personNkniendAkneelingAKnieNnkneeN 1Downloadablefromhttpgoogl7KG2UUsingderivationalknowledgeforsmoothingraisesthequestionofhowsemanticallysimilarthelemmaswithinafamilyreallyareFortunate ID: 128785
Download Pdf The PPT/PDF document "offfromwordstosemanticclasses,eitheradop..." 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.
offfromwordstosemanticclasses,eitheradoptedfromaresourcesuchasWordNet(Resnik,1996)orinducedfromdata(PantelandLin,2002;Wangetal.,2005;Erketal.,2010).Similarly,distributionalfeaturessupportgeneralizationinNamedEntityRecognition(Finkeletal.,2005).Althoughdistributionalinformationisoftenusedforsmoothing,toourknowledgethereislittleworkonsmoothingdistributionalmodelsthem-selves.Weseetwomainprecursorstudiesforourwork.Bergsmaetal.(2008)buildmodelsofse-lectionalpreferencesthatincludemorphologicalfeaturessuchascapitalizationandthepresenceofdigits.However,theirapproachistask-specicandrequiresa(semi-)supervisedsetting.AllanandKu-maran(2003)makeuseofmorphologybybuildinglanguagemodelsforstemming-basedequivalenceclasses.Ourapproachalsousesmorphologicalprocessing,albeitmoreprecisethanstemming.3AResourceforGermanDerivationDerivationalmorphologydescribestheprocessofbuildingnew(derived)wordsfromother(basis)words.Derivedwordscan,butdonothaveto,sharethepart-of-speech(POS)withtheirbasis(oldA!oldishAvs.warmA!warmV,warmthN).Wordscanbegroupedintoderivationalfamiliesbyform-ingthetransitiveclosureoverindividualderivationrelations.Thewordsinthesefamiliesaretypicallysemanticallysimilar,althoughtheexactdegreede-pendsonthetypeofrelationandidiosyncraticfac-tors(bookN!bookishA,Lieber(2009)).ForGerman,thereareseveralresourceswithderivationalinformation.Weuseversion1.3ofDERIVBASE(Zelleretal.,2013),1afreelyavailableresourcethatgroupsover280,000verbs,nouns,andadjectivesintomorethan17,000non-singletonderivationalfamilies.Ithasaprecisionof84%andarecallof71%.Itshighercoveragecom-paredtoCELEX(Baayenetal.,1996)andIMSLEX(Fitschen,2004)makesitparticularlysuitablefortheuseinsmoothing,wheretheresourceshouldincludelow-frequencylemmas.ThefollowingexampleillustratesafamilythatcoversthreePOSesaswellasawordwithapre-dominantmetaphoricalreading(tokneel!tobeg):knieenV(tokneelV),beknieenV(tobegV),KniendeN(kneeling personN),kniendA(kneelingA),KnieNn(kneeN) 1Downloadablefrom:http://goo.gl/7KG2UUsingderivationalknowledgeforsmoothingraisesthequestionofhowsemanticallysimilarthelem-maswithinafamilyreallyare.Fortunately,DE-RIVBASEprovidesinformationthatcanbeusedinthismanner.Itwasconstructedwithhand-writtenderivationrules,employingstringtransformationfunctionsthatmapbasislemmasontoderivedlem-mas.Forexample,asufxationruleusingtheafx“heit”generatesthederivationdunkel–Dunkel-heit(darkA–darknessN).Sincederivationalfam-iliesaredenedastransitiveclosures,eachpairofwordsinafamilyisconnectedbyaderivationpath.Becausetherulesdonothaveaperfectpre-cision,ourcondenceinpairsofwordsdecreasesthelongerthederivationpathbetweenthem.Tore-ectthis,weassigneachpairacondenceof1=n,wherenisthelengthoftheshortestpathbetweenthelemmas.Forexample,bekleiden(enrobeV)isconnectedtoVerkleidung(disguiseN)throughthreestepsviathelemmaskleiden(dressV)andverklei-den(disguiseV)andisassignedthecondence1/3.4ModelsforDerivationalSmoothingDerivationalsmoothingexploitsthefactthatderiva-tionallyrelatedwordsarealsosemanticallyrelated,bycombiningand/orcomparingdistributionalrep-resentationsofderivationallyrelatedwords.Thedenitionofaderivationalsmoothingalgorithmconsistsoftwoparts:atriggerandascheme.Notation.Givenawordw,weusewtodenoteitsdistributionalvectorandD(w)todenotethesetofvectorsforthederivationalfamilyofw.Weassumethatw2D(w).ForwordsthathavenoderivationsinDERIVBASE,D(w)isasingletonset,D(w)=fwg.Let (w;w0)denotethecon-denceofthepair(w;w0),asexplainedinSection3.Smoothingtrigger.Asdiscussedabove,thereisnoguaranteeforhighsemanticsimilaritywithinaderivationalfamily.Forthisreason,smoothingmayalsodrownoutinformation.Inthispaper,wereportontwotriggers:smoothalwaysalwaysperformssmoothing;smoothifsim=0smoothsonlywhentheunsmoothedsimilaritysim(w1;w2)iszeroorunknown(duetow1orw2notbeinginthemodel).Smoothingscheme.Wepresentthreesmoothingschemes,allofwhichapplytothelevelofcompletefamilies.Thersttwoschemesareexemplar-basedschemes,whichdenethesmoothedsimilarityforawordpairasafunctionofthepairwisesimilaritiesbetweenallwordsofthetwoderivationalfamilies. Table1:Resultsonthesemanticsimilaritytask(r:Pearsoncorrelation,Cov:Coverage) SmoothingtriggerSmoothingschemerCov% DM.DE,unsmoothed.4458.9 DM.DE,smoothalwaysavgSim.3088.0maxSim.4388.0centSim.4488.0 DM.DE,smoothifsim=0avgSim.4388.0maxSim.4288.0centSim.4788.0 BOWbaseline.3694.9 increasescorrelationsomewhattor=0:47.Thedifferencetotheunsmoothedmodelisnotsignif-icantatp=0:05accordingtoFisher's(1925)methodofcomparingcorrelationcoefcients.Thebag-of-wordsbaseline(BOW)hasagreatercoveragethanDM.DEmodels,butatthecostoflowercorrelationacrosstheboard.TheonlyDM.DEmodelthatperformsworsethantheBOWbaselineisthenon-conservativeavgSim(averagesimilarity)scheme.Weattributethisweakper-formancetothepresenceofmanypairwisezerosimilaritiesinthedata,whichmakestheavgSimpredictionsunreliable.Toourknowledge,therearenopreviouspub-lishedpapersondistributionalapproachestomod-elingthisdataset.ThebestpreviousresultisaGermaNet/Wikipedia-basedmodelbyZeschetal.(2007).Itreportsahighercorrelation(r=0:59)butaverylowcoverageat33.1%.ResultsforSynonymChoice.TheresultsforthesecondtaskareshowninTable2.Theun-smoothedmodelachievesanaccuracyof53.7%andacoverageof80.8%,asreportedbyPad´oandUtt(2012).Smoothingincreasesthecover-agebyalmost6%to86.6%(forexample,aques-tionitemforinferiorbecomescoveredafterback-ingofffromthetargettoInferiorit¨at(inferiority)).Allsmoothedmodelsshowalossinaccuracy,al-beitsmall.Thebestmodelisagainaconservativesmoothingmodel(sim=0)withalossof1.1%ac-curacy.Usingbootstrapresampling(EfronandTib-shirani,1993),weestablishedthatthedifferencetotheunsmoothedDM.DEmodelisnotsigni-cantatp0:05.Thistime,theavgSim(averagesimilarity)smoothingschemeperformsbest,withtheprototype-basedschemeinsecondplace.Thus,theresultsforsynonymchoicearelessclear-cut:derivationalsmoothingcantradeaccuracyagainstTable2:Resultsonthesynonymchoicetask(Acc:Accuracy,Cov:Coverage) SmoothingtriggerSmoothingschemeAcc%Cov% DM.DE,unsmoothed(Pad´o&Utt2012)53.780.8 DM.DE,smoothalwaysavgSim46.086.6maxSim50.386.6centSim49.186.6 DM.DE,smoothifsim=0avgSim52.686.6maxSim51.286.6centSim51.386.6 BOW“baseline”56.998.5 coveragebutdoesnotleadtoaclearimprovement.Whatismore,theBOW“baseline”signicantlyoutperformsallsyntacticmodels,smoothedandunsmoothed,withanalmostperfectcoveragecom-binedwithahigheraccuracy.6ConclusionsandOutlookInthispaper,wehaveintroducedderivationalsmoothing,anovelstrategytocombatingsparsityinsyntacticvectorspacesbycomparingandcom-biningthevectorsofmorphologicallyrelatedlem-mas.Theonlyinformationstrictlynecessaryforthemethodsweproposeisagroupingoflemmasintoderivationallyrelatedclasses.Wehavedemon-stratedthatderivationalsmoothingimprovestwotasks,increasingcoveragesubstantiallyandalsoleadingtoanumericallyhighercorrelationinthesemanticsimilaritytask,evenforvectorscreatedfromaverylargecorpus.Weobtainedthebestre-sultsforaconservativeapproach,smoothingonlyzerosimilarities.Thisalsoexplainsourfailuretoimprovelesssparseword-basedmodels,whereveryfewpairsareassignedasimilarityofzero.Acomparisonofprototype-andexemplar-basedschemesdidnotyieldaclearwinner.Theestima-tionofgenericsemanticsimilaritycanprotmorefromderivationalsmoothingthantheinductionofspeciclexicalrelations.Infuturework,weplantoworkonothereval-uationtasks,applicationtootherlanguages,andmoresophisticatedsmoothingschemes.Acknowledgments.Authors1and3weresup-portedbytheECprojectEXCITEMENT(FP7ICT-287923).Author2wassupportedbytheCroatianScienceFoundation(project02.03/162:“Deriva-tionalSemanticModelsforInformationRetrieval”).WethankJasonUttforhissupportandexpertise.