Luận án tiến sĩ: Korean Zero Pronouns: Analysis and Resolution

In Korean sentences, arguments such as subject and object enjoy a great degree of dom in terms of the form in which they are realized: a noun head followed by a sub-ject/object/adverbial

KOREAN ZERO PRONOUNS: ANALYSIS AND RESOLUTION

Na-Rae Han

A DISSERTATION

inLinguistics

Presented to the Faculties of the University of Pennsylvania in PartialFulfillment of the Requirements for the Degree of Doctor of Philosophy

UMI Number: 3211080

INFORMATION TO USERS

The quality of this reproduction is dependent upon the quality of the copysubmitted Broken or indistinct print, colored or poor quality illustrations and photographs, print bleed-through, substandard margins, and improper alignment can adversely affect reproduction.

In the unlikely event that the author did not send a complete manuscriptand there are missing pages, these will be noted Also, if unauthorizedcopyright material had to be removed, a note will indicate the deletion

®UMI

UMI Microform 3211080Copyright 2006 by ProQuest Information and Learning Company.All rights reserved This microform edition is protected againstunauthorized copying under Title 17, United States Code

ProQuest Information and Learning Company

300 North Zeeb RoadP.O Box 1346 Ann Arbor, MI 48106-1346

To my dear brother Baek-Kyoung.

First and foremost, I thank my two advisors, Ellen Prince and Martha Palmer I am eternallyindebted to them for their encouragement and intellectual and spiritual! guidance, whichshaped me as a researcher throughout my long journey towards graduation Ellen, whom I

have the honor of being the last student of, has been a great source of inspiration to me forher sharp intellect and wisdom, as well as her unparalleled zest for life I owe Martha notonly for her generous financial support over the years, but also for her guardianship — shetruly took me under her wings

I also thank Robin Clark for his teachings during my early years of graduate school;Bill Poser, whose love and knowledge of everything linguistic has always shown me thegreatness to aspire to; and last but not the least, my committee, Maribel Romero and Ar-avind Joshi, who provided valuable feedbacks on the directions of the thesis, which notunlike any other theses started out as a entangled jumble of grand ideas

XRCE and LDC are two great institutions I was fortunate enough to have opportunities

to conduct research at I thank Lauri Karttunen, Ken Beesley and Mike Maxwell for theirmentorship during my times there

I would like to thank Justin Mott, Cassie Creswell, Tom Morton, Andy Schein, ShamKakade, and Heejong Yi, with whom I shared the most cherished memories of my graduateschool years They stayed close and bore witness to my years of battle with my thesis;without their cheering, I might well be writing it still

There are many other friends and colleagues at Penn, who over the years made it muchmore than just a school but rather like a home to me I had many lunches, dinners, parties,and heated discussions with them: John Bell, Alexis Dimitriadis, Kieran Snyder, Ron Kim,Sophia Malamud, Uri Horesh, Eva Banik, Tom McFadden, Elsi Kaiser, Sandhya Sundare-san, Rashmi Prasad, Eleni Miltsakaki, Seungyen Yang, Eon-Suk Ko, Chunghye Han andJiyoon Lee.

My friends and mentors from my previous school, the linguistics program at SeoulNational University, have been always there for me, sometimes in the States and othertimes across the Pacific Ocean I thank Jinyoung Choi, Seunghun Lee and Hyunjoo Kimfor their friendship (and also for being great “juniors” to me); Yoonshin Kim for being myclose friend (and a great “senior”); Sookhee Chae, Shijong Ryu and Chulwoo Park for theirkind mentoring And finally Prof Chungmin Lee for the very first linguistics class I tookand for guiding me into the field He taught me during my SNU linguistics years, and heencouraged me to pursue this degree in the U.S

Also many special thanks to Gene Buckley and Amy Forsyth for their wonderful helpregarding administrative matters; as I fumbled through the maze of graduation, their knowl-edge and attention on intricate procedural details saved me more than once from crucialmistakes I cannot forget Mrs Carole Lingle, whose warm presence filled the departmentoffice in earlier years of my graduate school Staff at IRCS also has been the quiet yetessential organizing force behind that great institution, to which I owe much of my profes-sional development

Finally, my deepest gratitude goes to my parents and my sister, without whose constantsupport and encouragement none of my academic goals would have been realized Theirfaith in me and my love for them kept me going at times of difficulty and anguish

KOREAN ZERO PRONOUNS: ANALYSIS AND RESOLUTION

Na-Rae Han

Supervisors: Ellen F Prince and Martha Palmer

Zero pronouns, or dropped arguments, are a remarkably frequent phenomenon in Korean.This single syntactic form is made up of diverse subcategories, each of which is character-ized by distinct semantic and pragmatic properties More widely acknowledged types arethose that depend on other linguistic expressions for their reference: such text-dependenttypes include anaphoric and discourse-deictic zero pronouns Other text-independent typesare deictic zero pronouns, generic and specific indefinite zero pronouns and situationalzero pronouns, although it is possible for some of these text-independent types to entercoreferential relations with other nominal expressions in their surroundings Previous re-search focusing on anaphoric zero pronouns, most notably that based on Centering Theory,claims that information-theoretic notions such as saliency govern their felicitous use andinterpretation While the general insight holds true, various efforts to encapsulate it byway of precise formulation of Cf-ranking or other hierarchies fall short, largely due to thefact that the notion is encoded by heterogeneous linguistic factors whose relations cannot

be expressed single-dimensionally From a language processing point of view, the diversenature of Korean zero pronouns presents the unique challenge of blending the tasks of cat-egorization and identification of their antecedents In this dissertation, using MaximumEntropy as the machine learning method of choice, various statistical models for Koreanzero pronoun resolution have been successfully trained and tested on two Korean Treebankcorpora These Models serve as a valuable opportunity for empirically testing various the-oretical claims and observations made on Korean zero pronoun anaphora Features used

in constructing the models and making predictions on zero pronoun reference encode

lin-guistic properties surrounding zero pronouns and their potential antecedents The featuresfound to have a particularly strong contribution are indeed those that encode the linguisticaspects that are commonly cited in the linguistic literature as playing a crucial role in Ko-rean Zero pronoun usage, such as topic-hood, subject-hood and the nullness of form Whilethe relative importance of such features does not directly translate to linguistic hierarchies,

it nevertheless provides support to some of the specific criteria used in them

Acknowledgements iii

Abstract M Contents vii

List of Tables xi

Introduction 1

I Analysis 4

1 Previous Work on Zero Pronouns 5

1.1 Previous Sentence-Level Work on Zero Pronouns 6

1.1.1 proin Government and Binding Theory 6

1.1.2 Huang’s (1983, 1984, 1989) Work on Chinese, Japanese and Ko-TEAN PYO a ÈẼ 1]

1.1.3 Optimality Theory Approaches to Zero Pronouns 14

1.2 Centering Theory: A Discourse-Oriented Approach to Pronouns 22

1.2.1 The Centering Theory: An Overview 22

1.2.2 Centering Theory Across Languages 26

1.2.3 The Zero Pronoun and Centering 30

2 Analysis of Korean Zero Pronouns 35 2.1 Defining the Object of the Study 2 2.0.2.0 0.00.0 00048 35 2.1.1 Why Zero “Pronouns” 0 0.0 ee ee 35 2.2 Overt Pronounsin Korean 0.0 eee ee ee 38 2.3 The Problem of Identification: Where to Find the Invisible 43

2.4 Korean Zero Pronouns by Reference Types c 45 2.4.1 General Situational Zero Pronouns 46

2.4.2 Delctic ZeroPronouns ee es 51 2.43 Indefinite Personal Zero Pronouns 56

2.4.3.1 Specific Indefinite Zero Pronouns 57

2.4.3.2 +human Semantic Restriction on Generic Zero Pronouns 60 2.4.3.3 Coreference in Generic Zero Pronouns 63

2.4.4 Discourse-Anaphoric Zero Pronouns 66

2.5 Semantic Interpretation of Anaphoric Relations 68

2.5.1 Discourse(Textual)-Deictic Zero Pronouns 72

2.6 Deictic and/or Anaphoric: the Fuzzy Distincion 76

3 The Centering Theory and Korean Zero Anaphora 80 3.1 Criteria for Cf Ranking: What EncodesSalence? 81

3.2 Establishing Cf Ranking for Korean 2 2 ee 86 3.3 Topic-Marked NPs vs Zero Pronouns 99

3.4 The Centering Theory and Zero Pronoun Resolution 115

II Resolution 118 Overview 119

1 Previous Work 121 1.1 Pronoun Resolution: An Overview 2.0.0.0 0000 ee eee nae 121

1.1.2 The Traditional Approach 2.220005 123 1.13 The Statistical Approach 2 2 ee 124 1.1.4 The Knowledge-Poor Approach 00004 125

1.1.5 Resolution of Zero Pronouns: The Case of Spanish , 126

1.2 Centering Theory and Pronoun Resolulon 128

1.2.1 Brennan, Friedman and Pollard(1987) 128

1.2.2 Strube’s (1998)“S-lst”Approach 129

1.2.3 Left-Right Centering by letreault(1999) 131

1.2.4 Resolution of Zero Pronouns Using Centering: The Case of Thai 134

1.3 Optimality Approaches to Anaphora Resolulon 137

1.3.1 Beaver’s (2002) CenteringinOT 138

1.3.2 Hong (2002) and Kim (2003): OT-Based Korean Anaphora Reso-lution © nu kg và ky 141 2 The Data: The Penn Korean Treebank Corpora 145 2.1 Overviewofthe Corpora HQ va ko 145 2.2 Annotatingthe Data 2 Q HQ ung vo 151 2.2.1 Zero Pronouns with a Intra-sentential Antecedent 151

2.2.2 Determining the Categories 155 2.2.3 The AnnotationScheme Ặ 00002000 157

3 A Rule-Based Approach: Variations of the Hobbs Algorithm

3.1 The Hobbs Algorithm

3.2 Variations of Hobbs on Korean Zero Pronouns

3.3 Significance of Syntactic Environments

4.5 Performance Scores: A Round-Up

Conclusions

Bibliography

161162

164

168170

171

172172

175180183191196

199204

205208

209

210212

219

222

227

List of Tables

1.1

1.2

1.3

1.4

1.5

1.6

1.7

1.8

1.9

1.10

2.1

2.2

2.3

2.4

11

1.2

1.3

Dropped topic subject in Italian (cantare(x), x=lui, x=topic) 16

Overt non-topic subject in Italian (cantare(x),x=lui) 16

Overt topic subject in English (sing(x), x=he, x=topic) 17

Overt non-topic subject in English (sing(x),x=he) 17

Overt topic subject in Yiddish (sing(x), x=he, x=topic) 18

Overt topic subject in Yiddish (ranedQ) 18

Thai input with overt subject © 0 co ca 21 Thai input with Prosubject 2 LH HH KV 21 Thai input with embedded Pro subJect 21

Centering transition states 2 ko 25 Pronominal system of Korean, 1st and 2ndperson 39

Pronominal system of Korean, 3rdperson 40

Classification of Japanese zero anaphor by Kameyama (1985) 45

Classification of Korean zero pronouns 46

Pronoun resolution algorithms for New York Times 134 Pronoun resolution algorithms for fictional texts 2.0 0.0000 134 COHERE and ALIGN produce preference ranking of 4 centering transition

TYPES / aAAAaqaaaŨŨŨ 140

2.1 Zero pronoun frequencies in KTB landKTB2 147

2.2 pro and PRO frequencies in the Penn Chinese Treebank5.1 147

2.3 Zero pronoun frequencies by grammaticalroles 147

2.4 Zero pronoun frequencies bytype ce ee 148 2.5 Overt pronouns inKTB 1, lstand2nd person 149

2.6 Overt pronouns in KTB 1, 3rd person and other 149

2.7 Overt pronouns in KTB 2, lstand 2ndperson 150

2.8 Overt pronouns in KTB 2, 3rd personandother 150

3.1 Naive Hobbs algorithm on Korean zero pronouns 165

3.2 Breakdown ofantecedentNPsinKTB2 166

3.3 Hobbs algorithm on Korean zero pronouns, adverbial NPs not considered 167 3.4 Hobbs algorithm on Korean zero pronouns, argument antecedents only 167

3.5 Hobbs algorithm on Korean zero pronouns, performance by clausal groups 169 4.1 Binary classification for zero pronoun and its potential antecedent NP in 06:2182 4.2 Binary classification of coreference for zero pronounin06:2 182

4.3 Binary classification of coreference for zero pronounin06:2 183

4.4 Feature vector of coreference events for two pro subjects, partially repre-sented 2 ằẶẼ‡£<ÄẶặẶặaẻ 192 4.5 Numbers of targets/events generated per7Ør2 193

4.6 Numbers of coreferential targets/events generated perpro 193

4.7 Training and testing set sizes forthe two corpora 193

4.8 Featuresetsizes 2 nu ng Q g kg V v ki ki k va 194 4.9 Performance of models as a binary classifer 194

4.10 Confusion matrix for binary classifiers, KTBI 195

4.11 Confusion matrix for binary classifiers, KTB2 195

4.13

4.14

4.15

4.16

4.17

4.18

4.19

4.20

4.21

4.22

4.23

4.24

4.25

4.26

4.27

4.28

4.29

4.30

4431

4.32

4.33

4.34

4.35

Performance of coreference resolution models 196

Performance of coreference resolution models, averaging on 10 cross-fold validation 2 4 196 Accuracy by zero pronoun type,KTBI, 197

Accuracy by zero pronoun type,KTB2 197

Prediction pattern for anaphoric (a) zero pronouns, KTB2 198

Prediction pattern for deictic-speaker (1) zero pronouns, KTB2 199

Prediction pattern for generic (g) zero pronouns, KTB2 199

Confusion matrix for category classification model (Phase 1), KTB 1 200

Combined performance of Phase 1 and Phase2, KTB1 201

Comparison of two approaches, KTB 1 201

Confusion matrix for category classification model (Phase 1), KTB2 202

Combined performance of Phase 1 and Phase2, KTB2 203

Comparison of two approaches, KTB2 203

Numbers of coreferential NPs per NP-anaphorlcøo 207

Performance of coreference resolution models 207

Performance of coreference resolution models, averaging on 10 cross-fold validation 2 ee 207 Performance degradation by removing Íeatureset 209

Performance degradation by removing featureset 210

Performance of individual features: “nearest subject”, “nearest topic” 211

Performance of combinationoffeaties 212

Top 20 positively weighted features 2 0 0.0000 ee eee 214 Top 20 negatively weighted features 215

Performance of Maximum Entropy models on 3 clausal groups 218

Weights of 4 features: trained for pros ina matrix clause 218

4.36 Weights of 4 features: trained for pros in an adverbial clause 4.37 Weights of 4 features: trained for pros in an embedded clause 4,38 Key performance scores 2 ee

In Korean sentences, arguments such as subject and object enjoy a great degree of dom in terms of the form in which they are realized: a noun head followed by a sub-ject/object/adverbial case marker; a noun head followed by a special postposition marker,most notably the topic marker; a bare noun; a pronoun with or without a postpositionmarker; and finally, the null form Korean is only one of the group of languages whichpermits extensive use of null-form arguments These invisible arguments are believed to

free-be a substitute for pronouns! and have received terms that reflect such a view throughout

the literature, including zero anaphor, zero pronominal, and the more theory-specific pro.The term zero pronoun is adopted in this study; the abbreviated form pro is also used, butwithout implying the full set of theoretical assumptions which are commonly associated

with the term Below is a very typical example, illustrated by a question-answer pair:

Qu 42 FS ASS AA Sots?

John-un enu kwamok-ul ceyil cohaha-ni?

John-Top which subject-Acc most like-Q?

‘Which subject does John like the most?’

'See Section 2.1.1 and also Kameyama (1985), Chapter 3.3

Throughout this thesis, uj, u2, Un numbering notation is used instead of a, b, n for a sequence

of examples that constitutes a continuous discourse When such a discourse sequence involves multiple speakers, they will be denoted by A, B and so on, i.e., uy.A, ua.B.

u 0 Fae AD Sots.

— swuhak-ul ceyil cohahay.

(SBJ) math-Acc most like-Dec

‘(He) likes math the most.’

The use of zero pronouns in a discourse has been believed to be largely governed by course principles, as illustrated by the above example where the zero pronoun subject isunderstood to be referring to the entity that the subject NP of the preceding utterance de-notes In such uses of zero pronouns, saliency of their referents in the discourse context issaid to play a significant role However, zero pronouns are not always anaphoric, that is,their interpretation is not always based on some preceding /inguistic expression: those in(2a) and (2b) below are zero deictic pronouns that directly refer to an entity in the givenspatio-temporal context On the other hand, generic indefinite zero anaphors in (2c) refer

dis-to the people in general, as in some uses of English they, one and you, as well as Germanman and French on General situational zero anaphors (2d) are zero subjects that refer totime, weather, and a general situation, a usage akin to English dummy subject it

0 san-ey ka-eya 9 holangi-lul cap-ci.

(SBJ) mountain-to go-only-if (SBJ) tiger-Acc catch-PresDec

‘Only if (one) goes to the mountains (one) catches a tiger.’

d @ WA Balch.

Ú pelsse yel-si-ngi-ta.

(SBJ) already ten-o’clock-Cop-Dec

‘(It) is 10 o’clock already.’

These dropped arguments, or zero pronouns, are a remarkably frequent phenomenon

in Korean, as we will see in a later chapter Apparently, their use is not only permitted inthe language but also has to be preferred in some way, as their very occurrences clearlyattest More importantly, correct interpretation of these null forms is essential in successfulcommunication, which Korean speakers are obviously capable of doing

Given this, two questions naturally arise about Korean zero pronouns: (1) what are they,and (2) what can we do about them The former is a theoretical question, which seeks toaddress the fundamental motivations and conditions behind the collective phenomenon ofKorean zero pronoun The latter is from an engineering point of view, which is aimed atfinding out how successfully one can replicate the human understanding of these pronouns

by utilizing the knowledge gained from answering question (1)

The goal of this thesis is, then, twofold One is to gain better understanding of this

highly context-oriented phenomenon by observing them in naturally occurring discourse

In the course of doing so, we aim to discover the distinctive types and distributive patterns

of Korean zero pronouns and propose a classification and annotation scheme based on thefindings The other goal is to build a system that can resolve zero pronouns in Korean,

by either assigning them to correct categories or identifying antecedent NPs in the contextfrom which their reference can be derived In what follows, the two goals are presented in

two separate parts.

Part I

Analysis

Chapter 1

Previous Work on Zero Pronouns

The phenomenon of zero pronouns first entered the scene of heated theoretical debate when

it was given a critical role in Government and Binding Theory by Chomsky (1981, 1982,1986) Under the tenet of Universal Grammar (UG), the theoretical issue at question was

what parameters in the UG allow zero pronoun subjects to occur in only certain languages

Much of the work that followed within the GB theory framework sought to precisely

cap-ture the cross-linguistic variation pattern of zero pronouns by refining and reformulatingthe theory of Empty Categories (EC), of which the zero pronoun is one, by focusing on thesyntactic behavior of zero pronouns in particular language settings

The 1990’s saw the surge of Optimality Theory, initiated by Prince and Smolensky(1993) for the sub-discipline of phonology, which was then adopted for the topic of zeropronouns Again focusing on the cross-linguistic variation of their syntactic admissibility,researchers sought to generalize such cross-linguistic variations in the forms of ordered andviolable constraints in the Universal Grammar

While the two schools of research were essentially concerned with syntactic conditions

of zero pronouns, the need to look beyond the syntactic levels and to the larger domain ofdiscourse to ensure proper semantic interpretation of the pronoun category were being rec-

ognized within the formal semantics and pragmatics community The theories of DynamicSemantics and Discourse Representation Theory (Heim, 1982; Kamp and Reyle, 1993)were born out of such a realization, which provided formalization of semantic interpreta-tions of sentences within a discourse which are linked by anaphoric relations of the nounexpressions involved.

Centering Theory (Grosz, Joshi and Weinstein, 1983, 1995, and many others) furthersthe notion of involvement of discourse factors in the use of pronouns, and casts informationtheoretic notions at the center of their theory of pronoun usage Zero pronouns in manylanguages have been studied extensively within the Centering framework, under the beliefthat zeros in these languages are much less constrained grammatically, but are governedrather by pragmatic and specific contextual considerations in determining whether a givenzero is linguistically or non-linguistically bound

1.1 Previous Sentence-Level Work on Zero Pronouns

Ever since Chomsky introduced the term pro in his theory of government and binding(1981, 1982, 1986), many languages which allow the subject of the sentence to be emptyhave been collectively referred to as pro-drop languages In this section, we present asurvey of the syntactic literature on the general phenomenon of zero pronouns in some ofthe prominent pro-drop languages including Korean

1.1.1 pro in Government and Binding Theory

In Government and Binding Theory (Chomsky, 1981, 1982, 1986), zero pronouns areanalyzed as a case of empty categories In this theory, nominal empty categories alongwith their overt counterparts can be represented as a combination of two binary features:+pronominal and +anaphor, as follows:

(3) The relationship of overt and empty categories

overt empty anaphor pronominal

reflexive pronouns NP-trace + —

personal pronouns pro — +

r-expressions wh-trace — —

N/A PRO + +

The theory was essentially concerned with the referential properties of such nominal ements, overt or empty Three Binding Principles were given as governing their distribution(Chomsky, 1986):

el-(4) The Binding Principles

e A an anaphor is bound in a local domain

e 8B.a pronominal is free in a local domain

e C anr(eferring)-expression is free

where binding 1s defined as:

a binds @ if and only if (1) œ c-commands @ and (2) a and @ are co-indexed

Chomsky (1981) divides the non-pronominal empty category (EC) to introduce yetanother EC, namely the variable The variable is locally bound by an element that isnot in an argument position (i.e., subject, object, etc.), while an anaphor is locally bound

by an argument:

(5) a An EC is a pronominal if and only if it is free or locally bound by an element

with an independent thematic role, and a non-pronominal otherwise

b A non-pronominal EC is an anaphor if and only if it is locally A-bound, and avariable if locally A-bound.

Reflexive pronouns (+anaphor, —pronominal) and personal pronouns (—anaphor, inal) are known to exhibit complementary distribution in their ability to co-refer with otherelements in the sentence, as shown in contrasting examples (6a) and (6b) below The in-sight is captured by the Binding Principles A and B It also appears that referring expres-sions such as names can never have c-commanding antecedents anywhere in the sentence,illustrated with (6c), which is captured by the Principle C.

+pronom-(6) a Bill, likes himself;,,;

b Bill, likes him,;;

c *He; thinks Bill likes John

The last category PRO has a positive value for both anaphor and pronominal features,which derives from the fact that it sometimes behaves like +anaphoric elements such asreflexive pronouns (in a) and like pronominal one in others (in b):

(7) a John; wants PRO,/himself; to win

b PRO/for one to leave would be a mistake

This leads to an apparent contradiction where PRO has to be both bound and free in itsgoverning category It is solved by simply stipulating that PRO lacks a proper govern-ing category This null element PRO exhibits fairly complex referential and distributionalproperties Unlike other types of nominals, it is restricted in its distribution to the position

of subject of an infinitive, as shown in the contrasting examples (a), (b) and (c) below Insome contexts, PRO is obligatorily coindexed with a particular c-commanding NP in thehigher clause, in some cases the subject NP (in a, d) and the object NP in others (in e),which is believed to be decided by the lexical property of the matrix verb Yet in other con-texts, no such obligation for coindexation exists In (f) where PRO occurs in an infinitivalsubject clause, and in (g) with PRO in an interrogative object clause, the PROs lack specificreference and are understood instead as arbitrary

(8) (examples from Harbert (1995))

a John, tried PRO; to win

b *Mary hoped for the president to apoint PRO

c *I believe that PRO will stay home

d Mary; promised me PRO; to be home early

e John forced me; PRO; to stay late

f PRO,;; to leave early would be inexcusable

g John doesn’t know how PRO„;; to behave oneself at parties

pro, defined as the empty nominal element with features +pronominal and —anaphor,was established from the observation that some languages such as Italian and Spanish allowzero subjects in declarative sentences, while other languages like English cannot This factreflects one parameter of Universal Grammar, it was argued, which is that languages divideinto two sets in relation to the availability of null subjects (Perlmutter, 1971) A language inwhich pro exists is called a pro-drop language As the classification table (3) above shows,zero pronouns and regular pronouns were thought to be essentially the same, except thatthe former lack phonetic contt An example in Spanish:

(9) Dijo que [e] vendria mañana

he said [e] would-come tomorrow

‘He said (he) would come tomorrow.’

The typical co-occurrence of pro with rich inflectional systems caught attention early

on, and the explanation given was that pro is licensed only by a rich enough INFL(inflection)node, richness being normally dependent on the presence of a person specification (Tarald-sen, 1980; Rizzi, 1986; Borer, 1986) Hence, pro was seen as obligatorily requiring propergovernment (by the AGR (agreement) node) The differences between a non-pro-drop

language like English and a drop language like Italian or Spanish is that only in languages is the INFL node a proper governor for the subject The rationale behind thistheorizing is clear: a rich inflection system renders a zero subject “recoverable” while

pro-a poor one does not Furthermore, subject-object pro-asymmetry wpro-as being widely reportedacross languages in allowing zero realization of an argument, which was believed as ren-dering a supporting piece of evidence to the claim: only the subject position enters into

an agreement relation with the verb in Italian and Spanish, hence pro is unavailable in theobject position in these languages Also, Huang (1989) finds further supporting evidence inPashto, which allows null object pronouns in perfect aspect sentences where there is objectagreement morphology on the verb

This theoretical picture soon proved falling short on two accounts First, the binarydistinction between pro-drop vs non-pro-drop is too simplistic a typology: there is a widevariability in which pro-drop languages allow only certain types of subjects to drop Forexample, semantic properties of subjects proved to be relevant: pro-drop in languages such

as German and Icelandic is limited to pleonastic subjects, while Italian and Spanish allowzero realization for both referential pronouns and pleonastic types (Safir, 1985):

(10) a *@ will zu Hause bleiben

Ú want at home to-stay.

‘I(*pro) want to stay home.’

b @klar ist, daB er nicht kommen wird.

Qclearis thathenot come will.

‘(It) is clear that he will not come.’

Second, the claim that pro is only available in subject positions in languages with ject agreement is disputed in Rizzi (1986) He argues that null objects must be posited inthe syntactic representation of sentences like the following in Italian:

sub-(11) II bel tempo Imvoglia [pro] a [PRO restare]

the good weather induces # to _ stay.

“The good weather induces (pro) to (PRO) to stay.’

Lastly, it was pointed out that a group of languages, including Chinese, Japanese andKorean, allow zero subjects to occur rather freely, even though their systems for markingagreement are even more impoverished than that of English Huang’s (1983, 1984, 1989)work on Chinese pro distribution was aimed at adapting GB theory’s original take on theproblem so that the zero-pronoun behavior in these languages is properly explained Hiswork is examined closely in the next section

1.1.2 Huang’s (1983, 1984, 1989) Work on Chinese, Japanese and

Ko-rean pro

Huang draws from from Li and Thompson’s (1976) distinction of “topic-prominent” and

“subject-prominent” languages, and proposes a Universal Grammar parameter called topic for Chinese, Japanese and Korean in order to formulate the zero-pronoun behaviorsdisplayed by them In these topic-prominent languages which include Chinese, Japaneseand Korean, structural subjects are not a basic requirement of the sentence Furthermore,

zero-in such languages sentences of the form topic-comment abound and must count as basicforms in that they cannot be plausibly derived from other “more basic” forms Huang’smain argument is that most zero subjects in these languages are in fact a zero topic Morespecifically, he proposes that the relevant phenomena be derived jointly by the followingset of principles and assumptions:

(12) a The principle of recoverability

b The assumption that a zero pronoun is a pronoun

c The assumption that the agreement-marking AGR on a verb qualifies as a tential “antecedent” of a zero pronoun

po-d The binding theory of Chomsky (1981), in particular the condition of disjointreference (DJR) or condition (B)

e The Generalized Control Rule (GCR)

All of the items above are readily accepted ideas from the standard GB theory, exceptfor the last item, The Generalized Control Rule (GCR), which is Huang’s own contribution

It is basically Chomsky’s (1980) rule of control extended to cover both pro and PRO Huangclaims that the two are essentially the same entity and proposes to abolish the distinctionbetween the two categories originating in Chomsky (1982) by subjecting them to the singlerule of GCR:

(13) a Generalized Control Rule (GCR):

an empty pronominal is controlled in its control domain (if it has one)

b Control Domain:

œ 1s the control domain for ( if it is the minimal category satisfying (1) and (ii):(i) a 1s the lowest S or NP containing

(a) @, or

(b) the minimal maximum category containing Ø;

(ii) œ contains a SUBJECT accessible to 3

The key insight for collapsing the two categories comes from the observation that theabsence of subject-verb agreement plays a crucial role for both: null subjects are allowed inChinese which has no subject-verb agreement; in English, PRO only occurs in embeddedclauses that are not governed by AGR, a syntactic incarnation of the agreement feature inthe GB framework After examining various examples and evidence cross-linguistically,

he concludes that pro and PRO are indeed one entity which is purely pronominal

His other major theoretic contribution to the discussion of zero pronouns in the threeEast Asian languages is introducing the notion of topic in accounting for some of the zero-

pronoun phenomena that were not otherwise easily incorporated into the syntactic theory

of the GB More specifically, he follows Tsao (1977) in positing the process of “topic NPdeletion” for these languages, and argues that most zero subjects are instances of such topic

NP deletion Moreover, to explain the availability of zero objects in these languages, heargues that these are not zero pronouns but in fact variables locally A-bound by the zerotopic NP, whose reference is fixed in discourse

Kim (2003) examines Huang’s work and points out that it fails in three points First,Huang sets apart zero objects in embedded clauses from other zero pronouns and claimsthat they are variables bound by topic One of its implications is that they cannot be coin-dexed with subject or objects in the main clause (example in Chinese):

(14) a *Zhangsan, zhidao [Lisi mei banfa shuifu e;]

Zhangsan know Lisi no method persuade e;

“*Zhangsan; knows Lisi cannot persuade him,.’

b [TOP e,] Zhangsan zhidao [Lisi mei banfa shuifu e;]

[TOP e;] Zhangsan know Lisi no method persuade e;

‘Zhangsan knows that Lisi cannot persuade him,/her;.’

However, counter-examples are not difficult to contruct in Korean:

(15) &%9l7l €Ẵ@9]? 0, ARAL sac.

Toli;-ka Swuni-ka 9; kwoylophi-ess-ta-ko malha-ess-ta.

Toli;-Nom Swuni-Nom (OBJ;) tease-Past-Dec-Quote say-Past-Dec

‘Toli; said Swuni teased (him,).’

Second, zero-topic does not always coincide with the topic of the preceding utterance,

as Huang claims For example:

(16) m TA) Oh ALE [294 BS naJ,E AR PAT.

ecey apeci;-nun [kangaci han mali];-lul sa OSi-ess-ta

yesterday father;-Top [puppy one CLASS];-Acc buy come-Past-Dec

“Yesterday father; bought a puppy;.’

ở" 0„; Bol Si Beet.

Ú.,; tel-i hayah-ko manh-ess-ta.

(SBJ,;/;) fur-Nom white-And abound-Past-Dec

“(Ít¿/; was such that) fur was white and plenty.’

7) w 32¿E s42) eal Aol] 3sl,s REGATH.

Youngswu;-nun ecey tungkyokil-ey Younghi;-lul manna-ess-ta.Youngswu,-Top yesterday way-to-school-on Younghi;-Acc meet-Past-Dec

“Youngswu; met Younghi, on his way to school.’

Ug Dsi/j oj A 3] of) cy.

Dsi/; yecenhi yeyppu-ess-ta.

(SBJ,;/;) still pretty-Past-Dec

‘(She,;/;) was still pretty.’

Third, zeros in subject and object positions can be indexicals, a use of zero pronounsHuang does not address at all:

1.1.3 Optimality Theory Approaches to Zero Pronouns

According to Optimality Theory, “a language-particular grammar is a means of ing the conflicts among universal constraints” (Prince and Smolensky, 1993) Grimshaw(1997), Grimshaw and Samek-Lodovici (1998) are an attempt at capturing the cross-linguisticvariation in the distribution of subjects in terms of such competing universal constraintsregarding their expression Different languages resolve conflicts by employing different

resolv-ranking among the constraints, hence language variation arises They suggest the ing set of five constraints regarding expression of subjects:

follow-1 SUBJECT: The highest A-specifier in an extended projection must be filled (Grimshaw,1997) Failed by clauses without a subject in the canonical position

2 FULL-INT: (Full-Interpretation) Parse lexical conceptual structure (Grimshaw, 1997).Failed by expletives and auxiliary do

3 DROPTOPIC: Leave arguments coreferential with the topic structurally unrealized.Failed by overt constituents which are coreferential with the topic

4 ALIGNFOCUS: Align the left edge of focus constituents with the right edge of amaximal projection Failed by non-aligned foci

5 PARSE: Parse input constituents Failed by unparsed elements in the input

They first focus on the so-called pro-drop languages that allow referential dropped jects, most notably Italian Citing numerous previous works which show that the null

sub-subjects in these languages are possible only when licensed by a discourse antecedent with

topic status (Samek-Lodovici, 1996; Calabrese, 1986; Di Eugenio, 1990, 1995; Vallduvi,1992), they suggest the high-ranking status of the constraint DROPTOPIC in these lan-guages is responsible for the phenomenon The constraint requires a subject coreferentialwith the topic to be structurally unrealized; when a subject has no antecedent or its an-

tecedent is not a topic, the subject is obligatorily overt More specifically, they propose thatpro-drop follows from the constraint DROPTOPIC requiring arguments whose antecedent is

a topic to be unexpressed, or structurally “unparsed”, with the penalty of violating PARSE

as a result The constraint DROPTOPIC outranks the two constraints PARSE and SUBJECT

in these languages, which require subjects to be overtly expressed:

(19) Ranking for languages allowing referential null subjects:

DROPTOPIC >> PARSE >> SUBJECT

The tableau below illustrates the case of a grammatical null subject in Italian which iscoreferential with at topic in the sentence The candidate (a) ha cantato wins over the othercandidates (b) and (c), Jui ha cantato and ha cantato lui, which have overtly expressedpronoun subjects Note that FULL-INT is always satisfied since no candidates involveexpletive pronouns

| Candidates |, DROPTOPIC | PARSE | FULL-INT | SUBJECT

b lui ha cantato *Ị

c ha cantato lui *I

Table 1.1: Dropped topic subject in Italian (cantare(x), x=lui, x=topic)

Compare it with the following case of the same proposition, which, although not apparentwith the absence of surrounding context, is presumed to have a non-topic-marked subject.Here, both (b) Jui ha cantato and (c) ha cantato lui satisfy the highest-ranked DROPTOPICconstraint, and the former ultimately wins over the latter due to the fact that it does notviolate the SUBJECT constraint by “moving up” the subject to the highest specifier position,thereby bringing it up to the front of the sentence

| Candidates DROPTOPIC | PARSE | FULL-INT | SUBJECT

a ha cantato *|

t= b lui ha cantato

c ha cantato lui *|

Table 1.2: Overt non-topic subject in Italian (cantare(x), x=lui)

On the other hand, in languages such as English where subjects are required to beovert on all occasions, the relative ranking between DROPTOPIC and PARSE is reversed, soPARSE dominates DROPTOPIC:

(20) Ranking for English where null subjects are disallowed:

Table 1.3: Overt topic subject in English (sing(x), x=he, x=topic)

| Candidates PARSE | DROPTOPIC | SUBJECT | FULLINT |

a has sung *| j1 ®*

s# b, he has sung

c has sung he

Table 1.4: Overt non-topic subject in English (sing(x), x=he)

Grimshaw and Samek-Lodovici consider the case of Yiddish and Icelandic which

man-date zero realization of expletive subjects but do not allow topic-marked subjects to drop,

and propose the following ranking to generate such a system:

(21) Ranking for Yiddish and Icelandic with no expletives and no referential

pro-drop:

PARSE >> FULL-INT >> SUBJECT >> DROPTOPIC

First, the unavailability of topic-pro-drop indicates that PARSE outranks DROPTOPIC inthese languages A Yiddish example, where the subject is marked as the topic is given inTable 1.5 Also, obligatory zero-realization of the subject position in the two languageswhen the verb has propositional content can be captured by ranking FULL-INT higher thanSUBJECT As illustrated by Table 1.6, candidate (b) with an expletive subject violates

higher-ranked FULL-INT, while the optimal candidate (a) violates SUBJECT by having anon-realized subject position.

| Candidates || PARSE | FULL-INT | SUBJECT | DROPTOPIC |

a sung *!

t= b he sung

c expl sung *|

Table 1.5: Overt topic subject in Yiddish (sing(x), x=he, x=topic)

| Candidates || PARSE | FULL-INT | SUBJECT | DROPTOPIC |

t= a rained

b it rained *

Table 1.6: Overt topic subject in Yiddish (rainedQ)

Another notable OT-theoretic account of cross-linguistic variation of zero pronouns can

be found in Speas’s (2001) work She aims to cover some slightly different aspects of linguistic empirical data: Grimshaw and Samek-Lodovici’s (1998) zero pronouns were lim-ited to subject pronouns occurring in finite clauses, namely pro in the classic Chomskyan

cross-sense; Speas (2001), on the contrary, is mainly concerned with the availability of subject,

and more importantly, object pronouns,! in both finite and non-finite clauses She claims

that languages fall into three types with respect to the two parameters:

(22) The distribution of null pronouns

Type 1: Pro = subject of nonfinite clause

Type 2: Pro = subject of nonfinite clause and subject of finite clause

Type 3: Pro = subject of nonfinite clause, subject of finite clause, and object

Nonexistent: Pro = object only

‘Huang (1984) originally argued that null objects in Mandarin are variables bound by a null operator, and

not null pronouns, and proposed that object pro is universally impossible The claim has since been disputed for languages such as Korean (Yoon, 1985) and Thai (Hoonchamlong, 1991), as was done by Speas (2001).

Type 1 languages include English; Type 2 includes languages such as Spanish andMandarin Chinese, which allow subject Pro but not object Pro; Type 3 includes Thai andKorean, which have the most generous use of zero pronouns in all possible environments;the last type, not found among natural languages, permits the zero form in object positionsand not in subject positions.

Most notably, she posits Pro, the only kind of null pronoun in her theory, which is meant

to replace both pro and PRO in the standard accounts of GB theory She notes that manyauthors, including Bouchard (1984), Sportiche (1983), Manzini (1983), Borer (1989) andHuang (1984), have pointed out problems inherent in maintaining the standard distinctionbetween zero forms in subject position of infinitives and gerunds (PRO) from those in otherpositions (pro), and adopts the view of Borer and Huang in that the apparent distinction

is simply an artifact of the interaction of several different principles The typology oflanguages given above in (22) follows this theoretic assumption, and she argues that thecross-linguistic distribution and interpretation of Pro follows from different rankings ofviolable constraints in OT given below:

(23) 1 CONTROL (CTL): A featureless pronoun must be coindexed with the closest

c-commanding nominal element in its C(Control)-Domain

2 BINDING THEORY PRINCIPLE B (BTB): A pronoun must be free in its ing Domain

Bind-3 MAX(PRO): If Pro occurs in the input, then its output correspondent is Pro

4 AGR(X): +Tense has AGR features

5 No PHI-FTS: Avoid AGR features

6 CASE: An NP with phi features must receive Case

Instead of giving a full review of her theoretic accounts of the four language types, I willillustrate how the zero-pronoun distribution of Type 1 languages, of which Korean is one,

is derived in her theory The relevant constraints for this type of languages are the first threeabove, CONTROL, BINDING THEORY PRINCIPLE B (BTB), and MAX(PRO) The CON-TROL constraint has the effect of forcing a Pro, assumed featureless in the theory, to have

a closely located antecedent; BTB balances out CONTROL in that it prevents a pronounfrom having an antecedent within a binding domain MAX(PRO) is a key constraint andneeds some elaboration Unlike Grimshaw and Samek-Lodovici’s (1998) approach, Speasassumes that whether a pronoun is overt or not is specified in the input, which can then

be altered in the output by violating the MAX(PRO) constraint Recall that in Grimshaw

and Samek-Lodovici (1998) the input has information as to the topic status of subjects butdid not include information addressing the form of the subject; the form that a subject as-

sumes in the output is controlled by violation/non-violation of the DROPTOPIC constraint

Both approaches in effect capture competition between candidates with a zero/overt ject; Speas’s (2001) approach, however, does so while avoiding directly addressing the

sub-pragmatic aspects behind the choice of pronominal forms, by formulating it instead as a

constraint that oversees truthfulness between the Pro input and the output

Accordingly, the two tableaux below illustrate two contrasting cases of Thai, one with

an overt pronoun subject and the other with a Pro subject:

(24) a khaw hen Nit

he/she saw Nit

‘he/she saw Nit’

b Pro hen Nit

Pro saw Nit

‘(he/she) saw Nit’

For the example below with a zero subject in an embedded clause, the ranked set ofconstraints correctly predicts that the zero pronoun can be coreferential with the matrix

subject:

b khaw hen Nit *

Table 1.8: Thai input with Pro subject

(25) Nitboo waa Pro hen Noy

Nit speak say Pro say Noy

‘Nit said that he/she/they saw Noy’

s a, Nit boo waa Pro hen Noy ;

t# b, Nit; boo waa Pro; hen Nay

c Nit boo waa khaw hen Noy *Ị

Table 1.9: Thai input with embedded Pro subject

Both Grimshaw and Samek-Lodovici’s (1998) and Speas’s (2001) work sought to count for cross-linguistic distribution of zero pronouns by focusing on the syntactic avail-ability of various types of zero pronouns The former work tapped into the domain ofpragmatics by introducing such notions as “topic” into the framework, but it was largelyrepackaged as a syntactic element, thereby exempting the theory from looking beyondthe level of syntax The accounts largely focus on whether or not a certain realization ofpronominal element is permissible in a language, and as result they are able to offer onlylimited glimpses of how such pronominal elements are interpreted, as example (25) andtable (1.9) show Following these works, there have been numerous studies which aim atcasting OT as a full system of pronoun interpretation, and ultimately, anaphora resolution

ac-Rather than focusing on explicating cross-linguistic variation of zero pronoun distribution,these works aim to account for how an optimal candidate is chosen among competing in-terpretations of a pronoun within a particular language We will return to these works later

in Section 1.3 of Part II

12 Centering Theory: A Discourse-Oriented Approach

to Pronouns

So far we discussed how the zero pronoun was first perceived as a sentence-level nomenon and therefore a syntactic problem, only to earn recognition later as a problemthat is pragmatic in nature Centering Theory, grounded within the pragmatics discipline,provides a theoretical framework which enables the study of pronouns in a larger domain

phe-of discourse In this section, I present a brief review phe-of the theory and related works thatapply it to the problem of resolving pronouns, including zeros

1.2.1 The Centering Theory: An Overview

The Centering model (Grosz, Joshi, and Weinstein, 1983, 1995) is intended to capture therelationship between local coherence and the use of referring expressions Centering hasits computational foundations in the work of Grosz and Sidner (Grosz, 1977; Sidner, 1979;Grosz and Sidner, 1986) and was further developed by Grosz, Joshi and Weinstein (Grosz,Joshi and Weinstein, 1983; Grosz, Joshi and Weinstein, 1986; Joshi and Weinstein, 1981)

In Grosz and Sidner’s theory of discourse structure, discourses can be segmented based

on intentional structure, and a discourse segment exhibits both local and global coherence.Global coherence depends on how each segment relates to the overall purpose of the dis-course; local coherence depends on aspects such as the syntactic structure of the utterances

in that segment, the use of ellipsis, and the choice of referring expressions.

In modeling local coherence, Centering requires two constructs: a single looking center and a list of forward-looking centers Centers are semantic entities thatare part of the discourse model It is formalized as a system of constraints and rules thatgovern the interpretation of these centers It is assumed that discourses are composed ofconstituent segments (Grosz and Sidner, 1986), each of which consists of a sequence ofutterances Each utterance in a discourse has associated with it a set of discourse entitiescalled FORWARD-LOOKING CENTERS, Cf There is a unique special member in this setcalled the BACKWARD-LOOKING CENTER, Cb The Cb is the discourse entity that theutterance most centrally concerns, which also links the current utterance to the previousdiscourse The set of FORWARD-LOOKING CENTERS, Cf, is ranked according to discourse

backward-salience Grosz, Joshi and Weinstein (1995) notes that the items in the Cf list have to

be ranked according to a number of factors including grammatical role, text position, andlexical semantics and postulates the following ranking for Cfs in English, which is solelybased on grammatical roles:

(26) Cfranking for English by grammatical roles

subject > object(s) > other(s)

As we will see later, many varied versions of Cf rankings have been proposed for a number

of different languages, which typically utilize factors other than simple grammatical roles

to reflect language-specific aspects of Cfs Walker, Iida and Cote (1994) hypothesizes thatthe Cf ranking criteria are the only language-dependent factors within the centering model

In the Cf ranking, the highest ranked element is designated as the PREFERRED CENTER,

Cp The ranking reflects the assumption that Cp(U;), the PREFERRED CENTER of thecurrent utterance U;, will most likely be the Cb(U;+1), the BACK WARD-LOOKING CENTER

of the next utterance U;,1, and therefore represents a prediction about the Cb of the next

utterance The actual realization of Cb of the next utterance, of course, may or may not bearout this prediction: the BACK WARD-LOOKING CENTER of the next utterance Cb(Uj41) isdefined as the highest-ranked Cf in the current set of Cfs (e.g., Cf(U;)) that is actuallyrealized in the next utterance, and therefore may or may not coincide with the Cp(U;).With regard to the relation between these elements, Cfs, Cb and Cp, the theory of centeringspecifies the following set of constraints:

(27) Constraints:

For each utterance U; in a discourse segment Uj, , Un:

1 There is precisely one backward-looking center Cb

2 Every element of the forward centers list, Cf(U;), must be realized in U;

3 The center, Cb(U;), is the highest-ranked element of Cf(U;_ ¡) that is realized

There is no overt reference to Cp in this formulation, but (27 2.) implies the followingrelation between Cb(U,) and Cp(U;_1): “Cb(U;) is Cp(U;_1) if Cp(U;_1) is among Cf(U;)”.Since Cb and Cp serve the mirror-opposite roles, that is, one of linking the currentutterance to the previous discourse and one of signifying the preferred choice for the center

of the next utterance, the Cp and Cb in an utterance coupled together form the transitionstatus of attention across pairs of adjacent utterances Also, whether or not the currentutterance inherits the BACKWARD-LOOKING CENTER from the previous utterance forms

another axis in the following four-way distinction of transition types’:

These transitions differ from each other according to whether BACK WARD-LOOKING TERs of successive utterances are identical or not (CONTINUE, RETAIN vs SHIFT), andwhether they match the most highly ranked element of the current FORWARD-LOOKING

CEN-?The distinction between the two types of SHIFT was first introduced in Brennan, Friedman and Pollard

(1987).

Cb(U,)=Cb(U,_¡) | Cb(U;) # Cb(U;_¡)

or Cb(U;_¡) =[?]

Cb(U,) = Cp(U,) CONTINUE SMOOTH-SHIFTCb(Ö,) z Cp(U,) RETAIN ROUGH-SHIFT

Table 1.10: Centering transition states

CENTER list, the Cp(U;), or not (CONTINUE, SMOOTH-SHIFT vs RETAIN, ROUGH-SHIFT).Informally, the former factor encodes whether or not the speaker carried over the center of

the previous utterance into the current one by making the current utterance about the same

entity The latter signals the speaker’s intention as to whether or not she/he will continue to

talk about the same entity in the next utterance Some transitions between discourse

seg-ments are more coherent than others; discourse segseg-ments that continue centering the sameentity are more coherent than those that repeatedly shift from one center to another Theseobservations are encapsulated in two rules:

(28) Rules:

For each U; in a discourse segment U;, , Um:

1 Ifsome element of Cf(U;_1) is realized as a pronoun in Uj, then so is Cb(U;)

2 Transition states are ordered CONTINUE is preferred to RETAIN is preferred

to SMOOTH-SHIFT is preferred to ROUGH-SHIFT

Rule (1) captures the intuition that pronominalization is a mechanism by which course salience is conveyed Rule (1) implies that if there are multiple pronouns in anutterance, one of them must be the Cb Furthermore, it follows that if there is only one

dis-pronoun, then it must be the Cb Rule (2) states two ordered preferences: first, having thesame Cb as the previous utterance’s (Cb(U;)=Cb(U;_1)) is preferred; second, the prospect

of continuing with the same Cb in the next utterance (Cb(U;) = Cp(U;)) is preferred Thisordered preference for transition states represents the differing levels of local coherence