利用基于图互增理论的自举算法学习语义辞典.pdf

《利用基于图互增理论的自举算法学习语义辞典.pdf》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、Vol.34,No.10ACTAAUTOMATICASINICAOctober,2008LearningSemanticLexiconsUsingGraphMutualReinforcementBasedBootstrapping1111ZHANGQiQIUXi-PengHUANGXuan-JingWULi-DeAbstractThispaperpresentsamethodtolearnsemanticlexiconsusinganewbootstrappingmethodbasedongraphmutualreinforcement(GMR).Theapproachusesonl

2、yunlabeleddataandafewseedwordstolearnnewwordsforeachsemanticcategory.Di®erentfromotherbootstrappingmethods,weuseGMR-basedbootstrappingtosortthecandidatewordsandpatterns.Experi-mentalresultsshowthattheGMR-basedbootstrappingapproachoutperformstheexistingalgorithmsbothinin-domaindataandout-domaind

3、ata.Furthermore,itshowsthattheresultdependsonnotonlythesizeofthecorpusbutalsothequality.KeywordsSemanticlexicon,bootstrapping,graphmutualreinforcement(GMR)[1¡3]Inrecentyears,bootstrappingmethodshavere-bilemanufacturenamesandautomobilepartsextractedbyceivedconsiderableattentioninmanyapplication¯

4、elds,GMR-bootstrappingfromChinesecorpus(detailedinSec-andsemanticlexicons[4¡6]haveprovedusefulformanynat-tion2)wasbetterthanthequalityoflexiconsextractedbyurallanguageprocessingtasks.Althoughsupervisedmeth-Basilisk.Thereminderofthepaperisorganizedasfollows:odsusuallycanachievebetterresultsthant

5、hosebysemi-InSection1,weintroduceourbootstrappingstructureandsupervisedandunsupervisedmethods,theyarestronglyscoringfunctions.InSection2,experimentsaregiventoconstrainedbythenumberoflabeleddata.Usually,boot-showtheimprovements.Section3discussestherelatedstrappingmethodscanusebothasmallsizeoflab

6、eleddataworks.Section4concludesthepaper.andalargeamountofunlabeledsampleswithoutextracosttoobtainbetterresult.1GMR-bootstrappingSemanticlexiconshaveprovedtobeusefulformany[20]GMR-bootstrappingisaweaklysupervisedlearningnaturallanguageprocessingtasks,includingquestion[7¡8][9]method.Likeotherboot

7、strappingmethods,theinputsofanswering,informationextractionandsoon.Learn-GMR-bootstrappingarealargeamountofunlabeleddataingsemanticlexiconsisatasktoautomaticallyacquireandafewmanuallyselectedseedwordsforeachsemanticwordswithsemant