Knowledge graphs - The Alan Turing Institute

Introduction · Facilitate access to and integration of data sources; · Add context and depth to other, more data-driven AI techniques such as machine learning; ... Skiptomaincontent Researchareas Datastructures Databases Knowledgerepresentation Graphtheory Applications(Machinelearning) Introduction Knowledgegraphs(KGs)organisedatafrommultiplesources,captureinformationaboutentitiesofinterestinagivendomainortask(likepeople,placesorevents),andforgeconnectionsbetweenthem.IndatascienceandAI,knowledgegraphsarecommonlyusedto: Facilitateaccesstoandintegrationofdatasources; Addcontextanddepthtoother,moredata-drivenAItechniquessuchasmachinelearning;and Serveasbridgesbetweenhumansandsystems,suchasgeneratinghuman-readableexplanations,or,onabiggerscale,enablingintelligentsystemsforscientistsandengineers.  ThisinterestgroupwillfacilitateresearchandinnovationinacriticalareaofdatascienceandAI. Explainingthescience Theterm‘knowledgegraph’hasbeenintroducedbyGooglein2012torefertoitsgeneral-purposeknowledgebase,thoughsimilarapproacheshavebeenaroundsincethebeginningofmodernAIinareassuchasknowledgerepresentation,knowledgeacquisition,naturallanguageprocessing,ontologyengineeringandthesemanticweb.Today,KGsareusedextensivelyinanythingfromsearchenginesandchatbotstoproductrecommendersandautonomoussystems.Indatascience,commonusecasesarearoundaddingidentifiersanddescriptionstodataofvariousmodalitiestoenablesense-making,integration,andexplainableanalysis. InAI,knowledgegraphscomplementmachinelearningtechniquesto: reducetheneedoflarge,labelleddatasets; facilitatetransferlearningandexplainability; encodedomain,taskandapplicationknowledgethatwouldbecostlytolearnfromdataalone. Aknowledgegraphorganisesandintegratesdataaccordingtoanontology,whichiscalledtheschemaoftheknowledgegraph,andappliesareasonertoderivenewknowledge.Knowledgegraphscanbecreatedfromscratch,e.g.,bydomainexperts,learnedfromunstructuredorsemi-structureddatasources,orassembledfromexistingknowledgegraphs,typicallyaidedbyvarioussemi-automaticorautomateddatavalidationandintegrationmechanisms. Whiletherearemanydefinitionsofknowledgegraphsaround,mostofthemagreethatknowledgegraphsare: Graphs:unlikeknowledgebases,thecontentofKGsisorganisedasagraph,wherenodes(entitiesofinterestandtheirtypes),relationshipsbetweenandattributesofthenodesareequallyimportant.Thismakesiteasytointegratenewdatasetsandformatsandsupportsexplorationbynavigatingfromonepartofthegraphtotheotherthroughlinks. Semantic:themeaningofthedataisencodedforprogrammaticuseinanontology,whichdescribesthetypesofentitiesinthegraphandtheircharacteristicsandcanberepresentedasaschemasub-graph.Thismeansthatthegraphisbothaplacetoorganiseandstoredata,andtoreasonwhatitisaboutandderivenewinformation. Alive:knowledgegraphsareflexibleintermsofthetypesofdataandschemastheycansupport.They,includingtheirschemas,evolvetoreflectchangesinthedomainandnewdataisaddedtothegraphasitbecomesavailable. Someknowledgegraphsareusedprimarilywithintheorganisationthatcreatedthem.ThemostcommonexampleistheGoogleknowledgegraph,whichisusedinwebsearch,orAmazon’sproductgraph.Otherknowledgegraphsareopenlyavailable.TheseincludeDBpedia,Wikidata,WordNet,Geonames,etc. Aims TheUKhasbeenattheforefrontofseveralstrandsofworkthathavemadeKGsassuccessfulastheyaretoday,inareaslike:knowledgerepresentationformalisms;naturallanguageprocessing;machinelearning;methodologiestoconstruct,learnandmanageontologiesandknowledgebases;defactostandardontologiesandvocabularies;scalablereasoning,linkeddataetc. Themainobjectivesoftheinterestgroupare: Tostrengthenthisnationalcommunityofscholarsandinnovatorstocontinuetopursueworld-leadingresearchandexplorenoveltechnologiesrelatedtoandapplicationsofKGs, ToreinforcethesystematicuseofKGsinpracticaldatascienceandAIapplications. ToidentifyaportfolioofjointresearchprojectsforcurrentandfuturePGstudents. Thegroupwillencouragemembersto: Presentlatestideasandachievements; Shareideas,knowledgeandexperiences; Forgecollaborations; Alignandexpandexistingeducationandtrainingactivities; Reflectuponandraiseawarenessofspecificchallengesinequality,diversityandinclusioninthefield;and Poolresourcesandexpertisetocollectivelyunlocknewfundingandpartnershipopportunitiesthatarenotaccessibletomembersinisolation. Talkingpoints Constructingandmaintaininglarge-scale,yethigh-qualityknowledgegraphs Modernknowledgegraphsaretheresultofcomplexassembliesofmanualandautomaticmodellinganddataingestionpipelines.Stayingontopoftheseprocesseswhileensuringthattheinformationremainsuptodate,consistentandtrustworthyrequiresspecialisedsocio-technicalmethodsrangingfromknowledgeacquisitiontonaturallanguageprocessingtomachinelearningandhuman-computerinteraction. Knowledge-enhanceddata-driventechnologies Knowledgegraphsareneverusedinisolation.Theyco-existandcomplementdata-drivensolutions,includingnaturallanguageprocessing(textunderstanding,knowledgeextraction,textgeneration)anddatamanagement(semanticlabelling,recordlinking,datacleansingetc.).Recentresearchlooks,forinstance,attrade-offsbetweensemanticrepresentationsanddeeplearning,and,morebroadlyatdesigningsymbolic/connectionistAIarchitectures. Newformsofknowledgegraphs KnowledgegraphsneedtobeasrichastheAIstheyserve.WeneednewideasandformalismstogobeyondthetypesofknowledgecapturedincurrentopenandenterpriseKGs,whichtendtofocusonentitiesandtheirpropertiesandonfactualknowledge.Newareasincludecommon-senseknowledge,multipleviewpoints,aswellaseventsandothertypesoftemporalinformation,andcause-effectchains. Knowledgegraphsinhuman-AIsystems Knowledgegraphsareatthecoreofmanyhuman-facingtechnologies,suchassearch,questionanswering,dialogueandrecommenders.Itisimportanttoacknowledgetherequirementseachofthesescenariosposetohowgraphsarecreated,maintainedandused. Applicationsofknowledgegraphs Manyorganisations,suchashealthcareandfinancialserviceproviders,arefacedwithdatasilosacrosstheirorganisationalunits.Knowledgegraphscanhelpwith,butnotlimitedto,datagovernance,frauddetection,knowledgemanagement,search,chatbot,recommendation,aswellasintelligentsystemsacrossdifferentorganisationalunits. Howtogetinvolved Clickheretojoinusandrequestsign-up Recentupdates Theinformationaboutpastandfuture meet-ups,listof membersandrelevant teachingresourcescanbeaccessedfromourGitHubrepository. Organisers Dr Jeff Pan ReaderofKnowledgeGraphs,SchoolofInformatics,UniversityofEdinburgh Professor Elena Simperl ProfessorofComputerScienceatKing'sCollegeLondon Dr Ernesto Jiménez-Ruiz Lecturer,City,UniversityofLondon Professor Ian Horrocks TuringFellow Contactinfo [email protected] JumpTo Researchareas Introduction Explainingthescience Aims Talkingpoints Recentupdates Organisers Contactinfo