VERB SENSE DISAMBIGUATION BASED ON THESAURUS

OF PREDICATE-ARGUMENT STRUCTURE

An Evaluation of Thesaurus of Predicate-argument Structure for Japanese Verbs

Koichi Takeuchi, Suguru Tsuchiyama, Masato Moriya, Yuuki Moriyasu and Koichi Satoh

Graduate School of Natural Science and Technology, Okayama University, Tsushimanaka, Okayama-shi, Japan

Keywords:

Thesaurus, Predicate-argument structure, Japanese verbs, Word sense disambiguation.

Abstract:

This paper presents a system for word sense disambiguation based on a manually constructed thesaurus of

predicate-argument structure, which is an ontology on the linguistic side providing essential information for

mapping form texts to verb concepts. This system can be effective for word sense disambiguation even though

the target word sense system is different from the thesaurus. We applied the proposed word sense disam-

biguation system to the test corpus of SemEval-2010 Japanese tasks. Experimental results showed that the

thesaurus-based disambiguation system outperformed a CRFs-based system in recall rates of verb sense dis-

ambiguation. From the results of verb sense disambiguation, we clariﬁed that the abstracted verb classes (709

types) in our proposed system were effective sets for verb sense disambiguation.

1 INTRODUCTION

This paper presents a system for word sense disam-

biguation based on a manually constructed thesaurus

of predicate-argument structure. The system can be

effective for word sense disambiguation even though

the target word sense system is different from the the-

saurus.

We are manually constructing a verb thesaurus of

predicate-argument structure (Takeuchi et al., 2010)

to deal with verbal paraphrases as well as verb

sense disambiguation, such as “He employed/used a

method” or “He employed/hired Mr. Smith as an

accountant”.Since each verb is highly polysemous,

verbs can be categorized into several verb classes, and

verbs in each verb class have shared concepts.Thus,

our dictionary describes set verb classes with argu-

ment structures to make a correlation between their

arguments

. From the view of ontology, our thesaurus

can be regarded as an ontology on the linguistic side,

which can provide essential information for mapping

form texts to verb concepts (i.e., verb classes). Thus

one of the issues in creating a verb thesaurus is mak-

ing sure how effective the constructed thesaurus is for

verb sense disambiguation.

On the other hand, the ﬁrst Balanced Corpus of

The thesaurus is freely downloadable at

http://vsearch.cl.cs.okayama-u.ac.jp/.

Contemporary Written Japanese, BCCWJ, has

been developed (Maekawa, 2008), and SemEval-

2010 (Okumura et al., 2010) Japanese word sense

disambiguation tasks are organized based on the

sense annotated corpus of the BCCWJ. In regard

to verb sense, identifying verb sense indicates the

ﬁnding of a group of verb synonyms. Thus, the verb

sense tagged corpus can be regarded as a test bed of a

task to detect a verb synonym group. This task would

be the ﬁrst step in detecting the argument structures

of verbs for dealing with verbal paraphrases.

Given the abovebackground,in this paper we clar-

iﬁed the usefulness of the verb thesaurus by applying

a thesaurus-based shallow semantic analyzer to WSD

for verb senses, regarding the task of detecting a verb

class, i.e., a verb synonym group, in our thesaurus.

The deﬁnitions of verb classes in our thesaurus and

those of verb sense tags in SemEval-2010 are not

equivalent. Thus, our verb classes were converted to

verb sense tags by a conversion table constructed in-

dependently for this task. Since the SemEval-2010

corpus provides a training corpus and test corpus,

the performance of tasks of the thesaurus-based an-

alyzer was compared with the performance of Con-

ditional Random Fields (CRFs), a statistical learning

approach-based model. It was also compared with the

performance of the best WSD system for SemEval-

2010 tasks.

Experimental results of verb sense disambigua-

208

Takeuchi K., Tsuchiyama S., Moriya M., Moriyasu Y. and Satoh K..

VERB SENSE DISAMBIGUATION BASED ON THESAURUS OF PREDICATE-ARGUMENT STRUCTURE - An Evaluation of Thesaurus of Predicate-

argument Structure for Japanese Verbs.

DOI: 10.5220/0003639802080213

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (KEOD-2011), pages 208-213

ISBN: 978-989-8425-80-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

tion showed that the thesaurus-based analyzer outper-

formed CRFs in recall rates for all genres as well

as for white paper domains. For precision rates,

the thesaurus-based analyzer provided lower accuracy

than the best WSD system; however, most of the er-

rors were caused not by the analyzer, but by the con-

version table. Thus, we clariﬁed that the performance

of the thesaurus-based analyzer was almost equal to

that of the best WSD system when we discounted con-

version table errors.

2 BACKGROUND: LANGUAGE

RESOURCES FOR PREDICATE

ARGUMENT STRUCTURE

The position of our thesaurus of predicate-argument

structure for Japanese verbs can be described from

the following three viewpoints: language resources,

Japanese dictionaries and ontologies.

View from Language Resources. In English thor-

ough, well-organized language resources relating

to predicate-argument structure are constructed and

available, e.g., EVCA (Levin, 1993), Dorr’s LCS

(Dorr, 1997), PropBank (Palmer et al., 2005), Verb-

Net (Kipper-Schuler, 2005), WordNet (Fellbaum,

1998) and FrameNet (Baker et al., 1998). In addi-

tion, there is a research project (Pustejovsky and Mey-

ers, 2005) to ﬁnd a general description framework of

predicate-argument structure by merging several lex-

ical databases (i.e., PropBank, NomBank, TimeBank

and PennDiscouse TreeBank).

On the other hand, our thesaurus provides sev-

eral granularities of semantic verb classes with argu-

ment structure (see Section 3). Comparedto the above

language resources, our thesaurus corresponds partly

to each lexical database, i.e., Frame and FrameEle-

ments in FrameNet correspond to our verb class and

semantic role labels, and the way of organizing simi-

lar verb classes with a thesaurus corresponds to Word-

Net; however, these elements and the method of orga-

nization of similar verb classes with a thesaurus are

not exactly the same as those of our approach and our

proposed thesaurus.

View from Japanese Dictionaries. There are sev-

eral Japanese lexicons: IPAL (IPAL, 1986) was con-

structed focusing on morpho-syntactic classes but

IPAL is small

. EDR (EDR, 1995) is composed of a

large-scale lexicon and corpus. EDR is a well thought

out and wide ranging bilingual dictionary between

Japanese and English, but EDR’s semantic classes

It contains 861 verbs and 136 adjectives.

were not designed with syntactical lexical relations

between verbs, such as alternations and causative or

transitive relations between verbs. In contrast, our

thesaurus can deal with these relations.

Besides Japanese version of WordNet (Bond et al.,

2008) and FrameNet (Ohara et al., 2006) have been

proposed. Japanese WordNet are constructed by ma-

chine translation from English to Japanese and man-

ual revision, and then we ﬁnd that some of the ba-

sic verbs of Japanese (i.e., highly ambiguous verbs)

are wrongly assigned to unrelated synsets. Japanese

FrameNet currently has published fewer than 150

verbs, then it is much smaller than our thesaurus.

View from Ontologies. Previous upper ontology

studies have discussed how to describe processes and

events (Takeda, 2004) (Galton, 2010), and practical

ontologies such as SUMO (Niles and Pease, 2001)

and DOLCE (Masolo et al., 2002) have been pub-

lished; however, because they are upper ontologies,

they are too abstract to deal with practical events. In

contrast, a more concrete event ontology that can give

a formal framework to deal with relations between

Japanese verb concepts with description logic has

been proposed (Kaneiwa et al., 2007) (Kaneiwa and

Iwazume, 2010). An event ontology gives clear deﬁ-

nitions of different verb concepts; however, the target

of the ontology is not to deal with practical texts, but

to deal with logical semantic relations between con-

cepts. Thus, there is no information for verb sense

disambiguation that constitutes an essential technique

for mapping from texts to concepts. Compared to this

approach, our proposed verb thesaurus provides in-

formation on verb sense disambiguation as well as on

verb classes (i.e., concepts), and thus, our thesaurus

can bridge the gap between texts and event ontolo-

gies, which are well-organized according to descrip-

tion logic.

3 THESAURUS OF

PREDICATE-ARGUMENT

STRUCTURE

Since the details of our thesaurus are described in pre-

vious papers (Takeuchi et al., 2010) here we describe

the basic design of our thesaurus.

The proposed thesaurus consists of a hierarchy

of verb classes we deﬁned. A verb class, a concep-

tual class, indicates a shared meaning of verbs in a

verb class. A parent verb class includes concepts of

a child verb class; thus a child verb class is a de-

tailed concept of the parent verb class and multiple

inheritance is prohibited in the hierarchy. Meaning

VERB SENSE DISAMBIGUATION BASED ON THESAURUS OF PREDICATE-ARGUMENT STRUCTURE - An

Evaluation of Thesaurus of Predicate-argument Structure for Japanese Verbs

209

of a verb class is described as a semantic descrip-

tion that is a kind of semantic skeleton such as lexical

conceptual structure (Jackendoff, 1990) (Kageyama,

1996) (Dorr, 1997). Thus, core semantic relations be-

tween arguments are represented in the semantic de-

scription.

We allow a verb has several senses, i.e., a verb can

be a polysemous, and adopt verb senses deﬁned in

Lexeed (Fujita et al., 2006) as a verb senses repository

of our thesaurus. Thus each verb sense is assigned to a

verb class, and then a verb sense is linked to example

sentences. Every example sentence is analyzed into

its arguments and semantic role labels; the core argu-

ments are linked to the semantic description via vari-

ables. This allows that if semantic role labels cannot

capture the correct linking, the links of variables can

designate corresponding arguments in example sen-

tences (Figure 2). In addition, by linking one seman-

tic description to several example sentences of a verb

sense, our thesaurus can provide rich verb sense dis-

ambiguation information.

Here we explain this structure using verbs such

as buy, purchase, hire, rent, recapture

. Each verb

sense is designated by example sentences, e.g., “Hi-

roshi buys a bicycle to his son”, “Jiro purchases a car

form her”, “Taro hires a car”, “Kazuko rents a book”

and “Yoshio recaptures the top position”. As Figure 1

shows, all of the above verb senses are involved in the

verb class Moving One’s Possession From

. The se-

mantic description, which expresses the core meaning

of the verb class is

([Agent] CAUSE) BECOME [1] BE AT [2],

where the brackets [] denote variables that can be

ﬁlled with arguments in example sentences; [Agent]

is a semantic role label that can be annotated to all

example sentences; the parentheses () denote an com-

ponent. The semantic description consists of roughly

3 components describing causer, manner and (change

of) state. A manner component expresses various

kinds of complex meanings of a verb concept such

as condition, purpose, attitude, and so on

Figure 1 shows the children of the verb

class Moving One’s Possession From, e.g.,

The actual verbs are Japanese in our thesaurus, but to

make the explanation easy to understand we give their En-

glish translation.

The actual verb class names correspond to

the hierarchical system in the thesaurus, e.g., the

full length of Moving One’s Possession From is

Change of State/ Change of Position (Physical)/ Mov-

ing One’s Possession From, where ’/’ denotes a delimiter

between hierarchies. Then the verb class names in Figures

1 and 2 show the abbreviated verb class names.

Currently about 800 kinds of words are used to describe

a manner component.

Moving One’s Possession From/Buying, Mov-

ing One’s Possession From/Renting and Mov-

ing One’s Possession From/Repossessing. The verbs

buy and purchase are in the Buying class, while the

verbs rent and hire are in the Renting class. The

semantic descriptions in the children verb classes are

more detailed than those in the parent’s description.

!"#$%&'(%)*+',"++)++$"%'-."/

!!"#$%%&"'()*+,$%)-',$%%

',./$%',(*&/"',!

!"#$%&"'()*+,%

)-',$%',./!

!"#$%&'(%)*+',"++)++$"%'-."/01

234$%&

!"#$%&'(%)5+',"++)++$"%'-."/01

6)7"++)++$%&

!0123,./4%526789%%:85;<8%1=4%:8%2>%1?4"!

!0123,./4%526789"0:@%<82A7%;B%123,./4%%

!"#-.3%1=49%:85;<8%1=4%:8%2>%1?4!

!0123,./4%526789"0:@%<82A7%;B%123,./4%

',.C.3%1=4%9%:85;<8%1=4%:8%2>%1?4!

+,D*.C(%E,+('-&CF.!

G,'!%(H*++!

',(*&/"',!

!"#$%&'(%)5+',"++)++$"%'-."/01

6)%8%&

!0123,./4%526789"0:@%<82A7%;B%123,./4%

I,&F++,++-.3%01=4%J%K)-()%)*+%3F.,%F"/%FL%

1?4%99%:85;<8%1=4%:8%2>%1?4!

Figure 1: Example of verb classes, verbs and their semantic

descriptions.

A semantic description in the Renting class, i.e.,

([Agent] CAUSE)

(BY MEANS OF [Agent] renting [1])

BECOME [1] BE AT [2]

describes semantic relations between [1], [2] and

[Agent]. Since semantic role labels and variables are

annotated to all of the example sentences, correspond-

ing arguments can be linked via semantic role labels

and variables in the semantic description. As we show

!"#$%&'(%)*+',"++)++$"%'-."/01

234$%&

!!"#$%&'()*#+,-."!/0)1-#2,)34)"#$%&'())

5678&$)"9().)/-*31-)"9()/-)#:)";(!

<8=>?@8)))567?))A)58B7BC%)))))))'>)@8?)?>&D)

"#$%&'()))))))))))))":@%E%("9())"F%B8G8%&'(";(!

H8=>)))))))))))))G6=B@A?%?))))A))BA=))))))))))))I=>E)@%=D)

"#$%&'(";()))))))))))))))))))))))":@%E%("9())",>6=B%(!

Figure 2: Linking between semantic description and exam-

ple sentences.

in Figure 2, the semantic description contains only es-

sential arguments. Thus, the key arguments such as

[1] and [2] in example sentences have links, and an-

other argument such as [Source] does not have a link.

This indicates the two type of arguments we assumed,

KEOD 2011 - International Conference on Knowledge Engineering and Ontology Development

210

i.e., arguments that are essential to a verb class and ar-

guments that are not essential. This is because of the

current perspective on arguments in linguistics, that

is, there are several levels of arguments depending on

closeness to a verb concept such as core, non-core,

peripheral and adjunct arguments in FrameNet, and

constructions (Goldberg, 1995). Since the variety of

adjuncts and constructions is wide, the proposed de-

scription framework can deal with this variety, which

is impossible to pre-compile, by adding analyzed ex-

ample sentences. The current types of semantic roles

are organized into 71 for the results of analyses of

about 7,400 verb senses.

4 ARGUMENT STRUCTURE

ANALYZER

A predicate-argument structure analyzer, ASA

, was

constructed on the basis of our thesaurus of predicate-

argument structure. ASA identiﬁes the verb classes

and semantic role labels of their arguments. In our

thesaurus, for polysemous verbs, each verb sense cat-

egorized to a verb class has a few example sentences.

Since each verb sense in a verb class only has a few

example sentences, statistical learning methods do not

work well in the preliminary tests. Thus, as the basic

strategy for detecting the verb class (i.e., verb sense)

of an input sentence, we take a nearest neighbor ap-

proach: ﬁnd the most similar example sentence com-

pared to the input sentence, and take the verb class of

the example sentence as the word sense.

The similarity between an input sentence and an

example sentence is evaluated on the similarity of the

arguments between them; the similarity of the argu-

ments is evaluated on three features: shallow syntac-

tic position, noun categories, and surface words. Let

SimSnt be this similarity function, the disambiguation

of verb sense for an input sentence X is to detect the

verb class

C that gives the highest score of among the

example sentences Y

in a verb class C.

C = argmax

SimSnt(X, Y

). (1)

The details of SimSnt are described in (Takeuchi et al.,

2009). Since SimSnt() is calculated on the basis of the

example sentences in the thesaurus, the performance

of the ASA’s verb class disambiguation depends on

the quality and quantity of the thesaurus.

http://cl.cs.okayama-u.ac.jp/study/project/sea.html.

4.1 CRFs-based Word Sense

Disambiguation System

We applied Conditional Random Fields (Lafferty

et al., 2001) as a competitive alternative to the ASA.

Conditional Random Fields is a probabilistic model

for labeling sequence data, and we applied it to

word sense disambiguation. The parameters of CRFs

can be trained using the training corpus provided by

SemEval-2010.

CRFs selects the best output sequence, i.e., a se-

quence of word senses Y = (Y

, Y

, ..., Y

) given in-

put word sequence X = (X

, X

, ..., X

) by the follow-

ing equations:

P(Y|X) =

exp(Λ· F(Y, X))

, (2)

where Z

denotes a normalized factor, Y

denotes

possible label (i.e., word sense) candidates and

∑

denotes the sum of all possible word sense sequences

from an input word sequence X. Λ is a weight for the

feature vector F. For the word sense disambiguation

in Section 5 we apply as the features surface word,

part-of-speech and combinations of previous and fol-

lowing words and part-of-speeches at from -3 to +1

positions

according to the results of CoNLL shared

task

5 VERB SENSE

DISAMBIGUATION

EXPERIMENTS AND

DISCUSSIONS

SemEval-2010 annotated corpus consists of four gen-

res (books, newspaper articles, white papers, and doc-

uments from a Q&A site on the WWW) and we used

this corpus as the gold standard for a Japanese verb

sense annotated corpus.

Table 1: Precision and recall of verb sense disambiguation

in white papers.

Precision Recall

CRFs 0.971 (134/138) 0.244 (134/550)

ASA 0.660 (229/347) 0.416 (229/550)

Table 1 shows the results of verb sense disam-

biguation for the test data in white papers. The pa-

rameters in CRFs were trained on the training data

in white papers, and the training data excluded doc-

uments that overlapped with test data. In Table 1 the

See http://crfpp.sourceforge.net/.

http://www.clips.ua.ac.be/conll2000/.

VERB SENSE DISAMBIGUATION BASED ON THESAURUS OF PREDICATE-ARGUMENT STRUCTURE - An

Evaluation of Thesaurus of Predicate-argument Structure for Japanese Verbs

211

precision of CRFs is quite high; however, the number

of correctly detected verb senses is much lower than

that of the ASA—the ASA detected almost twice as

many verb senses as the CRFs did. The point of the

results is that the ASA does not use any domain de-

pendent information, but the CRFs do. If we apply

these systems to all genres, the differences become

much more clearer.

Figure 2 shows the results of verb sense disam-

biguation in all genres. In Figure 2 RALI-2, a naive

Table 2: Precision and recall of verb sense disambiguation

in books, newspaper articles, white papers and documents

from a Q&A site on the WWW.

Precision Recall

RALI-2 0.719 no published data

CRFs 0.940 0.0372

(608/648) (608/16332)

ASA 0.593 0.378

(6173/10417) (6173/16332)

Bayes model, denotes the best results of the word

sense disambiguation system in SemEval-2010 tasks.

Both CRFs and RALI-2 used training data documents

that overlapped with test data documents. The train-

ing data consisted of three genres i.e., books, news-

paper articles and white papers, and there was no

training data for Q&A documents. From the results,

we found that the ASA outperformed CRFs in recall

rates; CRFs seemed to have difﬁculty detecting cor-

rect verb senses in different genres. This tendencyi.e.,

simple statistical models such as naive Bayes over-

come more sophisticated statistical models such as

support vector machines (SVMs) and maximum en-

tropy models in word sense disambiguation, is also

reported in the results of SemEval-2010 task (Oku-

mura et al., 2010).

Table 2 shows that the ASA did not detect any

verb sense for 5915 (16332 - 10417) examples due

to the lack of verb entries in our thesaurus. From this

result we can estimate the coverage of the verb en-

tries in our thesaurus for all genres in Japanese text

was 63.8% (10417/16633).

The output verb senses of the ASA are verb

classes in our thesaurus, but the results of the ASA

in Tables 1 and 2 were evaluated based on the trans-

formed verb senses in SemEval-2010. To do this

transformation, we use the conversion table that was

manually created while doing the construction work

of the thesaurs; this construction work was indepen-

dently done for the purpose of this evaluation task.

The total number of detected verb classes by the ASA

in all genres was 10,417 words, but 1,637 words of

these (about 16%) were not converted to any word

sense because of the lack of sense identiﬁcations in

the conversion table. This data indicates that there is

room for further improvement of correctly detecting

verb sense by adding instances of verb senses to our

thesaurus.

If we exclude non-converted examples, the preci-

sion rate of the ASA would be 70% (6173/8780) for

all genres. This precision rate is almost the same as

that of RALI-2. From the view of the conversion ta-

ble, about 70% of verb classes in total can be success-

fully converted to SemEval-2010 word senses. The

number of types of all verb classes is currently 709;

and the verb classes are manually deﬁned by summa-

rizing over 7,400 verb senses in Lexeed (Fujita et al.,

2006). Thus the success of the conversion indicates

that the abstracted verb class is not too coarse but still

keeps a granularity that can discriminate verb senses

in SemEval-2010 tasks. Furthermore, since verb class

disambiguation is done on the basis of the analyzed

example sentences in the thesaurus, we can conclude

that the proposed thesaurus provides effective linguis-

tic data for verb sense disambiguation.

The methods of verb sense disambiguation de-

pend on the availability of linguistically annotated

resources. Table 3 shows the results of CRFs in a

white paper genre that used overlapping training data.

Comparing the results in Table 1, CRFs outperformed

Table 3: Precision and recall of verb sense disambiguation

by CRFs in a white paper genre that used overlapping train-

ing data.

Precision Recall

CRFs 0.993 0.434

the ASA in both precision and recall rates. The results

indicate that if we have enough sense annotated train-

ing data as test data for the same genre, a statistical

learning approach will work well. Thus, we need to

keep developing our thesaurus.

6 CONCLUSIONS

To evaluate a manually constructed Japanese verb the-

saurus that is an ontology on the linguistic side pro-

viding essential information for mapping form texts

to verb concepts, we constructed a system for word

sense disambiguation (ASA) based on the thesaurus

and applied the system to SemEval-2010 word sense

disambiguation tasks. The deﬁnitions of verb classes

in our thesaurus and verb sense tags in SemEval-

2010 are not equal. Thus, we converted our verb

classes to verb sense tags by a conversion table that

we constructed independently for this task. Since

the SemEval-2010 corpus provides training corpus

KEOD 2011 - International Conference on Knowledge Engineering and Ontology Development

212

and test corpus the performance of the ASA was

compared with that of Conditional Random Fields

(CRFs), a statistical learning approach-based model.

Experimental results of verb sense disambiguation

showed that the ASA outperformed CRFs in recall

rates for all genres as well as for the white paper do-

main. Regarding precision rates, the ASA provided

lower accuracy than the best system for SemEval-

2010 tasks; however, since most of the errors were

caused by the conversion table, we found that the pre-

cision rate of the ASA was almost equal to the level

(70%) of the best WSD system when we excluded

conversion table errors. From the recall rate of the

ASA for all genres the current coverage of the verb

entries in our thesaurus can be estimated at 63.8%.

The key to success for the ASA will be proper

working of the conversion table. Thus, the suc-

cess of the conversion indicates the abstracted verb

class is not too coarse, but still keeps a granularity

that can discriminate verb senses in SemEval-2010

tasks. In addition, since verb class disambiguation

is done based on analyzed example sentences in the

thesaurus, we can conclude that the proposed the-

saurus provides effective linguistic information for

verb sense disambiguation.

REFERENCES

Baker, C. F., Fillmore, C. J., and Lowe, J. B. (1998). The

Berkeley FrameNet project. In Proceedings of the

36th Annual Meeting of the Association for Compu-

tational Linguistics, pages 86–90.

Bond, F., Isahara, H., Kanzaki, K., and Uchimoto, K.

(2008). Construction of Japanese WordNet from

Multi-lingual WordNet. In Proceedings of the 14th

Annual Meeting of Japanese Natural Language Pro-

cessing, pages 853–856.

Dorr, B. J. (1997). Large-Scale Dictionary Construction for

Foreign Language Tutoring and Interlingual Machine

Translation. Machine Translation, 12(4):271–325.

EDR (1995). EDR: Electric Dictionary the Second Edition.

Japan Electronic Dictionary Research Institute, Ltd.

Fellbaum, C. (1998). WordNet an Electronic Lexical

Database. MIT Press.

Fujita, S., Tanaka, T., Bond, F., and Nakaiwa, H. (2006). An

implemented description of japanese: The lexeed dic-

tionary and the hinoki treebank. In COLING/ACL06

Interactive Presentation Sessions, pages 65–68.

Galton, A. (2010). On what goes on: The ontology of pro-

cesses and events. In Formal Ontology in Informa-

tion Systems: Proceedings of the Fourth International

Conference (FOIS2006), pages 4–11.

Goldberg, A. E. (1995). Constructions. The University of

Chicago Press.

IPAL (1986). IPA Lexicon of the Japanese Language for

Computers. IPA: Information-Technology Promotion

Agency, Japan.

Jackendoff, R. (1990). Semantic Structures. MIT Press.

Kageyama, T. (1996). Verb Semantics. Kurosio Publishers.

(In Japanese).

Kaneiwa, K. and Iwazume, M. (2010). An Event Ontology

for the Semantic Web. Computer Software, 27(5):1–

13. (in Japanese).

Kaneiwa, K., Iwazume, M., and Fukuda, K. (2007). An up-

per ontology for event classiﬁcations and relations. In

Proceedings of the Twentieth Australian Joint Confer-

ence on Artiﬁcial Intel ligence (AI2007), LNCS 4830,

Springer-Verlag, pages 394–403.

Kipper-Schuler, K. (2005). VerbNet: A broad-coverage,

comprehensive verb lexicon. PhD thesis, PhD Thesis,

University of Pennsylvania.

Lafferty, J., McCallum, A., and Pereira, F. (2001). Con-

ditional random ﬁelds: Probabilistic models for seg-

menting and labeling sequence data. In Proc. 18th

International Conf. on Machine Learning, pages 282–

289.

Levin, B. (1993). English Verb Classes and Alternations.

University of Chicago Press.

Maekawa, K. (2008). Balanced corpus of contemporary

written japanese. In Proceedings of the 6th Workshop

on Asian Language Resources (ALR), pages 101–102.

Masolo, C., Borgo, S., Gangemi, A., Guarino, N., Oltra-

mari, A., and Schneider, L. (2002). WonderWeb de-

liverable D17. the WonderWeb library of foundational

ontologies and the DOLCE ontology.

Niles, I. and Pease, A. (2001). Towards a standard upper

ontology. In Proceedings of the 2nd Interna-tional

Conference on Formal Ontology in Information Sys-

tems (FOIS-2001), pages 2–9.

Ohara, K. H., Fujii, S., Ohori, T., Suzuki, R., Saito, H.,

and Ishizaki, S. (2006). Frame-based contrastive lexi-

cal semantics and japanese framenet: The case of risk

and kakeru. In Proceeding of the Fourth International

Conference on Construction Grammar.

Okumura, M., Shirai, K., Komiya, K., and Yokono, H.

(2010). Semeval-2010 task: Japanese wsd. In Pro-

ceedings of the 5th International Workshop on Seman-

tic Evaluation, pages 69–74.

Palmer, M., Gildea, D., and Kingsbury, P. (2005). The

proposition bank: An annotated corpus of semantic

roles. Computational Linguistics, 31(1):71–105.

Pustejovsky, J. and Meyers, M. P. A. (2005). Merging

propbank, nombank, timebank, penn discourse tree-

bank and coreference. In Proceedings of the Workshop

on Frontiers in Corpus Annotation II: Pie in the Sky,

pages 5–12.

Takeda, H. (2004). Upper Ontology. Japanese Society of

Artiﬁcial Intelligence, 19(2):172–178. (in Japanese).

Takeuchi, K., Inui, K., Takeuchi, N., and Fujita, A.

(2010). A Thesaurus of Predicate-Argument Structure

for Japanese Verbs to Deal with Granularity of Verb

Meanings. In The 8th Workshop on Asian Language

Resources, pages 1–8.

Takeuchi, K., Tsuchiyama, S., Moriya, M., and Moriyasu,

Y. (2009). Construction of Argument Structure Ana-

lyzer Towards Searching Same Situation and Actions.

In Technical Report of IEICE, NLC-2009-33, pages 1–

6. (in Japanese).

VERB SENSE DISAMBIGUATION BASED ON THESAURUS OF PREDICATE-ARGUMENT STRUCTURE - An

Evaluation of Thesaurus of Predicate-argument Structure for Japanese Verbs

213