Improving Japanese Zero Pronoun Resolution
by Global Word Sense Disambiguation
Daisuke Kawahara and Sadao Kurohashi
Graduate School of Information Science and Technology, University of Tokyo
{kawahara,kuro}@kc.t.u-tokyo.ac.jp
Abstract
This paper proposes unsupervised word
sense disambiguation based on automati-
cally constructed case frames and its in-
corporation into our zero pronoun resolu-
tion system. The word sense disambigua-
tion is applied to verbs and nouns. We
consider that case frames define verb senses
and semantic features in a thesaurus define
noun senses, respectively, and perform sense
disambiguation by selecting them based on
case analysis. In addition, according to the
one sense per discourse heuristic, the word
sense disambiguation results are cached and
applied globally to the subsequent words.
We integrated this global word sense disam-
biguation into our zero pronoun resolution
system, and conducted experiments of zero
pronoun resolution on two different domain
corpora. Both of the experimental results
indicated the effectiveness of our approach.
1 Introduction
For a long time, parsing has been a central is-
sue for the area of natural language analyses.
In recent years, its accuracy has improved to
over 90%, and it became the fundamental tech-
nology that is applied to a lot of NLP applica-
tions, such as question answering, text summa-
rization, and machine translation. Accordingly,
anaphora resolution, which is positioned as the
next step of parsing, has been studied actively
(Ng and Cardie, 2002; Yang et al., 2003; Iida
et al., 2003; Isozaki and Hirao, 2003; Kawa-
hara and Kurohashi, 2004). Its performance,
however, is not satisfactory enough to benefit
the NLP applications. We investigated errors
of our Japanese zero pronoun resolution system
(Kawahara and Kurohashi, 2004), and found
that word sense ambiguity causes a major part
of errors.
Our zero pronoun resolution system uti-
lizesthegeneral-purposethesaurusNihongo Goi
Taikei (Ikehara et al., 1997) (hereafter, NTT
thesaurus) to do matching of example words.
In this thesaurus, one or more semantic features
are given to each word, and similarity between
words can be calculated by comparing closeness
of their semantic features in the thesaurus tree
(Appendix A). Multiple semantic features for
a word, i.e. word sense ambiguity, cause in-
correct matching, and furthermore deteriorate
accuracy of the zero pronoun resolution sys-
tem. For instance, in the thesaurus, “gobou”
(burdock/priest/temple) has four semantic fea-
tures: <crop>, <vegetable>, <priest> and
<temple>∗. If <priest> is used for “gobou” in
a cooking domain text, though “gobou” means
‘burdock’ (a genus of coarse biennial herbs) in
its context, “gobou” is identified as an agent.
That is, “gobou” is incorrectly analyzed as an-
tecedents of the following nominative zero pro-
nouns.
If such word sense ambiguity can be resolved,
incorrect matching decreases, and the anaphora
resolution system will improve. Word sense dis-
ambiguation is a basic issue of NLP. Recently,
the evaluation exercises for word sense disam-
biguation such as SENSEVAL-1 (Kilgarriff and
Palmer, 2000) and SENSEVAL-2 (Yarowsky,
2001) have been held, but word sense disam-
biguation has rarely been incorporated into
deep analyses like anaphora resolution.
This paper proposes unsupervised word sense
disambiguation based on automatically con-
structed case frames and its incorporation into
the zero pronoun resolution system. The word
sense disambiguation is applied to verbs† and
nouns. We consider that case frames define
verb senses and semantic features in the the-
saurus define noun senses, respectively. A verb
is disambiguated by selecting a corresponding
case frame to its context, and a noun is disam-
biguated by selecting an appropriate semantic
∗In this paper, <> means a semantic feature.
†In this paper, we use ‘verb’ instead of ‘verb, adjective
and noun+copula’ for simplicity.
NOUN
CONCRETE ABSTRACT
AGENT PLACE CONCRETE
HUMAN ORGANIZATION
ABSTRACTEVENTABSTRACT RELATION
TIMEPOSITIONQUANTITY. . . .
Figure 1: The upper levels of the NTT thesaurus.
feature from the ones defined for the noun in
the thesaurus. In addition, according to the one
sense per discourse heuristic, the disambigua-
tion results are cached and applied globally to
the following words in the same text.
The remainder of this paper is organized as
follows. Section2brieflydescribestheNTTthe-
saurus and the automatic construction method
of case frames. Section 3 outlines our zero pro-
noun resolution system. Section 4 describes the
method of word sense disambiguation and its
integration into the zero pronoun resolution sys-
tem. Section 5 presents the experiments of the
integrated system. Section 6 summarizes the
conclusions.
2 Resources
We consider that verb and noun senses corre-
spond to case frames and semantic features de-
fined in the NTT thesaurus, respectively. This
section describes the NTT thesaurus and the
case frames briefly.
2.1 NTT thesaurus
NTT Communication Science Laboratories con-
structed a semantic feature tree, whose 3,000
nodes are semantic features, and a nominal dic-
tionary containing about 300,000 nouns, each of
which is given one or more appropriate seman-
tic features. Figure 1 shows the upper levels of
the semantic feature tree.
The similarity between two words is defined
by formula (1) in Appendix A.
2.2 Automatically constructed case
frames
We employ the automatically constructed case
frames (Kawahara and Kurohashi, 2002) as
the basic resource for zero pronoun resolution
and word sense disambiguation. This section
outlines the method of constructing the case
frames.
The biggest problem in automatic case frame
construction is verb sense ambiguity. Verbs
which have different meanings should have dif-
ferent case frames, but it is hard to dis-
ambiguate verb senses precisely. To deal
with this problem, predicate-argument exam-
ples which are collected from a large cor-
pus are distinguished by coupling a verb and
its closest case component. That is, ex-
amples are not distinguished by verbs (e.g.
“tsumu” (load/accumulate)), but by couples
(e.g. “nimotsu-wo tsumu” (load baggage) and
“keiken-wo tsumu” (accumulate experience)).
This process makes separate case frames
which have almost the same meaning or usage.
For example, “nimotsu-wo tsumu” (load bag-
gage) and “busshi-wo tsumu” (load supply) are
similar, but have separate case frames. To cope
with this problem, the case frames are clustered.
Example words are collected for each case
marker, such as “ga”, “wo”, “ni” and
“kara”. They are case-marking postpositions
in Japanese, and usually mean nominative, ac-
cusative, dative and ablative, respectively. We
call such a case marker ‘case slot’ and example
words in a case slot ‘case examples’.
Case examples in a case slot are similar, but
have some incorrect semantic features because
of word sense ambiguity. For instance, “ni-
motsu” (baggage), “busshi” (supply) and “nise-
mono” (imitation) are gathered in a case slot,
and all of them are below the semantic feature
<goods>. On the other hand, “nisemono” be-
longs to <lie>. <lie> is incorrect for this case
slot, and possibly causes errors in case analysis.
We delete a semantic feature that is not similar
to the other semantic features of its case slot.
To sum up, the procedure for the automatic
case frame construction is as follows.
1. A large raw corpus is parsed by the
Japanese parser, KNP (Kurohashi and
Nagao, 1994b), and reliable predicate-
argument examples are extracted from the
parse results.
2. The extracted examples are bundled ac-
cordingtotheverbanditsclosestcasecom-
ponent, making initial case frames.
3. The initial case frames are clustered using a
similarity measure function. This similar-
ity is calculated by formula (5) in Appendix
B.
4. For each case slot of clustered case frames,
an inappropriate semantic feature that is
not similar to the other semantic features
is discarded.
We constructed two sets of case frames: for
newspaper and cooking domain.
The newspaper case frames are constructed
from about 21,000,000 sentences of newspaper
articles in 20 years (9 years of Mainichi news-
paper and 11 years of Nihonkeizai newspaper).
They consist of 23,000 verbs, and the average
number of case frames for a verb is 14.5.
The cooking case frames are constructed from
about 5,000,000 sentences of cooking domain
that are collected from WWW. They consist
of 5,600 verbs, and the average number of case
frames for a verb is 6.8.
In Figure 1, some examples of the resulting
case frames are shown. In this table, ‘CS’ means
a case slot. <agent> in the table is a general-
ized case example, which is given to the case
slot where half of the case examples belong to
<agent>. <agent> is also given to “ga” case
slot that has no case examples, because “ga”
case components are often omitted, but “ga”
case slots usually mean nominative.
3 The Outline of the Zero Pronoun
Resolution System
We have proposed a Japanese zero pronoun
resolution system using the case frames, an-
tecedent preference orders, and a machine learn-
ing technique (Kawahara and Kurohashi, 2004).
Its procedure is as follows.
1. Parse an input sentence using the Japanese
parser, KNP.
2. Process each verb in the sentence from left
to right by the following steps.
Table 1: Case frame examples.
CS case examples†
ga <agent>, group, party, ···youritsu (1)
wo <agent>, candidate, applicant(support)
ni <agent>, district, election, ···
ga <agent>youritsu (2)
wo <agent>, member, minister, ···(support)
ni <agent>, candidate, successor
... ... ...
orosu (1) ga <agent>
(grate) wo radish
ga <agent>orosu (2)
wo money(withdraw)
kara bank, post
... ... ...
itadaku (1) ga <agent>
(have) wo soup
ga <agent>itadaku (2)
wo advice, instruction, address(be given)
kara <agent>, president, circle, ···
... ... ...
†case examples are expressed only in English for
space limitation.
2.1. Narrow case frames down to corre-
sponding ones to the verb and its clos-
est case component.
2.2. Perform the following processes for
each case frame of the target verb.
i. Match each input case component
with an appropriate case slot of
the case frame. Regard case slots
that have no correspondence as
zero pronouns.
ii. Estimate an antecedent of each
zero pronoun.
2.3. Select a case frame which has the high-
est total score, and output the analysis
result for the case frame.
The rest of this section describes the above
steps (2.1), (2.2.i) and (2.2.ii) in detail.
3.1 Narrowing down case frames
The closest case component plays an important
role to determine the usage of a verb. In par-
ticular, when the closest case is “wo” or “ni”,
this trend is clear-cut. In addition, an expres-
sion whose nominative belongs to <agent> (e.g.
“<agent> has accomplished”), does not have
enough clue to decide its usage, namely a case
frame. By considering these aspects, we impose
the following conditions on narrowing down case
frames.
• The closest case component exists, and
must immediately precede its verb.
• The closest case component and the closest
case meet one of the following conditions:
– The closest case is “wo” or “ni”.
– The closest case component does
not belong to the semantic marker
<agent>.
• A case frame with the closest case exists,
and the similarity between the closest case
component and examples in the closest case
exceeds a threshold.
We choose the case frames whose similarity
is the highest. If the above conditions are not
satisfied, case frames are not narrowed down,
and the subsequent processes are performed for
each case frame of the target verb. The simi-
larity used here is defined as the best similarity
between the closest case component and exam-
ples in the case slot. The similarity between two
examples is defined as formula (1) in Appendix
A.
Let us consider “youritsu” (support) in the
second sentence of Figure 2. “youritsu” has the
case frames shown in Table 1. The input expres-
sion “kouho-wo youritsu” (support a candidate)
satisfies the above two conditions, and the case
frame “youritsu (1)” meets the last condition.
Accordingly, this case frame is selected.
3.2 Matching input case components
with case slots in the case frame
We match case components of the target verb
with case slots in the case frame (Kurohashi and
Nagao, 1994a). When a case component has a
case marker, it must be assigned to the case
slot with the same case marker. When a case
component is a topic marked phrase or a clausal
modifiee, which does not have a case marker, it
can be assigned to one of the case slots in the
following table.
topic marked phrases : ga, wo, ga2
clausal modifiees : ga, wo, non-gapping
The conditions above may produce multiple
matching patterns. In this case, one which has
the best score is selected. The score of a match-
ing pattern is defined as the sum of similarities
of case assignments. This similarity is calcu-
lated as the same way described in Section 3.1.
a0a2a1a4a3a6a5a4a7a9a8a4a10a11 a1a13a12
a7a6a5a13a10a7a6a1a13a14
a15a17a16a19a18 a20
a18
a11
a18a4a15a21a16a19a18
a22
a10 a14a4a7a24a23a26a25a4a11
a18
a25
a15a17a16
a5
a20
a1a13a14a27a11
a18
a25
a15
a7a9a8a26a10a11 a1a13a10a28a17a25
a7a9a8a4a10a29a30a10
a15
a7a13a25a31a28a17a25
a32a34a33
a16a19a18
a23
a18
a25a31a28a35a10a11 a7a31a25
a15
a7a31a25a31a28a17a25
a36
a18
a20
a25a30a29a21a10
a20
a18
a25a31a8
a18a4a15a17a16a19a18
a32a34a33a35a37
a5
a32a34a33
a14a26a10
a38
a7a9a8a26a10a8a26a5a4a28a39a5
a40a41a8a26a10a29a21a10
a15
a37
a5
a32a34a33a35a37
a5
a37
a5a43a42a31a44a6a14a4a10a45
a37
a5a31a46a47a44a6a14a4a10a48
a37
a5a13a48a49a44a6a14a26a10a42
a37
a5a4a45a49a44a9a14a26a10a50
a37
a5a4a50a27a44a6a14a26a10a51
a8a31a23
a18
a25a13a0a52a1a4a10
a15
a7a9a8a4a10a11 a5
a28a35a10
a20a9a20
a18
a25a31a8
a18a4a15a17a16a19a18
a32a34a33a35a37
a5 a40a41a8a26a10a29a21a10a7a6a1a4a14
a15
a8a53a1a13a14
a18
a8
a18
a25a31a7a6a8a4a10a14
a15a17a16a19a18
a20
a1a26a11a11 a1a13a10
a15
a7a6a8a4a10a11 a5
a37
a5a31a12
a54
a10 a0a52a10a14a53a11
a18
a25
a15a21a16
a5
a55a6a56
a55a13a57a35a58
a55a4a57a60a59
a55a4a57a62a61
a55a4a57a60a63
Figure 2: Analysis example.
The result of case analysis tells if the zero
pronouns exist. That is, vacant case slots in
the case frame, which have no correspondence
with the input case components, mean zero pro-
nouns. In this paper, we concentrate on three
case slots: “ga”, “wo”, and “ni”.
In the case of “youritsu” (support) in Figure
2 and the selected case frame “youritsu (1)”,
“wo” case slot has a corresponding case compo-
nent, but “ga” and “ni” case slots are vacant.
Accordingly, two zero pronouns are identified in
“ga” and “ni” case of “youritsu”.
3.3 Antecedent estimation
The antecedents of the detected zero pronouns
are estimated. Possible antecedents are exam-
ined according to the antecedent preference or-
der (Kawahara and Kurohashi, 2004). If a pos-
sible antecedent is classified as positive by a bi-
nary classifier and its similarity to examples in
its case slot exceeds a threshold, it is determined
as the antecedent.
For example, “youritsu” (support) in Fig-
ure 2 has zero pronouns in “ga” and “ni”
cases. The ordered possible antecedents for
“ga” are L7:“Minsyutou”, L14:“Jimintou”(φ
ga), L14:“Ishihara chiji”(φ wo), ···. The first
candidate “Minsyutou (similarity:0.73)”, which
is labeled as positive by the classifier, and whose
similarity to the case frame examples exceeds
a threshold (0.60), is determined as the an-
tecedent.
4 Global Word Sense
Disambiguation
We integrate a global method of word sense dis-
ambiguation into the zero pronoun resolution
system described in the previous section. The
word sense disambiguation is applied to verbs
and nouns based on the case frames. Further-
more, the word sense disambiguation results are
cached and applied globally by the subsequent
analyses based on the one sense per discourse
heuristic. In the rest of this section, we de-
scribe verb and noun sense disambiguation re-
spectively using examples of cooking domain.
4.1 Verb sense disambiguation
The case frames are specific enough to the diver-
sity of verb senses, because their meanings are
distinguished by the couple of the verb and its
closest case component (Kawahara and Kuro-
hashi, 2002). We regard the process of verb
sense disambiguation as the case frame selection
(Step (2.3) described in Section 3). In addition,
theverbsensedisambiguationresultsarecached
and applied globally in the same text. In other
words, the selected case frames are cached for
each verb, and only the case frames that are
similar to the cache are used for the same verb
following in the same text. The similarity mea-
sure for two case frames is stated in Appendix
B, and the threshold is set to 0.60 empirically.
Here is an example article that consists of three
sentences.
oroshi-gane-de
grater
kabura-wo
turnip
oroshite-ikimasu.
grate
(Let’s grate a turnip.)
kore-wa
this
ookii
big
kabura
turnip
desu.
be
(This is a big turnip.)
konoyouni
like this
oroshi-masu.
grate
(Grate like this.)
For “oroshite” (grate) in the first sentence,
the case frame “orosu (1)” (in Table 1), which
means “grate radish”, is selected, because the
closest case component “kabura” (turnip) ex-
ists, and is very similar to the “wo” case exam-
ple “daikon” (radish). This selected case frame
is cached for the verb “orosu”. For “oroshi”
(grate) in the third sentence, case frames are
not narrowed down for lack of the closest com-
ponent. The previous system performs the
antecedent estimation process for all the case
frames of “orosu”, and incorrectly estimates the
antecedent of “wo” zero pronoun as “oroshi-
gane” (grater)‡. On the other hand, our pro-
posed method deals with only the similar case
frames to the cached “orosu (1)”. That is, the
case frame “orosu (2)”, which means “with-
draw money from bank or post”, is not similar
to “orosu (1)”, and is not used. Accordingly,
the system certainly estimates the antecedent
of “wo” zero pronoun as “kabura” (turnip).
4.2 Noun sense disambiguation
We define the process of noun sense disambigua-
tion as selecting an appropriate semantic fea-
ture from the ones given to a noun in the NTT
thesaurus. This process is performed based on
the matching of the input case components and
the case frame decided by the step (2.3) de-
scribed in Section 3. For each input case compo-
nent, its semantic features are matched against
those of case examples of its corresponding case
slot, and the best matched one is selected. In
addition, this disambiguation result is applied
globallyliketheverbsensedisambiguation. The
determined semantic feature is cached for each
noun, and is given to the same noun following in
the same text, instead of reconsidering all of its
semantic features. Here is an example article.
mazuwa
first
osumashi-wo
clear soup
itadaki-masu.
have
(First, let’s have clear soup.)
honkakutekina
real
dashi-wo
stock
tori-mashita.
prepare
(We prepared real stock.)
“itadaki” (have) in the first sentence has
the closest case component “osumashi” (clear
soup), and the case frame “itadaku (1)” (in Ta-
ble 1) is selected, because its “wo” case example
“soup” is very similar to “osumashi”.
In the NTT thesaurus, “osumashi” (clear
soup) has three semantic features: <soup>,
<look> and <eccentric>. <eccentric> is lo-
cated below <agent> in the thesaurus, and
the previous system incorrectly estimates an-
tecedents of “ga” zero pronouns of the follow-
ing verbs as “osumashi” (because almost all the
‡In Japanese, “gane” of “oroshi-gane” (grater) ex-
actly matches with “kane” (money), the “wo” case ex-
ample of “orosu (2)”.
Table 2: Accuracy (newspaper).
precision recall F
baseline 515/924 (0.557) 515/1087 (0.474) 0.512
our method 526/911 (0.577) 526/1087 (0.484) 0.527
Table 3: Accuracy (cooking).
precision recall F
baseline 696/1092 (0.637) 696/1482 (0.470) 0.541
our method 713/1081 (0.660) 713/1482 (0.481) 0.556
case frames have <agent> in their “ga” case
slots). In our approach, each of the semantic
features are matched against the case example
“soup”, and only the best matched semantic
feature <soup> is given to “osumashi”.
5 Experimental Results and
Discussion
We conducted experiments of zero pronoun res-
olution on two different domain corpora. One
is newspaper articles of “Relevance-tagged cor-
pus” (Kawahara et al., 2002), and the other
is utterances of cooking TV programs. These
cooking utterances were handled by (Shibata et
al., 2003). They annotated various relations to
closed captions of the cooking utterances based
on the specification of the “Relevance-tagged
corpus” (Kawahara et al., 2002).
For newspaper domain, the antecedent pref-
erence and the classifier were trained with 1,841
sentences in the newspaper corpus, and the
newspaper case frames were used. The exper-
iment was performed on 633 sentences. For
cooking domain, we used 813 sentences (5 TV
programs), and conducted 5-fold cross valida-
tion using the cooking case frames.
We evaluated “ga”, “wo” and “ni” cases that
are the large majority of zero pronouns. The
experimental results are shown in Table 2 and
Table 3. The accuracies in these tables are cal-
culated by evaluating both detection and an-
tecedent estimation of zero pronouns together.
The baseline corresponds to our previous sys-
tem without word sense disambiguation.
From Table 2 and Table 3, we can see that
the system accuracy is improved by the global
word sense disambiguation. The improvement
is not big, but there are no analysis result that
changed for the worse. The improvement is
hardly contributed by the verb sense disam-
biguation, but mainly by the noun sense dis-
ambiguation. This is because appropriate case
frames are used in many cases without the verb
sense disambiguation, and this process did not
lead to the improvement. The number of case
frames of the verbs to which the verb sense dis-
ambiguation is applied decreased to 16%, and
this indicates that the analysis efficiency im-
proved significantly. In addition, we evaluated
randomly selected 100 nouns to which the noun
sense disambiguation is applied. 91 nouns were
disambiguated correctly, and quite high disam-
biguation accuracy was achieved.
6 Conclusion
This paper has incorporated a framework
of global word sense disambiguation into a
Japanese zero pronoun resolution system. The
word sense disambiguation is applied to verbs
and nouns. A verb is disambiguated by select-
ing a corresponding case frame to its context,
and a noun is disambiguated by selecting an
appropriate semantic feature. Furthermore, the
disambiguation results are cached and applied
globally. That is to say, it is utilized in the fol-
lowing analyses in the same text. In the future,
we will investigate the word sense disambigua-
tion errors further, and expect to improve the
system accuracy.
Appendix
A Similarity between examples
The similarity between two examples e1,e2 is
calculated using the NTT thesaurus as follows:
sim(e1,e2) = maxx∈s1,y∈s2 sim(x,y) (1)
sim(x,y) = 2Ll
x +ly
where x,y are semantic features, and s1,s2 are
sets of semantic markers of e1,e2 respectively.
lx,ly are the depths of x,y in the thesaurus, and
the depth of their lowest (most specific) com-
mon node is L. If x and y are in the same node
of the thesaurus, the similarity is 1.0, the max-
imum score based on this criterion.
B Similarity between case frames
Two case frames, F1 and F2, are first aligned ac-
cording to the agreement of case markers (case
slots). Suppose the result of the case slot align-
ment of F1 and F2 is as follows:
F1 : C11, C12, ··· C1l ··· C1m
arrowbothv arrowbothv arrowbothv
F2 : C21, C22, ··· C2l ··· C2n
where Cxx denotes a case slot which contains
several case examples. This result means that
l case slots are aligned between F1 and F2 and
(m − l) and (n − l) case slots remained in F1
and F2 respectively.
The similarity between two case slots, C1i and
C2i, is the sum of the similarities of case exam-
ples as follows:
sim(C1i,C2i) =P
e1∈C1i
P
e2∈C2i
√
|e1||e2|·sim(e1,e2)P
e1∈C1i
P
e2∈C2i
√
|e1||e2| (2)
where |e1| and |e2| represent the frequencies of
e1 and e2 respectively.
The similarities of case slots are summed up
with the weight of frequencies of case examples
as follows:
WSofCS =P
l
i=1
√
|C1i||C2i|·sim(C1i,C2i)P
l
i=1
√
|C1i||C2i| (3)
where
|C1i| = X
e1∈C1i
|e1|, |C2i| = X
e2∈C2i
|e2|
On the other hand, the ratio of aligned case
slots is calculated as follows:
RofACS =
sP
l
i=1|C1i|P
m
i=1|C1i|
×
Pl
i=1|C2i|P
n
i=1|C2i|
(4)
Finally, the similarity between case frames is
calculated as follows:
sim(F1,F2) = WSofCS×RofACS (5)

References

Ryu Iida, Kentaro Inui, Hiroya Takamura, and Yuji
Matsumoto. 2003. Incorporating contextual cues
in trainable models for coreference resolution. In
Proceedings of the 10th EACL Workshop on The
Computational Treatment of Anaphora, pages 23–30.

Satoru Ikehara, Masahiro Miyazaki, Satoshi Shirai,
Akio Yokoo, Hiromi Nakaiwa, Kentarou Ogura,
and Yoshifumi Oyama Yoshihiko Hayashi, editors.
1997. Japanese Lexicon. Iwanami Publishing.

Hideki Isozaki and Tsutomu Hirao. 2003. Japanese
zero pronoun resolution based on ranking rules
and machine learning. In Proceedings of the 2003
Conference on Empirical Methods in Natural Lan-
guage Processing, pages 184–191.

Daisuke Kawahara and Sadao Kurohashi. 2002.
Fertilization of case frame dictionary for robust
Japanese case analysis. In Proceedings of the 19th
International Conference on Computational Lin-
guistics, pages 425–431.

Daisuke Kawahara and Sadao Kurohashi. 2004.
Zero pronoun resolution based on automatically
constructed case frames and structural preference
of antecedents. In Proceedings of the 1st Inter-
national Joint Conference on Natural Language
Processing, pages 334–341.

Daisuke Kawahara, Sadao Kurohashi, and Kˆoiti
Hasida. 2002. Construction of a Japanese
relevance-tagged corpus. In Proceedings of the 3rd
International Conference on Language Resources
and Evaluation, pages 2008–2013.

Adam Kilgarriff and Martha Palmer. 2000. Intro-
duction to the special issue on SENSEVAL. Com-
puters and the Humanities, 34(1):1–13.

Sadao Kurohashi and Makoto Nagao. 1994a. A
method of case structure analysis for Japanese
sentences based on examples in case frame dic-
tionary. In IEICE Transactions on Information
and Systems, volume E77-D No.2.

Sadao Kurohashi and Makoto Nagao. 1994b. A syn-
tactic analysis method of long Japanese sentences
based on the detection of conjunctive structures.
Computational Linguistics, 20(4):507–534.

Vincent Ng and Claire Cardie. 2002. Improving ma-
chine learning approaches to coreference resolu-
tion. In Proceedings of the 40th Annual Meeting
of the Association for Computational Linguistics,
pages 104–111.

Tomohide Shibata, Daisuke Kawahara, Masashi
Okamoto, Sadao Kurohashi, and Toyoaki Nishida.
2003. Structural analysis of instruction utter-
ances. In Proceedings of Seventh International
Conference on Knowledge-Based Intelligent In-
formation and Engineering Systems (KES2003),
pages 1054–1061.

Xiaofeng Yang, Guodong Zhou, Jian Su, and
Chew Lim Tan. 2003. Coreference resolution us-
ingcompetitionlearningapproach. InProceedings
of the 41st Annual Meeting of the Association for
Computational Linguistics, pages 176–183.

David Yarowsky, editor. 2001. SENSEVAL-2: Sec-
ond International Workshop on Evaluating Word
Sense Disambiguating Systems. The Association
for Computational Linguistics.
