面向知識圖譜的構(gòu)建與推理研究進展

08/26/2023VS.VS.ZHEJIANGUNIVERSITYEmbeddingConcepts:TermsandClassHierEmbeddingConcepts:TermsandClassHierEmbeddingFacts:Entities,Attributes,RelaEmbeddingLogic:Axiomsand[1]HarnessingthePowerofLLMsinPractice:ASurveyonChatGPTandBeyond.20232知識缺失謬誤問題！知識缺失謬誤問題！ZHEJIANGUNIVERSITY語言的有限性--YannLeCun深度學(xué)習(xí)快慢雙系統(tǒng)--YoshuaBengio符號表示落地問題！ZHEJIANGUNIVERSITY語言的有限性--YannLeCun深度學(xué)習(xí)快慢雙系統(tǒng)--YoshuaBengio復(fù)雜推理問題！復(fù)雜推理問題！4浙江大學(xué)上海交通大學(xué)應(yīng)用......浙江大學(xué)上海交通大學(xué)應(yīng)用......ZHEJIANGUNIVERSITYLLAMA......5ZHEJIANGUNIVERSITY姆提出了知識工程的工智能中的重要地位ZHEJIANGUNIVERSITY8ZHEJIANGUNZHEJIANGUNIVERSITY數(shù)據(jù)源：自動檢索（1399）+人工查閱（353）à1727篇文獻(xiàn)ZHEJIANGUNIVERSITY任務(wù)任務(wù)summarization...ZHEJIANGUNIVERSITY技術(shù)技術(shù)retrieval、end-to-end...ZHEJIANGUNIVERSITY“>知識應(yīng)用（~79對話系統(tǒng)、問答系統(tǒng)、推薦系統(tǒng)、代碼搜索/補全、情感分析、因果>預(yù)訓(xùn)練模型（~79知識注入、知識探測、領(lǐng)域化>信息抽?。▇63實體識別、實體分類、關(guān)系抽取、事件抽取>場景擴展（~52知識編輯、持續(xù)學(xué)習(xí)、細(xì)粒度、在線學(xué)習(xí)、時空、跨語言、多模態(tài)、 ZHEJIANGUNIVERSITY 知識圖譜NamedEntityRecognition(NER)RelationRelationExtraction(RE)EventExtraction(EE)EntityEntityLinking(EL)......ZHEJIANGUNIVERSITYZHEJIANGUNIVERSITYZHEJIANGUNIVERSITY[1]UnifiedStructureGenerationforUniversalInformationExtraction.[2]InstructUIE:Multi-taskInstructionTuningforUnifiedInformationExtraction.2023ZHEJIANGUNIVERSITYTechGPTTechGPT:Technology-OrientedGenerativePretrainedTransformer強化：?以“知識圖譜構(gòu)建”為核心的關(guān)系三元組抽取等各類信息抽取任務(wù)?以“閱讀理解”為核心的各類智能問答任務(wù)?以“文本理解”為核心的關(guān)鍵詞生成等各類序列生成任務(wù)[1]Sequence-to-SequenceKnowledgeGraphCompletionandQuestionAnswering.ACZHEJIANGUNIVERSITYZHEJIANGUNIVERSITY[1]StructGPT:AGeneralFrameworkforLargeLanguageModeltoReasonoverStructuredData.2023Language-guidedAbstraLanguage-guidedAbstraZHEJIANGUNIVERSITYNeural-SymbolicEntangledFrNeural-SymbolicEntangledFr[1]Neural-SymbolicEntangledFrameworkforComplexQueryAnswering.NeurIPS2022[2]ComplexLogicalReasoningoverKnowledgeGraphsusingLargeLanguageZHEJIANGUNIVERSITYAgent：以LLM為核心，可以與agent、tool、human等協(xié)作完成任務(wù)Collaboration[1]LLMsforKnowledgeGraphConstructionandReasoning:RecentCapabilitiesandFutureOpportunities.2023[2]CollabKG:ALearnableHuman-Machine-CooperativeInformationExtractionToolkitfor(Event)KnowledgeGraphZHEJIANGUNIVERSITYPLM+KnowledgeGraph02處理對象擴展擴展：模型參數(shù)（隱式知識）PLM+KnowledgeGraphZHEJIANGUNIVERSITYPLMPLM+OntologicalKnowledge[1]DoPLMsKnowandUnderstandOntologicalKnowledge?ACL2023[2]BertNet:HarvestingKnowledgeGraphswithArbitraryRelationsfromPretrainedLanguageModels.ACL2023Findings02處理對象擴展擴展：模型參數(shù)（隱式知識）ZHEJIANGUNIVERSITYPLMPLM+CommonsenseKGPLMPLM+CommonsenseKG(CN)[1]SymbolicKnowledgeDistillation:fromGeneralLanguageModelstoCommonsenseModels.NAACL2022[2]CN-AutoMIC:DistillingChineseCommonsenseKnowledgefromPretrainedLanguageModels.EMNLP2022多模態(tài)實體鏈接多模態(tài)實體關(guān)系抽取多模態(tài)實體鏈接多模態(tài)實體關(guān)系抽取ZHEJIANGUNIVERSITY利用利用語言模型解決多模態(tài)實體、關(guān)系抽取鏈多模態(tài)知識圖譜推理多模態(tài)知識圖譜推理[1]RethinkingMultimodalEntityandRelationExtractionf[2]GenerativeMultimodal[3]MoSE:ModalitySplitandEnsembleforMultimodalKnowledgeGraphCompletion.EMNLP2022SpeechTriplesSpeechTriplesZHEJIANGUNIVERSITYSpeechRESpeechREZHEJIANGUNIVERSITY[1]AnnotatingColumnswithPre-trainedLanguageModels.SIGMOD2022[2]Fromtabulardatatoknowledgegraphs:Asurveyofsemantictableinterpretationtasksandmethods.JWS2023ZHEJIANGUNIVERSITY復(fù)雜本體建模復(fù)雜本體建模ZHEJIANGUNIVERSITY本體知識本體知識事件本體獲取事件本體獲取[1]LanguageModelAnalysisforOntologySubsumptionInference.ACL2023Findings[2]HarvestingEventSchemasfromLargeLanguageModels.2023利用語言模型獲取腳本知識為專業(yè)化小型模型賦予約束利用語言模型獲取腳本知識為專業(yè)化小型模型賦予約束ZHEJIANGUNIVERSITY腳本知識腳本知識[1]DistillingScriptKnowledgefromLargeLanguageModelsforConstrainedLanguagePlanning.ACL2023ZHEJIANGUNIVERSITYANALOGYKBANALOGYKB[1]ANALOGYKB:UnlockingAnalogicalReasoningofLanguageModelswithAMillion-scaleKnowledgeBase.2023ZHEJIANGUNIVERSITYMKGMKG_Analogy[1]MultimodalAnalogicalReasoningoverKnowledgeGraphs.ICLR2023ZHEJIANGUNIVERSITYZHEJIANGUNIVERSITYKoLAKoLA從知識的記憶、理解、應(yīng)用和創(chuàng)造4個層級，深度衡量大語言模型處理世界PLM+知識圖譜推理與問答PLM+知識圖譜推理與問答ZHEJIANGUNIVERSITYPLMPLM+知識圖譜構(gòu)建[1]Zero-ShotInformat[2]EvaluatingChatGPT’sInformationExtractionCapabilities:AnAssessmentofPerformance,Explainab[3]EvaluationofChatGPTasaquestionansweringsystemforansweringcomplexquestions.ISWC2023ZHEJIANGUNIVERSITYAgentBenchAgentBench[1]AgentBench:EvaluatingLLMsasAgents.2023ZHEJIANGUNIVERSITYZHEJIANGUNIVERSITYZHEJIANGUNIVERSITY螞蟻和螞蟻和OpenKG共建的工業(yè)級語義增強框架SPGZHEJIANGUZHEJIANGUNIVERSITYCCKSCCKS2023工業(yè)界論壇：8月27日上午ZHEJIANGUNIVERSITY）？任務(wù)驅(qū)動的知識建模與表示問題；參數(shù)化模型的符號知識定任務(wù)驅(qū)動的知識建模與表示問題；參數(shù)化模型的符號知識定[1]QAisthenewKR:Question-answerPair[3]EditingLargeLanguageModels:Problems,MethodsZHEJIANGUNIVERSITY《理想國》參數(shù)化模型的知識習(xí)得邊界問題；多智能體交互情境下的參數(shù)化模型的知識習(xí)得邊界問題；多智能體交互情境下的[2]InvestigatingtheFactualKnowledgeBoundaryofLargeLanguageModelswithRetrievalAugmentation,2023[3]EncouragingDivergentThinkinginLargeLanguageMo參數(shù)化模型的推理機理問題；符號引擎增強的可信參數(shù)化模型的推理機理問題；符號引擎增強的可信ZHEJIANGUNIVERSITY[1]KnowledgeableorEducatedGuess?RevisitingLanguageModelsasKnowledg[2]Chainofthoughtpromptingeli