Operazioni dell'Associazione per la Linguistica Computazionale, 1 (2013) 379–390. Redattore di azioni: Lillian Lee.
Submitted 6/2013; Revised 9/2013; Pubblicato 10/2013. C
(cid:13)
2013 Associazione per la Linguistica Computazionale.
Data-DrivenMetaphorRecognitionandExplanationHongsongLiMicrosoftResearchAsiahongsli@microsoft.comKennyQ.ZhuShanghaiJiaoTongUniversitykzhu@cs.sjtu.edu.cnHaixunWangGoogleResearchhaixun@google.comAbstractRecognizingmetaphorsandidentifyingthesource-targetmappingsisanimportanttaskasmetaphoricaltextposesabigchallengeformachinereading.Toaddressthisproblem,weautomaticallyacquireametaphorknowledgebaseandanisAknowledgebasefrombillionsofwebpages.Usingtheknowledgebases,wedevelopaninferencemechanismtorec-ognizeandexplainthemetaphorsinthetext.Toourknowledge,thisisthefirstpurelydata-drivenapproachofprobabilisticmetaphorac-quisition,recognition,andexplanation.Ourresultsshowsthatitsignificantlyoutperformsotherstate-of-the-artmethodsinrecognizingandexplainingmetaphors.1IntroductionAmetaphorisawayofcommunicating.Itenablesustocomprehendonethingintermsofanother.Forexample,themetaphor,Julietisthesun,allowsustoseeJulietmuchmorevividlythanifShakespearehadtakenamoreliteralapproach.Weutteraboutonemetaphorforeverytentotwenty-fivewords,oraboutsixmetaphorsaminute(Geary,2011).Specifically,ametaphorisamappingofconceptsfromasourcedomaintoatargetdomain(LakoffandJohnson,1980).Thesourcedomainisoftencon-creteandbasedonsensoryexperience,whiletar-getdomainisusuallyabstract.Twoconceptsareconnectedbythismappingbecausetheysharesomecommonorsimilarproperties,andasaresult,themeaningofoneconceptcanbetransferredtoan-other.Forexample,in“Julietisthesun,”thesunisthesourceconceptwhileJulietisthetargetconcept.Oneinterpretationofthismetaphoristhatbothcon-ceptssharethepropertythattheirexistencebringsaboutwarmth,life,andexcitement.Inametaphor-icalsentence,atleastoneofthetwoconceptsmustbeexplicitlypresent.Thisleadstothreetypesofmetaphors:1.Julietisthesun.Here,boththesource(sun)andthetarget(Juliet)areexplicit.2.Pleasewashyourclawsbeforescratchingme.Here,thesource(claws)isexplicit,whilethetarget(hands)isimplicit,andthecontextofwashisintermsofthetarget.3.Yourwordscutdeep.Here,thetarget(parole)isexplicit,whilethesource(possibly,knife)isimplicit,andthecontextofcutisintermsofthesource.Inthispaper,wefocusontherecognitionandex-planationofmetaphors.Foragivensentence,wefirstcheckwhetheritcontainsametaphoricexpres-sion(whichwecallmetaphorrecognition),andifitdoes,weidentifythesourceandthetargetcon-ceptsofthemetaphor(whichwecallmetaphorex-planation).Metaphorexplanationisimportantforunderstandingmetaphors.Explainingtype2and3metaphorsisparticularlychallenging,E,tothebestofourknowledge,hasnotbeenattemptedfornominalconcepts1before.Inourexamples,know-ingthatlifeandhandsarethetargetconceptsavoidstheconfusionthatmayariseifsourceconceptssunandclawsareusedliterallyinunderstandingthesen-tences.This,Tuttavia,doesnotmeanthatthesource1Nominalconceptsarethoserepresentedbynounphrases.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
3
5
1
5
6
6
6
8
3
/
/
T
l
UN
C
_
UN
_
0
0
2
3
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
380
conceptisauselessembellishment.Inthe3rdsen-tence,knowingthatwordsismappedtoknifeen-ablesthesystemtounderstandtheemotionorthesentimentembeddedinthetext.Thisisthereasonwhymetaphorrecognitionandexplanationisimpor-tanttoapplicationssuchasaffectionmining(Smithetal.,2007).Itisworthnotingthatsomeprefertoconsidertheverb“cut”,ratherthanthenoun“words”,tobemetaphoricinthe3rdsentenceabove.Weinsteadconcentrateonnominalmetaphorsandseektoex-plainsource-targetmappingsinwhichatleastonedomainisanominalconcept.Thisisbecauseverbsusuallyhavenominalarguments,aseithersubjectorobject,thusexplainingthesource-targetmappingofthenominalargumentcoversmost,ifnotall,caseswhereaverbismetaphoric.Inorderformachinestorecognizeandexplainmetaphors,itmusthaveextensivehumanknowl-edge.Itisnotdifficulttoseewhymetaphorrecog-nitionbasedonsimplecontextmodeling(e.g.,byselectionalrestriction/preference(Resnik,1993))isinsufficient.First,notallexpressionsthatviolatetherestrictionaremetaphors.Forexample,IhatetoreadHeideggerviolatesselectionalrestriction,asthecontext(embodiedbytheverbread)prefersanobjectotherthanaperson(Heidegger).Ma,Heideg-gerisnotametaphorbutametonymy,whichinthiscasedenotesHeidegger’sbooks.Second,noteverymetaphorviolatestherestriction.Forexample,lifeisajourneyisclearlyametaphor,butselectionalre-strictionorpreferenceishelplesswhenitcomestotheisAcontext.Existingapproachesbasedonhuman-curatedknowledgebasesfallshortofthechallenge.First,thescaleofahuman-curatedknowledgebaseisof-tenverylimited,whichmeansatbestitcoversasmallsetofmetaphors.Second,newmetaphorsarecreatedallthetimeandthechallengeistorec-ognizeandunderstandmetaphorsthathaveneverbeenseenbefore.Thisrequiresextensiveknowl-edge.Asaverysimpleexample,evenifthemachineknowsSportscarsarefireenginesisametaphor,itstillneedstoknowwhatisasportscarbeforeitcanunderstandMyFerrariisafireengineisalsoametaphor.Third,existinghuman-curatedknowl-edgebases(includingmetaphordatabasesandtheWordNet)arenotprobabilistic.Theycannottellhowtypicalaninstanceisofacategory(e.g.,arobinisamoretypicalbirdthanapenguin),orhowpopu-laranexpression(e.g.,abreathoffreshair)isusedasasourceconcepttodescribetargetsinanotherconcept(e.g.,younggirls).Unfortunately,withoutnecessaryprobabilisticinformation,notmuchrea-soningcanbeperformedformetaphorexplanation.Inthispaper,weaddresstheabovechallenges.WestartwithaprobabilisticisAknowledgebaseofmanyentitiesandcategoriesharnessedfrombillionsofwebdocumentsusingasetofstrictsyntacticpat-ternsknownastheHearstpatterns(Hearst,1992).WethenautomaticallyacquirealargeprobabilisticmetaphordatabasewiththehelpofbothsyntacticpatternsandtheisAknowledgebase(Section3).Finallywecombinethetwoknowledgebasesandaprobabilisticreasoningmechanismforautomaticmetaphorrecognitionandexplanation(Section4).Thispapermakesthefollowingcontributions:1.Toourknowledge,wearethefirsttointro-ducethemetaphorexplanationproblem,whichseekstorecovermissingorimpliedsourceortargetconceptsinanimplicitmetaphor.2.Thisisthefirstbig-datadriven,unsupervisedapproachformetaphorrecognitionandexpla-nation.Oneofthebenefitsofleveragingbigdataisthattheknowledgeweobtainislessbi-ased,hasgreatcoverage,andcanbeupdatedinatimelymanner.Moreimportantly,adatadrivenapproachcanassociatewitheachpieceofknowledgeprobabilitieswhicharenotavail-ableinhumancuratedknowledgebutareindis-pensableforinferenceandreasoning.3.Ourresultsshowtheeffectivenessbothintermsofcoverageandaccuracyofourapproach.Wemanagetoacquireoneofthelargestmetaphorknowledgebaseseverexistedwithapreci-sionof82%.Themetaphorrecognitionaccu-racysignificantlyoutperformsthestate-of-the-artmethods(Section5).2RelatedWorkExistingworkonmetaphorrecognitionandinterpre-tationcanbedividedintotwocategories:context-orientedandknowledge-driven.Theapproachpro-posedinthispapertouchesonbothcategories.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
3
5
1
5
6
6
6
8
3
/
/
T
l
UN
C
_
UN
_
0
0
2
3
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
381
2.1Context-orientedMethodsSomepreviousworkreliesoncontexttodifferentiatemetaphoricalexpressionsfromliteralones(Wilks,1978;Resnik,1993).Theselectionrestrictionthe-ory(Wilks,1978)arguesthatthemeaningofanex-pressionisrestrictedbyitscontext,andviolationsoftherestrictionimplyametaphor.Resnik(1993)usesKLdivergencetomeasuretheselectionalpreferencestrength(SPS),i.e.,howstronglyacontextrestrictsanexpression.Althoughhedidnotusethismeasuredirectlyformetaphorrecognition,SPS(andalsoarelatedmeasurecalledtheselectionassociation)iswidelyusedinmorere-centapproachesformetaphorrecognitionandinter-pretation(Mason,2004;Shutova,2010;Shutovaetal.,2010;Baumeretal.,2010).Forexample,Ma-son(2004)learnsdomain-specificselectionalprefer-encesandusethemtofindmappingsbetweencon-ceptsfromdifferentdomains.Shutova(2010)de-finesmetaphorinterpretationasaparaphrasingtask.Themethoddiscriminatesbetweenliteralandfig-urativeparaphrasesbydetectingselectionalprefer-enceviolation.TheresultofthisworkhasbeencomparedwithourapproachinSection5.Shutovaetal.(2010)identifyconceptsinasourcedomainofametaphorbyclusteringverbphrasesandfilter-ingoutverbsthathaveweakselectionalpreferencestrength.Baumer(2010)usessemanticrolelabelingtechniquestocalculateselectionalpreferenceonse-manticrelationsinsteadofgrammaticrelationsformetaphorrecognition.Alessrelatedbutalsocontext-basedworkisanalogyinterpretationbyrelationmapping(Turney,2008).Theproblemistogeneratemappingbetweensourceandtargetdomainsbycomputingpair-wiseco-occurrencesfordifferentcontextualpatterns.Ourapproachusesselectionalrestrictionwhenenrichingthemetaphorknowledgebase,andadoptscontextpreferencewhenexplainingtype2and3metaphorsbyfocusingonthenearbyverbsofapo-tentialsourceortargetconcept.2.2Knowledge-drivenMethodsAgrowingnumberofworksuseknowledgebasesformetaphorunderstanding(Martin,1990;Narayanan,1997;Barndenetal.,2002;VealeandHao,2008).MIDAS(Martin,1990)checksifasen-tencecontainsanexpressionthatcanbeexplainedbyamoregeneralmetaphorinahuman-curatedmetaphorknowledgebase.ATT-Meta(Barndenetal.,2002)performsmetaphorreasoningwithahuman-curatedmetaphorknowledgebaseandfirstorderlogic,anditfocusesonaffectiondetection(Smithetal.,2007;Agerri,2008;Zhang,2010).Kr-ishnakumaranandZhu(2007)usetheisArelationinWordNet(Mugnaio,1995)formetaphorrecognition.Gedigianetal.(2006)useFrameNet(Fillmoreetal.,2003)andProbank(KingsburyandPalmer,2002)totrainamaximumentropyclassifierformetaphorrecognition.TroFi(BirkeandSarkar,2006)rede-finesliteralandnon-literalastwosensesofthesameverbandprovidetwosenseswithseedsentencesfromhuman-curatedknowledgebaseslikeWord-Net,knownmetaphorandidiomsets.Foragivensentencecontainingtargetverb,itcomparesthesim-ilarityofthesentencewithtwoseedsetsrespec-tively.Ifthesentenceisclosertothenon-literalsenseset,theverbisrecognizedasnon-literalusage.Whiletheaboveworkallreliesonhumancu-rateddatasetsormanuallabeling,VealeandHao(2008)introducedthenotionoftalkingpointswhicharefigurativepropertiesofnoun-basedconcepts.Forexample,theconcept“Hamas”hasthefollow-ingtalkingpoints:isislamic:movementandgov-erns:gazastrip.TheyautomaticallyconstructedaknowledgebasecalledSlipNetfromWordNetandWebcorpus.ConceptsthatareconnectedontheSlipNetcan“slip”tooneanotherandarehenceconsideredrelatedinametaphor.However,straight-forwardtraversalontheSlipNetcanbecomecom-putationallyimpracticalandtheauthorsdidnotelab-orateontheimplementationdetails.Inpractice,theknowledgeacquiredinthispaperismuchlargerbutouralgorithmsarecomputationallymorefeasible.3ObtainingProbabilisticKnowledgeInthissection,wedescribehowtousealarge,general-purpose,probabilisticisAknowledgebaseΓHtocreateaprobabilisticmetaphordatasetΓm.ΓHcontainsisApairsaswellasscoresassociatedwitheachpair.ThemetaphordatasetΓmcontainsmetaphorsoftheform:(source,target),andaweightfunctionPmthatmapsametaphorpairtoaprobabilisticscore.ThepurposeofcreatingΓHis
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
3
5
1
5
6
6
6
8
3
/
/
T
l
UN
C
_
UN
_
0
0
2
3
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
382
tohelpcleanandexpandΓm,andtoperformproba-bilisticinferenceformetaphordetection.3.1IsAKnowledgeΓHΓH,ageneral-purpose,probabilisticisAknowl-edgebase,waspreviouslyconstructedbyWuetal.(2012).2ΓHcontainsisArelationsintheformof(X,hx),apairofhyponymandhypernym,forexam-ple,(SteveBallmer,CEOofITcompanies),andeachpairisassociatedwithasetofprobabilisticscores.Twoofthemostimportantscoresareknownastyp-icality:P(X|hx),thetypicalityofxofcategoryhx,andP(hx|X),thetypicalityofcategoryhxforin-stancex,whichwillbeusedinmetaphorrecogni-tionandexplanation.Bothscoresareapproximatedbyfrequencies,e.g.,P(X|hx)=#of(X,hx)inHearstextraction#ofhxinHearstextractionIntotal,ΓHcontains16millionuniqueisArela-tionships,and2.7millionuniqueconceptsorcate-gories(thehx’sin(X,hx)pairs).Theimportanceofbigdataisobvious.ΓHcontainsmillionsofcat-egoriesandprobabilisticscoresforeachcategorywhichenablesinferenceformetaphorunderstand-ing,aswewillshownext.3.2AcquiringMetaphorsΓmWeacquireaninitialsetofmetaphorsΓmfromsim-iles.Asimileisafigureofspeechthatexplicitlycomparestwodifferentthingsusingwordssuchas“like”and“as”.Forexample,thesentenceLifeislikeajourneyisasimile.Withouttheword“like,”itbecomesametaphor:Lifeisajourney.Thispropertymakessimileanattractivefirsttargetformetaphorextractionfromalargecorpus.Weusethefollowingsyntacticpatternforextraction:htargetiBE/VBlike[UN]hsourcei(1)whereBEdenotesis/are/hasbeen/havebeen,etc.,VBdenotesverbotherthanBE,andhtargetiandhsourceidenotenounphrasesorverbphrases.Notethatnoteveryextractedpairisametaphor.Poetryislikeanartmatchesthepattern,butitisnotametaphorbecausepoetryisreallyanart.WewilluseΓHtocleansuchpairs.Furthermore,duetothe2Datasetcanbefoundathttp://probase.msra.cn/.idiosyncrasiesofnaturallanguages,itisnottrivialtocorrectlyextractthehtargetiandthehsourceifromeachsentencethatmatchesthepattern.Weuseapostaggerandalemmatizeronthesentences,andwedeveloparule-basedsystemthatcontainsmorethantwodozenrulesforextraction.Forexample,aruleofhigh-precisionbutlow-recallis“htargetimustbeatthebeginningofasentenceorthebeginningofaclause(e.g.,followingthewordthat)”.Finally,from8,552,672sentencesthatmatchtheabovepattern(pattern1),weobtain1.2millionunique(X,sì)pairs,andafterfiltering,weareleftwithcloseto1millionuniquemetaphorpairs,whichformthestartingpointofΓm.3.3Cleaning,Expanding,andWeightingΓmThesimilepatternonlyallowsustoextractsomeoftheavailablemetaphorpairs.ToexpandΓm,weuseamoreflexiblebutalsonoisierpatterntoextractmorecandidatemetaphorpairsfrombillionsofsen-tencesinthewebcorpus:htargetiBE[UN]hsourcei(2)Theabove“isa”patterncoversmetaphorssuchasLifeisajourney.Butmanypairsthusextractedarenotmetaphors,forexample,Malaysiaisatropicalcountry.Thatis,pairsextractedbythe“isa”pat-terncontainsatleasttwotypesofrelations:thelit-eralisArelationsandthemetaphorrelations.Theproblemishowtodistinguishonefromtheother.Intheory,thesetofallIsArelations,IO,andthesetofallmetaphorrelations,M,donotoverlap,becausebydefinition,thesourceconceptandthetargetcon-ceptinametaphorarenotthesamething.Thus,ourintuitionisthefollowing.Thepairsproducedbythesimilepattern,calledS,isasubsetofM,whilethepairsextractedfromtheHearstpattern,calledH,isalsoasubsetofI.SinceMandIhardlyoverlap,SandHshouldhavelittleoverlap,too.Inpractice,veryfewpeoplewouldsaysomethinglikejourneyssuchaslife.Figure1illustratesthisscenario.Toverifythisintuition,werandomlysampled1,000sentencesandmanuallyannotatedthem.Ofthesesentences,40containanIsArelation,ofwhich27areenclosedinaHearst’spatternand13canbeextractedbythe“isa”pattern.Furthermore,28ofthese1000sentencescontainametaphorexpression,
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
3
5
1
5
6
6
6
8
3
/
/
T
l
UN
C
_
UN
_
0
0
2
3
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
383
(beast, sports car)(sports car, ferrari)(vehicle, ferrari)(beast, ferrari)Hearst patternIs-a relationSimile patternMetaphor relation“Is a” patternFigure1:Relationsamongdifferentsets.Dottedcirclesrepresentrelations(groundtruth).Solidcirclesrepresentpairsextractedbysyntacticpatterns.andwithinthe28metaphors,15areembeddedinasimilepattern.Moreimportantly,thereisnooverlapbetweentheIsArelationsandmetaphors(andhencethesimiles).Inalargerscaleexperiment,wecrawled1billionsentenceswhichmatchthe“isa”pattern(2)fromthewebcorpus.Fromthese,weextracted180millionunique(X,sì)pairs.24.8%ofΓHcanbefoundin“isa”patternpairs,while16.8%ofΓmcanbefoundin“isa”patternpairs.Furthermore,thereisalmostnooverlapbetweenΓHandΓm:1.26%ofΓHcanbefoundinΓm,and1.31%ofΓmcanbefoundinΓH.OurgoalistousetheinformationcollectedthroughthesyntacticpatternstoenrichthemetaphorrelationsorΓm.Armedwiththeaboveobservations,wemaketwoconclusions.First,IL(life,journey)pairweextractedfromlifeisajourneyismorelikelyametaphorsinceitdoesnotappearinthesetex-tractedfromHearstpatterns.Second,ifanyexistingpairinΓmalsoappearsinΓH,wecanremovethatpairfromΓm.Fromthe180millionunique(X,sì)pairsweex-tractedearlier,byfilteringoutlowfrequencypairs3andthosepairsinΓH,weobtain2.6millionoffreshmetaphors.Thisisalmost3timeslargerthaninitialmetaphorsetobtainedfromthesimilepattern.WefurtherexpandΓmbyaddingmetaphorsderivedfromΓmandΓH.Assume(X,sì)∈Γm,E(X,hx)∈ΓH,thenweadd(hx,sì)toΓm.Asanexample,if(Julie,sun)∈Γm,3Specifically,werandomlysamplepairsoffrequency1,2,…,10fromΓmandchecktheprecisionsofeachgroup.Wefilteroutpairswithfrequencylessthan5tooptimizetheprecision.thenweadd(personname,sun)toΓm,since(Julie,personname)∈ΓH.ThisenablesthemetaphordetectionapproachwedescribeinSection4.NotethatweignoretransitivityintheisarelationsfromΓHassuchtransitivityisnotalwaysreliable.Forexample,carseatisachair,andchairisfurni-ture,butcarseatisnotfurniture.HowtohandletransitivityinadatadrivenisAtaxonomyisachal-lengingproblem,andisbeyondthescopehere.Finally,wecalculatetheweightofeachmetaphor(X,sì).TheweightPm(X,sì)iscalculatedasfollows:Pm(X,sì)=occurrencesof(X,sì)inisApatternoccurrencesofisApattern(3)Theweightsofderivedmetaphors,suchas(personname,sun),arecalculatedasfollows:Pm(hx,sì)=X(X,hx)∈ΓHPm(X,sì)(4)4ProbabilisticMetaphorUnderstandingInthispaper,weconsidertwoaspectsofmetaphorunderstanding,metaphorrecognitionandmetaphorexplanation.Thelatterisneededfortype2and3metaphorswhereeitherthesourceorthetargetcon-ceptisimplicitormissing.Next,wedescribeaprob-abilisticapproachtoaccomplishthesetwotasks.4.1Type1MetaphorsInatype1metaphor,boththesourceandthetar-getconceptsappearexplicitly.Whenasentencematches“isa”pattern(pattern2),itisapotentialmetaphorexpression.Thefirstnouninthepatternisthetargetcandidate,whilethesecondnounisthesourcecandidate.Torecognizetype1metaphors,wefirstobtainthecandidate(source,target)pairfromthesentence.Then,wecheckifwehaveanyknowledgeaboutthe(source,target)pair.Intuitively,ifthepairexistsinthemetaphordatasetΓm,thenitisametaphor.Ifthepairex-istsintheis-AknowledgebaseΓH,thenitisnotametaphor.ButbecauseΓmisfarfrombeingcom-plete,ifapairexistsinneitherΓmnorΓH,thereisapossibilitythatitisametaphorwehaveneverseenbefore.Inthiscase,wereasonasfollows.ConsiderasentencesuchasMyFerrariisabeast.Assume(Ferrari,beast)6∈Γm,Ma(sportscar,
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
3
5
1
5
6
6
6
8
3
/
/
T
l
UN
C
_
UN
_
0
0
2
3
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
384
beast)∈Γm.Notethat(sportscar,beast)mayit-selfbeaderivedmetaphorwhichisaddedintoΓminmetaphorexpansion,andtheoriginalmetaphorex-tractedfromthewebdatais(Lamborghinis,beast).Inoltre,fromΓH,weknowFerrariisasportscar,thatis,(Ferrari,sportscar)∈ΓH,wecantheninferthatFerraritobeastisverylikelyametaphormapping.Specifically,let(X,sì)beapairweareconcernedwith.Wewanttocomputetheoddsof(X,sì)repre-sentingametaphorvs.anormalis-Arelationship:P(X,sì)1−P(X,sì)(5)whereP(X,sì)istheprobabilitythat(X,sì)formsametaphor.Now,combiningtheknowledgewehaveinΓH,wehaveP(X,sì)=X(X,hx)∈ΓHP(X,hx,sì)(6)Here,hxisapossiblesuperconcept,i.e.,apossibleinterpretation,forx.Forexample,ifx=apple,thentwohighlypossibleinterpretationsarecom-panyandfruit.InEq.(6),wewanttoaggregateonallpossibleinterpretations(allsuperconcepts)ofx.ThisispossiblebecauseofthemassivesizeoftheconceptspaceinΓH.WecanrewriteEq.(6)tothefollowing:P(X,sì)=X(X,hx)∈ΓHP(sì|X,hx)P(X|hx)P(hx)(7)Here,P(sì|X,hx)meanswhenxisinterpretedasanhx,theprobabilityofyasatargetmetaphoricalcon-ceptforhx.Thus,givenhx,yisindependentwithx,soP(sì|X,hx)canbesimplyreplacedbyP(sì|hx).WecanthenrewriteEq.(7)A:P(X,sì)=X(X,hx)∈ΓHP(sì|hx)P(X|hx)P(hx)=X(X,hx)∈ΓHP(hx,sì)P(X|hx)(8)ItisclearP(hx,sì)issimplyPm(hx,sì)inEq.(4)givenbythemetaphordatasetΓm.Furthermore,P(X|hx)isthetypicalityofxinthehxcategory,andP(hx)isthepriorofthecategoryhx.BothofthemareavailablefromtheisAknowledgebaseΓH.Thus,wecancalculateEq.(8)usinginformationinthetwoknowledgebaseswehavecreated.IftheoddsinEq.(5)isgreaterthanathresh-oldδ,whichisdeterminedempiricallytobeδ=P(metaphor)P(isa)4,wedeclare(X,sì)asametaphor.4.2ContextPreferenceModelingItismoredifficulttorecognizemetaphorswhenthesourceconceptorthetargetconceptisnotexplic-itlygiveninasentence.Inthiscase,werelyonthecontextinthesentence.Givenasentence,wefindmetaphorcandidatesandthecontext.Here,candidatesarenounphrasesinthesentencewhichcanpotentiallybethetargetorthesourceconceptofametaphor,whilecontextdenoteswordsthathaveagrammaticdependencywiththecandidate.Thedependencycanbesubject-predicate,predicate-object,ormodifier-header,etc.Thecontextcanbeaverb,anounphrase,oranad-jectivewhichhascertainpreferenceoverthetargetorsourcecandidate.Forexample,thewordhorseprefersverbssuchasjump,drinkandeat;thewordflowerprefersmodifierssuchasred,yellowandbeautiful.Inthiswork,wefocusonanalyzingtheprefer-encesofverbsusingsubject-predicateorpredicate-objectrelationbetweentheverbandthenounphrases.Weselect2,226mostfrequentverbsfromthewebcorpus.Foreachverb,weconstructthedis-tributionofnounphrasesdependontheverbinthesentencessampledfromthewebcorpus.ThenounphrasesarerestrictedtobethosethatoccurinΓH.Morespecifically,foranynounphraseythatap-pearsinΓH,wecalculatethefollowingPr(C|sì)=fr(sì,C)PCfr(sì,C)(9)wherefr(sì,C)meansthefrequencyofyandcon-textCwithrelationr.Notewecanbuildprefer-encedistributionforcontextotherthanverbssince,intheory,rcanbeanyrelation(e.g.modifier-headrelation).4.3Type2andType3MetaphorsIfasentencecontainstype2andtype3metaphors,eitherthesourceorthetargetconceptsinthesen-4Thisistheratiobetweenthenumberofmetaphorsandis-apairsinarandomsampleof“isa”patternsentences.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
3
5
1
5
6
6
6
8
3
/
/
T
l
UN
C
_
UN
_
0
0
2
3
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
385
tenceismissing.Foreachnounphrasexandacon-textCinsuchasentence,wewanttoknowwhetherxisofliteralormetaphoricuse.Itisametaphoricuseiftheselectionalpreferenceofsomey,whichisasourceortargetconceptofxinΓm,islargerthantheselectionalpreferenceofanysuper-conceptofxinΓH,byafactorδ.Formally,thereexistsaywhere(X,sì)∈Γmor(sì,X)∈Γm,suchthatP(sì|X,C)P(H|X,C)≥δ,∀(X,H)∈ΓH.(10)Tocompute(10),wehaveP(sì|X,C)=P(X,sì,C)P(X,C)=P(X,sì)P(C|X,sì)P(X,C)(11)Assumingxisatargetconceptandyisasourceconcept(aType3metaphor),wecanobtainP(X,sì)byEq.(8).5Inoltre,Cisindependentofxinatype2or3metaphor,sinceametaphorisanunusualuseofx(thetarget)withinagivencontext.ThereforeP(C|X,sì)=P(C|sì),whereP(C|sì)isavailablefromEq.(9).Allo stesso modo,wehaveP(H|X,C)=P(X,H)P(C|H)P(X,C)(12)whereP(X,H)isobtainedfromΓHandP(C|H)isfromthecontextpreferencedistribution.Toexplainthemetaphor,oruncoverthemissingconcept,y∗=argmaxy∧(sì,X)∈ΓmP(sì|X,C)=argmaxy∧(sì,X)∈ΓmP(sì,X)P(C|sì)Asaconcreteexample,considersentenceMycardrinksgasoline.Therearetwopossibletargets:carandgasoline.Thecontextforbothtargetsistheverbdrink.Letx=car.ByEq.(11),wefirstfindally’sforwhich(car,sì)∈Γmor(sì,car)∈Γm.Wegettermssuchaswoman,friend,gun,horse,etc.WhenwecalculateP(car,sì)byEq.(8),wealsoneedtofindhypernymsofcarinΓH,which5Type2metaphorscanbehandledsimilarly.mayincludevehicle,Prodotto,asset,etc.Foreachcandidatey,P(sì|car,C)iscalculatedbymetaphorknowledgeP(X,sì)andcontextpreferenceP(C|yi).Table1showstheresult.Sincetheselectionalpref-erenceofhorse(fromΓm)ismuchlargerthanotherliteralusesofcar,thissentenceisrecognizedasametaphor,andthemissingsourceconceptishorse.Table1:Logprobabilities(M:Metaphor,l:Literal).TypeyilogloglogP(yiP(yi,car)P(C|yi)|car,C)Lvehicle-6.2-∞-∞Lproduct-6.9-∞-∞Lasset-6.3-∞-∞Mwoman-8.5-2.8-11.3Mfriend-8.0-3.0-11.0Mgun-8.4-∞-∞Mhorse-8.2-2.4-10.6……………5ExperimentalResultWeevaluatetheperformanceofmetaphoracquisi-tion,recognitionandexplanationinoursystemandcompareitwithseveralstate-of-the-artmechanisms.5.1MetaphorAcquisitionFromthewebcorpus,wecollected8,552,672sen-tencesmatchingthe“islikea”pattern(pattern1)andweextracted932,621uniquehighqualitysim-ilemappingsfromthem.ThesesimilemappingsbecamethecoreofΓm.ΓHcontains16,736,068uniqueisApairs.Wealsocollected1,131,805,382sentencesmatchingthe“isa”pattern(pattern2),fromwhich180,446,190uniquemappingswereex-tracted.ThesemappingscontainbothmetaphorsandisArelations.Fromthere,weidentified2,663,127pairsofmetaphorsunseeninthesim-ileset.ThesenewmetaphorpairswereaddedtoΓm.Randomsamplesshowthattheprecisionsofthecoremetaphordatasetandthewholedatasetare93.5%and82%,respectively.Alloftheabovedatasets,asampleofcontextpreference,aswellasthetestsetsmentionedinthissectioncanbefoundathttp://adapt.seiee.sjtu.edu.cn/˜kzhu/metaphor.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
3
5
1
5
6
6
6
8
3
/
/
T
l
UN
C
_
UN
_
0
0
2
3
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
386
5.2Type1MetaphorRecognitionWecompareourtype1metaphorrecognitionwiththemethod(knownasKZ)byKrishnakumaranandZhu(2007).Forsentencescontaining“xisay”pat-tern,KZusedWordNettodetectwhetheryisahy-pernymofx.Ifnot,thenthissentenceisconsideredametaphor.Ourtestsetis200randomsentencesthatmatchthe“xBEay”pattern.Welabelasen-tenceinthesetasametaphorifthetwonounscon-nectedbyBEdonotactuallyhaveisArelation;oriftheydohaveisArelationbutthesentenceexpressedastrongemotion6.Table2:Type1metaphorrecognitionPrecisionRecallF1KZ13%30%18%OurApproach73%66%69%TheresultissummarizedinTable2.KZdoesnotperformaswellduetothesmallcoverageofWord-Nettaxonomy.Only33outof200sentencescon-tainaconceptxthatexistsinWordNetandhasatleastonehypernym.Andamongthese,only2sen-tencescontainaywhichisthehypernymancestorofxinWordNet.Clearly,thebottleneckisthescaleofWordNet.5.3Type2/3MetaphorRecognitionFortype2/3metaphorrecognition,wecompareourresultswiththreeothermethods.Thefirstcompet-ingmethod(calledSA)employstheselectionalas-sociationproposedbyResnik(1993).Selectionalassociationmeasuresthestrengthoftheconnectionbetweenapredicate(C)andaterm(e)by:UN(C,e)=Pr(e|C)logPr(e|C)Pr(e)S(C),(13)whereS(C)=KL(Pr(e|C)||Pr(e))=XePr(e|C)logPr(e|C)Pr(e)GivenanNP-predicatepair,ifitsSAscoreislessthanathresholdα(setto10−4byempirics),thenthepairisrecognizedasametaphorcontext.6Forexample,“thismanisananimal!”.Secondcompetingmethod(calledCP)isthecon-textualpreferenceapproach(Resnik,1993)intro-ducedinSection4.2.Toestablishcontextprefer-encedistributions,werandomlyselect100millionsentencesfromthewebcorpus,parseeachsentenceusingStanfordparser(Group,2013)toobtainallsubject-predicate-objecttriples,andaggregatethetriplestoget33,236,292subject-predicatepairsand38,890,877predicate-objectpairs.Theoccurrencesofthesepairsareusedascontextpreference.GivenapairofNP-predicatepair,ifitscontextpreferencescoreislessthanathresholdβ(setto10−5byem-pirics7),thenthepairisconsideredasmetaphoric.Thethirdcompetingmethod(calledVH)isavari-antofourownalgorithmwithΓmreplacedbyametaphordatabasederivedfromtheSlipNetpro-posedbyVealeandHao(2008),whichwecallΓVH.WebuiltaSlipNetcontaining21,451conceptnodesassociatedwith27,533distincttalkingpoints.Weconsidertwoconceptstobemetaphoriciftheyareatmost5hopsapartontheSlipNetThechoiceof5hopsisatrade-offbetweenprecisionandrecallforSlipNet.WethuscreatedΓVHwith5,633,760pairsofconcepts.Wesampled1,000sentencesfromtheBNCdataset(Clear,1993)asfollows.Wepreparealistof2,945frequentverbs(andtheirdifferentforms).Foreachverb,weobtainatmost5sentencesfromBNCdatasetwhichcontainthisverbasapredicate.Atthispoint,weobtainatotalof22,601sentencesandrandomlysample1,000sentencestoformatestset.Eachsentenceinthesetisthenmanuallyla-beledasbeing“metaphor”or“non-metaphor”.Welabelthemaccordingtothisprocedure:1.foreachverb,wecollecttheintendeduse,i.e.,thecategoriesofitsarguments(subjectorob-ject)accordingtoMarriamWebster’sdictio-nary;2.iftheargumentoftheverbinthesentencebe-longstotheintendedcategory,thesentenceislabeled“non-metaphor”;3.iftheargumentandtheintendedmeaningformametonymywhichusesapartoranattributeto7Theauthorsdidn’tspecifythechoiceofαandβ,andwepickvalueswhichoptimizetheperformanceoftheiralgorithms.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
3
5
1
5
6
6
6
8
3
/
/
T
l
UN
C
_
UN
_
0
0
2
3
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
387
representthewholeobject,thepairislabeledas“non-metaphor”;4.elsethesentenceislabeledas“metaphor”.Table3:Type2/3metaphorrecognitionPrecisionRecallF1SA23%20%21%CP50%20%26%VH11%86%20%OurApproach65%52%58%Theresultsfortype2and3metaphorrecogni-tionareshowninTable3.Ourknowledge-basedap-proachsignificantlyoutperformstheotherpeersbyF-1measure.AlthoughVHachievesagoodrecall,itsprecisionispoor.Thisisbecausei)SlipNetcon-structionmakesheavyuseofsiblingtermsontheWordNetbutsiblingtermsarenotnecessarilysimi-larterms;ii)manypairsgeneratedbyslippingovertheSlipNetareintheoryrelatedbutarenotcom-monlyutteredduetothelackofpracticalcontext.0%10%20%30%40%50%60%70%80%SPS (cid:1488)(2,3]SPS (cid:1488)(3,4]SPS (cid:1488)(4,5]F1 scoreSPS of verbsSACPVHOur approachFigure2:Metaphorrecognitionoftype2and3Fig.2comparesthefourmethodsonverbswithdifferentselectionalpreferencestrength,whichindi-cateshowstrongaverb’sargumentsarerestrictedtoacertainscopeofnouns.8Again,ourmethodshowsasignificantadvantageacrosstheboard.Weexplainwhyourapproachworksbetterus-ingtheexamplesinTable4.InsentenceAAU200,shattersisametaphoricusagebecausesilenceisnotathingthatcanbebrokenintopieces.SAandCPscoresforshatters-silencepairarehighbecausethiswordcombinationisquitecommon,8NotethatnoverbhasSPSlargerthan5.andhencethesemethodsincorrectlytreatitaslit-eralexpression.Thesituationissimilarwithstalk-companypairinABG2327.Ontheotherhand,forAN81309,manipulate-lifeisconsideredrarecom-binationandhencehaslowSAandCPscoresandisdeemedametaphorwhileinrealityitisaliteraluse.Asimilarcaseoccursforwork-concurpair.Inallthesecases,ourknowledgebasesΓmandΓHarecomprehensiveandaccurateenoughtocorrectlyidentifymetaphorsvs.non-metaphors.Onthecon-trary,themetaphordatabaseΓVHcoverswaytoomanypairsthatittreatseverypairasametaphor.Besidesourowndataset,wealsoexperimentonTroFiExampleBase9,whichconsistsof50verbsand3,736sentencescontainingtheseverbs.Eachsentenceisannotatedasliteralandnonliteraluseoftheverb.Ouralgorithmisusedtoclassifythesub-jectsandtheobjectsoftheverbs.WeuseStanforddependencyparsertoobtaincollapsedtypeddepen-denciesofthesesentences,andforeachsentence,runouralgorithmtoclassifythesubjectsandobjectsrelatedtotheverb,iftheverbactsasapredicate.Resultsshowthatourapproachachieves77.5%pre-cisionbutjustunder5%inrecall.Thelowrecallisbecause,io)non-literalusesintheTroFidatasetin-cludenotonlymetaphorbutalsometonymy,ironyandotheranomalies;ii)ourapproachcurrentlyfo-cusesonsubject-predicateandpredicate-objectde-pendenciesinasentenceonly,butthetargetverbsdonotactaspredicateinmanyoftheexamplesen-tences;iii)theStanforddependencyparserisnotro-bustenoughsohalfofthesentencesarenotparsedcorrectly.5.4MetaphorExplanationInthisexperiment,weusetheclassiclabeledmetaphoricsentencesfrom(LakoffandJohnson,1980).Lakoffprovided24metaphoricmappings,andforeachmappingthereareabouttenexamplesentences.Intotal,thereare214metaphoricsen-tences.Amongthem,wefocuson83sentenceswhosemetaphorisexpressedbysubject-predicateorpredicate-objectrelation,asthispaperfocusesonverbcentriccontextpreferences.Weevaluatetheresultsofcompetingalgorithms9TroFiExampleBaseisavailableathttp://www.cs.sfu.ca/˜anoop/students/jbirke/.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
3
5
1
5
6
6
6
8
3
/
/
T
l
UN
C
_
UN
_
0
0
2
3
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
388
Table4:MetaphorrecognitionforsomeexamplesentencesfromBNCdataset(HM:Umano,M:Metaphor,l:Literal).IDSentenceHMSACPVHOursAAU200Road-blocksalvoshattersBucharest’sfragilesilence.MLLMMABG2327Obstructionandprotectionismdonotstalkonlybigcompanies.MLLMMAN81309Butwhenscienceproposestomanipulatethelifeofahumanbaby,LMMMLACH1075Nevertheless,recentworkonMosleyandtheBUFhasconcurredLMMMLabouttheirbasicunimportance.bythefollowinglabelingcriteria.Weconsideranoutput(i.e.apairofconceptmapping)asamatch,iftheproducedpairexactlymatchesthegroundtruthpair,ofifthepairissubsumedbythegroundtruthpair.Forexample,thegroundtruthforthesentenceLetthatideasimmeronthebackburnerisideas→foodsaccordingtoLakoff(LakoffandJohnson,1980).Ifouralgorithmoutputsidea→stew,thenitisconsideredamatchsincestewbelongstothefoodcategory.Anoutputpairisconsideredcorrectifitisnotamatchtothegroundtruthbutisotherwiseconsideredmetaphoricbyatleast2ofthe3humanjudges.Givenasentence,sinceouralgorithmreturnsalistofpossibleexplanationsforthemissingconcept,rankedbytheprobability,weevaluatetheresultsbythreedifferentmetrics:MatchTop1:resultconsideredcorrectifthereisamatchwiththetopexplanation;MatchTop3:resultconsideredcorrectifthereisamatchinthetop3rankedexplanations;CorrectTop3:resultconsideredcorrectifthereisacorrectinthetop3explanations.Table5:PrecisionofmetaphorexplanationusingdifferenmetaphordatabasesMatchTop1MatchTop3CorrectTop3ΓVH26%49%54%Γm43%67%78%ComparisonwithSlipNetWecomparetheresultofouralgorithm(fromSection4.3)againstthevariantwhichusesΓVHob-tainedinSection5.3.Table5summarizestheprecisionsofthetwoal-gorithmsunderthreedifferentmetrics.SomeofthesesentencesandthetopexplanationsgivenbyouralgorithmarelistedinTable6.Theconcepttobeexplainedisitalicizedwhiletheexplanationthatisamatchorcorrectisboldedorbold-italicized,re-spectively.Theexplanationsareorderedfromlefttorightbythescore.ComparisonwithparaphrasingWhilewedefinemetaphorexplanationasatasktorecoverthemissingnoun-basedconceptinasource-targetmapping,analternativewaytoexplainametaphor(Shutova,2010)istofindtheparaphraseoftheverbinthemetaphor.Hereweevaluatepara-phrasingtaskonverbsinmetaphoricsentencebyShutovaetal(Shutova,2010).ForametaphoricverbVinasentence,Shutovaetal.selectasetofverbsthatprobabilisticallybestmatchesgrammarrelationsofV,andthenfilteroutthoseverbsthatarenotrelatedtoVaccordingtotheWordNet,andeventuallyre-rankremainingverbsbasedonselec-tionassociation.Insomesense,Shutova’sworkusesasimilarframeworkasours:firstrestrictthetargetpara-phrasingsetusingaknowledge,thenselectthemostproperwordbasedonthecontext.Thedifferenceisthatthetargetof(Shutova,2010)istheverbinsentence,whileourapproachfocusesonthenoun.ToimplementalgorithmbyShutova,weextractandcounteachgrammarrelationin1billionsen-tences.Thesecountsareusedtocalculatecon-textmatchingin(Shutova,2010),andarealsousedtocalculateselectionassociation.WeperformShutova’sparaphrasingonverbsin83sentences,ofwhichonly25findsagoodparaphrasesinShutova’stop3results.Afterremoving17sentenceswhichcontainlightverbs(e.g.,take,give,put),thealgo-
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
3
5
1
5
6
6
6
8
3
/
/
T
l
UN
C
_
UN
_
0
0
2
3
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
389
Table6:MetaphorsentencesexplainedbythesystemMetaphormappingSentenceExplanationIdeasarefoodLetthatideasimmeronthebackburner.stew;carrot;onionWedon’tneedtospoon-feedourstudentseggroll;acorn;wordwithknowledge.EyesarecontainersHiseyesdisplayedhiscompassion.window;symbol;tinycameraHiseyeswerefilledwithanger.hollowball;waterballoon;balloonEmotionaleffectisHismother’sdeathhithimhard.enemy;monsterphysicalcontactThatideabowledmeover.punch;stew;onionLifeisacontainer.Herlifeiscrammedwithactivities.tapestry;beach;danceGetthemostoutoflife.game;journey;prisonrithmfinds21goodparaphrasesintop3results.OnereasonforthelowrecallisthatWordnetisin-adequateinprovidingcandidatemetaphormapping.Thisisalsothereasonwhyourmetaphorbaseisbetterthanthemetaphorbasegeneratedbytalkingpoints.6ConclusionKnowledgeisessentialforamachinetoidentifyandunderstandmetaphorsInthispaper,weshowhowtomakeuseoftwoprobabilisticknowledgebasesau-tomaticallyacquiredfrombillionsofwebpagesforthispurpose.Thisworkcurrentlyrecognizesandex-plainsmetaphoricmappingsbetweennominalcon-ceptswiththehelpofselectionalpreferenceofjustsubject-predicateorpredicate-objectcontexts.Animmediatenextstepistoextendthisframeworktomoregeneralcontextsandafurtherimprovementwillbetoidentifymappingsbetweenanysourceandtargetdomains.7AcknowledgementsKennyQ.ZhuwaspartiallysupportedbyGoogleFacultyResearchAward,andNSFCGrants61100050,61033002and61373031.ReferencesRodrigoAgerri.2008.Metaphorintextualentailment.InCOLING(Posters),pages3–6.JohnBarnden,SheilaGlasbey,MarkLee,andAlanWallington.2002.Reasoninginmetaphorunder-standing:theatt-metaapproachandsystem.InCOL-ING’02,pages1–5.EricP.S.Baumer,JamesP.White,andBillTomlinson.2010.Comparingsemanticrolelabelingwithtypeddependencyparsingincomputationalmetaphoriden-tification.InCALC’10,pages14–22.JuliaBirkeandAnoopSarkar.2006.Aclusteringap-proachfornearlyunsupervisedrecognitionofnonlit-erallanguage.InInProceedingsofEACL-06,pages329–336.JeremyH.Clear.1993.Thedigitalword.chapterTheBritishnationalcorpus,pages163–187.CharlesJ.Fillmore,ChristopherR.Johnson,andMiriamR.L.Petruck.2003.BackgroundtoFrameNet.InternationalJournalofLexicography,16.3:235–250.JamesGeary.2011.IisanOther:TheSecretLifeofMetaphorandHowItShapestheWayWeSeetheWorld.Harper.MattGedigian,JohnBryant,SriniNarayanan,andBra-nimirCiric.2006.Catchingmetaphors.InInWork-shopOnScalableNaturalLanguageUnderstanding.StanfordNLPGroup.2013.TheStanfordparser.http://nlp.stanford.edu/software/lex-parser.shtml.MartiA.Hearst.1992.Automaticacquisitionofhy-ponymsfromlargetextcorpora.InCOLING’92,pages539–545.PaulKingsburyandMarthaPalmer.2002.Fromtree-banktopropbank.InInLanguageResourcesandEvaluation.SaisureshKrishnakumaranandXiaojinZhu.2007.Huntingelusivemetaphorsusinglexicalresources.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
2
3
5
1
5
6
6
6
8
3
/
/
T
l
UN
C
_
UN
_
0
0
2
3
5
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3
390
InProceedingsoftheWorkshoponComputationalApproachestoFigurativeLanguage,pages13–20,Rochester,NewYork,April.ACL.GeorgeLakoffandMarkJohnson.1980.MetaphorsWeLiveBy.UniversityofChicagoPress,Chicago,USA.J.H.Martin.1990.AComputationalModelofMetaphorInterpretation.AcademicPressProfessional,Inc.ZacharyJ.Mason.2004.Cormet:acomputational,corpus-basedconventionalmetaphorextractionsys-tem.Comput.Linguist.,30:23–44,March.GeorgeA.Miller.1995.Wordnet:alexicaldatabaseforenglish.Commun.ACM,38:39–41,November.SrinivasSankaraNarayanan.1997.Knowledge-basedactionrepresentationsformetaphorandaspect(karma).Technicalreport.PhilipStuartResnik.1993.Selectionandinformation:aclass-basedapproachtolexicalrelationships.Ph.D.thesis.EkaterinaShutova,LinSun,andAnnaKorhonen.2010.Metaphoridentificationusingverbandnouncluster-ing.InCOLING’10,pages1002–1010.EkaterinaShutova.2010.Automaticmetaphorinterpre-tationasaparaphrasingtask.InHLT’10,pages1029–1037.CatherineSmith,TimRumbell,JohnBarnden,BobHendley,MarkLee,andAlanWallington.2007.Don’tworryaboutmetaphor:affectextractionforcon-versationalagents.InACL’07,pages37–40.P.D.Turney.2008.Thelatentrelationmappingengine:Algorithmandexperiments.JournalofArtificialIn-telligenceResearch,33(1):615–655.TonyVealeandYanfenHao.2008.Afluidknowledgerepresentationforunderstandingandgeneratingcre-ativemetaphors.InCOLING,pages945–952.YorickWilks.1978.Makingpreferencesmoreactive.ArtificialIntelligence,11(3):197–223.WentaoWu,HongsongLi,HaixunWang,andKennyQiliZhu.2012.Probase:aprobabilistictaxonomyfortextunderstanding.InSIGMODConference,pages481–492.LiZhang.2010.Metaphorinterpretationandcontext-basedaffectdetection.InCOLING(Posters),pages1480–1488.
