theta role acording to Rizi. This is so because it is an element that refers to participants in the event described by the verb. 200 lbs, on the other hand, is an expresion that qualifies something, not something that participates in the event. I wil basicaly follow Rizi?s proposal here, and test the weak-island status of Japanese Conditional clauses refering to the notion of referentiality. Constructions Sensitive to Weak Islands We have briefly sen that referentiality is a crucial property of deciding whether an element can be extracted out of WIs or not. In this subsection, let us se more concretely which constructions exhibit WI efects and which do not. First, as we have already sen, arguments and adjuncts show diferences. Normaly arguments can be extracted from WIs but adjuncts cannot (Chomsky 1986; Huang 1982; Lasnik and Saito 1992). (55) a. ?Which man are you wondering whether to invite t? b. *How are you wondering whether to behave t? (Szabolsci & den Dikken 1999) 67 It has been observed however that among the types of adjuncts, why and how show strong WI sensitivity, when has an intermediate status, but adjuncts such as where do not (Chomsky 1986b; Huang 1982; Lasnik and Saito 1992; Szabolcsi and den Dikken 1999 among others). Szabolsci and den Dikken cite the following examples. (56) a. *Why did Quinn ask whether to do this t? b. *How did Quinn ask whether to do this t? c. ? When did Quinn ask whether to do this t? d. Where did Quinn ask whether to read this book t? These diferences among adjuncts suggest that WI-sensitivity does not depend just on whether the extracted elements are subcategorized for or not. Rather, the paterns ilustrated in (5) and (56) are compatible with Rizi?s dichotomy of extracted elements based on referentiality. Acording to Rizi, although manner or reason phrases may be arguments, they do not receive referential theta roles. Thus, in this sense they are similar to the amount phrases that we have sen in (53). An amount phrase may also be an argument but it does not receive a referential theta role. Whether the extracted elements are D-linked or not is another case that afects extractability. If the wh-phrase is D-linked, it can be extracted from WIs in contrast to Non D-linked wh-phrases (Cinque 1990; Lasnik and Saito 1992; Pesetsky 1987 among others). Szabolcsi and den Dikken cite the following minimal pair in (57). In (57b), even though the wh-phrase is an amount phrase, it is D-linked in the sense that it involves a contextualy established set of books. On the other hand, (57a) is not D- 68 linked in this sense. As the minimal pair suggests, if a non D-linked wh-phrase is extracted from a weak island, the sentence becomes unaceptable. (57) a. *How many books are you wondering whether to write t next year? b. How many books on the list are they wondering whether to publish t next year? Another support for the claim that D-linking is a significant factor for extraction comes from examples of so-caled aggresively non D-linked wh-phrases. Pesetsky points out that wh-the-hel phrases are not compatible with overt markers of D-linking such as which, and they indeed cannot be D-linked, i.e., it cannot refer to the member of a prestablished set of entities. (58) a. What the hel book did you read that in? b. *Which the hel book did you read that in? Because they cannot be D-linked, the straightforward expectation is that these wh- phrases cannot be extracted from WIs, and this inded sems to be correct. (59) a. ?Who the hel are you wondering whether to invite t? b. Which man are you wondering whether to invite t? 69 Acording to Cinque (1990), D-linking and referentiality are the same notion. Both of them crucialy refer to the notion of prestablished set, a certain presupposition. Thus, examples of the-hel question also show the referentiality asymmetry. Finaly, let us se some examples from scope reconstruction cases. It is wel known that a wh-phrase like how many books has wide or narow scope with respect to a verb like want (Cinque 1990; Kroch 1989; Rulmann 1995). This is ilustrated by the following example cited from Rullmann (1995). The wide scope reading of (60) is paraphrased in (60a), where it is asumed that there are certain number of books which Chris wants to buy and the speaker asks how many such books there are. On the other hand, the narow scope reading is paraphrased as (60b). Under this reading, it is not asumed that there is any specific set of books that Chris wants to buy, but it is asumed that Chris wants to buy a certain number of books. Whether the wh-phrase refers to the prestablished set of entities or not is the crucial notion to distinguish these two reading. (60) How many books does Chris want to buy? a. What is the number n such that there are n books that Chris wants to buy. b. What is the number n such that Chris wants it to be the case that there are n books that he buys? In the literature, it is observed that this how-many question is sensitive to WIs (Cinque 1990; Kroch 1989; Rullmann 1995). If the how-many phrase moves across a 70 WI, the narow scope reading becomes unavailable, i.e., the non-referential reading becomes unavailable. To ilustrate this point, Rulmann (1995) cites the following example. (61) How many books did no student want to buy? a. What is the number n such that there are n books that no student wants to buy? b. *What is the number n such that no student wants it to be the case that there are n books that s/he buys? The negation no is known to create WIs (Beck 1996; Ross 1984; Rullmann 1995 among others), and if how-many is extracted from under negation, the narow scope reading is no longer available anymore. Only the presuppositional, wide scope reading is available. This is understood as that the WI induced by the negation blocks the reconstruction of the how-many phrase, and as a result narow scope non- referential interpretation becomes unavailable. Let us summarize the discussion so far. We have sen that referentiality is the crucial notion for distinguishing the elements that can be extracted from WIs and those that cannot. Following this line of argument, we have sen at least four cases of extraction that are sensitive to WIs. Cinque argues that the major diference betwen WIs and Strong Islands is the following. WIs alow extraction of referential elements but disalow the extraction of non-referential elements. On the other hand Strong Islands do not alow any 71 extraction. Given this property, we can test whether a domain is a WI or not by making use of the referentiality asymmetry. In other words, if movement of a referential element is alowed but non referential element is not alowed from a certain domain, we can cal this domain a weak island. Folowing this way of thinking, we can test whether Japanese Conditionals are WI or not. 3.2.2. Mosi as a Weak Island Inducer In the discussion so far, I have shown cases of extraction that are sensitive to WIs. From now, I wil take advantage of the WI sensitivity of these elements and show that Japanese Conditionals are WIs. In the course of the discussion, I wil further show that the WI efects shown by Japanese Conditionals are induced by the CA, mosi. First let us se the simple argument-adjunct asymmetry. In (62), two sentences are compared where (62a) contains the in-situ argument wh-phrase dono-ringo ?which apple? inside the conditional clause, and (62b) contains the in-situ adjunct wh-phrase naze ?why.? In this comparison, we can se a clear argument/adjunct asymmetry. 4 The in-situ naze in a mosi conditional is totaly unaceptable, but in-situ dono-ringo is quite aceptable. 4 Note that recently, it becomes clearer that the interpretation of naze and its interaction with islands are heavily influenced by prosodic patern (Kitagawa 206 and references therein). In this study, I kep presenting the data based on the judgments from native speakers I interviewed without much consideration in the influence of prosody, just for sake of the simplicity of the argument. However, a careful examination of the examples containing wh-phrases is, inded, required. 72 (62) a. Quinn-wa mosi Stillman-ga dono-ringo-o Q-top mosi S-nom which-apple-acc tabeta-ra okorimasu-ka? eat-cond get-angry-Q Lit. ?Which apple wil Quinn get angry if Stilman eats?? b. *Quinn-wa mosi Stillman-ga naze ringo-o Q-top mosi S-nom why apple-acc tabeta-ra okorimasu-ka? eat-cond get-angry-Q ?Why 1 wil Quinn get angry [if Stilman eats an apple t 1 ]?? As we have sen in the earlier sections, mosi in Japanese conditionals is an optional adverb. Interestingly, if mosi is absent from conditional sentences, the extraction of both arguments and adjuncts becomes aceptable. The Examples in (63) are exactly the same as the examples in (62) except that mosi is not present in both examples. (63) a. Quinn-wa Stillman-ga dono-ringo-o tabeta-ra Q-top S-nom which-apple-acc eat-cond okorimasu-ka? get-angry-Q Lit. ?Which apple wil Quinn get angry if Stilman eats?? 73 b. ? Quinn-wa Stillman-ga naze ringo-o tabeta-ra Q-top S-nom why apple-acc eat-cond okorimasu-ka? get-angry-Q ?Why 1 wil Quinn get angry [if Stilman eats an apple t 1 ]?? For some speakers, naze in conditional clauses is stil not perfect. However, the remarkable fact is that comparing examples like (62b) and (63b), (63b) is much beter than (62b). This fact indicates that the presence or absence of CA mosi is crucial for the weak islandhood of Japanese conditional clauses. Diferently put, we can plausibly consider that mosi is an inducer of WI efects. Argument wh-phrases can also be overtly moved out of conditional clauses as the examples in (64) suggest. On the other hand, because long-scrambling of adjunct wh-phrases in Japanese is prohibited even out of complement clauses as in (65) (Nemoto 1993; Saito 1985) we cannot test whether the overt movement of naze creates the same results or not. However, the cases of long-scrambling suggest that as long as the extracted elements are referential arguments, conditional clauses are not islands. (64) a. dono-ringo-o 1 Quinn-wa [mosi Stillman-ga t 1 which-apple-acc Q-top mosi S-nom tabeta-ra] okorimasu-ka? eat-cond get-angry-Q Lit. ?Which apple wil Quinn get angry if Stilman eats?? 74 b. dono-ringo-o 1 Quinn-wa [Stillman-ga t 1 which-apple-acc Q-top S-nom tabeta-ra] okorimasu-ka? eat-cond get-angry-Q Lit. ?Which apple wil Quinn get angry if Stilman eats?? (65) *Naze 1 Quinn-wa Stillman-ni [Virginia-ga t 1 why Q-top S-dat V-nom ringo-o tabe-ta to] it-ta no? apple-acc eat-past comp say-past Q ?Why did Quinn say to Stilman that Virginia ate the apple?? Let us turn our atention to referentiality. Miyagawa (1998) points out that two types of presumably non-referential elements are sensitive to weak islands in Japanese. One of his examples involves the so-caled Floated Numeral Quantifier (FNQ). As quantity arguments in English are not referential, Numeral Quantifiers (NQ), modifiers denoting quantity, are also not referential. Acording to Miyagawa, NQ can be moved to the beginning of a sentence as in (66a). However, if the movement of NQ crosses a WI inducer, in his example it is the focus marker sika ?only? (Hagstrom 1998; Hoji 1985; Pesetsky 2000; Takahashi 1990; Tanaka 2003a among many others) the sentence becomes unaceptable. In example (66b), the NQ futa-tu moves over the WI inducer sika and the example becomes unaceptable. 75 (66) a. Futa-tu 1 Quinn-ga ringo-o t 1 tabe-ta (koto). 2-cl (thing) Q-nom apple-acc eat-past (fact) ?Two, Quinn ate apples? b. *Futa-tu 1 Quinn-sika ringo-o t 1 tabe-nakat-ta 2-cl (thing) Q-Foc apple-acc eat-not-past (koto). (fact) ?Two, only Quinn ate apples? In exactly the same way, overt movement of an indefinite quantifier shows WI sensitivity. Thus, if the indefinite quantifier dareka ?some? is moved across sika the sentence becomes unaceptable. (67) a. Virginia-ga gakusee-o dareka yon-da V-nom student-acc some invite-past (koto). (fact). ?Hanako invited some student.? b. Dareka 1 Virginia-ga gakusee-o t 1 yon-da some V-nom student-acc invite-past (koto). (fact). ?Virginia invited some student.? 76 c. *Dareka 1 Virginia-sika gakusee-o t 1 some V-foc student-acc yoba-nakat-ta (koto). invite-not-past(fact). ?Only Virginia invited some student.? Given the WI sensitivity of these extractions, our expectation is the following. If mosi conditionals are WIs, extraction of these elements should be prohibited. First let us se cases of FNQ. There is a remarkable diference in aceptability betwen a case where a quantifier moves across mosi and one that does cross mosi or one without mosi i.e., among the examples in (68), only (68b) is unaceptable, where the quantifier moves across mosi. This contrast is exactly what we expect. (68) a. Quinn-wa [(mosi) Virginia-ga ringo-o futa-tu Q-top (mosi) V-nom apple-acc 2-cl (thing) tabe-ta-ra] okoru-daroo. eat-past-cond get-angry-will ?Quinn wil get angry if Virginia eats two apples.? b. *Quinn-wa [futa-tu 1 mosi Virginia-ga ringo-o t 1 Q-nom 2-cl (thing) mosi V-nom apple-acc tabe-ta-ra] okoru-daroo. eat-past-cond get-angry-will ?Quinn wil get angry if Virginia eats two apples.? 77 c. Quinn-wa [mosi futa-tu 1 Virginia-ga ringo-o t 1 Q-nom mosi 2-cl (thing) V-nom apple-acc tabe-ta-ra] okoru-daroo. eat-past-cond get-angry-will ?Quinn wil get angry if Virginia eats two apples.? d. Quinn-wa [futa-tu 1 Virginia-ga ringo-o t 1 Q-nom 2-cl (thing) V-nom apple-acc tabe-ta-ra] okoru-daroo. eat-past-cond get-angry-will ?Quinn wil get angry if Virginia eats two apples.? The same contrast emerges in examples involving indefinite quantifiers. Like Miyagawa?s examples with sika, if the indefinite quantifier dareka moves over mosi, the example becomes unaceptable. (69) a. Quinn-wa [(mosi) Virginia-ga tomodati-o dareka T-top mosi H-nom friend-acc some party-ni yon-da-ra] naki-dasu daroo. party-dat invite-past-cond cry-start will. ?Quinn wil start crying if Virginia invites some friend to the party.? 78 b. *Quinn-wa [dareka mosi Virginia-ga tomodati-o t 1 Q-top some mosi V-nom friend-acc party-ni yon-da-ra] naki-dasu daroo. party-dat invite-past-cond cry-start will. ?Quinn wil start crying if Virginia invites some friend to the party.? c. Quinn-wa [mosi dareka Virginia-ga tomodati-o t 1 Q-top mosi some V-nom friend-acc party-ni yon-da-ra] naki-dasu daroo. party-dat invite-past-cond cry-start will. ?Quinn wil start crying if Virginia invites some friend to the party.? d. Quinn-wa [dareka Virginia-ga tomodati-o t 1 T-top some H-nom friend-acc party-ni yon-da-ra] naki-dasu daroo. party-dat invite-past-cond cry-start will. ?Quinn wil start crying if Virginia invites some friend to the party.? Taken together, the extraction of non-referential elements further supports the view that mosi conditionals in Japanese are WIs. Specificaly the contrast betwen extraction of the referential argument and NQs is the same contrast that we have sen in the English examples, i.e., referentiality maters for extraction out of mosi conditionals. Let us turn to the D-linking test. As we have sen, examples from English suggest that D-linked wh-phrases are referential and thus they can move out of WIs but non-D-linked wh-phrases are not referential and not extractable from WIs. In our 79 earlier discussion, we have sen that this diference becomes clear if we se the contrast betwen the D-linked wh-phrases and aggresively non-D-linked wh-phrases. Pesetsky (1987) suggests that the so-caled itai wh-phrases in Japanese show the same behavior as aggresively non-D-linked wh-phrases in English (se also Lasnik and Saito 1992 for related discussion). Details aside, if itai wh-phrases behave in the same way as aggresively non-D-linked wh-phrases, we can simply expect that itai wh-phrases cannot be extracted out of WIs but D-linked wh-phases can. This expectation can be tested using WI efects induced by sika. The following examples show that under the scope of sika, a D-linked wh-phrase is aceptable but both a non- D-linked wh-phrase and an itai wh-phrase are marginal at best. (70) a. Quinn-sika dono-hon-o yoma-nai-no? Q-only which-book-acc read-neg-Q ?Which book wil only Quinn read?? b. ? Quinn-sika nani-o yoma-nai-no? Q-only what-acc read-neg-Q ?What wil only Quinn read?? c. ?/ *Quinn-sika ittai nani-o yoma-nai-no? Q-only ittai what-acc read-neg-Q ?What the hel wil only Quinn read?? These examples suggest that, as in the English examples, there is a D-linked/non-D- linked asymmetry with respect to extraction from WIs. Now, let us se if this D- linked/non-D-linked asymmetry holds in the case of conditionals. If mosi induces a 80 WI efect, we expect that the extraction of D-linked wh-phrases should be alowed from mosi conditionals but that extraction of non-D-linked wh-phrases or itai wh- phrases should create degradation. Although a simple wh-phrase does not create severe degradation, this prediction sems to be confirmed. Even though the efects do not sem to me to be so strong, stil there is a diference in aceptability betwen the case of the D-linked and Non-D-linked wh-phrases in (71). 5 (72) summarizes the examples without mosi, and al the examples are aceptable. (71) a. Quinn-wa [mosi Virginia-ga dono-hon-o Q-top mosi V-nom which-book-acc yon-da-ra] bikkuri-simasu-ka? read-past-cond be-surprised-Q ?Which book wil Quinn be surprised if Virginia reads? ? b. ? Quinn-wa [mosi Virginia-ga nani-o Q-top mosi V-nom what-acc yon-da-ra] bikkuri-simasu-ka? read-past-cond be-surprised-Q ?What wil Quinn be surprised if Virginia reads?? 5 The strength of the efects is diferent among speakers. Some detected clear diferences but some do not. The author conducted aceptability judgment with six native speakers, including myself. Two of them detected clear diferences but four speakers recognized the diferences in the same direction, but for them the efect is rather mild. Further examination of the data is required. 81 c. ? Quinn-wa mosi Virginia-ga ittai nani-o Q-top mosi V-nom ittai what-acc yon-da-ra bikkuri-simasu-ka? read-past-cond be-surprised-Q ?What the hel wil Quinn be surprised if Virginia reads?? (72) a. Quinn-wa Virginia-ga dono-hon-o yon-da-ra bikkuri-simasu-ka? b. Quinn-wa Virginia-ga nani-o yon-da-ra bikkuri-simasu-ka? c. Quinn-wa Virginia-ga ittai nani-o yon-da-ra bikkuri-simasu-ka? A note is in order on the aceptability of the case of simple wh-phrases. It is true that for some speakers, a simple wh-phrase in a mosi conditional sounds bad. However, at least to my ear, the example does not sound as bad as (71c), the example containing the aggresively-non-D-linked wh-phrase. Why does this patern emerge? Here, I would like to present a potential acount. Pesetsky (1987) offers one possible answer for this question. He argues that an argument wh-phrase such as what in English can be D-linked even though it does not have the form of which. My conjecture here is that what Pesetsky argues holds true for Japanese wh-phrases also. In other words, the example (71b) sounds good because the simple wh-phrase nani can potentialy be D- linked. This conjecture leads us to the folowing prediction: if we can make simple wh-phrase non-D-linked, the example becomes worse. This prediction sems to be true. As we have sen, if we atach itai to the wh-phrases, the examples become bad 82 regardles of whether the wh-phrase is D-linked or non-D-linked. Thus, the contrast betwen (71b) and (71c), can be understood that a simple wh-phrase like nani can be D-linked. Eexample (71c) is bad because the wh-phrase is forced to be non-D-linked because of the presence of itai. To support this claim, we can consider cases of D- linked wh-phrases. Even wh-phrases with a D-linked form like dono-x ?which-x? become unaceptable if itai is atached to it under mosi conditionals. Recal also that a wh-phrse like naze ?why?, which cannot be D-linked/referential regardles of whether itai is atached or not, is bad under mosi conditionals. These contrasts can be easily acounted for under our asumption that only referential wh-phrases can be extracted out of WIs, i.e., simple wh-phrases can be D-linked and referential, but itai wh-phrases and adjunct wh-phrases cannot be D-linked and thus cannot be referential. Thus, the simple wh-phrases are extractable out of WIs, but the itai wh-phrases are not. Although there may be several possible alternatives, if we adopt Pesetsky?s idea outlined here, we can have a coherent acount for the data we have. (73) ? Quinn-wa mosi Virginia-ga ittai dono-hon-o T-top mosi H-nom ittai which-book-acc yon-da-ra bikkuri-simasu-ka? read-past-cond be-surprised-Q Finaly, let us examine how-many questions in Japanese. As in English counterparts, Japanese how many phrases show an ambiguity. Consider (74) which is ambiguous. In the reading (74a), the existential quantifier 'n-many articles' scopes above 'should'. This reading involves a presupposition: for instance, the commite 83 has 10 topics on which 10 articles are supposed to be writen. (74a) asks about the number of articles that Quinn is required to write by the commite. By contrast, the other reading in (74b), this sentence asks about the number of the articles that Quinn should write, say, by the end of this month. If we scramble the how many phrase to the beginning of the sentence, the same ambiguity is observed. (75) is an example with scrambling of the how many phrase. With this much in mind, let us turn to an example that contains a WI inducer, such as (76). (74) Sono iinkai-wa [Quinn-ga [nan-bon-no kizi-o] that committee-top[Q-nom [how many article-acc] kakubekida to] hookokusita nodesuka? should-write comp] reported Q 'How many articles did that commite report that Quinn should write?' a. For what number n: there are n-many articles x such that that commite reported that Quinn should write x (wide reading) b. (?) For what number n: that commite reported that it is necesary for there to be n many articles x such that Quinn writes x (reconstructed reading) (75) [nan-bon-no kizi-o] 1 sono iinkai-wa [how many article-aC] that committee-NOM [Quinn-ga t 1 kakubekida to] hookokusita nodesuka? [Q-NOM should write COMP] reported Q 'How many articles did that commite report that Taro should write?' 84 a. For what number n: there are n-many articles x such that that commite reported that Quinn should write x (wide reading) b. (?) For what number n: that commite reported that it is necesary for there to be n many articles x such that Quinn writes x (reconstructed reading) (76) is minimaly diferent from (75). In (76), the matrix subject is asociated with the quantificational expresion hotondo dono ?almost every.? On the other hand, in the(76), the matrix subject is not asociated with an overt quantifier. This type of quantifier is another instance that induces WI efects in Japanese (Hoji 1985; Miyagawa 1998 among others). If this type of WI inducer is inserted, the example does not show the ambiguity anymore. Thus in (76), only the presuppositional reading is available, and the example does not have the reading (76b). This efect can be understood in the same way as we have sen in the English examples, i.e., because of the WI inducer, the amount reading, a non-referential reading, becomes unavailable. (76) [nan-bon-no kizi-o] 1 hotondo dono iinkai-mo [how many article-acc] almost every committee [Quinn-ga t 1 kakubekida to] hookokusita nodesuka [Q-nom should write COMP] reported Q 'How many articles did almost every commite report that Quinn should write?' 85 a. For what number n: there are n-many articles x such that almost every commite reported that Quinn should write x. (wide) b. *For what number n: almost every commite reported that it is necesary for there to be n-many articles x such that Quinn writes x. (reconstructed) Now our question is whether the same type of disambiguation takes place if we locate the how many question in a mosi conditional. The answer sems to be positive. Let us examine the following example. A subtle judgment is required. In (77) the presuppositional reading (77a) is available. However, the quantity reading (77b) is not available. On the other hand, if mosi is not present , both of the readings are available as ilustrated in (78). (77) Nan-satu-no-hon-ni1 gakubuchoo-wa [mosi how-many-books-dat department-chair-top mosi Virginia-ga t1 shohyoo-wo kakeba] shoogakukin-ga V-nom review-acc write scholarship-nom das-eru-no? offer-CAN-Q? Lit. "How many books can the department chair offer the scholarship to Virginia if she wrote a review to?" a. For what number n: there are n-many books x such that the department chair can offer the scholarship to Virginia if she writes a review to x? 86 b. *For what number n: the department chair can offer the scholarship to Virginia if there are n-many books x such that she writes a review to x? (78) Nan-satu-no-hon-ni1 gakubuchoo-wa how-many-books-dat department-chair-top [Virginia-ga t1 shohyoo-wo kakeba] shoogakukin-ga V-nom review-acc write scholarship-nom das-eru-no? offer-CAN-Q? Lit. "How many books can the department chair offer the scholarship to Virginia if she wrote a review to?" a. For what number n: there are n-many books x such that the department chair can offer the scholarship to Virginia if she writes a review to x? b. For what number n: the department chair can offer the scholarship to Virginia if there are n-many books x such that she writes a review to x? The patern in (77) paralels the English examples that we have sen, where the presuppositional reading is not available because of the wh-island. Therefore, this suggests that mosi conditionals indeed behave like WIs. On the other hand, the contrast betwen (77) and (78) confirms that the presence of mosi is a crucial factor for the WI efects of conditionals. 87 3.2.3. Sumary Our main observations so far are the following. Mosi conditionals show the signature properties of WIs, i.e., mosi conditionals show a referential/non-referential asymmetry. Thus, referential elements are alowed to move out of mosi conditionals but non-referential elements are not. To establish this claim we have examined four diagnostic cases: argument/adjunct asymmetry; NQ movement; D-linked/non-D- linked asymmetry; and how many reconstruction. Furthermore, we have sen that these asymmetries go away if mosi is absent from conditional clauses. Based on these observations, we can draw the following conclusions. Japanese conditional clauses show various WI efects. However, these WI efects are not induced by conditional clauses per se, rather the Conditional Adverb mosi creates these efects. Thus, there are asymmetries betwen mosi conditionals and conditionals without mosi. Given this conclusion, a question arises: Why does an item like mosi create WI efects? In the following sections we wil try to answer this question. 3.3. The Distribution of Mosi In this section, I wil investigate the distributional properties of mosi. I wil establish the following claim: the conditional adverbial mosi is licensed by a finite clause. Furthermore, I wil show that an investigation of the distribution of mosi gives us a window into the internal structure of Japanese conditional clauses. 88 3.3.1. Japanese Nominal Conditionals Japanese conditional marking is typicaly realized as an inflection on the verb. For example, in (79), the morpheme that marks the conditional (nara) appears after the tense marker of the embedded verb. There are, however, cases in which the same conditional marker is atached to NPs (1b). (79) a. Quinn-ga ringo-o tabe-ta-nara Q-nom apple-acc eat-tense-cond ?If Quinn eats an apple? b. Quinn-nara gakkoo-ni iru-hazu-da. Q-cond school-at be-must-copula ?If you are talking about Quinn, he is at school.? In this section, I wil concentrate on examples like (79b), which I cal Nominal Conditionals. In doing so, I have two goals in mind; first, to reveal the distributional properties of mosi, and second, to figure out the internal syntax of conditional clauses. I wil show that nominal conditionals give us important clues towards achieving these two goals. The basic line of argumentation that I wil pursue is that the two constructions, nominal conditionals and standard conditionals, are derived from the same base structure. I wil show that Rizi?s (1997) Split CP hypothesis gives us an important insight in this regard. What I wil specificaly claim is that a nominal conditional is a residue of clausal elipsis (specificaly the elipsis of FinP) applied to a standard conditional clause. 89 3.3.2. Some Basics of Nominal Conditionals First let us review some basic properties of nominal conditionals. In Japanese traditional gramar, the marker nara as used in nominal conditionals has been caled Teidai Joshi (Topic Particle) (Takubo & Masuoka 1992). As the name suggests, this particle can be used basicaly in the same context as a typical topic marker such as wa (Kuno 1973). For example, in (80), both nara and wa are used in a typical context for a topic. In this context, the topicalized NP Tanaka-san must have some topic marker, and if a Nominative case marker such as ga is atached the sentence sounds bad. Both wa and nara can be used naturaly in this context as opposed to ga. (80) A: Quinn mi-na-katta-kai? T see-neg-past-Q ?Have you sen Quinn?? B: Quinn-nara/wa/*ga toshokan-de Q-nara/wa library-at benkyoo-site-ta-yo. study-do-prog-part ?Quinn was studying at the library? Besides its use as a topic marker, the nominal conditional has another interesting property. As can be sen in (79), nominal conditionals use the same morphology as standard conditional clauses. Japanese conditional clauses can host conditional adverbs (CAs) such as mosi. Interestingly, nominal conditionals also can host mosi, but in a limited environment, namely when there is a pronoun sore 90 refering to a clausal element. The intuition behind the presence of mosi is something like the folowing: if mosi is inserted in a nominal conditional the focus of the sentence is on the proposition in the context. On the other hand, if mosi is absent, the focus is on the noun or the entity that is marked by nara. An interesting fact is that mosi requires the presence of the pronoun sore but this pronoun does not require the presence of mosi. This fact suggests that mosi is dependent upon the presence of sore but not vice versa. (81) a. mosi Quinn-ga ringo-o tabe-ta-nara mosi Q-nom apple-acc eat-past-cond ?if Quinn eats an apple? b. *(sore-ga) mosi Quinn-nara it-nom mosi Q-cond ?if it is Quinn? c. sore-ga Quinn-nara it-nom T-cond ?if it is Quinn? Given examples like (81b), an obvious question arises. Why does mosi in nominal conditional requires the pronoun sore? This brings us back to the bigger questions about what properties compose the diferences and similarities betwen standard conditional clauses and nominal conditionals. In the following section, I wil answer these questions. Furthermore, I wil show that answering these questions gives us important clues to the internal syntax of Japanese conditional clauses. 91 3.3.3. The Conditional Focus Link To clarify the licensing condition on mosi in nominal conditionals, in this section I wil examine the syntax of the clausal pronominal sore in Japanese. Example (81b) suggests that the presence of the pronoun sore is crucial for the licensing of mosi in nominal conditionals. Therefore, revealing the nature of sore naturaly leads us to the nature of mosi licensing. Once I reach the licensing of mosi, I wil reveal the internal structure of Japanese conditionals using the distribution of mosi as a clue. In so doing, I wil show that there is a tight connection betwen conditional clauses and various focus constructions in Japanese. 3.3.3.1. The Syntax of the Clausal Pronoun sore in Japanese 3.3.3.1.1. Nakao & Yoshida (2005) on Pronominal Sluicing To understand the nature of sore, I wil first review a previous study by Nakao & Yoshida (2005). Their analysis of sore is the basis for my analysis of nominal conditionals. Nakao and Yoshida (henceforth, N&Y) observe an interesting property of sore. Acording to N&Y, the clausal pronoun sore has the properties of both pronouns (or nominals) and clauses. The intuition behind this view is that clausal structure is hiding behind the pronominal expresion. To capture this intuition, N&Y analyze sore as the residue of elipsis of CP. More specificaly, they argue that the sluicing construction that contains sore is derived from the so-caled specificational cleft construction through deletion of the presuppositional clause, in which sore occupies the place where the presuppositional clause is originaly generated. N&Y 92 base their argument on the properties of sore in Japanese sluicing constructions. They cal the sluicing construction containing sore Japanese Pronominal Sluicing (JPS). N&Y?s basic observations on JPS can be summarized in (82). Let us consider each of the observations. (82) a. The wh-remnant in JPS shows connectivity efects. b. The copular construction in JPS induces honorification agrement with sore but not with the wh-remnant. c. JPS does not alow the inversion of presuppositional clause and focused elements. Connectivity Efects The wh-remnant in JPS shows variety of connectivity efects. (83a) is an example of binding and Case connectivity efects, and (83b) is an example of postposition connectivity efects. There are two important points in example (83a). Acusative Case in Japanese can be analyzed as a structural Case (Saito, 1982; Takezawa, 1985, among others). I asume, following Saito (1982), or Takezawa (1985), that it is asigned to an NP that is the sister of the verb, the direct object. Therefore, it is most plausible to think that the acusative case on the remnant phrase, zibun-zisin (self) in (83a) is asigned by the verb in the sister relation, not by any other means. Second, the verb siru (know) is not an ECM type verb (Hoji, Takano, Hiraiwa, Tanaka Takezawa among others). Thus it is not likely that siru or its negative form sira-nai asigns acusative case to the remnant NP. These two points 93 strongly suggest that the remnant NP is asigned acusative case by the verb that is in the elided site, i.e., seme-ta ?blame.? Turning to the binding connectivity, the local anaphor zibun-zisin (self) is bound by the subject in the first conjunct. Zibun-zisin, however, requires a local c-commanding antecedent (ref) as example (84b) indicates. Thus, it is not plausible to think that the subject in the first conjunct?which does not c-command the local anaphor?directly binds it. Rather, this example strongly suggests that the anaphor is bound by an antecedent in the elided structure. The combination of the two connectivity efects, Case and local anaphor licensing, strongly suggest that there is hidden clausal structure in JPS. The example in (83b), the connectivity efect involving postpositions, points to exactly the same conclusion. The postpositional phrase in (83b) is selected by a particular clas of verb such as okuru (send), and it is not compatible with a verb like wakar (know) in Japanese. Thus, as in the case of Case connectivity, the remnant postpositional phrase should be selected by the verb that is in the elided structure. (83) a. Quinn 1 -ga dareka-o seme-ta ga, watasi-wa Q-nom someone-acc blame-past but, I-top sore-ga zibun-zisin 1 -o kadooka sira-nai. it-nom self-acc whether know-not ?Quinn 1 blamed someone, but I don?t know whether it was himself 1 .? 94 b. Quinn-ga dokoka-kara nimotu-o Q-nom somewhere-from baggage- acc okuttarasii ga, watasi-wa sore-ga Tokyo-kara sent-seem but I-top it-nom Tokyo-from kadooka wakara-nai. whether know-not ?It sems that Quinn sent his baggage to us from somewhere, but I don?t know from where? (84) a. Quinnm-wa Virginia-o*/ga ok tensai-dearu kadooka Q-nom V-acc/nom genius-be whether sira-nai. know-not ?Quinn does not know whether Virginia is a genius or not.? b. Quinn 1 -wa Stillman 2 -no hahaoya 3 -ga Q-top S-gen mother-nom zibun-zisin 1*/2*/3 ok -o semeta-to itta. self-acc blame-comp said ?Quinn said that Stilman?s mother blamed herself.? Honorification Agrement The second property of JPS is its agrement property. In Japanese generative gramar, honorification has been treated as a gramatical agrement phenomenon (Harada 1976, Boeckx and Ninuma 2003 among others). Although this analysis contradicts the general view that Japanese is a language without agrement (Fukui 95 1995, Kuroda 1988), the reason that honorification is treated as an agrement phenomenon is mainly because it is sensitive to a certain gramatical function, and induces a certain type of locality efect. Before examining honorification in JPS, let us briefly review this point. Harada pointed out that what he cals propositional honorifics are sensitive to gramatical functions such as subject or object. Basicaly, both subject honorification and object honorification induce special morphology on the predicate. Subject honorification is marked by the form of o-predicate-ni-naru if the subject is a person socialy superior to the speaker and object honorification is marked by the form of o- predicate suru if the object is a person socialy superior to the speaker, respectively. They are ilustrated in the folowing examples. (85) a. Subject Honorification Sasaki-sensei-wa watasi-ni koo S-teacher-top I-dat this-way o-hanasi ni nat-ta. speak-past ?Mr. Sasaki told me this way? b. Object Honorification Watasi-wa Sasaki-sensei-ni koo I-top S-teacher-dat this-way o-hanasi si-ta. speak-past ?I told Mr. Sasaki this way.? 96 An important point of the paradigm above is that subject honorification cannot be used in the context of object honorification or vice versa. If honorification were not a gramatical phenomenon, this sensitivity to the gramatical functions would not be expected. This point can be easily shown by switching the subject and the object of each example in (85). In (86a), verb has the subject honorification marker despite the fact that the subject is not a person socialy superior to the speaker, and in (86b) object honorification is induced but the object is not the person socialy superior to the speaker. Both examples are indeed unaceptable. (86) a. *Watasi-wa Sasaki-sensei-ni koo I-top S-teacher-dat this-way o-hanasi-ni nat-ta. speak-past ?I told Mr. Sasaki this way.? b. *Sasaki-sensei-wa watasi-ni koo S-teacher-top I-dat this-way o-hanasi si-ta. speak-past ?Mr. Sasaki told me this way.? Another supporting argument for the position that honorification is a gramatical phenomenon comes from the fact that it is subject to a certain type of locality condition (Boeckx and Ninuma 2003). Acording to Boeckx and Ninuma, object honorification is an agrement relation betwen the smal V and an object. 97 They further argue that this agrement relation is disrupted if there is an ofending intervener betwen the smal V and the object NP. To ilustrate this point, they cite the examples in (87). The patern ilustrated by these examples is that the verb can have object honorification when the indirect object is a person socialy superior to the speaker, but it cannot if the direct object is the honorification inducer. The comparison betwen (87b) and (87c) further supports this point. If the indirect object is not present, the sentence sounds much beter. (87) a. Virginia-ga Tanaka-sensei-ni Quinn-o V-nom T-teacher-dat Q-acc go-syookai si-ta. introduce-past ?Virginia introduced Quinn to Mr. Tanaka.? b. *Virginia-ga Quinn-ni Tanaka-sensei-o V-nom Q-dat T-teacher-acc go-syookai si-ta. introduce-past ?Virginia introduced Mr. Tanaka to Quinn.? c. Virginia-ga Tanak-sensei-o go-syookai si-ta. V-nom T-teacher-acc introduce -past ?Virginia introduced Mr. Tanaka (to someone).? Boeckx and Ninuma argue that this patern indicates that object honorification is subject to locality constraints. Simply put, if there is an indirect object, the direct 98 object cannot control the object honorification. They argue that honorification agrement is established by the operation Agre, and Agre is subject to locality conditions, the so-caled Defective Intervention Constraint (Chomsky 2000). Asuming that the indirect object is base generated at the higher position than the direct object, i.e., the indirect object c-commands the direct object, the indirect object intervenes v and the direct object. Schematicaly, (88) ilustrates this point, where the indirect object is intervening betwen v and the direct object. (88) vP VP v IO V? DO V Boeckx and Ninuma cite various types of supporting evidence for their position, and argue that honorification is best analyzed as being established by the operation Agre rather than by movement, thus supporting the validity of a syntactic theory which incorporates Agre. Putting the technical details aside, for us the most crucial point in both Harada?s and Boeckx and Ninuma?s studies is that honorification is a gramatical phenomenon, and best treated as an agrement relation betwen a verb and its object. With the discussion so far in mind, let us turn to honorification agrement in JPS. The crucial example for us is in (89). In (89), the copula in the second conjunct shows subject honorification. The wh-word dare is the object position and does not 99 have the form of a person socialy superior to the speaker. If the wh-word is used for a socialy superior person, it has the form of dotira or dono-kata but not dare. Thus, it is clear that the honorification is induced by sore but not by the wh-remnant. This fact, that if sore is not present, the example is not aceptable, shows that the agrement is established betwen sore and the copula. (89) ? Quinn-ga aru kata-ni o-ai si-ta Q-nom certain person-dat meet(Obj.Hon) -past rasii ga watasi-wa sore-ga dare-ni seem but I-top it-nom who-dat de-irassyat-ta ka sira-nai. be(Subj.Hon)-past Q know-not ?It sems that Quinn met a certain person, but I don?t know who it was.? This agrement patern can be captured if we asume there is a hidden clausal structure in the guise of the pronoun sore as ilustrated in (90). (90) is a cleft construction that Nakao and Yoshida asume to underlie JPS. In (90), the presuppositional clause occupies the subject position. Basicaly the presuppositional clause can be analyzed as a nominalized clause that refers to the NP in the focus position. If this is correct, the presuppositional clause can be honorificational, and this is plausible because the gap corresponds to the focalized wh-phrase, dare-ni ?to whom? is honorified by the object honorificational verb in the presuppositional clause. The copular in the second conjunct in (89) agres with this hidden 100 presuppositional clause, and because it is in the subject position, it induces subject honorification. (90) ?watasi-wa [ CP [ IP Quinn-ga t 1 o-ai si-ta] no]-ga I-top Q-nom meet(Obj.Hon) comp-nom dare 1 -ni de-irassyat-ta ka sira-nai. who-dat be(Subj.Hon.)-past Q know-not ?I don?t know who it was who Quinn met.? If we do not asume that the presuppositional clause is underlying sore, it is not clear why subject honorification can take place in JPS. Putting the use as a clausal pronoun aside, the most basic use of sore is to refer to an inanimate object. Thus sore itself does not have a feature that can induce honorification agrement. If we asume that sore is a simple pronoun without any clausal structure, and that it can be coreferential to its antecedent or interpreted in the way that deep anaphora is interpreted (Hoji 1995), then we have to asume that somehow the honorificational feature as wel as an animacy feature is asigned to the pronoun via a coreference relation or a deep anaphoric relation because, as we have sen, honorification is an agrement phenomenon. However, such feature asignment does not sem to be observed in the other environments. On the other hand, under Nakao and Yoshida?s analysis, the honorification phenomenon in JPS is given a straightforward explanation, i.e., there is a hidden presuppositional clause that has a feature for honorification. 101 Finaly, Nakao and Yoshida point out that the same kind of agrement patern can be sen in the English specificational pseudocleft construction. (91) shows that if a pseudocleft construction is interpreted specificationaly, the copular agres in number with the pre-copular presuppositional clause, rather than with the post- copular focus phrase. (91) [What you have bought] is fake jewels. This paralelism betwen JPS and English specificational pseudoclefts support the validity of Nakao and Yoshida?s analysis. Ban on Inversion The fourth property of JPS is the ban on inversion of the pre- and post-copular elements. For some unknown reasons, Japanese specificational constructions do not alow inversion of the presuppositional clause and the focused phrase. (92) a. [Quinn-ga katta no]-wa hon-o Q-nom bought comp-top book-acc san-satu da. three-classifier be ?What Quinn bought was thre books ? 102 b. *Hon-o san-satu-wa [Quinn-ga book-acc three-cassifier-top Q-nom katta no] da. bought comp be ?What Quinn bought is thre books ? In the same way, JPS does not alow the inversion of the pronoun sore and focused phrase. (93a) is an example of JPS with the regular order of the presuppositional clause and the focused phrase. In (93b), on the other hand, they are inverted, and the example is not aceptable. (93) Quinn-ga nanika-o katta-rasii ga Q-nom something-acc bought-seem but a. watasi-wa sore-ga nani-o (da) ka sira-nai. I-top it-nom what-acc be Q know-not b. *watasi-wa nani-ga sore (da) ka sira-nai. I-top what-nom it be Q know-not ?It sems that Quinn bought something, but I don?t know what it is.? The ban on inversion also indicates a strong paralelism betwen the specificational construction and JPS. Under the asumption that the unsen presuppositional clause is represented, the ban on inversion is given a straightforward explanation, i.e., because the presuppositional clause and the focused phrase cannot be inverted, the pronoun sore and the focused phrase cannot be inverted either. 103 3.3.3.2.Sumary So far I have shown various reasons to believe that some clausal structure, the presuppositional clause of the pseudocleft construction, underlies sore in JPS. The connectivity efects indicate that there is a clausal structure in sore, and the paralelisms betwen JPS and pseudocleft suggest that these two are actualy the same elements, and thus support the claim that sore corresponds to the presuppositional clause. 3.3.3.3. Sore in Nominal Conditionals Based on the discussion so far, I wil argue that sore in nominal conditionals is exactly the same element as the one found in JPS. To this end, I wil show that al the properties of JPS and sore in JPS that we have reviewed hold true for nominal conditionals with sore. The properties of sore in JPS that we have sen so far are summarized in (94). As I sumarize in the chart in 0), al of the properties of sore in JPS are sen in the sore in nominal conditionals. Let us examine these properties one by one. (94) a. The wh-remnant in JPS shows connectivity efects. b. The copular in JPS induces honorification agrement with sore but not with the wh-remnant. c. JPS does not alow the inversion of pre- and post-copular elements. 104 Properties (95) Conectivity Honorification Inversion JPS ? ? ? Nominal Conditionals ? ? ? The examples in (96) show that sore in nominal conditionals exhibits connectivity efects of Case, local anaphor licensing and postpositions. Exactly like the examples of JPS, sore in nominal conditionals shows connectivity efects of local anaphor licensing and Case in (96a) where the anaphor zibun-zisin is bound by the subject in the first conjunct, and acusative case is asigned to the remnant, and the connectivity efect of the postposition kara (from) in (96b). (96) a. Quinn 1 -ga dareka-o semete-iru rasii ga, Q-nom someone-acc blame-prog seem but mosi *(sore-ga) zibunzisin 1 -o nara mosi it-nom self-acc cond yameta hoo-ga yoi. stop it-is-better ?It sems that Quinn is blaming someone but if it is himself, he had beter stop doing that.? 105 b. Quinn-ga dokoka-kara nimotu-o Q-nom somewhere-from baggage-acc okutta-rasii ga mosi *(sore-ga) Tokyo-kara nara sent-seem but mosi it-nom Tokyo-from cond sugu-todoku daroo. soon-arrive will ?It sems that Quinn sent the baggage from somewhere, but if it is from Tokyo, the baggage wil arive soon.? The pronominal nominal conditional shows a similar honorification patern to JPS. In (97) the copular shows subject honorification even though the honorifiable person, Yamada-sense (Mr. Yamada) is in the object position. Thus, as in JPS, we can conclude that the copular agres with sore, and so the same explanation of honorification agrement holds for pronominal nominal conditionals. (97) Quinn-wa Yamada-sensee ka Sasaki-sensee-ni Q-top Y-teacher or S-teacher-dat oai-sita rasii ga, mosi sore-ga Yamada-sensee-ni meet(Obj.Hon.) seem but mosi it-nom Y-teacher-dat de-irasshatta-nara kitto kinchoo-sita daroo. cop(Subj.Hon)-cond certainly nervous-did may ?It sems that Quinn met Mr. Yamada or Mr. Sasaki, but if it was Mr. Yamada, Quinn was certainly nervous.? 106 Finaly inversion of sore and the remnant phrase is not alowed in pronominal nominal conditionals, as (98) shows. (98) Quinn-wa ringo ka mikan-o tabe-ta-rasii ga Q-top apple or orange-acc eat-past-seem but a. mosi sore-ga ringo-nara mosi it-nom apple-cond b. *mosi ringo-ga sore-nara mosi apple-nom it-cond oisi-katta hazu-da. tasty-past must-copular ?Quinn ate an apple or an orange, but if it was an apple that he ate, it must have been tasty.? These paralelisms betwen JPS and pronominal nominal conditionals strongly suggest that sore in these constructions is the same element. If we analyze these elements in diferent ways we cannot acount for the paralelism of these two sores. 3.3.3.4. Sore and Mosi The striking paralelism betwen JPS and pronominal nominal conditionals strongly suggests that there is a clausal structure in the guise of a pronominal in these constructions. The examples of connectivity efects strongly suggest the existence of clausal hierarchical structure in sore in these examples. Relatedly, JPs and nominal 107 conditionals show yet another similarity, i.e., the connectivity efects are not legitimate if sore is not present. We can se this point in the examples in (99). In both JPS and nominal conditionals, the examples without sore are degraded with the local anaphor, i.e., the local anaphor is not bound in the examples without sore. (99) Connectivity efects a. Quinn 1 -ga dareka-o seme-ta ga, watasi-wa Q-nom someone-acc blame-past but, I-top *(sore-ga) zibun-zisin 1 -o kadooka sira-nai. it-nom self-acc whether know-not ?Quinn 1 blamed someone, but I don?t know whether it was himself 1 .? b. Quinn 1 -ga dareka-o semete-iru rasii ga, Q-nom someone-acc blame-prog seem but *(sore-ga)zibunzisin 1 -o nara yameta hoo-ga yoi. it-nom self-acc cond stop it-is-better ?It sems that Quin is blaming someone but if it is himself, he beter stop doing that.? This correlation betwen the presence of sore and connectivity efects can be easily understood if there is indeed a clausal structure in the second conjunct when sore is present, but there is no such hidden clause when sore is absent. Interestingly, the distribution of mosi also observes an analogous efect of connectivity. In other words, if sore is not present, mosi is not licensed, in much the same way that the local anaphor is not licensed. 108 (100) Quinn 1 -ga dareka-o semete-iru rasii ga, Q-nom someone-acc blame-prog seem but mosi *(sore-ga) zibunzisin 1 -o nara yameta mosi it-nom self-acc cond stop hoo-ga yoi. it-is-better ?It sems that Quinn is blaming someone but if it is himself, he beter stop doing that.? It should be clear that when mosi is licensed, mosi clearly requires the clausal element that is replaced by the pronoun sore. If not, the unaceptability of the above example is not easily understood. Thus, I conclude that mosi is licensed by a clause that can be replaced by the clausal pronoun sore. Now the question is the identity of the ?clausal element? that is replaced with sore. If we get a beter idea of the structure of this clausal element, we can get a beter understanding of the licensing condition on mosi. 3.3.3.5.A Problem Remains Before turning to the investigation of the identity of the clausal element, let us note one problem with the current analysis that sore replaces the presuppositional clause of pseudocleft. 109 As it is wel known, NPs in Japanese can be phonologicaly empty. The inanimate pronoun sore, for example, can be empty. In (101b), the object that corresponds to the pronoun sore in A?s utterance is omited. 6 (101) a. A: Quinn-wa sore-o tabe-ta-no? Q-top it-acc eat-past-Q? ?Did you eat that, Quinn?? b. B: Un, ? tabe-ta-yo. yes, eat-past ?Yes, I ate that.? Given this property of empty argument licensing in Japanese, the most straightforward expectation with respect to the clausal pronoun sore is that it can also be empty. However, the fact that connectivity efects cannot be observed in the sluicing construction or nominal conditionals in the absence of sore suggests to us the opposite. If empty clausal pronoun sore exists, we expect no diferences betwen Japanese standard sluicing constructions or nominal conditionals without sore and JPS and pronominal nominal conditionals. Even if we asume some empty element is 6 Various analyses of this phenomenon have ben proposed. These proposed analyses can be clasified into two types. One is known as the pro analysis (Hoji 199), and the others are VP-elipsis analyses (Otani and Whitman 191). The treatment of this type of construction is stil controversial. I wil leave this problem open. For us, it is important that the object coresponding to the inanimate pronoun sore, or the larger constituent that contains sore, can be omited. 110 there, we are forced to conclude that this empty element is diferent from the clausal pronoun sore. 3.3.3.6. Toward the Licensing Condition of Mosi So far, I have shown that the clausal pronoun sore in JPS and the pronominal nominal conditional are the same element by showing the paralelisms betwen these two constructions. The conclusion we have reached is that mosi requires a clausal structure that can license it. In this subsection, I wil show that by adopting the analysis of focus construction in Japanese proposed by Hiraiwa and Ishihara, we can succesfully derive these paralelisms, and go closer to the nature of the Conditional Adverb mosi. In this subsection, I wil establish the following two claims. First, Japanese conditional clauses have an articulated CP structure that can be best analyzed by Rizi?s (1997) Split-CP analysis. Second, mosi is licensed by FinP under the Split-CP analysis. In the course of the discusion, I wil further propose a general licensing condition on a certain type of clausal elipsis in Japanese. 3.3.3.6.1. Rizi (1997) The heart of Rizi?s (1997) Split-CP hypothesis is that just as there is evidence that IP is beter analyzed by spliting it up into multiple projections, there is also evidence that CP should be split up into several projections. Specificaly, he claims that CP has the internal makeup ilustrated in (102). 111 (102) [ ForceP Force [ TopP* Top [ FocP Foc [ TopP* Top [ FinP Fin [ IP ?]]]] Rizi shows numerous pieces of empirical evidence for the split-CP system, but before reviewing his empirical arguments, let us look at his conceptual motivation for spliting CP up into multiple functional categories. It has been traditionaly argued that CP has selectional relations with the verb that takes it as a complement and with the clause that it takes as a complement. Viewed from outside, diferent types of clauses are selected by diferent verbs. For example, believe selects declarative clauses, wonder selects interogative clauses, and know selects either declarative clauses or interogative clauses. In this way, CP marks diferent clause types in acord with the semantics of the verb that takes the CP as its complement (Bresnan, Chen, Chomsky, etc). Rizi cals this function of CP marking clause types the force system. (103) a. Quinn believes that Virginia went to school. b. *Quinn believes who went to school. (104) a. *Quinn wonders that Virginia went to school. b. Quinn wonders who went to school (105) a. Quinn knows that Virginia went to school. b. Quinn knows who went to school. On the other hand, C itself selects a certain type of clause. Just as a specific type of verbs takes a specific type of CP, a specific complementizer takes a specific 112 type of clauses. Taking an example from English, the complementizer that takes a finite IP as its complement, but for takes an infinitival IP as its complement. He cals this a finitenes system. (106) a. Quinn thinks that Virginia went to school. b. *Quinn thinks that Virginia to go to school. (107) a. Quinn wants very much for Virginia to go to school. b. *Quinn wants very much for Virginia went to school. Thus, it sems that CP is responsible for two diferent functions. Rizi?s claim is that these two diferent functions can be best expresed as properties of diferent heads in ?CP.? Besides his conceptual argument, Rizi offers many empirical arguments for the existence of multiple layers in CP. Below, I pick several of his arguments to discuss so that we can se how things difer under the split-CP framework. Rizi bases his arguments for the existence of FocP and TopP on the relative order of focalized and topicalized elements with respect to each other and to ?complementizers.? Let us start from the relative orderings of topicalized phrases and complementizers. In Italian, the complementizer di that introduces non-finite embedded clauses follows topicalized phrases. On the other hand, another complementizer, che, that introduces finite embedded clauses, precedes them. 113 (108) a. Credo che loro apprezerebbero molto il tuo libro. ?I believe that they would appreciate your book very much.? b. Credo che il tuo libro, loro lo apprezerebbero molto. that your book ?I believe that your book, they would appreciate it a lot.? c. *Credo, il tuo libro, che loro lo apprezerebbero molto. your book that ?I believe, your book, that they would appreciate it a lot. ? (109) a. Credo di apprezare molto il tuo libro. ?I believe of to appreciate your book very much.? b. *Credo di il tuo libro, apprezarlo molto. of your book ?I believe of your book, to appreciate it a lot. ? c. Credo il tuo libro, di apprezarlo molto. your book of ?I believe your book of to appreciate it a lot.? The relative ordering of complementizers and topicalized phrases cannot be easily acommodated by a system with unique C position. However, the split-CP system that Rizi is adopting can easily capture their distribution by asuming that che is the Force 0 element and di is Fin 0 element respectively as is ilustrated in (110). (110) [ ForceP che [ TopicP ? [ FinP di [ IP ?]]] 114 Rizi also examines the relative order of wh-words and topicalized phrases, to make the same point. What he specificaly shows is that wh-words in relative clauses must precede topicalized phrases, but interogative wh-phrases must follow them. Schematicaly the ordering can be ilustrated as in (111). Relevant examples are (112) and (113), both from Italian. (111) ? Relative-Wh ? Topic Phrases ? Interogative-Wh ? IP (112) Wh-words in Relative Clauses a. Un uomo a cui, il premio Nobel, lo daranno senz?altro. to whom the Nobel Prize ?A man to whom, the Nobel Prize, they wil give it undoubtedly.? b. *Un uomo, il premio Nobel, a cui lo daranno senz?altro. the Nobel Prize to whom ?A man, the Nobel Prize, to whom they wil give it undoubtedly? (113) Interogative Wh-words a. *A chi, il premio Nobel, lo daranno? to whom the Nobel Prize ?To whom, the Nobel Prize, wil they give it?? b. Il premio Nobel, a chi lo daranno? the Nobel Prize to whom ?The Nobel Prize, to whom wil they give it?? 115 These examples suggest that wh-phrases in relative clauses occupy the specifier of the highest phrase, and that interogative wh-phrases occupy the specifier of a lower phrase than TopP, and this in turn suggests that interogative wh-phrases occupy the specifier of FocP. To support this claim, Rizi further shows that interogative wh- phrases cannot cooccur with focalized phrases, but relative wh-phrases can and must precede them. Rizi cites the examples in (114) and (115). The examples in (114) suggest that interogative wh-phrases and focus phrases compete for the same spot. Asuming that there is just one position for focus, at least in Italian, the contrast in (114) can be easily captured under the split-CP system, i.e., interogative wh-phrases occupy specifier of FocP that is lower than TopP. Turning to (115), we se that relative wh-phrases are compatible with focused phrases and, additionaly, that they must precede focus phrases. This is predicted by the split-CP system that Rizi asumes, if we hypothesize that relative wh-phrases occupy the specifier of ForceP and that FocP is located lower than ForceP. (114) Focus is not compatible with interogative wh-phrases a. *A GIANI che cosa hai deto (, non a Piero)? to Gianni what Focus Wh ?TO GIANI what did you tel (not to Piero)?? 116 b. *Che cosa A GIANI hai deto (, non a Piero)? what to Gianni Wh Focus ?What TO GIANI did you tel (not to Piero)?? (115) Focus is compatible with relative wh-phrases a. Eco un uomo a cui IL PREMIO NOBEL dovrebbero dare (non il to whom THE NOBEL PRIZE Rel.Wh Focus premio X). ?Here is a man to whom THE NOBEL PRIZE they should give (not prize X).? b. *Eco un uomo IL PREMIO NOBEL a cui dovrebbero dare? THE NOBEL PRIZE to whom Focus Rel.Wh Taken together, the distribution of these elements can be acurately captured by the split-CP system, but not by traditional analyses asuming that there is single CP projection. Rizi shows many other types of supporting evidence for his split-CP system. For us, however, the above should be sufficient to se the motivations for spliting CP up into multiple layers. 117 3.3.3.6.2. Focus Constructions in Japanese: Hiraiwa & Ishihara (2001) Based on Rizi?s split-CP system, Hiraiwa and Ishihara offer an analysis of several Japanese focus constructions. Specificaly, they pick thre focus constructions that have been known to have a lot of similarities, but for which no analyses were previously offered which could capture these similarities, and they show that the properties of these constructions can be readily derived by Rizi?s articulated CP- system. In this subsection, I wil briefly review their study, and build a bridge to the analysis of conditional clauses. Properties of Focus Constructions In Japanese there are several focus constructions that share significant similarities. Hiraiwa and Ishihara cal them the cleft construction, the no-da in-situ focus construction, and the sluicing construction, respectively. Examples of these constructions are summarized in (116). (116) a. Cleft [Quinn-ga tabe-ta no]-wa kono ringo-o Q-nom eat-past comp-top this apple-acc (mit-tsu) da. three-cl cop ?It is this apple that Quinn ate.? 118 b. No-da [Quinn-ga kono ringo-o tabe-ta no] da. Q-nom this apple-acc eat-past comp cop ?It is that Quinn ate this apple.? c. Sluicing Quinn-ga Virginia-ni nanika-o Q-nom V-dat something-acc age-ta arasii ga boku-wa [nani-o (da) ka] give-past seem but I-top what-acc cop Q wakara-nai. know-not. ?Quinn gave something to Virginia but I don?t know hat.? Let us first summarize the important properties of the cleft construction, and se to what extent the properties of clefts are shared by the other two constructions. Properties of Clefts Hiraiwa and Ishihara report five main properties of clefts, which I list in (117). (117) a. The possibility of multiple foci b. Island efects c. The complementizer no cannot be substituted with a pronoun/NP d. The unavailability of Nominative/Genitive Conversion (NGC) 119 e. Clause mate condition on multiple foci As is wel known, Japanese focus constructions (including cleft constructions), alow multiple foci (Koizumi 1995). In (118a), two NPs are in the focus position. Interestingly, if any of these focused elements are not case-marked, multiple foci are not alowed. This is, however, not true when there is just one focus element. If there is just one focus element, the case marker can be dropped. The cleft construction containing the focus element without case is sometimes caled a Pseudocleft (Kuroda 1999) or a Non-Case Marked-Cleft (Hoji 1990). Here folowing Hiraiwa and Ishihara?s notation, I wil refer to it as a pseudocleft. (118) a. [Quinn-ga age-ta no]-wa Virginia-ni Q-nom give-past comp-top V-dat ringo-o (mit-tsu) da. apple-acc three-cl cop ?(Lit.) It is (thre) apples to Virginia that Quinn gave.? b. *[Quinn-ga age-ta no]-wa Virginia*(-ni) Q-nom give-past comp-top V-dat ringo*(-acc) (mit-tsu) da. apple-acc three-cl cop 120 (119) [Quinn-ga Virginia-ni age-ta no]-wa ringo da. Q-nom V-dat give-past comp-top apple-? cop. ?It is an apple that Quinn gave to Virginia.? The second property is island sensitivity. As we can se in the contrast in (120), clefting but not pseudoclefting is sensitive to islands. In the cleft sentence, (120a), the focus element is extracted out of a complex NP and the sentence is not aceptable. On the contrary, the pseudocleft sentence in (42b) is aceptable even though the focus element is extracted out of a complex NP. This contrast is taken to indicate that clefting involves movement whereas pseudoclefting does not. 121 (120) a. *[Quinn-ga [[e 1 e 2 kai-ta] hito 1 ]-o Q-nom write-past person-acc hihansi-ta no]-wa kono-ronbon-o 2 da. criticize-past comp-top this-paper-acc cop ?(Lit) It is this paper that Quinn criticized the person who wrote.? b. [Quinn-ga [[e 1 e 2 kai-ta] hito 1 ]-o T-nom write-past person-acc hihansi-ta no]-wa kono-ronbon 2 da. criticize-past comp-top this-paper-? cop ?(Lit) It is this paper that Quinn criticized the person who wrote.? The third property is relevant to the properties of the complementizer involved in clefting. In a cleft construction, the complementizer -no cannot be substituted with an NP. In a pseudocleft, however, such alternation is possible. (121) a. [Quinn-ga tabe-ta no/*mono/*kudamono]-wa Q-nom eat-past comp/thing/fruit -top ringo-o (mit-tu) da. apple-acc three-cl be ?It was (thre) apples that Quinn ate? 122 b. [Quinn-ga tabe-ta no/mono/kudamono]-wa ringo Q-nom eat-past comp/thing/fruit-top apple-? (mit-tu) da. three-cl be ?It is (thre of) apples that Quinn ate? The fourth property is concerned with the status of the complementizer no that is used in cleft and pseudocleft constructions. As we have reviewed in section 2 in this chapter (relative clause section), a nominative NP in Japanese optionaly converts to a genitive NP in a structure in which a verb bears adnominal form (se Hiraiwa 2002; Watanabe 1996). This is caled Nominative-Genitive Conversion. Even though both in cleft and pseudocleft constructions the predicate bears adnominal form, cleft resists nominative-genitive conversion whereas pseudocleft alows it. The following examples ilustrate this contrast. (122) a. Cleft [Quinn-ga/ ? no tabe-ta no]-wa kono ringo-o Q-nom/ ? gen eat-past comp-top this apple-acc (mit-tu) da. three-cl cop ?It is (thre of) these apples that Quinn ate.? 123 b. Pseudocleft [Quinn-ga/no tabe-ta no]-wa kono ringo Q-nom/ ? gen eat-past comp-top this apple-? (mit-tu) da. three-cl cop ?It is (thre of) these apples that Quinn ate.? Finaly, the fifth property of clefts that Hiraiwa and Ishihara point out is the clausemate condition. As we have sen, unlike pseudoclefting, clefting alows multiple foci. The cleft with multiple foci, which is caled a multiple cleft, is known to be subject to the so-caled Clause-Mate Condition: The elements undergoing Multiple Clefting must be clausemates. (123b) and (123c) are derived from the same base sentence (123a). (123b) violates the clausemate condition and the sentence is unaceptable, but (123c) satisfies the condition and it is aceptable. (123) a. Base sentence Virginia-ga sensei-ni [Quinn-ga kono ringo-o V-nom teacher-dat Q-nom this apple-acc tabe-ta to] iituketa. eat-past comp told ?Virginia told the teacher that Quinn ate this apple.? 124 b. Clausemate Condition violation *[Virginia-ga e 1 [Quinn-ga e 2 tabe-ta to] V-nom Q-nom eat-past comp told iituketa no]-wa sensee-ni 1 kono ringo-o 2 da. comp-top teacher-dat this apple-acc cop ?(Lit.) it is the teacher, this apple that Virginia told that Quinn ate.? c. Satisfying the clausemate condition [Virginia-ga sensee-ni [e 1 e 2 tabe-ta to] V-nom teacher-dat eat-past comp iituketa no]-wa Quinn-ga 1 kono ringo-o 2 da. told comp-top Q-nom this apple-acc cop ?(Lit.) It is Quinn, this apple that Virginia told the teacher that ate.? Properties of No-da Constructions and Sluicing Constructions Now, let us turn to the other two focus constructions and se the similarities and diferences among these thre constructions. Hiraiwa and Ishihara point out that no-da constructions show significant syntactic paralelism with cleft constructions. Other than the island sensitivity and the clausemate condition, the no-da construction shows al the properties of cleft constructions: It alows multiple foci, it resists substitution of the complementizer with an NP, and nominative-genitive conversion is not alowed. However, the focused phrase can be inside an island, and if there are two focused phrases, they are not necesarily clausemates. These two properties presumably derive from constraints on overt movement, and focused phrases in the no-da construction do not sem to move 125 overtly. Given that the no-da construction does not contain any word order permutations, it is thus rather natural that it should difer from the cleft constructions on exactly these two properties. (124) a. Multiple Foci [QUINN-ga KONO RINGO-o tabe-ta no] da Q-nom this apple-acc eat-past comp cop focus focus ?(Lig.) It is Quinn, this apple that (he) ate.? b. NP-substitution [QUINN-ga kono ringo-o tabe-ta Q-nom this appke-acc eat-past no/*mono/*kudamono] da. comp/thing/fruit cop ?It is Quinn who ate this apple.? c. Nominative-Genitive Conversion [QUINN-ga/*no kono ringo-o tabe-ta no] da Q-nom/*gen this apple-acc eat-past comp cop ?(Lig.) It is Quinn, this apple that (he) ate.? d. Island [Quinn-ga [[e 1 KONO RONBUN-o kai-ta] Q-nom this paper-acc write-past hito 1 ]-o hihansi-ta no] da. person-acc criticize-past comp cop 126 e. Clausemate Condition [Virginia-ga SENSEE-ni [Quinn-ga KONO RINGO-o V-nom teacher-dat Q-nom this apple-ac tabe-ta to] iituke-ta no] da. eat-past comp told comp cop ?Virginia told the teacher that Quinn ate this apple.? Next let us look at sluicing. Hiraiwa and Ishihara show that sluicing and pseudosluicing in which the remnant NP does not bear any case markers or post- positions 7 (Fukaya and Hoji 1999), show exactly the same type of contrast that can be sen betwen clefting and pseudoclefting, except for the lack of a clausemate condition. Although I wil not go into any details about the clausemate condition, I would like to note that they suggest that the clausemate condition does not hold in sentences with wh-phrases, and that sluicing is crucialy a focus construction with wh-phrases. Thus there may be independent reasons for the lack of the clausemate condition, (for more details, se Hiraiwa and Ishihara). Putting the mater of the clausemate condition aside, acording to Hiraiwa and Ishihara the only diference betwen cleft and sluicing is the application of elipsis and elipsis-specific properties (if any). The relevant cases are the following: Sluicing 7 Fukaya and Hoji (199) use diferent terms for these constructions. They cal the former case-marked sluicing and the later non-case-marked sluicing. Here again, I folow Hiraiwa and Ishihara?s (202) terminology. Fukaya and Hoji show varieties of diferences betwen these two constructions. For more details, se Fukaya and Hoji and also Fukaya (202). 127 alows multiple foci but pseudosluicing does not; sluicing shows island efects but pseudosluicing does not. The examples below (125)-(126) show these points. (125) Multiple foci Quinn-ga dareka-ni nanika-o age-ta rasii ga? Q-nom someone-dat something-acc give-past seem but? ?It sems that Quinn gave something to someone but? a. Sluicing boku-wa [dare-ni nani-o] (da) ka wakara-nai. I-top who-dat what-acc cop Q know-not. ?I don?t know what he gave to whom.? b. Pseudosluicing *boku-wa [dare-? nani-?] (da) ka wakara-nai. I-top who-dat what-acc cop Q know-not. ?I don?t know who, what.? 128 (126) Island sensitivity 8 Quinn-wa [[otooto-ni nanika-o okuttekita]hito]-o Q-top brother-dat something-acc sent person-acc syootai-sita rasii ga? invite-past seem but ?It sems that Quinn invited a person who had sent something to his brother, but?? a. Sluicing *boku-wa [nani-o (da) ka] sira-nai I-top what-acc cop Q know-not ?I don?t know what? b. Pseudosluicing boku-wa [nani-? (da) ka] sira-nai. I-top what cop Q know-not ?I don?t know what.? 8 Extraction out of a complement clause does not show the degradation. (i) Taro-wa Hanako-ga ototo-ni nanika-o okutekita-to T-top H-nom brother-dat something-ac sent-comp ita-ga, boku-wa nani-o da-ka sira-nai. said-but, I-top what-ac cop-Q know-not. ?Taro said that Hanako sent something to his brother, but he does not know hat.? 129 The example relevant to the clausemate condition is the following: In (127) the remnants of sluicing are clearly not clausemates, and stil the sentence is perfectly aceptable. (127) Virginia-ga dareka-ni [Quinn-ga nanika-o H-nom someone-dat T-nom something-acc tabe-ta to] iituke-ta rasii ga, boku-wa dare-ni eat-past comp tell-past seem but I-top who-dat nani-o (da) ka what-acc cop Q wakara-nai. know-now. ?It sems that Virginia told someone that Quinn ate something, but I don?t know to whom what.? Finaly, I would like to note that Hiraiwa and Ishihara report the example in (128a) as an example that shows a ban on NP-substitution, and they argue that this also is a similarity betwen cleft and sluicing. Acording to their judgment, the example is not good. However, as we have discussed earlier, this type of sentence is an example of JPS, and I have used the same type of example to show Case connectivity in JPS. To my ear, this example does not sound bad, and acording to the native speakers of Japanese who I have interviewed, this example is as good as its pseudocleft counterpart. Given my earlier argument, and the judgment of this sentence, this suggests that the pronoun in this example corresponds to the elided 130 clause itself rather than the complementizer. Note though that, although I disagre with their treatment of this particular pair of examples, this does not diminish the validity of their broader point. The many other similarities and diferences of sluicing and cleft must stil be explained. (128) Virginia-ga nanika-o kat-ta rasii ga? V-nom something-acc buy-past seem but ?It sems that Virinia bought something, but?? a. Sluicing boku-wa [sore-ga nani-o (da) ka] wakara-nai. I-top it-nom what-acc cop Q know-not ?I don?t know what it is.? b. Pseudosluicing boku-wa [sore-ga nani (da) ka] wakara-nai. I-top it-nom what-? cop Q know-not ?I don?t know what it was.? Japanese Focus Constructions Under a Split-CP analysis So far we have sen that thre types of focus constructions in Japanese show significant similarities. To capture these similarities Hiraiwa and Ishihara ofer an analysis incorporating Rizi?s split-CP hypothesis. Their argument can be summarized as follows: Clefting and sluicing are derived from the underlying no-da construction through movement of focalized phrases folowed by topicalization or 131 elipsis of the clause headed by no. Thus, the thre constructions show similarity because they share the same base structure. On the other hand, their diferences are derived from the operation responsible for deriving each surface form. If elipsis takes place, properties related to elipsis show up; if movement takes place, the properties related to movement show up. Let?s se their analysis. First, they asume Split CP system. Acording to their analysis, sentence-final particles are hierarchicaly distributed to the heads of thre CP-projections: FinP/CP, FocP and TopP. (129) is the schematic representation of the split-CP system that they adopt. (129) [ TopP XP-wa ?[ FocP ? [ FinP ? [ IP ? [ VP ?V]-tense]-no]-da] Under this split-CP system, the derivation of each construction can be ilustrated as in (130) for clefts and (131) for sluicing. (130) The derivation of clefts a. Step 0 [ TopP [ FocP [ FinP [ TP NP-ga [ VP NP-o V]-tense]-no]-da] b. Step 1: Focus movement [ TopP [ FocP NP-o [ FinP [ TP NP-ga [ VP t NP-o V]-tense]-no]-da] 132 c. Step 2: Topic movement of FinP [ TopP [ FinP [ TP NP-ga [ VP t NP-o V]-tense]-no]-wa [ FocP NP-o t FinP -da] (131) The derivation of sluicing a. Step 0 [ TopP [ FocP [ FinP [ TP NP-ga [ VP wh-o V]-tense]-no]-da] b. Step 1: Focus movement [ TopP [ FocP wh-o [ FinP [ TP NP-ga [ VP t NP-o V]-tense]-no]-da] c. Step 2: Elipsis of FinP [ TopP [ FocP NP-o [ FinP [ TP NP-ga [ VP t NP-o V]-tense]-no]-da]] -ka As can be sen clearly in each derivation, the crucial step for driving clefting or sluicing is focus movement. From these derivations, it is obvious why clefting and sluicing are sensitive to islands but not the no-da construction: There is overt focus movement in the two former constructions but not in the no-da construction. An indication of the topicalization of FocP in cleft construction is the resulting word order, as wel as the topic marker on FocP. Note that if we do not adopt this analysis, it is very dificult to capture the similarity and diferences among these constructions. 133 A Modification: The Treatment of Pronominal Sluicing Here I would like to argue that Hiraiwa and Ishihara?s analysis of sluicing requires some modification. In the earlier section, I showed that JPS is best analyzed as the elipsis of the presuppositional clause of cleft construction. Based on the observation that JPS and the cleft construction show similarities, it is natural to extend the analysis of clefting to JPS. Under the split-CP analysis that we have just outlined, the derivation of JPS is something like (132). (132) The Derivation of JPS a. Step 0 [ TopP [ FocP [ FinP [ TP NP-ga [ VP wh-o V]-tense]-no]-da] b. Step 1: Focus movement [ TopP [ FocP wh-o [ FinP [ TP NP-ga [ VP t NP-o V]-tense]-no]-da] c. Step 2: Topic movement of FinP [ TopP [ FinP [ TP NP-ga [ VP t NP-o V]-tense]-no]-wa [ FocP wh-o t FinP -da] 134 d. Step 3: Elipsis and Pronominalization 9 of FinP [ TopP [ FinP [ TP NP-ga [ VP t NP-o V]-tense]-no]-wa [ FocP wh-o t FinP -da] sore Unlike the derivation of sluicing that we have sen before, in this derivation, FinP undergoes topic movement before the elipsis takes place. The necesity of this step is suggested by the observation that the sore or the presuppositional clause of the cleft construction cannot stay in-situ. 9 Nakao and Yoshida asume a diferent derivation from the one I review here. For them a cleft is formed by a base generated focus element and nul operator movement inside the presupositional clause. They argue that sore is derived from the ?specific presupositional clause? with the genitive demonstrative so-no, with the structure of [ DP so-no [ CP ? ]. Adopting the nul operator analysis, they can sidestep the problem of specificity island violations by the focused phrase. Furthermore, they have a principled acount of pronominalization. Acording to them, if a CP deletion takes place, so-no is stranded. However, in Japanese, since the genitive demonstrative canot stand alone, it must be converted to the form of sore for language specific morphophonological reasons. If we asume that JPS is derived from the cleft construction, and if we adopt the split-CP analysis, Nakao and Yoshida?s aproach raises a problem: Focus movement is not constrained by the specificity island. If, on the other hand, we discard the analysis that sore is derived from the specific presupositional clause, then we lose a principled acount of pronominalization. In this study, however, I adopt the split-CP analysis, which has the potential to acount for not only the nature of focus constructions, but also that of nominal conditionals, and leave this problem open. I simply stipulate that FinP can be converted to sore. For more details of this theory of pronominalization, se Nakao and Yoshida (206). 135 (133) Quinn-ga nanika-o tabe-ta rasii ga? Q-nom something-acc eat-past seem but ?It sems that Quinn ate something, but?? a. Wh-Cleft Boku-wa [Quinn-ga tabe-ta no]-ga/wa I-top Q-nom eat-past comp-nom/top nani-o (da) ka what-acc cop Q wakara-nai. know-not ?I don?t know what it is that Quinn ate.? b. Wh-Cleft with scrambling *Boku-wa nani-o [Quinn-ga tabe-ta no]-ga/wa I-top what-acc Q-nom eat-past comp-nom/top (da) ka wakara-nai. cop Q know-not ?I don?t know what it is that Quinn ate.? c. JPS Boku-wa sore-ga/wa nani-o (da) ka wakara-nai. I-top it-nom/top what-acc cop Q know-not ?I don?t know what it is.? 136 d. JPS with scrambling *Boku-wa nani-o sore(-ga/wa) (da)ka wakara-nai. I-top what-acc it(-nom/wa) cop Q know-not ?I don?t know what it is.? Now the question is whether the step of topic movement before elipsis is specificaly necesary for JPS or is a more general requirement. There is no obvious reason to asume that JPS and sluicing are derived diferently, except for the necesity of pronominalization for JPS. Thus, I would like to propose that the elipsis of the FinP is only possible if it undergoes topic movement. I therefore asume that even in the standard sluicing construction, FinP is topic-moved before being elided. 3.3.3.7. The Analysis of Nominal Conditionals The stage is now set for our analysis of nominal conditionals in Japanese, after our necesarily long and complicated review of Rizi and Hiraiwa and Ishihara. What I wil show here is that exactly the same properties of clefts that Hiraiwa and Ishihara try to derive hold true for nominal conditionals, and that the split-CP system is the best tool to capture the properties of nominal conditionals. 3.3.3.7.1. Further Properties of Nominal Conditionals There are five properties of clefts that Hiraiwa and Ishihara try to derive from their analysis. They are summarized in (134). 137 (134) a. The possibility of multiple foci b. Island efects c. The complementizer no cannot be substituted with a pronoun/NP d. The unavailability of Nominative/Genitive Conversion (NGC) e. Clausemate condition on multiple foci In this subsection, I show that the same contrast sen betwen clefts and pseudoclefts can be sen in the underlying form of nominal conditionals, i.e., there is a bifurcation betwen case-marked and non-case-marked conditionals. In conditional constructions, the standard conditional in (135a) coresponds to the no-da construction, in which no word order permutation occurs. From this sentence, we can make a conditional sentence containing a cleft-like focus construction. Let us cal it the cleft conditional. We can understand this construction as the conditional version of the cleft construction. As the example (135b) indicates, a focused phrase without a Case marker is also a possible option. Let us cal this the pseudocleft conditional. (135) a. mosi [[[[Quinn-ga ringo-o Virginia-ni age-ta mosi Q-nom apple-acc V-dat give-past no] de-are/nara] ba]? comp cop cond ?If Quinn gave apple to Virginia?? 138 b. mosi [[[[Quinn-ga Virginia-ni age-ta mosi Q-nom V-dat give-past no]-ga ringo(-o) (mit-tu) de-are/nara] ba]? comp-nom apple-acc three-cl cop cond ?If it is (thre apples) that Quinn gave to Virginia?? Now, I wil show that this cleft conditional shows al the signature properties of clefts in (134). First, the cleft conditional alows multiple foci. On the other hand, the pseudocleft conditional does not alow multiple foci. (136) a. The Cleft Conditional mosi [Quinn-ga e 1 e 2 age-ta no]-ga mosi Q-nom give-past comp-nom Virginia-ni 1 ringo-o 2 (mit-tu) nara-ba. V-dat apple-acc three-cl cop cond ?(Lit.) If it is to Virginia, (thre) apples that Quinn gave?? b. The Pseudocleft Conditional mosi [Quinn-ga e 1 e 2 age-ta no]-ga mosi T-nom give-past comp-nom Virginia*(-ni) 1 ringo*(-o) 2 (mit-tu) nara-ba. V-dat apple*-acc three-cl cop cond ?(Lit.) If it is to Virginia, (thre) apples that Quinn gave?? 139 Second, the cleft conditional is sensitive to island constraints. Thus, in (137) the focused phrase is extracted from a complex NP island, and the sentence is severely degraded. In the pseudocleft conditional, on the other hand, extraction out of a complex NP is possible. (137) a. mosi [Quinn-ga [ RC [Virginia-ga e 1 ringo-o mosi Q-nom V-nom apple-acc age-ta] hito 1 ]-ni at-ta nara ba] give-past person-dat meet-past cop cond ?If Quinn mets the man who Virginia gave the apple to?? b. *mosi [[Quinn-ga [ RC [Virginia-ga e 1 e 2 age-ta] mosi Q-nom V-nom give-past hito 1 ]-ni at-ta] -no] -ga ringo-o 2 person-dat meet-past-comp-nom apple-acc (mit-tu) nara-ba? three-cl cop-cond ?(Lit.) If it is (thre) apples that Quinn met the man who Virginia gave to?? 140 c. mosi [[Quinn-ga [ RC [Virginia-ga e 1 e 2 age-ta] mosi Q-nom V-nom give-past hito 1 ]-ni at-ta] -no] -ga ringo 2 (mit-tu) person-dat meet-past-comp-nom apple-? three-cl cop-cond nara-ba? ?(Lit.) If it is (three) apples that Quin met the man who Virginia gave to?? Third, the complementizer no cannot be substituted with an NP in the cleft conditional, but it is possible in the pseudocleft conditional. (138) a. Cleft Conditional mosi [Quinn-ga tabe-ta no/*mono/*kudamono]-ga mosi Q-nom eat-past comp/thing/fruit-nom ringo-o (mit-tu) nara-ba? apple-acc three-cl cop-cond? ?If it was (thre) apples that Quinn ate?? b. Pseudocleft Conditional mosi [Quinn-ga tabe-ta no/mono/kudamono]-ga mosi Q-nom eat-past comp/thing/fruit-nom ringo (mit-tu) nara ba? apple-? three-cl cop cond? ?If it was (thre) apples that Quinn ate?? 141 The fourth property is the possibility of Nominative/Genitive Conversion. The cleft conditional, like its cleft counterparts, resists nominative/genitive conversion, in contrast to the pseudocleft conditional. (139) a. Cleft Conditional mosi [Quinn-ga/*no tabe-ta no]-ga ringo-o mosi Q-nom/gen eat-past comp-nom apple-acc (mit-tu) nara-ba? three-cl cop-cond ?If it was (thre) apples that Quinn ate?? b. Pseudoleft Conditional mosi [Quinn-ga/no tabe-ta no]-ga ringo mosi Q-nom/gen eat-past comp-nom apple-? (mit-tu) nara-ba? three-cl cop-cond ?If it was (thre) apples that Quinn ate?? Finaly, the cleft conditional with multiple foci is sensitive to the clausemate condition. 142 (140) a. mosi [Virginia-ga sensee-ni [Quinn-ga ringo-o mosi V-nom teacher-dat Q-nom apple-acc tabe-ta-to] iituke-ta nara-ba] eat-past-comp tell-past cop-cond ?If Vrirginia tels the teacher that Quinn ate the apple?? b. *mosi [Viriginia-ga e 1 [Quinn-ga e 2 tabe-ta-t mosi V-nom Q-nom eat-past-comp iituke-ta no]-ga sensee-ni 1 rigo-o 2 tell-past comp-nom teacher-dat apple-acc (mit-tu) nara-ba three-cl cop-cond ?(Lit.) If it was to the teacher, thre apple that Virginia told that Quinn ate? ? Looking at these five properties, it is clear that cleft and the cleft conditional are quite similar creatures. Now let us turn to the nominal conditional. Our expectation here is that the nominal conditional corresponds to sluicing in Hiraiwa and Ishihara?s paradigm. To test this point, what I have to show is whether the contrast betwen sluicing and pseudosluicing can be sen betwen the nominal conditional with a remnant bearing case-markers and the one with a remnant without case-markers. What Hiraiwa and Ishihara show was the following two properties: sluicing alows multiple foci but pseudosluicing does not; sluicing shows island efects but pseudosluicing does not. Nominal conditionals, indeed exhibit exactly the same contrast as sluicing and 143 pseudosluicing, as the following examples show. Example (141) shows that the case- marked nominal conditional alows multiple foci but pseudocleft conditionals do not. (141) Multiple Foci Quinn-ga dareka-ni nanika-o age-ta rasii ga? Q-nom someone-dat something-acc give-past seem but ?It sems that Quinn gave something to someone, but? a. Case Marked Nominal Conditional (mosi sore-ga) Virginia-ni ringo-o (mit-tu) mosi it-nom V-dat apple-acc three-cl nara-ba? cop-cond ?(Lit.) if it was (thre) apples to Virginia ?? b. Non Case Marked Nominal Conditional (mosi sore-ga) Virginia ringo (mit-tu) mosi it-nom V-? apple-? three-cl nara-ba? cop-cond ?(Lit.) if it was (thre) apples to Virginia ?? Example (142) shows that the case-marked nominal conditional is sensitive to the complex NP island, but the non-case-marked nominal conditional is not. This, in turn, suggests that movement is not involved in the non-case-marked nominal conditional. 144 (142) Island Sensitivity Quinn-wa [[otooto-ni nanika-o okuttakita]hito]-o Q-nom brother-dat something-acc send person-acc syootai-sita rasii ga? invite-past seem but? ?It sems that Quinn invited a person who had sent something to his brother, but?? a. Case Marked Nominal Conditionals *(mosi sore-ga) ringo-o (mit-tu) nara-ba mosi it-nom apple-acc three-cl cop-cond ?If it was (thre) apples?? b. Non Case Marked Nominal Conditionals (mosi sore-ga) ringo-o (mit-tu) nara-ba mosi it-nom apple-? three-cl cop-cond ?If it was (thre) apples?? Summarizing the discussion so far, the conditional clauses reviewed in this subsection show significant paralelism with the no-da, cleft, and sluicing constructions that Hiraiwa and Ishihara examined. Furthermore, we have sen that JPS and the pronominal nominal conditional also show significant similarities. From these observations, it is quite plausible to conclude that they are basicaly the same type of construction, and thus should be analyzed in the same fashion. 145 3.3.3.7.2. The Derivation of Nominal Conditionals The basic line of analysis of the nominal conditional that I wil outline from this point is the same as the analysis of sluicing that is adopted by Hiraiwa and Ishihara. However, I have not yet described the internal makeup of conditional clauses under the split-CP system. Let us start from this point. From the discussion so far, it is easy to se that many parts of the complex verbal morphology of conditional clauses are shared by the no-da construction, even though some of them have diferent surface forms. Simply put, the right edge of conditional clauses is the same as that of cleft constructions, except for the following two points: first, the top nodes of these constructions are occupied by diferent elements; second, the morphemes that compose conditional morphology can be omited relatively frely. Compared with the no-da construction, the structure of conditionals can be ilustrated as in (143b). In the embedded context, the no-da construction must be folowed by complementizers like -to if it is declarative or -ka if it is interogative. On the other hand, in conditionals, the rightmost element is -ba and it sems that this -ba marks the clause type of a conditional, even though it can be omited. (143) a. No-da Constructions [ ForceP [ TopP [ FocP [ FinP [ IP [ VP ?V]-tense]-no]-da]-to/ka] b. Conditional [ ForceP [ TopP [ FocP [ FinP [ IP [ VP ? V]-tense](-no)]-(na)-ra/de-are](-ba)] 146 Based on the optionality of some of the morphemes, conditional verbs can have seven possible surface forms. (144) shows al seven possibilities with the verb taberu (eat). These examples show how the omision of conditional morphemes can take place. The most radical patern is (144g) where al the morphemes are omited except the force marker -ba. (144) a. tabe-ta-no-de-are-ba b. tabe-ta-ra-ba c. tabe-ta-ra d. tabe-ta-no-nara-ba e. tabe-ta-nara-ba f. tabe-ta-nara g. taber-e-ba A note on the traditional treatment of conditional clauses is in order. In the paradigm in (144) we can se thre of the four paterns of conditional forms that are traditionaly treated as diferent expresions: re-ba conditionals, ta-ra conditionals, na-ra conditionals respectively. Although there are some subtle semantic or pragmatic diferences among these thre forms of conditionals, in Japanese traditional gramar these are arbitrary distinctions. Under the traditional treatment, we cannot capture the intuition that they can al have the same basic meaning and, above al, a patern like (144a) where al the morphemes are expresed cannot be captured. With the structure in (143b) in mind, let us se the derivation of nominal conditionals. As in the derivation of sluicing, the nominal conditional results from the 147 base sentence through the cleft conditional structure and elipsis of FinP. (145) ilustrates the stepwise derivation. (145) a. Step 0 [ ForceP [ TopP [ FocP [ FinP [ IP NP-ga [ VP NP-o V]-tense]-no]-nara]-ba] b. Step 1: Focus Movement [ ForceP [ TopP [ FocP NP-o [ FinP [ IP NP-ga [ VP t NP-o V]-tense]-no]-nara]-ba] c. Step 2: Topic Movement of FinP [ ForceP [ TopP [ FinP [ IP NP-ga [ VP t NP-o V]-tense]-no]-ga [ FocP NP-o t FinP - nara]-ba] d. Step 3: Elipsis/Pronominalization of FinP [ ForceP [ TopP [ FinP [ IP NP-ga [ VP t NP-o V]-tense]-no]-ga [ FocP NP-o t FinP - nara]-ba] sore-ga Under this analysis, we can succesfully capture the similarities betwen conditional constructions and focus constructions. As in the case of focus constructions, the standard conditional, the cleft conditional and the nominal conditional show similarities because they are derived from the same underlying structure. Furthermore, we can derive the paralelisms betwen JPS and nominal conditionals with sore. Both of them are residues of FinP elipsis. The crucial 148 diferences betwen focus constructions and conditional constructions are their surface verbal morphology, and Force 0 elements. 3.3.3.7.3. The Licensing Condition on Mosi The discussion up to this point has made clear the nature of sore in the nominal condition. In our analysis sore replaces FinP. In the earlier discussion, I have shown that the CA mosi in the nominal conditional is legitimate only when sore is present. Now we have a beter understanding of the licensing condition of mosi. Under the proposed analysis mosi is licensed by FinP. If we asume the notion of Criterion adopted by Rizi, mosi is licensed by Fin 0 through a Spec-Head relation. 3.3.4. Conclusion In this section, I have revealed two important aspects of conditional clauses in Japanese. One is the internal syntax of the clause, and the other is the distributional properties of mosi. What we have specificaly sen is that there is a tight connection betwen Japanese focus constructions and conditional constructions, and this connection gives us a clue to understanding the distributional properties of mosi, i.e., mosi is licensed by FinP. In the following sections, I wil try to show why in the Japanese conditional clause an adjunct clause is not a strong island but rather is a weak island. The discussion here wil give us an important insight into understanding the weak- islandhood of conditional clauses. Below, I wil show that mosi acts like a focus marker. Based on this observation, I wil argue that mosi creates a so-caled 149 intervention efect. To capture this intervention efect, I wil argue that mosi is generated in the specifier of FinP and induces a Relativized Minimality efect (Rizi 1990). 3.4. The Question of Mosi and Weak Island Efects In the earlier discussion, we have sen that the Conditional Adverb (CA) mosi is an inducer of Weak Island (WI) efects. I now turn to the question of why mosi induces WI efects. Answering this question wil lead us to a beter understanding of mosi and CAs in general. The claim that I wil establish in the course of this discussion is that mosi is a conditional scope marker. 3.4.1. Weak Island Inducers To answer the question, it is worth thinking about what kind of elements induce WIs, and what properties they share so that we can test whether mosi is one of these elements or it is a totaly diferent element. In previous studies on WIs, there is consensus among researchers that certain quantificational elements tend to induce WIs (Beck 1996; Hagstrom 1998; Harada 1972; Harada 1973b; Honcoop 1998; Hornstein and Uriagereka 2002; Kim 2002; Miyagawa 2002; Pesetsky 2000; Rizi 1990; Ross 1984; Rullmann 1995; Tanaka 2003a among many others). The WI efects induced by these quantificational elements are sometimes caled intervention efects (Beck 1996 among many others). As the name suggests, the basic patern of the WIs induced by quantificational elements can be ilustrated as that a harmful 150 quantifier intervenes betwen the links of an operator-variable chain. There are various analyses proposed in the literature, but basicaly they al try to capture this basic patern. As an example of the intervention efects, it has been widely observed that negation induces an island efect (Beck 1996; Harada 1972; Linebarger 1987; Rizi 1990; Ross 1984; Rullmann 1995). The examples in (146) are from English, showing that negation blocks the overt movement of non-referential wh-elements. (147) summarizes examples from German, showing that negation blocks covert movement. (147b) suggests that just containing a negation does not make a sentence bad. Rather, if the in-situ wh-phrase is c-commanded by the negation, the sentence becomes bad. Importantly the two sentences in (147) have basicaly the same meaning. (146) a. Who don?t you think Quinn saw? b. *How don?t you think Quinn behaved? (147) a. Wer hat wo niemanden agetroffen? who has where nobody met ?Who didn?t met anybody where?? b. *Wer hat niemanden wo angetroffen? Who has nobody where met ?Who didn?t met anybody where?? As we have sen in the earlier discussion, intervention efects have also been observed in Japanese (Harada 1972; Hoji 1985; Miyagawa 2002; Takahashi 1990; 151 Tanaka 2003b). Miyagawa, for example, observes that, like in other languages, scope- bearing elements block movement. Aside from the examples that we have sen already, universal quantifiers and existential quantifiers, as wel as a certain clas of focus particles, are known to induce intervention efects. Let us se some of these examples. The examples in (148) show that universal quantifiers in Japanese block covert wh-movement. (148) a. ? Dono hito-mo nani-o kat-ta-no? every person-foc what-acc buy-past-Q ?What did every person buy?? b. ? *Hotondo dono hito-mo nani-o kat-ta-no? almost every person-foc what-acc buy-past-Q ?What did almost every person buy?? c. Quinn-ga nani-o kat-ta-no? Q-nom what-acc buy-past-Q ?What did Quinn buy?? Having sen these examples, we can conclude that Quantificational Elements are an inducer of intervention efects. In addition, based upon observations of a wide range of intervention/WI efects in Japanese, Miyagawa (2002) concludes that al WIs should be analyzed as being induced by Quantifiers. In other words, WIs in Japanese are Quantifier Induced Bariers (QIBs). This position is, however, controversial given the current debates on WIs among researchers (Szabolcsi and den Dikken 1999 152 Tomioka 2005). 10 However, it sems to be true that WI inducers are always quantificational elements. Thus, I wil basicaly asume that WIs are induced by quantificational elements. 3.4.2. Mosi as a Scope Marker Given the above discussion, the question naturaly arises whether the WI inducer mosi is also a quantificational element or not. In the following discussion, I wil show that there are some reasons to believe that mosi is a quantificational element. 3.4.2.1.The Basic Observations One of the prominent properties of mosi is its fredom of location inside a conditional clause. For example, in an example like (149), mosi can appear at any of the underlined positions. (149) Quinn-wa [ Virginia-ga Stillman-ni tegami-o Q-top V-nom S-dat letter dasita-ra] naki-dasu daroo. send-cond cry-start will ?Quinn wil start crying if Virginia send a leter to Stilman.? 10 Tomioka, for example, sugests that WIs should not be explained purely by syntax. He argues for a semantic/pragmatic aproach to WIs. Szabolcsi also argues for a semantic acount of WIs. 153 It is also worth noting that mosi cannot be licensed outside of a conditional clause. For instance, in (150) the conditional clause is fronted to the beginning of the sentence, and mosi is left outside of it. A mosi that is stranded out of a conditional clause normaly creates a wild unaceptability. Thus, mosi?s distribution is fre but not totaly fre. It is fre inside a conditional clause. As we have sen in the earlier discussion, mosi must be licensed by a conditional clause, and thus must be inside the licensing conditional clause. (150) *[Virginia-ga Stillman-ni tegami-o dasita-ra] V-nom S-dat letter-acc send-cond Quinn-wa mosi naki-dasu daroo. Q-nom mosi cry-star will Intended interpretation: ?Quinn wil start crying if Virginia sends a leter to Stilman.? Note also that mosi must be licensed by a conditional clause, but must be licensed by a clause-mate conditional marker. Let?s take a look at the following pair of sentences. (151b) is severely degraded compared to (151a). In (151b) mosi is not in the same clause as the conditional marker -ra. On the other hand, in (151a), they are in the same clause. 154 (151) a. Quinn-wa [Virginia-ga sono-ringo-o mosi Q-top V-nom that-apple-acc mosi tabeta-ra] okori-dasu daroo. eat-cond get-angry-start will ?Quinn wil get angry if Virginia eats that apple.? b. *Quinn-wa [Stillman-ga [Virginia-ga Q-top S-nom V-nom sono-ringo-o mosi tabeta-to] itta-ra] that-apple-acc mosi eat-comp say-cond okori-dasu daroo. get-angry-start will ?Quinn wil get angry if Stilman says that Virginia ate that apple.? This contrast, thus, suggests that mosi must be licensed by a clause-mate conditional. 1 1 The clause-mate condition, however, can be relaxed in certain environments. For example, when so- caled bridge expresions (Kato 1985) intervene betwen mosi and the conditional complementizer, the clause-mate condition can be violated as ilustrated in (i). (i) Virginia-ga sono hon-o mosi yonda-koto-ga are-ba? V-nom that bok-ac mosi read-fact-nom exist-cond ?If Virginia has an experience of reading that bok?? Interestingly bridge expresions can also mediate the relation betwen negation and negative polarity items in Japanese such as -sika, ?only?, which also is subject to the clause-mate condition in a normal situation (Kato 1985). (i) a. Quin-wa ringo-sika tabe-nakat-ta. 155 There is another interesting property of mosi that is related to its relative fredom of positioning, namely, there are some diferences in interpretation of elements that are under and outside the scope of mosi. 12 Let us take examples containing a numeral quantifier. Al the examples in (152) share basicaly the same meaning. However, in the examples containing mosi, depending on mosi?s position, the interpretation is diferent. The diference is related to what is focused in the conditional clause. In (152b) both gakuse ?student?, the quantifier or action of eating can be focused as in (152a), which does not contain mosi. On the other hand, in (152c), in its most natural interpretation, gakuse is not focused, but the quantifier or the action can be focused. Thus (152c) can be paraphrased using clefts such as (152c.i) or (152c.i). They are most naturaly available but (152c.ii) is somewhat Q-top aple-only eat-not-past ?It was only aples that Quin ate.? b. *Quin-wa [Virginia-ga ringo-sika tabe-ta-to]iwa-naka-ta. Q-top V-nom aple-only eat-past-comp say-neg-past. ?Quin said it was only aples that Virginia ate.? c. Taro-wa [ringo-sika tabe-ta-koto]-ga nai. T-top aple-only eat-past-fact-nom not ?It is only aples that Taro has an experience of eating.? Note that sika is also an example of a focus marker. The similarity betwen mosi and sika can also suport the claim that mosi is a focus marker. However, it is not at al clear in exactly which contexts the clause-mate condition can be relaxed. Tomohiro Fuji (p.c.) sugests that in environments that resemble raising or control constructions (so-caled restructuring environment), the clause-mate condition can be relaxed. In this study, however, I wil leave this point open. 12 This was sugested by Hiromu Sakai (p.c.). I would like to thank him for directing my atention to this type of phenomenon. 156 degraded. A tricky part of this is that if one puts a heavy focus stres on gakuse, the interpretation (152c.ii) becomes readily available. 13 However, the most natural interpretations is (152c.i) or (152c.i) depending on the focus intonation inside the scope of mosi. We can support this point by using negation. In (152b) we can negate any of the elements to the right of mosi. However, it is odd if someone says, ?no it is teachers not students? after an utterance like (152c). This oddity does not arise in the case where someone negates the elements to the right of mosi as in ?it is 5 not 3? or ?it is if they leave not if they come?. These interpretive diferences suggest that elements on the right-hand side of mosi are readily counted as the member of alternative set in the sense of Rooth (1985). Note that this negation test also can be side stepped by manipulating the intonation. However, the important point for us is 13 This might be related to a widely acknowledged phenomena, the expansion of the domain of focus (Chomsky 1972; Jackendof 1972 among many others). For example, in English, if emphatic stres fals on the object, the sentence can be interpreted in several diferent ways. This is ilustrated in the folowing example. In (i) the emphatic stres fals on the object NP the car but other elements than the object, including the whole sentence, can be focused. (i) John fixed THE CAR a. What did Quin fix? b. What did Quin do? c. What hapened? Taking this expansion of focus as a general phenomenon, it is not surprising that something on the left of mosi or the whole clause containing the focus marker can be interpreted as focused as wel as the ones on the right. The same focus expansion can, inded, be sen in examples containing other focus particles in Japanese (se Aoyagi 198; Kuroda 1965 among others for detailed discusion on this point). 157 even though the efect is not so strong, stil there are diferences in the most natural interpretations depending on the position of mosi. Exactly the same patern holds true for (152d). Here, only the verb is on the right of mosi, and thus only the action described by the verb can be focused. (152) a. gakusee-ga 3-nin kita-ra ? student-nom 3-cl (human) come-cond ?If thre students come?? b. mosi gakusee-ga san-nin kita-ra mosi student-nom 3-cl (human) come-cond c. gakusee-ga mosi san-nin kita-ra student-nom mosi 3-cl (thing) come-cond i. ?If it is 3 (students) (rather than 4 or 5) that comes? ? i. ?If it is coming of 3 (students) that studends do ?? ii. ? ?If it is students that comes (thre of them)? d. ringo-o mit-tu mosi tabere-ba apple-acc 3-cl (thing) mosi eat-cond i. ?If it is 3 (apples) (rather than 4 or 5) that you eat? ? i. ? ?If it is the eating of 3 (apples) that you do ?? ii. ? ?If it is apples that you eat (thre of them)? Summarizing the discussion so far, we have sen that the position of mosi afects the interpretation of a conditional sentence. The basic patern is that elements 158 on the right hand side of mosi can be easily focused, but those on the left hand side are not. 3.4.2.2. Disjunction Scope in English and Focus Particles in Japanese Now the question is how we can analyze the behavior of mosi. To gain clues for the analysis of mosi, I wil pick up an example from English that show similar behavior to mosi: the so-caled disjunctive scope marker either. Under the asumption that the conjunction or the disjunction is a scope-taking element (se Moltmann 1992; Munn 1993 among others for detailed discussion), there have been various studies on either in the literature (Larson, 1985; Schwartz 1999 among many others). Most previous studies on either have been devoted to capturing the correlation betwen its distribution and its interpretation. Let us summarize the most basic results of the previous studies. The most basic distributional property of either can be summarized as the following: either can occur in a clause that contains the disjunction or. More precisely, Either must be in a tensed clause that contains or but it can go outside of an infinitival clause containing or (Han and Romero 2004; Larson 1985; Schwarz 1999). This distributional patern is ilustrated in (153) and (154). (153) ilustrates the fact that either can be inside or outside of an infinitival clause containing disjunction. On the other hand, (154) clearly shows that it cannot be outside of the finite clause. (153) a. Quinn pretended [to either be looking for a burglar or a thief]. b. Quinn either pretended [to be looking for a burglar or a thief]. 159 (154) a. Quinn believes [that Stilman said [that Virginia was either drinking or playing video games]. b. Quinn believes [that Stilman said [that either Virginia was drinking or playing video games]. c. ? Quinn believes [that Stilman said either [that Virginia was drinking or playing video games]. d. ? Quinn believes [that Stilman either said [that Virginia was drinking or playing video games]. e. *Either Quinn believes [that Stilman said [that Virginia was drinking or playing video games]. Now let us turn to the interpretive aspects of either. Larson claims that either is a scope indicator for the disjunction. He cites the examples in (155) and (156). Larson points out that there are thre possible interpretations regarding the disjunction in (155) (Larson 1985; Rooth and Parte 1982). One interpretation that is represented in (155a) is the so-caled de dicto reading in which Mary is searching for a servant and would be satisfied with any individual x that mets the description ?x is a maid or x is a cook?. (Larson 1985: 218). Another reading of (155b) is the de re reading. Under the de re reading there is a particular individual who is either a maid or a cook such that Mary is seking that individual (Larson 1985:218). The third reading in (155c) is the following: Mary is looking for an individual x satisfying the description ?x is a maid? or else she is looking for any individual x satisfying the description ?x is a cook? (Larson 1985:218). 160 (155) Virginia is looking for a maid or a cook. a. Virginia is looking for (a maid) or (a cook). b. for some x, a maid or a cook, Virginia is looking for x. c. Virginia is looking for (a maid) or Virginia is looking for a cook. Larson points out that (156a) has al thre of the readings in (155). In (156b), on the other hand, the interpretation in (155a) is not available, i.e., or cannot take the wide scope. Also in examples (156c) and (156d), the narow scope readings become unavailable. (156) a. Virginia is looking for either a maid or a cook. b. Either Virginia is looking for a maid or a cook. c. Virginia is either looking for a maid or a cook. d. Virginia either is looking for a maid or a cook. The examples in (156) strongly suggest that the scope of or is afected by the position of either. If the either occurs adjacent to the disjunctive constituent, the narow scope reading becomes available. However, if the either is displaced from or, the narow scope interpretation becomes unavailable. Considering these interpretive aspects of either Larson concludes that it is a scope marker of disjunction. The crucial similarities betwen mosi and either are: (i) there is a corelation betwen their positioning and their interpretation, i.e., both elements sem to indicate the upper-bound of the scope of the operator, either or or conditional; and (i) they 161 have a local relation to their licensors, i.e., both of them are basicaly in a clause-mate relation to their licensors. Based on these similarities, it is plausible to understand that mosi has a similar function as either, i.e., marking the scope of its licensor. This, in turn, means that these elements are scope-bearing elements. Having sen the discussion so far, the real question is how we can capture the distribution of the scope markers. At this point, however, I would like to just asume the following two points. First, these scope markers are adverbs resembling the English focus particle even (Hendriks 2003), and their distribution of either is regulated by whatever principle regulates the distribution of even. I would like to leave the actual mechanism open, and just pursue the descriptive generalization. 3.4.3. Intervention Efects In the discussion so far, we have sen that mosi can be analyzed a scope- bearing element, more specificaly the scope marker of conditional operator. Now, we have a clue for why mosi induces WI efects. The simplest answer is that mosi is a quantificational element that becomes a harmful intervener if it appears betwen the moved phrase and its underlying position. In other words, mosi induces intervention efects in exactly the same way as other scope-bearing elements. 4. Aproaches to Adjunct (Non)Islands Thus far, we have established that conditional clauses in Japanese are not islands. A closer look at other adjunct domains suggests this claim is not limited to conditional clauses. Rather, it is generaly true for al adjunct clauses (for the related 162 discussion, se Saito & Fukui 1998; Ishii 1997; Mihara 1994 among others). Let us consider some examples. In the folowing, I present examples of Because Clauses and After Clauses as wel as RCs, Coordinate clauses and Complement clauses. 14 These are to be understood as baseline examples. (157) Because clauses Quinn-wa [Virginia-ga dono-gakusee-ni present-o Q-top V-nom which-student-dat present-acc ageta-node/-kara] nakidasita-no? gave-because/-because cried-Q? Lit. ?Which student does Quinn get angry because Virginia gave a present to?? a. Scrambling Dono-gakusee-ni 1 Quinn-wa [Virginia-ga t 1 which-student-dat Q-top V-nom present-o ageta-node/-kara] nakidasita-no? present-acc gave-because/-because cried-Q? 14 Examples of cordination are cited here because relativization in Japanese is normaly imune to island constraints except the Cordinate Structure Constraint, as we have sen in the discusion in section 2 of this chapter. Thus, examples of relativization in (1b) and (158b) should be compared with the example in (1b) and (1d). 163 b. Relativization [ NP [ CP Quinn-ga [Virginia-ga gap 1 present-o Q-nom V-nom present-acc ageta-node/-kara] nakidasita]gakusee 1 ] gave-because/-because cried student Lit. ?The student who Quinn cried because Virginia gave a present to.? (158) Temporal Clause: After Clause Quinn-wa [Virginia-ga dono-gakusee-ni present-o Q-top V-nom which-student-dat present-acc ageta ato] nakidasita-no? gave after cry-Q Lit. ?Which student did Quinn cry after Virginia gave a present to?? a. Scrambling out of After clause Dono-gakusee-ni 1 Quinn-wa [Virginia-ga t 1 which-student-dat T-top H-nom present-o ageta ato] nakidasita-no? present-acc gave after cry-Q b. Relativization out of After clause [ NP [ CP Quinn-ga[Virginia-ga t 1 present-o ageta Q-nom V-nom present-acc gave ato]nakidasita] gakusee 1 ] after cry student Lit. ?The student who Quinn cried after Virginia gave a present to? 164 (159) a. Scrambling out of Relative Clauses *Dono-gakusee-ni 1 Quinn-wa [ NP [ CP Virginia-ga t 1 which-student-dat Q-top V-nom t 2 ageta] present 2 ]-ga suki-nano? gave present-nom like-Q Lit. ?Which student does Quinn like the present that Virginia gave to?? b. Relativization out of Coordinations *Quinn-ga [ NP [ CP [ IP Virginia-ga [ VP [ VP Stillman-ni Q-nom V-nom S-dat hon-o age] [ VP t 1 [ VP present-o age]]-ta]] book-acc give present-acc give -past gakusee 1 ]-ni atta. student-dat met Lit. ?Quinn met the student who Virginia gave a book to Stilman and gave a present to.? c. Scrambling out of Complement Clauses Stillman-ni 1 Quinn-wa [Virginia-ga t 1 present-o S-dat Q-top V-nom present-acc ageta-to] omotteiru. gave-comp think ?Quinn thinks that Virginia gave a present to Stilman.? 165 d. Relativization out of Complement Clauses [ NP [ CP Quinn-ga [ CP Virginia-ga t 1 present-o Q-nom V-nom present-acc ageta-to] omotteiru] gakusee 1 ] gave-comp think student ?The student that Quinn thinks that Virginia gave a present to? The examples in (157) and (158) are as good as extraction out of complement clauses, as in (159c) and (159d). In particular, they sound much more aceptable when compared to extraction out of RCs or Coordination. Based on this comparison, we take it to be the case that extraction out of adjunct clauses is generaly good in Japanese. The aim of this section is to investigate why adjunct islands are not operative in Japanese. The first point I would like to make is that the non-islandhood of adjunct clauses is not a property specific to Japanese language. Instead, I show that at least two other languages (Korean and Malayalam) patern like Japanese. Furthermore, I show that English and Spanish also exhibit adjunct non-islands in some limited environments (Etxepare 1999, and Hornstein 2001). Based on these facts, I discuss two possible approaches to adjunct islands. The first approach suggests the possibility that the ?islandhood? of adjunct clauses is dependent upon the choice of predicates in the matrix clause and the positions of the adjunct clauses in a sentence. Etxepare (1999) and Hornstein (2001) show that in Spanish and English, extraction out of preposed adjunct clauses is generaly beter than extraction out of post-verbal adjunct clauses. We discuss whether the same holds true in Japanese. The second approach is 166 a parametric approach to adjunct islands. We pay atention to the fact that al the thre languages share a typological property, so-caled wh-in-situ. We discuss a possibility that adjunct non-islands can be derived from this particular property of these languages. In the following discussion, the argument is mainly based on data from conditional clauses, but esentialy the same holds true for other adjunct clauses. 4.1. Adjunct Islands and Left/Right Asymmetry 4.1.1. Spanish and English Etxepare (1999) points out that wh-phrases can be extracted from conditional clauses in Spanish if the conditional clause is in the complement position of a reporting verb, such as say, and if the conditional clause is located imediately after the complementizer. Conversely, if a conditional clause is in the complement of a verb like desire, interpret and mention, wh-extraction is not possible. He cites the following examples from Spanish. (160) a. Qu? libro 1 dijiste [que [si Ricardo leia alguna vez] abandonaria la Linguistica]? ?Which book did you say that if Ricardo ever read he would abandon linguistics?? b. *Qu? libro quieres [que [si algun le] abandone la linguistica? ?Which book do you desire that if anyone read, he would abandon linguistics?? 167 Hornstein (2001) points out that the same patern holds true in English. He cites the examples in (161). As evident in these examples, the possibility of extraction is afected by the type of matrix verb in the same way as in Spanish. (161) a. Which book 1 did you say [that [if Quinn ever read t 1 ] he would abandon linguistics]? b. *Which book 1 did you desire [that [if anyone read t 1 ] he would abandon linguistics]? Hornstein further showed that if the conditional clause is located after the embedded verb, extraction becomes impossible, even if the conditional clause is inside the complement clause of reporting verbs. (162) *Which book 1 did you say [that Quinn would abandon linguistics [if he ever read t 1 ]? Their acounts are based on the theory of sideward movement (Nunes 1995; Hornstein 1999, 2001). Etxepare and Hornstein argue that the patern of wh- extraction is expected under their theory of sideward movement. In this study, however, let us put the technical details aside, and investigate the descriptive generalization they formulated. Informaly, their findings can be summarized as the following. First, extraction is possible only if the matrix verb is a reporting verb, such as say. Second, extraction is possible only if a conditional clause is located higher 168 than the embedded subject. In their terms, if an if-clause is adjoined to the FP generated betwen the embedded CP and IP. Thus, as Hornstein (2001) points out, if the conditional clause is located at the post-verbal position, extraction is not possible; we cal this second generalization the Left/Right Asymmetry. 4.1.2. Scrambling from Conditionals and the Left/Right Asymmetry Let us now examine what the descriptive generalization can tel us about the adjunct non-islands in Japanese. The important point of the argument is if the two descriptive generalizations that Etxepare and Hornstein formulated are generaly true cross-linguisticaly, we would observe the same patern in languages other than Spanish or English. More specificaly, based on their descriptive generalizations, we predict that preposed adjunct clauses are islands if they are in the complement clause of non-reporting verbs, such as desire, mention or interpret. Furthermore, we predict that if the adjunct clauses are located post-verbaly, extraction is prohibited in general. Next, we examine these two points. First let us test whether the extraction out of adjunct clauses in Japanese are afected by the type of matrix verb. In order to test this, however, we must consider one caveat. The type of the complementizer that is selected by the non-reporting verbs that Etxepare mentions are not the same as the complementizer selected by say- type verbs in Japanese. Say-type reporting verbs in Japanese normaly select the declarative complementizer -to; however, non-reporting verbs, such as mention, interpret or desire, require the nominalizing complementizer -koto. Thus, we have to make sure that the nominalizing complementizer does not independently create island 169 efects. The examples in (163) show that scrambling out of koto-clauses does not create any detectable dificulty. Specificaly, comparing it with scrambling out of complement clauses or RCs, it becomes clear that it does not show island efects induced by RCs (for further details of koto-clauses se Uchibori 2000 and Watanabe 1996). (163) a. Scrambling out of Koto-clauses Dono-gakusee-ni 1 Quinn-wa [Virgina-ga t 1 which-student-dat Q-top V-nom present-o ageta-koto]-ni kizuita-no? present-acc gave-fact-dat realized-Q ?Which student did Quinn realize that Virginia gave a present to?? b. Scrambling out of Complement Clauses Dono-gakusee-ni 1 Quinn-wa [Virginia-ga t 1 which-student-dat Q-top V-nom present-o ageta-to] itta-no? present-acc gave-comp said-Q ?Which student Quinn said that Virginia gave a present to?? c. Scrambling out of Relative Clauses *Dono-gakusee-ni 1 Quinn-wa [ NP [ CP Virginia-ga t 1 which-student-dat Q-top V-nom ageta]present]-ga sukina-no gave present-nom like-Q Lit. ?Which student does Quinn like the present that Virginia gave to?? 170 We established that there is no problem with scrambling out of koto-clauses. Now, let us test the prediction of Etxepare-Hornstein generalization. The prediction is that extraction out of conditional clauses in the complement of reporting verbs is aceptable, but extraction from the complement of such verbs as mention, interpret and desire is degraded. The relevant Japanese examples are in (164). (164a) is an example of reporting verb, iu ?say.? (164b) is an example using the verb nozomu ?desire?. (164c) is example of scrambling out of multiply embedded complement clauses and (164d) is scrambling out of RCs, which are to be understood as the baseline example. (164) a. Reporting Verbs Dono-gakusee-ni 1 Quinn-wa [Stillman-ga which-student-dat Q-top S-nom [ cond mosi Virginia-ga t 1 present-o ageta-ra] mosi V-nom present-acc give-cond naki-dasu-daroo-to] itta-no? cry-start-will-comp said-Q ?Which student did Quinn say that if Virginia gave a present to Stilman would cry?? 171 b. Desire Dono-gakusee-ni 1 Quinn-wa [Stillman-ga which-student-dat Q-top S-nom [ cond mosi Virginia-ga t 1 present-o ageta-ra] mosi V-nom present-acc give-cond naki-dasu-koto]-o nozonde-iru-no? cry-start-fact-acc desire-be-Q ?Which student does Quinn desire that if Virginia gives a present to Stilman cries?? c. Multiply Embedded Complement Clause Dono-gakusee-ni 1 Quinn-wa [Stillman-ga which-student-dat Q-top S-nom [Virginia-ga t 1 present-o ageta-to] itta-to] V-nom present-acc gave-comp said-comp omotteiru-no? think-Q ?Which student does Quinn think that Stilman said that Virgia gave a present to?? 172 d. Multiply Embedded Relative Clause *Dono-gakusee-ni 1 Quinn-wa [Stillman-ga which-student-dat Q-top S-nom [ NP [ CP Virginia-ga t 1 t 2 ageta] present 2 ]-ga V-nom gave present-nom sukida-to] itta-no? like-comp said-Q Lit. ?Which student did Quinn say that Stil likes the present that Virginia gave to?? There are two points to be noted. First, because of the basic word order of Japanese, the type of examples that we are looking at contains center-embedding structures. Thus, we have to prepare an appropriate baseline that contains a multiply center- embedding construction such as (164c) or (164d). Because of the complexity of the center-embedded structure, it is more dificult to judge them than others. However, if we compare (164a and b) to an appropriate baseline, like (164d), we can se that extraction out of conditionals are much more aceptable than extraction out of RCs, and if compared to (164c), they are as good as extraction out of complement clauses. Second, because of the head-finality of Japanese, the complementizer comes at the end of each embedded clause, and we do not have any indicator of the left-edge of embedded clauses. Thus, to ensure that conditional clauses are located in higher positions they are located right below the embedded subject. We can locate conditional clauses to the left of the embedded subject too. However, in this case the sentence is ambiguous with respect to the position where the conditional clause is 173 interpreted. It can be either in the matrix clause or embedded clause. Stil, under the intended interpretations in which conditional clauses are interpreted in the embedded clause, the same judgements sem to come out. The relevant examples are presented in (165). However, we do not know if these are realy good test cases because of the ambiguity we noted. (165) a. Reporting Verbs Dono-gakusee-ni 1 Quinn-wa [[ cond mosi Virginia-ga which-student-dat Q-top mosi V-nom t 1 present-o ageta-ra] Stillman-ga present-acc give-cond S-nom naki-dasu-daroo-to] itta-no? cry-start-will-comp said-Q ?Which student did Quinn say that if Virginia gave a present to Stilman would cry?? b. Desire Dono-gakusee-ni 1 Quinn-wa [[ cond mosi which-student-dat Q-top mosi Virginia-ga t 1 present-o ageta-ra] V-nom present-acc give-cond Stillman-ga naki-dasu-koto]-o nozonde-iru-no? S-nom cry-start-fact-acc desire-be-Q ?Which student does Quinn desire that if Virginia gives a present to Stilman cries?? 174 These examples suggest the following two points. First, extractability is not determined by the type of matrix predicate. Thus, long-scrambling out of conditionals is aceptable either from the complement of reporting verbs or that of desire-type verbs. The comparison betwen the examples in (164) and (165) suggest this point. Aceptability of these examples, thus, disconfirms the first generalization of Etxepare and Hornstein. Second, the position of conditional clauses does not afect their extractability either. The examples in (164) may be counter examples against the second generalization such that conditional clauses must be located in a position high enough, where they are adjoined to FP. At this point, we do not have clear tests that can diagnose where conditional clauses are located in examples like (164a) or (164b). It is not clear if conditionals are adjoined to FP or not. This is because it is not clear where Nominative subjects are generated in Japanese (Fukui 1995; Kuroda 1988; Saito 1982: Takezawa 1987 among others). It may be in the specifier of VP or it may be in the specifier of IP. However, what is clear from the examples in (164) is that conditionals are located below the subject. This is in contrast to the English and Spanish examples. Thus, it is plausible to asume that they are adjoined to VP rather than a projection higher than IP. Consequently, extractability in these examples does not straightforwardly follow from the Etxepare-Hornstein generalization, which states that an element can be extracted from a conditional clause that is adjoined to the FP. However, the data from Japanese are not necesarily counterexamples to the cross-linguistic generalization of Etxepare-Hornstein. Rather, it is possible that the aceptability of scrambling out of conditionals is atributable to some special properties of Japanese, such as head-finality. Thus, to test the cross-language 175 generality of the generalization, we have to check other languages that show adjunct island efects. We return to this point later. Let us turn to the second generalization, the Left/Right Asymmetry. The important point that Hornstein is raising is if a conditional clause is located to the right of the verb, extraction is not possible. The relevant minimal pair is presented in (166). (166) a. Which book 1 did you say [that [if he ever read t 1 ] Quinn would abandon linguistics]. b. *Which book 1 did you say [that Quinn would abandon linguistics [if he ever read t 1 ]. We can capture the contrast in (166) in two ways. One is that extraction is not alowed because the conditional clause is located to the right of the verb. The other is that extraction is not alowed because the adjunct is adjoined to VP rather than IP. At this point, it is dificult to test these two hypotheses. It is not clear whether they make diferent predictions. Simply taken, it as if the adjunct is located right of the verb, extraction is not possible. What are the predictions of the hypotheses above when extended to Japanese adjunct clauses? We predict that an element cannot be extracted out of an adjunct clause if it is located to the right of the verb. On the other hand, if it is located to the left of the verb, extraction wil be alowed. The problem is whether Japanese has such an operation that can locate an adjunct clause to the right of the verb. One possible 176 test case would be the ?so-caled? right-dislocation construction. In right-dislocation, we can locate constituents to the right of the verb. In any of the other constructions, it does not sem to be possible to generate constituents to the right of the verb. Let us examine some examples of right-dislocation. (167b) is an example of right-dislocated complement clauses. 15 A prominent property of this construction in Japanese is that a special particle is atached to the matrix verb. If right-dislocation is applied, the special particle, so-caled shuu-josi ?end particle? is atached to the matrix verb. (167) a. Quinn-wa [Virginia-ga Stillman-ni present-o Q-top V-nom S-dat present-acc ageta-to] itta. gave-comp said ?Quinn said that Virginia gave a present to Stilman.? b. Quinn-ga __ 1 itta-yo, [Virginia-ga Still-ni present-o said-part ageta-to] 1 . As we have sen in various places, complement clauses normaly alow scrambling. However, once right-dislocation applies, scrambling is no longer licit. The minimal pair in (168) clearly shows this point. 15 The original position of the right-dislocated constituent is indicated by the underscore _. 177 (168) a. present-o 1 Quinn-wa [Virginia-ga Stillman-ni t 1 present-acc Q-top V-nom S-dat ageta-to] itta. 16 gave-comp said ?Quinn said that Virginia gave a present to Stilman.? b. *present-o 1 Quinn-wa __ 2 itta-yo, [Virginia-ga Stillman-ni t 1 ageta-to] 2 . In (168b), the acusative NP is extracted out of the right-dislocated complement clause, and the sentence is unaceptable. This unaceptability is surprising given that, in Japanese, even moved constituents are not islands (Kikuchi et al. 1994) as the examples in (169) show. In (169b) the acusative NP is scrambled out of the scrambled complement clause. Both examples in (169) are aceptable, but (169b) is even easier to interpret because it avoids the interfering efect from multiple center- embedding. 16 In this example, I am using a sentence containing the scrambling of an acusative NP. This is simply because the dative NP can be interpreted as the matrix argument, and thus may create unecesary ambiguity. However, scrambling of the dative argument does not make any diference (in aceptability?). To kep the examples in this study as similar as posible, I continue using dative verbs in the embeded clause. 178 (169) a. present-o 1 Quinn-wa [ CP Stillman-ga present-acc Q-top S-nom [ CP Virginia-ga otooto-ni t 1 ageteta-to] V-dat brother-dat gave-comp itta-to] omotteiru. said-comp think ?Quinn thinks that Stilman said that Virginia gave a present to her brother.? b. present-o 1 Quinn-wa [ CP [ CP Virginia-ga t 1 ageta-to] 2 Stillman-ga t 2 itta-to] omotteiru. The sharp contrast betwen (169b) and (168b) indicates that something special is going on in the right-dislocation construction. Turning our atention to adjunct clauses, exactly the same patern holds. As the examples in (170) indicate, scrambling out of conditional clauses in their original position is much more aceptable than that out of the right-dislocated conditionals. 179 (170) a. Quinn-wa [Stillman-ga __ 1 naki-dasu-daroo-to] Q-top S-nom cry-start-will-comp itta-yo, [mosi Virginia-ga sono-gakusee-ni said-part mosi V-nom that-student-dat present-o ageta-ra] 1 (-ne). present-acc give-cond(-part) ?Quinn said that Stil wil cry if Virginia gives a present to that student.? b. *present-o 1 Quinn-wa [Stillman-ga __ 2 naki-dasu-daroo-to] itta-yo, [mosi Virginia-ga sono-gakusee-ni t 1 ageta-ra] 2 (-ne) c. present-o 1 Quinn-wa [Stillman-ga [mosi present-acc T-top J-nom mosi Virgnia-ga sono-gakusee-ni t 1 ageta-ra] V-nom that-student-dat gave-cond naki-dasu-daroo-to] itta. cry-start-will-comp said Given these examples showing that the right-dislocated constituents are islands, it sems that the simple version of the Left/Right Asymmetry is supported. However, the story is not so simple. As we have sen, it sems something special is going on in the examples of right-dislocation. Therefore, in order to draw any conclusions, we have to investigate the syntax of right-dislocation and confirm that 180 there is no special factor that can independently block extraction out of dislocated constituents. A possible acount for the islandhood of the dislocated phrase is the following. Extraction is not alowed because the dislocated constituent is not part of the sentence. 17 Let us examine this possibility. The intuition behind this acount is that the right-dislocated constituent is a special element that is not a part of the structure of the other part of the sentence. Put diferently, it is disconnected from the matrix clause. Theoreticaly, this intuition might be captured by postulating that the right-dislocated structure has its own root in a sentence, in the a manner similar to the ?so-caled? parenthetical constructions under certain analyses (McCawley 1982). 18 This analysis alows us to analyze right- dislocation constructions in the same way as parenthetical constructions. A possible acount for the islandhood of the dislocated phrase is the following. If the right- dislocated constituent has its own root, then movement out of it means an element is moved out of a ?sentence? to the beginning of another ?sentence?. As far as I know, such inter-sentential movement operation is not alowed. Thus, acording to this acount, extraction out of the dislocated phrase is also not alowed. There are, however, at least two problems with this approach. First, it is not clear whether Japanese has parenthetical constructions. We do not have any good diagnoses that can distinguish parentheticals from the other constructions. Thus, we 17 This posibility is sugested by Ilhan Cagri. I thank her for leading my atention to this posibility. 18 For further details of multiple roted structures and shared constituency, se Citko (205), Moltman (192), Wilder (199) among others. Guimar?es (204) also contains a detailed discusion on this point. 181 cannot test whether they are similar or not. Second, there is evidence against a structural disconnection betwen the main clause constituents and right-dislocated constituents. One such argument comes from binding connectivity and subjacency efects and another from an example of discontinuous constituency. It is reported in the literature that right-dislocated phrases show binding connectivity efects and subjacency efects. Both are typical of constructions involving movement. Tanaka (2001) shows various paterns of connectivity efects exhibited by right-dislocation constructions. Examine the following pair of examples. (171) is the case of Binding Condition C. The right-dislocated constituent shows a binding connectivity efect in (171c) in which the R-expresion in the relative clause is bound by the right-dislocated subject pronoun. The example is unaceptable. (171) a. *Kare 1 -ga [ NP [ CP Virgnina-ga Quinn 1 -ni okutta] he-nom V-nom Q-dat sent tegami]-o mada yonde-inai. letter-acc yet read-neg ?He 1 has not yet read the leter that Virginia sent to Quinn 1 .? b. [ NP [ CP Virginia-ga Quinn 1 -ni okutta]tegami]-o 2 V-nom Q-dat sent letter-acc kare 1 -ga t 2 mada yonde-inai. he-nom yet read-not 182 c. *__ 1 [ NP [ CP Virginia-ga Quinn 2 -ni okutta] V-nom Q-dat sent tegami]-o 3 mada yonde-inai-yo, kare 2 -ga 1 . letter-acc yet read-neg-part he-nom This connectivity efect, as wel as other connectivity efects that Tanaka is citing, potentialy shows that the dislocated element is moved rightward, and holds a syntactic relation with the matrix clause. Second, right-dislocation shows a certain locality efect that can be presumably subsumed under the Subjacency Condition. (172c) exhibits severe unaceptability. This paralels with scrambling out of RCs in (172b). As opposed to right-dislocation out of RCs, extraction out of complement clauses does not show such degradation. (172) a. Quinn-ga [ NP [ CP Virginia-ga Stillman-ni Q-nom V-nom S-dat ageta]hon]-o nusunda. give book-acc stole. ?Quinn stole the book that Virginia gave to Stilman? b. *Stillman-ni 1 Quinn-ga [ NP [ CP Virginia-ga t 1 ageta]hon]-o nusunda. c. ? *Quinn-ga [ NP [ CP Virginia-ga __ 1 ageta]hon]-o nusunda-yo, Bill-ni 1 . 183 d. Quinn-ga [Virginia-ga __ 1 hon-o ageta-to] Q-nom V-nom book-acc gave-comp itta-yo, Stillman-ni 1 . said-part S-dat ?John said that Mary gave a book to Bil.? These examples potentialy show that dislocated phrases are moved rightward. However, Tanaka argues that there is another explanation. He argues that right- dislocation is a special case of scrambling with elipsis under identity. Acording to him, there is an elided structure that maintains the paralelism with the matrix clause. The dislocated element is actualy scrambled inside the hidden clausal structure. Thus, Tanaka argues that right-dislocation and scrambling show close paralelisms, and these locality and connectivity efects are induced by scrambling rather than rightward movement. Although Tanaka does not touch on the relation betwen the matrix clause and dislocated phrases 19 and does not argue against the structural 19 Endo (196) argues that right-dislocation constructions are derived from cordinated clauses via movement of the dislocated constituent and deletion. Thus, his analysis is basicaly similar to the one that Tanaka is exploring. In his (whose?) analysis, the relation betwen the matrix clause and the dislocated constituent is clear. However, because his analysis is apealing to a scrambling-like leftward movement of the dislocated constituent, it wil face the same problem as Tanaka?s as we wil se shortly. However, if we folow Endo?s aproach and asume that right-dislocation constructions have the structure of Cordination, it is posible to analyze the islandhod of the dislocated phrases as a violation of the Cordinate Structure Constraint. An obvious prediction of this aproach is if we extract elements in an acros-the-board fashion, extraction is alowed. Although this is potentialy an 184 disconnection either, this approach is compatible with the structuraly disconnected analysis of right-dislocation. Under this analysis, the structure of a right-dislocated sentence looks something like (173). (173) [ IP NP __ 1 V]-yo, [ IP NP 1 [ IP NP t 1 V]. A closer look at the examples of right-dislocation, however, reveals that this approach is not perfect, although it can capture many of the basic properties of the construction. A potential counter-argument against the scrambling approach comes from the possibility of discontinuous constituency. The example in (174b) suggests that right-dislocation can disconnect a relative clause from its head noun. 20 (174) a. Quinn-ga [ NP [ CP Virginia-ga katta] hon]-o Q-nom V-nom bought book-acc yonda. read ?Quinn read the book that Virginia bought.? b. Quinn-ga [ NP __ 1 hon-o] yonda-yo, [ CP Virginia-ga katta] 1 . interesting posibility, we are not adopting Endo?s analysis because of the problem that the analysis faces. If you?re not adopting it, then is al this necesary? 20 This is pointed out to me by Tomohiro Fuji (p.c.). I thank him for leting me know about this type of example. 185 To the best of my knowledge, there are two constructions that can split a relative clause and its head noun in Japanese: right-node raising and right-dislocation. However, no other constructions alow such detachment of relative clauses. For example, neither scrambling nor clefting results in an aceptable sentence if a relative clause alone is moved. (175a) is an example of scrambling and (175b) is an example of clefting. (175) a. *[ CP Virginia-ga katta] 1 Quinn-ga [ NP t 1 hon]-o V-nom bought Q-nom book-acc yonda. read ?Quinn read the book that Virginia bought.? b. *Quinn-ga [t 1 hon]-o yonda-no-wa, Q-nom book-acc read-comp-top [ CP Virginia-ga katta] 1 -da V-nom bought-cop Lit.?It is that Virginia bought that Quinn read the book.? The contrast betwen scrambling and right-dislocation cannot be captured by Tanaka?s scrambling and elipsis approach. The simplest way to reconcile Tanaka?s observations and the contrast betwen scrambling and right-dislocation is to asume that right-dislocated phrase is indeed rightward moved as in the extraposition construction or heavy-shift construction in English. As it is known that English extraposition from NP also can detach a relative clause from its head noun (Johnson 186 1986; Coopmans & Roovers 1986 among others). Thus, we can asume that right- dislocation in Japanese is a variant of rightward movement that is found in English. (176) [ NP Virginia reviews __ 1 ] have been published [ CP which criticized Quinn?s book] 1 . If right-dislocation is a type of rightward movement, then the connectivity efects should be taken as that the main clause R-expresion and the dislocated pronoun in (171c) has a direct c-command relation at some point of derivation. Thus it supports the position that the right-dislocated phrases take part in the main clause structure in some way, i.e., they are not realy disconnected. This entails that the multiple root analyses that potentialy disalow the communication betwen clauses with diferent roots cannot capture the direct relation betwen the main clause element and the dislocated phrase. If we do not appeal to a discourse/pragmatic explanation for which we do not have any clear arguments at this point 21 , the examples that we have sen so far may suggest that Left/Right-Asymmetry is an appropriate generalization that can acount for the islandhood of dislocated constituents. In particular, given that derived position islands (Merchant 2001; Takahashi 1993; Wexler and Culicover 1981 among others) are not operative in Japanese as we have sen, the Left/Right-Asymmetry sems to be the right generalization. In other words, it sems that moving to the right of the verb 21 Kuno (1973, 1978) is investigating the interpretive aspects of right-dislocation. However, it is not discused how the interpretive properties of right-dislocation is related to the extractability. 187 is a crucial factor for a constituent to be an island. Now the question is how we can explain this generalization. The explanation of Left/Right-Asymmetry is beyond the scope of this study. Thus, I would like to leave this problem open for the future research. 188 4.2. A Parametric Study on Adjunct Islands In the previous discussion, we concluded that the Left/Right-Asymmetry is the right generalization cross-linguisticaly. However, we have also sen that, as long as adjunct clauses are generated to the left of the clause, extraction out of adjunct clauses are alowed in Japanese even though their position is lower than what Etxepare and Hornstein argue, i.e., the position betwen the embedded subject and the complementizer. In this subsection, we try to explain this diference betwen Japanese and Spanish /English. 4.2.1. Languages that Alow Extraction out of Adjuncts At the outset of this discussion, we have sen that at least two languages show the same patern as Japanese with respect to the extraction out of adjunct clauses. Let us first take a look at data from Korean and Malayalam and try to figure out what is similar among these languages. The following are example of scrambling out of conditional clauses in Korean. (177) Conditional Clauses Quinn-un [manyak Virginia-ka Stillman-hanthey Q-top [cond-adverb V-nom S-dat senmwul-ul cwu-myen] wul-keya. present-acc gave-cond] cry-will. ?Quinn wil cry if Virginia gives a present to Stilman.? 189 a. Scrambling of Referential NP Bill-hanthey 1 John-un [manyak Mary-ka t 1 B-dat J-top [cond-adverb M-nom senmwul-ul cwu-myen] wul-keya. present-acc gave-cond]cry-will. b. Scrambling of wh-phrase out of conditionals Etten-haksayng-hanthey 1 Quinnn-un [manyak Which-student-dat1 Q-top [cond-adverb Virginia-ka t 1 senmwul-ul cwu-myen wul-ul-ka? V-nom present-acc gave-cond] cry-will-Q Lit. ?Which student wil Quinn cry if Virginia gives a present to?? (178) Relative Clauses Quinn-un [ NP [ CP Virginia-ka Stillman-hanthey cwu-n Q-top V-nom S-dat gave-adnom] senmwul-ul coahay. present]-acc like. ?Quinn like the present that Virginia gave to Stilman.? a. Scrambling out of Relative Clauses *Stillman-hanthey 1 Quinn-un [ NP [ CP Virginia-ka t 1 B-dat J-top M-nom cwu-n senmwul-ul coahay. gave-adnom] present]-acc like. 190 b. WH-scrambling out of Relative Clauses *Etten-haksayng-hanthey 1 Quinn-un Which-student-dat Q-top [ NP [ CP Virginia-ka t 1 cwu-n senmwul]-ul coahay? V-nom gave-adnom]present]-acc like Lit. ?Which student does Quinn like the present that Virginia gave to?? (179) Complement Clause Quinn-un [ CP Virginia-ka Stillman-hanthey senmwul-ul Q-top V-nom S-dat present-acc cwuessta-ko] saynggakha-n-ta. gave-comp think ?John thinks that Mary gave a rsent to Bil? a. Scrambling out of Complement Clauses Stillman-hanthey 1 Quinn-un [Virginia-ka t 1 S-dat Q-top [V-nom senmwul-ul cwuessta-ko] saynggakhay. present-acc gave-comp] think b. Wh-scrambling out of Complement Clauses Etten-haksayng-hanthey 1 Quinn-un [ CP Virginia-ka Which-student-dat Q-top V-nom t 1 senmwul-ul cwuessta-ko saynggakhay? present-acc gave-comp] think The following are relevant data from Malayalam. 191 (180) a. Conditional Clauses Quinn [Virginia Stillman-inu sammaanam Q-nom V-nom S-dat present-acc kodu-thaal] karayum. give-cond cry-will ?Quinn wil cry if Virginia gives a present to Stilman.? b. Scrambling out of Conditional Clauses Stillman-inu 1 Quinn [Virginia t 1 sammaanam S-dat Q-nom V-nom present-acc kodu-thaal] karayum. give-cond cry-will (181) a. Relative Clauses [ NP [ CP Quinn [ CP Virginia Stillman-inu t 1 Q-nom V-nom S-dat kodu-thennu] viswasikkane] pustakam 1 gave-comp believe-adnom book ?The book that John believes that Mary gave to Bil? b. Scrambling out of Relative Clauses *Stillman-inu 1 [ NP [ CP Quinn [ CP Virginia t 1 t 2 S-dat Q-nom V-nom kodu-thennu] viswasikkane] pustakam 2 gave-comp believe-adnom book 192 (182) a. Complement Clauses Quinn [ CP Virginia Stillman-inu sammaanam Q-nom V-nom S-dat present-acc kodu-thennu] viswasikkunu gave-comp believe ?Quinn believes that Virginia gave a present to Stilman.? b. Scrambling out of Complement Clauses Stillman-inu 1 Quinn [ CP Virginia t 1 sammaanam S-dat Q-nom V-nom present-acc kodu-thennu] viswasikkunu give-comp believe As we can se, both of these languages show exactly the same patern as Japanese. In al these cases, extraction out of conditional clauses is much beter than that out of relative clauses. Based on the data we have, we can conclude that there are at least two types of languages. In the first type, adjunct clauses are islands if they are not adjoined to FP in the embedded clause. In the second type, adjuncts are not islands as long as they are generated to the left to the verb. 2 2 If we consider languages such as Swedish or Danish, we can have the other type of language in which even relative clauses, as wel as adjunct clauses, are not islands (Engdahl 1986, 1982). In al the other languages at hand, relative clause islands are respected. Considering the fourth type of languages is inded interesting, but it is beyond the scope of this study. Thus, I leave this problem open for future research. 193 4.2.2. Parametric Features The question is what is the feature that distinguishes languages like English or Spanish and Japanese, Korean and Malayalam. Are there any typological features that can distinguish these two types of languages? Do the cross-linguistic similarities or diferences with respect to the extraction out of adjunct clauses systematicaly follow from those features? Even if we can find some significant similarities among these thre languages, if these similarities do not have any connection to the extraction out of conditional clauses, the problem of extractability remains unresolved. Thus, it is desirable to find typological features that are shared by these languages and are related in some way to extraction out of adjunct clauses. In the following discussion, I try to answer the first question, but the answer for the second question remains open. One of the prominent features that Japanese, Korean and Malayalam share is head-finality. Can we derive the extraction paterns from the head-finality? There is reason, however, to believe that head-finality is not the crucial feature that derives the adjunct island asymmetry, i.e., there are head-final languages that show adjunct island efects in the same way as Spanish and English. Basque is one such language. 23 In Basque, a conditional clause is an island if it is in the complement of non-reporting verbs such as desire or want, or if it is located to the right of the embedded verb. If they are located in a higher position, overt wh-movement is alowed. We can se this patern in the following examples in (183). 23 I would like to expres my gratitude to Leticia Pablos for native speaker?s judgment about the Basque sentences. 194 (183) a. Zein liburu pro esan zenuen [Rikardok 1 which-book say aux-past R-erg irakurtzen (baldin) bazuen] pro 1 read if if-aux hizkuntzalaritza utziko zuela? linguistics leave-future aux-comp ?Which book do you say that if Rikardo read he would abandon linguistics?? b. *Zein liburu pro esan zenuen pro 1 which book say aux hizukunzelaritza utziko zuela [Ricardok 1 linguistics leave-future aux-comp R-erg irakurtzen (baldin) bazuen]? read if if-aux c. *Zein liburu pro nahi zenuen [Rickardok 1 which book want aux-past R-erg irakurtzen (baldin) bazuen] pro 1 read if if hizkuntzalaritza utziko zeula? linguistics leave-aux aux-comp ?Which book do you desire that if Rikardo read he would abandon linguistics?? 195 As is evident from the examples, matrix verbs precede complement clauses in Basque. Thus, Basque is not a strongly head-final language like Japanese. However, the language alows head-final configurations, and stil extraction from adjunct clauses paterns very much like Spanish or English. Therefore, we can plausibly conclude that head-finality is not the right feature, or we cannot explain the paradigm only by head-finality. Looking at these languages that alow extraction from adjunct clauses, it becomes clear that there are other similarities, in addition to the head-finality. In particular, these languages share such properties as scrambling, wh-in-situ and an indeterminate system, as wel. These properties, in particular, sem to distinguish these thre languages from languages like Basque, English or others. Thus, it is worthwhile taking a look at these properties and check whether we can succesfully derive the cross-linguistic diferences betwen two types of languages. Let us first review these properties. As it is wel known, both in Korean and Malayalam, wh-phrases remain in-situ 24 , while they can be scrambled to the beginning of a clause, as ilustrated in (177) through (182). Examples of simple wh- questions are in (184). 24 Note that Jayaselan (201) sugests that non-cleft wh-question in Malayalam involves IP-internal focus movement even though the surface word order is the same as basic word order, SOV. In this study, however, we basicaly treat Malayalam as wh-in-situ language in the sense that wh-phrases do not obligatorily move to the sentence initial position, and they are pronounced in their base position. 196 (184) a. Korean Wh-Question (Beck & Kim 1997) Suna-ka mu?s-?l sa-ss-ni? Sun-nom what-acc buy-past-Q ?What did Suna buy?? b. Malayalam Wh-Questions (Jayaselan 2001) nii ent? aaN? tiinn-at?? you what is ate-nominalizer? ?What did you eat?? Second, in al the thre languages, Japanese, Korean and Malayalam, wh- words are used as so-caled indeterminate pronouns (Hiraiwa 2002; Kuroda 1965; Nishigauchi 1990; Shimoyama 2001; Watanabe 1992a, 1992b, 1996, 2001, 2002, 2003). In these languages, wh-words are used in various ways depending on the focus particles they atach to. The full paradigms of the indeterminate system in these languages are summarized in the table below. As we can se, if a wh-phrase is asociated with the conjunction marker, which is also used to expres also in Japanese, then they are interpreted as universal quantifiers. If they are asociated with the disjunction marker, they are existential quantifiers. 197 (185) Indeterminate system Japanese Korean 1 Malayalam Wh Wh-conj Wh-disj Wh Wh-conj Wh-disj 1 Wh Wh-conj Wh-disj dare (who) dare-mo (anyone) dare-ka (somebody) nwukwu (who) nwukwu- wa (anyone) nwukwu-na (somebody) 1 ar (who) ar-um (anyone) ar-o (somebody) nani (what) nani-mo (anything) nani-ka (something) mwues (what) mwues- kwa (anything) mwues-ina (something) 1 ent (what) ent-um (anything) ent-o (something) doko (where) doko-mo (anywhere) doko-ka (somewhere) eti (where) eti-wa (anywhere) eti-na (somewhere) 1 e? (where) e?-um (anywhere) e?-o (somewhere) itu (when) itu-mo (always) itu-ka (sometime) ence (when) ence-na (always) encenka (sometime) 1 epozh (when) epozh- um (always) epozh-o (sometime) The other thre languages we have sen (English, Spanish and Basque) do not share al of these properties, though they share some of them. Taking English as an example, it is a head-initial language and does not have the indeterminate system that Japanese, Korean or Malayalam does. In English, wh-phrases and quantifiers are formed in diferent ways. For example, a ?person? wh-phrase is who, but the universal quantifier for the person is not who-and, or who-also but someone, an independently used nominal element like one and a determiner like some is used to form an existential quantifier. The same holds true for Basque (Hualde and Urbina 2003). As we have shown in the case of the head-finality, we can also show that either the indeterminate system or wh-in-situ alone is not a crucial feature for the extractability out of adjunct clauses. For example, Russian has the same type of 198 indeterminate system as Japanese, Korean and Malayalam (Watanabe 2002), but Russian shows adjunct island sensitivity (Stepanov 2001). Crucialy, Russian is wh- movement language. On the other hand, the possibility of wh-in-situ is not sufficient to derive the patern of adjunct non-islandhood either. Malay alows wh-in-situ as wel as overt wh-movement and partial wh-movement (Cole and Hermon 1998). However, Malay overt wh-movement exhibits adjunct island efects (Cole and Hermon 1998). Finaly, the availability of scrambling does not sem to be the crucial factor either. As we have discussed, Russian shows adjunct island efects for overt movements even though Russian is indeed a language that alows scrambling (Bailyn 2001, 2002 among many others). These cross-linguistic paterns suggest that the combination of these features (head-finality, wh-in-situ, indeterminate system and scrambling) sems to be crucial in deriving the adjunct (non)-islandhood. In other words, if we take the strongest position based on the data at hand, we can conclude that only the languages that have these four properties alow extraction out of adjunct clauses. A note is in order. As it shal be clear from the data we have presented and discussion thus far, this conclusion is probably too strong. We have not checked other properties that are shared by these languages. However, the prediction of the theory of cross-language diferences adopting this conclusion is quite clear. If there is a language that shares al these four properties, it should alow extraction out of pre- verbal adjunct clauses. 199 The remaining question is how we can derive the adjunct non-islandhood from these parametric features. In so doing, we have to answer how these features are related to extraction. At this point, I leave this problem open. 5. Conclusions In this chapter, we have discussed the syntax of RCs and Conditional clauses. By applying various gramatical tests, we have revealed the internal syntax and external syntax of these two constructions. We have also tackled the problem of the non-islandhood of Japanese adjunct clauses. Based on a cross-linguistic study, we found that there is a cluster of parametric features that sems to be responsible for the non-islandhood of adjunct clauses. However, the question of how we can derive the cross-linguistic paterns has been left open. 200 CHAPTER 3. RELATIVE CLAUSE PREDICTION IN JAPANESE 1. Introduction Current studies in sentence procesing suggest that human sentence procesing is incremental, in the sense that the procesor incorporates input words into a gramatical analysis as soon as they are encountered, and thus structural commitments are made and the interpretation of sentences is constructed from left to right without delay (Marslen-Wilson 1973). With this as a background, a widely adopted asumption is that the human sentence procesor is strongly incremental, i.e., the sentence procesor builds a fully connected structure as the procesor receives each input word, where the procesor does not alow unstructured or partialy structured input that is stored as unconnected pieces (Frazier and Rayner 1987; Gibson 1991; Gorrel 1995; Inoue and Fodor 1995; Stabler 1994). Recently, it has been suggested in the literature that a strongly incremental parser requires a powerful predictive mechanism exploiting ?extra-lexical? knowledge (Lombardo and Sturt 2002; Sturt and Crocker 1996; Schneider, 1999; among others). The aim of this study is to show that the human sentence procesing mechanism is indeed equipped with a powerful predictive mechanism, through experimental studies on various aspects of relative clauses in Japanese. We wil argue that the predictive mechanism in the human sentence procesor has the ability to project the sentence structure that goes beyond the local structure that can be constructed based on the information available from each lexical item. 201 2. Necesity of extra-lexical knowledge in sentence procesing The necesity of extra-lexical knowledge by an incremental parser is typicaly ilustrated by examples like (1). In (1), a strongly incremental parser incorporates the adjective ?steper,? into the current representation without waiting for other lexical items to become available. In order to do so, however, the parser has to build the structure of the NP and IP categories, which do not have overt heads at the point where the word ?steper? is encountered. The parser has to project an NP node to host the adjective in this environment. Furthermore, the IP node is necesary because the NP that is predicted by the adjective must be atached as a specifier of IP. At the point of "steper" the most likely position for the NP in this partial context is the subject of an embedded clause. Building this structural skeleton requires the projection of IP based on the predicted NP node, for an atributive adjective must be licensed by an NP. While it is plausible to asume that the presence of an atributive adjective can set a prediction for the upcoming NP node, how the IP node be predicted is not at al obvious. This is so because there is no direct gramatical relation betwen an adjective and an IP node 25 . In other words, there is no requirement that an adjective must be licensed by an IP node or vice versa. The only conceivable requirement imposed by an adjective in a sentence like (1) is that it must be asociated with an NP. Thus, based solely on the gramatical requirement of an adjective, there is no reason that the parser should project up to an IP node upon encountering an adjective. Hence, 25 Prediction of the CP node is not so problematic because a verb like ?think? takes a clausal complement. Therefore, it is plausible to asume that the selectional information of the verb alows the prediction of CP. 202 this example, in turn, suggests that there are situations in which incremental structure building has to commit global syntactic inferences that can project the structure that goes beyond the local structure that can be projected using gramatical requirements stored in an input lexical item. (1) a. He thinks [ CP [ IP [ NP [ Adj steper] prices] have come about because ?]] (Lombardo and Sturt 2002: 138) b. . A theoretical question that an example such as (1) raises is what kind of parsing algorithm can succesfully proces the sentence without sacrificing incrementality. In the history of psycholinguistics, various parsing algorithms have been proposed in an efort to formulate a psychologicaly plausible sentence procesing mechanism. Taking one example, the so-caled left-corner strategy (Abney and Johnson 1991; Aho and Ulman 1972; Johnson-Laird 1983; Resnik 1992; Stabler 1994) has often been considered as a psychologicaly plausible model of sentence 203 procesing in the sense that it enable us to maintain incremental sentence procesing with a reasonable power of prediction of the upcoming structure. The basic intuition behind the left-corner strategy is the following. Initialy, structure is built in a bottom- up fashion using the information from an incoming lexical item. After any constituent is completed, however, its parent node is built and potential daughter nodes of the completed constituent are predicted. Although, the left-corner algorithm maintains incremental sentence procesing, even a sentence procesing mechanism incorporating this algorithm cannot incrementaly parse the sentence ilustrated in (1). This is so because the left- corner algorithm creates a prediction for an upcoming node based on the bottom-up information of the input lexical items. In (1), the left-corner strategy alows the parser to predict the NP node and its potential siblings, based on the gramatical information from the adjective ?steper.? However, the prediction of the IP node is beyond its reach for the reasons we have discussed above. Therefore, a sentence like (1) raises an interesting theoretical question about the nature of a psychologicaly plausible parsing algorithm or the nature of the predictive mechanism of the human sentence procesor. From a diferent perspective, an example like (1) raises another theoreticaly important question, i.e., whether the human sentence procesing mechanism should be understood as incremental at al. If the human parser does not maintain incrementality, and thus waits until a crucial lexical item appears in the input, the problem of non-local projection does not arise in the first place. Al else being equal, such a powerful mechanism should be avoided if we do not have a realy strong 204 motivation. The so-caled head-driven model or delay model in more general terms (Abney 1989; Pritchet 1991a, 1992a) alows us to handle a sentence like (1) without postulating such powerful predictive mechanisms. If structure building is driven entirely by lexicaly encoded information that is recovered from overt material in the input, then the procesing of sentences like (1) requires a certain delay. This is so because the lexical item ?steper? does not provide any specific cue for the IP node as we have observed. Put diferently, if the parser delays building the structure until the gramatical information of the input lexical items become available, it should not incorporate the atributive adjective into the structure imediately upon encountering it. The parser should then construct the sentence structure when it encounters the host noun for the atributive adjective, for instance. Given this alternative scenario, we are forced to ask the question whether the parser delays the procesing of a sentence until crucial lexical items appear in the input or whether it builds the structure dynamicaly by exploiting detailed extra- lexical syntactic knowledge with a powerful predictive mechanism. 3. On the Prediction of Japanese Relative Clauses For the two isues that we have discussed so far, Japanese may provide a good testing ground. Because of the strong head-finality of its basic sentence structures, the online sentence procesing of Japanese creates situations similar to (1) in almost every sentence. In Japanese, the head of each phrase folows al other elements in the phrase, both modifiers and arguments. Acordingly, the sentence procesor is not normaly 205 provided with reliable cues about the upcoming sentence structures until it encounters the head at the end of each phrase (Hirose 1999; Inoue 1991; Mazuka and Lust 1988, 1990; Mazuka and Itoh 1995; Miyamoto 2002, 2003). Among such head-final constructions in Japanese, relative clauses have atracted interest from researchers atention because they create notorious garden path efects (Hirose 1999; Inoue 1991; Mazuka and Lust 1988, 1990; Mazuka and Itoh 1995; Miyamoto 2002, 2003; Nakamura 1999/2000, 2003; Yamashita et al. 1993; Yamashita 1995). It is generaly asumed in the literature that garden path efects are created because in Japanese an upcoming relative clause structure is extremely dificult to predict. Let us discuss this point more in detail. It has been observed that the parser shows a bias to construct a simple clause or a complement clause structure whenever possible (Inoue 1991; Mazuka and Lust 1988, 1990; Mazuka and Itoh 1995; Miyamoto 2002, 2003; Yamashita et al. 1993; Yamashita 1995). Thus, a relative clause structure is not counted as the initialy preferable structure. From this initial preference for a complement clause analysis, it follows that when a disambiguating cue from the relative head or morphology of the verb is provided, a garden path efect may arise because the parser has to reanalyze the initial complement clause structure to a relative clause structure. The garden path efect is caused because no syntactic or morphological cues for the presence of the relative clause structure are provided at an earlier point in parsing, unlike relative clauses in English where the presence of relative pronouns can mark the beginning of a relative clause. This implies, however, that this garden path efect may be resolved if an upcoming relative clause structure can be predicted in advance. 206 A careful examination of the syntactic properties of Japanese noun phrases, however, reveals that a certain clas of modifiers of NPs, the so-caled numeral clasifiers, may provide an unambiguous cue for an upcoming relative clause structure. The structure in (2) ilustrates this schematicaly. As we shal se in more detail below, in a specific context where a clasifier cannot be asociated with its linearly adjacent NP, it must be asociated with the head of a relative clause. Under this circumstance the material that intervenes betwen the clasifier and the host NP must be a relative clause. (2) This relation betwen the clasifier and the relative head NP suggests, in turn, that if the parser can employ this information in the course of parsing, it can create a prediction for the upcoming relative clause structure. In other words, in certain environments the clasifier can provide an unambiguous cue for an upcoming relative clause during online sentence procesing. Although it is a plausible consideration that numeral clasifiers may indicate an upcoming relative clause structure to the parser, how the parser creates the prediction is not at al obvious. This is because clasifiers do not have any direct gramatical relation to relative clauses. A clasifier modifies a noun phrase but it 207 does not modify a relative clause. In the same way a relative clause modifies a noun phrase, but it does not modify a clasifier. Thus, even though a relative clause can intervene betwen a clasifier and its host NP, it does not imply that the clasifier and the relative clause have any direct relation to each other. They are both asociated with a noun phrase but not with each other. In other words, the presence of a clasifier does not gramaticaly imply the presence of a relative clause. With the discussion so far in mind, let us consider how a strongly incremental parser might project an upcoming relative clause structure. If the parser is strongly incremental, it has to project the following structural skeleton based on the encounter with a numeral clasifier that cannot be asociated with its adjacent NP in order to incorporate these elements into a sentence structure. First, the position occupied by the clasifier has to be projected. Clasifiers are licensed by an NP, so an NP node has to be projected. Second, an IP node is required, where the NP containing the clasifier is incorporated. Furthermore, a CP node must be projected in which the IP is inserted. Finaly, the CP has to be connected to the NP as a relative clause. In (3), this proces is ilustrated (dotted lines indicate predicted nodes). (3) 208 If this is the proces that the parser employs, several problems arise. Esentialy they are the same problems as we have sen in the discussion of the English example. First the problem of non-headed structures arises as we have sen in (1). In (3), IP, CP and NPs have to be projected without the cues from their heads. Second, it is not clear how the clasifier, a modifier of the NP can specificaly provide a cue for the relative clause structure, as we have discussed. The lexicaly encoded information of the clasifier does not provide specific cues for the structure of relative clauses. The requirement that any numeral clasifier demands is that it must be asociated with an appropriate noun phrase. Given that a numeral clasifier does not have any conceivable direct relation to relative clauses, it is not plausible to asume that the presence of a clasifier directly cues the presence of an upcoming relative clause. Thus, how the relative clause structure is predicted is not obvious from the information encoded in/with each lexical item such as numeral clasifiers. Finaly, if a prediction for an upcoming relative clause is ever possible, the parser has to be equipped with a powerful predictive mechanism. This predictive mechanism must alow the parser to project the NP, IP, CP nodes and how the CP is connected to the NP, upon encountering a situation in which a clasifier cannot be asociated with its adjacent NP. Now it should be clear that these problems do not arise if the parser delays the procesing of sentences until the crucial lexical item appears in the input. The parser just builds the structure when the unambiguous cues from the relative head NP or the embedded verb become available. Hence, in this view, headles predicted projections 209 are not needed. This, in turn, means that the parser does not need to have the powerful predictive mechanism while the incrementality is sacrificed. The goal of this study is to investigate the two points we have discussed above through experimental studies on Japanese relative clauses. Through a detailed experimental examination of various aspects of relative clauses in Japanese, we wil provide supporting evidence for the position that the parser maintains the strong incrementality and that it is equipped with a powerful predictive mechanism of the sort that we have briefly discussed above. There are two major findings in this study. The first is that the parser is able to create a prediction for an upcoming relative clause structure upon encountering a numeral clasifier that is not semanticaly compatible with its adjacent NP. This finding provides evidence for the view that the parser can project structure beyond the highly local structures that can be projected based on the input lexical information. Secondly, we wil show that, in addition to the prediction for a relative clause structure, the parser can compute at least one consequence of this prediction, namely the islandhood of the relative clause. We wil se that computing the islandhood of relative clauses requires abstract and detailed syntactic inferences. Thre experimental results are reported that additionaly support the existence of a powerful predictive parsing mechanism and that provide further evidence against delay models of sentence procesing. 210 4. The Procesing of Japanese Relative Clauses Now, let us briefly summarize more concretely the basic isues in the procesing of Japanese relative clauses that we are concerned with. Since Inoue's (1991) studies on Japanese parsing, it has generaly been agred that the beginning of embedded clauses in Japanese is hard to detect on-line. Japanese lacks obvious markers of the beginning of an embedded clause. Comparative syntactic studies betwen Japanese and English help us to understand this point more clearly. Let us cite some clear cases. In Japanese, a complementizer appears at the end of each clause. On the other hand, in English it comes at the beginning of the embedded clause. Japanese lacks relative pronouns or relative complementizers that can mark the left-edge of embedded clauses, while English has them (Fukui 1995; Kuroda 1988). Therefore, it sems that almost no reliable cues are provided for the beginning of embedded clauses that the parser can make use of. Recently, Miyamoto (2002), following Inoue?s (1991) insight, experimentaly showed that Case Markers can induce clause boundaries in Japanese. Miyamoto argues that the second Nominative NP in (4) can mark the onset of the embedded clause. A Nominative NP typicaly marks the onset of a tensed clause. From this fact, it folows that the second occurrence of the Nominative NP implies the presence of a tensed clause, and therefore it can tel the reader that an embedded clause begins at this position. 26 26 Note that the lexical semantics of Nominative NPs may afect the clause boundary induction. Se Muraoka and Sakamoto 203 for a detailed discusion on this point. 211 (4) Obasan-ga [ RC yoboyobo-no toshiyori-ga guuzen-ni kousaten-de mita] Woman-Nom feble old-man-Nom by chance intersection-Loc saw onnanoko-ni isoide koe-o kaketa. girl-Dat hury caled. ?The woman hurriedly caled the girl who the feble old-man saw at the intersection by chance.? Miyamoto's study shows that sometimes the parser can detect the beginning of an embedded clause. However, as far as relative clauses are concerned, it does not help much. A number of previous studies report that a sequence of NPs is likely to be interpreted as the arguments of a single verb whenever possible (Inoue 1991; Mazuka and Lust 1988, 1990; Mazuka and Itoh 1995; Miyamoto 2002, 2003; Yamashita et al. 1993; Yamashita 1995). For example, in (5a) the parser initialy proceses the thre NPs with diferent Case makers as being asociated with a single verb, in this case "ageta," gave, since the sequence of NPs, [NP-Nominative, NP-Dative, NP- Acusative], is typicaly asociated with a ditransitive verb. Because of this preference, if the verb's argument structure is not compatible with those NPs and the structure turns out to be a relative clause, a garden path efect arises because the parser has to reanalyze the sentence structure as ilustrated in (1a). 212 (5) a. Marco-ga Kity-ni ringo-o ageta. Marco-Nom Kity-Dat apple-Ac gave ?Marco gave an apple to Kity? a. Marco-ga Kity-ni ringo-o tabeta inu-o miseta. Marco-Nom Kity-Dat apple-Ac ate dog-Ac show ?Marco showed the dog that ate an apple to Kity.? (Inoue 1991; Mazuka and Itoh 1995) The situation is similar even if the cues for the beginning of the embedded clause are provided by Case markers. The second occurrence of a Nominative NP can induce a clause boundary. However, a Nominative NP is equaly compatible with a complement clause and a relative clause. Based on the observations in the literature it is plausible that the parser?s preference for interpreting NPs as arguments of a single verb forces the complement clause analysis rather than the relative clause analysis even though the beginning of the embedded clause is signaled. What we can conclude from the discussion so far is that the relative clause structure is one of the least preferable structures for the parser?s initial analyses. Moreover, it sems to be clear that the parser cannot normaly construct a relative clause structure until crucial information such as the head of the relative clause becomes available. So far, ways of unambiguously marking the left-edge of embedded clauses have not been reported in the literature. Here we show that clasifiers in Japanese have the potential to provide an unambiguous cue for an upcoming relative clause. In an example like (6), a numeral clasifier bearing Genitive Case, which we wil refer 213 to as a Genitive Numeral Clasifier, must be unambiguously asociated with the relative head NP, "hon" [book]. Although there is an NP, ?gakuse? [student] adjacent to the clasifier, it cannot be asociated with the clasifier because of the gramatical constraints on Genitive Numeral Clasifiers that we wil review shortly. (6) 3-satu-no gakuse-ga yonda hon. 3-clasifier (printed mater) student-Nom read book ?3 books that the student read.? This type of numeral clasifier has the following general gramatical requirements. (7) a. A Genitive Numeral Clasifier must be asociated with a structuraly adjacent Host NP. b. A Numeral Clasifier and its host NP must be semanticaly compatible. The structure of a noun phrase containing a Genitive Numeral Clasifier that can satisfy the above requirements is something like (8). In (8) the Genitive Numeral Clasifier is left adjoined to the NP that it modifies. (8) 214 This analysis implies that as long as the structural adjacency condition is satisfied, any modifiers or arguments of the head noun can intervene linearly betwen the Genitive Numeral Clasifier and the host NP. As we can se in the examples in (9), any modifiers or arguments of the head noun including a relative clause can indeed intervene betwen a Genitive Numeral Clasifier and its host NP without changing the relation betwen them. (9) a. Inserting an adjective [ NP 3-satu-no [ NP nagai hon]] 3-cl(printed mater)-Gen long book ?Thre long books? b. Inserting a possesive NP [ NP 3-satu-no [ NP Marco-no hon]] 3-cl(printed mater)-Gen Marco-Gen book ?Marco?s thre books? c. Inserting a modifier [ NP 3-satu-no [ NP gengogaku-no hon]] 3-cl(printed mater)-Gen linguistics-Gen book ?Thre books about linguistics? 215 d. Inserting modifiers and arguments [ NP 3-satu-no [ NP Marco-no nagai gengogaku-no hon]] 3-cl(printed mater)-Gen Marco-Gen long linguistics-Gen book ?Marco?s thre long books about linguistics? e. Inserting a relative clause [ NP 3-satu-no [ NP [ RC Marco-ga yonda] hon]] 3-cl(printed mater)-Gen Marco-Nom read book ?Thre books that the student read.? Now let us se the following examples. Both of the examples in 0) are aceptable sentences. In (10a) the numeral clasifier is asociated with the subject NP of the relative clause. On the other hand, in 0b) the clasifier must be asociated with the relative head NP because the linearly adjacent NP, the subject of the relative clause, is semanticaly incompatible. The only possible host NP is the NP headed by the relative head noun, i.e., the whole NP containing the relative clause. The structures of each example are in ilustrated in Fig. 1. 216 (10) a. 3-nin-no gakuse-ga Yonda Hon 3-cl(human)-Gen student-Nom Read Bok ?The bok that the thre students read? b. 3-satu-no gakuse-ga Yonda Hon 3-cl(printed mater)-Gen student-Nom Read Bok ?The thre boks that the student read? 217 Fig. 1. A. B. From the discussion so far, it should be clear that when the clasifier semanticaly mismatches with an imediately following nominative NP the only possible structure is a relative clause. Thus, if the parser can notice that a mismatching Clasifier can be licensed only in a relative clause structure, it may be possible to predict an upcoming relative clause. However, here the very problem that we have discussed earlier arises, namely, how specificaly the relative clause structure can be predicted. The examples that we have examined show that Numeral Clasifiers do not have any lexical requirement to have a relative clause as an intervening element. As we have sen, the intervening element can be other modifiers or arguments of the asociated noun. The only requirement that the Genitive Numeral Clasifier has is that it must be licensed by a structuraly adjacent Noun Phrase that is semanticaly compatible. In short, although the mismatching Genitive Numeral Clasifier may provide a cue for an upcoming relative clause structure, there is a gap 218 betwen the information available from the mismatch betwen the genitive numeral clasifier and its adjacent NP and the actual structure of the relative clause that the parser has to build. If the parser can fil this gap and can project the upcoming relative clause structure based on the cue, it means that the parser can dynamicaly exploit the relevant gramatical information without using the information from the heads. In what follows, we wil report the results from thre experiments. The results of the experiments wil show that the parser makes use of the indirect cue from the mismatching numeral clasifier and creates a prediction upon encountering a mismatch betwen a Clasifier and a Nominative NP. These results suggest that the parser is strongly incremental and that the parser is equipped with a powerful predictive mechanism. The first experiment, an off-line sentence fragment completion experiment, is designed to answer the following two questions. The first question is whether native speakers of Japanese generate a relative clause structure when they are provided with mismatching numeral clasifiers. If the mismatching numeral clasifier can provide a cue for an upcoming relative clause, then Japanese speakers should complete sentence fragments with relative clause structures. If, on the other hand, the mismatching numeral clasifier is not used as a cue for an upcoming relative clause structure, Japanese speakers should not complete the sentence fragments with a relative clause, rather, they should complete sentences with complement clauses following their general tendencies. 219 5. Experiment 1 An off-line sentence fragment completion test was conducted as an initial test of whether Japanese speakers are able to use numeral clasifiers as a cue to generate relative clauses, using a task without time restrictions. Participants were presented with sentence fragments consisting of a sequence of 3 NPs in a paper-and-pencil task and asked to write completions for the sentences that semed natural to them. The experiment manipulated two factors in a 2 x 3 factorial design. The first dependent variable that we are concerned with is the semantic compatibility betwen the numeral clasifiers and the adjacent nominative NPs, in order to ases whether speakers would recognize a mismatching clasifier-noun sequence as a cue for a relative clause boundary. Participants 120 Japanese speakers participated in the experiment, al of whom were students at Shizuoka University or Shizuoka Sangyo University, Shizuoka, Japan or at Meiji Gakuin University, Tokyo, Japan. Al gave informed consent and were paid the equivalent of $5.00 for their participation in the experiment, which lasted about 30 minutes. Design The experiment followed a 2 x 3 factorial design, which manipulated the match betwen a genitive numeral clasifier and a following nominative NP (match vs. mismatch) and the case-marking and position of a wh-phrase that appeared earlier 220 in the fragment than the (mis-)matching NP (initial-dative vs. medial-dative vs. initial-nominative). The wh-phrase manipulation was included in order to investigate the impact of clasifier mismatches on the procesing of filer-gap dependencies, an isue that is more relevant to the later discussion on the island efect of relative clauses, and thus is not discussed in detail at this point. A sample set of experimental conditions is shown in Table 1. 221 Table 1 Sample Set of Experimental Conditions for Experiment 1 a. Clasifier match/initial dative wh-phrase Dono-seto-ni tanin-wa 3-nin-no tosioita sense-ga which -student-Dat clas-teacher-Top 3-Cl(human)-Gen aged teacher-Nom b. Clasifier mismatch/initial dative wh-phrase Dono-seto-ni tanin-wa 3-satu-no tosioita sense-ga which -student-Dat clas-teacher-Top 3-Cl(bok)-Gen aged teacher-Nom c. Clasifier match/medial dative wh-phrase Tanin-wa dono-seto-ni 3-nin-no tosioita sense-ga clas-teacher-Top which -student-Dat 3-Cl(human)-Gen aged teacher-Nom d. Clasifier mismatch/medial dative wh-phrase Tanin-wa dono-seto-ni 3-satu-no tosioita sense-ga clas-teacher-Top which -student-Dat 3-Cl(bok)-Gen aged teacher-Nom e. Clasifier match/initial nominative wh-phrase Dono-seto-ga tanin-ni 3-nin-no tosioita sense-ga which -student-Nom clas-teacher-Dat 3-Cl(human)-Gen aged teacher-Nom f. Clasifier mismatch/initial nominative wh-phrase Dono-seto-ga tanin-ni 3-satu-no tosioita sense-ga which -student-Nom clas-teacher-Dat 3-Cl(bok)-Gen aged teacher-Nom The experimental materials consisted of 18 sets of 6 conditions distributed among six lists in a Latin Square design. Each participant saw exactly one of the lists of 36 target items intermixed with thirty-six filer items in a random order. The filer items included a variety of forms of numeral clasifiers and wh-phrases, and were thus intended to prevent participants from noticing the regularity in the form of the 222 target items. The length of the fragments and the anticipated complexity of the completions was matched across target and filer items. Some notes on the materials are in order. First, in this and al subsequent experiments in this article, the critical nominative embedded subject NP always denoted a human. The match betwen this NP and the preceding clasifier was varied by manipulating the animacy of the clasifier. The matching conditions used the clasifier for humans ?nin, and the mismatching conditions used clasifiers for inanimate objects such as books. It was necesary for the mismatching clasifiers to be inanimate, in order to avoid a potential ambiguity asociated with mismatching genitive numeral clasifiers that denote humans. As shown in (11), an animate clasifier may sometimes be understood as refering to the possesor of a following inanimate NP. Japanese speakers find this interpretation easier to obtain when the genitive clasifier is followed by a demonstrative article such as ano ?that?, but it is at least marginaly available even without the demonstrative. In contrast, since inanimate objects are highly implausible possesors, the same ambiguity does not arise with inanimate clasifiers. Acordingly, (12) is judged to alow only a reading where the clasifier is construed with the head of the relative clause. Therefore, the mismatching clasifiers were always inanimate, in order to preserve the status of the clasifier mismatch as a cue for a relative clause. Second, in order to maintain the naturalnes of the fragments, we inserted adjectives betwen the clasifiers and the nominative embedded subject NPs. Al the adjectives can modify only these embedded subjects, and thus semanticaly they are compatible only with animate NPs. 223 (11) 3-nin-no hon-ga hihan-sita gengogakusha 3-Cl (human) -Gen book-Nom criticized linguist Interpretation 1: A linguist who the book(s) by the thre authors criticized. Interpretation 2: Thre linguist who a book/book(s) criticized. (12) 3-satu-no gengogakusha-ga hihansita hon 3-cl (bok like things) -Gen linguist-Nom criticized book Interpretation1: Thre books that the linguist criticized. Interpretation2: #The book that the linguists in the thre books criticized. Results This study yielded a total of 2082 codable sentence fragment completions. Table 1 summarizes the efect of the clasifier match factor on the frequency of production of relative clauses in the fragment completions. Table 2 summarizes the counts and percentages of the types of completions across conditions. The impact of both the clasifier match factor and the wh-phrase type factor on the production of filer-gap dependencies is presented later. A completion was clasified as containing a ?matching relative clause? if it included a relative clause whose head was semanticaly compatible with the numeral clasifier in the sentence-initial fragment. Al other completions were clasified as ?other?. This included both completions that contained no relative clause, and completions that contained a relative clause whose head did not semanticaly match the clasifier. 224 Table 2 Rates of relative clause completions in Experiment 1. Construction Types Matching Relative Clause Other Total Conditions N % N % Clasifier Match 1 0.1 1018 9.9 1019 Clasifier Mismatch 851 80.05 212 19.95 1063 2082 225 Table 3 Rates of relative clause completions in Experiment 1 across 6 conditions Construction Type Matching Relative Clause Other Conditions N % N % total GNC Matching/ Scrambling 0 0 314 10 314 GNC Mismatching/ Scrambling 21 70.1 94 29.8 315 GNC Matching/ Non Scrambling 1 0.2 375 9.7 376 GNC Mismatching/ Non Scrambling 35 87.9 46 12.07 381 GNC Matching/ Wh-Nominative 0 0 329 10 329 GNC Mismatching/ Wh-Nominative 295 80.3 72 19.6 367 2082 As shown in Table 1, the match betwen the clasifier and the following NP had a large impact upon the proportion of matching relative clauses that participants generated. In the conditions with mismatching clasifiers 80% of completions contained a matching relative clause, whereas there was just a single instance (0.1%) generated in the conditions with matching clasifiers. A  2 test showed that the proportion of relative clause completions was significantly diferent betwen the matching and mismatching conditions ( 2 (1)=1375.84, p < .01). This patern holds 226 true across al conditions. In the Matching Scrambling Condition there were no matching relative completions, in the Matching Non Scrambling Condition there was just one, and in the Matching Wh-Nominative Condition there were no matching relative clauses again. On the other hand, in al of the Mismatching Conditions there were many more matching relative clause completions. The Mismatching Scrambling Condition contained 70.1% matching relative clause completions, the Mismatching Non Scrambling Condition contained 87.9%, and the Mismatching Wh-Nominative Condition contained 80.3%. A  2 test showed that the proportion of relative clause completions was significantly diferent betwen the matching and mismatching conditions within each level ( 2 (5)=1398.397, p < .01). This contrast shows very clearly that mismatching clasifiers were efective in raising the proportion of relative clauses that were generated, but the 19.9% (212/1063) of trials in the clasifier mismatch conditions that did not elicit a relative clause is also notable, since a relative clause should be required in order to complete these conditions in a gramaticaly aceptable fashion. These trials consisted of a combination of blanks, gibberish or ungramatical completions. An example of such an ungramatical completion is shown in (13). (13) is ungramatical both because it lacks a host NP for the clasifier and because it has only one predicate, despite having two subject NPs. (13) Tannin-wa 3-satsu-no [tosioita sense-ga ita-rasi]. clas-teacher 3-Cl bok-like -gen aged teacher-Nom be-there-sems ?As for the clas teacher, there sem to be thre (book-like) old teachers.? 227 Discussion The analysis of clause types in the sentence fragment completion data showed that the presence of a clasifier mismatch dramaticaly increased the likelihood that a Japanese speaker would treat a subject NP as the subject of a relative clause. Mismatching clasifiers elicited relative clause completions in over 80% of trials, indicating that despite the indirect nature of the cue that mismatching clasifiers provide, they are nevertheles efective cues for relative clauses, at least when participants have unlimited time to complete the task. In the next experiment we proced to the question of whether Japanese speakers are able to use the clasifier mismatch cue in the more time-sensitive environment of an on-line task. 6. Experiment 2 Building upon the finding in Experiment 1 that a mismatch betwen a clasifier and an adjacent NP provides a strong cue for relative clauses in an off-line seting, Experiment 2 was designed to investigate whether clasifier mismatches are efective cues in an on-line seting. As discussed earlier, Japanese speakers show a general bias to parse embedded subject NPs as the subject of a complement clause. Previous studies have shown that this can lead to a mild-to-moderate garden path efect if the embedded clause is subsequently disambiguated as a relative clause. There are potentialy thre elements that can indicate the presence of a relative clause and therefore induce a garden path efect: one is the relative head NP after the embedded verb; another is the 228 gap before the verb, although this cue is compromised by the fact that Japanese relatively frely alows argument omision; the other is an embedded verb without the complementizer -to. In the literature, some researchers have concluded that the first two elements are reliable disambiguating cues (Inoue, 1991; Mazuka, 1995; Yamashita, 1995; Yamashita, 1993), although the efectivenes of the bare verb cue has not been extensively tested. In one study of simple clause/relative clause ambiguities (Mazuka & Itoh, 1995) it has been argued that the relative head NP provides the only reliable disambiguating information. On the other hand, studies of the gapped/gaples relative clause ambiguity (Yamashita, 1995, Yamashita et al., 1993), suggest that the gap in the argument position before the verb provides disambiguating information, although the efects are typicaly not observed until the relative head NP position. Finaly, the third possibility has not been investigated in the literature, but it is also possible that a verb without the complementizer -to also works as a reliable disambiguator. In a relative clause the embedded verb cannot bear the complementizer -to. Thus, if the parser independently notices that a clause is an embedded clause based on such information as case markers, it is possible that the bare embedded verb may provide a reliable cue for the presence of a relative clause, and thus trigger the same type of garden path efect. We expect the same garden-path efect based on any of these thre factors in on-line reading times for sentences containing a relative clause whose subject is preceded by a matching numeral clasifier. On the other hand, if the subject of the relative clause is preceded by a mismatching numeral clasifier, and if the clasifier mismatch is an efective cue in an on-line task, then we predict a diferent patern. The NP that mismatches the 229 clasifier should be unexpected and should lead to an initial slowdown in reading- times, but if the mismatch is then succesfully used to recognize a relative clause boundary, then reading-times should be faster at the clause-final regions where the relative clause is normaly disambiguated, since the garden-path efect should be avoided. On the other hand, if Japanese speakers are unable to exploit the clasifier mismatch cue on-line or if it delays structure building until it reaches disambiguating information, then the clasifier mismatch should not lead to facilitation in reading times at the end of the relative clause. Participants Sixty-thre native speakers of Japanese participated in the experiment. Al were students at Shizuoka University, Shizuoka Sangyo University, Shizuoka, Japan, or Meiji Gakuin University, Tokyo, Japan. They gave informed consent and were paid the equivalent of $5 for their participation in the experiment, which lasted about 30 minutes. Materials and design Twenty-four sets of two conditions each were used in the experiment. Both conditions contained an embedded relative clause, but varied with regard to the semantic compatibility betwen the subject of the relative clause and a preceding numeral clasifier. In both conditions a main clause subject marked with the topic marker ?wa was followed by a genitive numeral clasifier, which in turn was followed by an adjective and a nominative NP that indicated the start of an embedded 230 clause. In the clasifier match condition the clasifier and the following NP were semanticaly compatible, and in the clasifier mismatch condition the clasifier was semanticaly incompatible with the following NP and was only compatible with the head of the relative clause that appeared a number of regions later. The head of the relative clause always corresponded to the direct object of the embedded verb, and thus the relative clause verb always appeared without an overt direct object. The relative clause verb was always a verb that canonicaly appears with both a direct and indirect object, such as a ditransitive verb like ageru ?give?, watasu ?pas?, or a causative verb. The relative head position and the matrix verb position were separated by two elements, an indirect object and a locative adverbial phrase. This additional material was inserted in order to separate any efects on reading times created at the relative head position from the reading time slowdown typicaly observed at the end of a sentence, i.e., the so-caled wrap-up efect. For this manipulation, we used a verb that is canonicaly asociated with two objects as the matrix verb. The twenty-four pairs of items were distributed betwen two lists in a Latin Square design. Each participant saw exactly one of the lists intermixed with seventy- two filer items in a random order. The filer items were matched with the target items in overal length and complexity, and were counterbalanced with regard to the distribution of genitive and other numeral clasifiers in main and embedded clauses. A sample set of experimental items is shown in (14). 231 (14) a. Clasifier Match Condition 1 2 3 4 5 6 Tanin-wa san-nin-no tosioita sense-ga Atarasi kocho-ni Clas-teacher-Top thre-cl(human)-Gen Aged teacher-Nom New president-Dat 7 8 9 10 1 12 yorokonde Okuta hon-o aru-seto-ni kyositu-de yomase-masita. Gladly Gave bok-Ac a-student-Dat clas-rom-at made-read. ?The teacher made a student read the bok that thre aged teachers gladly gave to the new president at the clasrom.? b. Clasifier Mismatch Condition 1 2 3 4 5 6 Tanin-wa san-satu-no tosioita sense-ga atarasi kocho-ni Clas-teacher-Top thre-cl(bok)-Gen Aged teacher-Nom New president-Dat 7 8 9 10 1 12 yorokonde Okuta hon-o aru-seto-ni kyositu-de yomase-masita gladly Gave bok-Ac a-student-Dat clas-rom-at made-read. ?The teacher made a student read thre boks that an aged teacher gladly gave to the new president at the clasrom.? Procedure The experiment was conducted on Del laptop computers running the Linger software developed at MIT (Rohde 2001-2003). Participants were timed in a phrase- by-phrase self-paced non-cumulative moving window reading task (Just et al. 1982). Sentences were presented using Japanese characters with the font MS Gothic 16 point, with target items segmented acording to the scheme in (14). Al sentences, 232 including the filer items, were presented on a single line. Stimulus segments initialy appeared as a row of dashes, and participants presed the space bar of the keyboard to reveal each subsequent region of the sentence. In order to ensure that participants atended to the stimuli, an argument-verb matching task was presented after each trial. It was not practical to ask yes/no comprehension questions, since many of the experimental sentences could themselves be understood as questions. Following each trial, a verb was displayed on the computer scren followed by two NPs, corresponding to NPs from the experimental sentence, and participants had to indicate which of the NPs was the subject of the verb in the sentence just read by presing one of two keys on the keyboard. In the comprehension task, the two NPs were displayed without case markers, in order to exclude the possibility of answering the question based on the case markers on the NPs. This task was adopted from Nagata (1993) and Aoshima et al. (2004) with a slight modification. In the previous studies only subject NPs were used in this task, but in our experiment both subject NPs and non-subject NPs were presented. This modification was adopted both in order to encourage participants to atend to al words in a sentence, and also because some of the filer sentences had quantificational wh-phrase subject NPs, which were not suitable for the comprehension task. In order to familiarize participants with the subject-verb matching comprehension task, an of-line practice sesion was included before the experiment. In this practice sesion, only thre instances of numeral clasifiers were included among the twenty practice sentences. 233 Results Comprehension acuracy and reading times at each region were analyzed using a repeated-measures ANOVA, with semantic compatibility betwen numeral clasifiers and the adjacent subject NP (match vs. mismatch) as a within-subjects factor. Al data from participants whose comprehension task acuracy was below 70% in total were discarded (n = 9, 16 %). Reading times longer than 3000ms were discarded, afecting 1% of trials. The means and analyses presented below are based on the remaining trials. The average comprehension acuracy among the fifty-four participants who were included in the analysis was 82.5%. Mean acuracy scores did not difer significantly betwen the two conditions. (Fs < 1). Average reading times for each region are summarized in Figure 1. There were significant diferences in reading times betwen the two conditions at Region 4 (embedded subject NP), Region 5 (adjective), and Region 8 (embedded verb), with a reversal in the patern of dificulty betwen Regions 4-5 and Region 8. There were no significant diferences at any other regions (al Fs < 1). 234 Fig. 2. Reading times per region, Experiment 2. At the embedded subject NP in Region 4 there was a slowdown in the clasifier mismatch condition (F1(1,39) = 15.3, MSe = 4897196, p < .01; F2(1,23) = 4.10, MSe = 3113357, p < .05). This efect continued to the following adjective in Region 5, where the slowdown was significant in the participants analysis but not in the items analysis (F1(1,39) = 5.05, MSe = 1207737, p < .05; F2(1,23) = 2.18, MSe = 768905, p = 0.15). At the embedded verb in Region 8, on the other hand, reading times in the clasifier match condition were on average 51 ms slower than in the clasifier mismatch condition (F1(1,39) = 4.15, MSe = 403055, p < .05; F2(1,23) = 4.19, MSe = 580423, p < .05). No other diferences were significant. 235 Discussion The aim of this study was to investigate whether clasifier mismatches could serve as a cue for relative clause structures in an on-line seting, and acordingly whether the parser is able to build head-final relative clause structures incrementaly, rather than defering structure building until the end of the relative clause. The slowdown at the embedded subject NP in the clasifier mismatch condition, which persisted to the adjective in the following region, is unsurprising. A mismatching clasifier-noun sequence is very rare in Japanese, and increased dificulty in reading times should be predicted relative to a matching clasifier-noun sequence, irespective of whether a predictive or a head-driven strategy is adopted for parsing NPs. The more interesting finding of this study is the reversal in the reading-time paterns at the embedded verb in Region 8, where the clasifier match condition was read more slowly than the clasifier mismatch condition. The experimental conditions are identical except for the clasifier manipulation in Region 2, so the reading-time diference at the embedded verb must be a consequence of the clasifier match manipulation. The embedded verb provides two sources of disambiguating evidence for the relative clause structure. First, the lack of the complementizer suffix ?to on the verb indicates that the verb is in a relative clause. Second, we consistently used verbs that select both a direct and an indirect object, and therefore the fact that the verb is not preceded by an acusative-marked direct object provides further support for a relative clause parse (cf. Yamashita, 1995; Yamashita et al, 1993). Of course, the lack of an overt direct object is not an unambiguous cue for a relative clause, since 236 Japanese alows null objects, typicaly in situations where the object is a highly acesible discourse referent. If the nominative NP in Region 4 is analyzed as the subject of a complement clause in the clasifier match condition, as should be predicted based on previous results (e.g., Miyamoto, 2002; Aoshima et al., 2004), then a garden-path efect is predicted when the relative clause verb is encountered. The fact that the verb is read more quickly in the clasifier mismatch condition suggests that the relative clause structure was not unexpected in this condition. This in turn suggests that the clasifier mismatch was sufficient to initiate construction of a relative clause in the on-line seting. The fact that reading times at the disambiguating embedded verb position were facilitated by the clasifier mismatch is consistent with the predictions of an incremental structure building acount of Japanese, in which the relative clause structure can be initiated before the embedded verb is reached. It is les easy to reconcile these results with a head-driven acount, in which the presence of a head of a phrase is necesary for integrating its sub-constituents (Abney 1989; Pritchet 1991a, 1992b). Such an acount would predict that construction of the embedded clause would be delayed until the embedded verb was procesed, and would not readily predict facilitation at that point in the clasifier mismatch condition. The fact that Japanese speakers appear to be able to use the clasifier mismatch cue to initiate building of relative clause structures begs the question of what parsing mechanism alows them to exploit this unambiguous yet highly indirect structural cue. This isue is taken up in more detail in the General Discusion section 237 below. First, however, we consider in more detail the content of the representation that Japanese speakers construct prior to the verb in the clasifier mismatch condition. The evidence from Experiment 2 indicates that a verb that disambiguates a relative clause structure is les surprising following a mismatching clasifier-noun sequence, and we suggested that this reflects earlier initiation of the relative clause structure. However, there are a couple of limitations on this conclusion. First, since the evidence is based on a reading-time facilitation at the verb position, this provides only indirect evidence that the relative clause structure was initiated prior to the verb position. Second, the facilitation at the verb position in this experiment suggests that the relative clause verb is more easily integrated into the existing structure in the clasifier mismatch condition than in the clasifier match condition, but provides only limited information on the extent to which the existing structure in the clasifier mismatch condition already has the structural properties of a relative clause. The next experiments addres both of these limitations by investigating the impact of clasifier- noun mismatches on the procesing of filer-gap dependencies in Japanese, in particular whether mismatching clasifiers activate the island property of relative clauses. 7. Relative Clause Islands in Japanese Japanese, like English, alows long-distance filer-gap dependencies. For example, the dative-marked indirect object of the embedded verb ageta ?gave? canonicaly appears inside the embedded clause as in (15a), but may also appear in sentence-initial position as a result of a proces known as scrambling (15b). Phrases 238 may also be scrambled to other positions betwen the sentence-initial position and the canonical position. (15) a. [Kity-wa [ CP otokonoko-ga Marco 1 -ni hon-o ageta-to] omotta]. Kity-Top boy-Nom Marco-Dat book-Ac gave-Comp thought. b. Marco 1 -ni [Kity-wa [ CP otokonoko-ga ___ 1 hon-o ageta-to] omotta]. Marco-Dat Kity-Top boy-Nom book-Ac gave-Comp thought. ?Kity thought that the boy gave the book to Marco.? Scrambling may target both referential NPs and wh-phrases, as in (15) or in (16). Importantly, unlike in English, the position of a wh-phrase in Japanese does not indicate its scope. As shown in (17), direct and indirect questions are distinguished in Japanese by the presence of a question particle on the main verb or embedded verb, respectively, and not by the position of the wh-phrase. (16) a. [Kity-wa [ CP otokonoko-ga dare 1 -ni hon-o ageta-to] omota-no]? Kity-Top boy-Nom who-Dat book-Ac gave-Comp thought-Q b. Dare 1 -ni [Kity-wa [ CP otokonoko-ga ___ 1 hon-o ageta-to] omotta-no]? Who-Dat Kity-Top boy-Nom book-Ac gave-Comp thought-Q ?Who did Kity think that the boy gave the book to?? (17) a. [Kity-wa Marco-ni [ CP otokonoko-ga dare-ni hon-o ageta-ka] ita]. Kity-Top Marco-Dat boy-Nom who-Dat book-Ac gave-Q said ?Kity said to Marco who did the boy gave the book to.? 239 b. [Kity-wa Marco-ni [ CP otokonoko-ga dare-ni hon-o ageta-to] Kity-Top Marco-Dat boy-Nom who-Dat book-Ac gave-Comp ita-no]? said-Q ?Who did Kity say to Marco that the boy gave the book to?? Scrambling, like wh-movement in English, can create potentialy unbounded filer-gap dependencies. However, scrambling is also subject to a number of the island constraints on dependency formation that restrict wh-movement in English and other languages. In particular, both scrambling and wh-movement may create dependencies that span a complement clause boundary (18a), but may not create dependencies that span a relative clause boundary (18b) (Ross 1967; Saito 1985). Following standard linguistic terminology, relative clauses are known as ?islands? for scrambling. (18) a. Marco 1 -ni [Kity-wa [otokonoko-ga __ 1 hon-o ageta-to] ita]. Marco-Dat Kity-Top boy-Nom book-Ac gave-Comp said ?Kity said that the boy gave the book to Marco.? b.* Marco 1 -ni [Kity-wa [ RC otokonoko-ga __ 1 ageta] hon]-ga sukida]. Marco-Dat Kity-Top boy-Nom gave book-Ac like. ?Kity likes the book that the boy gave to Marco? 240 The islandhood of relative clauses is succesfully captured in many diferent acounts of unbounded dependencies (Kaplan and Zaenen 1989; Kroch and Joshi 1985; Pollard and Sag 1994; Stedman 1996 among many others), and the choice among these acounts does not mater for the argument in this paper. However, as an example, Chomsky?s wel-known subjacency condition (Chomsky 1973) states that filer-gap dependencies may not span more than one bounding node, where the bounding nodes are asumed to be NP and S. A dependency that spans a relative clause boundary violates this constraint, because the relative clause contains an S category and the combination of the relative clause with its head forms an NP category, and thus at least two bounding nodes are crossed. Whatever acount of relative clause islands ultimately proves to be correct, the relevant point for our curent concerns is that the islandhood of relative clauses depends on the specific structural properties of relative clauses. Therefore, the islandhood of relative clauses can be used as a diagnostic of whether detailed relative clause structures are constructed folowing mismatching clasifier-noun sequences. We next turn to a discussion of filer-gap dependency procesing in Japanese, which wil provide a measure of islandhood in Japanese. 8. Procesing Filer-Gap Dependencies in Japanese Much evidence in English and similar languages indicates that when the parser encounters a fronted phrase (i.e., a filer), there is a preference to asociate the fronted phrase with the first potential gap position, leading to a general bias for shorter filer-gap dependencies. This bias has been confirmed using a variety of 241 diferent measures, including the filed gap efect in reading-time studies (Bourdages 1992; Crain & Fodor, 1985; Stowe, 1986), plausibility manipulations in eye-tracking studies (Traxler & Pickering, 1996), ERP studies (Garnsey, Tanenhaus, & Chapman, 1989) and speeded gramaticality judgment studies (McElre & Grifith, 1998). In the filed gap paradigm ilustrated in (19), for example, readers exhibit a slowdown in reading times upon encountering an overt NP (e.g. us) in a position where a gap had been expected. Such efects are taken to indicate that the parser actively posits a gap at the first possible position, without waiting for bottom-up evidence for an empty argument position (Crain & Fodor, 1985; Stowe, 1986). (19) a. My brother wanted to know if Ruth wil bring us home to Mom at Christmas. b. My brother wanted to know who Ruth wil bring us home to __ for breakfast? The bias for shorter filer-gap dependencies observed in English could, in principle, be due to either of a couple of diferent sources. On the one hand, it may reflect a constraint that explicitly favors creating a gap as soon as possible after encountering a filer, irespective of what other constraints are satisfied by positing that gap. Alternatively, the bias may reflect the fact that creation of a gap alows other linguistic constraints to be satisfied, such as confirmation of thematic role asignments, and thus the bias may reflect the parser?s goal of satisfying these other constraints as soon as possible. Although it is dificult to distinguish these alternatives 242 using evidence from English, Aoshima et al. (2004) provide evidence from Japanese in favor of the second alternative, by showing that under certain circumstances Japanese speakers favor longer filer-gap dependencies. A dative wh-phrase that appears at the beginning of a two-clause sentence in Japanese clearly does not occupy its canonical position, and is temporarily ambiguous betwen one of two possible gap positions. It may be asociated with a main clause gap (20a), which would alow for early completion of the filer-gap dependency, but would delay confirmation of a thematic role until the final word of the sentence, due to the verb-final property of Japanese. Alternatively, it may be asociated with an embedded clause gap (20b), which would delay completion of the filer-gap dependency, but would alow for earlier confirmation of the filer?s thematic role, and would make it possible for the question particle to appear on the embedded clause verb. A question particle in Japanese must appear at least as high in the sentence structure as the thematic position of a wh-phrase that it is asociated with. (20) a. Dare 1 -ni [ CP Kity-wa __ 1 [ CP Marco-ga sono-hon-o ageta-to] ita-no]? who-Dat Kity-Top Marco-Nom that-book-Ac gave-Comp said-Q ?To whom did Kity say that Marco gave that book to someone?? b. Dare 1 -ni [ CP Kity-wa [ CP Marco-ga __ 1 sono-hon-o ageta -to] ita-no]? who-Dat Kity-Top Marco-Nom that-book-Ac gave-Comp said-Q ?Who did Kity say that Marco gave the book to ?? 243 Aoshima et al. (2004) show using a number of diferent measures that Japanese speakers prefer to asociate fronted wh-phrases with a gap in the embedded clause. This suggests that the driving force behind filer-gap dependency formation is the satisfaction of gramatical requirements, rather than creation of a gap as an end in itself. One of the measures used to demonstrate this bias, which we draw upon in the current study, is a Japanese counterpart of the Filed Gap Efect paradigm. The examples in (21a-b) both contain two dative-marked NPs, one in the main clause and a second in the embedded clause. It is not unnatural for a Japanese sentence to contain two dative NPs. What is les natural is for a Japanese sentence to include two dative NPs within a single clause. Therefore, the logic of Aoshima and colleagues? Experiment 2 is that if Japanese speakers try to interpret the fronted dative wh-phrase in (21a) as an argument of the embedded clause, and if they do so as soon as they encounter the subject NP that marks the onset of the embedded clause, then they should be surprised to encounter a second, overt dative NP inside the embedded clause (i.e., katyoo-ni, ?asistant?). This is the counterpart of the English Filed Gap Efect observed when readers encounter an overt NP in a position where they had already posited a gap (Crain & Fodor, 1985; Stowe, 1986). The Japanese Filed Gap Efect would be sen as a slowdown in reading times at the embedded dative NP in (21a), relative to the same region in (21b), in which the first dative NP (i.e., senmu-ni, ?managing director?) is not unambiguously scrambled, and is therefore likely to be interpreted as a main clause argument (Kamide & Mitchel, 1999). Using a self-paced reading paradigm, Aoshima and colleagues observed a significant Filed Gap Efect at the second dative NP in (21b), providing further support for the bias for 244 longer filer-gap dependencies in Japanese, and also providing evidence that filer-gap dependencies are constructed in advance of the verb in Japanese. (21) a. Scrambled Condition Dono-syain-ni senmu-wa syatyo-ga katyo-ni syokyu-o yakusokusita-to imasita-ka Which employe-Dat managing director-Top president- Nom asistant manager-Dat raise-Ac promised-DeclC told-Q 'To which employe did the managing director tel that the president promised a raise to the asistant manager?' b. Control Condition Dono-syain-ga senmu-ni syatyo-ga katyo-ni syokyu-o yakusokusita-to imasita-ka Which employe-Nom managing director-Dat President- Nom asistant manager-Dat raise-Ac promised-DeclC told-Q 'Which employe told the managing director that the president promised a raise to the asistant manager?' 9. Filer-gap Dependencies and Relative Clause Islands There is a potential conflict betwen the islandhood of Japanese relative clauses and the finding that Japanese speakers favor longer filer-gap dependencies. Aoshima et al. (2004) demonstrated a bias for Japanese speakers to asociate fronted ambiguous wh-phrases with an embedded clause gap, presumably because this alows for earlier satisfaction of thematic or scope requirements, and they showed that filer- gap dependencies are formed before the embedded clause verb is reached. In that study al of the embedded clauses were complement clauses, which frely alow long- distance scrambling, and thus al of the embedded clause gaps turned out to be gramaticaly aceptable. However, the bias for longer-distance filer-gap 245 dependencies creates a potential danger, since a speaker may form a long-distance dependency into an embedded clause that turns out to be a relative clause. Relative clauses are islands for scrambling (Saito, 1985), and a number of studies have shown that comprehenders avoid forming filer-gap dependencies that cross island boundaries in English (Stowe, 1986; Traxler & Pickering, 1996 among others). However, Japanese speakers may inadvertently construct filer-gap dependencies that violate an island constraint, because relative clauses typicaly cannot be identified until the end of the relative clause, after the point when filer-gap dependencies are constructed. In what follows, we investigate whether this occurs and, more importantly, whether the presence of a clasifier-noun mismatch provides a sufficient cue to block the formation of a long-distance filer-gap dependency. Using the Japanese version of the Filed Gap Efect paradigm it should be possible to use the islandhood of relative clauses to test whether clasifier mismatches activate sufficiently rich relative clause structure for inhibiting filer-gap dependency formation, and whether this structure is available prior to the relative clause verb. We first investigate the interaction of clasifier mismatches and filer-gap dependency formation using an off-line sentence fragment completion paradigm, and then turn to an on-line reading-time study using the Filed Gap Efect paradigm. 10. Experiment 1b This study is an analysis of another dependent measure from the sentence fragment completion study described above in Experiment 1a. In Experiment 1a, we were concerned with the first dependent measure, namely an analysis of how 246 efective clasifier-noun mismatches were as cues for relative clauses. In Experiment 1b we focus on how participants complete filer-gap dependencies, and on how this interacts with the generation of relative clauses, which are islands for long-distance dependency formation. Aoshima et al. (2004, Experiment 3) used a sentence fragment completion task as one measure of the bias for longer filer-gap dependencies in Japanese, in a study where al embedded clauses were potential complement clauses. In the current study we predict that this finding should be be replicated in cases where a matching clasifier-noun sequence introduces an embedded clause, making it compatible with a complement clause analysis. In cases where the embedded clause is introduced by a mismatching clasifier-noun sequence, however, it should be gramaticaly impossible to construct a long-distance filer-gap dependency that spans the relative clause boundary. The aim of this study is to determine whether Japanese speakers are sensitive to this constraint in an off-line generation task. Participants, Materials and Design The details of the experimental design are presented in Experiment 1a above. The participants were the same participants. Here we summarize the second dependent measure that we are concerned with. Materials consisted of 18 sets of six conditions, organized in a 2 x 3 factorial design, as shown in Table 4. In addition to the manipulation of the match betwen the embedded clause subject and a preceding clasifier, we manipulated the form and position of a wh-phrase. In the initial dative condition a dative wh-phrase appeared in sentence initial position in the fragment. In this position it must obligatorily be understood as scrambled, and it is compatible 247 with either a main clause or an embedded clause gap. This corresponds to the condition that was most likely to yield an embedded clause gap in the studies by Aoshima et al. (2004). In the medial dative condition a dative wh-phrase appeared in second position in the fragment, betwen an initial topic-marked NP and the embedded subject NP. This phrase could, in principle, be interpreted either as a main clause in-situ argument or as a phrase that has undergone short-distance scrambling to the front of the embedded clause, but previous evidence suggests that speakers prefer the in-situ analysis (Kamide & Mitchel, 1999; Aoshima et al., 2004). Finaly, in the initial nominative condition the fragment started with a nominative wh-phrase, which can only be interpreted as a main clause argument, due to the ban on scrambling of nominative wh-phrases in Japanese (Saito 1985). Table 4 Sample Set of Experimental Conditions for Experiment 1 a. Clasifier match/initial dative wh-phrase Dono-seto-ni tanin-wa 3-nin-no tosioita sense-ga which -student-Dat clas-teacher-Top 3-Cl(human)-Gen Aged teacher-Nom b. Clasifier mismatch/initial dative wh-phrase Dono-seto-ni tanin-wa 3-satu-no tosioita sense-ga which -student-Dat clas-teacher-Top 3-Cl(bok)-Gen Aged teacher-Nom c. Clasifier match/medial dative wh-phrase Tanin-wa dono-seto-ni 3-nin-no tosioita sense-ga clas-teacher-Top which -student-Dat 3-Cl(human)-Gen Aged teacher-Nom d. Clasifier mismatch/medial dative wh-phrase Tanin-wa dono-seto-ni 3-satu-no tosioita sense-ga clas-teacher-Top which -student-Dat 3-Cl(bok)-Gen Aged teacher-Nom 248 e. Clasifier match/initial nominative wh-phrase Dono-seto-ga tanin-ni 3-nin-no tosioita sense-ga which -student-Nom clas-teacher-Dat 3-Cl(human)-Gen Aged teacher-Nom f. Clasifier mismatch/initial nominative wh-phrase Dono-seto-ga tanin-ni 3-satu-no tosioita sense-ga which -student-Nom clas-teacher-Dat 3-Cl(bok)-Gen Aged teacher-Nom Following Aoshima et al. (2004) sentence fragment completions were analyzed for two measures of where the wh-phrase was interpreted. First, we analyzed the distribution of question particles in the completions. Al wh-phrases must be asociated with a question particle that marks the scope of the question, as either a direct or an indirect question. If the question particle appears on the embedded verb, then the wh-phrase must be asociated with an embedded clause gap. The converse is not true, however, as a main clause question particle may be asociated with either a main clause or an embedded clause gap. Second, we analyzed the distribution in the completions of ditransitive verbs that obligatorily select a dative-marked NP. In al sentence fragments a single dative NP appeared overtly in the main clause. Although it was always posible to generate an embedded ditransitive verb in the completion, in no condition was this a gramatical requirement. Therefore, if any condition shows an increased proportion of embedded ditransitive verbs, this likely reflects an increased number of dative NPs that were interpreted as being asociated with an embedded clause gap. If fronted dative wh- phrases are preferentialy asociated with an embedded clause gap, then we expect to 249 find increased numbers of completions with embedded ditransitive verbs in the initial dative/clasifier match condition. However, if participants respect the islandhood of relative clauses, then the number of embedded ditransitive verbs should not be increased in the initial dative/clasifier mismatch condition. Results First, let us examine the number of Q-particles in the embedded clauses and the matrix clauses. We excluded from the analysis the fragments that were not completed, and completions containing Q-particles in both the embedded and matrix clause. The results are summarized in Table 5. Table 5 Summary of Question Type Question Type Embeded Main Conditions N % Count N % Count GNC Matching 21 23.7 931 591 63.4 931 GNC Mismatching 18 1.7 1054 92 94.1 1054 Total 239 12.04 1985 1583 79.7 1985 In the analysis of question-type, 12.04 % of fragments (239 trials) were completed as indirect questions, with a question marker on an embedded verb only. The proportion of trials in which a question particle was provided on the embedded clause verb was 23.7% for the Matching Conditions (221/931 trials), and 1.7% for the Mismatching 250 Conditions (18/1054 trials). A  2 test showed that the proportion of embedded Q- particle completions was significantly diferent betwen the Matching and the Mismatching Conditions ( 2 (1) =225.21, p < .01). A summary of question-type within each condition is in Table 6. Table 6 Summary of Question-type Question-type Embeded Main Conditions N % Count N % Count GNC Matching/ Scrambling 73 2.8 319 249 63.9 319 GNC Mismatching/ Scrambling 3 0.84 354 34 94.9 354 GNC Matching/ Non Scrambling 19 38.6 308 204 4.8 308 GNC Mismatching/ Non Scrambling 10 2.8 349 36 94.2 349 GNC Matching/ Wh-Nominative 29 9.5 304 138 81.9 304 GNC Mismatching/ Wh-Nominative 5 1.4 351 32 95.1 351 Total 239 12.04 1985 1583 79.7 1985 The proportion of trials in which a question particle was provided on the embedded clause verb was 22.8% (73/319 trials) for the Matching/Scrambling condition, 0.84% (3/354 trials) for the Mismatching/Scrambling condition, 36.8% (119/308 trials) for 251 the Matching/Non Scrambling condition, 2.8% (9/349 trials) for the Mismatching/Non Scrambling Condition, 9.5% (29/304 trials) for the Matching/Nominative Wh condition and 1.4% (5/351 trials) for the Mismatching/ Nominative Wh condition. 2 tests showed that the proportion was significantly diferent betwen Scrambling, Non Scrambling and Wh-Nominative conditions within the Matching condition ( 2 (2)=26.64, p<.01). Pairwise comparison showed that the proportion of Q-particles on embedded verbs was higher in (i), the Non Scrambling condition than in the Scrambling condition ( 2 (1)=15.48, p<.01 and in (i), in the Non Scrambling condition than in the Wh-nominative condition ( 2 (1)=19.80, p<.01). There was no significant diference among conditions within GNC Mismatching conditions. The analyses of verb argument structure are shown in Table 7. I counted the verbs that can take dative NPs as their arguments but not as benefactive phrases. Table 7 Summary of Verb Argument Structure Dative Argument Structure Embeded Main Conditions N % Count N % Count GNC Matching 219 23.5 931 793 85.1 931 GNC Mismatching 84 7.96 1054 1015 96.2 1054 Total 303 15.2 1985 1808 91.08 1985 In the analysis of verb argument structure, 15.3% of fragment completions (303 trials) contained an embedded ditransitive verb that obligatory takes a dative argument. The 252 proportion of trials in which a ditransitive verb was provided in the embedded clause was 23.5% for the GNC Matching Conditions (219/931 trials), and 7.96% for the GNC Mismatching Conditions (84/1054 trials). A  2 test showed that the proportion of embedded ditransitive verb completions was significantly diferent betwen the GNC Matching and the GNC Mismatching Conditions ( 2 (1) =93.67, p < .01). In the analysis of verb argument structure, trials in which a Q-particle was provided on the embedded verb only were separated from the trials in which a Q- particle was provided on the main clause verb. This reflects the general fact that a Wh-NP must be asigned its thematic role and licensed by a Q-particle. The results of verb argument structure with embedded Q-particles only are summarized in Table 8. 253 Table 8 Dative Argument Structure: Embedded Q-particle Only Dative Argument Structure (Embeded Q-particle only) Embeded Matrix Conditions N % Count N % Count GNC Matching/ Scrambling 43 40.1 107 43 58.9 73 GNC Mismatching/ Scrambling 3 60 5 2 6.6 3 GNC Matching/ Non Scrambling 59 41.5 142 72 60.5 19 GNC Mismatching/ Non Scrambling 2 12.5 16 6 60 10 GNC Matching/ Wh-Nominative 16 26.6 60 19 65.5 29 GNC Mismatching/ Wh-Nominative 3 50 6 4 80 5 Total 126 37.5 36 146 61.08 239 The proportion of embedded dative verbs was 40.1% (43/107 trial) for the Matching/Scrambling condition, 60% (3/5) for the Mismatching/Scrambling Condition, 41.5% (59/142) for the Matching/Non-Scrambling condition, 12.5% (2/16) for the Mismatching/Non-Scrambling condition, 26.6% (16/60 trials) for the Matching/Wh-Nominative condition, and 50% (3/6) for the GNC Mismatching/Wh- 254 Nominative condition. 2 tests showed that none of the diferences betwen conditions are reliable. The results of dative argument structure with matrix Q-particle only are summarized in Table 9. 255 Table 9 Dative Argument Structure: Matrix Q-particle Only Dative Argument Structure (Matrix Q-particle only) Embeded Matrix Conditions N % Count N % Count GNC Matching/ Scrambling 31 15.1 204 23 97.8 238 GNC Mismatching/ Scrambling 29 8.6 36 34 98.8 38 GNC Matching/ Non Scrambling 21 15.2 138 158 98.1 161 GNC Mismatching/ Non Scrambling 15 4.6 32 32 98.1 328 GNC Matching/ Wh-Nominative 35 14.0 249 273 97.5 280 GNC Mismatching/ Wh-Nominative 26 7.7 34 325 97.0 35 Total 157 9.91 1583 1645 97.9 1680 The proportion of embedded dative verbs was 15.1% (31/204 trial) for the Matching/Scrambling condition, 8.6% (29/336) for the Mismatching/Scrambling Condition, 15.2% (21/138) for the Matching/Non-Scrambling condition, 4.6% (15/322) for the Mismatching/Non-Scrambling condition, 14.0% (35/249 trials) for the Matching/Wh-Nominative condition, and 7.7% (26/334) for the Mismatching/Wh-Nominative condition. 2 tests showed none of the diferences betwen conditions were reliable. 256 Finaly, we combined information from the analyses of Q-particles and verb argument structure, in order to provide a composite estimate of the proportion of trials in which the dative wh-phrase was interpreted in the embedded clause. The composite figure was based on the proportion of responses in which there was either a Q-particle on the embedded clause verb or an embedded verb that takes a dative argument. Table 10 summarizes this. 257 Table 10 Composite Estimate Composite Estimate Embeded Q-particles and Dative Verbs Embeded Conditions N % Count GNC Matching/ Scrambling 16 36.3 319 GNC Mismatching/ Scrambling 6 1.6 354 GNC Matching/ Non Scrambling 178 57.7 308 GNC Mismatching/ Non Scrambling 12 3.4 349 GNC Matching/ Wh-Nominative 45 14.8 304 GNC Mismatching/ Wh-Nominative 8 2.2 351 Total 365 18.3 1985 The estimated proportion of fronted Dative wh-phrases that were interpreted in the embedded clause was 36.3% (116/319) for the Matching/Scrambling condition, 1.6% (6/354) for the Mismatching/Scrambling Condition, 57.7% (178/308) for the Matching/Non-Scrambling condition, 3.4% (12/349) for the Mismatching/Non- Scrambling condition, 14.8% (45/304 trials) for the Matching/Wh-Nominative condition, and 2.2% (8/351) for the Mismatching/Wh-Nominative condition. 258 A 2 test showed that the estimated proportion of embedded interpretations of wh-phrases was significantly higher in the Matching/Non-Scrambling condition than in the Matching/Scrambling Condition ( 2 (1)=28.89, p<.01). Also  2 tests showed that the estimated proportion of embedded interpretations of wh-phrases was significantly higher in the Matching/Non-Scrambling condition than in the Matching/Scrambling Condition ( 2 (1)= 147.09, p<.01), and in the Matching/Scrambling condition than in the Matching/Wh-Nominative condition( 2 (1)= 22.50, p<.01). There were no other reliable diferences. Discussion There were two main findings in this experiment. Those are summarized in (22). (22) a. Fronted Dative Wh-phrases are not interpreted in the embedded clauses when mismatching clasifiers are provided. b. Wh-phrases are interpreted more often in the embedded clause in the Matching/Non Scrambling condition than in the Matching/Scrambling condition. Let us discuss these findings in turn. First, (22a) is an expected result. This finding is consistent with what we have found in previous experiments, i.e., (i) the mismatching clasifier creates an expectation for an upcoming RC structure, and (i) if an upcoming RC is expected a wh-gap dependency is not created inside the RC to avoid 259 the violation of the RC island. As we have sen, acording to every measurement, the number of wh-phrases that are interpreted in the embedded clause within the mismatching clasifier conditions is smaler than in the other conditions. Thus, this finding supports our view that mismatching clasifiers provide a cue for the RC structure and further that this cue is strong enough to create an expectation for the islandhood of RCs. The second finding in (22b), however, was not expected. Specificaly, the results showing that the number of wh-phrases that are interpreted in the embedded clauses was larger in the Non Scrambling Conditions than in the Scrambling Condition. This finding does not support the findings in either Kamide and Mitchel (1999) or Aoshima et al. (2004). The comparison betwen the Scrambling Condition and the Wh-Nominative condition shows that wh-phrases are interpreted in the embedded clauses in the Scrambling Condition, and are thus compatible with our results in the on-line experiment, i.e., readers prefer the long distance scrambling analysis of the fronted Dative wh-phrases. However, the results of the Non- Scrambling condition conflict with the results of on-line experiment. In our on-line experiment, experiment 3, we created the base line condition using the type of sentences that correspond to the sentences in the Non Scrambling condition. As we have sen in the previous section, there was no Filed Gap Efect in the base line condition. This result suggests that wh-phrases located betwen the matrix subject and the embedded subject are not analyzed as scrambled phrases, rather they are treated as matrix indirect objects. 260 We note two points about this problematic result. The first is that there may be diferences betwen the on-line and off-line experiments. The other is the presence of clasifiers. The first point is relevant to the manner of presenting sentences in the two diferent experiments. In Experiment 1 and Experiment 3 we used highly similar sentences. However, in each experiment the way of presenting the sentences was diferent. In the on-line experiment, sentences were presented in a phrase-by-phrase fashion, and previous words disappeared from the computer scren. Readers could not re-read the previous words. On the other hand, in the off-line sentence fragment completion task, participants were able to read sentence fragments while they completed the sentences. My conjecture here is that such diferences in experiments may have created the diferent results. However what creates those diferences is totaly not clear. More studies are needed to make this point clearer. The second point is relevant to the design or our experiments. There is a possibility that the clasifier creates some problems. Unlike Aoshima et al.?s study or Kamide and Mitchel?s study, our experiments contain clasifiers. The design of the experiment was not exactly the same as in the two previous studies. Thus, to test the validity of those studies? findings, we would have to run another experiment excluding the clasifier. Therefore, it is not reasonable to conclude that Aoshima et al. and Kamide and Mitchel?s study were not correct on the basis of our study. Rather, we need to run diferent experiments excluding the clasifier. In the revised experiment, sentence fragments that do not contain clasifier should be provided. If the existence of a clasifier is the confound, and if the findings of Aoshima et al?s 261 findings are real, we would find a greater number of completions that indicate that the long-distance scrambled wh-phrases are interpreted in the embedded clause and wh- phrases in the scrambling condition than in the non scrambling condition. 11. Experiment 3 The aim of Experiment 3 was to test whether mismatching clasifier-noun sequences are sufficient to induce islands for filer-gap dependencies in an on-line task. If clasifier mismatches can block long-distance dependency formation, this suggests that sufficiently rich structure is built to activate the islandhood of relative clauses. Furthermore, by using the Japanese Filed Gap Efect paradigm in this study, we aimed to test for evidence of relative clause structure being built in advance of the embedded clause verb. Participants Ninety-eight native speakers of Japanese participated in the experiment, al of whom were students at Shizuoka University or Shizuoka Sangyo University, Shizuoka, Japan, or at Meiji Gakuin University, Tokyo, Japan. They al gave informed consent and were paid the equivalent of $5 for their participation in the experiment, which lasted about 30 minutes. Materials and design The experimental materials consisted of twenty-four sets of 4 conditions, organized in a 2 x 2 factorial design that manipulated the semantic congruity betwen 262 the embedded subject NP and a preceding genitive numeral clasifier (clasifier match vs. clasifier mismatch) and the position of a dative-marked wh-phrase (initial dative vs. medial dative). In the initial dative condition a dative wh-phrase appeared in sentence initial position, whereas in the medial dative condition the dative wh-phrase appeared in the second region of the sentence, betwen the topic-marked main clause subject and the embedded clause subject. The sentence-initial dative NP is obligatorily analyzed as scrambled, and was previously shown to be preferentialy asociated with an embedded clause gap (Aoshima et al., 2004). In contrast, the medial dative NP could be interpreted as either an in-situ main clause argument, or as a localy scrambled embedded clause argument. Previous findings suggest that if a Japanese NP can be interpreted as an in-situ argument, speakers prefer this analysis over a scrambling analysis (Kamide & Mitchel, 1999). Therefore, there should be no reason for a Filed Gap Efect in the medial dative conditions, and it serves as a baseline for the initial dative conditions. The twenty-four sets of items were distributed among four lists in a Latin Square design. Each participant saw exactly one of the lists intermixed with seventy- two filer items in a random order. The filer items were matched with the target items in overal length and complexity, and Genitive Numeral Clasifiers, other numeral clasifiers and wh-phrases, in main and embedded clauses were equaly distributed across stimuli and filers. A sample set of experimental materials is shown in Table 11. 263 Table 11 Sample set of experimental materials, Experiment 3 a. Clasifier mismatch/Initial dative Condition 1 2 3 4 5 6 Dono-sense-ni tanin-wa 3-satu-no [tosioita sense-ga atarasi Which-student-Dat clas-teacher-Top 3-Cl(bok)-Gen aged teacher-Nom new 7 8 9 10 1 12 kocho-ni yorokonde okuta] hon-o kyositu-de yomasemasita-ka? President-Dat gladly gave bok-Ac clas-rom-at read-made-honorific-Q ?Which student did the clas teacher made read thre boks at the clasrom that the old teacher gladly gave to the new president?? b. Clasifier mismatch/Medial dative Condition 1 2 3 4 5 6 Tanin-wa dono-sense-ni 3-satu-no [tosioita sense-ga atarasi clas-teacher-Top which-student-dat 3-Cl(bok)-Gen aged teacher-Nom new 7 8 9 10 1 12 Kocho-ni yorokonde okuta] hon-o kyositu-de yomasemasita-ka? President-Dat gladly gave bok-Ac clas-rom-at read-made-honorific-Q ?Which student did the clas teacher made read thre boks at the clasrom that the old teacher gladly gave to the new president?? 264 c. Clasifier match/Initial dative Condition 1 2 3 4 5 6 Dono-sense-ni tanin-wa [3-nin-no tosioita sense-ga] atarasi Which-student-Dat clas-teacher-Top 3-Cl(human)-Gen aged teacher-Nom New 7 8 9 10 1 12 Kocho-ni yorokonde okuta] hon-o kyositu-de yomasemasita-ka? President-Dat gladly gave bok-Ac clasrom-at read-made-honorific-Q ?Which student did the clas teacher made read boks at the clasrom that thre old teacher gladly gave to the new president?? d. Clasifier match/Medial dative Condition 1 2 3 4 5 6 Tanin-wa dono-sense-ni [3-nin-no tosioita sense-ga] atarasi clas-teacher-Top which-student-dat 3-Cl(human)-Gen aged teacher-Nom new 7 8 9 10 1 12 Kocho-ni yorokonde okuta] hon-o kyositu-de yomasemasita-ka? President-Dat gladly gave bok-Ac clasrom-at Read-made-honorific-Q ?Which student did the clas teacher force to read boks at the clasrom that thre old teachers gladly gave to the new president?? Al of the target conditions consisted of twelve regions, segmented into phrases as shown in Table 4. In al conditions the numeral clasifier appeared in the third region, the embedded subject in the fifth region, and the second dative NP in the seventh region. An adverb appeared betwen the embedded dative NP region and the embedded verb region in order to make it possible to diferentiate any delayed reading time slowdown efect due to the embedded dative NP from any efect caused 265 by the embedded verb. As in Experiment 2 the embedded verbs were typical ditransitive verbs like ageru ?give? or watasu ?pas? that normaly appear with both indirect and direct objects. The head of the relative clause in the target conditions always corresponded to the direct object of the embedded verb, and thus the ditransitive verb in the embedded clause always appeared without an overt direct object. Procedure The experiment used a self-paced moving window task using identical parameters to Experiment 2, including the use of an argument-verb matching comprehension task and an off-line practice sesion before the main experiment. Results Comprehension acuracy and reading times at each region were entered into a repeated-measure ANOVA, with clasifier match (match vs. mismatch) and the position of the dative wh-phrase (initial vs. medial) as within-subjects factors. Data from 10 participants (10%) whose comprehension task acuracy was below 70% in total were discarded, leaving eighty-eight participants who were included in the analysis. The average comprehension acuracy among the remaining participants was 81.9%. Mean acuracy scores were not significantly diferent among the four conditions (Fs < 1). Reading times for individual regions that were longer than 3000ms were discarded, afecting 2.5% of trials. Average reading times are shown in Figures 3-4. 266 Fig. 3. Average reading times in the clasifier mismatch conditions, Experiment 3. Fig. 4. Average reading times in the clasifier match conditions, Experiment 3. 267 The reading times in regions 1 through 4 did not show any significant diferences (al Fs < 1). There were either no main efects or interactions. At the embedded subject NP in Region 5 there was a main efect of clasifier match, due to reading times that were 60 ms. slower in the clasifier mismatch conditions (F1(1,87) = 23.5, MSe = 7592484, p < .01; F2(1,23) = 12.5, MSe = 6912526, p < .01). Pairwise comparisons showed that this efect was significant within each level of the clasifier match factor (F1 (1,87) = 5.71, MSe = 1443168, p <. 01; F2 (1,23) = 4.16, MSe = 1653686, p<. 0.5). On the other hand there were no significant efects of the position of the dative wh-phrase, either in the main ANOVA or in pairwise comparisons within each level of the clasifier match factor (al Fs < 1), nor was there an interaction of the clasifier match and the position of the wh-phrase. There were no significant reading time diferences in the other regions (Fs < 1). At the position of the embedded dative NP in the clasifier match conditions there was no significant efect of the position of the wh-phrase at the embedded dative NP in Region 7, there were no significant diferences in reading times (Fs < 1). However, at the adverb position in the imediately following region, region 8, there was a significant main efect of word order in the main ANOVA (F1(1,87) = 4.0, MSe = 309963, p < .05; F2(1,23) = 4.45, MSe =462981, p < .05). However there was no significant main efect of Clasifier Match factors. There was no interaction. Pairwise comparisons of the word order factor revealed that there was a significant slowdown in reading times for Scrambling/GNC Matching condition (F1(1, 87)=3.73, MSe = 404738, p < .05; F2 (1, 23)=5.35, MSe = 651668, p < .05). In the other 268 regions, we did not find any significant diferences (Fs < 1). In the GNC Mismatching conditions, we found no significant diferences at any region betwen the reading times for the two conditions a (Fs < 1). Discussion There were thre main findings in this experiment: The replication of the clasifier mismatch efect; a Filed-gap efect at the preverbal position in the clasifier Matching condition; and no filed-gap efect in the clasifier Mismatching conditions. Let us discuss them in turn. First we replicated the clasifier mismatch efect that we observed in the second experiment. This finding support the claim that Japanese readers can calculate the lexicaly encoded semantic information imediately upon encountering it. Secondly we replicated the Japanese filed-gap efect observed by Aoshima et al (2004). In the Clasifier Matching Condition, there is no indication of the upcoming relative clause. Thus, the parser's initial analysis should be that the embedded clause is a complement clause following its general preference for the complement clause analysis. Because it is analyzed as a complement clause, the fronted Dative Wh-phrase is likely to be interpreted in the most deeply embedded clause as Aoshima et al's study suggests. One concern with this result is the actual position of this reading time slowdown. The eighth region is the region imediately following the Dative NP region. This region is occupied by an adverb. In our design, however, this region is occupied by the same adverb in al conditions. For this reason, it is unlikely that the reading time diferences were caused by some factor intrinsic to 269 each condition. We therefore conclude that this is a spil over efect proceses that took place at the seventh region, the embedded Dative NP region. Let us discuss the consequences of this finding. First, this result replicated Aoshima et al.'s finding in the following two senses. First the fronted Wh-Dative-NP was interpreted in the most deeply embedded clause. Second the Filed-Gap Efect appeared at a preverbal region. These two points, together with Aoshima et al.'s findings, support the position that filer-gap dependencies are constructed in advance of the verb, and that filer-gap dependency formation is motivated by the requirements of the filer. This is, in turn, provides support for the Incrementality Hypothesis. First, the sentence structure is built even before the embedded verb becomes available in the input. This suggests that the parser does not wait for the information from the crucial lexical head. Second, the fronted NP is likely to be interpreted in the embedded clause. This observation suggests that the parser tries to satisfy the requirements of the filer, the thematic requirement, as early as possible. Now let us turn to the finding in the Clasifier Mismatching Conditions. Both in the scrambling and non-scrambling conditions, we do not observe the reading time slowdown at the embedded Dative NP region or at the embedded adverb region. In other words, the filed-gap efect that took place in the Clasifier Matching Conditions is not found in the Clasifier Mismatching Conditions. This result indicates that the parser avoided the relative clause island violation. Let us elaborate this point. Aside from the clasifier mismatch there were no diferences betwen these four conditions. Additionaly, the Clasifier Matching Conditions suggest that the indication of long-distance dependency formation comes 270 at the embedded adverb in region 8. Thus, if the parser created a filer-gap dependency inside the embedded clause, there should be a filed-gap efect compared to the control condition at the embedded Dative NP position or the embedded adverb position as in the Clasifier Matching Conditions. The lack of a reading time slowdown suggests that the parser takes advantage of the Clasifier Mismatch at the earlier point to avoid creating a filer-gap dependency inside the embedded domain, i.e., the parser detected the presence of the relative clause in advance, and further it could compute the consequence of the relative clause structure, i.e., the islandhood of the relative clause. This is so because before the embedded verb or the relative head, there is no indication of the relative clause structure other than the clasifier mismatch. Everything is the same across al the conditions. Thus, we can plausibly conclude that the parser avoided the island violation because of the Clasifier Mismatch that took place at the onset of the embedded clause. One potential alternative to our interpretation of the result is that because of the clasifier mismatch efect participants abandoned procesing the sentence structure thereafter, and just mechanicaly presed the space-bar. There are at least two reasons to question this conclusion. The first is that after the clasifier mismatch region in region 5 and the potential spilover region in region 6 there were no significant reading time diferences among conditions. If mechanical button presing without reading the sentence took place, we might expect that there would have been significant diferences in the reading patern after the clasifier mismatch region, significantly faster reading times would be the most likely case. However, there were no such diferences in reading paterns. The second reason is that there were no 271 significant diferences in acuracy for the comprehension questions. Again, if participants did not read the sentences and just presed the key we should have obtained lower acuracy for the Clasifier Mismatch Conditions. Nonetheles, there were no diferences in acuracy among the four conditions. Taken together, the results strongly suggest that the parser avoided the island violation by means of the presence of the clasifier mismatch at the earlier point. Now let us discuss the theoretical consequences of these findings. Overal, the results of Experiment 3 support the position that the sentence procesor is both strongly incremental and stil gramaticaly precise. The fact that the parser avoided the island violation on-line supports the view that a fully elaborated relative clause structure is built imediately after the Mismatching Clasifier. Furthermore, if the human parser is incremental, it is easy to imagine that it might sacrifice gramatical precision and just build whatever structures based on the localy available information. If this were true, we would obtain the results that indicated that the parser violated the relative clause island. This is so because, as we have sen, local lexical information does not tel the parser that the current structure is a relative clause, until the embedded verb or the relative head becomes available. In summary, our results strongly suggest that the sentence procesor is strongly incremental and it is equipped with a powerful predictive mechanism. With those two properties, the human parser is designed to satisfy both incrementality and gramatical precision at the same time. 272 12. General Discusion 12.1. Procesing of relative clauses The goal of this paper was to take advantage of the gramatical properties of Japanese Genitive Numeral Clasifiers to investigate the extent to which the sentence procesor is strongly incremental and equipped with a powerful predictive mechanism that uses impoverished local information to project the full structural skeleton of an upcoming structure. We have argued that the curious gramatical property of the Genitive Numeral Clasifier potentialy creates a prediction for an upcoming relative clause structure when the clasifier is semanticaly incompatible with its imediately following Nominative NP. Because a Mismatching Clasifier can only be licensed in a relative clause configuration, it may provide an unambiguous cue for an upcoming relative clause structure. While, at the same time, if it is the case the parser can ever create a prediction for an upcoming relative clause structure based on the Mismatching Clasifier, the parser has to employ a rich and complex syntactic inference involving multi-step syntactic computations. Experiments 1 and 2 showed that Japanese readers make use of the Mismatching Clasifier to forese an upcoming relative clause structure. In the of- line sentence completion study in Experiment 1 the participants generated a significantly greater number of sentences containing relative clauses in the Clasifier Mismatch Conditions than in the Clasifier Match conditions. This result indicates that the mismatching clasifier robustly biases Japanese native speakers to generate relative clauses. The on-line self-paced moving window experiment in Experiment 2 showed that Japanese readers make use of the clasifier mismatch to predict the 273 upcoming relative clause structure. The faster reading times obtained at the embedded verb position in the Mismatching Condition indicated that readers projected the relative clause structure upon encountering the clasifier mismatch. This finding also suggests that a relative clause verb, which does not bear the complementizer -to, provides disambiguating information, i.e., readers can notice that a sentence contains a relative clause as soon as they encounter a bare embedded verb that does not have the complementizer. Thus, we observed reading time diferences at the embedded verb position, the first possible disambiguation position for relative clauses. Experiment 1b and 3 showed that the parser avoids the relative clause island violation by means of the clasifier mismatch. This suggests that the parser can compute an elaborated relative clause structure based on the impoverished cue from the clasifier mismatch. The investigations on the second dependent measure in Experiment 1b revealed that the fronted wh-dative NP is rarely interpreted in the embedded clause when the mismatching Genitive Numeral Clasifier is provided. Thus, we obtained asignificantly smaler number of sentence completions that indicated that the wh- dative NP was interpreted in the embedded clause. Experiment 3 used an on-line self-paced moving window experiment to verify that the parser can avoid island violation on-line on the basis of the cue from the clasifier mismatch. In the Clasifier Matching Conditions the scrambling condition showed a filed gap efect at the embedded adverb position compared to the control condition. On the other hand there was no such efect in the Clasifier Mismatching Conditions. The results showed that the parser did not try to create a filer-gap 274 dependency inside the embedded domain. This lack of efect indicates that the parser avoided the relative clause island violation. Putting these pieces together, our results indicate that Japanese readers project the full structural skeleton imediately when they encounter a semantic mismatch betwen the Genitive Numeral Clasifier and an embedded Nominative NP. Crucialy, the parser sems to create sufficiently rich structure in the embedded clause to prevent the fronted NP from being asociated with the embedded verb. This means that the parser preserves both incrementality and gramatical precision at the same time. There can be, in principle, a number of ways to deal with the structural uncertainty of Japanese sentences, in the context of on-line sentence procesing. Taking an example from filer-gap dependencies, if priority is placed on incrementality, a parser can blindly create a gap where the requirement of the filer can be satisfied as early as possible. In this case, the gramatical precision of structure building could be sacrificed. For the sentence procesor, regardles of whether it is inside a relative clause island or not, the embedded verb is always the earliest position where a fronted constituent can receive its thematic role. On the other hand, if the parser's priority is gramatical precision, the parser might defer positing the gap inside an embedded clause until it has clear evidence that the embedded clause is not an island. Under this approach, the parser can avoid the danger of creating a dependency that might turn out to be ungramatical. In this case, however, the incrementality of structure building would be sacrificed. 275 Our experimental results indicate that the parser does not sacrifice either full incrementality or gramatical precision. Rather, by dynamicaly exploiting rich gramatical inferences based on relatively impoverished information from the input, the parser manages to achieve incremental and gramaticaly precise structure building. 12.2. Incremental structure building with a powerful predictive mechanism Although, we have emphasized the necesity of a powerful predictive mechanism that should underlie the incremental sentence procesor, of the specific mechanism that the parser is equipped with has not been discussed in detail so far. Any approach in which the parser can aces gramatical knowledge of Genitive Numeral Clasifiers and relative clauses can, in principle, handle the parser's behavior we have sen so far (se Inoue, 1991 or Inoue and Fodor 1995 for example). Stil, however, the problem of how and at what point in time gramatical information is consulted and the elements in the input are incorporated into a syntactic tre structure remains unclear. As is clearly argued in Lombardo & Sturt (2002), certain extra- lexical knowledge has to be exploited in order to achieve incremental procesing of sentences that we have sen so far. The crucial step in the procesing of these sentences containing Mismatching Clasifiers is that the parser projects a full structural representation for the relative clause at the exact position of the clasifier mismatch, as ilustrated in Fig. 5. 276 Fig. 5. One approach that has extensive power of prediction is the so-caled left-corner parser (Abney and Johnson 1991; Aho and Ulman 1972; Babyonyshev and Gibson 1995; Gibson 1991; Schneider 1999; Stabler 1994 among others). The left-corner parser has various atractive features as a model of the human sentence procesor. However, even a left-corner parser does not capture such masive prediction of upcoming material. Based on lexicaly encoded gramatical information, a left-corner parser can project the two independent nodes of NP and IP based on the Clasifier Mismatch. Even though a left-corner parser can predict the upcoming sister nodes of the given input, as far as this prediction is dependent on the bottom-up information from each lexical item, it is not possible to project and incorporate the nodes that are necesary for the relative clause structure. Thus, for the parser to acomplish the predicted structure in Fig. 5 there must be a syntactic inference in which the localy predicted NP node and the IP node are connected to each other via a CP node, i.e., a prediction based on predicted nodes is necesary, a form of recursive prediction. A chalenge for this type of predictive mechanism is how to constrain the predictive 277 power in such a way that the mechanism does not overgenerate possible structures. Although a mechanism that incorporates recursive prediction is currently not available, our results suggest that the human sentence procesor is designed in such a way. 13. Conclusion In this chapter we have sen evidence for strong incrementality. In the course of discussion, we have argued that to achieve the strong incrementality, the parser is equipped with a powerful predictive mechanism. We have discoverted a potential trigger for the predictive mechanism of sentence procesing. What we have discovered was, the clasifier mismatch, a relatively an indirect cue can be a trigger for the predictive mechanism. We have also discovered that the representation that is built by the predictive mechanism ust be rich enough to represent the islandhood of RCs. We have argued that the predictive mechanism can project the detailed structure of the RC that encodes the information of the islandhood of the RC upon encountering the clasifier mismatch. In the later chapters, we wil se how a specific mechanism of sentence procesing can capture the prediction made by the clasifier mismatch. 278 CHAPTER 4. CONDITIONALS AND LONGER DEPENDENCY FORMATION 1. Introduction The goal of this chapter is to further investigate the nature of the predictive mechanism underlying the human sentence procesor. In the previous chapter, we have sen that the parser makes use of a relatively indirect cue, the clasifier mismatch efect, to predict an upcoming relative clause structure. Based on the results of a series of experiments, we have established the following claims. First, we have sen in the previous chapter that there is a potential trigger for the predictive mechanism, the clasifier mismatch. The experimental results demonstrated that even an indirect cue such as the clasifier mismatch can be used by the parser to predict the upcoming structure. Second, we have sen the representation that the predictive mechanism aceses. Through the investigation of the interaction betwen the Longer Dependency Bias (LDB) and relative clause islands, we have argued that the representation built by the predictive mechanism must be such that it can derive the island efects of relative clauses. In this chapter, we wil turn to another environment where we can test the nature of the predictive mechanism, namely conditional clauses in Japanese. This chapter is devoted to investigate the following two points. First we wil se the nature of the LDB. We wil se what is the driving force behind LDB. The first half of the chapter, therefore, is spent investigating this point. Second, based on the finding of the first half, we wil investigate the procesing of conditional clauses, where LDB is 279 blocked. We wil se that looking into the blocking efect of LDB in conditional clauses wil tel us about the nature of the predictive mechanism. 2. On the Motivations for the Longer Dependency Bias In chapter 2, we have examined how the longer dependency is constrained. We have confirmed that LDB is best understood in terms of the human sentence procesor?s general property of completing a dependency as early as possible. We also confirmed that the LDB can be best captured by the constraint satisfaction model (Boland et al. 1995; Boland and Boehm-Jernigan 1998; Macdonald et al. 1994a; MacDonald et al. 1994b; Tanenhaus et al. 1989 among many others) in such a way that the gramatical requirement on the fronted element is satisfied as early as possible (Aoshima et al. 2004). Although the general direction has been made clear in the previous chapter, we have not discussed what is actualy the ?gramatical requirement? on the fronted phrase that needs to be satisfied as early as possible. As is pointed out in Aoshima et al. (2004), there are several possibilities. Bearing in mind that the fronted NPs are al wh-phrases bearing Dative Case in the series of experiments that we have sen, it is obvious that there are at least four posibilities: it could be the thematic requirement (theta criterion: Chomsky 1981); it could be wh- feature satisfaction (the requirement of wh-Q binding: Harada 1972; Saito 1989); it could be Case feature satisfaction (structural Case asignment: Chomsky 1981; Saito 1982, 1983; Takezawa 1987); or it could be a combination of these features. One of the goals of this chapter is to examine which feature(s) on the fronted NP drives the formation of the longer dependency. Through two sets of experimental 280 studies, we wil conclude that the motivation for the LDB is the establishment of an asociation betwen a wh-phrase and a Question Particle (Q-particle) as soon as possible. The first set of experiments is concerned with a manipulation of the features of the fronted NPs. We wil se what happens if we change the fronted NP from a wh- NP to a non-wh-NP (referential NP). We wil review the results of two experiments and show that only the wh-NP creates the LDB. The second set of experiments wil be concerned with the predicted verbal morphology. Specificaly, we wil ask whether the LDB is created if the predicted verbal morphology cannot host a Q-particle. To examine this point, we wil consider the procesing of Japanese conditional constructions. Another purpose of this chapter is to se how the structure is predicted in head-final constructions during online procesing, through a detailed examination of the experimental findings on conditional clauses. I wil show that the interaction betwen long-distance scrambling and the syntactic properties of conditional clauses gives us a clue to exactly what is predicted by what kind of element, i.e., we can se a trigger of the predictive mechanism, and the representation that the predictive mechanism aceses. Experiments 1 and 2 are concerned with the driving force behind the longer dependency bias. In two experiments, we test the behavior of the sentence procesor when a property of the fronted materials is manipulated, incorporating Aoshima et al.?s (2004), and Aoshima?s (2003) Japanese filed-gap paradigm as a probe for the LDB. Specificaly, we test the wh-scrambling construction, and construction containing the scrambling of referential NPs. The results of these experiments 281 demonstrate that the wh-feature plays a crucial role in creating the longer dependency. Experiments 3 through 6 are concerned with the trigger and the representation employed by the predictive mechanism. The main purpose of these experiments is to test the behavior of the sentence procesor when the predicted upcoming structure does not license the wh-feature of the fronted phrase. For this purpose, Japanese conditional clauses provide a good testing ground. By taking advantage of the gramatical properties of Japanese conditional clauses, we wil first show what kind of bottom-up information can be a trigger for the structural prediction of conditional clauses, and what is actualy predicted by the predictive mechanism. Two experiments (an online and an ofline experiment) are devoted to answering these questions. The results of these experiments wil reveal that the conditional adverb mosi ?by any chance? sets a prediction for the upcoming conditional verbal morphology. After establishing the trigger and the contents of the prediction, we wil test how the parser behaves when the predicted upcoming structure is not compatible with the feature caried by the fronted phrase. In particular, by taking advantage of the fact that the conditional verbal morphology is not compatible with the Q-particle and the fact that Japanese conditional clauses are not islands, we can test whether a longer dependency is preferentialy formed when the conditional verbal morphology is predicted by means of the adverb mosi. The other two experiments are devoted to testing this point. The results of these experiments demonstrate that when the 282 upcoming verbal morphology that cannot host Q-particle is predicted, the LDB is blocked. Taken together, these two sets of experiments reveal a potential trigger for the predictive mechanism and the representation acesed by the predictive mechanism, as wel as the driving force for the LDB in general. 3. The Source of the Bias for Longer Filer-Gap Dependencies in Japanese 27 3.1. Longer Dependency Bias To understand the logic of the experiments, it is helpful to review the motivation for the LDB again here. First let us summarize Aoshima et al.?s (2004) basic asumptions. Their acount of the LDB appeals to the folowing two important asumptions concerning the sentence procesing mechanism. The first asumption is principle-based parsing, whereby gramatical principles are directly incorporated into the mechanism of sentence procesing in order to build structure (Berwick 1991; Crocker 1994, 1996; Pritchet 1992b). The second asumption is the so-caled constraint satisfaction model (Boland et al. 1995 among many others) in a general sense, in which al categories are asociated with lexical constraints, and al categories can initiate atempts to satisfy those constraints. Thus, each lexical item in the input can potentialy initiate the structure building needed in order to satisfy the lexical constraints that it introduces. 27 This section is a sumary of a paper with the same title (Aoshima et al., 205) read at the 18 th Anual CUNY Conference on Human Sentence Procesing, at the University of Arizona, Tucson. I am grateful to Sachiko Aoshima and Colin Philips for alowing me to include these studies in this disertation. 283 Within this view, gramatical constraints such as the theta criterion or wh-feature satisfaction are understood as lexical constraints of lexical items in the input. Adopting these two asumptions, Aoshima and colleagues argue that the interaction of the basic word order of Japanese sentences, i.e., head-finality and the possibility of long-distance scrambling plays a crucial role for creating the LDB. Because of the basic word order of Japanese sentences, the most deeply embedded verb is normaly 28 the first verb to be encountered during the procesing of a sentence from left-to-right, as the schematic structure in (1) shows. (1) [Subject [Subject Object Verb-comp] verb] 28 Note that there is always a posibility that an embeded clause is post-posed as in right dislocation construction as in (i). However, this type of sentences requires special focus interpretation and special intonation paterns. Thus it is not likely that parser?s initial choice would be the right-dislocated structure rather than the simple embeded clause structure. (i) Walt-ga ita(-yo), Sue-ga sukida-te. W-nom said S-nom like-comp ?What walt said was that he likes Sue.? A careful examination is of course necesary about what can be the parser?s initial choice and what canot. But in this study, I continue to asume that the parser?s initial choice is always a complement clause if there is no obvious cue for other alternatives. 284 This, in turn, means that the position of the most deeply embedded verb is the first position where gramatical constraints conveyed by each lexical item can potentialy be satisfied or confirmed by an explicit lexical head. Second, if a long-distance scrambled NP is involved, it creates a temporary ambiguity with respect to its potential gap position. Given the string in (2) for example, during the online procesing of a sentence, there can be two possible gap positions where the Dative Wh-phrase can potentialy receive its theta role. One is the indirect object position in the matrix clause. A verb like iu ?say? can take a dative NP as its indirect object as wel as a clausal complement. Thus, a gap in the matrix clause as in (2a) is a possible option. The other possibility is to locate the gap in the embedded clause. In this case, the embedded verb is most likely to be a verb like ageru ?give? that can take a Dative NP as its object. 285 (2) Wh-NP-Dat [Subject [Subject ? a. Wh-NP-Dat [Subject [gap] [Subject Object V] SAY-Q] b. Wh-NP-Dat [Subject [Subject [gap] Object GIVE-Q] V-Q] If the first option is taken, the Q-particle that must be asociated with the fronted wh- phrase can only be atached to the matrix verb, and the sentence is unambiguously interpreted as a direct wh-question (if the sentence-initial clause is confirmed as the matrix clause). Japanese wh-interogative sentences have a wel-known restriction such that the Q-particle which is understood as a complementizer must c-command 286 the wh-phrase that it licenses (Hagstrom 1998; Harada 1972; Saito 1989). Thus, because of this gramatical constraint, the Q-particle that is asociated with the fronted wh-phrase cannot be atached to the embedded verb. On the other hand, if the gap is located in the embedded clause, the Q-particle that licenses the wh-phrase can be atached to either the embedded verb or the matrix verb, depending on the interpretation of the sentence. If the Q-particle is atached to the embedded verb, the sentence is interpreted as an indirect wh-question. If it is atached to the matrix verb, the sentence becomes a direct wh-question. The most important diference betwen these two options is the possibility of ataching the Q-particle to the embedded verb. If the gap is located in the matrix clause, the Q-particle cannot be atached to the embedded verb, but if it is in the embedded clause, the Q-particle can be atached to the embedded verb. Now, let us look at this temporary ambiguity created by long-distance scrambling and head finality from the perspective of the gramatical constraint satisfaction model. From the viewpoint of the gramatical constraint satisfaction model, the fronted dative wh-phrase can initiate an atempt to satisfy its gramatical constraints. It should be clear that this element is asociated with at least thre types of requirements that we can rephrase in terms of gramatical features: a thematic feature; a Case feature; and a Wh-feature. If any of these features triggers structure building so that they can be satisfied, the earliest point where they are satisfied is the most deeply embedded verb position because it is the first verb to be encountered. From this, the LDB naturaly follows. Because it is the most deeply embedded verb that can satisfy and confirms the gramatical features on the fronted element at the 287 earliest point, gap creation in the most deeply embedded position is the prefered option. 3.2. Manipulating the Trigger In the above discussion, it becomes clear that in Aoshima et al.?s acount, it is the features that are conveyed in the fronted element that drive structure building, so that these features wil be satisfied. Now, let us return to our original question: what is the real motivation for the LDB? As we have sen, the fronted dative wh-NP has to cary at least thre features: a theta role; case; and a wh-feature. Thus, if the LDB is driven by gramatical feature satisfaction, it is one of these features or a combination of these features that triggers the longer dependency formation. Given that if an NP is not case marked it cannot be scrambled, and that gap creation is obviously triggered by the non-canonical word order that is signaled by the order of case-marked NPs, it is clear that Case feature plays an important role in initiating structure building (se Aoshima 2004 for detailed discussion on this point). Thus, it should be the case that the motivation for the LDB is a combination of the Case feature and either the theta role or the Wh-feature, or both. Now the question is how we can test these possibilities. Recaling that it is not the case that only wh-phrases can be scrambled in Japanese, the answer should be quite simple: we can just manipulate the wh-feature on the scrambled NP. In other words, we make a minimal pair of wh-scrambling and the scrambling of a referential NP, and we can test whether the wh-feature is the 288 crucial feature for the LDB or not. Let us se the basic logic by taking a look at the following examples with schematic structures. (3) a. Which-NP-Dat [Subject [Subject ? Verb-Q] Verb] Dono-gakusee-ni [Walt-wa [Master Yhudi-ga which-student-dat W-top M-Y-nom ringo-o ageta-ka] tazuneta. apple-acc gave-Q ask-past ?Walt asked which student Master Yuhudi gave the apple to.? b. (That-)NP-dat [Subject [Subject ? Verb-Q] Verb] (Sono-)gakusee-ni[Walt-wa [Master Yuhdi-ga that-student-dat W-top M-Y-nom ringo-o ageta-ka] tazuneta. apple-acc gave-Q ask-past ?Taroo asked whether Jiroo gave the apple to that student or not.? (3a) is the same construction as Aoshima et al. originaly tested. There, the fronted phrase is a wh-phrase. (3b), on the other hand, contains a fronted phrase that does not bear a wh-feature. This construction is diferent from (3a) only in the wh-feature of the fronted phrase. Given these two constructions, what we expect is the following. First, in (3a) we expect that a longer dependency be preferentialy formed as Aoshima et al. originaly observed. On the other hand, in (3b) we expect that the longer dependency should be formed if the motivation for the LDB is confirmation of the thematic role of the fronted phrase because both in (3a) and in (3b), the fronted phrase 289 must bear some theta roles due to the Theta Criterion (Chomsky 1981). However, if the wh-feature, or a combination of the wh-feature and the theta role is crucial for the LDB, we do not expect a longer dependency to be formed in (3b), simply because the fronted NP does not bear a wh-feature. 3.3. The Japanese Filed-Gap Paradigm and the Motivation for the Longer Dependency Bias As we did in the earlier studies on RCs, we wil adopt the so-caled Japanese Filed-Gap Paradigm (Aoshima et al. 2004). In order to se the specific predictions, it is helpful to briefly review the Japanese Filed-Gap Paradigm. Also, we wil se another question that is related to the design of the original Japanese filed-gap experiment that is not our main concern, but stil an important question to be investigated. 3.3.1. Aoshima et al., (2004) and Aoshima (2003) First of al, the basic intuition behind the Japanese Filed-Gap Paradigm is as follows. If a longer dependency is created, the gap should be created in the most deeply embedded clause during the procesing of a sentence. If there is an additional dative phrase besides the fronted dative phrase in the embedded clause, a surprise efect occurs. To test this, Aoshima et al., made use of the structure in (4). In this structure, there is a dative NP in the embedded clause in addition to the fronted wh- dative-NP. In this structure, if the gap is created and it is asociated with the fronted Wh-dative, the embedded clause has to contain two dative NPs. In Japanese such a 290 construction containing two dative NPs in one clause is not common. Thus, if the fronted wh-dative is interpreted in the embedded clause, Japanese readers do not expect another dative NP to be present in the embedded clause, resulting in a reading time slowdown upon encountering the second dative NP. 291 (4) Wh-Scrambling Condition Dono-shain-ni senmu-wa shachoo-ga ? which-worker-dat senior-director-top president-nom ? Aoshima et al. and Aoshima (2003) observed this efectby comparing the structure in (4) with two types of baseline conditions (5). (5) a. Wh-Nominative Condition Dono-shain-ga senmu-ni shachoo-ga? which-worker-nom senior-director-dat president-nom 292 b. Medial Wh-Dative Condition Senmu-wa dono-shain-ni shachoo-ga ? senior-director-top which-worker-dat president-nom (5a) is a straightforward baseline condition. Given that Japanese does not normaly alow scrambling of Nominative NPs (Saito 1985), there is no moved element involved in this structure. Therefore, we do not expect any long-distance filer-gap dependency formation, and we do not expect any surprise efects asociated with moved phrases. Thus, we predict that sentences with the structure of (4) wil show significant reading time slow down at the second dative NP position compared with (5a). Aoshima (2005) obtained the same results in the comparison of (4) and (5b). (5b), however, is a litle bit trickier as a base-line condition compared to (5a). (5b) contains a wh-dative phrase that is located in betwen two subject NPs. It has been independently observed that in this construction the wh-dative-NP is not interpreted in the embedded clause even though it is gramaticaly possible (Kamide and 293 Mitchel 1999). In other words, although a string like (6) can potentialy be parsed in two ways with respect to the position of the dative wh-phrase (the matrix argument analysis and embedded argument analysis respectively), somehow the parser pursues the matrix argument analysis in (6a). (6a) is the structure of the matrix analysis where the dative wh-phrase is understood as the indirect object of the matrix verb such as iu ?say?. On the other hand in (6b), the dative wh-phrase is understood as a scrambled indirect object of an embedded verb like ageru ?give?. 294 (6) Senmu-wa dono-shain-ni shachoo-ga? senior-director-top which-worker-dat president-nom a. Matrix analysis b. Embedded analysis For this reason, we do not expect that a dependency betwen the dative wh-phrase and a gap is formed in the medial wh-dative condition, and thus we do not expect any efects asociated with the moved phrase either. The results reported in Aoshima?s studies confirm the prediction we have sen above. The construction containing a fronted dative wh-phrase at the beginning of the sentence showed slower reading time at the position of the second dative NP compared to both of these baseline conditions. 295 3.3.2. Manipulating the Features on the Fronted Phrase To test whether the wh-feature is the crucial feature for the LDB, we can make use of the same type of baseline conditions as Aoshima et al. and Aoshima (2005) did. First, we can use the thre conditions above, the wh-scrambling condition in (4), the wh-nominative condition in (5a) and the wh-medial condition in (5b), to replicate Aoshima?s original results i.e., the wh-scrambling condition showed slower reading time at the second dative NP position compared to the baseline conditions. Furthermore, the comparison betwen (5a) and (5b) wil alow us to clarify Kamide and Mitchel?s observation that a medial dative-phrase is not interpreted as a scrambled embedded argument. Our specific prediction about this comparison is that there wil be no significant diference in reading times at the second dative NP position betwen (5a) and (5b). By manipulating the wh-feature of these thre conditions, we can gain the conditions that do not contain wh-phrases. In so doing, we just have to delete the wh- determiner dono ?which? as in (7). (7) a. Scrambling Condition Josee-shain-ni senmu-wa female-worker-dat senior-director-top shachoo-ga ? presidento-nom 296 b. Medial Referential NP condition senmu-wa Josee-shain-ni senior-director-top female-worker-dat shachoo-ga ? presidento-nom c. Nominative Referential NP condition Josee-shain-ga senmu-ni female-worker-nom senior-director-dat shachoo-ga ? presidento-nom Looking at the examples (7b) and (7c), it is obvious that they are the same type of constructions except for the particles on the subjects: one is a nominative Case marker and the other is a topic marker. Despite the diferences in the markers on the subjects, the sentence initial NPs in both of the examples are most likely to be interpreted as matrix subjects. Thus, if we are not interested in the diferences betwen nominative-marked NPs and topic-marked NPs, the comparison betwen (7b) and (7c) is not informative for our current purpose. In order to test whether the wh-feature is crucial for the LDB, it is sufficient to compare (7a) and (7b). If the wh- feature based hypothesis or the wh-feature and thematic requirement based hypothesis is correct, we do not predict any slowdown at the second dative NP position in (7a) compared to (7b). If, on the other hand, the thematic requirement hypothesis is correct, we do expect the slowdown at the second dative NP position in (7a) compared to (7b). 297 3.4. Experiments 3.4.1. Experiment 1 This experiment was designed to test how wh-phrases are interpreted in wh- scrambling structures, using Aoshima et al.?s (2004) and Aoshima?s (2005) Japanese Filed-Gap paradigm. The aim was to determine whether a fronted wh-phrase in Japanese is preferentialy interpreted in the main clause or in the embedded clause, clarifying the results from Aoshima and colleagues, and the aim was also to determine whether a medial wh-phrase is preferentialy interpreted in the main clause or in the embedded clause, clarifying Kamide and Mitchel?s observation. If a fronted wh-phrase is preferentialy interpreted in the embedded clause, we should observe the same filed-gap efect as in Aoshima?s series of studies. The same holds true for the medial wh-phrase. If it is interpreted in the embedded clause, we should observe the filed-gap efect also. Participants Thirty-seven native speakers of Japanese participated in the experiment. Al of them were students at Hiroshima University, Japan. They were paid $5.00 or its equivalent for their participation in the experiment, which lasted about 30 minutes. Materials and design Twenty-four sets of thre conditions each were used in the experiment, which manipulated the position of the wh-phrase (scrambled vs. medial vs. nominative). The 298 24 sets of items were distributed among thre lists in a Latin Square design. Each participant saw exactly one of the lists intermixed with 72 filer items in a random order. The filer items were matched with the target items in overal length and complexity. A sample set of experimental items is shown in (8) where superscripts indicate the regions used in presenting the sentences. 299 (8) A sample set of experimental items a. Wh-Scrambling Dono-shain-ni 1 senmu-wa 2 shachoo-ga 3 which-worker-dat 1 senior-director-top 2 president-nom 3 kaigi-de 4 buchoo-ni 5 shookyuu-o 6 meeting-at 4 department-director-dat 5 promotion-acc 6 yakusoku-sita-to 7 sirase-masi-ta-ka? 8 promice-did-comp 7 notice-hon-past-Q 8 ?Which worker did the senior director noticed that the president promised the promotion to the department director at the meting?? b. Wh-Medial Senmu-wa 1 dono-shain-ni 2 shachoo-ga 3 senior-director-top 1 which-worker-dat 2 president-nom 3 kaigi-de 4 buchoo-ni 5 shookyuu-o 6 meeting-at 4 department-director-dat 5 promotion-acc 6 yakusoku-sita-to 7 sirase-masi-ta-ka? 8 ?Which worker did the senior director notified that the president promised the promotion to the department director at the meting?? 300 c. Wh-Nominative Dono-shain-ga 1 senmu-ni 2 shachoo-ga 3 which-worker-nom 1 senior-director-dat 2 president-nom 3 kaigi-de 4 buchoo-ni 5 shookyuu-o 6 meeting-at 4 department-director-dat 5 promotion-acc 6 yakusoku-sita-to 7 sirase-masi-ta-ka? 8 promice-did-comp 7 notify-hon-past-Q 8 ?Which worker notified the senior director that the president promised the promotion to the department director at the meting?? In the wh-scrambling condition and the wh-medial condition, a main clause subject NP marked with the topic-marker -wa was followed by an embedded clause introduced by a nominative marked NP. In the wh-nominative condition, the sentence initial wh-phrase bore nominative case, and the dative marked wh-phrase followed it. The dative marked wh-phrase was followed by an embedded clause, which began with a nominative-marked NP. We adopt the wel-supported supposition that the nominative marked NP provides a strong cue for the onset of an embedded clause (Inoue 1991; Miyamoto 2002 among others). The position of the embedded clause in these conditions is perfectly natural in Japanese, because Japanese is a head-final language, the main clause verb normaly appears at the end of the sentence, following the embedded clause. Al the conditions contained an overt dative-marked NP. The verb of the embedded clause was marked with a declarative complementizer, and the verb of the main clause with a question particle, indicating that the sentence should be interpreted as a matrix wh-question. 301 In al thre conditions, the dative NP in the embedded clause was the second dative NP in the sentence. In the scrambled condition, the dative wh-phrase at the beginning of the sentence should also have been asociated with the embedded clause, if a longer dependency is preferentialy formed. If the fronted wh-phrase in the scrambled condition is asociated with the embedded clause, then readers should be surprised to encounter the second dative NP. In the other two conditions, the two sentence initial NPs were matched to the scrambled condition in the respect that there was one wh-phrase and one dative NP. However, in these two conditions, we did not expect that either of the two NPs in the matrix clause would be asociated with the embedded clause. In the wh-medial condition, the wh-phrase was located after the matrix subject NP and the dative marked NP in this position was not expected to be asociated with the embedded clause for the reason we have sen above (se Kamide and Mitchel 1999 for details). In the wh-nominative condition, on the other hand, the sentence initial nominative wh-phrase should not be asociated with the embedded clause, since Japanese normaly does not alow the scrambling of nominative NPs (Saito 1985 among others). Procedure The experiment was conducted on Del laptop computers running Linger developed by Douglas Rohde at MIT (Rohde 2001-2003). Participants were timed in a phrase-by-phrase self-paced non-cumulative moving-window reading task (Just et al. 1982). The segmentation indicated with superscripts in (8) was used in the 302 presentation. The complementizer and Q-particle were presented together with the verbs because both the complementizer and Q-particle are bound morphemes in Japanese. Sentences were presented using Japanese characters. Stimulus segments initialy appeared as a row of dashes, and participants presed the space bar of the keyboard to reveal each subsequent region of the sentence. In order to guarante that participants atended to the stimuli, a subject-verb matching task was presented after each trial. This task was adopted from Nagata (1993), and similar to the task used in the study by Miyamoto and Takahashi (2000, 2003). A verb was displayed on the computer scren followed by two agent NPs, corresponding to the topic-marked NP and the nominative-marked NP in the target sentence, and participants had to decide which of the NPs was the subject of the verb in the sentence just read by presing one of two keys on the keyboard. In the comprehension task, the two agent NPs were displayed without case-marking, in order to exclude the possibility of case-based question answering strategies. In order to familiarize participants with the subject-verb matching task, an off- line practice sesion was included prior to the self-paced reading task. Data Analysis Analyses were conducted on comprehension task response acuracy and reading times. Al data from participants whose comprehension task acuracy was below 80% in total were discarded. 303 Comprehension task accuracy Among the 37 participants included in the analysis, the average comprehension acuracy was 82.5%. Mean acuracy scores did not difer significantly across the thre conditions (al Fs < 1). Self-paced reading The reading time analysis yielded the following results. Reading times for al thre conditions are shown in Fig. 6, and a comparison of the scrambling condition and the wh-nominative condition is shown in Fig. 7, and a comparison of the wh- medial condition and the wh-nominative condition is shown in Fig. 8. At the third region, there were significant diferences betwen reading times in the wh-scrambling and wh-nominative condition (F1(1, 36)=6.48, p<0.01 ; F2(1, 23)=2.86, p=1). Other than that, at al regions prior to the fifth region, there were no significant diferences in reading times among al thre conditions. At the following regions we observed a Filed-Gap Efect, reflected in longer reading times for the embedded dative NP, only in the scrambled word order. At the embedded dative NP (region 5), there was a significant main efect of word order type (F1 (1, 36)=3.79, p < 0.05; F2 (1, 23)=4.03, p < 0.05). Pairwise comparison revealed that the dative NP was read significantly more slowly in the Wh-scrambling condition than in the Wh-nominative condition (F1 (1, 36)=3.79, p < 0.05; F2 (1, 23)=4.03, p < 0.05). On the other hand, the comparison betwen either the Wh-scrambling condition and the Wh-medial condition, or betwen the Wh- 304 medial condition and the Wh-nominative condition showed no coresponding slowdown (al Fs < 1). There were no other significant diferences in reading times. Fig. 6. Fig. 7. 305 Fig. 8. Discussion The main result of this experiment is that a slowdown in reading time was observed in the wh-scrambling condition at the embedded dative NP (region 5), relative to the reading time for the same NP in the Wh-nominative condition. Aoshima et al.?s (2004) results are thus replicated by this result. Following Aoshima et al., we interpret this slowdown as the Japanese counterpart of the Filed Gap Efect (Crain, 1985; Stowe, 1986). The comparison betwen either the Wh-scrambling condition and the Wh- medial condition, or betwen the Wh-medial condition and the Wh-nominative condition does not show a slowdown at the second dative NP position. The comparison betwen the wh-scrambling and wh-medial conditions is problematic. The results that we found in the relative clause experiments are not replicated. We do not have a clue for why the diference betwen the wh-medial condition and wh- scrambling condition was not significant. On the other hand, the comparison betwen 306 the wh-medial condition and the wh-nominative condition suggests that the dative wh-phrase in the medial position is not interpreted in the embedded clause, replicating Kamide and Mitchel?s finding and compatible with Aoshima?s results. The significant slowdown at the embedded subject position (region 3) in the wh-nominative condition is apparently problematic. However, this problem is only apparent. In this condition, both subject NPs are Nominative NPs, but in the wh- scrambling condition the first NP is a topic-marked NP. As is observed in various places, a sequence of two nominative NPs generaly creates dificulty in reading (Babyonyshev and Gibson 1995; Gibson 1998; Lewis 1996; Miyamoto 2002; Uehara and Bradley 2002 among others). Thus, it is likely that this slowdown reflects the fact that the condition contains two nominative subjects. Finaly, the Wh-nominative condition did not show a slowdown at the second dative position compared to the other conditions. This result indicates that the referential dative NP located in betwen two subject NPs is not interpreted as a scrambled NP. This again, supports the claim by Kamide and Mitchel that a dative NP in that position is not asociated with the embedded clause. This result indeed supports Kamide and Mitchel?s position, but there is a concern regarding the diference betwen the medial dative wh-phrase and medial referential dative NP. As we have sen, we did not observe a significant diference in reading time betwen the wh-scrambling condition and the wh-medial condition at the second dative NP position. Because we did not observe a significant diference betwen the wh-medial condition and the wh-nominative condition either, we cannot state anything definitely, but the above observation may suggest that if the medial dative NP bears 307 the wh-feature, it may potentialy be asociated with the embedded clause. However, at this point, we do not have any strong evidence that supports this view. 3.4.2. Experiment 2 This experiment was designed with the goal of investigating how long- distance scrambled referential NPs are procesed. It is important to se long-distance scrambled referential NPs, in order to understand what is the motivation for the LDB. The findings in the previous experiment suggest that if a fronted phrase is a wh- phrase, the LDB is created and the Japanese Filed Gap Efect is observed as a consequence. The question here is whether the same result is obtained if the fronted NP does not bear a wh-feature. If the fronted referential NP creates a LDB, it suggests that the wh-feature is not the most crucial factor for the LDB. On the other hand, if the LDB is not created by a fronted referential NP, it suggests that the wh-feature plays a more important role for creating the LDB than the other features that the fronted NP conveys, or at least it plays a crucial role. It may be the case that two features are equaly important for LDB. 308 Participants Twenty-seven native speakers of Japanese participated in the experiment. Al of them were students at Hiroshima University. They were paid $5.00 or its equivalent for their participation in the experiment, which lasted about 30 minutes. Materials and Design Experimental materials consisted of 24 sets of sentences with two conditions each, which we refer to as the scrambled condition and the control condition, respectively. (9) shows one set of conditions used in the experiment. (9) A sample set of experimental conditions a. Scrambled Condition Josee-shain-ni 1 senmu-wa 2 shachoo-ga 3 female-worker-dat 1 senior-director-top 2 president-nom 3 kaigi-de 4 buchoo-ni 5 shookyuu-o 6 meeting-at 4 department-director-dat 5 promotion-acc 6 yakusoku-sita-to 7 osieta. 8 promice-did-comp 7 noticed 8 ?The senior director notified to a female worker, that the president promised a promotion to the department head at the meting.? 309 b. Control Condition Senmu-wa 2 Josee-shain-ni 1 shachoo-ga 3 senior-director-top 2 female-worker-dat 1 president-nom 3 kaigi-de 4 buchoo-ni 5 shookyuu-o 6 meeting-at 4 department-director-dat 5 promotion-acc 6 yakusoku-sita-to 7 osieta. 8 promice-did-comp 7 noticed 8 ?The senior director notified to a female worker that the president promised the promotion to the department head at the meting.? In both conditions, an embedded clause follows the two main clause NPs at the start of the sentence. The embedded clause contains an overt dative marked NP. The embedded verb is marked with a declarative complementizer. In both conditions, the dative NP in the embedded clause was the second dative NP in the sentence. In the scrambled condition, if the fronted dative NP is asociated with the embedded clause, readers should be surprised to encounter a second dative NP, giving rise to a Filed Gap Efect. On the other hand, if the sentence initial dative NP is not asociated with the embedded clause, we do not expect the surprise efect. In the control condition, the two sentence-initial NPs were matched with the scrambled condition in the respect that there was one dative NP and one nominative NP. As we have discussed earlier, in this condition, we have no expectation that either of these NPs should be asociated with the embedded clause. The sentence initial nominative NP cannot be analyzed as a scrambled phrase, and the dative NP in 310 this position, in betwen the matrix and embedded subject, is not asociated with the embedded clause, as Kamide and Mitchel and our previous experiment suggest. Therefore, the second dative NP in the control condition should not be interpreted as if it were scrambled from the embedded clause. Procedure The self-paced reading procedure and the comprehension task were identical in format to that used in Experiment 1. Data analysis Analyses were conducted on comprehension task response acuracy, item acuracy, and reading times. Al data from participants whose comprehension task acuracy was les than 80% were discarded (n = 4, 12.9 %). Results Comprehension acuracy and reading times at each region were entered into a repeated-measures ANOVA, with word order as the within-subject factor. Comprehension task accuracy Among the participants who were included in the analysis, the average comprehension acuracy was 85.5%. The average correct response percentage did not difer significantly betwen the two conditions (F < 1). Self-paced reading 311 Reading times for al regions are shown in Fig 4. At al regions there were no significant diferences betwen reading times in the scrambled and control conditions (Fs < 1). Fig. 9. Discussion The main result of this experiment is that a slowdown in reading time was not observed in the scrambled condition at the embedded dative NP (region 5), relative to the reading time for the same NP in the control condition. We interpret this absence of a slowdown as showing that the fronted NP is not asociated with the embedded clause, and therefore did not result in a Filed Gap Efect. This result support Kamide and Mitchel?s claim, and also strengthens the view that a nominative NP is not preferentialy interpreted as a scrambled phrase, because Nominative NPs cannot be long-distance scrambled in Japanese. 312 3.5. General Discusion For our purposes, the most important finding in these two experiments is the contrast betwen the scrambling of wh-phrases and that of referential NPs with respect to Filed Gap Efect. As should be clear from the discussion above, when the fronted material is a wh-phrase, a Filed Gap Efect is observed. However, if the fronted NP is a referential NP, the Filed Gap Efect is not observed. This contrast suggests that wh-feature on the fronted NP is an important pre-condition for the LDB. Looking at this contrast from the perspective of the Gramatical Constraint Satisfaction model, the result can be interpreted as follows. The driving force behind longer dependency formation is wh-feature satisfaction. In other words, the parser tries to asociate a fronted wh-phrase with the licensing Q-particle as early as possible. As we have discussed, the verb in the most deeply embedded clause is the first verb that can host a Q-particle, so the parser tries to create a gap in the most deeply embedded clause if the fronted phrase is a wh-phrase. This result clearly contradicts the prediction under the view that the parser?s structure building is driven by thematic role satisfaction (Pritchet 1992 among many among others). If the parser tries to satisfy the thematic requirement on the fronted NP, we do not expect the contrast betwen wh-scrambling and referential NP scrambling because for both cases the most deeply embedded verb is the first position where the thematic relation can be confirmed. On the other hand, the result is compatible with the view that the combination of the wh-feature and thematic role is crucial for the LDB. The above two experiments do not tease these two possibilities apart. Later in this chapter, we wil se that there is an environment where we can test 313 which is the more appropriate view. To test the prediction of these two possibilities (the wh-feature alone or the combination of the wh-feature and the thematic requirement) we have to find an environment in which either the theta role can be asigned but wh-feature cannot be satisfied in the embedded clause or the wh-feature can be satisfied but the theta role cannot be asigned. Shortly we wil se that Japanese conditional clauses are indeed an environment that is compatible with the first possibility where the thematic requirement can be satisfied but the wh-feature cannot be satisfied. Finaly let us briefly discuss why a medial wh-phrase is not asociated with the embedded clause. The question is why the thematic requirement on the fronted phrase does not sem to drive longer dependency formation. Here I would like to point out one possibility, namely that the thematic role of an NP can be read off from the Case marker. Japanese NPs, in most cases, bear overt Case particles. These Case particles can encode which Case an NP is asigned. However, such information as gramatical functions or thematic roles sems to be imposed on the overt Case particle too. For example, if there is a sequence of NPs that bear Nominative Case and Dative Case respectively, readers can gues which verb may follow. If there is no other material than, say, an Acusative marked NP as in (10a), a verb like ageru ?give? is most likely to follow these NPs. If, on the other hand, clausal material follows these NPs, the matrix verb is most likely to be a verb like iu ?say?. In both of these cases, however, the theta role of the dative NP is something like the goal or the recipient. Thus, just looking at the Case maker can tel readers a lot about thematic role of the NPs. 314 (10) a. Walt-ga Sue-ni ringo-o ageta. W-nom S-dat apple-acc gave ?Walt gave an apple to Sue.? b. Walt-ga Sue-ni [Subject ?V-to] itta. W-nom S-dat V-comp said ?Walt said to Sue that ?? Now, looking at this acount from the perspective of the Gramatical Constraint Satisfaction model, it follows naturaly that the thematic role does not drive longer dependency formation. When the parser hits the dative NP and the nominative NP, the parser creates a scrambling structure (Aoshima et al. 2004 and Aoshima 2003 among many others). At this point, the structure with the basic word order in Japanese is recovered in which the nominative NP is folowed by the dative NP. As we have sen, the thematic role of the dative phrase can be reliably guesed. Thus, the thematic requirement is virtualy satisfied at this point, and there is no reason that the parser has to do something more to resolve the thematic requirement of the fronted NP. Although this is an important discussion to think about the nature of structure building, this point is not directly related to our main purpose, i.e., the motivation for the LDB and the nature of predictive mechanisms. So I wil not go into further details about this topic and leave this part informal. For more detailed discussion in the reader is refered to Aoshima et al., (2006). 315 4. Conditionals and Long-Distance Dependency Formation 29 This section investigates the following two points. First, we wil try to resolve the remaining problem that arose in the previous section about what is the real driving force behind the LDB: whether it is the wh-feature on the fronted NP or whether it is the wh-feature and the thematic role. The second isue is to investigate the nature of the predictive mechanism. In particular, we wil investigate what is the trigger of the mechanism and what is the representation that is acesed. In order to tackle these two isues, we wil take a detailed look at Japanese conditional clauses. 4.1. The Question Let us first briefly review our problem. There are several possible motivations for the LDB. We have considered the following possibilities: the thematic requirement, Case licensing, the asociation of wh-phrases with Q-particles, or a combination of these. Based on two experiments, we drew the tentative conclusion that the wh-feature of the fronted NP is crucial for the LDB. However, the experiments could not tease apart whether the wh-feature is the crucial feature or whether it is the combination of the wh-feature and the thematic requirement that is crucial. This is the primary problem that we are trying to resolve. We pointed out that to test the predictions of each possibility we would need to have an environment where either the theta role can be asigned but the wh-feature cannot be satisfied in 29 This section is a sumary of a poster with the same title (Yoshida et al., 206) presented at the 19th CUNY Conference on Human Sentence Procesing, at the City University of New York, New York. I am grateful to the co-authors Sachiko Aoshima, Hajime Ono and Colin Philips for alowing me to include this material in this work. 316 the embedded clause or where the wh-feature can be satisfied but not the theta role. We wil se that Japanese Conditional Clauses are appropriate testing grounds for this. 4.2. Japanese Conditionals from a Real Time Perspective Let us briefly summarize the properties of Japanese Conditional Clauses and why they are appropriate for testing the two competing possibilities. In addition, let us point out another and more important problem for us that Japanese Conditionals raise, namely the problem of the nature of predictive mechanisms. The details of the syntax of Japanese Conditionals are explored in Chapter 4. In this section, we wil se thre properties that are most directly related to our study. First, let us point out the non-islandhood of Japanese conditionals. Despite being adjuncts, Japanese conditionals are not islands. Taking a look at the following examples, it becomes clear that scrambling out of conditionals is not degraded compared to the other cases such as extraction from RCs. (11) summarizes examples of extraction out of thre diferent domains: Conditionals, RCs and Complement Clauses. In these examples, scrambling of dative marked wh-phrases is shown because it is most relevant for our purposes, but the same patern holds for the scrambling of referential NPs. 317 (11) a. Dono-gakusee-ni 1 Walt-wa [(mosi) Sue-ga t 1 which-student-dat W-top mosi S-nom purezento-o ageta-ra] naki-dasu-no? present-acc give-cond cry-start-Q Lit. ?Which student wil Walt cry if Sue gives the present to?? b. *Dono-gakusee-ni 1 Walt-wa [ NP [ RC Sue-ga t 1 which-student-dat W-top S-nom ageta] purezento]-ga suki-na-no? gave present-nom like-be-Q Lit. ?Which student does Walt like the present that Sue gave to?? c. Dono-gakusee-ni 1 Walt-wa[ CP Sue-ga t 1 purezento-o which-student-dat W-top S-nom present-acc ageta-to] omotte-iru-no? gave-comp think-ing-Q? Lit. ?Which student does Walt think that Sue gave the present to?? The fact that NPs can be extracted out of conditional clauses means that the extracted NP can receive its thematic role from the embedded verb in the conditional clause. Second, conditional clauses can be introduced by the optional conditional adverb mosi ?by any chance? as (11a) shows. Importantly, this adverb can only be licensed by conditional clauses. Thus, if mosi is present, the clause introduced by mosi must be a conditional clause. 318 Third, conditional clauses are uniquely distinguished from other clauses by their verbal morphology. In other words, if a clause is a conditional clause, the verb has to have a conditional marker. Conditional markers are suffixes such as -ra, -nara, -ba, -naraba and others. These sufixes corespond to the complementizer if in English. The second and the third properties consequently derive the following property that if a clause is introduced by mosi, it must have one of the conditional markers on the embedded verb. To capture this relation betwen mosi and the conditional adverb, we can rephrase their relation in a way like mosi is licensed by the conditional marker. Most importantly, the conditional verb, or the verb bearing the conditional marker cannot host question particles. In a complement clause, the verb can bear a Q- particle if the embedded clause is an interogative clause. However, the question particle cannot co-occur with the conditional markers even in a case where a wh- phrase is contained in the conditional clause. 319 (12) a. [mosi Walt-ga nani-o tabeta-ra(*-ka)] mosi W-nom what-acc eat-cond(*-Q) Lit. ?if Walt eats what? b. [ CP Walt-ga nani-o tabeta-ka] W-nom what-acc ate-Q Lit. ?? what Walt ate? c. d. We can understand this property morpho-syntactialy. The question particle and the conditional markers are in complementary distribution. We can capture this complementary distribution by asuming that the question particle and the conditional marker occupy the same slot in the complementizer system. Now it should be clear that these four properties of conditional clauses can provide us with a good testing ground for the question we raised, i.e., which feature is the driving force for the LDB. From the perspective of real time sentence procesing, conditional constructions containing wh-scrambling such as (11a) create a situation in which the thematic feature of the fronted wh-phrase can be satisfied at the embedded verb position in the conditional clause, but the wh-feature cannot. First, in an example like (11a), the long-distance scrambling of the wh-phrase creates a temporary ambiguity with respect to the position of the gap, i.e., the matrix 320 clause analysis and the embedded clause analysis, as we have sen before in the case of complement clauses. The structure of each analysis is in (13). 321 (13) a. Matrix Clause Analysis Wh-NP-dat Subject [gap] [mosi Subject Object V-ra] V-Q b. Embedded Clause Analysis Wh-NP-dat Subject [mosi Subject [gap] Object V-ra] V-Q In the same way as in the embedded clauses in our previous studies, the embedded verb is the first verb to be encountered during the procesing of a sentence. In an example like (11a), the conditional clause is located after the matrix subject. 322 The crucial diference from the earlier cases is, however, that in the conditional clauses the embedded verb cannot host the Q-particles. This means that the wh-Q asociation cannot be established in the conditional clauses. Furthermore, if the conditional clause is introduced by the adverb mosi, it must be licensed by the conditional-marked verb, which is not compatible with a Q-particle. Thus, if the presence of mosi can signal a conditional-marked verb, and if the speakers can compute this consequence imediately, it becomes clear at the position of mosi that the wh-Q asociation cannot be established in the conditional clause. Asuming that mosi can signal the upcoming conditional structure, we can make the following prediction with respect to formation of the longer dependency. If the LDB is motivated by the thematic requirement of the fronted NP, a longer dependency can be formed because the thematic requirement can be satisfied in the conditional clause due to the non-islandhood of conditionals. On the other hand, if the wh-feature drives longer dependency bias, we expect that LDB wil be created in a sentence like (11a). As we have discussed, a conditional verb cannot host a Q-particle. Thus, as long as the parser does not atempt yet another embedding, such that an interogative complement clause is constructed inside the conditional clause, the wh-Q asociation cannot be established in the conditional clause. Given the dificulty of multiply center-embedded structures in Japanese (Babyonyshev and Gibson 1995 among others), it is not likely that the parser should try to construct another embedded clause inside the conditional clause. Thus, in this case, the LDB is not created due to the fact that wh-Q asociation cannot be established in the embedded clause. 323 The acount above incorporates two important asumptions. First, the adverb mosi can set a prediction of the upcoming conditional clause structure. Second, the prediction or the predicted structure refers to the conditional verbal morphology. Only if these two conditions are fulfiled does the acount above hold. If the LDB is not observed, then the parser makes a prediction for the upcoming conditional verb. If the LDB is observed, then either the prediction is not made, or the hypothesis that the wh-feature is the motivation for the LDB is wrong. If the later is the case, we are forced to reconsider the interpretation of the experimental results we have sen earlier. 324 4.3. Experiments There are four experiments that investigate the points outlined in the previous subsection. Two of them, one offline and one online experiment, are devoted to testing whether the adverb mosi helps the parser to predict an upcoming conditional clause. Building upon the results of these first two experiments, the remaining experiments were designed to probe the behavior of the parser in sentences that contain fronted dative wh-phrases and conditional clauses. 4.3.1. Methodology and Predictions In order to understand the logic behind the experiments, it is always helpful to review the specific predictions along with the methodologies that we employ as we have done before. The methodologies involved in the first set of experiments are a sentence fragment completion task and a phrase-by-phrase self-paced reading task. In the offline experiment, we provided pairs of fragments of sentences. One member of the pair contained the adverb mosi, and the other contained standard adverbs that do not have any special relation to specific constructions. If the presence of mosi helps native speakers of Japanese create a prediction for conditional clauses, we expect they complete the fragments with conditional clauses. On the other hand, standard adverbs are not expected to create such a bias because they do not bear a special relation to conditional clauses. 325 (14) a. Subject mosi Subject ______ b. Subject Adv Subject ______ Basicaly the same logic applies to the online experiment. If the adverb mosi creates a prediction for an upcoming conditional structure during online sentence procesing, we expect to observe facilitation efects or surprise efects at the conditional verb position. Because only the verbal morphology can unambiguously distinguish conditionals from other embedded clauses, we expect to se a reading time diferences at the conditional verb position. Specificaly, we predict that the conditional verb wil be read more quickly in the mosi condition compared to the control condition, where mosi is not provided. (15) a. Subject [mosi Subject Object Verb-Cond] V b. Subject [Adv Subject Object Verb-Cond] V The remaining two experiments are devoted to testing the interaction of long- distance scrambled wh-phrases and the prediction for a conditional clause. The offline experiment is designed to test whether a fronted wh-phrase is preferentialy asociated with an embedded conditional clause or not. If the wh-phrase is asociated with a conditional clause, we expect that sentence fragments should be completed with conditional clauses in which the embedded verbs are those that can take the fronted wh-phrase as one of their arguments. This bias should be observed in the adverb/scrambling condition due to the LDB, which is basicaly the same type of 326 condition that Aoshima et al. originaly tested. Because there is no special adverb in this condition, it should be preferentialy interpreted as a construction containing a complement clause. This can be tested by comparing the completions betwen the types of fragments in (16). (16) a. Wh-NP-Dat Subject mosi Subject ______ b. Wh-NP-Dat Subject Adv Subject ______ Another online experiment is designed to test whether a fronted wh-phrase is preferentialy interpreted in the embedded conditional clause or not. Adopting the Japanese Filed Gap Paradigm, we expect that if the adverb mosi creates a prediction for an upcoming conditional verb, and if the driving force for the longer dependency bias is the wh-feature, we do not expect to observe a Filed Gap Efect at the second dative NP position compared to the control condition. Schematicaly the sentence looks like (17a). However, if the driving force is the thematic requirement, a Filed Gap Efect is expected. Thus, in this way we can pin down the crucial feature that motivates the LDB. On the other hand, if mosi is not provided, we do not expect that the conditional clause should be predicted. Thus, we expect to observe a Filed Gap Efect in this case. A schematic representation of the relevant sentences is shown in (17b). (17) a. Wh-NP-Dat Subject [mosi Subject NP-dat NP-ac V-cond] V-Q b. Wh-NP-Dat Subject [Adv Subject NP-dat NP-ac V-cond] V-Q 327 4.3.2. Experiment 3: Sentence Fragment Completion Task An offline sentence fragment completion task was conducted in order to test whether the presence of mosi creates a bias toward the generation of conditional clauses. If adverbs like mosi work as reliable cues for an upcoming conditional structure, we expect that there should be significantly more conditional clause completions compared to a control condition where such adverbs are not provided. Participants 40 native speakers of Japanese participated in the experiment. Al of them were students at Hiroshima University, Japan, who were paid $5.00 or its equivalent for their participation in the experiment, which lasted about 30 minutes. Materials and design Experimental materials consisted of 12 sets of 2 conditions each (the mosi condition and the control condition respectively). The fragments in these conditions consisted of a topic-marked NP, an adverb, an adjective and a nominative-marked NP. The mosi condition contained conditional adverbs such as mosi and others that are only legitimate in the environment of conditional clauses. The control condition, on the other hand, contained a degre adverb that modified the imediately following adjective that was asociated with the nominative-marked NP. The nominative- marked NP was provided in order to signal the onset of an embedded clause. Previous studies have shown that a nominative NP can strongly indicate the beginning of an 328 embedded clause (Inoue 1991; Miyamoto 2002 among others). Thus we expected that al the completions of the target conditions should have involved biclausal structures. The 12 sets of target items were distributed betwen two lists in a Latin Square design. Each participant saw exactly one of the lists intermixed with twenty- four filer items in a random order. The filer items were designed in such a way to avoid speakers? potential strategy to use the same structure in al completions. (18) Sample set of experimental conditions for Experiment 3 a. Mosi Condition NP-top mosi Adj NP-nom Sono-buchoo-wa mosi hunarena buka-ga that-department-chief-top mosi uninitiated subordinate-nom b. Control Condition 30 NP-top Adv Adjective NP-nom Sono-buchoo-wa shooshoo hunarena buka-ga that-department-chief-top a-litle-bit uninitiated sub.-nom Results This study yielded a total of 456 codable sentence fragment completions. The results from two participants had to be discarded due to the fact that more than half of their responses were left blank or ungramatical. The completions were clasified 30 Although, in roman transcription, the number of characters or length of the adverb loks diferent in the two conditions, in Japanese characters using Kanji characters, the length or number of characters are perfectly matched. 329 acording to the number of conditional clauses used to complete the sentence fragment. The results are summarized in Table 12. Table 12 Overal, 226 out of 456 completions contained an embedded conditional clause (49.5%). The proportion of conditional completions was higher in the Mosi condition (88.5%, 202/228 trials) than in the Control condition (10.5%, 24/228 trials). A ? 2 test showed that the diference was significant (?????????? 2 (1)=277.95, p < 0.01). This result confirms that mosi biases speakers to generate conditional clauses. Completions clasified as ?others? in al conditions contained blanks (18.6%, 43/230), simple clauses (7.3%, 17/230) simple complement clauses (49.1%, 113/230 trials), relative clauses (6.9%, 16/230 trials), other adjunct clauses than conditionals (13.9%, 32/230 trials), embedded wh-interogative construction (0.8%, 2/230 trials) and matrix wh-interogative construction (0.4%, 1/230 trials). Conditions Clause Type Conditionals Others N % N % Mosi Condition 202 88.5 26 11.4 Control Condition 24 10.5 204 89.4 Total 226 49.5 230 50.4 330 Discussion The main finding of this study is that Japanese speakers show a strong preference to generate conditional clauses if adverbs like mosi are provided. This result thus supports that claim that mosi sets a prediction for an upcoming conditional clause. 4.3.3. Experiment 4: Self-Paced Reading Task Building upon the finding in Experiment 3 that mosi creates a strong bias for the completion of a conditional clause, Experiment 4 was designed with the goal of investigating the time-course of the prediction of conditional clauses. We informaly stated that mosi creates a bias for conditional clause completion in the previous subsection. More specificaly, this means that if mosi is provided, the embedded verb bears conditional morphology because only the verbal morphology can distinguish conditional clauses from the other embedded clauses. Now, from the perspective of real-time sentence procesing, we expect the following reading-time patern. If mosi sets a prediction for an upcoming conditional verb, a conditional verb in a sentence containing mosi should be read more easily than a conditional verb in a sentence that does not contain mosi, because the sentence completion measure indicated that other adverbs than mosi do not create a bias for a conditional clause completion and thus they are not likely to set a prediction for an upcoming conditional structure. 331 Participants Twenty-six native speakers of Japanese participated in the experiment. Al of them were students at Hiroshima University, Japan. They were paid $5.00 or its equivalent for their participation in the experiment, which lasted about 30 minutes. Materials and design Sixten sets of two conditions each were used in the experiment, manipulating the type of adverbs (Mosi and Degre Adverb). The 16 sets of items were distributed betwen two lists in a Latin Square design. Each participant saw exactly one of the lists intermixed with 72 filer items in a random order. The filer items were matched with the target items in overal length and complexity. A sample set of experimental items is shown in (19). Again the superscripts indicate regions. (19) a. Mosi Condition Sono-kakarichoo-wa 1 mosi 2 hunare-na 3 that-senior-director-top 1 mosi 2 uninitiated 3 buka-ga 4 torihikisai-ni 5 songai-o 6 subordinate-nom 4 customer-dat 5 casualty-acc 6 ataeta-ra 7 sonoba-de 8 jishoku-o 9 give-cond 7 immediately 8 resignation-acc 9 negaideru-daroo. 10 ask-will. 10 332 ?That senior director wil imediately ask for his resignation if his somewhat inexperienced employee causes casualties to the customer.? b. Control Condition Sono-kakarichoo-wa 1 shooshoo 2 that-senior-director-top 1 a-little-bit 2 hunare-na 3 buka-ga 4 torihikisai-ni 5 uninitiated 3 subordinate-nom 4 customer-dat 5 songai-o 6 ataeta-ra 7 sonoba-de 8 casualty-acc 6 give-cond 7 immediately 8 jishoku-o 9 negaideru-daroo. 10 resignation-acc 9 ask-will. 10 ?That senior director wil imediately ask for his resignation if his a somewhat inexperienced employee causes casualties to the customer.? In these two conditions, a main clause topic-marked subject was imediately followed by an embedded clause, which began with either an adverb of mosi-type or a degre adverb that modified the adjective asociated with the embedded subject NP. In both of the conditions, the embedded subjects were preceded by adjectives and the embedded verbs bear conditional morphemes. In order to ensure that any reading time efects asociated with the embedded verb would not be confounded with efect at the main verb, an adverbial phrase and an acusative-marked NP were inserted in betwen the embedded and matrix verbs. 333 Procedure The self-paced reading procedure and the comprehension task were identical in format to that used in Experiment 1 and 2. Data analysis Analyses were conducted on comprehension task response acuracy, and reading times. Al data of participants whose comprehension task acuracy was les than 80% in total were discarded (n = 4, 13, 3 %). Reading times longer than 2500ms were discarded. This procedure afected 2.0 % of trials. The means and analyses presented below are based on the remaining trials. Results Comprehension acuracy and reading times at each region were entered into a repeated-measures ANOVA, with adverb type (mosi and control) as within-subject factors. Comprehension task accuracy Among the subjects who were included in the analysis, the average comprehension acuracy was 85.4 %. The average correct response percentage did not difer significantly across the two conditions (F < 1). 334 Self-paced reading Reading times for al regions are shown in Fig. 10. Except for the second, seventh and tenth regions, there were no significant diferences betwen reading times in the Mosi condition and the control condition (al Fs < 1). Fig. 10. At the adverb in region 2, the Mosi condition was read 34 ms more slowly than the control condition The efect was significant both in the participant analysis and item analysis (F1(1,25)=5.62, P < 0.05; F2(1, 15)=4.41, P < 0.05). The region 7 yielded a significant diference in reading time (the diference is 27 ms.) in participant analysis but not in the items analysis (F1(1, 25)=4.56, P < 0.05; F2(1, 15)=1.62, P < 1). This reading pater goes to the opposite direction from that observed 335 at the region 2. Finaly, the last region, region 10, showed a significant diference in reading time (the diference is 186 ms.) both in participant and item analysis (F1(1,25)=11.25, P < 0.05; F2(1, 15)=5.3, P < 0.05). Region 7, the critical region for us, which contains the conditional verbs, yielded a significant diference in reading time in the participants analysis but not in the items analysis. Although there is a possibility that region 8, the adverb region, reflect spil over efect from the imediately preceding region, we did not obtained a significant efect there either. However, stil the control condition yielded numericaly slower reading times than the Mosi condition. Taking this fact into consideration, we calculated the reading time diferences betwen the two conditions by combining the two regions, region 7 and region 8. Our expectation is if these two regions crucialy reflect the efect of the manipulation of Mosi, we would be able to gain a more powerful efect by combining these two regions. However, the result did not reach a significant efect in the item analysis (F1(1, 25)= 4.86, P < 0.05; F2(1, 15)= 1.47, P < 0.2). Discussion There were two main findings in this study. First, the conditional adverbs of mosi-type were read slower than the degre adverbs. Second, this reading patern reversed at the embedded verb region, where the Mosi condition was read more quickly than the control condition. We interpret this result as the reflection of the prediction of the upcoming conditional structure by means of the adverb mosi. In other words, the adverb mosi efectively sets a prediction for an upcoming conditional 336 structure, and as a result the verb bearing the conditional morphology was read more easily. The slowdown at the adverb mosi reflects either the fact that it a rare adverb compared to degre adverbs, or the fact that conditional clauses are expected. In any case, the overal reading time profile is best understood as showing that mosi provides an efective cue for an upcoming conditional structures. Note that this result acords wel with the result from the offline sentence fragment completion experiment. Taken together, we can conclude that mosi is an efective cue for an upcoming conditional structure. Although the final region also yielded a significant efect, it is dificult to infer anything from this region because it is the matrix verb region and the end of the sentence where the so-caled wrap-up efect typicaly takes place. For this reason, it is not so clear whether this region reflect the manipulation on the adverb at the beginning of the embedded clause. 4.3.4. Experiment 5: Self-Paced Reading Task: Japanese Filed Gap Paradigm Now the stage is set for investigating our main isue, i.e., testing the motivation for the LDB, and the representation employed by the predictive mechanism. Building on the findings in Experiment 3 and 4 that the conditional adverb mosi provides an early warning for an upcoming conditional structure, Experiment 5 was designed with the goal of investigating whether fronted wh-phrases in Japanese are preferentialy asociated with an embedded conditional clause or not. 337 The finding in the previous experiment that conditional verbs were read faster in the mosi condition than in the control condition indicates that the abstract syntactic structure of the conditional clause and the morphology of the conditional verb were foresen by the parser, by means of the presence of mosi. However, it is not obvious whether the consequences of the conditional morphology can also be anticipated, specificaly the fact that the conditional verbal morphology is not compatible with Q- particles. What we expect is the following. If the parser calculates the consequence of the prediction of the upcoming conditional verbal morphology, and if the driving force for longer dependency formation is the wh-feature we should not observe the LDB in conditional clauses. If on the other hand, either of the above hypotheses is not true, we may observe the LDB in conditional clauses. Even if it is the case that the wh-feature is the driving force behind the LDB, if the parser cannot calculate the consequence of the prediction, it is also possible that the parser may try to asociate fronted wh-phrases with the embedded clause, expecting that a Q-particle may be present on the embedded verb. On the other hand, if the driving force is the thematic feature, rather than the wh-feature, it is again posible that the parser might try to asociate a fronted wh-phrase with the embeded clause in order for the fronted phrase to receive its theta role, because conditional clauses are not islands for long- distance scrambling (se Chapter 2 for details). In order to test these competing hypotheses, this experiment adapted the Japanese Filed Gap paradigm. We used Filed Gap efects to probe for gap creation. 338 If the gap is created, we wil observe Filed Gap Efect by providing a lexical dative phrase in the embedded clause. Participants Sixty native speakers of Japanese participated in the experiment. Al of them were students at Hiroshima University, Japan. They were paid $5.00 or its equivalent for their participation in the experiment, which lasted about 30 minutes. Materials and design Twenty sets of four conditions each were used in the experiment, in a 2 X 2 factorial design, which manipulated the case-marking of the wh-phrase (Dative vs. Nominative), and the type of adverb (Mosi vs. Degre Adverb). The 20 sets of items were distributed among four lists in a Latin Square design. Each participant saw exactly one of the lists intermixed with 72 filer items in a random order. The filer items were matched with the target items in overal length and complexity. A sample set of experimental items is shown in (20). (20) a. Dative Mosi Condition Dono-joosi-ni 1 sono kachoo-wa 2 mosi 3 which-boss-dat 1 that senior-director-top 2 mosi 3 mijukuna 4 buka-ga 5 torihikisaki-ni 6 uninitiated 4 subordinate-nom 5 customer-dat 6 songai-o 7 ataeta-ra 8 sugusama 9 jishoku-o 10 casualty-acc 7 give-cond 8 immediately 9 resignation-acc 10 negaidemasu-ka? 1 339 ask-Q 1 ?Which boss wil the senior director ask for the resignation imediately if his somewhat inexperienced employee causes casualties to the customer?? b. Nominative Mosi Condition Dono-joosi-ga 1 sono kachoo-ni 2 mosi 3 which-boss-nom 1 that senior-director-dat 2 mosi 3 mijukuna 4 buka-ga 5 torihikisaki-ni 6 uninitiated 4 subordinate-nom 5 customer-dat 6 songai-o 7 ataeta-ra 8 sugusama 9 kubi-o 10 casualty-acc 7 give-cond 8 immediately 9 resignation-acc 10 iiwatasimasu-ka? 1 tell-Q 1 ?Which boss wil fire the senior director imediately if his somewhat inexperienced employee causes trouble to the customer?? c. Dative Adverb Condition Dono-joosi-ni 1 sono kachoo-wa 2 shooshoo 3 which-boss-dat 1 that-senior-director-top 2 little 3 mijukuna 4 buka-ga 5 torihikisaki-ni 6 uninitiated 4 subordinate-nom 5 customer-dat 6 songai-o 7 ataeta-ra 8 sugusama 9 jishoku-o 10 casualty-acc 7 give-cond 8 immediately 9 resignation-acc 10 negaidemasu-ka? 1 ask-Q 1 340 ?Which boss wil the senior director ask for the resignation imediately if his somewhat inexperienced subordinate causes troubles to the customer?? d. Nominative Adverb Condition Dono-joosi-ga 1 sono kachoo-ni 2 shooshoo 3 which-boss-nom 1 that-senior-director-dat 2 little 3 mijukuna 4 buka-ga 5 torihikisaki-ni 6 uninitiated 4 subordinate-nom 5 customer-dat 6 songai-o 7 ataeta-ra 8 sugusama 9 kubi-o 10 casualty-acc 7 give-cond 8 immediately 9 resignation-acc 10 iiwatasimasu-ka? 1 tell-Q 1 ?Which boss wil fire the senior director imediately if his somewhat inexperienced employee causes troubles to the customer?? In al conditions the matrix clause material at the beginning of the sentence were folowed by either mosi or a degre adverb. The embedded subject in al conditions was preceded by an adjective. The embedded subjects were al nominative-marked so that they could provide a strong cue for the onset of an embedded clause. In order to ensure that any reading time efects asociated with the embedded verb would not be confounded with efects at the main verb, an adverbial phrase and an acusative-marked NP were inserted betwen the embedded and matrix verbs. The embedded verb in al conditions bore a conditional morpheme. These adjustments were adopted from experiment 4. 341 The only diferences among conditions involved the Case-marking of the wh- phrase and the type of adverb at the beginning of the embedded clause. In the Dative conditions, the topic marked matrix subject NPs were imediately followed by the embedded clause material, either mosi or a degre adverb. On the other hand, in the Nominative conditions, the matrix subject was located at the beginning of the sentence, and the dative-marked NPs were imediately followed by the embedded clause material, again either mosi or a degre adverb. Procedure The self-paced reading procedure and the comprehension task were identical in format to that used in Experiment 1, 2 and 4. Data analysis Analyses were conducted on comprehension task response acuracy and reading times. Al data from participants whose comprehension task acuracy was below 80% in total were discarded. 8 participants failed to met this criterion (11.8%). Trials read slower than 3000ms were discarded, corresponding to les than 2.3% of trials. The means and analyses presented below are based on the remaining trials. 342 Results Comprehension acuracy and reading times at each region were entered into a repeated-measures ANOVA, with Case-marking (Dative, Nominative) and adverb (mosi, degre adverb) as within-subjects factors. Comprehension task accuracy Among the 50 participants included in the analysis, the average comprehension acuracy was 85.5%. Mean acuracy scores did not difer significantly across the four conditions (al Fs < 1). Self-paced reading The reading time analysis yielded the following results. Reading times for the Mosi conditions are shown in Fig. 11 and those for Degre Adverb conditions (Adverb conditions) in Fig. 12. Fig. 11. mosi NP-dat NP-ac 343 Fig. 12. At al regions prior to the seventh region (the acusative NP region), there were no significant diferences betwen reading times in the Dative and Nominative conditions (al Fs < 1). Comparisons betwen the Mosi and Adverb conditions did not yield any significant diference either (al Fs < 1). Although there was no significant efect at the critical region (region 6) (al Fs < 1), at the following regions, region 7 and region 8, we observed a slowdown, reflected in longer reading times for the embedded dative NP only in Adverb conditions. At the acusative-marked NP (Region 7), there was a significant main efect of Case-marking in the participants analysis, and items analysis showed marginaly significant efect (F1 (1, 59) =4.17, P < 0.05; F2; (1, 19) = 4.06, P = 0.06). The Adv NP-dat NP-ac 344 interaction of Case-marking type and Adverb type was not significant (al Fs < 1). Pairwise comparisons revealed that within the Adverb conditions the acusative- marked NPs were read significantly more slowly in the Dative condition than in the Nominative condition. The efect was significant in participant analysis, and in the item analysis there was a marginaly significant efect (F1(1, 59)=6.44, P< 0.01; F2(1, 19)=3.49, P = 0.06). On the other hand, the same comparison for the Mosi conditions showed no corresponding slowdown (al Fs < 1). Because there are no diference among conditions other than the Case- marking of the wh-phrases and the type of adverbs, it is not likely that the slowdown that we observed reflects factors other than our own manipulations. Thus, we interpret this slowdown as a spil-over efect of a Filed Gap Efect created by the dative NP in region 6, the critical region. The embedded verb position (the region 8) also yielded a significant main efect, but only in the participants analysis (F1 (1, 59) = 4.35, P < 0.05; F2 (1, 19)= 2.45, P = 0.15). The interaction of Case-marking type and Adverb type was not significant (al Fs < 1). Pairwise comparisons showed that within the Adverb condition the embedded verbs were read slower in the Dative condition than in the Nominative condition. The efect was significant in the participants analysis, and in the items analysis, the efect was marginaly significant (F1 (1, 59)=5.0, P < 0.05; F2 (1, 19)=2.33, P =0.07). Again the same logic applies to this region. There is no reason other than the manipulation we made in the target items that may cause the slowdown. One might be concerned with the fact that the region contains the embedded verb. Remembering, however, that al the conditions have conditional 345 verbs at the embedded verb position, it is unlikely that the procesing of the conditional verb was responsible for any of the diferences in reading times. Thus, again we interpret this efect as a spil-over efect that caried over from region 6, the critical region. Given that it is likely that the reading time slowdown at these two regions reflects a Filed Gap Efect, we may se a clearer efect if we can combine these two regions. The combining of region 7 and the region 8 showed a significant main efect of the Case-marking type (F1 (1, 59) = 10.2, P < 0.01; F2 (1, 19)=5.13, P < 0.05). Pairwise comparisons revealed that within the Adverb conditions the combined region was read more slowly in the Dative Condition than in the Nominative Condition (F1(1, 59)=10.05, p< 0.01; F2(1, 19)=4.17, p < 0.05). Because al the items in these regions were the same across-the-board, it was not problematic to combine these two regions. Therefore we take this efect as indicating that a Filed Gap Efect took place at the critical region (region 6). Note that the combination of the same regions in the comparison in the Mosi conditions did not yield a significant diference in reading times (al Fs < 1). This result indicates that in the Mosi conditions, the second dative NP did not create a surprise efect. Other subsequent regions did not yield any significant efects (al Fs < 1). Discussion There were two findings in this experiment. First, the degre adverb asociated with the embedded subject basicaly did not afect the longer dependency formation bias. Even if the embedded subject NP is modified by an adjective with a 346 degre adverb, the fronted dative wh-NP is preferentialy asociated with the embedded clause. Thus, we basicaly replicated Aoshima an colleagues results. Although we obtained slightly weaker result than Aoshima et al.?s original study, our study supports the view that when a construction contains a fronted dative wh-phrase, it is preferentialy asociated with the most deeply embedded clause. Second, the presence of Mosi blocks the LDB. If Mosi is present at the beginning of the embedded clause, the fronted phrase is not preferentialy interpreted in the embedded clause (the conditional clause). This result suggests the folowing two points. First, the results support the hypothesis that the longer dependency is formed solely because of the parser?s desire to establish a Wh-Q asociation. In other words, the parser does not create a gap in the embedded clause if it cannot support the Q-particle on the embedded verb. Second, the parser appears to forese that the embedded verb of the conditional clause cannot host the Q-particle by means of the conditional adverb Mosi. In other words, upon encountering Mosi the parser predicts the structure of the upcoming conditional clause, and as a result it can recognize that the embedded verb bearing the conditional form cannot host the Q-particle. If this acount is correct, it implies that the parser is equipped with a powerful predictive mechanism that alows it to pre-proces the structural skeleton of an upcoming conditional clause because it predicts the morphology of the embedded verb which reflects a complex structural relation among V, I and C in the conditional clause structure (se the discussion in Chapter 4 for details), not just the phrase structure. 347 4.3.5. Experiment 6: Sentence Fragment Completion Task An additional offline sentence fragment completion task was conducted with thre purposes. The first purpose was to examine whether the presence of Mosi regulates the bias for the generation of a conditional clause structure even if the sentence initial wh-phrases are provided in the fragments. The second purpose was to provide an additional test of Japanese speakers? preference not to produce a Q-particle in the embedded clauses if Mosi is provided, even in the cases where a wh-phrase is provided in the fragments. The third and last purpose of this study is to provide another asesment for testing whether fronted wh-phrases are preferentialy asociated with the embedded clauses if Mosi is provided in the fragments. By testing these thre points, we can provide further support for the findings in the previous experiment. Participants 52 native speakers of Japanese participated in the experiment. Al of them were students at Hiroshima University. They were paid $5.00 or its equivalent for their participation in the experiment, which lasted 30 minutes. Materials and design The experimental materials consisted of 12 sets of 4 conditions each. The Four conditions independently manipulated the type of Case-marking for wh-phrases 348 (Dative vs. Nominative) and the type of Adverb (Mosi vs. Degre adverb). A sample set of experimental conditions for this study is summarized in (21) (21) a. Dative/Mosi Dono-buka-ni sono buchoo-wa mosi which-subordinate-dat that department-chief-top mosi hijoosiki-na sinnyuu-shain-ga ? thoughtless new-recruit-nom b. Dative/Adverb Dono-buka-ni sono buchoo-wa which-subordinate-dat that department-chief-top shooshoo hijoosiki-na sinnyuu-shain-ga ? little thoughtless new-recruit-nom c. Nominative/Mosi Dono-buka-ga sono buchoo-ni mosi which-subordinate-nom that department-chief-top mosi hijoosiki-na sinnyuu-shain-ga ? thoughtless new-recruit-nom d. Nominative/Adv Dono-buka-ga sono buchoo-ni which-subordinate-nom that department-chief-top shooshoo hijoosiki-na sinnyuu-shain-ga ? little thoughtless new-recruit-nom 349 Al four conditions contained a wh-phrase at the beginning of each fragment. The Dative condition began with a dative-marked wh-phrase, and the Nominative condition with a nominative-marked wh-phrase. In the Dative conditions, the dative-marked wh-phrase was followed by a topic-marked NP. The topic-marked NP was followed by either Mosi or by a degre adverb that modified the adjective that imediately followed it. On the other hand, in the Nominative conditions the nominative-marked wh-phrase was followed by a dative-marked NP, and the remaining material was the same as in the Dative conditions. In al four conditions, each fragment ended with a nominative-marked NP. Participants were asked to complete the remaining part of each fragment. The phrases used in the fragments were taken from the target items used in Experiment 5. The 12 sets of items were distributed among four lists in a Latin Square design. Each participant saw exactly one of the lists intermixed with thirty-six unrelated items in a random order. The filer items were designed in such a way that we could prevent speakers from adopting a strategy of using the same structure in al completions. Results This study yielded a total of 476 codable sentence fragment completions. The completions were clasified acording to the number of conditional clauses used to complete the sentence fragment, the position of the question particles that indicated the scope of the wh-phrase, and the argument structure of the embedded clause verb. Responses were clasified as conditional clauses if a conditional-marked verb was 350 provided. Responses were clasified as embedded questions if exactly one question marker on the embedded verb was provided. A Q-particle could also be placed on the main verb or on both the main and embedded verbs. Responses were also coded based on the argument structure of the verbs provided in the completions, in order to provide an additional criterion for the interpretation of the dative-marked wh-phrase. If the wh-dative was interpreted in the embedded clause, for example, there should be a verb that can take the dative NP as its argument. In the analysis of embedded clause type, 61.3% of fragments (292 trials) were completed as embedded conditional clauses with a conditional morpheme on an embedded clause. There were large diferences across conditions in the number of conditional clause responses. The proportion of trials in which a conditional verb was provided in the embedded clause for the Dative/Mosi condition was 86.1% (124/144 trials), for the Dative/Adv condition was 16.6% (24/144), for the Nominative/Mosi condition was 87.5% (126/144), and for the Nominative/Adverb condition was 12.5% (18/144). Other completions in the Mosi conditions contained either blank or ungramatical completions in which no conditional marker was provided. A ? 2 test showed that the proportion of embedded conditional completions was significantly diferent betwen the Mosi conditions and Adverb conditions (? 2 (1)= 300.502, p<0.01). ? 2 tests showed that there were significant diference betwen the Dative/Mosi condition and the Dative/Adverb condition (? 2 (1)=138.996, P<0.01), betwen the Dative/Mosi condition and the Nominative/Adverb condition (? 2 (1)=156.086, P<0.01 ), betwen the Nominative/Mosi condition and the Dative/Adverb condition (? 2 (1)= 144.751, 351 P<0.01), and betwen the Nominative/Mosi condition and the Nominative/Adverb condition (? 2 (1)= 162.0 P<0.01). There were no other reliable diferences in the other comparisons. The proportion of conditional clauses provided in the completions in each condition is summarized in Table 13. Table 13 In the analysis of question type, 9.7% of fragments (46/476 trials) were completed as indirect questions, with a question marker on an embedded verb only. There were diferences across conditions in the number of indirect question completions. The proportion of trials in which a question particle was provided on the embedded verb was 1.56 % for the Dative/Mosi condition (2/128 trials), 27.0% for the Dative/Adverb condition (26/96 trials), 2.27% for the Nominative/Mosi condition (3/132 trials), and 13.3% for the Nominative/Adverb condition (15/112 trials). ? 2 tests Conditions Embedded Clause-Type Conditional Others N % N % Dative/Mosi 124 86.1 20 13.8 Dative/Adverb 24 16.6 120 83.3 Nominative/Mosi 126 87.5 18 12.5 Nominative/Adverb 18 12.5 126 87.5 Total 292 50.6 284 49.3 352 showed that the proportion of embedded question particle completions was significantly diferent betwen the Dative/Mosi condition and Dative/Adverb condition (? 2 (1)=39.304, P<0.01), and betwen the Dative/Adverb condition and the Nominative/Mosi condition (? 2 (1)=37.407, p<.01). The other comparisons did not show significant diferences. Results of the question-type analysis are summarized in Table 14. Table 14 In order to check where the Dative Wh-phrase is interpreted, especialy in the Mosi conditions where Q-particles are not used in the embedded clause, we analyzed verb argument structure. In this analysis, we included those trials in which a Q- particle was provided on the matrix verb only. This reflects the fact that in the Mosi conditions if the Q-particle was not provided in the matrix clause, such trials were Conditions Question Type Embedded Main Both N % N % N % Dative/Mosi 2 1.56 112 87.5 0 0 Dative/Adverb 26 27.0 62 64.5 19 13.0 Nominative/Mosi 3 2.27 120 90.9 0 0 Nominative/Adverb 15 13.3 102 91.0 12 8.33 Total 46 9.82 396 78.8 31 5.38 353 likely to be ungramatical because the Wh-phrase must be licensed by the Q-particle, and the possibility that a Q-particle was provided in the embedded clause is quite low. In the analysis of verb argument structure, 39.7% of completions with conditional morphemes contained a dative argument taking embedded verb (116/292). This proportion was higher in the Dative conditions (96/292) than in the Nominative conditions (20/292), a diference that was significant (? 2 (1)=79.208, P<0.01). ? 2 tests showed that there were significant diferences in the proportion of the dative-taking conditional verbs betwen the Dative/Mosi condition and the Nominative/Mosi condition (? 2 (1)=40.648, P<0.01) and betwen the Dative/Mosi condition and the Nominative/Adverb condition (? 2 (1)=13.820, p<.01). The other comparisons did not show significant diferences. The significant diference betwen the Dative/Mosi condition and the Nominative/Mosi condition is informative for us. This strongly indicates that the wh-dative NP in the Mosi conditions is preferentialy asociated with an embedded clause. Totals for the analyses of verb argument structure are shown in Table 15. Table 15 Counts and percentages of verbs that select a dative-marked argument in an embedded conditional Condition Embedded Only Main Only N % N % Dative/Mosi 84 67.7 24 22% Dative/Adverb 12 50.0 4 30% Nominative/Mosi 16 12.6 44 89% Nominative/Adverb 4 22.2 14 53% Total 116 86 354 Discussion There were thre main findings in this study. First, the efect of mosi in biasing the sentence generation toward conditional clauses was again observed. This is a replication of experiment 3. Thus, we can confidently conclude that mosi creates a strong bias toward upcoming conditional clauses. Second, it is not so a strong measure, but the Question type analysis suggests that Wh-datives tended to be asociated with the embedded clause. Specificaly the comparison betwen the Dative/Adverb condition and the Nominative/Adverb condition is the most informative in this respect. The analysis showed that more Q- particles were produced in the completions in the Dative/Adverb condition than in the Nominative/Adverb condition. This suggests that the dative-marked wh-phrase was more likely to be interpreted in the embedded clause compared to a condition that does not have such possibilities. The analysis of Q-particle also supports a relation betwen the conditional morphology and the Q-particle. The fact that the Mosi conditions contained very few embedded Q-particles suggests that Mosi worked as a reliable cue for the upcoming conditional clause, and that Japanese speakers do not try to embed an additional clause inside conditional clause in order to host a Q-particle asociated with the fronted wh-phrase either, supporting our earlier speculation. These two findings help us to fulfil two of the purposes of this study. First, we could confirm that Mosi indeed creates a strong bias toward conditional clauses. Second, Fronted wh-phrases sem to be preferentialy asociated with an embedded clause. Note, however, that the conclusion that we can draw from the analysis of Q- 355 particle positions was not strong because we could not find a significant diference betwen the Dative/Adverb condition and the Nominative/Adverb condition, the comparison we used in the online experiments. Let us think about the third purpose of the study. Based on the results of our Filed Gap study, we expected that the offline experiment would also show the same tendency, i.e., the fronted dative-marked Wh-phrase should not be preferentialy asociated with the embedded clause if Mosi provides an early warning for an upcoming conditional clause structure because it is not compatible with the motivation for the LDB. However, the analysis of argument structure revealed that this was not the case in the offline sentence completion task. In the Dative/Mosi condition there were significantly more dative taking verbs in the completed conditional structures. This efect cannot be related to the presence of Mosi alone, because it exists in the Nominative/Mosi conditions also but they are not completed with embedded verbs that a take dative argument. Rather, this efect should be interpreted as being created by the fronted wh-dative phrases. This result directly contradicts the finding in the Filed-Gap experiment. Remember, however, that we have emphasized that Japanese conditional clauses are not islands for long-distance scrambling based on native speakers? aceptability judgments. As long as native speakers? intuitions are reliable, we have a way to interpret a semingly problematic result in our ofline experiment. First of al, if conditional clauses are not islands as the aceptability judgments suggest, it is not surprising if even a long-distance scrambled phrase is interpreted in the conditional 356 clauses. Rather, the situation in online experiment is surprising from this point of view. Now what acount can we give about this online offline contrast? Although informal, we can give the following reasoning for why a longer dependency might not be blocked in the offline seting. There sems to be a crucial diference betwen online sentence procesing and offline judgments or sentence completions. That is, the diference in resources that readers can aces. Typicaly, in online sentence procesing, the parser can aces only limited resources. Because in an online self-paced reading context readers have to read sentence from left-to-right, they cannot re-aces the left-context once each phrase is read. Thus, if a sentence contains dependent elements such as displaced elements or elements that must be asociated with another elements, readers have to store them in memory during the procesing of the remaining part of the sentence, until they are licensed (Gibson 2000; Lewis 2001; among many others). In this online situation, it is plausible if the parser tries to establish a dependency involving these elements as soon as possible, so as to release these elements from working memory. If the licensing relation is easy for the parser to infer, due to, for example, an overt morphological marking, the parser can establish a licensing structure and release the element from working memory. For example, we have sen that the thematic requirement on the fronted dative phrase sems to be established easily by the parser. What we have discussed was that the thematic role of the dative phrase can be read off from the overt Case-marking and basic word order in Japanese sentences. Thus, the parser does not need to asociate the fronted NP with the lexical verb in order to confirm the actual thematic role that it receives, as long as 357 the thematic requirement is concerned. The findings from our first two experiments also suggest the same point. On the other hand, if the licensing structure is not obvious from the curent input items alone, the situation becomes diferent. Wh-Q-particle asociation is just such a case. As we have sen, the scope of a wh-phrase cannot be determined based on the position of the wh-phrase. Rather it is fixed by the Q-particle that is afixed to the verbs. Because of this, the word order, or the Case-marker does not help the parser to establish the licensing structure for a wh-phrase. The presence of the lexical verb is crucialy necesary, and the embedded verb is always the first verb to be encountered. Thus, the parser that has a desire to satisfy gramatical features of each lexical item tries to asociate the wh-phrase with the embedded verb in order to confirm its scope at the earliest point. By doing so, the parser can release the wh- phrase from working memory as early as possible. The intuition behind our conclusion in the previous subsection was if Mosi is provided in a sentence, then a wh-Q-particle dependency should not be established in the embedded clause because of the conditional morphology. In this case the most deeply embedded verb is not the first position for Q-particle, tather the verb that imediately follows the embedded verb is the first possible host for Q-particle. Going back to the ofline situation, during the ofline judgment or sentence completion task, native speakers can aces much richer resources than in the online situation. They have more time for reading each sentence or they can even repeat the sentence and examine the structure in detail. In this situation, native speakers do not need to establish gramatical relations in such a hurry. There is no strong driving 358 force urging them to establish sentence structure. If this is the case, we lose the reason to block longer dependency formation even when Mosi is provided. If such requirement for earlier dependency formation is motivated by resource limitations, then in the ofline context, the dependency blocking efect is also not motivated. Thus, it is not surprising that the longer dependency is not to be blocked. However, this reasoning does not provide an indication of why longer dependency sems in fact to be prefered in the Mosi condition. This is clearly a puzzling situation, and we do not have any clear answer at this point. 359 5. General Discusion: The Silhouete of the Predictive Mechanism 5.1. The Nature of the Predictive Mechanism (or the Things that the Parser should be Able to do) At the outset of this chapter, we laid oud the aims of this study. They were: [i] to investigate the real motivation behind the LDB; and [i] to investigate the nature of the predictive mechanism underlying the human sentence procesor. The second aim can be rephrased as the following two questions. (i) What counts as a potential trigger for the predictive mechanism (the trigger problem)? (i) What is the representation that the predictive mechanism generates (the representation problem). There is another important problem: What is the algorithm employed by the predictive mechanism (the algorithm problem)? We wil discuss the algorithm problem in Chapter 5 in detail. So this problem wil be put aside temporarily. In this section, we wil pay atention to the first thre problems: the motivation for the LDB; the possible triggers for the predictive mechanism; and the representation that the predictive mechanism aceses. First let us talk about the motivation for the LDB. The results from thre online experiments suggest that the LDB is solely motivated by the wh-feature on a fronted dative-marked wh-phrase. Experiments 1 and 2 revealed that if the wh-feature is stripped from the fronted dative NP, a longer dependency was not formed. Experiment 5 further strengthened this view. In experiment 5, we created a situation where the wh-feature on the fronted materials cannot be satisfied by the most deeply embedded verb bearing the 360 conditional morpheme. We manipulated the parser?s prediction for upcoming structures by means of the conditional adverb Mosi. If Mosi is present, it provides an early warning for an upcoming conditional clause, and as a consequence of this early warning, the LDB is blocked. Under the asumption that the parser avoids extra embedding whenever possible, and if the most deeply embedded verb cannot bear the Q-particle, the earliest posible position where the verb bearing the Q-particle can appear is the position imediately following the most deeply embedded verb. The schematic structure ilustrated in (22) clearly shows this point. (22) In a structure like (22), upon encountering mosi the parser notices that the embedded verb cannot bear a Q-particle, and that the earliest position where the Q-particle can appear is the matrix verb position. Additionaly, the gramar alows wh-Q-particle asociation at the first gap position, thus there is no reason that the parser should 361 further stretch the dependency to go into the embedded clause. If we do not take this acount, it is not clear why the fronted wh-phrase should not be asociated with the embedded clause even though conditional clauses are not islands. One might argue that conditional clauses show a mild island efect, and that this mild island efect is sufficiently strong to block the LDB. However, this acount has a shortcoming. In the offline experiment, Experiment 6, we have found some evidence that the fronted wh-phrase is preferentialy asociated with the embedded conditional clause. The mild island acount thus cannot explain this ofline result as wel as native speakers? judgments. Taken together, this study shows that the LDB is solely motivated by the wh- feature conveyed by the fronted dative Wh-phrase. In the previous discussion, we informaly proposed that mosi creates a prediction for an upcoming conditional clause structure. However, we left two important questions unresolved: how the predictive mechanism is triggered and what representation is built by the predictive mechanism? These are the questions we wil try to answer here. Let us discuss ?the trigger problem? first. This is the question of what kind of information is used to create a prediction for an upcoming structure? In the previous chapter, we discussed that the parser can use a relatively indirect cue to anticipate the upcoming structure, such as the Clasifier Mismatch. In this study on conditional clauses, we have found that a dependent element such as a conditional adverb like mosi can be used to forese an upcoming conditional clause. This prediction proces is quite straightforward compared to the case of Clasifier Mismatch. 362 As is extensively discussed in Chapter 4, syntacticaly Mosi has a strong relation with the conditional clause. As we have discussed in various places, Mosi can only be licensed in a conditional clause. Therefore, if Mosi is present, the conditional clause must be present too. Thus mosi can signal an upcoming conditional structure. We wil se more formal and detailed discusion on this point in the later chapter, Chapter 5. We have understood that mosi can create a prediction of some sort for an upcoming conditional structure. Now the question is what is actualy predicted, i.e., the representation problem. A series of experiments that we have described have made it obvious that the predicted structure must be sufficiently informative that it can block LDB. However, it is not clear what kind of structural representation counts as a sufficiently informative representation. The key to revealing this point lies in the motivation for the LDB. As we have sen, to block the LDB, it must be clear that the verb cannot host a Q-particle. This, in turn, means that when the LDB is blocked, the parser calculates that a Q-particle cannot be hosted by the embedded clause. This information is morphological information. Asuming that Mosi is structuraly licensed by the conditional morpheme (se Chapter 4 for a detailed discussion), the structure built by the predictive mechanism should represent the phrase structure of the conditional clause, and the morphology of the conditional verb that licenses Mosi should not host the Q-particle. If this is true, the predictive mechanism should be able to build the full phrase structure of a conditional clause, going al the way down to the head of VP where the conditional morpheme suffixes and al the way up to CP where the conditional morpheme -ra, a complementizer, ultimately sits. Both the VP 363 structure and the CP structure sem to be necesary. Because of its afixal nature, the conditional morpheme requires a verb as a host. A CP is also necesary because the conditional sufix -ra or its variants is understood as a head of CP. Thus if -ra is predicted, the CP projection should also be predicted. (23) Finaly a question arises, namely what kind of sentence procesing algorithm alows such prediction of the upcoming structure? We wil explore this question in Chapter 5 5.2. Reconsidering the Procesing of Relative Clause Islands (or the Things that the Parser should be Able to Forese) So far, we have sen what the parser should be able to do when an adverb like mosi is provided. In the discussion above, we have sen that the prediction of verbal morphology is a necesary component of the blocking of the LDB. Asuming that this morphological prediction is true, it raises an interesting question about the apparent online island efect created by RCs. In the previous chapter, we concluded that LDB 364 was blocked in the environment of predicted RCs because RCs are islands in Japanese. However, the above discussion on conditional clauses casts doubt on this view of the online procesing of RC islands. Looking closely at the morphological property of RC verbs, it becomes clear that RC verbs also cannot host a Q-particle. As discussed in Chapter 4, Japanese RC verbs must bear the so-caled Predicate Adnominal form (Hiraiwa, 2000; Hiraiwa, 2001), and the Predicate Adnominal form is not compatible with the Q-particle. The incompatibility of the Adnominal form and the Q-particle can be understood in a way that the empty C and Q-particle are competing the same slot, exactly in the same way as the conditional morpheme and Q-particle do. In Japanese, if a CP is an imediate constituent of an NP, the embedded verb inside the CP must bear the Adnominal form. Based on this observation, we can interpret the online island efect of RCs from a diferent perspective, i.e., because the Adnominal form is predicted, the LDB is blocked. We have sen that a Genitive Numeral Clasifier (GNC) requires a host NP. Therefore, the GNC can potentialy create a prediction for an upcoming NP. On the other hand, a Nominative NP can create a prediction for a clausal projection (Miyamoto 2002). Presumably S/IP and CP nodes are predicted by means of a nominative NP due to the fact that a nominative NP is typicaly licensed by a finite Infl (Chomsky 1981, 1986a, 1986b; Takezawa 1987), and the fact that the finite IP is dominated by a CP (Chomsky 1986a, 1986b). If a Clasifier Mismatch takes place, the two predicted nodes, the NP and the CP are to be connected to each other in some way, under models of sentence procesing incorporating a fully incremental structure building mechanism (Aoshima et al. 2004; 365 Bader and Laser 1994; Crocker 1996; Schneider 1999; Stabler 1994; Sturt and Lombardo 2005 among others). If so, the predicted CP must be incorporated into the structure as a complement or an adjunct (an imediate constituent) of the predicted NP. In either way, if it becomes clear that the CP must be atached to the NP, the gramar of Japanese tels the parser that the embedded verb inside the CP must bear the Adnominal form, and thus the CP must be headed by an empty C. Asuming that the parser can calculate the consequences of the prediction, i.e., that the Q-particle cannot appear in the predicted CP because it has to host an empty C, the parser has sufficient information to stop creating the longer dependency. This is so because the wh-feature cannot be satisfied at the embedded verb position because the Q-particle cannot appear there. The reading time diference at the embedded verb position that we have observed in the second experiment in Chapter 2, indeed supports this view. Because the verbal morphology is predicted, a reading time diference appeared at the embedded verb position. Under this acount, again the parser has to predict al the way down to the VP projection and al the way up to the CP projection upon encountering the clasifier mismatch. However, if the parser incorporates a predictive mechanism that can predict the verbal morphology in an RC, we can derive the online RC island efects. Under this acount, the blocking of the LDB in RCs and conditional clauses can be understood in the same way. Note, however, that the same question arises as we asked in the previous section, namely whether there is an algorithm that alow the prediction that we have ilustrated so far or not. In the next chapter, Chapter 5, an algorithm wil be proposed that can appropriately capture the blocking of the LDB. 366 CHAPTER 5. ON THE NATURE OF PREDICTIVE MECHANISMS 1. Introduction In this chapter, we are concerned with the remaining questions involving the nature of the predictive mechanism, namely the Algorithm Problem. So far we have sen some evidence that the parser is equipped with a powerful predictive mechanism that can predict detailed upcoming gramatical structures. Our intuition is that the parser projects a detailed syntactic structural skeleton based on bottom-up information. What we have sen was that the parser is able to use indirect or direct cues to build the upcoming structure. Furthermore, by looking through the cases where the Longer Dependency Bias (LDB) is blocked, we have investigated the representation that the predictive mechanism generates. The remaining problem for us is what is the actual mechanism of sentence procesing that can derive these predictions. In this chapter, a parsing algorithm wil be proposed that can explain the predictive mechanism of sentence procesing as wel as various wel-known properties of the human sentence procesor. 2. The Requirements for the Predictive Mechanism In Chapters 2 through 4 we have sen aspects of the predictive mechanism underlying the sentence procesor through detailed investigations on Japanese Relative Clauses (RCs) and Conditional Clauses. In the course of the discussion we have paid special atention to two aspects of the predictive mechanism. One is the trigger of the predictive mechanism. What kind of bottom-up information is used to 367 initiate a prediction for upcoming structures. We have sen two types of triggers: indirect cues and direct cues. When the parser predicts upcoming head-final RCs in Japanese a relatively indirect cue is used. We have confirmed in the earlier chapter that a Clasifier Mismatch that does not have any direct gramatical relation to a specific construction like RCs can indeed alow the parser to predict an upcoming RC structure. Let us first clarify the specifics of the clasifier mismatch cue. As we have sen, a Genitive Numeral Clasifier (GNC) must be asociated with its host NP. We asumed the structural relation betwen the GNC and its host NP shown in (24). Because of the semantic requirement of the clasifier, the head of the host NP must be a noun that is semanticaly compatible with the clasifier. Furthermore, we asumed that GNC and the host NP must be structuraly, not linearly, adjacent to each other. Betwen the GNC and NP, an RC can intervene as in (25), but it does not afect these requirements of the GNC, and also the presence of the GNC does not afect the relation betwen the RC and the NP that it modifies. The semantic compatibility betwen the clasifier and the head of the host NP and the structural requirement that is ilustrated in (24) are two requirements that the GNC has to fulfil. As should be clear, there is no specific relation betwen the GNC and the RC. Both are optional modifiers for the NP, and neither requires the presence of the other. 368 (24) (25) Only in the type of configurations in (25), however, the clasifier mismatch can come about. In this configuration, the clasifier and its linearly adjacent NP, for example the subject of the RC, can semanticaly mismatch, and the semantic mismatch is only possible in this type of complex NP configuration. And when the parser recognizes this mismatch, the parser notices that the mismatch can only be legitimate in the complex NP configuration such as (25), and this information leads to the prediction of the RC structure. NEither the clasifier nor the NP that is acidentaly adjacent to the clasifier bears a direct signal for the RC itself. Rather, the semantic mismatch is the crucial signal that alows the parser to forese the upcoming RC structure. For this reason, the cue from the clasifier mismatch is indirect, i.e., it has no direct relation to the RC structure. In the study of conditional clauses, we have discovered another type of trigger, i.e., the conditional adverb Mosi. This is a more direct cue for the predictive mechanism. We have sen that gramaticaly Mosi stands in a tight one-to-one 369 relation with the conditional clause. If Mosi is there, a conditional clause must be present because Mosi can only be licensed in the conditional clause. By means of this tight gramatical relation, Mosi alows the parser to predict an upcoming conditional clause structure. Now let us turn to the representation built by the predictive mechanism. We have discussed that both in the case of RCs and conditional clauses, the representation created by the predictive mechanism must be something that can block longer dependency bias. In the case of RC prediction, this semed to be straightforward at first glance. Because RCs are islands in Japanese a longer dependency that enters an island domain is not legitimate. Thus, a possible approach was that a representation that can encode the islandhood of RCs is built by the parser, e.g., the representation of subjacency. However, the studies on conditional clauses provided us with another way that can lead us to the online island efect. The studies on conditional clauses revealed that the LDB is blocked during online sentence procesing because Mosi sets a prediction of the upcoming conditional morphology, which is not compatible with the Q-particle. A series of experiments revealed that the real motivation for the LDB is wh-Q-particle asociation. Thus, if the conditional verb cannot host the Q-particle, there is no reason to pursue the longer dependency. Thus, prediction of the embedded verbal morphology is crucial for blocking the LDB. Within the approach where we put an import on the morphological aspect of the predicted structure, there is a possibility that the online island efect of the RCs can also be treated in the same way. A detailed syntax of RCs has revealed that the 370 embedded verb in RCs also bears special verbal morphology, the so-caled Predicate Adnominal form (Hiraiwa 2000, 2001 among others). What we have confirmed was that the Adnominal form is also not compatible with the Q-particle. Thus, if we think that the Clasifier Mismatch leads to the prediction of the Adnominal morphology, the fact that the LDB is not sen in the predicted RC structure may be sufficiently explained. In the following discussion, I wil summarize both subjacency approaches and morphology based approaches. 371 2.1. Subjacency Aproaches In this subsection, let us summarize advantages and disadvantages of subjacency-based approaches. First we have to note that it does not sem to be the case that al the complex NPs are islands in Japanese. For example the following contrast suggests that a complement clause of a noun is not an island or at least not as strong an island as RCs. (26) a. Noun Complement Clauses Dono-spy-ni tantee-wa [ NP [ CP hannin-ga joohoo-o which-spy-dat detective-top suspect-nom information-acc nagasita toiu] shuchoo]-o tenkai-sita-no? leaked comp claim-acc made-Q? Lit. ?Which spy did the detective make a claim that the suspect leaked the information to?? b. Relative Clauses *Dono-spy-ni tantee-wa [ NP [ CP hannin-ga nagasita]joohoo]-o which-spy-dat detectie-top suspect-nom leaked info-acc nyuushu-sita-no? obtained-Q? Lit.?Which spy 1 did the detective obtain the information 2 that the suspect leaked t 2 to t 1 ?? 372 This contrast suggests that structure specific to RCs is crucial to the islandhood of RCs. The fact that a domain is a complex NP is not sufficient to derive its islandhood. The crucial diference betwen RCs and noun complement clauses is whether operator movement is involved or not (for the details about movement approaches to Japanese RCs, se Chapter 2). In a regular head-external RC in Japanese, empty operator movement is crucialy involved in its derivation. Thus, in head-external RCs there is a gap that corresponds to the relative head. On the other hand, noun complement clauses do not involve operator movement of the type involved in RCs. Given this diference, we can derive the islandhood of the RCs from a bounding theory incorporating the clasical subjacency condition (Chomsky 1973, 1981; Rizi 1982 among many others), in which NP and IP are the bounding categories and movement crossing two of these nodes in one step creates a subjacency violation. Specificaly, asuming that the specifier of CP in head-external RCs is occupied by an empty operator coresponding to the argument gap inside the RC, A?-movement out of an RC cannot pas through the CP-spec position and it has to move across IP and NP in one step. On the other hand, asuming that in the noun complement clauses, the CP-spec position is available for A?-movement, and thus the subjacency violation can be side-stepped. 373 (27) a. Relative Clauses b. Noun Complement Clauses With this discussion in mind, let us think about subjacency approaches to online RC island efects. In order for subjacency approaches to derive the online island efect, the presence of the empty operator in the CP-spec position is crucial. Thus, if the existence of two bounding nodes and the presence of the empty operator in the CP-spec are predicted, then the islandhood of the upcoming structure is also predicted. This is equivalent to asuming that the full structural skeleton of the RC structure, including the empty operator, is projected when the upcoming RC structure is predicted. Although incorporating such information is possible, there are some dificulties with this type of approach from the perspective of online structure building. The most crucial chalenge is the dificulty in recognizing the location and the identity of the gap during parsing In other words, until the verb?s argument structure or the relative head becomes available, sometimes the parser cannot determine whether the gap is contained in the RC or which argument slot corresponds 374 to the gap. To understand this problem we have to understand the typology of RCs and also the ways in which an RC can be procesed. There are various types of RCs in Japanese. Above al, one of the most ecentric types of RCs is the so-caled Gaples RCs and their variants. In Japanese, certain RCs do not contain an argument gap position (Hoshi 1995, 2004; Inoue 1976; Kamio 1983; Kuno 1973; Kuroda 1974, 1975/76, 1976/77, 1999; Matsuda 1993; Murasugi 1991 among many others). Let us se some examples. 31 The important property of the examples in (28a) and (29a) is that the strings constituting the embedded clauses can be stand-alone matrix clauses if they are not incorporated into the NP environments.Put diferently, no argument sems to be mising in the embedded clauses, and thus they do not contain explicit gap positions. (28) a. [ NP San-satu-no [ NP [ RC yuumee-na dezainaa-ga 3-cl-gen famous designer-nom Rosa(-no-tame)-ni cover-o dezain-sita]hon]]. R(-for)-dat cover-acc designed book ?Thre books which a famous designer designed its cover to for Rosa.? b. Yuumee-na dezainaa-ga Rosa(-no-tame)-ni famous designer-nom R(-for)-dat cover-o dezain-sita. cover-acc designed. ?A famous designer designed the cover (for something) for Rosa.? 31 In this study, we wil losely clasify gaples RCs as RCs that superficialy do not contain a gap that coresponds to the relative head. Thus, we include some RCs that are not analyzed as gaples RCs in Japanese syntax. 375 (29) a. [ NP San-nin-no [ NP [ RC Sidney-ga Rosa-ni three-cl-gen S-nom R-dat otooto-o shookai-sita] yuujin]]. younger-brother-acc introduced friend Lit. ?Thre friends whose brother Sidney introduce to Rosa.? b. Sidney-ga Rosa-ni otooto-o shookai-sita. S-nom R-dat younger-brothers-acc introduced ?Sidney introduced his younger brothers to Rosa.? The possibility of gaples RCs creates potential problem for incremental structure building under subjacency-based approaches. When these RCs are read from left-to- right, it is not clear where the gap is located or which argument could correspond to the gap in the RCs until the parser encounters the embedded verb or the head of the relative clauses because there is always a possibility that the predicted structure turns out to be a gaples RC like (28) or (29) (se Yamashita et al. 1993; Yamashita 1995 for related discussion). Even though the clasifier mismatch can indicate that some NP is located outside of the embedded clause, it does not help the parser to recognize the gap in these cases because, as we have sen, even in these cases the gaples RC can be there. Let us examine the point above more in detail under the subjacency-based acount. Under the subjacency-based approach, what the parser has to do when it encounters the clasifier mismatch is to insert an empty operator in the embedded CP- spec position. If the empty operator is inserted, the parser can project the structure that can derive the subjacency efect. Because the CP-spec position is occupied, the 376 displaced elements at the beginning of the sentence cannot be asociated with the embedded clause without violating the subjacency condition. Thus, as long as online structure building respects the mental gramar, the LDB should be blocked. (30) Wh-NP-Dat ? [ NP GNC [ NP [ CP Op [ IP Subject ? ]] NP (host) ] This approach, however, involves an important asumption. That is, the operator indicates the presence of a gap position corresponding to an argument inside the embedded clause. 32 Therefore, under this approach the parser?s initial choice upon encountering the clasifier mismatch is a gapped RC analysis rather than a gaples RC analysis. From this it follows that if the RC turns out to be a gaples RC, it is predicted that there wil be a surprise efect asociated with properties related to gaples RCs, e.g., Filed Gap Efect because al of the argument slots are filed by overt lexical elements. Although it has not been tested whether this prediction is correct or not, it does not sem that there is an independent reason that the gaples RC analysis should not be an initial choice for the parser. Specificaly given that examples like (29) are quite natural (they are read as natural as gapped RCs), and do not sem to induce any dificulty at least at the level of intuitive judgment, it is not clear why the gaples RCs should create any problems. Of course, we have to 32 Note that the semantic information from the clasifier also sugests that there is an argument gap. For example the clasifier for boks, can tel the parser that the relative head NP is headed by a noun bok. However, this information is not so decisive because the type of the RCs as in (28) exists in Japanese. 377 consider various factors such as the frequency of such gaples RCs or aceptability diferences betwen gaples RCs and gapped RCs, and careful conclusion must be drawn. Although we do not have any experimental results that can tel us whether gaples RCs are problematic or not, there is no compeling argument for the view that the parser?s initial choice should be gapped RC analysis, and we do not have any strong evidence that supports it either. If so, it is plausible to asume that the gaples RCs do not create any troubles. If it is true, we can say that those two options are equaly plausible as first choices for the parser. Under this hypothesis, what is the representation that the predictive mechanism should create? It is possible to Insert an empty operator corresponding to the argument gap upon encountering the clasifier mismatch. However, in this case the potential eror in choosing a gapped RC analysis arises. Given this the safest strategy is to insert the operator after encountering an explicit bottom-up information that can tel the parser which argument is mising in the embedded clause. Thus under this approach, it is possible that the parser initialy cannot project the representation that can satisfactorily represent the islandhood of the RC structure. The islandhood of RC becomes clear only after the verb or other input confirms which argument slot can be the gap position. Thus, this approach predicts that until the embedded verb position or the relative head position a sentence containing a clasifier mismatch could be treated as non-island. This approach, in turn, predicts further that even in the clasifier mismatch condition, the Filed Gap efect could take place, contrary to our findings. 378 To maintain a subjacency-based approach, we have to asume that an empty operator is inserted in anyways, without considering the possibility of gaples RCs. This is not totaly desirable option but not at al implausible option. 2.2. Morphology-Based Aproaches In this subsection, we wil turn to Morphology-Based Approaches and discus their advantages and disadvantages. Let us turn to the other alternative, the morphology-based approach. Under the morphology-based approach, the LDB is blocked if the predicted embedded verb is not compatible with the Q-particle, which can confirm the scope of the fronted wh- phrase. Thus, this approach does not appeal to a subjacency-based acount. In other words, even if the predicted upcoming structure does not fully represent the structure of the RC, the blocking efect of the LDB is expected. The question for this approach is what is necesary for the predictive mechanism to predict the upcoming verbal morphology. As we have sen in Chapter 2, the Predicate Adnominal form requires an empty C that is created by the morphological amalgamation of V, T and C (Hiraiwa 2000, 2001). Thus, it is sufficient for the parser to recognize that the Adnominal form requires the empty C in order to block the LDB. If the Complementizer is occupied by the empty C, it follows that the Q-particle, which is another instance of Complementizer, cannot appear in the same slot. Additionaly, they are only legitimate if the constituent containing the Adnominal-Inflected verb is an imediate constituent of an NP (when they are structuraly adjacent to an NP in the structure of adjunction or the complement of an NP). For these reasons, the information that the 379 predicted structure is incorporated into the structure of NP in some way is crucial for the Adnominal morphology to be predicted. To fulfil these requirements, what the parser has to project upon encountering the clasifier mismatch is the full-fledged phrase structure of CP headed by the empty C that is incorporated into the structure of NP either as an adjunct or as a complement. (31) The result of Experiment 2 in Chapter 3 indeed supports this claim. In the online reading experiment in Experiment 2, we obtained the result that the embedded RC verb was read more easily when a clasifier mismatch is provided compared to the control condition. If the upcoming verbal morphology was not predicted, we would not expect such a result at the embedded verb position. In other words, because the morphological form of the verb is predicted, the embedded verb was easier to read. There are several advantages of this morphology-based approach. First of al, this approach does not appeal to the presence or absence of an empty operator to explain the online islandhood of the predicted RCs. Thus, the potential problems asociated with the empty operator do not arise. Second, this approach predicts that if the morphology of the embedded verb that is not compatible with the Q-particle is 380 predicted by the parser, the longer dependency bias is blocked. This means that we can explain the blocking of the LDB in the predicted RCs and Conditionals in the same fashion. In the case of conditional clauses, the prediction of the verbal morphology was crucial too. Also, this approach predicts that the longer dependency is blocked in Noun Complement clauses too, if they are predicted in some way because they have the same Adnominal morphology on the embedded verb. Third, as long as the parser has a mechanism that can predict the upcoming verbal morphology upon encountering the clasifier mismatch, this approach does not sacrifice incrementality. This, however, does not naturaly follow fron anything. Thus we need to cal for an explicit predictive mechanism. This problem leads us to the Algorithm problem. 2.3. On the Requirements for the Predictive Mechanism: Sumary In the discussion so far, we have compared two potential approaches to the online island efect of the predicted RCs. We have sen several advantages of the morphology-based approach over the subjacency-based approach. 381 3. Perspectives on Online Sentence Procesing: A Quest for Incrementality In the previous section, we have raised the question of what type of algorithm alows the parser to predict the structures of RCs or Conditional clauses that satisfy the requirements that we have ilustrated above. To answer this question, we wil first review a general property of human sentence procesing, incrementality, and discuss parsing algorithms that have been proposed in the literature. The main purpose of this discussion is to specify what is a psychologicaly plausible design for the human sentence procesor that can acommodate the predictive mechanism that we have discused. 3.1. Incrementality So far we our discussion loosely presupposed the notion of incrementality. In this subsection, let us briefly summarize the notion of incrementality of human sentence procesing. Current evidence from psycholinguistic research suggests that human sentence procesing is largely incremental, in the sense that syntactic structures are built and interpretations become available on a word-by-word basis without delay (Marslen- Wilson 1973). So-caled garden path phenomena (Bever 1970; Pritchet 1988 among many others) have often been understood as evidence for incrementality. Garden path phenomena tel us that when the parser faces an ambiguous input, it does not delay its decision until crucial disambiguating information becomes available. Rather the parser makes a commitment to an initial analysis that may turn out to be an incorrect analysis. When the parser is garden-pathed, consciously or experimentaly detectable 382 surprise efects arise at the point where the initial misanalysis is resolved. Let us take an example from English. The string before the main verb sank in (1) can be parsed as a simple independent clause. When the main verb is encountered, a surprise efect arises, and readers experience dificulty in reading. If the parser waits until the element that can confirm the structure of the sentence, this garden path efect is not expected. The fact that readers experience the surprise efect suggests that the parser commits itself to the initial main clause analysis before the main verb is encountered, and the verb sank forced the parser to reanalyze the structure from the main clause structure to the reduced relative clause structure, resulting in the garden path efect. (32) The boat floated down the river sank. To acount for the garden path phenomena, many researchers emphasized the importance of the licensing head. For example, the theory proposed by Abney (1987) or Pritchet (1992), the so-caled head-driven parsing, explained the garden path phenomena in a way that the sentence procesing is driven to satisfy the requirements of the licensing heads, typicaly verbs. Under this approach, an example like (1) creates garden path efect because the materials before the verb sank can satisfy al the requirements that the verb floated conveys, such as thematic requirements. Head-driven strategies imply that in a head-final language, e.g., Japanese, the procesor waits until the final word of a phrase before building the phrase structure and making a commitment to an analysis. In other words, head-driven theories predict that the input materials are not integrated into a syntactic structure before the 383 licensing head becomes available in the input (Pritchet 1991b, 1992a, 1992b among others). However, there are various lines of evidence against this particular view either from head-final languages like Dutch (Frazier 1987) or Japanese (Aoshima et al. 2004; Inoue 1991; Inoue and Fodor 1995; Mazuka and Itoh 1995; Miyamoto 2002) or even from a head-initial language like English (Sturt and Lombardo 2005). Let us take a look at some examples from Japanese. Inoue (1991) cites the following example from Japanese. (33) Sidney-ga Rosa-ni ringo-o tabeta inu-o ageta. S-nom R-dat apple-acc ate dog-acc gave ?Sideny gave Rosa the dog which ate the apple.? Japanese native speakers experience some dificulty dealing with this sentence when they reach the verb tabeta ?ate?. This dificulty in reading is not expected if the parser delay the structure building until it encounters the verb. Because, when the embedded verb ?ate? becomes available, its argument structure information also becomes available. Based on the argument structure information, the parser can notice that the previously encountered thre NPs are not compatible with the verb?s argument slots, and conclude that the string should be analyzed as a relative clause rather than a simple clause. Thus, the dificulty indicates that the parser commits an analysis before it encounters the embedded verb. Based on the case-particles that each NP bears, the thre NPs are initialy postulated as coarguments of the same clause. When the 384 embedded verb is encountered, it becomes clear that the initial analysis is failed, resulting in the garden path efect. Inoue?s argument is based on native speakers? intuitive judgment, but some experimental studies support the same point too. Miyamoto (2002) found that the upcoming relative clause structure can be facilitated when a sentence contains a sequence of NPs with the same Case Markers. He examined the following pair of sentences. (34) a. Ofisu-de shokuin-ga kakaricho-o[ RC t 1 ocha-o office-at employeee-nom manager-acc tea-acc dasita] josei-ni teineini shookai-sita. served woman-dat politely introduced ?At the office, the employee politely introduced the manager to the woman who served the tea.? b. Ofisu-de shokuin-ga kakaricho-ni[ RC t 1 ocha-o office-at employeee-nom manager-dat tea-acc dasita] josei-o teineini shookai-sita. served woman-acc politely introduced ?At the office, the employee politely introduced the woman who served the tea to the manager.? Miyamoto makes two important observations. First, when there are two acusative NPs in a sentence as in (34a), the second acusative NP creates a disruption. Second, in (34a), the relative head noun was read more easily than in (34b). Acording to 385 Miyamoto these two observations are not explained by head-driven approaches. Under the head-driven approaches, the disruptive efect of the second acusative NP is not predicted because these NPs come into the input before the verbs and thus they should not be procesed. Furthermore, head-driven approaches do not predict the facilitation efect at the relative head position in(34a). In both of the conditions in (34), everything is the same except for the acusative NPs. Thus, the facilitation efect cannot be atributed to any other factors than the presence of the two acusative NPs in (34a). Japanese has the so-caled Double Acusative Constraint (Harada 1973a, 1974; Hiraiwa 2002; Kuroda 1965 among many others) that excludes a clause containing two acusative NPs. In (34a) the parser has to insert a clause boundary betwen two acusative NPs in order to analyze the string gramaticaly in satisfying the Double Acusative Constraint. Once a clause boundary is inserted, the embedded clause structure is created and it becomes clear that the subject position in the embedded clause is empty. Thus, the parser can predict the upcoming relative clause structure, resulting in the facilitation of the relative head position. Acording to this acount, the acusative NPs before the embedded verb must be procesed in advance of the embedded verb. Thus this finding is not compatible with head-driven approaches, and supports the incremental structure building. Let us se another piece of evidence for incrementality that is directly relevant to our study. As we have sen several times already, when a sentence contains a fronted dative marked wh-phrase, it is asociated with the most deeply embedded clause. To probe for this LDB, we used Aoshima et al.?s Japanese Filed Gap 386 paradigm. In Aoshima et al.?s original experiments and also in our own experiments, the Filed Gap efect was observed at the embedded dative position before the verb is encountered. This preverbal Filed Gap efect cannot be explained under the head- driven approaches. If the sentence structure is not built before the verb is encountered, we do not expect the Filed Gap efect to take place before the verb position. However, we observed a slowdown at the second occurrence dative position which is before the verb. Thus Filed Gap efect also supports the incremental model of sentence procesing, and reject the head-driven approaches. We can understand the incrementality as a general property of the human sentence procesing. Now the question is what kind of mechanism can derive the incrementality. From now our discussion turns to psychologicaly plausible sentence procesing strategies. 3.2. Mechanisms for Sentence Procesing In the following subsections, we wil review some of the previously proposed sentence procesing mechanisms informaly. Our focus here is which type of algorithm can appropriately capture the incrementality of human sentence procesing without problems. To investigate this point, we wil discuss the advantages and problems of thre strategies: Purely botom-up strategies; purely top-down strategies; and mixed strategies. 387 3.2.1. Purely Bottom-Up Strategy and Their Problems Bottom-up parsers are driven by the words in the sentence, using the rules of the gramar to combine the words into constituent structure in a ?bottom-up? fashion. Let us se how a purely bottom-up parser proceses a sentence ?the man left?. First, let us asume simple phrase structure gramar in (35). (35) S - > NP VP Det - > the, a, every ? NP - > N N - > John, Mary, man, woman, book ? NP - > Det N P - > on, with, about ? NP - > NP P V - > run, say, put, open, read, give P - > P NP VP - > V VP - > V NP VP - > VNP P (36) ilustrates how the sentence ?the man left? is procesed. 388 (36) At step 1, a determiner is found, so the Det node can be projected. At the step 2, a noun ?man? is found, the rule of the gramar alows the projection of N node. If there is a rule in which these categories appear on the right-hand-side, the algorithm can combine them to the category in the left-hand-side of the same rule. At this point we have two categories, det and N that are compatible with the right-hand-side of the second NP rule. Thus at step 3, the algorithm combines these categories into one NP. That is, they are combined using the rule NP-> Det N rule. The same proces continues until al the words are procesed. At the end of the proces, the algorithm projects the S node, and we are left with a complete sentence. 3 It is important to note that step 2 and step 3 in (36) can be applied at any point. Within each step, there can always be alternatives if a word has an ambiguity. 3 This is quite a simple and informal ilustration, but for our purposes this informal ilustration is suficient. Basicaly this ilustration, and the folowing ilustrations of the top-down and left-corner algorithm are adopted from Crocker (199). For the details, readers are recomended to se Crocker (199) and reference therein. 389 Thus if an ambiguous word like bank which can be either N or V is in the input, projecting the N or projecting V are the alternatives for this word. Also there can be alternatives if more than one phrase structure rule has a right-hand-side that matches the categories. The advantage of this algorithm is that the structure is built using the bottom up information of each lexical item in the input, and there is no necesity of combining the category, the algorithm does not make severe erors. The categories can be combined when the item is found in the input that can specify which rule should be applied to project the category. Although it has an advantage, there are two problems with this algorithm. First, within this algorithm the sentence is not procesed incrementaly. This problem becomes obvious when we look at a sentence with a right-branching structure like (37). (37) [ IP The dog [ VP saw [ NP the cat [ CP that [ IP [ VP chased [ NP the mouse]]]. Within the botom-up algorithm, the CP node cannot be atached to the NP node until the word ?mouse? at the end of the sentence is procesed. This is so because [ CP that ?] cannot be projected before the [ IP [ VP chased [ NP the mouse]] is projected, the [ IP [ VP chased ?]] cannot be projected before the [ VP chased [ NP the mouse] is asembled, and the [ VP chased ?] cannot be projected before the [ NP the mouse] is asembled. Thus in the right-branching structure, the algorithm has to delay the proces until the very end of the sentence. 390 The second problem is related to the first problem. Because of the first problem ilustrated above, the algorithm has to store the completed categories uncombined until the end of the sentence in the right-branching structure. Thus, this algorithm predicts the dificulty with the purely right-branching structure, which normaly does not create dificulty (Babyonyshev and Gibson 1995; Chomsky and Miler 1963 Gibson 1991; Yngve 1960 among others). For this reason, the purely bottom-up algorithm is not a reasonable algorithm for human sentence procesing. 3.2.2. Purely Top-Down Strategies and Their Problems Another possible strategy is a purely top-down strategy. In contrast with the bottom-up algorithm, the top-down algorithm starts procesing from postulating the top node, S. Thus, the proces proceds from top to bottom. (38) ilustrates how the sentence ?the man left? is procesed using the gramar in (35). 391 (38) At step 1, the algorithm initiates the proces by postulating the S node. The algorithm can find the rule in which the left-hand side corresponds to the postulated category. So at step 2, the algorithm applies the S - > NP VP rule, and expands the tre. In this way, the algorithm repeats applying the rule for the first daughter of each node until the bottom of the tre is reached. When the bottom of the tre is reached, the algorithm looks back up the tre, and finds another node that can be expanded in the same way. The parser repeats this proces until al the nodes are expanded and reach the bottom of the tre. An obvious advantage of this algorithm is its incrementality. Because the algorithm proceds strictly from top-to-bottom and left-to-right, the sentence is 392 procesed incrementaly. When the parser reaches the terminal ?the? in the subject of the sentence in (38), al the necesary nodes are procesed and integrated into the structure. Although incrementality is preserved, there are some problems with this strategy. First this strategy predicts dificulty with the left-branching structure like (39). Because the proces proceds from the first daughter of S to right, and the algorithm does not look back up the tre before it reaches the terminal of the first daughter, the parser has to store many incomplete NP node before it reaches the terminal My in (39). Thus, the more left-recursion occurs in the sentence, the more NP nodes are there for the parser to keep track with, and the left-recursion structure is predicted to be dificult to proces contrary to fact (Abney and Johnson 1991; Babyonyshev and Gibson 1995; Gibson 1991 among many others). (39) [ NP [ NP [ NP [ NP My cousin?s] aunt?s] dog?s] tail] fel of. On the other hand, this strategy does not have problems procesing the right- branching structure as in (40). In the right-branching structure, the terminal of each node ?comes soon? without many intervening incomplete categories. (40) [ IP [ NP the dog][ VP saw [ NP the cat [ CP that [ IP [ VP chased [ NP the mouse]]]]. The algorithm incorporating the purely top-down strategy, therefore, predicts that the left-branching structure is more dificult than the right-branching structure if the 393 number of embedded is the same. Furthermore, the algorithm predicts that predominantly left-branching languages like Japanese are dificult to proces. However, there is empirical evidence that left-embedding structures are not dificult to proces for humans, and they are certainly easier than, for example, center- embedding constructions (Abney and Johnson 1991; Babyonyshev and Gibson 1995; Gibson 1991, 1998; Mazuka et al. 1989; Miler and Chomsky 1963 among many others). Thus, a purely top-down strategy is not a reasonable strategy for human parser. 3.2.3. The Left-Corner Strategy Because of the problems with purely bottom-up and purely top-down parsing algorithm, it has been proposed that human parser operates in partialy top-down and partialy bottom-up fashion. One such mixed algorithm is a left-corner algorithm or left-corner strategy (Abney and Johnson 1991; Aho and Ulman 1972; Babyonyshev and Gibson 1995; Gibson 1991; Johnson-Laird 1983; Kimbal 1973, 1975; Resnik 1992; Stabler 1994). The left-corner algorithm proceses a sentence in a way in which it proceses the leftmost daughter of the right-hand side of a rule of the gramar in a bottom-up fashion, and applies the rest of the rule from the top-down fashion. Let us se how a left-corner parser proceses the sentence ?the man read the book.? 394 (41) The man read the book. In (41), when the algorithm finds the word the, it alows a Det node to be built. Det is the left-most category at the right-hand side of the second NP rule in (35). The second NP rule tels that it can be expanded as Det and N (NP - > Det N). Thus the sibling of the NP, the N node is projected. The word man comes into the input, and it alows the N node to be built, which can be atached as the right-hand member of the NP, and NP is completed by ataching the newly projected N into the predicted N node. When the NP is completed, the S rule in (35) tels that NP is the left-most category of the right-hand side of S. Thus, the NP alows the S node and its sibling, VP-node to be 395 built. The word read comes into the input, which can project up to V and VP (the step 6). VP is the right most member of the currently constructed S node. So the newly projected VP is atached to the predicted VP (the step 7). The word the comes into the input and as we have sen it can project an NP node and an N node, the sibling of the NP can be predicted (the step 8 and 9). The word book comes into the input, projecting the N node. The newly projected N node is incorporated into the predicted N node (step 10 and 11). The newly projected NP is the right-most category of the current VP node, thus the newly projected NP can be incorporated into the predicted NP node (step 12) and the sentence structure is completed. An advantage of this strategy is that it alows for more incremental parsing than purely bottom-up parsing because it alows some constituents to be integrated into the structure before the constituent has been completed. Furthermore, because initial rule application is driven by bottom-up information, a left-branching structure does not create a problem. Additionaly, a left-corner parser can resolve a problem with a bottom-up algorithm such as the categories are stored unatached until the end of the sentence in the right-branching structure. Because a left-corner algorithm alows a certain prediction, and categories are integrated into the structure as soon as it becomes clear that the predicted node and the newly projected node are the same categories, the parser does not need to store the categories unatached until the end of the sentence even in the right-branching structures. 34 34 This part is treated diferently depending on the type of left-corner algorithm. For example, the so- caled Arc Eager algorithm alows for the parser to integrate even incomplete categories into the 396 Because it can resolve the problems with purely bottom-up or purely top- down strategies, left-corner strategies are understood as one of the most psychologicaly plausible parsing algorithms. However, this algorithm is not totaly fre from problems. Schneider (1999) points out one of the problems with the standard left-corner algorithm. The problem is concerned with the procesing of head- final languages. As we have sen earlier, a head-final language like Japanese shows masive temporary ambiguity because of its head-finality and the lack in the left-edge markers of constituents (Hirose 1999). Because of the left-edge ambiguity, in languages like Japanese, the same string of words can be continued in a various ways as ilustrated in (42). (42) Sideny-ga Rosa-ni ringo-o? S-nom R-dat apple-acc a. ? ageta. (Simple Clause) ? gave ?Sideny gave the apple to Rosa.? b. ? ageta-to itta. (Complement Clause) ? gave-comp said. ?Sideny said that he gave the apple to Rosa.? structure. On the other hand, the so-caled Arc Standard algorithm does not alow the incomplete categories to be integrated into the structure and the same type of problem arises as the botom-up algorithm. In this disertation, we wil adopt the arc eager algorithm, and not talk about the other type of algorithms. Se Abney and Johnson (191), Stabler (194) and Schneider (199) for detailed related discusion. 397 c. ? ageta-node yorokonda. (Adjunct Clause) ? gave-because be-pleased. ?Sidney is pleased because he/someone gave the apple to Rosa.? d. ? ageta hito-o shookai-sita. (Relative Clause) ? gave person-acc introduced. ?Sideny introduced to Rosa the person who he gave the apple to.? Recal that even in a head-final language like Japanese, the string in (42) is interpreted before a licensing head like a verb is encountered. This, in turn, suggests that the parser commits itself to a particular analysis before the verb is encountered. However, if it is true, the left-corner parser predicts that (42b, c, and d) create dificulties because any of these continuation requires some reanalysis. When the string in (42) is encountered, the left-corner parser projects the structure in (43). If there is no other bottom-up information, the algorithm does not project the other structure. However, if the parser commits to this particular structure, when the verb with the complementizer, or the relative head is encountered, the structure must be reanalyzed from a simple clause structure in (43) to other structures such as a complement clause structure, an adjunct clause structure or a relative clause structure. If such reanalysis creates dificulty (Pritchet 1992b; Sturt 1997), the continuations in (42b, c, and d) should be dificult to read. However, none of these continuations are particularly dificult. 398 (43) Given that al the continuations in (42) are read easily, how can we derive this fact without sacrificing the advantages of Left-Corner Algorithm? 3.2.4. SPARSE (Schneider 1999): Incrementality and Flexibility Facing the problem of the left-edge ambiguity discussed above, Schneider (1999) proposes a variant of left-corner parsers that has sufficient flexibility in predicting the upcoming structure. This parsing algorithm is caled SPARSE. As we wil se shortly, SPARSE resolves al the major problems of top-down and bottom-up parsing in the same way as the left-corner algorithm, and it further resolves the problem of left-edge ambiguity without sacrificing incrementality. Thus we wil adopt SPARSE as the syntactic structure building algorithm. One of the key features of Sparse is that it is not forced to predict a specific syntactic category. Rather, SPARSE make a prediction of the upcoming structure on the basis of features. This is made possible by adopting Bare-Phrase Structure (Chomsky 1995). In other words, in SPARSE, the minimal building block acesible to the parser is the syntactic feature rather than syntactic category. SPARSE predicts 399 the upcoming structure in a particular manner. The parser is alowed to predict structure only if a fully connected tre cannot be built otherwise. Thus the predicted features are required to license the currently procesed word into the structure. Let us se how a sentence like (44) is procesed with SPARSE. 35 (44) Dorothy wil se munchkins. When the parser encounters the word Dorothy, the parser does not build any structure. However, the features conveyed by the word are activated. (45) Dorothy: [Cat: Noun], [Case: Nom, Ac], [Num: Sg], [Person: 3] This feature bundle can be read as ?the word Dorothy conveys the categorical feature ?Noun?, the Case feature either Nominative or Acusative, Number feature ?singular? and the person feature ?third person??. As we can se clearly, each lexical item is understood as the bundle of features in SPARSE folowing the formalism of Bare Phrase Structure. The next word, wil, is encountered. The lexical entry of wil contains at least the ones in (46). 35 The discusion here basicaly folows Schneider (199)?s Chapter 2. 400 (46) Wil Inherent Features Licensing Features [Cat: V] [Case: Nom, Left] [Num: {Sg, Pl}, Left] [Person: {1,2,3}, Left] [Cat: V, Right] [VForm: Infin, Right] The feature structure in (46) ilustrates another important feature of SPARSE. SPARSE distinguishes two types of features that can be stored in the lexical entry for each word: the inherent features and the licensing features. The inherent features include a type of the feature such as Case or Category, and the value of the feature, e.g., Nominative, Acusative or Dative for Case, and so on. Inherent features specify the gramatical category of a word, its morphological features and so on. Licensing features, on the other hand, define the relations such as Case asignment, selectional relations, or theta role asignment. Because directionality of these relations is diferent from language to language (acording to, for example, the head directionality parameter (Travis 1989), the licensing features include such information of directionality also in a way as [Case: Nom, Left]. When the parser encounters a new ord, its licensing features are looked up in order to determine whether they can license the previously encountered word, Dorothy. If they can, Dorothy is atached to wil. In this case, wil can license the 401 [Cat: Nom] feature of Dorothy by the licensing feature [Cat: Nom, Left]. Now the structure in (47) is built. (47) wil Dorothy wil [Case: Nom, Ac] Licensing Feature Inherent Feature [Num: SG] [Case: Nom, Left] [Num: {SG, PL}] [Person: 3] [Num: {SG, PL}, Left] [Person: {1,2,3}] [Person {1,2,3}, Left] In SPARSE, such relations as complement or specifier are expresed by the non-null intersection of the values of the features that are shared by two combined elements. For example, in (47), the value ?singular? or the value ?third person? is chosen from the set of number features or person features for wil as a result of an unambiguous Number or Person feature on the noun Dorothy. (47) shows features that are evaluate at this point. The underscored values in the feature sets in (47) are checked and specified by intersecting the set of the features of the combined heads. The parser encounters the next word, se. When se is encountered, the parser examines its features to determine whether it can be atached to the previous word, wil. The inherent features of se are [Cat: verb] and [VForm: Infin]. It can be licensed by wil, and combined as a complement of it. This step is ilustrated in (48). 402 (48) wil Dorothy wil wil se [Category: verb, Right] [Cat: verb] [VForm: Infin, Right] [VForm: Infin] Finaly the munchkins comes into the input. Other than the number feature, it has the same features Dorothy in (45). (49) munchkins, [Case: Nom, Ac], [Theta: Theme] The inherent feature of munchkins, [Case: Nom, Ac] is intersected with the licensing feature of the previous verb se, [Case: Ac, Right]. So munchkins is integrated into the structure as the complement of se. The resulted structure is in (50) 403 (50) wil Dorothy wil wil se se munckins [Case: Ac, Right] [Case: Nom, Ac] [Theta: Theme, Right] [Theta: Theme] So far we have sen how the structure of simple sentence can be parsed by SPARSE. Now let us look at how SPARSE handles left-edge ambiguity. The important point in the following discussion is when and how the structure is predicted. Schneider proposes that heads are predicted only when they are needed to alow for an incremental parse. In other words, predicted heads are posited only when the structure cannot be built incrementaly if they are not posited. First let us se how the predicted heads are posited using relatively a simple example. Schneider cites (51) from German. (51) NP-nom NP-ac Vtrans ? dass er den Hund sah ? that he-nom the dog-acc saw ?? that he saw the dog? (e.g., I know that he saw the dog.) When the nominative pronoun er is encountered, it cannot be atached into the main tre, because the previously encountered complementizer, dass does not asign nominative case. Because er cannot be atached into the existing structure, the 404 distinguished feature for er, nominative case, is used as the basis for a licensing head. Now the lexicon is searched for heads that can asign nominative case, and T head is the only head that asigns nominative case. Therefore, a T head is posted as the licenser for the nominative pronoun, and the entire pronoun-tense complex is atached as the complement of the complementizer dass. (52) Existing Structure das Incoming Material er [Case: Nom] In this case, only one type of head can license the nominative pronoun. Thus it can be guaranted that a T head wil license er in the tre. When the determiner den is encountered, again this word cannot be directly atached to the existing tre. Therefore the search for the licensing head is initiated. In this case however, there are multiple possible heads that can license the determiner because, the inherent features are the [Case, Ac, Dat]. The head that can license this determiner is either a verb or a postposition. The only shared properties of these heads are that they asign case. Thus SPARSE posits an underspecified head that 405 contains only the licensing feature [Case: {Ac, Dat}, Left]. and the determiner- predicted head complex is incorporated into the existing structure. (53) Existing Structure Incoming Material den [Case: Ac, Dat] In this case, the predicted head is underspecified head with only the licensing features, therefore, either the verb or postposition is compatible with the predicted head and the desired flexibility is obtained. I wil skip the remaining parts here. Schneider also cite an example from Japanese and shows how a head-final relative clause structure in (54) could be parsed. (54) Hirosi-ga [ NP [ CP [ IP ? Masao-o mita]] otoko]-o ? H-nom M-acc saw man-acc 406 This type of structure can be parsed without conscious dificulty (se Mazuka and Itoh 1995 for related discussion and observation). At the point imediately before the NP otoko-o ?man-ac? is encountered as in (55), it has the structure of a simple SOV clause as in (56). (55) Hirosi-ga Masao-o mita H-nom M-acc saw ?Hirosi saw Masao.? (56) T Hirosi T saw T Masao saw When the NP otoko-o is encountered, it can only be interpreted as the head of a RC. To build an RC, the entire VP-T constituent must be reanalyzed from a position inside the matrix clause to a position inside an RC. The subject NP Hirosi is left unafected. When the NP otoko-o is encountered, it cannot be atached directly to the existing tre. The search for predicted heads is initiated. A null C head for an RC along with the asociated operator is found, that adjoins to the left of an NP. Schneider speculate that the C head of an RC selects for an NP, so it wil be returned by the search for elements that select for NPs. Once this predicted structure is built, the VP-T constituent from the existing tre can be stripped off from the simple clause structure and can be taken as the complement of the CP. As a consequence of positing an operator, a trace is inserted as the subject of saw. 407 (57) Existing Structure Hirosi To Existing (Cont?d) Incoming Material man To Incoming (Cont?d) Existing Strictire (Cont?d) Hirosi Incoming Material (Cont?d) 408 Let us summarize the discussion so far. The most important aspect of SPARSE is that it preserves incrementality as wel as the flexibility of the prediction. Because of its flexibility, SPARSE can succesfully deal with the left-edge ambiguity without problems. In this sense, SPARSE is one of the most reasonable parsing algorithms for head-final languages. Therefore, in this work, we adopt SPARSE as the basic structure building algorithm. 4. How to Block the Longer Dependency? From now, we wil investigate how SPARSE can derive a prediction for an upcoming RC structure or conditional clause structure. In the course of the discussion it wil be made clear that we have to modify SPARSE to capture the predictions of these structures. In this section we wil concentrate on showing how the non-interogative complementizers can be predicted. In the course of the discussion, we wil se a specific way to enhance the predictive component of the parsing algorithm. For the sake of simplicity, I wil keep most of the discussion here informal. 4.1. Dealing with the Indirect Cue First let us discuss the case of RC prediction. As we have discussed, to capture the fact that the longer dependency bias is blocked in a predicted RC, it is sufficient to predict the Adnominal form of the embedded verb because it is not compatible with the motivation for the LDB, the wh-feature. In the discussion on the syntax of 409 RC, it has made clear that the Adnominal morphology is formed by V-T (finite) -C (empty) . Thus, the parser has to predict this verbal complex by means of the clasifier mismatch. With this discussion in mind, let us se whether SPARSE can predict the Adnominal morphology. First let us examine an example of simple NP that does not contain a clasifier mismatch as in (58). (58) san-nin-no gakusee-ga ? three-cl (person) -gen student-nom ?Thre students ?? First the genitive marked clasifier is encountered and its feature structure is activated. (59) san-nin-no 36 [Cat: Clas], [Case: Gen], [Sem: Person] Then the second word comes into the input. It has the following feature structure. 36 In order to kep the discusion as simple as posible, we wil not be concerned with the mater of how each head are atached to the case-particle, or how each clasifier is atached to numeral or genitive case particle. 410 (60) gakuse-ga: Inherent Features Licensing Features [Cat: N] [Case: Gen, Left] [Case: Nom] [Cat: Clas, Left] [Sem: Person] [Sem: Person, Left] Some notes are in order. We have to specify the licensing feature so that we can capture the relation betwen the clasifier and its host noun. Basicaly clasifiers are licensed by a certain head noun, and they must have specific semantic relations to the head noun. We can capture such relations by postulating a licensing feature [clas: person, left] for example. The feature for clasifiers must specify the semantic features, and this feature and noun?s inherent semantic feature must be the same. N heads also license genitive case. Thus in the same fashion, we can postulate a licensing feature for the genitive case like [Case: Gen, Left]. Now the parser inspects the licensing feature of gakuse-ga to determine whether it can be combined with the previously encountered clasifier. The clasifier is atached to gakuse-ga as a result of licensing the [Case: Gen] feature and [Sem: Person]. (61) 411 Let us turn to an example of an RC containing clasifier mismatch in (62). (62) san-satu-no [ NP [ CP [ IP gakusee-ga [ VP gap yon]-da]?] hon] three-cl (bok) -gen student-nom read-past-? book ?Thre books that the student read.? What we have sen is that when the semantic incompatibility betwen the clasifier san-satu-no ?thre-clasifier-gen? and its adjacent nominative NP, gakuse-ga ?student-nom? is encountered, the parser can predict the morphology of the embedded verb. When the parser encounters the genitive-marked numeral clasifier, its feature structure is activated. Because clasifiers have narowly specified semantic information, we can add semantic features to its inherent features. (63) san-satu-no [Cat: Clas], [Case: Gen], [Sem: Book (or book like object)] The next NP gakuse-ga is then encountered. It has the features as we have sen above, repeated here as (64). 412 (64) gakuse-ga: Inherent Features Licensing Features [Cat: N] [Case: Gen, Left] [Case: Nom] [Cat: Clas, Left] [Sem: Person] [Sem: Person, Left] Now the parser inspects the licensing features of gakuse-ga in order to determine whether it can license the previously encountered clasifier. Because, the [Case: Gen] and [Cat: Clas] on the clasifier can be checked, the parser atempts to combine these two. However, when they are combined, it becomes clear that [Sem: book] cannot be checked by [Sem: Person, Left]. This semantic mismatch initiates the parser?s atempt to reanalyze the structure. A necesary asumption here is the parser tries to combine these two elements because there can be non-null intersection of the values in the two sets of features. Under this view, the clasifier mismatch efect is a reflection of the parser?s atempt to combine these two elements. (65) 413 Now the reanalysis takes place, and initialy combined elements are separated. Because of this reanalysis, these two elements cannot be atached into the structure, and the search for predicted heads is initiated at this point. First, the lexicon is searched for heads that can license the clasifier. This search returns only one type of head. Because the clasifier bears genitive case the head that can license genitive case is searched for. 37 It is a Noun. Furthermore, the clasifier is semanticaly specified, in this case it is specified for books or ?book like objects?. For the sake of exposition, here let us asume that the clasifier is unambiguously specified for books. Thus, the noun with the semantic value of book is the one that can license this clasifier. The clasifier, then projects the structure in (66). 37 Note that this prediction involves the prediction of adjunction structure because it is the structural licensing relation betwen the clasifier and the host NP as we have sen in Chapter 3. 414 (66) Next, the nominative-marked NP cannot be incorporated into the structure either. Thus the predicted head is searched in the lexicon. Asuming that nominative Case in Japanese is licensed uniquely by finite T head (Takezawa 1987), this search also returns a single head, the finite T. Thus, the following predicted structure can be built. 38 38 Both the prediction of the finite T for the nominative NP and N head for genitive clasifier are made based on overtly marked case information. In Japanese, if an NP is not explicitly case marked, it is not posible to tel which Case it bears, without considering such information as verb?s argument structure. Thus it is plausible to asume that in a situation where NPs do not bear Case markers, these specific predictions canot be made. Rather the underspecified heads should be predicted in such a case. A note, however, is in order. Although these asumptions are inded plausible, it is an interesting question to ask what is predicted if NPs do not bear Case. In order to investigate this question, Japanese is not a god language. In Japanese Case droping sems to be wel-constrained, although its nature (se Kageyama 193; Yoshida and Yoshida 193 among many others). Wel-know generalizations are the folowing. The nominative Case on the subject NP and Genitive Case are normaly not omited. Given these properties of Japanese Case particles, it is dificult to test the situation where Case particles are droped. In this regard, a language that alows Case droping or that does not employ explicit case morphology (such as Chinese) is interesting language to test. The same question is adresed in Hsu (206). 415 (67) Now, we have two predicted structures. The parser then tries to incorporate these two predicted structures. However, N head does not license the category T. Thus these two structures cannot be combined, and the proces is doomed to breakdown. 416 (68) Existing Structure Existing Structure For these two structures to be combined gramaticaly, ?the structure of an RC? must be predicted, and it means that the CP on the top of the predicted IP must be predicted, and it must be atached to the predicted NP as an adjunct. 4.2. Enhancing the Predictive Component It shal be clear now that the predictive power of SPARSE is too limited to acommodate the prediction of the RC structure. The question here, then, is whether we can modify the algorithm so that we can acommodate the RC prediction or the algorithm should be discarded altogether and we should sek a new algorithm. The key points are how can the CP structure be predicted in addition to the predicted IP, and how it can be incorporated into the structure as an adjunct of the predicted NP. In addition to these key points, we have to sufficiently restrict the 417 power of the predictive mechanism so that the predictive mechanism does not become too powerful. Let us se how we can modify SPARSE to alow for a prediction of CP on top of the predicted IP. The modification that we propose is to alow the parser to aces the information of the predicted head. As we have sen, Nominative NPs in Japanese are licensed by the finite T. Thus, it is plausible to asume that if an NP bearing the Nominative Case is procesed, SPARSE can predict the finite T rather than an underspecified licensing head. If a finite T head is predicted, its intrinsic features should also be predicted. Asuming that a finite T must be selected and thus be dominated by a C or CP (Chomsky 1986b), we can incorporate this information as finite T head?s intrinsic feature, i.e., if a finite T is predicted, it is also predicted that there must be a C head that selected the finite T. This, in turn, implies that if the predicted T is not a finite T, it does not have such an intrinsic feature. So for example, a non-finite T can be selected by C (in control construction), V (in ECM construction) or P (in ECM construction) (Chomsky 1986a). Thus, to preserve flexible structure building and avoid potential erors, the parser should predict an underspecified head, if non-finite T is specificaly predicted. Consequently, the environment in which C is predicted is quite limited. A question arises with respect to the recursive prediction. We have sen that the predicted finite T alows for a further prediction of C. Now, the question is whether the predicted C head alows for a further prediction of the structure or not. Considering the selectional requirements of C, the answer is no. A C or CP is selected by a wide variety of categories. For example, C is selected by a V such as think, by an N such as claim and so on. Furthermore, it can be a root category 418 also. Thus, as in the case of non-finite T, a prediction set by the predicted C should be an underspecified head or given the possibility of root-C, it should not predict anything. Thus, we can plausibly asume that the predicted C head can predict an underspecified head that licenses C. Given that the predicted C can only predict the underspecified head, the recursive prediction should be stopped at this point. It is not plausible to asume that the predicted underspecified head can predict some category. Thus, when the underspecified head is predicted, the recursive prediction is terminated. In this way, we can alow the parser to predict the C head, and we can stil sufficiently restrict the predictive power of the parser. The prediction of a finite T alows also for a prediction of the V head. Either finite T or infinite T requires V. We can asume that the licensing feature for V is specified in T?s lexical information. Thus if T is predicted, V is also predicted. Now let us turn to the second part, i.e., the question of how the parser ataches the predicted CP as an adjunct of the predicted N. There are at least two plausible ways to consider. The first possibility is to alow the parser to just concatenate the predicted C and the predicted N. The other possibility is to make use of the information of the predicted head. Simple Merger Approach Let us discuss the first possibility. Because SPARSE incorporates Bare Phrase Structure (Chomsky 1995) as a theory of phrase structure, it can simply merge any two existing elements. Thus, the predicted C head and the predicted N head can be simply merged. 419 If we adopt this approach, what is the course of structure building that the parser should follow? At the point when these two heads are merged, however, the projection or the label of the resulted complex cannot be determined. In other words, this parser does not specify whether the C is atached as an adjunct or a complement at the point of merge. The projection can be determined when the lexical N head comes into the input, and its licensing features are acesed by the parser. If the N head takes a complement clause, the C is analyzed as a complement, but if the N head does not take any complement, the C is analyzed as an adjunct. This possibility, therefore, appeals to the benefit of the flexible structure building to the utmost extent. 39 Note however that in this approach we have to asume that there is not a preference of the parser for the complement clause analysis over the adjunct clause analysis. If there is such a preference, and if we incorporate such a preference in our mechanism in a way that ataches C to N as a complement whenever possible, we predict that the building of a RC structure always involves a reanalysis from the prefered complement clause analysis to the les prefered adjunct analysis. Therefore, if there is no complement clause preference, this approach alows for a right prediction. There is, however, a problem with this approach. Under this approach, the structure of the RC can be succesfully built. However, the morphology of the embedded verb, key information for blocking the longer dependency bias, is not 39 Discusion with Amy Weinberg and Jef Lidz led me to this posibility. I would like to expres my gratitude to them for leting me realize this simpler aproach. 420 predicted. From the simple merger of C and N, the morphology of the embedded verb does not follow. Thus, if we do not find a way to predict the predicate adnominal form of the embedded verb, this approach is not appealing. Another problem of this approach is if the parser alows simple merger without taking into consideration the features of predicted heads, it alows for the merger of the predicted T and N without projecting the C head. The resulting structure is obviously problematic. Asuming that the structure of RCs involves adjunction of C to N, this approach alows for a gramaticaly non-legitimate structure. Thus, to maintain simple merger approach, we have to resolve these problems, which at this point, we do not have any way. Acesing the Lexical Contents of the Predicted Head Another possibility is to alow the parser to make use of the information of the predicted head. Based on selectional and licensing properties of clasifiers, it is possible that when the clasifier is procesed, a specific N head is predicted as we have sen. For example, if there is a clasifier for books, satu, in a sentence, then there must be an NP headed by the noun book. Thus, the clasifier creates the prediction for a specific N head. Because a specific noun like book is predicted, its semantic information and licensing features should be acesible by the parser. A noun like book does not take a complement clause. Thus, it does not have a licensing feature for the C. Because of this information, the option of ataching C as a complement of N is excluded. This, in turn, means that if the predicted N can take a 421 complement clause, the predicted C can be atached as a complement. On the other hand, book or any noun can normaly take CP as an RC. RCs or any clausal element are optional elements for a noun. A way to incorporate this information in the system is to asume that each N has licensing feature for C as an optional element. Thus, no mater what the lexical semantic property of the predicted N head might be, if an N is predicted, it can also be predicted that C is licensed by an optional element. 40 Here, let us stipulate that optional licensing features of each head are always acesible but need not to be checked. This is a plausible way to acommodate the optionality of elements such as modifiers, complements of NPs and so on. Under this view, if the category N can be predicted, the optional features of the predicted N head are also predicted. Based on the optional licensing feature for C, the predicted C and the predicted N can be combined. A further modification is required for this approach. As we discussed earlier, for us the morphology of the embedded verb is crucial. Thus, we have to modify the parser so that it can predict the morphology of the embedded verb. The morphological requirement that we have to capture is that an RC must have predicate adnominal morphology on its embedded verb (or any predicate that is incorporated into an NP must have adnominal morphology). This means that a N and the form of the predicate that is asociated with the N has a one-to-one relation. Given this one-to-one relation, we can plausibly asume that each N head has a specification for adnominal form. We 40 The parser adopted by Gibson (191) and Babyonyshev (196) has the same type of mechanism. Their parser alows for the prediction of the optional element that can be ?ignored? if the lexical item compatible with the predicted optional node does not come into input. 422 can capture this requirement by the licensing features, i.e., N has an optional feature that licenses predicate adnominal form. In the case of RCs, that the embedded verb has the predicate adnominal form eans that the complementizer is empty. Therefore, we can asume that each N head has an optional licensing feature for empty C. Modified SPARSER Let us se how these modifications work in predicting and incrementaly structuring the upcoming RC in the environment of the clasifier mismatch. We take an example of (62), repeated here as (69). (69) san-satu-no [ NP [ CP [ IP gakusee-ga [ VP gap yon]- da]?] hon] three-cl (bok) -gen student-nom read- past-? book ?Thre books that the student read.? The problem of this construction arises after the parser commits to reanalysis. The crucial step is when the parser reaches (70). 423 (70) Existing Structure Existing Structure At (70), the N head that can license the clasifier is predicted, i.e., the N head with the semantic feature of book. Additionaly, the finite T head is predicted by means of the nominative-marked NP. The modified SPARSE can aces the licensing features of the predicted N, including its optional licensing features. But no licensing features can license the predicted T head. Therefore, these two existing structures cannot be combined. Because these two heads cannot be combined, the further prediction must be set in which a head that can license T is searched for in the lexicon. As we have discussed, finite T must be licensed by C. Therefore, the lexicon returns a C head. The C is projected and combined with T. At the same point, the licensing feature of the finite T alows for the top-down prediction of the V head. Note that the category N, as the complement of V, cannot be predicted because the V?s subcategorization information is not clear if a lexical V is not inserted. Thus, even though the verb can 424 be predicted, the complement of the verb cannot. Put diferently, the predicted V in this case is not specific. Now, the modified SPARSE can combine the predicted C and N. The optional licensing features of N can license the existing C head. Thus, the predicted C head can be combined with the predicted N head. The result is (71). (71) Existing Structure Existing Structure For the sake of explication, let us summarize the proces from (70) to (71) step-by-step. 425 (72) Step 1: Search for features that can license [Case: Genitive]. Step 2: Project N head Step 3: Search for features that can license [Case: Nom] Step 4: project T head. Step 5: Search for features that can license [T: finite] Step 6: Project C head Step 7: Check optional licensing feature [C: left] on the existing N head against existing C head Step 8: Combine N head and C head. An obvious prediction of this analysis is if the lexical verb is encountered and if it hosts an overt complementizer, the parser has to reanalyze the structure, resulting in a surprise efect. However, if a verb without a complementizer is encountered, the parser incorporates it into the structure without any problem. This is exactly what we saw in the case of clasifier mismatch, i.e., we observed the efect at the embedded verb position. We can ask if this enhancement is too powerful or not. If the predictive component is too powerful, it makes a wrong prediction. This enhancement is, however, wel-constrained. Even though it alows for recursive predictions, it is not always possible. As is explicitly stated, these extra predictions are possible only if a specific head is predicted, and an atempt to combine the predicted structures is sure to failed. 426 The question we have to ask now is whether the predicted representation is sufficient to block LDB. Presumably the fronted wh-phrase can predict the following structure. (73) Other details aside, the C head predicted by means of fronted Wh and the C head predicted by the clasifier mismatch does not match, specificaly in the feature [CForm]. The interogative C must have a Q-particle, and it should be specified in the inherent features. Because of this mismatch, the C head predicted by the clasifier mismatch cannot check the features in C. Thus, the structure in (73) and (71) cannot be combined. As a consequence, LDB is blocked. 4.3. Dealing with the Direct Cue Let us turn to how the SPARSE can manage a prediction of the upcoming conditional clauses. What we want to capture is how the conditional adverb mosi sets a prediction of the upcoming conditional verbal morphology. There are two key points. One is that Mosi has a direct relation to C head. The second key point is that Mosi is also licensed by T head. As it has been extensively explored in the syntax of conditionals, Mosi requires T head. Thus if we can sufficiently specify these relations 427 in terms of features, it wil lead us to understand how Mosi predicts the conditional clauses and how the predicted conditional structure can block the LDB. With the above discussion in mind, let us first se how the adverbs are treated in SPARSE. Schneider (1999) adopts the asumption from Categorial Gramar (Stedman 1996, 2000) in which an adjunct selects its host. This information can be represented by specifying the licensing features of each adverb. For example, a temporal adverb like tomorrow has the following licensing feature. (74) tomorrow: *[Cat: V, Left] (74) can be read as that tomorrow is licensed by the verb on its left. The asterisk on the licensing feature indicates that the lexical entry does not project. This feature specification is, however, not sufficient to capture the distribution of an adverb like tomorrow. As the example (75) shows, it can occur at the sentence final position. By the time when tomorrow is encountered, the parser has built the structure in (76). (75) Dorothy wil se munchkins tomorrow. (76) 428 Because of its licensing features, tomorrow cannot be atached to the most recently procesed word, munchkins, which is an N. When a new word cannot be atached to the structure based on its licensing features, the parser searches the tre upwards along the right edge of the tre for a possible host for the word. 41 The first constituent that is encountered by the search, is the projection of se. Upon encountering se, the parser atempts to compare the features of tomorrow against the features on se. The licensing feature *[Cat: Verb, Left] of tomorrow alows the parser to combine it with se, and as a result, se project the new constituent. This step is ilustrated in (77). 41 For the details of the search proces, se Schneider (199). We wil not go into details of this part because our discusion is no directly related to the isue of how a sentence final materials are incorporated into the structure. 429 (77) Existing Structure Incoming Materials tomorrow *[Cat: V, Left] This treatment of adverbs, however, has a problem with sentence initial adverbs as in (78), as Kazanina (2005) points out. (78) Tomorrow Carmen wil dance flamenco. To acommodate sentence initial adverbs, Kazanina proposes that adverbs have the following features. (79) Tomorrow: *[Cat: {[V, Left], [T, Right]}] By revising the licensing features of adverbs in this way, Kazanina proposes that when the sentence initial adverb is encountered, it projects the T-head by means of 430 *[Cat: T, Right] feature, this is because there is no projection of V in the sentence when tomorrow is encountered in (78). Because of the head-final property, Japanese adverbs always precede its host constituent. Thus, the situation is similar to the sentence initial adverbs in English. Following Kazanina?s proposal, we can capture the licensing relation of Mosi and its hosts by specifying the licensing features. A possible way to do is as in (80). In (80), Mosi?s licensing features specify two types of licensing relations, i.e., the relation with specific type of C and the relation with the Finite Clause. Again we can specify a feature related to semantics. Here we adopt ?clause type? as such semantic features. Let us cal it ?Type?. Other than that the features we use are the same type of features we have used in the previous discussion. (80) Mosi: [Cat: C, Right], [Type: Cond, Right] [CForm: -ra, Right], [Cat: T, Right] [CForm: Fin, Right] A potential problem of this solution is that it alows for the parser to predict multiple heads at the same time when the parser encounters Mosi. This paralel predicted heads can be ilustrated as in (81). 431 (81) Existing Structure Existing Structure The problem of this approach is that we need to constrain when the paralel prediction is possible and when it is not. However, we do not have any wel-motivated constraints. Furthermore, alowing for this type of paralel prediction is almost equivalent to alow for the paralel structure building. Once we alow for paralel structure building in this case, we do not have any reason not to alow for that in the other cases. If the paralels structure building is alowed in general, it is possible to capture the left-edge ambiguity by the paralel structure building. If we can capture the left-edge ambiguity in this way, we do not need to postulate the flexible structure building mechanism like SPARSE in the first place. However, if we alow for the paralels structure building mechanism for ambiguous cases, we have to have a theory of which predicted structure is preferentialy picked up by the parser, and which is not. This approach opens up to an important and interesting isue of the architecture of sentence procesing mechanism. However, this approach may add unnecesary 432 complication to the overal architecture of the parser. So if possible it is beter not going for this type of approach. If we do not appeal to the paralel prediction, how can we capture the licensing relation betwen Mosi and two heads, T and C? Looking closely at the syntax of conditionals, it sems to be true that conditional C can only asociate with a finite T. If so we can employ the same type of mechanism that we proposed in the case of indirect cue, i.e., using the licensing feature of the predicted head. Because Mosi has realy a tight connection to the conditional C, the possible range of the forms of conditional Cs is quite limited. Thus, we can asume that when Mosi predict the upcoming C head, the predicted C head is sufficiently specified. If so, the parser cal for the licensing features of predicted C, exactly in the same way as what N head that is predicted by the clasifier mismatch did in the prediction of RC structure. In order to acommodate this type of prediction, we have to just specify the licensing feature of Mosi and conditional C as in (82). (82) a. Mosi: Licensing features *[Cat: C, Right], [Type: Cond, Right], [CForm: -ra, right] b. C: Inherent features, [Cat: C] [Type: Cond] ,[CForm: -ra] Licensing features [Cat: T, Left], [TForm: Fin, Left] 433 The hidden asumption here is that C has at least thre inherent features, the categorical feature, the semantic feature (the clause type), and specific form. If it is the case that just C is predicted and these features are not specified, the feature representation should look like C: [Cat: C], [Type: Declarative, Interogative, Conditional ?], [CForm: ?, -to, -ka ?]. Because of the number possible clause types and CForms are restricted, it is not problematic to list al of these values in the lexicon. Let us take an example in (83) and se how the parsing proceds with these features. (83) Mosi gakusee-ga kaetta-ra mosi student-nom leav-cond ?If the student leaves ?? First, the parser encounters Mosi. The features of Mosi in (82a) are activated. At this point the C head is projected based on the licensing features of Mosi. Next, because the predicted C is sufficiently specified, its licensing features can be acesed by the parser. Based on C?s licensing features, the T can be projected as the complement of C. 434 (84) Mosi *[Cat: C, Right] [Type:Cond, Right] [CForm: -ra, right] Based on the information on Mosi the parser can build this rich structure before the Subject is encountered. In this system, again the specific verbal morphology is predicted. Thus, if the newly incoming lexical V does not have the predicted form, the parser has to initiate the reanalysis, resulting in the surprise efect. In the experimental studies on Conditional Clauses, we have sen that the efect took place at the position of the conditional verb. This system correctly predicts such efect at the conditional verb position. With the discussion so far in mind, let us examine whether this structure can represent sufficient information to block the Longer Dependency or not. Let us asume that the fronted wh-phrase can project the following structure. (85) 435 Looking at the features of the predicted C, it is clear that the Inherent Features are not compatible with the predicted Conditional C. Specificaly, the clause type feature and CForm features are not matched. If so, the C head predicted by wh-NP-dat cannot be asociated with the C head predicted by Mosi. As a result, the C head predicted by Mosi and C head predicted by Wh-phrase cannot be combined. Thus it is possible to block the LDB using this incompatibility of the predicted C heads. 436 5. Conclusions In this chapter we have se how we can modify SPARSE in order to predict sufficiently rich structure to block the LDB. We have revised SPARSE so that the parser can operate powerful recursive predictions. By this modification we made it possible to predict the specific verbal morphology that is not compatible with Q- particle. We have sen that this prediction of the verbal morphology in principle alows the parser to block the LDB in the predicted RCs or Conditional Clauses by means of the incompatibility of predicted Cs. The remaining isue is how the LDB is actualy blocked. To explore this isue we have to understand how the wh-feature drives the longer dependency. Aoshima?s (2003) system is designed so that al the fronted dative NPs create LDB. However as we have sen, only the wh-phrases create longer dependency. How can we capture this distinction? This question is beyond the scope of this study. So we would like to leave the isue open here. 42 42 This isue is explored in Aoshima et al. (206). Readers are recomended to refer to Aoshima et al.?s work for further details. 437 CHAPTER 6. GENERAL CONCLUSIONS The preceding chapters investigated isues related to the strong incrementality of human sentence procesing. I have argued that the strong incrementality is supported by a powerful predictive mechanism. Throughout this thesis, we have asked the following thre questions. What are potential triggers for the predictive mechanism? What are the representations that the predictive mechanism generates? What is the appropriate algorithm that is responsible for the predictive mechanism? To answer these questions, we have investigated the interactions betwen wh- dependency formation and two specific constructions in Japanese, Relative Clauses and Conditional Clauses from a theoretical viewpoint (Chapter 2), a perspective of real-time sentence procesing (Chapter 3 & 4), and a computational modeling (Chapter 5). We have claimed that the parser of wh-dependencies is driven by the satisfaction of syntactic constraints. Specificaly, we have claimed that wh-feature satisfaction is the driving force behind wh-dependency formation in real-time sentence procesing. Based on this finding, we have investigated how wh-dependency formation interacts with the predictive mechanism. Through a series of experimental studies, we have discovered that when the upcoming RC structure and conditional clause structure are predicted, the long-distance wh-gap dependency is not created. On the basis of these findings, we tried to figure out how the human sentence procesing can be strongly incremental. In chapter 2, we have sen examined detailed syntactic properties of RCs and conditionals in Japanese. In this chapter we have revealed the following properties of 438 RCs and conditionals. We have argued that the derivation of RCs contains the movement of the empty operator on the one hand, and the structure of RC involves CP rather than just IP on the other. The first point, the movement of the empty operator is supported by various observations on the so-caled connectivity efects. The second point, the CP analysis, is motivated by the formation of the so-caled Predicate Adnominal morphology. In this chapter, the detailed internal and external syntax conditional clauses have also been revealed. There were thre important discoveries in this chapter. First, Japanese conditional clauses are not islands in Japanese, even though they are adjunct clauses. Second, conditional adverb Mosi is uniquely licensed by a finite clause and conditional morpheme. Finaly, by examining the complex conditional morphology, we have revealed that conditional clause has multi-layered CP structure. In chapter 3, we have sen the first piece of evidence for the strong incrementality. We have sen that the semantic incompatibility betwen the genitive- marked clasifier and its adjacent NP can set a prediction of the upcoming RC structure. Based on the discovery that the wh-gap dependency does not penetrate into the predicted RC structure, we have argued that the predicted structure must be rich enough to represent the islandhood of RCs. In chapter 4, we have discovered the second piece of evidence for the strong incrementality. We have sen that the upcoming conditional clause is predicted when the conditional adverb Mosi is provided. Based on the observation that the LDB is blocked also in the predicted conditional clauses, we have argued that Mosi sets a 439 prediction of the upcoming conditional morphology that is not compatible with the motivation for the LDB. In chapter 5, we have discussed what type of algorithm can derive the strong incrementality without sacrificing flexible structure building. We have argued that a type of Left-Corner algorithm is psychologicaly more plausible algorithm for human sentence procesing. We adopted a particular parsing mechanism caled SPARSE. We argued that SPARSE can incrementaly build the structure without sacrificing the necesary flexibility. However, even SPARSE cannot predict the full structure of the upcoming RCs or Conditional Clauses. In order to capture the prediction of the upcoming RCs and Conditionals, we have modified the predictive component of SPARSE. Our modification was as follows. A top-down prediction based on the predicted head is possible only if the parser can predict a specific head with specific semantic content. With this modification, the parser can set recursive prediction only in limited environment. This work has focused on the procesing of relative clauses and conditional clauses, and we have sen that sentence procesing is strongly incremental. We have further argued that the strong incrementality is supported by a powerful predictive mechanism. However, there are some remaining isues in this work. We have not made clear how the long-distance wh-gap dependency can be stopped in the predicted RCs and Conditionals. We have shown how these structures are predicted but we have not shown how these predicted structures interact with the LDB. To resolve this isue, we have to understand how the longer dependency formation is created. To 440 achieve the understanding of the LDB it is needed to discover the mechanism of the longer dependency formation. Additionaly, to examine empirical validity of the proposed parsing mechanism, we have to experimentaly investigate the environment where the unambiguous structure is predicted before the head is encountered. In this respect, more studies on head-final languages as wel as head-initial languages are necesary. 441 Bibliography Abney, Steven P, and Johnson, Mark. 1991. Memory Requirements and Local Ambiguities of Parsing Strategies. Journal of Psycholinguistic Research 20:233-250. Abney, Steven. 1989. A Computational Model of Human Parsing. Journal of Psycholinguistic Research 18:129-144. Aho, A, and Ulman, J. 1972. The Theory of Parsing, Translation, and Compiling (Vol. 1). Englewoods Clifs, NJ: Prentice-Hal. Aoshima, Sachiko, Philips, Colin, and Weinberg, Amy. 2004. Procesing Filer-Gap Dependencies in a Head-Final Language. Journal of Memory and Language 51:23. Aoyagi, Hiroshi. 1998. On the Nature of Particles in Japanese and Its Theoretical Implications, University of Southern Calrifornia: Ph.D. disertation. Auster, Paul. 1985. City of Glass. Los Angels, CA: Sun and Moon Pres (Reprinted in Auster, Paul. 1987. The New York Trilogy) Auster, Paul. 1986. Ghosts. Los Angels CA: Sun and Moon Pres (Reprinted in Auster, Paul. 1987. The New York Trilogy) Ausster, Paul. 1987. The New York Trilogy. New York, NY: Penguin Books Auster, Paul. 1989. Moon Palace. New York, NY: The Viking Pres. Auster, Paul. 1994. Mr. Vertigo. New York, NY: The Viking Pres. Auster, Paul. 2004. Oracle Night. New York, NY: Henry Holt & Company. 442 Babyonyshev, Maria, and Gibson, Edward. 1995. Procesing Overload in Japanese. MIT Working Papers in Linguistics 26:1-35. Bader, Markus, and Laser, Ingeborg. 1994. German Verb-Final Clause and Sentence Procesing: Evidence for Imediate Atachment. In Perspectives on Sentence Procesing, eds. Charles Clifton Jr., Lyn Frazier and Keith Reiner, 225-242: Lawrence Erlbaum. Beck, Sigrid. 1996. Quantified Structures as Bariers for Lf Movement. Natural Language Semantics 4:1-56. Berwick, Robert C. 1991. Principles of principle-based parsing. In Principle-based parsing: Computation and psycholinguistics, ed. R.C. Berwick, 1-37: Kluwer Academic Publishers. Bever, Thomas G. 1970. The Cognitive Basis for Linguistic Structures. In Cognition and the Development of Language, ed. J. R. Hayes, 279-360. New York: Wiley. Bhat, Rajesh, and Pancheva, Roumyana. 2001. Conditionals. Ms. University of Texas, Austin, and USC. Bianchi, Valentina. 2000. The Raising Analysis of Relative Clauses: A Reply to Borsley. Linguistics Inquiry 31:123-140. Boeckx, Cedric, and Ninuma, Fumikazu. 2003. Conditions on agrement in Japanese. Natural Language and Linguistic Theory 22:453 - 480. Boland, Julie E., and Boehm-Jernigan, Heather. 1998. Lexical constraints and prepositional phrase atachment. Journal of Memory and Language 39:684. 443 Boland, Julie E., Tanenhaus, Michael K., Garnsey, Susan M., and Carlson, Greg N. 1995. Verb Argument Structure in Parsing and Interpretation: Evidence from Wh-Questions. Journal of Memory and Language 34:774. Boskovic, Zeljko, and Takahashi, Daiko. 1998. Scrambling and Last Resort. Lingustic Inquiry 29:347-366. Browning, Marguerite. 1987. Null Operator Constructions. Ph.D. Disertation, Masachusets Institute of Technology. Chomsky, Noam. 1972. Studies on Semantics in Generative Gramar. The Hague: Mouton. Chomsky, Noam. 1973. Conditions on Transformations. In A Festschrift for Moris Hale, eds. Stephen Anderson and Paul Kiparsky, 232-286. New York: Holt Rinehart & Winston. Chomsky, Noam. 1977. On Wh-Movement. In Formal Syntax, eds. Peter Culicover, Thomas Wasow and Adrian Akmajian, 71-132. New York: Academic Pres. Chomsky, Noam. 1981. Lectures on Government and Binding. Dordrecht: Foris. Chomsky, Noam. 1986a. Knowledge of Language: Its Nature, Origin, and Use. New York: Praeger. Chomsky, Noam. 1986b. Bariers. Cambridge, Mas.: MIT Pres. Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mas.: The MIT Pres. Cinque, Guglielmo. 1984. A-Bar Bound Pro Vs. Variable. Ms. University of Venice. Cinque, Guglielmo. 1990. Types of A'-Dependencies: Linguistic Inquiry Monographs; 17. Cambridge, Mas.: MIT Pres. 444 Citko, Barbara. 2005. On the Nature of Merge: External Merge, Internal Merge, and Paralel Merge. Linguistc Inquiry 36:475-497.Collins, Chris. 1998. A Note on Extraction from Conditionals. In Cornel Working Papers in Linguistics, eds. N. Adisasmito-Smith and T. Doeleman, 57-66. Ithaca, NY: CLC, Cornel University. Collins, Chris. 1998. A Note on Extraction from Conditionals. In Cornel Working Papers in Linguistics, eds. N. Adisasmito-Smith and T. Doeleman, 57-66. Ithaca, NY: CLC, Cornel University. Crocker, Mathew W. 1994. On the Nature of Principle-Based Sentence Procesor. In Perspectives on Sentence Procesing, eds. Charles Clifton Jr., Lyn Frazier and Keith Rayner, 245-266. Hilsdale, New Jersey: Lawrence Earlbaum Asociates, Publishers. Crocker, Mathew W. 1996. Computational Psycholinguistics: An Interdisciplinary Approach to the Study of Language: Studies in Theoretical Psycholinguistics; V. 20. Dordrecht; Boston: Kluwer Academic Publishers. Frank, Robert. 2002. Phrase Structure Composition and Syntactic Dependencies. Cambridge, Mas.: MIT Pres. Frazier, Lyn, and Rayner, Keith. 1987. Resolution of Syntactic Category Ambiguities: Eye Movements in Parsing Lexicaly Ambiguous Sentences*1. Journal of Memory and Language 26:505. Frazier, Lyn. 1987. Syntactic Procesing: Evidence from Dutch. Natural Language and Linguistic Theory 5:519-559. 445 Fukui, Naoki, and Takano, Yuji. 2000. Nominal Structure: An Extension of the Symmetry Principle. In The Derivation of Vo and Ov, ed. Peter Svenonius, 210-254. Amsterdam: John Benjamins. Fukui, Naoki. 1995. Theory of Projection in Syntax: Studies in Japanese Linguistics. Stanford, Calif.: CSLI Publications; Tokyo: Kurosio Publishers. Gibson, Edward. 1991. A Computational Theory of Human Linguistic Procesing: Memory Limiations and Procesing Breakdown, Carnegie Melon University. Gibson, Edward. 1998. Linguistic Complexity: Locality of Syntactic Dependencies. Cognition 68:1-76. Gorrel, Paul. 1995. Syntax and Parsing: Cambridge University Pres. Grimshaw, Jane. 1991. Extended Projection, Brandeis University: Ms. Guimar?es, Maximiliano. 2004. Derivation and Representation of Syntactic Amalgams. Ph.D. Disertation, University of Maryland. Hagstrom, Paul. 1998. Decomposing questions. Ph.D. Disertation, Masachusets Institute of Technology. Han, Chung-Hye, and Romero, Maribel. 2004. The Syntax of Whether/Q.Or Questions: Elipsis Combined with Movement. Natual Language and Linguistic Theory 22:527-564. Harada, Kazuko Ito. 1972. Constraints on Wh-Q Binding. In Studies in Descriptive and Applied Linguistics 5, 180-206. Tokyo: Division of Language, International Christian University. Harada, Shin-Ichi. 1971. Ga-No Conversion and Idiolectal Variation in Japanese. Gengokenkyu 60:25-38. 446 Harada, Shin-Ichi. 1973. Counter Equi Np Deletion. In Annual Bulletin, Research Institute of Logopedics and Phoniatrics, 133-143: University of Tokyo. Harada, Shin-Ichi. 1973a. Constraints on Binding. Studies in English Literature:41- 72. Harada, Shin-Ichi. 1973b. Counter Equi NP Deletion. In Annual Bulletin, Research Institute of Logopedics and Phoniatrics, 133-143: University of Tokyo. Harada, Shin-Ichi. 1974. Remarks on relativization. Annual Bulletin, Reserach Institute of Logopedics and Phoniatrics 7:113-147. Harada, Shin-Ichi. 1976. Ga-No Conversion Revisited. Gengokenkyu 70:23-38. Harada, Shin-Ichi. 1976. Honorifics. In Syntax and semantics 5: Japanese generative gramar, ed. Masayoshi Shibatani, 499-561. New York: Academic Pres. Hendriks, Petra. 2003. "Either" as a Focus Particle. Ms. Center fo Language and Cognition, University of Groningen. Hinds, John. 1973a. Some Remarks on Soo Su-. Papers in Japanese Linguistics 2:18- 30. Hinds, John. 1973b. Some Remarks on Soo Su-. Papers in Japanese Linguistics 2:18- 30. Hiraiwa, Ken. 2000. On Nominative-Genitive Conversion. MIT Working Papers in Linguistics 39. Hiraiwa, Ken. 2001. Nominative-Genitive Conversion Revisited, MIT: ms. Hiraiwa, Ken. 2002. Facets of Case: On the Nature of the Double-O Constraint, MIT: ms. 447 Hirose, Yuki. 1999. Resolving Reanalysis Ambiguity in Japanese Relative Clauses, Graduate Faculty in Linguistics, City University of New York. Hoji, Hajime, Miyagawa, Shigeru, and Tada, Hiroaki. 1989. Np-Movement in Japanese. Ms. University of Southern California, Los Angels, Ohio State University, Columbus, and Masachusets Institute of Technology, Cambridge. Hoji, Hajime. 1985. Logical Form Constraints and Configurational Structure in Japanese, University of Washington: Doctoral Disertation. Honcoop, Martin. 1998. Excursions in Dynamic Binding. Ph.D. Disertation, Leiden University. Hornstein, Norbert, and Uriagereka, Juan. 2002. Reprojections. In Derivation and Explanation in the Minimalist Program, eds. David Samuel Epstein and T. Daniel Sely, 106-132. Oxford, UK Hoshi, Koji. 1995. Structural and Interpretive Aspects of Head-Internal and Head- External Relative Clauses, University of Rochester: Ph.D. disertation. Hoshi, Koji. 2004. Parametrization of the External D-System in Relativization. Language, Culture and Communication 33:1-50. Huang, Cheng-Teh James. 1982. Logical Relations in Chinese and the Theory of Gramar, Masachusets Institute of Technology: PhD. Inoue, Atsu, and Fodor, Janet. D. 1995. Information-Paced Parsing of Japanese. In Japanese Sentence Procesing, eds. R. Mazuka and N. Nagai, 9-36. Hilsdale, NJ: Lawrence Erlbaum Asociates. 448 Inoue, Atsu. 1991. A Comparative Study of Parsing in English and Japanese, Linguistics, The University of Connecticut. Inoue, Kazuko. 1976. Henkee Bunpoo to Nihongo [Transformational Gramar and Japanese Language]. Tokyo: Taishuukan Shoten. Ishii, Toru. 1997. An Asymmetry in the Composition of Phrase Structure and Its Consequences. Ph.D. Disertation, University of California, Irvine. Ishii, Yasuo. 1991. Operators and Empty Categories in Japanese, University of Connecticut: Ph.D. disertation. Jackendoff, Ray. 1972. Semantic Interpretation in Generative Gramar: Studies in Linguistics Series, 2. Cambridge, Mas.: MIT Pres. Jayaselan, K. A. 1990. Incomplete Vp Deletion and Gapping. Linguistic Analysis 20:64-81. Johnson-Laird, P. N. 1983. Mental Models: Towards a Cognitive Science of Language, Inference, and Consciousnes: Cognitive Science Series; 6. Cambridge, Mas.: Harvard University Pres. Just, Marcel Adam, Carpenter, Patricia A., and Woolley, Jacqueline D. 1982. Paradigms and Proceses in Reading Comprehension. Journal of Experimental Psychology 111:228-238. Kamide, Yuki, and Mitchel, Don C. 1999. Incremental Pre-Head Atachment in Japanese Parsing. Language and Cognitive Proces 14:631-662. Kamio, Akio. 1983. Meisiku-No Koozoo [the Structure of Noun Phrases]. In Nihongo-No Kihonkoozoo [the Basic Structure of Japanese], ed. Kazuko Inoue. Tokyo: Sanseidoo. 449 Kaplan, Ronald M., and Zaenen, Annie E. 1989. Long-Distance Dependencies, Constituent Structure, and Functional Uncertainty. In Alternative Concepts of Phrase Structure, eds. Mark Baltin and Anthony S. Kroch, 17-42. Chicago: University of Chicago Pres. Katada, Fusa. 1991. The Lf Representation of Anaphors. Linguistic Inquiry 22:287- 313. Kato, Takaomi. 2005a. Neg-Raising, Last Resort, and the Nature of CSC. Harvard Working Papers in Linguistics 11:79-102. Kato, Takaomi. 2005b. A Case against the Representational Approach to the Coordinate Structure Constraint. In Procedings of the Thirty-Fifth Annual Meting of the North East Linguistic Society (NELS 35), eds. Leah Bateman and Cherlon Usery. Amherst: GLSA. Kato, Yasuhiko. 1985. Negative Sentences in Japanese. Sophia Linguistica 19. Kayne, Richard S. 1994. The Antisymmetry of Syntax. Cambridge: Mas: MIT Pres. Kikuchi, Akira, Oishi, Masayuki, and Yusa, Noriaki. 1994. Scrambling and Relativized L-Relatednes. In Mit Working Papers in Linguistics 24: Formal Approaches to Japanese Linguistics 1, 141-158. Cambridge: Mas: MIT Pres. Kim, Shin-Sook. 2002. Intervention Efects Are Focus Efects. In Japanese and Korean Linguistics 10, eds. Noriko Akatsuka and Susan Strauss, 615-628. Stanford: CSLI Publications. Kimbal, John P. 1973. Seven Principles of Surface Structure Parsing in Natural Language. Cognition 2:15-47. 450 Kimbal, John P. 1975. Predictive Analysis and over-the-Top Parsing. In Syntax and Semantics, ed. John P. Kimbal. New York, NY.: Academic Pres. Kitagawa, Yoshihisa (2006) Naze. Ms. Inidiana University. Kroch, Anthony S. 1989. Amount Quantification, Referentiality and Long Wh- Movement. Ms. University of Pennsylvania. Kroch, Anthony S., and Joshi, Aravind K. 1985. The Linguistic Relevance of Tre Adjoining Gramar. University of Pennsylvania Department of Computer and Information Science Technical Report MS-CIS-85-16. Kuno, Susumu. 1973. The Structure of the Japanese Language. Cambridge, Mas.: MIT Pres. Kuroda, S-Y. 1988. Whether We Agre or Not: A Comparative Syntax of English and Japanese. In Papers from the Second International Workshop on Japanese Syntax, ed. W. Poster, 103-143: CSLI. Kuroda, Shige-Yuki. 1965. Generative Gramatical Studies in the Japanese Language, MIT: Ph.D. disertation. Kuroda, Shige-Yuki. 1974. Pivot-independent relative clauses in Japanese i. Papers in Japanese Linguistics 3:59-93. Kuroda, Shige-Yuki. 1975/76. Pivot-Independent Relative Clauses in Japanese Ii. Papers in Japanese Linguistics 4:85-96. Kuroda, Shige-Yuki. 1976/77. Pivot-independent relative clauses in Japanese, ii. Papers in Japanese Linguistics 5. 451 Kuroda, Shige-Yuki. 1986a. What Happened Agter the Movement of Nps in La Jolla? In Working Papers from the First Sdf Workshop in Japanese Syntax, ed. Shige-Yuki Kuroda, 71-96. La Jolla: University of California San Diego. Kuroda, Shige-Yuki. 1986b. Movement of Noun Phrases in Japanese. In Isues in Japanese Linguistics, eds. Takashi Imai and Mamoru Saito, 229-271. Dordrecht: Foris. Kuroda, Shige-Yuki. 1999. Syuyoobunaizaigata Kankeisetu [Head Internal Relative Clauses]. In Kotoba No Kaku to Shuuen [Core and Periphery of Language], eds. Shige-Yuki Kuroda and Masaru Nakamura. Tokyo: Kurosio Publisher. Lakoff, George, and Ross, John Robert. 1976. Why Can't Do So into the Sink? In Syntax and Semantics, Volume 7: Notes from the Linguistic Underground, ed. James D. McCawley, 101-111. New York: Academic Pres. Larson, Richard K. 1985. On the Syntax of Disjunction Scope. Natural Lanaguage and Linguistic Theory 3:217-264. Lasnik, Howard, and Saito, Mamoru. 1992. Move ?: Conditions on Its Application and Output. Cambridge: The MIT Pres. Lasnik, Howard, Depiante, Marcela A., and Stepanov, Arthur. 2000. Syntactic Structures Revisited: Contemporary Lectures on Clasic Transformational Theory. Cambridge, Mas.: MIT Pres. Lasnik, Howard. 1981. Restricting the theory of transformations: A case study. In Explanation in linguistics, ed. Norbert & David Lightfoot Hornstein, 152-173: Longman. 452 Lasnik, Howard. 1981a. Restricting the Theory of Transformations: A Case Study. In Explanation in Linguistics, ed. Norbert & David Lightfoot Hornstein, 152- 173: Longman. Lasnik, Howard. 1981b. Restricting the Theory of Transformation: A Case Study. In Expanations in Linguistics, eds. Norbert Hornstein and David Lightfoot, 152- 173. London: Longmans. Lewis, Richard L. 1996. Interference in Short-Term Memory: The Magical Number Two (or Thre) in Sentence Procesing. Journal of Psycholinguistics Research 25:93-115. Linebarger, Marcia C. 1987. Negative Polarity and Gramatical Representation. Linguistics and Philosophy 10:325-387. Lombardo, Vincenzo, and Sturt, Patrick. 2002. Incrementality and Lexicalism; a Trebank Study. In The Lexical Basis of Sentence Procesing: Formal, Computational and Experimental Isues, eds. Paola Merlo and Suzanne Stevenson. Amsterdam/ Philadelphia: John Benjamins. Macdonald, Maryelen C, Pearlmutter, Neal, and Seidenberg, Mark S. 1994a. Syntactic Ambiguity Resolution as Lexical Ambiguity Resolution. In Perspectives on Sentence Procesing, eds. Charles Clifton Jr., Lyn Frazier and Keith Rayner, 123-154. Hilsdale, New Jersey: Lawrence Earlbaum Asociates, Publishers. MacDonald, Maryelen C., Pearlmuter, Neal J., and Seidenberg, Mark S. 1994b. Lexical nature of syntactic ambiguity resolution. Psychological Review 101:676-703. 453 Marslen-Wilson, Wiliam. 1973. Linguistic Structure and Speech Shadowing at Very Short Latencies. Nature 244:522-533. Masuoka, Takashi, and Takubo, Yukinori. 1992. Kiso Nihongo Bunpoo [the Basic Japanese Gramar]. Tokyo: Kuroshio Publishers. Masuoka, Takashi, and Takubo, Yukinori. 1992. Kiso Nihongo Bunpoo [the Basic Japanese Gramar]. Tokyo: Kuroshio Publishers. Matsuda, Yuki. 1993. On the So-Caled Japanese Headles Relative Clauses: Complementation Analysis. Ms. University of Southern California. Mazuka, Reiko, and Itoh, Kenji. 1995. Can Japanese Speakers Be Led Down the Garden Path? In Japanese Sentence Procesing, eds. Rekio Mazuka and Noriko Nagai: Lawrence Erlbaum Asociates. Mazuka, Reiko, and Lust, Barbara. 1988. Why Is Japanese Not Dificult to Proces? A Proposal to Integrate Parameter Seting in Universal Gramar and Parsing. Paper presented at NELS. Mazuka, Reiko, and Lust, Barbara. 1990. On Parameter Seting and Parsing: Predictions for Cross-Linguistic Diferences in Adult and Child Procesing. In Language Procesing and Language Acquisition, ed. Lyn Frazier and Jil de Viliers, 163-205. Dordrecht: Kluwer Academic Publishers. Mazuka, Reiko, Itoh, Kenji, Kiritani, Shigeru, Niwa, Shin-Ichi, Ikejiri, Kayo, and Naitoh, Kazuo. 1989. Procesing of Japanese Garden Path, Center-Embedded, and Multiply-Left-Embedded Sentences: Reading Time Data from an Eye Movement Study. Annual Bulletin, Research Institute of Logopedics and Phoniatrics 23:187-212. 454 Mazuka, Reiko. 1991. Procesing of Empty Categories in Japanese [MAY]. Journal of Psycholinguistic Research 20:215-232. McCawley, James D. 1982. Parentheticals and Discontinuous Constituent Structure. Linguistic Inquiry 13:91-106. Mihara, Kenichi. 1994. Nihongo-No Toogo Koozoo: Sese Bunpoo Riron to Sono Ooyoo [Syntactic Structures of Japanese: The Theory of Generative Gramar and Its Application]. Tokyo: Shohakusha. Mikami, Akira. 1960. Zoo Wa Hana Ga Nagai. Tokyo: Kurosi Syuppan. Miler, George A, and Chomsky, Noam. 1963. Finitary Models of Language Users. In Handbook of Mathematical Psychology Volume 2, ed. R Luce, R. Bush & E. Galanter, 419-491: John Wiley. Miyagawa, Shigeru. 1989. Structure and Case Marking in Japanese: Syntax and Semantics; V. 22. San Diego: Academic Pres. Miyagawa, Shigeru. 1993. Lf Case-Checking and Minimal Link Condition. In Papers on Case and Agrement. Mit Working Papers in Linguistics, ed. Colin Philips. Cambridge, MA: MIT Working Papers in Linguistics. Miyagawa, Shigeru. 1998. On Islands. Ms. Masachusets Institute of Technology. Miyagawa, Shigeru. 2002. The Nature of Weak Islands. Ms. Masachusets Institute of Technology. Miyamoto, Edson, and Takahashi, Shoichi. 2002. The Procesing of Wh-Phrases and Interogative Complementizers in Japanese. Paper presented at Procedings of the 10th Japanese/Korean Linguistic Conference. 455 Miyamoto, Edson. T. 2002. Case markers as clause boundary inducers in Japanese. Journal of Psycholinguistic Research 31:307-347. Miyamoto, Edson. T. 2003. Reanalysis of clause boundaries in Japanese as a constraint-driven proces. Language and Speech 46:23-52. Moltmann, F. 1992. Coordination and Comparatives. Ph.D. Disertation, Masachusets Institute of Technology. Munn, Alan B. 1993. Topics in the Syntax and Semantics of Coordinate Structures. Ph.D. Disertation, University of Maryland. Muraoka, Satoru, and Sakamoto, Tsutomu. 2003. The influence of lexical information on the clause boundary procesing in japanese. [nihongo-ni okeru setsu- kyoukai kaiseki-ji-no goi-joohoo-no ekyoo]. Paper presented at Paper presented at Japanese Cognitive Science Society, University of Electro- Communications. Murasugi, Keiko. 1991. Noun Phrases in Japanese and English: A Study in Syntax, Learnability and Acquisition, University of Connecticut, Storrs: Ph.D. disertation. Nakamura, Michiko. 1999/2000. Procesing of Relative Clauses: A Psycholinguistic Study of Adult Native Speakers of Japanese. In University of Hawai at Manoa Working Papers in Linguistics, 59-71. Nakamura, Michiko. 2003. Procesing of Multiple Filer-Gap Dependencies, The Graduate Division, The University of Hawai. Nakau, Minoru. 1973. Sentential Complementation in Japanese. Tokyo: Kaitakusha. 456 Nemoto, Naoko. 1993. Chains and Case Positions: A Study from Scrambling in Japanese, University of Connecticut: Ph.D. disertation. Nishigauchi, Taisuke. 1990. Quantification in the Theory of Gramar. Dordrecht: Kluwer. Ochi, Masao. 2001. Move F and Ga/No Conversion in Japanese. Pesetsky, David Michael. 2000. Phrasal Movement and Its Kin: Linguistic Inquiry Monographs; 37. Cambridge, Mas.: MIT Pres. Pesetsky, David. 1987. Wh in Situ: Movement and Unselective Binding. In The Representation of (in)Definitenes, eds. Eric J. Rueland and Alice G. B. ter Meulen. Cambridge, Mas: MIT Pres. Pollard, Carl Jese, and Sag, Ivan A. 1994. Head Driven Phrase Structure Gramar. Chicago: University of Chicago Pres. Pritchet, Bradley L. 1991a. Subjacency in a Principle-Based Parser. In Principle- Based Parsing: Computation and Psycholinguistics, eds. Robert C. Berwick, Steven. P. Abney and Carol Le Tenny, 301-345. Dordrecht: Kluwer Academic Publisher. Pritchet, Bradley L. 1991b. Head Position and Parsing Ambiguity. Journal of Psycholinguistic Research 20:251-270. Pritchet, Bradley L. 1992a. Gramatical Competence and Parsing Performance. Chicago, IL: The University of Chicago Pres. Pritchet, Bradley L. 1992a. Parsing with Gramar: Islands, Heads, and Garden Paths. In Island Constraints, ed. Helen Goodluck, & Michael Rochemont, 321-349: Kluwer Academic Publishers. 457 Pritchet, Bradley. 1988. Garden Path Phenomena and the Gramatical Basis of Language Procesing. Language 64:539-576. Resnik, Philip. 1992. Left-Corner Parsing and Psychological Plausibility. In The Procedings of the Fiftenth International Conference on Computational Lniguistics, Coling-92, 191-197. Rizi, Luigi. 1982. Violations of the wh island constrain and the subjacency condition. In Isues in italian syntax, ed. Luigi Rizi. Dordrecht: Foris. Rizi, Luigi. 1990. Relativized Minimality: Linguistic Inquiry Monographs; 16. Cambridge, Mas.: MIT Pres. Rochemont, Michael S. 1986. Focus in Generative Gramar. Amsterdam; Philadelphia: J. Benjamins Pub. Co. Rohde, Douglas. 2001-2003. Linger: A Flexible Platform for Language Procesing Experiments: MIT. Rooth, Mats, and Parte, Barbara Hal. 1982. Conjunction, Type Ambiguity and Wide Scope Or. In Procedings of the First West Coast Conference on Formal Linguistics (Wcfl 1), eds. M acken and N Wiegand. Ross, John R. 1967. Constraints on Variables in Syntax, Linguistics, MIT: Ph.D. Ross, John Robert. 1984. Inner Islands. In Procedings of the Tenth Annucal Meting of the Berkeley Linguistic Society, eds. C. Brugmann and M. Macaulay. Berkley, California: Berkely Linguistic Society. Rullmann, Hotze. 1995. Maximality in the Semantics Os Wh-Constructions. Amherst, MA: GLSA (Graduate Linguistic Student Asociation), Department of Linguistics, University of Masachusets. 458 Sag, Ivan A. 1976. Deletion and Logical Form, Ph.D. Disertation, Masachusets Institute of Technology. Saito, Mamoru. 1982. Case Marking in Japanese: A Preliminary Studies, MIT: Generals paper. Saito, Mamoru. 1983. Case and Government in Japanese. Procedings of WCFL 2:247-259. Saito, Mamoru. 1985. Some Asymmetries in Japanese and Their Theoretical Implications, MIT: Ph.D. disertation. Saito, Mamoru. 1989. Scrambling as Semanticaly Vacuous a'-Movement. In Alternative Conceptions of Phrase Structure, eds. Mark Baltin and Anthony S. Kroch, 182-200. Chicago: The University of Chicago Pres. Sakai, Hiromu. 1994. Complex Np Constraint and Case-Conversions in Japanese. In Current Topics in English and Japanese, ed. Masaru Nakamura, 179-203. Tokyo: Hituzi Syobo. Schneider, David Andrew. 1999. Parsing and Incrementality, University of Delaware: Ph.D. disertation. Schwarz, Bernhard. 1999. On the Syntax of Either.Or. Natural Language and Lingusitic Theory 17:339-370. Shibatani, Masayoshi. 1978. Nihongo No Bunseki [the Analysis of Japanese]. Tokyo: Taishuukan. Shimoyama, Junko. 2001. Wh-Constructions in Japanese, University of Masachusets: Ph.D. disertation. 459 Stabler, Edward P. 1994. The Finite Connectivity of Linguistic Structure. In Perspectives on Sentence Procesing, eds. Charles Jr. Clifton, Lyn Frazier and Keith Rayner, 303-336. Hilsdale, New Jersey: Lawrence Earlbuam Asociates, Publishers. Stedman, Mark. 1996. Surface Structure and Interpretation: Linguistic Inquiry Monographs; 30. Cambridge, Mas.: MIT Pres. Stedman, Mark. 2000. The syntactic proces. Cambridge, Masachusets: The MIT Pres. Stepanov, Arthur V. 2001. Cyclic Domains in Syntactic Theory, University of Connecticut: Ph.D. disertation. Sturt, Patrick, and Crocker, Mathew W. 1995. Incrementality and Monotonicity in Syntactic Parsing. Edinburgh Working Papers in Cognitive Science 11:23-66. Sturt, Patrick, and Crocker, Mathew W. 1996. Monotonic Syntactic Procesing: A Cross-Linguistic Study of Atachment and Reanalysis. Language and cognitive proceses 11:449-494. Sturt, Patrick, and Lombardo, Vincenzo. 2005. Procesing coordinate structures: Incrementality and connectednes. Cognitive Science:291-305. Sturt, Patrick. 1997. Syntactic reanalysis in human language procesing, University of Edinburgh: Ph. D. Disertation. Szabolcsi, Anna, and den Dikken, Marcel. 1999. Islands. GLOT International 4-6:3-9. Takahashi, Daiko. 1990. Negative Polarity, Phrase Structure, and the ECP. English linguistics 7:129-146. 460 Takahashi, Daiko. 1993. Movement of Wh-Phrases in Japanese. Natural Language and Linguistic Theory 11:655-678. Takano, Yuji. 2004. Coordination of Verbs and Two Types of Verbal Inflection. Linguistic Inquiry 35:168-171. Takezawa, Koichi. 1987. A Configurational Approach to Case Marking in Japanese, University of Washington. Tanaka, Hidekazu. 2003a. Remarks on Beck's Efect: Linearity in Syntax. Linguistic Inquiry 34:314-323. Tanenhaus, Michael K., Boland, Julie, Garnsey, Susan M., and Carlson, Greg N. 1989. Lexical structure in parsing long-distance dependencies. Journal of Psycholinguistic Research 18:37-50. Tateishi, Koichi. 1994. The Syntax of "Subjects": Studies in Japanese Linguistics; [3]. Stanford, Calif. Tokyo: Center for the Study of Language and Information; Terada, Michiko. 1990. Incorporation and Argument Structure in Japanese. Doctoral. disertation, University of Masachusets. Travis, Lisa. 1989. Parameters of Phrase Structure. In Alternative Conceptions of Phrase Structure, eds. Mark Baltin and Anthony S. Kroch, 263-279. Chicago: The University of Chicago Pres. Uehara, Keiko, and Bradley, Dianne C. 2002. Center-Embedding Problem and the Contribution of Nominative Case Repetition. In Sentence Procesing in East Asian Languages, ed. Mineharu Nakayama, 257-287. Staonford, California: CSLI Publications. 461 Watanabe, Akira. 1992a. Wh-in-Situ, Subjacency and Chain Formation, MIT: MIT Ocasional Papers in Linguistics. Watanabe, Akira. 1992b. Subjacency and S-Structure Movement of Wh-in-Situ. Journal of East Asian Lingusitics 1:255-291. Watanabe, Akira. 1994. A Cross-Linguistic Perspective on Japanese Nominative- Genitive Conversion and Its Implications for Japanese Syntax. In Current Topics in English and Japanese, ed. Masaru Nakamura, 341-369. Tokyo: Hitsuji Shobo. Watanabe, Akira. 1994. A Cross-Linguistic Perspective on Japanese Nominative- Genitive Conversion and Its Implications for Japanese Syntax. In Current Topics in English and Japanese, ed. Masaru Nakamura, 341-369. Tokyo: Hitsuji Shobo. Watanabe, Akira. 1996. Case Absorption and Wh-Agrement: Studies in Natural Language and Linguistic Theory; V. 37. Dordrecht; Boston: Kluwer Academic Publishers. Watanabe, Akira. 1996. Nominative-Genitive Conversion and Agrement in Japanese: A Cross-Linguistic Perspective. Journal of East Asian linguistics 5:373-410. Watanabe, Akira. 2002. Parametrization of Quantificational Determiners and Head- Internal Relatives, University of Tokyo: Ms. Watanabe, Akira. 2003. Wh and Operator Constructions in Japanese. Lingua 113:519-558. 462 Wilder, Chris. 1999. Right Node Raising and the LCA. In The Procedings of West Coast Conference on Formal Linguistics 18. Yamashita, Hiroko, Stowe, Laurie, and Nakayama, Mineharu. 1993. Procesing of Japanese Relative Clause Construction. Paper presented at Japanese Korean Linguistics. Yamashita, Hiroko. 1995. Verb Argument Information Used in a Pro-drop Language: An Experimental Study in Japanese [Sept]. Journal of Psycholinguistic Research 24:333-347.