deep_learning_for_information_retrieval

《deep_learning_for_information_retrieval》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、SIGIR2016TutorialPisaItalyJuly17,2016DeepLearningforInformationRetrievalHangLi&ZhengdongLuHuaweiNoah’sArkLabOutlineofTutorial•Introduction•Part1:BasicsofDeepLearning•Part2:FundamentalProblemsinDeepLearningforIR•Part3:ApplicationsofDeepLearningtoIR•SummaryOverviewofInformationRetri

2、evalContent:Documents,Images,Intent:RelationalTablesKeyWords,QuestionQueryInformationInformationandRetrievalSystemKnowledgeBaseRelevantResultKeyQuestions:HowtoRepresentIntentandContent,HowtoMatchIntentandContent•Ranking,indexing,etcarelessessential•InteractiveIRisnotparticularlyco

3、nsideredhereApproachinTraditionalIRDocument:Query:StarWars:EpisodeVIIstarwarstheforceawakensreviewsThreedecadesafterthedefeatoftheGalacticEmpire,anewthreatarises.qd11f(q,d)00q,df(q,d)VSM

4、

5、q

6、

7、

8、

9、d

10、

11、01•Representingqueryanddocumentastf-idfvectors•Calculatin

12、gcosinesimilaritybetweenthem•BM25,LM4IR,etccanbeconsideredasnon-linearvariantsApproachinModernIRQuery:Document:starwarstheforceawakensreviewsStarWars:EpisodeVIIThreedecadesafterthedefeatoftheGalacticqdEmpire,anewthreatarises.(starwars)vv(theforceawakens)q1f(q,d)d1(reviews)vvqmd

13、n•Conductingqueryanddocumentunderstanding•Representingqueryanddocumentasfeaturevectors•Calculatingmultiplematchingscoresbetweenqueryanddocument•Trainingrankerwithmatchingscoresasfeaturesusinglearningtorank“Easy”ProblemsinIR•Search–Matchingbetweenqueryanddocument•QuestionAnsweringf

14、romDocuments–Matchingbetweenquestionandanswer•Wellstudiedsofar•DeepLearningmaynothelpsomuch“Hard”ProblemsinIR•ImageRetrieval–Matchingbetweentextandimage–Notthesameastraditionalsetting•QuestionAnsweringfromKnowledgeBase–Complicatedmatchingbetweenquestionandfactinknowledgebase•Gener

15、ation-basedQuestionAnswering–Generatinganswertoquestionbasedonfactsinknowledgebase•Notwellstudiedsofar•DeepLearningcanmakeabigdealHardProblemsinIRQuestionAnsweringNameHeightWeightQ:HowtallisYaoMing?fromKnowledgeBaseYaoMing2.29m134kgLiuXiang1.89m85kg(Notagonimages)ImageRetrievalQ:A

16、dogcatchingaballGeneration-basedT