知识图谱_梳理

知识图谱需要的技术知识图谱架构•知识图谱一般架构:[来源自百度百科]•复旦大学知识图谱架构:•早期知识图谱架构•知识图谱一般架构:[来源自百度百科]架构讨论数据检索预处理构建关系矩阵网络图谱参数调整可视化数据规范化处理结果导读•早期知识图谱架构知识抽取•实体概念抽取•实体概念映射•关系抽取•质量评估KDD2014TutorialonConstructingandMiningWeb-scaleKnowledgeGraphs,NewYork,August24,2014Asamplerofresearchproblems•••••••••••••Growth:knowledgegraphsareincomplete!Linkprediction:addrelationsOntologymatching:connectgraphsKnowledgeextraction:extractnewentitiesandrelationsfromweb/textValidation:knowledgegraphsarenotalwayscorrect!Entityresolution:mergeduplicateentities,splitwronglymergedonesErrordetection:removefalseassertionsInterface:howtomakeiteasiertoaccessknowledge?Semanticparsing:interpretthemeaningofqueriesQuestionanswering:computeanswersusingtheknowledgegraphIntelligence:canAIemergefromknowledgegraphs?AutomaticreasoningandplanningGeneralizationandabstraction9关系抽取定义:常见手段:语义模式匹配[频繁模式抽取,基于密度聚类,基于语义相似性]层次主题模型[弱监督]KDD2014TutorialonConstructingandMiningWeb-scaleKnowledgeGraphs,NewYork,August24,2014Methodsandtechniques•••SupervisedmodelsSemi-supervisedmodelsDistantsupervision2.Entityresolution•Singleentitymethods•Relationalmethods3.Linkprediction••••Rule-basedmethodsProbabilisticmodelsFactorizationmethodsEmbeddingmodels80Notinthistutorial:•Entityclassification•Group/expertdetection•Ontologyalignment•Objectranking1.Relationextraction:KDD2014TutorialonConstructingandMiningWeb-scaleKnowledgeGraphs,NewYork,August24,2014•Extractingsemanticrelationsbetweensetsof[grounded]entities•Numerousvariants:•••••Undefinedvspre-determinedsetofrelationsBinaryvsn-aryrelations,facetdiscoveryExtractingtemporalinformationSupervision:{fully,un,semi,distant}-supervisionCuesused:onlylexicalvsfulllinguisticfeatures82RelationExtractionKobeBryantLALakersplayForthefranchiseplayerofonceagainsavedmanofthematchfortheLakers”histeam”LosAngeles”“KobeBryant,“Kobe“KobeBryant?KDD2014TutorialonConstructingandMiningWeb-scaleKnowledgeGraphs,NewYork,August24,2014Supervisedrelationextraction•Sentence-levellabelsofrelationmentions••AppleCEOSteveJobssaid..=(SteveJobs,CEO,Apple)SteveJobssaidthatApplewill..=NIL•Traditionalrelationextractiondatasets•••ACE2004MUC-7Biomedicaldatasets(e.gBioNLPclallenges)••Learnclassifiersfrom+/-examplesTypicalfeatures:contextwords+POS,dependencypathbetweenentities,namedentitytags,token/parse-path/entitydistance83KDD2014TutorialonConstructingandMiningWeb-scaleKnowledgeGraphs,NewYork,August24,2014Semi-supervisedrelationextraction•Genericalgorithm(遗传算法)1.2.3.4.5.Startwithseedtriples/goldenseedpatternsExtractpatternsthatmatchseedtriples/patternsTakethetop-kextractedpatterns/triplesAddtoseedpatterns/triplesGoto2•••••Manypublishedapproachesinthiscategory:DualIterativePatternRelationExtractor[Brin,98]Snowball[Agichtein&Gravano,00]TextRunner[Bankoetal.,07]–almostunsupervisedDifferinpatterndefinitionandselection86founderOfKDD2014TutorialonConstructingandMiningWeb-scaleKnowledgeGraphs,NewYork,August24,2014Distantly-supervisedrelationextraction88•••Existingknowledgebase+unlabeledtextgenerateexamplesLocatepairsofrelatedentitiesintextHypothesizesthattherelationisexpressedGoogleCEOLarryPageannouncedthat...SteveJobshasbeenAppleforawhile...Pixarlostitsco-founderSteveJobs...IwenttoParis,Franceforthesummer...GoogleCEOcapitalOfLarryPageFranceAppleCEOPixarSteveJobsDistantsupervision:modelinghypothesesTypicalarchitecture:1.Collectmanypairsofentitiesco-occurringinsentencesfromtextcorpus2.If2entitiesparticipateinarelation,severalhypotheses:1.Allsentencesmentioningthemexpressit[Mintzetal.,09]“BarackObamaisthe44thandcurrentPresidentoftheUS.”(BO,employedBy,USA)89KDD2014TutorialonConstructingandMiningWeb-scaleKnowledgeGraphs,NewYork,August24,2014KDD2014TutorialonConstructingandMiningWeb-scaleKnowledgeGraphs,NewYork,August24,2014Sentence-levelfeatures●●●●●Lexical:wordsinbetweenandaroundmentionsandtheirparts-of-speechtags(conjunctiveform)Syntactic:dependencyparsepathbetweenmentionsalongwithsidenodesNamedEntityTags:forthementionsConjunctionsoftheabovefeaturesDistantsupervisionisusedontolotsofdatasparsityofconjunctiveformsnotanissue92Distantsupervision:modelinghypothesesTypicalarchitecture:1.Collectmanypairsofentitiesco-occurringinsentencesfromtextcorpus2.If2entitiesparticipateinarelation,severalhypotheses:1.2.Allsentencesmentioningthemexpressit[Mintzetal.,09]Atleastonesentencementioningthemexpressit[Riedeletal.,10]“BarackObamaisthe44thandcurrentPresidentoftheUS.”(BO,employedBy,USA)“ObamaflewbacktotheUSonWednesday.”(BO,employedBy,USA)95KDD2014TutorialonConstructingandMiningWeb-scaleKnowledgeGraphs,NewYork,August24,2014Distantsupervision:modelinghypothesesTypicalarchitecture:1.Collectmanypairsofentitiesco-occurringinsentencesfromtextcorpus2.If2entitiesparticipateinarelation,severalhypotheses:1.2.3.Allsentencesmentioningthemexpressit[Mintzetal.,09]Atleastonesentencementioningthemexpressit[Riedeletal.,10]Atleastonesentencementioningthemexpressitand2entitiescanexpressmultiplerelations[Hoffmannetal.,11][Surdeanuetal.,12]“BarackObamaisthe44thandcurrentPresidentoftheUS.”(BO,employedBy,USA)“ObamaflewbacktotheUSjustWedne