Dependency Based Embeddings for Sentence Classification Tasks

《Dependency Based Embeddings for Sentence Classification Tasks》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、DependencyBasedEmbeddingsforSentenceClassificationTasksAlexandrosKomninosSureshManandharDepartmentofComputerScienceDepartmentofComputerScienceUniversityofYorkUniversityofYorkUKUKak1153@york.ac.uksuresh@cs.york.ac.ukDraft14March2016.ToappearinNAACLal.,2011;Ki

2、m,2014).WordembeddingsprovideHLT2016.bettergeneralizationtounseenexamplessincetheyAbstractcancapturegeneralsemanticandsyntacticproper-tiesofwords.OneofthemostpopularmethodsofWecomparedifferentwordembeddingsfromlearningwordembeddingsistheskipgrammodelofastan

3、dardwindowbasedskipgrammodel,Mikolovetal.(2013a;2013b)whereembeddingsaskipgrammodeltrainedusingdependencyaretrainedbymakingpredictionsofcontextwordscontextfeaturesandanovelskipgramvariantthatutilizesadditionalinformationfromde-appearinginawindowaroundatarge

4、tword.pendencygraphs.Weexploretheeffective-Thestandardskipgrammodelignoressyntaxandnessofthedifferenttypesofwordembeddingsonlypartiallytakesintoconsiderationthesequen-forwordsimilarityandsentenceclassificationtialstructureoftext,butstillcapturescertainsyn-ta

5、sks.Weconsiderthreecommonsentenceclassificationtasks:questiontypeclassifica-tacticpropertiesofwords.AsignificantamounttionontheTRECdataset,binarysentimentofpreviousresearchhasexploredmethodsfordi-classificationonStanford’sSentimentTree-rectlytakingsyntaxintoacc

6、ountforwordembed-bankandsemanticrelationclassificationondinglearning(Baronietal.,2015;ChengandKart-SemEval2010dataset.Foreachtaskweusesaklis,2015;Hashimotoetal.,2014).Onesimplethreedifferentclassificationmethods:aSup-methodisbasedontraditionalcount-baseddistr

7、ibu-portVectorMachine,aConvolutionalNeu-tionalsemanticspacesandutilizeswordswithsyn-ralNetworkandaLongShortTermMemoryNetwork.Ourexperimentsshowthatdepen-tactictypesfromadependencyparsegraphascon-dencybasedembeddingscanoutperformstan-textfeatures(PadoandLapa

8、ta,2007;Baroniand´dardwindowbasedembeddingsinmostoftheLenci,2010).Thismethodhasalsobeenappliedtotaskswhileusingdependencycontextembed-skipgrammodels,wherewordspredictdependencydingsasadditionalfeatures