vine rst second Representation Heads Modi ers Representation Heads Modi ers Representation Heads Modi ers Representation Heads Modi ers FirstOrderFeatureCalculation ArcLengthByPartofSpeech ArcLeng ID: 519137
Download Pdf The PPT/PDF document "Preview:Coarse-to-FineCascades" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Preview:Coarse-to-FineCascades vine rst second Representation Heads Modiers Representation Heads Modiers Representation Heads Modiers Representation Heads Modiers First-OrderFeatureCalculation ArcLengthByPart-of-Speech ArcLengthByPart-of-Speech ArcLengthByPart-of-Speech ArcLengthExamples ArcLengthExamples ArcLengthExamples ArcLengthExamples ArcLengthHeatMap BandedMatrix OuterArc OuterArc Coarse-to-Fine vine Coarse-to-Fine vine rst second InferenceQuestionsquestions:HowdowereduceinferencetimetoO(n)?Howdowedecidewhicharcstoprune?VineParsing(EisnerandSmith,2005) First-OrderParsing First-OrderParsing First-OrderParsing First-OrderParsing VineParsing VineParsing VineParsing VineParsing VineParsing ArcPruningPrunearcsbasedonmax-marginals.maxmarginal(a)=maxy:a2y(yw)Cancomputeusinginside-outsidealgorithm.Genericalgorithmusinghypergraphparsing. Max-MarginalPruninggoal:Deneathresholdonmax-marginalscore.Validationparameter tradeso betweenspeedandaccuracy.t (w)= maxy(yw)+(1� )1 jAjXa2Amaxmarginal(a;w)Highestscoringparseupperboundsanymax-marginal.Assumeaverageofmax-marginalsislowerthangold. PruningThreshold PruningThreshold PruningThreshold PruningThreshold PruningThreshold StructuredCascadeTraining StructuredCascadeTraining StructuredCascadeTraining StructuredCascadeTraining StructuredCascadeTraining StructuredCascadeTraining ImplementationInferenceExperimentsuseahighly-optimizedC++implementation.Baselinerst-orderparserprocesses2000tokens/sec.Hypergraphparsingframeworkwithsharedinference.ModelFinalmodelstrainedwithhamming-lossMIRA.Fullcollectionofdependencyparsingfeatures(Koo,2010).First-,second-,andthird-ordermodelsmatchstate-of-the-art. Speed/AccuracyExperiments:First-OrderParsing Speed/AccuracyExperiments:Second-OrderParsing EmpiricalComplexity:First-OrderParsing EmpiricalComplexity:Second-OrderParsing MultilingualExperiments:First-OrderParsing MultilingualExperiments:Second-OrderParsing Specialthanksto:RyanMcDonald,HaoZhang,MichaelRinggaard,TerryKoo,KeithHall,KuzmanGanchev,YoavGoldberg,AndreMartins,andtherestoftheGoogleNLPteam