挖掘技術(shù)textmining以提供文獻(xiàn)內(nèi)容梗概為目的不加評(píng)論和補(bǔ)充

1 Anexpressofacertain withoutanyexnationsandcomment.It'sunnecessarytoknowwhowritesthesummary.(ANSI)Aconciseandaccurateexpressofthewithoutanyex nationandcomment.Asummaryisindependentontheauthorofthesummary.(ISO214-1976(E))2 性/指示性 以性的形式表述文獻(xiàn)中信息價(jià)值較高的部分，而以指示性的形式表述其余部分的。 3SummaryClassifiedbyuser'sGenericSummarizationUser-querySummarizationClassifiedbytext ClassifiedbySummarizationBasedonExtractionSummarizationBasedonUnderstandingClassifiedbyneedSupervisedSummarizationUnsupervisedSummarization4 5ABriefHistoryofDUC,holdbyNIST,fromByMarkT.MayburyandInderjeetMani6 Vivisimo公司 78內(nèi)部評(píng)價(jià)方法(IntrinsicMethods)：在提供參考的前提下，以參考為基準(zhǔn)評(píng)價(jià)系統(tǒng)的質(zhì)量。通常情況下，系統(tǒng)摘要與參考越吻合，其質(zhì)量越高。外部評(píng)價(jià)方法(ExtrinsicMethods)：不需要提供參考，利用文檔代替原文檔夠提高應(yīng)用性能的被認(rèn)為是質(zhì)量好的摘9Evaluation-- 得與目標(biāo) (coselectionrate)。 Evaluation-- “?！薄埃骸薄埃弧薄?！”“？”；為使專(zhuān)家與機(jī)械具有可比性，只專(zhuān)家和機(jī)械的句子都按照在原文 n

RecallNhm/ PrecisionNhm ? UnderstandingConference(DUC)isaseriesofsummarizationevaluationsthathavebeenconductedbytheNationalInstituteofStandardsandTechnology(NIST)since2001.Itsgoalistofurtherprogressinautomatictextsummarizationandenableresearcherstoparticipateinlarge-scaleexperimentsinboththedevelopmentandevaluationofsummarizationSince2008, DUChasmovedtotheTextysisConference(TAC)QuestionAnswering;RecognizingSummarizationKeyword ymostEasytoHardtoread,poorrepresentationofSentenceExtractkeySummariesoftendon’treadNaturallanguageunderstanding/HardtodoSomethingbetweenthelasttwo SummarizationWebThisis

inText

Content:weight=Sumoftopicterms’weights;Length:weight/=Position:weight*=positionConten;Stigmawords;Stigma:Conten;Stigmawords;Redundancy:weight*=&Repair e.g:Edmundson的 詞(Bonus [Luhn,1958]：能夠指示文章 實(shí)詞的個(gè)數(shù)來(lái)計(jì)算句子的權(quán)值。[V.A.Oswald]主張句子的權(quán)值應(yīng)按其所含 ANES(AutormaticNewsExtractionSystem)作所謂的目次性，這也很受歡統(tǒng)計(jì)表明：大部分科技文獻(xiàn)的標(biāo)題都能基本反映葉句，刪句留主干句的方 (如氣象預(yù)報(bào)等)。 基于理解的自動(dòng)常包含語(yǔ)法分析、語(yǔ)義分析、信息提取和生成，作者應(yīng)屬于此。分層的，理解比低層理解更為手工人員在編制時(shí)并不一定通結(jié)束語(yǔ)及其論題句，以發(fā)現(xiàn)其，再挑選句子并修飾稍加組織生成。文獻(xiàn)，不同用戶(hù)點(diǎn)和觀察角度可能不同，的結(jié)果應(yīng)當(dāng)不同。SentenceSentenceRepresenteachsentenceasafeatureComputescorebasedonPresentinorderinwhichtheyoccurinPostprocessingtomakesummarymoreEliminateredundantDeletesubordinateclauses,SentenceImportance-- [G. Sentence 冗余性消除 ? ? A Sigir95paperonsummarization A TrainablesentenceProposedalgorithmisappliedtoitsowndescription(thepaper)FeatureFixed-phraseCertainphrasesindicatesummary,e.g.“inParagraphParagraphinitial/finalmorelikelytobeThematicwordRepetitionisanindicatorofUppercasewordUppercaseoftenindicatesnamedentities.Sentencelengthcut-Summarysentenceshouldbe>5Hand-labelsentencesintrainingset(good/badsummarysentences)Trainclassifiertodistinguishgood/badsummarysentencesModelused:Na?veCanranksentencesaccordingtoscoreandshowtopntouser. EvaluationofBaseline(choosefirstnsentences):Overallperformance(42-44%)notveryHowever,thereismorethanoneClusteringbasedClustering Sentences 及MMRQuery-SpecificAgenericsummarymakesnoassumptionaboutthereader’sinterests.Query-specificsummariesarespecializedforasingleinformationneed,thequery.Summarizationismucheasierifwehaveadescriptionofwhattheuserwants.Recallfromlast -typeexcerpts–simplyshowincontextMMR le ????? -??1- i 術(shù)和IGR(InformationGainRatio)技術(shù)結(jié)合起來(lái)，稱(chēng)為MMI-MS MMRArgmax[Sim1(Di,Q)-(1-)maxSim2(Di,DjDiR\ DMMI-MSArgmax[Imp(Si)-(1-)maxSimj(Si,SjSiSS\ IGR_sum(w,D) CCset(D

gain_r(w,C(alMarginalRelevance 關(guān)方法:MMR-SS(SemanticSimilaritybased alMarginal time-sequence(D

mintime p x2D

*[Pjw

Q(A,B)

i 2x2y2 i

；TypesofMDSingle trackedoveralongtimeElizabethTaylor’sboutwith GiveextraweighttoMayneedto eMultipleeventsofasimilarMarathonrunnersandMorebroadbrush,ignoreAnissuewithrelatedGunIdentifykeyconceptsandselects