Word Semantic Representations using Bayesian Probabilistic Tensor Factorization

《Word Semantic Representations using Bayesian Probabilistic Tensor Factorization 》由會員上傳分享，免費在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。

1、WordSemanticRepresentationsusingBayesianProbabilisticTensorFactorizationJingweiZhangandJeremySalwenMichaelGlassandAl?oGliozzoColumbiaUniversityIBMT.J.WastonResearchComputerScienceYorktownHeights,NY10598,USANewYork,NY10027,USA{mrglass,gliozzo}@us.ibm.com{jz2541,jas2312}@columbia.

2、eduAbstractworddistributions.Forinstance,theyarebelievedtohavedif?cultydistinguishingantonymsfromManyformsofwordrelatednesshavebeensynonyms,becausethedistributionofantonymousdeveloped,providingdifferentperspec-wordsareclose,sincethecontextofantonymoustivesonwordsimilarity.Weintr

3、oducewordsarealwayssimilartoeachother(Moham-aBayesianprobabilistictensorfactoriza-madetal.,2013).Althoughsomeresearchclaimstionmodelforsynthesizingasinglewordthatincertainconditionstheredoexistdiffer-vectorrepresentationandper-perspectiveencesbetweenthecontextsofdifferentantony-

4、lineartransformationsfromanynumbermouswords(Scheibleetal.,2013),thedifferencesofwordsimilaritymatrices.Theresult-aresubtleenoughthatitcanhardlybedetectedbyingwordvectors,whencombinedwiththesuchlanguagemodels,especiallyforrarewords.per-perspectivelineartransformation,ap-Anotherim

5、portantclassoflexicalresourceforproximatelyrecreatewhilealsoregulariz-wordrelatednessisalexicon,suchasWord-ingandgeneralizing,eachwordsimilarityNet(Miller,1995)orRoget’sThesaurus(Kipfer,perspective.2009).Manuallyproducingorextendinglexi-Ourmethodcancombinemanuallycre-consismuchm

6、orelaborintensivethangenerat-atedsemanticresourceswithneuralwordingVSMwordvectorsusingacorpus.Thus,lex-embeddingstoseparatesynonymsandiconsaresparsewithmissingwordsandmulti-antonyms,andiscapableofgeneraliz-wordtermsaswellasmissingrelationshipsbe-ingtowordsoutsidethevocabularyoft

7、weenwords.Consideringthesynonym/antonymanyparticularperspective.Weevaluatedperspectiveasanexample,WordNetanswerslessthewordembeddingswithGREantonymthan40%percentofthetheGREantonymques-questions,theresultachievesthestate-of-tionsprovidedbyMohammadetal.(2008)di-the-artperformance.

8、rectly.Moreover,binaryentriesinlexiconsdonotind