KOREAN ZERO PRONOUNS: ANALYSIS AND RESOLUTION

Na-Rae Han

A DISSERTATION

in

Linguistics

Presented to the Faculties of the University of Pennsylvania in Partial

Fulfillment of the Requirements for the Degree of Doctor of Philosophy

2006

Ellen F. Prince
Supervisor of Dissertation

Martha Palmer
Co-Supervisor of Dissertation

Eugene Buckley

Graduate Group Chairperson

To my dear brother Baek-Kyoung.

ii
Acknowledgements

First and foremost, I thank my two advisors, Ellen Prince and Martha Palmer. I am eternally

indebted to them for their encouragement and intellectual and spiritual guidance, which
shaped 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 for
her sharp intellect and wisdom, as well as her unparalleled zest for life. I owe Martha not

only for her generous financial support over the years, but also for her guardianship she
truly 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 the

greatness 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 not
unlike 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 their

mentorship during my times there.

I would like to thank Justin Mott, Cassie Creswell, Tom Morton, Andy Schein, Sham
Kakade, and Heejong Yi, with whom I shared the most cherished memories of my graduate
school 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.

iii
 There are many other friends and colleagues at Penn, who over the years made it much
more 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 and
Jiyoon Lee.
My friends and mentors from my previous school, the linguistics program at Seoul
National University, have been always there for me, sometimes in the States and other

times across the Pacific Ocean. I thank Jinyoung Choi, Seunghun Lee and Hyunjoo Kim
for their friendship (and also for being great juniors to me); Yoonshin Kim for being my
close friend (and a great senior); Sookhee Chae, Shijong Ryu and Chulwoo Park for their
kind mentoring. And finally Prof. Chungmin Lee for the very first linguistics class I took

and for guiding me into the field. He taught me during my SNU linguistics years, and he
encouraged me to pursue this degree in the U.S.
Also many special thanks to Gene Buckley and Amy Forsyth for their wonderful help
regarding 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 crucial
mistakes. I cannot forget Mrs. Carole Lingle, whose warm presence filled the department
office in earlier years of my graduate school. Staff at IRCS also has been the quiet yet
essential 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 constant
support and encouragement none of my academic goals would have been realized. Their
faith in me and my love for them kept me going at times of difficulty and anguish.

iv
 ABSTRACT

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 are
those that depend on other linguistic expressions for their reference: such text-dependent
types include anaphoric and discourse-deictic zero pronouns. Other text-independent types

are deictic zero pronouns, generic and specific indefinite zero pronouns and situational
zero pronouns, although it is possible for some of these text-independent types to enter
coreferential 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 and
interpretation. While the general insight holds true, various efforts to encapsulate it by
way of precise formulation of Cf-ranking or other hierarchies fall short, largely due to the
fact that the notion is encoded by heterogeneous linguistic factors whose relations cannot
be expressed single-dimensionally. From a language processing point of view, the diverse

nature of Korean zero pronouns presents the unique challenge of blending the tasks of cat-
egorization and identification of their antecedents. In this dissertation, using Maximum
Entropy as the machine learning method of choice, various statistical models for Korean
zero pronoun resolution have been successfully trained and tested on two Korean Treebank

corpora. 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-

v
guistic properties surrounding zero pronouns and their potential antecedents. The features
found to have a particularly strong contribution are indeed those that encode the linguistic
aspects 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. While
the 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.

vi
Contents

Acknowledgements iii

Abstract v

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 pro in Government and Binding Theory . . . . . . . . . . . . . . . 6
1.1.2 Huangs (1983, 1984, 1989) Work on Chinese, Japanese and Ko-

rean pro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
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

vii
 1.2.3 The Zero Pronoun and Centering . . . . . . . . . . . . . . . . . . . 30

2 Analysis of Korean Zero Pronouns 35

2.1 Defining the Object of the Study . . . . . . . . . . . . . . . . . . . . . . . 35
2.1.1 Why Zero Pronouns . . . . . . . . . . . . . . . . . . . . . . . . 35
2.2 Overt Pronouns in Korean . . . . . . . . . . . . . . . . . . . . . . . . . . 38

2.3 The Problem of Identification: Where to Find the Invisible . . . . . . . . . 43

2.4 Korean Zero Pronouns by Reference Types . . . . . . . . . . . . . . . . . 45
2.4.1 General Situational Zero Pronouns . . . . . . . . . . . . . . . . . . 46
2.4.2 Deictic Zero Pronouns . . . . . . . . . . . . . . . . . . . . . . . . 51

2.4.3 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 Distinction . . . . . . . . . . . . . . . 76

3 The Centering Theory and Korean Zero Anaphora 80

3.1 Criteria for Cf Ranking: What Encodes Salience? . . . . . . . . . . . . . . 81
3.2 Establishing Cf Ranking for Korean . . . . . . . . . . . . . . . . . . . . . 86

3.3 Topic-Marked NPs vs. Zero Pronouns . . . . . . . . . . . . . . . . . . . . 99

3.4 The Centering Theory and Zero Pronoun Resolution . . . . . . . . . . . . . 115

viii
II Resolution 118

Overview 119

1 Previous Work 121

1.1 Pronoun Resolution: An Overview . . . . . . . . . . . . . . . . . . . . . . 121
1.1.1 Early Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
1.1.2 The Traditional Approach . . . . . . . . . . . . . . . . . . . . . . 123
1.1.3 The Statistical Approach . . . . . . . . . . . . . . . . . . . . . . . 124

1.1.4 The Knowledge-Poor Approach . . . . . . . . . . . . . . . . . . . 125

1.1.5 Resolution of Zero Pronouns: The Case of Spanish . . . . . . . . . 126
1.2 Centering Theory and Pronoun Resolution . . . . . . . . . . . . . . . . . . 128
1.2.1 Brennan, Friedman and Pollard (1987) . . . . . . . . . . . . . . . . 128

1.2.2 Strubes (1998) S-list Approach . . . . . . . . . . . . . . . . . . 129

1.2.3 Left-Right Centering by Tetreault (1999) . . . . . . . . . . . . . . 131
1.2.4 Resolution of Zero Pronouns Using Centering: The Case of Thai . . 134
1.3 Optimality Approaches to Anaphora Resolution . . . . . . . . . . . . . . . 137

1.3.1 Beavers (2002) Centering in OT . . . . . . . . . . . . . . . . . . . 138

1.3.2 Hong (2002) and Kim (2003): OT-Based Korean Anaphora Reso-
lution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

2 The Data: The Penn Korean Treebank Corpora 145

2.1 Overview of the Corpora . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
2.2 Annotating the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

2.2.1 Zero Pronouns with a Intra-sentential Antecedent . . . . . . . . . . 151

2.2.2 Determining the Categories . . . . . . . . . . . . . . . . . . . . . 155
2.2.3 The Annotation Scheme . . . . . . . . . . . . . . . . . . . . . . . 157

ix
3 A Rule-Based Approach: Variations of the Hobbs Algorithm 161
3.1 The Hobbs Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
3.2 Variations of Hobbs on Korean Zero Pronouns . . . . . . . . . . . . . . . . 164

3.3 Significance of Syntactic Environments . . . . . . . . . . . . . . . . . . . 168

3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

4 A Maximum Entropy Reference Resolution System 171

4.1 The Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
4.1.1 Classification and Resolution . . . . . . . . . . . . . . . . . . . . . 172
4.1.2 Two-Phased vs. Single-Phased System . . . . . . . . . . . . . . . 175

4.1.3 Maximum Entropy as a Ranking Method . . . . . . . . . . . . . . 180

4.1.4 Selection of Features . . . . . . . . . . . . . . . . . . . . . . . . . 183
4.2 Single-Phased System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
4.2.1 Performance by Type . . . . . . . . . . . . . . . . . . . . . . . . . 196
4.3 Two-Phased System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199

4.4 A Resolution System for Anaphoric Zero Pronouns . . . . . . . . . . . . . 204

4.4.1 A Maximum-Entropy Model . . . . . . . . . . . . . . . . . . . . . 205
4.4.2 Evaluating Features . . . . . . . . . . . . . . . . . . . . . . . . . . 208
4.4.2.1 The Ablation Method . . . . . . . . . . . . . . . . . . . 209

4.4.2.2 Feature Opt-In Models . . . . . . . . . . . . . . . . . . 210

4.4.2.3 Feature Weights . . . . . . . . . . . . . . . . . . . . . . 212
4.5 Performance Scores: A Round-Up . . . . . . . . . . . . . . . . . . . . . . 219

Conclusions 222

Bibliography 227

x
List of Tables

1.1 Dropped topic subject in Italian hcantare(x), x=lui, x=topici . . . . . . . . 16

1.2 Overt non-topic subject in Italian hcantare(x), x=luii . . . . . . . . . . . . 16

1.3 Overt topic subject in English hsing(x), x=he, x=topici . . . . . . . . . . . 17
1.4 Overt non-topic subject in English hsing(x), x=hei . . . . . . . . . . . . . 17
1.5 Overt topic subject in Yiddish hsing(x), x=he, x=topici . . . . . . . . . . . 18

1.6 Overt topic subject in Yiddish hrained()i . . . . . . . . . . . . . . . . . . . 18

1.7 Thai input with overt subject . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.8 Thai input with Pro subject . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.9 Thai input with embedded Pro subject . . . . . . . . . . . . . . . . . . . . 21

1.10 Centering transition states . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.1 Pronominal system of Korean, 1st and 2nd person . . . . . . . . . . . . . . 39

2.2 Pronominal system of Korean, 3rd person . . . . . . . . . . . . . . . . . . 40
2.3 Classification of Japanese zero anaphor by Kameyama (1985) . . . . . . . 45
2.4 Classification of Korean zero pronouns . . . . . . . . . . . . . . . . . . . . 46

1.1 Pronoun resolution algorithms for New York Times . . . . . . . . . . . . . 134

1.2 Pronoun resolution algorithms for fictional texts . . . . . . . . . . . . . . . 134
1.3 C OHERE and A LIGN produce preference ranking of 4 centering transition
types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

xi
2.1 Zero pronoun frequencies in KTB 1 and KTB 2 . . . . . . . . . . . . . . . 147
2.2 pro and PRO frequencies in the Penn Chinese Treebank 5.1 . . . . . . . . . 147
2.3 Zero pronoun frequencies by grammatical roles . . . . . . . . . . . . . . . 147

2.4 Zero pronoun frequencies by type . . . . . . . . . . . . . . . . . . . . . . 148

2.5 Overt pronouns in KTB 1, 1st and 2nd person . . . . . . . . . . . . . . . . 149
2.6 Overt pronouns in KTB 1, 3rd person and other . . . . . . . . . . . . . . . 149
2.7 Overt pronouns in KTB 2, 1st and 2nd person . . . . . . . . . . . . . . . . 150
2.8 Overt pronouns in KTB 2, 3rd person and other . . . . . . . . . . . . . . . 150

3.1 Naive Hobbs algorithm on Korean zero pronouns . . . . . . . . . . . . . . 165

3.2 Breakdown of antecedent NPs in KTB 2 . . . . . . . . . . . . . . . . . . . 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 pronoun in 06:2 . . . . . . . . 182
4.3 Binary classification of coreference for zero pronoun in 06:2 . . . . . . . . 183

4.4 Feature vector of coreference events for two pro subjects, partially repre-
sented . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
4.5 Numbers of targets/events generated per pro . . . . . . . . . . . . . . . . . 193
4.6 Numbers of coreferential targets/events generated per pro . . . . . . . . . . 193
4.7 Training and testing set sizes for the two corpora . . . . . . . . . . . . . . 193

4.8 Feature set sizes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

4.9 Performance of models as a binary classifier . . . . . . . . . . . . . . . . . 194
4.10 Confusion matrix for binary classifiers, KTB 1 . . . . . . . . . . . . . . . . 195
4.11 Confusion matrix for binary classifiers, KTB 2 . . . . . . . . . . . . . . . . 195

xii
4.12 Performance of coreference resolution models . . . . . . . . . . . . . . . . 196
4.13 Performance of coreference resolution models, averaging on 10 cross-fold
validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196

4.14 Accuracy by zero pronoun type, KTB 1 . . . . . . . . . . . . . . . . . . . 197

4.15 Accuracy by zero pronoun type, KTB 2 . . . . . . . . . . . . . . . . . . . 197
4.16 Prediction pattern for anaphoric (a) zero pronouns, KTB 2 . . . . . . . . . 198
4.17 Prediction pattern for deictic-speaker (i) zero pronouns, KTB 2 . . . . . . 199
4.18 Prediction pattern for generic (g) zero pronouns, KTB 2 . . . . . . . . . . 199

4.19 Confusion matrix for category classification model (Phase 1), KTB 1 . . . . 200
4.20 Combined performance of Phase 1 and Phase 2, KTB 1 . . . . . . . . . . . 201
4.21 Comparison of two approaches, KTB 1 . . . . . . . . . . . . . . . . . . . 201
4.22 Confusion matrix for category classification model (Phase 1), KTB 2 . . . . 202

4.23 Combined performance of Phase 1 and Phase 2, KTB 2 . . . . . . . . . . . 203

4.24 Comparison of two approaches, KTB 2 . . . . . . . . . . . . . . . . . . . 203
4.25 Numbers of coreferential NPs per NP-anaphoric pro . . . . . . . . . . . . . 207
4.26 Performance of coreference resolution models . . . . . . . . . . . . . . . . 207

4.27 Performance of coreference resolution models, averaging on 10 cross-fold

validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
4.28 Performance degradation by removing feature set . . . . . . . . . . . . . . 209
4.29 Performance degradation by removing feature set . . . . . . . . . . . . . . 210
4.30 Performance of individual features: nearest subject, nearest topic . . . 211

4.31 Performance of combination of features . . . . . . . . . . . . . . . . . . . 212

4.32 Top 20 positively weighted features . . . . . . . . . . . . . . . . . . . . . 214
4.33 Top 20 negatively weighted features . . . . . . . . . . . . . . . . . . . . . 215
4.34 Performance of Maximum Entropy models on 3 clausal groups . . . . . . . 217

4.35 Weights of 4 features: trained for pros in a matrix clause . . . . . . . . . . 218

xiii
4.36 Weights of 4 features: trained for pros in an adverbial clause . . . . . . . . 218
4.37 Weights of 4 features: trained for pros in an embedded clause . . . . . . . . 218
4.38 Key performance scores . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220

xiv
Introduction

In Korean sentences, arguments such as subject and object enjoy a great degree of free-

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 postposition
marker; and finally, the null form. Korean is only one of the group of languages which

permits extensive use of null-form arguments. These invisible arguments are believed to
be a substitute for pronouns1 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, but

without 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:2

(1) u1 . >

r r
Q
# 3q
l
` j{
] a
9 ~
%   m?
John-un enu kwamok-ul ceyil cohaha-ni?
John-Top which subject-Acc most like-Q?
Which subject does John like the most?
1
See Section 2.1.1 and also Kameyama (1985), Chapter 3.3.
2
Throughout this thesis, u1 , 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, e.g., u1 .A, u2 .B.

1
 u2 .
`
< ]
 j{a
9 ~
% K.
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 dis-
course principles, as illustrated by the above example where the zero pronoun subject is
understood 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 is

said to play a significant role. However, zero pronouns are not always anaphoric, that is,
their interpretation is not always based on some preceding linguistic expression: those in
(2a) and (2b) below are zero deictic pronouns that directly refer to an entity in the given
spatio-temporal context. On the other hand, generic indefinite zero anaphors in (2c) refer

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

(2) a. (looking at an elephant)


3
% A
' 1
W.
emcheng ku-ney.
(SBJ) very big-Excl.
(It)s very big!

b. \
|
9 m?
cip-ey ka-ni?
(SBJ) home-to go-Q?
Are (you) going home?

c. \

 |
s
\

 t.
san-ey ka-eya 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.

2
 d. +
O
Z \ r
P .
pelsse yel-si-ngi-ta.
(SBJ) already ten-oclock-Cop-Dec.
(It) is 10 oclock 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 in
the language but also has to be preferred in some way, as their very occurrences clearly
attest. More importantly, correct interpretation of these null forms is essential in successful

communication, 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 to
address the fundamental motivations and conditions behind the collective phenomenon of

Korean zero pronoun. The latter is from an engineering point of view, which is aimed at
finding 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 the
findings. 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 context

from which their reference can be derived. In what follows, the two goals are presented in
two separate parts.

3
 Part I

Analysis

4
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 reformulating
the theory of Empty Categories (EC), of which the zero pronoun is one, by focusing on the
syntactic behavior of zero pronouns in particular language settings.

The 1990s 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 zero
pronouns. Again focusing on the cross-linguistic variation of their syntactic admissibility,
researchers sought to generalize such cross-linguistic variations in the forms of ordered and
violable 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 of
discourse to ensure proper semantic interpretation of the pronoun category were being rec-

5
ognized within the formal semantics and pragmatics community. The theories of Dynamic
Semantics 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 noun
expressions involved.
Centering Theory (Grosz, Joshi and Weinstein, 1983, 1995, and many others) furthers
the notion of involvement of discourse factors in the use of pronouns, and casts information
theoretic notions at the center of their theory of pronoun usage. Zero pronouns in many

languages have been studied extensively within the Centering framework, under the belief
that zeros in these languages are much less constrained grammatically, but are governed
rather by pragmatic and specific contextual considerations in determining whether a given
zero 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 empty
have been collectively referred to as pro-drop languages. In this section, we present a

survey of the syntactic literature on the general phenomenon of zero pronouns in some of
the 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 are

analyzed as a case of empty categories. In this theory, nominal empty categories along
with their overt counterparts can be represented as a combination of two binary features:
pronominal and anaphor, as follows:

6
 (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 el-
ements, overt or empty. Three Binding Principles were given as governing their distribution
(Chomsky, 1986):

(4) The Binding Principles

A. an anaphor is bound in a local domain

B. a pronominal is free in a local domain

C. an r(eferring)-expression is free

where binding is defined as:

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

Chomsky (1981) divides the non-pronominal empty category (EC) to introduce yet

another EC, namely the variable. The variable is locally bound by an element that is
not 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 a

variable if locally A-bound.

7
 Reflexive pronouns (+anaphor, pronominal) and personal pronouns (anaphor, +pronom-
inal) are known to exhibit complementary distribution in their ability to co-refer with other
elements 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.

(6) a. Billi likes himselfi/j .

b. Billi likes himi/j .

c. *Hei thinks Bill likes Johni .

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 as

reflexive pronouns (in a) and like pronominal one in others (in b):

(7) a. Johni wants PROi /himselfi 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 its
governing category. It is solved by simply stipulating that PRO lacks a proper govern-
ing category. This null element PRO exhibits fairly complex referential and distributional
properties. 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. In
some contexts, PRO is obligatorily coindexed with a particular c-commanding NP in the
higher 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 infinitival
subject clause, and in (g) with PRO in an interrogative object clause, the PROs lack specific
reference and are understood instead as arbitrary.

8
 (8) (examples from Harbert (1995))

a. Johni tried PROi to win.

b. *Mary hoped for the president to apoint PRO.

c. *I believe that PRO will stay home.

d. Maryi promised me PROi to be home early.

e. John forced mei PROi to stay late.

f. PROarb to leave early would be inexcusable.

g. John doesnt know how PROarb 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 allow

zero subjects in declarative sentences, while other languages like English cannot. This fact
reflects one parameter of Universal Grammar, it was argued, which is that languages divide
into two sets in relation to the availability of null subjects (Perlmutter, 1971). A language in
which 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 that
the former lack phonetic contt. An example in Spanish:

(9) Dijo que [e] vendra manana.

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 proper
government (by the AGR (agreement) node). The differences between a non-pro-drop

9
language like English and a pro-drop language like Italian or Spanish is that only in pro-
languages is the INFL node a proper governor for the subject. The rationale behind this
theorizing is clear: a rich inflection system renders a zero subject recoverable while

a poor one does not. Furthermore, subject-object asymmetry was being widely reported
across 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 the
object position in these languages. Also, Huang (1989) finds further supporting evidence in

Pashto, which allows null object pronouns in perfect aspect sentences where there is object
agreement morphology on the verb.
This theoretical picture soon proved falling short on two accounts. First, the binary
distinction between pro-drop vs. non-pro-drop is too simplistic a typology: there is a wide

variability in which pro-drop languages allow only certain types of subjects to drop. For
example, 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 allow
zero realization for both referential pronouns and pleonastic types (Safir, 1985):

(10) a. * will zu Hause bleiben.

want at home to-stay.
I(*pro) want to stay home.

b. klar ist, da er nicht kommen wird.

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

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

(11) Il bel tempo invoglia [pro] a [PRO restare].

the good weather induces to stay.

10
 The good weather induces (pro) to (PRO) to stay.

Lastly, it was pointed out that a group of languages, including Chinese, Japanese and
Korean, allow zero subjects to occur rather freely, even though their systems for marking

agreement are even more impoverished than that of English. Huangs (1983, 1984, 1989)
work on Chinese pro distribution was aimed at adapting GB theorys original take on the
problem so that the zero-pronoun behavior in these languages is properly explained. His
work is examined closely in the next section.

1.1.2 Huangs (1983, 1984, 1989) Work on Chinese, Japanese and Ko-

rean pro

Huang draws from from Li and Thompsons (1976) distinction of topic-prominent and
subject-prominent languages, and proposes a Universal Grammar parameter called zero-

topic for Chinese, Japanese and Korean in order to formulate the zero-pronoun behaviors
displayed by them. In these topic-prominent languages which include Chinese, Japanese
and Korean, structural subjects are not a basic requirement of the sentence. Furthermore,
in such languages sentences of the form topic-comment abound and must count as basic
forms in that they cannot be plausibly derived from other more basic forms. Huangs

main argument is that most zero subjects in these languages are in fact a zero topic. More
specifically, he proposes that the relevant phenomena be derived jointly by the following
set 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 po-

tential antecedent of a zero pronoun

11
 d. The binding theory of Chomsky (1981), in particular the condition of disjoint
reference (DJR) or condition (B)

e. The Generalized Control Rule (GCR)

All of the items above are readily accepted ideas from the standard GB theory, except

for the last item, The Generalized Control Rule (GCR), which is Huangs own contribution.
It is basically Chomskys (1980) rule of control extended to cover both pro and PRO. Huang
claims that the two are essentially the same entity and proposes to abolish the distinction
between the two categories originating in Chomsky (1982) by subjecting them to the single
rule of GCR:

(13) a. Generalized Control Rule (GCR):

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

b. Control Domain:
is the control domain for if it is the minimal category satisfying (i) and (ii):

(i) is the lowest S or NP containing

(a) , or

(b) the minimal maximum category containing ;

(ii) contains a SUBJECT accessible to .

The key insight for collapsing the two categories comes from the observation that the

absence of subject-verb agreement plays a crucial role for both: null subjects are allowed in
Chinese which has no subject-verb agreement; in English, PRO only occurs in embedded
clauses that are not governed by AGR, a syntactic incarnation of the agreement feature in
the 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 three
East Asian languages is introducing the notion of topic in accounting for some of the zero-

12
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 NP
deletion 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, he

argues that these are not zero pronouns but in fact variables locally A-bound by the zero
topic NP, whose reference is fixed in discourse.
Kim (2003) examines Huangs work and points out that it fails in three points. First,
Huang sets apart zero objects in embedded clauses from other zero pronouns and claims

that 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. *Zhangsani zhidao [Lisi mei banfa shuifu ei ].

Zhangsan know Lisi no method persuade ei .
*Zhangsani knows Lisi cannot persuade himi .

b. [TOP ei ] Zhangsan zhidao [Lisi mei banfa shuifu ei ].

[TOP ei ] Zhangsan know Lisi no method persuade ei .
Zhangsan knows that Lisi cannot persuade himi/heri .

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

(15) 
[s
t i  Hs
  i \
v+
 I


.
Tolii -ka Swuni-ka i kwoylophi-ess-ta-ko malha-ess-ta.
Tolii -Nom Swuni-Nom (OBJi ) tease-Past-Dec-Quote say-Past-Dec.
Tolii said Swuni teased (himi).

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

as Huang claims. For example:

(16) u1 . Q
#j] Q
 !ti H [y
 
t 
 
o]j \
 
 $
4 .
ecey apecii -nun [kangaci han mali]j -lul sa osi-ess-ta.
yesterday fatheri -Top [puppy one CLASS]j -Acc buy come-Past-Dec.
Yesterday fatheri bought a puppyj .

13
 u2 . i/j s
O
_  



.
i/j tel-i hayah-ko manh-ess-ta.
(SBJi/j ) fur-Nom white-And abound-Past-Dec.
(Iti/j was such that) fur was white and plenty.

(17) u1 .
%

i
H
 Qj
# ] x
p
1 U\



% Bj \




z.
Youngswui-nun ecey tungkyokil-ey Younghij -lul manna-ess-ta.
Youngswui-Top yesterday way-to-school-on Younghij -Acc meet-Past-Dec.
Youngswui met Younghij on his way to school.

u2 . i/j
# y
 \V.I
.
i/j yecenhi yeyppu-ess-ta.
(SBJi/j ) still pretty-Past-Dec.
(Shei/j ) was still pretty.

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

(18) a. Q
# n m?
eti ka-ni?
(SBJ) where go-Q?
Where are (you) going?

b. /{
? 
9 a

.
nayil ka-lkkey.
(SBJ) tomorrow go-Prom.
(I) will go tomorrow.

1.1.3 Optimality Theory Approaches to Zero Pronouns

According to Optimality Theory, a language-particular grammar is a means of resolv-

ing the conflicts among universal constraints (Prince and Smolensky, 1993). Grimshaw

(1997), Grimshaw and Samek-Lodovici (1998) are an attempt at capturing the cross-linguistic
variation in the distribution of subjects in terms of such competing universal constraints
regarding their expression. Different languages resolve conflicts by employing different

14
ranking among the constraints, hence language variation arises. They suggest the follow-
ing set of five constraints regarding expression of subjects:

1. S UBJECT: The highest A-specifier in an extended projection must be filled (Grimshaw,

1997). Failed by clauses without a subject in the canonical position.

2. F ULL -I NT: (Full-Interpretation) Parse lexical conceptual structure (Grimshaw, 1997).

Failed by expletives and auxiliary do.

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

4. A LIGN F OCUS: Align the left edge of focus constituents with the right edge of a
maximal 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 sub-

jects, most notably Italian. Citing numerous previous works which show that the null
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; Vallduv,
1992), they suggest the high-ranking status of the constraint D ROP TOPIC in these lan-
guages is responsible for the phenomenon. The constraint requires a subject coreferential

with 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 that
pro-drop follows from the constraint D ROP TOPIC requiring arguments whose antecedent is
a topic to be unexpressed, or structurally unparsed, with the penalty of violating PARSE

as a result. The constraint D ROP TOPIC outranks the two constraints PARSE and S UBJECT
in these languages, which require subjects to be overtly expressed:

15
 (19) Ranking for languages allowing referential null subjects:
D ROP TOPIC PARSE S UBJECT

The tableau below illustrates the case of a grammatical null subject in Italian which is

coreferential with at topic in the sentence. The candidate (a) ha cantato wins over the other
candidates (b) and (c), lui ha cantato and ha cantato lui, which have overtly expressed
pronoun subjects. Note that F ULL -I NT is always satisfied since no candidates involve
expletive pronouns.

Candidates D ROP TOPIC PARSE F ULL -I NT S UBJECT

a. ha cantato * *
b. lui ha cantato *!
c. ha cantato lui *! *

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

Compare it with the following case of the same proposition, which, although not apparent
with the absence of surrounding context, is presumed to have a non-topic-marked subject.
Here, both (b) lui ha cantato and (c) ha cantato lui satisfy the highest-ranked D ROP TOPIC
constraint, and the former ultimately wins over the latter due to the fact that it does not

violate the S UBJECT constraint by moving up the subject to the highest specifier position,
thereby bringing it up to the front of the sentence.

Candidates D ROP TOPIC PARSE F ULL -I NT S UBJECT

a. ha cantato *! *
b. lui ha cantato
c. ha cantato lui *!

Table 1.2: Overt non-topic subject in Italian hcantare(x), x=luii

On the other hand, in languages such as English where subjects are required to be
overt on all occasions, the relative ranking between D ROP TOPIC and PARSE is reversed, so
PARSE dominates D ROPTOPIC:

16
 (20) Ranking for English where null subjects are disallowed:
PARSE D ROP TOPIC

The result of this ranking is that the candidate with an overtly expressed subject prevails,

whether it is co-indexed with a topic (Table 1.3) or not (Table 1.4):

Candidates PARSE D ROP TOPIC S UBJECT F ULL I NT

a. has sung *! *
b. he has sung *
c. has sung he * *!

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

Candidates PARSE D ROP TOPIC S UBJECT F ULL I NT

a. has sung *! *
b. he has sung
c. has sung he *!

Table 1.4: Overt non-topic subject in English hsing(x), x=hei

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 F ULL -I NT S UBJECT D ROP TOPIC

First, the unavailability of topic-pro-drop indicates that PARSE outranks D ROP TOPIC in
these languages. A Yiddish example, where the subject is marked as the topic is given in

Table 1.5. Also, obligatory zero-realization of the subject position in the two languages
when the verb has propositional content can be captured by ranking F ULL -I NT higher than
S UBJECT. As illustrated by Table 1.6, candidate (b) with an expletive subject violates

17
higher-ranked F ULL -I NT, while the optimal candidate (a) violates S UBJECT by having a
non-realized subject position.

Candidates PARSE F ULL -I NT S UBJECT D ROP TOPIC

a. sung *! *
b. he sung *
c. expl. sung *! *

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

Candidates PARSE F ULL -I NT S UBJECT D ROP TOPIC

a. rained *
b. it rained *!

Table 1.6: Overt topic subject in Yiddish hrained()i

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

be found in Speass (2001) work. She aims to cover some slightly different aspects of cross-
linguistic empirical data: Grimshaw and Samek-Lodovicis (1998) zero pronouns were lim-
ited to subject pronouns occurring in finite clauses, namely pro in the classic Chomskyan
sense; Speas (2001), on the contrary, is mainly concerned with the availability of subject,

and more importantly, object pronouns,1 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

1
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).

18
 Type 1 languages include English; Type 2 includes languages such as Spanish and
Mandarin Chinese, which allow subject Pro but not object Pro; Type 3 includes Thai and
Korean, 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 positions
and 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 many
authors, including Bouchard (1984), Sportiche (1983), Manzini (1983), Borer (1989) and

Huang (1984), have pointed out problems inherent in maintaining the standard distinction
between zero forms in subject position of infinitives and gerunds (PRO) from those in other
positions (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 of

languages given above in (22) follows this theoretic assumption, and she argues that the
cross-linguistic distribution and interpretation of Pro follows from different rankings of
violable constraints in OT given below:

(23) 1. C ONTROL (CTL): A featureless pronoun must be coindexed with the closest

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

2. B INDING T HEORY P RINCIPLE B (BTB): A pronoun must be free in its Bind-

ing Domain

3. M AX (P RO ): If Pro occurs in the input, then its output correspondent is Pro

4. AGR (X): +Tense has AGR features

5. N O P HI -F TS: Avoid AGR features

6. C ASE: An NP with phi features must receive Case

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

19
is derived in her theory. The relevant constraints for this type of languages are the first three
above, C ONTROL, B INDING T HEORY P RINCIPLE B (BTB), and M AX (P RO ). The C ON -
TROL constraint has the effect of forcing a Pro, assumed featureless in the theory, to have

a closely located antecedent; BTB balances out C ONTROL in that it prevents a pronoun
from having an antecedent within a binding domain. M AX (P RO ) is a key constraint and
needs some elaboration. Unlike Grimshaw and Samek-Lodovicis (1998) approach, Speas
assumes that whether a pronoun is overt or not is specified in the input, which can then
be altered in the output by violating the M AX (P RO ) constraint. Recall that in Grimshaw

and Samek-Lodovici (1998) the input has information as to the topic status of subjects but
did 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 D ROP TOPIC constraint.
Both approaches in effect capture competition between candidates with a zero/overt sub-

ject; Speass (2001) approach, however, does so while avoiding directly addressing the
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 of
constraints correctly predicts that the zero pronoun can be coreferential with the matrix
subject:

20
 BTB M AX (P RO ) C ONTROL
a. Pro hen Nit *!
b. khaw hen Nit

Table 1.7: Thai input with overt subject

BTB M AX (P RO ) C ONTROL
a. Pro hen Nit *
b. khaw hen Nit *!

Table 1.8: Thai input with Pro subject

(25) Nit bOO waa Pro hen NOy

Nit speak say Pro say Noy
Nit said that he/she/they saw Noy

BTB M AX (P RO ) C ONTROL
a. Nit bOO waa Pro hen NOy *
b. Niti bOO waa Proi hen NOy *
c. Nit bOO waa khaw hen NOy *!

Table 1.9: Thai input with embedded Pro subject

Both Grimshaw and Samek-Lodovicis (1998) and Speass (2001) work sought to ac-
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 of
pragmatics by introducing such notions as topic into the framework, but it was largely
repackaged as a syntactic element, thereby exempting the theory from looking beyond
the level of syntax. The accounts largely focus on whether or not a certain realization of

pronominal element is permissible in a language, and as result they are able to offer only
limited glimpses of how such pronominal elements are interpreted, as example (25) and
table (1.9) show. Following these works, there have been numerous studies which aim at
casting OT as a full system of pronoun interpretation, and ultimately, anaphora resolution.

21
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.

1.2 Centering Theory: A Discourse-Oriented Approach

to Pronouns

So far we discussed how the zero pronoun was first perceived as a sentence-level phe-

nomenon and therefore a syntactic problem, only to earn recognition later as a problem
that 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
of discourse. In this section, I present a brief review of the theory and related works that
apply 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 the
relationship between local coherence and the use of referring expressions. Centering has
its 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 Sidners 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

22
in that segment, the use of ellipsis, and the choice of referring expressions.
In modeling local coherence, Centering requires two constructs: a single backward-
looking center and a list of forward-looking centers. Centers are semantic entities that

are part of the discourse model. It is formalized as a system of constraints and rules that
govern the interpretation of these centers. It is assumed that discourses are composed of
constituent segments (Grosz and Sidner, 1986), each of which consists of a sequence of
utterances. Each utterance in a discourse has associated with it a set of discourse entities
called FORWARD - LOOKING CENTERS, Cf. There is a unique special member in this set

called the BACKWARD - LOOKING CENTER, Cb. The Cb is the discourse entity that the
utterance most centrally concerns, which also links the current utterance to the previous
discourse. The set of FORWARD - LOOKING CENTERS, Cf, is ranked according to discourse
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, and
lexical semantics and postulates the following ranking for Cfs in English, which is solely
based on grammatical roles:

(26) Cf ranking 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 that
the 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(Ui ), the PREFERRED CENTER of the
current utterance Ui , will most likely be the Cb(Ui+1 ), the BACKWARD - LOOKING CENTER

of the next utterance Ui+1 , and therefore represents a prediction about the Cb of the next

23
utterance. The actual realization of Cb of the next utterance, of course, may or may not bear
out this prediction: the BACKWARD - LOOKING CENTER of the next utterance Cb(Ui+1 ) is
defined as the highest-ranked Cf in the current set of Cfs (i.e., Cf(Ui )) that is actually

realized in the next utterance, and therefore may or may not coincide with the Cp(Ui ).
With regard to the relation between these elements, Cfs, Cb and Cp, the theory of centering
specifies the following set of constraints:

(27) Constraints:
For each utterance Ui in a discourse segment U1 , ... , Un :

1. There is precisely one backward-looking center Cb.

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

3. The center, Cb(Ui ), is the highest-ranked element of Cf(Ui1 ) that is realized

in Ui .

There is no overt reference to Cp in this formulation, but (27 2.) implies the following
relation between Cb(Ui ) and Cp(Ui1 ): Cb(Ui ) is Cp(Ui1 ) if Cp(Ui1 ) is among Cf(Ui ).
Since Cb and Cp serve the mirror-opposite roles, that is, one of linking the current
utterance 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 transition
status of attention across pairs of adjacent utterances. Also, whether or not the current
utterance inherits the BACKWARD - LOOKING CENTER from the previous utterance forms
another axis in the following four-way distinction of transition types2 :

These transitions differ from each other according to whether BACKWARD - LOOKING CEN -

TERs of successive utterances are identical or not (CONTINUE , RETAIN vs. SHIFT), and
whether they match the most highly ranked element of the current FORWARD - LOOKING
2
The distinction between the two types of SHIFT was first introduced in Brennan, Friedman and Pollard
(1987).

24
 Cb(Ui )=Cb(Ui1 ) Cb(Ui ) 6= Cb(Ui1 )
or Cb(Ui1 ) =[?]
Cb(Ui ) = Cp(Ui ) CONTINUE SMOOTH - SHIFT
Cb(Ui ) 6= Cp(Ui ) RETAIN ROUGH - SHIFT

Table 1.10: Centering transition states

CENTER list, the Cp(Ui ), 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 speakers 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 segments that continue centering the same
entity are more coherent than those that repeatedly shift from one center to another. These

observations are encapsulated in two rules:

(28) Rules:
For each Ui in a discourse segment Ui , ... , Um :

1. If some element of Cf(Ui1 ) is realized as a pronoun in Ui , then so is Cb(Ui ).

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 dis-
course salience is conveyed. Rule (1) implies that if there are multiple pronouns in an
utterance, one of them must be the Cb. Furthermore, it follows that if there is only one
pronoun, then it must be the Cb. Rule (2) states two ordered preferences: first, having the

same Cb as the previous utterances (Cb(Ui )=Cb(Ui1 )) is preferred; second, the prospect
of continuing with the same Cb in the next utterance (Cb(Ui ) = Cp(Ui )) is preferred. This
ordered preference for transition states represents the differing levels of local coherence.

25
1.2.2 Centering Theory Across Languages

Different languages have different inventories of grammatical means and strategies avail-

able for their speakers to choose from in formulating noun expressions. Therefore, it is only
natural that the relationship between local coherence and the use of referring expressions
has different manifestations across languages. Since the conception of the theory, there
have been numerous studies re-examining and modifying the details of Centering Theory
in order to capture language-dependent aspects.

As noted earlier, the ranking of FORWARD - LOOKING CENTERs was the first to be noted
by many studies as a prominent source for language variation. These studies identify dif-
ferent factors in the language under consideration that affect Cf ranking and propose a Cf
ranking that is unique to the language. For Japanese, many researchers (Kameyama, 1985;

Walker, Iida and Cote, 1990, 1994) have pointed out that empathy and topic marking affect
Cf ordering. Walker, Iida and Cote (1990, 1994) place the two on the very top of the Cf
ordering for Japanese:

(29) Cf ranking for Japanese (Walker, Iida and Cote, 1994)

(grammatical or zero) topic > empathy > subject > object2 > object > other(s)

Empathy is a notion first proposed by Kuno (1976), which represents the speakers po-
sition or identification in describing a situation. It is an essential grammatical notion often
involved in describing events with a sense of directionality, such as giving and receiving,

and coming and going. The following example from Walker, Iida and Cote (1994) demon-
strates how placing empathy over subject in the Cf ranking helps explain Japanese speakers
interpretation of the zero pronouns contained in the sequence:

(30) u1 . Hanako wa kuruma ga kowarete komatteimasita.

Hanako Top/Nom car Nom broken at-a-loss-was.
Her car broken, Hanako was at a loss.

26
 Cb: HANAKO
Cf: [HANAKO, CAR]

u2 . Taroo ga sinsetu-ni te o kasite-kuremasita.

Taroo Nom (OBJ2/EMP) kindly hand Acc lend-gave.
Taroo kindly did (Hanako) a favor in helping her.
Cb: HANAKO
Cf: [HANAKO (EMPATHY), TAROO (SUBJ)]

u3 . Tugi no hi eiga ni sasoimasita.

next of day (SBJ) (OBJ) movie to invited.
Next day (Hanako) invited (Taroo) to a movie.
Cb: HANAKO
Cf1: [HANAKO, TAROO] CONTINUE

Cf2: [TAROO, HANAKO] RETAIN

There is an ambiguity regarding interpretation of the two zero pronouns in utterance (u3 ),
which leads to the two alternative Cf rankings shown in Cf1 and Cf2. When native speakers

of Japanese were polled, they showed a clear preference towards the interpretation corre-
sponding to Cf1 (16 speakers favoring Cf1 compared to 2 speakers for Cf2). If Hanako,
an entity marked as empathy by the empathy-loaded Japanese verb kure-ru, is placed in
the Cf-ranking above Taroo, the subject of the utterance, the pronoun interpretation in Cf1
constitutes a CONTINUE transition whereas the interpretation in Cf2 constitutes a RETAIN.

This overall scheme conforms to the rule of preference in centering transitions, given pre-
viously in (28). Conversely, if the empathy-marked Hanako was to be placed below the
subject Taroo in the Cf-ranking, Cf1 constitutes a RETAIN and Cf2 a CONTINUE, and one
is left without an explanation as to why speakers prefer a RETAIN transition in Cf2 over a

CONTINUE in Cf1.
As for the grounds for TOPIC outranking EMPATHY in the Cf ranking, Walker, Iida and
Cote (1994) gives the following example:

27
 (31) u1 . Mitiko wa kanai o gityoo ni osite-kuremasita.
Mitiko Top/Nom wife Acc/Emp chairman Dat recommended-gave.
Mitiko did my wife a favor in recommending her as chairperson.

u1 . Mitiko ga kanai o gityoo ni osite-kuremasita.

Mitiko Nom wife Acc/Emp chairman Dat recommended-gave.
Mitiko did my wife a favor in recommending her as chairperson.

u2 . asu no kaihyoo-kekka o tanosimi-ni siteimasu.

(SBJ) tomorrow of results Acc look-forward doing-is.
A: (Mitiko) is looking forward to tomorrows results.

B: (My wife) is looking forward to tomorrows results.

Here, (u1 ) and (u1 ) are minimally different in that Mitiko is a grammatical topic with a
topic-marker wa in (u1 ) whereas it is a plain subject with a nominative marker ga in (u1 ).
The zero subject in the following sentence has an ambiguous interpretation, one referring

to Mitiko in translation A and the other to my wife in B. There is a strong preference for the
interpretation where the zero refers back to the topic-marked Mitiko in the case of (u1 u2 )
sequence; the (u1 u2 ) sequence, however, permits the second interpretation as well, where
the zero referring back to the empathy-marked my wife. This suggests that empathy-marked
NP my wife is not as salient as the TOPIC.

Turan (1995) examines the notion of empathy for Turkish data, and concludes that it is
the hierarchy of thematic roles that are more relevant in Cf-ranking in Turkish, by noting
that the object of some psychological verbs rank higher than the subjects.

(32) u1 . Ahmeti Aliyik kasten yumuatkItI.

Ahmet Ali-Acc deliberately soft-Caus-Past-Past.
Ahmet calmed down Ali deliberately.

u2 . i onuk saatlerce ikna etmeyeugratI.

(SBJi ) he-Acc hour-Plu-Adv persuatsion do-Inf-Dat try-Past.
(He) tried to persuade him for hours.

28
 (33) u1 . Ahmet karIsInIi farkIna varmadan yumuatmItI.
Ahmet wife-Poss awareness arrive-Neg-Abl soft-Caus-Past-Past.
Ahmet calmed down his wife without being aware.

u2 . i artIk eskisi gibi cok cabuk sinirlenmiyordu.

(SBJi ) anymore old-one like very fast nervous-Neg-Prog-Past.
(She) didnt become nervous as quickly as she used to.

In the two contrasting examples above, the first utterances are minimally different in
terms of the thematic roles assigned to the subject. In (32u1 ), the presence of the adverb
deliberately forces the subject Ahmet to have the agent role; in (33u1 ), the adverbial phrase

without being aware renders the subject with the theme role. In the subsequent utterances,
the zero subjects are understood to refer to different entities, namely the subject of the
preceding utterance in (32u1 ) and the object of the preceding utterance in (33u1 ).
Turan notes that Kunos empathy can help explain (32u1 ) but not (33u1 ): in (33u1 ), the

speaker is expected to empathize with the subject Ahmet rather than with an expression
that is dependent on it, i.e., his wife; this predicts that the former and not the latter will
be salient. However, as seen above, the dependent expression in object position must rank
higher in the Cf list because it is the Cb, realized with a zero subject, in the subsequent
utterance. This is due to the fact that the subject, Ahmet, is assigned a theme role owing

to the presence of the adverbial phrase without being aware, while his wife takes the role
of experiencer, which outranks theme in thematic hierarchy. Turan (1995) proposes the
following Cf ranking rule for Turkish based on thematic hierarchy:

(34) Ranking of Cf-list by thematic roles (Turan, 1995):

agent > experiencer > (inalienable) possessor > theme

Similarly to the case of Japanese, many studies conducted for Korean share the position
that topic plays a prominent role in Cf ranking hierarchy (No, 1991; Jang, 1986, 1994; Kim,
1994; Ryu, 2001): a topic-marked element in an utterance outranks any other grammatical

29
argument in an utterance, including the subject. Ryu (2001) formulates his ranking thusly:

(35) Center hierarchy for Korean by Ryu (2001)

r,

topic ( H /un,nun/) > subject > object2 > object > others


There are a few studies which include discourse participants such as the speaker and
the hearer in the ranking. Choi & Lee (1999) is one such study:

(36) Center hierarchy for Korean by Choi & Lee (1999)

speaker, hearer > subject > indirect object > direct object > others

Various center hierarchies suggesed for Korean will be examined in detail later in Section
3.2.

1.2.3 The Zero Pronoun and Centering

While English has pronominalization as the only means of encoding salience of an entity,
other languages have more at their disposal: the so-called pro-drop languages famously

allow a zero form. The role of zero pronouns in the framework of Centering Theory has
attracted attention of many researchers working in such pro-drop languages as Japanese,
Italian, Turkish, among others. In this section, we review studies on the zero pronouns in
the three languages: Walker, Iida and Cote (1990, 1994), Di Eugenio (1990, 1996, 1998),

and Turan (1995). They aim to discover the governing rules underlying the use of zero
pronouns in the three respective languages and find the motivation in types of centering
transitions. The studies offer differing details across languages, depending on the particu-
lars of the theoretical stance that is assumed in each study.

While the grammar of Japanese has both overt and zero forms of pronouns available,
Walker, Iida and Cote (1994) mostly disregard the overt form and treat the zero form as
analogous to the overt pronoun in other languages. The decision is based on the observation

30
that the use of the overt pronoun in Japanese is rare in normal speech, and is limited even
in written text, which historically comes from the fact that the overt pronouns such as
kare (he) and kanozyo (she) were introduced into Japanese in order to translate gender-

insistent pronouns in foreign languages (Martin, 1976). After examining evidence from
Japanese native speakers interpretation of zero pronouns in their data, they found a strong
correlation between interpretation of zeros and the CONTINUE transition. They formulate
it as what they call the Zero Topic Assignment (ZTA) rule:

(37) Zero topic assignment (Walker, Iida and Cote, 1994)

When a zero in Ui+1 represents an entity that was the Cb(Ui ), and when no other
CONTINUE transition is available, that zero may be interpreted as the ZERO TOPIC

of Ui+1 .

Recall that ZERO TOPIC ranks higher in their formulation of Cf ranking of Japanese (shown

in 29). This rule combined with the Cf ranking therefore means that the entity that was pre-
viously the Cb is predicted to continue as the Cb even if it is not in the subject position. The
authors further conjecture that the ZTA rule is applicable in all free word-order languages
with zeros.
Unlike Japanese, the overt pronoun in Italian is fully functional and utilized in its

pronominal system, which presents an Italian speaker with a choice of null versus strong,
overt pronoun. Di Eugenio (1990, 1996, 1998) argues that the choice between the two
forms can be accounted for in terms of centering transitions. She proposes:

(38) The function of null subjects in Italian (Di Eugenio, 1990, 1996, 1998):

a. Typically, a null subject signals a CONTINUE, and a strong pronoun a RETAIN

or a SHIFT.

b. A null subject can be felicitously used in cases of RETAIN or SHIFT if Ui pro-

vides syntactic features that force the null subject to refer to a referent different

31
 from Cb(Ui1 ). Moreover, it is the syntactic context up to and including the
verbal form(s) carrying tense and/or agreement that makes the reference felic-
itous or not.

The first line states that there is a strong preference for null subjects in the case of CON -

TINUE transitions. The second provides further discourse conditions that render use of the
null form felicitous in other transitions, such as when the context up to and including the
verbal forms marked for tense and agreement provides early enough clues on the refer-
ence of the zero form. Furthermore, after examining a small corpus of naturally occurring

data, Di Eugenio (1998) finds that taking the transition preceding a CONTINUE into ac-
count provides an elegant explanation for some strong pronouns used in CONTINUEs. A
CONTINUE preceded by a RETAIN behaves differently from one preceded by a CONTINUE
or by a SHIFT, and use of overt pronouns in such CONTINUESs is often felicitous. This

is due to the fact that RETAIN predicts that the center will shift, which is not substantiated
when it is then followed by a CONTINUE transition: the use of overt pronouns therefore can
be interpreted as a means of alerting the hearer, who is expecting a shift in center, that the
prediction is in fact countered.

Turkish is another pro-drop language which allows both overt and zero pronouns. Both
forms necessarily agree with the number and person morphology on the verb, while third-
person singular agreement morpheme is null and third-person plural morpheme is optional.
Gender is not morphologically marked in Turkish. Turan (1995) reports a strong correlation
between CONTINUE transitions and use of zero subjects in Turkish, a similar result reported

for the aforementioned languages. Also, when a zero pronoun occurs in a SMOOTH - SHIFT
transition, it tends to refer back to the subject of the previous utterance, which she refers to
as SMOOTH - SHIFT- SUBJ .

32
 (39) The function of null subjects in Turkish:
Null subjects are used to encode CONTINUE and SMOOTH - SHIFT- SUBJ transitions,
i.e., null subjects pick up either Cps or Cbs.

She concludes that null subjects encode a CONTINUE or SHIFT- SUBJ transitions, i.e.,
they realize the previous Cp or Cb. The Cb in a CONTINUE transition is always realized by
a null pronoun in Turkish, unless it is a stressed/focused pronoun or the speaker wants to
communicate some other information such as the start of a new discourse segment. Overt
pronoun subjects mostly occur in SHIFT- TO - OBJECT transitions; they can occur in a CON -

TINUE if and only if the entity they refer to is phonetically prominent, and is the focus of
the utterance.
Lee and Lee (2000) propose what the authors call the Controlled Information Packag-
ing Theory (CIPT) for zero pronouns in Korean. CIPT is an extended version of Vallduvs

(1994) Information Packaging Theory, in which a fifth information structure (40 e) is in-
troduced in addition to the original four (40a-d) previously established in the theory:

(40) a. Link Tail Focus structure (L-T-F structure)

b. Link Focus structure (L-F structure)

c. Tail Focus structure (T-F structure)

d. Focus structure (F structure)

e. Slotlink Focus structure (SL-F structure)

The SlotlinkFocus structure can be exemplified in the following exchange, first noted by
Vallduv (1994: 16):

(41) a. A: Why dont you go to the theater more often?

b. B: TICKETS are expensive.

33
He notes that the sentence in (41 b) is not about any particular referent. Lee and Lee (2000)
takes a step further and recognize it as a type of information structure in its own right,
adapting it from Minskys (1975) notion of frameslots relation. The idea is that, when the

theater is uttered, by way of our cognitive knowledge it acts as the frame to which many
objects associated with it get activated in the slots that it commands.
This Slotlink Focus information structure, Lee and Lee (2000) claims, is what is
behind most zero anaphors. Coupling it with the assumption that the notion of center
put forth in Centering Theory is useful in establishing the antecedent of zero anaphors in

Korean, they propose the following revised hierarchy for the ranking of the forward-looking
centers for zero anaphors with the SL-component placed at the highest:

(42) Cf ranking for Korean by Lee and Lee (2000)

SL-component > {Speaker, Hearer} > Subject > Indirect Object > Direct Object

> Others

They argue that they can calculate the reference of zero anaphora in any form of domain-
restricted dialogues with this hierarchy, and they demonstrate its validity using examples
found in a corpus consisting of dialogues concerning hotel reservations.

34
Chapter 2

Analysis of Korean Zero Pronouns

In this chapter, we present detailed accounts of Korean zero pronouns. We first define and
defend the notion of zero pronouns as the object of the study, and then present detailed
analysis of various subtypes of Korean zero pronouns.

2.1 Defining the Object of the Study

Working with a linguistic construct which is characterized by its lack of form makes it
a particularly important task to precisely formulate and properly motivate the object of
the study. In this section, we first present the justification behind the term pronoun by
demonstrating that the Korean zero pronouns, at least a large subclass of them, serve the

same function that unstressed pronouns in English do. Next, the problem of identifying the
zero element is discussed, in both theoretical and practical terms.

2.1.1 Why Zero Pronouns

When we adopt the term zero pronoun for the subject of the study it is implied that it

is essentially a pronoun which is distinguished by its overt counterpart by its null formal

35
factor. As we will see in more detail in upcoming sections, this concept is somewhat
misleading. While it is true that some elements that are covered by the term in this study
indeed largely share the same property with overt pronouns, there are many others that do

not and therefore cannot be instantiated with overt pronouns.

Historically, the term zero pronoun derived from the observation that unstressed pro-
nouns in languages such as English are most natural when left unexpressed in being trans-
lated into other languages. Kuroda (1965) argues that there is a functional correspondence
between English unstressed pronouns and NP ellipsis in Japanese. Among the three exam-

ples below which contain repeated references to the same entity George, only (b) and (c)
make a good style of English by avoiding repeatedly using full NPs for the same reference.

(43) (emphasis added)

a. (bad) George does Georges work when George feels like it.

b. (good) George does his work when he feels like it.

c. (good) He does his work when he feels like it.

Corresponding Japanese sentences, where he is replaced by Japanese 3rd person pro-

noun kare, are all stylistically bad: repeating full NPs results in unnaturalness, and so does
a succession of repeated pronouns.

(44) (emphasis added)

a. (bad) George wa George ga si-ta-i toki ni George no sigoto

(bad) George Top George Nom do-want-RelEnd time at George Gen work
o su-ru.
Acc do-Pres.
George does Georges work when George feels like it.

b. (bad) George wa kare ga si-ta-i toki ni kare no sigoto o

(bad) George Top he Nom do-want-RelEnd time at he Gen work Acc
su-ru.
do-Pres.

36
 George does his work when he feels like it.

c. (bad) Kare wa kare ga si-ta-i toki ni kare no sigoto o

(bad) he Top he Nom do-want-RelEnd time at he Gen work Acc
su-ru.
do-Pres.
He does his work when he feels like it.

A natural and good style in Japanese would have a zero form instead of the overt pronoun
for the second and third mention of the entity:

(45) (emphasis added)

George/Kare wa si-ta-i toki ni sigoto o su-ru.

George/he Top (SBJ) do-want-RelEnd time at (GEN) work Acc do-Pres.

George/he does (his) work when (he) feels like it.

An exactly parallel case can be made for Korean. The Korean counterparts of examples
in (44) with a Korean third-person pronoun
(/ku/, he) are all unnatural-sounding:

(46) a. (bad) GeorgeH George

 
z`

 :
M \
(bad) George-nun George-ka ha-ko-siph-ul ttay-ey
(bad) George-Top George-Nom do-AuxEnd-want-RelEnd time-At
George_ { `
9
 

.
George-uy il-ul ha-nta.
George-Gen work-Acc do-PresDec.
George does Georges work when George feels like it.

b. (bad) GeorgeH
  
z`

 :
M \ _
(bad) George-nun ku-ka ha-ko-siph-ul ttay-ey ku-uy
(bad) George-Top he-Nom do-AuxEnd-want-RelEnd time-At he-Gen
`
{
9
 

.
il-ul ha-nta.
work-Acc do-PresDec.
George does his work when he feels like it.

37
 c. (bad)
H
  
z`

 :
M \ _ { `
9

(bad) ku-nun ku-ka ha-ko-siph-ul ttay-ey ku-uy il-ul
(bad) he-Top he-Nom do-AuxEnd-want-RelEnd time-At he-Gen work-Acc


.
ha-nta.
do-PresDec.
He does his work when he feels like it.

As with the Japanese case, the most natural Korean translation is the one with the second

and third mentions left empty:

(47) George/H
 
z`

 :
M \ `
{
9

George/ku-nun ha-ko-siph-ul ttay-ey il-ul
George/he-Top (SBJ) do-AuxEnd-want-RelEnd time-At (GEN) work-Acc


.
ha-nta.
do-PresDec.
George/He does his work when he feels like it.

This demonstrates that some Korean zero pronouns are functionally equivalent to the un-
stressed pronouns in English. Many previous Centering-Theory-based studies on Korean

zero anaphora, including Kim (1994, 2003), Hong (1999), and Ryu (2001), take up this
position in assuming functional equivalence between the Korean zero pronoun and English
unstressed pronoun.

2.2 Overt Pronouns in Korean

Korean does have a system of overt pronouns, but it is known to be underdeveloped and
under-used (Lee et al., 1997). While a paradigm can be constructed which corresponds to
that of pronominal systems found in Western languages, it makes up an incomplete and

less robust system, characterized by item-specific arbitrariness in actual usage.

Korean has grammaticalized means of encoding politeness and honorification, and they
are heavily reflected in its personal pronoun system. While Western languages display

38
 singular plural
1st person non-polite 


o
/na/ /wuli/
polite 
$ 
$ B([ t)

/ce/ /cehi(tul)/
singular plural
2nd person non-polite -, 
 1W -
 B([ t), 
1 W[t

/ne/, /caney/ /nehi(tul)/, /caneytul/
polite
{

, o, #
Q 

{
[
 t,
o[
t, #
Q [
 t

/tangsin/, /tayk/, /elusin/ /tangsintul/, /tayktul/, /elusintul/

Table 2.1: Pronominal system of Korean, 1st and 2nd person

a clear-cut dichotomy between honorific and non-honorific second-person pronouns such

as French tu vs. vous, German du and Sie, Russian ty and vy, the case of Korean show-
cases a system that is at the same time more complex and less complete (Table 2.1). The
only subcase where the distinction is clearly established is the first-person pronouns. The
subsystem of the second-person pronouns is considered incomplete in many ways. In ad-
dressing an honored person, using a pronoun is generally avoided; instead, common noun

q

terms which represent their social status such as _
t  (/sensayngnim/, teacher), #
Q
Q_
 (/emenim/, mother), 
_

 (/sacangnim/, company-owner), whenever possible,
are used. All second-person pronouns are highly restricted in their use and are acceptable
within a very specific set of socio-linguistic contexts. {


 /tangsin/ is often cited as a

polite second-person pronoun, but it has the curious property of being mostly associated
with pointedly rude manners used in confrontational situations or altercations. The two
o (/tayk/, home) and #

other terms given in the place of polite second-person pronouns Q


 (/elusin/, senior), also rooted in common nouns, are only used in highly specialized

contexts: #
Q

 (/elusin/, senior) is typically used to address elders in their 70s and 80s
in a traditional family-like settings. All other pronouns in this category are burdened with
similar restrictions. 1
1
For example, 
W
1 (/caney/, you) is often used by a senior person to someone belonging to his off-

39
 proximal medial distal
3rd person human non- E , s
s i,
> , ., s , $
F , $ s
polite /yay/, /ii/ /kyay/, /ku/, /kunye/, /kui/ /cyay/, /ce/, /cei/
polite s  r

r
 
$ r

/ipwun/ /kupwun/ /cepwun/
object s
, s 



$

/i/, /ikes/ /kukes/ /cekes/
place 
# l 
 l 
$ l
/yeki/ /keki/ /ceki/

Table 2.2: Pronominal system of Korean, 3rd person

The third-person pronouns are given in Table (2.2).2 They follow the 3-way distinction
of distance, proximal, medial and distal, which comes from the fact that the pronouns
are composed of the three Korean pre-nominal modifiers encoding the distinction s
 (/i/),

(/ku/), $
 (/ce/) combined with dependent nouns 
s (/ai/, kid), s
 (/i/, person),
r


(/pwun/, person (honorific)), 
(/kes/, thing), and others. Lim and Chang (1995, p. 131)
notes that the composite nature of the 3rd person pronouns reflects its late development in
the history of the language. The three-way distinction of distance encodes deictic relations

with regard to discourse participants: the proximal reference is used to refer to a target
close to the speaker; the medial reference a target close to the hearer; the distal reference a
target that is away from both the speaker and the hearer.
 /i/,
s /ku/, $
 /ce/ alone can function as a pronoun without combining with a de-

pendent noun head, but each is used to different effects.

/ku/ and its plural form
[t

/kutul/ are used to refer to a third-person in written texts only (example 48a). While s


/itul/, the plural form of s
 /i/, can be used to refer to a group of people (example 48b), its
springs generation in his adulthood with whom he has authoritative relationship. An elderly teacher might
use 
W
1 (/caney/, you) towards his student in his 20s; a mother-in-law might use the term to address her
son-in-law.
2
The plural forms are omitted in the table: they are available for the human and object classes and are
systematically derived by attaching the plural suffix morpheme [
t /tul/.

40
 3
singular counterpart s
 /i/ can only be used as a discourse deixis, that is, to refer to an
event-like semantic content denoted by a previous clausal segment (example 48c). $
 /ce/
and its plural form $
[t /cetul/ are used as personal pronouns in archaic texts (example

48d), such as the Bible (Lim and Chang, 1995).

(48) a.
 tH
 
>
. H {
 
9 l\
 %


George-nun keyulu-ta. ku-nun ilha-ki-lul mwues-pota-to
George-Top lazy-Dec. he-Top work-Nomin-Acc anything-than-Even
#
z
 Q
 .
silheha-nta.
hate-PresDec.
George is lazy. He hates working more than anything.

b. /
@Z O
"
r
|
 
a>
' [t\
/
@ 


\
 f"
taypepwen-un saken kwankyeyca-tul-ey tayha-n phankyel-eyse
supreme-court-Top case associate-Plu-Dat regard-RelEnd sentence-At
[
s t\
> 3ws
l  \ 
O  

 rt
: %
3 .
i-tul-eykey calmos-i eps-ta-ko kyellon-cis-ess-ta.
this-Plu-Dat guilt-Nom lack-Dec-Quote conclusion-make-Past-Dec.
The Supreme court concluded in its sentence on case associates that there is

no guilt to these people.

c.
h D # Q;s
2 _
 r  s
4  

 t 

e HX
 <,
yosay elini-tul-uy silyek-i nappaci-ko iss-nuntey,
recently child-Plu-Gen sight-Nom deteriorate-AuxEnd be-CoordEnd,

s H
 
TV r
'A_


% s
.
i-nun olayn TV sicheng-uy yenghyang-i-ta.
this-Top long TV watch-Gen influence-Cop-Dec.
Childrens eyesight is deteriorating these days, and this is the influence of long
TV watching.

d. 
$[ts
 
o _`

 g0

M A\f
" = J#
Q /
? 
4  .
ce-tul-i wuli cwunim-ul wangwi-eyse kkul-e nayli-ess-tota.
that-Plu-Nom our lord-Acc throne-From pull-AuxEnd down-Past-Dec.
Those people (they) pulled down our lord from the throne.
3
In Bonnie Webbers (1991) sense.

41
 There is a general agreement that pronominalization in Korean is neither a wide-spread
nor strict phenomenon. Lee et al. (1997) states that there is no rule in Korean that forces
pronominalization of reference to a previously mentioned entity, and that it is often the case

that a full noun phrase occurs where a pronoun would be appropriate in English. Hence, it
is regarded natural to have a full noun phrases occurring repeatedly in the following text:

(49) (from Lee et al., 1997, p235)

u1 .

H

\
f" I Q #z
.
nwuna-nun pwusan-eyse tayena-ess-ta.
sister-Top Pusan-At born-Past-Dec.
Sister was born in Pusan.

u2 .





`


*


 r

nwuna-ka kohyang-i-n pwusan-ul ttena-n kes-un
sister-Nom hometown-Co-RelEnd Pusan-Acc leave-RelEnd that-Top
1
x
p
< \

 \
%
a 
O f"%
i .
kotunghakkyo-lul colepha-kose-i-ess-ta.
highschool-Acc graduate-After-Cop-Past-Dec.
It was after graduating high school that sister left hometown Pusan.

u3 . 
H
 
 r H
 \
t w
l
3 %
i .
ku hwu nwuna-nun tasi-nun pwusan-ey ka-ci mosha-ess-ta.
that later sister-Top again-Top Pusan-To go-AuxEnd cannot-Past-Dec.
After that, sister couldnt go back to Pusan again.

u4 . Q
`

 




_ E &

r
r



kulena kohyang-ul hyangha-n nwuna-uy ayceng-un hansi-to
but hometown-Acc face-RelEnd sister-Gen affection-Top moment-Even


d r
s
h
&  O
3
\ %
 .
sik-un cek-i eps-ess-ta.
cool-RelEnd time-Nom not-exist-Past-Dec.
But sisters affection towards the hometown never cooled off for a moment.

Here, the repetition of the full noun phrase


(/nwuna/, sister) is preferred over use of

use of pronouns
/ku/ or its female counterpart
. /kunye/: the pronouns are mostly
found in highly stylized texts such as novels, and given the speakers presumed familiarity

42
with the referent, they would have sounded odd and overly formal. Lee et al. (1997) does
not consider the other obvious alternative, however: zero pronominalization of sister,
which would result in natural-sounding discourse as well. In fact, there is one case of zero

pronominalization already present in the sequence: the subject of the verb a

\
% 
O  /colepha/
to graduate is left empty. If this position were filled with the overt noun phrase, it would
have felt unnecessarily repetitive, as with the case we have seen before in example (46a).

2.3 The Problem of Identification: Where to Find the In-

visible

Zero pronouns, being an invisible linguistic construct, present a fundamental problem to

be addressed prior to their treatment, namely their identification. In fact, the problem of
identification has direct consequences on the theory of zero pronouns in that it can be
approached in a way that facilitates treatment of a particular problem of zero pronouns, for

example, zero pronoun resolution.

In order for an unexpressed element to be identified, it has to be predictable from the
structure of the sentence in which it is contained. Therefore, identification of zero pronouns
crucially relies on the specific representation of the clause and sentence structures. The

most significant indicator is, arguably so in any theory, the predicate. Predicates have their
own subcategorization schemes, and when an argument is missing from such a scheme,
one can reasonably deduce the presence of a zero pronoun in the place.
The Korean Propbank Annotations (Ryu et al., 2006) is a collection of annotation on the

syntactic and semantic relations between predicates and their arguments found in the Penn
Korean Treebank corpora. It encodes such relations with framesets, which are attributed to
predicate roots. The following is an example frameset for the verb o
v
 (/kalikhita/
to point, indicate):

43
 (50) frameset kalikhi.01 indicate:
Roleset:

Arg0: indicator

Arg1: thing indicated

pred: kalikhita

arg0: sbj
arg1: obj

Example:

r
 >H

y

&
\
10r
 o v
e
3 %
 .
sikyey-nun cengkak 10-si-lul kalikhi-ko iss-ess-ta.
clock-Top sharp 10-hour-Acc indicate-AuxEnd is-Past-Dec.

The clock was indicating 10 oclock sharp. Arg0-nun: clock

Arg1-lul: 10 oclock sharp

Rel: indicate
Aux-ko: is-Past-Dec

Given such a subcategorization frame for a verb, identifying a zero argument is equivalent
to determining that an argument in the frame is missing and, furthermore, which argument
it is, if the predicate warrants more than one arguments. Therefore, when the following
example with the same verb is encountered:

(51) f
"A

` o ( 
.
ceccok-ul kalikhi-ess-ta.
west-Acc indicate-Past-Dec.
(Something/someone) pointed to the west.

there is only one argument when two is required by the subcategorization frame; the present
argument "
fA

` (/ceccok-ul/, West-Acc) is Arg1, which can be verified by the ac-

44
cusative case marker `
/ul/; therefore a zero pronoun is identified in the subject position,

marked by Arg0.

2.4 Korean Zero Pronouns by Reference Types

Despite sharing the same form factor, Korean zero pronouns are made up of several sub-
groups, each of which are characterized by distinct semantic and pragmatic roles they play.
Korean zero pronouns can be divided into four types according to the type of reference that
they constitute.

In a categorization proposed by Kameyama (1985) for Japanese zero pronouns, four

subcategories are recognized, which are grouped into two major types, text-dependent
use and text-independent use. Three subcategories belong to the type of text-independent
use, which cover non-anaphoric usage of zero pronouns in Japanese. She calls those zero
pronouns with text-dependent reference the discourse anaphoric zero anaphors. The zero

pronouns belonging in this category are truly anaphoric in nature: they derive their refer-
ents (usually 3rd person) from what has been expressed before. This classification is given
in Table 2.3.

text dependent use discourse anaphoric zero anaphor

zero deictic and zero indexical

text independent use indefinite personal zero anaphor
general situational zero anaphor

Table 2.3: Classification of Japanese zero anaphor by Kameyama (1985)

It is imperative to note the inconsistency in Kameyamas use of the term anaphor/anaphoric.

In Levinsons (1983) definition, an anaphoric usage is where some term picks out as ref-
erent the same entity (or class of objects) that some prior term in the discourse picked

45
out. Lexicon of Linguistics 4 defines anaphoric pronoun as a pronoun which refers back
to another constituent in the sentence, and more generally anaphor as an element which
depends for its reference on the reference of another element. Since her indefinite per-

sonal zero anaphor and general situational zero anaphor fall under text-independent use,
they cannot be anaphoric according to these definitions. Rather, she is using anaphor as
loosely equivalent to pronominal or pronoun. In her use of the term anaphoric in discourse
anaphoric zero anaphor, however, she means precisely what is stated in the definitions
given above, i.e., an anaphoric usage of a pronoun whose referent is dependent on some

previous mention in the context.

For Korean, we adopt the following classification scheme with modified terminology:

anaphoric zero pronoun

text dependent use
discourse-deictic zero pronoun
deictic zero pronoun
text independent use indefinite personal zero pronoun
general situational zero pronoun

Table 2.4: Classification of Korean zero pronouns

After conducting a corpus study of naturally occurring Korean data, the classification is
found to be suited to Korean zero pronouns. Detailed discussion on the four types of
Korean zero pronouns is presented in the following sections.

2.4.1 General Situational Zero Pronouns

Kameyama (1985) defines general situational zero pronouns as zero subjects that refer
to time, weather, and a general situation, as English expletive subject it does. A group
of Korean zero pronouns display a parallel usage. The following shows some typical,
constructed examples:
4
http://www2.let.uu.nl/UiL-OTS/Lexicon/

46
 (52) a. +
O
Z \ r
P .
pelsse yel-si-ngi-ta.
(SBJ) already ten-oclock-Cop-Dec.
(It) is 10 oclock already.

b.

.
cal tway-ss-ta.
(SBJ) well become-Past-Dec.
(It) turned out well.

English expletive pronouns are said to lack in semantic content: The verbs that license
them (such as weather verbs rain) do not project a thematic argument to the subject position
that they occupy. If this particular type of Korean zero pronouns are true counterparts of
the English expletive pronoun it, then it follows that they must lack semantic content as

well.
Given that the subject slots of the two sentences are both phonetically and semantically
null, one might ask: Is it necessary to posit a zero syntactic element for them in the first
place? In positing a zero pronoun in the subject slot, are we drawing on a particular linguis-

tic theory such as GB? 5 Naturally, one can subscribe to a syntactic theory for Japanese or
Korean which, unlike a theory for languages such as English, simply does not require verbs
to have a syntactic subject. We will not pursue this matter further here, as it involves a full-
scale debate of the syntactic theory of Korean. Instead, we offer an argument in support
of situational zero pronouns from a more practical perspective. For an application dealing

with zero pronouns, it is necessary at one stage or another to examine all possible argument
slots of a predicate and determine whether it is filled or not, before deciding whether the
zero pronoun is semantically present. This is true because the subject slot of those pred-
icates with situational zero subjects is under no syntactic obligation that they should be
5
Kim (2003) casually notes that it is more natural to leave the subject positions empty in these sentences
and suggests that it is more reasonable to view the sentences as not having a subject in the first place rather
than viewing the subject as having been dropped.

47
empty, as seen in the examples below:

(53) a. 
tF K r s
 Z+
O \ r
P .
cikum sikan-i pelsse yel-si-ngi-ta.
now time-Nom already ten-oclock-Cop-Dec.
The present hour is 10 oclock already.

b.
{ s
9 

.
ku il-i cal tway-ss-ta.
that matter-Nom well become-Past-Dec.
That matter turned out well.

In a way, the very fact that the predicates are capable of taking overt noun phrase subjects
indicate that the zero pronouns are not semantically null after all. Rather, they refer to
general time, general situation and so forth, which tend to be left unexpressed. On
the other hand, in some idiomatic expressions involving verbs which permit situational

subjects, forcing an overtly expressed subject achieves varied levels of acceptability. For
example:

(54) a. -
H
 j
s ] 


1 .
ne-nun icey ka-eto kwaynchan-nta.
you-Top now go-Vend ok-PresDec.
You can go now. Literally: (It) is ok even if you go.

-
b. ? H
 j
s ] 
H>
 


1 .
?ne-nun icey ka-eto motunkes-i kwaynchan-nta.
?you-Top now go-Vend everything-Nom ok-PresDec.
Everything is ok even if you go.

(55) a. -
H
 j
s ]  a
)
.
ne-nun icey ka-eto twoy-nta.
you-Top now go-Vend become-PresDec.
You can go now. Literally: (It) is well even if you go.

-
b. ?? H
 j
s ] 
H>
  a
)
.
??ne-nun icey ka-eto motunkes-i twoy-nta.
??you-Top now go-Vend everything-Nom become-PresDec.
Everything is well even if you go.

48
 Given these facts, it is clear that whether or not an argument has semantic content is
not an a priori matter in Korean. Therefore, it is reasonable for an application to start
with every argument position that is unfilled and determine whether there is a null element

that has a semantic content. This, in our opinion, can be achieved without committing to
any specific theory of syntax. In our thesis, then, the term zero pronoun is used free of
the theoretical implications that the term usually evokes; it is best rephrased as unfilled
argument position.
There is another distinct usage of zero pronouns that is frequently encountered, which

we group together under the general situational zero pronoun category. It involves id-
iomatic expressions with a verbal element, such as the following:

(56) @
a. ...\ /Kf "
...ey tayha-ese
...Dat regard-VEnd
regarding ...

b. ...\ aK
' f "
...ey kwanha-ese
...Dat regard-Vend
regarding ...


c. ...\  f"
...ey ttalu-ese
...Dat follow-Vend
according to ...

0

d. ...` AKf "
...ul wiha-ese
...Acc care-Vend
for the sake of ...

(57) o
^
= H
 
t
r
: o
\ /
@ Kf " 
O
[ "


.
chelswu-nun cikwuonnanhwa-ey tayha-ese selmyeng-ha-ess-ta.
chelswu-Top global-warming-Dat regard-Vend explanation-do-Past-Dec.
Chelswu explained regarding global warming.

49
 In the example above, the verb @
/ (/tayha/, to regard) takes an oblique argument t


r
: o
 (/cikwuonnanhwa/, global warming) and an empty subject to form an adver-

O
bial clause which modifies the main clause verb [ "

 (/selmyengha/, to explain). The

empty subject, then, is the verbal content that is being modified by the adverbial clause: it
is the action of explaining that is the semantic content of the relation of regarding. This
can be verified by relativization facts:

(58) 
t
r
: o
\ /
@ 
=
o
^ _  
O
[

"
cikwuonnanhwa-ey tayha-n chelswu-uy selmyeng
global-warming-Dat regard-RelEnd chelswu-Gen explanation
Chelswus explanation which regards global warming

This type of zero pronoun use is decidedly not a main-clause phenomenon. The verbal
construction cannot be used as a main verb, as shown by the awkwardness of Example
(59a); to achieve a similar sense, an equative construction has to be used which involves

the dependent functional noun head 
(/kes/, that; thing) as in (59b):

(59) =
o
a. ?!^ _  
O
[ "

r

t
r
: o
\ /
@ 
.
?!chelswu-uy selmyeng-un cikwuonnanhwa-ey tayha-ess-ta.
?!chelswu-Gen explanation-Top global-warming-Dat regard-Past-Dec.
Chelswus explanation regarded global warming.

b. o
^
= _ 
O
[ "

r

t
r
: o
\ /
@ 

chelswu-uy selmyeng-un cikwuonnanhwa-ey tayha-n
chelswu-Gen explanation-Top globalwarming-Dat regard-RelEnd
s

 %
3 .
kes-i-ess-ta.
thing-Cop-Past-Dec.
Chelswus explanation was one regarding global warming.

All the expressions given in (56) share these properties. We suspect that this type of

zero pronouns display a set of characteristics that are distinct enough from those of the
previous cases of situational zero pronouns that in some theories they are better served
with a category of their own. For the present study, they are grouped under the situational

50
zeros with their differences duly noted, on the grounds that they share the common property
of resisting instatiation by overt referring expressions.

2.4.2 Deictic Zero Pronouns

Zero deictic pronouns are those that directly refer to an entity in the given spatio-temporal

context. Their referent can be an object or a third person present in the scene of the dis-
course as in (60), or, more frequently, discourse participants including the speaker and the
hearer as in (61).

(60) (looking at an elephant)


3
% A
' 1
W.
emcheng ku-ney.
(SBJ) very big-Excl.

(It)s very big!

(61) (a dialogue between A and B)

u1 .A Q
# n m?
eti ka-ni?
(SBJ) where go-Q?
Where are (you) going?

u2 .B
6,
x f
" a\
' .
ung, tosekwan-ey ka-e.
yeah, (SBJ) library-To go-Dec.
Yeah, (I) am going to the library.

u3 .A
!
3 s

 .
kulem kathi ka-ca.
then (SBJ) together go-Exhor.
Then let (us) go together.

The first and the second person deictic pronouns are often closely tied to grammatical
elements of the sentence or the semantics of the verb that they are an argument of, which

51
provides cues for their recovery. Some grammatical moods mandate either the first or the
second deictic subjects; psychological verbs, stating intention or psychological status of
the subject, require the first or the second person subject. Three grammatical moods are

strongly tied to particular types of deictic references: Imperatives require the hearer as the
subject of the sentence (example 62a); exhortatives require the speaker and the hearer as
the subject (62b); and promissives the speaker (62c). In Korean, moods are expressed by
verbal ending suffixes: #
Q /ela/, 
 /ca/, `
a
 /lkkey/ are verbal endings that encode the
imperative, exhortative, and promissive moods, respectively.

(62) a. Imperative

d
h
&
 " #
3 Q.
cemsim mek-ela.
(SBJ) lunch eat-Imp.
(You) eat lunch.

b. Exhortative

s

 .
kathi ka-ca.
(SBJ) together go-Exhor.
Let (us) go together.

c. Promissive

{
 ?
9 /{
9 a
+
.
ku il nayil ha-lkkey.
(SBJ) that work tomorrow do-Prom.
(I) will do the work tomorrow.

Kim (2003) claims that the subject positions of clauses with such moods sound more
natural when left empty, and this is due to the fact that the subject is readily identifiable
from the information carried by the mood suffixes. This line of argument is similar to those
given for pro subjects in languages such as Spanish and Italian whose recoverability is said

52
to be due to their rich verbal agreement morphology. She argues that overt subjects instead
of zero subjects in such constructions carry a contrastive sense in the following examples:

(63) (adapted from 29.a of Kim (2003))

a. -
H
 j
s ] .
ne-nun icey ka-ela.
you-Top now go-Imp.
You go already.

b.
o H s
 j ] .
wuli-nun icey ka-ca.
we-Top now go-Exhor.
Let us go already.

It is true that the sentences carry a contrastive sense: (63a) is appropriate in a situation
where A is ordering B to leave while implying C can stay. However, the contrastiveness
effect is hardly due to the overtness of the subject itself: Korean postposition marker
r/
H

(/un/nun/), commonly viewed as the topic marker, are also known to mark contrastiveness

(Lee, 2000, 2003, among others) in some contexts. Indeed, when a bare pronoun NP is
substituted in place of the nun-marked pronoun NP, the sentences are neutral without the
sense of contrastiveness:6

(64) a. Imperative

 &
- d
h
 " #
3 Q .
ne cemsim mek-ela.
you lunch eat-Imp.
You eat lunch.

6
The subject of exhortative mood can be second-person as long as it is semantically clear that the action
denoted by the verb is taken together by the speaker and the hearer. This can be achieved by including
adverbials such as
s
 /kathi/ togetherwith adverbial phrases such as

|
/na-lang/ me-with as in: 
-
|


s
 
. (You me-with together go.)

53
 b. Exhortative


o s
 .
wuli kathi ka-ca.
we together go-Exhor.
Let us go together.

c. Promissive

 {  ?
9 /{
9 a
+
.
na ku il nayil ha-lkkey.
I that work tomorrow do-Prom.
I will do the work tomorrow.

But what about the subject pronoun marked with the nominative case marker s
/
/i/ka/? Case markers in Korean, both nominative and accusative, are known to function as
a contrastive focus marker in some environments where an open proposition of the form
[Pred(x)] is cognitively active in the context and the identity of the entity x is in question

(Lee, 2000, 2003, among others). Therefore, the following example, where the second-
person pronoun is overt and is accompanied by the nominative marker  /ka/, is only
acceptable where the question who will eat lunch? is active in both speakers and hearers
minds.

(65) Imperative (overt pronoun subject with the nominative case marker)

W
1 d
h
&
 " #
3 Q .
ney-ka cemsim mek-ela.
you-Nom lunch eat-Imp.

You eat lunch.

In fact, what distinguishes deictic subjects occurring with such fixed-subject moods
from others is the very availability of the contrastive readings. In the example (66 u2 .B)
below, which is used in a neutral context and is in the declarative mood, the contrastive-

54
ness effect is notably absent with both the topic-marked overt pronoun subject and the
nominative-marked overt pronoun subject.

(66) u1 .A
 A Qb
# G>
 t / ?m?
kulay ettehkey cinay-ni?
so (SBJ) how do-Q?
So, how are (you) doing?

u2 .B
6, /

x /
 H/?
 / h
D 7 
.
ung, /na/na-nun/nay-ka yosay com pappu-e.
yeah (SBJ)/I/I-top/I-nom recently little busy-Dec.
Yeah, (I)/I am a little busy these days.

In summary, in sentences with the three deictic-subject-oriented moods, the topic/nominative-

marked overt pronoun subject always carries the contrastive sense, while in other moods

they can be used in neutral contexts.

Chang (1986) notes that zero subjects accompanying psychological predicates such as

l 
 /kipputa/ is happy, a

 
  /culkepta/ is joyous,


 /pankapta/ is delighted
are interpreted as the first-person in the declarative mood while they are interpreted as the

second-person in the interrogative:

(67) a. 

G
D >s

h  
s f" 
l .
hankwuk thim-i iki-ese kippu-ta.
korea team-Nom win-CoordEnd (SBJ) happy-Dec.
(I) am happy that the Korean team won.

b. 

G
D >s

h  
s f" 
l t?
hankwuk thim-i iki-ese kippu-ci?
korea team-Nom win-CoordEnd (SBJ) happy-Q?
Are (you) happy that the Korean team won?

Generally speaking, declaring an emotional state of someone other than the speaker her-
self is avoided in Korean; instead, Korean has corresponding verbs for such psychological
adjectives, formed by verbalization suffix #
Q (/eha/), which lets the speaker phrase the

55
event as an observed outward display of emotion (e.g., She rejoiced in example 68) and
avoid directly stating the persons inward feelings (e.g., She was happy).7

(68) 

G
D >s

h  
s f"


% BH
 
l 
.
hankwuk thim-i iki-ese yenghi-nun kippu-eha-ess-ta.
korea team-Nom win-CoordEnd yenghi-Top happy-Vbzx-Past-Dec.
Yenghi rejoiced that the Korean team won.

2.4.3 Indefinite Personal Zero Pronouns

Indefinite personal zero anaphors are usually a subject and refer to people in general. There

are a number of distinct environments where such indefinite personal zero pronouns occur:
a modal environment (69a), complementized clauses including the gerund construction
(69b), a subject position followed by a scene-setting adverbial (69c).

(69) a. |
s
\

 9 
 \

 
 .
holangi-lul cap-ulye-myen san-ey ka-ya-ha-nta.
(SBJ) tiger-Acc catch-Intend-If (SBJ) mountain-Des go-must-PresDec.
If (one) wishes to catch a tiger, (one/he) must go to the mountains.

b. +

`
{
9
 '
+ 
p H



ha-l il-ul twi-lo milu-nun kes
(SBJ) (SBJ) do-RelEnd task-Acc later-Des postpone-RelEnd Comp
(For one) to postpone a task (for one/him) to do.

c. 2 @
// @\f "H
 Q
# 
" @
 /\
 6
 x
H?

2 taytay-eyse-nun etten hochwultayho-lul sayongha-ess-nunka?
2 battalion-at-Top (SBJ) what call-sign-Acc use-Past-Q?
In the 2nd Battalion what call signs do (they) use?

These pronouns correspond to a large variety of pronouns of arbitrary interpretation

across languages. Generic zeros in conditional/modal environments are analogous to lex-

7
An obvious exception is in narrative texts such as novels, where the narrating voice assumes an omni-
scient stance.

56
ical impersonal pronouns in European languages such as English one, German man and
French on.

(70) a. John doesnt know how PROarb to behave oneself at parties.

b. PROarb Leaving early would be inexcusable.

Generic zeros in complementized clauses, on the other hand, are analogous to uncon-
trolled PRO subjects of infinitival clause and subject of gerund construction, as in Example
(70) above. The crucial difference between the English and the Korean cases, however, lies

in the following: while English PROs appear in uninflected clauses, corresponding envi-
ronments in Korean lack any discerning morphological or syntactic characteristics, and, as
such, the zero pronoun slot can always be occupied by an overt nominal expression. Korean
pros, therefore, subsume both pro and PRO in the standard linguistic theory of Government

and Binding (see Section 1.1.1).

Finally, in English and Russian (Malamud, 2004) the third-person plural pronoun they
and the zero pronoun respectively encode generic reference of the subject accompanied by
a scene-setting adverbial:

(71) a. They speak English in America.

b. Na galerke zataili dyhanie.

On gallery held-3pl breath.
In the gallery, (they) held their breath.

2.4.3.1 Specific Indefinite Zero Pronouns

Not all indefinite zero pronouns are generic in terms of their meaning. Sometimes, it is
understood from the context that there has to be a specific entity that a zero pronoun refers
to, although it is not mentioned anywhere in the text or is not locally accessible. Take the
following sentence, for example:

57
 (72) indefinite specific zero pronouns

H
 N
B
/
<
%

o
_ 

d
` =
y
n 9
kyupu-nun kongsangkwahakyenghwa-uy thul-ul pilli-e
The-Cube-Top (SBJ) SF-movie-Gen frame-Acc adopt-Aux

p _

1 s
 H
 

o\ %


o

H

milo-uy thalchwul-ilanun sinhwa-lul yenghwa-lo mantu-n
labyrinth-Gen escape-Quote myth-Acc movie made-RelEnd


s
 .
cakphwum-i-ta.
work-Cop-Dec.

Cube is a work such that (someone) made the myth of an escape from a labyrinth

into a movie by adopting a SF-movie framework.

It is implied that there exists someone who is the maker of the movie (i.e., the director),
which is the referent of the subject zero pronoun. However, a mention of the director occurs
only at a later part of the text, after several intervening paragraphs. 8 Contrast this case with

the following example, where the zero pronoun is referring to an arbitrary entity:

(73) indefinite generic (non-specific) zero pronouns



~ _ Qn
# u q\

5 f" U `

 `
{
9
 : t
M  o H

pam-uy etwum sok-eyse kil-ul ilh-l ttay cili-nun
(SBJ) night-Gen darkness midst-at way-Acc lose-RelEnd when geography-Top
8
A related concept is bridging references, the phenomenon of which was extensively investigated in
Poesio, Vieira and Teufel (1997), Vieira and Poesio (2000). Bridging references are defined as those uses of
definite descriptions based on previous discourse which require some reasoning in the identification of their
textual antecedent. For example, the definite description the suffering people are going through is linked
via inference to the previously occurring expression last weeks earthquake (Poesio, Vieira and Teufel,
1997). Note that in the case of Korean, the distinction between definiteness/indefiniteness is not inherent, and
the case retains some aspects of either: it is like a definite description in the sense that a unique referent is
identifiable from the context through semantic reasoning; it is nevertheless an entity that was never addressed
in the prior discourse, which is a property of indefiniteness. We prefer leaving it as an indefinite usage, albeit
a specific one.

58
 B
 o
#   .
maywu phenliha-ta.
very convenient-Dec.

When (one) gets lost in the darkness of the night, geography comes in very handy.

Notably, the indefiniteness of such zero pronouns is closely tied to the semantic in-

terpretation of the sentence in which they occur. The sentence often involves a modal
environment such as conditionals and the irrealis tense.
Cinques (1988) take on third-person plural pronouns with arbitrary interpretation can
be extended to these cases of indefinite Korean zero pronouns. He treats them as a kind

of indefinite, i.e., a free variable, which is bound through different mechanisms in episodic
sentences and generic sentences. In the former, it is the existential quantifier over events
that unselectively binds the free variable introduced by the indefinite; in the latter, the uns-
elective binding is done by the generic operator, yielding a quasi-universal interpretation.
This suits the two cases of Korean zero pronouns seen above in (72) and (73), since the

embedded clause in the former example has episodic interpretation, and the modal envi-
ronment in the latter gives rise to a generic meaning.
This elegant approach is not without controversies, however. Cabredo-Hofherr (2002)
notes that the distribution of existential and universal readings does not coincide with the

episodic/generic distinction. Two examples in Spanish, involving zero pronouns:

(74) a. En este parque juegan futbol en la tarde.

in this park play football in the afternoons.
In this park, (they) play football in the afternoons.

b. Ayer en Espana celebraron el da del trabajo.

yesterday in Spain celebrated the day of-the work.
In Spain, (they) celebrated Labour Day yesterday.

In (74a), the zero pronoun receives an existential interpretation even though the sen-
tence has a generic meaning. On the other hand in (74b) which has an episodic reading, the

59
zero pronoun receives a universally quantified interpretation. These types of zero pronoun
usages can be found in abundance in the Penn Korean Treebank corpora. The following ex-
ample illustrates a Korean zero pronoun with a preferred universal interpretation occurring

in an episodic sentence:

(75) u1 .A 2 @// @\f "H




\
 # Q "

2 taytay-eyse-nun ku cakcen-ey etten
2 battalion-at-Top (SBJ) operation-at what call-sign-Acc

@
 /\
 6
 x

H?

hochwultayho-lul sayongha-ess-nunka?
use-Past-Q?
In the 2nd Battalion what call signs did (they) use in the operation?

u2 .B 1
x`
l
 6
 x
_
vm
.
chwutong-ul sayongha-ess-upnita.
(SBJ) chwutong-Acc use-Past-PolDec.
(They) used chwutong.

This leads to the conclusion that Cinques (1988) approach alone does not provide a
sufficient account for the arbitrary/indefinite and generic interpretations of the Korean zero

pronouns reviewed in this section. Indeed, the very idea that the universal/existential read-
ings of the zero pronouns stem from their allegedly indefinite nature comes under question.
We do not pursue this matter further here, leaving detailed discussions to future work.

2.4.3.2 +human Semantic Restriction on Generic Zero Pronouns

Notably, generic zero pronouns are subject to a semantic restriction which limits their refer-

ents to humans. Even in the presence of a scene-setting adverbial which triggers appropriate
semantic associations, zero pronoun interpretations with non-human referents for generic
pronouns are hard to come by. In (76a) and (76b), the scene-setting adverbial in zoos
should help invoking the entities that are naturally expected to exist in the scene, i.e., zoo

animals, thereby producing a bridging effect which lets readers associate the referent of

60
the generic zero subjects with them as salient entities. In reality, however, the -human in-
terpretation is unnatural (example 76a), whereas the +human interpretation where the zero
subjects refer to zookeepers are acceptable (76b).

(76) x
l
a. ?1 t
" \
f"H
 
\ z s
3
" \
 
3
" H
 .
?tongmwulwon-eyse-nun halu-ey twu-kki meki-lul mek-nunta.
?zoo-at-Top (SBJ) day-in two-Class fodder-Acc eat-PresDec.
?In zoos (they, e.g., animals) eat fodder twice a day.

b. l
1
x t
" \
f"H
 
\ z s
3
" \
 r

.
tongmwulwon-eyse-nun halu-ey twu-kki meki-lul cwu-unta.
zoo-at-Top (SBJ) day-in two-Class fodder-Acc give-PresDec.
In zoos (they, e.g., zookeepers) give fodder twice a day.

The following examples illustrate that the generic zero subjects cannot be inanimate

objects either. The zero subjects following the scene-setting adverbials in this airport
cannot refer to airplanes, which arguably are high on the saliency scale given the scene,
while airport staff is a perfectly fine choice for their referent.

(77)  /
a. ?s N
B \
f " H
 
\
VY H

.
?i konghang-eyse-nun halu-ey swupayk-chalyey ttu-nta.
?this airport-at-Top (SBJ) day-in hundreds-times takeoff-PresDec.
?In this airport (they, i.e., airplanes) take off hundreds of times a day.

b. 
s / N
B \
f " H
 
\ "

_

i konghang-eyse-nun halu-ey swuman-myeng-uy
this airport-at-Top (SBJ) day-in 10s-of-thousands-CLASS-Gen
x
p
5 o[
t

` 
% o

 .
sungkayk-tul-ul cheliha-nta.
passenger-Plu-Acc process-PresDec.
In this airport (they, i.e., airport staff) process tens of thousands of passengers

a day.

For anaphoric zero subjects, by contrast, -human referents are freely allowed. Example
(78) illustrates a case of zero subjects (in u2 ) anaphorically referring back to the plural

61
+human subject of the previous sentence. Likewise in (79), the zero subject in the second
sentence is naturally understood as coreferential with the -human -animate subject of the
previous sentence, buildings in Philadelphia.

(78) u1 .

 ' H\




e H
 
 p
r7 
|[
t
r

layngkhastekwun-ey sal-ko iss-nun amisicok salam-tul-un
lancaster-county-in live-VerbEnd is-RelEnd Amish-clan person-Plu-Top
:

x&

h 
 t
q
~d
`



 .
centhongcekin saynghwalpangsik-ul koswuha-nta.
traditional lifestyle-Acc uphold-PresDec.
The Amish people living in Lancaster county are upholding traditional way of
life.

u2 . l

< 
r
@
&
/& 
h _
#
\
 ^
{
9
cenki-wa kasu kath-un hyuntaycek phyenuy-lul ilchey
(SBJ) electricity-and gas like-RelEnd modern convenience-Acc entirely

  9 j
18[ l 

[
t_
 _
v@
/
kepwuha-mye 18-seyki senco-tul-uy mosup-taylo
reject-While (SBJ) 18-century ancestor-Plu-Gen appearance-as




e .
sal-ko iss-ta.
live-VerbEnd is-Dec.
(They) are living exactly like their ancestors in the 18th century as (they) reject
modern conveniences such as electricity and gas altogether.


(79) u1 . 9
 4 qx
S  \H
 
p G
D

&

l \ Q
t # 

phillateylphia-ey-nun mikwuk cengchak choki-ey cieci-n
Philadelphia-in-Top America settlement beginning-in built-RelEnd


| t
t
[s  !
# e

.
kenmwul-tul-i yeles iss-ta.
building-Plu-Nom several exist-Dec
In Philadelphia, many buildings exist which were built in the early days of
American settlement.

u2 . r
{
_  

|
d
`

 /
@  H

tangsi-uy kenchwuk-yangsik-ul tayphyoha-nun
(SBJ) that-time-Gen architecture-style-Acc represent-RelEnd

62
 )

 
H

o s
 .
kwicwungha-n mwunhwa-yusan-i-ta.
valuable-RelEnd culture-legacy-Cop-Dec
(They) are valuable cultural legacy which represent the architectural style of

the time.

The inability of Korean generic zero pronouns to take -human reference is not an iso-
lated phenomenon: similar semantic restrictions are documented for generic pronouns in
various languages. For example, in French, on cannot take non-human reference (Prince,

2003). A similar phenomenon is reported for Yiddish men (Prince, 2003).

(80) a. Ou est Jeani /mon livrej ? Ili/j /Oni/j est la.

where is Johni/my bookj ? Hei /Itj /Onei/j is there.
Where is Johni/my bookj ? Hei /Itj /Onei/j is there.

b. Ayer gelti /Menj iz avek, fun vanen esi/j /meni/j iz gekumen.

your moneyi/onej is away, from where iti/j /onei/j is come.
Your moneyi/Onej went away as iti/j /onei/j came.

2.4.3.3 Coreference in Generic Zero Pronouns

While generic zero pronouns are understood to refer to no one in particular, coreferential
readings between two or more such zero pronouns are favored in some cases. Consider:

(81) a. |
s

\
 9 
 \

 
 .
holangi-lul cap-ulye-myen san-ey ka-ya-ha-nta.
(SBJ) tiger-Acc catch-Intend-If (SBJ) mountain-Des go-must-PresDec.
If (one) wishes to catch a tiger, (one/he) must go to the mountains.

b. 


1  H
 g
 @
y>
p
0  H
<
`

 

p
paopapnamwu-nun cachic nuckey son-ul ssu-myen kuttay-n
paopap-tree-Top (SBJ) per-chance late hand-Acc use-If then-Top

& 
% u +

>
O
\  a
)
.
cengmal chechiha-l swu eps-key twoy-nta.
really (SBJ) remove-Rel possibility not-exist-AuxEnd become-Dec.

63
 Baobab tree becomes impossible (for one) to remove if (one/he) treats it too
late.

c. +

`
{
9
 '
+ 
p H



ha-l il-ul twi-lo milu-nun kes
(SBJ) (SBJ) do-RelEnd task-Acc later-Des postpone-RelEnd Comp
(For one) to postpone a task (for one/him) to do

In these examples, all the zero pronouns refer to generic people or an arbitrary some-

one. Therefore they can be categorized as an indefinite generic use of a zero pronoun.
However, note that the two pronouns within each sentence are understood to be coreferen-
tial: in example (81a), the person who wishes to catch a tiger and the person who has to
go to the mountains are the same one, whoever it might be. This fact is expressed in the
English translation for each sentence, where the second occurrence of one is substitutable

with he. Semantically, one of the two zero pronouns is a bound variable, whose referent is
determined through the other one.9
Coreference can also occur across sentential boundaries. In this case, however, a certain
parallelism effect seems to be required, such as continuation of a modal environment (82)

or question-answer pairs (83).

(82) u1 g7 |
s
\

 9 
 g7 \

 
 .
g7 holangi-lul cap-ulye-myen g7 san-ey ka-eyaha-nta.
(SBJg7 ) tiger-Acc catch-Intend-If (SBJg7 ) mountain-Des go-Must-PresDec.
If (oneg7 ) wishes to catch a tiger, (one/heg7 ) must go to the mountains.

u2


& 
, g7 x
l
1 t
" \
  
 .
kukes-to antoy-myen, g7 tongmwulwen-ey-lato ka-eyaha-nta.
it-Also not-do-if, (SBJg7 ) zoo-to-even go-Must-PresDec.
If that wont do, (one/heg7 ) must still go to a zoo.

9
Alternatively, both are variables bound by an operator which resides higher up in the tree, giving them
generic interpretation.

64
 (83) u1 A. 2 @// @ \f" H
 g1 Q
# 

 \


 9

2 taytay-eyse-nun g1 yele mwucenso-lul hochwulha-lyemyen
2 battalion-At-Top (SBJg1 ) multiple station-Acc call-Intend
g1 Q
# 
" @
 /\
 6
 x
H?

g1 etten hochwultayho-lul sayongha-nunka?
(SBJg1 ) what call-sign-Acc use-Q?
In the 2nd battalion, if (theyg1 ) want to call multiple stations, what call sign
do (theyg1) use?

u2 B. g1



HX
 <\
 \

 :<
M

>
r

g1 mwucenso hankwuntey-lul pwulu-l ttay-wa ttokkath-un
(SBJg1 ) station one-place-Acc call-RelEnd time-With same-RelEnd


d g1 m
+
.
sik-ulo g1 ha-pnita.
way-In (SBJg1 ) do-PolEnd.
(Theyg1 ) do in the same way as when (theyg1 ) call one station.

Coreference interpretation between these zero pronouns is not obligatory and depends
largely on availability of coherent semantic interpretation. Generic zero pronouns in the
two conjoined clauses are not necessarily co-referential as seen in (84); disjoint interpreta-

tion is forced in (85), where the verb phrase @

/
 
 (/taysin hata/, do instead/substitute)
is used instead of p

 (/milwuta/, postpone):

(84) gx |
s
\

 9 
 gx \

 
, gy
gx holangi-lul cap-ulye-myen gx san-ey ka-eyaha-ko, gy
(SBJgx ) tiger-Acc catch-Intend-If (SBJgx ) mountain-To go-Must-And, (SBJgy )
t

l\

9 
 gy 
  \  
 .
mwulkoki-lul cap-ulye-myen gy pata-ey ka-eyaha-nta.
fish-Acc catch-Intend-If (SBJgy ) sea-To go-Must-PresDec.
If (onegx ) wishes to catch a tiger (one/hegx ) must go to the mountains, and if
(onegy ) wishes to catch fish (onegy ) must go to the sea.

(85) g1 g2 +

`
{
9
 /
@  K
 
H



g1 g2 ha-l il-ul taysin ha-e cwu-nun kes
(SBJg1 ) (SBJg2 ) do-RelEnd task-Acc instead do-Aux give-RelEnd Comp
(For oneg1 ) to substitute for a task (for someone elseg2 ) to do

65
2.4.4 Discourse-Anaphoric Zero Pronouns

Discourse-anaphoric zero pronouns, or simply anaphoric zero pronouns, are those zero

pronouns which find their referents in what has been said earlier in the text. In Levinsons
(1983) definition, an anaphoric usage is where some term picks out as referent the same
entity (or class of objects) that some prior term in the discourse picked out. Lexicon of
Linguistics 10 defines anaphoric pronoun as a pronoun which refers back to another con-
stituent in the sentence, and more generally anaphor as an element which depends for its

reference on the reference of another element. Therefore, an anaphoric relation consists of

two elements: the anaphor, or the anaphoric expression, at hand, and the antecedent from
11
which the anaphor derives its reference.
In the following example (Korean translation of The Little Prince), the zero subject of

(u2 ) is an anaphor, and it depends for its reference on the subject of the preceding sentence
(u1 ), i.e.,


(/poapaym/, boa constrictors), which is its antecedent.

(86) anaphoric zero pronoun

u1 .

r
s
3
" \
 t
9 
xP

: :


 
 .
poapaym-un meki-lul ccip-ci-anh-ko thongccaylo samkhi-nta.
boa-snake-Top prey-Acc chew-Aux-Neg-And as-whole swallow-PresDec.
Boa constrictors swallow their prey whole, without chewing.

10
http://www2.let.uu.nl/UiL-OTS/Lexicon/
11
It occurs more commonly that the anaphor follows the antecedent in linear order, which is the etymo-
logical basis of the terms anaphor and antecedent. These cases are designated as retrospective anaphora, to
distinguish them from anticipatory anaphora, which has the relative order between the antecedent and the
anaphor reversed. While anaphor and anaphora are generally used to cover both cases, some use the terms
cataphora/cataphor for the anticipatory cases and endophora/endophor for the retrospective cases in partic-
ular. While the term antecedent implies its linear precedence relative to the anaphor, its generalization to
the anticipatory case is standardly accepted.

66
 u2 . 

 6
 "
s
3  o| c :
M t `


Kule-n taum meki-ka sohwatoy-l ttay-kkaci cam-ul
(SBJ) do-so-RelEnd next prey-Nom digest-RelEnd time-until sleep-Acc


.
ca-nta.
sleep-PresDec.
After that (they) sleep until the prey is digested.

In the Penn Korean Treebank, two syntactic types are allowed to act as an argument
of a verb: the noun phrase (NP) and the clause (S). It stands to reason, then, that Korean

pronouns, either overt or zero, can take either nominal or sentential elements as their an-
tecedents. The zero subject of the main clause in Example (87) below takes as its antecedent
the preceding clause, which can also be substantiated by the overt pronoun subject s
 /i/,
this. Example (88) illustrates that the empty pronoun subject in the boa constrictors
example above can be replaced by an overt pronoun.

(87)
h D # Q;s
2 [ t_
 r  s
4  

t 

e HX
 <,
yosay elini-tul-uy silyek-i nappaci-ko iss-nuntey,
recently child-Plu-Gen sight-Nom deteriorate-AuxEnd be-CoordEnd,

/s H
 
TV r
' A_


% s
 .
/i-nun olayn TV sicheng-uy yenghyang-i-ta.
(SBJ)/this-Top long TV watch-Gen influence-Cop-Dec.
Childrens eyesight is deteriorating these days, and (it)/this is the influence of long
TV watching.

(88) u2 . [ t
r

 6
 "
s
3  o| c :
M t `


ku-tul-un Kule-n taum meki-ka sohwatoy-l ttay-kkaci cam-ul
it-Plu-Top do-so-RelEnd next prey-Nom digest-RelEnd time-until sleep-Acc


.
ca-nta.
sleep-PresDec.
After that they sleep until the prey is digested.

67
2.5 Semantic Interpretation of Anaphoric Relations

In this section, we explore in detail the nature of semantic interpretations holding in anaphoric
relations. Recall the two definitions of anaphora/anaphor in the previous section. Levin-

son (1983) offers a more strict version of how an anaphor is interpreted: he defines an
anaphoric usage as one where one can pick out as referent the same entity that some prior
term in the discourse picked out. The other definition offered by Lexicon of Linguistics
is more vague: an anaphor is an element which depends for its reference on the reference

of another element. In many simple cases, Levinsons definition of coreference holds: the
boa constrictors example in (86) is one such case. In other cases, which are presented
below, the anaphoric expression is interpreted as referring to an entity that is semantically
closely related to the antecedent.
First, a prototypical example of coreference relation between an anaphor and its an-

tecedent:

(89) coreference between NPs

u1 . o
^
= 1  Qj
# ] 2 \

 
.
chelswu1 -ka ecey cha2 -lul sa-ess-ta.
chelswu1 -Nom yesterday car2 -Acc buy-Past-Dec.
Chelswu1 bought a car2 yesterday.

u2 . 1 Z
t

 _
 2
#?
HX
 <, 2 

1 onul na-hanthey 2 poyecwu-ess-nuntey, 2 acwu
(SBJ1 ) today I-To (OBJ2 ) show-Past-CoordEnd, (SBJ2 ) very


" &
 .
mecci-ess-ta.
nice-Past-Dec.
(He1 ) showed (it2 ) to me today, (it2 ) was very nice.

In the example above, there are two anaphoric relations: one between Chelswu and the first
zero pronoun subject of the second utterance (1 ); another one between car and the other
two zero pronoun occurances in the second utterance (marked as 2 ). In both cases, the

68
anaphors refer to the same entity that their respective antecedents do: the anaphors and the
antecedents are coreferential.
An anaphoric relation does not have to be that of strict coreference. One example of

such a semantic relation is that of kind-token relation. Consider:

(90) kind-token relation between anaphorically linked NPs

u1 .A
o  sH^
 u
\
 -
 z #
 QK.
wuli ai-nun kimchi-lul nemwu silheha-e.
our kid-Top kimchi-Acc too hate-Dec.
My kid hates Kimchi so much.

u2 .B Q
#$ a 
o| \
9 f "H



3
" ~X
 <.
ecekey wuli-cip-eyse-nun cal-man mek-tentey.
yesterday (SBJ) our-house-at-Top (OBJ) well-Only eat-RetroDec.
(He) was eating (some) happily in my house yesterday.

Presumably, the kimchi mentioned by A refers to the kind; the kimchi that was being hap-
pily consumed by As child in Bs house was some particular dish of kimchi that was on Bs

dinner table. They do not refer to the same entity very closely related semantically, but
not the exactly same one nevertheless. The antecedent denotes the generic kind, and the
anaphor type-shifts to refer to a specific incident that belongs to the kind. It is important
to note that neither the zero pronoun nor the antecedent noun, a bare noun phrase in form,
carry any information with regard to this type of semantic interpretation: it is the mood that

is essentially responsible for it. The utterance (u1 .A) has the irrealis/present tense, which
contributes to the individual-level predicate reading of the sentence, which then enables the
generic reading of the noun phrase kimchi, while the retrospective tense in (u2 .B) forces
the zero pronoun to have a specific reading.

The following example, on the other hand, illustrates a case of what is known as sloppy
anaphora.

69
 (91) sloppy anaphora

u1 .A Q
#j] 

  #
Q.
ecey na cha sa-ess-e.
yesterday I car buy-Past-Dec.
I bought a car yesterday.

u2 .B
&

? 



 HX
 <!
cengmal? na-to sa-ess-nuntey!
really? I-Also (OBJ) buy-Past-Excl!
Really? I bought (one) too!

The car referred to in As utterance is a specific car that A bought; the zero pronoun object
in Bs utterance refers to a car as well, albeit a different one purchased by B. That is, the

specific-indefinite reference of the antecedent noun is continued by the zero anaphor with
a specific-indefinite reference with a disjoint interpretation. Notably, an overt pronoun
cannot substitute for the zero anaphor, as shown below.
 /kuke/ it can only refer to
the very car that the antecedent refers to, which results in semantic incoherence. A similar

distinction is found between two English pronouns one and it.

(92) overt pronoun not permitted in sloppy anaphora

u2 .B
&

? 

*   
HX
 <!
cengmal? na-to *kwuke sa-ess-nuntey!
really? I-Also *it buy-Past-Excl!
Really? I bought *it too!

By contrast, an overt pronoun anaphor is acceptable when the antecedent noun refers to a
particular brand, while maintaining a disjoint reference:

(93) u1 .A Q
#j] 


D  r c 
#
Q.
ecey na honta sibik sa-ess-e.
yesterday I honda civic buy-Past-Dec.
I bought a Honda Civic yesterday.

70
 u2 .B
&

? 

 / 
HX
 <!
cengmal? na-to kwuke/ sa-ess-nuntey!
really? I-Also it/(OBJ) buy-Past-Excl!
Really? I bought it/(it) too!

As with overt pronouns and other referring expressions, Korean zero pronouns are ca-

pable of denoting a wide range of semantic types. When they denote entity types, which
is the predominant case, their antecedents are realized as a noun or a noun phrase, as the
examples so far illustrate. In addition, clausal units can act as antecedents. A clausal unit
offers two distinct semantic types of interpretation: the propositional content, as theorized

in standard predicate logic, and the event, as put forth in event semantics and situational
logic. Therefore, a zero pronoun with a clausal antecedent can refer to a propositional
content (94 u2 .B) or an event (94 u2 .B). 12

(94) u1 .A o
^
=  
t r
+>\
%
@
3 /.
chelswu-kkaci-to ku sihem-ey pwuth-ess-tay.
Chelswu-even-also that exam-At pass-Past-Reportive.
(I heard that) Even Chulswu passed that exam.

u2 .B
/ z
 s
  ?
kuke/ sasil-i-ya?
that/(SBJ) true-Cop-Q?
Is that/(that) true?

u2 .B
/ '


 s
{
9  .
kuke/ tahayngsulep-n il-i-ya.
that/(SBJ) lucky-RelEnd event-Cop-Dec.
That/(That) is lucky.

Exactly which semantic type the zero pronoun denotes is determined by the property of the
predicate: only a propositional content can be true or false (in 94 u2 .B), and only an event
or a happening can be said to be lucky or otherwise (in 94 u2 .B).
12
Zero pronouns that are linked to a prior clausal element are particularly susceptible to a wide range of
semantic interpretation, which will be explored in detail in the following section.

71
2.5.1 Discourse(Textual)-Deictic Zero Pronouns

Recall that the category of deictic zero pronoun reviewed in section 2.4.2 was classified

under text-independent use. Those deictic zero pronouns refer to entities that are present
in the scene of discourse, most commonly the discourse participants. The deictic zero pro-
nouns reviewed in this section are the other kind: they are text-dependent, in the sense that
the discourse or the text itself provides the source of referent. Stirling & Huddleston (2002)
provides the following examples of discourse deixis involving the English demonstrative

pronoun that:

(95) a. Take the bus to Murwillumbah. Shall I spell that for you?

b. A: It was a great luck. B: Is that another Australianism?

c. A: Is this a dagger that I see before me? B: That was too loud.

That in (95a) refers to Murwillumbah the word and not Murwillumbah the town. Similarly,
the zero subject in (96) below refers to the name ^


o /kimnwuli/ Nwuli Kim and not

the bearer of the name. Also, the zero subject in (97) refers to the preceding sentence as
the precise section of text:13 The preceding sentence is, then, not the antecedent but the
referent of the zero pronoun: this is a case of discourse deixis, a case of deixis where the
referent is not physically present in the situation of utterance but is located in the discourse
itself (Stirling & Huddleston, 2002).

(96) zero pronoun as discourse deixis

u1 .A j
] ^ 
o{m
9  .
ce-ka kim-nwuli-i-pnita.
I-Nom Kim-Nwuli-Cop-PolDec.
I am Kim Nwuli.
13
This is analogous to what Lyons (1977) calls pure textual deixis.

72
 u2 .B
/ 
 \ Vr
2
s s
1W .
kuke/ acwu yeyppu-n ilum-i-neyyo.
That/(SBJ) very pretty-RelEnd name-Cop-Dec.
That/(That) is a very pretty name.

(97) u1 . &
u
& 
h  )
 0 A H
 r
f
 1
 x
l



@\ ~



cengchicekin hewi somwun-un sahul dongan-man sinloy-lul pat-umyen
political false rumor-Top 3days while-Only credit-Acc get-If

& 
s
 a
) .
khetalan towun-i toy-nta.
big help-Nom become-PresDec.
False political rumors are very useful if they are believed for as short as 3
days.

u2 . 
s
r/
X
<_ 

" s
 ~ 
 B
n jnu_ 
ikes-un khwuteytha-uy myengin-i-ten ti meytichi-uy
this-Top/(SBJ) coup-detat-Gen master-Co-PastRelEnd Di Medici-Gen
s
.
mal-i-ta.
words-Cop-Dec.
That/(That) is Di Medicis words, a master of coup detat.

By contrast, the zero pronoun subject in the following example, used anaphorically,
refers to the event represented by the clausal antecedent in the preceding sentence:

(98) zero pronoun as an anaphor with a clausal antecedent

u1 . 3{ 7
9 x
r

\
f"H

{ t
  



3-il cungkwensicang-eyse-nun khosutak ciswu-ka cennal-pota
3-day stockmarket-At-Top Kosdaq index-Nom previous-day-Than


r
270.40` l
 2

.
nac-un 270.40-ul kilokha-ess-ta.
low-RelEnd 270.40-Acc record-Past-Dec.
In the stock market on the 3rd, the Kosdaq index recorded 270.40 which is
lower than the previous day.

u2 . {
`



 
{
9
 [ ts

pwutam-ul nukki-n ilpan thwucaca-tul-i
(SBJ) pressure-Acc feel-RelEnd general investor-Plu-Nom

73
 e
 z
&
 \
r
:

 f

" l

e  :
M Hs
 .
chaik-silhyen-ey ponkyek nase-ko iss-ki ttaymwun-i-ta.
profit-realization-At actively start-AuxEnd be-NomEnd cause-Cop-Dec.
(It) is because general investors are setting out for profit realization, who are

feeling pressure.

The use of a zero pronoun as discourse deixis is most clearly distinct from its anaphoric
usage when the reference is to a word or a phrase, as seen in (96). More subtle are cases like
the following, in which the zero pronoun subjects in speaker Bs slightly varied utterances

(u2 .B u2 .B) are all linked to the entire preceding utterance made by speaker A (u1 .A):

(99) (adapted from Webber (1991) example 5)

u1 .A o
^
=  
t r
+>\
%
@
3 /.
chelswu-kkaci-to ku sihem-ey pwuth-ess-tay.
Chelswu-even-also that exam-At pass-Past-Reportive.
(I heard that) Even Chulswu passed that exam.

u2 .B
/ z
 s
  ?
kuke/ sasil-i-ya?
that/(SBJ) true-Cop-Q?
Is that/(that) true?

u2 .B
/ '


 s
{
9  .
kuke/ tahayngsulep-n il-i-ya.
that/(SBJ) lucky-RelEnd event-Cop-Dec.
That/(That) is lucky.

u2 .B
/ f

st?
kuke/ kecismal-i-ci?
that/(SBJ) lie-Cop-Q?
That/(That) is a lie, isnt it?

u2 .B
/ 
 u _@  H
 _

d 
1W.
kuke/ machi uywoy-i-lanun sik-uy malthwu-i-ney.
that/(SBJ) (SBJ) as-if surprise-Co-RelEnd way-Gen speech-Cop-Dec.
That/(That) is a way of talking as if (it) is a surprise.

74
As seen earlier in example (94), the zero pronoun subjects in (99 u2 .B) and (99 u2 .B) are
anaphoric; their interpretations, as a propositional content and an event respectively, are
dependent on that of their antecedent, i.e., the utterance in (99 u1 .A). The zero pronoun

subjects in (99 u2 .B) and (99 u2 .B), however, are discourse-deictic: some linguistic
element provided by (99 u1 .A) acts as the referent of these zeros. More specifically, the
zero pronoun (99 u2 .B) refers to the speech act in (99 u1 .A); in (99 u2 .B), the zero
pronoun subject in the main clause refers to the linguistic description expressed by (99
u1 .A) (the zero subject in the embedded clause is anaphoric and refers to the event denoted
14
by (99 u1 .A).).
14
A note on terminology is in order here. All of these four cases in the original example in English
presented in Webber (1991) are labeled as instances of discourse deixis. This differing view on terminology
is in part due to the fact that the parallel cases in English involve demonstrative pronouns this and that,
whose function is rooted in its origin as deictic expressions. She did not want to take the position that
separate processes (i.e., anaphoric reference to events, propositions or other linguistic entities) are involved
in interpretation of the demonstrative pronouns; rather, she views them as a deictic phenomenon in nature
and argues for a unified account where appropriate referents are generated from what is pointed to (the
demonstratum) by a referring function, an approach she adopted from Nunberg (1979) and others.
In this work, I follow Stirling & Huddleston (2002) in assuming that a pronoun can derive its reference
from a previously mentioned linguistic expression which denotes a propositional content and/or an event, and
in this case their usage is anaphoric. On the other hand, when a pronoun refers to a portion of text or any
other linguistic aspect of the text such as the speech act, the pronoun is seen as directly referring to the part
of discourse and therefore is considered deictic in nature. In particular, they illustrate the distinction with the
following examples (p. 1461), which correspond to the Korean examples (99) and (99) presented above:

(1) a. A: Kim has been falsifying the accounts. B: Thats terrible. [anaphoric]

b. A: Kim has been falsifying the accounts. B: Thats a lie. [discourse-deictic]

75
2.6 Deictic and/or Anaphoric: the Fuzzy Distinction

In Levinsons (1983) definition, an anaphoric usage is where some term picks out as refer-
ent the same entity (or class of objects) that some prior term in the discourse picked out.

In Lyonss (1979) simple phrase, deixis is identification by pointing. Specifically, deictic

15
terms point out of the current discourse to facts concerning the context of the utterance.
Upon a closer look, then, these definitions are not mutually exclusive: an expression can
refer to an entity present in the discourse scene and have a coreferential antecedent at the

same time. This is especially true for first and second person pronouns, whose referents are
necessarily a part of the discourse scene. Even the third person pronouns can sometimes be
used deictically, when used to directly refer to someone other than the speaker or the hearer
present at the discourse scene. Therefore, distinguishing between the two distinct usages
of such zero pronouns may not always be clear-cut. Consider the subject zero pronoun in

the second sentence. It refers to an entity that is present in the discourse scene (i.e., the
speaker), therefore it is deictic; it takes the subject I in the preceding utterance (u1 ) as its
coreferential antecedent, which makes it anaphoric:

(100) u1 . 
H


7

>
>K  
s H
 |
 `





nan-un com ttokttokhay poi-nun salam-ul manna-l
I-Top (SBJ) somewhat smart look-RelEnd person-Acc meet-RelEnd
:
M ,





 m~ 
 /
? a
> j
]
ttay-mata, hangsang phwum-ko tani-ten nay kulim cey
time-each, (SBJ) always carry-AuxEnd go-RelEnd my drawing number
1\ 
 /? \
r
 + >K

H
4 

.
1ho-lul kkenay-e ku-lul sihumha-e poko-n ha-ess-ta.
1-Acc pull-out-AuxEnd he-Acc test-AuxEnd try-AuxEnd do-Past-Dec.
Whenever (I) met a smart-looking person, I tested him whipping out my draw-
15
Lexicon of Linguistics (http://www2.let.uu.nl/UiL-OTS/Lexicon/) defines deixis as the
phenomenon where elements in a language may have a reference which is dependent on the immediate
context of their utterance, for example, personal pronouns (I, you, he), demonstratives (this, that), special
expressions like here and there, temporal expressions like yesterday and now.

76
 ing No.1 which (I) always carried around.

u2 . 

& 
s K

4 

e H
 |



ku-ka cengmal ihaylyek iss-nun salam-i-nka
(SBJ) he-Nom really understanding exist-RelEnd person-Cop-If


z
3
 %
 
~ s

  .
al-ko siph-ess-ten kes-i-ta.
know-AuxEnd desire-Past-RelEnd that-Cop-Dec.
(I) wanted to know whether he really possessed understanding.

One can resort to certain kinds of precedence rules here: whether a zero pronoun is
deictic or anaphoric relies on presence or absence of a prior mention, i.e., an antecedent.
This criterion, unfortunately, results in a distinction that is quite accidental. It is simple
enough when the referent entity is never mentioned in the discourse: then all incidents of

the zero pronouns in the discourse referring to that entity will be branded as deictic. If
an explicit mention occurs somewhere in the discourse, however, they will all be deictic
up until and including the first mention, and then uniformly anaphoric from that point on.
This serves as a rather poor categorization system, which offers very little insight to link

the motivation behind the use of zero form and its linguistic surroundings: zero forms
occurring in the same linguistic context can receive two different types, simply from their
different relative location to a mention of their referent, which might well have occurred in
a non-local manner. There is also a further concern in this approach, namely that of how
to treat an antecedent that is a zero form itself. The definition of prior mention becomes

dubious when it is stretched to include a null form.

In an alternative approach, the first person and the second person pronouns, zero or
explicit, are classified as deictic in all environments, since their interpretation in theory
can be established without depending on the reference of an antecedent. However, this

approach suffers the same arbitrariness that plagued the first option in that it ignores the
linguistic context in which pronouns occur. It also goes against the typical pattern of chain
of co-reference: Levinson (1983) notes that deictics or other definite referring expressions

77
are often used to introduce a referent, and anaphoric pronouns are used to refer to the same
entity thereafter. This is necessarily true in the case of an expression that refers to an object
in the scene: the first mention is in the form of indexicals such as this and that; from the

second mention and on, an anaphoric pronoun it is used.

The third option is a fundamentalist approach: a zero pronoun is anaphoric if its in-
terpretation necessarily depends on the interpretation of an antecedent; it is deictic if the
use of null form was motivated by a reason not directly related to the structure of the text,
for example, the referents prominence in the physical setting of the discourse scene. This

approach forces us to look beyond the apparent linguistic environment of zero pronouns,
investigating the underlying principles behind the use of null forms. There are two issues
with this approach. First, for most cases the distinction is not at all transparent. The sec-
ond problem is that this approach makes it necessary for us to assume too much about the

theoretical framework the one which presumably offers a comprehensive theory of zero
pronouns. This, we believe, is not desirable at the annotation stage and carries the risk of
biasing the process.
There is a simple solution to this conundrum, which is to allow an expression to be

anaphoric and deictic at the same time. Indeed, Lyons (1977) points out, and Levinson
(1983) agrees, that it is perfectly possible for a deictic term to be used both anaphorically
and deictically. Lyons (1977) gives the following example to illustrate the point:

(101) I was born in London and have lived there ever since.

In this sentence, there refers back to whatever place London refers to, but simultaneously

contrasts with here on the deictic dimension of space, locating the geographical point of
utterance outside London. Stirling and Huddleston (2002) also notes that a term may be si-
multaneously deictic and anaphoric. They demonstrate the close relation between anaphora
and deixis with the following example:

78
(102) a. Whats that hes got in his hand? (deictic)

b. He wants $30, but thats too much. (anaphoric)

In (102a) that will be interpreted deictically as referring to some object present in the
situation of utterance and relatively distant from the speaker, whereas in (102b) it obtains
its interpretation anaphorically, from the antecedent $30. They note that the two uses of
that are clearly related, and it is plausible to regard the anaphoric use as derivative from the
deictic and in fact it retains some residual deictic meaning.

In light of these points, I adopt the point of view that the deictic and anaphoric usages
of pronouns are closely related, hence an expression can be both anaphoric and deictic.
Therefore, the zero pronouns in the example (100) above are treated as being simultane-
ously deictic and anaphoric.

79
Chapter 3

The Centering Theory and Korean Zero

Anaphora

One of the fundamental assumptions underlying most studies on Korean zero anaphora
within the Centering Theory framework has been that Korean zero pronouns are the stand-
ins for the overt pronouns of the Indo-European languages in the theory. It follows, then,

that zero pronouns in Korean coincide with the center of the utterance in which they ap-
pear, just like their overt counterparts do by way of Rule 1 (28). Given that zero pronouns
constitute Cbs, their interpretation is tied to the transition type preference dictated by Rule
2: since CONTINUE, which entails that a zero pronoun Cb is coreferential with the Cb of

preceding utterance, is preferred over SHIFT, which entails the opposite, it is predicted that
a zero pronoun refers to the previous Cb. Under this set of assumptions, the exact cali-
bration of the Center Hierarchy became the focus of research. Discovery of the correct
Cf-ranking hierarchy ensures successful prediction of coreference patterns for zero pro-
nouns; in turn, observed coreference patterns of zero pronouns are used as evidence for

establishing particulars of Cf-ranking hierarchies.

As noted earlier, Grosz, Joshi and Weinstein (1995), while relying solely on grammati-

80
cal roles for their original formulation of Cf-ranking for English, suggest that the items in
the list can be ranked by other factors as well, including text position and lexical semantics.
It turns out for many languages that salience, which the Cf-ranking hierarchy is assumed to

encode, can be manifested by various aspects of language use. There have been a number
of efforts to incorporate more diverse notions into Cf-ranking; a few lead to more direct
approaches which aim to formulate Cf-ranking in more information-theory-oriented terms.
In the next few sections, we review these efforts and their implications on the role of center
hierarchy and zero pronoun anaphora.

3.1 Criteria for Cf Ranking: What Encodes Salience?

As noted previously in Section 1.2, Grosz, Joshi and Weinstein (1983, 1995) based their
formulation of the FORWARD - LOOKING CENTER (Cf) hierarchy for English on the notion
of grammatical roles, while citing other factors such as text position on lexical semantics as

potentially relevant. In much of the research that followed, additional factors governing the
ranking have been explored for various languages. For German, Rambow (1993) considers
word order as the defining factor in determining Cf ranking. Gordon, Grosz and Gilliom
(1993) claim that word order, as well as grammatical role, plays a role in Cf ranking in

English. Thematic role has been suggested as the Cf ranking criterion for languages such
as Turkish (Turan, 1995, 1998). Pragmatic concepts such as empathy and topic have been
placed high in the Cf ranking (Example 29) for Japanese (Kameyama, 1985, 1986; Walker,
Iida and Cote, 1990, 1994).

Notably, the linguistic form of the referring expressions within the Cf set gained support
from some researchers as a crucial factor over the years. Kameyama (1996, 1998) takes this
position. In addition to the original hierarchy based on grammatical role, she introduces an
additional hierarchy based on nominal expression types in English. The former is dubbed

81
the grammatical function hierarchy (GF ORDER) and the latter the nominal expression type
hierarchy (EXP ORDER).

(103) GF ORDER:

Given a hierarchy [Subject > Object > Object2 > Others], an entity realized by a
higher-ranked phrase is normally more salient in the output attentional state.

(104) EXP ORDER:

Given a hierarchy [Zero Pronominal > Pronoun > Definite NP > Indefinite NP],

an entity realized by a higher-ranked expression type is normally more salient in

the input attentional state.

She views the attentional state as having two stages of input and output, at time points
before and after the current utterance. The two orders are to make predictions for each one:

GF ORDER for the output attentional state, and EXP ORDER for the input attentional state.
Crucially, a single label of Center replaces Cp and Cb, which is determined by the EXP

ORDER and not the GF ORDER. Another crucial point is that it is the EXP ORDER of ui .,
the current utterance, which determines the Center in the output state of ui .: in the standard
Centering Theory, it is the grammatical-role-based Cf ranking of the previous utterance

ui1 which determines Cb(ui ). She stipulates it in the following defeasible preference:

(105) EXP CENTER:

An expression of the highest-ranked type in exp order normally realizes the Center
in the output attentional state.

The standard Centering Theory, in fact, does incorporate the formal feature of refer-
ring expressions in its theory: it is manifested in the form of Rule 1, which states that a
pronominal form, when found, is the Cb in a sentence. It has been observed that Rule 1
and Constraint 3 run into conflicts. Kameyama gives the following example:

82
(106) u1 . John went to Jims party.

u2 . He was very pleased to see John again.

u3 . He had just recovered from a stressful week at work.

The He in (u2 ) obviously refers to Jim; the interpretation of the He in (u3 ) is of interest
here, for which the choice of Cb of (u2 ) becomes relevant. The standard approach paints
two conflicting pictures: Constraint 3 (repeated below1), coupled with the grammatical-

role-based Cf ranking, yields John as Cb(u2 ), while Rule 1, focusing on the formal prop-
erty of the nominal expressions, yields Jim, the pronominal subject of (u2 ), as Cb(u2 ).
Kameyamas formulation, on the other hand, singularly picks Jim as the C(u2 ) according
to the EXP CENTER preference. Given that Jim is the more natural referent of the pronoun

subject in (u3 ) and therefore is the Center of (u3 ), a theory that selects Jim as Cb(u2 ) is
more successful since it renders the transition type between the two utterances CONTINUE,
while selecting John as Cb(u2 ) will result in the less preferred transition type of SMOOTH -
SHIFT. Center selection based on Rule 1, then, is more successful, and so is Kameyamas
which replaces the conflicting approaches of Rule 1 and Constraint 3 of the standard cen-

tering theory with a uniformly form-based ranking of EXP ORDER in the process of Center
selection.
The interplay between the two orders can be summarized as follows. Since the output
1

(1) Constraint 3:
For each utterance Ui in a discourse segment U1 , ... , Un :
The center, Cb(Ui ), is the highest-ranked element of Cf(Ui1 ) that is realized in Ui .

(2) Rule 1:
For each Ui in a discourse segment Ui , ... , Um :
If some element of Cf(Ui1 ) is realized as a pronoun in Ui , then so is Cb(Ui ).
Rule 2:
Transition states are ordered. CONTINUE is preferred to RETAIN is preferred to SMOOTH - SHIFT is
preferred to ROUGH - SHIFT.

83
attentional state at the end of utterance Ui constitutes the input attentional state of the next
utterance Ui+1 , the combination of GF ORDER and EXP ORDER and their stipulated relation
with the respective attentional state achieves the following effect. The discourse entities

in Ui+1 inherit the salience ranking from the output attentional state of Ui , which follows
from the grammatical roles that they occupied in Ui , and are realized in linguistic forms
that are appropriate for the salience ranking; however, they are free to occupy different
grammatical roles in Ui+1 this time, which in turn shifts their salience ranking in the output
attentional state of Ui+1 . This is how the flow of salience ranking among discourse entities

across utterances is modeled in her theory. What the EXP CENTER preference achieves is,
then, to guarantee a pronominal expression, if any, the Center status, so it is preferred as
the antecedent of subsequent pronoun occurrences regardless of its grammatical role.
The givenness hierarchy (Gundel, Hedberg and Zacharski, 1993; Gundel, 1998) is an-

other well-known theoretical construct that correlates the salience scale of referring expres-
sions to their linguistic form. An attempt to integrate Centering Theory and the givenness
hierarchy is reported in Gundel (1998). She argues for a Cf-ranking of English based on
the givenness hierarchy, formulated as below:

(107) The Givenness Hierarchy

in focus > activated > familiar > uniquely identifiable
it that that NP the N

> referential > type-identifiable

indefinite, this N aN

Another closely-related linguistic hierarchy is found in Princes (1981) familiarity scale,

shown in (108) below, which predicts the relative degrees of accessibility of referents.
Strube (1998, also reviewed later in Section 1.2.2, Part II) derives from this hierarchy in
his effort to adapt Centering Theory in anaphora resolution algorithm. Strubes model re-
places Cf-list and Cb with what he calls the S-list (salience list), which describes the

84
attentional state of the hearer at any given point in processing a discourse. Unlike Cf-
list in the original Centering Theory which relied on grammatical role as the criteria for
its ordering, the S-list borrows notions from theories of information structure, such as the

hearer-old/hearer-new status of discourse entities from Princes (1981) familiarity scale and
information status in Prince (1992).

(108) Princes (1981) Familiarity Scale

Assumed Familiarity

New Inferrable Evoked

(Non- Containing (Textually) Situationally

Brand-new Unused
containing) Inferrable Evoked Evoked
Inferrable
Brand-new Brand-new
(Unanchored) Anchored

The elements of the S-list belong to three information sets: hearer-old discourse enti-
ties (OLD), meditated discourse entities (MED) and hearer-new discourse entities (NEW).
The three sets of expressions have their own subdivisions. OLD consists of evoked (E)
and unused entities; MED consists of inferrables (I), containing inferrables (Ic ), and an-
chored brand-new discourse intrasentential (BNa entities); NEW consists solely of brand-

new (BN) entities. The S-list is ordered by the information status of the above sets. OLD

entities are preferred to MED entities which are preferred over NEW entities. Within each
set, the ordering is by utterance and position in utterance.
In a way, the transition of the discussions surrounding what determines the Cf-ranking

signals a trend that gravitates towards favoring linguistic notions that are more pragmatic in
essence. The inadequacy of the grammatical role as the precise measure for the degree of
salience of discourse entities was apparent from the very early stage of the theory. After all,

85
salience is a highly context-oriented theoretical concept, and grammatical role is foremost
a syntactic category. The move towards favoring the form factor of referring expressions
over grammatical role, as seen in Kameyama (1996, 1998) and others, therefore, represents

this trend.
Strubes (1998) S-list theory represents by far the most extreme approach that attempts
to encode salience in more direct terms. In organizing his S-list, he more or less directly
incorporates the taxonomy from Princes (1992) familiarity scale, thereby including such
distinctions as hearer-new discourse entities and anchored brand-new. The problem for

his approach is, then, that the categories in the ranking themselves need a set of criteria to
determine whether a particular referring expression qualifies for the category. For example,
how does one determine if an NP in a discourse refers to a hearer-old entity or a hearer-
new entity? Surely the answer should be based on various factors that are linguistically

and contextually realized. By formulating his rankings in more cognitively oriented terms,
Strube (1998) abstracts away from the simplistic limitations of previous formulations, but
the burden of precisely capturing how the degree of salience of an expression can be de-
termined is in a way relegated to an auxiliary procedure which still has to deal with more

surface-oriented linguistic features.

3.2 Establishing Cf Ranking for Korean

With the discussion in mind, let us now turn to the topic of Cf-ranking for Korean and
Korean anaphora. Much of the discussion in previous work on Korean zero anaphora within

the Centering Theory framework focuses on calibrating the Center hierarchy suitable for
the language. The most common position held by many of these studies is the highest
ranking of topic (No, 1991; Jang, 1986, 1994; Kim, 1994; Ryu, 2001), not unlike the Cf
ranking proposed for Japanese by Walker, Iida and Cote (1994) and earlier by Kameyama

86
(1985, 1986). In these studies, a topic-marked element in an utterance outranks any other
grammatical argument in an utterance, including the subject.

(109) Cf ranking for Japanese (Walker, Iida and Cote, 1994)

(grammatical or zero) topic > empathy > subject > object2 > object > other(s)

Ryu (2001) and No (1991) both add topic-marked NPs at the top of what is essentially
a Cf hierarchy analogous to the original grammatical-role-based Cf ranking proposed for
English. Ryus (2001) formulation:

(110) Center hierarchy for Korean by Ryu (2001)

r,

topic ( H /un,nun/) > subject > object2 > object > others


No (1991) departs slightly from the common approach, by stipulating that the high ranking
2
subject and object NPs must be bare NPs without a case marker:

(111) Center hierarchy for Korean by No (1991)

topic-marked element > case-marker-less subject > case-marker-less object > oth-
ers

Note that Ryus (2001) ranking, while it might look at first glance like a grammatical-
role-based criterion topped with a heterogeneous discourse-level notion of topic, is in fact

formed around a single coherent linguistic feature: since grammatical roles of NPs and
2
The hierarchy based on Nos (1991)s original formulation is given below.

(1) Cf Ranking algorithm in Korean

a. the entity realized by the subject NP if that NP is not nominative-marked

b. the entity realized by the object NP if there is one and if that NP is not accusative-marked

c. an element of Cf(Un1 ) that is realized in Un

87
topic-hood of NPs are all marked by post-position suffixes in Korean, his ranking relies on
the single linguistic aspect of morphological marking. Nos (1991), on the other hand, indi-
cates a shift towards including formal considerations in Cf ranking: dropped case markers

is one of the many options for nominal expression forms available for speakers of Korean
to choose from.
There are, however, a few studies which disagree on the role of topic in the hierarchy.
Hong (2002) argues that the relative ranking between topic and subject cannot be deter-
mined for Korean, a position shared by Cha et al. (1997) and Choi & Lee (1999). Hong

(2002) offers two sets of dialogue, each of which he argues supports one criterion outrank-
ing the other; therefore, he concludes, the ranking between the two cannot be determined.
The first of the two dialogue segments shown below purportedly illustrates topic outranking
subject in Cf ranking:

(112) un1 . o
^
= H




 s% 
i .
chelswu-nun hwaltalha-n ai-i-ess-ta.
Chelswu-Top outgoing-RelEnd child-Cop-Past-Dec.
Chelswu was an outgoing kid.
Cf: [Chelswu], Cb: ?

un . o
^
= H
 \

f" 
  j{
] a
9 ~
% 
.
chelswu-nun pan-eyse minswu-ka ceyil coh-ess-ta.
Chelswu-Top class-in Minswu-Nom most like-Past-Dec.
Chelswu liked Minswu the most in the class.

topic > subject Cf: [Chelswu > Minswu], Cb: Chelswu CONTINUE

un+1 . 
6


  ~
%
a 
.
ku taum-ulo cinswu-ka coh-ess-ta.
(SBJ) that next-at Cinswu-Nom like-Past-Dec.
(He) liked Cinswu next.
Cf: [pro=Chelswu, Cinswu], Cb: pro=Chelswu CONTINUE

88
 un+2 . 
r\ f" H



% B j{
] a
9 ~
% 
.
kyohoy-eyse-nun yenghuy-ka ceyil coh-ess-ta.
(SBJ) church-at-Top Yenghuy-Nom most like-Past-Dec.
In church, (he) liked Yenghuy the best.

Cf:[pro=Chelswu, Yenghuy], Cb: pro=Chelswu

Hong (2002) takes Chelswu to be the topic and Minswu the subject in (un ). The zero
pronoun in (un+1 ) is most naturally interpreted as Chelswu. Under the topic outranks
subject assumption, Chelswu is the Cp in (un ); and since Cb(un+1 ) is Chelswu as well,

it makes the transition type of the utterance CONTINUE, the most favored one of the four
transition types. This supports the topic > subject ranking, since it leads to the most
natural and coherent discourse transition type.
But that result is contradicted in the following slightly varied example, he says:

(113) un1 . o
^
= H




 s% 
i .
chelswu-nun hwaltalha-n ai-i-ess-ta.
Chelswu-Top outgoing-RelEnd child-Cop-Past-Dec.
Chelswu was an outgoing kid.
Cf: [Chelswu], Cb:?

un . o
^
= H
 \

f" 
  j{
] a
9 ~
% 
.
chelswu-nun pan-eyse minswu-ka ceyil coh-ess-ta.
Chelswu-Top class-in Minswu-Nom most like-Past-Dec.
Chelswu liked Minswu the most in the class.

topic > subject Cf: [Chelswu, Minswu], Cb: Chelswu CONTINUE

subject > topic Cf: [Minswu, Chelswu], Cb: Chelswu RETAIN

un+1 . 6
 }


.
maumssi-ka chakha-ess-ta.
(SBJ) heart-Nom kind-Past-Dec.
(He) had a kind heart.
topic > subject Cb: [pro=Minswu], Cb: pro=Minswu ROUGH - SHIFT

subject > topic Cb: [pro=Minswu], Cb: pro=Minswu SMOOTH - SHIFT

89
 un+2 . xs

7 ]l
X  
.
congicepki-to calha-ess-ta.
(SBJ) origami-Also do-well-Past-Dec.
(He) was good at origami too.
Cf: [pro=Minswu], Cb: pro=Minswu CONTINUE

Here, presumably, the two zero pronoun subjects in (un+1 ) and (un+2 ) are most naturally

interpreted as referring to Minswu. Because of that, its the subject > topic arrangement
that yields the better discourse transition types this time: it leads to RETAIN in (un+1 ) and
SMOOTH - SHIFT in (un+2 ), as opposed to CONTINUE in (un+1 ) and ROUGH - SHIFT in (un+2 )
under the opposite ranking. Since the two examples support conflicting relative ranking of

the two criteria, Hong (2002) claims, this points to the conclusion that the ranking between
topic and subject cannot be determined for Korean.
There are several flaws in this argument. First of all, it is debatable whether Minswu
in (un ) is the true subject of the sentence. Hong (2002) himself briefly notes the point

in his footnote, but dismisses the concern by declaring it as not crucial: it is in fact a
glaring problem which demands closer examination. It seems that in his efforts to pick a
grammatical construction which has both topic and subject appear in the matrix clause, he
ended up choosing one involving a psych-verb predicate that is open to varied linguistic
analyses. The predicate argument structure in (un ) can be schematized as (114a) below:

(114) a. A B ~
%
a .
A-ka B-ka coh-ta.
A-Nom B-Nom good-Dec.
literally: A is B is good, meaning: A likes B

b. B ~
%
a .
B-ka coh-ta.
B-Nom good-Dec.
B is good.

90
 c. A a

B\ ~
% 
 .
A-ka B-lul cohaha-nta.
A-Nom B-Acc likes-PresDec.
A likes B.

Note that (114a) has two subject NPs, i.e., NPs with the nominative postposition marker 
/ka/. The first NP is expressed with the topic marker
H /nun/ in Hongs examples, which

replaces the nominative and accusative case markers. Under a standard linguistic analysis,
the first NP A-Nom, which I will refer to as NP1 , is said to be the subject of the entire

sentence while the second NP B-Nom, NP2 , is said to be the subject of what is referred
to as the predicate clause, which is formed by the B-Nom good-Dec phrase. Note that
NP2 , when appearing as the sole argument of a sentence in (114b), can function as the
matrix subject of the adjectival predicate a
~ (/coh/, is good). There is an opposing point
%

of view, however, which disfavors the notion of predicate clause and views NP2 not as a
subject but rather as an internal argument of the adjective. In this view, the construction
in (114a) receives an analysis that is identical to the one given to a related construction
shown in (114c), which involves the verb a
~
% 
(/cohaha/, like) which is derived from
the adjective a
~ (/coh/, is good) by a verbalization suffix. In this construction, NP2 is
%

realized with the accusative case marker \

/lul/. The systematic relation between the

two constructions is what the class of psych-verbs or psych-predicates has in common,
which includes predicates that attribute emotional qualities such as a
~ (/coh/, is good),
%
V
\ (/yeyppu/, is pretty),
[O
 (/mwusep/, is scary). In either analysis, NP2 is not the

subject of the sentence: in the former predicate clause account, it is merely the subject of
the predicate clause; in the latter, it is a VP-internal argument of the predicate.
Secondly, the two elements do not appear to be competing on equal grounds. There
are other semantic and discourse-oriented inferences at play here which contribute to the

91
allegedly favored interpretations of the zero pronouns. In (112), the structural parallelism3
displayed in (un , un+1 , un+2 ) influences heavily towards the preferred coreference interpre-
tation of all three subjects including the two zero subjects in (un+1 ) and (un+2 ). There is

a lexical cue present as well: the adverbial that next in (112 un+1 ) can only reasonably
mean next to Minswu, therefore it provides a strong cue that the zero pronoun subject is
not Minswu but Chelswu. These cues make it semantically and pragmatically incoherent
to interpret the zero pronouns as referring to Minswu: Chelswus being favored as the ref-
erent of zero pronouns therefore cannot be seen as soley the result of the topic > subject

Cf ranking.
In (113), on the other hand, there is another discourse-structure-oriented inference at
work. There is a strong tendency to interpret the utterance (un+1 ) as serving giving evi-
dence role to the preceding segment (un ). In this interpretation, (un ) describes how Chel-

swu liked Minswu, (un+1 ) elaborates next on the point and provides a reason why: the
subject of (un+1 ) is Minsu in this case.
It is, indeed, not a trivial task to construct examples where two different interpretations
compete on equal grounds except for the exact feature to be tested. The reason why Hong

(2002) had to use the double-nominative/psych verb construction is that he needed to be

able to have both topic and subject appear as matrix level arguments. Why cant one, then,
simply test the ranking between topic and subject with a sentence where an object is made
into the topic with the topic marker, as in the following example?

(115) o
^
=  Qj
# ] 

H



z.
chelswu-ka ecey minswu-nun manna-ess-ta.
Chelswu-Nom yesterday Minswu-Top meet-Past-Dec.
Chelswu met Minswu yesterday.
3
The point is also acknowledged by Hong (2002) in a footnote.

92
 The first problem is that now Minswu has arguably as two roles: while it is topic-
marked, it semantically retains its object role even though the accusative marker is sup-
pressed by the topic marker and is no longer visible on the surface. This problem can

be overcome by simply assuming that the role of topic is more prominent and therefore
overrides its object role. The larger problem, however, is that the topic-marked Minswu
is not a pure form of topic: it is what has been labeled (Lee, 2000, 2003, among many
others) contrastive topic, which involves invocation of an alternative set. By use of the
topic marker, the example above strongly implies that there is someone else or others that

Chelswu did not meet. The distribution of contrastive topic is largely linked to the topic
marker attached to an expression in a non-sentence-initial position. Then, why not con-
struct an example where an object not only bears the topic marker but also is fronted, as
below? This way, the fronted topic-marked Minswu now acts as a pure topic.

(116) 

 H
 =
o
^  Qj
# ]


z.
minswu-nun chelswu-ka ecey manna-ess-ta.
Minswu-Top Chelswu-Nom yesterday meet-Past-Dec.
Minswu Chelswu met yesterday. (== Chelswu met Minswu yesterday.)

This, however, introduces another source of disparity: Chelswu is the subject, but Min-

swu has undergone two processes: topic-marking AND fronting. When we let the two NPs
compete for their capacity for functioning as a readily-accessible antecedent, we cannot be
sure if we are weighing between a subject and a topic-marker-bearing element or a sub-
ject and a fronted element. Simply put, it is an extremely daunting task, if not downright

impossible, to create an example where topic and subject compete on equal grounds.
Another criterion that has often been featured prominently in Cf-ranking for Korean is
that of discourse deixis: the speaker and the hearer. Choi and Lee (1999) lists speaker,
hearer at the top of the hierarchy in place of topic. They claim that the pragmatic notion

of speaker and hearer is relevant in determining the ranking in addition to the others that

93
are based on grammatical roles, as seen in their formulation of Cf ranking in (117) below.
Lee and Lee (2000) join them in listing speaker and hearer over the grammatical roles,
under the SL-component placed on top (see Section 1.2.3 for a detailed discussion of SL-

component):

(117) Center hierarchy for Korean by Choi and Lee (1999)

speaker, hearer > subject > indirect object > direct object > others

(118) Cf ranking for Korean by Lee and Lee (2000)

SL-component > {speaker, hearer} > subject > indirect object > direct object >
others

Presumably, the motivation behind this move is that the speaker and the hearer, even
when not overtly expressed as an argument of a sentence, seem to be favored over any argu-
ment NPs as the antecedents of subsequently occurring zero pronouns. Consider the matrix

zero pronoun in utterance (u2 ): the speaker of the utterances is favored over Chelswu, the
subject argument of (u1 ), even though it was not expressed at all as an argument in (u1 ).

(119) u1 . o
^
=  Qj
# ] +
r >\
f"
h&`



.
chelswu-ka ecey sihem-eyse paykcem-ul mac-ess-ta.
Chelswu-Nom yesterday exam-At 100score-Acc receive-Past-Dec.
Chelswu scored 100% in the exam yesterday.

u2 . 
 _ @ _ s
{
9   ty
q


.
acwu uywoy-uy il-i-lako sayngkakha-ess-ta.
(SBJ) (SBJ) very surprise-Gen event-Cop-That think-Past-Dec.
(I/He? ) thought (it) was very surprising.

This approach is countered in Hong (2000) and Kim (2003). Hong (2000) expresses
doubts on comparability of grammatical notions as subject/object and pragmatic notions as

speaker/hearer, and argues for removal of these criteria and generic zero pronouns from the
study of Korean zero pronouns within the Centering Theory framework. Kim (2003) uses
the following illustration in her attempt to support this position:

94
(120) u1 . 
t j H
 O
ZY
 Us .
cijoj -nun chaykpelley-i-ta.
Cijoj -Top bookworm-Cop-Dec.
Cijo is a bookworm. Cf: Ciho

u2 . Q
#j] I j f&
" \
h f "


zHX
 <,
ecey-to I j secem-eyse manna-ess-nuntey,
yesterday-also (SBJI ) (SBJj ) bookstore-At meet-Past-And,
(II ) met (himj ) at the bookstore yesterday,

speaker included in Cf: Cf: [I, Ciho], Cb: Ciho

speaker excluded in Cf: Cf: [Ciho], Cb: Ciho

u3 . I/j



{
9
e
3 %
 .
I/j chayk-man ilk-ko iss-ess-ta.
(SBJI/j ) book-only read-AuxEnd is-Past-Dec.
(*I/he) was just reading books.
speaker included in Cf: Cf: [I], Cb: I SMOOTH SHIFT

speaker excluded in Cf: Cf: [Ciho], Cb: Ciho CONTINUE

She contends that the favored coreference interpretation, i.e., the one where the zero pro-
noun subject of (u3 ) refers to Ciho and not the speaker, results in the favored CONTINUE

transition type only when the speaker is forcibly removed from the Cf-list. Hence, she
concludes, the speaker of the utterance, while assuming an argument role in (u2 ), must

be excluded from the Cf-list. While I find the line of reasoning considerably weakened
by the same flaws I pointed out earlier for the examples by Hong (2002)4 , the pattern of
zero-pronoun coreference displayed in the example above, coupled with the contradictory
pattern shown in the earlier example (119), should be taken as evidence demonstrating that
4
Preference for Ciho as the referent of the zero pronoun in (u3 ) can be explained by the semantic cue
provided earlier by the statement that Ciho is a bookworm: Ciho is the bookworm, therefore it makes sense
that he is the one who is reading the book. In addition, whether or not SMOOTH SHIFT constitutes a signif-
icant enough offense compared to CONTINUE as to warrant removal of a NP argument (although not overtly
expressed) from a Cf-list is a debatable matter. SMOOTH SHIFTS are, after all, known to occur.

95
the roles that deictic personal pronouns occupy in coreference are tangential to those of
other anaphoric pronouns. To say the least, discourse deictic pronouns seem to be rather
free in presenting themselves out of nowhere as the referent of a zero pronoun (119) or

excusing themselves from the potential candidate list (120). Chang (1994) also asserts that
dropping of first-person pronouns and zero pronoun anaphora are two distinct phenomena.
The two discussions laid out so far illustrate the difficulty and inherent danger in at-
tempting to incorporate linguistic features which are diverse in nature into a one-dimensional
ranking. The examination of Hongs (2000) examples shows that topic and grammatical

roles such as subject and object are not in fact commensurable; the above discussion of
the place of discourse participants speaker and hearer in Cf-ranking proves the same point.
This leads to the conclusion that any attempt to formulate the Cf-ranking with elements
that are not exclusively encoded along a single linguistic dimension immediately incurs

this type of risks.5

Kims (2003) proposal for Cf-ranking in Korean sets itself from others by imposing
structured layers on the ranking:
5
The Cf ranking for Japanese proposed by Walker, Iida and Cote (1994) suffers in this regard: their iden-
tification of topic relies on two distinct modes, i.e., nullness of the pronominal form and the morphological
topic marking; empathy is a property largely expressed by the semantic features of predicates; the rest of the
elements in the Cf ranking are made up of grammatical roles, i.e., subject, object, etc. Notably, all Cf-rankings
for Korean reviewed here rely solely on the morphological marking of topic postposition as the identifier of
topic, thereby eschewing Walker, Iida and Cotes (1994) problem of dual criteria. By equating topic with the
morphological topic marker, these Cf-rankings can be viewed as based on a single observable feature, that
is, the postposition morpheme, as the grammatical roles such as subject and object are largely marked by
case postposition markers as well. However, grammatical roles and topic-hood are not truly commensurable
concepts, as illustrated by the detailed examination on Hongs (2002) take on topic.

96
(121) Cf ranking for Korean by Kim (2003)

a. Discourse-initial utterance
H /nun/) > other given information > new

topic (realized with topic marker

information

b. Second utterance and on

given information (zero > pronoun > definite expression > repeated noun)

> new information (subject > object > locative > oblique)

Hers is a blend of information-theoretic concepts and grammatical features. At the top-

level, she partitions the ranking into two groups: given information ranked over new
information. The given information group and the new information group are then inter-

nally ranked with two distinct criteria: the former with the formal properties of the referring
expression at hand, and the latter based on grammatical roles. The decision to use formal
properties of referring expressions, such as overtness/covertness, pronominal form and def-
inite expression, as the primary ranking criterion within the given information group is
influenced by Gundels work (1998) and Kameyamas EXP ORDER (1996, 1998). Note

that these works do not explicitly refer to the bi-partite grouping of given-new informa-
tion: their scales, instead, constitute a single continuum which ranges from most given to
least given. The rankings use of the given/new information dichotomy resembles Princes
(1992) hearer-old/hearer-new distinction. For entities that are newly introduced to the dis-

course, their formal properties are irrelevant, and it is the grammatical role that they assume
that is used as the ranking criterion.
H /nun/ is given the topmost ranking

Additionally, the topic postposition marker
only when it is found in the discourse-initial utterance. She contends that: (1) only given

information can be marked with the topic marker, and (2) at the beginning of a discourse,
elements with the topic postposition marker realize the topic of the sentence, which has

97
the highest degree of salience. From this, she deduces the highest placement of the topic-
marked element in the ranking for discourse-initial utterances. For non-discourse-initial
utterances, it is the zero pronoun that occupies the top slot of the ranking. In the literature,

the zero pronoun has been treated more than once as equivalent to the topic-marked NP:
recall that Walker, Iida and Cote (1994) famously equated it with the topic-marked NP, call-
ing it zero-topic. Kims (2003) formulation suggests that the topic-marked NP and the zero
pronoun play more or less the same role as marking the highest degree of saliency, albeit
in complementary distribution depending on the position of the utterance in discourse. 6 In

fact, she stops short of labeling zero pronouns as zero-topic: she merely calls them center
expression.
She argues that the zero pronoun is strongly correlated with the Center of an utterance.
With the zero form seen as the highest ranking form in the Cf-ranking, she echoes the po-

sition of Kameyama (1996, 1998), who argues for adopting formal properties of referring
expressions as the deciding factor in Center selection. More specifically, a referring ex-
pression realized in the form of a zero pronoun in an utterance is favored as the Center
over another expression that occupied a higher-ranked grammatical role in the preceding

utterance. To illustrate this, she suggests the following example7, which is analogous to an
earlier one given for English by Kameyama (in 106):

(122) u1 .
%

i
H
 Qj
# ] 
t ^=\
o f " &

Bj \




z.
yengswui-nun ecey cihachel-eyse cenghij -lul manna-ess-ta.
Yengswui-Top yesterday subway-at Cenghij -Acc meet-Past-Dec.
Yengswui met Cenghij yesterday in the subway.
6
She does not address the issue of topic markers used in non-discourse-initial utterances. She also effec-
tively excludes the issue of placement of zero pronouns in discourse-initial utterances suggesting that only
given information can be addressed by zero pronouns, and that the referent of zero pronouns in a discourse-
initial position is supplied situationally, in other words, their use is non-linguistically governed.
7
The example was slightly altered in order to remove some empty pronouns included in an embedded
clause which might add complications not pertaining to the discussion at hand.

98
 u2 . j j 

\

% i \

 '


j j olaynman-ey manna-n yengswui-lul mwuchek
(SBJj ) (SBJj ) a-while-in meet-RelEnd Yengswui-Acc very
0

> 
.
pankaweha-ess-ta.
delight-at-Past-Dec.
(Shej ) was delighted at Yengswui who (shej ) met after a long while.

u3 . j j r

 E  l\
   6 %
\
i f " ? /
4 .
j j camsi yayki-lul ha-taka taum yek-eyse nali-ess-ta.
(SBJj ) (SBJj ) briefly talk-Acc do-And next station-at get-off-Past-Dec.
(Shej ) chatted briefly and (shej ) got off at next station.

There is a strong preference to interpret the two main-clause zero pronoun subjects in
(u2 ) and (u3 ) as referring to Cenghi.8 According to Constraint 3 and Cf-ranking of the

standard Centering Theory, Yengswu is the Cb(u2 ). On the other hand, Kims (2003) Cf-
ranking coupled with Kameyamas (1996, 1998) EXP CENTER rule establishes Cenghi, the
referent of the zero pronoun of the utterance (u2 ), as its Center C(u2 ). Cb(u3 ) and C(u3 )
are assigned to Cenghi under both accounts. Since C(u2 ) and C(u3 ) are the same under

Kims (2003) approach, the transition between the two utterances is CONTINUE; under
the traditional centering approach, the transition involves shifting of Cbs and therefore
constitutes the less desirable SMOOTH SHIFT.

3.3 Topic-Marked NPs vs. Zero Pronouns

Let us go back to the discussion of the role of topic-marked NPs and zero pronouns. Kim
(2003) placed both on the top of the Cf-ranking in complementary distribution so the former
is top-ranked in discourse-initial segments and the latter in all other positions of discourse.
Topic-marked NPs are omitted in the Cf-ranking for non-discourse-initial utterances, and
8
The two zero pronoun subjects of the embedded clauses in (u2 ) and (u3 ), which are not discussed here,
also refer to Cenghi.

99
zero pronouns are effectively excluded from the other ranking reserved for discourse-initial
utterances (see Footnote 6), rendering the pair functional equivalents in complementary
distribution.

Topic-marked NPs, however, can be found in non-discourse-initial segments. Consider

the following example:

(123) u1 .
%

i
H
 Qj
# ] 
t ^=\
o f " &

Bj \




z.
yengswui-nun ecey cihachel-eyse cenghij -lul manna-ess-ta.
Yengswui-Top yesterday subway-at Cenghij -Acc meet-Past-Dec.
Yengswui met Cenghij yesterday in the subway.

u2 .
&

Bj
H
 j 

\

% i \

 '


cenghij -nun j olaynman-ey manna-n yengswui -lul mwuchek
Cenghij -Top (SBJj ) a-while-in meet-RelEnd Yengswui -Acc very
0

> 
.
pankaweha-ess-ta.
delight-at-Past-Dec.
Cenghij was delighted at Yengswui who (shej ) met in a while.

u3 . j j 

o
 \


f"  6%
\
i f "
j j cenhwapenho-lul cwu-konase taum yek-eyse
(SBJj ) (SBJj ) phone-number-Acc give-After next station-at
/
? 
4 .
nali-ess-ta.
get-off-Past-Dec.
(Shej ) gave phone number and (shej ) got off at the next station.

This segment is exactly the same as the previous example (122) with the same semantic
interpretation, except that the matrix subject of (u2 ), previously represented by a zero pro-
noun, is now instantiated with an overt subject with the topic marker
H /nun/. However,


this example (123) is perceived as more natural than the previous (122): among 10 native
speakers of Koreans that were polled, all agreed that (123) is more natural than (122).
Note that there is no ambiguity in interpreting the subject of (122 u2 ) even when left
empty: because of the mention of Yengsu in the object position, the subject is forced to

100
be interpreted as referring to Cenghi in order to obtain a coherent meaning. Why, then,
does use of an overt topic-marked NP in place of a zero pronoun result in a more felicitous
discourse? I propose that the use of zero pronoun in the subject position is associated with

the CONTINUE transition, while topic-marked NPs in the same position signal establishing
of a new Center.9 The proposal is formulated as below:

(124) The functions of topic-marked NPs and zero pronouns in the subject position
in Korean:

1. zero pronouns and topic-marked NPs in the subject position mark the Center
of the current utterance, i.e., C(ui ).

2. zero pronoun subjects signal a CONTINUE transition between ui1 and ui , i.e.,
C(ui ) = C(ui1 ), while topic-marked NP subjects signal a newly established
Center, i.e., C(ui ) 6= C(ui1 ) (SHIFT transition) or a new C(ui ) where i=1.

A corollary can be derived from this rule regarding zero pronoun reference: zero
pronouns are co-referential with the top-ranked element in the previous utterance, i.e.,

Cb(ui1 ):

(125) Zero pronoun subject reference rule in Korean:

Zero pronoun subjects refer to the Center of the previous utterance, i.e., C(ui1 ).

There are a few questions to be examined. First of all, use of zero pronouns for a Center

which is different from the Center of the previous utterance is penalized, as explained
above; is use of the topic-marked NP construction in an environment where current Center
is the same as the previous Center disfavored as well? The examples shown above can be
9
As noted earlier, topic-marked NPs in non-subject positions are known to carry a distinct function of
contrastive topic, which prohibits the possibility of extending a similar contrast with zero pronouns to other
positions.

101
altered as follows to create such an environment. 10

(126) u1 .
%

i
H
 Qj
# ] 
t ^=\
o f " &

Bj \




z.
yengswui-nun ecey cihachel-eyse cenghij -lul manna-ess-ta.
Yengswui-Top yesterday subway-at Cenghij -Acc meet-Past-Dec.
Yengswui met Cenghij yesterday in the subway.

u2 . %

i H/i

@
? s
9


>  `


yengswui-nun/i olaysman-i-lamye pankapkey mal-ul
Yengswui-Top/(SBJi ) a-while-Cop-Quote cheerfully talk-Acc

3
 %
 .
kel-ess-ta.
start-Past-Dec.
Yengswui /(Hei ) cheerfully said hello saying its been a while.

u3 .
%

i H/i
 i 

o
 \


f"  6 %
\
i f "
yengswui-nun/i i cenhwapenho-lul cwu-konase taum yek-eyse
Yengswui-Top/(SBJi ) (SBJi ) phonenumber-Acc give-After next station-at
/
? 
4 .
nali-ess-ta.
get-off-Past-Dec.
Yengswui /(Hei ) gave phone number and (hei ) got off at next station.

Surprisingly, the prediction does not bear out, and the use of an overt NP subject with
the topic marker is not disfavored. The reason, we believe, is to be found in the follow-
ing. While Yengsu, the Center of (u1 ), is favored as the subject of subsequent utterances,
10
There are a dative argument of say hello in (u2 ) and also a dative argument of the adverbial clause give
phone number in (u3 ), both referring to Cenghi. If they exist in addition to the matrix subject zero pronouns,
two questions are raised: first, how their interpretations are determined; second, how their Cf-rankings are
determined in relation to the zero subject pronouns. I will assume that the subject zero pronouns are given
precedence over the dative zero pronouns in both questions. First, the interpretations of the dative zero
pronouns are contingent on those of their respective subject zero pronouns, freeing them from the requirement
of being coreferential with the Center of the preceding utterance. Second, given both pronouns have the same
formal property, the subject zero pronoun is ranked higher than the dative zero pronoun, obtaining the Center
status. The dative zero pronouns are omitted here for the sake of simplicity, but will be examined further later
in the section.

102
this does not mean that ambiguity is absent altogether: one cannot entirely rule out the
possibility that Cenghi is the one who said hello, and therefore, use of an overt NP in the
subject position in fact helps remove the uncertainty. While more people pointed Yengswu

as the referent of the zero subjects in (u2 ) and (u3 ), some regarded Cenghi as a possible
referent. This predicts that in the absence of alternatives, repeated use of topic-marked NPs
instead of a zero pronoun will be more clearly penalized. Indeed, in the following example
where Yengswu is made the only choice for the center status of the first utterance, everyone
agrees that the overt topic-marked subjects in (u2 ) and (u3 ) below make the segment sound

unnatural and that zero pronoun subjects are much preferred.

(127) u1 .
%

i
H
 {
B 
9 
t ^=`
o 
  

H
 
 .
yengswui-nun mayil cihachel-ul tha-ko chwulkunha-nta.
Yengswui-Top everyday subway-Acc ride-And commute-PresDec.
Yengswui commutes on subway everyday.

u2 . ?%

H/i
 
6r \ `
|
9
 f

"f", 7r A\

?yengswu-nun/i 6si-ey cip-ul nase-se 7si-ccum-ey
?Yengswu-Top/(SBJi ) 6oclock-At home-Acc leave-And 7oclock-about-At


f \



.
cikcang-ey tochakna-nta.
work-At arrive-PresDec.
?Yengswu/(Hei ) leaves home at 6 and arrives at work around 7.

u3 . ?% H/i
 
t ^=s
o  q

H
 
r `


?yengswu-nun/i cihachel-i pwumpi-nun sikan-ul
?Yengswu-Top/(SBJi ) subway-Nom crowded-RelEnd time-Acc

x 9 
 X
s >
O  { n
9 

 HK
 
 

.
phiha-lyemyen ilehkey ilccik chwulkunha-eya ha-nta.
avoid-Intend this-way early commute-AuxEnd must-PresDec.
?Yengswu/(Hei ) must commute early like this in order to avoid the hours when
subway is crowded.

Eliminating competition entirely clearly renders the recurring use of topic-marked NPs
sub-optimal. What about, then, a competition where the winner outranks its alternatives
with a more substantial margin? The following segment depicts such a scenario:

103
(128) u1 .
%

i
H
 Qj
# ] 
t ^=\
o f " # Qw =h
n &
 ;
2 j \


yengswui-nun ecey cihachel-eyse elilcek chinkwuj -lul
Yengswui-Top yesterday subway-at childhood friendj -Acc



z.
manna-ess-ta.
meet-Past-Dec.
Yengswui met a childhood friendj yesterday in the subway.

u2 . ?%

i
H/i

@
? s
9


>  `


?yengswui-nun/i olaysman-i-lamye pankapkey mal-ul
?Yengswui -Top/(SBJi ) a-while-Cop-Quote cheerfully talk-Acc

3
 %
 .
kel-ess-ta.
start-Past-Dec.
Yengswui /(Hei ) cheerfully said hello saying its been a while.

u3 . ?% i H/i
 i 

o
 \


f"  6 %
\
i f "
?yengswui-nun/i i cenhwapenho-lul cwu-konase taum yek-eyse
?Yengswui -Top/(SBJi ) (SBJi ) phonenumber-Acc give-After next station-at
/
? 
4 .
nali-ess-ta.
get-off-Past-Dec.
Yengswui /(Hei ) gave phone number and (hei ) got off at next station.

In this example, Cenghi is replaced with a childhood friend. While Cenghi must be
someone whom the speaker is familiar with, Yengswus childhood friend is more likely to
be a newly introduced entity from the hearers point of view. The two NPs, then, occupy the
two points at the either end of the Cf-ranking hierarchy: the degree by which Yengswu-Top

outranks its competitor is far greater than the one between Yengswu-Top and Cenghi-Acc
in the previous examples. As predicted, there is very little ambiguity as to the referent of
the subject of the two subsequent utterances, and the zero pronoun subjects are preferred
over overt topic-marked NP subjects for them.

Let us consider another example which illustrates the roles of zero pronouns and topic-
marked NPs and how the choice between them interacts with various factors presented by
the specifics of the discourse. In the following, two discourse connectives are used in the

104
second utterance which serve two distinct functions: kuliko is used to signal continuation
of an action, which typically involves retaining the same subject in the previous utterance;
kuleca, on the other hand, is predominantly used when a switch in reference between the

preceding and the following utterances occurs (Choi and Joh, 2004)11 . Because of the
lexical cues supplied by these connectives, there is no ambiguity in interpretation of the
two respective zero pronoun subjects: in (u2 .a) unambiguously refers to Yengswu, the
subject of (u1 ); in (u2 .b) refers to Cheli, the object of (u1 ).

(129) u1 .
%

i  =s
o
^ j \
 
l @
 / M : 
4 .
yengswui-ka chelij -lul kapcaki han-tay ttali-ess-ta.
Yengswui-Nom Chelij -Acc suddenly one-hit hit-Past-Dec.
Yengswui hit Chilij all of a sudden.

u2 .a  o

i/j /% i H
 
 
o\ t
 l
kuliko i/j /yengswui-nun makwu soli-lul cilu-ki
and (SBJi/j )/Yengswui-Top intensely voice-Acc shout-Nominal

r

.
sicakha-ess-ta.
start-Past-Dec.
And (hei/j )/Yengswui started shouting loudly.

u2 .b Q  i/j /^
=s
o j H
 
 
o\ t
 l
kuleca i/j /chelij -nun makwu soli-lul cilu-ki
then (SBJi/j )/Chelij -Top intensely voice-Acc shout-Nominal

r

.
sicakha-ess-ta.
start-Past-Dec.
Then (hei/j )/Chelij started shouting loudly.

From a minimalistic line of reasoning, one might hypothesize that the zero pronoun subjects

will be preferred in both cases since the two discourse connectives themselves provide
sufficient information on the reference of the subject. In reality, native speakers of Korean
11
Choi and Joh (2004) claim that kuliko and kuliko-nun are associated with the constraint Cb(ui ) =
Cp(ui+1 ), with the latter subject to an additional constraint Cp(ui ) = Cp(ui+1 ). kuleca, which they call
reference switching marker, carries the constraint Cp(ui ) 6= Cp(n+1).

105
unanimously find the overt NP subject in (u2 .b) preferred over the zero pronoun subject12 ;
most prefer the zero pronoun subject in (u2 .a) but a few also find the overt NP subject
more than acceptable. The dominant preference for the overt NP subject in (u2 .b) implies

that felicitous use of zero pronouns is not sufficiently conditioned by the simple measure
of recoverability as has been suggested throughout literature: it is governed by a more
strict condition that mandates that C(ui1 ) = C(ui ), and therefore a choice of overt noun is
preferred in the environments where the condition is not met. The acceptability of the overt
subject in (u2 .a) is in line with the similar judgment shown for (126): both illustrates a

case of an overt NP form judged acceptable where use of the less informative zero pronoun
form would have resulted in the intended interpretation on the hearers part. The use of the
overt NP subjects are, therefore, redundant. Seemingly illogical, this type of redundancy
in natural language usage is in fact reported to be a prevalent phenomenon (Walker, 1993).

Presumably, at some point in the scale the amount of redundancy itself becomes excessive
as to render the extra information undesirable: the unnaturalness of the overt NP subjects
in the earlier example (127) serves this point.
Another important question to consider is: is marking of shift in Centers uniquely done

with topic-marked NPs or can it be achieved with other overt NP types? If shift of refer-
ence in the sequence of two Centers demands a more marked way of specifying the referent
of the second Center, any NP types that are more informative than a zero pronoun might



do for clarification purposes. Besides & B
H Cenghi-Top, possible choices of referring

expressions are numerous: pronoun followed by either the topic marker suffix or a nomina-

.
tive suffix ( 
H/
 . she-Top/she-Nom), demonstrative + common noun followed

by either the topic marker, a nominative suffix or in the bare form ( #
H/
 #

/
#
 that woman-Top/that woman-Nom/that woman), proper noun followed by a
12
Choi and Joh (2004) also finds that there is a strong correlation between the use of the connective kuleca
and an overt NP subject.

106



nominative suffix (& B Cenghi-Nom), and bare proper noun (&


B Cenghi). The
point of interest here is the overt pronoun, since other NP forms can contain enough infor-
mation unto themselves to uniquely identify their referents. Turan (1995) found that overt

and zero pronouns in Turkish play distinct roles in center transitions (discussed earlier in
39): null subjects encode a CONTINUE or SHIFT- SUBJ transitions, that is, they realize the
previous Cp or Cb; Overt pronoun subjects mostly occur in SHIFT- TO - OBJECT transitions.
Di Eugenio (1990, 1996, 1998) reports similar findings for the case of Italian zero and
overt pronouns. The exact same contrast does not extend to Korean zero and overt pro-

nouns, however: as noted in Section (2.2), overt Korean pronouns in the 3rd person are a
marked phenomenon with their usage highly restricted to a certain type of genre, and there-
fore a fair comparison between overt and covert pronouns is not readily available. There
are two parameters in the choice of referring expressions: first the nominal form (proper

name, demonstrative + common noun or pronoun) and the choice of suffix (topic marker,
case marker or none). In this case, the choice of nominal form is irrelevant, and the topic-
marked demonstrative + common noun will achieve the same effect as the topic-marked
proper noun. For the rest of the choices of suffixes, the bare form is not suited for the

position, but the case-marked form is.

In the following example, (u2 ) shows all three possible NP forms plus the ungrammat-
ical bare nominal form for the subject position side by side for comparison; here, the zero
dative pronoun is represented for both (u2 ) and (u3 ). Two interpretations of (u3 ) are rep-
resented in (u3 .a) and (u3 .b) which vary in terms of references assigned to the two subject

pronouns and the dative pronoun. The point of interest is whether the differing forms of
(u2 )s subject affect the interpretation of (u3 ). Under the assumption that the zero subject of
(u3 ) is coreferential with the Center of (u2 ), prevalence of interpretation (u3 .a) means that
Cenghi outranks Yengswu to become the C(u2 ); prevalence of interpretation (u3 .b) means

exactly the opposite ranking.

107
(130) u1 .
%

i
H
 Qj
# ] 
t ^=\
o f " &

Bj \




z.
yengswui-nun ecey cihachel-eyse cenghij -lul manna-ess-ta.
Yengswui-Top yesterday subway-at Cenghij -Acc meet-Past-Dec.
Yengswui met Cenghij yesterday in the subway.

u2 .
&

Bj
H/&


Bj /j /*&

Bj
@
? s
9 i
cenghij -nun/cenghij -ka/j /*cenghij olaysman-i-lamye i
Cenghij -Top/Cenghij -Nom/(SBJj )/*Cenghij a-while-Cop-Quote (DATi )



>  `

 
3
 %
 .
pankapkey mal-ul kel-ess-ta.
cheerfully talk-Acc start-Past-Dec.
Cenghij /(Shej ) cheerfully said hello (to Yengswui ) saying its been a while.

u3 .a [Noun-Topj /Noun-Nomj /zeroj > zeroi; SBJj > DATi]

j j i 

o
 \


f"  6 %
\
i f "
j j i cenhwapenho-lul cwu-konase taum yek-eyse
(SBJj ) (SBJj ) (DATi ) phonenumber-Acc give-After next station-at
/
? 
4 .
nali-ess-ta.
get-off-Past-Dec.
(Shej ) gave phone number (to himi) and (shej ) got off at the next station.

u3 .b [zeroi > Noun-Topj /Noun-Nomj /zeroj ; DATi > SBJj ]

i i j 

o
 \


f"  6 %
\
i f "
i i j cenhwapenho-lul cwu-konase taum yek-eyse
(SBJi ) (SBJi ) (DATj ) phonenumber-Acc give-After next station-at
/
? 
4 .
nali-ess-ta.
get-off-Past-Dec.
(Hei ) gave phone number (to herj ) and (hei ) got off at the next station.

The most obvious preference is with the zero pronoun subject in (u2 ): only the reading
(u3 .a) is deemed possible. In fact, the interpretation of the zero pronoun subject in (u3 )

covaries with that of the zero pronoun subject in (u2 ): whether or not they refer to Yengswu
or Cenghi in (u1 ), the strong intuition is to keep them coreferential. On the other hand,
the judgment on the two other cases of the topic-marked and nominative-marked subject is

108
less clear. While people favor the interpretation (u3 .a) when pressed to pick one, there also
exists a reluctance in ruling out (u3 .b) altogether. Therefore, the Cf-ranking that (u3 .a) sup-
ports in this case, i.e., [Noun-Top/Noun-Nom > zero], is only weakly confirmed.

To gain insights as to why this is the case, consider the following example, where the
subject and the dative arguments of (u2 ) are reversed:

(131) u1 .
%

i
H
 Qj
# ] 
t ^=\
o f " &

Bj \




z.
yengswui-nun ecey cihachel-eyse cenghij -lul manna-ess-ta.
Yengswui-Top yesterday subway-at Cenghij -Acc meet-Past-Dec.
Yengswui met Cenghij yesterday in the subway.

u2 . i
@
? s
9


& Bj \
 >


> `

 
3
 %
 .
i olaysman-i-lamye cenghij -eykey pankapkey mal-ul kel-ess-ta.
(SBJi ) a-while-Cop-Quote Cenghij -To cheerfully talk-Acc start-Past-Dec.
(Hei ) cheerfully said hello to Cenghij saying its been a while.

u3 .a [Noun-Datj > zeroi; DATj > SBJi ]

j j i 

o
 \


f"  6 %
\
i f "
j j i cenhwapenho-lul cwu-konase taum yek-eyse
(SBJj ) (SBJj ) (DATi ) phonenumber-Acc give-After next station-at
/
? 
4 .
nali-ess-ta.
get-off-Past-Dec.
(Shej ) gave phone number (to himi) and (shej ) got off at the next station.

u3 .b [zeroi > Noun-Datj ; SBJi > DATj ]

i i j 

o
 \


f"  6 %
\
i f "
i i j cenhwapenho-lul cwu-konase taum yek-eyse
(SBJi ) (SBJi ) (DATj ) phonenumber-Acc give-After next station-at
/
? 
4 .
nali-ess-ta.
get-off-Past-Dec.
(Hei ) gave phone number (to herj ) and (hei ) got off at the next station.

For this case, there is little room for ambivalence: (u3 .b) is strongly favored over (u3 .a),
which gets no vote at all.

109
 The reason for the contrast between (130) and (131), I believe, is that both the formal
factor and the grammatical role play into the Cf-ranking. (131) has a clear winner since
the zero subject in (u2 ) emerges as the favored choice on both criteria; (130), on the other

hand, has a conflict of interests at hand, since the zero pronoun in (u2 ), which has the
advantage of scaling high in terms of its formal property, suffers from being a lower-ranked
grammatical role. Therefore, a Centering Theory which bases its Cf-ranking solely on the
measure of grammatical role or that of the referring expressionss formal property is bound
to fail to explain the contrast between the two examples. Kims (2003) is an attempt to

incorporate both measures into one Cf-ranking; however, hers is limited in the way the two
are compartmentalized into non-overlapping environments, i.e., the form-based ranking to
given information and the grammatical-role-based ranking to new information, and cannot
capture the interaction of the two criteria as displayed by the examples. Other Cf-rankings

proposed for Korean, which freely borrow parts of the two measures to construct a single
Cf-ranking, suffer from arbitrary orderings which are hard to prove with neutral examples.
Clearly, it stands to reason that one dimension of linguistic aspect can be ordered in terms of
the salience their members carry: within the range of grammatical roles, a subject wins over

an indirect object an NP can be either one at a time after all. However, when comparing
two elements with properties that are of different linguistic dimensions, it is not entirely
obvious that a clear winner should even emerge: comparing a zero pronoun and a subject
would be one such case.
We conclude that there can be many linguistic scales that are representative of salience

of NP elements, but determining the relative rankings between specific items from different
measures is neither possible nor theoretically well motivated. We claim, therefore, that
there is no single absolute hierarchy for determining Cf ordering. Rather, Cf rankings
are determined by a number of linguistic scales that are potentially relevant in measuring

salience, which include the following well known hierarchies:

110
(132) Linguistic Hierarchies for Cf ranking:

a. grammatical role

subject > object > other arguments

b. thematic role
agent > experiencer > (inalienable) possessor > theme

c. clausal embedding level

main-clause arguments > arguments in 1-level embedding > arguments in 2-
level embedding > . . .

d. discourse participants
speaker, hearer > others

e. linear order
sentence-initial > others

f. recency

mentioned more recently > mentioned less recently

g. noun form: post-position marker

topic-marked noun > case-marked noun > bare noun

h. noun form: overtness

zero form > overt form (proper noun, demonstrative + common noun, pronoun)

Following Kameyama (1996, 1998) and Kim (2003), we posit a single notion of Center
which replaces both Cb and Cp, under the assumption that a backward-looking center (Cb)
of an utterance also functions as the preferred forward-looking center (Cp). As argued
earlier in (124), zero pronouns and topic-marked overt NPs in the subject position mark the

Center of the current utterance, while differing in terms of the transition types they signal.
The zero pronoun subjects in Korean, then, refer to the Center of the previous utterance, as
presented in (125). The two claims are repeated below:

111
(133) (= 124) The functions of topic-marked NPs and zero pronouns in the subject
position in Korean:

1. zero pronouns and topic-marked NPs in the subject position mark the Center

of the current utterance, i.e., C(ui ).

2. zero pronoun subjects signal the CONTINUE transition between ui1 and ui ,

i.e., C(ui ) = C(ui1 ), while topic-marked NP subjects signal a newly estab-

lished Center, i.e., C(ui ) 6= C(ui1 ) (i.e., SHIFT transition) or a new C(ui )
where i=1.

(134) (= 125) Zero pronoun subject reference rule in Korean:

Zero pronoun subjects refer to the Center of the previous utterance, i.e., C(ui1 ).

This leaves zero pronouns in non-subject positions unaccounted for. It is easily verified
that zero pronouns in such positions also prefer the previous Center as their referent. In
the following example, Chelswu, the Center of the preceding utterance, is favored as the
referent of the two zero pronouns in the dative positions in (u2 ):

(135) u1 . A: o^
= i H
 Qj
# ]

% j
< 
 

` >
chelswui -nun ecey yengswuj -wa ssaum-ul khekey
Chelswui -Top yesterday Yengswuj -With fight-Acc hugely
@

/.
ha-ess-tay.
do-Past-Reportive.
Chelswui had a big fight with Yengswuj yesterday (reportedly).

u2 . B:
A? ? /  i 

oK f" 
 t7 x

` i
kulay? nay-ka i cenhwaha-ese cachocicong-ul i
yeah? I-Nom (DATi ) call-And situation-Acc (DATi )
t#

Q<  x
.
mwulepo-ayakess-ta.
ask-will-Dec.
Yeah? I should call (himi) and ask (himi) about the situation.

112
 Special attention is needed, however, in those cases containing multiple instances of
zero pronouns. Let us revisit our previous example (=126, dative zero pronouns represented
this time):

(136) u1 .
%

i
H
 Qj
# ] 
t ^=\
o f " &

Bj \




z.
yengswui-nun ecey cihachel-eyse cenghij -lul manna-ess-ta.
Yengswui-Top yesterday subway-at Cenghij -Acc meet-Past-Dec.
Yengswui met Cenghij yesterday in the subway.

u2 . %

i H/i

@
? s
9 j


>  `


yengswui-nun/i olaysman-i-lamye j pankapkey mal-ul
Yengswui-Top/(SBJi ) a-while-Cop-Quote (DATj ) cheerfully talk-Acc

3
 %
 .
kel-ess-ta.
start-Past-Dec.
Yengswui /(Hei ) cheerfully said hello (to herj ) saying its been a while.

u3 . %

i H/i
 i j 

o
 \


f"  6

yengswui-nun/i i j cenhwapenho-lul cwu-konase taum
Yengswui-Top/(SBJi ) (SBJi ) (Datj ) phonenumber-Acc give-After next
\
i
% f " ? /
4 .
yek-eyse nali-ess-ta.
station-at get-off-Past-Dec.
Yengswui /(Hei ) gave phone number (to herj ) and (hei ) got off at next station.

Here, the i s in (u2 ) and (u3 ) are in the subject position, and they are interpreted as
coreferential with the Center of their previous utterance. On the other hand, the j s in
the dative argument position, subject to the semantic constraint which forces their referents

to be disjoint from that of their respective subjects, refer to Cenghi which never achieves
the Center status. While it seemingly contradicts the findings from the previous example
(135), it should be treated as analogous to the parallel case containing overt pronoun sub-
jects instead, that is, as a case in which the requirement for coherent semantic interpretation

renders the pronoun reference unambiguous, under the assumption that the interpretation of
zero pronoun subjects takes precedence over that of other zero pronouns. Note that, how-
ever, a zero pronoun subject forced to refer to a non-Center due to a semantic consideration

113
is best replaced with an overt topic-marked NP subject (Example 123): no such preference
exists for the case of zero pronouns in other positions.
What about non-subject zero pronouns ability to establish themselves as the Center?

In the following segment, the zero pronoun dative argument in (u2 ) is competing for the
Center status with the subject argument expressed in three different forms: topic-marked
NP, zero pronoun, and the subject-case-marked NP:

(137) u1 .
%

i
H
 Qj
# ] 
t ^=\
o f " &

Bj \




z.
yengswui-nun ecey cihachel-eyse cenghij -lul manna-ess-ta.
Yengswui-Top yesterday subway-at Cenghij -Acc meet-Past-Dec.
Yengswui met Cenghij yesterday in the subway.

u2 . %

i
H/i /%


i 
@
? s
9 j


> 
yengswui-nun/i/yengswui -ka olaysman-i-lamye j pankapkey
Yengswui-Top/(SBJi )/Yengswu-Nom a-while-Cop-Quote (DATj ) cheerfully
`

 
3
 %
 .
mal-ul kel-ess-ta.
talk-Acc start-Past-Dec.
Yengswui /(Hei )/Yengswui cheerfully said hello (to herj ) saying its been a

while.

u3 . i i j 

o
 \


f"  6 %
\
i f "
i i j cenhwapenho-lul cwu-konase taum yek-eyse
(SBJi ) (SBJi ) (Datj ) phonenumber-Acc give-After next station-at
/
? 
4 .
nali-ess-ta.
get-off-Past-Dec.
(Hei ) gave phone number (to herj ) and (hei ) got off at next station.

In all three cases, the subject zero pronoun in (u3 ) is more naturally interpreted as referring

to Yengswu in the subject position. It can be deduced, therefore, that the zero pronoun
dative argument is not the Center of (u2 ).
To sum up, a zero pronoun argument in a non-subject position patterns with subject
zero pronouns in that it favors the Center of the preceding utterance as its referent; it does

114
not, however, establish itself as the Center. The Zero pronoun reference rule in Korean
presented above (125, 134) is thus generalized to include non-subject zero pronouns:

(138) (= 125, 134) Zero pronoun reference rule in Korean:

Zero pronouns refer to the Center of the previous utterance, i.e., C(ui1 ).

3.4 The Centering Theory and Zero Pronoun Resolution

Given the proposals (133, 138) for zero pronoun interpretation within the Centering Frame-
work, the question arises: how suited are they for the task of zero pronoun resolution? The

short answer is that zero pronoun resolution often does not reduce to the seemingly simple
process of identifying the Center of the preceding utterance, as the zero pronoun reference
rule in (138) prescribes. There is a multitude of reasons.
The first and more obvious one has to do with the fact that Centering Theory makes

predictions on only the referring-type zero pronouns, i.e., the anaphoric zeros. All other
types of zero pronouns including explicit zero pronouns, generic zero pronouns and deictic
zero pronouns are not motivated by discourse-related factors and therefore are not subject
to the principles of the Centering Theory which is foremost meant to be a model for local

coherence of a discourse.
Secondly, it is widely acknowledged that semantic considerations routinely disrupt the
coreference pattern predicted by the rules presented above. The effect is manifested as
an absolute constraint in some cases, such as when a subcategorization frame of a verb
requires certain semantic features for its arguments, excluding higher-ranked NPs without

such features as potential antecedent as a results. For example:

(139) u1 . o
^
= c
H
 Z
t &
d
h 
xp \ 
  3
3
" %
 .
chelswuc -nun onul cemsim-ulo phicap -lul sa-e mek-ess-ta.
Chelswuc -Top today lunch-As pizzap -Acc buy-And eat-Past-Dec.
Chelswuc ate pizzap for lunch today.

115
 u2 . c/p 
z Kf" Z 
> % 
i .
c/p nukkiha-ese pyello-i-ess-ta.
(SBJc/p ) greasy-And mediocre-Cop-Past-Dec.
(Itc/p ) was greasy and mediocre.

A more subtle case is when a referring expression other than the Center is preferred

as the antecedent of a later instance of zero pronoun in order to obtain a more probable
semantic interpretation of the discourse. In the following example, Cheli, not Yengswu
which is the Center of (u1 ), is preferred as the referent of the zero pronoun subjects in
(u2 .b) because he is more likely to be the one who is/is not hurting or crying due to the

fact that hes the one who suffered by the attack described in (u1 ). By contrast, when the
second utterance depicts a continued act of aggression (in u2 .a), its subject is more naturally
interpreted as Yengswu.

(140) u1 .
%

i  =s
o
^ j \
 
l @
 / M : 
4 .
yengswui-ka chelij -lul kapcaki han-tay ttayli-ess-ta.
Yengswui-Nom Chelij -Acc suddenly one-hit hit-Past-Dec.
Yengswui hit Chelij all of a sudden.

u2 .a 
A   Ht
  i/?j 
_ \ 
 #
Q 
.
kulayto mocala-nunci i/?j uyca-lul ketecha-ess-ta.
(SBJ) yet not-suffice-If (SBJi/?j ) chair-Acc kick-Past-Dec.
(Hei/?j ) kicked a chair as if (it) still wasnt enough.

u2 .b ?i/j 
A 
rt
 ?i/j t




 .
?i/j kulayto an aphu-nci ?i/j wul-ci anh-ess-ta.
(SBJ?i/j ) yet not hurt-If (SBJ?i/j ) cry-AuxEnd didnt-Past-Dec.
(He?i/j ) didnt cry as if (he?i/j ) wasnt hurt.

In fact, zero pronouns that are truly ambiguous, with no lexical cues present around
them to bias their interpretation towards a certain element, are quite rare in actual usage.13
When faced with a potentially significant level of ambiguity in reference interpretation
13
They are also hard to construct: all the examples used in the previous section were carefully constructed
to remove such biases through multiple revisions.

116
associated with using a zero pronoun, a user tends to opt for an overt NP expression rather
than leaving its resolution solely to the measure of salience, as shown earlier with Example
(126).

Lastly, rejecting a single Cf-ranking hierarchy in favor of a potentially open set of var-
ious linguistic hierarchies, as noted in (132), leaves the process of Center selection largely
unspecified. The rule (133) states that a zero pronoun or a topic-marked NP in the subject
position, when present, is the Center of the utterance; when neither is present, however,
there is no written rule that dictates whether the case-marked subject, a top element on the

grammatical role scale, or the speaker in a dative argument position, a top element on the
discourse-participant scale, should win. Presumably, their respective high-ranked statuses
in the two hierarchies contribute towards their Center-hood, as do many other factors rep-
resented by the remaining hierarchies. In the end, the degrees of contribution, or weight,

of the factors involved decide which nominal expression emerges as the Center of an ut-
terance and the preferred antecedent for later zero pronoun occurrences. Discovering and
assigning appropriate amounts of weight to each rank in the set of hierarchies is arguably
best achieved by means of replicating statistical patterns in observed data: a task which

is naturally suited to the methodology of machine learning. In a later chapter of Part II,
we explore in detail how machine learning techniques can be successfully employed in
modeling such rankings.

117
 Part II

Resolution

118
Overview

Having given the problem of Korean zero pronouns linguistic-theoretic attention, we now

turn to a slightly different perspective and devote the second half of the thesis to the engi-
neering side of the problem: building a system that can resolve Korean zero pronouns. In
undertaking the process, we make an effort to incorporate the knowledge and understanding
gained from the deeper linguistic analyses, discussed in the previous part.

In many ways, we are lucky to have a perfect vehicle for empirically testing out the
theoretical insights: the two Penn Korean Treebank corpora, which are syntactically anno-
tated collections of Korean texts with zero pronouns pre-identified in their correct syntactic
positions. After we augmented them with reference annotations, they provided a set of

valuable data that are naturally suited for the task of constructing statistical models for
pronoun resolution.
In designing and building resolution models, we focus primarily on identifying and
combining many pieces of linguistic factors that seemingly contribute to correct interpreta-
tion of zero pronouns, and less on striving to produce the very best performance by experi-

menting with various cutting-edge statistical techniques. To the best of our knowledge, ours
is the first study of its own kind, one which aims to develop a general-purpose resolution
system on Korean zero pronouns trained on a large set of annotated data.
The rest of Part II is organized as follows. First, previous work in the field of pronoun

resolution is reviewed. In the next chapter, the data set used in the study, the two Penn

119
Korean Treebank corpora, is introduced, as well as the details and design of the reference
annotations made on the data. Next, we present the results from applying rule-based ap-
proaches to the data, in particular the naive Hobbs algorithm (Hobbs, 1977, 1978) and its

modified versions. Finally, machine-learning-based resolution models are discussed, where

several different approaches are tried for building systems that target either all types of zero
pronouns or the NP-anaphoric type only.

120
Chapter 1

Previous Work

The problem of treating zero pronouns has attracted the attention of many researchers in
the field of computational linguistics working with so-called pro-drop languages. The need
for resolving zero pronouns was recognized early on, as the problem proved essential in
developing many key natural language processing technologies, such as machine transla-

tion and text summarization. At present day, zero pronoun resolution is mostly seen as a
part of the larger problem of reference resolution. We first present a brief survey of current
studies in anaphora resolution, and then present a study on Spanish as a case of pronoun
resolution work that tackles the problem of zero pronouns in particular.

1.1 Pronoun Resolution: An Overview

Most pronoun anaphora resolution systems deal with resolution of anaphors which have
noun phrases as their antecedents, partly because identifying anaphors which have verb
phrases, clauses, sentences or even discourse segments as antecedents is seen as a separate

problem from the overall anaphora resolution task. Typically, all noun phrases preceding
a pronoun anaphor are initially regarded as potential candidates for antecedents. Usually,

121
a search scope has to be set: most approaches look for NPs in the current and preceding
sentence. Once a set of antecedent candidates is established, a resolution system then
attempts to track down the correct antecedent employing a number of strategies and factors.

1.1.1 Early Works

Early work on anaphora resolution made use of very restricted linguistic knowledge and
incorporated it into their strategies. Winograd (1972) was the first to develop procedures
for pronoun resolution in his SHRDLU system. He collects all the antecedent candidates
for a given pronoun and rates their plausibility based on their syntactic position such that

subject is favored over object, which is in turn favored over object of a preposition. In
addition, focus elements were favored, which were determined from the answers to WH-
questions and from indefinite noun phrases in yes-no questions. Hobbs (1977, 1978) was
one of the first systems that fully cast syntactic knowledge as the measure for ranking be-

tween antecedent candidates. Hobbss algorithm is based on various syntactic constraints

on pronominalization which are used to search sentence trees. Although it is regarded as
too simplistic by todays standards, that very characteristic means computational economy
and therefore counts as the systems biggest appeal. As a result, Hobbs algorithm is still

extensively compared and used as a benchmark by many recent systems. It was imple-
mented as a part of our own study, which is presented later in Chapter 3.
The late 1980s and early 1990s saw a surge of research on pronoun resolution. These
studies can be grouped into three approaches: the traditional approach, which extends
the earlier methods and integrates knowledge sources and factors in searching for the best

candidate, the statistical approach, which relies largely on statistical techniques based
on corpus data, and finally, the knowledge-poor approach, which discards linguistic and
domain knowledge in favor of simple methods based on information readily available on
the surface of text.

122
1.1.2 The Traditional Approach

Carbonell and Browns work (1988) belongs in the first camp. The authors propose a

general framework for anaphor resolution based on the integration of multiple knowledge
sources: sentential syntax, case-frame semantics, dialogue structure and general world
knowledge. They recast these knowledge sources in the forms of constraints and preference
rules. The constraints are first used in reducing the number of candidate antecedents for the
anaphor in question, and then preferences are applied to each of the remaining candidates.

The PUNDIT processing system proposed by Dahl and others (1986, 1990) used a local
focusing mechanism to suggest referents for a range of noun phrases including pronouns,
elided NPs, one-anaphora, and context-dependent full NPs. PUNDITs focusing algorithm
uses a single focus list, called the FOCUSLIST, to record elements of the discourse and to

suggest referents. After processing a sentence, the elements of that sentence are prepended
to the focusing list in the following order: the sentence, the direct object, the subject and
finally any objects of prepositions. If there is a pronoun in the sentence, its referent goes
at the front of the list. One important feature to note is that the elements are prepended

to the focus list on the basis of surface syntactic information rather than thematic roles, an
approach also adopted by Centering Theory (reviewed in Section 1.2).
Lappin and Leass (1994), also employing a traditional approach, rely on salience
measures derived from syntactic structure and a simple dynamic model of attentional state
to select from the set of candidates. They propose a set of parameters including grammati-

cal roles, parallelism of grammatical roles, frequency of mention, proximity, and sentence
recency as measures of salience.

123
1.1.3 The Statistical Approach

The entire field of computational linguistics saw a monumental shift in its methodology

with the introduction of statistical/machine-learning methods, and the topic of anaphora

resolution was no exception to this sweeping new trend. The appeal of statistical methods
is two-fold: first the labor-intensive process of constructing rules and preferences can be
avoided; second, statistically obtained models are more easily adaptable to new genres of
text. Crucially, these systems rely on large-scale, annotated corpora from which statistical

patterns of coreference are extracted, which serve as the building blocks of the resolution
models.
Nasukawa (1994) relies on collocation patterns as the main selectional constraint to
determine how eligible a candidate for antecedent is. When a pronoun in a text is sub-

mitted for pronoun resolution, the text is scanned for the existence of identical collocation
patterns. The antecedent candidates with a similar collocation pattern as exhibited by the
pronoun and its verb are preferred over others. Dagan and Itai (1990) also use collocation
information as the most important restriction factor, but unlike Nasukawa (1994) they ex-

tracted the patterns from an independent, external corpus and not from the text on which
resolution is currently being performed.
Connolly, Burger and Day (1994) is one of the first studies that employed a full-scale
machine learning approach for anaphora resolution. The idea is to cast anaphoric reference
as a classification problem for which a classifier can be discovered empirically using tradi-

tional learning methods, which became a standard methodology among machine-learning-

based systems of anaphora resolution. The approach adopted is to decompose the candidate
selection problem into separate two-class classification problems. Each classification prob-
lem is defined on a pair of candidates and an anaphor, where the classes correspond to

choosing one of the candidates as a better antecedent of the anaphor. Once a classifier
has been constructed, it is used by the reference resolution algorithm to select the best

124
candidate.

1.1.4 The Knowledge-Poor Approach

Most of the approaches outlined so far rely heavily on linguistic and domain knowledge.

One of the disadvantages these approaches share is that their methods are labor-intensive
and time-consuming: the rule-based approaches require carefully crafting rules and con-
ditions; the statistical approaches often require building large-scale annotated corpora for
training. Consequently, the need for inexpensive and robust systems, possibly suitable for
unrestricted texts, arose in the research community. We discuss three such efforts.

Kennedy and Boguraev (1996) aim to produce a system that extends the one previously
developed by Lappin and Leass (1994), but without expensive and knowledge-rich sources
such as full scale syntactic parsing. The system mostly parallels that of Lappin and Leass
(1994) but differs in that it works on the output of a part-of-speech tagger as well as on

some low-level grammatical information such as grammatical functions. The evaluation

results reported a poorer performance compared to that of Lappin and Leass (1994), but
the authors were able to show that their system is highly adaptable across a wide variety of
genres of text, including documents from the World Wide Web.

Similarly, Mitkovs (1996, 1997) approach eliminates the need for a parser or syntactic
information on the text input. Only a part-of-speech tagger is needed for identifying noun
phrases around the anaphor; the process of determining the best antecedent candidate is
based on simple heuristic rules that do not require any further syntactic knowledge.
Baldwins (1997) CogNIAC is designed based upon the observation that a significantly

large proportion of pronoun anaphors in fact do not require complex resolution strategies
but rather can be successfully resolved by simple ones. It operates on a set of core rules
designed to offer strategic resolution tailored for different subclasses of anaphors and spe-
cific conditions in which the pronouns occur. These core rules are applied to pronouns

125
in a successive manner, and those pronouns that none of the rules resolve are simply left
unresolved. The result is a high precision (92%) and a low recall (64%) score. When some
low-precision rules were applied to the remainder of pronouns, the precision dropped to

77.9%.

1.1.5 Resolution of Zero Pronouns: The Case of Spanish

There are several studies on anaphora resolution which specifically target zero pronouns in
so-called pro-drop languages. We review a study of zero pronoun resolution in Spanish as
a case study.

While there have been numerous works on pronoun resolution in Spanish, Ferrandez
& Peral (2000) is the first to focus specifically on the resolution of zero pronouns. They
note that the problem consists essentially of two sub-problems: detection and resolution.
In Spanish, they observe, zero pronouns occur only in subject positions, unlike many other

languages which also permit use of zero pronouns in other argument slots. Detection of
zero pronouns in Spanish is therefore facilitated by this fact they limit the search space to
clausal boundaries. They adopt a partial-parsing strategy which is designed to provide all
the necessary information for resolving anaphora without resorting to full syntactic pars-

ing, as described in Ferrandez et al. (1998). That study shows that only a small set of con-
stituents are needed for anaphora resolution: co-ordinated prepositional and noun phrases,
pronouns, conjunctions, and verbs. Zero pronoun subjects are then discovered by scanning
in each verb and determining whether or not it has an overt subject.
Once a zero pronoun is detected and their system inserts the pronoun in the appropriate

position, the pronoun is processed for resolution using two steps: applying a set of con-
straints, and then applying a set of preference rules. Working with this particular language
has an advantage: Spanish verbs are marked with morphological suffixes which indicate
agreement information, including Person and Number agreement, and Gender agreement

126
in the cases where the clausal verb is the copula. This information plays a key role in
filtering out unsuitable NP antecedent candidates, along with a few other criteria such as
c-command constraints and semantic consistency. If the application of restrictions fails to

isolate a single antecedent candidate, the preference rules are applied next. Ferrandez &
Peral (2000) use ten such preference rules: some of them encode varying levels of proxim-
ity between the pronoun and the proposed antecedent, while others promote the candidates
that meet certain criteria such as being a proper noun or having previously appeared in
the text with the same verb. To discover the optimal ordering between these preferences,

they hand-craft a corpus with 106 zero pronouns and train their Slot Unification Parser for
Anaphora Resolution (SUPAR) system on them.
Evaluation is carried on separately with the two subtasks, detection and resolution, us-
ing 1,599 verbs found in two sets of corpora consisting of diverse genres. Ferrandez &

Peral (2000) report that 46% of all verb tokens in the corpora have a zero subject. Among
those verbs with a zero subject, they report 98% of them were identified correctly. Per-
formance on verbs with overt subjects was poorer in comparison, with an 80% accuracy
rate, resulting in a lower overall combined accuracy of 88% for both types. For the task

of resolution, they considered only 228 anaphoric zero pronouns among the 734 zero pro-
nouns. The solution output by their SUPAR system was compared against judgments by
two human experts, and was deemed successful when it was in agreement with them. The
reported success rate was 75%, which the authors compare with the performance of their
own implementation of Hobbs (1977), which yielded 49.1%. They also compare the result

with their own systems performance on resolution of overt pronouns, which scored 82.2%
84% accuracy, and conclude that its poorer performance on zeros is mainly due to the
lack of gender information on zero pronouns.

127
1.2 Centering Theory and Pronoun Resolution

In this section, we review a few studies that cast Centering Theory in the domain of pronoun
resolution. Centering theory was not originally intended as a model for solving the problem

of pronoun reference but rather a model for local coherence. However, if one makes an
assumption about a given discourse that it was produced by a rational speaker following
the centering constraints and rules with the aim of maximizing local coherence, then the
theory can be reconstructed to be used as a model for interpreting the pronouns contained

in the discourse.

1.2.1 Brennan, Friedman and Pollard (1987)

Brennan, Friedman and Pollard (1987) was the first study that applied the theory of center-
ing to pronoun resolution. Their anaphora resolution algorithm consists of three basic steps

summarized as follows:

(141) 1. Generate all possible Cb-Cf combinations.

2. Filter combinations by constraints, e.g., contra-indexing1, centering rules and

constraints.

3. Rank remaining combinations by transition orderings.

Basically, they translate the pronounantecedent relation as the one holding between a
BACKWARD - LOOKING CENTER and a FORWARD - LOOKING CENTER that is coreferential
with it. Once all possible pairs of Cb-Cf are extracted, the algorithm mostly uses rules and
constraints of Centering Theory to first rule out some candidates, and then rank them based

on what kind of transition type they constitute.

1
Contra-indexing is a constraint which specifies the impossibility of coreference of the two pronouns in
configurations such as He likes him.

128
 Walker (1989) conducted a comparison study of this algorithm and Hobbss (1977)
anaphora resolution algorithm, using hand-evaluation on three kinds of small corpora. The
outcome was that Brennan, Friedman and Pollard (1987) and Hobbs (1977) performed

about the same over the fictional domain of 100 utterances (90% and 88% respectively), and
Hobbs outperformed Brennan, Friedman and Pollard (BFP henceforth) over the domains
consisting of newspaper articles (89% to 79%) and tasks (51% to 49%). Walker (1989)
observes that Hobbss outperformance on the task-oriented domain is mainly due to the
bigger role that global focus plays in this kind of genre; Centering does not provide a

means to explicitly deal with global focus. She proposes an extended version of the BFP
algorithm which gives more weight to global focus, and notes that the extended version
now outperforms Hobbs.

1.2.2 Strubes (1998) S-list Approach

The Brennan, Friedman and Pollard (1987) approach was more or less a direct recast of
Centering Theory as a model of anaphora resolution: it had several problems, as noted by
many, including its difficulty in dealing with incrementality, unsure status of intra-sentential
coreference, as well as computational overhead. Strubes (1998) S-list approach was

partially motivated by these issues.

Strubes model replaces Cf-list and Cb with what he calls the S-list (salience list),
which describes the attentional state of the hearer at any given point in processing a dis-
course. This definition is quite similar to that of a Cf-list at first glance. However, the two
differ in ranking and composition. First, the S-list can contain elements from current utter-

ance as well as the previous utterance, a move intended to target intra-sentential corefer-
ence. Second, unlike Cf-list in the original Centering Theory which relied on grammatical
role as the criteria for its ordering, the S-list borrows notions from theories of information
structure, such as hearer-old/hearer-new status of discourse entities from Princes (1981)

129
familiarity scale and information status in Prince (1992), as well as surface order.
The elements of the S-list belong to three information sets: hearer-old discourse enti-
ties (OLD), meditated discourse entities (MED) and hearer-new discourse entities (NEW).

The three sets of expressions have their own subdivisions. OLD consists of evoked (E)
and unused entities; MED consists of inferrables (I), containing inferrables (Ic ), and an-
chored brand-new discourse intrasentential (BNa entities); NEW consists solely of brand-
new (BN) entities.
The S-list is ordered by the information status of the above sets. OLD entities are pre-

ferred to MED entities which are preferred over NEW entities. Within each set, the ordering
is by utterance and position in utterance. Basically, an entity of an utterance x is preferred
over an entity of an utterance y if x follows y. If the entities are in the same utterance, they
are ranked by position in the sentence an entity close to the beginning of the sentence is

preferred to one that is farther away. The resolution algorithm processes a text from left to
right in the following manner:

(142) 1. If a referring expression is encountered,

(a) if it is a pronoun, test the elements of the S-list in the given order until
the test succeeds;

(b) update S-list; the position of the referring expression under consideration
is determined by the S-list-ranking criteria which are used as an insertion
algorithm.

2. If the analysis of utterance U is finished, remove all discourse entities from

the S-list which are not realized in U.

In this approach, pronoun resolution amounts to a simple lookup operation in the S-

list, and the need to consult transition types is eliminated. Strube performs a small-scale
comparison test of his algorithm with the BFP algorithm on three short stories from a

130
Hemingway novel and three articles from the New York Times, totaling 576 pronouns con-
tained in 386 sentences. His method performed better than BFP modified with Kameyamas
(1998) intra-sentential centering scheme, at over 85% correctness compared to the latter al-

gorithms 76%.

1.2.3 Left-Right Centering by Tetreault (1999)

Left-Right Centering (Tetreault, 1999) builds upon Centering Theorys constraints and
rules with some modifications that aim to improve the theorys handling of incremental
processing of pronouns. One of the main criticisms of the BFP algorithm is its lack of pro-

vision for incremental resolution of pronouns, which limits it as a cognitive model (Kehler,
1997). Psycholinguistic research supports the claim that listeners process utterances one
word at a time, so when a listener hears a pronoun he/she will try to resolve it immediately.
Only when new information is later presented which makes the resolution improbable, the

listener will back-track and find a correct antecedent.

The Left-Right Centering algorithm works by first searching for an antecedent in the
current utterance. If one is not found, then the previous Cf-lists (starting with the closest
utterance) are searched left-to-right for an antecedent. The algorithm is summarized as

below:

(143) 1. Preprocessing- from previous utterance: Cb(Un1 ) and Cf(Un1 ) are avail-
able.

2. Process Utterance- parse and extract incrementally from Un all references to

discourse entities. For each pronoun do:

(a) Search for an antecedent intrasententially in Cf-partial(Un )2 that meet

2
Cf-partial is a list of all processed discourse entities in Un .

131
 feature and binding constraints. If one is found proceed to the next pro-
noun with utterance. Else go to (b).

(b) Search for an antecedent intersententially in Cf(Un1 ) that meets feature

and binding constraints.

3. Create Cf- create Cf-list of Un by ranking discourse entities of Un according

to grammatical function. (Their implementation uses a left-right breadth-first
walk of the parse tree to approximate sorting by grammatical function.)

4. Identify Cb- the backward-looking center is the most highly ranked entity
from Cf(Un1 ) realized in Cf(Un ).

5. Identify Transition- with the Cb and Cf resolved, use the criteria from BFP
1987 to assign the transition.

Tetreault (2001) conducts a comparison study on 4 pronoun resolution algorithms, in-

cluding the three centering-based ones reviewed above: Hobbss (1978), Brennan, Fried-
man and Pollard (1987), Strubes (1998) S-list approach (henceforth S-list), his own Left-
Right Centering (1999: henceforth LRC). Two domains of text were used, both of them

taken from the Penn Treebank corpus: 195 newspaper articles originally appeared on the
New York Times containing 3900 utterances, and three fictional texts containing 553 ut-
terances. The sentences were annotated, first set by Ge, Hale and Charniak (1998) and the
second by the author, with reference information for testing purposes, in addition to the

syntactic annotation that is part of the Penn Treebank annotation.

Some decisions on the evaluation method as well as modifications on some of the algo-
rithms were made in order to compare the algorithms on equal footing, as the algorithms
differ in details of their operations. Conjoined sentences were broken up into their sub-
utterances following Kameyamas (1998) approach, but otherwise complex sentences were

kept as one unit. Following Strube and Hahns (1999) evaluation methods, no world knowl-

132
edge was assumed, and only agreement criteria (gender, number) as well as binding criteria
were applied. Crucially, the notion of discourse segment, defined as a paragraph in some
studies (Walker, 1989) was discarded: instead, each article was considered as constituting

a whole discourse.
The BFP algorithm had to be modified slightly to compensate for underspecifications
in its intrasentential resolution. The original S-list approach incorporates both semantics
and syntax in its familiarity ranking scheme; a shallow version which only uses syntax was
implemented in the study, which means that inferrables are not represented and entities re-

mentioned as NPs may be underrepresented in the ranking. Both S-list and BFP algorithms
were modified so that they have the ability to look back through all past Cf/S-lists, so that
they have the same amount of search space as the LRC and Hobbs. Hobbs (1977) makes
use of selectional constraints to help redefine the search space for neutral pronouns such as

it, but they were not used in the study due to unavailability of such semantic information.
The evaluation results are presented in Table (1.1) and Table (1.2). On both domains,
LRC-F was the best performer overall. LRC-F is a version of LRC, with a modified Cf-list
ranking, which is still based on grammatical function but with a special ordering stipulation

that involves moving entities in a prepended phrase to the back of the Cf-list. BFP was out-
performed by the two other centering-based algorithm and the naive algorithm by Hobbs,
which is a consistent result with what was reported in other previous studies (Strube, 1997).
S-list performed worse than LRC on NYT corpus, but it fared better than LRC on the novel
texts. This is due to the high density of pronouns in the novel text, which rewards S-lists

strong preference towards pronouns in the Cf-ranking. S-list ranks pronouns higher in its
salience list since they are hearer-old.
The amount of difference in performance between the LRC-F algorithm and others is
quite noticeable. The key difference that sets the LRC-F algorithm apart from others is that

it takes into consideration the clausal hierarchy within a sentence in the ranking of Cf-list.

133
 Algorithm Right %Right %Right Intra %Right Inter
BFP 1004 59.04 75.1 48.0
S-list 1211 71.7 74.1 67.5
Hobbs 1298 76.8 74.2 82.0
LRC 1268 74.9 72.0 82.0
LRC-F 1362 80.4 77.7 87.3

Table 1.1: Pronoun resolution algorithms for New York Times

Algorithm Right %Right %Right Intra %Right Inter

BFP 241 46.4 81.8 43.8
S-list 337 74.2 84.4 56.5
LRC 372 72.1 84.3 64.2
LRC-F 420 81.1 86.0 76.2

Table 1.2: Pronoun resolution algorithms for fictional texts

LRC-Fs implementation of Cf-ranking is based on the grammatical role and the surface
order, but with a modification called prepended movement: it is an operation that takes

all entities in the Cf-list that are in prepended phrases, i.e., non-subject surface initial
positions, behind those that are contained in the main clause. This has the effect of favoring
the entities located in the main clause over the ones that precede them in surface order but
are embedded more deeply, while preserving preferences for entities in main-clause and

sentence-initial subject positions. It achieves the same effect as looking in the main clause
before the intersentential search. Tetreault (2001) concludes that better pronoun resolution
can be achieved by marking entities in prepended phrases as less salient.

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

Thai is another pro-drop language, and it can have a zero subject in declarative sentences,

as well as zero objects. Aroonmanakun (1997, 1999, 2000) proposes a model of extended
Centering that incorporates discourse structure for resolution of Thai zero pronouns. The
motivation behind this extension is Centering Theorys inadequacy in handling antecedent-

134
pronoun relations which span non-adjacent sentences. He claims that in many of these
cases the two sentences, although not next to each other in linear order, can be rendered
as belonging in adjacent segments when hierarchical discourse structure is imposed on the

text.
He incorporates the notion of discourse structure as put forth by Rhetorical Structure
Theory (Mann and Thomson, 1987; Fox, 1987) to extend Centering Theory along the fol-
lowing lines. Given that a discourse is analyzed as a hierarchical structure of discourse
units (Ui ), each discourse unit has one BACKWARD - LOOKING CENTER (Cb) and a set of

FORWARD - LOOKING CENTERs (Cf). A discourse unit can be either a single utterance or
multiple utterances. Cf is an ordered list of discourse entities realized in that unit. The Con-
straints and Rules in Centering Theory remain largely the same, with the unit U redefined
as discourse unit rather than utterance:

(144) a. Constraints:
For each discourse unit Ui in a discourse:

1. There is precisely one backward-looking center Cb.

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

3. The center, Cb(Ui ) is the highest-ranked element of Cf(Ui1 ) that is real-

ized in Ui .

b. Rule 1: If any element of Cf(Ui ) is realized by a (zero) pronoun in Ui+1 , then

the Cb(Ui+1 ) must be realized by a (zero) pronoun also.

Rule 2: Sequences of continuation are preferred over sequences of retaining;

and sequences of retaining are to be preferred over sequences of shifting.

The notions of precedence and left adjacency are also redefined to suit discourse unit
as the new unit of computation:

135
(145) Precedence
Ui1 precedes unit of Ui iff either

1. Ui1 is the left adjacent unit of Ui .

2. Ui1 is the left adjacent unit of Uk and Ui is the left-most unit under Uk .

Left adjacency

Ui is the left adjacent unit of Uj if Ui and Uj has the same parent, Uk , and there is
no other unit between Ui and Uj , and Ui occurs before Uj .

For the ranking of Cfs in Thai, Aroonmanakun (2000) follows the standard approach in

Centering Theory of using grammatical relations. He postulates the following ordering in

Cf similar to that proposed for Japanese:

(146) Cf ranking for Thai by Aroonmanakun (2000):

topic > subject > object(s) > others

Aroonmanakun (2000) conducts a comparative study of the original Centering algo-

rithm and his own extended version, using naturally occurring Thai data, with the aim of
verifying whether or not hierarchical structure of discourse contributes to zero pronoun
resolution in Thai. A total of 719 zero pronouns were identified in twenty texts consisting
of 15,949 words. The corpus was analyzed by twelve native speakers of Thai with regard
to discourse structure and coreference. The results of the test indicate that the extended

centering algorithm can indeed resolve more instances of zero pronouns than the existing
centering algorithm: out of the 159 cases where the existing algorithm failed, the extended
version was able to successfully resolve 42 instances.
However, the author correctly notes that the success of the extended algorithm on this

simple comparison is owed to the fact that it has a wider scope of reference resolution than
the existing centering algorithm: the search scope for antecedents is defined as the preced-
ing discourse unit in this theory, which has a larger scope than a single utterance. When

136
search spaces for either algorithm were adjusted to be comparable in size, no significant
difference in the performance was found. He speculates that the extended centering did not
perform better because most of the antecedents of zero pronouns in his corpus were found

in the immediately preceding utterance: 87.32% of all zero pronoun instances had an an-
tecedent in the immediately preceding utterance; 7.89% had an antecedent two utterances
ahead; only 3.11% had an antecedent that was more than two utterances away. Because of
this pattern, he concludes, the existing centering algorithm can resolve most zero pronouns
in naturally occurring text, but the number of examples used in the study is too small to

support a strong claim.

1.3 Optimality Approaches to Anaphora Resolution

Previously in Section (1.1.3), we showed how optimality-theoretic approaches were adopted

in accounting for cross-linguistic distribution patterns of zero pronouns (Grimshaw and

Samek-Lodovici, 1998; Speas, 1991). The basic architecture of Optimality Theory, which
views surface-realization of a linguistic representation as a best candidate which is cho-
sen from a pool of all potential alternatives, rather naturally translates to the task of anaphora
resolution, which also involves a process of selecting a right antecedent among many po-

tential candidates. We review here two works which actively cast OT as the theoretical
framework for anaphora resolution. Beaver (2002) is an attempt to recast the Centering-
based anaphora resolution algorithm put forth in Brennan, Friedman and Pollard (1987,
henceforth BFP), reviewed in Section 1.2.1 as an OT-based anaphora resolution system.

Kim (2003) proposes a set of constraints and rankings for interpretation of anaphoric zero
pronouns in Korean. It should be noted that unlike the systems reviewed so far these two
works do not in fact implement a working system or empirically test their theories on
naturally occurring data. Instead, as with most work done within the linguistic-theoretic

137
tradition, they build and validate the details of their theories upon carefully constructed
data, the result of which serve as the linguistic basis for an anaphora resolution scheme.

1.3.1 Beavers (2002) Centering in OT

Beaver (2002) proposes an OT-based model of anaphora resolution inspired by Centering

Theory, which he dubs Centering in OT (COT). While descriptively equivalent to the

BFP model, his COT model is stated declaratively unlike the inherently procedural nature
of the BFP algorithm, and makes clearer the status of the various constraints used in the
theory. It should be noted that its connection with the original Centering Theory is only
indirect, as it draws on the work of BFP, which in itself is an adaptation of Centering

Theory. In his reformulation, the first step Generate of the three central steps of the
BFP algorithm Generate-Filter-Rank (shown in example 141) naturally corresponds to OTs
GEN, the function/algorithm that creates the candidate set, in this case the set of possible
antecedents for a given pronoun. What Filter and Rank achieved in the BFP model is

replaced by a set of six violable OT constraints:

(147) a. AGREE Anaphoric expressions agree with their antecedents in terms of number

and gender.

b. D ISJOINT Co-arguments of a predicate are disjoint.

c. P RO -TOP The topic is pronominalized.

d. FAM -D EF Each definite NP is familiar. This means both that the referent is
familiar, and that no new information about the referent is provided by the
definite.

e. C OHERE The topic of the current sentence is the topic of the previous one.

f. A LIGN The topic is in subject position.

which are ranked as follows:

138
(148) Ranking of COT constraints by Beaver (2002):
AGREE, D ISJOINT > P RO -TOP > FAM -D EF > C OHERE > A LIGN

The relative ranking of constraints loosely reflect the temporal precedence in the BFP
algorithm: AGREE, D ISJOINT, P RO -D ROP correspond to the Filter stage, and C OHERE
and A LIGN the Rank stage.
Among the six, FAM -D EF is the only COT constraint which does not correspond to a

BFP constraint. The first two constraints AGREE and D ISJOINT encode the Filter step
of the BFP algorithm: they prune out antecedent candidates which produce grammati-
cal violations, including agreement violations and the Binding Theory Principle B viola-
tions. P RO -TOP has essentially the effect of Centerings Rule 1 (see example 28) which

postulates: if there are pronouns in the current sentence, then one of them refers to the
BACKWARD - LOOKING CENTER of the current sentence. The two remaining constraints
C OHERE and A LIGN jointly encode the final Rank step of BFP, which ranks remaining
candidates according to preferred transition types, with the terminology topic replacing the

role played by BACKWARD - LOOKING CENTER in Centering Theory.

His model of Centering in OT (COT), which consists of the set of six constraints
along with their relative ranking, is predictively equivalent to the BFP centering model in
a strong, formal sense, he claims. Having filtered out resolutions that for some reason are
considered improper, the final resolution preference in BFP comes down to a competition

between alternative transitions, which are ordered in terms of preference: CONTINUE, RE -

TAIN , SMOOTH - SHIFT, ROUGH - SHIFT . Having C OHERE outrank A LIGN gives the desired
effect of the preference, as the tableau in (1.3) illustrates.
In further discussions, Beaver (2002) suggests the possibility of different languages

employing different sets of constraints suited for the language-specific mechanisms of en-
coding topic and salience. For Japanese where NPs marked as empathetic and NPs marked
with the morpheme wa- are highly salient, the two extra language-specific constraints might

139
 Transition C OHERE A LIGN
CONTINUE
RETAIN *
CONTINUE
SMOOTH SHIFT *
CONTINUE
ROUGH SHIFT * *
RETAIN *
SMOOTH SHIFT *
RETAIN *
ROUGH SHIFT * *
SMOOTH SHIFT *
ROUGH SHIFT * *

Table 1.3: C OHERE and A LIGN produce preference ranking of 4 centering transition types

be relevant:

(149) a. S ALIENT WA: If in the previous sentence discourse entity was realized by
a wa-marked form, and discourse entity was also realized in that sentence,
then is more salient than .

b. S ALIENT E MPATHY: If in the previous sentence discourse entity was marked

as empathetic, and discourse entity was not, then is more salient than .

He also suggests two possibly universal constraints which are related to salience:

(150) a. S ALIENT F ORM: If in the previous sentence discourse entity was realized by
a more minimal form than discourse entity , then is more salient than .

b. R ECENCY: One discourse entity is more salient than another if it was referred

to in a later clause.

140
1.3.2 Hong (2002) and Kim (2003): OT-Based Korean Anaphora Res-

olution

Hong (2002) and Kim (2003) respectively propose OT-based resolution models of Korean,
drawing from previous OT theories of zero pronouns and anaphora resolution. Five univer-
sal constraints are employed in Hongs (2002) work:

(151) 1. *F EATURE M ISMATCH: A linguistic element must match its licensor in (phono-

logical, morphological, syntactic and semantic) features.

2. P RINCIPLE B: If arguments of the same semantic relation are not marked as

being identical, interpret them as being distinct.

3. DOAP(Dont Overlook Anaphoric Possibilities): Opportunities to anaphorize

text must be seized.

4. C ONTINUE: The center of a previous utterance continues as the center in the

current sentence.

5. PARALLELISM: As the antecedent of a anaphoric expression, choose a (logi-

cally, structurally, or thematically) parallel element from the preceding clause.

which are ranked as follows:

(152) *F EATURE M ISMATCH > P RINCIPLE B > DOAP > C ONTINUE > PARALLELISM

*F EATURE M ISMATCH is analogous to Beavers (2002) AGREE in that it rules out

antecedent candidates which are grammatically or semantically incompatible with the pro-

noun in question. P RINCIPLE B is a familiar constraint also utilized in Speas (1997, 2001),
Hendriks and de Hoop (2001), and also in Beaver (2002) as the D ISJOINT constraint. PAR -
ALLELISM was first proposed in Hendriks and de Hoop (2001) as a constraint which pro-
motes candidates that exhibit structure parallel to the pronouns, and was extended by Hong

141
(2002) to cover logical and thematic parallelism. DOAP, also initially proposed by Hen-
driks and de Hoop (2001), is needed to capture referential candidates that are not textually
represented, such as discourse participants and generic referents. C ONTINUE obviously

encodes the preference for the CONTINUE transition in Centering Theory.

Most of the constraints employed in Kims (2003) work reflect similar insights. Her
formulation departs from that of Hongs (2002) in including two novel constraints R EL -
EVANCE P RINCIPLE and R ECENCY P REFERENCE C ONSTRUAL , and also in the relative
ranking of C ONTINUE below PARALLELISM. The seven constraints are as follows:

(153) a. R ELEVANCE PRINCIPLE (BR):

Be Relevant.

b. T HETA - ROLE M ATCH (TM):

The argument must match with semantic selectional restriction required by its
predicate.

c. M ORPHOLOGICAL M ATCH (MM):

The argument must match with a specific morpheme which plays the role of
the AGREE feature.

d. D ISJOINT I NTERPRETATION (DI):

The arguments subcategorized by the same predicate must not be coreferential

with each other.

e. R ECENCY P REFERENCE C ONSTRUAL (RPC):

The entities realized in an immediate preceding utterance are the most promi-
nent.

f. F UNCTIONAL PARALLELISM (FP):

The parallelism in the grammatical, the thematic and the functional roles.

142
 g. C ENTER -C ONTINUE (CC):
The Cp(Un-1) is realized to Cb(Un) in the rule of center-establishment of
Centering.

which has the following ranking:

(154) BR > TM > MM > DI > RPC > FP > CC

Most notable in her set of constraints is the highest-ranked R ELEVANCE P RINCIPLE,

whose motivation can be found in Grices theorem (1985) and Sperber and Wilsons (1986)
Relevance Theory. Kim (2003) cites this constraint as the most important in anaphora res-
olution in Korean which outranks all others, and gives the following contrasting examples

(155) and (156) to support her claim:

(155) u1 .
%

H
 Qj
# ] =
o
^ 1
W \ t

Z Q 

y .
Yengswu-nun ecey Chelswu ney-ey nol-le ka-ess-ta.
Yengswu-Top yesterday Cheslswu home-to play-End go-Past-Dec.
Yengswu went to Chelswus to play yesterday.

u2 .
Q
\
|
9 O
3
\ %
 .
kulena cip-ey eps-ess-ta.
but (SBJ) house-at not-in-Past-Dec.
But (he) was not home.

The key factor in anaphoric interpretation of the zero subject in the second utterance is the
connective
Q

(/kulena/, but). It implies that an expected outcome was not realized;
=
o
the party who is expected to be home in this case is ^ (Chelswu), which yields the
only relevant interpretation. Compare it with a parallel example below, only with a different

connective this time:

(156) u1 .
%

H
 Qj
# ] =
o
^ 1
W \ t

Z Q 

y .
Yengswu-nun ecey Chelswu ney-ey nol-le ka-ess-ta.
Yengswu-Top yesterday Cheslswu home-to play-End go-Past-Dec.
Yengswu went to Chelswus to play yesterday.

143
 u2 .
Af " \
|
9 O
3
\ %
 .
kulayse cip-ey eps-ess-ta.
therefore (SBJ) house-at not-in-Past-Dec.
Therefore (he) was not home.

In this case,
Af
" (/kulayse/, therefore) expresses causality between the two utterances.
Yengswus visit to Chelswus home can only reasonably cause Yengswus not being home
and not Chelswus absence. The R ELEVANCE P RINCIPLE constraint rules in favor of the
interpretation of zero pronoun as referring to Yengswu.

The obvious problem with the R ELEVANCE P RINCIPLE is that it is given an all-trumping
power yet cannot be formulated in precise, measurable terms. With its presence, it is un-
clear why all other constraints are even needed: the ability to judge whether or not a certain
coreference assignment results in a relevant, or coherent, interpretation is only one step

away from being able to evaluate the relative degrees of such relevance, which then can be
used to single out the optimal antecedent. At the very least, the constraint renders Kims
(2003) OT-based resolution system unimplementable, since it assumes world-knowledge
as well as a sophisticated inference system based on it.

144
Chapter 2

The Data: The Penn Korean Treebank

Corpora

2.1 Overview of the Corpora

Two sources of annotated corpora are used in building and testing models for Korean zero

pronoun reference resolution: The Korean English Treebank Annotations Corpus (Palmer
et al., 2002; Han et al., 2002) and The Penn Korean Treebank Version 2.0 (Han et al., to
be published). The former consists of Korean and English bilingual texts extracted from
military training manuals; it is a relatively small corpus of 5,083 sentences and 54,071
words1 . The latter is a second volume of the corpus, and it contains texts from the news

domain. It is significantly larger, consisting of 647 news articles in 112 files which contain
132,040 words and 5,010 sentences. (From this point on, the two will be referred to as the
Penn Korean Treebank 1 and the Penn Korean Treebank 2, or even shorter as KTB 1 and
1
Word counts for the Penn Korean Treebank 1 and 2 were measured post-tokenization: the numbers
therefore include symbols and punctuation marks that have been tokenized apart from the words they were
attached to in the raw text.

145
KTB 2, respectively.)
The two corpora differ sharply in genre and style. The Penn Korean Treebank 1 is in
the military domain and largely consists of dialogue sequences between military personnel

depicting various military strategies and intelligence gathering activities. The sentences
are relatively short at about 10.6 words long, and the dialogues contain many question-and-
answer pairs in colloquial style. The Penn Korean Treebank 2, on the other hand, is in the
news domain and contains short news reports on world events and politics. Its sentences
are more formal, and generally longer at about 26.4 words per sentence.

In the corpora, zero pronouns are pre-identified and denoted by the string *pro*. To-
gether, they contain 12,522 instances of zero pronouns: 4,361 in the Penn Korean Treebank
1 and 8,161 in the Penn Korean Treebank 2.2 The relative frequencies of zero pronouns in
the two corpora are charted in Table 2.1. The Penn Korean Treebank 1 shows a higher

ratio of zero pronoun instances per word. The Penn Korean Treebank 2 exhibits higher fre-
quency of zero pronouns measured per sentence, but its sentences are 2.5 times longer than
those of KTB 1. This frequency pattern confirms the widely-held belief that zero pronoun
use is more prevalent in speech than in written texts.

Some parts of the Penn Korean Treebank 1 are isolated sentences which do not form
coherent discourse segments: while some of the zero pronouns used in these sentences have
intra-sentential antecedents, they cannot be used with the general scheme of coreference
resolution and were excluded from this study. The figures in the third row were obtained
exclusively from those zero pronouns found in discourse blocks.

At one pro per every 12 or so words (KTB 1) or one pro per every 16 words (KTB 2),
zero pronouns in Korean are a remarkably frequent phenomenon. This is partly owing to
2
Excluding 134 and 26 *pro* tokens for KTB 1 and KTB 2 respectively, which were determined during
our reference annotation process as not genuine cases of zero pronouns but rather errors made in the Treebank
annotation process.

146
 source words sentences pro pro/word pro/sentence
KTB 1 (all) 54,503 5,083 4,361 0.080 0.858
KTB 1 (in discourse) 28,607 2,689 2,468 0.086 0.918
KTB 2 132,040 5,010 8,161 0.062 1.63

Table 2.1: Zero pronoun frequencies in KTB 1 and KTB 2

the fact that they encompass a variety of distinct usages. To put this fact in a cross-linguistic
perspective, we examined Chinese data for comparison. The Penn Chinese Treebank 5.1

corpus distinguishes PRO in uninflected clauses in addition to pro in inflected clauses. Even
with the two classes combined, the corpus showcases a much lower frequency of either pro
or PRO per every 28 words or so (Table 2.2). While these estimates themselves cannot be
taken at face value due to language-specific details that are sure to abound, it nevertheless
supports the conclusion that Korean pros are indeed a particularly prolific phenomenon.

(For a discussion of how Korean pro corresponds to both PRO and pro in other languages,
see Sections 2.4.3 and 1.1.1.)

source words pro PRO pro+PRO pro+PRO/word

Chinese Treebank 5.1 507,000 7,481 10,950 18,431 0.036

Table 2.2: pro and PRO frequencies in the Penn Chinese Treebank 5.1

In terms of grammatical role, subject zero pronouns account for the vast majority of
zero pronoun cases, followed by object zero pronouns and then others. Notably, KTB

1 exhibits 1.7 times the frequency of object zero pronouns compared to that of KTB 2
(5.61% vs. 3.33%).

source subject (%) object (%) other (%) total

KTB 1 4,074 (93.4%) 245 (5.61%) 42 (0.96%) 4,361
KTB 2 7,820 (95.8%) 273 (3.33%) 69 (0.85%) 8,161

Table 2.3: Zero pronoun frequencies by grammatical roles

147
 The corpora contain rich linguistic annotation, including part-of-speech tags and syn-
tactic structures. Empty elements such as zero pronouns and NP traces are already identi-
fied in their proper location (see Sections 2.3 and 2.2.1). On top of the existing annotation,

zero pronouns and their antecedents were annotated according to the annotation scheme de-
scribed earlier in Section 2.2.3. KTB 1 was annotated by the author; KTB 2 was annotated
by a third party.

source anaphoric propo- deictic generic expletive total

sitional i y w
KTB 1 1,251 68 588 304 30 1,630 490 4,361
(28.7%) (1.55%) (13.5%) (6.97%) (0.68%) (37.4%) (11.2%) (100%)
KTB 2 5,270 18 267 19 56 513 2018 8,161
(64.6%) (0.22%) (3.27%) (0.23%) (0.68%) (6.28%) (24.7%) (100%)

Table 2.4: Zero pronoun frequencies by type

Table 2.4 shows frequencies of zero pronouns by their types. The distribution patterns
of the two corpora exhibit sharp differences across all types: the percentage of anaphoric

zeros in KTB 2 (64.6%) is more than double the percentage of anaphoric zeros in KTB 1
(28.7%); all other types, including deictic, generic and propositional pronouns, show much
lower frequencies in KTB 2, with the exception of expletive zeros which are found almost
2.5 times more frequently in KTB 2 compared to KTB 1. The striking contrast can be
attributed to the different genres of the two texts, as well as some text-specific peculiari-

ties of the Penn Korean Treebank 1 data. As noted earlier, the Penn Korean Treebank 1
consists of dialogue sequences, which explains the overwhelming presence of deictic zero
pronouns (21.1% in total): references to discourse participants, including the speaker and
the hearer, are prevalent throughout. The Penn Korean Treebank 2, on the other hand, is a

collection of news articles: references to discourse participants are rarer (4.3% total for all
three types), the majority of which are found inside quoted speech. The higher frequency
of expletive zero pronouns in KTB 2 can also be attributed to genre differences: the news

148
text of KTB 2 includes a large amount of instances of preposition-like verbal elements
meaning regarding and according to (shown earlier in 56), which are associated with
a more technical style of writing. The high frequencies of propositional zeros and generic

zeros, on the other hand, seem to be the result of the specifics of the content of the KTB
1 text. The dialogues consist largely of interrogation sequences, which produced frequent
references to propositional contents or events contained in a clausal unit, e.g., Is (thatp )
true?. Similarly, the interrogation sequences included many types of questions and an-
swers regarding actions and logistics of various military units, which are often expressed

with scene-setting adverbials referring to a military unit which are then followed by generic
zero pronoun references, e.g., In battalion 1, (theyg ) do....

singular plural
1st person non-polite
/na/
 45 
o /uli/ 221
polite  /ce/
$ 433 
$ B([ t) /cehi(tul)/
366
singular plural
2nd person non-polite - /ne/
 118 -
 B([ t) /nehi(tul)/
5
1
 W /caney/ 1

{

 /tangsin/ 1

) a /kwikwan/
' 100
polite
{

 /tangsin/ 4

Table 2.5: Overt pronouns in KTB 1, 1st and 2nd person

proximal medial distal

3rd person human non-polite /ku/
6
[
s t /itul/ 2
[t /kutul/ 18

1
W /kuney/ 21
polite
r /kupwun/ 2

object s /i/ 18 /ku/ 20

s
/ikes/ 50 
/kukes/ 197
place 
# l /yeki/ 53  l /keki/ 58 $
 l /ceki/ 2
other (wh-expressions, quantified expressions) 687

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

149
 singular plural
1st person non-polite
/na/
 41 
o /uli/ 130
polite
singular plural
2nd person non-polite
polite
{

 /tangsin/ 4 Q
# 
r /yelepwun/ 4


Table 2.7: Overt pronouns in KTB 2, 1st and 2nd person

proximal medial distal

3rd person human non- /ku/
274
polite .
 /kunye/ 17
[
s t /itul/ 116
[t /kutul/
9 $
[ t /cetul/ 1

polite
r /kupwun/ 2

object s /i/ 369 /ku/ 16

s
/ikes/ 19 
/kukes/ 6
place 
# l /yeki/ 9  l /keki/ 2
other (wh-expressions, quantified expressions) 36

Table 2.8: Overt pronouns in KTB 2, 3rd person and other

Compared to the prevalence of zero pronouns, a relatively small number of overt pro-
nouns are found in the two Korean Treebank corpora. KTB 1 has a total of 2,428 overt
pronoun occurrences, which is less than half of the 5,083 zero pronoun count; KTB 2 has

a smaller total number of 1,055 overt pronouns, which amounts to about 20% of its 5,010
zero pronoun count. Tables (2.5 2.8) detail the composition of the pronouns.
There are a few key differences in the distribution patterns across the two corpora.
First, KTB 1 exhibits higher per-word frequency of overt pronouns than KTB 2. Second,
first and second person pronouns make up the majority of the overt pronouns in KTB 1

(Table 2.5), while in KTB 2 they account for a far smaller proportion (Table 2.7). The
dialogue genre of KTB 1, we believe, is largely responsible for the first two points. Third,
substantial numbers of the two standard Korean 3rd person overt pronouns
(/ku/, he)
and
. (/kunye/, she) are observed in KTB 2. As discussed in Section 2.2, Part I, these

150
overt pronouns are mostly confined to formal texts: their prevalence in KTB 2, which is a
news corpus and sparsity in the dialogue-based KTB 2 conform to the point. Also, a large
number of s
 /i/ this is observed in KTB 2, a pronoun with a highly specialized function

of discourse deixis (see Example 48c, Part I). Again, this is a characteristic of the text genre
of KTB 2.3

2.2 Annotating the Data

The two corpora were annotated using the proposed classification system for Korean zero

pronouns. Zero pronouns were identified and categorized in the text, with their potential
antecedents identified where applicable. Some issues arose in the course of studying the
corpus, which were reflected in designing the annotation scheme used for this study.

2.2.1 Zero Pronouns with a Intra-sentential Antecedent

Having the zero pronoun as a valid construct in a syntactic theory offers conflicting takes on

two or more predicates that seemingly share an argument. A sentence with two predicates
which share one subject can receive two distinct structural analyses: one with VP coordi-
nation with a shared subject or another with two coordinated clauses with a zero pronoun,
coindexed with the overt subject, posited as the subject in one of them.

(157) Syntactic Structures for We had dinner and came back.

a. VP coordination: shared NP subject

3
There are a few additional minor yet interesting observations to be made about the overt pronouns in the
corpora. In KTB 1, {


 /tangsin/, you, a polite pronoun by standard account, is seen used in an impolite
manner, as previously noted in Section 2.2, Part I. Also, a highly domain-specific second person pronoun )

a /kwikwan/, dear-officer occurs with an unusual frequency in KTB 1, a corpus in the military domain.
'

151
 (S (NP-SBJ /NPN+/PAU)

wuli nun

we Top

(VP (VP (NP-OBJ /NNC+/PCA)

cenyek ul

dinner Acc

/VV+/ECS)

mek ko

eat And

(VP /VV+/EPF+/EFN))

tolao ass ta

come-back Past Dec

./SFN)

b. S coordination: zero pronoun represented

(S (S (NP-SBJ /NPN+/PAU)

wuli nun

we Top

(VP (NP-OBJ /NNC+/PCA)

cenyek ul

dinner Acc

/VV+/ECS))

mek ko

eat And

(S (NP-SBJ *pro*)

zero-pronoun

(VP /VV+/EPF+/EFN))

tolao ass ta

152
 come-back Past Dec

./SFN)

The Penn Korean Treebank corpora offer strictly defined guidelines (Han et al., 2001;
Han et al., 2005) for annotating such constructions. It allows for coordination at all levels:
VV (verb part-of-speech level), VP (verb-phrase level) and S (clause level), whilst recom-

mending coordination at the lowest level possible for a given sentence. However, only two
morphemes out of a large number of Korean connective verbal endings are recognized as
/-ko/ and -
capable of heading a VP coordination structure: - 
/-ken/, standard con-
junctive and disjunctive connectives in Korean. Clauses headed by all other connectives,

if the option of VV coordination is unavailable, will have to be analyzed as involving an

S coordination structure and, therefore, a zero pronoun in the subject position in one of
the clauses. This annotation decision partly contributed to a large number of subject zero
pronouns in a coordinated S structure with intra-sentential coreference.

Still a structural ambiguity remains even after two full clausal units are assigned to such
sentences: it is yet to be decided which predicate has an overt subject and which one a zero
subject. The decision will lead to two distinct syntactic trees: one with a coordinated S
structure and the other with a subordinate S modifying the main clause VP, as illustrated in

f
the following example (with a slightly different connective - " (/-kose/, and then) this
time):
In addition to the coordinated S structure where the overt subject is associated with the
first predicate of the sequence, associating the overt subject with the final VP will place the
zero subject within an intervening subordinate clause which modifies the main clause VP. 4

(158) Syntactic Structures for We had dinner and came back.

a. zero pronoun in main clause

4
There is an obvious third possibility, where the zero pronoun linearly precedes the overt NP subject. As
logical as it may be, this option is highly unintuitive and therefore is not considered here.

153
 (S (S (NP-SBJ /NPN+/PAU)

wuli nun

we Top

(VP (NP-OBJ /NNC+/PCA)

cenyek ul

dinner Acc

/VV+/ECS))

mek kose

eat And-Then

(S (NP-SBJ *pro*)

zero-pronoun

(VP /VV+/EPF+/EFN))

tolao ass ta

come-back Past Dec

./SFN)

b. zero pronoun in subordinate clause

(S (NP-SBJ /NPN+/PAU)

wuli nun

we Top

(VP (S (NP-SBJ *pro*)

zero-pronoun

(VP (NP-OBJ /NNC+/PCA)

cenyek ul

dinner Acc

/VV+/ECS))

mek kose

eat And-Then

154
 (VP /VV+/EPF+/EFN))

tolao ass ta

come-back Past Dec

./SFN)

One important implication of these two distinct structures is the relative placement of
the overt NP and the zero pronoun in terms of clausal depth. Given that the final verb in

Korean projects the matrix clause, the structure in (158a) places the overt NP subject in the
subordinate clause and the zero pronoun in the main clause, while the relation is reversed
in (158b). If one is to view the overt subject NP as the antecedent of the zero pronoun,
the analysis in (158b) will appeal to many since the antecedent is occupying a structurally
higher position than the bound zero pronoun. Adding more support to this position, clausal

depth of an argument plays a significant role in any theory which attempts to represent level
of saliency or discourse accessibility of discourse entities. From this theoretical stance, it
seems more reasonable to place the full NP subject at the main clause level for an easy back
reference in subsequent discourse.

The Guidelines for the Penn Korean Treebank recommend (Han et al., 2001) an in-
cremental parsing approach (Chapter 22.3), i.e., associating the overt NP subject with
the lefthand side (subordinate clause) VP and assigning a zero pronoun to the latter (main
clause) VP. However, it also notes that this is just a guideline to be followed when the

annotator is in doubt.

2.2.2 Determining the Categories

Once a zero pronoun is identified, the next step in corpus annotation is to determine its type,

according to the classification scheme previously presented in Table 2.4. In this section,
we discuss some issues that arise in categorization, which eventually lead to some design
decisions made for the annotation scheme.

155
 In Section 2.6, Part I, we saw that pronoun usage can be both anaphoric and deictic
(example 100). To capture this insight, we adopt an annotation scheme in which: Whenever
a zero pronoun refers to an entity that is understood to be present in the discourse scene,

mostly in first person singular/plural or a second person, it will bear a tag linking it to a
discourse entity, in addition to a number which coindexes it with an antecedent NP in the
text if present. The previously seen example (100) is annotated thusly:

(159) u1 . 
H5 i5


7

>
>K  
s H
 |
 `





nan-un5 i5 com ttokttokhay poi-nun salam-ul manna-l
I-Top5 (SBJi5 ) somewhat smart look-RelEnd person-Acc meet-Rel
:
M , i5





 m~ 
 /5
? a
> j
]
ttay-mata, i5 hangsang phwum-ko tani-ten nay5 kulim cey
time-each, (SBJi5 ) always carry-AuxEnd go-RelEnd my5 drawing number
1\ 
 /? \
r
 +>K

H
4 

.
1ho-lul kkenay-e ku-lul sihumha-e poko-n ha-ess-ta.
1-Acc pull-out-AuxEnd he-Acc test-AuxEnd try-AuxEnd do-Past-Dec.
Whenever (Ii5 ) met a smart-looking person, I5 tested him whipping out my5
drawing No.1 which (Ii5 ) always carried around.

u2 . i5 

& 
s K

4 

e H
 |



i5 ku-ka cengmal ihaylyek iss-nun salam-i-nka
(SBJi5 ) he-Nom really understanding exist-RelEnd person-Cop-If


z
3
 %
 
~ s

  .
al-ko siph-ess-ten kes-i-ta.
know-AuxEnd desire-Past-RelEnd that-Cop-Dec.
(Ii5 ) wanted to know whether he really possessed understanding.

In addition, generic indefinite zero pronouns can have a coreferential relation, most
frequently with another generic indefinite zero, as seen previously in Section 2.4.3.3. These
instances are treated similarly, with numeric indices denoting coreference relations:

(160) g7 |
s
\

 9 
 g7 \

 
 .
g7 holangi-lul cap-ulyemyen g7 san-ey ka-eyaha-nta.
(SBJg7 ) tiger-Acc catch-IntendIf (SBJg7 ) mountain-Des go-Must-PresDec.
If (oneg7 ) wishes to catch a tiger, (one/heg7 ) must go to the mountains.

156
(161) g3 g3 +

`
{
9
 '
+ 
p H



g3 g3 ha-l il-ul twi-lo milu-nun kes
(SBJg3 ) (SBJg3 ) do-Rel task-Acc later-Des postpone-RelEnd Comp
(For oneg3 ) to postpone a task (for one/himg3) to do.

2.2.3 The Annotation Scheme

Annotation of the two corpora with zero pronoun reference is done in the following man-
ner. For each zero pronoun identified within a text, all noun phrases with the same referent,
in either a full NP, a pronominal or even the null form, are identified across the text and

given a unique numerical index. All non-anaphoric types and propositional-anaphoric type
zero pronouns are marked with their category. For reasons explained in the previous sec-
tions, deictic and generic zero pronouns can also bear numerical indices, further indicating
a coreference relation between them and other NPs. For propositional-anaphoric zero pro-

nouns, their potential antecedents are not NPs but rather clausal units such as S nodes,
which were identified and marked with an index.
The following summarizes the coding scheme used in annotation:

(162) A coding system for zero pronoun reference

(i) NP-anaphoric relations are indicated by a numerical index

(ii) All other categories are marked with an alphabetic index:

deictic speaker: i

deictic hearer: y
deictic speaker and hearer: w
indefinite generic: g
indefinite specific: s*

situational: x
propositional anaphoric: p

157
 It should be noted that the indefinite specific (s) category, while present in the an-
notation scheme, did not find actual usage in the two Penn Korean Treebank corpora that
were annotated using the scheme. We are not sure whether this is due to the real absence

of applicable pro usages or the difficulty involved in distinguishing it from the closely re-
lated indefinite generic (g) category. The s category, therefore, was excluded from our
resolution models as a categorical target, as we shall see in the next section.
Example (163) shows a sample piece of coreference annotation using the scheme, and
(164) shows the exact same segment of conversation annotated within the Korean Tree-

bank file, which is pre-annotated with syntactic information. For each zero pronoun, the
coreference indices are given on its phrasal projection node. Noun phrases which are in a
coreference relation with a zero pronoun are also given a numerical index.
Note that there is a speaker turn between (163 06:02.A) and (163 06:03.B): as a result,

the entity bearing index 2 is the hearer in (163 06:02.A) and therefore is also marked with
deictic index y, whereas it bears deictic index i in (163 06:03.B) as it is the speaker of the
utterance.

(163) 06:01.A 
, ^  :

2 ,
y2

.
ca, kim thukmwucang2, y2 anc-ula.
well, kim 1SG2 , (SBJy2 ) sit-Imp.
All right, 1SG Kim2 , (youy2 ) sit down.

06:02.A y2 $3 7
- "
3
3 %

 ?
y2 mwe3 com mek-ess-na?
(SBJy2 ) what3 little eat-Past-Q?
Did (youy2 ) eat something3 ?

06:03.B V
\, i2 3 3
3
" %_
 vm
 .
yey, i2 3 mek-ess-supnita.
yes, (SBJi2 ) (OBJ3 ) eat-Past-PolDec.
Yes, (Ii2 ) ate (something3 ).

158
(164) ;;A;06:1: , , .
(S (INTJ /IJ)

ca

well

,/SCM

(NP-VOC#2 /NPR

kim

Kim

/NNC)

thukmwucang

sergeant

,/SCM

(S (NP-SBJ#y2 *pro*)

zero-pronoun

(VP /VV+/EFN))

anc ula

sit Imp

./SFN)

;;A;06:2: ?

(S (NP-SBJ#y2 *pro*)

zero-pronoun

(VP (NP-OBJ#3 /NPN)

mwe

something

(VP (ADVP /ADV)

com

little

159
 (VP /VV+/EPF+/EFN)))

mek ess na

eat Past Q

?/SFN)

;;B;06:3: , .

(S (INTJ /IJ)

yey

yes

,/SCM

(S (NP-SBJ#i2 *pro*)

zero-pronoun

(VP (NP-OBJ#3 *pro*)

zero-pronoun

/VV+/EPF+/EFN))

mek ess supnita

eat Past Polite-Dec

./SFN)

160
Chapter 3

A Rule-Based Approach: Variations of

the Hobbs Algorithm

As mentioned in earlier sections (1.1.1 and 1.2), Hobbss (1977, 1978) rule-based algorithm
for the resolution of pronominal coreference in English is widely adopted as a baseline
method in many studies on pronoun resolution in English. In this chapter, we discuss the

applicability of the Hobbs algorithm to resolution of Korean zero pronouns. By implement-

ing Hobbs, we intend to achieve two main goals. First, results from the Hobbs algorithm
establish a point of comparison between the phenomenon of Korean zero pronoun anaphora
and pronoun resolution in other languages, and highlight those aspects that are unique to

the Korean case, if any. Second, with the absence of comparable precedent studies on Ko-
rean zero pronoun resolution, the Hobbs algorithm provides a reasonable baseline to be
expected from a rule-based approach.

161
3.1 The Hobbs Algorithm

The Hobbs algorithm operates on syntactically parsed sentences, and therefore our zero
pronoun coreference annotation on the two Penn Korean Treebank corpora is readily suited

to the task. Its original formulation on English pronouns means that its application to our
study has to be limited to anaphoric (a) zero pronouns only: only those zero pronouns
that are purely anaphoric, that is, those that do not have coreference relations with any
non-anaphoric categories such as deictic or generic throughout the text, are considered in

this study. This effectively removes all first and second-person zero pronouns.
The Hobbs algorithm operates on one sentence at at time, scanning parsed tree nodes
for a suitable antecedent for a pronoun. The algorithm in its most basic form is subject-
biased, and relies on left-to-right and breadth-first searches through syntactic nodes, which
is reproduced in (165). It starts with the tree in which the pronoun is found and expands

the scope to each preceding tree when it fails to find a suitable candidate. In the initial
steps 1 through 4 when it looks for an intrasentential antecedent, the search moves further
and further up the tree to the left of pronoun; when it fails to identify a suitable antecedent
the search will move up to the preceding sentence. For pronouns occurring close to the

left edge of the sentence, this means that inter-sentential resolution is favored; for those
occurring in other positions, antecedent candidates located within the same sentence are
given precedence.

(165) Naive Hobbs Algorithm

1. Begin at NP node immediately dominating the pronoun.

2. Go up the tree to the first NP or S node encountered. Call this node X. Call

the path to reach X p.

3. Traverse all branches below node X to the left of path p, in left-to-right,

breadth-first manner. Propose as the antecedent any NP node that is encoun-

162
 tered that has an NP or an S node between it and X.

4. If node X is the highest S node in the sentence, traverse the parse trees of
previous sentences in order of recency (the most recent first), from left-to-
right, breadth-first and propose as antecedent the first NP encountered.

5. From node X go up the tree to the first NP or S node encountered. Call this
new node X, and call the path traversed to reach it from the original X p.

6. If X is an NP node and if the path p to X did not pass through the N-bar node
that X immediately dominates, propose X as the antecedent.

7. Traverse all branches below node X to the left of path p, left-to-right, breadth-
first. Propose any NP node encountered as the antecedent.

8. If X is an S node, traverse all branches of node X to the right of path p, left-

to-right, breadth-first, but do not go below any NP or S encountered. Propose
any NP node encountered as the antecedent.

9. Go to step 4.

It must be noted that the Hobbs algorithm further relies on some semantic constraints.
First, it relies on gender and number agreement: the latter can for the most part be easily

deduced by morphological reflections; the former requires some level of world-knowledge

such as the preferred gender associated with proper names. Since Korean zero pronouns
lack either pieces of information, this provision was not implemented. Second, in his orig-
inal configuration Hobbs adopts the use of selectional constraints to filter out semantically

inconsistent antecedents. He illustrates this point with the following example:

(166) The castle in Camelot remained the residence of the king until 536 when he moved
it to London.

163
 In this example, the algorithm proposes first 536 and then the castle as the an-
tecedent of the pronoun it, both of which are ruled as semantically implausible by the se-
lectional constraints as the entities denoted by the NPs are not movable. The third choice,

the residence, passes the selectional constraint and therefore is selected as the antecedent
of the pronoun. While it is feasible to employ such semantic reasoning in a small-scale
hand-coded evaluation, it requires a higher level of knowledge representation in automated
evaluation, which most computerized systems do not have access to. The semantic con-
straints, therefore, were not implemented in our study; instead, the bare-bones algorithm

illustrated in (165), often referred to as the naive Hobbs algorithm, was employed. With-
out the constraints, the search scope is effectively limited to the sentence immediately pre-
ceding the sentence where the pronoun is located: the algorithm needs not look any further
than the immediately preceding sentence, as it always succeeds in designating one of the

NPs in the sentence without the semantic constraints in place to disqualify some of them.

3.2 Variations of Hobbs on Korean Zero Pronouns

On English pronouns, reported success rates of the Hobbs algorithm range between 49% to
91.7%, mostly centered around upper 80s%. In his 1978 paper, Hobbs tests his algorithm

on the English pronouns he, she, it and they, and reports that the bare algorithm
correctly resolved the pronouns for 88.3% of the cases, and adding the above mentioned
selectional constraints improved the performance to 91.7%. Walker (1989) tests the algo-
rithm on different genres of texts, and finds that it performs worst on task-oriented corpora

(49%) and performs better in other domains such as newspaper articles (89%) and fiction
(88%). A relatively poorer performance is reported for other languages, including Chinese:
Converse (2005) applies the algorithm on Chinese pronouns in the Penn Chinese Treebank
corpora, and finds its success rates to be 63.2% for third person overt pronouns, 53.2% for

164
all anaphors, and 49.0% for zero pronouns. On Spanish zero pronouns, Ferrandez & Peral
(2000) finds its success rate to be 49.1%.
The naive version of the Hobbs algorithm was implemented and applied to the two Penn

Korean Treebank corpora. The accuracy of the algorithm was 46.1% and 62.2% for KTB
1 and KTB 2 respectively (shown in Table 3.1).

source pro count correctly resolved % accuracy

KTB 1 879 405 46.1%
KTB 2 5270 2283 62.2%

Table 3.1: Naive Hobbs algorithm on Korean zero pronouns

The results indicate that the naive Hobbs algorithm performs rather poorly on the Ko-

rean zero pronoun data. After close examination of errors, it is revealed that adverbial noun
phrases (NP-ADV) in Korean pose a problem for Hobbs. Their English counterparts are
often prepositional phrases with a noun phrase contained in them; these noun phrases are
therefore placed deeper in the tree, and do not get traversed before the subject NP of the
sentence even when the PP occurs sentence-initially (example 167). In Korean, by con-

trast, the semantic functions of English prepositions are mostly carried out by post-position
morphemes, and as a result adverbial noun phrases are ubiquitous. When such adverbial
noun phrases occur sentence-initially, they are encountered and proposed as the antecedent
before the subject gets a chance (example 168).

(167) English: Hobbs picks our class before Wednesday

(S (PP-TMP (IN On)

(NP (NNP Wednesday)))

(, ,)

(NP-SBJ (PRP$ our)

(NN class))

165
 (VP (VBD went)

(PP (IN on)

(NP (DT a)

(NN field)

(NN trip))))

(. .))

(168) Korean: Hobbs picks {
\
9 (Wednesday-on) before
o

r (our class)

(S (NP-ADV /NNC+/PAD)

(S (NP-SBJ /NPN

/NNC+/PAU)

(VP (NP-OBJ /NNC

/NNC+/PCA)

/VV+/EPF+/EFN))

./SFN)

In fact, adverbial noun phrases (NP-ADV) are the least common antecedent NP type
in our data. Table 3.2 presents the breakdown of all noun phrases annotated as an antecedent

in the Penn Korean Treebank 2 by their syntactic types. With the exception of one vocative
NP (NP-VOC), the NP-ADV category sports the smallest number of antecedents.

antecedent NP count %
NP-SBJ 11821 77.8%
NP (mostly genitive case) 1117 7.4%
noun 833 5.5%
NP-OBJ 771 5.1%
NP-COMP 517 3.4%
NP-ADV 128 0.8%
NP-VOC 1 0.0%
total 15188 100%

Table 3.2: Breakdown of antecedent NPs in KTB 2

166
 source pro count correctly resolved % accuracy
KTB 1 879 425 48.4%
KTB 2 5270 3324 63.1%

Table 3.3: Hobbs algorithm on Korean zero pronouns, adverbial NPs not considered

source pro count correctly resolved % accuracy

KTB 1 879 454 51.6%
KTB 2 5270 3368 63.9%

Table 3.4: Hobbs algorithm on Korean zero pronouns, argument antecedents only

As can be predicted from the observation, removing adverbial NPs altogether from
antecedent candidacy improves the performance. The revised Hobbs algorithm performs at

the accuracy rates of 48.4% and 63.1% for the two corpora (Table 3.3).
Another experimental modification was tried on the Hobbs algorithm: further elimina-
tion of the bare NP type from candidacy, leaving only the three argument types (NP-SBJ,
NP-OBJ and NP-COMP) to be considered as potential antecedents. In the Penn Korean

Treebank, most bare NPs are part of a compound NP which itself is an argument; these
bare NPs, therefore, will not get a chance to be proposed as an antecedent by the algorithm,
since the higher-sitting parent argument NP is traversed first. There are two exceptions: (1)
bare NPs embedded in an NP-ADV are now candidates, because its parent NP-ADV node

has been rendered ineligible with introduction of the above modification; (2) bare NPs with
a copula suffix, which projects up to a verb phrase node (VP). Bare NPs in these two en-
vironments are exposed to tree traversal, and will be picked up by the algorithm in some
cases. While bare NPs in general account for a larger number of antecedent NPs than ob-
ject and complement NPs (Table 3.2), they make a poor candidate for an antecedent since

their overall frequency is far greater than those of the two argument-type NPs. Adding this
revision again improves the performance: now 51.6% and 63.9% of pronouns in Korean
Treebank 1 and 2 corpora, respectively, are resolved successfully (Table 3.4).

167
3.3 Significance of Syntactic Environments

The performance of the revised algorithm was further analyzed with respect to the syntactic
levels of the clauses in which zero pronouns are located. Three clausal locations are dis-

tinguished: matrix clauses, adverbial clauses which modify the matrix clause, and finally
embedded clauses which include complementized or relativized clauses and also adver-
bial clauses on a sub-matrix level. In complex sentences, the matrix or the main clause is
the one located at the right edge, reflecting the head-final nature of the Korean language.

The syntactic configurations associated with the three groups of clauses are illustrated in
example (169) below.

(169) Three groups of syntactic levels

(S (S1 (NP-SBJ ...)

(VP ...))

(S2 (NP-SBJ ...)

(VP (NP-OBJ (S3 (WHNP ...)))))

(S4 (NP-SBJ ...)

(VP (S5-COMP ...)))

.)

S4: matrix clause

S1, S2: adverbial clause
S3: relativized embedded clause; S5: complementized embedded clause

Table (3.5) presents the performance analysis by the clausal groups. The results in-
dicate that there are significant degrees of performance variance across the three clausal
environments for both corpora, in which zero pronouns in adverbial clauses perform the

168
strongest (62.7% and 74.7%), followed by those in matrix clauses (51.7% and 64.8%), and
finally the poorest performance by zeroes in an embedded clause (42.6% and 62.1%).

corpus sentence level pro correctly resolved % accuracy

KTB 1 Matrix clause 665 344 51.7%
Adverbial clause 309 194 62.7%
Embedded clause 413 176 42.6%
KTB 2 Matrix clause 494 320 64.8%
Adverbial clause 643 480 74.7%
Embedded clause 4133 2568 62.1%

Table 3.5: Hobbs algorithm on Korean zero pronouns, performance by clausal groups

The findings are in partial agreement with results on Chinese pronouns reported by

Converse (2005), where wider performance gaps were found among pronouns belonging
to different groups of syntactic levels. Her classification method, however, differs from the
one employed in the current study. She groups pronouns into three classes M1, M2 and S,
where M1 stands for pronouns that appear as matrix subjects, and S for pronouns that

were found in any kind of subordinate construction. M2 is an interim category which

bears a closer relation with M1, reserved for pronouns that appear as subjects of parallel,
independent clauses in multi-clause sentences which are specific to the Chinese data. Under
this classification, our Matrix clause category mostly falls under her M1, with a few

falling under M2; our Adverbial clause and Embedded clause together correspond to
her S, with a few belonging to our Adverbial clause falling under M2. In her results, it
was the matrix-clause (M1) group that was the best performer (73.3%), followed by M2
(63.0%) and finally those in S (42.5%).
The two sets of results are not directly comparable due to the above mentioned disparity

in classification. Additionally, the size of the Chinese data, totaling at 155 zero pronouns, is
rather too small to warrant strong conclusions. Keeping these risks in mind, they neverthe-
less yield a few points of interesting observations. First, the two studies have in common

169
that the zero pronouns in embedded environments, i.e., Embedded clauses in the Korean
data and S in the Chinese data, are served the worst by the Hobbs algorithm. Second,
the two analyses exhibit contradictory results on the performance of adverbial clauses that

are not embedded: zero pronouns in the Adverbial clause environment in the Korean
data perform the best, while their Chinese counterparts, falling under the S class, are the
worst performer in Hobbs. Overall, while it is true in the case of Korean data that the Hobbs
algorithm produces differing degrees of success for zero pronouns in the three syntactic en-
vironments, the performance gaps are not as pronounced as those reported for the Chinese

data.

3.4 Summary

The naive Hobbs algorithm was implemented and tested on NP-anaphoric pros in the two
Korean Treebank corpora, and was found to be accurate for 46.1% (KTB 1) and 62.2%

(KTB 2) of the time. When more restrictions were placed on the algorithm limiting an-
tecedents to be argument NPs only, which includes subject, object and other complements,
the performance was improved to 51.6% and 63.9%. Compared to the performance of
resolution systems that were statistically trained, which will be presented in the next sec-

tion, we find the performance of these simple heuristics-based algorithms to be more than
decent. Especially, the algorithms are more effective on those zero pronouns located in a
higher level such as the main clause or the adverbial clause than those located in embedded
environments.

170
Chapter 4

A Maximum Entropy Reference

Resolution System

Rule-based algorithms such as the Hobbs algorithm typically operate on a few numbers
of hand-picked linguistic contexts, whose interaction must be carefully formulated by a
linguist. The advantage of statistically trained systems, on the other hand, is that they are

flexible in designs in the way that the training process itself can learn what conditions are
predictive to what extent, presented with a large number of such conditions, or features,
that have not been pre-screened for their relevance.
In this chapter, we set out to build statistical models for resolution of Korean zero

pronouns, using two sets of annotated corpora. In particular, we employ the Maximum
Entropy modeling framework, which has been proven to be flexible enough to be used in
the creation of a wide variety of natural language processing components. Our goals of
this chapter are twofold. One is to demonstrate that the Maximum Entropy framework can
be successfully employed in providing solutions to the problem of Korean zero pronoun

resolution, which is faced with the unique challenge of combining numerous linguistic
factors to make predictions for different subtypes of zero pronouns whose natures and

171
behaviors vary widely from one another. The second one, which concerns the narrow class
of NP-anaphoric zero pronouns, is to explicate the contribution of various linguistic factors
to resolution models, in the hopes of shedding light on the phenomenon of zero pronoun

anaphora in Korean.
The chapter is organized as follows. First, the overall design of the resolution systems
being built is presented, including the use of the Maximum Entropy framework as a ranking
method and the issue of combining the two separate tasks of classification and resolution.
Then, the set of linguistic features used in our study is introduced. As for the resolution

systems, we have built two kinds: a cross-categorical one handling all types of pros and the
other focusing on NP-anaphoric pros only. For the former, we outline two methods, one
that combines the classification and the resolution tasks all in one phase, and the other that
deals with them in two separate stages. In building resolution systems that are intended for

NP-anaphoric zero pronouns only, we present results from various experiments which are
aimed to reveal predictive powers of certain groups of linguistic features.

4.1 The Design

4.1.1 Classification and Resolution

From a language processing point of view, the diverse nature of Korean zero pronouns
poses a unique challenge. Pronoun resolution for Korean cannot simply rely on tradi-
tional approaches for anaphor resolution, but rather the problem itself has to be redefined

to suit each different type of zero pronoun. Most approaches to resolution of English pro-
nouns concern third person pronouns only, and a successful resolution means locating an
antecedent NP in the surrounding text. In the case of Korean zero pronouns, however, a
reference resolution system must deal with two disparate tasks: classification of the zero

172
pronoun and identification of an antecedent. Moreover, each type of zero pronoun has a
different definition of successful resolution.
For some types, the resolution stops at classification: once a zero pronoun is classified

as an x-type, i.e., an expletive zero pronoun, there is no further step to be taken. Other
types involve identification of its referent in some way. Deictic zero pronouns need to be
resolved to an entity present in the discourse scene, in most cases a discourse participant.
Anaphoric pronouns can be treated in the same manner as English pronouns by resolving
to a coreferential noun phrase in the text if it is NP-anaphoric, or by finding a clausal unit

which it refers if it is propositional.

The case of the indefinite zero pronoun is less clear. By definition, we are unlikely to
be able to recover the referent of specific indefinite zero pronouns from the environment.
One can also argue that the referent of generic indefinites need not be further resolved, as

they can be left as vague as required. However, the referent of a generic indefinite is often
restricted by the domain of discourse, which is usually signified in a nearby context. The
resolution strategies for each type are summarized in (170):

(170) Resolution strategy by zero pronoun types

i. NP anaphoric: find an NP antecedent in surrounding text

ii. propositional anaphoric: find a clausal antecedent in surrounding text

iii. deictic: find the referent among discourse entities, usually discourse partici-
pants

iv. expletive: no further resolution

v. specific indefinite: no further resolution

vi. generic indefinite: no further resolution

In the current study, the definition of successful resolution is expressed in more spe-
cific terms. In doing so, we generalize the strategies stated above to allow for partial suc-

173
cesses, especially discovery of coreference relations with other pronouns, including zero
forms. For all non-anaphoric categories that can enter coreferential relations with nomi-
nal expressions such as indexical and generic, discovery of such coreferential relations via

identification of an antecedent NP, overt or null, is counted as successful. For the proposi-
tional anaphoric category, we simply stop at correct categorization without making further
attempts to identify clausal or eventual antecedents for them. This decision is motivated in
part by the small sample size of the type (86 tokens found in the two corpora, 79 excluding
those in the Penn Korean Treebank 1 sections that are not contiguous discourses), as well

as their seemingly simple behavior in so far as the data at hand is concerned: in the two cor-
pora, over 90% of the propositional anaphoric pros are found to refer to the propositional
or eventual content of the preceding sentence. However, a few number of propositional
anaphoric pros, 4 in all, are found to have NP antecedents in addition to clausal ones (see

Tables 4.14 and 4.15 in Section 4.2.1). The clausal antecedents are not included in the
resolution target sets but the NP antecedents are: as long as these NP antecedents are con-
cerned, resolution of these particular propositional-anaphoric pros proceeds in the same
manner as NP-anaphoric pros. The definition of successful resolution by zero pronoun

type is presented below:

(171) Definition of successful resolution by zero pronoun types

i. NP anaphoric (a):
find an NP antecedent, overt or null, in surrounding text

ii. propositional anaphoric (p):

(a) correct classification as type
(b) (rare) identification of a coreferential NP, overt or null, in surrounding text

iii. deictic (i, y, w):

(a) correct classification as type

174
 (b) identification of a coreferential NP, overt or null, in surrounding text

iv. expletive (x):

correct classification as type

v. * specific indefinite (s):

correct classification as type

vi. generic indefinite (g):

(a) correct classification as type
(b) identification of a coreferential NP, overt or null, in surrounding text

As reported earlier in Section 2.2.3, the specific indefinite (s) type did not occur in

our data. The category was therefore dropped from the pool of categorical targets in the
actual model building processes.

4.1.2 Two-Phased vs. Single-Phased System

The nature of Korean zero pronouns presented so far naturally implies a two-phased ar-

chitecture in a resolution system: for each pronoun, the system first attempts to assign it a
correct category, and then it will proceed to determine an antecedent where necessary. In
the first phase, the target set will consist of the zero pronoun categories given in (171); in
the second phase, the target set will consist of NPs found in nearby text.
However, a real-life resolution system does not have to be modeled as a direct reflection

of the nature of the problem to be solved. An alternative is a system which collapses the
two phases into one where it simultaneously tries to predict the category and the antecedent.
Arguably, such an approach has some advantages over the more principled two-phase ap-
proach. First, reduced error: a single-phased prediction system is likely to have less er-

ror propogation and therefore a reduced error rate over a multi-phased prediction system,
though this point remains to be empirically tested. Second, elimination of redundancy. In

175
each phase, the decision making process will be based on a collection of information ex-
tracted from the context, and there is a significant overlap between the information used
for determining the category of a zero pronoun and the information needed for locating an

antecedent of a zero pronoun, such as grammatical and lexical features. By combining two
phases, reuse of such information can be avoided. Third, certain zero pronouns can be non-
anaphoric and anaphoric at the same time, which may be the case with deictic and generic
pronouns, as explained earlier in sections 2.6, 2.4.3.3. A single unified phase facilitates
treatment of such cases, as we will see shortly.

Combining the two prediction components can be accomplished by collapsing the two
pools of predicted outcomes into one. That is, instead of the set of zero pronoun classes
for phase one and the set of antecedent candidates for phase two, a single set of targets
is generated for each zero pronoun including the closed number of zero pronoun class

indices given in (171) and an unbounded number of potential antecedents extracted from
the nearby context. A ranking algorithm will then pick a winning choice from among them.
An important consequence of this approach is that there is a direct competition between
zero pronoun class indices and NP antecedent candidates: for example, an x-class index

can win over NP antecedent candidates with the effect of classifying a zero pronoun as
an expletive; or conversely, an NP antecedent candidate can win over all class indices and
other NP antecedent candidates thereby ruling the pronoun as anaphoric.
As an illustration, the following is a sample target set generated for the zero pronoun
subject of sentence 06:2 in the previous examples (163, 164) (repeated below):

(172) 06:01.A 
, ^  :

2 ,
y2

.
ca, kim thukmwucang2, y2 anc-ula.
well, kim 1SG2 , (SBJy2 ) sit-Imp.
All right, 1SG Kim2 , (youy2 ) sit down.

176
 06:02.A y2 $3 7
- "
3
3 %

 ?
y2 mwe3 com mek-ess-na?
(SBJy2 ) what3 little eat-Past-Q?
Did (youy2 ) eat something3 ?

06:03.B V
\, i2 3 3
3
" %_
 vm
 .
yey, i2 3 mek-ess-supnita.
yes, (SBJi2 ) (OBJ3 ) eat-Past-PolDec.
Yes, (Ii2 ) ate (something3 ).

(173) ;;A;06:1: , , .
(S (INTJ /IJ)

ca

well

,/SCM

(NP-VOC#2 /NPR

kim

Kim

/NNC)

thukmwucang

sergeant

,/SCM

(S (NP-SBJ#y2 *pro*)

zero-pronoun

(VP /VV+/EFN))

anc ula

sit Imp

./SFN)

177
;;A;06:2: ?

(S (NP-SBJ#y2 *pro*)

zero-pronoun

(VP (NP-OBJ#3 /NPN)

mwe

something

(VP (ADVP /ADV)

com

little

(VP /VV+/EPF+/EFN)))

mek ess na

eat Past Q

?/SFN)

;;B;06:3: , .

(S (INTJ /IJ)

yey

yes

,/SCM

(S (NP-SBJ#i2 *pro*)

zero-pronoun

(VP (NP-OBJ#3 *pro*)

zero-pronoun

/VV+/EPF+/EFN))

mek ess supnita

eat Past Polite-Dec

./SFN)

178
(174) sample target set for zero pronoun in sentence 06:2

target description
p propositional
i speaker
y* hearer

w speaker + hearer
g generic indefinite
x expletive (situational)
NP-VOC06:01 * vocative NP in preceding sentence

NP-SBJ06:01 * subject NP in preceding sentence

NP-OBJ06:02 object NP in current sentence

Note that there are some assumptions that have been made to simplify the resolution
scheme. First only three targets are listed for the deictic category: i, y and w, repre-

senting the speaker, the hearer and we including the speaker and the hearer. This restricts
resolution of deictic zero pronouns to the three discourse participant entities. Also, in this
particular target set, no clausal-unit antecedents are generated for the zero pronoun, and
instead the categoric index p is included.
Given the target set, a successful resolution has occurred if the ranking algorithm picked

as the winner either of the three coreferential items marked with *: the y-index, the voca-
tive NP or the subject NP from the preceding sentence. As discussed earlier, this shows how
a zero pronoun referring to a discourse participant has chances of success on two different
levels, from its deictic index and from its previous mentions, if any. It is a matter of debate

whether resolving the pro to the subject NP of the preceding sentence, itself a zero pronoun,
should be regarded as a successful resolution. Ideally, a linguistic expression with a fuller
semantic content should be recovered by a resolution system. In a full-fledged reference

179
resolution system that not only resolves zero pronoun references but also keeps track of
all coreference relations across a text, it can be accomplished by creating and maintaining
coreference chains, within which the reference is shared in a transitive fashion among its

members.

4.1.3 Maximum Entropy as a Ranking Method

Since its re-discovery as a valuable NLP technique, maximum entropy models (Ratna-
parkhi, 1997, 1998; McCallum, 2002) have kept their popularity as the machine learning
method of choice amongst many researchers in the field. Their strengths are twofolds: their

flexibility and linguistically intuitive nature. Their main component is features, which are
easy to design, construct and manipulate, and can be derived from various linguistic aspects
of the problem to be solved. Maximum entropy offers an elegant and computationally ef-
fective way to combine diverse pieces of contextual evidence, encoded in features, in order

to estimate the probability of a certain linguistic class occurring with a certain linguistic
context. For example, to estimate the probability of a word assigned a particular part-of-
speech tag, many pieces of contextual information can be translated into features, such as
the word at hand and other surrounding words. For the purpose of coreference resolution,

the classification task is binary: given a pair of a pronoun and another NP, a prediction is
to be made between two classifications 0 and 1, which stand for not coreferential and
coreferential. Therefore, a Maximum Entropy feature function in this study is of the
following form:

1 if a = a and cp(b) = true

(
f cp,a (a, b) =
0 otherwise
where a is a particular prediction on as class and cp is a contextual predicate, i.e., a
piece of contextual information, from the context b. One example of a feature that might

be useful for reference resolution is:

180
 1 if a = 1 (i.e., coreferential) and pro-is-matrix-subject(b) = true
(
fcp,a (a, b) =
0 otherwise

Presented with a set of features, a maximum entropy model can produce probability
estimates toward each possible outcome by combining parameters associated with the fea-
tures. As with other machine learning techniques, maximum entropy models, often called

classifiers, are learned through a training phase: parameters for each feature are dis-
covered from observed data through theory-independent algorithms such as Generalized
Iterative Scaling. Once parameters j are learned, the probability of a context b being
classified as class a can be estimated by the following formula:

k
1 Y f (a,b)
p(a|b) = j
Z(b) j=1 j
P Qk f (a,b)
where Z(b) = j=1 j j is a normalization constant which ensures that p sums to 1
over a.
A training stage is typically followed by a testing phase, which consists of presenting

the trained model with a set of known class outcomes and their features, and evaluating the
predictions made by the model to measure its accuracy.
While the Maximum-Entropy based method of coreference resolution is a binary clas-
sification method, it can be recast as a ranking method by making use of the likelihood

scores the classifier assigns towards each of the two classes. In other words, all (pronoun,
NP) pairs generated for a pronoun and its potential antecedents can be ranked by the degree
of how likely it is to be classified as coreferential.
First, let us illustrate the binary classification process with the case of the zero pro-
noun subject in the previously seen example (164). All pairs of the zero pronoun and its

antecedent candidate NPs are given either class 0, not coreferential or 1, coreferential:

181
 (pronoun, targetNP)
(NP-SBJ06:02 , p)
(NP-SBJ06:02 , i)
(NP-SBJ06:02 , y)
(NP-SBJ06:02 , w)
(NP-SBJ06:02 , g)
(NP-SBJ06:02 , x)
(NP-SBJ06:02 , NP-VOC06:01 )
(NP-SBJ06:02 , NP-SBJ06:01 )
(NP-SBJ06:02 , NP-OBJ06:02 )
class description
0 pronoun, propositional
0 pronoun, speaker
1 pronoun, hearer
0 pronoun, speaker + hearer
0 pronoun, generic indefinite
0 pronoun, expletive
1 pronoun, vocative NP in 06:01
1 pronoun, subject NP in 06:01
0 pronoun, object NP in 06:02

Table 4.1: Binary classification for zero pronoun and its potential antecedent NP in 06:2

Based on contextual information in the form of features associated with each pair, a

maximum entropy classifier produces the following set of estimations for the probability
of each of the two classes, 0 and 1. The class with the higher probability score is the class
decision made by the classifier, marked in bold: it is compared against the actual class
for evaluation. The only incorrect prediction made by the classifier involves the deictic

reference y: the classifier assigns a higher score for class 0 thereby ruling it disjoint with
the pronoun, when in fact the pronoun refers to the hearer.

(pronoun, targetNP) P(0) P(1) class evaluation

(NP-SBJ06:02 , p) 0.937 0.062 0 correct
(NP-SBJ06:02 , i) 0.991 0.008 0 correct
(NP-SBJ06:02 , y) 0.655 0.344 1 incorrect
(NP-SBJ06:02 , w) 0.984 0.015 0 correct
(NP-SBJ06:02 , g) 0.718 0.281 0 correct
(NP-SBJ06:02 , x) 0.943 0.056 0 correct
(NP-SBJ06:02 , NP-VOC06:01 ) 0.488 0.511 1 correct
(NP-SBJ06:02 , NP-SBJ06:01 ) 0.471 0.528 1 correct
(NP-SBJ06:02 , NP-OBJ06:02 ) 0.980 0.019 0 correct

Table 4.2: Binary classification of coreference for zero pronoun in 06:2

The ranking-based approach, on the other hand, pays no attention to the coreferen-
tial/not coreferential decision made on each pair, but instead ranks them based on the

182
probability score assigned to the class coreferential. The above list is therefore ranked
as below, successfully placing the three coreferential pairs at the top. The ranking selects
the subject NP of the preceding utterance as the most likely antecedent of the zero pro-

noun. Note that the pair with the deictic reference y, which was incorrectly predicted as
not coreferential with the binary classifier above, proves itself as a worthy contender this
time, ranking just below the two other coreferential pairs.

rank (pronoun, targetNP) P(0) P(1) class

1 (NP-SBJ06:02 , NP-SBJ06:01 ) 0.471 0.528 1
2 (NP-SBJ06:02 , NP-VOC06:01 ) 0.488 0.511 1
3 (NP-SBJ06:02 , y) 0.655 0.344 1
4 (NP-SBJ06:02 , g) 0.718 0.281 0
5 (NP-SBJ06:02 , p) 0.937 0.062 0
6 (NP-SBJ06:02 , x) 0.943 0.056 0
7 (NP-SBJ06:02 , NP-OBJ06:02 ) 0.980 0.019 0
8 (NP-SBJ06:02 , w) 0.984 0.015 0
9 (NP-SBJ06:02 , i) 0.991 0.008 0

Table 4.3: Binary classification of coreference for zero pronoun in 06:2

4.1.4 Selection of Features

As explained in the previous section, a machine learning classifier needs a set of features

associated with a case in order to make a classification decision for it. The features are
essentially pieces of contextual information which weigh towards each of the possible out-
comes. In terms of the task of reference resolution, then, relevant features are those that
present evidence towards either the category of a zero pronoun or whether or not a par-

ticular NP is an antecedent of a zero pronoun. The features can be categorized into three
groups: the first concerns the environments of the zero pronoun itself; the second con-
cerns the antecedent NP; the last group of features includes those that capture the relation
between the zero pronoun and the antecedent NP. The exhaustive lists of features in the

183
respective groups are presented in (175), (176) and (177).
Within each group, the features are organized around familiar linguistic properties, such
as the grammatical roles and formal factors of the nominal elements, verbal semantics and

morphology. Each feature is binary with a value of 0 or 1: a pro is assigned the value 1 for
the feature pro is a main clause argument if it is a main clause argument, and 0 if it not.
Some of the features are lexical: head noun of NP is X, predicate of pro/NP is X and
predicate of pro/NP has connective verbal ending X, where X represents lexical forms
of predicates and verbal endings. While seemingly multi-valued, these are nevertheless

deconstructed into arrays of binary features: for all the lexical values found in a model
for these features, each feature is represented as a binary value. To be more specific: if
1,000 predicate types have been found to occur with a pro in a given training material for
more times than a threshold number (5 in the present study), the predicate of pro/NP is X

feature expands to a collection of 1,000 individual features on each of which pro is assigned
the value of either 0 or 1. As a result, the size of the feature space is in fact a lot larger than
enumerated in the three tables below, and varies depending on the training material used.

(175) Features on pro

grammatical role

pro is subject

pro is object

pro is other type of argument

formal factors

none

location: level of embedding

pro is a main clause argument

184
 pro is in an adverbial clause modifying the main clause

pro is an embedded clause

location: linear

pro is in a sentence-initial position

type of embedding

pro is in a relative clause

pro is in a nominalized clause

pro is in a complementized clause

verbal semantics

predicate of pro is X

predicate of pro is is because

pro is the subject of an equative (copula) construction, i.e., pro is NP.

pro is the subject of a preposition-like verb

pro is the subject of a psych-verb, e.g.,


 /alta/ know,

 /komapta/
is thankful, 

 /palata/ hope

verbal morphology: modality

predicate of pro has morphology expressing deontic modality

predicate of pro has morphology expressing epistemic modality

verbal morphology: tense

predicate of pro has morphology expressing irrealis tense

predicate of pro has morphology expressing future tense

predicate of pro has morphology expressing past tense

predicate of pro has morphology expressing retrospective tense

185
 predicate of pro has morphology expressing progressive tense

verbal morphology: mood

predicate of pro has morphology expressing interrogative mood

predicate of pro has morphology expressing imperative mood

predicate of pro has morphology expressing declarative mood

predicate of pro has morphology expressing exhortative mood

predicate of pro has morphology expressing promissive mood

predicate of pro has morphology expressing conditional mood

verbal morphology: level of speech

predicate of pro has morphology expressing honorification

predicate of pro has morphology expressing politeness

verbal morphology: connective ending

predicate of pro has connective verbal ending X

other

pro is in a quoted speech

(176) Features on NP antecedent

grammatical role

NP is subject

NP is object

NP is other type of argument

NP is possessive (genitive case)

NP is vocative

NP is locative

186
 NP is non-argument

formal factors

NP is a zero pronoun

NP is a pronoun

NP is topic-marked

NP is in a bare form

location: level of embedding

NP is a main clause argument

NP is in an adverbial clause modifying the main clause

NP is an embedded clause

location: linear

NP is in a sentence-initial position

noun semantics

head noun of NP is X

verb semantics

predicate of NP is X

predicate of NP is is because

NP is the subject of an equative (copula) construction, i.e., NP is NP

NP is the complement of an equative (copula) construction, i.e., NP is

NP

NP is the subject of a preposition-like verb

NP is the subject of a psych-verb, e.g.,


 /alta/ know,

 /komapta/

is thankful, 

 /palata/ hope

187
 verbal morphology: modality

predicate of NP has morphology expressing deontic modality

predicate of NP has morphology expressing epistemic modality

verbal morphology: tense

predicate of NP has morphology expressing irrealis tense

predicate of NP has morphology expressing future tense

predicate of NP has morphology expressing past tense

predicate of NP has morphology expressing retrospective tense

predicate of NP has morphology expressing progressive tense

verbal morphology: mood

predicate of NP has morphology expressing interrogative mood

predicate of NP has morphology expressing imperative mood

predicate of NP has morphology expressing declarative mood

predicate of NP has morphology expressing exhortative mood

predicate of NP has morphology expressing promissive mood

predicate of NP has morphology expressing conditional mood

verbal morphology: level of speech

predicate of NP has morphology expressing honorification

predicate of NP has morphology expressing politeness

verbal morphology: connective ending

predicate of NP has connective verbal ending X

(177) Relational features on pro and target NP

pro and NP are in the same sentence

188
 pro and NP are in the same clause

pro and NP have the same grammatical role

pro and NP (not in the same clause) are arguments of the same lexical verb

pro and NP are arguments of two lexical verbs that are antonyms (e.g., e



O
/issta/ exist and \  /epsta/ do not exist)

pro and NP are located in two coordinated clauses

NP precedes pro in linear order

NP immediately c-commands pro

NP is the head of the relative clause in which pro is found

NP and pro are subjects of a question-answer pair

NP is the scene-setting adverbial preceding pro

NP binds pro syntactically/semantically

Nearest Subject: NP is the subject NP which is located the closest on the

lefthand-side of the pro

Nearest Topic : NP is the topic-marked NP which is located the closest on

the lefthand-side of the pro

An event, defined on a pro and target pair, is represented as a vector of all of these binary
features: it is essentially a collection of pieces of contextual and linguistic information
which are relevant to deciding whether or not the pro-target pair is coreferential.
Constructing events on a zero pronoun and a NP antecedent pair simply follow from the

feature sets presented above. Special attention is needed, on the other hand, for coreference
events which pair up a pro with one of the six categorical targets. First of all, since there is
no antecedent NP involved, the second and the third group of features which are extracted

189
from the antecedent NP and its surrounding environments are irrelevant: the events simply
take on negative values for all features belonging in the two groups. The only observable
information comes from the surroundings of the zero pronoun itself: the first group of

pro-based features is applicable for these events, but these features will have to be treated
themselves as some sort of relational features, which encode the relations between the zero
pronoun and the particular categorical choice of reference at hand. Therefore, the pro-
based features are duplicated for each of the six categories, but they are altered to reflect
the particular category. To be more specific, the terms of the first few features are changed

to include a reference to the particular choice of categorical NP reference:

(178) Features on pro for generic indefinite (g) interpretation

grammatical role

pro is g and subject

pro is g and object

pro is g and other type of argument

...

As a result, the seven categories, including the NP-anaphoric events and the six non-NP-
anaphoric ones, each occupy feature spaces that are completely disjoint. In a feature vector

which incorporates all the features specific for any category, it means that all events be-
longing in one category will uniformly take on the negative value for every feature meant
for other categories.
Let us illustrate the scheme by revisiting a previous example (163), reproduced here:

(179) (=163, 172)

06:01.A 
, ^  :

2 ,
y2

.
ca, kim thukmwucang2, y2 anc-ula.
well, kim 1SG2 , (SBJy2 ) sit-Imp.

190
 All right, 1SG Kim2 , (youy2 ) sit down.

06:02.A y2 $3 7
- "
3
3 %

 ?
y2 mwe3 com mek-ess-na?
(SBJy2 ) what3 little eat-Past-Q?
Did (youy2 ) eat something3 ?

06:03.B V
\, i2 3 3
3
" %_
 vm
 .
yey, i2 3 mek-ess-supnita.
yes, (SBJi2 ) (OBJ3 ) eat-Past-PolDec.
Yes, (Ii2 ) ate (something3 ).

When the search space is restricted to the present and the immediately preceding sentence,
the pro subjects in (06:02.A) and (06:03.B) yield 9 coreference events respectively, 3 of
them NP-anaphoric and 6 non-NP-anaphoric. Each event is represented as a binary feature

vector which is sized around 2,500 for models built for the Penn Korean Treebank 1; only
a part of the vector is shown in the Table 4.4.

4.2 Single-Phased System

When building events, a limited search space needs to be specified from which prospective
antecedent NPs are collected. For this study, only the utterance in which the zero pronoun
is contained and its immediately preceding utterance were considered. All noun phrases
included in the scope other than the zero pronoun to be resolved, including those that are

zero pronouns themselves, are considered targets: nouns that are not on the phrasal level are
not considered. In addition to these NP targets, there are the categorical ones: the 6 non-NP-
anaphoric classes i, y, w, g, x, p. The total number of events generated per zero pronoun
equals the total number of targets. Table 4.5 shows the total number of targets generated

for the two corpora and the average number of coreference events per zero pronoun. The
target sets turn out to be quite large for both corpora: a zero pronoun in KTB 1 is expected
to have a target set of 20, while one in KTB 2 has a target set of twice the size. Out of the

191
 (NP-SBJ06:02 , NP-VOC06:01 )

(NP-SBJ06:02 , NP-OBJ06:02 )

(NP-SBJ06:03 , NP-OBJ06:02 )
(NP-SBJ06:03 , NP-OBJ06:03 )
(NP-SBJ06:02 , NP-SBJ06:01 )

(NP-SBJ06:03 , NP-SBJ06:02 )
(pronoun, target)

(NP-SBJ06:02 , i)

(NP-SBJ06:02 , w)
(NP-SBJ06:02 , g)

(NP-SBJ06:02 , p)

(NP-SBJ06:03 , i)

(NP-SBJ06:03 , w)

(NP-SBJ06:03 , x)
(NP-SBJ06:03 , p)
(NP-SBJ06:02 , y)

(NP-SBJ06:02 , x)

(NP-SBJ06:03 , y)

(NP-SBJ06:03 , g)
feature
pro is i and SBJ 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
pro is y and SBJ 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
pro is w and SBJ 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
pro is g and SBJ 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0
pro is x and SBJ 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0
pro is p and SBJ 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0
pro is i and OBJ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

pro is SBJ 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 1 1 1
pro is OBJ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
pro is other role 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
pro has predicate "
 to eat
3 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 1 1 1
pro has predicate
to sit
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

NP is SBJ 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0
NP is OBJ 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 1
NP is pro 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 1
NP is topic-marked 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
NP is sentence-initial 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0
NP has predicate " to eat
3 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 1 1 1
NP has predicate to sit
0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0

pro and NP are in same sentence 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1
pro and NP have same predicate 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0

Table 4.4: Feature vector of coreference events for two pro subjects, partially represented

192
20 or so targets, less than 2 are coreferential for KTB 1, making the ratio of true (i.e.,
coreferential) events quite sparse at 9.3%. True events are more sparse for KTB 2: about
1 out of 20 events is a coreferential one (Table 4.6).

source pro targets targets per pro

KTB 1 2,468 51,117 20.7
KTB 2 8,161 345,435 42.3

Table 4.5: Numbers of targets/events generated per pro

source pro coreferential coreferential % of corefer-

targets targets per pro ential targets
KTB 1 2,468 4,730 1.90 9.3%
KTB 2 8,161 17,861 2.19 5.2%

Table 4.6: Numbers of coreferential targets/events generated per pro

In order to facilitate 10-fold cross validation, each corpus was partitioned into 10 sub-
sets, each of which is of a roughly equal size based on their count of coreference events.
At each evaluation, 9 parts were used as the training material and the remaining 1 as the
testing material (Table 4.7). The performance of a coreference resolution model was then
measured by adding up the 10 evaluations.

source pro total # of events training set size testing set size
KTB 1 2,468 51,117 46,005 5,112
KTB 2 8,161 345,435 310,883 34,552

Table 4.7: Training and testing set sizes for the two corpora

In training, any feature that occurs less than 5 times in the training data was dropped
from model building. As a result, the average number of features actually used in building

a model is far smaller than the total number of features found in the training data: almost
60% and 50% of features are discarded in KTB 1 and KTB 2, respectively (Table 4.8).

193
The values of feature parameters are estimated using 100 iterations of Generative Iterative
Scaling (GIS) method, using Ratnaparkhis (1997, 1998) implementation.

source features extracted from corpus features used in model

KTB 1 6,121 2,512
KTB 2 21,406 10,917

Table 4.8: Feature set sizes

As explained previously in (4.1.3), the initial maximum entropy model obtained from

training on the training data is a binary classifier: given a coreference event, it predicts
whether it is coreferential or not. Because the overwhelming majority of events are non-
coreferential, the baseline performance of binary classifiers, one that always predicts non-
coreferential, is high to begin with: 90.7% and 92.6% respectively. The trained classifiers

only slightly improve upon the baselines: the average accuracies of binary classifiers as a
result of 10-fold cross validation are 93.2% and 95.7% respectively.

source defaulting to 0 binary classification performance

KTB 1 90.7% 93.2%
KTB 2 92.6% 95.7%

Table 4.9: Performance of models as a binary classifier

The high baselines mean that the classifiers are highly biased towards making one par-
ticular outcome, in this case 0 not coreferential. The following confusion matrixes, com-

piled across the 10 evaluation runs, show that the majority of correct predictions comes
from predicting the 0 outcome:
While the vast majority of 0 coreference events are predicted correctly, a smaller por-
tion of coreferential events are correctly predicted with outcome 1: the recall figures of

the coreferential events of the two corpora are 49.0% (KTB 1) and 36.3% (KTB 2). The
precision figures are higher: out of all events that are predicted 1 by the models, 70.8%
(KTB 1) and 65.8% (KTB 2) are indeed coreferential.

194
 model predicts 0 model predicts 1 total
outcome 0 43036 (88.5%) 944 (1.9%) 43980
outcome 1 2381 (4.9%) 2288 (4.7%) 4669
total 45417 3232 48649

Table 4.10: Confusion matrix for binary classifiers, KTB 1

model predicts 0 model predicts 1 total

outcome 0 324212 (93.9%) 3362 (1.0%) 327574
outcome 1 11382 (3.3%) 6479 (1.9%) 17861
total 335594 14744 345435

Table 4.11: Confusion matrix for binary classifiers, KTB 2

As explained earlier, our zero pronoun resolution scheme is ultimately a ranking model
constructed from the binary Maximum Entropy classifiers. According to the scheme, all
potential targets for a given pro are rounded up and ranked based on the probability scores

with which they are predicted to be classified as coreferential. Resolution of a zero

pronoun is ruled successful when the top-ranked candidate indeed represents a coreferential
pro-NP or pro-category pair. According to this resolution scheme, 69.7% and 64.7% of
zero pronouns are successfully resolved for KTB 1 and KTB 2, respectively. Some zero
pronouns are found to be unresolvable: they either do not have any antecedents at all in

the text or have an antecedent NP outside of their search scope, i.e., the current sentence
and the one immediately preceding it. These performance scores are obtained by totaling
the number of successful resolutions across the 10 test runs; alternatively, mean accuracy
scores obtained by averaging them were the same at 69.7% and 64.7%, with the KTB 1

data showing larger degrees of variance across different runs (Table 4.13).

195
 source total zero unresolvable correctly accuracy accuracy among
pronouns resolved resolvable pros
KTB 1 2468 113 1720 69.7% 73.0%
KTB 2 8161 174 5282 64.7% 66.1%

Table 4.12: Performance of coreference resolution models

source mean lowest highest

accuracy accuracy accuracy
KTB 1 69.7% 64.4% 76.7%
KTB 2 64.7% 61.1% 68.6%

Table 4.13: Performance of coreference resolution models, averaging on 10 cross-fold val-

idation

4.2.1 Performance by Type

Assessing the performance of the zero pronoun resolution system by the types of zero
pronouns is complicated by the fact that some zero pronouns are categorized as generic or
deictic and have textual antecedents at the same time. For example, a case of a generic

pronoun failing to pick up the generic target candidate cannot be ruled as unsuccessful,
since it is possible that it also has a coreferential NP antecedent in the text which the
resolution system might have ranked as the most likely candidate. In the following, the
resolution systems performance is examined by the types of zero pronouns. In doing so,

generic and deictic zero pronouns that have NP antecedents are treated primarily as non-
anaphoric; only the zero pronouns that are purely anaphoric, i.e., those that do not have
any coreferential relations with a generic or deictic entity, are handled in the anaphoric
category.

Tables 4.14 and 4.15 show the performance of the system by zero pronoun type. The
NP-anaphoric row indicates the number of zero pronouns in non-NP-anaphoric cate-
gories with NP antecedents in the text. A few in the propositional category are found
to have NP antecedents, while larger proportions of i and y zero pronouns have NP an-

196
tecedents. All three deictic categories exhibit higher NP-anaphoricity in KTB 2, w over
80% and i and y in the mid-90 and high-90 ranges. Zero pronouns in w and g categories
exhibit drastically different frequencies of NP-anaphoricity between the two corpora: 4.3%

vs. 83.9% for w and 74.9% vs. 32.9% for g.

source anaph- propo- deictic generic exple- total

oric sitional i y w tive
total count 879 61 392 214 23 597 302 2468
NP-anaphoric N/A 3 245 158 1 447 0 854
% NP-anaphoric N/A 4.9% 62.5% 73.8% 4.3% 74.9% 0% N/A
correct prediction 452 42 347 163 18 488 210 1720
% correct 51.4% 68.9% 88.5% 76.2% 78.3% 81.7% 69.8% 69.7%

Table 4.14: Accuracy by zero pronoun type, KTB 1

source anaph- propo- deictic generic exple- total

oric sitional i y w tive
total count 5270 18 267 19 56 513 2018 8161
NP-anaphoric N/A 1 265 18 47 169 0 500
% NP-anaphoric N/A 5.6% 99.3% 94.7% 83.9% 32.9% 0% N/A
correct prediction 3250 12 224 6 33 147 1610 5282
% correct 61.7% 66.7% 83.9% 31.6% 58.9% 28.7% 79.8% 64.7%

Table 4.15: Accuracy by zero pronoun type, KTB 2

Several observations can be made about the systems performance between the two
data set across the seven categories. First, KTB 1 exhibits consistent performance levels

among the non-NP-anaphoric types which ranges between 68.9% to 88.5%, while in KTB
2 larger amounts of fluctuation are seen among the 6 types, which range from 28.7% to
83.9%. Second, the anaphoric and the expletive types are better performers in KTB 2,
while propositional, deictic and generic types fare better in KTB 1 data; the systems higher

accuracy of the KTB1 data is largely due to the high performance of the deictic and generic
categories. Third, with the exception of deictic-hearer (y) pros whose size in KTB 2 is too
small to take its performance results as statistically significant, the generic pros (g) display

197
the largest performance gap between the two corpora, at 81.7% correct for KTB 1 and only
28.7% correct for KTB 2.
In assessing the performance by zero pronoun type, one should bear in mind the possi-

bility that these observations are the result of their relatively smaller sizes. In both corpora,
the majority of zero pronouns are NP-anaphoric, and the rest are unevenly split among
the rest of the categories, leaving some with inadequately small numbers of samples. The
small size of a certain category alone is not a predictor for poor performance: for example,
KTB 2 contains 18 instances of propositional-anaphoric zero pronouns and 19 instances

of deictic-hearer zero pronouns, but the former achieves a 66.7% accuracy while the lat-
ter merely 31.6%. It is reasonable, however, to regard performance measured on smaller
sample sizes as less reliable.
Let us now examine in detail the prediction patterns made for some of the types. NP-

anaphoric (a) zero pronouns, deictic-speaker (i) zero pronouns and generic (g) zero pro-
nouns in KTB 2 are chosen, shown in Tables 4.16, 4.17 and 4.18:

prediction anaphoric proposi- deictic generic expletive total

(a) tional (p) i y w (g) (x)
(a) # of predictions 4912 13 42 2 5 74 222 5270
correct among (a) 3250 0 0 0 0 0 0 3250
% correct 66.2% 0% 0% 0% 0% 0% 0% 61.7%
% among total
correct predictions 100% 100%

Table 4.16: Prediction pattern for anaphoric (a) zero pronouns, KTB 2

Again, recall that those pronouns that are generic or deictic but with an NP antecedent are

not identified as anaphoric, but instead counted primarily as generic or deictic. As expected,
the classifier mis-classifies many purely anaphoric (a) zero pronouns as belonging to other
categories, such as deictic, generic, propositional, and even expletive (Table 4.16). Tables
(4.17) and (4.18) show that more deictic-speaker (i) pronouns and generic (g) pronouns are

198
 prediction anaphoric proposi- deictic generic explet- total
(a) tional (p) i y w (g) ive (x)
(a) # of predictions 242 0 22 0 0 3 0 267
correct among (a) 202 0 22 0 0 0 0 224
% correct 83.5% 100% 0% 83.9%
% among total
correct predictions 90.2% 9.8% 100%

Table 4.17: Prediction pattern for deictic-speaker (i) zero pronouns, KTB 2

prediction anaphoric proposi- deictic generic expletive total

(a) tional (p) i y w (g) (x)
(a) # of predictions 441 0 2 1 1 47 21 513
correct among (a) 100 0 0 0 0 47 0 147
% correct 22.7% 0% 0% 0% 100% 0% 28.7%
% among total
correct predictions 68.0% 32.0% 100%

Table 4.18: Prediction pattern for generic (g) zero pronouns, KTB 2

matched with an NP antecedent, some of which turned out to be successful. In fact, 90.2%
and 68.0% of all successful resolution of deictic-speaker and generic pronouns respectively

are attributed to NP antecedents. In the case of deictic-speaker pronouns in particular, the

overwhelming majority of which have NP antecedents (Table 4.15: 99.3%), favoring an
NP antecedent over the i categorical candidate is a successful strategy that contributes to
its overall high accuracy level.

4.3 Two-Phased System

As discussed earlier in Section 4.1.2, an alternative to the single-phased system is a two-

phased system of pronoun resolution, one that involves the first step of determining cate-
gories on zero pronouns and then the second step of picking out NP antecedents for those

pronouns judged in step 1 as NP-anaphoric. This approach was implemented using the
same set of features. In a nutshell, the approach led to a slight performance gain for KTB

199
 category a p i y w g x
total
prediction
% recall 879 61 392 214 23 597 302
a 659 7 96 48 3 182 67
75.0%
p 7 42 11
68.9%
i 21 260 18 15
66.3%
y 23 16 109 7 1
50.9%
w 19 1
82.6%
g 150 2 20 38 1 389 10
65.2%
x 19 10 1 4 213
70.5%

Table 4.19: Confusion matrix for category classification model (Phase 1), KTB 1

2 while causing a significant performance hit for KTB 1.

Phase 1 consists of building a classifier which predicts one of the seven pronoun cate-
gories: a, p, i, y, w, g, x. Table 4.19 shows evaluation results for the categor-
ical classifier trained and tested on KTB 1. The figures along the diagonal line represent
the number of zero pronouns that have been correctly categorized; with the exception of

the zero pronouns in the a NP-anaphoric category, these zero pronouns have been suc-
cessfully resolved and the resolution process has been concluded with them (see first row
of Table 4.20). The shaded row in Table 4.19 (also second row in Table 4.20) represents
the zero pronouns that are ruled as a NP-anaphoric by the classifier: resolution of them

is not concluded until a correct antecedent NP is found in Phase 2. Note that many non-
NP-anaphoric zeroes are categorized as a and are therefore submitted to Phase 2. Some
of them, such as the expletive zeros in category x, are bound to fail in Phase 2 since they

200
typically do not have NP antecedents; some others in the deictic i, y, w and the generic
g category have NP antecedents and will have a second chance at successful resolution in
Phase 2.

category a p i y w g x all
total
879 61 392 214 23 597 302 2468
resolved in Phase 1 42 260 109 19 389 213 1032
% recall 68.9% 66.3% 50.9% 82.6% 65.2% 70.5%
submitted to Phase 2 659 7 96 48 3 182 67 1062
resolved in Phase 2 416 0 64 29 1 99 0 609
% recall 47.3% 0% 16.3% 13.6% 4.3% 16.6% 0%
total resolved 416 42 324 138 20 488 213 1641
% total recall 47.3% 68.9% 82.7% 64.5% 87.0% 81.7% 70.5% 66.5%

Table 4.20: Combined performance of Phase 1 and Phase 2, KTB 1

category a p i y w g x all
resolved total
% recall 879 61 392 214 23 597 302 2468
1-phase 452 42 347 163 18 488 210 1720
approach 51.4% 68.9% 88.5% 76.2% 78.3% 81.7% 69.8% 69.7%
2-phase 416 42 324 138 20 488 213 1641
approach 47.3% 68.9% 82.7% 64.5% 87.0% 81.7% 70.5% 66.5%

Table 4.21: Comparison of two approaches, KTB 1

Table 4.20 shows combined performance of Phase 1 and 2. Out of 1,641 pros that have
been correctly resolved, 63% of them (1,032 pros) were resolved in Phase 1, and the rest
37% (609 pros) through phase 2. Comparison of evaluation results for the two approaches

by category types is given in Table 4.21. Categories w and x perform better under the two-
phase approach, but by negligible margins. The rest of the categories perform the same or
significantly better under the single-phase approach, pushing the single-phase approach to
win over the two-phase approach by 3.2% performance gain. The category with the largest

201
performance gap is the deictic-hearer y, with a recall number lower by over 11%; it is the
largest contributing factor to the lower performance result of the two-phase approach. In
addition, we can see in Table 4.19 that too few NP-anaphoric pronouns are submitted to

Phase 2 in the first place: only 75.3% of them are classified as a in Phase 1, and even fewer
of them are correctly matched with a coreferential NP antecedent in Phase 2.

category a p i y w g x
total
prediction
% recall 5270 18 267 19 56 513 2018
a 5023 3 189 16 49 449 350
95.3%
p 2 12
66.7%
i 32 67 4 1
25.1%
y 3 3 2 3
10.5%
w 7 3 1 3 1
1.8%
g 55 6 1 42 5
8.2%
x 148 2 1 1 15 1662
82.4%

Table 4.22: Confusion matrix for category classification model (Phase 1), KTB 2

The evaluation results are somewhat different with KTB 2. Tables 4.22 and 4.23 show
the two phases of resolution experiments conducted on KTB 2. What is immediately no-

ticeable is the divide between the types of non-a zero pronouns that can also have NP
antecedents and those that do not: the three deictic types (i, y, w) and the generic g,
belonging in the former group, rarely get resolved in Phase 1 (21.5%, 10.5%, 1.8%, 8.2%
respectively) while the propositional anaphoric type p and the expletive type x, belonging
in the latter group, exhibit higher percentage of resolution in Phase 1 (66.7% and 82.4%

202
 category a p i y w g x all
total
5270 18 267 19 56 513 2018 8161
resolved in Phase 1 12 67 2 1 42 1662 1786
% recall 66.7% 25.1% 10.5% 1.8% 8.2% 82.4%
submitted to Phase 2 5023 3 189 16 49 449 350 6079
resolved in Phase 2 3355 0 148 5 28 94 0 3630
% recall 63.7% 0% 55.4% 26.3% 50.0% 18.3% 0%
total resolved 3355 12 215 7 29 136 1662 5416
% total recall 63.7% 66.7% 80.5% 36.8% 51.8% 26.5% 82.4% 66.4%

Table 4.23: Combined performance of Phase 1 and Phase 2, KTB 2

respectively). This divide was not present with the two-phase approach with KTB 1: all six
non-a categories, regardless of their NP-anaphoric capability, were heavily biased towards
categorical resolution in Phase 1 (Table 4.20). In addition, a higher percentage (95.3%)
of NP-anaphoric (a) pronouns, compared to 75.0% of KTB 1, are correctly classified in

Phase 1, which means most a-type pros get a chance at correct resolution in Phase 2. In a
way, then, the classification phase of the two-phase approach is more successful for KTB
2, resolving NP-anaphoric-incapable types such as x and p in the phase while sending
other NP-anaphoric-capable types into the next phase. Given the higher NP-anaphoricity

rates for non-a types in KTB 2 (see Tables 4.14 and 4.15), this behavior of the system is
beneficial to its performance.

category a p i y w g x all
resolved total
% recall 5270 18 267 19 56 513 2018 8161
1-phase 3250 12 224 6 33 147 1610 5282
approach 61.7% 66.7% 83.9% 31.6% 58.9% 28.7% 79.8% 64.7%
2-phase 3355 12 215 7 29 136 1662 5416
approach 63.7% 66.7% 80.5% 36.8% 51.8% 26.5% 82.4% 66.4%

Table 4.24: Comparison of two approaches, KTB 2

203
 Table 4.24 presents a comparison between the performances of the two approaches for
KTB 2. The overall performance is slightly improved in the two-phase approach, from
64.7% to 66.4%. In contrast to KTB 1, the performance of pronoun categories remains

relatively unchanged between the two approaches: apparent exceptions are y and w, but
their sample sizes are too small to be considered significant. Overall, the performances
of the two largest categories a and x improved in the two-phase approach, leading to its
outperforming the single-phase approach. In conclusion, the two-phase approach produces
better performance for KTB 2, while the single-phase approach outperforms the two-phase

approach for KTB 1.

4.4 A Resolution System for Anaphoric Zero Pronouns

In addition to the cross-categorical resolution schemes presented in the previous sections,

we constructed additional resolution systems targeting the NP-anaphoric subcategory of

Korean zero pronouns. They are explored in detail in this chapter. In the previous models,
the process of evaluating and analyzing the contribution of linguistic contexts is compli-
cated by the fact that the feature sets employed interact with different zero pronoun cate-
gories in different manners. Certainly, a zero pronouns particular syntactic position, such

as whether or not it is a main clause argument, bears different relations with predictions for
different categories, for example a and x. By limiting the scope of model building to the
case of NP-anaphoric zeroes, we aim to render the relation between linguistic contexts and
their contribution to these models more transparent, thereby enabling us to conduct further

analyses in a more straightforward manner.

Limiting the scope of a pronoun resolution system to a more homogeneous, core anaphoric
cases is an approach that is commonly taken. In studies on overt pronouns, this is some-
times done by opting for certain types of pronouns in a pronominal system, such as third-

204
person singular he and she whose usages rarely deviate from the traditionally anaphoric
ones, and excluding others such as it and they which can be used non-referentially or in
opaque contexts. In a system dealing with zero pronouns, however, it should be noted that

such a selective approach comes with a theoretical and computational overhead since the
selection process is not a simple one that relies on the formal factor alone but one that
requires applying a deeper level of analysis, such as human judgment or the category clas-
sification step of our two-phase resolution system. From a practical point of view, it is
unreasonable to expect input to zero pronoun resolution systems to have been pre-screened

to include NP-anaphoric zeros only.1

With this reservation in place, we nevertheless believe that NP-anaphora specific mod-
els are useful inasmuch as they lead to clearer linguistic analyses as well as make it possible
to compare their performance and behavior to published systems for other languages that

target anaphoric pronouns. In what follows, we present results from training and testing
Maximum Entropy models for NP-anaphoric zero pronouns, followed by descriptions of
several experiments designed to assess contribution of individual features.

4.4.1 A Maximum-Entropy Model

For the purpose at hand, we picked out zero pronouns in the Penn Korean Treebank 2 cor-
pus that are co-indexed with noun phrases only, namely the ones in the a category. This
removes all expletive, generic zero pronouns, pronouns referring to propositional contents
as well as first and second person zero pronouns from the model. The initial models are
built for both Treebank corpora; for additional experiments that manipulate particular fea-

ture groups, we focused on the Penn Korean Treebank 2 data in order to further streamline
our analysis.
1
Obviously, the input used in this study, one that has been pre-parsed with zero pronouns pre-identified in
the right place, is already assuming a sophisticated level of knowledge representation.

205
 The methods and features used in constructing these models remain essentially the same
as those used in the previous experiments with a few key differences. The methodology of
training a Maximum-Entropy classifier to be used as a ranking scheme, described in Section

4.1.3, is retained. The two key differences, which depart from what is described in Section
4.1.4, are the following. First, only noun phrases that are found in surrounding texts are
included in the target set, and the 6 categorical targets, namely p, i, y, w, g, x, are no
longer generated. This means that the first 6 targets in the potential antecedent pool shown
in the previous example (4.1) are excluded (although, the zero pronoun in the example itself

is excluded from the present model since it is a deictic y pronoun). Second, the group of
features that captures the environments of pros, i.e., the first of the three groups of features
listed in (175), is dropped from the models. As explained earlier in Section 4.1.4, the main
contribution made by this group of features is towards distinguishing different categories

of pros. This is because the resolution scheme ultimately partitions the binary true/false
(coreferential/non-coreferential) event space into subsets of coreference events that belong
to each zero pronoun token, and the ranking of events matters only within the subgroups.
To be more specific, any context that relates to the zero pronoun alone will not have an

impact on rankings between targets that include NP antecedent candidates, since all events
for these candidates will be applied the same weights with regard to the features on the
pro. This does not mean that the contexts of pros are not taken into account at all: they
are significant in so far as they are expressed as relational to specific NP candidates, such
as whether or not an NP candidate and the pro are located in the same sentence, which is

captured by the third group of relational features (shown in (177)).

After removing all non-a pros, the count of remaining a pros in KTB 1 is reduced
by almost two-thirds (Table 4.25), mainly owing to the large number of first- and second-
person pronouns in the corpus. Reduction of data size is not as significant for KTB 2, for

which NP-anaphoric pros account for 65% of all pros. As can be expected, the average

206
number of targets per pro is exactly 6 less than those observed for pros in all categories (cf.
4.5). Table 4.25 also shows that the proportions of coreferential targets over all targets have
increased over the previous models for all categories (cf. 4.6).

source pro NP NP targets coreferential coreferential % of corefer-

targets per pro targets targets per pro ential targets
KTB 1 879 12012 13.7 1303 1.48 10.8%
KTB 2 5270 191419 36.3 13292 2.52 6.9%

Table 4.25: Numbers of coreferential NPs per NP-anaphoric pro

As before, evaluation was carried out by means of 10-fold cross-validation: 10 models

were trained on 9 parts of 10 equally proportioned partitions, which were then tested on

their respective remaining partition. The accuracy figures for the two corpora were found
to be 64.2% for KTB 1 and 66.4% for KTB 2 (Table 4.26), measured by summing up
the numbers of successful resolutions across all 10 models. The accuracy score obtained
by averaging success rates of the 10 individual test runs were close to those at 64.0% and

66.4% (Table 4.27); KTB 1 exhibits larger variance between runs that range from the lowest
57.4% to the highest 75.8%, which should not come as a surprise given its small data size.

source total zero unresolvable correctly accuracy accuracy among

pronouns resolved resolvable pros
KTB 1 879 113 564 64.2% 73.6%
KTB 2 5270 174 3499 66.4% 68.7%

Table 4.26: Performance of coreference resolution models

source mean lowest highest

accuracy accuracy accuracy
KTB 1 64.2% 57.4% 75.8%
KTB 2 66.4% 61.7% 71.3%

Table 4.27: Performance of coreference resolution models, averaging on 10 cross-fold val-

idation

207
 Compared to the Naive Hobbs algorithm (3.1), which were 46.1% and 62.2% for the
KTB 1 and 2 corpora, the trained models managed to achieve higher performance levels.
The scale of improvement, however, was not as great as we hoped it to be. The performance

gain for the KTB 1 corpus was larger at about 8%, while the KTB 2 corpus showed a more
moderate amount of improvement of about 4%.

4.4.2 Evaluating Features

With a narrowed focus on anaphoric zero pronouns, we are in a better position to take up
the task of ultimate interest: investigating how various features, each of which is encod-

ing a particular linguistic context, contribute to zero pronoun resolution. In doing so, we
especially hoped to be able to demonstrate that the various linguistic concepts which are
discussed in Part I as relevant to the Korean zero pronoun phenomenon do, in fact, serve as
predictive features in the statistically trained models.

Evaluating the contribution of individual features is not a straightforward task, much of

which stems from the way Maximum Entropy models handle features. One fundamental
underlying assumption of the Maximum Entropy method is independence among features;
in most real-life applications, including the present study, this assumption simply cannot

hold true. While Maximum Entropy models are known to be adept at compensating for
feature dependence, it nevertheless creates complications in the way features interact with
each other in their collective contribution towards particular outcomes, and, eventually, the
way their parameters are approximated during training. Simply put: evaluating the contri-
bution of individual features involves excluding the influence of all others; since features

are not independent, such evaluations are likely to be burdened with many interfering fac-
tors.
With these complications in mind, there are nevertheless a few ways to get a glimpse
into the roles of individual features. The first one involves building alternate models with

208
each feature or feature group dropped, and then observing a resulting decrease in perfor-
mance. The second method involves building alternate models with only a handful of fea-
ture groups at a time and comparing their performances. The last one is to look at feature

weights assigned during model training. The three methods are outlined in the following
sections.

4.4.2.1 The Ablation Method

One possible way to measure the contribution of each feature is to note the effect on perfor-
mance when it is removed from the full model consisting of all the features, often referred

to as the ablation method. The resulting decrement reflects a contribution unique to the
feature that cannot be compensated for by the other features. We constructed a series of
varied models, each omitting a specific group of features which encode a particular linguis-
tic aspect. For example, to build a maximum entropy resolution system that is blind to the

formal factors of NP antecedents, we removed the features that indicate whether or not an
NP is topic-marked, pronominal, a zero pronoun, or in a bare form. The outcome of these
experiments is given in Table 4.28 below.

# correctly resolved performance feature(s) removed

3224 61.2% nearest sbj
3387 64.3% form
3446 65.4% same lexical verb
3448 65.4% grammatical role
3450 65.5% nearest topic
3451 65.5% co-clausal
3457 65.6% lexical verb
3460 65.7% co-sentential
3464 65.7% syntactic level
3474 65.9% sentence-initial
3499 66.4% all features

Table 4.28: Performance degradation by removing feature set

209
 The nearest subject feature was the most outstanding performer of all: its removal
resulted in the largest amount of performance loss, with the second-highest ranked feature
a distant runner up. The set of formal features caused the second-largest performance hit.

It should be noted that the features nearest subject and grammatical role, and also
nearest topic and form are not independent: in fact, the two formers are highly spe-
cialized subcases of the two latters. This means that even when the nearest topic feature
is removed, the information regarding topic-hood of an NP is still present as a part of the
form set of features. To completely rid a model of formal information, both have to be

removed. Removing nearest subject and grammatical role features shows a greater
level of performance loss (2nd row in Table 4.29), and so does removing nearest topic
and form (third row in Table 4.29). Finally, removing both groups of features, i.e., all
formal and grammatical-role-related, results in an even greater amount of performance hit

(top row in Table 4.29).

# correctly resolved performance feature(s) removed

3124 59.3% grammatical role, form, nearest sbj/topic
3216 61.0% grammatical role, nearest sbj
3356 63.7% form, nearest topic
3499 66.4% all features

Table 4.29: Performance degradation by removing feature set

4.4.2.2 Feature Opt-In Models

Another common approach to assessing the contribution of each feature is to build a model
with only that feature present and observe system performance compared to the baseline,
that is, the opposite of the ablation method. This approach has limitation in the present

study, however, due to the ranking-based nature of our system. Recall that our system does
not stop at classification of coreference events, but applies further rankings among them in

210
order to specify the top candidate: when there are not enough features to base the rankings
on, it may fail to select a unique winning candidate. For example, a model which builds
on grammatical role features alone will not be able to place rankings between NPs with

the subject grammatical role, when there are multiples of them found in the context. For
this reason, evaluation of models using the feature opt-in method is not universally applied
but rather starts with two features which are guaranteed to fire on no more than a single
NP within a candidate pool, thereby letting a unique wining candidate surface: the nearest
subject and the nearest topic features.

A nearest subject NP is present for almost every zero pronoun, except for a handful
of cases where a zero pronoun occurs very close to the beginning of a discourse segment:
5,250 out of the 5,270 anaphoric pros, or 99.6%, have such an NP. The nearest topic fea-
ture is not as guaranteed: in general, topic-marked NPs are not as ubiquitous as subject NPs,

which lead to a smaller total of 4,941 anaphoric NPs with a nearest topic NP antecedent
candidate, which is 93.8% of the time. Since these features are positively correlated to the
outcome 1, i.e., coreferential, resolution models with either feature to models that pick
the NP with the feature as the antecedent of a pro, while picking a random candidate when

there is no such NP around. In this sense, maximum-entropy based machine learning is

not an essential part of the model: the same effect can be achieved by simple algorithmic
procedures. The performance of the two models with either feature is shown in Table 4.30:

# correctly resolved performance feature(s) used

3413 64.8% nearest subject
2910 55.3% nearest topic
364 6.9% none (baseline: one NP randomly picked)
3499 66.4% all features

Table 4.30: Performance of individual features: nearest subject, nearest topic

The predictive power of the nearest subject feature is most impressive: simply by

211
picking the nearest subject for each zero pronoun, a resolution system is expected to suc-
ceed almost 65% of the time. In fact, its success rate is very close to that of the full-feature
model, which was only 1.5% higher at 66.4%. The feature wins over the Naive Hobbs al-

gorithm as well, which was shown to be successful 62.2% of the times (3.1). The nearest
topic feature proved a strong predictor as well, achieving 55.3% success rate, although not
quite as strong as the nearest subject feature.
We further constructed a model that has access to more information, including the two
features above as well as formal factors and grammatical roles of antecedent NPs. The

performance of the resulting model is seen as improved over the one employing the nearest
subject feature alone, which is now only slightly worse than the full-feature model, but by
a margin that is too small to be taken as statistically significant (Table 4.31):

# correctly resolved performance feature(s) used

3490 66.2% form, grammatical role, nearest subject/topic
3499 66.4% all features

Table 4.31: Performance of combination of features

4.4.2.3 Feature Weights

One obvious source of information for the individual features contribution is the param-
eters, or weights, assigned to each of the features during the training process. Recall that
the probability of an event b being classified as class a can be estimated by the following
formula:

k
1 Y f (a,b)
p(a|b) = j j
Z(b) j=1

f (a,b)
where j j represents the parameter for feature j associated with a particular outcome,
or class, a. Therefore, each feature is assigned a weight for each outcome that it is associ-

212
ated with; in our case, each feature is assigned two weights, one towards the outcome 0
(non-coreferential) and the other 1 (coreferential). A feature exhibiting a large weight
towards outcome 1 means that it is positively correlated with an NP being an antecedent;

a feature with a low parameter value towards the same outcome means that it is negatively
correlated with an NP being an antecedent. A feature with the most predictive power, one
that occurs with almost all true antecedent NPs and rarely with NPs that are not antecedents,
is expected to have a high parameter value towards outcome 1 and a low value towards
0.

The above formula can be re-stated using natural logarithms, as below:

1 Pkj=1 jfj (a,b)

p(a|b) = e
Z(b)

Here, the weight represents the power value in the natural logarithm representation of
parameters, such that = e . This means that: a positive value of towards an outcome
results in the increased overall probability of that outcome as obtained from the above
formula; a negative value of towards an outcome decreases the overall probability of that

outcome; finally, the value 0 of towards an outcome means that the feature is neutral to
the outcome, that is, it does not affect its probability.
The list of features with the most positive and the most negative weights are presented
in the Tables 4.32 and 4.33. Note that the features referring to specific lexical properties of

target NPs, such as the the head noun of NP is X and the predicate of NP is X, were
excluded from the list, leaving in only those features that encode more general linguistic
properties.
The top halves of the 40 selected features with extreme parameter values, either positive
or negative, tend to be correlated with the outcome 1 rather than 0.

Many features are noted for their inclusion in both lists. The following features exhibit a

213
 rank weight () outcome feature on target NP
1 21.947 1 NP is possessive
2 11.233 1 NP is a zero pronoun
3 10.409 1 NP is the nearest subject
4 9.703 1 pro and NP are arguments of the same lexical verb
5 9.370 1 NP binds pro
6 7.990 1 pro and NP are in coordinated clauses
7 7.717 1 NP is the nearest topic
8 6.553 1 NP is the scene-setting adverbial preceding pro
9 4.519 1 NP linearly precedes pro
10 3.950 1 NP is in a sentence-initial position
11 3.601 0 predicate of NP is with promissive morphology
12 2.918 0 NP is in a main clause
13 2.740 1 NP immediately c-commands pro
14 2.405 0 NP is in an embedded clause
15 2.352 1 pro and NP has the same part-of-speech
16 2.346 0 NP is the head of the relative clause in which pro is found
17 2.129 0 NP is a subject
18 2.121 0 NP is in a bare form
19 2.117 1 NP is a pronoun
20 1.977 0 NP is an object

Table 4.32: Top 20 positively weighted features

high parameter value towards coreference while exhibiting a low value for non-coreference:

NP is a zero pronoun, NP is the nearest subject, NP binds pro, pro and NP are in
coordinated clauses, pro and NP are arguments of same lexical verb, and NP is the
nearest topic. These features, then, are highly effective in discerning coreferential events
from non-coreferential ones, by heavily promoting the likelihood of the outcome 1 and
at the same time penalizing the likelihood of the other outcome. Some other features ex-

hibit the exact opposite pattern: a high parameter value towards non-coreference and a low
value towards coreference. The features include: NP is in an embedded clause, NP is a
subject, NP is an object, and predicate of NP is with promissive morphology. These
features work exactly the opposite of the other group of features, that is, by increasing the

overall probability score of an NP being a non-antecedent and decreasing its probability of

214
 rank weight () outcome feature on target NP
1 -62.503 1 NP is the subject of a preposition-like verb
2 -50.209 1 NP is locative
3 -43.725 1 NP is non-argument
4 -32.524 1 NP immediately c-commands pro
5 -31.404 1 NP is an object
6 -25.068 1 NP is other type of argument
7 -22.334 1 NP is in an embedded clause
8 -17.618 1 pro and NP are in the same clause
9 -16.498 1 predicate of NP is with promissive morphology
10 -16.275 0 NP binds pro
11 -15.962 1 NP is the subject of a copula construction
12 -15.775 1 NP is a subject
13 -14.060 1 NP is in a bare form
14 -12.576 1 NP is the complement of a copula construction
15 -11.523 0 NP is the nearest subject
16 -11.214 0 pro and NP are in coordinated clauses
17 -7.896 0 pro and NP are arguments of the same lexical verb
18 -5.987 0 NP is the nearest topic
19 -4.732 0 NP is a zero pronoun
20 -3.944 1 pro and NP are in the same sentence

Table 4.33: Top 20 negatively weighted features

being an antecedent. These are, then, good predictors for non-coreferential NPs.
The former group includes many features that encode familiar linguistic aspects of zero
pronoun antecedents that are widely believed to be relevant in zero pronoun anaphora, such
as the subject grammatical role, topichood, zero-form, locality, and lexical parallelism.

Regarding the latter group, it is not surprising that NP is object should be a good predictor
for non-antecedent NPs, given that only 5% of NP antecedents in KTB 2 are objects (Table
3.2). Inclusion of the feature NP is a subject in this group, on the other hand, might come
as a surprise; 78% of antecedent NPs in KTB 2 are, after all, subjects. The answer to this
features unexpected behavior can be found in its relation to another feature, NP is the

nearest subject. It has a dependent relation with NP is a subject: whenever the former
feature is on, the latter is on as well. The key point here is that NP is the nearest subject

215
is, in fact, one of the most powerful features, if not the single most predictive, as we have
seen in the previous sections. Since there are many more subjects that are not antecedents
than ones that are antecedents, and since those antecedent subjects are well covered by the

more specific NP is the nearest subject feature, NP is a subject feature is relegated to

become a predictor for non-antecedent NPs in order to account for its association with the
vast majority of non-antecedent subject NPs.
The presence of the feature NP is possessive at the top of the positively weighted
feature list is mystifying for two reasons. First, the possessive case is not among the group

of linguistic aspects that are commonly cited in the literature as a quality of antecedent NPs,
yet it outweighs all other features that are in the group, such as NP is nearest Subject and
NP is pro. Second, despite its top rank in the positively weighted feature list, the feature
does not appear in the other list at all. In fact, its weight towards the outcome 0 is close

to 0 at mere -0.092. This suggests that the feature is useful in so far as promoting an NP
as an antecedent, when it is present; in judging an NP as a non-antecedent, it is close to
being neutral. Again, the answer for the features weights can be found in its relation with
other features with which it often co-occurs. While a very small number of antecedent

NPs are possessives (refer to Table 3.2), when a possessive NP is indeed an antecedent it
typically occurs with mostly negative features such as NP is non-argument, without the
company of other typically positive features such as NP is topic or NP is pro and so
on. Therefore, in order to counter the negative weights from other features and the lack of
other positive weights, the NP is possessive feature ends up with a highly positive weight

towards the coreference outcome.

As should be clear by now, the numeric values of the weights of the features in the tables
cannot be taken as a definite representation of their predictive powers. In other words: their
varying levels of strength cannot be quantified in absolute terms, which is largely due to

their inter-dependent nature and the way Maximum Entropy handles it. Even the notion of

216
comparative degrees of strength by which NP is a zero pronoun is rated stronger than
NP is the nearest subject in Table 4.32 must be approached with caution, as illustrated
by the cases of the two features NP is a subject and NP is possessive. Given the weights

of the features and the orderings that follow from them, it is natural to want to interpret them
as something analogous to such linguistic hierarchies as Cf-ranking of Centering Theory
or constraint rankings in Optimality Theory that we discussed earlier. While it is fair to
say that there are general regularities and tendencies to be found in the mappings between
the feature weights and the two hierarchies, it must be stressed that the rankings presented

in the two tables above do not directly translate to such hierarchies. The weights and the
orderings of the features in Maximum Entropy models are a product of taking into account
the organic relations present in the network of features, and are therefore much more than
binary ordering relations that constitute the hierarchies.

Comparison with linguistic hierarchies aside, the feature weights are nevertheless what
makes individual Maximum Entropy models unique. Inspired by the varying degrees of
performance of the Hobbs algorithm when applied to subsets of pros (Table 3.5), we trained
separate Maximum Entropy models for the three groups of zero pronouns found in the

clausal environments: the matrix clause, the adverbial clause, and the embedded clause.
The goal was to observe shifts in feature weight assignments. The overall performance
degraded slightly as a result (Table 4.34). While we did not make an attempt to exactly
pinpoint the source of the degradation, we believe it is not unreasonable given the decreased
sizes of the training sets as a result of breaking up the original data set.

group # pro # correctly resolved % accuracy

matrix clause 494 343 69.4%
adverbial clause 643 473 73.6%
embedded clause 4133 2601 62.9%
total 5270 3417 64.8%

Table 4.34: Performance of Maximum Entropy models on 3 clausal groups

217
 Upon examining the assigned feature weights, four features emerged as among the
most noticeable in shifts in their relative rankings across the three models: NP is a zero
pronoun, NP is the nearest subject, NP is the nearest topic, and NP is bound. Their

weights towards the outcome coreferential are listed in Tables 4.35, 4.36, and 4.37.

weight () outcome feature on target NP

13.018 1 NP is a zero pronoun
7.856 1 NP is the nearest topic
7.533 1 NP binds pro
5.183 1 NP is the nearest subject

Table 4.35: Weights of 4 features: trained for pros in a matrix clause

weight () outcome feature on target NP

12.578 1 NP is a zero pronoun
9.273 1 NP binds pro
6.719 1 NP is the nearest subject
5.556 1 NP is the nearest topic

Table 4.36: Weights of 4 features: trained for pros in an adverbial clause

weight () outcome feature on target NP

11.237 1 NP is a zero pronoun
10.723 1 NP is the nearest subject
8.783 1 NP binds pro
8.453 1 NP is the nearest topic

Table 4.37: Weights of 4 features: trained for pros in an embedded clause

The feature NP is a zero pronoun is seen consistently topping all three lists. In the
model for embedded pros, however, the gap between its weight and the second runner-up
NP is the nearest subject is not as significant (by a difference of 0.5) as in other models

(difference of 5+ in the main clause pro model and 3+ in the adverbial clause pro model).
This is due to the fact that NP is the nearest subject is unusually strong in the embedded

218
clause pro model. The NP is the nearest topic feature, on the other hand, is strongest in
the main clause pro model. The NP binds pro feature is the strongest in the adverbial
clause pro model, although the distinction is not as pronounced as the other features.

The patterns indicate that zero pronouns found in the three distinct syntactic environ-
ments have different preferences for their antecedent NPs. The topichood of a potential
antecedent carries more weight for zero pronouns in a main clause; the subjecthood, or
the subject grammatical role, of antecedent NPs is more important for zero pronouns in em-
bedded clauses; finally, zero pronouns in adverbial clauses prefer as their antecedents those

NPs that are in a position that syntactically binds them. The findings lend support for Milt-
sakakis (2002) claim that different mechanisms are needed for inter-sentential anaphora,
most of which involve zero pronouns on matrix level, and intra-sentential anaphora, which
most often occur with zero pronouns in dependent or embedded clauses.

4.5 Performance Scores: A Round-Up

Some key performance scores of various resolution systems that have been presented so far
are summarized in Table 4.38 below.
In terms of sheer numbers, the performance scores indicate a moderate level of success

in tackling the problem of zero pronoun resolution in Korean. While other reference reso-
lution studies report higher levels of success, it should be noted that most of these studies
not only differ in terms of the language that they target but also in the exact nature of the
problem that each of them seeks to address. Most studies on pronoun resolution in En-

glish are concerned with overt 3rd-person pronouns only: Hobbs (1977; 88% for fictional
domain, 89% for newspaper articles and 51% for task domains), Brennan, Friedman and
Pollard (1987; 90% for fictional domain, 79% for newspaper articles and 49% for task
domains), Strube (1998; 85%), and Tetreault (1999; 80.4% for news articles, 81.1% for

219
 data accuracy system description
KTB 1, cross-categorical 69.7% 1-phase Maximum Entropy model
KTB 1, NP-anaphoric pro 64.2% Maximum Entropy model
KTB 1, NP-anaphoric pro 46.1% naive Hobbs algorithm
KTB 1, NP-anaphoric pro 51.6% modified Hobbs algorithm

data accuracy system description

KTB 2, cross-categorical
KTB 2, NP-anaphoric pro
KTB 2, NP-anaphoric pro
KTB 2, NP-anaphoric pro
KTB 2, NP-anaphoric pro
KTB 2, NP-anaphoric pro
66.4% 2-phase Maximum Entropy model
66.4% Maximum Entropy model
62.2% naive Hobbs algorithm
63.9% modified Hobbs algorithm
64.8% Nearest subject heuristics
66.2% ME model, trained on form, grammatical role,
nearest subject/topic features

Table 4.38: Key performance scores

fictional texts) are among those studies. Ferrandez & Perals (2000) work is the only study

among reviewed that target zero pronouns, and they report 75% accuracy on anaphoric zero
pronouns in Spanish. However, their study was conducted on a much smaller scale: their
set of data consisted of 228 anaphoric zero pronouns.
More interesting than comparison with resolution studies in other languages is the per-
formance of various models that are built on different approaches. The level of success

that cross-categorical resolution models achieved proves that machine-learning methods

can be successfully employed in constructing resolution systems that are faced with the
unique challenge of handling multiple categories of zero pronouns at the same time. As a
statistical system, these Maximum-Entropy-based resolution systems are adept at combin-

ing heterogeneous pieces of information in order to successfully utilize them in complex

decision-making tasks. Hand-crafting heuristic rules that purport to deal with all categories
of zero pronouns, on the other hand, will no doubt prove a daunting task, having to con-
dition and formulate correlation between numerous aspect of linguistic contexts with the

many categories.

220
 With the resolution targets narrowed down to the more homogeneous group of NP-
anaphoric pros, the various models built for them give us an opportunity to compare mod-
els built on machine-learning methods and others that rely on simple heuristics. For both

KTB 1 and KTB 2 data, statistically trained models are shown to outperform those relying
on rules. While the performance gain from incorporating a rich set of features was sig-
nificant for the KTB 1 data, it was not as large as one would expect for the KTB 2 data.
Especially, simple heuristics-based approaches to anaphoric zeros in KTB 2 are found to
achieve comparable levels of success: the naive Hobbs algorithm succeeds 62.2% of the

time, and an algorithm that picks out the nearest subject NP as the antecedent for each
zero pronoun encountered is successful 64.8% of the time, which are only a couple per-
centage points below the accuracy of the statistical model trained on the full set of features.
These are highly localized relational features that draw on the interaction between a zero

pronoun and its would-be antecedent NP: when they are fired, they do so on a single NP
around a pro, with a high likelihood for actual coreference. Lastly, a maximum entropy
model trained on a select few number of features, which include the nearest subject and
nearest topic features and ones encoding formal factors and grammatical roles of NP an-

tecedents, was found to achieve a level of success rate close to one employing a full set
of features encoding rich linguistic information. It is a well-known fact that in statistical
models built on a large number of features a saturation point will be reached in the course
of adding more and more features. We conjecture that these pieces of information are by
and large sufficient in guaranteeing good performance of a zero pronoun resolution system

in Korean, and additional information on other linguistic contexts does not further enhance
the performance by a significant margin, although their addition surely will help in a few
cases.

221
Conclusions

In investigating the phenomenon of Korean zero pronouns in this thesis, we aimed to com-

bine two approaches: one that seeks to provide a linguistic analysis, and another that seeks
to develop a solution to the problem of their resolution. The two tasks are correlated in
some fundamental ways. Gaining insights into the nature and motivations behind how and
why zero pronouns occur is a prerequisite to engineering a working system that can sim-

ulate human understanding of zero pronouns. Conversely, constructing resolution systems

and testing them out on a large set of naturally occurring examples provide a valuable
chance to attest the various theoretical claims and observations in an empirical manner.
In Part I, we focused on analysis of Korean zero pronouns. In Chapter 1, previous work

on zero pronouns was outlined, starting with such syntax-oriented views as Government
and Binding Theory and Optimality Theory, moving on to the theories that are grounded in
pragmatic points of view, such as Centering Theory. In Chapter 2, Korean zero pronouns
were categorized by reference types: anaphoric zero pronouns and discourse-deictic zero
pronouns as text-dependent usage, and deictic zero pronouns, indefinite personal pronouns

(generic and specific), and situational zero pronouns as the types that are text-independent
in their usage. The various types of zero pronouns are then shown to possess distinct
properties in their interpretation and the contexts in which they appear.
In Chapter 3, we discussed the applicability of Centering Theory to Korean zero anaphora.

Especially, existing theories that promote different strategies for precisely formulating Cf-

222
ranking for Korean were closely examined. Ever since the original Cf-ranking for English
was published which was based on grammatical role, many Cf-rankings have been pro-
posed that combine elements from various linguistic dimensions such as grammatical role,

topic-hood, empathy and other formal properties of nominal expressions. Additionally,

some Cf-rankings directly incorporate higher-level linguistic concepts such as salience and
given/new information status. After careful examination of these rankings and the argu-
ments behind them, we concluded that Cf-rankings that draw on disparate linguistic dimen-
sions are not feasible, and also that those Cf-rankings that rely on more refined theoretic-

linguistic concepts are not desirable as a basis of resolution systems, which must operate
on more observable, surface-level characteristics of zero pronouns and their surroundings.
In essence, the very nature of Cf-ranking as a strictly single-dimensioned hierarchy makes
it ill-equipped when it comes to accounting for the web of contributions that are made by a

host of heterogeneous linguistic factors.

In Part II, we set out to construct a resolution system for Korean zero pronouns. First,
past research on pronoun resolution work was reviewed in Chapter 1, which ranges from
the more traditional rule-based approaches in its early days to the more recent trend of

statistical approaches that rely on various machine-learning technologies. In Chapter 2,

the data set used for the resolution study, the two Penn Korean Treebank corpora, is in-
troduced, along with the details on the annotation strategies employed in annotating the
data for coreference. In Chapter 3, we tested the Hobbs algorithm, one of the most widely
adopted among rule-based approaches, on the Korean data. The result indicates that the

algorithm, while simplistic in its formulation, achieves a good level of success for Korean
zero pronoun resolution.
In Chapter 4, we built statistically trained models, based on the Maximum Entropy
method. The first set of resolution systems are cross-categorical: they aim to resolve all

categories of zero pronouns. Two approaches are tried, in which (1) the category prediction

223
and the selection of antecedent NPs for anaphoric pros proceed all in one phase and (2) the
tasks of category prediction and antecedent NP selection are carried out sequentially, with
the latter applying to those NPs that have been ruled as anaphoric through the first step.

The former strategy produced a better performance for the Korean Treebank 1 corpus,
while the latter was shown to be more effective for the Korean Treebank 2. The second
set of resolution systems was constructed exclusively targeting NP-anaphoric pros, which
showed improved performance scores compared to those of Hobbs.
Ultimately, we were interested in finding out the contributions that individual features,

each capturing a particular linguistic property, made to the models. The query was carried
out in three ways: one involving ablation studies, one in which feature opt-in models
were experimented on, and finally one involving examining feature weights. One of the
findings from these studies is that the linguistic contexts that are often discussed as playing

a crucial role in the phenomenon of Korean zero pronouns, such as topic-hood, zero-form,
grammatical role, semantic and syntactic control and structural parallelism, are in fact sig-
nificant predictors for NP antecedents in the statistically trained models. Especially, the
nearest subject feature was found to be the single most predictable feature, which is rem-

iniscent of the strategy favored by the Hobbs algorithm. More specifically, however, the
nearest subject feature is most relevant in cases of zero pronouns occurring in embedded
clauses: since about 80% of all anaphoric zero pronouns are found in such an environ-
ment, the feature is found dominant in the general model. In a model that targets pros at
a matrix level, by contrast, the nearest topic feature was found to be of relatively higher

importance. Due to the inter-dependent nature of features and the way Maximum Entropy
handles them, their varying levels of strengths cannot be quantified in absolute terms. They
do not, therefore, directly translate to hierarchical linguistic-theoretic constructs such as
Cf-ranking of Centering Theory or constraint rankings in Optimality Theory, and cannot

be used to prove or disprove particular orderings in such rankings. We believe that this is

224
in fact a desirable outcome: the features encode diverse linguistic dimensions that are often
inter-dependent, and it is not necessarily possible, or even theoretically sound, to impose a
single hierarchical ordering among them. Nevertheless, the predictive powers of individual

features as observed in the resolution models can be seen as empirical validation of their
significance in linguistic theories of Korean zero pronouns.
Overall, this thesis has made several contributions to the problem of Korean zero pro-
nouns. First, it provided a comprehensive linguistic analysis on the phenomenon of Korean
zero pronouns as a whole. While there have been many studies of Korean zero pronouns

that focus on the most prominent type of NP-anaphoric zero pronouns, this study is one of
the first that attempts to cast light on the phenomenon in general, with detailed accounts of
the categories and the interplay between them. Second, it provided a critique of existing
theories of Korean zero pronouns by demonstrating limitations of each theory. Finally, the

thesis suggested a working model for resolving Korean zero pronouns. It is the first of its
kind: a system that is statistically trained on a large set of annotated corpora.
Clearly, many related areas of interest have been left unexplored in this project, which
we hope to pursue in future work. First of all, there are a few NLP techniques that have

high potential for improving the performance of our resolution system. One such promis-
ing technique comes from incorporating semantic knowledge into the system, such as the
predicate argument subcategorization information presented in the Korean Propbank Anno-
tations (Palmer et al., 2006). Semantic information for a zero pronoun can be deduced from
the subcategorization information on the lexical verb of which it is an argument, which can

then be utilized in assessing compatibility of antecedent candidates. In addition, applying

more sophisticated techniques of feature selection is likely to help optimize the system with
the most relevant and powerful set of features and enhance its performance, as well as help
to streamline it.

In terms of inquiries that are of theoretical interest, we believe that an in-depth compar-

225
ison between the two sets of models on each of the two Treebank corpora will yield inter-
esting results, hopefully shedding light on usage differences of zero pronouns in dialogue
and text genres. Finally, we recognize the problem of zero pronoun resolution as a part of

a larger problem, that of coreference resolution. The resolution scheme of NP-anaphoric

pros in the present study is limited in its scope in that it focuses on pair-wise coreference
between a zero pronoun and an NP. As a practical system, this approach often leaves us
with an insufficient amount of information for the ultimate recovery of the reference of a
zero pronoun, as the antecedent NP selected by the system is frequently a zero pronoun

itself or a pronominal expression. Our zero pronoun resolution system is not expected to
operate as a stand-alone application, but rather is conceived as part of a full-fledged refer-
ence resolution system that not only resolves zero pronoun references but also is capable
of discovering sequences of coreference relations among all referring expressions across a

text. Interpretation of a zero pronoun is truly accomplished in such a system: its referent
can be determined via another nominal expression in its coreference chain, whose fuller
linguistic content allows for unique identification of its referent entity.

226
Bibliography

Aroonmanakun, Wirote (1997) Referent resolution for zero pronouns in Thai, in Arthur

Abramson, ed., Southeast Asian Linguistic Studies in Honour of Vichin Panupong.

Aroonmanakun, Wirote (1999) Extending Focusing for Zero Pronoun Resolution in Thai,

Doctoral dissertation, Georgetown University.

Aroonmanakun, Wirote (2000) Zero pronoun resolution in Thai: a centering approach, in

et al. Burnham, Denis, ed., Interdisciplinary Approaches to Language Processing: The

International Conference on Human and Machine Processing of Language and Speech,
127147.

Baldwin, Breck (1997) CogNIAC: high precision coreference with limited knowledge
and linguistic resources, in Proceedings of the ACL Workshop on Operational Factors
in Practical, Robust Anaphora Resolution for Unrestricted Texts, 3845.

Beaver, David (2004) The optimization of discourse anaphora, Linguistics and Philoso-
phy 27:1, 356.

Borer, Hagit (1986) I-subjects, Linguistic Inquiry 17, 375416.

Borer, Hagit (1989) Anaphoric Agr, in Osvaldo Jaeggli and Kenneth Safir, eds., The Null
Subject Parameter, Kluwer, Dordrecht, 69109.

Bouchard, Denis (1984) On the Content of Empty Categories, Foris, Dordrecht.

Brennan, Susan E, Marilyn W. Friedman, and Carl J. Pollard (1987) A centering approach

227
 to pronouns, in Proceedings of the 25th Annual Meeting of the Association for Com-
putational Linguistics, ACL, 155162.

Cabredo-Hofherr, Patricia (2002) Arbitrary readings of 3pl pronominals, Sinn and Be-
deutung 7.

Calabrese, Andrea (1986) Pronomina, in Naomi Fukui, Tova Rapoport, and Elizabeth
Sagey, eds., Papers in Theoretical Linguistics, MIT Working Papers in Linguistics, MIT,
146.

Carbonell, James G., and Ralf D. Brown (1988) Anaphora resolution: a multi-strategy

approach, in Proceedings of the 12th International Conference on Computational

Linguistics (COLING88), Budapest, Hungary, 96101.

Chierchia, Gennaro (1995) The variability of impersonal subjects, in Emon Bach, E. Je-
linek, A. Kratzer, and B. H. Partee, eds., Quantification in Natural Languages, Kluwer,
Boston.

Choi, Jinyoung, and Yoon-Kyoung Joh (2004) Discourse functions of Korean sentential
conjunctions: a centering approach, in Proceedings of the 9th International Pragmat-
ics Conference.

Chomsky, Noam (1980) On binding, Linguistic Inquiry 11:1, 146.

Chomsky, Noam (1981) Lectures on Government and Binding, Foris Publications, Dor-
drecht, The Netherlands.

Chomsky, Noam (1982) Some Concepts and Consequences of the Theory of Government
and Binding, MIT Press, Cambridge, MA.

Chomsky, Noam (1986) Knowledge of Language: Its Nature, Origin and Use, Praeger,

Westport, Connecticut.

Choo, Miho (1994) A Unified Account of Null Pronouns in Korean, Doctoral dissertation,
University of Hawaii.

228
Cinque, Guglielmo (1988) On Si constructions and the theory of arb, Linguistic Inquiry
19, 521582.

Connoly, Dennis, John D. Burger, and David S. Day (1994) A machine learning approach
to anaphoric reference, in Proceedings of the International Conference New Methods
in Language Processing, Manchester, United Kingdom, 255261.

Converse, Susan P. (2005) Resolving pronominal references in Chinese with the Hobbs
algorithm, in Proceedings of the Fourth SIGHAN Workshop on Chinese Language
Processing.

Dagan, Ido, and Alon Itai (1990) Automatic processing of large corpora for the resolu-
tion of anaphora references, in Proceedings of the 13th International Conference on

Computational Linguistics (COLING90), Helsinki, Finland.

Dahl, Deborah A. (1996) Focusing and reference resolution in PUNDIT, in Proceedings

of the 5th National Conference on Artificial Intelligence, Philadelphia, PA.

Dahl, Deborah A., and Catherine N. Ball (1990) Reference resolution in PUNDIT, in

P. Saint-Dizier and S. Szpakowicz, eds., Logic and Logic Grammars for Language
Understanding, Ellis Horwood Limited.

Di Eugenio, Barbara (1990) Centering theory and the Italian pronominal system, in Pro-
ceedings 13th International Conference on Computational Linguistics (COLING90),
Helsinki, Finland, 270275.

Di Eugenio, Barbara (1996) The discourse functions of Italian subjects: a centering ap-
proach, in Proceedings 16th International Conference on Computational Linguistics,
COLING96, Copenhagen, Denmark.

Di Eugenio, Barbara (1998) Centering in Italian, in Marilyn A. Walker, Aravind K.

Joshi, and Ellen F. Prince, eds., Centering Theory in Discourse, Oxford University
Press, Oxford, 138160.

229
Dimitriadis, Alexis (1996) When pro-drop languages dont: overt pronominal subjects
and pragmatic inference, in Proceedings of CLS 32.

Donnellan, Keith (1966) Reference and definite descriptions, Philosophical Review 75:3,
281304.

Eckert, Miriam (1998a) Discourse Deixis and Null Anaphora in German, Doctoral disser-
tation, University of Edinburgh.

Eckert, Miriam (1998b) The German topic position and null anaphora, in Proceedings
of DAARC2 (Discourse, Anaphora and Reference Colloquium).

Eckert, Miriam, and Michael Strube (1999) Resolving discourse deictic anaphora in dia-
logues, in Proceedings of the 9th Conference of the European Chapter of the Associa-
tion for Computational Linguistics.

Eckert, Miriam, and Michael Strube (2001) Dialogue acts, synchronising units and
anaphora resolution, Journal of Semantics 17:1, 5189.

Ferrandez, A., M. Palomar, and L. Moreno (1998) Anaphor resolution in unrestricted texts
with partial parsing, in Proceedings of the 36th Annual Meeting of the Association for
Computational Linguistics.

Ferrandez, Antonio, and Jesus Peral (2000) A computational approach to zero-pronouns

in Spanish, in Proceedings of the 38th Annual Meeting of the Association for Compu-
tational Linguistics (ACL2000), Hong Kong, China, 166172.

Fox, Barbara A. (1987) Discourse Structure and Anaphora: Written and Conversational
English, Cambridge University Press, Cambridge, MA.

Ge, Niyu, John Hale, and Eugene Charniak (1998) A statistical approach to anaphora

resolution, in Proceedings of the Sixth Workshop on Very Large Corpora.

Gordon, Peter C., Barbara J. Grosz, and Laura A. Gilliom (1993) Pronouns, names and
the centering of attention in discourse, Cognitive Science 17:3, 311348.

230
Grice, Herbert Paul (1985) Logic and conversation, in A. P. Martinich, ed., The Philos-
ophy of Language, Oxford University Press, Oxford, 159170.

Grimshaw, Jane (1997) Projection, heads, and optimality, Linguistic Inquiry 28:3, 373
422.

Grimshaw, Jane, and Vieri Samek-Lodovici (1998) Optimal subjects and subject univer-
sals, in Pesetsky et al., ed., Is the Best Good Enough? Optimality and Competition in
Syntax, MIT Press, Cambridge, Mass., 193219.

Grosz, Barbara J. (1977) The representation and use of focus in a system for under-
standing dialogs, in Proceedings of the 5th International joint conference on artificial
intelligence, volume 1, Cambridge, MA, 6776.

Grosz, Barbara J., Aravind K. Joshi, and Scott Weinstein (1983) Providing a unified ac-
count of definite noun phrases in discourse, in Proceedings of the 21st Annual Meeting

of the ACL. Association of Computational Linguistics, Cambridge, Massachusetts, 44

50.

Grosz, Barbara J., Aravind K. Joshi, and Scott Weinstein (1986) Towards a computational
theory of discourse interpretation, unpublished manuscript.

Grosz, Barbara J., Aravind K. Joshi, and Scott Weinstein (1995) Centering: a framework
for modelling the local coherence of discourse, Computational Linguistics 21:2, 203
226.

Grosz, Barbara J., and Candace L. Sidner (1986) Attention, intentions, and the structure
of discourse, Computational Linguistics 12:3, 175204.

Gundel, Jeanette K. (1998) Centering theory and the givenness hierarchy: A proposed
synthesis, in M. A. Walker, A. K. Joshi, and E. Prince, eds., Centering Theory in
Discourse, Oxford University Press.

231
Gundel, Jeanette K., Nancy Hedberg, and Ron Zacharski (1993) Cognitive status and the
form of referring expressions in discourse, Language 69:2, 274307.

Hahn, Udo, and Michael Strube (1997) Centering in-the-large: computing referential
discourse segments, in Proceedings of the 35th Annual Meeting of the Association for
Computational Linguistics and of the 8th Conference of the European Chapter of the
Association for Computational Linguistics, Madrid, Spain, 104111.

Han, Chung-hye, and Na-Rae Han (2001) Part of speech tagging guidelines for Penn
Korean treebank, Technical report, IRCS, University of Pennsylvania.

Han, Chung-hye, Na-Rae Han, and Eon-Suk Ko (2001) Bracketing guidelines for Penn
Korean treebank, Technical report, IRCS, University of Pennsylvania.

Han, Chung-hye, Na-Rae Han, Eon-Suk Ko, and Martha Palmer (2002a) Development
and evaluation of a Korean treebank and its application to NLP, in Proceedings of the

3rd International Conference on Language Resources and Evaluation (LREC-2002).

Han, Chung-hye, Na-Rae Han, Eon-Suk Ko, Martha Palmer, and Heejong Yi (2002b)

Penn Korean treebank: development and evaluation, in Proceedings of PACLIC 16,

University of Pennsylvania.

Han, Na-Rae, and Shijong Ryu (2005) Guidelines for Penn Korean Treebank version 2.0,
Technical report, IRCS, University of Pennsylvania.

Harbert, Wayne (1995) Binding theory, control and pro, in Gert Webelhuth, ed., Gov-
ernment and Binding Theory and the Minimalist Program, Blackwell.

Heeman, Peter A., and Graeme Hirst (1995) Collaborating on referring expressions,
Computational Linguistics 21:3.

Heim, Irene (1982) The semantics of definite and indefinite noun phrases, Doctoral disser-

tation, University of Massachusetts, Amherst.

232
Hendriks, Petra, and Helen de Hoop (2001) Optimality theoretic semantics, Linguistics
and Philosophy 24, 132.

Hobbs, Jerry (1976) Pronoun resolution, Technical report, City College, City University
of New York.

Hobbs, Jerry (1977) Examples of elusive antecedents, Research Report 77-2, City Col-
lege, City University of New York.

Hobbs, Jerry (1978) Resolving pronoun references, Lingua 44, 311338.

Hong, Minpyo (2002a) Optimality Theory in Semantics and the Anaphora Resolution in
Korean: An Adumbration, Language and Information 6:2, 129152.

Hong, Minpyo (2002b) A Review on Zero Anaphora Resolution Theories in Korean,

Studies in Modern Grammar 29, 167186.

Hoonchamlong, Yuphaphann (1991) Some issues in Thai anaphora: a government and

binding approach, Doctoral dissertation, University of Wisconsin-Madison.

Huang, Cheng-Te James (1983) A note on binding theory, Linguistic Inquiry 14, 554

561.

Huang, Cheng-Te James (1984) On the distribution and reference of empty pronouns,

Linguistic Inquiry 15, 531574.

Huang, Cheng-Te James (1989) Pro drop in Chinese: a generalized control approach,

in Osvaldo Jaeggli and Kenneth J. Safir, eds., The Null Subject Parameter, Studies in
natural language and linguistic theory 15, Reidel, Dordrecht, The Netherlands, 185
214.

Huang, Cheng-Te James (1991) Movement structures and bounding theory, university
of California, Irvine.

Huang, Cheng-Te James (1992) Complex predicates in control, in Richard K. Larson,

Sabine Iatridou, Utpal Lahiri, and James Higginbotham, eds., Control and Grammar,

233
 Studies in linguistics and philosophy 48, Kluwer, Dordrecht, The Netherlands, 109
147.

Joshi, Aravind K., and Scott Weinstein (1981) Control of inference: role of some aspects
of discourse structure and centering, in proceedings of the 7th international joint

conference on artificial intelligence, Vancouver, B.C., 385387.

Kameyama, Megumi (1985) Zero Anaphora: The Case of Japanese, Doctoral disserta-
tion, Stanford University. Available from University Microfilms at the University of
Michigan.

Kameyama, Megumi (1986) A Property-sharing constraint in centering, in Proceedings

of the 24th Annual Meeting of the Association of Computational Linguistics, New York,
NY, 200.

Kameyama, Megumi (1996) Indefeasible semantics and defeasible pragmatics, in Quan-

tifiers, Deduction, and Context, CSLI, Stanford, CA.

Kameyama, Megumi (1997) Recognizing referential links: an information extraction

perspective, Technical report, AI Center, SRI International.

Kameyama, Megumi (1998) Intrasentential centering: a case study, in Marilyn A.

Walker, Aravind K. Joshi, and Ellen F. Prince, eds., Centering Theory in Discourse,

Oxford University Press, Oxford, U.K., 89112.

Kamp, Hans, and Uwe Reyle (1993) From Discourse to Logic, Kluwer, Dordrecht, Hol-
land.

Kaplan, David (1989) Demonstratives, in Joseph Almog, John Perry, and Howard
Wettstein, eds., Themes From Kaplan, Oxford University Press, 481563.

Kayne, Richard (1975) French Syntax, MIT Press, Cambridge: Massachusetts.

Kehler, Andrew (1997) Current theories of centering for pronoun interpretation: a critical
evaluation, Computational Linguistics 23:3, 467475.

234
Kennedy, Christopher, and Branimir Boguraev (1996a) Anaphora for everyone: pronom-
inal anaphora resolution without a parser, in Proceedings of the 16th International
Conference on Computational Linguistics.

Kennedy, Christopher, and Branimir Boguraev (1996b) Anaphora in a wider context:

tracking discourse referents, in Proceedings of the 12th European Conference on Ar-
tificial Intelligence (ECAI-96).

Kim, Boome (1991) An indefinite analysis of proarb , MITWPL 14, 147163.

Kim, Mi-Young (2003) An optimality approach to the referential interpretation of zero

anaphora in Korean, Doctoral dissertation, Seoul National University, Seoul, Korea.

Kuno, Susumu (1976) Subject, theme, and speakers empathy: a re-examination of rela-
tivization phenomena, in Charles Li and Sandra A. Thompson, eds., Subject and Topic,
Academic Press, New York.

Kuroda, Shige-Yuki (1965) Generative grammatical studies in the Japanese language,

Doctoral dissertation, MIT.

Lappin, Shalom, and Herbert J. Leass (1994) An algorithm for pronominal anaphora,
Computational Linguistics 20:4.

Lee, Chungmin (2000) Contrastive topic in Chinese and Korean, in Proceedings of

International Symposium on Topic and Focus in Chinese, Linguistic Society of Hong

Kong & the Hong Kong Polytechnic University.

Lee, Chungmin (2003) Contrastive topic and/or contrastive focus, in Bill McClure, ed.,

Japanese/Korean Linguistics, volume 12, CSLI, Stanford.

Lee, Ik-Hwan, and Minhaeng Lee (2000) Anaphora resolution and discourse structure: a
controlled information packaging approach, Language and Information 4:1, 6782.

Levinson, Stephen C. (1983) Pragmatics, Cambridge University Press, Great Britain.

235
Li, Charles, and Sandra A. Thompson (1976) Subject and Topic: a New Typology, Aca-
demic Press, New York.

Lyons, John (1977) Semantics, volume 1 and 2, Cambridge University Press.

Lyons, John (1979) Deixis and anaphora, in T. Myers, ed., The Development of Conver-
sation and Discourse, Edinburgh University Press.

Malamud, Sophia A. (2004) Arbitrariness: a definite account, in Proceedings of WC-

CFL23.

Mann, William C., and Sandra A. Thompson (1987) Rhetorical structure theory: a theory
of text organization, Technical Report ISI/RS-87-190, ISI, University of Southern
California.

Manzini, Maria Rita (1983) On control and control theory, Linguistic Inquiry 14, 421
446.

Martin, Samuel (1976) A Reference Grammar of Japanese, Yale University Press.

McCallum, Andrew Kachites (2002) MALLET: A Machine Learning for Language

Toolkit, http://mallet.cs.umass.edu.

Miltsakaki, Eleni (2002) Toward an aposynthesis of topic continuity and intrasentential

anaphora, Computational Linguistics 28:3.

Minsky, Marvin (1975) A framework for representing knowledge, in Patrick Henry

Winston, ed., The Psychology of Computer Vision, McGraw-Hill, New York.

Mitkov, Ruslan (1996) A corpus-based approach to automatic abstracting, in Proceedings

of the AISB Workshop on Language Engineering for Contents Analysis and Information
Retrieval, Brighton, UK.

Mitkov, Ruslan (1997) Pronoun resolution: the practical alternative, in Simon Botley and
Tony McEnery, eds., Discourse Anaphora and Anaphor Resolution, University College
London Press.

236
Mitkov, Ruslan (1998) Robust pronoun resolution with limited knowledge, in Proceed-
ings of COLING-ACL, 869875.

Murata, Masaki, Hitoshi Isahara, and Makoto Nagao (1999) Pronoun resolution in
Japanese sentences using surface expressions and examples, in ACL99 Workshop
on Coreference and Its Applications.

Nakagawa, Hiroshi (1992) Zero pronouns as experiencer in Japanese discourse, in Pro-

ceedings of the Fourteenth International Conference on Computational Linguistics.

Nakaiwa, Hiromi (1997a) Automatic extraction of rules for anaphora resolution of

Japanese zero pronouns from aligned sentence pairs, in Proceedings of the Fifth work-
shop on Very Large Corpora (WVLC-5).

Nakaiwa, Hiromi (1997b) Automatic extraction of rules for anaphora resolution of

Japanese zero pronouns from aligned sentence pairs, in Proceedings of ACL-97/EACL-

97 Workshop on Operational Factors in Practical, Robust, Anaphora Resolution for

Unrestricted Texts.

Nakaiwa, Hiromi, Francis Bond, Takahiro Uekado, and Yayoi Nozawa (1996) Resolving
zero pronouns in texts using textual structure, in Proceedings of the International
Conference on New Methods in Natural Language Processing: NeMLaP-2.

Nakaiwa, Hiromi, and Satoshi Shirai (1996) Anaphora resolution of Japanese zero pro-
nouns with deictic references, in Proceedings of the 16th International Conference on
Computational Linguistics.

Nakaiwa, Hiromi, and Setsuo Yamada (1997) Automatic identification of zero pronouns
and their antecedents within aligned sentence pairs, in Proceedings of the 3rd Annual
Meeting of the Association for Natural Language Processing.

Nasukawa, Tetsuya (1994) Robust method of pronoun resolution using full-text informa-

237
 tion, in Proceedings of the 15th International Conference on Computational Linguis-
tics (COLING94).

No, Young-Kyoon (1991) A centering approach to the *[CASE] [TOPIC] restriction in

Korean, Linguistics 29.

Nunberg, Geoffrey (1979) The non-uniqueness of semantic solutions: polysemy, Lin-

guistics and Philosophy 3:2, 143184.

Okumura, Manabu, and Kouji Tamura (1996) Zero pronoun resolution in Japanese dis-
course based on centering theory, in Proceedings of COLING96, 871876.

Palmer, Martha, Chung-Hye Han, Na-Rae Han, Eon-Suk Ko, Hee-Jong Yi, Alan Lee, Chris
Walker, John Duda, and Nianwen Xue (2002) Korean English Treebank Annotations,
Linguistic Data Consortium (LDC) catalog number LDC2002T26 and ISBN 1-58563-
236-8.

Palmer, Martha, Shijong Ryu, Jinyoung Choi, Sinwon Yoon, and Yeongmi Jeon (2006)
Korean Propbank, Linguistic Data Consortium (LDC) catalog number LDC2006T03

and ISBN 1-58563-374-7.

Palmer, Martha, Carl Weir, Rebecca J. Passonneau, and Tim Finin (1993) The KERNEL

text understanding system, Artificial Intelligence 63, 1968.

Park, Hyun-Seok, Dania Egedi, and Martha Palmer (1994) Recovering empty arguments

in Korean, in Proceedings of the Joint Conference of the 8th Asian Conference on Lan-
guage, Information, and Computation and the 2nd Pacific Asia Conference on Formal
and Computational Linguistics.

Paul, Michael, Kazuhide Yamamoto, and Eiichiro Sumita (1999) Corpus-based anaphora
resolution towards antecedent preference, in Proceedings of the 37th ACL, Workshop
Coreference and Its Applications.

238
Perlmutter, David (1971) Deep and Surface Structure Constraints in Syntax, Holt, Rinehart
& Winston, New York.

Poesio, Massimo, Renata Vieira, and Simone Teufel (1997) Resolving bridging descrip-
tions in unrestricted text, in Proceedings of. ACL-97 Workshop on Operational Factors

in Practical, Robust, Anaphora Resolution For Unrestricted Texts, 16.

Prasad, Rashmi (2000) A corpus study of zero pronouns in Hindi: an account based on
centering transition preferences, in Proceedings of Discourse Anaphora and Reference
Resolution Conference (DAARC).

Prasad, Rashmi, and Michael Strube (2000) Discourse salience and pronoun resolution in
Hindi, Penn Working Papers in Linguistics 6.3.

Prince, Alan, and Paul Smolensky (1993) Optimality theory: constraint interaction in gen-
erative grammar, Technical report, Rutgers University Center for Cognitive Science.

Prince, Ellen (2003) The Yiddish impersonal pronoun men one in discourse, presented

at the Annual Meeting of the LSA, Atlanta.

Prince, Ellen F. (1981) Towards a taxonomy of given-new information, in Peter Cole,

ed., Radical Pragmatics, Academic Press, New York, NY, 223255.

Prince, Ellen F. (1992) The ZPG letter: subjects, definiteness, and information-status,

in Discourse Description: Diverse Linguistic Analyses of a Fund-raising Text, John

Benjamins, Amsterdam, 295325.

Prince, Ellen F. (1998) Subject-prodrop in Yiddish, in P. Bosch and R. van der Sandt,
eds., Focus: Linguistic, Cognitive, and Computational Perspectives, Cambridge Univ.
Press, 82104.

Radford, Andrew (1981) Transformational Syntax, Cambridge University Press.

Rambow, Owen (1993) Pragmatic aspects of scrambling and topicalization in German: a

centering approach, manuscript, University of Pennsylvania.

239
Ratnaparkhi, Adwait (1997) A simple introduction to Maximum Entropy models for nat-
ural language processing, Technical report, IRCS, University of Pennsylvania.

Ratnaparkhi, Adwait (1998) Maximum Entropy Models for Natural Language Ambiguity
Resolution, Doctoral dissertation, University of Pennsylvania.

Rizzi, Luigi (1986) Null objects in Italian and the theory of pro, Linguistic Inquiry 17,
501557.

Safir, Kenneth (1985) Syntactic Chains, Cambridge University Press, Cambridge.

Samek-Lodovici, Vieri (1996) Constraints on Subjects: An Optimality Theoretic Analysis,

Doctoral dissertation, Rutgers University, New Brunswick, NJ.

Sidner, Candace L. (1979) Towards a computational theory of definite anaphora com-

prehension in English discourse, Technical Report TR-537, MIT AI Lab, Cambridge,
Massachusetts.

Speas, Margaret (1997) Optimality theory and syntax: movement and pronunciation, in
D. et al. Archangeli, ed., Optimality Theory, Blackwell.

Speas, Margaret (2001) Constraints on null pronouns, in Geraldine Legendre et al., ed.,
Optimality-Theoretic Syntax, The MIT Press.

Sperber, Dan, and Deidre Wilson (1986) Relevance: Communication and Cognition,
Blackwell, Oxford, UK.

Sportiche, Dominique (1983) Structural Invariance and Symmetry in Syntax, Doctoral

dissertation, MIT.

Stirling, Lesley, and Rodney Huddleston (2002) Deixis and Anaphora, chapter 17 in
Rodney Huddleston and Geoff Pullum, eds., The Cambridge Grammar of the English
Language, Cambridge University Press, 14491564.

Strube, Michael (1998) Never look back: an alternative to centering, in Proceedings of

COLING-ACL 98, 12511257.

240
Sturbe, Michael, and Udo Hahn (1999) Functional centering grounding referential co-
herence in information structure, Computational Linguistics 25:309344.

Takada, Shingo, and Norihisa Doi (1994) Centering in Japanese: a step towards better
interpretation of pronouns and zero-pronouns, in Proceedings of the 15th conference
on Computational linguistics.

Taraldsen, Knut Tarald (1980) The theoretical interpretation of a class of

markedextractions, in Proceedings of the Third GLOW Conference.

Tetreault, Joel R. (1999) Analysis of syntax-based pronoun resolution methods, in Pro-

ceedings of ACL 99.

Tetreault, Joel R. (2001) A corpus-based evaluation of centering and pronoun resolution,

Computational Linguistics 27:4.

Tsao, Feng-fu (1977) A functional study of topic in Chinese: the first step toward discourse,
Doctoral dissertation, USC, Los Angeles, California.

Turan, Umit D. (1995) Null vs. Overt Subjects in Turkish Discourse: A Centering Analysis,
Doctoral dissertation, University of Pennsylvania.

Turan, Umit D. (1998) Ranking forward-looking centers in Turkish: universal and lan-
guage specific properties, in Marilyn A. Walker, Aravind K. Joshi, and Ellen F. Prince,
eds., Centering Theory in Discourse, Oxford University Press, Oxford, 138160.

Vallduv, Enric (1990) The Informational Component, Doctoral dissertation, University of

Pennsylvania, Philadelphia, PA.

Vallduv, Enric (1992) The Informational Component, Garland, New York.

Vallduv, Enric (1994) The dynamics of information packaging, in Elisabet Engdahl,

ed., Integrating Information Structure into Constraint-Based Categorial Approaches,
HCRC publications, University of Edinburgh, Edinburgh, 426.

241
Vallduv, Enric, and Elisabet Engdahl (1996) The linguistic realization of information
packaging, Linguistics 34, 459519.

Vieira, Renata, and Massimo Poesio (2000) Processing definite descriptions in corpora,
in S. Botley and T. McEnery, eds., Corpus-based and Computational Approaches to

Anaphora, UCL Press.

Walker, Marilyn A. (1989) Evaluating discourse processing algorithms, in Proceedings

of 27th Annual Meeting of the Association of Computational Linguistics.

Walker, Marilyn A. (1993) Redundant affirmation, deliberation and discourse, Penn

Review of Linguistics.

Walker, Marilyn A. (1998) Centering, anaphora resolution, and discourse structure, in

Marilyn A. Walker, Aravind K. Joshi, and Ellen F. Prince, eds., Centering Theory in
Discourse, Oxford University Press.

Walker, Marilyn A., Masayo Iida, and Sharon Cote (1990) Centering in Japanese dis-

course, in Proceedings of the 13th International Conference on Computational Lin-

guistics, Helsinki.

Walker, Marilyn A., Aravind K. Joshi, and Ellen F. Prince (1998) Centering in naturally
occurring discourse: an overview, in Marilyn A. Walker, Aravind K. Joshi, and Ellen F.

Prince, eds., Centering Theory in Discourse, Oxford University Press, Oxford, 128.

Walker, Marilyn A., Masayo Iida, and Sharon Cote (1994) Japanese discourse and the
process of centering, Computational Linguistics 20:2, 193232.

Webber, Bonnie Lynn (1991) Structure and ostension in the interpretation of discourse
deixis, Language and Cognitive Processes 6:2, 107135.

Winograd, Terry (1972) Understanding Natural Language, Academic Press, New York.

Yoon, James Hye-Suk (1985) On the treatment of empty categories in topic prominent
languages, manuscript, University of Illinois.

242
Yoshimoto, Kei (1988) Identifying zero pronouns in Japanese dialogue, in Proceedings
of COLING-88, 779784.

p

^ %



(1994) G#
D Q{
o

_d
 
o, Masters thesis, "
f@
 /

< .

(2001)

 G#
D Q{
o


_
d
 
o
<%

+
A 6
x

& , Masters thesis, "
f@
 /

<
.

e
s O
[, s
 %
, G
3 a (1997)
-
 G_
D 
#
 Q,


H
 o
.

<

e
 ,

"
(1995)
G#
D Q
HZ
 
O r 1,
:
 G~
D 5
x
x
:@
 /

< .


3  (1986) 
 o6x:
r

 ,

1
 
.


3  (1994) :
 x

+ HZ
 
O r- {
: o

<o6
x,

1
 
.

|

r, 5
x

, ~
F
1
q (1997)
p
 G#
D Q/
@6x

t|
q K

`
 0
 A
G
p'
' a
As
r_
: &

h

x, in ]

6 j 9
r/
Jx

9
 G#
D Q&

%o/
@r

< t@
/
r1

H
7 H
|,
9 @
/

< .

jF

, s
 '
 &
(1999) , in y
h #

3 et al., ed., +
Ad
_
p
r
: 

G#
D Ql
t,

 


H

o
, "
f, 157205.



< 
 (1999)

 G#
D Q%

+@
A /"


<
Gp'
' a
As
r, in
:
 G
D #
 Q&


< r.


< 
 (2000) G
p'
' a
As
r
: @
/o^
\f
"_
H
7
t|
q

& ,
t
 
< 11:2.

243