3 : 21 [6] [7] (DiscoursePar- sing) RST-DT(RhetoricalStruc- turetheory-discoursetreebank RST-DT) (Term Frequency-InverseDocumentFrequency) PDTB (Penn

Size: px

Start display at page:

Download "3 : 21 [6] [7] (DiscoursePar- sing) RST-DT(RhetoricalStruc- turetheory-discoursetreebank RST-DT) (Term Frequency-InverseDocumentFrequency) PDTB (Penn"

秩费莫
5 years ago
Views:

1 JOURNAL OFCHINESEINFORMATION PROCESSING Vol.27No.3 May2013 : (2013) ( ; ) : 篇章作为词和句子之后的一种文本分析粒度在自然语言理解和自然语言生成中起到至关重要的作用该文从计算语言学角度出发对中英文篇章分析技术的研究现状进行了综述介绍了中英文篇章分析技术在自然语言处理中的应用并分别从篇章理论篇章语料库及评测篇章分析器的自动构建等方面详细阐述了中英文篇章分析技术最后归纳出篇章分析技术后续研究的几个方向 : 篇章 ; 篇章分析 ; 语料库 ; 评测 :TP391 :A SurveyofDiscourseAnalysisMethods 1 : XU FanZHU QiaomingZHOU Guodong (SchoolofComputerScience & TechnologySoochow UniversitySuzhouJiangsu215006China; NaturalLanguageProcessingLabSoochow UniversitySuzhouJiangsu215006China) Abstract:Discourseakindoftextanalysisgranularitybeyondwordandsentenceplaysacrucialroleinnaturallan- guageunderstandingandgeneration.thispapersurveysthestate-of-the-artresearchesinchineseandenglishdis- courseanalysisundertheperspectiveofcomputationallinguisticsincludingtheapplicationsofchineseandenglish discourseanalysistheprocessofconstructingafulchineseandenglishdiscourseparseraccordingtodiferentdis- coursetheoriesdiscoursecorpusandevaluationaswelasalgorithmsanddetailedimplementation.alsothispa- peroutlinesseveraldirectionsforfurtherresearchesondiscourseanalysis. Keywords:discourse;discourseanalysis;corpus;evaluation (DiscourseAnalysis) ( (Centering) [1] ) (RhetoricalStructureTheory RST) [2] (DiscourseLexi- calizedtree AdjoiningGrammar D-LTAG) [3] [4-5] : : : ( ); (BK ); (11KIJ520003); (CXZZ ) : (1979 ) ; (1963 ) Web ; (1967 )

2 3 : 21 [6] [7] (DiscoursePar- sing) RST-DT(RhetoricalStruc- turetheory-discoursetreebank RST-DT) (Term Frequency-InverseDocumentFrequency) PDTB (Penn Discourse TreeBank PDTB) [13-15] [13] [14-15] NLP(NaturalLanguageProcessing) [12] SMT (TextSummarization TS) TF-IDF (QuestionAnswering QA) 10 ACL EMNLP COLING : ; [16-20] [16] [17] [18] Why 2 ;3 ;4 PDTB RST-DT [19] ;5 SchemaProcess [20] 2 (InformationExtraction IE) NLP 1 [21-22] (Statistical Machine Transla- tion SMT) [23-24] [8-12] [8-9] [10-11] 1

3 (InformationRetrieval IR) : ; IR 3 [25] [26] D-LTAG RST PDTB [31] RST-DT [32] (N-gram ) (K-means 3.1 ) [23] BOW(Bag-Of- Words) [24] (Computer-Assisted Assess D-LTAG ment CAA) D-LTAG (Lexi- calizedtreeadjunctgrammars LTAG) NLP [27-28] [27] (Predicate-Argument) (Syntactic-Semantic) [33] [28]CAA D-LTAG LTAG N :(1)LTAG (SentimentAnalysis SA) ( 1 ) D-LTAG D c * subconj conn [29] 1(a) ;(2) LTAG 1(b) so (Authorship Atribution 1(c) so AA) 1(d) [30] 1 D-LTAG D-LTAG (1)D-LTAG

4 3 : 23 [34] D-LTAG T1 1 D-LTAG 1.a LTAG T2 1.b LT- AG γ However & (1.a)Johnwenttothezoo. (1.b)Howeverhetookhiscelphonewithhim. 1 However John wenttothezoo hetookhiscelphonewith him 2 :γ T1 T2 & & T2 τ1 T1 τ2 T2 2 1D-LTAG RST RST HongKonglayinthereportofacommiteeap- pointedbythegovernorin November1980tore- [35] 1 RST (Elemental Dis- viewthescopeofpost-secondaryandtechnicaledu- courseunits EDUs) cationtohongkong.(2.b1)initsreportofjune (Span) 1981itrecommended (2.b2)thatthereshouldbea generalandsubstantialincreaseinthenumberof :CircumstanceElab- placesfortertiary education; (2.b3)one ofthe orationbackgroundenablementand Motivation measuresforachievingthiswastheestablishment OtherRelations RST ofasecondpolytechnic. :(1) : 2 4 (2.a2.b12.b22.b3) ;(2) : 2.a2.b1-b3 2.a ; 2.b1-b3 ;2.b1 2.b2-b3 (3) : RST 2.b2-b3 2.b1 ;2.b22.b3 2.b2 2.b3 RST 3 5 :(1) ;(2) ;(3) ;(4) : ;(5) : RST EDUs 3 RST [36] PDTB RST 2 1 RST (2.a)TheGenesisoftheCityPolytechnicof 2 RST htp:// downloads. html(rstbibliographies)

5 LDC(Lin guisticdataconsortium) PDTB AltLex :(1) ;(2)Alter- nativelexicalization(altlex): ; RST-DT (3)Entity-basedCoherenceRelation(EntRel): LDC 2002 RST-Tool ;(4)No Rela- ( ) tion(norel): PDTB Alt- Lex : 1 Temporal Contingency ComparisonExpansion 4 1 PDTBRST-DT PDTB RST-DT PDTB RST-DT PDTB RST-DT (1) (2) WSJ D-LTAG #Docs=385 #Corse-relations=18 #Fine-relations=78 #EDUs=21789 (1) ( (1) );(2) (Argu- EDUs; ments);(3) Alt- #Docs=2159 (2) RST Lex ( #Relations= EDUs ); (4) AltLex EntRelNoRel ;(5) :#Docs ;#Relations AltLex EntRelNoRel ;#Corse-rela- tions ;#Fine-relations ;#EDUs EDUs 3.3 Accuracy= TruePositive+TrueNegative (1) Al TruePositive ( ;TrueNegative ) ;Al F1 (Precision) (Recal) (2) F1= 2 Precision Recal Precision+Recal (2) Precision= TruePositive TruePositive+FalsePositive ( 3) TruePositive Recal= TruePositive+FalseNegative ( 4) F1 (1) FalsePositive ( );FalseNegative 4 1 PDTBRST-DT htp://

6 3 : 25 lierwhosaidshegavethem similarhelp.(con- 4.1 PDTB (Argument PDTB 1Argument2) Ar- gument1 Argument2 Ar- PDTB gument1argument2 3 4 [31] [37] Argument1 Argument2 Argument2 1 WSJ Argument1 3 Inadditionits machinesaretypicalyeasier tooperatesocustomersrequirelessassistancefrom software.(contingency:cause:result)(wsj 1887) 4 MrsYearginislying.Implicit=BECAUSE Theyfoundstudentsinanadvancedclassayearear- TINGENCY:PragmaticCause:justification)(WSJ 0044) [37] PDTB Argument : (Headword) (1) ;(2) (Argument) ; (3) ;(4) AltLex Argument1 69.8%Accuracy PDTB Argument Argument Argument (Argument) AltLex [38] AltLex Selfcategory Parentcategory ( 96.26%Accuracy94.19%F1 ) [39] 93% Accuracy [37] [38] AltLex [41] AltLex (PartofSpeech POS) PDTB 97.34%Accuracy95.76%F % Accuracy 93.62%F (Argument) Argument 97.95%F %F1 ; Argument Argument 86.24%53.85%F %40.37% F1 [40]Argument Argument Argument 1 PDTB Argument1 Argu- ment2

7 [37] [37] AltLex 39.63%F % F1 Atribute AltLex 79.68% [42-46] [44] tribute [47] [42] 4.2 RST-DT 40% Accuracy PDTB 44.58%Ac EDUs curacy [47] [48] 84% F1 [49] 21.3%Accuracy [42-47] [42-47] 57.34%42.59%F1 At- [43-44] PDTB RST- DT RST-DT PDTB 49.95%Accura- cy35.10%f1 [45] PDTB ;(2) 40.2% Accuracy [46] [46]EntRelNoRel [50] ;(2) [42-47] PDTB [51] PDTB ( :Section2-21; POS :Section22; :Section23) 94%F1 [434446] Section Section0-1 Section21-22; [42] Section2-22 Section23-24; [4547] Section23 Section %F1 PDTB RST :(1)EDUs PDTB EDUs p(b wt)(w t b { :(1) }) 84.7% F1 86.9% F1 [47] RST-DT : 18.9% F [49] PDTB AltLex : [50]

8 3 : ;(2) EDUs ( ; ) ;(3) [52] 48.1% F1 [49] [51-52]: [49] [5052] [58] 5 [59] 5.1 [53-55] : ; ( ) [56]5.1.2 [7] RST ; ; RST :(1) [57] RST ( ) ( );(2) : [6] ; ; [59] :(1)

9 [62] [60] [63] ; ; [61] : 2 2 (1) ; (2) : : [64] ; RST ; PDTB RST [67] [65] [66] 56.8% [68] 100 { 5.2 } :

10 3 : RST 6 [69] RST RST-Tool (Kappa 0.638) RST : (1) PDTB PDTB RST-DT [70] : : PDTB 40% [71] SinicaTreebank3.1 : Subordi natingcoordinating Argument 82% : Argument DiscourseadverbialArgument ; ; AltLex EntRelNoRel Kappa Kappa ; Kappa : 5.3 [71] RST 88.28% Accuracy WordNet ( 63.69%F1 (2) ) BOW 1 htp://ling.ccnu.edu.cn:8089/jiansuo/testfuju.jsp

11 ; (SemanticRoleLabeling SRL) SRL Argument ; [1] GroszB JJoshi A K Weinstein S.Centering:A Frameworkfor ModelingtheLocalCoherenceofDis- course[j].computationallinguistics199521(2): [2] Mann W CThompson S A.Rhetorical Structure (3) Theory:Towardafunctionaltheoryoftextorganiza- tion[j].text19888(3): [3] Webber B.D-LTAG:extendinglexicalized TAG to discourse[j].cognitivescience200428(5): [4] JerryR H.Onthecoherenceandstructureofdis- course[r].usa:stanfordca1985. [5] WolfFGibsonE.Representingdiscoursecoherence: acorpus-basedanalysis[c]//proceedingsofthe20th International Conference on Computational Linguis- tics.morristown:associationforcomputationallin- guistics2004: [6]. [M]. : 2000: [7]. [M]. : 2001: ( [8] MeyerT.Disambiguating Temporal-Contrastive Dis- courseconnectivesformachinetranslation[c]//pro- ) ; ceedingsofthe49thannualmeetingoftheassociation for Computational Linguistics: Human Language Technologies.Morristown:AssociationforComputa- tionallinguiscs2011: [9] MeyerTBelisA P.MultilingualAnnotationandDis- ambiguation of Discourse Connectives for Machine ; Translation[C]//Proceedingsofthe10th AnnualSig- NLP dialmeetingon Discourseand Dialogue.Morristown: AssociationforComputationalLinguiscs2011: [10] NagardR LKoehnP.AidingPronounTranslation with Co-Reference Resolution[C]//Proceedings of WorkshoponSMTand MetricsMATR.Morristown: (4) AssociationforComputationalLinguiscs2010: [11] HaeneltK.TowardsaQualityImprovementin Ma- chinetranslation:modelingdiscoursestructureand IncludingDiscourseDevelopmentintheDetermination oftranslation Equivalents[C]//Proceedingsofthe 4th International Conference on Theoretical and MethodologicalIssuesin Machine Translation.Mor-

12 3 : 31 ristown: Association for Computational Linguiscs 1992: [12] MitkovR.Howcouldrhetoricalrelationsbeusedin cessingand Management200339(6): [27] Mohler MBunescu R Mihalcea R.Learning to machinetranslation (and atleasttwo open ques- tions)[c]//proceedingsofacl WorkshoponIn- tentionality and Structure in Discourse Relations. Morristown: Association for Computational Lin- guiscs1993: [13]. 2011: [J] (4): [14]. Datasetand MethodforAutomaticalyGradingESOL [J] (2):1-7. Texts[C]//Proceedingsofthe49th AnnualMeeting [15]. ofthe Association for Computational Linguistics. [J] (3): Morristown:AssociationforComputationalLinguis- [16] ChaiJJing R. Discourse Structure for Context tionandretrievalalgorithms[j].information Pro- GradeShortAnswerQuestionsusingSemanticSimi- larity Measuresand Dependency Graph Alignments [C]//Proceedingsofthe49thAnnualMeetingofthe Associationfor Computational Linguistics. Morris- town: Association for Computational Linguistics [28] Yannakoudakis HBriscoe TMedlock B.A New tics2011: [29] SomasundaranSNamataGWiebeJetal.Super- QuestionAnswering[C]//Proceedingsofthe Work- shopon Pragmaticsof Question Answering atthe Human Language Technology Conference of the NorthAmericanChapteroftheAssociationforCom- putationallinguistics.morristown: Associationfor ComputationalLinguistics2004: [17] Sun MChaiJY.Discourseprocessingforcontext questionansweringbasedonlinguisticknowledge[j]. Knowledge-basedSystems200720(6): (2): [19] [J] (4): [31] PrasadRMiltsakakiEDineshNetal.ThePenn [20]. [J]. Discourse Treebank 2.0 Annotation Manual[R] (3): [21] HutunenSVihavainenAEterP Vetal.Rele- vancepredictionininformationextractionusingdis- courseandlexicalfeatures[c]//proceedingsofthe 18thNordicConferenceofComputationalLinguistics. Latvia2011: [22] CimianoPReyle USaricJ.Ontology-drivendis- courseanalysisforinformationextraction[j].data & KnowledgeEngineering2005(55): [23]. adjoininggrammar[j].journaloflogiclanguage [J] (2): andinformation200112(3): [24]. [34] JoshiA KSchabesY.Tree-Adjoing Grammarand [J]. Lexicalized Grammars [R]. USA: University of (4): Pennsylvania1991. mationretrievalapproachbasedondiscoursetype tionsystem.springer.2006: [26] MoratoJLlorensJGenovaGetal.Experiments visedand Unsupervised Methodsin Employing Dis- courserelationsforimprovingopinionpolarityclas- sification[c]//proceedingsofthe2009conferenceon EmpiricalMethodsin NaturalLanguageProcessing. Morristown:AssociationforComputationalLinguis- tics2009: [30] EscalanteHJSolorioT.LocalHistogramsofChar- actern-gramsforauthorship Atribution[C]//Pro- USA:UniversityofPennsylvania2008. [18]. ceedingsofthe49th AnnualMeetingoftheAssocia- Why [J] tionforcomputationallinguistics.morristown:as- sociationforcomputationallinguistics2011:288- [32] CarlsonLMarcu DOkurowski M E.Buildinga Discourse-TaggedCorpusintheFrameworkofRhe- toricalstructuretheory[c]//proceedingsofthean- nual Sigdial Meeting on Discourse and Dialogue Morristown: Association for Computational Lin- guiscs2001: [33] ForbesKMiltsakakiEPrasadRetal.D-LTAG System:DiscourseParsing withalexicalized Tree- [35] Taboada MMann W C.ApplicationsofRhetorical StructureTheory[J].DiscourseStudies20068(4): [36] () (). [M]. : 2002: [25] WangD YLukR W PWongK Fetal.AnInfor- [C]//Proceedingsofthe11thInternationalConfer- enceonapplicationsofnaturallanguagetoinforma- indiscourseanalysisimpactoninformationclassifica- [37] LinZHNgH TKan M Y.A PDTB-styledend-to- enddiscourseparser[r].singapore:nationaluni-

13 versityofsingapore2010. [38] PitlerENenkovaA.UsingSyntaxtoDisambiguate ExplicitDiscourseConnectivesinText[C]//Proceed- ingsofthejointconferenceofthe47thannualmeet- ingoftheaclandthe4thinternationaljointconfer- enceonnaturallanguageprocessingoftheafnlp. Morristown:AssociationforComputationalLinguis- tics2009: [39] PitlerERaghupathy MMehta Hetal.EasilyI- dentifiable Discourse Relations[C]//Proceedingsof the22ndinternationalconferenceon Computational Linguistics.Morristown:Associationfor Computa- tionallinguistics2008: [40] WelnerBPustejovskyJ.AutomaticalyIdentifying theargumentsof Discourse Connectives[C]//Pro- ceedingsofthe2007jointconferenceon Empirical MethodsinNaturalLanguageProcessingandCompu- tationalnaturallanguagelearning.morristown:as- sociationfor Computational Linguistics2007: [41] PrasadRJoshiAWebberB.RealizationofDis- courserelationsby Other Means:AlternativeLexi- calizations[c]//proceedingsofthe23rdinternational Conference on Computational Linguistics. Morris- town: Association for Computational Linguistics 2010: [42] Wang WTSuJTanC L.KernelBasedDiscourse RelationRecognition with TemporalOrderingInfor- mation[c]//proceedingsofthe49thannualmeeting ofthe Association for Computational Linguistics. Morristown:AssociationforComputationalLinguis- tics2010: [43] ZhouZMXuYNiuZYetal.PredictingDiscourse ConnectivesforImplicitDiscourseRelation Recogni- tion[c]//proceedingsofthe23rdinternationalcon- ferenceon ComputationalLinguistics. Morristown: Association for Computational Linguistics 2010: [44] ZhouZMLan MNiuZYetal.TheEfectsof Discourse Connectives Prediction on Implicit Dis- logue.morristown: Association for Computational Linguiscs2010: [45] LinZHKan M YNg H T.RecognizingImplicit DiscourseRelationsinthePenn DiscourseTreebank town: Association for Computational Linguistics 2009: [46] PitlerELouis ANenkova A.Automatic Sense PredictionforImplicit Discourse Relationsin Text courserelation Recognition[C]//Proceedingsofthe 9th AnnualSigdial Meeting on Discourseand Dia- [C]//Proceedingsofthe2009ConferenceonEmpiri- calmethodsinnaturallanguageprocessing.morris- [C]//ProceedingsoftheJointConferenceofthe47th AnnualMeetingoftheACLandthe4thInternational JointConferenceon NaturalLanguageProcessingof theafnlp.morristown:associationforcomputa- tionallinguistics2009: [47] Hernault HBolegala DIshizuka M. A Semi- SupervisedApproachtoImproveClassificationofIn- frequent Discourse Relations using Feature Vector Extension[C]//Proceedingsofthe2010 Conference onempiricalmethodsin NaturalLanguageProcess- ing.morristown:associationforcomputationallin- guistics2010: [48] TofiloskiMBrookeJTaboada M.ASyntacticand Lexical-BasedDiscourseSegmenter[C]//Proceedings ofthejointconferenceofthe47thannualmeetingof theaclandthe4thinternationaljointconference on Natural Language Processing of the AFNLP. Morristown:AssociationforComputationalLinguis- tics2009: [49] SoricutRMarchD.SentenceLevelDiscoursePar- sing UsingSyntacticand LexicalInformation[C]// Proceedings of the Human Language Technology ConferenceoftheNorthAmericanChapteroftheAs- sociationforcomputationallinguistics.morristown: Association for Computational Linguistics 2003: tational Linguistics.Morristown: Association for ComputationalLinguistics2004: [50] LeThanh HAbeysinghe GHuyck C.Generating DiscourseStructuresforWritenTexts[C]//Proceed- ingsofthe20thinternationalconferenceoncompu- [51] HernaultHBolegalaDIshizuka M.A Sequential ModelforDiscourseSegmentation[C]//Proceedings ofthe11th International Conference oninteligent TextProcessingandComputationalLinguistics.Mor- ristown:associationforcomputationallinguistics 2010: serbasedonsupport Vector MachineClassification [52] DuVerleD APrendingerH.A NovelDiscoursePar- [C]//ProceedingsoftheJointConferenceofthe47th AnnualMeetingoftheACLandthe4thInternational JointConferenceon NaturalLanguageProcessingof theafnlp.morristown:associationforcomputa- tionallinguistics2009: [53]. [J] : [54]. [J] : ( 55)

14 3 : 55 andoptimization ofconfusion networkfor LVCSR [J] [11] L ManguE BrilA Stolcke.Findingconsensusin speechrecognition:worderrorminimizationandoth- erapplicationsofconfusionnetworks[j].incomput- erspeechandlanguage200014(4): [12] JXueYZhao.ImprovedConfusionNetworkAlgo- rithm and ShortestPath Searchfrom Word Latice [J].ICASSP : : [16] BCJMooreBR Glasberg ArevisionofZwicker s loudness model [J].Acustica Acta. Acustica : [17] M Slaney Auditory Toolbox Version2 Interval Research Corporation TechnicalReport1998no [18]. [J]. ( )200444(1): mentposterior probability for automatic Mandarin pronunciationevaluation[j].icassp : [14]. [J]. [20]. [J] (4): (2): [15] P D PatersonK RobinsonJ Holdsworthetal. [13] JZhengC HuangM Chuetal.Generalizedseg- Complexsoundsandauditoryimages [C].inAudi- toryandperception.oxforduk:y CazalsL De- manyand K Horner (Eds)Pergamon Press [19] JLiFZhengW H Wu.Context-independentChi- neseinitial-finalacoustic modeling[j].iscslp 00 Oct :23-26Beijing. [21]. [D]. : [22]. HMM [J] (6): 櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚櫚 ( 32) [55]. [J] ( )200922(3): [56]. [J] (4): ( )20084: [57] WebsterJJ. [J] : [58]. [M]. : 1984: [59]. 1 [M]. : 1962: [60]. [M]. : 1956: [61]. [M]. : 1985: [62]. [M]. : 1993: [70] Xue Nianwen.Annotating Discourse Connectivesin thechinese Treebank[C]//Proceedings of Corpu- [63]. [M]. : RelationRecognition[C]//Proceedingsofthe5thIn- 2007: [65]. [J]. [66]. [D]. : [67]. [J] (1): [68]. [J] (3):1-8. [69]. [J] (4): sanno.morristown:associationfor Computational Linguistics2005: [71] Hen-Hsen HuangHsin-HisChen.ChineseDiscourse ternationaljointconferenceonnaturallanguage. [64]. [72] Processing.Asian Federation of Natural Language [J] (11):1527- Processing.2011:

ognitiongedr) CDC JHU ACE2005 CDC ACE2005 IBM CDC ACE2005 CDC 1 ACE [35-8] CDC CDC [ ] NIST ACE2008 GEDR CDC 1 1 CDC John Smiths [

ognitiongedr) CDC JHU ACE2005 CDC ACE2005 IBM CDC ACE2005 CDC 1 ACE [35-8] CDC CDC [ ] NIST ACE2008 GEDR CDC 1 1 CDC John Smiths [ 29 1 2015 1 JOURNAL OFCHINESEINFORMATION PROCESSING Vol.29No.1 Jan.2015 1003-0077(2015)01-0057-10 (1. 215006) (2. 215006) (CrossDocumentCoreferenceCDC) CDC ACE2005 ACE CDC TP391 A ConstructionofInformationExtraction-orientatedChinese