Quick viewing(Text Mode)

Automatic Semantic Role Labeling

Automatic Semantic Role Labeling

HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

AutomaticSemanticRoleLabeling

ScottWentauYihKristinaToutanova MicrosoftResearch

1

NaturalLanguageUnderstanding QuestionAnswering

WHOM WHAT WHO WHEN

Kristina hit Scott with a baseball yesterday

 Who hitScottwithabaseball?  Whom didKristinahitwithabaseball?  What didKristinahitScottwith?  When didKristinahitScottwithabaseball?

2

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SyntacticAnalysis(1/2)

3

SyntacticAnalysis(2/2)

4

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SyntacticVariations

Yesterday , Kristina hit Scott withabaseball Scott washitby Kristina yesterday withabaseball Yesterday , Scott washit withabaseball by Kristina Withabaseball , Kristina hit Scott yesterday Yesterday Scott washitby Kristina withabaseball Kristina hit Scott withabaseball yesterday

Agent, hitter Thing hit Instrument Temporal adjunct

5

SemanticRoleLabeling– GivingSemanticLabelstoPhrases

 [AGENT John] broke [THEME thewindow]

 [THEME Thewindow] broke

 [AGENT Sotheby’s] .. offered [RECIPIENT theDorrance heirs] [THEME amoneybackguarantee]

 [AGENT Sotheby’s] offered [THEME amoneybackguarantee] to [RECIPIENT theDorrance heirs]

 [THEME amoneybackguarantee] offered by [AGENT Sotheby’s]

 [RECIPIENT theDorrance heirs] will[ARMNEG not] be offered [THEME amoneybackguarantee]

6

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

WhyisSRLImportant– Applications

 QuestionAnswering  Q:WhenwasNapoleondefeated?

 Lookfor: [PATIENT Napoleon] [PRED defeat-synset ] [ARGMTMP *ANS*]

 MachineTranslation English(SVO) Farsi(SOV) [AGENT Thelittleboy] [AGENT pesar koocholo] boylittle [PRED kicked ] [THEME toop germezi] ballred [THEME theredball] [ARGMMNR moqtam] hardadverb [ARGMMNR hard] [PRED zaad-e]hitpast

 DocumentSummarization  PredicatesandHeadsofRolessummarizecontent

 InformationExtraction  SRLcanbeusedtoconstructusefulrulesforIE 7

QuickOverview

 PartI.Introduction  WhatisSemanticRoleLabeling?  Frommanuallycreatedgrammarstostatisticalapproaches  EarlyWork  Corpora– FrameNet,PropBank,ChinesePropBank,NomBank  TherelationbetweenSemanticRoleLabelingandothertasks  PartII.GeneraloverviewofSRLsystems  Systemarchitectures  Machinelearningmodels  PartIII.CoNLL05sharedtaskonSRL  Detailsoftopsystemsandinterestingsystems  Analysisoftheresults  ResearchdirectionsonimprovingSRLsystems  PartIV.ApplicationsofSRL

8

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

MovingtowardStatisticalApproaches

 Earlywork [Hirst 87][Dolan,Richardson,Vanderwende,93&98]

 Availablecorpora  FrameNet [Fillmoreetal.01] Main Focus  http://framenet.icsi.berkeley.edu  PropBank [Palmeretal.05]  http://www.cis.upenn.edu/~mpalmer/project_pages/ACE.htm  Corporaindevelopment  ChinesePropBank  http://www.cis.upenn.edu/~chinese/cpb/  NomBank  http://nlp.cs.nyu.edu/meyers/NomBank.html

9

EarlyWork [Hirst 87]

 SemanticInterpretation “Theprocessofmappinga syntacticallyanalyzedtext of naturallanguagetoa representation ofitsmeaning.”  Absity – semanticinterpreterbyHirst  Basedonmanuallycreatedsemanticrules

 Input: Nadia subj boughtthebook obj fromastoreinthemall.  Output: (a?u (buy?u (agent=(the?x(person?x (propername =“Nadia”)))) (patient=(the?y(book?y))) (source=(a?z(store?z (location=(the?w(mall?w)))))))

Exampletakenfrom[Hirst 87] 10

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

EarlyWork [Dolan,Richardson,Vanderwende,93&98]

 MindNet:  Agraphoflabeledwithsemanticrelationsautomatically acquiredfromonlinedictionariesandencyclopedias  MindNetidentifies24labeledsemanticrelationsbasedonmanually createdsemanticrules  Relationsareweightedbasedonvertexfrequency

http://research.microsoft.com/mnex 11

FrameNet [Fillmoreetal.01]

 SentencesfromtheBritishNationalCorpus(BNC)  Annotatedwith framespecific semanticroles  Variousparticipants,props,andotherconceptualroles

Frame:Hit_target Lexicalunits(LUs): (hit , pickoff , shoot ) Wordsthatevoketheframe Agent Means (usuallyverbs) Target Place Core NonCore Instrument Purpose Frameelements(FEs): Manner Subregion Theinvolvedsemanticroles Time

[Agent Kristina ] hit [Target Scott ][Instrument withabaseball ][Time yesterday ].

12

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

FrameNet – Continued

 MethodologyofconstructingFrameNet  Define/discover/describe frames  Decidetheparticipants( frameelements )  List lexicalunits thatinvoketheframe  Findexamplesentencesinthecorpus(BNC)andannotatethem  Corpora  FrameNet I– BritishNationalCorpusonly  FrameNet II– LDCNorthAmericanNewswirecorpora  Size  >8,900lexicalunits,>625frames,>135,000sentences http://framenet.icsi.berkeley.edu

13

PropositionBank(PropBank) [Palmeretal.05]

 Transfersentencestopropositions  Kristina hit Scott → hit( Kristina ,Scott )

 Penn → PropBank  AddasemanticlayeronPennTreeBank  Defineasetofsemanticrolesforeachverb  Eachverb’srolesarenumbered

…[A0 thecompany]to… offer [A1 a15%to20%stake][ A2 tothepublic] …[A0 Sotheby’s]… offered [A2 theDorrance heirs][ A1 amoneyback guarantee] …[A1 anamendment] offered [A0 byRep.PeterDeFazio]… …[A2 Subcontractors]willbe offered [A1 asettlement]…

14

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

PropositionBank(PropBank) DefinetheSetofSemanticRoles

 It’sdifficulttodefineageneralsetofsemantic rolesforalltypesofpredicates(verbs).  PropBankdefinessemanticrolesforeachverb andsenseintheframefiles.

 The(core)argumentsarelabeledbynumbers.  A0– Agent;A1– PatientorTheme  Otherarguments– noconsistentgeneralizations

 Adjunctlikearguments– universal toallverbs  AMLOC,TMP,EXT,CAU,DIR,PNC,ADV,MNR, NEG,MOD,DIS

15

PropositionBank(PropBank) FrameFiles

 hit.01“strike”  A0:agent,hitter;A1:thinghit; A2:instrument,thinghitbyorwith AMTMP Time [A0 Kristina ] hit [A1 Scott ][A2 withabaseball ] yesterday .

 look.02“seeming”  A0:seemer;A1:seemedlike;A2:seemedto

[A0 It ] looked [A2 toher ]like [A1 hedeservedthis ].  deserve.01“deserve” Proposition:  A0:deservingentity;A1:thingdeserved; Asentenceand A2:inexchangefor atargetverb

Itlookedtoherlike [A0 he ] deserved [A1 this ].

16

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

PropositionBank(PropBank) AddaSemanticLayer

A0

A1 A2 AMTMP

[A0 Kristina ] hit [A1 Scott ][A2 withabaseball ][AM-TMP yesterday ].

17

PropositionBank(PropBank) AddaSemanticLayer– Continued

A1 CA1

A0

[A1 Theworstthingabouthim ] said [A0 Kristina ][C-A1 ishislaziness ].

18

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

PropositionBank(PropBank) FinalNotes

 Currentrelease (Mar4,2005) :PropositionBankI  VerbLexicon:3,324framefiles  Annotation:~113,000propositions http://www.cis.upenn.edu/~mpalmer/project_pages/ACE.htm

 Alternativeformat:CoNLL04,05sharedtask  Representedintableformat  Hasbeenusedasstandarddatasetfortheshared tasksonsemanticrolelabeling http://www.lsi.upc.es/~srlconll/soft.html

19

CorporainDevelopment

 ChinesePropBank http://www.cis.upenn.edu/~chinese/cpb/  SimilartoPropBank,itaddsasemanticlayeronPenn ChineseTreebank  Aprereleaseversionhas250Kwordsand10,364 sentences;~55%

 NomBank http://nlp.cs.nyu.edu/meyers/NomBank.html  LabelargumentsthatcooccurwithnounsinPropBank

[A0 Her ][ REL gift ] of [A1 abook ][ A2 toJohn ]

 CurrentRelease:Sep.2005  93,809instancesofnouns;2,805differentwords;~80%  Highfrequency(>600)nounshavebeencompleted

20

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

QuickOverview

 PartI.Introduction  WhatisSemanticRoleLabeling?  Frommanuallycreatedgrammarstostatisticalapproaches  EarlyWork  Corpora– FrameNet,PropBank,ChinesePropBank,NomBank  TherelationbetweenSemanticRoleLabelingandothertasks  PartII.GeneraloverviewofSRLsystems  Systemarchitectures  Machinelearningmodels  PartIII.CoNLL05sharedtaskonSRL  Detailsoftopsystemsandinterestingsystems  Analysisoftheresults  ResearchdirectionsonimprovingSRLsystems  PartIV.ApplicationsofSRL

21

RelationtoOtherTasks

 Informationextraction

 Semanticforspeechdialogues

 Deepsemanticparsing

 PennTreebankfunctiontagging

 Predictingcasemarkers

 Aspectsofcomparisons Coverage Depth of semantics Direct application SRL Broad Shallow No

22

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

RelatedTask:InformationExtraction

 Example(HUBEvent99evaluations, [Hirschmanetal.99])  Asetofdomaindependent templettes ,summarizinginformation abouteventsfrommultiplesentences := INSTRUMENT London[gold] AMOUNT_CHANGE fell[$4.70]cents CURRENT_VALUE $308.45 DATE: daily

Timeforour daily marketreportfromNASDAQ. Londongold fell$4.70cents to $308.45.

 Manyothertaskspecifications:extractinginformationabout products,relationsamongproteins,authorsofbooks,etc.

23

InformationExtractionversus SemanticRoleLabeling

Characteristic IE SRL Coverage narrow broad

Depthofsemantics shallow shallow

Directlyconnectedto sometimes no application

 Approachestotask:diverse  Dependsontheparticulartaskandamountofavailabledata  HandwrittensyntacticsemanticgrammarscompiledintoFSA  Sequencelabelingapproaches(HMM,CRF,CMM)  Surveymaterials: http://scottyih.org/IEsurvey3.htm [Appelt &Israel99],[Muslea99] 24

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

RelatedTask:SpeechDialogs

 SpokenLanguageUnderstanding: extractthesemanticsfrom anutterance  Mustdealwithuncertainlyanddisfluenciesinspeechinput  Example:tasksetupinanarrowflightreservationsdomain (ATISevaluations, [Price90] )

Seattle Boston

Sentence: “ShowmeallflightsfromSeattletoBoston”

25

ATISParsingversus SemanticRoleLabeling

Characteristic ATIS SRL Coverage narrow broad

Depthofsemantics deeper shallow

Directlyconnectedto yes no application

 ApproachestoATISparsing(overviewin [Wangetal.05] ):  Simultaneoussyntactic/semanticparsing[Milleretal.96],knowledge basedapproach[Ward94,Dowdingetal.93]  Currentbest:smallsemanticgrammarandasequencelabelingmodel (nofullsyntacticparsinginformation) Error 3.8% ([Wangetal.06]).

26

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

RelatedTask:SemanticParsingfor NLInterfacestoDatabases

 Example:GeoQuery Domain(adomainoffactsforUS geography) [Zelle &Mooney96] Sentence: HowmanycitiesarethereintheUS? Meaning Representation: answer(count(city(loc_2(countryid(usa)))))

 Characteristics:  Arestricteddomainforwhichwehaveacompletedomainmodel  Sentencesareusuallyshortbutcouldbeungrammatical  Syntaxoftargetrepresentationismorecomplexcomparedtothe ATIStask  Needtorepresentquantifiers(thelargest,themostpopulated,etc.)

27

SemanticParsingforNLInterfacesto DatabasesversusSemanticRoleLabeling

Characteristic NL interfaces to DB SRL Coverage narrow broad Depthofsemantics deep shallow Directlyconnectedto yes no application

 Approaches  Handbuiltgrammars [Androutsopoulos etal.05] (overview)  Machinelearningofsymbolicgrammars– e.g. [Zelle &Mooney96]  Learnedstatisticalsyntactic/semanticgrammar [Ge &Mooney05] (supervised); [Zettlemoyer &Collins05],[Wong&Mooney06] (unsupervised) 28

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

RelatedTask:DeepParsing

 Handbuiltbroadcoveragegrammarscreatesimultaneous syntacticandsemanticanalyses  TheCoreLanguageEngine[Alshawi92]  LexicalFunctionalGrammarLFG([Bresnan01], [Maxwell&Kaplan93])  HeadDrivenPhraseStructureGrammar([Pollard&Sag94], [Copestake&Flickinger00])  Modelmorecomplexphenomena  Quantifiers,quantifierscope,notjustverbsemantics,anaphora, aspect,tense  Asetofanalysesispossibleforeachsentenceaccording tothegrammar:needtodisambiguate  Untilrecently:nopubliclyavailabledatasetsor specificationsforsemantics  Difficulttocreateandexpand

29

DeepParsingversus SemanticRoleLabeling

Characteristic Deep Parsing SRL

Coverage broad broad Depthofsemantics deep shallow Directlyconnectedto no no application

 Approach  Handbuildgrammar(possiblyexpandautomatically)  Treatedasaparsingproblem(jointsyntacticandsemantic disambiguation)  ForLFG([Riezler etal.02])  ForHPSG([Toutanovaetal.04],[Miyao&Tsujii05]) 30

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

RelatedTask:PredictionofFunctionTags [Blaheta&Charniak 00]

ThePennTreebankcontainsannotationoffunctiontagsfor somephrases: subject,logicalsubject,adjuncts(temporal, locative,etc.)

SlidefromDonBlaheta03thesisdefense 31

PredictionofFunctionTagsversus SemanticRoleLabeling

Characteristic Predicting SRL Function Tags Coverage broad broad Depthofsemantics shallower shallow Directlyconnected no no toapplication

 Approach:aclassifierbasedonvotedperceptionsand otherMLtechniques  UsingrichsyntacticinformationfromPennTreebankparsetrees  GrammaticaltagsF196.4,othertagsF183.8 [Blaheta03]

32

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

RelatedTask:PredictingCaseMarkers

 Somelanguageshavecasemarkers  Theyindicatethesyntacticosemantic relationbetweenaphraseand thephraseitmodifies  NeededforMachineTranslation, foreignlanguagelearning  InJapanese,casemarkersindicatee.g subject, object,location.  Moresimilartofunctiontagsthantosemanticrolelabels  Goodnews:noannotateddataisrequired!  Thecasemarkersarepartofthesurfacestring

33

PredictingCaseMarkersversus SemanticRoleLabeling Characteristic Predicting Case SRL Markers Coverage broad broad Depthofsemantics shallower shallow Directlyconnectedto yes no application

 Approaches  Usingcontentwordsfromthetargetlanguageonlyplus dependencyinformation [Uchimoto etal.02]  Usingsyntacticandfeaturesfromthesourceand targetlanguages [Suzuki&Toutanova06] ;percasemarkererror usingautomaticparses:8.4% 34

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SummaryofPartI– Introduction

 WhatisSemanticRoleLabeling?  CorporaforSemanticRoleLabeling  WewilldiscussmainlyPropBank.  RelatedtaskstoSRL  Informationextraction  Deepsemanticparsing  PennTreebankfunctiontagging  Predictingcasemarkers

 Next part: overview of SRL systems

35

QuickOverview

 PartI.Introduction  WhatisSemanticRoleLabeling?  Frommanuallycreatedgrammarstostatisticalapproaches  EarlyWork  Corpora– FrameNet,PropBank,ChinesePropBank,NomBank  TherelationbetweenSemanticRoleLabelingandothertasks  PartII.GeneraloverviewofSRLsystems  Systemarchitectures  Machinelearningmodels  PartIII.CoNLL05sharedtaskonSRL  Detailsoftopsystemsandinterestingsystems  Analysisoftheresults  ResearchdirectionsonimprovingSRLsystems  PartIV.ApplicationsofSRL

36

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

PartII:OverviewofSRLSystems

 DefinitionoftheSRLtask  Evaluationmeasures  Generalsystemarchitectures  Machinelearningmodels  Features&models  Performancegainsfromdifferenttechniques

37

DevelopmentofSRLSystems

 Gildea&Jurafsky2002  FirststatisticalmodelonFrameNet

 7+papersinmajorconferencesin2003  19+papersinmajorconferences2004,2005

 3sharedtasks  Senseval3(FrameNet)– 8teamsparticipated  CoNLL 04(PropBank)– 10teamsparticipated  CoNLL 05(PropBank)– 19teamsparticipated

38

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

TaskFormulation

 Mostgeneralformulation:determinealabelingon(usually butnotalwayscontiguous) substrings (phrases )ofthe sentence s, givena p

[A0 Thequeen] broke [A1 thewindow] .

[A1 Byworkinghard] , [A0 he ] said , [CA1 youcangetexhausted].

 Everysubstring c canberepresentedbyasetofword indices  Moreformally,asemanticrolelabelingisamappingfrom thesetofsubstringsofstothelabelset L. L includesall argumentlabelsand NONE.

39

Subtasks

 Identification:  Veryhardtask:toseparatetheargumentsubstringsfromthe restinthisexponentiallysizedset  Usuallyonly1to9(avg. 2.7 )substringshavelabelsARGand theresthaveNONEforapredicate  Classification:  Giventhesetofsubstringsthathavean ARG label,decidethe exactsemanticlabel  Coreargument semanticrolelabeling:(easier)  Labelphraseswithcoreargumentlabelsonly.Themodifier argumentsareassumedtohavelabelNONE.

40

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

EvaluationMeasures

Correct: [A0 Thequeen] broke [A1 thewindow][ AMTMP yesterday ]

Guess: [A0 Thequeen ] broke the [A1 window][ AMLOC yesterday ] Correct Guess {Thequeen}→A0 {Thequeen}→A0 {thewindow}→A1 {window}→A1 {yesterday}>AMTMP {yesterday}>AMLOC allother→ NONE allother→ NONE

 Precision,Recall,FMeasure{ tp =1, fp =2, fn =2} p=r=f=1/3  Measuresforsubtasks  Identification(Precision,Recall,Fmeasure){tp=2,fp=1,fn=1} p=r=f=2/3  Classification(Accuracy)acc=.5(labelingofcorrectlyidentifiedphrases)  Corearguments(Precision,Recall,Fmeasure){tp=1,fp=1,fn=1} p=r=f=1/2

41

PartII:OverviewofSRLSystems

 DefinitionoftheSRLtask  Evaluationmeasures  Generalsystemarchitectures  Machinelearningmodels  Features&models  Performancegainsfromdifferenttechniques

42

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Terminology:LocalandJointModels

 Local models decidethelabelofeachsubstring independentlyofthelabelsofothersubstrings  Thiscanleadtoinconsistencies  overlappingargumentstrings

By [A1 working [A1 hard ] ,he ] said ,youcanachievealot.  repeatedarguments

By [A1 working ] hard, [A1 he ] said ,youcanachievealot.  missingarguments

[A0 Byworkinghard,he ] said , [A0 youcanachievealot ].

 Joint models takeintoaccountthedependencies amonglabelsofdifferentsubstrings

43

BasicArchitectureofaGenericSRLSystem

Localscoresfor phraselabelsdonot dependonlabelsof otherphrases

Jointscorestake intoaccount dependencies amongthelabels ofmultiplephrases

44

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

AnnotationsUsed

 SyntacticParsers  Collins’,Charniak’s(mostsystems) CCGparses ([Gildea&Hockenmaier 03],[Pradhanetal.05]) TAGparses ([Chen&Rambow03])

 Shallowparsers

[NP Yesterday],[ NP Kristina][ VP hit][ NP Scott][ PP with][ NP a baseball].

 Semanticontologies(WordNet,automaticallyderived), andnamedentityclasses (v) hit (causetomovebystriking) WordNet hypernym propel, impel (causetomoveforwardwithforce )

45

AnnotationsUsed Continued

 Mostcommonly,substringsthathaveargumentlabels correspondtosyntacticconstituents  InPropbank,anargumentphrasecorrespondstoexactlyoneconstituentinthe correctparsetree for 95.7 %ofthearguments;  whenmorethanoneconstituentcorrespondtoasingleargument (4.3 %),simplerulescanjoinconstituentstogether(in80%ofthese cases,[ Toutanova05 ]);  InPropbank,anargumentphrasecorrespondstoexactlyoneparse treeconstituentin Charniak’sautomaticparsetree forapprox 90.0% ofthearguments.  Somecases(about30%ofthemismatches)areeasilyrecoverablewith simplerulesthatjoinconstituents([ Toutanova05 ])  InFrameNet,anargumentphrasecorrespondstoexactlyoneparse treeconstituentinCollins’ automaticparsetreefor 87% ofthe arguments.

46

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

LabelingParseTreeNodes

 Givenaparsetree t,label thenodes(phrases)inthe treewithsemanticlabels A0  Todealwithdiscontiguous arguments NONE  Inapostprocessingstep, joinsomephrasesusing simplerules  Useamorepowerful labelingscheme,i.e.CA0 forcontinuationofA0 Anotherapproach:labelingchunked sentences.Willnotdescribeinthissection. 47

LocalScoringModels

 Notation:aconstituentnode c,atree t,apredicate node p, featuremapforaconstituent  Targetlabels

c p

 Two(probabilistic)models  Identificationmodel

 Classificationmodel

 Sometimesonemodel

48

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

WhySplittheTaskinto IdentificationandClassification

 Differentfeaturesarehelpfulforeachtask  Syntacticfeaturesmorehelpfulforidentification,lexicalfeatures morehelpfulforclassification  Example:theidentityofthepredicate,e.g. p=“hit” ismuchmore importantforclassificationthanforidentification ([Pradhanetal.04]):  Identificationallfeatures:93.8nopredicate:93.2  Classificationallfeatures:91.0nopredicate:82.4  Somefeaturesresultinaperformancedecreaseforoneandan increasefortheothertask [Pradhanetal.04]

 Splittingthetaskincreasescomputationalefficiencyin training  Inidentification,everyparsetreeconstituentisacandidate( linear inthesizeoftheparsetree,avg.40 )  Inclassification,labelasmallnumberofcandidates( avg.2.7 )

49

CombiningIdentificationand ClassificationModels Step 1 . Pruning. Usingahand specifiedfilter.

Step 2. Identification. Identificationmodel Step 3. Classification. (filtersoutcandidates Classificationmodel withhighprobabilityof assignsoneofthe NONE) argumentlabelstoselected A0 A1 nodes(orsometimes possiblyNONE)

50

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

CombiningIdentificationand ClassificationModels – Continued

or

One Step . Simultaneously A0 identifyandclassify A1 using

51

CombiningIdentificationand ClassificationModels – Continued

 [Gildea&Jurafsky 02]  Identification + Classification forlocalscoring experiments  One Step forjointscoringexperiments  [Xue&Palmer 04]and[Punyakanoketal.04,05]  Pruning + Identification + Classification  [Pradhanetal.04]and[Toutanovaetal.05]  One Step

52

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

JointScoringModels

A0 AMTMP NONE

A1 AMTMP

 Thesemodelshavescoresforawholelabelingofatree (notjustindividuallabels)  Encodesomedependenciesamongthelabelsofdifferentnodes

53

CombiningLocalandJointScoring Models

 Tightintegrationoflocalandjointscoringinasingle probabilisticmodelandexactsearch [Cohn&Blunsom 05] [Màrquez etal.05],[Thompsonetal.03]  Whenthejointmodelmakesstrongindependenceassumptions

 Rerankingorapproximatesearchtofindthelabeling whichmaximizesacombinationoflocalandajointscore [Gildea&Jurafsky 02][Pradhanetal.04][Toutanovaetal.05]  Usuallyexponentialsearchrequiredtofindtheexactmaximizer

 Exactsearchforbestassignmentbylocalmodelsatisfying hardjointconstraints  UsingIntegerLinearProgramming [Punyakanoketal04,05](worst caseNPhard)

 Moredetailslater 54

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

PartII:OverviewofSRLSystems

 DefinitionoftheSRLtask  Evaluationmeasures  Generalsystemarchitectures  Machinelearningmodels  Features&models  ForLocalScoring  ForJointScoring  Performancegainsfromdifferenttechniques

55

Gildea&Jurafsky(2002)Features

 Keyearlywork  Futuresystemsusethese featuresasabaseline

 ConstituentIndependent  Targetpredicate(lemma)  Voice  Subcategorization  ConstituentSpecific Target broke  Path Voice active  Position( left,right ) Subcategorization VP→VBDNP  PhraseType Path VBD↑VP↑S↓NP  GoverningCategory (S or VP ) Position left  HeadWord PhraseType NP GovCat S HeadWord She 56

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

EvaluationusingCorrectand AutomaticParses

Fora correctparse ,95.7%of argumentscorrespondtoasingle constituentandtheirboundaries areeasytoconsider

Foran automaticparse (Charniak’s parser), about90%ofthearguments correspondtoasingleconstituent; theargumentsforwhichtheparser madeabracketingerroraredifficultto Wrong! get additionally,attachmenterrorsand labelingerrorsmakethetaskmuch harder

57

PerformancewithBaselineFeatures usingtheG&JModel

 algorithm: interpolationofrelative frequencyestimatesbasedonsubsetsofthe7features introducedearlier 100 90 82.0 80 69.4 Automatic 70 Parses FrameNet 59.2 60

Results 50

40 Id Class Integrated 100 90 82.8 79.2 Automatic 80 Propbank Parses 67.6 Results 70 Correct Parses 60 53.6 50 40 Class Integrated Justbychangingthelearningalgorithm67.6→ 80.8 using

SVMs [Pradhanetal.04] ), 58

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Surdeanuetal.(2003)Features

 ContentWord(differentfrom headword) Headword Contentword  HeadWordandContentWord POStags  NElabels(Organization, Location,etc.)

baseline features added features 100 89.0 Headword ContentWord 90 84.6 83.7 78.8 80

70

60

50 40 Id Class Gainsfromthenewfeaturesusingcorrectparses; 28% error

reductionforIdentificationand 23% errorreductionforClassification 59

Pradhanetal.(2004)Features

 Morestructural/lexicalcontext( 31% errorreductionfrom baselineduetothese+Surdeanuetal.features)

Lastword/POS Firstword/POS Parentconstituent PhraseType/ Leftconstituent HeadWord/POS Rightconstituent PhraseType/ PhraseType/ HeadWord/POS HeadWord/POS

60

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Pradhanetal.(2004)Results

baseline features added features added features 100 93.8 91.0 100 90.4 87.9 90.0 e 86.7

90 e r 90 86.0 80.8 r u 79.4 80 u s 80 s

a

a

e 70

e 70

m

m

- 60

- 60

f

f 50 50

40 40 Id Class Integrated Id Class Integrated Resultsoncorrectparsetrees Resultsonautomaticparsetrees BaselineresultshigherthanGildeaandJurafsky’s duetoa differentclassifier SVM These are the highest numbers on Propbank version July 2002 61

Xue&Palmer(2004)Features

 Addedexplicitfeature np_give_NP_np conjunctionsina np_give_CURR_np MaxEntmodel,e.g. predicate+phrasetype np_v_NP_np  Syntacticframefeature (helps a lot )  HeadofPPParent (helps a lot )  Iftheparentofa constituentisaPP, theidentityofthe preposition(feature goodforPropBank Feb04)

62

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Xue&Palmer(2004)Results

baseline features added features

100 93.0 88.1 88.5 90 82.9 Anewerversionof 80 Propbank– February 70 60 2004 50 40 Class Integrated AllArgumentsCorrectParse

correct parse automatic parse correct parse automatic parse

100 95.0 100 93.0 90.6 88.5 90 90 78.2 80 80 76.2

70 70 60 60

50 50 40 40 Class Integrated Class Integrated CoreArguments AllArguments

Resultsnotbetterthan[Pradhanetal.04],butcomparable. 63

MachineLearningModelsUsed

 Backofflatticebasedrelativefrequency models ([Gildea&Jurafsky 02],[Gildea&Palmer02])

 Decisiontrees ([Surdeanuetal.03])

 SupportVectorMachines ([Pradhanetal.04])

 Loglinearmodels ([Xue&Palmer 04][Toutanovaetal.05])

 SNoW ([Punyakanoketal.04,05])  AdaBoost,TBL,CRFs,…

64

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

JointScoring:EnforcingHard Constraints

 Constraint1:Argumentphrasesdonotoverlap

By [A1 working [A1 hard ] ,he ] said ,youcanachievealot.  Pradhanetal.(04)– greedysearchforabestsetofnon overlappingarguments  Toutanovaetal.(05)– exactsearchforthebestsetofnon overlappingarguments(dynamicprogramming,linearinthesize ofthetree)  Punyakanoketal.(05)– exactsearchforbestnonoverlapping argumentsusingintegerlinearprogramming  Otherconstraints ([Punyakanoketal.04,05])  norepeatedcorearguments(goodheuristic)  phrasesdonotoverlapthepredicate  (morelater )

65

GainsfromEnforcingHard Constraints

 Argumentphrasesdonotoverlap  Pradhanetal.(04)goodgainsforabaselinesystem: 80.8 → 81.6 correctparses  Toutanovaetal.(05)asmallgainfromnonoverlappingfora modelwithmanyfeatures 88.3 → 88.4 correctparses

 Otherhardconstraints(norepeatingcorearguments,set oflabeledargumentsallowable,etc.)  Punyakanoketal.(04)evaluationofthisaspectonlywhenusing chunkedsentences(notfullparsing) 87.1 → 88.1 correctparses 67.1 → 68.2 automaticparses

66

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

JointScoring:IntegratingSoft Preferences

A0 AMTMP

A1 AMTMP

 Therearemanystatisticaltendenciesforthesequence ofrolesandtheirsyntacticrealizations  Whenbotharebeforetheverb,AMTMPisusuallybeforeA0  Usually,therearen’tmultipletemporalmodifiers  Manyotherswhichcanbelearnedautomatically

67

JointScoring:IntegratingSoft Preferences

 GildeaandJurafsky(02)– asmoothedrelativefrequencyestimate oftheprobabilityofframeelementmultisets:

 Gainsrelativetolocalmodel 59.2 → 62.9 FrameNetautomaticparses

 Pradhanetal.(04)– alanguagemodelonargumentlabel sequences(withthepredicateincluded)

 Smallgainsrelativetolocalmodelforabaselinesystem 88.0 → 88.9 on corearguments PropBankcorrectparses

 Toutanovaetal.(05)– ajointmodelbasedonCRFs witharichset ofjointfeaturesofthesequenceoflabeledarguments( morelater)  GainsrelativetolocalmodelonPropBankcorrectparses 88.4 → 91.2 (24%errorreduction);gainsonautomaticparses 78.2 → 80 .0

68

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

CombiningAnnotationsand CombiningSystems

 Punyakanoketal.(05)combineinformationfromsystemstrainedon top n parsetreesproducedbyCharniak’sparserandCollins’ parser.  Effectivelyconstituentsfromalltreescanbeselectedasarguments  ConstraintsfornonoverlapandotherconstraintsareenforcedthroughILP  Gains 74.8 → 77.3 onautomaticparses(CoNLL 05devset)

 Haghighi etal.(05)combinetop n Charniakparsetrees  ThisisachievedinaBayesianway:sumovertheparsetrees approximatedbymax  Gains 79.7 → 80.3 onautomaticparses(CoNLL 05testset)  Pradhanetal.(05)combinedifferentsyntacticviews  Charniaksyntacticparse,CombinatoryCategorialGrammarparse  Gains 77.0 → 78.0 onautomaticparses(CoNLL 05devset)

 OthersystemsinCoNLL 2005

 Morelateronallofthese 69

SummaryofPartII– SystemOverview

 IntroducedSRLsystemarchitecture:  annotations,localscoring,jointscoring  Describedmajorfeatureshelpfultothetask  showedthatlargegainscanbeachievedbyimprovingthefeatures  Describedmethodsforlocalscoring,combiningi dentification and classification models  Describedmethodsforjointscoring  gainsfromincorporating hard constraints  gainsfromincorporating soft preferences  Introducedtheconceptofcombiningsystemsandannotations  significantgainspossible  Next part :moredetailsonthesystemsinCoNLL 2005

70

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Break!!

[A0 We ][ AM-MOD will ]see[ A1 you ][ AM-TMP afterthebreak ].

71

QuickOverview

 PartI.Introduction  WhatisSemanticRoleLabeling?  Frommanuallycreatedgrammarstostatisticalapproaches  EarlyWork  Corpora– FrameNet,PropBank,ChinesePropBank,NomBank  TherelationbetweenSemanticRoleLabelingandothertasks  PartII.GeneraloverviewofSRLsystems  Systemarchitectures  Machinelearningmodels  PartIII.CoNLL05sharedtaskonSRL  Detailsoftopsystemsandinterestingsystems  Analysisoftheresults  ResearchdirectionsonimprovingSRLsystems  PartIV.ApplicationsofSRL

72

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

PartIII:CoNLL05SharedTaskonSRL

 Detailsoftopsystemsandinteresting systems  Introducethetop4systems  Describe3spotlightsystems  Analysisoftheoverallresults  Generalperformance  Systemproperties  Perargumentperformance  DirectionsforimprovingSRLsystems

73

DetailsofCoNLL05Systems

 Topperformingsystems #3 Màrquezetal.(TechnicalUniversityofCatalonia) #4 Pradhanetal.(UniversityofColoradoatBoulder) #1 Punyakanoketal.(U.ofIllinoisatUrbanaChampaign) #2 Haghighi etal.(StanfordUniversity) Kristina’ssystem Scott’ssystem  Spotlightsystems  Yi&Palmer– integratingsyntacticandsemanticparsing  Cohn&Blunsorn – SRLwithTreeCRFs  Carreras– systemcombination

74

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SRLasSequentialTagging [Màrquezetal.]

 Aconceptuallysimplebutcompetitivesystem  SRListreatedasaflatsequentiallabeling problemrepresentedintheBIOformat.

 Systemarchitecture  Preprocessing(sequentialization)

 FP CHA :fullparse,basedonCharniak’sparser

 PP UPC :partialparse,basedonUPCchunker&clauser  LearningusingAdaBoost  Greedycombinationoftwosystems

75

Sequentialization – FullParse [Màrquezetal.]– Continued

 Explorethesentenceregionsdefinedbytheclauseboundaries.

 Thetopmostconstituentsintheregionsareselectedastokens.

 Equivalentto[Xue&Palmer 04]pruningprocessonfullparsetrees

S Kristina BA0 hit O NP VP Scott BA1 A0 NP PP NP with BA2 A1 A2 AM-TMP a NP baseball yesterday BAMTMP Kristina hit Scott withabaseball yesterday

76

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Sequentialization – PartialParse [Màrquezetal.]– Continued

 Onlyclausesandbasechunksareavailable.

 Chunkswithinthesameclauseareselectedastokens.

Kristina BA0

hit O A0 Scott BA1 A1 A2 with BA2 AM-TMP a IA2 A2 Baseball yesterday BAMTMP

77

GreedyCombination [Màrquezetal.]– Continued

 Jointhemaximumnumberofargumentsfrom theoutputofbothsystems  MoreimpactonRecall

 Differentperformanceondifferentlabels

 FP CHA :betterforA0andA1;PP UPC :betterforA2A4

 Combiningrule

1. AddingargumentsA0andA1fromFP CHA 2. AddingargumentsA2,A3,andA4fromPP UPC 3. RepeatStep1&2forotherarguments  Dropoverlapping/embeddingarguments

78

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Results [Màrquezetal.]– Continued

 Overallresultsondevelopmentset

F1 Prec. Rec.

PP UPC 73.57 76.86 70.55

FP CHA 75.75 78.08 73.54 Combined 76.93 78.39 75.53

 Finalresultsontestsets  WSJ23(2416sentences)

 77.97(F 1),79.55(Prec.),76.45(Rec.)  Brown(426sentences;crossdomaintest)

 67.42(F 1),70.79(Prec.),64.35(Rec.)

79

SemanticRoleChunkingCombining ComplementarySyntacticViews [Pradhanetal.]

 Observation:theperformanceofanSRLsystemdepends heavilyonthesyntactic view  Syntacticparsetreesgeneratedbyfullparsers  Charniak’s,Collins’,…  Partialsyntacticanalysisbychunker,clauser,etc.

 Usageofsyntacticinformation  Features(e.g.,path,syntacticframe,etc.)  Argumentcandidates(mostlytheconstituents)

 Strategytoreducetheimpactofincorrectsyntacticinfo.  BuildindividualSRLsystemsbasedondifferentsyntacticparsetrees (Charniak’sandCollins’)  Usethepredictionsasadditionalfeatures  BuildafinalSRLsysteminthesequentialtaggingrepresentation

80

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

ConstituentViews [Pradhanetal.]– Continued

ParseTree#1 ParseTree#2 S S

NP VP NP VP

NP

NP PP PP

NP NP

Kristina hit Scott withabaseball Kristina hit Scottwithabaseball

A0 A1 A2 A0 A1

81

ChunkView [Pradhanetal.]– Continued

 Sequentialization usingbasechunks [Hacioglu&Ward 03]  Chunker: Yamcha [Kudo&Matsumoto 01]  http://chasen.org/~taku/software/yamcha/

Chunks TrueLabel Pred #1 Pred #2 Kristina BA0 BA0 BA0 hit O O O Scott BA1 BA1 BA1 with BA2 BA2 IA1 a IA2 IA2 IA2 Baseball

82

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Algorithm [Pradhanetal.]– Continued

 GeneratefeaturesfromCharniak’sandCollins’ parsetrees  Addafewfeaturesfromonetotheother,and constructtwoSRLsystems  RepresenttheoutputassemanticBIOtags,and usethemasfeatures  Generatethefinalsemanticrolelabelsetusing aphrasebasedchunkingparadigm

83

Architecture [Pradhanetal.]– Continued

Charniak Collins Words

Phrases

BIO BIO BIO Features Chunker

BIO

Semantic Role Labels

SlidefromPradhanetal.(CoNLL 2005) 84

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Results [Pradhanetal.]– Continued

 Overallresultsondevelopmentset

System F1 Prec Rec Charniak 77 80 75 Collins 76 79 74 Combined 78 81 76

 Performance(F 1)onTestsets  Submittedsystem:WSJ2377.4,Brown67.1  Bugfixedsystem:WSJ2378.6,Brown68.4

 Software: ASSERT (Automatic Statistical SE mantic Role Tagger) http://oak.colorado.edu/assert 85

GeneralizedInference [Punyakanoketal.]

 Theoutputoftheargumentclassifieroftenviolatessome constraints,especiallywhenthesentenceislong.

 Usetheintegerlinearprogramminginferenceprocedure [Roth&Yih 04]  Input:thelocalscores(bytheargumentclassifier),and structuralandlinguisticconstraints  Output:thebestlegitimateglobalpredictions  Formulatedasanoptimizationproblemandsolvedvia IntegerLinearProgramming.  Allowsincorporatingexpressive(nonsequential) constraintsonthevariables(theargumentstypes).

86

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

IntegerLinearProgrammingInference [Punyakanoketal.]– Continued

 Foreachargument ai andlabel t  SetupaBooleanvariable: ai, t ∈ {0,1}  indicatingif ai isclassifiedas t  Goalistomaximize

 Σ i score( ai = t ) ai,t  Subjecttothe(linear)constraints  AnyBooleanconstraintcanbeencodedthisway.

 Ifscore( ai = t)=P( ai = t),thentheobjectiveis  Findtheassignmentthatmaximizestheexpected numberofargumentsthatare correct  Subjecttotheconstraints.

87

ExamplesofConstraints [Punyakanoketal.]– Continued

 Noduplicateargumentclasses AnyBooleanrulecanbeencoded Σ ≤ asasetoflinearconstraints. a ∈ POT ARG x{a = A0 } 1

 CARG IfthereisaCarg phrase,thereisan arg beforeit ∀a’’’ ∈ POT ARG , Σ ≥ (a ∈ POT ARG ) Λ (a isbefore a’’’ ) x{a = A0 } x{a’’’ = CA0 }

 Manyotherpossibleconstraints:  Nooverlappingorembedding  IftheverbisoftypeA,noargumentoftypeB  hit cantakeonlyA0A2but NOT A3A5  Relationsbetweennumberofarguments

JointinferencecanbeusedalsotocombinedifferentSRLSystems.

88

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Results [Punyakanoketal.]– Continued

 Char:Charniak’sparser(5besttrees)

 Col:Collins’ parser F1

Col WSJ Char Char-2 79.44 Char-3 Char-4 Brown Char-5 67.75 Combined

50 60 70 80 90 OnlineDemo: http://l2r.cs.uiuc.edu/~cogcomp/ srldemo.php

89

AJointModelforSRL [Haghighi etal.]

 Themainideaistobuildarichmodelforjointscoring, whichtakesintoaccountthedependenciesamongthe labels ofargumentphrases. Onepossiblelabelingsuggestedbylocalmodels

A0 AMTMP

A1 AMTMP

90

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

JointDiscriminativeReranking [Haghighi etal.]– Continued

 Forcomputationalreasons:startwithlocalscoringmodel withstrongindependenceassumptions

 FindtopNnonoverlappingassignmentsforlocalmodel usingasimpledynamicprogram [Toutanovaetal.05]  SelectthebestassignmentamongtopNusingajoint loglinearmodel [Collins00]  Theresultingprobabilityofacompletelabeling L ofthe treeforapredicate p isgivenby:

91

JointModelFeatures [Haghighi etal.]– Continued

A0 AMTMP

A1 AMTMP

Repetition features: countofargumentswithagivenlabelc( AMTMP )=2 Complete sequence syntactic-semantic features for the core arguments: [NP_ A0 hitNP_ A1 ],[NP_ A0 VBDNP_ A1 ](backoff) [NP_ A0 hit](leftbackoff) [NP_ ARG hitNP_ ARG ](nospecificlabels) [1hit1](countsofleftandrightcorearguments) 92

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

UsingMultipleTrees [Haghighi etal.]– Continued

 UsingthebestCharniak’sparse,ondevelopmentset

 LocalModel:74.52(F 1);JointModel:76.71(F 1)  FurtherenhancedbyusingTopKtrees  Fortop k treesfromCharniak’sparsert1,t2 ,L,tk find correspondingbestSRLassignmentsandchoosethe L1,L , L k treeandassignmentthatmaximizethescore(approx.joint probabilityoftreeandassignment)

 FinalResults:

 WSJ23:78.45(F 1),79.54(Prec.),77.39(Rec.)

 Brown:67.71(F 1),70.24(Prec.),65.37(Rec.)

 Bugfixedpostevaluation:WSJ2380.32(F 1)Brown68.81(F 1)

93

DetailsofCoNLL05Systems

 Topperformingsystems  Màrquezetal.(TechnicalUniversityofCatalonia)  Pradhanetal.(UniversityofColoradoatBoulder)  Punyakanoketal.(U.ofIllinoisatUrbanaChampaign)  Haghighi etal.(StanfordUniversity)

 Spotlightsystems  Yi&Palmer– integratingsyntacticandsemanticparsing  Cohn&Blunsom – SRLwithTreeCRFs  Carreras– systemcombination

94

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

TheIntegrationofSyntacticParsingand SemanticRoleLabeling [Yi&Palmer]

 ThebottleneckoftheSRLtask:parsing  With[Xue&Palmer 04]pruning,givendifferentparsers: 12%~18%argumentsarelost(DevelopmentSet:WSJ22)  Whatdowewantfromsyntacticparsing?  Correctconstituentboundaries  Correcttreestructures:expressingthedependencybetweenthe targetverbanditsarguments(e.g.,the path feature)

 Theproposedapproach:  Combinesyntacticparsing&argumentidentification(different cutofthetask)  Trainanewparseronthetrainingdatacreatedbymergingthe PennTreebank&thePropBank(sec0221)

SlidefromYi&Palmer (CoNLL 2005) 95

DataPreparation&BaseParser [Yi&Palmer]– Continued

 Datapreparationsteps  StripoffthePennTreebankfunctiontags  2typesofsublabelstorepresentthePropBankarguments  AN:corearguments  AM:adjunctlikearguments  Trainnewmaximumentropyparsers [Ratnaparkhi99]

S

NPAN VP

NPAN PPAN NPAM

NP

Kristina hit Scott withabaseball yesterday 96 96 BasedonYi&Palmer’s slides(CoNLL 2005)

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Results&Discussion [Yi&Palmer]– Continued

 Overallresultsondevelopmentset

F1 Prec. Rec. ANparser 67.28 71.31 63.68 AMparser 69.31 74.09 65.11 Charniak 69.98 76.31 64.62 Combined 72.73 75.70 69.99

 FinalF 1 – WSJ23:75.17,Brown:63.14

 WorsethanusingCharniak’sdirectly  Becauseofweakerbaseparser?  Hurtbothparsingandargumentidentification?

97

SRLwithTreeCRFs [Cohn&Blunsom]

 Adifferentjointmodel– applytreeCRFs  GeneratethefullparsetreeusingCollins’ parser  Prunethetreeusing[Xue&Palmer 04]  Labeleachremainingconstituentthe semantic role or None  LearntheCRFs model

 EfficientCRFinferencemethodsexistfortrees  MaximumLikelihoodTraining:sumproductalgorithm  FindingthebestinTesting:maxproductalgorithm

98

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

TreeLabeling [Cohn&Blunsom]– Continued

None

V A0

A1 A2 AM-TMP

None

99

ModelandResults [Cohn&Blunsom]– Continued

1 = λ  DefinitionofCRFs p(y | x) exp∑∑ k fk (c,y c ,x) Z(x) c∈C k  Maximumloglikelihoodtraining

E~p(x,y)[ fk ]− E p(x,y)[ fk ] = 0

 Usesumproducttocalculatemarginal E p(x,y)[ fk ]  Inference  Usemaxproducttofindthebestlabeling

 Results: FinalF 1 – WSJ23:73.10,Brown:63.63  Findings [Cohn&Blunsom CoNLL05slides] :  CRFs improvedovermaxent classifier(+1%)  Charniakparsesmoreuseful(+3%)  Veryfewinconsistentancestor/dependentlabelings  Quiteanumberofduplicateargumentpredictions

DatafromCohn&Blunsom’s slide(CoNLL 2005) 100

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SystemCombination [Carrerasetal.]

 Howmuchcanwegainfromcombiningdifferent participatingsystemsatargumentlevel?  Eachsystemproposesarguments,scoredaccording

tooverallF 1 ondevelopment  Thefinalscoreforanargumentisthesumofscores givenbysystems  GreedySelection  Repeat,untilnomoreargumentsinthecandidatelist  Selectargumentcandidatewiththebestscore  Removingoverlappingargumentsfromcandidatelist

101

Results&Discussion [Carrerasetal.]– Continued

WSJ23 F1 Prec. Rec. punyakanok+haghighi+pradhan 80.21 79.10 81.36 punyakanok 79.44 82.28 76.78

Brown F1 Prec. Rec. haghighi+marquez+pradhan+tsai 69.74 69.40 70.10 punyakanok 67.75 73.38 62.93

 Thegreedymethodofcombingsystems increasesrecallbutsacrificesprecision.

 ThegainonF 1 isnothuge.

102

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

PartIII:CoNLL05SharedTaskonSRL

 Detailsoftopsystemsandinteresting systems  Introducethetop4systems  Describe3spotlightsystems  Analysisoftheoverallresults  Generalperformance  Systemproperties  Perargumentperformance  DirectionsforimprovingSRLsystems

103

ResultsonWSJandBrownTests

F1:70%~80% Smalldifferences

Everysystem suffersfrom crossdomain test(~10%)

FigurefromCarreras&Màrquez’s slide(CoNLL 2005) 104

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SystemProperties

 LearningMethods  SNoW,MaxEnt,AdaBoost,SVM,CRFs,etc.  Thechoiceoflearningalgorithmsislessimportant.

 Features  Allteamsimplementmoreorlessthestandardfeatures withsomevariations.  Amustdoforbuildingagoodsystem!  Aclearfeaturestudyandmorefeatureengineeringwill behelpful.

105

SystemProperties– Continued

 SyntacticInformation  Charniak’sparser,Collins’ parser,clauser,chunker,etc.  TopsystemsuseCharniak’sparserorsomemixture  Qualityofsyntacticinformationisveryimportant!

 System/InformationCombination  8teamsimplementsomelevelofcombination  Greedy,Reranking,Stacking,ILPinference  Combinationofsystemsorsyntacticinformationisa goodstrategytoreducetheinfluenceofincorrect syntacticinformation!

106

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

PerArgumentPerformance CoNLL05ResultsonWSJTest

 CoreArguments  Adjuncts(Freq.~30%) (Freq.~70%) BestF 1 Freq. TMP 78.21 6.86% BestF 1 Freq. A0 88.31 25.58% ADV 59.73 3.46% A1 79.91 35.36% DIS 80.45 2.05% A2 70.26 8.26% MNR 59.22 2.67% A3 65.26 1.39% LOC 60.99 2.48% A4 77.25 1.09% MOD 98.47 3.83% CAU 64.62 0.50% NEG 98.91 1.36% Argumentsthatneed tobeimproved

107 DatafromCarreras&Màrquez’s slides(CoNLL 2005)

GroupsofVerbsinWSJTest

 BytheirfrequenciesinWSJTrain 0 120 21100 101500 5011000 Verbs 34 418 359 149 18 Props 37 568 1098 1896 765 Args. 70 1049 2066 3559 1450

 CoNLL05ResultsonWSJTest– CoreArguments 0 120 21100 101500 5011000 Args.% 0.9 12.8 25.2 43.4 17.7

BestF 1 73.38 76.05 80.43 81.70 80.31

Argumentsoflowfrequency verbsneedtobeimproved 108 DatafromCarreras&Màrquez’s slides(CoNLL 2005)

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

PartIII:CoNLL05SharedTaskonSRL

 Detailsoftopsystemsandinteresting systems  Introducethetop4systems  Describe3spotlightsystems  Analysisoftheoverallresults  Generalperformance  Systemproperties  Perargumentperformance  DirectionsforimprovingSRLsystems

109

DirectionsforImprovingSRL

 Betterfeatureengineering  Maybethemostimportantissueinpractice

 Jointmodeling/inference  Howtoimprovecurrentapproaches?

 Finetunedlearningcomponents  Canamorecomplicatedsystemhelp?

 Crossdomainrobustness  ChallengetoapplyingSRLsystems

110

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

BetterFeatureEngineering

Gildea&Jurafsky ’02 Surdeanuetal’03 Xue&Palmer ’04 • Targetpredicate • ContentWord • Featureconjunctions • Voice • HeadWordPOS • Syntacticframe • Subcategorization • ContentWordPOS • HeadofPPParent • Path • NamedEntity • Position(left,right) Pradhanetal’04 • PhraseType • PhraseType/HeadWord/POSof • GoverningCategory Left/Right/Parentconstituent • HeadWord • First/Lastword/POS

 Individualfeaturecontributionisnotclear  Everysetoffeaturesprovidesomeimprovement,but…  Differentsystem,differentcorpus,differentusage

111

JointModel/Inference

 Unlesspurelocalmodelreachesprefectresults,joint model/inferenceoftencanimprovetheperformance

 Greedyrules  Fast&Effective  Withnoclearobjectivefunction  Oftenincreaserecallbysacrificingprecision

 Integerlinearprogramminginference [Roth&Yih 04]  Withclearobjectivefunction  Canrepresentfairlygeneral hard constraints  Moreexpensivetointegrate soft(statistical) constraints

 JointModel [Toutanovaetal.05][Cohn&Blunsom 05]  Capturestatisticalandhardconstraintsdirectlyfromthedata  Needrerankingtoavoidcomplexityproblems [Toutanovaetal.05]  Captureonlylocaldependency [Cohn&Blunsom 05] 112

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

FinetunedLearningComponents

 Separatecoreargumentsandadjuncts  Adjunctsareindependentofthetargetverb  Performancemaybeenhancedwithspecificfeatures  Pradhanetal.(2005)didfeatureselectionforeachargumenttype

 Trainsystemsfordifferent(groupsof)verbs  Verbs(orsenses)mayhaveverydifferentrolesets  Example:stay.01 (remain) vs.look.02 (seeming)

[A1 Consumerconfidence] stayed [A3 strong] inOctober.

[A0 Thedemand] looked [A1 strong] inOctober.

113

CrossDomainRobustness

 TheperformanceofSRLsystemsdrops significantlywhenappliedonadifferentcorpus

 ~10%F 1 fromWSJtoBrown  Theperformanceofallthesyntactictaggersand parsersdropssignificantly  AlltrainedonWSJ

 Maynotbuildarobustsystemwithoutdata  Semisupervisedlearning  Activelearning

114

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SummaryofPartIII: CoNLL05SharedTaskonSRL

 DescribedthedetailsoftopperformingSRLsystems  Implementgenerallyall standard features  Usegoodsyntacticinformation– Charniak’s parser&more  Deploysystem/informationcombinationschemes

 Achieve~80%F 1 onWSJ,~70%F 1 onBrown

 Introducedsomeinterestingsystems  Trainsyntacticparserandargumentidentifiertogether  ApplyTreeCRFs model  Investigatetheperformanceofalargesystemcombination

115

SummaryofPartIII: CoNLL05SharedTaskonSRL– Continued

 AnalyzedtheresultsoftheCoNLL05systems  Generalperformance  PerformanceonWSJisbetween70%and80%  Thedifferencesamongsystemsaresmall

 Everysystemsuffersfromcrossdomaintest;~10%F 1 droponBrowncorpus  Perargumentperformance  CoreargumentsA1andA2andsomefrequentadjunct argumentsneedtobeimproved  Argumentsoflowfrequencyverbsneedtobeimproved

116

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SummaryofPartIII: CoNLL05SharedTaskonSRL– Continued

 DirectionsforimprovingSRLsystems  Performcarefulfeaturestudy  Designbetterfeatures  Enhancecurrentjointmodel/inferencetechniques  Separatemodelsfordifferentargumentsets  Improvecrossdomainrobustness

 Next part :ApplicationsofSRLsystems

117

QuickOverview

 PartI.Introduction  WhatisSemanticRoleLabeling?  Frommanuallycreatedgrammarstostatisticalapproaches  EarlyWork  Corpora– FrameNet,PropBank,ChinesePropBank,NomBank  TherelationbetweenSemanticRoleLabelingandothertasks  PartII.GeneraloverviewofSRLsystems  Systemarchitectures  Machinelearningmodels  PartIII.CoNLL05sharedtaskonSRL  Detailsoftopsystemsandinterestingsystems  Analysisoftheresults  ResearchdirectionsonimprovingSRLsystems  PartIV.ApplicationsofSRL

118

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

PartIV:Applications

 InformationExtraction  Reducedevelopmenttime  Summarization  Sentencematching  QuestionAnswering  Understandquestionsbetter  TextualEntailment  Deepersemanticrepresentation

119

SRLinInformationExtraction [Surdeanuetal.03]

 InformationExtraction( HUBEvent99evaluations, [Hirschmanetal99] )  Asetofdomaindependent templettes ,summarizing informationabouteventsfrommultiplesentences

:= INSTRUMENT London[gold] AMOUNT_CHANGE fell[$4.70]cents CURRENT_VALUE $308.45 DATE: daily

Timeforour daily marketreportfromNASDAQ.

Londongold fell$4.70cents to $308.45. 120

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SRLinInformationExtraction [Surdeanuetal.03]Continued

 Findpredicateargumentrelationsand mapresultingstructuresintotemplettes S viahandwrittensimplerules VP NP ARG1 and MARKET_CHANGE_VERB => NP

INSTRUMENT PP ARG2 and (MONEY or PERCENT or QAUNTITY) and MARKET_CHANGE_VERB => NorwalkbasedMicroWarehouse fell5¼to34½ AMOUNT_CHANGE ARG1 ARG2 ARGMDIR

(ARG4 or ARGM_DIR) and NUMBER and INSTRUMENT AMNT_CHANGE CURR_VALUE MARKET_CHANGE_VERB=> CURRENT_VALUE

121

SRLinInformationExtraction [Surdeanuetal.03]Continued

 100 Results 91.3 90  SRL1 82.8  Identification71.9 80 72.7 SRL 1 68.9  Classification78.9 70 67.0 SRL 2 58.4 FSA  SRL2 60

 Identification89.0 50

 Classification83.7 40  FSAisatraditional Market Change Death finitestateapproach BetterSRLleadstosignificantly betterIEperformance. TheFSAapproachdoesbetterbutrequiresintensivehuman effort(10persondays). ThesystemsusingSRLrequire2hoursofhumaneffort. 122

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SRLinSummarization (SQUASH, [Melli etal.05]SFU )

 Thetaskistogeneratea250wordsummaryfrom multipledocuments  Givenaspecifiedtopicandlevelofdetail(specific,general) Title: AmericanTobaccoCompaniesOverseas Narrative: Intheearly1990's,Americantobaccocompaniestriedtoexpand theirbusinessoverseas.Whatdidthesecompaniesdoortrytodoand where?Howdidtheirparentcompaniesfare? Granularity :specific

 ThesystemusesSRLextensivelyfor:  Estimatingasignificancescoreforasentence  whichentitiesparticipateinwhichsemanticrelations  Estimatingsentencesimilarity  whichentitiesparticipatinginwhich semanticrelations arecontained intwosentences

123

SRLinSummarization (SQUASH, [Melli etal.05] Continued )

 ItisnotpossibletoremovejusttheSRLcomponentfrom thesystemsinceSRLisusedthroughout  ImprovingtheSRLsystemimprovesSummarization performance(ROUGE2scoresonthedevelopmentset)  NaïveSRL 0.0699  ASSERTSRL 0.0731  Thisisaprettylargeimprovementconsideringthe impactofothersuccessfulfeatures  Biastowardthefirstsentences 0.0714 → 0.0738  TheoverallplacementofanearlierversionofS QUASH was7 th outof25systemsinDUC2005

124

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SRLinQuestionAnswering [Narayanan&Harabagiu04]

 ParsingQuestions Q:WhatkindofmaterialswerestolenfromtheRussiannavy?

PAS(Q):What[ A1 kindofnuclearmaterials]were[Predicate: stolen ] [A2 fromtheRussianNavy]?

 ParsingAnswers A(Q) :Russia’sPacificFleethasalsofallenpreytonucleartheft;in1/96,approximately7kg of HEUwasreportedlystolenfromanavalbaseinSovetskaya Gavan.

PAS(A(Q)):[ A1(P1) Russia’sPacificFleet]has[ AMDIS(P1) also] [P1:fallen][ A1(P1) preytonucleartheft]; [AMTMP(P2) in1/96],[ A1(P2) approximately7kgofHEU] was[ AMADV(P2) reportedly][P2: stolen ] [A2(P2) fromanavalbase][ A3(P2) inSovetskawa Gavan]

 Result:exactanswer= “approximately7kgofHEU”

SlidefromHarabagiuandNarayanan(HLT2004) 125

SRLinQuestionAnswering [Narayanan&Harabagiu04]Continued

 ParsingQuestions Q:WhatkindofmaterialswerestolenfromtheRussiannavy?

FS(Q):What[ GOODS kindofnuclearmaterials]were[ TargetPredicate stolen ] [VICTIM fromtheRussianNavy]?

 ParsingAnswers A(Q) :Russia’sPacificFleethasalsofallenpreytonucleartheft;in1/96,approximately7kg of HEUwasreportedlystolenfromanavalbaseinSovetskaya Gavan.

FS(A(Q)):[ VICTIM(P1) Russia’sPacificFleet]hasalsofallenpreyto[ GOODS(P1) nuclear] [TargetPredicate(P1) theft];in1/96,[ GOODS(P2) approximately7kgofHEU] wasreportedly[ TargetPredicate(P2) stolen ] [VICTIM(P2) fromanavalbase][ SOURCE(P2) inSovetskawa Gavan]

 Result:exactanswer= “approximately7kgofHEU”

SlidefromHarabagiuandNarayanan(HLT2004) 126

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SRLinQuestionAnswering [Narayanan&Harabagiu04]Continued

 Evaluationofgainsduetopredicateargument information. Structure Used Percent of Questions AnswerHierarchy 12% PropBankanalyses 32% FrameNetanalyses 19%

 Percentofquestionsforwhichthecorrectanswertypewas identifiedthroughusingeachstructure.  Question: Whatistheadditionalvaluecomparedto matchingbasedonsyntacticanalyses?

127

SRLinTextualEntailment [Brazetal.05]

 Doesagiventext S entailagivensentence T  S:Thebombershadnotmanagedtoenterthebuilding  T:Thebombersenteredthebuilding  Evaluatingentailmentbymatchingpredicate argumentstructure

 S1:[ ARG0 Thebombers]had[ ARGM_NEG not]managedto [PRED enter][ ARG1 thebuilding]

 T1:[ ARG0 Thebombers][ PRED entered][ ARG1 thebuilding] S doesnotentail T becausetheydonothavethe samesetofarguments

128

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

SRLinTextualEntailment [Brazetal.05]Continued

 SRLformsthebasisofthealgorithmfordeciding entailment.

 Itisalsoextensivelyusedin rewriterules whichpreserve semanticequivalence.

 NotpossibletoisolatetheeffectofSRLandunknown whetherasyntacticparseapproachcandosimilarlywell.

 ResultsonthePASCALRTEchallenge2005

 Wordbasedbaseline: 54.7

 SystemusingSRLandsyntacticparsing: 65.9  Thesystemplaced4 th outof28runsby16teamsinthe PASCALRTEChallenge 129

SummaryofPartIV:Applications

 InformationExtraction  SRLhasadvantagesindevelopmenttime;goodSRL → goodIE  FSAsystemsarestillabout10%better.  Summarization  SophisticatedsentencematchingusingSRL  ImprovingSRLimprovessummarization.  QuestionAnswering  Havingmorecomplexsemanticstructuresincreasesthenumberof questionsthatcanbehandledabout3times.  TextualEntailment  SRLenablescomplexinferenceswhicharenotallowedusing surfacerepresentations.  Action item :evaluatecontributionsofSRLvs.syntacticparsing  Noneofthesystemsperformsacarefulcomparison

130

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Conclusions

 SemanticRoleLabelingisrelativelynewbuthas attractedalotofinterest  Largecorporawithannotateddataareavailable  FrameNet,PropBank  Itprovidesanovel broadcoverage levelof semanticinterpretation  Shallowerthansomealternatives(DeepParsingfor limitedandbroaddomains)  Deeperthanothers(PennTreebankanalyseswith functiontags)  TaskswhichprofitfromPennTreebanksyntactic analysesshouldprofitfromthissemanticlayer

131

Conclusions CurrentStateoftheArtsystems

 Achieveabout 80% perargumentFmeasure A0 (60% wholepropositionscorrect) A1  Performanceisrespectablebutstillthereisalotof roomforimprovement  Interannotatoragreementis 99% for all nodesgiven goldstandard syntacticparses(chanceagreementis 88% );notcomparableto systemresults  Buildonthestrengthofstatisticalparsingmodels  Performpoorlywhenthesyntacticparsersdoso  Usesyntacticinformationextensively  Havemechanismsforincreasingrobustnesstoparsererror  Usepowerfulmachinelearningtechniques  Modeldependenciesamongargumentlabels

132

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Conclusions DirectionsforImprovingSRL

 Increaserobustnesstosyntacticparsererror  Findwaystocollectadditionalknowledge  Useunlabeleddata  Shareinformationacrossverbs  CanapplicationscreatemoredataforSRL automatically?  Improvethestatisticalmodels  Otherfeatures,otherdependencies  Improvesearch/inferenceprocedures

133

Conclusions MajorChallenges

 NeedtoconnectSRLtonaturallanguage applications  Studytheadditionalvalueofsemanticlabels comparedtosurfacerepresentationsandsyntactic analyses  ApplySRLtootherapplications  MoreInformationExtractionapplications  ATISlabelingandNLinterfacestodatabases  Havewedefinedthecorporawell?  Validatetheannotationstandardsthroughapplicationdomains  WhatlevelofaccuracyisneededinorderforSRLto beuseful?

134

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

Final Remarks

 SemanticRoleLabelingisanexcitingareaofresearch!  Progressisfast  Thereisstillroomforlargecontributions  Providesrobustbroadcoveragesemanticrepresentations  Easyintegrationwithapplications(InformationExtraction, QuestionAnswering,Summarization,TextualEntailment)  Goodresultsintasks  ToolsavailableonlinethatproduceSRLstructures  ASSERT (Automatic Statistical SE mantic Role Tagger) http://oak.colorado.edu/assert  UIUCsystem (http://l2r.cs.uiuc.edu/~cogcomp/srldemo.php )

135

Acknowledgments

 We’dliketothankthefollowingpeople,whokindlyprovidedtheir slidestousorhelpedusunderstandtheirsystems.  LucyVanderwende,SameerPradhan,XavierCarreras,Lluís Màrquez, SzutingYi,Mihai Surdeanu,Anoop Sarkar,Srini Narayanan,Sanda Harabagiu,andMarkSammons.

 WeareverygratefultoJoshuaGoodman,whogaveusmany valuablecommentsandhelpedustopreparethematerials.

 Wearealsothankfultoourcolleaguesandfriendswhoattendedour practicaltalkandgaveususefulfeedback.

 Finally,wethanktheaudienceofourtutorialfortheirinterestand alsothequestionsanddiscussions.

136

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

References:Introduction

 Hiyan Alshawi,editor.Thecorelanguageengine. MITPress,1992.  IonAndroutsopoulos,GraemeRitchie,andPeterThanisch.Natural languageinterfacestodatabases anintroduction.In JournalofNatural LanguageEngineering1(1),1995.  DouglasAppelt andDavidIsrael.IntroductiontoInformationExtraction technology. Tutorialat IJCAI1999.  DonBlahetaandEugeneCharniak.Assigningfunctiontagstoparsedtext. In ProceedingsofNAACL2000.  DonBlaheta.Functiontagging. PhDThesis,BrownCSDepartment,2003.  JoanBresnan.Lexicalfunctionalsyntax. Blackwell,2001.  AnnCopestake andDanFlickinger.Anopensourcegrammardevelopment environmentandbroadcoverageEnglishgrammarusingHPSG.In ProceedingsofLREC2000.  JohnDowding,JeanGawron,DougAppelt,JohnBear,LynnCherny, RobertMoore,andDouglasMoran.Gemini:anaturallanguagesystemfor spokenlanguageunderstanding.In ProceedingsofACL1993. 137

References:Introduction

 CharlesJ.Fillmore,CharlesWooters,andCollinF.Baker.Buildingalarge lexicaldatabankwhichprovidesdeepsemantics. In Proceedingsofthe PacificAsianConferenceonLanguage,InformationandComputation2001.  Ruifang Ge andRaymondMooney.Astatisticalsemanticparserthat integratessyntaxandsemantics.In ProceedingsofCoNLL 2005.  GraemeHirst.Semanticinterpretationandtheresolutionofambiguity (Studiesinnaturallanguageprocessing). CambridgeUniversityPress, 1987.  LynetteHirschman,PatriciaRobinson,LisaFerro,NancyChinchor,Erica Brown,RalphGrishman,andBethSundheim. Hub4Event99general guidelinesandtemplettes,1999 .  JohnT.MaxwellandRonaldM.Kaplan.Theinterfacebetweenphrasaland functionalconstraints.In ComputationalLinguistics,19(4),1993.

138

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

References:Introduction

 IonMuslea.ExtractionpatternsforInformationExtractiontasks:asurvey. In ProceedingsoftheAAAIWorkshoponMachineLearningforIE,1999.  YusukeMiyaoandJun'ichiTsujii.Probabilisticdisambiguationmodelsfor widecoverageHPSGparsing.In ProceedingsofACL2005.  ScottMiller,RobertBobrow,RobertIngria,andRichardSchwartz.AFully statisticalapproachtonaturallanguageinterfaces.In ProceedingsofACL 1996.  MarthaPalmer,DanGildea,andPaulKingsbury.ThePropositionBank:An annotatedcorpusofsemanticroles.In ComputationalLinguistics,31(1), 2005.  CarlPollardandIvanA.Sag.HeadDrivenPhraseStructureGrammar. UniversityofChicagoPress,1994.  PattiPrice.Evaluationofspokenlanguagesystems:theATISdomain. In ProceedingsofthethirdDARPASpeechandNaturalLanguageWorkshop, 1990.

139

References:Introduction

 StefanRiezler,TracyH.King,RonaldM.Kaplan,RichardCrouch,JohnT. MaxwellIII,andMarkJohnson.ParsingtheWallStreetJournalusinga LexicalFunctionalgrammaranddiscriminativeestimationtechniques. In ProceedingsofACL2002.  HisamiSuzukiandKristinaToutanova.Learningtopredictcasemarkersin Japanese.In ProceedingsofACLCOLING 2006.  KristinaToutanova,PenkaMarkova,andChristopherD.Manning.Theleaf projectionpathviewofparsetrees:ExploringstringkernelsforHPSGparse selection.In ProceedingsofEMNLP2004.  Kiyotaka Uchimoto,SatoshiSekine andHitoshi Isahara.Text generation from keywords.In Proceedings ofCOLING2002.  YeYiWang,JohnLee,MilindMahajan,andAlexAcero.Combining statisticalandknowledgebasedspokenlanguageunderstandingin conditionalmodels. In ProceedingsofACLCOLING2006.

140

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

References:Introduction

 YeYiWang,LiDeng,andAlexAcero. Spokenlanguageunderstanding:An introductiontothestatisticalframework. In IEEESignalProcessing Magazine,Vol 27No.5.2005.  WayneWard.RecentImprovementsintheCMUspokenlanguage understandingsystem.In ProceedingsofHumanLanguageTechnology Workshop,1994.  YukWah WongandRaymondMooney.Learningforsemanticparsingwith statistcial machinetranslation.In ProceedingsofHLT/NAACL2006.  JohnZelle andRaymondMooney.Learningtoparsedatabasequeries usinginductivelogicprogramming.In ProceedingsofAAAI1996.  LukeZettlemoyer andMichaelCollins.Learningtomapsentencestological form:structuredclassificationwithprobabilisticCategorialGrammars.In ProceedingsofUAI2005.

141

References:OverviewofSRLSystems

 JohnChenandOwenRambow.Useofdeeplinguisticfeaturesforthe recognitionandlabelingofsemanticarguments.InProceedingsofEMNLP 2003.  XavierCarrerasand LluísMàrquez.IntroductiontotheCoNLL2005shared task:Semanticrolelabeling.In ProceedingsofCoNLL 2005 .  TrevorCohnandPhilipBlunsom.Semanticrolelabelling withtree ConditionalRandomFields. In ProceedingsofCoNLL 2005.  DanielGildeaandDanielJurafsky.Automaticlabelingofsemanticroles.In ComputationalLinguistics,28(3),2002 .  DanielGildeaandMarthaPalmer.Thenecessityofparsingforpredicate argumentrecognition.In ProceedingsofACL2002 .  DanielGildeaandJuliaHockenmaier.Identifyingsemanticrolesusing CombinatoryCategorialGrammar.In ProceedingsofEMNLP2003.

142

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

References:OverviewofSRLSystems

 AriaHaghighi,KristinaToutanova,andChristopherManning.Ajointmodel forsemanticrolelabeling. In ProceedingsofCoNLL 2005.  LluísMàrquez,Pere Comas,Jesús Giménez,andNeus Català. Semanticrolelabelingassequentialtagging.In ProceedingsofCoNLL 2005.  SameerPradhan,WayneWard,Kadri Hacioglu,JamesH.MartinandDan Jurafsky.Semanticrolelabelingusingdifferentsyntacticviews.In ProceedingsofACL2005.  SameerPradhan,WayneWard,Kadri Hacioglu,JamesMartin,andDan Jurafsky.ShallowsemanticparsingusingSupportVectorMachines.In ProceedingsofHLT2004.  VasinPunyakanok,DanRoth,WentauYihandDavZimak.Semanticrole labelingviaIntegerLinearProgramminginference.In Proceedingsof COLING2004.

143

References:OverviewofSRLSystems

 VasinPunyakanok,DanRoth,andWentauYih.Thenecessityofsyntactic parsingforsemanticrolelabeling.In ProceedingsofIJCAI2005 .  Mihai Surdeanu,SandaHarabagiu,JohnWilliams,andPaulAarseth.Using predicateargumentstructuresforInformationExtraction.In Proceedingsof ACL2003 .  KristinaToutanova.Effectivestatisticalmodelsforsyntacticandsemantic disambiguation. PhDThesis,StanfordCSDepartment,2005.  KristinaToutanova,AriaHaghighi,andChristopherD.Manning.Joint learningimprovessemanticrolelabeling.In ProceedingsofACL2005 .  Nianwen XueandMarthaPalmer.Calibratingfeaturesforsemanticrole labeling.In ProceedingsofEMNLP2004 .

144

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

References: CoNLL05SharedTaskonSRL

 XavierCarrerasand LluísMàrquez.IntroductiontotheCoNLL2005shared task:Semanticrolelabeling.In ProceedingsofCoNLL 2005 .  TrevorCohnandPhilipBlunsom.Semanticrolelabelling withtree ConditionalRandomFields. In ProceedingsofCoNLL2005.  MichaelCollinsandTerryKoo.Discriminativererankingfornatural languageparsing.In ComputationalLinguistics31(1),2005.  DanielGildeaandDanielJurafsky.Automaticlabelingofsemanticroles.In ComputationalLinguistics,28(3),2002 .  Kadri Hacioglu andWayneWard.Targetworddetectionandsemanticrole chunkingusingSupportVectorMachines.In ProceedingsofHLTNACCL 2003.  AriaHaghighi,KristinaToutanova,andChristopherManning.AJointmodel forsemanticrolelabeling. In ProceedingsofCoNLL2005.

145

References: CoNLL05SharedTaskonSRL

 VasinPunyakanok,DanRoth,andWentauYih.Generalizedinferencewith multiplesemanticrolelabelingsystems.In ProceedingsofCoNLL2005.  Taku Kudo andYujiMatsumoto.ChunkingwithSupportVectorMachines.In ProceedingsofNAACL2001.  LluísMàrquez,Pere Comas,Jesús Giménez,andNeus Català. Semanticrolelabelingassequentialtagging.In ProceedingsofCoNLL 2005.  SameerPradhan,WayneWard,Kadri Hacioglu,JamesMartin,andDan Jurafsky.ShallowsemanticparsingusingSupportVectorMachines.In ProceedingsofHLT2004.  SameerPradhan,Kadri Hacioglu,WayneWard,JamesH.Martin,andDaniel Jurafsky.Semanticrolechunkingcombiningcomplementarysyntacticviews. In ProceedingsofCoNLL 2005.  DanRothandWentauYih.ALinearProgrammingformulationforglobal inferenceinnaturallanguagetasks.In ProceedingsofCOLING2004 .

146

WentauYih&KristinaToutanova HLTNAACL06Tutorial AutomaticSemanticRoleLabeling

References: CoNLL05SharedTaskonSRL

 Mihai Surdeanu,SandaHarabagiu,JohnWilliams,andPaulAarseth.Using predicateargumentstructuresforInformationExtraction.In Proceedingsof ACL2003 .  KristinaToutanova,AriaHaghighi,andChristopherD.Manning.Joint learningimprovessemanticrolelabeling.In ProceedingsofACL2005 .  SzutingYiandMarthaPalmer.Theintegrationofsyntacticparsingand semanticrolelabeling.In ProceedingsofCoNLL 2005.  Nianwen XueandMarthaPalmer.Calibratingfeaturesforsemanticrole labeling.In ProceedingsofEMNLP2004 .

147

References:Applications

 RodrigodeSalvoBraz,RoxanaGirju,VasinPunyakanok,DanRoth,and MarkSammons.Aninferencemodelforsemanticentailmentinnatural language.In ProceedingsofAAAI2005 .  LynetteHirschman,PatriciaRobinson,LisaFerro,NancyChinchor,Erica Brown,RalphGrishman,andBethSundheim. Hub4Event99general guidelinesandtemplettes,1999 .  GaborMelli,YangWang,Yudong Liu,Mehdi M.Kashani,Zhongmin Shi, Baohua Gu,Anoop Sarkar andFredPopowich.DescriptionofSQUASH,the SFUquestionansweringsummaryhandlerfortheDUC2005summarization task.In ProceedingsofDUC2005 .  SriniNarayananandSandaHarabagiu.Questionansweringbasedon semanticstructures.In ProceedingsofCOLING2004.  Mihai Surdeanu,SandaHarabagiu,JohnWilliams,andPaulAarseth.Using predicateargumentstructuresforInformationExtraction.In Proceedingsof ACL2003 .

148

WentauYih&KristinaToutanova