0705论文13.pptxVIP免费

贪心科技|让每个人享受个性化教育服务Review:PaperReadingPosition-awareAttentionandSupervisedDataImproveSlotFilling范老师2020/07/05贪心科技|让每个人享受个性化教育服务Content•Introduction•Method•TheTACRelationExtractionDataset•Experiments•RelatedWork•Conclusion贪心科技|让每个人享受个性化教育服务Introduction1.大背景：知识图谱构建时需要抽取关系2.槽填充：找到指定的概念在句子中对应的实例Upuntilnow,automaticknowledgeextractionhasprovensufficientlydifficultthatmostofthefactsintheseknowledgegraphshavebeenbuiltupbyhand.贪心科技|让每个人享受个性化教育服务Method贪心科技|让每个人享受个性化教育服务TheTACRelationExtractionDatasetExistingrelationextractiondatasetsdatasetarelessuseful.Thisismainlybecause:(1)thesedatasetsarerelativelysmallforeffectivelytraininghigh-capacitymodels(seeTable2),and(2)theycaptureverydifferenttypesofrelations.Forexample,theSemEvaldatasetfocusesonsemanticrelations(e.g.,Cause-Effect,Component-Whole)betweentwonominals.Datacollection:usecrowdsourcingtomarkthesubjectandobjectentityspansandtherelationtypes.Datasetstratification:Intotalwecollect106,264examples.Splitdatasetbyyearfordifferentexperimentsteps.Discussion:I.moretypesofrelationsII.ReuseentityandrelationtypesoftheTACKBPtasksIII.AnnotateallnegativeinstancesIV.Enlargethelengthofaveragesentencelength贪心科技|让每个人享受个性化教育服务Experiments–BaselineModelsTACKBP2015winningsystemWecompareagainstStanford’stopperformingsystemontheTACKBP2015coldstartslotfillingtask(Angelietal.,2015).Atthecoreofthissystemaretworelationextractors:apattern-basedextractorandalogisticregression(LR)classifier.Convolutionalneuralnetworks1.CNN(extractsentencefeatures+max-pooling+full-connection+softmax)2.Positionalembeddings(relativepositionofeachwordtothesubjectandobjectentities)Dependency-basedrecurrentneuralnetworks1.Shortestdependencypathsbetweenentitiesareoftenusedasinputtotheneuralnetworks.Eachdependencypathisdividedintotwosub-paths.2LSTM贪心科技|让每个人享受个性化教育服务Experiments–ImplementationDeta...