鲁棒性的汉语人称代词消解

《鲁棒性的汉语人称代词消解》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。

1、维普资讯http://www.cqvip.com1000·9825／2005／16(05)0700~2005JournalofSoitware软件学报Vo1．16，No．5鲁棒性的汉语人称代词消解王厚峰，梅铮(北京大学计算机科学技术系，北京100871)RobustPronominalResolutionwithinChineseTextWANGHou-Fen，MEIZheng(DepartmentofComputerScienceandTechnology，PekingUniversity，Beijing

2、100871，China)+Correspondingauthor：Phn：+86·10—62753081ext106，E·mail：wanghf@pku．edu．ca，http：／／www．ic1．pku．edu．caReceived2004·-06·27；Accepted2004·-08·10WangHF，MeiZ．RobustpronominalresolutionwithinChinesetext．JournalSoftware，2005，16(5)：700-707．DOI：10．1360／jos1

3、60700Abstract：AnaphoraResolutionisplayingmoreandmoreimportantroleinNaturalLanguageProcessing．Thereisanincreasingneedforthedevelopmentofeffectiveandrobuststrategiesofanaphoraresolutiontomeetthedemandsofpracticalapplications．However,traditionalapproachestoan

4、aphoraresolutionrelyheavilyonmultilevellinguisticknowledge，suchassyntactic，semantic，contextualanddomainknowledge．Itisundoubtedlydificulttoacquiresuchknowledgeatpresent．Thispaperpresentsatwo-stepapproachwithlimitedknowledgetoresolvepronominalanaphorawithinC

5、hinesetext，whichonlyusesnumberfeatures，genderfeaturesandthefeaturesofgrammaticalroles．Inthisapproach，afilterisfirstlyusedtoeliminatethoseexpressionswhosefeaturesareinconsistentwiththepronoun，andthusformasetofpotentialantecedentcandidates；then，ascoringalgor

6、ithmisemployedtocalculatescoreofthecandidates，andthecandidatewiththehighestscoreisselectedastheresultantantecedent．Thealgorithmdoesnotexamineeachcandidateintheset，butautomaticallydeterminewhethertoendthecalculationornotbydynamicallytestingaterminationcondi

7、tion，thereforethecomputationalcomplexityislow．Inaddition，theapproachdoesnotneedadeepanalysisofthetext，andCaneasilybeimplemented．Experimentshowstheresultissatisfactory．Keywords：pronominalanaphoraresolution；antecedent；feature；filter；scorealgorithm摘要：指代消解在自然语

8、言处理中起着越来越重要的作用．许多自然语言处理应用系统都需要高效、鲁棒的指代消解策略．然而，传统的指代消解方法需要用到句法知识、语义知识、上下文知识，甚至领域知识等多级知识，在目前的自然语言处理水平下，要有效荻取这些知识是相当困难的．结合汉语的特点，提出了一种弱化语言知识的人称代词消解方法，仅仅用到了单复数特征、性别特征和语法角色特征．该方法主要分为两步，首先，利用这3种特征的简单约束关系，过滤与人称代词特征不一