IntroductionProblem definition and propertiesMain Computational Resources and SystemsState-of-the-artArchitectureFeature engineeringSRL systems in detailEmpirical evaluation and lessons learnedProblems and challengesGeneralization to new DomainsDependence on SyntaxSRL systems in applicationsConclusionsSemantic Role LabelingPast, Present and FutureLlu´ıs M`arquezTALP Research CenterTecnhical University of CataloniaTutorial at ACL-IJCNLP 2009Suntec – SingaporeAugust 2, 2009—Version from August 3, 2009—1Introduction2State-of-the-art3Empirical evaluation and lessons learned4Problems and challenges5ConclusionsIntroduction: 3Tutorial Overview1IntroductionProblem definition a nd propertiesMain Computational Resources and Sy stems2State-of-the-art3Empirical evaluation and lessons learned4Problems and challenges5ConclusionsIntroduction: Problem definition and properties 4Tutorial Overview1IntroductionProblem definition a nd propertiesMain Computational Resources and Sy stems2State-of-the-art3Empirical evaluation and lessons learned4Problems and challenges5ConclusionsIntroduction: Problem definition and properties 5Semantic Role Lab eling: The ProblemSRLdef= detecting basic event structures such aswho did what towhom, when and where [IE point of view]Introduction: Problem definition and properties 6Semantic Role Lab eling: The ProblemSRLdef= detecting basic event structures such aswho did what towhom, when and where [IE point of view]The luxury auto maker last year sold 1,214 cars in the U.S.PPNPVPNPNPPA0 AM−TMP AM−LOCPredicateA1ObjectAgentSTemporalMarkerLocativeMarkerIntroduction: Problem definition and properties 7Semantic Role Lab eling: The ProblemSRLdef= identify thearguments of a given verb and assign themsemantic labels describing the roles they play in the predicate(i.e., identify predicate argument s tructures)[CL point of view ]The luxury auto maker last year sold 1,214 cars in the U.S.PPNPVPNPNPPA0 AM−TMP AM−LOCPredicateA1ObjectAgentSTemporalMarkerLocativeMarkerIntroduction: Problem definition and properties 8Semantic Role Lab eling: The ProblemSyntactic variationsTEMPz }| {Yesterday,HITTERz }| {Kristina hitTHING HITz }| {ScottINSTRUMENTz }| {with a baseballScott was hit by Kristina yesterday with a baseballYesterday, Scott was hit wit h a baseball by KristinaWith a basebal l, Kristina hit Scott yesterdayYesterday Scott was hit by Kristina with a baseballKristina hit Scott with a baseball yesterdayExample from (Yih & Toutanova, 2006)Introduction: Problem definition and properties 9Semantic Role Lab eling: The ProblemSyntactic variationsTEMPz }| {Yesterday,HITTERz }| {Kristina hitTHING HI Tz }| {ScottINSTRUMENTz }| {with a baseballScott was hit by Kristina yesterday with a baseballYesterday, Scott was hit wit h a baseball by KristinaWith a basebal l, Kristina hit Scott yesterdayYesterday Scott was hit by Kristina with a baseballKristina hit Scott with a baseball yesterdayExample from (Yih & Toutanova, 2006)Introduction: Problem definition and properties 10Semantic Role Lab eling: The ProblemStructural viewMapping from input to output structures:Input is text (enriched with morpho-syntactic information)Output is a sequence of labeled argumentsSequential segmenting/labeling problem“ Mr. S mith sent the report to me this morning . ”[Mr. Smith]AGENTsent [the report]OBJto [me]RECIP[this morning]TMP.Mr.B−AGENTSmithIsent theB−OBJreportItoOmeB−RECIPthisB−TMPmorningI.OIntroduction: Problem definition and properties 11Semantic Rol e Labeling: The ProblemStructural viewMapping from input to output structures:Input is text (enriched with morpho-syntactic information)Output is a sequence of labeled argumentsSequential segmenting/labeling problem“ Mr. S mith sent the report to me this morning . ”[Mr. Smith]AGENTsent [the report]OBJto [me]RECIP[this morning]TMP.Mr.B−AGENTSmithIsent theB−OBJreportItoOmeB−RECIPthisB−TMPmorningI.OIntroduction: Problem definition and properties 12Semantic Rol e Labeling: The ProblemStructural viewMapping from input to output structures:Input is text (enriched with morpho-syntactic information)Output is a sequence of labeled argumentsSequential segmenting/labeling problem“ Mr. S mith sent the report to me this morning . ”[Mr. Smith]AGENTsent [the report]OBJto [me]RECIP[this morning]TMP.Mr.B−AGENTSmithIsent theB−OBJreportItoOmeB−RECIPthisB−TMPmorningI.OIntroduction: Problem definition and properties 13Semantic Rol e Labeling: The ProblemStructural Viewhad.IthathatawithtrappedcatThetheratSNP PP NP NP NP VPNP VPSAGENTPATIENTINSTRUMENTAGENTPAT.R−PAT.Output is a hierarchy of labeled argumentsIntroduction: Problem definition and properties 14Semantic Rol e Labeling: The ProblemStructural Viewhad.IthathatawithtrappedcatThetheratSNP PP NP NP NP VPNP VPSAGENTPATIENTINSTRUMENTAGENTPAT.R−PAT.Output is a hierarchy of labeled argumentsIntroduction: Problem definition and properties 15Semantic Rol e Labeling: The ProblemLinguistic nature of the problemArgument identification is strongly related to syntaxMarkerThe luxury auto maker last year sold 1,214 cars in the U.S.PPNPVPNPNPPA0 AM−TMP AM−LOCPredicateA1ObjectAgentSTemporalMarkerLocativeRole labeling is a semantic ta s ke.g., selectional preferences should play an important roleIntroduction: Problem definition and properties 16Semantic Rol e Labeling: The ProblemLinguistic nature of the problemArgument identification is strongly related to syntaxMarkerThe luxury auto maker last year sold 1,214 cars in the U.S.PPNPVPNPNPPA0 AM−TMP AM−LOCPredicateA1ObjectAgentSTemporalMarkerLocativeRole labeling is a semantic ta s ke.g., selectional preferences should play an important roleIntroduction: Problem definition and properties 17Semantic Rol e Labeling: ApplicationsIs SRL really useful for NLP a pp lications?1Information Extraction (Surdeanu et al., 2003 ; Frank et al., 2007)2Question & Answering (Narayanan and Harabagiu, 20 04)3Automatic Summarization (Melli et al., 2005)4Coreference Resolution (Ponzetto and Strube, 2006)5Machine Translation (Boas, 2002; Gim´enez and M`arquez, 2007;Wu and Fung, 2009a;2009b)6etc. [more on SRL and applications in the last section]Introduction: Problem definition and properties 18Semantic Rol e Labeling: ApplicationsIs SRL really useful for NLP a pp lications?1Information Extraction (Surdeanu et al., 2003 ; Frank et al., 2007)2Question & Answering (Narayanan and Harabagiu, 20 04)3Automatic Summarization (Melli et al., 2005)4Coreference
View Full Document