2008-PHD-Global Inference for Sentence Compression An Integer Linear Programming Approach.pdf

《2008-PHD-Global Inference for Sentence Compression An Integer Linear Programming Approach.pdf》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、GlobalInferenceforSentenceCompression:AnIntegerLinearProgrammingApproachJamesClarkeNIVERUSEITHYTOHFGERDUINBDoctorofPhilosophyInstituteforCommunicatingandCollaborativeSystemsSchoolofInformaticsUniversityofEdinburgh2008AbstractInthisthesiswedevelopmodelsforsentencecompression.Thistex

2、trewritingtaskhasrecentlyattractedalotofattentionduetoitsrelevanceforapplications(e.g.,sum-marisation)andsimpleformulationbymeansofworddeletion.Previousmodelsforsentencecompressionhavebeeninherentlylocalandthusfailtocapturethelongrangedependenciesandcomplexinteractionsinvolvedintex

3、trewriting.Wepresentasolu-tionbyframingthetaskasanoptimisationproblemwithlocalandglobalconstraintsandrecastexistingcompressionmodelsintothisframework.Usingtheconstraintsweinstillsyntactic,semanticanddiscourseknowledgethemodelsotherwisefailtocap-ture.Weshowthattheadditionofconstrain

4、tsallowrelativelysimplelocalmodelstoreachstate-of-the-artperformanceforsentencecompression.Thethesisprovidesadetailedstudyofsentencecompressionanditsmodels.Thedifferencesbetweenautomaticandmanuallycreatedcompressioncorporaareassessedalongwithhowcompressionvariesacrosswrittenandspok

5、entext.Wealsodis-cussvarioustechniquesforautomaticallyandmanuallyevaluatingcompressionoutputagainstagoldstandard.Modelsarereviewedbasedontheirassumptions,trainingre-quirements,andscalability.Weintroduceageneralmethodforextendingpreviousapproachestoallowformoreglobalmodels.Thisisach

6、ievedthroughtheoptimisationframeworkofIntegerLinearProgramming(ILP).Wereformulatethreecompressionmodels:anunsuper-visedmodel,asemi-supervisedmodelandafullysupervisedmodelasILPproblemsandaugmentthemwithconstraints.Theseconstraintsareintuitiveforthecompressiontaskandarebothsyntactica

7、llyandsemanticallymotivated.Wedemonstratehowtheyimprovecompressionqualityandreducetherequirementsontrainingmaterial.Finally,wedelveintodocumentcompressionwherethetaskistocompressev-erysentenceofadocumentandusetheresultingsummaryasareplacementfortheoriginaldocument.Fordocument-based

8、compressionweinvestigatedi