PROCESSING FACTORS IN LANGUAGE COMPREHENSION AND  PRODUCTION: THE CASE OF CANTONESE DATIVE CONSTRUCTIONS 

CHEUNG Ki Shun Antonio  M. Phil. THESIS 

THE UNIVERSITY OF HONG KONG 

JUNE 2006 

Abstract of thesis entitled Processing factors in language comprehension and production:    the case of Cantonese dative constructions  Submitted by CHEUNG Ki Shun Antonio  for the degree of Master of Philosophy at the University of Hong Kong in June 2006

Chinese linguistics favours pragmatic factors to explain alternatives to the canonical SVO order, such as topicalization and the BA-construction. However, the typological configuration of Cantonese, SVO with prenominal relative clauses (RCs), is predicted to cause processing difficulty for its speakers, especially for the dative construction: [V Theme(T) Recipient(R)] when the theme is modified by a relative clause. A previous study by Cheung (2004, 2005) reported that, as opposed to heavy NP shift (HNPS) in English, Cantonese speakers more frequently use alternative constructions such as the double-bei2 construction [bei2 T bei2 R] or the zoeng1-construction [zoeng1 T bei2 R] (which is equivalent to the Mandarin BA-construction) while the theme NP is more complex. This study employs a dual-task paradigm to compare the comprehension and production of the zoeng1-construction and the canonical double-object dative. The zoeng1-construction is found to be read faster in a masked self-paced reading task, and is imitated more accurately than the canonical construction in the elicited imitation task. To account for such effects, it is suggested that (i) the lexical domain for the verb bei2 is shorter in the zoeng1-construction than the canonical construction, consistent with the principle Minimize Domains (Hawkins 2004), and (ii) the theta roles and thus the argument structure can be assigned early in the zoeng1-construction, consistent with the principle Maximize Online Processing (Hawkins 2004).

Processing factors in language comprehension and production:    the case of Cantonese dative constructions 

by

CHEUNG Ki Shun Antonio  BCogSc, H.K. 

A thesis submitted in partial fulfilment of the requirements for the degree of Master of Philosophy at The University of Hong Kong

June 2006

ਲ਼۰ٙͽд౦Ă଀౦҃Ԟਲ਼ć! Ꮶ۰ٙͽд‫׉‬Ă଀‫҃׉‬ԞᏦć! ֏۰ٙͽдຍĂ଀ຍ҃Ԟ֏Ą! Ӎщ଀͈Ԟ֏̝ˠ҃ᄃ̝֏‫ݙ‬Ċ! Įఄ̄đᗔቔđγ‫ۏ‬ௐ˟Ȉ̱į The fish trap exists because of the fish; once you've gotten the fish, you can forget  the  trap.  The  rabbit  snare  exists  because  of  the  rabbit;  once  you've  gotten  the  rabbit,  you  can  forget  the  snare.  Words  exist  because  of  meaning;  once  you've  gotten  the  meaning,  you  can  forget  the  words.  Where  can  I  find  a  man  who  has  forgotten words so I can have a word with him?  –  Zhuangzi  Ch.  26  External  things 

Declaration 

I declare that this thesis represents my own work, except where due acknowledgement is made, and that it has not been previously included in a thesis, dissertation or report submitted to this University or to any other institution for a degree, diploma or other qualifications.

Signed ______________________________________ Cheung Ki Shun Antonio

i 

 

Acknowledgements  First and foremost, I would like to thank my supervisor, Dr. Stephen Matthews (Steve). Steve is always kind, patient, helpful and resourceful – with such qualities combined, he must be one of the supervisors of everybody’s dreams. I am highly indebted to him for his advice and support on everything related to this thesis and more. Although he has become burdened with administrative work recently, and he no longer can afford private weekly meetings for more than 3 hours (luckily we still managed to leave some time for afternoon tea), he is always supportive and provides me with constructive comments despite being so busy. Dr. Virginia Yip has also given me valuable advice on research, career and more general matters as well, and her guidance will be well remembered. I am also grateful to Dr. Conrad Perry for his constructive comments on methodology and analysis, and for his speedy response with the rate of 50 emails per day between us (maximum). Also, I thank Dr. Eddy Lam for his suggestions on the statistical treatment to my data which has such a large variance (which is consistent with the predictions but unexpected). I would also like to thank Dr. K.K. Luke, Dr. Zhang, Gao Hua and Xiaoling for their helpful comments at the departmental seminar. I would also like to thank the Graduate School of HKU for its support, especially the indispensable Postgraduate Studentship. I am also privileged to get to know many people I have met at the 2005 LSA Linguistic Institute, which I find to be stimulating and enjoyable. I would especially like to thank: Dr. Kay Bock for her insightful suggestions to the experimental methodology and her fun course, Sentence Production: Making Syntax of Sense; Dr. Adele Goldberg for her course, Constructions in Context, which opened my mind about Construction Grammar; and Dr. Ted Gibson for his course Sentence Comprehension and for his useful comments on my experimental paradigm. Without meeting them this thesis would not look like this. Also, I am honoured to be awarded the Chinese Linguistics Fellowship and I would like to thank the Chiang Ching-Kuo Foundation for that. Furthermore, I thank Prof. John Hawkins at RCEAL of Cambridge for his helpful comments, and for his landmark work of the integration of typology and processing, which has influenced the most important work of my undergraduate years – the Final Year Project. His work opened my mind for an alternative to Generative Grammar.

ii 

 

Thanks to the participants in the experiment, although they have produced varied results in comprehension, who volunteered without asking for repay, during the reading week when they were not required to come to school. Life with Cognians is always wonderful, with the friendship supporting my undergrad and postgrad years, especially Louis and Alan; and thanks those I have tutored in Introduction to Linguistics and Introduction to Cognitive Science, who have discuss some related issues with me. We Cognians are deeply indebted to Prof. Laurence Goldstein, as he contributed enormously to the establishment of BCogSc at HKU. Helen must be thanked too, as she is working on topicalization, so her work has influenced my study as well. Thanks for the academic discussions (and also those of entertainment value)! Thanks to Kevin and Alice, who have also been giving comments and asking questions during lab meetings and helped me fine-tune aspects of the thesis. Also, I will never forget the warm atmosphere at the Department of Linguistics. Thanks Rosa and Percy for the refreshments and snacks in the tutors’ office. Sincere thanks to you all =) More thanks go to my family, especially my parents who are always supportive of my studies. They have been tolerating a missing son for weeks, with telephone conversations and emails as the only means of communication. I love you mum and dad =) Moreover, I would like to thank die Mannschaft (the German national football team) for spectacular goals and wins in the 2006 FIFA World Cup, which I have been watching while writing up this thesis. Danke schön! Last but not least, muchas gracias to Elaine my soulmate, without you, my life would never be the same. Aloha nui loa! Thank you again, God, for giving us wonderful and interesting minds and language.

iii 

 

Table of contents  DECLARATION ................................................................................................................................................ I ACKNOWLEDGEMENTS.............................................................................................................................. II TABLE OF CONTENTS ................................................................................................................................ IV LIST OF TABLES AND FIGURES ............................................................................................................... VI ABBREVIATIONS ........................................................................................................................................VII 1.

INTRODUCTION AND OVERVIEW ..................................................................................................1

2.

RELATIONSHIP BETWEEN GRAMMAR AND PROCESSING .....................................................4 2.1.

DISMISSAL OF PROCESSING FACTORS BY CHOMSKYAN LINGUISTICS ..........................................................4

2.2.

LANGUAGE PROCESSING ....................................................................................................................................7

2.2.1.

Early Immediate Constituents (EIC, Hawkins 1994) .................................................................. 8

2.2.2.

Dependency Locality Theory (Gibson 1998, 2000) ....................................................................10

2.2.3.

New metrics proposed by Hawkins (2004) ...................................................................................12

2.3. 3.

4.

THE “CHINESE PROBLEM”............................................................................................................. 26 3.1.

THE TYPOLOGICAL RARITY OF CHINESE ......................................................................................................26

3.2.

THE DIFFICULTY IN PROCESSING CHINESE SENTENCES WITH RCS .........................................................29

3.3.

CANTONESE AND DATIVES .............................................................................................................................30

ALTERNATIVE CONSTRUCTIONS................................................................................................ 33 4.1.

ALTERNATIVES IN AND ACROSS LANGUAGES ..............................................................................................33

4.1.1.

English and Japanese ..............................................................................................................................33

4.1.2.

Cantonese .....................................................................................................................................................34

4.2.

5.

LANGUAGE COMPREHENSION AND PRODUCTION .......................................................................................19

MORE ON ALTERNATIVE CONSTRUCTIONS IN CANTONESE ......................................................................36

4.2.1.

Historical perspective .............................................................................................................................36

4.2.2.

Various accounts of alternative constructions ............................................................................39

4.2.3.

Developmental data and processing factors ................................................................................42

4.2.4.

Different predictions for alternative constructions...................................................................46

4.2.5.

The BA‐construction – the zoeng1‐construction in Cantonese ............................................49

EXPERIMENTAL STUDY ................................................................................................................. 56 5.1.

METHODOLOGY ...............................................................................................................................................56

5.2.

COMPREHENSION RESULTS ...........................................................................................................................62

5.2.1.

Statistical analysis....................................................................................................................................64

5.2.2.

Discussion and qualitative analysis..................................................................................................64

iv 

  5.3. 5.3.1.

Statistical analysis....................................................................................................................................68

5.3.2.

Discussion and qualitative analysis..................................................................................................69

5.4.

FURTHER STATISTICAL INVESTIGATIONS ....................................................................................................71

5.5.

CORRELATION BETWEEN COMPREHENSION AND PRODUCTION ..............................................................76

5.6.

GENERAL DISCUSSION ....................................................................................................................................77

5.6.1. 6.

PRODUCTION RESULTS ...................................................................................................................................67

The canonical double object construction (DOC).......................................................................78

IMPLICATIONS AND CONCLUSION.............................................................................................. 80 6.1.

POSSIBLE FOLLOW‐UPS ..................................................................................................................................80

6.2.

SUMMARY .........................................................................................................................................................82

6.3.

IMPLICATIONS..................................................................................................................................................83

6.4.

FINAL CONCLUSIONS .......................................................................................................................................84

APPENDIX..................................................................................................................................................... 86 SETTINGS OF THE EXPERIMENTAL STUDY .....................................................................................................................86 A.

Settings in file preferences..........................................................................................................................86

B.

Settings in file preferences2 .......................................................................................................................86

C.

Settings in file introduction........................................................................................................................86

D.

Settings in file items.......................................................................................................................................88

BIBLIOGRAPHY........................................................................................................................................... 96

v 

 

List of tables and figures  TABLE 1 RELATIONSHIP BETWEEN VO‐OV AND RELN‐NREL (580 LANGUAGES, DRYER 2005).........................27 TABLE 2 POSSIBLE CONSTRUCTIONS WITH GIVE IN ENGLISH (WASOW 2002)..........................................................33 TABLE 3 DISTRIBUTION OF DITRANSITIVE CONSTRUCTIONS IN DJAMOURI (2001) ..................................................41 TABLE 4: ADULT’S CHILD‐DIRECTED BEI2‐DATIVES IN CANCORP, SUMMARIZED FROM CHAN (2003)..............43 TABLE 5 DISTRIBUTION OF BEI2 DATIVES IN CHILD‐PARENT INTERACTIONS (CHOW 2005)..................................44 TABLE 6 ACCURACY OF ACT‐OUT TASK SUMMARIZED FROM HO (2004) ....................................................................45 TABLE 7 DISTRIBUTION OF RESPONSES OBTAINED IN CHEUNG (2004, 2005).........................................................48 TABLE 8 UNASSIGNMENT DIFFERENCE BETWEEN DOC AND BA WITH AN UNMODIFIED THEME ............................53 TABLE 9 ALTERNATIVE ANALYSIS OF UNASSIGNMENT DIFFERENCE BETWEEN DOC AND BA WITH AN  UNMODIFIED THEME...................................................................................................................................................54

TABLE 10 UNASSIGNMENT DIFFERENCE BETWEEN DOC AND BA WITH RC‐MODIFIED THEME .............................55 TABLE 11 STATISTICS FOR THE DISTRIBUTION OF RTS FOR SENTENCE‐FINAL WORDS .............................................63 TABLE 12 AVERAGE RTS (MS) FOR NON SENTENCE‐FINAL WORDS IN 6 CONDITIONS ..............................................64 TABLE 13 AVERAGE RT (MS) FOR SENTENCE‐FINAL WORDS IN 6 CONDITIONS .........................................................64 TABLE 14 OVERALL ACCURACY (%) OF THE RESPONSES IN THE 6 CONDITIONS ........................................................68 TABLE 15 ACCURACY OF REPETITION (%) IN THE CHOICE OF CONSTRUCTION IN THE 6 CONDITIONS ...................68 TABLE 16 ACCURACY (%) IN REPEATING THE SENTENCE ELEMENTS IN THE 6 CONDITIONS...................................69 TABLE 17 DISTRIBUTION FOR TYPE OF RESPONSES TO THE 6 CONDITIONS ................................................................69 TABLE 18 DISTRIBUTION OF RESPONSES TO DOC INPUT OBTAINED IN CHEUNG (2005) .......................................70 TABLE 19 THEME LENGTHS AND TDS OF THE 6 CONDITIONS .......................................................................................73 TABLE 20 PARTIAL CORRELATIONS BETWEEN LANGUAGE COMPREHENSION AND PRODUCTION METRICS ............77 FIGURE 1 EXAMPLE OF RIGHT AND LEFT BRANCHING LANGUAGES USING AN ENGLISH LEXICON .............................10 FIGURE 2 COMPONENTS IN THE LANGUAGE PRODUCTION SYSTEM (BOCK 1995:185)............................................19 FIGURE 3 INTERACTION MODEL OF DIALOGUE IN A SCHEMATIC REPRESENTATION OF THE STAGES OF  COMPREHENSION AND PRODUCTION PROCESSES (PICKERING & GARROD 2004:176 FIGURE 2)...............21

FIGURE 4 PROCESSING ARCHITECTURE PROPOSED BY JACKENDOFF (2002:199) ....................................................23 FIGURE 5 A SIMULATION OF SCREEN OUTPUT OF A MASKED SENTENCE .......................................................................60 FIGURE 6 AVERAGE RTS PER WORD OVER ALL CONDITIONS OF STIMULUS SENTENCES.............................................62 FIGURE 7 FREQUENCY DISTRIBUTION OF RTS FOR THE SENTENCE‐FINAL WORDS.....................................................63 FIGURE 8 READING TIMES FOR BARE THEME NP CONDITIONS [T] ...............................................................................65 FIGURE 9 READING TIMES FOR ADJECTIVE‐MODIFIED THEME NP CONDITIONS [ADJ T]...........................................66 FIGURE 10 READING TIMES FOR THE HEAVILY‐MODIFIED THEME NP CONDITIONS [RC T].....................................66

vi 

 

Abbreviations  3sg

3rd person singular pronoun

A

Agent (Subject of transitive verb)

Adj

Adjective

AdjP

Adjective Phrase

Adv

Adverb

Asp

Aspect marker

BA

Pretransitive marker/co-verb: ba (把) in Mandarin; zoeng1 (將) in Cantonese

bei2

The Cantonese verb (畀, also written as 俾) “give”

CL

Classifier

Comp

Complement

CRD

Constituent Recognition Domain

Det

Determiner

DOC

Double-object construction

DLT

Dependency Locality Theory

EIC

Early Immediate Constituents

FGD

Filler-Gap Domain

HNPS

Heavy NP Shift

HPD

Head-Pronoun Domain

HPSG

Head-driven Phrase Structure Grammar

IC

Immediate Constituents

L

Language

LD

Lexical Domain

LU

Language Universals

MaOP

Maximize Online Processing

MiD

Minimize Domains

MiF

Minimize Forms

N

Noun

NP

Noun Phrase

NRel

Postnominal relative clause NP[N Rel]

OP

On-line Property

OV

Object-before-verb order

P

Preposition (or postposition for postpositional languages)

PCD

Phrasal Combination Domain (cf. CRD)

vii 

 

PFV

Perfective marker

PP

Preposition Phrase

Pron

Pronoun

Prt

Linking particle: Mandarin de (的), Cantonese ge3 (嘅)

R

Recipient (IO, indirect object)

RC

Relative clause

RD

Right-dislocation [S’ [XP]]

RelN

Prenominal relative clause [Rel N]NP

S

Sentence / Sentential Complement

SFP

Sentence final particle

SOV

Subject-Object-Verb order

SPLT

Syntactic Prediction Locality Theory (cf. DLT)

SVC

Serial Verb Construction

SVO

Subject-Verb-Object order

T

Theme (DO, direct object)

TDD

Total Domain Differential

TMA

Tense, modality and aspect

UG

Universal Grammar

UP

Ultimate Property

V

Verb

VP

Verb Phrase

VO

Verb-before-object order

WM

Working memory

viii 

1 Introduction and overview 

1. Introduction and overview  Among the important issues in the field of linguistics are: ƒ

How are performance and grammar related? Whose approach to grammar is more realistic, the Formalists’ or the Functionalists’?

ƒ

How are language comprehension and production related in language processing? Do they overlap?

ƒ

Should Chinese linguistics consider processing factors in addition to the traditional accounts of alternative constructions?

The significance of these issues is high, as the answers to these issues would be of concern to theoretical linguistics, psycholinguistics and Chinese linguistics. Not only are these issues important, but it appears that they are interconnected, as this study attempts to show. The relationship between these issues would be explained in further detail below. Generative linguistics, following Chomsky, has eschewed performance and processing factors in favour of intuitive judgements. Functionalist accounts have argued the contrary, for example, Performance-Grammar Correspondence Hypothesis (PGCH, Hawkins 2004) suggested that grammar is shaped by performance preferences. The relationship between grammar and processing, comprehension and production, as well as processing theories such as Hawkins (2004) and DLT (Gibson 2000) are discussed in Chapter 2. By studying processing factors affecting Cantonese, this study investigates if there is any experimental support to the PGCH and the processing principles proposed by Hawkins (2004). Chinese languages 1 such as Cantonese provide an ideal testing case for these issues, as they have an unusual configuration which is rare among the languages of the world. The processing of Chinese has begun to receive research attention in the psycholinguistic literature (see Kwan 2000; Matthews & Yeung 2001; Hsiao & Gibson 2003; Cheung 2004, 2005). Combining their findings with the typological observations, it appears that the Chinese type of configuration demands explanation from models of language processing such as Hawkins (1994, 2004) and Gibson (1998, 2000). Chapter 3 discusses the typological problem of Chinese, termed the “Chinese Problem” (Cheung 2005), which arises from centre-embedding. The typological configuration of Cantonese, SVO with prenominal relative clauses (RelN), is predicted to cause processing difficulty for its speakers. The “Chinese Problem” arises in a transitive sentence already, but

                                                         The distinction made here between languages and dialects is based on mutual intelligibility, and is thus of a purely academic distinction and bears no political implication whatsoever. 1

1 

1 Introduction and overview  when there is even more material after the verb, as in the double-object construction (DOC), it becomes more serious when the theme is modified by a relative clause. This introduces an inefficient order when processing ditransitives whose objects are marked with a pre-nominal relative clause according to Minimize Domains (MiD, Hawkins 2004). Processing theories are applied to the configuration of Chinese, especially Cantonese dative constructions. Such inefficiency appears to be evidence against Hawkins (2004), but a possible solution is proposed, following Matthews and Yeung (2001) and Cheung (2004, 2005). While Chinese linguistics favours pragmatic factors to explain alternatives to the canonical SVO order, such as topicalization and the BA-construction, a previous study by Cheung (2004, 2005) reported that, Cantonese 2 speakers more frequently use alternative constructions such as the double-bei2 construction [bei2 T bei2 R] or the zoeng1-construction [zoeng1 T bei2 R] (which is equivalent to the Mandarin BA-construction) while the theme NP is more complex, as opposed to Heavy NP Shift (HNPS) in English. The factors affecting the choice of alternative structures or orderings are discussed in further detail in Chapter 4. Previous work done on sentence processing includes eye-movement studies (cf. Sedivy et al. 1999), corpus-based studies (Wasow 2002) and other reading tasks that study the reading time of sentences. Levelt (1989) has also employed different methodologies in studies of speech production. Such research has largely been done in English, which has modifiers before the noun but the relative clause following the noun, as in the NP in example 1: 1.

a Det

[small] boy [that likes swimming and eating] [Adj] N [RC]

Other studies of Heavy Noun Phrase Shift (HNPS) in English also employed corpus analysis, elicited imitation, elicited production and experimentation (such as reading time studies) using reaction time as a measure (cf. Hawkins 1994; V. Ferreira 1996; Stallings et al. 1998; Smith & Wheeldon 1999; Pearlmutter & Gibson 2001; Jackendoff 2002; Wasow 2002; F. Ferreira 2003; also see a study of “long-before-short” preference in Japanese, Yamashita & Chang 2001). Investigation of the online comprehension of the Cantonese dative is not found in the literature, and little work has been done on the production aspect of the Cantonese dative constructions. Building on the findings in Cheung (2004, 2005) and Hawkins (2004), I shall investigate how processing factors affect both language comprehension and production. To account for the typological rarity of the Chinese type, it was hypothesized that the

                                                         Cantonese is represented in this thesis by Chinese characters and the Romanization used is the LSHK Jyutping Cantonese Romanization Scheme. For more information, see http://www.lshk.org.

2

2 

1 Introduction and overview  BA-construction and the other alternative constructions allow the “Chinese Problem” to be alleviated or even bypassed (Cheung 2004, 2005), following the predictions of the PGCH (Hawkins 2004). In Cantonese the use of zoeng1-construction (the disposal construction) is perhaps like HNPS in English. An experimental study employing a dual-task paradigm is designed in order to compare the comprehension and production of the zoeng1-construction and the canonical double-object construction (DOC) dative. The results are analyzed and discussed in Chapter 5. Finally, Chapter 6 concludes this study by summing up relevant factors and also by discussing theoretical and methodological issues relating to the results of the present study.

3 

2 Relationship between grammar and processing 

2. Relationship between grammar and processing  Views of the relationship between grammar and processing are largely split into two camps. One being formalism and the other being functionalism. One of the main advocates of formalism is Chomskyan linguistics. Chomsky is famous for his contribution to the field of linguistics and cognitive science in general, particularly the rejection of behaviourism. He has contributed much in various debates, such as with Skinner and Piaget, against behaviourism. However, I believe that both behaviourism and Chomskyan linguistics are extreme positions. While Chomskyan linguistics revived mental representations, it has perhaps introduced too many abstract levels of representations, many of which perhaps are without psychological reality (cf. Culicover & Jackendoff 2005). It also abandoned ‘careful attention to the sources and quality of the data used to support theoretical claims’ (Wasow 2002:xi) and rigorous statistical testing of data.

2.1. Dismissal of processing factors by Chomskyan linguistics  In general, generative grammar within formalism considers that the grammatical competence and performance are very different things. Competence, the knowledge of grammar, is innate and internal (Universal Grammar, UG), and the ideal parser is independent of processing limitations (Chomsky 1965). Performance, on the other hand, is external as it may be affected by external factors. On this basis, generative grammar assumes that performance factors have no impact on grammar, as commented by Hawkins (2004). Chomsky (2000:117) distinguished two different components of the language faculty – a performance system and a cognitive system, citing dissociations of the two systems. Wasow (2002) also referred to Chomsky’s assumption in 1965 that performance considerations could be safely ignored, as ‘investigation of performance will proceed only so far as understanding of competence permits’ (Chomsky 1965:10). The distinction between competence and performance might have been important to Chomsky as behaviourism was prevalent. Furthermore, Wasow (2002) summarized that in generative grammar, performance factors were used as ‘a means of discounting lots of [quantitative] data’ (p.140) and that ‘performance models are considered to be informed by competence theories but not vice versa’ (p.148). On the other side, functionalists disagree. Haspelmath (2002) argued for the attractiveness of functional explanations: ƒ ƒ ƒ

‘they are more general, because they link linguistic facts to non-linguistic facts.’ ‘they are cognitively more plausible than a highly specified innate UG with thousands of OT constraints’ ‘functional explanations are testable, unlike the innateness claims’

4 

2 Relationship between grammar and processing  ƒ

functional explanations are ‘compatible with generative analyses’, but ‘often generative analyses are made redundant by functional explanations, so ignoring functional explanations is irresponsible.’ (Haspelmath 2002:6)

Haspelmath (2002) summarized the characteristics of functionalism as follows: ƒ ƒ ƒ ƒ ƒ

‘Functionalists claim that the explanation of language form involves appreciating the regularities of language function.’ ‘Functionalists try to explain competence, often on the basis of performance, because they think that competence can be affected by performance.’ ‘Everyone recognizes that grammars are full of arbitrariness, but functionalists emphasize non-arbitrariness and attempt explanation wherever possible.’ ‘Functionalists only reject the idea that syntax should be studied as if functional explanation were impossible (just as formal linguists reject the idea that syntax should be studied as if systematic arbitrary generalizations were impossible).’ ‘Functionalists don’t think of description as explanatory, hence they do not emphasize the properties of the descriptive tools they work with.’ (Haspelmath 2002:2-3)

As pointed out by Haspelmath (2002), languages are clustered around certain highly probable types. There are many structures that can be acquired but that are rarely attested in actual languages, probably because they are not useful and thus are likely to be eliminated in language change. Cognitively impossible languages are, presumably, unattested (cf. Newmeyer 2005). Although speakers do not intend to create well-designed grammars, there is an ‘invisible hand’ process in language change (Haspelmath 1999), suggesting that functional forces like iconicity, efficiency and clarity affect grammar. The invisible hand process is argued by Croft (2000) to arise from language use in adults, because speech produced by learners and children is not prestigious enough to spread to other speakers. Speakers are, although not fully conscious of their language at all levels, intentional, as they want to be understood, so both efficiency and clarity affect their choices over a vast array of possible utterances (Haspelmath 2002). Haspelmath (2002) also explained that functional forces, like structure-concept iconicity and parsing, are derivative of economy and clarity in this sense. Efficient parsing (including structure-concept iconicity) is good for the hearer, but to some extent also for the speaker. Iconicity of the sequence (“the first things get the first words”) is good for the speaker, but also for the hearer. Perhaps Chomskyan linguistics has also gone too far and has become a pure intellectual game. According to Wasow (2002:160), ‘[i]n practice, … [Chomsky’s] own work and that of almost all other generative grammarians makes virtually no reference to any type of evidence other than well-formedness judgments’. Wasow (2002:160-1) also criticized Chomsky’s attempt to reject the use of carefully controlled methods such as laboratory experiments as a source of insight into mental representations, and the exclusive reliance of generative grammar on introspective judgments. He argued that this view is largely mistaken, since

5 

2 Relationship between grammar and processing  grammaticality judgments are indeed a form of language performance, and he has observed that generative grammarians themselves have frequently produced what they consider ungrammatical in their publications. He also suggested that acceptability judgments are not absolute, as judgments vary between speakers. This implies that grammaticality is perhaps a matter of degree, and combined with experimental results in psycholinguistics, it appears that grammar should also be affected by processing factors. Example 2 is a doubly-embedded Mandarin relative clause used in a reading time study by Hsiao and Gibson (2003:8), 2.

[[邀邀 [勾勾 法法] 的 富富] 的 法官] 心心心心 [yaoching [gojie faguan] de fuhao] de guanyuan shinhuaibugui invite conspire judge Prt tycoon Prt official have-bad-intentions “The official who invited the tycoon who conspired with the judge has bad intentions” (Hsiao & Gibson 2003:8 Example 4b)

Although the sentence structure is grammatical, some participants fail to understand the sentence at some point and judge it as ungrammatical. The debate between the two schools is not over. It is especially notable that even in a debate started by Newmeyer’s (2003) article in the journal Language, “Grammar is grammar and usage is usage”, he acknowledges that ‘functional linguists … have provided (to my mind) incontrovertible evidence that grammars are shaped in part by performance considerations’ (p.683). More recent syntactic theories such as HPSG have also been designed to be compatible with performance models (Sag et al. 2003), as the belief that empirical facts about various aspects of performance can and should inform the theory of linguistic competence is gaining ground. Designed to be compatible with performance, especially the fact that processing is rapid and incremental, HPSG is surface oriented and posits no abstract structures (Sag et al. 2003). HPSG is constraint-based and strongly lexicalist, and localized most grammatical and semantic information within lexical entries, and as a result there are no transformation rules. Sag et al. (2003) also attacked traditional transformational grammar on the grounds of bias as transformational derivations have directionality built into them in the form of transformational operations, leading to the obscurity of the relationship between grammar and its role in processing. Since production involves going from a meaning to an utterance and comprehension involves going from an utterance to a meaning, a grammar that is used in both processes should not favour one order over the other. Sag et al. (2003) proposed that this is reflected by a predictably close relationship between production and comprehension of a natural language by any language user. As what can be produced by language users is similar to what a language user can comprehend, Sag et al. (2003) argued that this simple observation suggests rather that the differences in performance should be explained by a theory that posits distinct processing regimes making use of a single language

6 

2 Relationship between grammar and processing  description, instead of opposite views, such as that the production grammar is independent from the comprehension grammar. On the other hand, language typology studies language structures based inductively on investigation of a wide range of languages. This approach to linguistics studies language features and investigates if languages can vary without limit. Typologists would like to know if there are regularities (universals) across languages and whether these regularities correlate with each other. For example, Greenberg (1966) has proposed a list of language universals, and Keenan and Comrie (1977) have proposed a Noun Phrase Accessibility Hierarchy (NPAH). Frequently, typologists are functionalists, and develop explanations for such patterns, basing on the reasoning “If variation patterns exist in the world’s 6000 languages, there might be a reason for it”. Hawkins (1990, 1994, 2004) has suggested that the relationship between grammar and performance can be easily summarized: grammars conventionalize performance preferences. The definition of the Performance-Grammar Correspondence Hypothesis (PGCH, Hawkins 2004) is given below: Performance‐Grammar Correspondence Hypothesis (PGCH)  Grammars have conventionalized syntactic structures in proportion to their degree of preference in performance, as evidenced by patterns of selection in corpora and by ease of processing in psycholinguistic experiments. (Hawkins 2004:3) The support for the PGCH comes from a variety of phenomena, such as alternative relative clauses with and without an explicit relativizer as in English [the Danes (whom/that/ ø) he taught], heavy noun phrase (NP) shift (HNPS), the resumptive pronoun strategy as in a variety of languages such as Chinese and Hebrew, and that centre embedding is avoided in most languages of the world. This is based cross-linguistically on observed patterns in the corpora, which include daily conversations and written materials, and also results from observing ease of processing in psycholinguistic experiments.

2.2. Language processing  With Greenberg’s (1966) universals and observations in mind, Hawkins (1990, 1994) proposed his own theory to account for language preferences in word orders that used a more functional approach. PGCH also seems to be a plausible solution to the question Hawkins asked in 1994: ‘we need to ask how it is possible for language users to successfully produce or comprehend a given structure in real time’ (Hawkins 1994:16). On the other hand, Gibson proposed Dependency Locality Theory (DLT, Gibson 2000; also see Syntactic Prediction

7 

2 Relationship between grammar and processing  Locality Theory, SPLT, Gibson 1998), basing his observations in psycholinguistics in attempt to understand the human parser, studying why certain structures are difficult to process, including ambiguous structures such as garden-path sentences and unambiguous centre-embedded structures. New principles in language processing proposed by Hawkins (2004) will also be discussed, followed by a discussion in section 2.3 on the relationship between comprehension and production in language processing.

2.2.1. Early Immediate Constituents (EIC, Hawkins 1994)  Hawkins (1994) proposed the principle of Early Immediate Constituents 3 (EIC, later subsumed by Minimize Domains, MiD, in Hawkins 2004). He suggests that the strategy used by the parser is to maximize the proximity (hence “early” and “immediate”) of the heads of all constituents to the verb (Hawkins 1994). The principle is given below, and subsequently an illustration is provided (see example 3). Constituent Recognition Domain (CRD)  The CRD for a phrasal mother node M consists of the set of terminal and non-terminal nodes that must be parsed in order to recognize M and all ICs of M, proceeding from the terminal node that constructs the last IC on the right, and including all intervening terminal nodes and the non-terminal nodes that they construct. (Hawkins 1994:58) 3.

I [give [the new book] [to him]]

In the example, the VP is give the new book to him, and upon give, the VP node is constructed. Then, book and to would be the heads of the Immediate Constituents (IC) (NP the new book, and PP to him) of the give-VP, and upon encountering these two words, the recognition of the VP is complete. In other words, it is necessary for the parser to encounter to in order to assign the thematic/grammatical properties to the NP and PP of the VP, and the CRD of this VP starts at the word give and ends at to, leading to a CRD of 5 words. Calculating IC‐to‐non‐IC ratios  The IC-to-non-IC ratio for a CRD is calculated by dividing the number of ICs in the domain by the total number of non-ICs (or words alone) in that domain, expressing the result as a percentage. The ratio for a whole sentence is the aggregate of the scores for all the CRDs within the sentence. (Hawkins 1994:76-7) Using example 3 again, the IC-to-word ratio of the VP would be 3 ICs (give as V, NP and PP) to 5 words, i.e. 3/5 = 60%. The ratio for the whole sentence would be the average of the

                                                         3

To illustrate, P and NP would be the I(mmediate) C(onstituents) of a PP.

8 

2 Relationship between grammar and processing  scores for the CRDs within the sentences, that is, the average of the ratios for the CRDs of the VP, NP and PP. Early Immediate Constituents (EIC) (Expanded)  The human parser prefers linear orders that maximize the IC-to-non-IC ratios of constituent recognition domains. Orders with the most optimal ratios will be preferred over their non-optimal counterparts in the unmarked case; orders with non-optimal ratios will be more or equally preferred in direct proportion to the magnitude of their ratios. For finer discriminations, IC-to-non-IC ratios can be measured left-to-right. (Hawkins 1994:78-9) As Wasow (2002) calculated that the number of words and the number of nodes are highly correlated (with a coefficient of 0.94), in effect, IC-to-word ratios can be applied. Hawkins (1994) applied EIC to explain how HNPS and extraposition are motivated by parsing efficiency, as exemplified by examples 4 to 7 below: 4.

5.

Canonical prepositional dative [V T to R] I [give [the book that I have borrowed] [to him]] The CRD of the VP of give extends from give to to (the head P of the PP), IC-to-word ratio is 3 ICs (V NP PP) to 8 words, and the efficiency is calculated by 3/8 = 37.5% HNPS Example [V to R T] I [give [to him] [the book that I have borrowed]] The CRD of the VP of give extends from give to book (the head N of the NP), IC-to-word ratio is 3 ICs (V PP NP) over 5 words, and the efficiency is calculated by 3/5 = 60%

The canonical dative is less preferred than the HNPS example, as the HNPS example has a higher efficiency of 60% as opposed to 37.5%. Hawkins (1994) argued that even the difference of 1% would have an effect on the choice of construction. Extraposition of sentential subjects offers another set of examples: 6. 7.

Sentential subject: [[That Bill is frightened] [surprised [Mary]]] Extraposed: [It [surprised [Mary] [that Bill was frightened]]]

The VP domains of both 6 and 7 are 100% efficient, as [surprised Mary] in 6 has 2/2 IC-to-word ratio, and [surprised Mary that] in 7 has also a 100% (3/3) IC-to-word ratio. However, for the S domain, the dummy subject it in 7 allows efficient recognition of the S domain, as It surprised has 2/2 IC-to-word ratio, leading to an aggregate of 100% for 7, but the sentential subject version (That to surprised) is penalized by a low IC-to-word ratio of 2/5 = 40% efficiency, leading to a lower efficiency of 70% (= (2/2 + 2/5) / 2). Production research on HNPS (such as Stallings et al. 1998; Arnold et al. 2000; Wasow 2002) has argued that HNPS could also be motivated by difficulty in utterance planning and production. According to typology, left and right branching languages can be recognized as predominantly head-initial or head-final, as in Figure 1:

9 

2 Relationship between grammar and processing 

VP                                           VP      /       \                                    /          \    V             PP                              PP           V        /        \                    /      \       P       NP                 NP          P    /____\            /____\              go to French movies              French movies to go  Figure 1 Example of right and left branching languages using an English lexicon 

According to Hawkins (1994), further support is supplied by left-branching languages (for example, SOV languages such as Japanese and Korean) that exhibit a tendency to put heavy constituents at the beginning of sentences (Yamashita & Chang 2001; Yamashita 2002) rather than at their ends as in English. He argued that this could be predicted by the parser-based theory of weight effects, and that the intermediate structures including [go [French movies] to] are dispreferred due to processing reasons. In particular, he suggested that for different types of languages, different kinds of parsing are used. In head-initial languages, immediate constituents (IC) are attached backwards during top-down parsing, whereas for head-final languages, the first IC often does not provide adequate information to identify the mother node, and thus they are attached forward. Head-final languages require a look-ahead strategy, forcing the parser to recognize CRD only at the right periphery of the domain. Bottom-up parsing is thus used. EIC can be applied to Japanese according to experimental studies (Yamashita & Chang 2001) and corpus analyses (Yamashita 2002) – long phrases tend to be shifted ahead of short ones, due to the structure of Japanese being left-branching (see Hawkins 1994:137-54). In languages that permit phrases to appear pre- and post-verbally, conceptual and form-related factors should differentially influence these regions (cf. Yamashita 2002:613-9).

2.2.2. Dependency Locality Theory (Gibson 1998, 2000)  Another parsing theory, Syntactic Prediction Locality Theory (SPLT, Gibson 1998) is proposed to account for syntactic complexity in parsing, focusing especially on the activation of related syntactic items in the working memory (WM). It is later superceded by Dependency Locality Theory (DLT, Gibson 2000), but in general it is similar to EIC with respect to working memory constraints on the human parser. DLT also supercedes two parsing heuristics in ambiguity resolution proposed by Frazier (1987): Minimal Attachment and Late Closure. According to DLT, the ease of processing is defined by: ƒ

processing of words;

10 

2 Relationship between grammar and processing  ƒ

integration of words with existing structures;

ƒ

maintaining incomplete structures in memory.

Its focus is on the integration of words with existing structures and the costs associated with maintaining incomplete structures in the human WM. DLT put forward two kinds of costs which tax the available pool of resources in WM: a) Integration of items into the structure built thus far; b) Storage / Expectations: Predicting categories to complete the current structure In most cases, the Dependency Locality Theory is similar to the EIC, but it puts more focus on intervening materials, especial new discourse referents which have to be assigned dependency relations. This is based mostly on observations in psycholinguistics. Information status may have a role to play in processing. For example, for a centre-embedded (nested) structure, a first or second person pronoun is easier than a third-person: 8. 9.

The boy [who the girl [that I know] lives next to] is tall. The boy [who the girl [that Peter knows] lives next to] is tall.

The first person pronoun I is nested in example 8, while in example 9 it is a third-person proper name. Experimental results (Warren & Gibson 2002) show that 8 is easier than 9, supporting DLT. Furthermore, the Dependency Locality Theory demonstrated how referents interfere with parsing by providing support with examples such as relative clauses and garden-path sentences, which causes a temporally misjudgement of grammaticality. The following is a famous example of a garden path sentence: 10. The horse raced past the barn fell. In example 10, it is difficult to integrate the main verb fell with its argument horse, as ƒ

the parser has since encountered intervening material and the activation of horse have decreased;

ƒ

there is interference of similar items (such as barn) in-between, making horse hard to choose.

Recall the Chinese relative clause example 2 (Hsiao & Gibson 2003:8), which is repeated below for convenience: [勾結 法官] 的 富豪] 的 官員] 心懷不軌 11. [[邀請 [yaoching [gojie faguan] de fuhao] de guanyuan shinhuaibugui invite conspire judge Prt tycoon Prt official have-bad-intentions “The official who invited the tycoon who conspired with the judge has bad intentions” (Hsiao & Gibson 2003:8 Example 4b)

11 

2 Relationship between grammar and processing  Example 11 is shown to be very difficult to process. When the parser reaches the nested dependencies, the increasing number of intervening objects takes up cognitive resources, hindering parsing efficiency (as exemplified in the slow reading time near 1000ms/word, Hsiao & Gibson 2003).

2.2.3. New metrics proposed by Hawkins (2004)  Both EIC and DLT have cross-linguistic and psycholinguistic support, for example, in NP fronting in Japanese (Yamashita & Chang 2001; Yamashita 2002); a large range of languages of different typological types including Finnish, Polish, Turkish, Hungarian and Korean (Hawkins 1994); heavy NP shift in English; reading time and eye-tracking studies. Hawkins (2004) has proposed 3 principles to account for processing efficiency and complexity, with some concern paid to new findings as well as what EIC could not predict. They are: ƒ

Minimize Domains (MiD);

ƒ

Minimize Forms (MiF); and

ƒ

Maximize Online-Processing (MaOP)

Minimize Domains (MiD)  MiD is basically the same as EIC, but with new names for the terms. The definitions for CRD and EIC are repeated here for convenience: Constituent Recognition Domain (CRD)  The CRD for a phrasal mother node M consists of the smallest set of terminal and non-terminal nodes that must be parsed in order to recognize M and all I(mmediate) C(onstituents) of M. Early Immediate Constituents (EIC)  The human parser prefers linear orders that minimize CRDs (by maximizing their IC-to-nonIC [or IC-to-word] ratios), in proportion to the minimization difference between competing orders. (Hawkins 2004:32) Later in Hawkins (2004), as he generalized the preferences of MiD over both comprehension and production, CRD is relabelled as Phrasal Combination Domain (PCD), and EIC also updated 4 . This is because Hawkins (2004) attempts to extend his processing theories to both language comprehension and production.

                                                         Combination occurs when two items are in the same mother phrase, such as the verb and the particle in phrasal verbs such as put out. Dependency occurs when an item requires another for its assignment of syntactic or semantic properties with respect to the other, such as pronouns and their referents (such as Peter washed himself, where himself is dependent on Peter).

4

12 

2 Relationship between grammar and processing  Phrasal Combination Domain (PCD)  The PCD for a mother node M and its I(mmediate) C(onstituent)s consists of the smallest string of terminal elements (plus all M-dominated non-terminals over the terminals) on the basis of which the processor can construct M and its ICs. Early Immediate Constituents (EIC)  The human processor prefers linear orders that minimize PCDs (by maximizing their IC-to-nonIC [or IC-to-word] ratios), in proportion to the minimization difference between competing orders. (Hawkins 2004:107) And thus MiD is defined as follows: Minimize Domains (MiD)  The human processor prefers to minimize the connected sequences of linguistic forms and their conventionally associated syntactic and semantic properties in which relations of combination and/or dependency are processed. The degree of this preference is proportional to the number of relations whose domains can be minimized in competing sequences or structures, and to the extent of the minimization difference in each domain. (Hawkins 2004:31) Hawkins (2004) noted that small PCDs reduce simultaneous processing in WM, and it is intended that this holds across both comprehension and production. He also holds that EIC predictions still hold the same across different preferred orders in languages and typological preferences. To account for the existence of multiple domains and preferences for the processing of syntactic and semantic properties, Hawkins (2004) proposed a metric that can assess their combined effect within a given structure. Combining the different domains (such as the phrasal combination domains and the lexical domains) in the same metric, Total Domain Differentials is proposed. Total Domain Differentials (TDD)  The TDD is the collective minimization difference between two competing sequences measured in words and calculated on the basis of the phrasal combination domains, lexical domains, or other domains required for the processing of syntactic or semantic relations within these sequences. TDD Performance Prediction  Sequences with the highest collective minimization differences will be those that are preferably selected, in proportion to their relative TDDs. (Hawkins 2004:120) Following Wasow (2002), Hawkins assumed that the number of nodes (or only phrasal nodes) could well be used. He also argues that the relative strengths of the different domains may be reversed. To allow for such flexibility, the relative strengths of the different domains are not formulated in TDD, and a simple aggregate of domain lengths (TD) is used in the study.

13 

2 Relationship between grammar and processing  Minimize Forms (MiF)  MiF and MiD are both based on structural “economy”, an idea which has many precedents. The focus of MiF is on the economy of form, as in certain constructions and portmanteau morphemes. MiF is defined as follows: Minimize Forms (MiF)  The human processor prefers to minimize the formal complexity of each linguistic form F (its phoneme, morpheme, word, or phrasal units) and the number of forms with unique conventionalized property assignments, thereby assigning more properties to fewer forms. These minimizations apply in proportion to the ease with which a given property P can be assigned in processing to a given F. (Hawkins 2004:38). As quoted, MiF suggests a tendency to simplify all kinds of linguistic units, and the more property assignments are paired to a linguistic form, the better. More efficient structures would be produced more frequently, resulting in higher predictability, leading to reduction in form (Haspelmath 2002) 5 . Hawkins (2004) cites the following as support for MiF: ƒ

the use of pronouns instead of full NPs;

ƒ

compounding;

ƒ

coordinate deletions (perhaps also sluicing);

ƒ

resumptive pronouns;

ƒ

auxiliary verbs reduced and contracted (such as shortened auxiliary verbs in contact languages such as pidgins and creoles, e.g. the anterior and progressive markers ti and pe in Mauritian creole may be derived from dialectal French était and après)

MiF can perhaps also be illustrated by Zipf’s Law – words of higher frequency are shorter in length (Zipf 1949). By mapping the more efficient forms to the more frequent materials, processing is easier. MiF correctly predicts the occurrence of portmanteau morphemes or words, such as –s in English, Spanish verb suffixes, French aux, English words alphabet (alpha and beta among other letters) bit (binary digit), pixel (picture element), smog (smoke and fog), etc. Hawkins (2004) noted that there is an interesting interaction between MiF and MiD. With no intervening materials, relativizer-omission, resulting in reduced RCs, is more frequent than

                                                         The predictions made by MiF are also compatible with the claim made by Construction Grammar (CxG) that a construction can bear meaning on its own, which can lead to certain constructions being used to express the same meaning but in a more efficient way. For example, a verb with two objects will form a construction with a prototypical meaning of give, as in example 1: 1. I’ll ICQ / MSN you the file. (“I’ll give you the file through ICQ/MSN”) ICQ and MSN are both newly-coined verbs deriving from (relatively new) computer programs which allow file transfer, and the double object construction (DOC) is used to express the dative meaning instead of explicitly saying give. The high productivity of constructions also lends support to CxG (see also Goldberg 1995). 

5

14 

2 Relationship between grammar and processing  cases where intervening materials are present, suggesting MiF is in effect, but is overridden by MiD when intervening material is present. Maximize Online‐Processing (MaOP)  The last principle proposed by Hawkins (2004) differs from the other two. MiD and MiF are motivated by economy “Express the most with the least”, while MaOP is captured by “Express it earliest” (Hawkins 2004:25). It is also noted that these three principles often overlap, but may compete at times. MaOP is defined as follows: Maximize Online‐Processing (MaOP)  The human processor prefers to maximize the set of properties that are assignable to each item X as X is processed, thereby increasing O(n-line) P(roperty) to U(ltimate) P(roperty) ratios. The maximization difference between competing orders and structures will be a function of the number of properties that are unassigned or misassigned to X in a structure/sequence S, compared with the number in an alternative. (Hawkins 2004:51) The longer the delay in the assignment of properties in the online parse string, the more time the parse string will take to finish, and the more resources it would require for the parse string to be held in the memory. The earlier the ultimate properties are assigned, the better, as the parser can build its ultimate representations sooner. Following these predictions, in short, it is predicted that the distribution should be skewed towards early property assignments. Unassignment factors  a) the number of words and phrases that undergo some temporary unassignment of properties on-line, compared with an alternative structure/sequence in which the relevant properties are immediately assignable; b) the number of any mother-daughter attachments that are temporarily unassignable to the words and phrases in (a); c) the number of any relations of combination or dependency that are temporarily unassignable to the words and phrases in (a). (Hawkins 2004:52) The following examples 12 to 15 demonstrate unassignment factors in MaOP and how MiF interacts with MaOP: 12. 13. 14. 15.

I believe [the boy is sick]. I believe that [the boy is sick]. I believe [the boy knows the answer]. I believe that [the boy knows the answer]. (Examples 14 and 15 from example 3.19, Hawkins 2004:52)

Relativizer-omission due to MiF can be seen in 12 and 14. MaOP can explain temporary ambiguity as in DLT, as it is similar to storage costs in DLT, although they are not exactly the same. While 12 and 13 are relatively easier to process than 14 and 15, 12 and 14 are harder than 13 and 15 because the boy could also be assigned as the patient of believe, which

15 

2 Relationship between grammar and processing  may lead to a garden path effect at is or knows. In 13 and 15, the boy is correctly assigned as the agent of the embedded clause since the parser expects it, thus 13 and 15 are easier to process even though the length of the sentence is longer, as predicted by MaOP. Misassignment factors  a) the number of words and phrases that undergo some temporary misassignment of properties on-line; b) the number of any additional dominating nodes that must be introduced into the syntactic tree when correcting the misassignments in (a); c) the number of any mother-daughter attachments that are temporarily misassigned to the words and phrases in (a); d) the number of any relations of combination or dependency that are temporarily misassigned to the words and phrases in (a); e) the number of any mother-daughter attachments that replace those misassigned in (c); f) the number of relations of combination or dependency that replace those misassigned in (d). (Hawkins 2004:53-4)

  Hawkins further explained that when these factors are applied to examples 14 and 10, I believe the boy knows the answer and The horse raced past the barn fell, example 10 is a more severe garden path. More misassignment are made to words and phrases on-line (a) in this typical garden path sentence, more additional nodes must subsequently be introduced (b), more misassigned relations of attachment (c), combination and dependency (d), and more relations are made to replace those misassigned (e and f). As Hawkins (2004) explains it, The horse raced past the barn is ‘parsed as a simple main clause on-line, and are then reanalyzed as a head plus reduced relative clause (i.e. the horse that was raced past the barn) once the predicate fell is encountered’ (p. 53). A metric similar to the earlier IC-to-word ratio is proposed: On‐line property to ultimate property ratio (OP‐to‐UP ratio)  The OP-to-UP ratio is calculated at each word X within a connected set of words <…X…> whose property assignments differ between two competing structures S and S’. The cumulative number of properties assignable at each X is divided by the total number of properties to be assigned in the connected set in the ultimate representation of each structure, and the result is expressed as a percentage (e.g. 4/20 = 20%, 8/20 = 40%, 10/20 = 50% for successive words, etc.). The higher these on-line percentages, the more efficient is structure S or S’, since more properties are assigned earlier. An Unassignment Difference between the two structures S and S’ can then be quantified by summing up OP-to-UP ratios for all words being compared and assessing whether S or S’ has the higher overall score. (Hawkins 2004:55-6)

16 

2 Relationship between grammar and processing  In a description of HPSG, Sag et al. (2003) commented about garden path sentences as a disruption to processing 6 : If the processor makes a decision about how to resolve a local ambiguity, but information later in the sentence shows that the decision was the wrong one, we would expect processing to be disrupted, as in garden path sentences. (Sag et al. 2003:296) According to HPSG, an important aspect is that language processing rapidly draws on all available types of linguistic and non-linguistic information as such information is needed: listeners use whatever information is available to them, as soon as it becomes available to them, to infer the speaker's intentions. Ambiguity resolution is done using any information available, such as plausibility assessment. Context and relevant non-linguistic information can eliminate the garden path effect. Competing interpretations exist in parallel, but are active to varying degrees. A particular alternative interpretation is active to the extent that evidence is available to support it as the correct interpretation of the utterance being processed, supported by evidence that implausible interpretations are not conscious. This suggested the importance of online processing factors to grammar. Similarly, it was contended by Hawkins that although multiple constraints might lead to complications, they account for the data in a better manner. DLT and EIC are theories that account for complexity/preferences solely using a locality/weight account, and they are bound to make certain wrong predictions. For example, Hawkins (2004) pointed out that ergative languages pose a challenge to DLT (and also rather likely to EIC). For nominative-accusative languages (e.g. German, Finnish): subject-before-object is less costly than object-before-subject. This order is also preferred, as predicted by a dependency-based account like DLT. It is predicted that for ergative-absolutive languages like Avar and Hindi, subject-before-object ordering is more costly. On encountering an agent of a transitive verb, the ergative case marker leads to an expectation of an up-coming object, and the opening and maintenance of a new node taxes our WM. However, subject-before-object ordering is still more frequent in ergative languages, suggesting that the advantage of correct assignment of agent and patient is important. For languages with object-before-subject ordering, ergative languages are prevalent while accusative languages are less common. This is argued by Hawkins (2004: 233 fn 4) to be further support for MaOP 7 . The correct predictions made by

                                                         6 Processing

plays a major role in the design of HPSG. Combined with the lexicalist approach, Construction Grammar can be implemented in HPSG. Sag et al. (2003) also advanced the view that linguistic and non-linguistic constraints on the interpretation are interleaved in real time.  7 An argument for all ergative languages to be object-before-subject does not work, since in fact the preference for subject-before-object ordering is stronger: The assignment for the marked lower cases, that is, the ergative (or accusative) case, can be recognized as it is encountered, without requiring the access to higher cases. In this manner, an object-before-subject ergative language would also be efficiently processed in terms of MaOP.

17 

2 Relationship between grammar and processing  MaOP point towards a multi-factor model where MaOP is also considered, and may be at times more important than working memory demands. This has rather important implications, namely that sometimes the human processor would prefer structures that is associated with a higher WM load, as in Hawkins (2004): [D]omain minimization applies in the normal case, unless it is in competition with independently motivated processing preferences…. This means that constituents and operations that add to the processing load of filler-gap dependencies will always be minimized, whereas those that facilitate processing may not be. (Hawkins 2004: 211) MaOP is also intended to apply in a similar manner to topic-before-predication, including topicalization in Chinese and long-before-short ordering in head-final languages (see Japanese scrambling discussed earlier, Yamashita & Chang 2001). Such languages are often termed as topic-prominent languages. Hawkins (2004) argued that the primary reason for topic-prominent languages to have productive topics, which are not always given 8 , is that there are multiple dependency relations holding between topic and predication, and it is the predicate that depends on the topic. Topic-before-predication, as a result, enjoys a better OP-to-UP ratio. This should hold for all languages, but for RelN languages, the head noun would be closer to the predicate, thus favouring MiD. Although MaOP appears to be biased towards comprehension, production would also benefit. V. Ferreira (1996) found that verbs that have higher syntactic flexibility are more easily produced, supporting an incremental model of sentence production. The consequence for an incremental model is that syntactic processing cannot occur on later phrases before earlier phrases. This is good news for MaOP: the earlier the properties are assigned to the linguistic material to be uttered, the earlier the latter phrases can be prepared for speech. In an alternative analysis, topicalized phrases tend to be information rich, which is likely to be accompanied by extra information such as RCs. Topics are thus more salient and their activations may be higher, possibly due to the different ways it may be expressed. Topic phrases are also accompanied by a topic marker -wa in Japanese, an optional topic particle in Cantonese, or in the case of English and Cantonese, a prosodic pause. With the help of these cues, ultimate properties can be assigned earlier and easily to the topic phrase, while in the canonical order there might be a higher rate of unassignment or misassignment.

                                                         8

It is also noted that given-before-new is not universal (Hawkins 1994: 224-243).

18 

2 Relationship between grammar and processing 

2.3. Language comprehension and production  Although processing factors are gaining significance in the study of grammar, views on language processing are still divided. Comprehension and production have long been studied separately due to the two groups of researchers looking at different things which have not really converged at all, and also due to the different nature of the tasks – the goal of comprehension is to reconstruct the meaning from a string of words, while the goal of production is to produce a sequence of words from a concept. It should be borne in mind that reverse engineering, such as language comprehension, which involves reconstructing a structure from serial input, is a difficult process, yet evolution has allowed us to perform these tasks with limited effort. It is even possible (for those who believe in modules) to consider that the two tasks employ two different grammars and are in two separate modules, as the different nature of the tasks may suggest very different algorithms. However, comprehension and production studies have been observing different aspects of language. For example, comprehension research has mostly been studying reaction times, while production studies pay more attention to the products, such as speech error data patterns (e.g. Levelt 1989). Levelt (1989) considered production to be necessarily linear and uni-dimensional. Following this idea, Bock (1995) proposed the following model, which separates language production into three broad components: a message component, a grammatical component, and a phonological component (see Figure 2).

Figure 2 Components in the language production system (Bock 1995:185) 

19 

2 Relationship between grammar and processing  She noted that views of the relationship between both tasks range from closely identified (‘production being simply “comprehension turned upside down”’, Bock 1995:205) to substantial dissociation between them. Also presented is some scepticism about the close association of the two tasks: different tasks are under different constraints, and production needs to satisfy certain syntactic constraints where such demands are possibly lower in comprehension, but on the other hand, the input to production is clear while the input to comprehension is degraded and often ambiguous. However, it was also recognized that linguistic resources employed in comprehension cannot be entirely independent of those for production – communication can only be successful when words and their arrangements mean the same thing regardless of the task, this implies that both tasks have to be accounted for by performance theories for general descriptions of the language knowledge called upon in both tasks, and the role of performance data in theories of linguistic knowledge. Bock (1995) argued that processing theories, such as the early version of EIC (Hawkins 1990), need to consider production data, since speakers are the proximate cause of the features of language, and whenever they talk, they create linguistic structures. Bock (1995) also noted that comprehension research has been playing a more substantial role than production research in informing competence theories, as studying comprehension alone may avoid the largely unknown pathway from thought to language. Comprehension research has also been tied closely to language acquisition, where the acquisition problem is mainly a problem of making sense of linguistic input. Bock et al. (in press) mentioned that speaking seems to be more difficult than listening in several aspects: ƒ

various cues 9 are available to the listener;

ƒ

retrieval is more of a challenge in production;

ƒ

speaking a foreign language is the commonplace standard of bilingual competence;

ƒ

public speaking is many people’s greatest fear, while public listening inspires no terror.

On the other hand, arguments for the contrary are also provided in Bock et al. (in press): ƒ

listening involves segmenting a continuous stream of sound into identifiable components;

ƒ

ambiguity is more of a problem in comprehension;

ƒ

to correctly produce an utterance in a foreign language like où est la toilette? is easier than comprehending a fluently spoken response (and a pointed gesture is preferred).

                                                         For example, Arnold et al. (2003) reported that disfluencies in speech are used by listeners to help them interpret speakers’ intentions, as descriptions of new referents are likely to be more complex, and hence contain more disfluencies than descriptions of given/old objects.

9

20 

2 Relationship between grammar and processing  Another argument comes from the psychology of dialogue, where the proposal of an interactive alignment model (Pickering & Garrod 2004, see Figure 3 below) assumes that the human processor draws upon the same representations constructed for the purposes of comprehension or production can be used for the other task.

Figure 3 Interaction model of dialogue in a schematic representation of the stages of  comprehension and production processes (Pickering & Garrod 2004:176 Figure 2)

Pickering and Garrod (2004) argued that it would explain priming in production when speaker has produced (or cross-modally, heard) the prime, as well as why we can complete one another’s utterances and get the syntax, semantics, and phonology correct. They also argued that parity is also increasingly advocated outside language. Vigliocco and Hartsuiker (2002) have suggested that the lack of converging results between language comprehension and production may just be due to the differences in the task nature and the different degrees of interactivity of the two tasks, but there remains a need for a shared system for both tasks: [B]ecause of the differences in their goals, production and comprehension may require different degrees of temporal overlap of different processes and different processing strategies. These different requirements, however, do not imply that the systems’ architecture is different. In fact, we argue that assuming the same architecture is parsimonious. Bidirectional flow of information, we argue, is essential to achieve stable learning. In adult sentence comprehension, then, it may serve functions such as reducing ambiguity at each level; in production, instead, bidirectional flow of information may serve the function of ensuring accuracy, aiding a production-internal monitoring system. (Vigliocco & Hartsuiker 2002:408) To further complicate matters, even if comprehension and production are not fully shared, there are still different proposals as to how they are related. There are proposals that speaker altruism is epiphenomenal, that is, speakers do not actually take listeners into account when planning an utterance (V. Ferreira & Dell 1996), but it also appears that we listen and comprehend what we say, while when we read there also seems to be a mental narration going on. Levelt (1989) suggested similarities and interconnections between the two systems

21 

2 Relationship between grammar and processing  exist as a certain degree of consciousness is involved in the self-monitoring in speech production, and listener accommodation is a by-product of a speaker’s comprehension of his own speech. One wonders if the production system is embedded with the comprehension system (Levelt 1989) or the other way round. If the model for the language faculty proposed by Jackendoff (2002) is correct, then a single system shared across both tasks takes in either a concept or a string of words then produces the other one as output, and much of the processing architecture is shared as a result. Building on arguments against a syntax-driven view of linguistics, Jackendoff (2002) maintains the view that phonology, syntax and semantics are equally generative, and these components form a parallel architecture and they communicate with each other through “interface” components. Lexical items, which are not limited to words but may be idioms or abstract constructions 10 , may emerge from the interfaces among components. He also argued for a more integrated theory of competence and performance, as motivated by linguistic theory but also independently informed by psycholinguistics. A thought experiment is provided in Jackendoff (2002): Presumably memory limitations account for the impossibility of producing a 6000-word sentence, every part of which is locally grammatical. One could not possibly keep all that in f-mind at once. For a more extreme case, one would die before completing a 60-million-word sentence. (Jackendoff 2002:30) Jackendoff (2002) also cited the contributions of speech error research to linguistic theory, and reasoned that theories of competence and performance should line up. While he acknowledged that comprehension and production produce parts in different orders, he has cited evidence for a more integrated approach such as LFG, including interactions between phonology and morphosyntax 11 . In his remarks on HPSG and CxG, Jackendoff (2002) noted the primacy of signs (lexical items) to be complexes of sound, meaning and grammatical properties. Thus, language can be viewed as a whole to be a mapping between sounds and meanings, being a collection of structures of independent-but-linked levels, linked together by sets of interface constraints including words and other lexical items. These interfaces are important, as seen the emphasis of lexical items by HPSG (Jackendoff 2002; Sag et al. 2003). This leads to a parallel constraint-based architecture which is logically non-directional: it is not biased towards parsing or production (Jackendoff 2002). Starting from any structure

                                                         Jackendoff (2002) provided examples for constructions such as [V NP PRT]: V NP[time period] away means “spend NP V-ing” (Jackendoff 2002:175). 11 Jackendoff (2002) suggested that a possible interface constraint is that the syntactic subject corresponds to agent, and the object corresponds to patient (see p. 133), following the generalization that the number of syntactic arguments is almost always equal to or smaller than the number of semantic arguments. Also, objects can be usually be omitted without noticeably changing meaning, as omitted arguments usually are recoverable from context or are irrelevant to the message being conveyed (see Jackendoff 2002; Goldberg 2006). 10

22 

2 Relationship between grammar and processing  (phonological, semantic or syntactic), one can construct the corresponding elements in other components. He also stressed the implications of constraints on processing that come from the implementation of the human processor in the brain. A possible implementation is illustrated by Jackendoff (2002), with emphasis on the interfaces between the phonological, syntactic and conceptual components (see Figure 4).

Figure 4 Processing architecture proposed by Jackendoff (2002:199) 

The role of working memory    While the field of psycholinguistics has not yet decided on the architecture of our language processing system, there exists more support, both empirical and theoretical, that the tasks are not that separate after all. It is worth mentioning that the role of working memory (WM) has been gaining attention in the psycholinguistics research literature (cf. comprehension in Caplan & Waters 1999; Gordon et al. 2002; Fedorenko et al. 2004; Bock et al. 2005; Bock et al. in press). Gordon et al. (2002) reports experimental results suggesting that the verbal WM resources used for linguistic integration are not domain specific, which is different from the claim by Caplan and Waters (1999). Also, heavy noun phrase shift (HNPS) is an exemplification of WM demands (although other factors might come to play, cf. Wasow 2002; Hawkins 2004). In Japanese, there is a long-before-short preference instead (Yamashita & Chang 2001), and it is argued to affect the WM load associated with production (Yamashita 2002). One possibility suggested by Yamashita (2002) is that the prominence of the semantic content makes the speaker say a constituent earlier in the sentence, as it attracts attention (Levelt 1989). A scrambled constituent with a referential phrase is argued to be highly available in the speaker’s mind because it has been activated from the previous discourse or inferred by context. Yamashita (2002) argued that the preferences in English and Japanese allow the speaker to produce one clause at a time, which is assumed to involve a lighter WM load than working with two clauses at the same time, as the elements of one clause must be kept in a buffer while the second clause is processed (c.f. Hawkins 1994; Gibson 1998).

23 

2 Relationship between grammar and processing  Furthermore, Bock et al. (2005) reported that structural priming works across both tasks, suggesting that at least some part of our language faculty responsible for structure is shared between language comprehension and production, or even that there is a more general-purpose architecture coupling comprehension and production as priming effects in comprehension can be carried over to production processes. Bock et al. (in press) have obtained results showing that both comprehension and production tax cognitive resources in a dual task paradigm, with the primary task being driving and the secondary task which includes language comprehension and production. They observed that shortly before a complex structure is to be produced and shortly after a complex structure is heard, (simulated) driving performance is affected. Processing resources including WM seems to influence both tasks, as in HNPS and syntactic priming. This argues for the convergence of comprehension and production. Classical models usually assume separate modalities for comprehension and production, principally informed by the discovery of Broca’s area, from a study of a patient with problems producing language, and the discovery of Wernicke’s area, from a study of a patient who could not comprehend language. However, Stowe et al. (2005) suggested that comprehension and production cannot be split up in this way. In a summary of imaging studies, they have pointed out that there are activations in both areas when people perceive and produce language, suggesting that Broca’s and Wernicke’s areas contribute to both comprehension and production. Activation of Wernicke’s area during production was attributed to covert monitoring while activation of Broca’s area during comprehension was attributed to covert production or to the use of WM in rehearsal to aid in task demands. The summary in Stowe et al. (2005) led to the authors’ conclusion that both areas are necessary to the normal function of the language, with support from lesion studies indicating that most aphasics exhibit a mixed behaviour with both comprehension and production deficits in sentence processing, and cannot be easily classified unambiguously in one of the groups. The relevant questions thus become how much the two tasks overlap, where they overlap, and how processing demands affect both these components and tasks. As discussed in section 2.1, Hawkins (2004) suggests that grammar and performance are correlated – those forms allowed in grammar are the ones that are more efficient, and in turn the efficient forms are grammaticalized and accepted – when innovation and borrowing provide new ways to express an idea, or when the typology of a certain language changes (such as changing from SOV to SVO), there would be an invisible hand process to maximize the processing ease (PGCH). Hawkins (1994, 2004) cited corpus data in support of his theory that performance in both comprehension and production is limited by processing factors including WM load. The processing account by Hawkins (2004) is clearly intended to cover both comprehension and production (and perhaps language acquisition as well). The

24 

2 Relationship between grammar and processing  three processing principles MiD, MiF and MaOP are designed with both language comprehension and production in mind (Hawkins 2004), and the combined model tries to predict both comprehension and production, as they both share a same pool of memory resources. The three principles can be seen as how different factors of efficiency interact with each other in different levels: ƒ

objective message transmission, which may be abstracted from processing demands;

ƒ

how languages of the world fall into different types and how these languages are learned;

ƒ

how these languages are processed in real-time in language comprehension and production (e.g. whether they are processed differently).

This study would also test the hypothesis that comprehension and production share (most of) the same language processing mechanism and they converge (see Stowe et al. 2005). Verbal WM is assumed to be a single subsystem of human WM (Caplan & Waters 1999) and there is even evidence for a general WM (instead of domain specific WMs, Gordon et al. 2002). One may assume WM to be shared by comprehension and production. The predictions are as follows: reading time at a complex region will be slower when complexity increases; and increase in complexity will correspond with shifting (increase in frequency of another syntactic alternative). Since the same person has the same set of WM resources, the magnitude of reading time difficulty should correlate with shifting. MiD will be used to access the efficiency of the experiment items and the responses at recall. The proposed experiment attempts to correlate the processing demands of comprehension and production by requiring the same participant to comprehend and produce sentences. In light of the separate researches in comprehension and production, this experiment attempts to play the devil’s advocate in suggesting that comprehension and production actually converge rather than diverge, since the language we comprehend and the language we produce are more or less the same. The same grammar and the processing machinery appear to be shared by both comprehension and production (Stowe et al. 2005). So, the two processes seem to be at least related together by the same processing demands and syntactic rules, and the overlap may be larger than we think. Also, the results, if successful, might function as a bridge between the formalist and functionalist traditions by suggesting the innate mechanism is related to the nature of processing.

25 

3 The “Chinese Problem” 

3. The “Chinese Problem”  In a previous study by Cheung (2005), the “Chinese Problem” is described as a typological problem, explained in further detail in sections below.

3.1. The typological rarity of Chinese  Typologically speaking, Chinese languages such as Cantonese are rare and typologically marked in that they have disharmonic headedness (Dryer 1992, 2003, 2005). Most languages of the world are consistently left or right branching, and the rarity of the Chinese type and the trend for languages to avoid head-initial VPs and head-final NPs is taken by Dryer (2003) to imply the existence of some causal factors against languages to become the Chinese type. Greenberg (1966) provided a landmark proposal in typology to describe universals of grammar using 45 rules (Language Universals, LU) with particular reference to constituent order of 30 languages, but several of these are problematic when applied to Chinese collectively, for example see Greenberg (1966): ƒ ƒ

ƒ

LU 2: In languages with prepositions, the genitive almost always follows the governing noun, while in languages with postpositions it almost always precedes it. (Greenberg 1966:78) LU 22: If in comparisons of inequality the only order, or one of the alternative orders, is standard-marker-adjective, then the language is postpositional. With overwhelmingly more than chance frequency, if the only order is adjective-marker-standard, the language is prepositional. (Greenberg 1966:89) LU 24: If the relative expression precedes the noun [Rel N]... either the language is postpositional or the adjective precedes the noun or both. (Greenberg 1966:91)

Given that Chinese appears to be primarily prepositional, we have the following expected observations with respect to these LUs: ƒ

Re LU 2: Chinese genitives precede the head noun.

ƒ

Re LU 22: The comparisons are in the order marker-standard-adjective in Mandarin, and adjective-marker-standard in Cantonese.

ƒ

Re LU 24: The order in the NP is adjective-noun and RelN.

Greenberg’s universals holds largely true for most languages of the world, despite the fact that the sample consists of only 30 languages. However, if they do hold for Chinese, then Chinese would have been both prepositional and postpositional. To complicate matters, Chinese has rather blurry syntactic categories, and there are views arguing for Chinese as a prepositional language (Li 1990, but also see Li & Thompson 1976) and by some for Chinese to be a postpositional one (Ernst 1988, see also Matthews & Yip 1994 and Dryer 2003 for discussions on this topic). Dryer (2003) also suggested that it is best not to say that

26 

3 The “Chinese Problem”  the language is primarily prepositional or primarily postpositional, since Chinese languages have an unexpected number of postpositions as a VO language. Dryer (2003) further commented on the Chinese type of languages being a rare typological type as follows: ƒ

Chinese also has a rare word order of prenominal modifiers [XP N] while the verb phrases are head-initial.

ƒ

The preverbal manner adverbs and prenominal relative clauses (RelN) are unusual for VO languages.

ƒ

Among a sample of 199 VO languages, only Mandarin, Cantonese and Hakka have a rare possible order of PP V.

Most of the languages of the world are consistently head-initial or head-final (see Figure 1 on page 10), but Huang and Li (1995: 54) observed that Chinese has disharmonic headedness, stating that ‘Chinese is an essentially head-final language, except for the objects of verbs, objects of prepositions and a limited number of complements’. Dryer (1992) stated that Chinese languages are the only VO languages that have RelN order (also in Dryer 2003). Dryer (1992) noted that his Branching Direction Theory predicts that in VO languages modifiers should tend to follow the noun, stating in a footnote: In many languages, “adjectives” are really verbs, so that, as modifiers of nouns they are really relative clauses. As relative clauses they are instances of a fully recursive phrasal category and thus the [Branching Direction Theory] predicts that they too should tend to follow the noun in VO languages and precede it in OV languages. I assume that this prediction is correct: since they are relative clauses, they should exhibit the same pattern as relative clauses, preceding the noun more often in OV languages than in VO languages, though often following the noun in OV languages, just as relative clauses do. (Dryer 1992:119 fn 19) After examining a larger sample of 580 languages, Dryer (2005) has found that Chinese languages are still almost the only VO languages to have RelN order 12 . Dryer (1992, 2003, 2005) observed that while for OV languages the distribution of RelN and NRel is rather balanced, the VO languages are largely skewed towards NRel. VO languages are consistently NRel, and the Chinese type, being SVO and RelN, accounts only for slightly more than 1% of the VO languages. Table 1 summarizes the findings from Dryer (2005). Languages OV VO

RelN 109 5

NRel 96 370

Table 1 Relationship between VO‐OV and RelN‐NRel (580 languages, Dryer 2005)

                                                         Bai, a Tibeto-Burman language of China, and Amis, an Austronesian language of Taiwan, also have the same configuration, but they are influenced by Chinese (Dryer 2005) as they are both in close geographical proximity to Chinese languages. 12

27 

3 The “Chinese Problem”  Dryer (2003) further elaborated that such rarity implies some causal factors discouraging languages of such a typological configuration from arising, and encouraging such languages to change to some other types if they do arise, for example, a change from OV to VO is followed by the language becoming NRel and V-PP either simultaneously or shortly afterwards. It is thus exceptionally notable that Chinese has been VO and RelN for a long time. This seems to be in violation of several language universals (cf. Dryer 1992, 2003, 2005; Hawkins 1994). Hawkins (1990, 1994, 2004) made several comments on Chinese languages: ƒ

ƒ

VO languages almost invariably employ postnominal rather than prenominal relative clauses. Mandarin Chinese is an obvious exception here, though it is typologically mixed, having both VO and OV, and prepositions and postpositions, in addition to RelN. (Hawkins 1990: 255) An early verb in VP and a final noun in NP is typologically quire rare … and it results in some very long PCDs for VP when the relative clause is complex. (Hawkins 2004:141)

VPs are one particular case of interest (Hawkins 1994:268): there are two structures attested only in Sinitic languages, one being preverbal PP in the VP, as in example 16: PPs [PP P [[RC] N]NP]: ‘Chinese has many [VP PP V NP] orders, and is apparently quite unique in this regard’ (Hawkins 1994:281) 16. 我哋 要 早啲 同 結婚 嘅 對象 講 ngo5dei6 jiu3 zou2di1 [PP[tung4 [git3fan1 ge3 deoi3zoeng6]] gong2]VP. we need earlier with marry Prt partner talk “We need to talk in advance with the partner who is getting married” (Matthews 2004)

ƒ

The other being the combination of VO word order and prenominal relative clauses (RelN), which is almost unique to Sinitic languages, as in example 17: ƒ VP[V [[RC] N]NP] 17. 我 畀返 琴日 借 嘅 錢 你 ngo5 bei2faan1 [kam4jat6 ze3 ge3 cin2 ]NP lei5 I give back yesterday borrow Prt money you “I’m returning the money I borrowed yesterday to you” (Matthews & Yeung 2001) Datives as such will be discussed in further detail in section 3.3, Cantonese and datives. The combination of RelN and VO also contributes to Hsiao and Gibson’s (2003) observation that Chinese is special in terms of subject relatives being harder to process than object relatives, against NPAH proposed by Keenan and Comrie (1977). The “Chinese Problem” – typology  The combination of SVO and RelN is rare. Either the Chinese order [V RelN] is actually not problematic, despite the fact that most languages avoid this order; or if there are causal factors behind this dispreference (Dryer 2003), why is this order retained by Chinese languages while other OV languages becoming VO change from RelN to NRel?

28 

3 The “Chinese Problem” 

3.2. The difficulty in processing Chinese sentences with RCs  Dryer (2003) postulated that there are causal factors discouraging language to be of the Chinese type. In a processing account, the configuration of Chinese does cause processing difficulties, as explained by Hawkins (1994). EIC can account for many typological generalizations across languages, but some aspects of Chinese appear to be an exception to the theory, as certain basic principles are commonly violated and explanations are needed to explain the discrepancy. As mentioned in Cheung (2005), the consequences of being a language with SVO and RelN are that it inevitably results in centre-embedding, which is problematic for the parser (see also Hawkins 1994). This is supported by findings that it is difficult to process Cantonese phrases embedded with heavy elements such as RCs (see psycholinguistic literature concerning the processing of Chinese, for example, production experiment investigating Cantonese datives, Cheung 2004, 2005; reading experiment investigating relative clauses of Mandarin, Hsiao & Gibson 2003; reading time of Cantonese zoeng1-construction, Kwan 2000; reading experiments of topicalization in Cantonese, Matthews & Yeung 2001, Matthews et al. 2005). Processing typology (Hawkins 2004) would suggest that the PGCH, in terms of its prediction of word order typology, is supported by the following: ƒ

the avoidance of centre-embedding by most languages of the world;

ƒ

the markedness of the Chinese type of languages with both SVO and RelN;

ƒ

the trend for languages to avoid the Chinese type by changing their configuration after converting from OV to VO (Dryer 2003).

As mentioned in the previous section (p.28), Hawkins (2004) commented that SVO and RelN would allow very long VP domains when the RC is complex. For example, when the object of a transitive VP is modified by a RC, the sentence structure becomes the following: NP VP[V [[RC] N]NP] The RC is centre-embedded in the VP, and if the RC is increasingly complex, the recognition of the VP becomes increasingly difficult as its PCD (underlined) becomes longer. Pinker (1994) recognizes the problem of centre embedding – the parser does not know what to do when a verb (V) is centrally embedded in a verb phrase (VP) with heavy preverbal and postverbal phrases, so that the VP is virtually impossible for the parser to understand – the person would even think that the sentence is ungrammatical. An example is a

29 

3 The “Chinese Problem”  doubly-embedded Mandarin relative clause in a reading time study by Hsiao and Gibson (2003:8), when processing theories such as EIC and DLT predict that there would be processing difficulties, for example as in example 2, which is repeated below: 18. [[邀請 [勾結 法官] 的 富豪] 的 官員] 心懷不軌 [yaoching [gojie faguan] de fuhao] de guanyuan shinhuaibugui invite conspire judge Prt tycoon Prt official have-bad-intentions “The official who invited the tycoon who conspired with the judge has bad intentions” (Hsiao & Gibson 2003:8 Example 4b) In 18, a RC [gojie faguan] is centre-embedded within another RC [yaoching [gojie faguan] de fuhao] 13 . The reading times at critical sites (faguan, fuhao and de guanyuan) are found to be extremely slow (near 1000ms), indicating parsing problems (Hsiao & Gibson 2003). Sometimes the parser may fail, leading the speaker to misinterpret the sentence or judge it as ungrammatical. The explanation by MiD is as follows: the LD of yaoching, the PCD of the agent NP (yaoching…de guanyuan) and the FGD of its head noun guanyuan become extreme long; and according to DLT, there are multiple dependencies and intervening items in the whole agent NP. To sum up, SVO and RelN introduce centre embedding, as exemplified by the rarity and dispreference of such a configuration in typology. Combined with empirical findings that centre-embedding is difficult to process, the “Chinese Problem” is confirmed to present a processing problem. The Chinese configuration is thus a bad configuration according to models such as MiD and DLT, and it presents a problem for accounts that predict hard-to-process structures are impossible structures. The long-lasting configuration as observed by Dryer (2003) challenges these models of language processing. The “Chinese Problem” – processing  The combination of SVO and RelN results in a very long VP domain when the RC is complex (Hawkins 2004). This configuration is also typologically marked, as supported by predictions by processing theories like DLT or MiD. Why is this order retained for such a long time by Chinese languages (Dryer 2003)? How can processing stay efficient given the inefficient order?

3.3. Cantonese and datives  According to Bauer and Matthews (2003), areal diffusion is believed to cause a North-South cline where northern Chinese languages/dialects such as Mandarin are more head-final, while in the South, the head-initial type like Cantonese is more common. Cantonese has

                                                         13

The Romanization is that of Hsiao & Gibson (2003), following Taiwanese usage.

30 

3 The “Chinese Problem”  certain head-initial features that are not common in Chinese languages, and there are parallel constructions that co-exist with a Mandarin-like counterpart. Bauer and Matthews (2003) also cited sources that support the idea of language contact that various Tai-Kadai, Austronesian and Sino-Tibetan languages coexisted in the region that Cantonese is spoken, and these languages are in intensive contact that may have affected Cantonese. The North-South cline is reflected in the different configurations for datives. Cantonese datives are of particular interest because it has a number of unusual word order patterns, including the dative construction: [V T(heme) R(ecipient)]. Chinese languages in Northern China like Mandarin has [V R T], while in Southern China, Chinese languages like Cantonese have the rarer order [V T R]. Combining the canonical order of the dative construction [V T R] and RelN, the problem of centre-embedding in the “Chinese problem” is compounded as centre-embedding becomes even more serious. As informed by processing theories such as MiD and DLT as well as empirical findings from Cheung (2004, 2005), Cantonese datives are more difficult to process especially when the theme is modified by a relative clause: VP[V [[RC] N] R] As seen above, the PCD of the VP and the LD of V span the whole length of the VP in order

to recognize all the arguments of the ditransitive verb V. Only when the modifier of the theme is short is the VP efficient in MiD terms. In processing terms, this can be illustrated as follows: English has SVO and NRel, so both the LD of give and the PCD of the VP are not lengthened by RCs modifying the theme: 19. I give you a book. 20. I give you a book that I use. Meanwhile, Mandarin has [V R T], and Cantonese has [V T R], and both lead to centre-embedding when RCs modify the arguments in the VP. Examples 21 and 22 are in Mandarin, which has the order [V R T]: 21. 我 給 你 一本書 I give you one CL book (see example 19) 22. 我 給 你 我用的一本書 I give you [I use Prt one CL book] (see example 20) In Mandarin examples 21 and 22, the recognition of the theme NP is delayed by the introduction of the RC modifying the theme. However, as Cantonese has [V T R], the RC [ 我用嗰本/嘅]書 “The book [that I use]” would delay both the theme and the recipient: 23. 我 畀 一 本 書 你 ngo5 bei2 [jat1 bun2 syu1] nei5 I give one CL book you (see example 19)

31 

3 The “Chinese Problem”  24. 我 畀 [我 用 嘅 一 本 書] 你 ngo5 bei2 [ngo5 jung6 ge3 jat1 bun2 syu1] nei5 I give [I use Prt one CL book] you (see example 20) As in example 24, in order to recognize the VP headed by bei2, the parser has to wait until the recipient, nei5, occurs, and the recognition of both ICs (T syu1 and R nei5) are delayed while several words of the RC [ngo5 jung6 ge3 jat1 bun2] have to be processed, inducing a processing load. As both T and R are separated from V, DLT would predict the Cantonese datives to be processed slower than a Mandarin one if the theme is modified by an RC. Thus, Cantonese datives should theoretically be more difficult to process than Mandarin datives, which are more difficult to process than English datives. Another point why Cantonese datives are worth investigating is that more alternative constructions are available (see Cheung 2004, 2005 and also chapter 4). The “Chinese Problem” – Cantonese‐type datives  Datives present a more interesting case, as the “Chinese Problem” is compounded. As datives are more complex VPs, heavier centre-embedding will cause a more serious problem for the human processor. Cantonese datives, with [V T R], are also considered more difficult to process since when the objects are modified by RCs. How can Cantonese speakers tolerate the super inefficient ditransitive VPs? Do they use special strategies?

32 

4 Alternative constructions 

4. Alternative constructions  4.1. Alternatives in and across languages  We can say things in many different ways but still express the same idea. This is not limited to specific languages that allow free word order, but also applies to languages with relatively fixed orders, such as English and Cantonese. Options include paraphrasing, topicalization, scrambling, etc. However, in this study, the concern is only with VPs of a more complicated type, namely dative constructions.

4.1.1. English and Japanese  Various alternative constructions can be found in English and Japanese. In experiments conducted in English, Wasow (2002) obtained various types of alternative constructions 14 , for example, “topicalized object” 15 , as summarized in Table 2 . Goal first Double object, give IO DO HNPS, give DO IO Passivized goal Double object, goal topicalized Prepositional, goal topicalized Prepositional Passivized theme Prepositional, theme topicalized

Give the white rabbit the carrot Give to the white rabbit the carrot The white rabbit should be given the carrot The white rabbit, give the carrot To the white rabbit, give the carrot Theme first Give the carrot to the white rabbit The carrot should be given to the white rabbit The carrot, give to the white rabbit

Table 2 Possible constructions with give in English (Wasow 2002)

Japanese, on the other hand, is considered a topic-prominent language (as are various other East-Asian languages including Chinese and Korean). It has case-marking and a topic marker –wa, and the order of theme and recipient are not restricted to a particular order. Yamashita and Chang (2001; cf. corpus study in Yamashita 2002) has obtained different types of scrambling, using the constrained production paradigm: 25. Taro-ga/wa Miki-ni ringo-o ageta Taro-nom/top Miki-dat apple-acc gave 26. Miki-ni Taro-ga/wa ringo-o ageta 27. Ringo-o Taro-ga/wa Miki-ni ageta 28. Taro-ga/wa ringo-o Miki-ni ageta “Taro gave Miki an apple” (Example 2a-d,Yamashita & Chang 2001)

                                                         Wasow (2002) also obtained extraposition from NP (discontinuous NP), such as A gun went off which I had cleaned. 15 Topicalization is common in Chinese languages (including Cantonese), and Cantonese-English bilingual children also uses topicalization in English (see Matthews 2004). 14

33 

4 Alternative constructions 

These sentences all mean Taro gave Miki an apple, and the fairly free word order and null pronouns allow Japanese speakers to use word order to mark the saliency of different arguments (Yamashita & Chang 2001). Having multiple alternative phrase orderings available can itself also lead to faster and less error-prone utterances (V. Ferreira 1996).

4.1.2. Cantonese  According to Matthews and Yip (1994), Cantonese has a word order about as fixed as in English since they both have little grammatical morphology, and they rely on word order to express grammatical relations. The basic order of Cantonese is SVO, with variations of different types, such as topicalization (when an element is preposed to the front of the sentence and made the topic of the sentence), right dislocation (where an element is put last, separate from the rest of the sentence), and the zoeng1-construction which is also possible in Mandarin (the BA-construction). The canonical order of Cantonese datives is different from that of English. The dative verb is bei2 (畀 give).The canonical order [bei2 T R] in Cantonese is cross-linguistically marked (according to Universal Grammar, Michaelis & Haspelmath 2003; also described as the “inverted double object construction” in Tang 1998), whereas the 2 alternative orders [bei2 R T] and [bei2 T bei2 R] are cross-linguistically unmarked (and acquired earlier by monolinguals and produced more frequently, as observed by Chan 2003). Cheung (2005) suggested a possible reason for this order to be prevalent in Cantonese: [bei2 T R] allows verb serialization [bei2 T R V2] such that R is the agent of V2; As discussed in Cheung (2004, 2005), most of the strategies in Wasow (2002) are attested. Several alternative structures are available for the dative constructions as listed below: ƒ

[V T bei2 R], the serial verb construction (SVC) 16 (Matthews & Yip 1994:138; cf. “prepositional dative construction” in Tang 1998:36) 17 with ling1/ning1 拎, lo2 攞,

                                                         SVC is a family of constructions (see Goldberg 2006), as Li and Thompson (1981) have suggested that SVCs may be analyzed into different types of coordination or subordination structures. 17 Michaelis and Haspelmath (2003) have argued that the borderline between SVC and a preposition marking the recipient (Indirect Object Construction) is fuzzy, depending on the grammaticalization of the serial verb, such as whether V2 can be negated or take TMA markers. For example, in Yoruba, Saramaccan and Haitian Creole (Lefebvre 1998), the verb give is attested in the V2 position of SVCs: Mwen pran liv bay Pòl. I take book give Paul “I gave the book to Paul” (Haitian Creole: Lefebvre 1998:291) The Cantonese bei2 is also considered to be a verb in V2 position (Cheung 2005), due to its behaviour resembling more closely as a second verb (cf. Matthews 2006). Matthews (2006) observed that although bei2 in V2-position is unable to take aspect marking, it can take verbal particles such as faan1 (返): 16

34 

4 Alternative constructions  gaau1 交, and bei2 畀 (see a discussion of the double-bei2 construction in Tang 1998) as possible Vs; ƒ

[zoeng1 T bei2 R], the zoeng1-construction (which retains a sense of displacement, cf. Matthews & Yip 1994:144, Bauer & Matthews 2003:153; cf. S. Cheung 1994,; and notes on BA-construction by Dryer 2003);

ƒ

[bei2 R (SFP) T], which may be HNPS or right dislocation (RD) 18 (see also Matthews & Yip 1994:137; Tang 1998:36-7; Bauer & Matthews 2003:153);

ƒ

[T, A bei2 R] topicalization (applied usually to T, see Matthews & Yeung 2001) 19 ;

ƒ

extraposition from NP, [bei2 [simplified NP of T] R (SFP) [modifier of T]], as in the following example:

[個 先生], [喺 上堂 用 嘅 英文 嗰 本] 29. 畀 [本 書] [bei2 [bun2 syu1] [go3 sin1saang1]], [hai2 soeng3tong4 jung6 ge3 jing1man2 go2 bun2] give [CL book] [CL teacher], [at class use Prt English Det CL] “give the book to the teacher, the one used in class” (Cheung 2004) Extraposition is also interesting because it is possible in some other languages that have less-than-optimal configurations, as in Thai 20 and German (which has centre-embedding sometimes [NRel V2]). A related observation is in Luke (1998), where it is observed that Cantonese has also post-modifiers in NPs, as in example 30: 30. 我 想 搵 個 細蚊仔 肥肥哋 嘅 ngo5 soeng2 wan2 go3 sai3man1zai2 fei4fei5dei2 ge2 I want find CL kid chubby Prt “I want to look for a child who is a bit chubby” (Luke 1998:48 Example 3a) This observation may be explained by a possibility of [N modifier] order, but it could also be an extraposition of the modifier of the noun, as in Cheung (2004, 2005). Luke (1998)

                                                                                                                                                              我 攞 啲 衫 畀返/*咗 佢 ngo5 lo2 di1 saam1 bei2-faan1/*-zo2 keoi5 I take CL clothing give-back/PFV 3sg “I brought her back some clothes” A double-give construction, using the verb give in both V1 and V2, is also attested in the language Fa d’Ambu (Michaelis & Haspelmath 2003). 18 For a discussion on RD, see Y.L. Cheung (1997). However, since the SFP is not always present, it is difficult to distinguish between HNPS and RD (Cheung 2004, 2005). 19 There were some relevant observations of topicalization in Cantonese as early as in 1941 by O’Melia (1941), If the sentence becomes clumsy or involved, e.g. when the relative clause modifies the object of the main verb, Chinese more commonly prefers to use the ‘Nominative Absolute’. Example: He would very much like to buy the house you are building. Theoretically this would be: He would very much like to buy you are building that house. In fact such a clause is more commonly transposed, e.g. The house you are building, he would very much like to buy; i.e. You build that house, he would very much like to buy, nei hei koh kaan uk, k’ui hou seung maai. [nei hei go gaan uk, keoi ho soeng maai.] (O’Melia 1941) 20 A Thai example would be: hay ngen khun saam baat give money you three baht “I'm giving you 3 Baht (of money)” (Thai: Matthews p.c.)

35 

4 Alternative constructions  observed that the RCs in such a position are non-restrictive. He further demonstrated that the meanings of [modifier N] and [N modifier] may be different: sometimes it is unusual not to use a postnominal modifier, but for certain situations it would be judged unacceptable. This phenomenon thus appears to be affected by semantic factors, but syntactic factors may also be involved. Another logical problem, observed by Tang (1998:36 fn 2) is that the canonical order in Cantonese [V T R] is actually ambiguous if both objects are animate. Tang (1998) postulates that for the canonical construction there is a null dative marker, since some judge it as slightly unnatural when the main verb is bei2, perhaps due to the repetition of bei2; but bei2 is obligatory for certain other verbs (such as ling1/ning1 拎, lo2 攞 and gaau1 交) as V: [V T {bei2 / ø} R] Matthews and Yip (1994) have reported that there is sometimes a redundant use of keoi5 (3sg pronoun) as object of the second verb in serial verb constructions (SVC) – this is compatible with the claim that the construction is possibly motivated by parsing efficiency (see section 4.2.2), as in example 31: 31. 要 將 啲 唔開心 嘅 嘢 gut 一聲 吞咗 佢 jiu1 zoeng1 [di1 m4hoi1sam1 ge2 je5] gut2jat1seng1 tan1-zo2 keoi5 need put CL unhappy Prt stuff “gut”-one-voice swallow-PFV it “You should take the unhappy things and swallow them in one gulp” (Matthews & Yip 1994)

4.2. More on alternative constructions in Cantonese  4.2.1. Historical perspective  The configuration of Cantonese, SVO, RelN and [V T R] and the alternative orders may be remnants of an older word order (Cheung 2005). The following alternative orders of the dative construction are possible in Archaic Chinese (Pulleyblank 1995) and Classical Chinese (Peyraube 2004): ƒ

the double object construction with the order [V R T];

ƒ

the yi (以, possibly a direct object marker) construction: [yi T V R] or [V R yi T] (restricted to [+give] [+say] and [+teach]), where the object pronoun in the yi phrase can be omitted (Pulleyblank 1995);

ƒ

the yu (于/於, the preposition to) construction: [V T yu R] (applicable to all verbs);

ƒ

topicalization, [T, (A) yi V PP]

36 

4 Alternative constructions  Djamouri (2001) provided a count of VO and OV orders in a description of the dominant order in Shang bone inscriptions. In the count, objects are mostly in the postverbal position, with 93.8% frequency (S)VO in contrast to 6.2% of (S)OV. The ditransitive can also be expressed in various ways. For example, the recipient (R) can be expressed either by an unmarked NP or a PP, and the theme can be omitted. [V yu R] and [yu R V] are possible when T is absent, and when both T and R are present, [V yu R T] and [V T yu R] are found. Topicalization is also found in the study. It is also observed that spatial locative phrases are usually postverbal and temporal locative phrases preverbal. Djamouri (2001) provided the note that word order changes ‘have to be interpreted with respect to the dominant order, since they often reveal a particular semantic intention’ (p. 146). This seems to influence, or is influenced by, certain current analyses that alternations to word order often have semantic-pragmatic factors. Herforth (2003), on the other hand, described Late Zhou Chinese (LZC, 475-221 BC) grammar. It was shown that Chinese has the order [(S) (A1) V (O) (A2)] (where A2 is the 2nd predicate) at that time. Topicalization and the yi construction are both found in corpus data. The item yi was described as an instrumental adposition, used in inversion when in questions or in focus and [S O/A2 V] occurs as a regular alternative to the basic [S V O A2] order determined by semantic factors (Herforth 2003). It is also noted that LZC is top-heavy as opposed to English (which is bottom-heavy), differing significantly in fundamental ways, i.e., semantic (wide-scope) or discourse-pragmatic (identificational/interrogative focus) O and A2 phrases must occur in the preverb, and the preverb position ‘accommodates gapless topics and wide-scope “any” phrases, adjuncts, object question phrases, and focus elements’ (Herforth 2003:69). Meanwhile, Peyraube (2004) reported that the BA-construction (把, which corresponds to the contemporary Cantonese zoeng1 將) appears in the form of [S ba T V R] around the sixth century in early Medieval Chinese. He hypothesized the construction to be a structure with ba as a preposition grammaticalized from serial verb constructions by analogy with the yi construction. More on the BA-construction is discussed later in section 4.2.5. Chappell (2001) summarized that in general, the modifier precedes the modified, and that can predict a lot of structures from this basic form, e.g.: ƒ

subordinate clauses precede main clauses;

ƒ

in VP syntax, adverbials of manner in general precede verbs apart from some notable exceptions in Southern Sinitic languages for time and frequency adverbs which may follow;

37 

4 Alternative constructions  ƒ

modal verbs precede main verbs;

ƒ

prenominal modification in NP syntax, i.e. relative clauses, possessors and attributives all preceding the head, with the structure: {(Dem)/(Num)} CL N

Yue (2003, 2004) also recognized the differences in the syntax of Chinese languages being understood ‘in the light of contact with neighbouring languages’ (Yue 2003:84). Yue (2003) noted that NPs and PPs are head-initial, where VP is head-middle. Her account of the VP is as follows: VP Æ (Adv) (PP (Prt)) V (Asp) (NP) (NP) (Comp)/(Adv) Yue (2004) described some changes in the history of Cantonese datives. Formerly, the DOC was mostly construed with [V T gwo3 R], with gwo3(過) as the indirect object marker, irrespective of verbal category, but DOC gradually came to have the recipient unmarked or marked with bei2 and not with gwo3 at all. The use of bei2 with R is now common usage. In short, Cantonese double object constructions (DOC) can be classified into two types: ƒ

for [+give] verbs: [V T {bei2/gwo3(過)} R];

ƒ

for [-give] verbs: [V R T] as in Mandarin

Yue postulated that the DOC probably appeared as a SVC with a [+give] verb in V2, where V1 is not one with the inherent feature of [+give]. The entire construction was one that stressed giving service. Later, bei2/gwo3 appeared in the V2 position even when the first verb carries the feature of [+give]. Since bei2 would then duplicate the meaning of give, it became gradually grammaticalized into an indirect object marker. Over the last hundred years, the DOC allowed more [+give] verbs to use the [V R T] order. She noted that [V R T] with [+give] is very bookish, and is possibly the result of influence by Mandarin and English. Language contact also plays a role. According to Thurgood (2003), Chinese languages have had various forms of contact with other languages such as Austroasiatic (Mon-Khmer), Altaic (e.g. Tungusic and Mongol), Tibeto-Burman languages, Tai-Kadai, Hmong-Mien, and Austronesian. These languages had a significant impact on word order (Dryer 2003). For example, Bauer and Matthews (2003) have stated that the orders peculiar to Cantonese are head-initial, when parallel constructions exist with a Mandarin counterpart. Matthews and Yip (2001) provided evidence for two strata in Cantonese RCs, and the Cantonese-native one is first acquired. Language contact is also a good solution to the question why Sinitic languages have a North-South cline (Ansaldo & Matthews 2001). Ansaldo and Matthews (2001) have described Chinese languages as “creoloids”, and suggested that language

38 

4 Alternative constructions  contact have been a significant factor of the characteristics found in Cantonese (see also section 3.3).

4.2.2. Various accounts of alternative constructions  Alternative constructions are usually attributed to semantic-pragmatic factors (such as information status or “topic prominence” as a characteristic of Chinese languages, Li & Thompson 1976; Matthews & Yip 1994) or lexical factors (Tang 1998). Such factors are also discussed in the English literature, such as information status (cf. Stallings et al. 1998; F. Ferreira 2003), prosodic factors (Fodor 1998; Jackendoff 2002), local ambiguity avoidance (Wasow & Arnold 2003), and lexical factors (V. Ferreira 1996; also how experiencer verbs are processed differently in F. Ferreira 2003:190-1). For a summary, see Wasow and Arnold (2003:147-150). The traditional view in Chinese (including Cantonese) linguistics is that the different forms are driven by different factors, mostly semantic, discourse, and/or pragmatic ones. For example, topicalization is held to be driven by discourse or pragmatic reasons such as establishing the topic or the given-before-new principle. Although these traditional accounts appear to converge with the constructionist approach that constructions have meanings of their own (see for example, Goldberg 1995), this is perhaps a remnant of structuralism. Wasow (2002:57,131) suggested that one should not take a simplistic view of structural ordering based on the following reasons: ƒ

the ultimate explanation of constituent ordering will be multi-dimensional;

ƒ

syntactic constraints are not categorical, can vary in strength, and their interactions can be complex;

ƒ

structural factors such as syntactic complexity interact with semantic factors, discourse factors, lexical biases and constructional biases.

The strength of processing factors can be seen in various phenomena. One such case is that of heavy NP shift (HNPS) for datives (see example 5, I give to him the book that I have borrowed). A processing account suggests that difficulty is caused by the complexity of the theme in the sentence, which affects the processing efficiency of the input sentence in EIC/MiD terms. V. Ferreira (1996) and Stallings et al. (1998) both used a constrained production task to probe the human production system. V. Ferreira (1996) found from an error study in dative alternation experiments that alternating datives (such as give which allows more alternative constructions, as opposed to the donate-type, which does not allow DOC) enjoy higher syntactic flexibility. Stallings et al. (1998) have demonstrated that in HNPS, long or heavy theme NPs are more likely to be moved to the end of sentences when compared to a shorter theme. Wasow (2002) carried out experiments involving 2 participants

39 

4 Alternative constructions  who have to ask and answer questions according to cue cards and as a result, the task appears more natural. Wasow (2002) observed that alternative constructions are preferred when the theme is more complex, that is, they are used significantly more often with long themes than with short themes (see Section 4.1.1, Table 2 on page 33). As previously mentioned, Japanese also shows a weight effect in scrambling, as a long-before-short preference in Japanese is obtained in experimental results (Yamashita & Chang 2001) and corpus data (Yamashita 2002), contrary to the short-before-long preference as seen in English (Stallings et al. 1998; Arnold et al. 2000; Wasow 2002). This result is predicted by both MiD and DLT (see Hawkins 1994, 2004; Gibson 2000). English verb-particle constructions offer one aspect where MiD makes more accurate predictions than DLT. Hawkins (2004) is explicit in being consistent with constructionist approaches (which can be implemented in phrase structure grammars such as HPSG, as opposed to context-free grammar), while Gibson (2005b) assumes grammar to be context-free. In verb-particle constructions where a dependency between the structure and the meaning of the construction exists, Hawkins (2004) has also defined a Lexical Domain (LD) formed by the verb, particle and the argument. Furthermore, Lohse et al. (2004) demonstrate that in production, people tend to minimize the lexical domains of the verb-particle construction (as predicted by MiD). This can be illustrated by the following examples 32 and 33: 32. Joe [looked up [the number of the ticket]] PCD of VP: 4 words 33. Joe [looked [the number of the ticket] up] PCD of VP: 7 words (Lohse et al., 2004) Lohse et al. (2004) explained that the immediate constituents of the VP look…up are the head of the VP look, the head of the NP number (or the, which will construct a NP/DP node) and the particle up, and these three heads will form the PCD of the VP, following MiD. The production experimental data and corpus data reveal that speakers prefer to minimize the domains. In contrast to MiD, DLT would predict no such preference since there is no intervening effect by up. The difference reflects the fact that the focus of DLT is the number of discourse referents, which has some experimental support but perhaps has downplayed some other evidence as in Lohse et al. (2004). The underlying reason may be that DLT is still context-free (see Gibson 2005b). Chinese linguistics  In the Chinese literature, factors affecting the use of alternative constructions are traditionally attributed to topic-prominence, while processing factors are rarely considered.

40 

4 Alternative constructions  In a description of the dominant order in Shang bone inscriptions, Djamouri (2001) provided a count of ditransitives. In ditransitives, the recipient (R) is expressed either by an unmarked NP or a PP. Various orders have different frequencies, as tabulated below:

T absent T present

Dominant order (frequency %) [V yu R] (78%) [(S) V (yu) R T] (70%)

Alternative order (frequency %) [(yu) R V] (22%) [(S) V T (yu) R] (30%)

Subtotal 100% 100%

Table 3 Distribution of ditransitive constructions in Djamouri (2001) 

The dominant orders for the ditransitives are the order [V yu R] when the theme is absent and [V (yu) R T] when both the theme and recipient are present. He also observed that sometimes word order variation of objects after the verb ‘does not clearly reveal any difference in semantic emphasis’, while the structure *[S V T Pron] is ungrammatical (Djamouri 2001:147 fn 6). This discovery is perhaps similar to the limitations to English verb-particle constructions – while She looked it up is grammatical, She looked up *it is ungrammatical (example taken from Lohse et al. 2004). Wasow (2002) observed that a similar case holds for HNPS - short postposed themes sound awkward or even ungrammatical, as is a heavy theme in the canonical English DOC. Cantonese appears to exhibit weight effects, and various observations provide supporting evidence that processing factors affects grammar. Tang (1998) noted that the acceptability of the structure [bei2 T bei2 R] is sensitive to the distance between the two tokens of bei2: 34. ??我 聽日 畀 錢 畀 佢 ngo5 ting1jat6 bei2 [cin4] bei2 keoi5 I tomorrow give [money] give 3sg 今年 賺 嘅 錢 畀 佢 35. ?我 聽日 畀返 ngo5 ting1jat6 bei2faan1 [gam1lin4 zaan6 ge3 cin2] bei2 keoi5 I tomorrow give back [this year earn Prt money] give 3sg “I’m giving the money (I have earned this year) back to him tomorrow” (Cheung 2004) Tang (1998) noted that datives with modified themes as in example 35 are judged to be grammatical more frequently than those without, as in example 34. Not only does this support the idea that grammaticality judgments are non-absolute, it also provides support to the claim that performance affects the choice of alternative constructions. In example 34, MiF would predict that the human processor prefers that DOC is used in order to have a even shorter PCD of the VP, but in example 35, MiD would predict a RC-modified theme would result in a long LD for bei2, and the reoccurrence of bei2 would re-activate the word bei2, and Cheung (2004) hypothesized that there might be a new minimized efficient LD for bei2 (underlined in the gloss). Yeung (1999) performed reading time studies on topicalization and found that topicalization is preferred for sentence with object modified by relative clauses, while for short object NPs

41 

4 Alternative constructions  there is no preference. This is in accordance with EIC. Objects modified by RCs are heavier and the human parser would prefer them to be topicalized. Topicalization would shorten the CRD, enhancing the recognition of sentence constituents and speeding up the time for understanding. Matthews and Yeung (2001) and a follow-up by Matthews et al. (2005) have shown that in comprehension, the preference for topicalization can be driven by syntactic weight. With these findings, Hawkins (2004) suggested that such cases are support for the PGCH. Bottom-up processing of the topic phrase, as in Japanese, allows a better configuration in terms of MiD. Hawkins (2004) also demonstrated that PGCH also predicts that the frequency of alternative structures would correlate to their relative efficiency in terms of MiD, MiF and MaOP, exemplified by the sensitivity of extraposition of German RCs to weight and also results from Matthews and Yeung (2001). Experimental study in Cheung (2004, 2005) has also established in elicited production that the choice of alternation constructions can be driven by weight, as in other languages. Matthews and Yeung (2001) and Cheung (2004, 2005) have suggested that alternative constructions in Cantonese (and Chinese in general) allow Cantonese speakers to express more efficiently with respect to the canonical order, which may be difficult to process with centre-embedding. Alternative constructions – a traditional account  Following Chomskyan linguistics or Chinese linguistics, alternative constructions are derived from a more “basic” order. Factors affecting these alternations include semantic factors, pragmatic factors, prosodic factors, and lexical factors. Processing factors are not considered. Alternative constructions – a processing account  Following the PGCH and Hawkins (2004), alternative constructions will be more preferred as their efficiency becomes better than their canonical counterparts. In other words, processing factors do have an effect in the choice of alternative constructions used. Alternative constructions may allow a “way-out” for the “Chinese Problem” by providing a more efficient construction (in terms of Hawkins 2004, Matthews & Yeung 2001; Cheung 2004, 2005).

4.2.3. Developmental data and processing factors  Cheung (2004, 2005) argued that developmental corpus data should also supply evidence that processing factors matter, following the elaboration of the idea that children with limited cognitive resources would have higher motivation to alternate the structure. Several studies on the development of the bei2 dative in children are provided below.

42 

4 Alternative constructions  Chan (2003) investigated child acquisition data of Cantonese, and corpus data points out that the alternative structures [bei2 R T] and [bei2 T bei2 R] appear before the canonical order. Children also omit the theme more often and have difficulties in using the canonical order. Chan (2003) also found that the theme is more often displaced than the recipient in adult’s child-directed speech. Similarly, influence of weight leading to displacement of a heavy direct object NP is also found in adult input, as summarized in Table 4. Structure Null T and null R [bei2] Null T [bei2 R] Null R [bei2 T] T in-situ [canonical: bei2 T R] Topicalized [T bei2 R] Postposed T [bei2 R T] Right-dislocated T [bei2 R PRT T] Serial verb structure [bei2 T bei2 R] T surfaced in other displaced positions

Frequency (total tokens: 1880) 9.15% 48.46% 5.05% 29.52% 3.19% 0.27% 1.81% 1.65% 0.90%

Table 4: Adult’s child‐directed bei2‐datives in CANCORP, summarized from Chan (2003) 

Chow (2005) noted that Chan (2003) did not report the number of tokens the children produced, and provided a count for the same corpus data: the children produced 62 datives with full arguments (i.e., A, T and R are all present), and 80.65% of these are non-canonical while the full canonical order only 7 times by children younger than 3. While Chan (2003) proposed that the early emergence of [bei2 R] was due to its high frequency in adult input, Chow (2005) believed that the input frequency account could not explain why children produce alternative constructions [bei2 R T], [T bei2 R] and [bei2 T bei2 R] with a relatively higher frequency than the canonical order [bei2 T R]. Chow (2005) also mentioned that the SVC [V T bei2 R V], being cognitively more complex, is also acquired later than [V T bei2 R] and [V T R] (where V is not limited to bei2). In a cross-sectional database containing parent -child interactions (with a total of 101 children, who are older than the ones in CANCORP), Chow (2005) counted 146 bei2 datives produced by the children (along with 33 other bei2 utterances excluded due to various reasons) and 664 bei2 datives produced by the parents. These 810 bei2 datives are classified and their distribution is summarized in Table 5:

43 

4 Alternative constructions  Dative constructions Full canonical forms (subtotal) [bei2 T R] [bei2 T R V] [V T bei2 R] [V T bei2 R V] [V bei2 R V T] Full non-canonical forms (subtotal) [bei2 R T] [T bei2 R] [bei2 T bei2 R] [bei2 T bei2 R V] [T V bei2 R] [T bei2 R V] [bei2 R V T] Non-full canonical forms (subtotal) [bei2] [bei2 T] [bei2 R] [bei2 R V] [V bei2 R] [V bei2 R V] [bei2 T V] [V bei2 T V]

Tokens from children (%) 60 (41.08%) 21 (14.38%) 1 (0.68%) 17 (11.64%) 21 (14.38%) 0 (0%) 12 (8.21%) 1 (0.68%) 6 (4.11%) 2 (1.37%) 0 (0%) 2 (1.37%) 0 (0%) 1 (0.68%) 74 (50.67%) 1 (0.68%) 7 (4.79%) 24 (16.44%) 21 (14.38%) 8 (5.48%) 11 (7.35%) 2 (1.37%) 0 (0%)

Tokens from parents (%) 358 (53.91%) 38 (5.72%) 15 (2.26%) 71 (10.69%) 181 (27.26%) 53 (7.98%) 21 (3.15%) 2 (0.30%) 4 (0.60%) 1 (0.15%) 1 (0.15%) 2 (0.30%) 7 (1.05%) 4 (0.60%) 285 (42.92%) 6 (0.90%) 32 (4.82%) 34 (5.12%) 48 (7.23%) 31 (4.67%) 123 (18.52%) 6 (0.90%) 2 (0.30%)

Table 5 Distribution of bei2 datives in child‐parent interactions (Chow 2005) 

Chow (2005) observed that parents: ƒ

produced the extended full-dative [V T bei2 R V] 27.26% of the time, while using a variant 47.88% of the time, where T may be displaced or omitted;

ƒ

produced similar but simpler non-full forms 17.02%, omitting the first or last V, as opposed to 18.52% for [V bei2 R V]. Sometimes the recipient was also omitted.

She also noted that parents omit arguments less than children (non-full forms 42.92% in adults, 50.67% in children). Chow (2005) commented that the children in her study are older and their production is therefore more adult-like than those studied in Chan (2003). Children also produce non-canonical order more frequently (8.21%) than adults (3.15%). More complex forms such as [V T bei2 R V] (14.38% for children, 27.26% for adults) and [V bei2 R V] (7.53% for children, 18.52% for adults) are used much less by children than by parents. This follows the predictions of a processing account, as older children have more cognitive resources available for language comprehension and production, but they still have fewer resources than adults for planning and producing the constructions as well as extracting and processing the input from parents, as Chow (2005) recognized. Chow (2005) summarized by saying that children between ages 3;01 and 4;07 have yet to acquire the more complex bei2-datives, as they ‘failed to manage the more extended dative form and resorted to the simpler forms during sentence production’ (Chow 2005:20).

44 

4 Alternative constructions 

This is echoed by Ho (2004), using an act-out task in a study with 60 children aged 3 to 8. It was found that bei2-datives with the agent missing 21 are understood nearly perfectly in all age-groups, while for those with the agent present, older children have better performance on the comprehension task regardless of the construction. A substantial amount of errors consist of children misinterpreting the stimulus such that they themselves are the agent of the datives. Accuracy of the children’s performance is summarized in Table 6.

Agent (A) missing

T and/or R missing

A, T, R all present

Structure [bei2 R] [bei2 T] [bei2 T R] [bei2 R T] [T bei2 R] [A bei2] [A bei2 R] [R A bei2] [A bei2 T] [T A bei2] [A bei2 T R] [A bei2 R T] [T A bei2 R] [A T bei2 R] [R A bei2 T]

3 100% 100% 90% 100% 93.3% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0%

4 100% 100% 100% 93.3% 100% 50% 40% 6.7% 40% 40% 40% 33.3% 33.3% 33.3% 0%

Average age (years-old) 5 6 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 40% 50% 40% 50% 3.3% 6.7% 40% 50% 40% 50% 36.7% 50% 40% 50% 40% 50% 40% 50% 0% 0%

7 100% 100% 100% 100% 100% 70% 70% 23.3% 70% 70% 70% 70% 70% 70% 0%

8 100% 100% 100% 100% 100% 100% 100% 33.3% 100% 100% 100% 100% 100% 100% 16.7%

Table 6 Accuracy of act‐out task summarized from Ho (2004) 

Ho (2004) also noted that when the recipient is topicalized, comprehension is especially poor. Put differently, the inclusion of agent and the topicalization of R before A increased the difficulty of the dative so much that even 8 year olds could only achieve rather low scores in such constructions 22 . The number of arguments is argued to affect the level of processing – when A is absent, performance is generally high; when A is present and T displaced, performance becomes worse; when A is present with a topicalized R, it is considerably more difficult; and when all arguments of the verb are present and R topicalized, even 8 year-olds can hardly respond correctly. 23 A follow-up revealed that most children at age 3 could already understand monotransitive VPs. Ho (2004) attributed the difficulty of datives to the overt agent in a ditransitive and the complexity of ditransitives, which take 3 arguments, and children may experience a limitation in processing capacity for three participants roles, and have to resort to selective on-line processing. She argued that being external to the VP, the

                                                         21 Missing arguments, whether they are agent, theme or recipient, can be omitted because they are provided in the task explanation before the stimulus sentence is presented. 22 However, whether the topicalization of the recipient is grammatical is debatable (see Xu & Peyraube 1997). 23 Error analysis in Ho (2004) also revealed that the younger children are able to infer that the more animate participant that follows the verb is the recipient.

45 

4 Alternative constructions  agent may be more vulnerable when the number of constituents to be manipulated exceeds children’s processing capacity, and older children have developed better cognitive processes to overcome the processing limitation. Their performance was also compared to the production of full passives by English-speaking children. To sum up, developmental data provide much support to the processing account, i.e., the late acquisition of more complex dative constructions and the higher frequency of use of simpler alternatives follow the predictions of processing theories such as EIC and DLT.

4.2.4. Different predictions for alternative constructions  Although Chinese linguistics consider the use of alternative constructions to be due to semantic/pragmatic factors (for example, see Y.L. Cheung 1997 on the discussion of Cantonese RD), which coincides with the approach of Construction Grammar (different constructions have different meanings), the more-traditional accounts suppose alternative constructions to be derived from transformations, an assumption which is rejected by Construction Grammar (Goldberg 1995, 2006). However, alternative constructions are still considered to enjoy a less favourable performance, due to displacement of arguments, a certain semantic or pragmatic focus, or a lower frequency of non-canonical orders, that there would be a possible processing penalty of being a non-canonical order. Goldberg (2006) also noted that there is a processing load for displaced arguments relative to their canonical expression and filler-before-gap places a burden on processing, especially when displacement of constituents cross over clause boundaries. The recipient (R) in ditransitives is also observed to resist being the filler in an unbounded dependency (as exhibited in the children’s responses in Ho 2004) despite it is an argument of the main clause. As a backgrounded element, the additional processing load caused by the displacement of R from canonical position combines with the pragmatic clash to result in unacceptability. Goldberg (2006) further suggested that both the functions of the constructions involved and the existence of processing constraints should be recognized. However, little support is found for a processing cost for alternative constructions. Matthews and Yeung (2001) expected a processing cost for topicalization, but topicalized OSV sentences with short theme NPs were not read slower than canonical SVO sentences. Cheung (2004, 2005) has also established that variation in syntactic frame (i.e. the use of alternative constructions) can be driven by weight, as in other languages (cf. Heavy NP Shift). However, there exists no other literature on the processing of Cantonese datives. According to predictions of the PGCH, alternative constructions of better TDDs will instead be preferred in proportion to their processing efficiency (Hawkins 2004). With respect to the “Chinese

46 

4 Alternative constructions  Problem”, Cheung (2004, 2005) postulated that different strategies are employed to maintain efficient parsing. Evidence from psycholinguistic experiments appears to support the PGCH. PGCH would predict that alternatives that have better TDDs will be processed faster and more accurately in both language comprehension and production. Alternative constructions – different predictions  Different predictions are made by different theories. “Traditional” accounts would predict that alternative constructions will be dispreferred; while PGCH would predict that alternative constructions that have better TDDs will be preferred. These predictions are assumed to hold in performance in both language comprehension and production. In Construction Grammar (CxG), the term construction is defined as a form-meaning pairing that is not derived from another basic order (Goldberg 1995, 2006). According to Goldberg (1995), limitations on certain verbs to preposing object and the ungrammaticality of pronouns in certain constructions (e.g. the sentence She looked up *it could not be derived from She looked it up) would be considered one of the arguments for alternative constructions not to be derived from other constructions. Also, CxG accepts processing factors and can be expressed in terms of Head-driven Phrase Structure Grammar (HPSG). Constructions may have different semantic properties, but they are inter-related to each other, and Goldberg (1995) suggested that when alternative constructions are viewed in their own terms, interesting generalizations will emerge. Goldberg (2006) also obtained experimental support from an offline sentence sorting experiment that constructions are as good a predictor of sentence meaning as the verb. Cross modal priming is also found, when speakers after hearing or saying one grammatical pattern are primed to produce other instances of the same construction. Alternative constructions, then, are better considered separately. As provided in section 4.1.2, various alternatives are observed in Cheung (2004, 2005) as the products of elicited production with the canonical DOC order [A bei2 T R] as input: ƒ

a high frequency of the canonical DOC order;

ƒ

[zoeng1 T bei2 R], the zoeng1-construction;

ƒ

double-bei2 construction [A bei2 T bei2 R];

ƒ

other [A V T bei2 R] responses, with V being ling1/ning1 拎, lo2 攞 or gaau1 交;

ƒ

HNPS or RD [A bei2 R (SFP) T];

ƒ

topicalization (discouraged by probe question) [T, A bei2 R] (see Bock 1996:415);

ƒ

extraposition from NP [A bei2 [bare T] R (SFP) [modifier for T]]

These constructions are elicited and introduced by the participants as they are only required to repeat the sentence meaning after hearing the stimuli. In other words, they are free to change the sentence structure while trying to preserve the sentence meaning. Cheung (2005)

47 

4 Alternative constructions  grouped the above responses observed in Cheung (2004) into DOC, the zoeng1-construction, the [V T bei2 R] construction and others. The distribution is summarized in Table 7: Frequency (%) DOC [bei2 T R] [V T bei2 R] (V includes bei2) BA-construction [zoeng1 T bei2 R] Others Subtotal

Bare T NP 94.05% 5.56% 0% 0.40% 100%

[Adj T] 81.75% 13.49% 0.40% 4.37% 100%

[RC T] 67.06% 20.24% 5.95% 6.75% 100%

Table 7 Distribution of responses obtained in Cheung (2004, 2005)

Cheung (2005) reanalyzed the findings in Cheung (2004), and found that the IC-to-word ratio could predict the frequency of alternative constructions produced by participants in the elicited production experiment. Each alternative construction elicited in the experiment is produced more frequently when weight increases. By providing evidence for processing factors in Cantonese, results in Cheung (2004, 2005) support the PGCH. Corpus data (see Cheung 2004) also suggest that heavy NPs in the dative construction sometimes shift in Cantonese. 90% of the bei2 datives are in the canonical order [bei2 T R], where the remaining 10% are in the HNPS order [bei2 R T]. The ratio of DOC to HNPS is very low where T is short (1:24 for themes of 2-3 syllables), but rises to 4:7 in the case of 4-5 syllables, and 2:1 in the case of T with 6-8 syllables. This trend is consistent with the predictions of the PGCH. Hypothesis on the “Chinese Problem”  PGCH (Hawkins 2004; see also Matthews & Yeung 2001; Cheung 2004, 2005) suggests that rearrangement and keeping relative clauses short allows a possible solution to the “Chinese Problem”. Also, alternative orders are preferred in degree of their efficiency. When an idea would be expressed inefficiently in the canonical order DOC (e.g. [bei2 [RC T] R]), the speaker uses an alternative order, such as the BA-construction [zoeng1 T bei2 R]. Combining CxG and the PGCH, and as a follow-up to Cheung (2004, 2005), this study considers two constructions in Cantonese, DOC and the zoeng1-construction. The two constructions have semantic differences (S. Cheung 1994), but not too different, and perhaps can be viewed in parallel to the “dative-alternation” in English 24 . In addition, even if a pronoun is used as the theme, the zoeng1-construction is grammatically acceptable (thus it is a better candidate than double-bei2, RD and extraposition from NP). HNPS is arguably not

                                                         As demonstrated in Goldberg (1995, 2006), the constructions [give R T] and [give T to R] are not derived from one another, as well as the pair load hay onto the truck and load the truck with hay, and active and passives. The term “dative-alternation” is used here for convenience only. 24

48 

4 Alternative constructions  an alternative construction, but an alternative ordering where the heavier element is displaced – this analysis is taken by CxG and HPSG.

4.2.5. The BA‐construction – the zoeng1‐construction in Cantonese  Peyraube (2004) reported that the BA-construction (把), which corresponds to the contemporary Cantonese zoeng1 將, appears in the form of [A ba T V R] around the sixth century in early Medieval Chinese. He hypothesized the construction to be a structure with ba as a preposition grammaticalized from serial verb constructions by analogy with the yi construction. A description of the BA-construction in Mandarin is provided by S. Cheung (1994) and it is described as “the disposal form”: ƒ

semantically, the construction indicates “disposal”;

ƒ

syntactically, the object has to be definite;

ƒ

it cannot be used for simple predicates, but OK if an aspectual marker is used;

ƒ

the BA-construction is stylistically preferable as it is ‘more natural and idiomatic to a native ear’ (S. Cheung 1994:459, 467-8) in certain cases, and it is even compulsory for ditransitives 25 , e.g. [BA T V PP], [BA [RC T] V]

Although S. Cheung (1994) assumed a transformational account, according to CxG (see Goldberg 1995, 2006), alternative constructions are not derived by transformation. It was also demonstrated in Goldberg (1995, 2006) that alternative orders such as DOC [give R T] and the prepositional dative [give T to R] in English are two separate constructions. The similar situation in which the BA-construction cannot be used for simple predicates provides an argument for the BA-construction and the canonical counterpart constructions to be different constructions. Other Chinese languages provide extra support for this argument for the BA-construction to have functional properties and form a family of constructions of their own. The BA-construction is not only attested in Mandarin, but in Taiwanese (Southern Min) as the ka-construction, and in Hakka as the lau-construction (Lai 2003). Lai (2003) noted that the Hakka lau-construction can be used for a variety of semantic roles, with lau functioning as a marker for patient, benefactive, goal, source, and comitative. It is also noted (Lai 2003) that these functions share a feature in common, namely that that more than one participant is inherently involved in the construction, following the observation made by Huang (1982) that the BA-construction occurs with a complex VP. Lai (2003) also noted that the lau-construction is often used with an action verb that causes the change of state in the lau phrase, and the result of change is indicated in the resultative complement. This patient

                                                         25

This point may indicate a processing preference (Matthews & Yeung 2001; see also Cheung 2004, 2005).

49 

4 Alternative constructions  function of lau is in accordance with the analysis of resultative construction in Goldberg (1995). Goldberg (1995:188) asserted that ‘resultatives can only be applied to arguments which potentially undergo a change of state as a result of the action denoted by the verb’. The statement coincides with the requirement of the lau-construction that signifies a theme undergoing change of state caused by the agent. The zoeng1(將)-construction is the Cantonese version of the BA-construction, with the order [zoeng1 T VP], and with a dative verb like bei2, it has [zoeng1 T bei2 R]. It is also attested in monotransitive verbs even when the direct object is a pronoun. Cheung (2005) suggested that it may be even less constrained than topicalization in this respect, not to mention that certain alternatives such as extraposition is impossible when the NP is a pronoun or a bare noun. Matthews and Yip (1994) observed that unlike the ba-construction in Mandarin, there are more restrictions to the zoeng1-construction in Cantonese, as it is not used with all transitive verbs, and is restricted to cases where motion takes place (also for abstract cases such as disposal). It is thus clear that at least for Cantonese, the canonical order and the zoeng1-construction are different constructions, consistent with CxG. To make the problem more complex, the syntactic category of BA (and the Cantonese zoeng1) is still debatable. This has implications to processing theories like EIC as the structure of the construction would be different depending on the category to which zoeng1 is assigned. Li and Thompson (1981) have noticed that BA cannot serve as a simple answer to questions, nor take aspect markers. This has perhaps influenced analyses in the literature which contains accounts of BA being a preposition or a case marker (Huang 1982), or being a light verb/co-verb/converb. However, there is a growing consensus for BA to be analyzed as a higher head (Paul 2002), either as a verb (e.g. Bender 2000) or as a functional category. Paul (2002) argued that the BA is a higher verbal head taking a VP-complement. In a comparison with verb-copying in Chinese (cf. Huang 1982), Paul (2002) explained her position as follows: [ba] is not a contentless F-category, but an intrinsically contentful head with features of its own. What appears on the surface as the object of ba in fact occupies the specifier position of the complement VP of ba, viz ba and the following [N]P do not form a constituent. …the [N]P generally called “ba-object” is selected and assigned the role of patient by the VP, not by ba…. The analysis proposed here contrasts sharply with the widespread assumption that ba is a preposition… (Paul 2002:162-3) It is also observed that agent-oriented adverbal expressions can only precede but not follow BA. Paul (2002), following Huang (1982), analyzed the VP in the BA-construction as a kind of complex predicate, and the object NP ‘occupies the specifier position in the VP-complement of [BA], this reflects the dependency of the [N]P on the complex predicate

50 

4 Alternative constructions  expressed by the verbal projection below [BA]’ (Paul 2002:168-9). Following this insight, I assume that BA (zoeng1 in Cantonese and ba in Mandarin) is a verb (V) which obligatorily takes both an object (T) and a complement VP (see Bender 2000). VP[BA

T VP]

Whether empty categories (e) exist is a complicating factor. Consider examples 36 and 37: 36. [In which box]i did you put the very large and beautifully decorated wedding cake bought from the expensive bakery ei? 37. ?# [Which box]i did you put the very large and beautifully decorated wedding cake bought from the expensive bakery in ei? (Gibson 2005a) Gibson (2005a) noted that if there are empty-categories mediating long distance dependencies, a distance-based complexity theory like the DLT does not predict the contrast between examples 36 and 37: it makes no predictions. However, MiD predicts that there is a filler-gap domain (FGD), and the gap will have to be co-indexed with the referent it represents (also in HPSG, see Sag et al. 2003). Such a long distance dependency is considered inefficient in processing, as acknowledged in Goldberg (2006). Citing observations on island constraints, she argued that processing accounts alone do not explain processing constraints on island extractions such as direct questions, or why it is that displacement of constituents, both across clause boundaries and unbounded ones (the recipient of the English dative), should present a problem to processing, while a constructional approach would generalize that backgrounded constructions are islands (BCI), and emphasized that it is ‘pragmatically anomalous to treat an element as at once backgrounded and discourse-prominent’ (Goldberg 2006:135). It is reasoned that the functions of constructions need to be recognized. This appears to be the case for the zoeng1-construction, as the object NP of the BA-construction cannot be extracted (see Paul 2002). In the case of the zoeng1-construction, one may not need to posit the existence of a FGH or e. One may also conclude that Wasow’s insight (2002) that multiple factors are interacting with each other is correct. Although processing factors are not widely posited within Chinese linguistics, an example of the BA-construction was mentioned in Hawkins (1990): 38. 我 把 你 給 我 的 書 丟掉 了 wo ba [ni gei wo de shu] diudiao le I BA you give me Prt book lose SFP Agent Theme V “I have lost the book you gave me” (Hawkins 1990:242) Arguably, such construction appears more like a SOV order than a SVO language. As evidenced by the more balanced distribution of RelN and NRel among SOV languages, SOV + RelN do not cause as much processing load as SVO + RelN (see Table 1 on page 27).

51 

4 Alternative constructions  Hawkins (1990) took it as an example supporting EIC. This appears to be one of the ways in which a SVO and RelN language can avoid centre-embedding, namely, to have a short object close to the verb (in front of the verb instead of after the other object NP) so that the LD of the V may be minimized. This also echoes the claim by Huang (1982) that Chinese restricts the number of elements that can occur postverbally: potential centre-embedding is difficult to produce and parse, and this is coherent with processing motivations affecting grammar (PGCH, Hawkins 2004). Kwan (2000) has established preliminary results based on the processing account for the motivation of the zoeng1-construction in Cantonese. EIC is used to explain the preference for Cantonese zoeng1-construction in sentences with heavy themes, as the construction is found to be processed more rapidly than SVO counterparts. In current terms (Hawkins 2004), MiD and MaOP would predict the BA-construction to have processing advantages in case of centre-embedding in the DOC. For unmodified themes and a pronominal recipient, MiD predicts that DOC is slightly more efficient than the BA-construction due to the extra word 26 : 39. DOC: VP[bei2 [CL T] R] 畀 [本 書] 我 bei2 [bun2 syu1] ngo5 PCD for VP: bei2 to ngo5 = 4 words LD for bei2 (underlined): bei2 to ngo5 = 4 words no PCD for embedded VP = 0 words TD: 8 words 40. BA: VP[zoeng1 [CL T] bei2 R] 將 [本 書] 畀 我 zoeng1 [bun2 syu1] bei2 ngo5 PCD for VP: zoeng1 to bei2 = 4 words LD for bei2 (underlined): syu1 to ngo5 = 4 words PCD for embedded VP: bei2 to ngo5 = 2 words TD: 10 words However, for modified themes, the head noun of the theme and the main verb would be closer in the BA-construction, forming a smaller LD for bei2, as in examples 41 and 42: 41. DOC: VP[bei2 [CL (XP) T] R] 畀 [本 得得 嘅 書] 我 bei2 [bun2 dak1ji3 ge3 syu1] ngo5 PCD for VP: bei2 to ngo5 = 6 words LD for bei2 (underlined): bei2 to ngo5 = 6 words no PCD for embedded VP = 0 words (constant) TD: 12 words 42. BA: VP[zoeng1 [CL (XP) T] bei2 R] 將 [本 得得 嘅 書] 畀 我 zoeng1 [bun2 dak1ji3 ge3 syu1] bei2 ngo5

                                                         26

Here, the term TD refers to the total size of the domains involved in the VP.

52 

4 Alternative constructions  PCD for VP: zoeng1 to bei2 = 6 words LD for bei2 (underlined): syu1 to ngo5 = 3 words (constant) PCD for embedded VP: bei2 to ngo5 = 2 words (constant) TD: 11 words With the addition of an AdjP consisting of 2 words (dak1ji3 ge3), Total Domain (TD) would be smaller for the BA-construction, as the RC increases both the PCD for the VP and the LD of bei2 in the DOC, while in the BA-construction only the PCD of the zoeng1 VP is increased. The trend favouring the BA-construction would continue when T has more words, such as being modified by an RC. However, if the PCD of the VP is very inefficient, such as with the introduction of a very complex RC (or many adjectives), then MiD would predict that the inefficient PCD would induce pressure for another alternative to be used (such as topicalization). In addition, according to MaOP, BA assigns its object as the theme; while in the double-object construction [V NP NP] there is the possibility of HNPS, or possible misassignment of the theme as the recipient. Examples 39 and 40 are repeated here for reference, and unassignment differences (see Hawkins 2004:56-7) can be calculated as in Table 8: 43. DOC: VP[bei2 [CL T] R] 畀 [本 書] 我 bei2 [bun2 syu1] ngo5 44. BA: VP[zoeng1 [CL T] bei2 R] 將 [本 書] 畀 我 zoeng1 [bun2 syu1] bei2 ngo5 “give me that book” DOC Categories Phrases Attachments 27 Relations 28 OP/UP ratio 29 BA Categories Phrases Attachments

bei2 V VP V[VP]

bun2 CL NP1 CL[NP1]

3/15 = 20% zoeng1 V VP1 V[VP1]

6/15 = 40% bun2 CL NP1 CL[NP1]

syu1 N N[NP1] NP1[VP] 9/15 = 60% syu1 N

ngo5 N NP2 N[NP2] NP2[VP] NP1=T-VP NP2=R-VP 15/15 = 100% Bei2 V VP2 V[VP2] VP2[VP1]

ngo5 N NP2 N[NP1] N[NP2] NP1[VP1] NP2[VP2] Relations NP1=T-VP NP2=R-VP OP/UP ratio 3/19 = 16% 6/19 = 32% 10/19 = 53% 14/19 = 74% 19/19 = 100% Unassignment 20-16 = 4% 40-32 = 8% 60-53 = 7% 100-74 = 26% +45%, DOC difference preferred Table 8 Unassignment difference between DOC and BA with an unmodified theme 

From the word-by-word calculation of unassignment difference, the DOC appears to be more efficient in terms of MaOP with an aggregate unassignment difference of 45%. However, for

                                                         X[Y] indicates that X is attached to Y. NP = X-VP indicates that the NP is assigned as the X argument role (agent, T, R, etc.) of the VP. 29 OP/UP ratio = UPs assigned online divided by the total number of UPs at each word 27 28

53 

4 Alternative constructions  the DOC, the assignment of T to syu1 has to wait until ngo5 is encountered, as there is the possibility of HNPS. As a result, the assignments of T to syu1 and R to ngo5 are both done at the final word in the DOC. For the zoeng1-construction, the assignment of syu1 as T can be done immediately as zoeng1 assigns T to its object, instead of waiting till ngo5 is encountered. A more aggressive prediction would be that when zoeng1 is encountered, the obligatory immediate constituents theme NP and VP would be constructed. The aggregate unassignment difference becomes -1% in favour of the zoeng1-construction, suggesting that the both constructions are of comparable efficiency in terms of MaOP, as in Table 9: However, this assumption is still speculative and requires further support. DOC OP/UP ratio BA Categories Phrases Attachments

bei2 20% zoeng1 V VP1 NP1 VP2 V[VP1] NP1[VP1] VP2[VP1]

bun2 40% bun2 CL

syu1 60% syu1 N

ngo5 100% bei2 V

ngo5 N NP2 CL[NP1] N[NP1] V[VP2] N[NP2] NP2[VP2] Relations NP1=T-VP NP2=R-VP OP/UP ratio 7/19 = 37% 9/19 = 47% 12/19 = 63% 14/19 = 74% 19/19 = 100% Unassignmen 20-37 = -17% 40-47 = -7% 60-63 = -3% 100-74 = 26% -1%, no t difference preference Table 9 Alternative analysis of unassignment difference between DOC and BA with an  unmodified theme 

Another point to note is that bei2 is polysemous and could mark the passive or mean allow, or give, thus bei2-phrases are prone to ambiguity. A verb embedded in an RC may be misassigned as the immediate VP daughter of the zoeng1-construction, leading to wasted efforts. However, it is more of a problem for the DOC as more reassignments have to be made at the end, when relevant syntactic or semantic cues are encountered. Embedded agents cause misassignment which can only be resolved late in the DOC, while the RC can be recognized early in the zoeng1-construction. Constraints of the zoeng1-construction may facilitate processing sometimes, as the embedded verb may be understood to belong to an RC when it could not be the VP of the zoeng1-construction. Preference for the zoeng1-construction would be even larger when there are more modifiers modifying the NPs, as embedded clauses would delay the assignment of argument roles. Consider examples 45 (DOC: VP[bei2 [N V Det CL T] R]) and 46 (BA-construction: VP[zoeng1 [N V Det CL T] bei2 R]) and the illustration of misassignment factors in Table 10: 45. 畀 bei2 give

[你 借 嗰 本 書] 我 [nei5 ze3 go2 bun2 syu1] ngo5 you borrow Det CL book me

54 

4 Alternative constructions  46. 將 [你 借 嗰 本 書] 畀 我 zoeng1 [nei5 ze3 go2 bun2 syu1] bei2 ngo5 BA you borrow Det CL book give me “give me the book you borrowed” DOC Categories Phrases Attachments

bei2 V VP1 V[VP1]

nei5 N NP1 N[NP1]

ze3 V VP2 V[VP2]

go2 Det NP2 Det[NP2]

3/29 = 10% zoeng1 V VP1 V[VP1]

6/29 = 21% nei5 N NP1 N[NP1]

9/29 = 31% ze3 V VP2 V[VP2]

12/29 = 41%

bun2 CL

syu1 N

ngo5 N NP3 S

CL[NP2]

N[NP2]

N[NP3] NP3[VP1] NP1[S] VP2[S] S[NP2] NP2[VP1] NP1=A-VP2 NP2=T-VP2 NP2=T-VP1 NP3=R-VP1

14/29 = 48% bun2 CL

16/29 = 55%

29/29 = 100%

syu1 N

CL[NP2]

N[NP2]

bei2 V VP3 V[VP3] VP3[VP1]

Relations

OP/UP ratio BA Categories Phrases Attachments Relations OP/UP ratio Unassignment difference

go2 Det NP2 S Det[NP2] NP1[S] VP2[S] S[NP2]

NP1=A-VP2 3/32 = 9% 10-9 = 1%

6/32 = 19% 21-19 = 2%

9/32 = 28% 31-28 = 3%

17/32 = 53% 41-53 = -12%

19/32 = 59% 48-59 = -11%

ngo5 N NP3 N[NP3] NP3[VP3] NP3=R-VP3

NP2=T-VP2 NP2=T-VP3 23/32 = 72%

27/32 = 84%

55-72 = -17%

100-84 = 16%

32/32 = 100% -18%, BA preferred

Table 10 Unassignment difference between DOC and BA with RC‐modified theme 

The aggregate unassignment difference is -18.10%, indicating that the zoeng1-construction would be preferred over the DOC since centre-embedding severely delays the recognition of the DOC. In the DOC, until the recipient ngo5 is parsed, bei2 nei5 ze3 bun2 syu1 may be parsed as a VP with bei2 meaning “allow”, so the string of words might be misread as “let you borrow one book”. The disambiguating word appears at the very end of the sentence, being the recipient ngo5, and the recognition of the correct structure is drastically delayed. On the other hand, the RC can be reliably recognized at the classifier bun2, thanks to the knowledge of the zoeng1-construction as [zoeng1 NP VP]. Alternative constructions – different predictions  “Traditional” accounts would predict that alternative constructions will be dispreferred; while PGCH would predict that alternative constructions that have better TDDs will be preferred. These predictions are assumed to hold in performance in both language comprehension and production.

55 

5 Experimental study 

5. Experimental study  Research done on processing of datives (and sentence processing in general) includes eye-movement studies, corpus study, experimentation using reaction time as a measure (such as reading time studies), various methods tapping speech production (such as elicited imitation and elicited production) or even a mixture of the above (Levelt 1989; Hawkins 1994; V. Ferreira 1996; Stallings et al. 1998; Sedivy et al. 1999; Smith & Wheeldon 1999; Arnold et al. 2000; Pearlmutter & Gibson 2001; Yamashita & Chang 2001; Yamashita 2002; Wasow 2002; F. Ferreira 2003). As mentioned in the previous sections, using Cantonese in sentence processing is beginning to receive research attention (Kwan 2000; Matthews & Yeung 2001; Cheung 2004, 2005; Matthews et al. 2005; cf. Mandarin RCs, Hsiao & Gibson 2003). However for datives, while many studies are done on English, only a production study in Cheung (2004, 2005) is found in the Chinese or Cantonese literature.

5.1. Methodology  To further investigate the PGCH, comprehension data is also necessary. Consequently, a dual-task paradigm is adopted here to investigate both comprehension 30 and production. To allow for easier comparison with results from previous studies, the masked moving-window reading paradigm is followed. For production, the elicited imitation paradigm is selected. On one hand, the task procedures used here should yield results similar to those from elicited production in Cheung (2004, 2005). Bock (1996) also provided support to this kind of “Simply Recall” tasks, as a great deal of difficulty to recall the precise wording is perceived by people in general, yet people have very little difficulty remembering the gist of what they have just heard [or seen]. This task also allows messages that are more abstract and elaborate than those that can be used in description of pictures or videos. Participants  18 university students at The University of Hong Kong aged between 18 and 24 with Cantonese as their L1 were recruited for the experiment, all of whom agreed to participate voluntarily. None had explicit knowledge or training in psycholinguistics or syntax. Fixed variables  Working memory (WM) capacity would be constant for each participant. As each participant would comprehend an item and reproduce it, the amount of cognitive resources in the WM available to the participant would be more or less constant. This allows a comparison of the

                                                         Ideally, an eye-tracking paradigm might reveal more by providing both reading times and subjects’ regressions, but such equipment is not available.

30

56 

5 Experimental study  two tasks to be done in a more direct way than to compare different performance across participant. Manipulated independent variables  Two variables are manipulated for the stimulus sentences: ƒ

Theme complexity: the modifier of the theme (heaviness) with three levels:

ƒ

[T]: bare T NP of 3 words [Det CL N] 嗰本書 that book • [Adj T]: adjective-modified T of 5 words [Det CL Adj ge N] 嗰本得得嘅書 that funny book • [RC T]: NP, modified by RC and Adj, of 7 words 31 [[A V] Det CL Adj ge N] 佢借緊嗰本得得嘅書 that funny book he is borrowing Construction: the construction used – both constructions are grammatical even for •

pronoun as theme: • •

the canonical order DOC [bei2 T R] the zoeng1-construction [zoeng1 T bei2 R] (coded as BA)

Items  12 sets of stimulus sentences, each with six variations according to the manipulated variables described above. The different sets all use the verb bei2, a simple recipient NP [CL N], and the head Ns are used in the theme and recipients are familiar objects. To reduce ambiguity and garden-path effects, the recipient is always a person and the theme is made as inanimate as possible. Other than items 8 and 9 where the recipients are of a proper name nature and have no CL, the number of words is the same for the corresponding conditions across items 32 . The different item sets use similar NPs as agents and recipients, and that would result in a lower probability for participants to recognize that the elements were seen in another condition of the same item set shown previously. Item set 1 is listed here as an illustration. 47. Item set 1 DOC [T]: 個 細細仔 畀咗 [嗰 本 書] 個 先生 go3 sai3lou6zai2 bei2.zo2 [go2 bun2 syu1] [go3 sin1saang1] BA [T]: 個 細細仔 將 [嗰 本 書] 畀咗 個 先生 go3 sai3lou6zai2 zoeng1 [go2 bun2 syu1] bei2.zo2 [go3 sin1saang1] (both “The kid has given that book to the teacher”) DOC [Adj T]: 個 細細仔 畀咗 [嗰 本 得得  書] 個 先生 go3 sai3lou6zai2 bei2.zo2 [go2 bun2 dak1ji3 ge3 syu1] [go3 sin1saang1] BA [Adj T]:

                                                         To distinguish between the two levels [Adj T] and [RC T] in terms of the number of words, it is decided that the [RC T] would be an object-extracted RC by adding an embedded agent and V to [Adj T]. 32 Also, the number of characters is controlled to be within the same range. 31

57 

5 Experimental study  個 細細仔 將 [嗰 本 得得  書] 畀咗 個 先生 go3 sai3lou6zai2 zoeng1 [go2 bun2 dak1ji3 ge3 syu1] bei2.zo2 [go3 sin1saang1] (both “The kid has given that funny book to the teacher”) DOC [RC T]: 個 細細仔 畀咗 [佢 借緊 嗰 本 得得  書] 個 先生 go3 sai3lou6zai2 bei2.zo2 [keoi5 ze3.gan6 go2 bun2 dak1ji3 ge3 syu1] [go3 sin1saang1]

BA [RC T]: 個 細細仔

將

[佢 借緊

嗰 本

得得  書]

畀咗 個 先生

go3 sai3lou6zai2 zoeng1 [keoi5 ze3.gan6 go2 bun2 dak1ji3 ge3 syu1] bei2.zo2 [go3 sin1saang1]

(both “The kid has given that funny book he is borrowing to the teacher”)

Balancing the item conditions and fillers:    The 6 conditions are distributed equally among participants by the experiment program, Linger 33 version 2.94. All participants see two conditions of the 6 variations, seeing each construction once with either bare T, modified by an adjective [Adj T], or an RC [RC T]. Complexity is manipulated by setting the program to show one theme complexity per construction for every set item, effectively splitting the participants randomly into three groups. Each participant sees the same number of the six conditions, although the conditions come from different item sets. Also, each condition of every set of items is seen by the same number of participants. This is due to feedback from the pilot experiment when the pilot participant reported that there is a possible learning effect after a condition of the same item set has been seen previously. On the other hand, the lower frequency of the zoeng1-construction in Cantonese would be compensated for by the use of fillers, which also serve to disguise the task nature. Stimuli from Matthews et al. (2005), which are topicalized OSV sentences and their SVO canonical counterparts, are used as fillers (henceforth fillers). Fillers are also in 3 levels of theme complexities. 96 additional fillers (henceforth filler-fillers) are also taken from Cheung (2004) and Matthews et al. (2005). The items are listed in the Appendix, starting on page 89. To reduce ordering effects, the program also automatically shuffles the stimuli and the fillers and therefore each participant receives the items in a different order, but the program always keep items, fillers and filler-fillers mixed. In other words, while filler-fillers can appear after one another, no test item will appear after another item and no filler will appear after another filler item. 11 more items are drafted as practice items. Procedures  The experiment is done in a quiet room on a Sony Vaio PCG-Z1GP laptop which has a 14.1” SXGA+ Colour TFT display. The participants performed the experiment one by one as to avoid different computers being a confounding factor. They are told to sit comfortably at a

                                                         Linger version 2.94 is developed by Doug Rohde. It can be easily customized to use different settings, such as running a self-paced or auto-paced reading or listening tasks. Other advantages include being able to show Chinese characters and that it can be run across different platforms. It can be obtained at http://tedlab.mit.edu/~dr/Linger/. 33

58 

5 Experimental study  distance so that they can read the sentences on LCD screen the way they are accustomed to. They are also told that they should read and understand the sentences presented to them word by word at their normal reading speed, while responding to comprehension questions and probes for sentence recall as quickly and accurately as possible. A hint is also given by the experimenter, stating that remembering the contents (instead of rote memory) will be helpful in the performance of the tasks. After the instructions are read on screen, 11 practice items are shown to allow participants to get used to the experiment. The experiment would start only after the experimenter is confident that the participants are accustomed to the moving-window display and that they respond properly to the comprehension questions and the recall cue. In each trial, they have to press the spacebar to see the next word. After pressing the spacebar, _

_

_

_ _

_

_

_

_

_

_ _

_

_

_

_

_

_ _

will become 個 _

_

_ _

which after another press of the spacebar, will become _ 細 細 仔 _

_

_

_ _

_

_

_

and so on, until every word of the sentence is shown. Figure 5 shows a simulated screen 34 showing the DOC [T] condition for item set 1 (example 48) with all Chinese characters masked, that is, before the participant presses the spacebar. 48. 個 細細仔 畀咗 嗰 本 書 個 先生 go3 sai3lou6zai2 bei2.zo2 go2 bun2 syu1 go3 sin1saang1 Det kid give-PFV Det CL book Det teacher “The kid has given that book to the teacher”

                                                         The words are separated in the examples, but in written Chinese (including Cantonese), words and characters are not separated by spaces. 34

59 

5 Experimental study            _

_

_

_ _

_

_

_

_

_

_ _

Figure 5 A simulation of screen output of a masked sentence 

For stimuli and some fillers, a recall probe [請重複一次?] (cing2 cung4fuk1 jat1 ci3 “Please recall (the sentence)”) would appear on the screen, and the participant would recall the sentence. The participants are told that the recall cue occurs randomly, but is in fact coded in the experiment settings (see the Appendix). This recall forms the production element of the experiment. This formed the dual-task nature of the experiment as the participants were reproducing sentences (a production task) while reading (a comprehension task). An Apple iPod attached with a Griffin iTalk is used to record the utterance reproduced by the participants. This is arguably better than recording with the computer due to two reasons: voice recording might hinder the accuracy of RT measured by the computer, and the sound recorded by the iPod appears to be better than that recorded by the computer with a microphone. For certain fillers, there were comprehension questions to monitor whether the participants were paying attention, and that would keep the participants’ attention to the materials instead of merely pressing the spacebar automatically. A response is given when the answer is wrong, but no positive responses are given. The comprehension questions also maintained the validity of the comprehension data, since if the comprehension questions are not correctly answered, it is unlikely that the participant is paying attention to the task. In between sentences, the participants could rest if they wanted to, but none took breaks since the time required for the experiment is short and was between 12 and 20 minutes.

60 

5 Experimental study  Dependent variables observed  ƒ Comprehension: reading time (RT) is collected as Linger is programmed to record the RT (in milliseconds) for each window (of Cantonese word), the time between a screen (of a word in a sentence especially) is shown on screen and a key is pressed; ƒ

Comprehension question accuracy;

ƒ

The construction produced in elicited production;

ƒ

General accuracy in elicited imitation (Accuracy), also both the accuracy of recalling the construction (Structure) and the sentence elements (Information), e.g. NP arguments.

A summary of the hypotheses:  On language comprehension: the zoeng1-construction alleviates poor IC-to-word ratio (according to MiD) and will show a length/heaviness effect especially for the VP with theme modified by relative clauses; and the construction would be read faster than DOC, with the increasing weight of the theme NP. On language production: production errors will show a length/heaviness effect and a construction effect, sensitive to both factors, with the zoeng1-construction being more accurately recalled than the DOC regardless of length, and sentences with heavier theme NPs experiencing lower accuracy. On language processing as a whole: the RT in comprehension task should be positively correlated with the rate of production errors (longer RT would reflect that a structure is more difficult to parse, and this in turns should result in a higher error rate), the cognitive resources for language processing being constant because it is the same person. (Section 5.5 will be on the hypothesis that the processing mechanism in language comprehension and production are related.) Pilot studies  As the experiment is of a novel dual task nature, pilot tests were carried out to evaluate whether participants would employ strategies in reading in order to remember the sentences. The reading data appeared to be largely similar to those from single-task studies, revealing that participants do not seem to be using specific strategies to memorize the sentence. One pilot participant reported that there is a possibility of learning effect, thus the task was modified so that participants saw only 2 conditions per item set instead of seeing all conditions of all item sets. It should be also noted that the production task is unlikely to be confounded since in production research (Bock p.c.), as all participants need to see the sentence they have to recall, in any case of elicited imitation.

61 

5 Experimental study 

5.2. Comprehension results  A moving window task allows the reading time for each word to be recorded. However, as no online metric for efficiency taking all of MiD, MiF and MaOP into account is provided by Hawkins (2004), the predictions of the hypotheses for comprehension data are tested separately for the RTs averaged across words in the same sentence. Treatment of data:  Initially, a 3 standard deviation (SD) cut-off was used for the reading times (RT) of each word in each condition (Construction * Theme complexity) for each participant, but large numbers of items had reaction times that were above the 3SD cut-off in certain sentence positions, especially the last word in the sentence (2.7% of the data, 16% of the sentence-final words). Upon examination, this appeared to be caused by a slower reading time in general for such syntactic positions, in accordance with the predictions of EIC and DLT, as the processing demands are different for each word. Therefore, the SD cut-offs were recalculated based on the RT for each word by each participant separately. Data cut off are not included in the calculations of averages and sentence reading times. The average RTs of each word in the sentences are provided below in Figure 6. Average reading time per window

1200 1000 800 RT (ms)

600 400 200 0

ACL

AN

V1

TA

TV

TDet TCL TAdj Tge

TN

(V2) RCL

RN

Average RT 382.1 404.6 398.7 411.4 477.5 411.2 390 444.5 419.9 432.4 391.8 399.6 1118

Syntactic categories

Figure 6 Average RTs per word over all conditions of stimulus sentences

The overall average RT for each word is 478.11 ms. As shown in Figure 6, the average reading time for the words in the sentences is about 409.36 ms for the non-final words, but for the last word the RT is systematically slower with an average of 1117.83 ms, with a

62 

 

5 Experimental study  standard deviation of 1785.43 ms. The distribution of RTs for the final word is plotted in Figure 7 and the statistics for the distribution are provided in Table 11:

0 800

2000 3200 4400 5600 6800 8000 9200 10400 1160012800 14000 1520016400 1760018800

Figure 7 Frequency distribution of RTs for the sentence‐final words 

Moments Mean Std Dev Std Err Mean upper 95% Mean lower 95% Mean N Sum Weight Sum Variance Skewness Kurtosis CV N Missing

Quantiles 100.0% 99.5% 97.5% 90.0% 75.0% 50.0% 25.0% 10.0% 2.5% 0.5% 0.0%

maximum

quartile median quartile

minimum

RT (ms) 19075 15358 6197 2072 1084 602 413 285 216 188 182

1117.8329 1785.4312 86.606133 1288.0638 947.60212 425 425 475079 3187764.5 5.6112084 41.195169 159.72254 7

Table 11 Statistics for the distribution of RTs for sentence‐final words 

The range of RTs for the sentence-final words spans from 182 ms to 19075 ms, but slow sentence-final RTs over 6000 ms are not outside the 3SD cut-offs and thus not considered outliers. The particular distribution is reflected in the positive skewness (5.61) and large kurtosis of (41.20). This calls for an explanation but they are consistently produced by the reading behaviour of the same participants. Besides, the non-final words and the final word would be averaged separately instead of averaging the RTs of all the words in the whole sentence, as the conditions have a different number of words.

63 

5 Experimental study 

5.2.1. Statistical analysis  Repeated measures ANOVA is used on participant means (F1) and item means (F2). The confidence interval is set at 95%. Average RTs for non sentence‐final words  The construction is a significant factor affecting the average RT for non sentence-final words in the participants analysis: F1(1,17)=4.933, p=.040; F2(1,33)=1.243, p=.241. The complexity of the theme NP is also a significant factor in the participants analysis: F1(1.293,21.973)=5.246 after the Greenhouse-Geisser correction, p=.024; F2(2,33)=1.245, p=.301. No interactions are found significant. Possible reasons for both factors to be non-significant in the items analysis might be the large variance in the participants’ individual reading times or the small number of items (N=12). The averages for non sentence-final words are tabulated in Table 12: RT (ms) [T] [Adj T] [RC T]

DOC 407.55 398.44 438.25

BA 387.34 389.97 422.59

Table 12 Average RTs (ms) for non sentence-final words in 6 conditions

Average RTs for sentence‐final words  Due to the large variation in the distribution of the RTs, a transformation of ln(RT) (natural log of RT) was performed. This allows the statistical model assumptions to be better followed. The construction is a significant factor affecting the average RT for sentence-final words in the participants analysis: F1(1,17)=6.205, p=.023; F2(1,33)=0.685, p=.414. The complexity of the theme NP is also a significant factor in the participants analysis: F1(2,34)= 7.88, p=.002; F2(2,33)=2.192, p=.128. No interactions are found to be significant. The averages for sentence-final words are tabulated in Table 13: RT (ms) [T] [Adj T] [RC T]

DOC 1110.46 1291.43 1311.86

BA 853.60 906.51 1236.32

Table 13 Average RT (ms) for sentence‐final words in 6 conditions 

5.2.2. Discussion and qualitative analysis  Although Hawkins (2004) provided no combined metric for online performance, reading times may reveal evidence of processing factors. The reading times for the condition bare theme NP are plotted in Figure 8:

64 

5 Experimental study  RT (ms) 450 440 430 420 410 400 390 380 370 360 350

ACL

AN

V1

TD

T CL

TN

DOC 384.9296 405.2254 401.8429 421.1944 390.8169 421.5571 BA

(V2)

RCL

RN

430.8333 1110.465

373.5286 391.7826 395.2647 375.8451 374.0857 410.8592 385.9429 392.2881 853.5972

Word

 

Figure 8 Reading times for bare theme NP conditions [T] 

V1 refers to bei2 in the DOC construction and zoeng1 in the BA-construction; while V2 is bei2 for the BA-construction, it does not exist for the DOC and is reflected by the broken line at the position marked V2. While theories based on canonical word order would predict the DOC to be read faster, MiD predicts that the BA-construction would not suffer a slower RT, as both constructions have a TD of 12 (refer to section 4.2.5, Table 8 to Table 10 starting on page 53, for the calculation of TDs). The data here is in favour of MiD over the other predictions, although the superior performance of the zoeng1-construction is not predicted. Comparing the TDs of the DOC and the zoeng1-construction, the difference increases when a modifier of the theme in DOC is centre-embedded, resulting in a TD of 16 for the DOC and 14 for the zoeng1-construction. This is reflected by the slightly slower RTs for the conditions [Adj T] (see Figure 9). The reading times for the conditions [Adj T] show less difference between the two lines but the BA-construction is still read faster than the DOC on average. In the theme NP, the DOC is slowed down to a larger degree with respect to the zoeng1-construction. In particular, the function words (Det, CL and ge3) are read faster when compared to other words, but the particle ge3 in the DOC is slow while the main verb bei2 is fast: the slow ge3 may be due to the accumulation of processing load in the theme NP, while the verb bei2 in the zoeng1-construction is read quickly perhaps due to the fact that the syntax of the BA-construction requires an obligatory verb and the construction of the VP node is facilitated.

65 

5 Experimental study  RT (ms)

450 440 430 420 410 400 390 380 370 360 350

ACL

AN

V1

TD

T CL

T Adj

T ge

TN

DOC 377.264 424.577 402.394 368.343 364.563 429.565 414.408 413.029 BA

(V2)

RCL

RN

391.593 1291.43

384.352 395.535 400.875 367.211 376.275 418.871 385.029 412.901 376.549 380.086 906.514

Word

Figure 9 Reading times for adjective‐modified theme NP conditions [Adj T] 

The zoeng1-construction is also read faster in the condition [RC T] in Figure 10. RT (ms) 550

500

450

400

350

ACL

AN

V1

TA

TV

TD

TCL TAdj

Tge

TN

DOC 374.6 391.8 398.1 426.4 486.9 497.3 402.3 472.8 460.2 500.3 BA

398.1 418.3 393.3 396.4 468.1 442.6 432.2 455.9 418.9

437

(V2) RCL

RN

410.4 1312 412.4 391.7 1236

Word

 

Figure 10 Reading times for the heavily‐modified theme NP conditions [RC T] 

The fact that the RTs are in general slower can be attributed to the fact that the TDs are longer than in the less complex conditions. The TD for the DOC is 20, while the TD is 16 for the zoeng1-construction. The impact of centre-embedding can be seen in both constructions at the embedded RC, where the RTs for the embedded verb are increased to over 460 ms. As more arguments are held in the working memory, the RTs are slower. More remarkably, the sentence-final RTs for the zoeng1-construction experienced a marked increase (from 906.51 ms in [Adj T] to 1236.32 ms in [RC T]), while there is no marked increase for the DOC

66 

5 Experimental study  construction over the three theme complexity conditions. The most plausible explanation is that the gain in TD by using the zoeng1-construction is offset by the increasing length of the PCD of the zoeng1 VP. Although the LD of bei2 is minimized, the distance between zoeng1 and the head noun of the theme also creates centre-embedding. While additional materials such as an RC increase the PCD for the VP in the zoeng1-construction, the processing load is increased. The longer sentence-final RT in the DOC and the [RC T] condition of the zoeng1-construction may well be a reflection of VP/S integration time (see Table 10), when the syntactic structure is finalized and the argument roles mapped to the NPs for the sentence meaning. In short, the processing factors proposed by Hawkins (2004) are supported by the Cantonese comprehension data.

5.3. Production results  Elicited imitation (simple recall) is performed after the appearance of the recall cue when the final word of the sentence is read by the participant. The responses are recorded on an iPod and later transcribed and coded. Treatment of data:  The responses of the elicited production task are transcribed and coded according to whether the sentence is recalled successfully (Accuracy). Further subscores are calculated based on whether the construction is preserved (Structure), and whether the sentence elements are preserved (Information). The analysis of the production task is straightforward. As no RT is measured, only the participants' responses are categorized and errors are noted and coded. Accuracy of the recall is measured. Two other measures are included: Structure for the correct construction used (1 for correct, 0 for error) and Information for correct information repeated (1 for correct, 0 for error). The types of responses include errors of various types such as forgetting elements of the sentence (nouns or embedded verb) or mixing up the arguments (scoring 0 in Information and Accuracy); and using a non-target construction (scoring 0 in Structure and Accuracy): using the BA-construction when the input is DOC, producing a DOC when the input is the BA-construction, the double-bei2 construction [bei2 T bei2 R] and other errors in sentence structure, such as recalling the RC and the main clause as a conjoined clause, or sometimes admitting failure. Restarts and repairs are tolerated but only severe problems are recorded as errors in the analysis if information is missed or the structure differs from that of the input.

67 

5 Experimental study 

5.3.1. Statistical analysis  Repeated measures ANOVA is used on participant means (F1) and item means (F2). The confidence interval is set at 95%. Overall accuracy  The BA-construction enjoys higher accuracy than the DOC, and the heavier the modifier of the theme, the lower the accuracy. Both factors are highly significant in both the by-participants and the by-items analyses. The construction is a significant factor affecting the overall recall accuracy in both the participants and items analyses: F1(1,17)=7.713, p=.013; F2(1,33)=27.484, p<.001. The complexity of the theme NP is also a significant factor in the participants and the items analyses: F1(1.375,23.367)=27.628 after the Greenhouse-Geisser correction, p<.001; F2(2,33)=69.172, p<.001. No interactions are found significant. The overall accuracy of the responses for the different conditions is shown in Table 14: Accuracy (%) [T] [Adj T] [RC T]

DOC 80.56% 69.44% 36.11%

BA 94.44% 93.06% 56.94%

Table 14 Overall accuracy (%) of the responses in the 6 conditions 

Structure (accuracy in the choice of construction)  The zoeng1-construction is repeated with higher accuracy than the DOC, and the heavier the modifier of the theme, the lower the accuracy. The construction is a highly significant factor affecting the overall recall accuracy in both the participants and items analyses: F1(1,17)=11.858, p=.003; F2(1,33)=42.049, p<.001. Meanwhile, the complexity of the theme NP is also a significant factor in both the participants and the items analyses: F1(1.509,25.653)=10.95 after the Greenhouse-Geisser correction, p=.001; F2(2,33)=9.603, p=.001. No interactions are found significant. The accuracy in the choice of construction in the responses for the different conditions is shown in Table 15: Structure (%) [T] [Adj T] [RC T]

DOC 80.56% 72.22% 54.17%

BA 95.83% 94.44% 86.11%

Table 15 Accuracy of repetition (%) in the choice of construction in the 6 conditions 

Information (accuracy in the repetition of sentence elements)  The more complex the modifier of the theme NP, the less accurate are the sentence elements repeated. The complexity of the theme NP is a significant factor in both the participants and the items analyses: F1(1.083,18.417)=26.296 after the Greenhouse-Geisser correction, p<.001; F2(2,33)=76.125, p<.001. Performances for the different constructions do not differ

68 

5 Experimental study  significantly. Construction is a non-significant factor in both the participants and items analyses: F1(1,17)=0.035, p=.854; F2(1,33)=0.032, p=.859. No interactions are found significant. Such results are consistent with the predictions, as there are more arguments in more complex theme NPs, while for less complex NPs the number of arguments may not exceed processing limitations. The accuracy in the repetition of sentence elements for the different conditions is shown in Table 16: Information (%) [T] [Adj T] [RC T]

DOC 100% 95.83% 65.28%

BA 98.61% 98.61% 65.28%

Table 16 Accuracy (%) in repeating the sentence elements in the 6 conditions 

5.3.2. Discussion and qualitative analysis  Responses can be separated into 4 main categories: ƒ

canonical double object construction DOC [bei2 T R];

ƒ

the BA-construction: [zoeng1 T bei2 R];

ƒ

the double-bei2 construction [bei2 T bei2 R] (see Tang 1998; Cheung 2004, 2005);

ƒ

other variants and errors

Generally speaking, the participants responded more accurately when the theme is less modified, and the zoeng1-construction is more accurately recalled as the double-bei2 construction is used in 13.89% (unmodified theme NP) to 23.61% (when the theme is modified by an RC) of the responses. Sometimes the zoeng1-construction is recalled as the DOC or a double-bei2 construction. This supports the use of the elicited imitation paradigm as it echoes the claim by Bock (1996) that people are good at remembering the meaning of a message but not the exact wordings. This also suggests that it is not likely that the participants are using rote-memory to perform the task, thus providing support to the validity of the RT data. The distribution of the responses is listed Table 17: Construction used in responses for condition (%)

DOC

BA

double-bei2

Others

Subtotal

DOC [T] DOC [Adj T] DOC [RC T] BA [T] BA [Adj T] BA [RC T]

58 (80.56%) 52 (72.22%) 39 (54.17%) 2 (2.78%) 3 (4.17%) 3 (4.17%)

4 (5.56%) 7 (9.72%) 12 (16.67%) 69 (95.83%) 68 (94.44%) 62 (86.11%)

10 (13.89%) 12 (16.67%) 17 (23.61%) 1 (1.39%) 1 (1.39%) 3 (4.17%)

0 (0%) 1 (1.39%) 4 (5.56%) 0 (0%) 0 (0%) 4 (5.56%)

72 72 72 72 72 72

Table 17 Distribution for type of responses to the 6 conditions 

69 

5 Experimental study 

As can be seen in the statistics for the response accuracy of structure, there is a higher accuracy for Structure (correct structures are bolded in Table 17) of the zoeng1-construction. Also, the implications from the frequency of zoeng1-construction in different theme NP complexities are valuable. While a higher complexity of the theme reduces the high accuracy of the task performance, the zoeng1-construction with its theme NP modified by an RC is still recalled more accurately than the canonical bare-T-NP DOC. This suggests that the zoeng1-construction has some advantage over the DOC, and following the predictions of MiD, the short LD of bei2 (4 words starting from the head noun of the theme NP to the head noun of the recipient NP) in the BA-construction would be at least a word shorter than the LD of bei2in the DOC as it spans from bei2 to the head noun of the recipient (6 words). The shifting pattern is compared with Cheung (2004, 2005), where participants were cued to repeat the semantics (sentence meaning instead of exact wordings) after being exposed to pre-recorded auditory stimuli in only the canonical order. Cheung (2004, 2005) can be viewed as a more naturalistic study of language production, and the comparison may answer some questions regarding whether the instruction to repeat the whole sentence is comparable to just the reproduction of the contents of the message, as well as determining the effects of syntactic priming. Data from Cheung (2005) is also listed for comparison in Table 18: Frequency (%) in condition DOC with short T DOC with long T DOC with complex T

DOC 94.05% 81.75% 67.06%

BA 0% 0.40% 5.95%

[V T bei2 R] 5.56% 13.49% 20.24%

Others 0.40% 4.37% 6.75%

Table 18 Distribution of responses to DOC input obtained in Cheung (2005) 35

With regard to the canonical DOC, the decrease in the frequency of using the same construction with theme complexity is similar across both studies. This suggests that the production data is valid and is not affected by the dual-task design. However, there seems to be a priming effect to a certain extent, as demonstrated by the higher frequency of the DOC in the participants’ responses in Cheung (2004, 2005). This suggests that follow-up studies to Cheung (2004, 2005) should consider using more filler items with various possible constructions, such that the priming effect of the canonical order of the stimuli and filler items does not carry over to other constructions. The occurrence of the double-bei2 construction is argued, following Cheung (2004, 2005), to be the evidence for the claim that there is a limited tolerance of the human processor to

                                                         It should be noted that while in Cheung (2004, 2005), [T] corresponds to short and [Adj T] corresponds to long, [RC T] has different structures than complex T, which includes various kinds of RCs. 35

70 

5 Experimental study  inefficient structures (Hawkins 2004). The increasing rate of using both the zoeng1-construction and the double-bei2 construction when the DOC suffers from heavier centre-embedding shows that a more efficient order would be tolerated by MiF, even though MiF favours the use of less lexical material for more syntactic operations and argument role assignments. The other type, being an error, also illustrates that limited cognitive resources and processing factors are correlated.

5.4. Further statistical investigations  In sections 5.2 and 5.3, the factors investigated include construction and theme complexity with 3 categorical types [T], [Adj T] and [RC T]. As these NPs are well-controlled in terms of word length, it is valid to test whether Theme length (number of words in the theme) is a significant factor, in place of the categorical Theme complexity (see page 57). Example 47 is repeated for convenience: 49.

Item set 1 DOC [T]: 個 細路仔 畀咗 [嗰 本 書] 個 先生 go3 sai3lou6zai2 bei2.zo2 [go2 bun2 syu1] [go3 sin1saang1] BA [T]: 個 細路仔 將 [嗰 本 書] 畀咗 個 先生 go3 sai3lou6zai2 zoeng1 [go2 bun2 syu1] bei2.zo2 [go3 sin1saang1] (both “The kid has given that book to the teacher”) DOC [Adj T]: 個 細路仔 畀咗 [嗰 本 得意 嘅 書] 個 先生 go3 sai3lou6zai2 bei2.zo2 [go2 bun2 dak1ji3 ge3 syu1] [go3 sin1saang1] BA [Adj T]: 個 細路仔 將 [嗰 本 得意 嘅 書] 畀咗 個 先生 go3 sai3lou6zai2 zoeng1 [go2 bun2 dak1ji3 ge3 syu1] bei2.zo2 [go3 sin1saang1] (both “The kid has given that funny book to the teacher”) DOC [RC T]: 個 細路仔 畀咗 [佢 借緊 嗰 本 得意 嘅 書] 個 先生 go3 sai3lou6zai2 bei2.zo2 [keoi5 ze3.gan6 go2 bun2 dak1ji3 ge3 syu1] [go3 sin1saang1] BA [RC T]: 個 細路仔 將 [佢 借緊 嗰 本 得意 嘅 書] 畀咗 個 先生 go3 sai3lou6zai2 zoeng1 [keoi5 ze3.gan6 go2 bun2 dak1ji3 ge3 syu1] bei2.zo2 [go3 sin1saang1]

(both “The kid has given that funny book he is borrowing to the teacher”)

As seen in the example, [T] consists of 3 words, [Adj T] of 5 words, and [RC T] of 7 words. Having the theme NP controlled for word length, TDs can also be calculated as follows (see also section 4.2.5) 36 : DOC [T]: 畀咗 [嗰 本 書] 個 先生 個 細細仔 go3 sai3lou6zai2 bei2.zo2 [go2 bun2 syu1] [go3 sin1saang1]

                                                         Item sets 8 and 9 use a proper name instead of a [CL N] NP, resulting in a word less in the recipient NP. The resulting PCDs and TDDs are given in the brackets. 36

71 

5 Experimental study  PCD for VP: bei2 to sin1saang1= 6 words (5 words); LD for bei2 (underlined): bei2 to sin1saang1 = 6 words (5 words); PCD for embedded VP: no embedded VP = 0 words TD: 12 words (10 words) BA [T]: 個 細細仔 將 [嗰 本 書] 畀咗 個 先生 go3 sai3lou6zai2 zoeng1 [go2 bun2 syu1] bei2.zo2 [go3 sin1saang1] PCD for VP: zoeng1 to bei2 = 5 words; LD for bei2 (underlined): syu1 to sin1saang1 = 4 words (3 words); PCD for embedded VP: bei2 to sin1saang1 = 3 words (2 words) TD: 12 words (10 words) DOC [Adj T]: 個 細細仔 畀咗 [嗰 本 得得 嘅 書] 個 先生 go3 sai3lou6zai2 bei2.zo2 [go2 bun2 dak1ji3 ge3 syu1] [go3 sin1saang1] PCD for VP: bei2 to sin1saang1= 8 words (7 words); LD for bei2 (underlined): bei2 to sin1saang1 = 8 words (7 words); PCD for embedded VP: no embedded VP = 0 words TD: 16 words (14 words) BA [Adj T]: 個 細細仔 將 [嗰 本 得得 嘅 書] 畀咗 個 先生 go3 sai3lou6zai2 zoeng1 [go2 bun2 dak1ji3 ge3 syu1] bei2.zo2 [go3 sin1saang1] PCD for VP: zoeng1 to bei2 = 7 words; LD for bei2 (underlined): syu1 to sin1saang1 = 4 words (3 words); PCD for embedded VP: bei2 to sin1saang1 = 3 words (2 words) TD: 14 words (12 words) DOC [RC T]: 畀咗 [佢 借緊 嗰 本 得得 嘅 書] 個 先生 個 細細仔 go3 sai3lou6zai2 bei2.zo2 [keoi5 ze3.gan6 go2 bun2 dak1ji3 ge3 syu1] [go3 sin1saang1] PCD for VP: bei2 to sin1saang1= 10 words (9 words); LD for bei2 (underlined): bei2 to sin1saang1 = 10 words (9 words); PCD for embedded VP: no embedded VP = 0 words TD: 20 words (18 words) BA [RC T]: 個 細細仔 將 [佢 借緊 嗰 本 得得 嘅 書] 畀咗 個 先生 go3 sai3lou6zai2 zoeng1 [keoi5 ze3.gan6 go2 bun2 dak1ji3 ge3 syu1] bei2.zo2 [go3 sin1saang1] PCD for VP: zoeng1 to bei2 = 9 words; LD for bei2 (underlined): syu1 to sin1saang1 = 4 words (3 words); PCD for embedded VP: bei2 to sin1saang1 = 3 words (2 words) TD: 16 words (14 words) One has to bear in mind that the use of TD implies that one has to assume a certain syntactic structure. However, the definition of syntactic categories in Chinese is unclear at times, such as the case of zoeng1. The choice of assuming zoeng1 to be a verb is arguably conservative, as the structure [V NP VP] implies a PCD of the VP from the verb zoeng1 to the embedded verb bei2, but a short analysis assuming that zoeng1 is an object marker is also presented to demonstrate that the BA-construction is still favourable in terms of MiD: DOC [T]: 個 細細仔 畀咗 [嗰 本 書] 個 先生 go3 sai3lou6zai2 bei2.zo2 [go2 bun2 syu1] [go3 sin1saang1] PCD for VP: bei2 to sin1saang1= 6 words; LD for bei2 (underlined): bei2 to sin1saang1 = 6 words; PCD for theme NP: go2 to syu1 = 3 words TD: 15 words BA [T]: 個 細路仔 將 [嗰 本 書] 畀咗 個 先生 go3 sai3lou6zai2 zoeng1 [go2 bun2 syu1] bei2.zo2 [go3 sin1saang1]

72 

5 Experimental study  PCD for VP: zoeng1 to sin1saang1 = 7 words; LD for bei2 (underlined): syu1 to sin1saang1 = 4 words; PCD for theme NP: zoeng1 to syu1 = 4 words TD: 15 words DOC [Adj T]: 個 細細仔 畀咗 [嗰 本 得得 嘅 書] 個 先生 go3 sai3lou6zai2 bei2.zo2 [go2 bun2 dak1ji3 ge3 syu1] [go3 sin1saang1] PCD for VP: bei2 to sin1saang1= 8 words; LD for bei2 (underlined): bei2 to sin1saang1 = 8 words; PCD for theme NP: go2 to syu1 = 5 words TD: 21 words BA [Adj T]: 個 細細仔 將 [嗰 本 得得 嘅 書] 畀咗 個 先生 go3 sai3lou6zai2 zoeng1 [go2 bun2 dak1ji3 ge3 syu1] bei2.zo2 [go3 sin1saang1] PCD for VP: zoeng1 to sin1saang1 = 9 words; LD for bei2 (underlined): syu1 to sin1saang1 = 4 words; PCD for theme NP: zoeng1 to syu1 = 6 words TD: 19 words DOC [RC T]: 個 細細仔 畀咗 [佢 借緊 嗰 本 得得 嘅 書]個 先生 go3 sai3lou6zai2 bei2.zo2 [keoi5 ze3.gan2 go2 bun2 dak1ji3 ge3 syu1] [go3 sin1saang1] PCD for VP: bei2 to sin1saang1= 10 words; LD for bei2 (underlined): bei2 to sin1saang1 = 10 words; PCD for theme NP: keoi5 to syu1 = 7 words TD: 27 words BA [RC T]: 個 細細仔 將 [佢 借緊 嗰 本 得得 嘅 書] 畀咗 個 先生 go3 sai3lou6zai2 zoeng1 [keoi5 ze3.gan2 go2 bun2 dak1ji3 ge3 syu1] bei2.zo2 [go3 sin1saang1]

PCD for VP: zoeng1 to sin1saang1 = 11 words; LD for bei2 (underlined): syu1 to sin1saang1 = 4 words; PCD for theme NP: zoeng1 to syu1 = 8 words TD: 23 words The above calculations show that the difference between the TDs for the two constructions does not change significantly when an alternative view of the syntactic category of zoeng1 is taken. Hence, it is assumed that zoeng1 is a verb, following the analysis of the Mandarin ba in Bender (2001). Theme length and TD of the 6 conditions are presented in Table 19: Condition DOC [T] DOC [Adj T] DOC [RC T] BA [T] BA [Adj T] BA [RC T]

Theme length (words) 3 5 7 3 5 7

TD (see fn 36, page 71) 12 (10) 16 (14) 20 (18) 12 (10) 14 (12) 16 (14)

Table 19 Theme lengths and TDs of the 6 conditions 

Further statistical testing was performed to investigate whether TD affects language comprehension and production. Simple ANOVA was performed on the comprehension data and nominal logistic regression was performed on the production data. As the individual RTs show huge variation especially in the sentence-final word, analysis was performed on individual observations instead of the usual practice to analyze the participants and items

73 

5 Experimental study  data separately as in psycholinguistics. It is argued that averaging over participants and over items will lead to data loss due to individual variation. For the comprehension data, the RTs for the average non sentence-final RTs and average sentence-final RTs are also separated, following section 5.2. The production data analyzed are the factors Accuracy, Structure and Information. As previously explained, there is huge variation for the RTs, and some transformation of the raw data is required to fulfil the model assumption of equal variance. The reciprocal of the RTs is chosen and it appears that the model assumptions are satisfied. To start with, the substitution of Theme complexity by Theme length is analyzed: Average RTs for non sentence‐final words  There are significant individual differences, F=49.718, p<.0001, and there is also a group effect, F=55.061, p<.0001. Theme length is a significant factor, F=8.642, p=.0035, and construction is also a significant factor, F=4.194, p=.0412. Item is not found to be significant, F=0.768, p=.672. Average RTs for sentence‐final words  There are significant individual differences, F=43.020, p<.0001, and there is also a group effect, F=89.033, p<.0001. Theme length is a significant factor, F=11.645, p=.0007, and construction is marginally a insignificant factor, F=3.831, p=.051. Item is not found to be significant, F=0.453, p=.931. Overall accuracy  The individual differences between participants are significant, χ2=48.244, p<.0001, but there is no group effect, χ2=2.259, p=.323. Theme length is a significant factor, χ2=59.123, p<.0001, and construction is also significant, χ2=27.202, p<.0001. Item is not a significant factor, χ2=9.145, p=.609. Structure (accuracy in the choice of construction)  The individual differences are significant, χ2=32.454, p=.0056, but there is no group effect, χ2=0.0697, p=.966. Theme length is significant, χ2=39.332, p<.0001, and construction is also significant, χ2=19.053, p<.0001. Item is not a significant factor, χ2=7.770, p=.734. Information (accuracy in the repetition of sentence elements)  The individual differences are significant, χ2=29.936, p=.012, but there is no group effect, χ2=0.402, p=.818. Theme length is a significant factor, χ2=41.809, p<.0001. Construction is not a significant factor, χ2=0.287, p=.592. Item is also not significant, χ2=9.791, p=.549.

74 

5 Experimental study  The manipulation of the theme NP seems to result in slower RTs and response accuracy across the two constructions except for Information, where there is no effect of construction but the length of the theme NP is still a significant factor. The predictions of MiD are supported. Further analysis by using TD as calculated above instead of Construction and Theme length (or Theme complexity) as two separate variables is given below. Average RTs for non sentence‐final words  There are significant individual differences, F=49.501, p<.0001, and there is also a group effect, F=54.820, p<.0001. TD is a significant factor, F=10.029, p=.0017. Item is not found to be significant, F=0.950, p=.492. Average RTs for sentence‐final words  There are significant individual differences, F=43.325, p<.0001, and there is also a group effect, F=89.844, p<.0001. TD is a significant factor, F=17.367, p<.0001. Item is not found to be significant, F=0.448, p=.934. Overall accuracy  The individual differences are significant, χ2=47.576, p<.0001, but there is no group effect, χ2=1.940, p=.379. TD is a significant factor, χ2=66.808, p<.0001. Item is not a significant factor, χ2=12.416, p=.333. Structure (accuracy in the choice of construction)  The individual differences are significant, χ2=30.775, p=.0094, but there is no group effect, χ2=0.067, p=.967. TD is a significant factor, χ2=37.642, p<.0001. Item is not a significant factor, χ2=11.062, p=.438. Information (accuracy in the repetition of sentence elements)  The individual differences are significant, χ2=27.622, p=.024, but there is no group effect, χ2=0.157, p=.925. TD is a significant factor, χ2=46.176, p<.0001. Item is not a significant factor, χ2=13.254, p=.925. The results using TD as a factor instead of both construction and length/complexity of the theme NP converge with the previous analyses. The results are also consistent with the predictions made by the hypotheses. By obtaining similar results, it also appears that the use of TD is fruitful, supporting Hawkins (2004). Other alternative constructions  Citing Matthews & Yeung (2001), Hawkins (2004:209) commented that ‘Cantonese preposes a RelN in its VO structure in proportion to the complexity of the relative and the strain on the PCD for VP’ The stimuli of Matthews et al. (2005) which include topicalized OSV and canonical SVO sentences are used in this experiment as fillers. As mentioned in

75 

5 Experimental study  section 5.1, these items are scattered between the DOC, zoeng1-construction and filler-fillers and they are seen by the participants. Analysis was performed on these items, but statistics reveal that only theme complexity was a factor affecting average RTs for the sentence-final word. However, a qualitative examination of the reading times showed consistency with the results from Matthews et al. (2005). It is predicted that with increased number of items and participants, the effects would be significant. However, as the focus of the investigation is the processing advantages of the zoeng1-construction, the potentials of topicalization as an alternative construction that may facilitate processing would have to be left for another study.

5.5. Correlation between comprehension and production  In the experimental study, presumably the same person has the same amount of cognitive resources (including WM) and the conditions of the items should pose the same processing difficulty across the sets of stimuli. Once the data has been cleaned up, the RTs can be plotted against the shifting frequencies. As demonstrated by the analyses in sections 5.2 and 5.3, performance of both language comprehension and production are affected by processing factors. RT is slower when syntactic complexity increases, and increase in complexity corresponds with higher probability of using an alternative construction. However, to investigate the correlation between language comprehension and production, it is necessary to control for syntactic complexity. As mentioned earlier in section 2.3, it is predicted that the same person has the same language faculty and is under the same WM constraints. If the two tasks language comprehension and production are handled by separate systems, then partialling out the factors affecting performance would allow a deeper understanding of how processing factors affects language processing. A correlation would lend support to Hawkins (2004), where processing principles MiD, MiF and MaOP are claimed to be equally applicable to both language comprehension and production. Reading difficulty, as reflected in slower RT and a higher tendency to use a different construction in recall, is to be partialled out. Partialling out the processing factors would investigate how language comprehension and production are related. This is done as follows. A partial correlation is performed on the metrics measured in comprehension and those in production, with the following factors partialled out. ƒ

Total Domain size of the stimuli (TD);

ƒ

NP length of the theme (number of words);

ƒ

Average RT for each participant (ms) (as each participant reads in a different speed)

76 

5 Experimental study  As Structure is contributing factor to Accuracy, a significant correlation between nonFinalRT and Structure would imply that the partial correlation between nonFinalRT and Accuracy to be significant too. The statistics are summarized in Table 20, and it is found that there are significant correlations between the following pairs of metrics: ƒ

Average RT for non sentence-final words (nonFinalRT) and Structure;

ƒ

Average RT for sentence-final words and Structure;

ƒ

nonFinalRT and Accuracy

Correlation [Significance (two-tailed), Structure Information Accuracy degrees of freedom = 420] Average RT for non sentence-final words 0.182 (p < 0.001) -0.026 (p > 0.05) 0.143 (p = 0.003) Average RT for sentence-final words -0.141 (p = 0.004) 0.046 (p > 0.05) -0.087 (p > 0.05) Table 20 Partial correlations between language comprehension and production metrics 

With the reading speeds of individuals and the complexity of the stimuli controlled, Accuracy is correlated with nonFinalRT, suggesting that the more time it takes, the higher the accuracy. This is not very surprising. However, Structure is differently correlated with the average RTs of non sentence-final (0.182) and sentence-final (-0.141) words, which suggests that while reading the sentence more careful reading (slower) will be more accurate, the sentence-final effect which is associated with a higher processing load (discussed in section 5.2) holds for both tasks – a slower sentence-final reading time is correlated with a lower probability in the same construction being recalled. This is predicted by the PGCH – a sentence which is difficult to parse would be difficult to reproduce. Although correlations of slightly over ±0.14 are not high, it is understandable as the two metrics are of a different nature, one being a measurement of time and the other one being a percentage (Accuracy). There is potentially a task effect too, as comprehension and production have rather distinct goals. If response latency is measured, the correlation is predicted to be higher.

5.6. General discussion  The better performance of the zoeng1-construction with respect to the canonical DOC order is reflected in the faster reading time per word in the masked self-paced reading task and a higher accuracy in the elicited imitation task. The fact that the zoeng1-construction is preferred despite being a non-canonical word order calls for an explanation. The comprehension data and the production data are consistent with the predictions of the PGCH, and are different from the predictions of the “traditional” theories. As spelt out in section 4.2.4, certain theories predict that canonical orders would be processed with the best performance, but results from this study provided evidence against

77 

5 Experimental study  such accounts. Instead, the results are in accordance with the predictions by the PGCH (Hawkins 2004), as the principles MiD and MaOP predict that the zoeng1-construction is processed more efficiently when compared to the canonical DOC, as the DOC poses a higher processing load with centre-embedding. First, the LD for the verb bei2 is shorter in the zoeng1-construction than the DOC, favouring MiD. MaOP also predicted that the zoeng1-construction to be more efficient as the argument roles could be immediately assigned to the theme and recipient when they are encountered (rather than at the end as in the DOC), resulting in a faster recognition of the semantic structure. Also, in production the zoeng1-construction is also predicted to be advantageous in production. The head-final NP of Cantonese is similar to a Japanese NP, having RelN, and as the long-before-short preference in Japanese (Yamashita & Chang 2001; Yamashita 2002) having the NP preposed would avoid centre-embedding by allowing the ICs of the VP to be closer, which results in a smaller LD of the verb. A heavier element is also more salient in production as it attracts attention by having more information. In production, MiD and MaOP provide an advantage to the zoeng1-construction, so that it is recalled with a higher accuracy. The production data also echoes with the comprehension data. The theme complexity effect (predicted by MiD) supports the claim that when a particularly low efficiency is reached, an alternative is chosen instead of the canonical order in order to improve efficiency (Hawkins 1994, 2004; cf. Cheung 2004, 2005). The experimental results suggest that processing factors (such as MiD and MaOP) may be important in the choice of construction used in a sentence (see Hawkins 2004). This support may be of importance to the PGCH, as the exceptional configuration of Chinese languages appeared to be problematic to Hawkins (1994, 2004), but the treatment of Chinese may be inadequate in Hawkins (1994, 2004). By establishing the effects of processing factors affecting also other constructions in Cantonese, the rare configuration of Chinese is not as much a challenge to processing theories in Hawkins (2004) anymore. The analysis of the filler items for the effects of topicalization provided additional support.

5.6.1. The canonical double object construction (DOC)  Although the DOC is the canonical construction for dative expressions, it experiences less favourable results in the experiment. A sensible question to ask at this stage, then, would be why the DOC could be the canonical expression at all in the first place, if it is dispreferred in terms of processing? First of all, one must consider that in natural situations, relative clauses are not necessarily used. As a result, the inefficient order [V RC O] does not arise every time there is a VP. Moreover, if the object NPs are not modified, the efficiency of the construction

78 

5 Experimental study  is still high [V [N] [N]]. When compared to the verb-copying construction ([V T V R], the double-bei2 construction), the DOC is more “economical” and is one word shorter than most alternative constructions such as the zoeng1-construction, and would be preferred by MiF. Tang (1998) postulated that the DOC might arise from the omission of the V2 bei2 in [bei2 T bei2 R], with the assumption that the former dative marker bei2 is replaced by a null marker. It might also be related to the change of the V2 from gwo3 (過) to bei2 as the dative V2, and bei2 might be omitted due to avoidance of unnecessary repetition, in accordance with the predictions of MiF. It would perhaps be reasonable to suggest that when V2 is changed to bei2, MiF is in force and leads the dative to be expressed as [V T R] instead of [V T V R] because the construction also encodes the dative meaning, just as the BA-construction encodes the disposal meaning. The DOC is only dispreferred when it is inefficient, which is the case when there are modifiers before the head nouns. The DOC being the canonical order is not implausible, if two things are taken into account: (i) the relatively low frequency of such modifiers leading to long or complex NPs, and (ii) the fact that the amount of working memory the average person has allows for a certain degree of inefficiency in language processing. Regarding the “Chinese Problem”, although some factors may have led to the exceptional word order in Cantonese (and Chinese), it is suggested that the existence of alternative constructions/orders may contribute to reduce the processing load, in addition to the avoidance of lengthy descriptions such as RCs. Another result obtained from this study is that non-canonical constructions, such as topicalization, may facilitate processing, besides making fine distinctions of meaning or providing pragmatic effects. Different structures may have entered Cantonese by creativity of language or language contact (for example, see account of ditaxia in Cantonese relative clauses, Matthews & Yip 2001), thus creating a variety of possible strategies, even with inefficient canonical orders – people mostly use short objects, especially pronouns, when the NPs are already established as given information 37 , while alternative constructions offer a way-out to centre-embedding.

                                                         Pinker (1994) points out that the human parser avoids heavy constituents in production; for example, a passive is more easily processed by the parser than a heavy noun phrase (NP), even if it is stylistically dispreferred. 37

79 

6 Implications and conclusion 

6. Implications and conclusion  6.1. Possible follow‐ups  To support the validity of the data obtained in this study, a possible adaptation would be the employment of the Rapid Successive Viewing Paradigm (RSVP). As a passive reading paradigm, it is slightly more natural and fairly similar to how we read subtitles on the television, for example. However, in that way the online data of comprehension is sacrificed. Another possible alternation is to change the modality of the comprehension task to be auditory. A self-paced moving window paradigm would still provide online processing data, while a RSAP (with auditory instead of viewing) is even more natural as it is exactly how we listen to others’ speech. Auditory presentation has also the advantage of being feasible for experimentation on children. One could also compare results from such variations with the results of Cheung (2004) where the presentation of the elicited production task is auditory but the recall is not required to be a repetition of the stimuli. Following the idea in Cheung (2005), another possible extension is to perform a similar task on children or the elderly, who have lesser cognitive resources as brain development in children is not yet complete and the elderly experience a reduced pool of available resources. They would be expected to have longer delays in RT and a higher frequency to shift to an alternative construction when processing difficulty increases. This also follows the finding that the individual participant is a significant factor in predicting variance in response. One possible reason is that each individual is different in terms of his/her language. Another would be that each individual might have a different level of cognitive resources available, such as working memory. A systematic testing of a person’s digit span, normal reading speed or reading behaviour (e.g. frequency of regressions in reading) may also be helpful. Further studies may further support this idea, including the alternative constructions of topicalization and verb-doubling. Extraposition of modifiers from NP in German and extraposition in Cantonese may also be similar as both languages have potential centre-embedding (following Hawkins 2004). Apart from looking at languages of different typological type, another potential testing area of the PGCH lies in historical linguistics. The examination of word order in relation to heaviness of the object in preverbal and postverbal objects in Archaic or Medieval Chinese, the analysis of parsing efficiency in structures now unattested in Contemporary Cantonese, etc. may turn out to support the PGCH.

80 

6 Implications and conclusion  Also, as Matthews and Yip (2001) report, certain constructions in Cantonese are ‘losing ground to head-final Mandarin-like constructions … the result is the progressive eradication of the distinctive features of Cantonese grammar through convergence with Mandarin’, (p. 270) – studies of a longitudinal nature could support the hypothesis if results reveal how certain structures crucial to keeping certain expressions of idea efficient are retained in the grammar. Replications of this study, whether on the methodology or in Chinese dative constructions, would be much welcomed. Early deployment of the dual task methodology similar to this study was fairly unsuccessful, according to Bock (p.c.). Improvements to the techniques necessary for conducting a dual-task experiment on language comprehension and production would provide much insight to the architecture of our language processing machinery. Furthermore, experimental studies of dative constructions in Cantonese or in other Chinese languages have been rare. Studies of the on-line processing of Chinese RCs (Hsiao & Gibson 2001) have obtained results contradicting the NPAH (Keenan & Comrie 1977). Further studies on the dative, especially the effects of centre-embedding, would inform typology and functional linguistics enormously, as Chinese languages has a typologically marked order. Various alternative orders remain untested, while studies on topicalization have obtained slightly differing results. For example, while Matthews and Yeung (2001) have obtained a weight effect of the theme NP, weight effect is not found significant in Matthews et al. (2005) 38 . Instead, there appears to be a general processing advantage for topicalization (Matthews et al. 2005), a point also raised by Hawkins (2004) which suggests topic-before-predication is beneficial to the human processor in terms of MaOP. Further studies with better control of the different factors would provide more information of how these factors interact in topicalization, for the processing of Cantonese in general, or even provide insight to the architecture of the language faculty. One of the alternative constructions that are worth investigating is the verb-doubling order, which seem likely to be affected by processing factors (see the sensitivity of double-bei2 construction to theme complexity, Tang 1998). The verb-doubling construction is sometimes even obligatory with fairly complex VPs in Cantonese, for example, 50. 我 洗 呢 件 衫 洗咗 兩 次 ngo5 sai2 ni1 gin6 saam1 sai2-zo2 loeng2 ci1 I wash this CL shirt wash-PFV two times “I’ve washed this shirt twice” (Matthews & Yip 1994:142)

                                                         Ching (p.c.) reports that: (i) while Matthews et al. (2005) reported no overall weight effect, there seems to be a significant interaction between weight and the participant; and (ii) the structure (topicalized OSV vs. canonical SVO) is a significant factor for the heaviest theme NPs. 38

81 

6 Implications and conclusion  51. 我 睇 書 睇 到 瞓着咗 ngo5 tai2 syu1 tai2 dou3 fan3zoek6-zo2 I read book read until fall-asleep-PFV “I fell asleep (as a result of) reading” The obligatory use of verb-doubling construction may be a result of performance factor affecting grammar, as extra complements (for example loeng2 ci1 “twice” in example 50) extend the PCD of the VP, and reduplicating the verb would result in 2 smaller VPs, which can be considered as a divide-and-conquer strategy, and is probably a more efficient construction following MiD, with a higher aggregate IC-to-word ratio (Hawkins p.c.). Further evidence supporting the PGCH may include the vanishing of certain constructions across time – these constructions may not facilitate processing or provide semantic or pragmatic distinctions, as well as a smaller variety of strategies being available to languages considered “more efficient” in processing terms. Languages with freer word orders have lesser available constructions. A follow-up would be to investigate changes of preference in the choice of the canonical order especially when typological changes have occurred, for example, comparing the frequency of the BA-construction with reference to theme NP length or its complexity. The findings from such studies would provide different kinds of new evidence supporting current models or demand changes to existing models to accommodate the differences in accordance with the PGCH by Hawkins (2004), thus improving our knowledge of language processing in the human mind.

6.2. Summary  Despite claims that different constructions exists because of semantic-pragmatic factors such as Chappell (2001) that object preposing is a contrastive device, or the proposal that speakers want to make finer conceptual semantic distinctions such as to disambiguate potentially ambiguity or to stress certain participant roles (see Bodomo et al. 2003), it is plausible to say that different strategies may serve Cantonese by compensating for the inefficient canonical order in Hawkins’ (1990; 1994) terms. Such alternative strategies also exist in English Wasow (2002). In terms of typology, the seemingly strange configuration of Chinese (and Cantonese) may be explained by areal features and language contact (such as an account of ditaxia in relative clauses, Matthews & Yip 2001). Different alternative constructions may have entered Cantonese in terms of language contact or may have been coined by the creativity of language, thus creating a variety of possible strategies, even with the inefficient canonical order with the unique typological configuration. Alternative structures can be

82 

6 Implications and conclusion  strategies to resolve the “Chinese Problem”, as shown by experimental studies in the earlier section (see also Matthews & Yeung 2001; Cheung 2004, 2005; Matthews et al. 2005). This could explain why the canonical double object order in Cantonese remains in existence. In other words, the fact that the Chinese configuration, SVO and RelN, is typologically marked and extremely rare among the languages of the world is no longer an unsolved problem to the PGCH. On the contrary, this study provided additional support to the PGCH by demonstrating that the alternative orders provide a way to alleviate processing load. In other words, this study provides experimental support to Hawkins (2004): Extraposition from NP [in German]… can reduce lengthy domains in the basic orders of VP. Similarly, RelN languages, motivated by MiD, tolerate misassignments and unassignments only to a certain extent, allowing MaOP to assert itself in the form of numerous otherwise unexplained typological properties. (Hawkins 2004:208)

6.3. Implications  As Hawkins (2004) admits, it is a truism that one cannot prove a theory right, but one may gather supportive evidence which cannot be predicted by alternative theories. By examining the questions raised in this study, more understanding of the language faculty may be gained: How do processing factors such as usage affect grammar?  A basis of Construction grammar is that constructions can be learned based on input, with children being able to make fast generalizations from the usually skewed input (Goldberg 2006). CxG also contends that surface grammar may be sufficient for language processing, as sometimes constructions may not specify that a position of the construction be occupied by a certain syntactic category. This approach appears to converge with the position of Hawkins (2004), where a minimal set of assumptions of structure is assumed – this also allows Hawkins (2004) to apply the performance model to different frameworks, such as HPSG. PGCH (Hawkins 2004) also assumes that grammar is shaped by convention, so that processing factors play a large role in grammaticalization. Combining the two views, it appears that the blurry syntactic categories in fact do not trouble Chinese speakers – as long as the phrases are processed correctly, the syntactic category of the word BA may not matter, because when the human processor encounters zoeng1, it can reliably predict that the BA-construction is used and an NP and a VP are upcoming. As the literature suggests, BA is undergoing grammaticalization and its syntactic category is under debate, but such functional words at the very least supply the human processor with information about the sentence and facilitate processing. Constructions that have their own meaning and category, and that would include pragmatic, semantic and phonological characteristics. The results

83 

6 Implications and conclusion  obtained in this study also suggest that it is time for theoretical linguistics to reconsider processing factors, as functional words with blurry syntactic categories have been demonstrated to affect processing efficiency. With an additional source of information, theoretical linguistics would benefit from the extra dimension, and greater collaboration between theoretical and functional linguistics would contribute even more to linguistics and the understanding of the human mind. Is there a cost associated with choosing an alternative?  While it is observed that displaced constituents across clause boundaries place a burden on WM (cf. Goldberg 2006:154), little support to the account that alternative constructions pose extra processing constraints is obtained so far, as the zoeng1-construction is not read slower than the DOC in the experimental study (see section 5.2). Also, Matthews and Yeung (2001) and Matthews et al. (2005) expected additional cost for topicalization but did not find any processing penalty for it. Contrary results to the cost account are obtained by Matthews et al. (2005), with results supporting a general processing advantage for topicalization. This would also appear to be at odds with traditional views that alternative constructions are derived from a canonical order by movement (using terms such as shifting), or that non-canonical constructions suffer from extra processing costs because they are less frequently used. A conclusion would be that surface-oriented approaches make better predictions and this could inform theoretical linguistics. If the PGCH is correct, then the alternative orders may function as syntactic strategies that benefit communication, in addition to the semantic and pragmatic factors traditionally associated with the use of alternative constructions. How are comprehension and production related? Are the two tasks handled by the  same components of our neural machinery?    Experimental results in this study obtained a correlation between the two tasks, suggesting that even though the tasks have very different goals, the tasks are correlated, suggesting an overlap of the processing machinery used in both tasks. Imaging studies may provide the necessary evidence to this theory, as suggested by Stowe et al. (2005). However, this area of research perhaps require better technology, as unintrusive brain scanning techniques such as fMRI and PET can still improve, while psycholinguistic experiments might also benefit from more creativity, careful control and more advanced experimental design.

6.4. Final conclusions  Language is evidently affected by a variety of factors, with processing factors demonstrated to play a role in the choice of constructions. This is rather intuitive because speakers have an arsenal of constructions to choose from, even for more rigid languages such as English and

84 

6 Implications and conclusion  Chinese. Functional forces in communication undeniably demand grammars to provide ways to provide semantic and pragmatic distinctions, and in addition, language processing, being constrained by limitations of our cognitive resources and architecture, is also a driving force in the shaping of grammars. Haspelmath (2002) has made an analogy with how people treat their money – on one hand, people often spend money for reasons of survival, but on the other hand, people are trying to save money in general. Spending cognitive resources to spell out words facilitate clarity of expression, but people also tend to minimize the difficulty so as to save up cognitive resources. Even on this seemingly simple decision on whether to spell out more or say less, competing motivations exist. This may also explain why different researchers looking at various aspects of language have obtained diverse results. For this reason, considering a problem from a variety of perspectives may well provide converging evidence. By assembling insights from different aspects of linguistics, more powerful generalizations that can encompass different aspects of language and cognition may be developed, providing another big step forward in cognitive science.

85 

Appendix 

Appendix  Settings of the experimental study  A. Settings in file preferences  set LangEncoding big5 encSource preferences2

B. Settings in file preferences2  set TextFont "-family {細細細_HKSCS} -size -36 -weight bold" set MaskColor grey set MaskJoins 0 set MaskSpaces 0 set BreakInterval 1000 set GoKey space set NoKey f set YesKey j set QuestionKeys "錯: \"[string toupper $NoKey]\". 對: \"[string toupper $YesKey]\"." set PreDelay 1000 set RightAnswer "" set WrongAnswer "錯, 邀請心"

C. Settings in file introduction  instruct {多多你多多呢個多多 你將你你你你上你到一你你你文你你你 但但 D 你文中你中中中中中, 好好好： ---------------每一你你你將你每每一每每你文中你你你上每每每每 當你當 SPACEBAR 時, 你你你你到你一每你文中 邀你請請請請請請請你請請細請請每你文中, 然然當 SPACEBAR 你看一每你文中 邀請請當 SPACEBAR 直直你請直你你中.} instruct {當你你請一你你你然, 你你你你要你你一你你於請你你你你你你你. 如你如如你如你對, 邀當 "YES" (J) 如你如如你如你錯, 邀當 "NO" (F) 你你請先將你你你你於 J 與 F 鍵上. 邀請請請請請你你.} instruct {你你上螢你你每每 "邀請請一次" 邀每聲將直你你中請請一次, 然然當 "YES" (J) 多多你實實你中實實

86 

Appendix  邀把你請請請你上 如如你要, 你你請你每一你你你開你你你你你息} instruct {請看但多多你以要: 1. 你你上一 D 你文你你中中中中中 你要當 SPACEBAR 令你你請睇到 D 你文中 2. 邀請請請請請你請請請請請, 睇請睇細請請你中 3. 你請直你你中然, 如時你如你你你. 邀當 對(J) 或 錯(F) 如如你你錯, 你你你請螢你, 請讀一 D 請請心一 D 4. 如時你你你要有請請你你一次. 邀請請請請你你一次, 然然當 (J) 表表請每 5. 請電你電電電表看一你你你 多多請實時, 你你將請英文表表多多請實. 邀請請多多請請請. 如如如你, 而而你請而多多請請請請每. 首先你如一首首請, 等你你請等請多多你等等.} command {首請開你...} createItem {# practice 1 而而|但|一|你|首請} presentItem practice:1:command {你第你...} createItem {# practice 2 今次|你你|你然|如|一|你|如你 ? 對實對如對如你呢? N} presentItem practice:2:createItem {# practice 3 而而|你|你請|睇到|你錯|你|點 ? 邀請得當錯(F)? Y} presentItem practice:3:instruct {跟中你你你你跟跟好多多用你你你... 如時你你你要有請請你你一次...} command {開你...} createItem {# practice 4 我|想|請|海海海海|玩} presentItem practice:4:createItem {# practice 5 佢|唔唔得|食|生如|架 ? 邀請請一次? Y} presentItem practice:5:createItem {# practice 6 我|搵|如|D|相|俾|你|睇} presentItem practice:6:-

87 

Appendix 

createItem {# practice 7 你|還還|食食|如|D|錢|俾|我 ? 邀請請一次? Y} presentItem practice:7:createItem {# practice 8 聽日|你你|你|落落 ? 天文天但唔但請聽日你好天? N} presentItem practice:8:createItem {# practice 9 如|本|書|我|覺得|幾好 ? 邀請請一次? Y} presentItem practice:9:createItem {# practice 10 做請|功功|佢|就|走|到|海海|玩} presentItem practice:10:createItem {# practice 11 個|present|悶得悶，|請請|都|走走 ? 個 present 如有有? N} presentItem practice:11:instruct {首請請實. 如如如你, 而而你請而多多請請請請每. 你你請開你.} command {多多請等開你...}

D. Settings in file items  Each item was tagged in the following format: # construction item_number condition Word@Tag | Word@Tag|... ? Comprehension question or probe? Correct answer

To specify the constructions the sentences are using, the following codes are used to code sentences with the following abbreviations DOC BA TOPc TOPt filler

Double-object construction zoeng1-construction Non-topicalized SVO sentence (from Matthews et al. 2005) Topicalized OSV sentence (from Matthews et al. 2005) Filler sentences (filler-fillers as described in section 5.1)

The condition position is checked by the program when it creates a random list of conditions from each item set. a b

unmodified NP NP modified by an adjective

88 

Appendix  c -

NP modified by an RC fillers: all filler items are shown

The tags used in coding the syntactic category of the Cantonese words used in the experiment are as follows: ACl AN V B TA TV TD TCl TAdj Tge TN RCl RN

Classifier of agent NP Head noun of agent NP Main verb (bei2 for experiment items) zoeng1 Embedded agent for theme NP Embedded transitive verb for theme NP Demonstrative go2 for theme NP Classifier of theme NP Adjective modifying theme NP Link particle ge3 for adjective of theme NP Head noun of theme NP Classifier of recipient NP Head noun of recipient NP

An example of a tagged set of item 1 is provided below: # antonio1 1 a 個@ACl|細細仔@AN|俾俾@V|如@TD|本@TCl|書@TN|個@RCl|先生@RN ? 邀請請一次? Y # antonio2 1 a 個@ACl|細細仔@AN|將@B|如@TD|本@TCl|書@TN|俾俾@V|個@RCl|先生@RN ? 邀請請一次? Y # antonio1 1 b 個@ACl|細細仔@AN|俾俾@V|如@TD|本@TCl|得得@TAdj|你@Tge|書@TN|個@RCl|先生@RN ? 邀請請一次? Y # antonio2 1 b 個@ACl|細細仔@AN|將@B|如@TD|本@TCl|得得@TAdj|你@Tge|書@TN|俾俾@V|個@RCl|先生@RN ? 邀請請一次? Y # antonio1 1 c 個@ACl|細細仔@AN|俾俾@V|佢@TA|借緊@TV|如@TD|本@TCl|得得@TAdj|你@Tge|書@TN|個@RCl|先生@RN ? 邀請請一次? Y

Experiment items  DOC 1 a

個|細細仔|俾俾|如|本|書|個|先生

? 邀請請一次? Y

BA 1 a

個|細細仔|將|如|本|書|俾俾|個|先生

? 邀請請一次? Y

DOC 1 b

個|細細仔|俾俾|如|本|得得|你|書|個|先生

? 邀請請一次? Y

BA 1 b

個|細細仔|將|如|本|得得|你|書|俾俾|個|先生

? 邀請請一次? Y

DOC 1 c

個|細細仔|俾俾|佢|借緊|如|本|得得|你|書|個|先生

? 邀請請一次? Y

BA 1 c

個|細細仔|將|佢|借緊|如|本|得得|你|書|俾俾|個|先生

? 邀請請一次? Y

DOC 2 a

個|先生|俾俾|如|份|功功|班|學生

? 邀請請一次? Y

BA 2 a

個|先生|將|如|份|功功|俾俾|班|學生

? 邀請請一次? Y

DOC 2 b

個|先生|俾俾|如|份|容容|你|功功|班|學生

? 邀請請一次? Y

BA 2 b

個|先生|將|如|份|容容|你|功功|俾俾|班|學生

? 邀請請一次? Y

89 

Appendix  DOC 2 c

個|先生|俾俾|佢|預預好|如|份|容容|你|功功|班|學生

? 邀請請一次? Y

BA 2 c

個|先生|將|佢|預預好|如|份|容容|你|功功|俾俾|班|學生

? 邀請請一次? Y

DOC 3 a

個|經經|俾俾|如|份|工|個|男仔

? 邀請請一次? Y

BA 3 a

個|經經|將|如|份|工|俾俾|個|男仔

? 邀請請一次? Y

DOC 3 b

個|經經|俾俾|如|份|低低|你|工|個|男仔

? 邀請請一次? Y

BA 3 b

個|經經|將|如|份|低低|你|工|俾俾|個|男仔

? 邀請請一次? Y

DOC 3 c

個|經經|俾俾|佢|做實|如|份|低低|你|工|個|男仔

? 邀請請一次? Y

BA 3 c

個|經經|將|佢|做實|如|份|低低|你|工|俾俾|個|男仔

? 邀請請一次? Y

DOC 4 a

個|老老|俾俾|如|份|報報|個|秘書

? 邀請請一次? Y

BA 4 a

個|老老|將|如|份|報報|俾俾|個|秘書

? 邀請請一次? Y

DOC 4 b

個|老老|俾俾|如|份|請要|你|報報|個|秘書

? 邀請請一次? Y

BA 4 b

個|老老|將|如|份|請要|你|報報|俾俾|個|秘書

? 邀請請一次? Y

DOC 4 c

個|老老|俾俾|佢|睇請|如|份|請要|你|報報|個|秘書

? 邀請請一次? Y

BA 4 c

個|老老|將|佢|睇請|如|份|請要|你|報報|俾俾|個|秘書

? 邀請請一次? Y

DOC 5 a

個|店官|俾俾|如|批|貨貨|個|顧顧

? 邀請請一次? Y

BA 5 a

個|店官|將|如|批|貨貨|俾俾|個|顧顧

? 邀請請一次? Y

DOC 5 b

個|店官|俾俾|如|批|天熱|你|貨貨|個|顧顧

? 邀請請一次? Y

BA 5 b

個|店官|將|如|批|天熱|你|貨貨|俾俾|個|顧顧

? 邀請請一次? Y

DOC 5 c

個|店官|俾俾|海公|入俾|如|批|天熱|你|貨貨|個|顧顧

? 邀請請一次? Y

BA 5 c

個|店官|將|海公|入俾|如|批|天熱|你|貨貨|俾俾|個|顧顧

? 邀請請一次? Y

DOC 6 a

個|侍侍|俾俾|如|碟|牛牛|個|顧顧

? 邀請請一次? Y

BA 6 a

個|侍侍|將|如|碟|牛牛|俾俾|個|顧顧

? 邀請請一次? Y

DOC 6 b

個|侍侍|俾俾|如|碟|全等|你|牛牛|個|顧顧

? 邀請請一次? Y

BA 6 b

個|侍侍|將|如|碟|全等|你|牛牛|俾俾|個|顧顧

? 邀請請一次? Y

DOC 6 c

個|侍侍|俾俾|大大|煎好|如|碟|全等|你|牛牛|個|顧顧

? 邀請請一次? Y

BA 6 c

個|侍侍|將|大大|煎好|如|碟|全等|你|牛牛|俾俾|個|顧顧

? 邀請請一次? Y

DOC 7 a

個|伯伯|俾俾|如|個|銀銀|個|警警

? 邀請請一次? Y

BA 7 a

個|伯伯|將|如|個|銀銀|俾俾|個|警警

? 邀請請一次? Y

DOC 7 b

個|伯伯|俾俾|如|個|真真|你|銀銀|個|警警

? 邀請請一次? Y

BA 7 b

個|伯伯|將|如|個|真真|你|銀銀|俾俾|個|警警

? 邀請請一次? Y

DOC 7 c

個|伯伯|俾俾|綁綁|跌|如|個|真真|你|銀銀|個|警警

? 邀請請一次? Y

BA 7 c

個|伯伯|將|綁綁|跌|如|個|真真|你|銀銀|俾俾|個|警警

? 邀請請一次? Y

DOC 8 a

阿|爸爸|俾俾|如|D|錢|媽媽

? 邀請請一次? Y

BA 8 a

阿|爸爸|將|如|D|錢|俾俾|媽媽

? 邀請請一次? Y

DOC 8 b

阿|爸爸|俾俾|如|D|剩低|你|錢|媽媽

? 邀請請一次? Y

BA 8 b

阿|爸爸|將|如|D|剩低|你|錢|俾俾|媽媽

? 邀請請一次? Y

DOC 8 c

阿|爸爸|俾俾|請同|還|如|D|剩低|你|錢|媽媽

? 邀請請一次? Y

BA 8 c

阿|爸爸|將|請同|還|如|D|剩低|你|錢|俾俾|媽媽

? 邀請請一次? Y

DOC 9 a

阿|細細|俾俾|如|份|報報|爸爸

? 邀請請一次? Y

BA 9 a

阿|細細|將|如|份|報報|俾俾|爸爸

? 邀請請一次? Y

DOC 9 b

阿|細細|俾俾|如|份|今日|你|報報|爸爸

? 邀請請一次? Y

BA 9 b

阿|細細|將|如|份|今日|你|報報|俾俾|爸爸

? 邀請請一次? Y

90 

Appendix  DOC 9 c

阿|細細|俾俾|佢|買|如|份|今日|你|報報|爸爸

? 邀請請一次? Y

BA 9 c

阿|細細|將|佢|買|如|份|今日|你|報報|俾俾|爸爸

? 邀請請一次? Y

DOC 10 a

個|教首|俾俾|如|個|機你|個|球官

? 邀請請一次? Y

BA 10 a

個|教首|將|如|個|機你|俾俾|個|球官

? 邀請請一次? Y

DOC 10 b

個|教首|俾俾|如|個|難得|你|機你|個|球官

? 邀請請一次? Y

BA 10 b

個|教首|將|如|個|難得|你|機你|俾俾|個|球官

? 邀請請一次? Y

DOC 10 c

個|教首|俾俾|大而|等緊|如|個|難得|你|機你|個|球官

? 邀請請一次? Y

BA 10 c

個|教首|將|大而|等緊|如|個|難得|你|機你|俾俾|個|球官

? 邀請請一次? Y

DOC 11 a

個|然生仔|俾俾|如|隻|DVD|個|朋朋

? 邀請請一次? Y

BA 11 a

個|然生仔|將|如|隻|DVD|俾俾|個|朋朋

? 邀請請一次? Y

DOC 11 b

個|然生仔|俾俾|如|隻|天熱|你|DVD|個|朋朋

? 邀請請一次? Y

BA 11 b

個|然生仔|將|如|隻|天熱|你|DVD|俾俾|個|朋朋

? 邀請請一次? Y

DOC 11 c

個|然生仔|俾俾|佢|拍|如|隻|天熱|你|VCD|個|朋朋

? 邀請請一次? Y

BA 11 c

個|然生仔|將|佢|拍|如|隻|天熱|你|VCD|俾俾|個|朋朋

? 邀請請一次? Y

DOC 12 a

個|乘顧|俾俾|如|D|散報|個|請朋朋

? 邀請請一次? Y

BA 12 a

個|乘顧|將|如|D|散報|俾俾|個|請朋朋

? 邀請請一次? Y

DOC 12 b

個|乘顧|俾俾|如|D|多多|你|散報|個|請朋朋

? 邀請請一次? Y

BA 12 b

個|乘顧|將|如|D|多多|你|散報|俾俾|個|請朋朋

? 邀請請一次? Y

DOC 12 c

個|乘顧|俾俾|佢|袋中|如|D|多多|你|散報|個|請朋朋

? 邀請請一次? Y

BA 12 c

個|乘顧|將|佢|袋中|如|D|多多|你|散報|俾俾|個|請朋朋

? 邀請請一次? Y

Filler items  The filler items (filler-fillers) used in this study include fillers from Cheung (2004) and all experimental items and fillers from Matthews et al. (2005). Some minor modifications are made to the sentences and/or their comprehension questions. Some comprehension questions are replaced by a probe to repeat the sentence. Experimental items from Matthews et al. (2005) (see also preliminary analysis on page 75): TOPc 1 a

你|幫|我|買|如|D|糖|丫

TOPt 1 a

如|D|糖|你|幫|我|買|丫

TOPc 1 b

你|幫|我|買|好好|如|D|糖|丫

? 邀請請一次? Y

TOPt 1 b

好好|如|D|糖|你|幫|我|買|丫

? 邀請請一次? Y

TOPc 1 c

你|幫|我|買|我|唔得|食|如|D|糖|丫

? 我但我唔唔得食糖? N

TOPt 1 c

我|唔得|食|如|D|糖|你|幫|我|買|丫

? 我但我唔唔得食糖? N

TOPc 2 a

我|唔唔|做|如|你|數|呀

TOPt 2 a

如|你|數|我|唔唔|做|呀

TOPc 2 b

我|唔唔|做|尋日|如|你|數|呀

? 邀請請一次? Y

TOPt 2 b

尋日|如|你|數|我|唔唔|做|呀

? 邀請請一次? Y

TOPc 2 c

我|唔唔|做|佢|上堂|教|如|你|數|呀

? 佢但我教數學堂? Y

TOPt 2 c

佢|上堂|教|如|你|數|我|唔唔|做|呀

? 佢但我教數學堂? Y

TOPc 3 a

我|好想|請|如|個|島|呀

TOPt 3 a

如|個|島|我|好想|請|呀

91 

Appendix  TOPc 3 b

我|好想|請|上次|如|個|島|呀

? 邀請請一次? Y

TOPt 3 b

上次|如|個|島|我|好想|請|呀

? 邀請請一次? Y

TOPc 3 c

我|好想|請|雜雜|介介|如|個|島|呀

? 我但我想請我我? Y

TOPt 3 c

雜雜|介介|如|個|島|我|好想|請|呀

? 我但我想請我我? Y

TOPc 4 a

我|好想|試|如|架|車|呀

TOPt 4 a

如|架|車|我|好想|試|呀

TOPc 4 b

我|好想|試|紅紅|如|架|車|呀

? 邀請請一次? Y

TOPt 4 b

紅紅|如|架|車|我|好想|試|呀

? 邀請請一次? Y

TOPc 4 c

我|好想|試|爸爸|俾|我|如|架|車|呀

? 爸爸但我爸俾隻爸俾我? N

TOPt 4 c

爸爸|俾|我|如|架|車|我|好想|試|呀

? 爸爸但我爸俾隻爸俾我? N

TOPc 5 a

我|決決|買|如|部|琴|啦

? 邀請請一次? Y

TOPt 5 a

如|部|琴|我|決決|買|啦

? 邀請請一次? Y

TOPc 5 b

我|決決|買|啡紅|如|部|琴|啦

? 我但我你買琴? Y

TOPt 5 b

啡紅|如|部|琴|我|決決|買|啦

? 我但我你買琴? Y

TOPc 5 c

我|決決|買|尋日|睇你|如|部|琴|啦

TOPt 5 c

尋日|睇你|如|部|琴|我|決決|買|啦

TOPc 6 a

我|聽日|要|用|如|本|書

? 邀請請一次? Y

TOPt 6 a

如|本|書|我|聽日|要|用

? 邀請請一次? Y

TOPc 6 b

我|聽日|要|用|檯檯|如|本|書

? 本書但我本睇俾? N

TOPt 6 b

檯檯|如|本|書|我|聽日|要|用

? 本書但我本睇俾? N

TOPc 6 c

我|聽日|要|用|但|書書|買|如|本|書

TOPt 6 c

但|書書|買|如|本|書|我|聽日|要|用

TOPc 7 a

我|聽日|要|首|如|首|歌

? 邀請請一次? Y

TOPt 7 a

如|首|歌|我|聽日|要|首

? 邀請請一次? Y

TOPc 7 b

我|聽日|要|首|好難|如|首|歌

? 首歌但我好難聽? N

TOPt 7 b

好難|如|首|歌|我|聽日|要|首

? 首歌但我好難聽? N

TOPc 7 c

我|聽日|要|首|但|婚婚|唱|如|首|歌

TOPt 7 c

但|婚婚|唱|如|首|歌|我|聽日|要|首

TOPc 8 a

我|唔唔俾|如|隻|爸|呀

? 邀請請一次? Y

TOPt 8 a

如|隻|爸|我|唔唔俾|呀

? 邀請請一次? Y

TOPc 8 b

我|唔唔俾|黑紅|如|隻|爸|呀

? 隻爸但我黑紅你? Y

TOPt 8 b

黑紅|如|隻|爸|我|唔唔俾|呀

? 隻爸但我黑紅你? Y

TOPc 8 c

我|唔唔俾|情請情|本|如|個|盒|呀

TOPt 8 c

情請情|本|如|個|盒|我|唔唔俾|呀

TOPc 9 a

我|好好|唔得|如|個|請

? 我但我唔唔如個請? N

TOPt 9 a

如|個|請|我|好好|唔得

? 我但我唔唔如個請? N

TOPc 9 b

我|好好|唔得|好肥|如|個|請

TOPt 9 b

好肥|如|個|請|我|好好|唔得

TOPc 9 c

我|好好|唔得|新新|講緊|如|個|請

? 邀請請一次? Y

TOPt 9 c

新新|講緊|如|個|請|我|好好|唔得

? 邀請請一次? Y

TOPc 10 a

我|好唔得|如|D|花|呀

? 我但我唔唔得花? N

TOPt 10 a

如|D|花|我|好唔得|呀

? 我但我唔唔得花? N

92 

Appendix  TOPc 10 b

我|好唔得|黃紅|如|D|花|呀

TOPt 10 b

黃紅|如|D|花|我|好唔得|呀

TOPc 10 c

我|好唔得|你|爸|俾|我|如|D|花|呀

? 邀請請一次? Y

TOPt 10 c

你|爸|俾|我|如|D|花|我|好唔得|呀

? 邀請請一次? Y

TOPc 11 a

我|尋日|睇俾|如|套|戲

? 我但我睇俾戲? Y

TOPt 11 a

如|套|戲|我|尋日|睇俾

? 我但我睇俾戲? Y

TOPc 11 b

我|尋日|睇俾|每成|如|套|戲

TOPt 11 b

每成|如|套|戲|我|尋日|睇俾

TOPc 11 c

我|尋日|睇俾|每成|推介|如|套|戲

? 邀請請一次? Y

TOPt 11 c

每成|推介|如|套|戲|我|尋日|睇俾

? 邀請請一次? Y

TOPc 12 a

我|終於|搵搵|如|件|衫

? 我但我搵緊衫? Y

TOPt 12 a

如|件|衫|我|終於|搵搵

? 我但我搵緊衫? Y

TOPc 12 b

我|終於|搵搵|請去|如|件|衫

TOPt 12 b

請去|如|件|衫|我|終於|搵搵

TOPc 12 c

我|終於|搵搵|而家|俾|我|如|件|衫

? 邀請請一次? Y

TOPt 12 c

而家|俾|我|如|件|衫|我|終於|搵搵

? 邀請請一次? Y

Mixed fillers are taken from Cheung (2004) and Matthews et al. (2005): 好多|問題|都|係|由|誤會|引起

? 請重複一次? Y

你|係|果度|好|喇|，|千祈|咪搞|佢|D|野

? 請重複一次? Y

我|細個|要|幫|媽咪|做|家務

? 請重複一次? Y

我|而家|返|屋企|先|再|出去

? 請重複一次? Y

我|無幾耐|之前|出左|部|零機價|電話

? 請重複一次? Y

老闆|話|你|做緊|果|份|文件|好緊要|架

? 請重複一次? Y

佢|係|出面|搵緊|野食|番黎

? 請重複一次? Y

聽朝早|去|對面街|果|間|茶樓|飲|早茶

? 請重複一次? Y

佢|話|果|幅|畫|係|佢|架

? 請重複一次? Y

今日|拍賣|果|幅|畫|係|佢|架

? 請重複一次? Y

我|以前|去過|山頂|果|間|屋

? 請重複一次? Y

山頂|果|間|屋|我|以前|去過

? 請重複一次? Y

我|搵|細個|果|D|相|俾|你|睇

? 請重複一次? Y

我|搵|我|同|你|影|果|D|相|俾|你|睇

? 請重複一次? Y

你|還番|果|D|錢|俾|我

? 請重複一次? Y

你|還番|上星期|借|果|D|錢|俾|我

? 請重複一次? Y

佢|話|俾|尋晚|果|D|粥|我

? 請重複一次? Y

佢|話|俾|你|今朝|煲|果|D|粥|我

? 請重複一次? Y

個|先生|準備好|聽日|要|去|中大|present|果|個|PowerPoint

? 佢聽日要去中大present? Y

我|個|朋友|撞車|入左|醫院

? 我個朋友係咪受左傷? N

本|書|係|關於|中國|既|歷史

? 本書係咪講耶穌既事? N

本|書|既|作者|係|我|個|中學|同學

? 本書係咪講耶穌既事? N

先生|叫|我|上堂|果|陣|唔好|傾偈

? 先生係咪話唔好傾偈? Y

先生|話|我|既|成績|有|進步

? 先生係咪話唔好傾偈? N

93 

Appendix  你|要|做|多|D|運動|呀

? 你係咪要訓多D? N

你|記住|每日|要|訓|多過|八|個|鐘

? 你係咪要訓多D? N

多謝|你|今日|陪|我|去|聽|音樂會

? 你係咪陪我去睇戲? N

多謝|你|份|生日禮物

? 你係咪陪我去睇戲? N

我|一陣間|要|去|學|普通話

? 我一陣係咪同哥哥一齊? N

我|一陣間|要|同|哥哥|去|學|游水

? 我一陣係咪同哥哥一齊? Y

我|好耐|冇去|泳池|游水|啦

? 我係咪日日去行山? N

我|好耐|冇同|媽咪|行街|買|衫|啦

? 我係咪日日去行山? N

我|今次|去|台灣|玩得|好開心|呀

? 我係咪去左台灣? Y

我|今次|旅行|玩得|好開心

? 我係咪去左台灣? N

我|今日|要|係|開放日|幫手

? 我係咪去派紀念品? N

我|要|係|開放日|果|日|派|紀念品|呀

? 我係咪去派紀念品? Y

我|重|未去|你|屋企|搵番|個|袋|呀

? 我係咪未還書? N

我|重|未去|圖書館|還|書|呀

? 我係咪未還書? Y

我|想|新年|果|陣|去|日本|玩

? 我係咪想去日本? Y

我|想|暑假|果|陣|去|美國|迪士尼

? 我係咪想去日本? N

我|好鍾意|同|爸爸|一齊|去|旅行

? 我係咪同爸爸去行山? N

我|好鍾意|打|籃球|同|游水

? 我係咪同爸爸去行山? N

我|好耐|都|冇|同|佢|傾偈|啦

? 我係咪成日見佢? N

我|好耐|都|冇見|佢|啦

? 我係咪成日見佢? N

我|聽日|要|交|份|作文

? 我係咪有野要交? Y

我|聽日|要|俾倒|份|稿|老細

? 我係咪有野要交? Y

我|重有|好多|野|未做完|呀

? 我係咪有野要做? Y

我|放左工|通常|會|約|朋友|食飯

? 我係咪有野要做? N

我|今日|係|地鐵站|撞倒|佢

? 我係咪係巴士站撞倒佢? N

我|今日|係|巴士站|附近|撞倒|佢

? 我係咪係巴士站撞倒佢? Y

我|今晚|約左|佢|去|睇戲|呀

? 我係咪約左佢? Y

我|今晚|要|陪|媽咪|去|睇|醫生|呀

? 我係咪約左佢? Y

我|大個左|要|做|一|個|好|老師|呀

? 我係咪要做護士? N

我|想|大個左|做|一|個|醫生

? 我係咪要做護士? N

我|今朝|食左|火腿|三文治

? 我係咪食左蛋治? N

我|今朝|食左|牛奶|麥皮|做|早餐

? 我係咪食左蛋治? N

我|覺得|佢|係|一|個|好|既|男仔

? 我係咪唔鍾意佢? N

我|覺得|佢|個|人|好自私|呀

? 我係咪唔鍾意佢? N

我|咁|大個|都|唔識|踩|單車

? 我係咪唔識用電腦? N

我|由細到大|都|好驚|貓|同|狗|架

? 我係咪唔識用電腦? N

我|今日|買左|件|白色|恤衫

? 我係咪買左衫? Y

我|今日|買左|件|衫|同埋|對|涼鞋

? 我係咪買左衫? Y

今日|我|同|媽咪|去|沙田|探|婆婆

? 我係咪想去動物園? N

今日|我|要|去|街巿|買肉|呀

? 我係咪想去動物園? N

我|遲D|會|同|屋企人|去|日本|玩

? 我係咪會去加拿大? Y

我|遲D|要|去|加拿大|讀書

? 我係咪會去加拿大? Y

94 

Appendix  我|而家|係|廚房|整緊|芒果蛋糕

? 我係咪整緊野食? Y

我|而家|睇緊|新聞報導|呀

? 我係咪整緊野食? N

我|鍾意|去|果|度|玩|過山車

? 我係咪鍾意玩海盜船? N

我|鍾意|去|海洋公園|玩|跳樓機

? 我係咪鍾意玩海盜船? N

我|好鍾意|飲|朱古力奶|架

? 我係咪鍾意飲可樂? N

我|好鍾意|飲|檸檬茶|同埋|橙汁

? 我係咪鍾意飲可樂? N

我|唔係好識|去|果|度

? 我係咪識去海洋公園? N

我|唔係好識|去|海洋公園

? 我係咪識去海洋公園? Y

我|考試|果|陣|時|瘦左|好多

? 我係咪覺得自己肥左? N

我|覺得|我|好似|肥左

? 我係咪覺得自己肥左? Y

我|聽日|要|去|中環|見工|呀

? 我係咪聽日去見工? Y

我|聽日|要|去|見工|呀

? 我係咪聽日去見工? Y

我|畢左業|想|做|先生

? 我畢業之後係咪會做野? Y

我|畢業|之後|諗住|再|讀|多|一|年

? 我畢業之後係咪會做野? N

佢|好記得|我|鍾意|食|咩|架

? 佢係咪好有記性? Y

佢|對|我|真係|好好|架

? 佢係咪好有記性? N

你|次次|約|佢|都|要|等|成|個|鐘|架

? 佢係咪唔守時? Y

佢|次次|約|人|都|會|遲到|架

? 佢係咪唔守時? Y

佢|讀完|就|要|返番|去|澳洲

? 佢係咪會留係香港? N

佢|讀完|要|去|爸爸|間|公司|幫手

? 佢係咪會留係香港? Y

爸爸|同|媽咪|去|台灣|旅行

? 爸爸係咪同我去旅行? N

爸爸|同|媽咪|想|再|去|多|次|韓國

? 爸爸係咪同我去旅行? N

爸爸|話|今晚|可能|要|夜少少|番

? 爸爸係咪會夜番? Y

爸爸|話|今晚|會|番黎|食飯

? 爸爸係咪會夜番? N

家姐|話|請|我|去|酒店|食|自助餐

? 家姐係咪有野俾我? N

家姐|話|買|手信|俾|我

? 家姐係咪有野俾我? Y

隊長|話|今次|係|唔輸得|架

? 隊長係咪想贏? Y

隊長|話|今次|係|最後|一|次|比賽

? 隊長係咪想贏? N

媽咪|叫|我|聽日|出街|著|多|件|衫

? 媽咪係咪叫我著多D衫? Y

媽咪|叫|我|聽日|著|多|件|衫

? 媽咪係咪叫我著多D衫? Y

媽咪|今日|好開心|喎

? 媽咪係咪唔開心? N

媽咪|好似|有|D|唔開心|喎

? 媽咪係咪唔開心? Y

醫生|話|我|要|休息|兩|三|日

? 醫生係咪話我發燒? N

醫生|話|我|要|戒|食|牛肉|同|雞蛋

? 醫生係咪話我發燒? N

95 

Bibliography 

Bibliography  Ansaldo, Umberto & Stephen J. Matthews (2001). Typical creoles and simple languages: the case of Sinitic. Linguistic Typology 5,2: 311-325. Arnold, Jennifer E., Thomas Wasow, Anthony Losongco & Ryan Ginstrom (2000). Heaviness vs. newness: The effects of complexity and information structure on constituent ordering. Language 76: 28-55. Arnold, Jennifer E., Maria Fagnano & Michael K. Tanenhaus (2003). Disfluencies signal theee, um, new information. Journal of Psycholinguistic Research 32,1: 25-36. Bauer, Robert & Stephen J. Matthews (2003). Cantonese. In Thurgood & LaPolla 2003. Bender, Emily M. (2001). The syntax of Mandarin ba: Reconsidering the verbal analysis. Journal of East Asian Linguistics 9, 2: 105–145. Bock, J. Kathryn (1995). Sentence production: From mind to mouth. In Joanna L. Miller & Peter D. Eimas, eds., Handbook of perception and cognition. Vol 11: Speech, language, and communication, 181-216. Orlando, FL: Academic Press. -. (1996). Language production: Methods and methodologies. Psychonomic Bulletin & Review 3,4: 395-421. Bock, J. Kathryn, Gary S. Dell, Franklin Chang & Kristine H. Onishi (2005). Persistent structural priming from language comprehension to language production. Manuscript submitted for publication. Bock, J. Kathryn, Gary S. Dell, Susan M. Garnsey, Arthur F. Kramer & Tate T. Kubose (in press). Car talk, car listen. In Antje S. Meyer, Linda R. Wheeldon & Andrea Krott, eds., Language and executive processes. Hove: Psychology Press. Bodomo, Adams B., Olivia S.C. Lam & Natalie S.S. Yu (2003). Double object and serial verb benefactive constructions in Cantonese. The 8th International Lexical-Functional Grammar Conference. Stanford, CA: Center for the Study of Language and Information (CSLI), 2003, 59-74. (Publication No. : 89941). Retrieved on 20th May, 2005 at http://cslipublications.stanford.edu/LFG/8/lfg03bodomoetal.pdf.

Caplan, David & Gloria S. Waters (1999). Verbal working memory and sentence comprehension. Brain and Behavioral Sciences 22: 77-126. Chan, Angel (2003). The development of bei2 dative constructions in Early Child Cantonese. MPhil thesis. Chinese University of Hong Kong. Chappell, Hilary (2001). Synchrony and diachrony of Sinitic languages: a brief history of Chinese dialects. In Chappell 2001. -., ed. (2001). Sinitic grammar: synchronic and diachronic perspectives. Oxford: Oxford University Press.

96 

Bibliography  Cheung, Antonio (2004). On the role of weight in constituent order in Cantonese: an experimental study of double object constructions. BCogSc thesis. The University of Hong Kong. -. (2005). Syntactic alternatives: investigating the effects of weight in Cantonese dative construction. Paper read at The 10th International Conference on Yue Dialects, CUHK, Hong Kong, 12 December. Cheung, Samuel H.N. (1994). A practical Chinese grammar. Hong Kong: Chinese University Press. Cheung, Yam Leung (1997). A study of right dislocation in Cantonese. MPhil thesis. Chinese University of Hong Kong. Chomsky, Noam (1965). Aspects of the theory of syntax. Cambridge, MA: MIT Press. -. (2000). New horizons in the study of language and mind. Cambridge: Cambridge University Press. Chow, Dorcas C.C. (2005). The effects of parental input on the early development of bei2 dative constructions in Cantonese children. BSc(Sp&HearSc) thesis. The University of Hong Kong. Croft, William (1995). Modern syntactic typology. In Masayoshi Shibatani & Theodora Bynon, eds. Approaches to language typology. Oxford: Oxford University Press. -. (2000). Explaining language change: An evolutionary approach. London: Longman. Culicover, Peter W. & Ray Jackendoff (2005). Simpler syntax. Oxford: Oxford University Press. Djamouri, Redouane (2001). Markers of prediction in Shang bone inscriptions. In Chappell 2001. Dryer, Matthew (1992). The Greenbergian word order correlations. Language 68,1: 81-138. -. (2003). Word order in Sino-Tibetan languages from a typological and geographical perspective. In Thurgood & LaPolla 2003. -. (2005). Relationship between the order of object and verb and the order of relative clause and noun. In Martin Haspelmath, Matthew Dryer, David Gil, & Bernard Comrie, eds., The world atlas of language structures, 390-1. Oxford: Oxford University Press. Ernst, Thomas (1988). Chinese postpositions again. Journal of Chinese Lingusitics 16,2: 219-245. Fedorenko, Evelina, Edward Gibson & Douglas Rohde (2004). Working memory in sentence comprehension. Proceedings of the 26th Annual Meeting of Cognitive Science Society. Mahwah, NJ: Lawrence Erlbaum. Ferreira, Fernanda (2003). Misinterpretation of noncanonical sentences. Cognitive Psychology 47: 164-203.

97 

Bibliography  Ferreira, Victor S. (1996). Is it better to give than to donate? Syntactic flexibility in language production. Journal of Memory and Language 35: 724–755. Ferreira, Victor S. & Gary S. Dell (1996). Do speakers choose their words cooperatively? Investigating the production-comprehension interface. Poster presented at The 37th Annual Meeting of the Psychonomics Society. Chicago, IL. Fodor, Janet D. (1998). Learning to parse? Journal of Psycholinguistic Research 27: 285-319. Frazier, Lyn (1985). Syntactic complexity. In David Dowty, Lauri Karttunen & Arnold Zwicky, eds., Natural language processing: Psychological, computational and theoretical perspectives, 129-189. Cambridge: Cambridge University Press. Gibson, Edward (1998). Linguistic complexity: locality of syntactic dependencies, Cognition 68,1: 1-76. -. (2000). The dependency locality theory: a distance-based theory of linguistic complexity. In Alec Marantz, Yasushi Miyashita & Wayne O'Neil, eds., Image, language, brain, 95-126. Cambridge, MA: MIT Press. -. (2005a). Lecture 7: Serial/parallel; Processing evidence for empty categories. LSA313 Sentence comprehension. Lecture notes. Cambridge, MA: LSA 2005. Retrieved 22nd July 2005 at http://tedlab.mit.edu/lsa-course/lect 7 syntax reps.pdf. -. (2005b). Lecture 8: Syntax: phrasal adjacency; Referential context. LSA313 Sentence comprehension. Lecture notes. Cambridge, MA: LSA 2005. Retrieved 22nd July 2005 at http://tedlab.mit.edu/lsa-course/lect 8 syntax & context.pdf.

Goldberg, Adele E. (1995). Constructions. Chicago, IL: The University of Chicago Press. -. (2006). Constructions at Work. Oxford: Oxford University Press. Gordon, Peter C., Randall Hendrick, & William H. Levine (2002). Memory-load interference in syntactic processing. Psychological Science 13: 425-430. Greenberg, Joseph H. (1966). Some universals of grammar with particular reference to the order of meaningful elements. In Joseph H. Greenberg, ed. Universals of language. Cambridge, MA: MIT Press. Haspelmath, Martin (1999). Optimality and diachronic adaptation. Zeitschrift für Sprachwissenschaft 18,2: 180-205. -. (2002). Functional and formal explanation in linguistics. Lecture notes. Düsseldorf: 2002 Düsseldorf Summer School. Retrieved 3rd May 2006 at http://email.eva.mpg.de/~haspelmt/Haspelmath9Handouts.pdf.

Hawkins, John A. (1990). A parsing theory of word order universals. Linguistic Inquiry 21,2: 223-262. -. (1994). A performance theory of order and constituency. Cambridge: Cambridge University Press.

98 

Bibliography  -. (2004). Efficiency and complexity in grammars. Oxford: Oxford University Press. Herforth, Derek (2003). A sketch of late Zhou Chinese grammar. In Thurgood & LaPolla 2003. Ho, Daphne S.L. (2004). On the acquisition of bei2-dative constructions in Cantonese. MA in Chinese Linguistics thesis. The Hong Kong Polytechnic University. Hsiao, Franny & Edward Gibson (2003). Processing relative clauses in Chinese, Cognition 90: 3-27. Huang, James C.T. (1982). Logical relations in Chinese and the theory of grammar. PhD dissertation, Massachusetts Institute of Technology. Huang, James C.T. & Audrey Li (1995). Recent generative studies in Chinese syntax. In James C.T. Huang. & Audrey Li, eds., New horizons in Chinese linguistics. Dordrecht: Kluwer Academic Publishers. Jackendoff, Ray (2002). Foundations of language. New York: Oxford University Press. Kwan, Fiona (2000). Parsing and processing of sentences with pretransitive construction in Cantonese. BCogSc thesis. The University of Hong Kong. Lai, Huei-ling (2003). Hakka lau constructions: A constructional approach. Language and Linguistics 4,2: 353-378. Lefebvre, Claire (1998). Creole genesis and the acquisition of grammar: the case of Haitian Creole. Cambridge: Cambridge University Press. Levelt, Willem J. (1989). Speaking: from intention to articulation. Cambridge, MA: MIT Press. Li, Audrey (1990). Order and constituency in Mandarin Chinese. Dordrecht: Kluwer Academic Publishers. Li, Charles & Sandra Thompson (1976). Subject and topic: a new typology of language. In Charles Li, ed., Subject and topic, 457-489. New York, NY: Academic Press. Lohse, Barbara, John A. Hawkins & Thomas Wasow (2004). Domain minimization in English verb-particle constructions. Language 80, 2: 238-261. Luke, K.K. (1998). Post-modifiers in the Cantonese NP. [粵語名詞組中的後置修飾語] Fangyan 1998,1: 48-52. Matthews, Stephen J. (2004). Lecture 4: Explaining word order universals. LING6021 Language Types and Universals. Lecture notes. Hong Kong: The University of Hong Kong. Retrieved 20th May, 2006 at http://www.hku.hk/linguist/program/Typology4.html. -. (2006). On serial verb constructions in Cantonese. In Alexandra Aikhenvald & Robert Dixon, eds., Serial verbs: a cross-linguistic typology, 69-87. Oxford: Oxford University Press.

99 

Bibliography  Matthews, Stephen J., Helen H.L. Ching & Virginia Yip (2005). Topicalization and the processing of relative clauses in Cantonese. Paper read at The 10th International Conference on Yue Dialects, CUHK, Hong Kong, 14 December. Matthews, Stephen J. & Louisa Y.Y. Yeung (2001). Processing motivations for topicalization in Cantonese. In Kaoru Horie & Shigeru Sato, eds., Cognitive-functional linguistics in an East Asian context, 81-102. Tokyo: Kurosio. Matthews, Stephen J. & Virginia Yip (1994). Cantonese: A comprehensive grammar. London: Routledge. -. (2001). Aspects of Contemporary Cantonese grammar: The structure and stratification of relative clauses. In Chappell 2001. Michaelis, Susanne & Martin Haspelmath (2003). Ditransitive constructions: Creole languages in a cross-linguistic perspective. Creolica 23/4/2003. Retrieved on 10th Dec, 2004 at http://www.creolica.net/michaelis.pdf. Newmeyer, Frederick J. (2003). Grammar is grammar and usage is usage. Language 79: 682-707. -. (2005). Possible and probably languages: a generative perspective on linguistic typology. Oxford: Oxford University Press. O’Melia, Thomas (1941). First year Cantonese. Hong Kong: Maryknoll House. Paul, Waltraud (2002). Proxy categories in Phrase Structure Theory and the Chinese VP. Cahiers de Linguistique - Asie Orientale 31, 2: 137-174. Pearlmutter, Neal & Edward Gibson (2001). Recency in verb phrase attachment. Journal of Experimental Psychology: Learning, Memory, and Cognition 27: 574–590. Peyraube, Alain (2004). Chapter 41 Ancient Chinese. In Roger Woodard, ed., The Cambridge encyclopedia of the world's ancient languages, 988-1014. Cambridge: Cambridge University Press. Pickering, Martin J. & Simon Garrod (2004). Toward a mechanistic psychology of dialogue. Behavioral and Brain Sciences 27: 169-226. Pinker, Steven (1994). The language instinct. London: Penguin Books. Pulleyblank, Edwin (1995). Outline of Classical Chinese grammar. Vancouver: UBC Press. Sag, Ivan A., Thomas Wasow & Emily M. Bender (2003): Syntactic theory: a formal introduction. Stanford, CA: Center for the Study of Language and Information Publications. Sedivy, Julie C., Michael K. Tanenhaus, Craig G. Chambers & Gregory N. Carlson (1999). Achieving incremental semantic interpretation through contextual representation. Cognition 71: 109-147. Smith, Mark & Linda Wheeldon (1999). High level processing scope in spoken sentence production. Cognition 73, 205-246.

100 

Bibliography  Stallings, Lynne, Maryellen MacDonald & Padraig O'Seaghdha (1998). Phrasal ordering constraints in sentence production: phrase length and verb disposition in heavy-NP shift. Journal of Memory and Language 39: 392-417. Stowe, Laurie A., Marco Haverkort & Frans Zwarts (2005). Rethinking the neurological basis of language. Lingua 115: 997-1042. Tang, Sze-Wing (1998). On the ‘inverted’ double object construction. In Stephen J. Matthews, ed., Studies in Cantonese linguistics, 35-52. Hong Kong: Linguistics Society of Hong Kong. Thurgood, Graham (2003). A subgrouping of the Sino-Tibetan languages: the interaction between language contact, change, and inheritance. In Thurgood & LaPolla 2003. Thurgood, Graham & Randy LaPolla, eds. (2003).The Sino-Tibetan languages. London: Routledge. Vigliocco, Gabriella & Robert J. Hartsuiker (2002). The interplay of meaning, sound, and syntax in sentence production. Psychological Bulletin 128,3: 442-472. Warren, Tessa & Edward Gibson (2002). The influence of referential processing on sentence complexity. Cognition 85: 79-112. Wasow, Thomas (2002). Postverbal behavior. Stanford, CA: CSLI Publications. Wasow, Thomas & Jennifer Arnold (2003). Post-verbal constituent ordering in English. In Günter Rhodenberg & Britta Mondorf, eds. Determinants of grammatical variation in English, 119-154. Berlin: Mouton de Gruyter. Xu, Liejiong & Alain Peyraube (1997). On the double object construction and the oblique construction in Cantonese. Studies in Language 21,1: 105-127. Yamashita, Hiroko & Franklin Chang (2001). “Long before short” preference in the production of a head-final language. Cognition 81: B45-B55. Yamashita, Hiroko (2002). Scrambled sentences in Japanese: linguistic properties and motivations for production. Text 22,4: 597-633. Yeung, Louisa Y.Y. (1999). An experimental study on sentence parsing: topicalization in Cantonese. BCogSc thesis. The University of Hong Kong. Yue, Anne (2001). The verb complement construction in historical perspective with special reference to Cantonese. In Chappell 2001. -. (2003). Chinese dialects: grammar. In Thurgood & LaPolla 2003. -. (2004). Materials for the diachronic study of the Yue dialects. In Feng Shi & Zhongwei Shen, eds., The joy of research: A festschrift in honor of Professor William S-Y. Wang on his seventieth birthday, 246-271. Tianjin: The Nankai University Press. Zipf, George K. (1949). Human behavior and the principle of least effort. Cambridge, MA: Addison-Wesley Press.

101