面向知识图谱的自然语言问答优秀课件.ppt

单击此处编辑母版标题样式,编辑母版文本样式,第二级,第三级,第四级,第五级,4,#,Question,Answering,Over,Knowledge,Graph,Lei,Zou,1,4,Knowledge,Graph,Google launches,Knowledge Graph,project,at,2,012.,2,1/3/2025,Knowledge,Graph,Essentially,KG,is,a,sematic,network,which,models,the,entities,(including,properties),and,the,relation,between,each,other.,3,1/3/2025,Resource,Description,Framework,(RDF),RDF,is,an,de facto standard,for,Knowledge,Graph,(KG).,RDF is a,language,for the conceptual modeling of information about web resources,A,building block,of semantic web,Make the information on the web and the interrelationships among them,Machine Understandable,4,1/3/2025,RDF,&SPARQL,Subject,Predicate,Object,Resident_Evil:_Retribution,type,film,Resident_Evil:_Retribution,budget,“6.5E7,Resident_Evil:_Retribution,director,Paul_W._S._Anderson,Paul_W._S._Anderson,type,director,Resident_Evil,director,Paul_W._S._Anderson,Paul_Anderson_(actor),type,actor,The_Revenant,strarring,Philadelphia,Priestley Medal,awards,Paul S.Anderson,Maclovia_(1948_film),distributor,Filmex,RDF,Datasets,SELECT?y WHERE,?x director Paul_W._S._Anderson.,?x type film.,?x budget?y.,SPARQL,5,“,What is the budget of the film directed by Paul Anderson?.,”,1/3/2025,Interdisciplinary Research,Knowledge,Engineering,KB,construction,Rule-based Reasoning,Machine,Learning,Knowledge,Representation,(Graph,Embedding),Natural,Language,Processing,Information,Extraction,Semantic,Parsing,Database,RDF,Database,Data,Integration,、,Knowledge,Fusion,6,1/3/2025,KG-based,Question,/,Answering,SPARQL,syntax,are,too,complex,for,ordinary,users,RDF,KG,is“,schema,-,less,”data,not,like,schema,-,first,relational,database,.,7,1/3/2025,An,Easy-to-Use,Interface to Access Knowledge Graph,It is interesting to both,academia,and,industry,.,Interdisciplinary,research,between database and NLP (natural,language processing)communities.,8,KG-based,Question,/,Answering,1/3/2025,“,(Researchers),They,must,invest,much,more,in,bold,strategies,that,can,achieve,natural,-,language,searching,and,answering,”,-Oren Etzioni,Search,needs,a,shake,up,NATURE,Vol 476,p25-26,2011.,Oren Etzioni,AAAI Fellow,9,KG-based,Question,/,Answering,1/3/2025,Facebook,Graph Search,“My friends who live in Canada”,“Facebook,Graph,Search”,-announced by Mark Zuckerberg on January 16,2013,10,1/3/2025,Facebook Graph Search,“,Photos of,my friends who live in Canada”,11,1/3/2025,EVI-(originally,True,Knowledge),Venture Capital,2007-09,1.2,Million,USD,2008-07,4,Million,USD,2012-01,Acquired,by,Amazon,William Tunstall-Pedoe:,True Knowledge:Open-Domain Question Answering using Structured Knowledge and Inference,.AI Magazine 31(3):80-92(2010),12,1/3/2025,Information Retrieval-based,Generate candidate answers,Ranking,Semantic Parsing-based,Translate NLQ to logical forms,Executing,13,KG-based,Question,/,Answering,1/3/2025,(Cite:Nan Duan,MSRA),14,1/3/2025,Information Retrieval-based,15,KG-based,Question,/,Answering,Paul.W.S.Anderson,film,director,Resident,_Evil,“,6.5E7,”,type,budget,type,“,What is the budget of the film directed by Paul Anderson?,”,director,1/3/2025,Information Retrieval-based,16,KG-based,Question,/,Answering,Paul.W.S.Anderson,film,director,Resident,_Evil,“,6.5E7,”,type,budget,type,“,What is the budget of the film directed by,Paul Anderson,?,”,Mentioned,entity,S,tep.,1,S,tep.,2,Candidate,Answer,Selection,(,within,2-hops),S,tep.,3,Ranking,Answers,“,6.5E7,”,director,1/3/2025,17,Question Answering with Subgraph Embeddings Bordes et al.EMNLP 2014,1/3/2025,18,Question Answering with Subgraph Embeddings Bordes et al.EMNLP 2014,Let,k:the dimension of the embedding space,N:,is the number of words,is the number of entities and relation types,Embedding a question q,is a sparse vector indicating the presence of words(usually 0 or 1).,1/3/2025,Embedding a candidate answer a,is a sparse vector representation of the answer,a,Single Entity,The answer is represented as a single entity:,is a 1-of-Ns coded vector with 1 corresponding the answer.,Path Representation,The answer is represented as a path from the entity mentioned in the question to the answer entity,a,.,is a 3-of-Ns(or 4-of-Ns)coded vector,expressing the start and the end entities of the path and the relation types(but not entities)in-between.,19,Paul.W.S.Anderson,film,director,Resident,_Evil,“,6.5E7,”,type,budget,type,director,2-hop paths,Question Answering with Subgraph Embeddings Bordes et al.EMNLP 2014,Candidate,Answer,1/3/2025,Embedding a candidate answer a,is a sparse vector representation of the answer,a,Subgraph Representation,The answer is represented both the path and 1-hop neighbors around the answer a.,20,Paul.W.S.Anderson,film,director,Resident,_Evil,“,6.5E7,”,type,budget,type,director,1-hop neighbors,Question Answering with Subgraph Embeddings Bordes et al.EMNLP 2014,Candidate,Answer,1/3/2025,21,The loss functionQuestion Answering with Subgraph Embeddings Bordes et al.EMNLP 2014,Scoring Function,candidate answer,question sentence,1/3/2025,Question Answering over Freebase with Multi-Column Convolutional Neural Networks Dong et al.,ACL 2015,22,1/3/2025,Question Answering over Freebase with Multi-Column Convolutional Neural Networks Dong et al.,ACL 2015,Scoring Function,answer path,answer context,answer type,candidate answer,question sentence,23,1/3/2025,MCCNNs for,Question Understanding,Let the question,The,look layer transform,every word into a vector,Question Answering over Freebase with Multi-Column Convolutional Neural Networks Dong et al.,ACL 2015,24,1/3/2025,MCCNNs for,Question Understanding,Let the question,The,max-pooling layer,The,convolutional layer,computes representation of the words in sliding windows.,Question Answering over Freebase with Multi-Column Convolutional Neural Networks Dong et al.,ACL 2015,25,1/3/2025,Embedding Candidate Answers,Answer Path,is a length-|R|binary vector,indicating the presence or absence of every relation in the answer path.,is the parameter matrix,Question Answering over Freebase with Multi-Column Convolutional Neural Networks Dong et al.,ACL 2015,26,1/3/2025,Embedding Candidate Answers,Answer Context,is a length-|C|binary vector,indicating the presence or absence of every entity or relation in the context.,is the parameter matrix,The 1-hop entities and relations connected to the answer path are regarded as the,answer context,.,Question Answering over Freebase with Multi-Column Convolutional Neural Networks Dong et al.,ACL 2015,27,1/3/2025,Embedding Candidate Answers,Answer Type,is a length-|T|binary vector,indicating the presence or absence of answer type.,is the parameter matrix,Type information is an important clue to score candidate answers.,Question Answering over Freebase with Multi-Column Convolutional Neural Networks Dong et al.,ACL 2015,28,1/3/2025,Model Training,For every correct answer a of the question q,we randomly sample k wrong a from the set of candidate answers,and use them as the negative instances to estimate parameters.,Question Answering over Freebase with Multi-Column Convolutional Neural Networks Dong et al.,ACL 2015,29,1/3/2025,30,Question Answering on Freebase via Relation Extraction and Textual Evidence,Xu et al.,ACL 2016,胡森,1/3/2025,31,Question Answering on Freebase via Relation Extraction and Textual EvidenceXu et al.,ACL 2016,Relation Extraction,1/3/2025,32,who plays ken barlow in coronation street?“,decompose,“who plays ken barlow”,+,“who plays in coronation street”,Question Answering on Freebase via Relation Extraction and Textual EvidenceXu et al.,ACL 2016,Question Decomposition,1/3/2025Information Retrieval-based,Generate candidate answers,Ranking,Semantic Parsing-based,Translate NLQ to logical forms,Executing,33,KG-based,Question,/,Answering,1/3/2025,Semantic,Parsing,Zettlemoyer,et,al.,UAI,05,Transforming natural language(NL)sentences into computer executable complete meaning representations(MRs),for domain-specic applications.,E.g.,“,Which states borders New Mexico?,”,Lambda-,calculus,Alonzo Church,1940,“,Simply typed,Lambda-calculus,can express varies database query languages such as,relational algebra,fixpoint logic and the complex object algebra.Hillebrand et al.,1996,34,1/3/2025,Semantic,Parsing,Manually constructed rules,Pedoe,AI magazine 2010,Grammar-based,e.g.,Combinatory Categorial Grammar,Zettlemoyer and Collins,UAI 2005,Supervised Learning,Berant and Liang,ACL 2014,Template,-,based,Approach,cite:,Weiguo,Zheng,Lei,Zou,et,al.,SIGMOD,15,35,1/3/2025,Semantic Parsing via Staged Query Graph Generation:,Question Answering with Knowledge Base Yih et al.,ACL 2015,36,1/3/2025,Semantic Parsing via Staged Query Graph Generation:,Question Answering with Knowledge Base Yih et al.,ACL 2015,Query Graph Generation,37,1/3/2025,Semantic Parsing via Staged Query Graph Generation:,Question Answering with Knowledge Base Yih et al.,ACL 2015,Query Graph Generation,38,1/3/2025,Semantic Parsing via Staged Query Graph Generation:,Question Answering with Knowledge Base Yih et al.,ACL 2015,Reward Function,39,1/3/2025,Semantic Parsing via Staged Query Graph Generation:,Question Answering with Knowledge Base Yih et al.,ACL 2015,Identifying Core Inferential Chain,(Relation Extraction),two neural networks,1)question,2)inferential chain,Compute Similarity,(e.g.cosine),40,1/3/2025,Language to Logical Form with Neural Attention,Dong et al.,ACL 2016,41,1/3/2025,Language to Logical Form with Neural Attention,Dong et al.,ACL 2016,42,1/3/2025,Using,graph,matching,-,based,method,Graph,Matching,-,based,Disambiguation,Combing,Disambiguation,and,Query,together,43,Our,Approach,-,Data,Driven&Relation-first framework,gAnswer,Zou,et,al,SIGMOD,14,1/3/2025,Semantic Query Graph,Our,Approach,-,Data,Driven&Relation-first framework,gAnswer,Zou,et,al,SIGMOD,14,44,1/3/2025,Our,Approach-,Data,Driven&Relation-first framework,gAnswer,Zou,et,al,SIGMOD,14,45,Besides KG,we require two dictionaries.,Entity Mention Dictionary,Relation Mention Dictionary,It helps the entity linking task,Spitkovsky et al.,LERC 12;Chisholm et al,TACL 15.,Mapping the natural language relation phrases to predicate in RDF dataset.,Nakashole et al.,EMNLP-CoNLL 2012,1/3/2025,Question Understanding,-Relation extraction,Relation Paraphrase Dictionary,Our,Approach-,Data,Driven&Relation-first framework,gAnswer,Zou,et,al,SIGMOD,14,46,1/3/2025,Question Understanding,-Find associated arguments,Our,Approach-,Data,Driven&Relation-first framework,gAnswer,Zou,et,al,SIGMOD,14,47,1/3/2025,Question Understanding,-Query,Graph,Assembly,Our,Approach,-,Data,Driven&Relation-first framework,gAnswer,Zou,et,al,SIGMOD,14,48,1/3/2025,Question Answering over Knowledge Graph,胡森,Query Execution,Our,Approach-,Data,Driven&Relation-first framework,gAnswer,Zou,et,al,SIGMOD,14,49,1/3/2025,Limitations,Still highly relied on parser and heuristic rules,Can not handle implicit relations,What is the budget of the film directed by Paul Anderson and starred by a,Chinese girl,Our,Approach,-,Data,Driven&Relation-first framework,gAnswer,Zou,et,al,SIGMOD,14,50,1/3/2025,Data Driven!,The structure of query graph can be modified in execution stage.,First recognize nodes.,Our,Approach-,Data,Driven&Node-first framework,gAnswer+,Hu,and Zou et,al,TKDE,17,51,1/3/2025,Hyper Query Graph,Extend SQG by,allowing false edges,.,query graph,semantic query graph,hyper query graph,Our,Approach-,Data,Driven&Node-first framework,gAnswer+,Hu,and Zou et,al,TKDE,17,52,1/3/2025,Question Understanding,-Node recognizing,entity extraction+,conflict resolution,-entity,type,literal,wildcard,constant,variable,modified,hidden information,Our,Approach-,Data,Driven&Node-first framework,gAnswer+,Hu,and Zou et,al,TKDE,17,53,1/3/2025,Question Understanding,Build structure of HQG,connect which two nodes?,Our,Approach-,Data,Driven&Node-first framework,gAnswer+,Hu,and Zou et,al,TKDE,17,54,1/3/2025,Our,Approach-,Data,Driven&Node-first framework,gAnswer+Hu,and Zou et,al,TKDE,17,55,1/3/2025,Question Understanding,Finding relations,Explicit,relation,Our,Approach-,Data,Driven&Node-first framework,gAnswer+,Hu,and Zou et,al,TKDE,17,56,1/3/2025,Question Understanding,Finding relations,Implicit relation,Our,Approach-,Data,Driven&Node-first framework,gAnswer+,Hu,and Zou et,al,TKDE,17,Locating,the,two,nodes,in,KG,and,finding,the frequent,predicate,between,them,.,57,1/3/2025,Query Executing,A top-down algorithm,Nave method,(1)Enumerate spanning subgraph of HQG,(2)Call algorithm SQG executing algorithm,(3)Sort and select top-k matches,Advanced method,Add to the candidate list of unsteady edges,Call algorithm 3,Our,Approach-,Data,Driven&Node-first framework,gAnswer+,Hu,and Zou et,al,TKDE,17,58,1/3/2025,Query Executing,-A top-down algorithm,Drawbacks,-Query graphs with higher scores may have no matches,s,p,o,var,Our,Approach-,Data,Driven&Node-first framework,gAnswer+Hu,and Zou et,al,TKDE,17,59,1/3/2025,Query Executing,A bottom-up algorithm,Intuition,Growing structures step by step,Keep correct structures when growing,Find matches of multi-label query graph(SQG),Drop useless candidates as early as possible,Our,Approach-,Data,Driven&Node-first framework,gAnswer+,Hu,and Zou et,al,TKDE,17,60,1/3/2025,Query Executing,-A bottom-up algorithm,Our,Approach-,Data,Driven&Node-first framework,gAnswer+,Hu,and Zou et,al,TKDE,17,61,1/3/2025,Query Executing,A bottom-up algorithm,Optimization,Call GraphExplore()only when adding unsteady edges,Design cost model to estimate the best explore order,Our,Approach-,Data,Driven&Node-first framework,gAnswer+,Hu,and Zou et,al,TKDE,17,62,1/3/2025,Experiments,QALD is a series of evaluation campaigns on question answering over linked data.,QALD-6,Competition,Results,63,1/3/2025,Experiments,WebQuestions is widely used in Question Answering literatures and does not contain golden SPARQL queries.,WebQuestions Results,64,1/3/2025,Online Demo,URL:,ganswer.gstore-?x WHERE,?x rdf:type Post.,?x:longitude?o.,?x:latitude?a.,ORDERY BY Dist(HERE,?o,?a),LIMIT 1,66,Is,it,Possible,?,1/3/2025,67,与深圳狗尾草公司合作,.,6600,万,Triples,Zhishi,.me,刘德华的女儿是？,柬埔寨首都在哪儿？,An,open-source,Graph,RDF,database,SPARQL,公子小白,1/3/2025,参考,文献：,Antoine Bordes,Nicolas Usunier,Alberto Garca-Durn,Jason Weston,Oksana Yakhnenko:,Translating Embeddings for Modeling Multi-relational Data.NIPS 2013:2787-2795,Luke S.Zettlemoyer,Michael Collins:,Learning to Map Sentences to Logical Form:Structured Classification with Probabilistic Categorial Grammars.UAI 2005:658-666,Pablo N.Mendes,Max Jakob,Christian Bizer:,DBpedia:A Multilingual Cross-domain Knowledge Base.LREC 2012:1813-1817,Fabian M.Suchanek,Gjergji Kasneci and Gerhard Weikum,Yago-A Core of Semantic Knowledge,16th international World Wide Web conference(WWW 2007),Kurt D.Bollacker,Colin Evans,Praveen Paritosh,Tim Sturge,Jamie Taylor:,Freebase:a collaboratively created graph database for structuring human knowledge.SIGMOD Conference 2008:1247-1250,Peter Buneman,Gao Cong,Wenfei Fan,Anastasios Kementsietsidis:Using Partial Evaluation in Distributed Query Evaluation.VLDB 2006:211-222,Yuk Wah Wong,Raymond J.Mooney:Learning for Semantic Parsing with Statistical Machine Translation.HLT-NAACL 2006,C.