Japan Advanced Institute of Science and Technology

Japan Advanced Institute of Science and Technology

JAIST Repository

https://dspace.jaist.ac.jp/

Title

析

Author(s)

Citation

Issue Date

Type

Text version

URL

Rights

Description

for information acquisition from Japanese

dictionary of EDR

FUJIWARA Shigeru

School of Information Science,

Japan Advanced Institute of Science and Technology

February 15, 1999

Keywords: EDR electronic dictionary,EDR thesaurus, Japanese concept explication,

N-gram statistics.

Recently, large scale of machine-readable dictionaries and corpora are often used in

NaturalLanguage Processing. EDRelectronic dictionaryisaverylargescaleof machine-

readable database of language, that has word dictionaries and thesaurus and corpora.

EDRthesaurusisapartofEDRelectronicdictionary,whichconsistsof 400,000concepts,

and concept explication explain each concept. We can get useful informationfrom EDR

thesaurus,e.g. hypernym,hyponym,synonym. Forusers needtheseinformation,the rich

concepts of EDR thesaurus isuseful.

However, for processes using EDR thesaurus, performance of EDR thesaurus is not

always enough. Forexample,

The distribution of concepts of EDR thesaurus is slanted.

The number of synonym given by EDR thesaurus is too much. More closely syn-

onyms are needed.

Theformercase aresolved byaddingnewconcepts byeditor. However, inthelattercase,

the building thesaurus which covering all requirements is not realistic because a level of

requirementisinconsistforevery users. Inthis paper, we propose methodof addingsuch

informationas distinguish dierenceamong synonyms.

We extract this information from concept explication. However, informationof con-

cept explication can not use by computer because they are written in natural language.

So we make morphological, syntactical, semantical analysis, for making machine handle

informationof concept explication.

Copyrightc 1999byFUJIWARA Shigeru

dictionaries,triedtogetinformationofsyntacticstructureusingsurfacefeatureofconcept

explication. For the reason, we get N-gram statistics from set of concept explication

which classied part of speech. As the result, dierent frequently N-grams every part of

speechare acquiredand itisknown that someof themgive informationwhichcontribute

syntactical and semantical analysis. (we call such N-gram key word.) In this paper, we

dene denite word, to-iu part, ni-oite part. These informationare extracted easily

by using informationof key word.

We explain the process of semantical analysis of concept explication. First, making

groups which consist of synonyms. Secondly, making groups whichconsist of same word

(wecallsuchgroupA-group)andmakinggroupswhichconsistofword havesamegovern

word and same case. (we call such group B-group.) After grouping process, do word

sensedisambiguationfordenitewordsandwordsofto-iupartandwordsofni-oitepartby

usingasemanticalrestrictiongiven bykey wordsandscoringbased information

of EDR thesaurus. Ifa A-group have a word whichxed sense already exist,x sense

of all words of this A-group. Words of B-group is decided sense by using information

that they have similar senses with each other. Words is not decided sense until

this process is decided sense by applying default methodbased information of frequency

of EDR Co-occurrence dictionary.

Forvericationofeectivenessofmethodisproposedus,wedidexperiment. Wemake

21 test sets which consist of leaf node of EDR thesaurus. (Sum of concepts of test set

is 535.) As the result, we get goodresult relatively for denite words and words of to-iu

part andwords of ni-oitewords,recall is56% andprecisionis64%. Andweget resultfor

other words by applying default method, recall is 60%, precision is 55%. Lastly, we get

result for allwords of test set, recall is 60%,precision is55%.

Especially, the methodfor denitewords and words of to-iupart and words of ni-oite

part get goodresult howeverit consist of low-cost methods.

Through the experiment, accuracy is not very high, but there is still room for im-

provement inthis score and ranking.