Outline 1 Motivation 2 3 JointModelofSegmentationandEntityResolutionSegmentationEntityResolutionJointModel 4 ExperimentsModelPerformanceBidirectionality Motivation PipelineApproach Cascadingerrorthro ID: 105625
Download Pdf The PPT/PDF document "Imperatively-DenedFactorGraphs(IDFs)" 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.
Outline 1 Motivation 2 Imperatively-DenedFactorGraphs(IDFs) 3 JointModelofSegmentationandEntityResolutionSegmentationEntityResolutionJointModel 4 ExperimentsModelPerformanceBidirectionality Motivation PipelineApproach CascadingerrorthroughthestagesCanbereducedbyusingN-bestlists,orsamplingy Unidirectionalinformationowdrucker,schapire,andsimardimprovingperformanceinneuralnetworksusingaboostingalgorithm,advancesinneuralinformationprocessingsystems5,1993,42-49.14 druckerharris,schapire,robert,andsimardpatrice1993. improvingperformanceinneuralnetworksusingaboosting algorithm.inadvancesinneuralinformationsprocessingsystems5,sanma-teo,ca.morgankaufmann.199342-49. Sutton&McCallumCoNLL2005 yFinkelet.al.EMNLP2006 S.Singh,K.Schultz,A.McCallum(UMass) Bi-directionalJointInference ECMLPKDD20099/29 Motivation PipelineApproach CascadingerrorthroughthestagesCanbereducedbyusingN-bestlists,orsamplingy Unidirectionalinformationowdrucker,schapire,andsimardimprovingperformanceinneuralnetworksusingaboostingalgorithm,advancesinneuralinformationprocessingsystems5,1993,42-49.14 druckerharris,schapire,robert,andsimardpatrice1993. improvingperformanceinneuralnetworksusingaboosting algorithm.inadvancesinneuralinformationsprocessingsystems5,sanma-teo,ca.morgankaufmann.199342-49. Sutton&McCallumCoNLL2005 yFinkelet.al.EMNLP2006 S.Singh,K.Schultz,A.McCallum(UMass) Bi-directionalJointInference ECMLPKDD20099/29 Motivation IteratedPipelineApproachx ClosetheloopofthepipelineBothtasksuseinformationfromeachotherReducescascadingerrorHowever,stillnoteliminatedN-bestlistscanbeusedtofurtherreducethiserrorz zWellneret.al.UAI2004 xHollingshead&RoarkACL2007 S.Singh,K.Schultz,A.McCallum(UMass) Bi-directionalJointInference ECMLPKDD200910/29 JointModel Segmentation Segmentation VariablesToken:ObservedvariablerepresentingawordinthementionLabel:Variablethatcantakeanyoftheeldtypesasavalue Field:ConsecutiveTokensthathavethesamelabeltype Factors:LabelToken,LabelPrev/NextToken,FieldFactor S.Singh,K.Schultz,A.McCallum(UMass) Bi-directionalJointInference ECMLPKDD200919/29 JointModel EntityResolution EntityResolution VariablesMention:VariablethattakesasingleEntityasitsvalueEntity:SetofMentionsthatarecoreferent Factors:AffinityandRepulsion S.Singh,K.Schultz,A.McCallum(UMass) Bi-directionalJointInference ECMLPKDD200920/29 JointModel JointModel ExampleModel S.Singh,K.Schultz,A.McCallum(UMass) Bi-directionalJointInference ECMLPKDD200922/29 Experiments ModelPerformance ModelPerformance Table:CoraEntityResolution:PairwiseF1andClusterRecall Method Prec/Recall F1 ClusterRec. Fellegi-Sunter 78.0/97.7 86.7 62.7 JointMLN 94.3/97.0 95.6 78.1 50-90mins IsolatedIDF 97.09/95.42 96.22 86.01 3mins JointIDF 95.34/98.25 96.71 94.62 18mins Table:CoraSegmentation:TokenwiseF1 Method Author Title Venue Total IsolatedMLN 99.3 97.3 98.2 98.2 }50-90mins JointMLN 99.5 97.6 98.3 98.4 IsolatedIDF 99.35 97.63 98.58 98.51 3mins JointIDF 99.42 97.99 98.78 98.72 18mins S.Singh,K.Schultz,A.McCallum(UMass) Bi-directionalJointInference ECMLPKDD200926/29 Experiments Bidirectionality Bidirectionality (a)SegmentationF1 (b)EntityResolutionF1Figure:F1ofthejointmodelasdifferenttypesoffactorsareadded,startingwiththebasemodelcontainingonlyisolatedmodelfactors.“Semi-Joint”referstothemodelcontainingweaklyjointfactorswhilethe“Fully-Joint”modelconsistsofbi-directionalhighly-coupledfactors. S.Singh,K.Schultz,A.McCallum(UMass) Bi-directionalJointInference ECMLPKDD200927/29 Thanks!SameerSingh,KarlSchultz,AndrewMcCallumUniversityofMassachusetts,Amherst{sameer,kschultz,mccallum}@cs.umass.eduVisitusatthepostersession