知识图谱技术综述_徐增林

454Vol.45No.420167JournalofUniversityofElectronicScienceandTechnologyofChinaJul.2016徐增林1，盛泳潘1，贺丽荣1，王雅芳2(1.6117312.250101)知识图谱技术是人工智能技术的重要组成部分，其建立的具有语义处理能力与开放互联能力的知识库，可在智能搜索、智能问答、个性化推荐等智能信息服务中产生应用价值。该文在全面阐述知识图谱定义、架构的基础上，综述知识图谱中的知识抽取、知识表示、知识融合、知识推理四大核心技术的研究进展以及一些典型应用。该文还将评论当前研究存在的挑战。知识融合;知识图谱技术;知识表示;开放互联;语义处理TP182Adoi:10.3969/j.issn.1001-0548.2016.04.012ReviewonKnowledgeGraphTechniquesXUZeng-lin1,SHENGYong-pan1,HELi-rong1,andWANGYa-fang2(1.StatisticalMachineIntelligence&Learning,UniversityofElectronicScienceandTechnologyofChinaChengdu611731;2.SchoolofComputerScienceandTechnology,ShandongUniversityJinan250101)AbstractKnowledgegraphtechnologyisacriticalpartofartificialintelligenceresearch.Itestablishesaknowledgebasewiththecapacityofsemanticprocessingandopeninterconnectioninordertoprovideintelligentinformationservice,suchassearch,question-answering,personalizedrecommendation,andsoon.Thisarticlefirstpresentsacomprehensivestudyondefinitionsandarchitecturesofknowledgegraphs.Thenwesummarizesrecentadvancesinknowledgegraphs,includingknowledgeextraction,knowledgerepresentation,knowledgefusion,andknowledgereasoning,withtypicalapplications.Finally,thisarticleconcludeswithfuturechallengesofknowledgegraphs.Keywordsknowledgefusion;knowledgegraphtechniques;knowledgerepresentation;openinterconnection;semanticprocessing20160515(61572111)(ZYGX2014J058)(1980).WebWeb1.0Web2.0Web3.0[1][2][3][4](knowledgegraph)Web3.02006[5]RDF(resourcedescriptionframework)(RDFschema)(WebontologylanguageOWL)2012517Google[6]45590[7]WebWeb11.1Google[8](,,)GERS12|E|{,,,}Eeee||E12|E|{,,,}Rrrr||RSERE121988-09-08()ID(attribute-valuepairAVP)1.21)(12)()Neo4j[9]TwitterFlockDB[10]sonesGraphDB[11]2)1(top-down)(bottom-up)Freebase[12]GoogleKnowledgeVault[13]4:59112WebRDFLOD(linkedopendata)[14]2.1LODFreebaseWikidataDBpediaYAGO41)FreebaseFreebase[15]MetawebGoogleGoogleFreebaseIMDBFlickr2014Freebase6800102420156FreebaseWikiData2)WikidataWikidata[16]WikipediaWikivoyageWikisource[17]Wikidata17003)DBpediaDBpedia[18]LODDBpediaDBpedia2014DBpedia30DBpedia4)YAGOYAGO[19](maxplanckinstituteMPI)YAGOWordNet[20]GeoNamesWordNet1005002012YAGOGeoNames[21]YAG02s10001.22.245592MusicBrainzIMDB1)IMDBIMDB(internetmoviedatabase)[22]20122IMDB21323834530159IMDB[23]2)MusicBrainzMusicBrainz[24]Last.fmGrooveSharkPandoraEchonestMusicBrainzMusicBrainzWebMusicBrainz[25]3)ConceptNetConceptNet[26]ConceptNetConceptNet5[27]ConceptNetGPLv333.1()33.1.1实体抽取(namedentitylearning)(namedentityrecognition)[28]3Web[12,29]1)[30][31]2)[32]KNNTwitter[33]MedlineGENIA70%3)4:593[34][35]3.1.2关系抽取[36](openinformationextraction,OIE)OIE[37]1)nKnowItAll[38]TextRunner[37][39]WikipediaOIEWOETextRunnerWOE[40]OIEReVerb[41]OIEOLLIE(openlanguagelearningforinformationextraction)OIEOIE40%n[42][43]nKPAKENReVerb2)MLN(Markovlogicnetwork)[44]OIE[45]StatSnowballOIEStatSnowball[37,46]EntSumCRFStatSnowball[37,47]MarkovTML(tractableMarkovlogic)TMLTML3.1.3属性抽取[37,48]WikipediaWordNet95%[49][50]3.2[51-53]3.2.1应用场景[53]455941)[54][55]2)[53]3.2.2代表模型1)[56](structuredembeddingSE)2)[57](singlelayermodelSLM)(,,)hrtT,1,2(,)()rtrhrtfhtglMMlTrkr()gtanh,1rM,2rMdkr3)(latentfactormodelLFM)[58-59](,,)hrtT(,)rhrtfhtlMlrMddrhltld[53][60]rMDISTMULT4)[61](,,)hrtT,1,2(,)()rrhrtrhrtrfhtglMlMlMlbTrkr()gtanhddkrM,1rM,2rMdkr[53]5)[62]RESACLRESCAL(,,)hrt10(,,)hrthrtXThrtlMl2Thrt||hrtLXlMl6)[63]TransErlhltl(,,)hrtTransEhrtlll(,)rfht12/||hrtLLlllhrlltl1L2LTransE4:5953.2.3复杂关系模型1-to-11-to-NN-to-1N-to-N4[63]1-to-NN-to-1N-to-N3TransE[53]1)TransH[64]TransHrlrwFhltlrwFTransH2)TransR[65]TransR(,,)hrtrrhrtlll[65]CTransRCTransRrhtllrcr3)TransDTransRTransR[66]TransD4)TransG[67]TransGTransGr(,,)hrt5)KG2E[68]KG2E(,,)hrthltlehtPll(,)hthrNrrP(,)rrNePrPKL3.2.4多源信息融合[53][69]DKRL(description-embodiedknowledgerepresentationlearning)FreebaseCBOW[70]CNN[71]CNNDKRL[64]word2vecTransE[63]45596word2vec3.3[72][73]3.3.1实体对齐(entityalignment)(entitymatching)(entityresolution)3[74]1)2)[75]3)[74]1)2)3)4)2)3)1)[76][77][78]②[79][80][81][82][83]SVMTAILOR[84][85][86]ALIAS[87]ActiveAtlas2)[88][74]TF-IDF[74]4:5973)[74][89-90][91]SiGMaSiGMa[74]LDA[92]CRF[93]Markov[94][92]LDA[85]CRF[93]CRFcanopy[94]MarkovMarkovMarkovMarkov3.3.2知识加工1)[95]IsA[96][12]MicrosoftProbaseIsAProbase92.8%[97]3[98]2[12,99]KnowltAll[38]TextRunner[37]NELL[100]Probase[101]1[102]2)[103]LDIFREVERRB45598[104]1000logisticGoogleKnowledgeVault[105]3.3.3知识更新[106][107]3.43.4.1基于逻辑的推理(firstorderlogic)(descriptionlogic)[108]/Tbox(terminologybox)ABox(assertionbox)[109]TBoxABox[110]OWL(Webontologylanguage)OWL[111]RDFPD*RDFsesamePD*[112]ORBO3.4.2基于图的推理[113]path-constraintrandomwalkpathranking[75]Tableau4:59944.11)[114]2)[115]31)2)226cm3)[7]GoogleSearch[6]BingSearch[116]CIA[117]BingSearch[116]Facebook[117]Twitter[118][119][120][7]4.2Paralex[121]Siri[122]Evi[123]NuanceTrueKnowledgeSiri[124]OASK[125]4.3Facebook2013GraphSearch[126][7]GraphSearch4.41)45600[127][127]2)Senselab[128]4)[129][7,130][7,130][131][7]5Google5.1()KnowItAllTextRunnerWOEReVerbR2A2KPAKENn[37]5.24:6011)1-to-11-to-NN-to-1