Paraphrase Acquisition via Crowdsourcing and Machine Learning基于众包和机器学习的释义获取

《Paraphrase Acquisition via Crowdsourcing and Machine Learning基于众包和机器学习的释义获取》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、ParaphraseAcquisitionviaCrowdsourcingandMachineLearningSTEVENBURROWSandMARTINPOTTHASTandBENNOSTEINWebTechnologyandInformationSystems,Bauhaus-UniversitätWeimarToparaphrasemeanstorewritecontentwhilstpreservingtheoriginalmeaning.Paraphrasingisimportantinfieldssuchastextreuseinjournali

2、sm,anonymisingwork,andimprovingthequalityofcustomer-writtenre-views.Thispapercontributestoparaphraseacquisitionandfocusesontwoaspectsthatarenotaddressedbycurrentresearch:(1)acquisitionviacrowdsourcing,and(2)acquisitionofpassage-levelsamples.Thechallengeofthefirstaspectisautomaticqua

3、lityassurance;withoutsuchameansthecrowdsourcingparadigmisnotef-fective,andwithoutcrowdsourcingthecreationoftestcorporaisunacceptablyexpensiveforrealisticorderofmagnitudes.Thesecondaspectaddressesthedeficitthatmostofthepreviousworkingeneratingandevaluat-ingparaphraseshasbeenconducted

4、usingsentence-levelparaphrasesorshorter;theseshort-sampleanalysesarelimitedintermsofapplicationtoplagiarismdetection,forexample.WepresenttheWebisCrowdParaphraseCorpus2011(Webis-CPC-11),whichrecentlyformedpartofthePAN2010internationalplagiarismdetectioncompetition.Thiscorpuscomprise

5、spassage-levelparaphraseswith4067positivesamplesand3792negativesamplesthatfailedourcriteria,usingAmazon’sMechanicalTurkforcrowdsourcing.Inthispaper,wereviewthelessonslearnedatPAN2010,andexplainindetailthemethodusedtoconstructthecorpus.Theempiricalcontributionsincludemachinelearning

6、experimentstoexploreifpassage-levelparaphrasescanbeidentifiedinatwo-classclassificationproblemusingparaphrasesimilarityfeatures,andwefindthatak-nearest-neighborclas-sifiercancorrectlydistinguishbetweenparaphrasedandnon-paraphrasedsampleswith0.980precisionat0.523recall.Thisresultimplies

7、thatjustunderhalfofoursamplesmustbediscarded(remaining0.477fraction),butourcost-analysisshowsthattheautomationweintroduceresultsina18%financialsavingandover100hoursoftimereturnedtotheresearcherswhenrepeatingasimilarcorpusdesign.Ontheotherhand,whenbuildinganunrelatedcorpusrequiringsa

8、y25%trainingdatafortheauto