Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 176–183,
New York, June 2006. c©2006 Association for Computational Linguistics
A Fully-Lexicalized Probabilistic Model
for Japanese Syntactic and Case Structure Analysis
Daisuke Kawahara∗and Sadao Kurohashi†
Graduate School of Information Science and Technology, University of Tokyo
7-3-1 Hongo Bunkyo-ku, Tokyo, 113-8656, Japan
{kawahara,kuro}@kc.t.u-tokyo.ac.jp
Abstract
We present an integrated probabilistic
model for Japanese syntactic and case
structure analysis. Syntactic and case
structure are simultaneously analyzed
based on wide-coverage case frames that
are constructed from a huge raw corpus in
an unsupervised manner. This model se-
lects the syntactic and case structure that
has the highest generative probability. We
evaluate both syntactic structure and case
structure. In particular, the experimen-
tal results for syntactic analysis on web
sentences show that the proposed model
significantly outperforms known syntactic
analyzers.
1 Introduction
Case structure (predicate-argument structure or log-
ical form) represents what arguments are related to
a predicate, and forms a basic unit for conveying the
meaning of natural language text. Identifying such
case structure plays an important role in natural lan-
guage understanding.
In English, syntactic case structure can be mostly
derived from word order. For example, the left ar-
gument of the predicate is the subject, and the right
argument of the predicate is the object in most cases.
Blaheta and Charniak proposed a statistical method
∗Currently, National Institute of Information and Communi-
cations Technology, JAPAN, dk@nict.go.jp
†Currently, Graduate School of Informatics, Kyoto Univer-
sity, kuro@i.kyoto-u.ac.jp
for analyzing function tags in Penn Treebank, and
achieved a really high accuracy of 95.7% for syn-
tactic roles, such as SBJ (subject) and DTV (da-
tive) (Blaheta and Charniak, 2000). In recent years,
there have been many studies on semantic structure
analysis (semantic role labeling) based on PropBank
(Kingsbury et al., 2002) and FrameNet (Baker et al.,
1998). These studies classify syntactic roles into se-
mantic ones such as agent, experiencer and instru-
ment.
Case structure analysis of Japanese is very differ-
ent from that of English. In Japanese, postpositions
are used to mark cases. Frequently used postposi-
tions are “ga”, “wo” and “ni”, which usually mean
nominative, accusative and dative. However, when
an argument is followed by the topic-marking post-
position “wa”, its case marker is hidden. In addi-
tion,case-markingpostpositionsareoftenomittedin
Japanese. These troublesome characteristics make
Japanese case structure analysis very difficult.
To address these problems and realize Japanese
case structure analysis, wide-coverage case frames
are required. For example, let us describe how to
apply case structure analysis to the following sen-
tence:
bentou-wa taberu
lunchbox-TM eat
(eat lunchbox)
In this sentence, taberu (eat) is a verb, and bentou-
wa (lunchbox-TM) is a case component (i.e. argu-
ment) of taberu. The case marker of “bentou-wa”
is hidden by the topic marker (TM) “wa”. The an-
alyzer matches “bentou” (lunchbox) with the most
176
suitable case slot (CS) in the following case frame
of “taberu” (eat).
CS examples
taberu ga person, child, boy, ···wo lunch, lunchbox, dinner, ···
Since “bentou” (lunchbox) is included in “wo” ex-
amples, its case is analyzed as “wo”. As a result, we
obtain the case structure “φ:ga bentou:wo taberu”,
which means that “ga” (nominative) argument is
omitted,and“wo”(accusative)argumentis“bentou”
(lunchbox). In this paper, we run such case structure
analysis based on example-based case frames that
are constructed from a huge raw corpus in an unsu-
pervised manner.
Let us consider syntactic analysis, into which our
method of case structure analysis is integrated. Re-
cently, many accurate statistical parsers have been
proposed (e.g., (Collins, 1999; Charniak, 2000) for
English, (Uchimoto et al., 2000; Kudo and Mat-
sumoto, 2002) for Japanese). Since they somehow
uselexicalinformationinthetaggedcorpus,theyare
called “lexicalized parsers”. On the other hand, un-
lexicalizedparsersachievedanalmostequivalentac-
curacy to such lexicalized parsers (Klein and Man-
ning, 2003; Kurohashi and Nagao, 1994). Accord-
ingly, we can say that the state-of-the-art lexicalized
parsers are mainly based on unlexical (grammatical)
information due to the sparse data problem. Bikel
also indicated that Collins’ parser can use bilexical
dependencies only 1.49% of the time; the rest of
the time, it backs off to condition one word on just
phrasal and part-of-speech categories (Bikel, 2004).
This paper aims at exploiting much more lexical
information, and proposes a fully-lexicalized proba-
bilistic model for Japanese syntactic and case struc-
ture analysis. Lexical information is extracted not
fromasmalltaggedcorpus,butfromahugerawcor-
pus as case frames. This model performs case struc-
ture analysis by a generative probabilistic model
based on the case frames, and selects the syntactic
structure that has the highest case structure proba-
bility.
2 Automatically Constructed Case Frames
We employ automatically constructed case frames
(Kawahara and Kurohashi, 2002) for our model of
Table 1: Case frame examples (examples are ex-
pressed only in English for space limitation.).
CS 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
... ... ...
itadaku (1) ga <agent>
(have) wo soup
ga <agent>itadaku (2)
wo advice, instruction, address(be given)
kara <agent>, president, circle, ···
... ... ...
case structure analysis. This section outlines the
method for constructing the case frames.
A large corpus is automatically parsed, and case
frames are constructed from modifier-head exam-
ples in the resulting parses. The problems of auto-
matic case frame construction are syntactic and se-
mantic ambiguities. That is to say, the parsing re-
sults inevitably contain errors, and verb senses are
intrinsically ambiguous. To cope with these prob-
lems, case frames are gradually constructed from re-
liable modifier-head examples.
First, modifier-head examples that have no syn-
tactic ambiguity are extracted, and they are dis-
ambiguated by a couple of a verb and its closest
case component. Such couples are explicitly ex-
pressed on the surface of text, and can be consid-
ered to play an important role in sentence mean-
ings. For instance, examples are distinguished not
by verbs (e.g., “tsumu” (load/accumulate)), but by
couples (e.g., “nimotsu-wo tsumu” (load baggage)
and “keiken-wo tsumu” (accumulate experience)).
Modifier-head examples are aggregated in this way,
and yield basic case frames.
Thereafter, the basic case frames are clustered
to merge similar case frames. For example, since
“nimotsu-wo tsumu” (load baggage) and “busshi-wo
tsumu” (load supply) are similar, they are clustered.
The similarity is measured using a thesaurus (Ike-
hara et al., 1997).
Usingthisgradualprocedure, weconstructedcase
frames from the web corpus (Kawahara and Kuro-
177
hashi, 2006). The case frames were obtained from
approximately 470M sentences extracted from the
web. They consisted of 90,000 verbs, and the aver-
age number of case frames for a verb was 34.3.
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 generalized example,
which is given to the case slot where half of the ex-
amples belong to <agent> in a thesaurus (Ikehara
et al., 1997). <agent> is also given to “ga” case
slot that has no examples, because “ga” case com-
ponents are usually agentive and often omitted.
3 Integrated Probabilistic Model for
Syntactic and Case Structure Analysis
The proposed method gives a probability to each
possible syntactic structure T and case structure L
of the input sentence S, and outputs the syntactic
and case structure that have the highest probability.
That is to say, the system selects the syntactic struc-
ture Tbest and the case structure Lbest that maximize
the probability P(T,L|S):
(Tbest,Lbest) = argmax
(T,L)
P(T,L|S)
= argmax
(T,L)
P(T,L,S)
P(S)
= argmax
(T,L)
P(T,L,S) (1)
The last equation is derived because P(S) is con-
stant.
3.1 Generative Model for Syntactic and Case
Structure Analysis
We propose a generative probabilistic model based
on the dependency formalism. This model considers
a clause as a unit of generation, and generates the
input sentence from the end of the sentence in turn.
P(T,L,S) is defined as the product of a probability
for generating a clause Ci as follows:
P(T,L,S) =
productdisplay
i=1..n
P(Ci|bhi) (2)
wherenisthenumberofclausesinS,andbhi isCi’s
modifying bunsetsu1. The main clause Cn at the end
1In Japanese, bunsetsu is a basic unit of dependency, con-
sisting of one or more content words and the following zero or
morefunctionwords. ItcorrespondstoabasephraseinEnglish,
and “eojeol” in Korean.
Figure 1: An Example of Probability Calculation.
of a sentence does not have a modifying head, but
we handle it by assuming bhn = EOS (End Of Sen-
tence).
For example, consider the sentence in Figure 1.
There are two possible dependency structures, and
for each structure the product of probabilities indi-
cated below of the tree is calculated. Finally, the
model chooses the highest-probability structure (in
this case the left one).
Ci is decomposed into its predicate type fi (in-
cluding the predicate’s inflection) and the rest case
structure CSi. This means that the predicate in-
cluded in CSi is lemmatized. Bunsetsu bhi is also
decomposed into the content part whi and the type
fhi.
P(Ci|bhi) = P(CSi,fi|whi,fhi)
= P(CSi|fi,whi,fhi)P(fi|whi,fhi)
≈ P(CSi|fi,whi)P(fi|fhi) (3)
The last equation is derived because the content part
in CSi is independent of the type of its modifying
head (fhi), and in most cases, the type fi is indepen-
dent of the content part of its modifying head (whi).
Forexample,P(bentou-wa tabete|syuppatsu-shita)
is calculated as follows:
P(CS(bentou-wa taberu)|te,syuppatsu-suru)P(te|ta.)
WecallP(CSi|fi,whi)generativemodelforcase
structure and P(fi|fhi) generative model for predi-
cate type. The following two sections describe these
models.
3.2 Generative Model for Case Structure
We propose a generative probabilistic model of case
structure. This model selects a case frame that
178
Figure 2: An example of case assignment CAk.
matches the input case components, and makes cor-
respondences between input case components and
case slots.
A case structure CSi consists of a predicate vi,
a case frame CFl and a case assignment CAk.
Case assignment CAk represents correspondences
between input case components and case slots as
shown in Figure 2. Note that there are various pos-
sibilities of case assignment in addition to that of
Figure 2, such as corresponding “bentou” (lunch-
box) with “ga” case. Accordingly, the index k of
CAk ranges up to the number of possible case as-
signments. By splitting CSi into vi, CFl and CAk,
P(CSi|fi,whi) is rewritten as follows:
P(CSi|fi,whi) = P(vi,CFl,CAk|fi,whi)
= P(vi|fi,whi)
×P(CFl|fi,whi,vi)
×P(CAk|fi,whi,vi,CFl)
≈ P(vi|whi)
×P(CFl|vi)
×P(CAk|CFl,fi) (4)
The above approximation is given because it is
natural to consider that the predicate vi depends on
itsmodifyingheadwhi,thatthecaseframeCFl only
dependsonthepredicatevi, andthatthecaseassign-
ment CAk depends on the case frame CFl and the
predicate type fi.
The probabilities P(vi|whi) and P(CFl|vi) are
estimated from case structure analysis results of a
large raw corpus. The remainder of this section il-
lustrates P(CAk|CFl,fi) in detail.
3.2.1 Generative Probability of Case
Assignment
Let us consider case assignment CAk for each
case slot sj in case frame CFl. P(CAk|CFl,fi)
can be decomposed into the following product de-
pending on whether a case slot sj is filled with an
input case component (content part nj and type fj)
or vacant:
P(CAk|CFl,fi) =productdisplay
sj:A(sj)=1
P(A(sj) = 1,nj,fj|CFl,fi,sj)
×
productdisplay
sj:A(sj)=0
P(A(sj) = 0|CFl,fi,sj)
=
productdisplay
sj:A(sj)=1
braceleftBig
P(A(sj) = 1|CFl,fi,sj)
×P(nj,fj|CFl,fi,A(sj) = 1,sj)
bracerightBig
×
productdisplay
sj:A(sj)=0
P(A(sj) = 0|CFl,fi,sj) (5)
where the function A(sj) returns 1 if a case slot sj
is filled with an input case component; otherwise 0.
P(A(sj) = 1|CFl,fi,sj) and P(A(sj) =
0|CFl,fi,sj) in equation (5) can be rewritten as
P(A(sj) = 1|CFl,sj) and P(A(sj) = 0|CFl,sj),
because the evaluation of case slot assignment de-
pends only on the case frame. We call these proba-
bilitiesgenerativeprobabilityofacaseslot,andthey
are estimated from case structure analysis results of
a large corpus.
Let us calculate P(CSi|fi,whi) using the ex-
ample in Figure 1. In the sentence, “wa” is
a topic marking (TM) postposition, and hides
the case marker. The generative probability of
case structure varies depending on the case slot
to which the topic marked phrase is assigned.
For example, when a case frame of “taberu”
(eat) CFtaberu1 with “ga” and “wo” case slots is
used, P(CS(bentou-wa taberu)|te,syuppatsu-suru)
is calculated as follows:
P1(CS(bentou-wa taberu)|te,syuppatsu-suru) =
P(taberu|syuppatsu-suru)
× P(CFtaberu1|taberu)
× P(bentou,wa|CFtaberu1,te,A(wo) = 1,wo)
× P(A(wo) = 1|CFtaberu1,wo)
× P(A(ga) = 0|CFtaberu1,ga) (6)
179
P2(CS(bentou-wa taberu)|te,syupatsu-suru) =
P(taberu|syuppatsu-suru)
× P(CFtaberu1|taberu)
× P(bentou,wa|CFtaberu1,te,A(ga) = 1,ga)
× P(A(ga) = 1|CFtaberu1,ga)
× P(A(wo) = 0|CFtaberu1,wo) (7)
Such probabilities are computed for each case frame
of “taberu” (eat), and the case frame and its cor-
responding case assignment that have the highest
probability are selected.
We describe the generative probability of a case
component P(nj,fj|CFl,fi,A(sj) = 1,sj) below.
3.2.2 Generative Probability of Case
Component
We approximate the generative probability of a
case component, assuming that:
• a generative probability of content partnj is in-
dependent of that of type fj,
• and the interpretation of the surface case in-
cluded in fj does not depend on case frames.
Taking into account these assumptions, the genera-
tiveprobabilityofacasecomponentisapproximated
as follows:
P(nj,fj|CFl,fi,A(sj) = 1,sj) ≈
P(nj|CFl,A(sj) = 1,sj) P(fj|sj,fi) (8)
P(nj|CFl,A(sj) = 1,sj) is the probability of
generating a content part nj from a case slot sj in a
case frame CFl. This probability is estimated from
case frames.
Let us consider P(fj|sj,fi) in equation (8). This
is the probability of generating the type fj of a case
component that has a correspondence with the case
slot sj. Since the type fj consists of a surface case
cj2, a punctuation mark (comma) pj and a topic
marker “wa” tj, P(fj|sj,fi) is rewritten as follows
2A surface case means a postposition sequence at the end of
bunsetsu, such as “ga”, “wo”, “koso” and “demo”.
(using the chain rule):
P(fj|sj,fi) = P(cj,tj,pj|sj,fi)
= P(cj|sj,fi)
×P(pj|sj,fi,cj)
×P(tj|sj,fi,cj,pj)
≈ P(cj|sj)
×P(pj|fi)
×P(tj|fi,pj) (9)
This approximation is given by assuming that cj
only depends on sj, pj only depends on fj, and tj
dependsonfj andpj. P(cj|sj)isestimatedfromthe
KyotoTextCorpus(Kawaharaetal.,2002),inwhich
the relationship between a surface case marker and
a case slot is annotated by hand.
In Japanese, a punctuation mark and a topic
marker are likely to be used when their belong-
ing bunsetsu has a long distance dependency. By
considering such tendency, fi can be regarded as
(oi,ui), where oi means whether a dependent bun-
setsu gets over another head candidate before its
modifying head vi, and ui means a clause type of
vi. The value of oi is binary, and ui is one of the
clause types described in (Kawahara and Kurohashi,
1999).
P(pj|fi) = P(pj|oi,ui) (10)
P(tj|fi,pj) = P(tj|oi,ui,pj) (11)
3.3 Generative Model for Predicate Type
Now, consider P(fi|fhi) in the equation (3). This is
the probability of generating the predicate type of a
clause Ci that modifies bhi. This probability varies
depending on the type of bhi.
When bhi is a predicate bunsetsu, Ci is a subor-
dinate clause embedded in the clause of bhi. As for
the typesfi andfhi, it is necessary to consider punc-
tuation marks (pi, phi) and clause types (ui, uhi).
To capture a long distance dependency indicated by
punctuation marks, ohi (whether Ci has a possible
head candidate before bhi) is also considered.
PV Bmod(fi|fhi) = PV Bmod(pi,ui|phi,uhi,ohi)
(12)
When bhi is a noun bunsetsu, Ci is an embedded
clause in bhi. In this case, clause types and a punc-
tuation mark of the modifiee do not affect the prob-
ability.
PNBmod(fi|fhi) = PNBmod(pi|ohi) (13)
180
Table 2: Data for parameter estimation.
probability what is generated data
P(pj|oi,uj) punctuation mark Kyoto Text Corpus
P(tj|oi,ui,pj) topic marker Kyoto Text Corpus
P(pi,ui|phi,uhi,ohi) predicate type Kyoto Text Corpus
P(cj|sj) surface case Kyoto Text Corpus
P(vi|whi) predicate parsing results
P(nj|CFl,A(sj) = 1,sj) words case frames
P(CFl|vi) case frame case structure analysis results
P(A(sj) = {0,1}|CFl,sj) case slot case structure analysis results
Table 3: Experimental results for syntactic analysis.
baseline proposed
all 3,447/3,976 (86.7%) 3,477/3,976 (87.4%)
NB→VB 1,310/1,547 (84.7%) 1,328/1,547 (85.8%)
TM 244/298 (81.9%) 242/298 (81.2%)
others 1,066/1,249 (85.3%) 1,086/1,249 (86.9%)
NB→NB 525/556 (94.4%) 526/556 (94.6%)
VB→VB 593/760 (78.0%) 601/760 (79.1%)
VB→NB 453/497 (91.1%) 457/497 (92.0%)
4 Experiments
We evaluated the syntactic structure and case struc-
ture outputted by our model. Each parameter is es-
timated using maximum likelihood from the data
described in Table 2. All of these data are not
existing or obtainable by a single process, but ac-
quired by applying syntactic analysis, case frame
construction and case structure analysis in turn. The
process of case structure analysis in this table is a
similarity-based method (Kawahara and Kurohashi,
2002). The case frames were automatically con-
structed from the web corpus comprising 470M sen-
tences, and the case structure analysis results were
obtained from 6M sentences in the web corpus.
The rest of this section first describes the exper-
iments for syntactic structure, and then reports the
experiments for case structure.
4.1 Experiments for Syntactic Structure
We evaluated syntactic structures analyzed by the
proposed model. Our experiments were run on
hand-annotated 675 web sentences 3. The web sen-
tences were manually annotated using the same cri-
teria as the Kyoto Text Corpus. The system input
was tagged automatically using the JUMAN mor-
phological analyzer (Kurohashi et al., 1994). The
syntactic structures obtained were evaluated with re-
3The test set is not used for case frame construction and
probability estimation.
gard to dependency accuracy — the proportion of
correct dependencies out of all dependencies except
for the last dependency in the sentence end 4.
Table 3 shows the dependency accuracy. In
the table, “baseline” means the rule-based syn-
tactic parser, KNP (Kurohashi and Nagao, 1994),
and “proposed” represents the proposed method.
Theproposedmethodsignificantlyoutperformedthe
baseline method (McNemar’s test; p < 0.05). The
dependency accuracies are classified into four types
according to the bunsetsu classes (VB: verb bun-
setsu, NB: noun bunsetsu) of a dependent and its
head. The “NB→VB” type is further divided into
two types: “TM” and “others”. The type that is most
related to case structure is “others” in “NB→VB”.
Its accuracy was improved by 1.6%, and the error
rate was reduced by 10.9%. This result indicated
that the proposed method is effective in analyzing
dependencies related to case structure.
Figure 3 shows some analysis results, where the
dotted lines represent the analysis by the baseline
method, and the solid lines represent the analysis by
the proposed method. Sentence (1) and (2) are in-
correctly analyzed by the baseline but correctly ana-
lyzed by the proposed method.
There are two major causes that led to analysis
errors.
Mismatch between analysis results and annota-
tion criteria
In sentence (3) in Figure 3, the baseline
method correctly recognized the head of “iin-wa”
(commissioner-TM) as “hirakimasu” (open). How-
ever, the proposed method incorrectly judged it as
“oujite-imasuga” (offer). Both analysis results can
be considered to be correct semantically, but from
4Since Japanese is head-final, the second last bunsetsu un-
ambiguously depends on the last bunsetsu, and the last bunsetsu
has no dependency.
181
a63 a63(1) mizu-ga takai tokoro-kara hikui tokoro-he nagareru.
water-nom high ground-abl low ground-all flow
(Water flows from high ground to low ground.)
a63 a63(2) ... Kobe shi-ga senmonchishiki-wo motsu volunteer-wo bosyushita ...
Kobe city-nom expert knowledge-acc have volunteer-acc recruited
(Kobe city recruited a volunteer who has expert knowledge, ...)
a63a63(3) iin-wa, jitaku-de minasan-karano gosoudan-ni oujite-imasuga, ... soudansyo-wo hirakimasu
commissioner-TM at home all of you consultation-dat offer window open
(the commissioner offers consultation to all of you at home, but opens a window ...)
Figure 3: Examples of analysis results.
Table4: Experimentalresultsforcasestructureanal-
ysis.
baseline proposed
TM 72/105 (68.6%) 82/105 (78.1%)
clause 107/155 (69.0%) 121/155 (78.1%)
the viewpoint of our annotation criteria, the latter is
not a syntactic relation, but an ellipsis relation. To
address this problem, it is necessary to simultane-
ously evaluate not only syntactic relations but also
indirect relations, such as ellipses and anaphora.
Linear weighting on each probability
We proposed a generative probabilistic model,
and thus cannot optimize the weight of each proba-
bility. Such optimization could be a way to improve
the system performance. In the future, we plan to
employ a machine learning technique for the opti-
mization.
4.2 Experiments for Case Structure
We applied case structure analysis to 215 web sen-
tences which are manually annotated with case
structure, and evaluated case markers of TM phrases
and clausal modifiees by comparing them with the
gold standard in the corpus. The experimental re-
sults are shown in table 4, in which the baseline
refers to a similarity-based method (Kawahara and
Kurohashi, 2002). The experimental results were re-
ally good compared to the baseline. It is difficult to
compare the results with the previous work stated in
the next section, because of different experimental
settings (e.g., our evaluation includes parse errors in
incorrect cases).
5 Related Work
There have been several approaches for syntactic
analysis handling lexical preference on a large scale.
Shirai et al. proposed a PGLR-based syntactic
analysis method using large-scale lexical preference
(Shirai et al., 1998). Their system learned lexical
preference from a large newspaper corpus (articles
of five years), such as P(pie|wo,taberu), but did
not deal with verb sense ambiguity. They reported
84.34% accuracy on 500 relatively short sentences
from the Kyoto Text Corpus.
Fujio and Matsumoto presented a syntactic anal-
ysis method based on lexical statistics (Fujio and
Matsumoto, 1998). They made use of a probabilistic
modeldefinedbytheproductofaprobabilityofhav-
ing a dependency between two cooccurring words
and a distance probability. The model was trained
on the EDR corpus, and performed with 86.89% ac-
curacy on 10,000 sentences from the EDR corpus 5.
On the other hand, there have been a number
of machine learning-based approaches using lexical
preference as their features. Among these, Kudo
and Matsumoto yielded the best performance (Kudo
and Matsumoto, 2002). They proposed a chunking-
based dependency analysis method using Support
Vector Machines. They used two-fold cross valida-
tionontheKyotoTextCorpus,andachieved90.46%
5The evaluation includes the last dependencies in the sen-
tence end, which are always correct.
182
accuracy 5. However, it is very hard to learn suffi-
cient lexical preference from several tens of thou-
sands sentences of a hand-tagged corpus.
There has been some related work analyzing
clausal modifiees and TM phrases. For exam-
ple, Torisawa analyzed TM phrases using predicate-
argument cooccurences and word classifications in-
duced by the EM algorithm (Torisawa, 2001). Its
accuracy was approximately 88% for “wa” and 84%
for “mo”. It is difficult to compare the accuracy
of their system to ours, because the range of tar-
get expressions is different. Unlike related work,
it is promising to utilize the resultant case frames
for subsequent analyzes such as ellipsis or discourse
analysis.
6 Conclusion
We have described an integrated probabilistic model
for syntactic and case structure analysis. This model
takes advantage of lexical selectional preference of
large-scale case frames, and performs syntactic and
case analysis simultaneously. The experiments indi-
cated the effectiveness of our model. In the future,
by incorporating ellipsis resolution, we will develop
an integrated model of syntactic, case and ellipsis
analysis.
References
Collin Baker, Charles Fillmore, and John Lowe. 1998. The
Berkeley FrameNet Project. In Proceedings of the 17th In-
ternational Conference on Computational Linguistics and
the 36th Annual Meeting of the Association for Computa-
tional Linguistics, pages 86–90.
Daniel M. Bikel. 2004. Intricacies of Collins’ parsing model.
Computational Linguistics, 30(4):479–511.
Don Blaheta and Eugene Charniak. 2000. Assigning function
tags to parsed text. In Proceedings of the 1st Meeting of
the North American Chapter of the Association for Compu-
tational Linguistics, pages 234–240.
Eugene Charniak. 2000. A maximum-entropy-inspired parser.
In Proceedings of the 1st Meeting of the North American
Chapter of the Association for Computational Linguistics,
pages 132–139.
Michael Collins. 1999. Head-Driven Statistical Models for
Natural Language Parsing. Ph.D. thesis, University of
Pennsylvania.
Masakazu Fujio and Yuji Matsumoto. 1998. Japanese depen-
dency structure analysis based on lexicalized statistics. In
Proceedings of the 3rd Conference on Empirical Methods in
Natural Language Processing, pages 88–96.
Satoru Ikehara, Masahiro Miyazaki, Satoshi Shirai, Akio
Yokoo, Hiromi Nakaiwa, Kentarou Ogura, Yoshifumi
Oyama, and Yoshihiko Hayashi, editors. 1997. Japanese
Lexicon. Iwanami Publishing.
Daisuke Kawahara and Sadao Kurohashi. 1999. Corpus-based
dependency analysis of Japanese sentences using verb bun-
setsu transitivity. In Proceedings of the 5th Natural Lan-
guage Processing Pacific Rim Symposium, pages 387–391.
DaisukeKawaharaandSadaoKurohashi. 2002. Fertilizationof
case frame dictionary for robust Japanese case analysis. In
Proceedings of the 19th International Conference on Com-
putational Linguistics, pages 425–431.
Daisuke Kawahara and Sadao Kurohashi. 2006. Case frame
compilation from the web using high-performance comput-
ing. In Proceedings of the 5th International Conference on
Language Resources and Evaluation.
Daisuke Kawahara, Sadao Kurohashi, and Kˆoiti Hasida. 2002.
Construction of a Japanese relevance-tagged corpus. In Pro-
ceedings of the 3rd International Conference on Language
Resources and Evaluation, pages 2008–2013.
Paul Kingsbury, Martha Palmer, and Mitch Marcus. 2002.
Adding semantic annotation to the Penn TreeBank. In Pro-
ceedings of the Human Language Technology Conference.
Dan Klein and Christopher D. Manning. 2003. Accurate un-
lexicalized parsing. In Proceedings of the 41st Annual Meet-
ing of the Association for Computational Linguistics, pages
423–430.
Taku Kudo and Yuji Matsumoto. 2002. Japanese dependency
analysis using cascaded chunking. In Proceedings of the
Conference on Natural Language Learning, pages 29–35.
Sadao Kurohashi and Makoto Nagao. 1994. A syntactic anal-
ysis method of long Japanese sentences based on the detec-
tion of conjunctive structures. Computational Linguistics,
20(4):507–534.
Sadao Kurohashi, Toshihisa Nakamura, Yuji Matsumoto, and
Makoto Nagao. 1994. Improvements of Japanese morpho-
logical analyzer JUMAN. In Proceedings of the Interna-
tional Workshop on Sharable Natural Language, pages 22–
28.
Kiyoaki Shirai, Kentaro Inui, Takenobu Tokunaga, and Hozumi
Tanaka. 1998. An empirical evaluation on statistical parsing
of Japanese sentences using lexical association statistics. In
Proceedings of the 3rd Conference on Empirical Methods in
Natural Language Processing, pages 80–87.
Kentaro Torisawa. 2001. An unsupervised method for canon-
icalization of Japanese postpositions. In Proceedings of the
6th Natural Language Processing Pacific Rim Simposium,
pages 211–218.
Kiyotaka Uchimoto, Masaki Murata, Satoshi Sekine, and Hi-
toshi Isahara. 2000. Dependency model using posterior
context. In Proceedings of the 6th International Workshop
on Parsing Technology, pages 321–322.
183
