METAPHOR AND SYMBOL,16(1&2),1–3 Copyright © 2001, Lawrence Erlbaum Associates, Inc. Introduction to the Special Issue on Metaphor and Artificial Intelligence John A. Barnden and Mark G. Lee School of Computer Science University of Birmingham Thisspecialissuearoseoutofcontributionsatasymposiumonmetaphor,artificial intelligence(AI),andcognitionheldaspartofthe1999ConventionoftheSociety fortheStudyofArtificialIntelligenceandtheSimulationofBehaviourinEdin- burgh,Scotland.Thearticlesinthisissuehaveinmostcasesundergonemajorrevi- sionasaresultbothofinteractionsatthesymposiumandofthejournal’speerre- viewing process. Themainorientationofthesymposiumwastowardcomputationalmodelsand psychologicalprocessingmodelsofmetaphoricalunderstanding.Thisorientation iswellreflectedintheselectionofarticlesinthisspecialissue.Twoareaboutim- plementedcomputationalsystemsforhandlingdifferentaspectsofmetaphorun- derstanding.OneofthesearticlesisbyThomasandMareschal(2001/thisissue) andtheotherisbyourselves(Lee&Barnden,2001/thisissue).Theycontrastin many respects, one of which is that the former is within the connectionist para- digm, whereas the latter is in the traditional symbolic paradigm. Two of the re- mainingarticles,thosebyvanGenabith(2001/thisissue)andbyVogel(2001/this issue),arelargelyabouthowmetaphorcanbeaccommodatedinacceptedlogical representationalframeworks.Theythereforehelptoshowthathandlingmetaphor computationallyisnotsomethingthatneedrequirerevolutionsincurrentpractice inAIorformalsemantics.Threearticlesare,indifferentways,onpsychological processes involved in metaphor understanding. These are by Bortfeld and McGlone (2001/this issue); Brisard, Frisson, and Sandra (2001/this issue); and Noveck,Bianco,andCastry(2001/thisissue).Thefirstrecommendsthatcurrently competing processing models in psychology could cooperate and complement eachotherratherthancompete.Thesecondprovidesevidenceinfavorofmeta- RequestsforreprintsshouldbesenttoJohnA.Barnden,SchoolofComputerScience,Universityof Birmingham, Edgbaston, Birmingham B15 2TT, England. E-mail: [email protected] 2 BARNDEN AND LEE phorical processing taking more time than literal processing when timings are takenwithinsentences,notjustattheendofsentences.Thethird,relatedly,points toevidencethatmetaphoricalprocessingcomeswithextracost,butthatthecost bringsadditionalbenefits.Finally,weincludeonearticle,byNeumann(2001/this issue),thatisdistinctlydifferentinflavorfromtheothers,asitisadetailedlinguis- ticstudy,usingdatafromGermanandJapanese,underpinningthecross-linguistic cognitive reality of conceptual metaphors. Thesymposiumproceedings(ProceedingsoftheAISB’99SymposiumonMeta- phor,ArtificialIntelligence,andCognition,1999)containsfurtherpapersofvarious types,includingexperimentalpsychologicalresults,resultsoflinguisticanalysisof examplesandlinguisticcorpora,observationsonmetaphorinartandarchitectural design, and steps toward the handling of metaphor in machine translation of lan- guages.Abstractsareavailableathttp://www.cs.bham.ac.uk/~jab/AISB-99. FromourownpointofviewasAIresearchersintometaphor,wefinditvaluable totakepartininterdisciplinaryforums,andwehopethatourauthorsfromother disciplinesdosotoo.Weweretouchedbytheextenttowhichvarioussymposium participantsfromoutsideAIweresurprisedattheveryexistenceofAIresearchers interestedinasubjectsuchasmetaphor.Undoubtedly,metaphoriscurrentlyami- norityconcernwithinAI(althoughitshouldbepointedoutthattheminorityhas beeninplacesinceearlyinthedevelopmentofAI).However,webelievethatthere isnewroomtohopeforgrowthofinterestinthesubjectwithinAI.Thereadersof thisjournalprobablydonotneedtobeconvincedoftheprevalenceandcentrality ofmetaphorineverydaytextandspeech.Becausetechnologicaldevelopmentsare makingitincreasinglypossibleandimportanttoincludeAIelementsinpublicly availableorcommercialsoftware,andnaturallanguageprocessingisanimportant aspect of user friendliness, issues such as metaphor are increasingly becoming loomingpracticalobstaclesasopposedtopiesinthedistantsky.Inparticular,the developmentoflargetextandspeechcorporaandoftoolscapableofdealingwith theirimmensesizemakeitreasonabletoembarkondevelopingmethodsforthe large-scale semiautomated analysis of metaphor in real discourse. InmanyareasofAI,notleastlanguageprocessing,AIresearchandpsychologi- calresearchmustinteractfortworatherdifferentreasons.Oneisthemoreobvious oneandisoftenpointedout:Studiesofhowthehumanmindoperatescouldsug- gestmechanismstoAIsystemdevelopers,andconversely,thedetailedcomputa- tionalorformalmodelingthatAIresearchersdo(whethertheyhaveconnectionist, symbolic,orotherorientations)cancontributetopsychologicaltheorizing—itcan suggestrichcomputationaldifficulties,abilities,subtleties,compromises,hybrids, and other possibilities. The second, less often considered reason is that if an AI systemistointeractwithpeople,notablybylanguage,itmusttosomeextentap- preciatethementalstatesandprocessesofthosepeople.Forexample,anAIsys- tem that understands metaphorical language—everyday language—must have someappreciationofhowthespeakersorwritersexpectorintendittobeunder- INTRODUCTION 3 stood.Psychologycanthrowlightonsuchexpectationsandintentionsandisthere- forerelevanttothedevelopmentoftheAIsystem,evenifthesystemitselfisnot intendedtobeapsychologicalmodel.Ofcourse,theextenttowhichthesystem mustbeabletoappreciatetheworkingsofpeople’smindsneedbenogreaterthan theextenttowhichweasordinarylanguageunderstanderscandoit,andthatex- tent is often small enough. Thearticlesinthisspecialissuewereselectedbythejournal’sstandardmecha- nismsofblindpeerreview.Thisappliedjustasmuchtoourownarticleastooth- ers.Wearegratefultooursmallbandofreviewers,especiallyAlbertKatz,fortheir immenselyhardwork;toMetaphor&Symbolforitsreceptiveness,patience,gen- eralguidance,andcarefulattentiontothecontentandstyleofthearticles;toallthe otherauthorsoftheincludedarticlesfortheirhardworkandtheirinterestincon- tributing to the symposium and special issue; and to authors whose articles we wereunabletoinclude,butwhoneverthelessmadeavaluablecontributiontothe symposium and enlarged our own knowledge of metaphor. REFERENCES Bortfeld,H.,&McGlone,M.S.(2001/thisissue).Thecontinuumofmetaphorprocessing.Metaphor and Symbol, 16, 75–86. Brisard,F.,Frisson,S.,&Sandra,D.(2001/thisissue).Processingunfamiliarmetaphorsinaself-paced reading task.Metaphor and Symbol, 16, 87–108. Lee,M.G.,&Barnden,J.A.(2001/thisissue).Reasoningaboutmixedmetaphorswithinanimple- mented artificial intelligence system.Metaphor and Symbol, 16, 29–42. Neumann,C.(2001/thisissue).Ismetaphoruniversal?Cross-languageevidencefromGermanandJap- anese.Metaphor and Symbol, 16, 123–142. Noveck,I.A.,Bianco,M.,&Castry,A.(2001/thisissue).Thecostsandbenefitsofmetaphor.Metaphor and Symbol, 16, 109–121. ProceedingsoftheAISB’99SymposiumonMetaphor,ArtificialIntelligence,andCognition.(1999). Brighton,England:UniversityofSussex,SocietyfortheStudyofArtificialIntelligenceandthe Simulation of Behaviour. Thomas,M.S.C.,&Mareschal,D.(2001/thisissue).Metaphorascategorization:Aconnectionistim- plementation.Metaphor and Symbol, 16, 5–27. vanGenabith,J.(2001/thisissue).Metaphors,logic,andtypetheory.MetaphorandSymbol,16,43–57. Vogel, C. (2001/this issue). Dynamic semantics for metaphor.Metaphor and Symbol, 16, 59–74. METAPHOR AND SYMBOL,16(1&2),5–27 Copyright © 2001, Lawrence Erlbaum Associates, Inc. Metaphor as Categorization: A Connectionist Implementation Michael S. C. Thomas Neurocognitive Development Unit Institute of Child Health Denis Mareschal Centre for Brain and Cognitive Development School of Psychology Birkbeck College Akeyissueformodelsofmetaphorcomprehensionistoexplainhow,insomemeta- phoricalcomparison“AisB,”onlysomefeaturesofBaretransferredtoA.Thefea- turesofBthataretransferredtoAdependbothonAandonB.Thisisthecentralthrust ofBlack’s(1979)well-knowninteractiontheoryofmetaphorcomprehension.How- ever,thistheoryissomewhatabstract,anditisnotobvioushowitmaybeimple- mentedintermsofmentalrepresentationsandprocesses.Inthisarticle,wedescribea simple computational model of online metaphor comprehension that combines Black’sinteractiontheorywiththeideathatmetaphorcomprehensionisatypeofcat- egorizationprocess(Glucksberg&Keysar,1990,1993).Themodelisbasedonadis- tributed connectionist network depicting semantic memory (McClelland & Rumelhart,1986).Thenetworklearnsfeature-basedinformationaboutvariouscon- cepts.Ametaphoriscomprehendedbyapplyingarepresentationofthefirstterm(A) tothenetworkstoringknowledgeofthesecondterm(B),inanattempttocategorizeit asanexemplarofB.TheoutputofthisnetworkisarepresentationofAtransformed by the knowledge of B. We explain how this process embodies an interaction of knowledgebetweenthe2termsofthemetaphor,howitaccordswiththecontempo- rarytheoryofmetaphorstatingthatcomprehensionforliteralandmetaphoricalcom- parisonsiscarriedoutbyidenticalmechanisms(Gibbs,1994),andhowitaccountsfor existingempiricalevidence(Glucksberg,McGlone,&Manfredi,1997)andgenerates RequestsforreprintsshouldbesenttoMichaelS.C.Thomas,NeurocognitiveDevelopment Unit, Institute of Child Health, 30, Guilford Street, London WC1N 1EH, England. E-mail: [email protected] 6 THOMAS AND MARESCHAL newpredictions.Inthismodel,thedistinctionbetweenliteralandmetaphoricallan- guage is one of degree, not of kind. Why use metaphor in language? Gibbs (1994) summarized three kinds of an- swers to this question (Fainsilber & Ortony, 1987; Ortony, 1975). First, the inexpressibility hypothesis suggests that metaphors allow us to express ideas thatwecannoteasilyexpressusingliterallanguage.Second,thecompactness hypothesissuggeststhatmetaphorsallowthecommunicationofcomplexcon- figurationsofinformationtocapturetherichnessofaparticularexperience.The use of literal language to communicate the same meaning would be cumber- some and inefficient. Third, the vividness hypothesis suggests that the ideas communicableviaametaphorareinfactricherthanthosewemayachieveusing literallanguage. Whenwereceiveinformationcodedintheformofametaphor(e.g.,notthat Richardisbrave,aggressive,etc.,butthat“Richardisalion”),howdoweprocess suchlanguagetoextractitsvividmeaning?Thetraditionalviewinphilosophyand linguisticswasthatlanguagecomprehensionandproductionarebuiltaroundlit- erallanguage,thatmetaphoricallanguageisbothhardertocomprehend(giventhat itisliterallyfalse;inourexample,Richardisnotalion)andrequiresspecialpro- cessingmechanismstodecode.Althoughitisdistinguishedbyitscommunicative advantages,metaphorwasseenasapurelylinguisticphenomenon(Grice,1975; Searle,1975).Morerecently,thisviewhasbeenchallengedontwogrounds(e.g., Gibbs,1994,1996;Lakoff,1993).First,itisclaimedthatmetaphorisconceptual ratherthanlinguistic.Second,itisclaimedthatmetaphorisnotanadd-ontothe moreprimaryliterallanguageprocessingsystem,butakeyaspectoflanguageit- self,sharingthesamekindofprocessingmechanisms.Inthisarticle,wefocuson the second of these claims. Theargumentthatmetaphorcomprehensiondoesnotrequirespecialprocess- ingmechanismshastwostrands(Gibbs&Gerrig,1989).Thefirstisthatonline processing studies suggest that (with appropriate contextual support) metaphors andliteralstatementstakethesameamountoftimetoprocess(e.g.,Inhoff,Lima, &Carroll,1984;Ortony,Schallert,Reynolds,&Antos,1978).Thisseemstorule out the possibility that metaphors are initially processed as literal statements, found to be false, and only then processed by metaphor-specific mechanisms. It doesnot,however,ruleoutthepossibilitythatliteralandmetaphoricalmeanings ofasentencemaybecomputedsimultaneouslyandinparallelbyseparatemecha- nisms.Thesecondstrandsuggeststhatliterallanguageprocessingisnoeasierthan metaphorical processing, given that both rely on a common ground between speakerandlistenertocomprehendwhatagivenutterancemeans(Gibbs,1994). Thatis,anapparentlyliteralstatementmaywellhaveanimplicatedmeaninggiven acertainsetofsharedassumptionsbetweenspeakerandlistener.Ifbothtypesof CONNECTIONIST MODEL OF METAPHOR AS CATEGORIZATION 7 language involve similar problems, it makes sense to see them as engaging the same sort of mechanisms. Black(1955,1962,1979)outlinedthreeviewsofhowthemetaphorcomprehen- sionprocessmaywork.Inthefirstofthese,thesubstitutionview,tounderstandthe metaphoricalcomparison“Richardisalion,”thiscomparisonmustinitiallybere- placedbyasetofliteralpropositionsthatfitthesamecontext(e.g.,Richardisbrave, Richardisaggressive).Inthecomparisonview,themetaphoristakentoimplythat the two terms are similar to each other in certain (communicatively relevant) re- spects.Forexample,bothRichardandthelionarebrave,aggressive,andsoforth. TheintentionofthecomparisonistohighlightthesepropertiesinthefirsttermRich- ard.Ineffect,thecomparisonisshorthandforthesimile“Richardislikealion.”In theinteractiveview,thecomparisonofthetwotermsinthemetaphorisnottakento emphasizepreexistingsimilaritiesbetweenthem,butitselfplaysaroleincreating thatsimilarity.Thetopic(firstterm)andvehicle(secondterm)interactsuchthatthe topicitselfcausestheselectionofcertainofthefeaturesofthevehicle,whichmay thenbeusedinthecomparisonwiththetopic.Inturn,thisparallelimplicationcom- plexmaycausechangesinourunderstandingofthevehicleinthecomparison. Althoughtheinteractionviewhasbeendescribedas“thedominanttheoryinthe multidisciplinarystudyofmetaphor”(Gibbs,1994,p.234),ithasneverthelessbeen criticized for the vagueness of its central terms. One of the key issues for psycholinguisticmodelsofmetaphorcomprehensionisexplanationofthenatureof theinteractionbetweentopicandvehiclethatconstrainstheemergentmeaningof thecomparison.Threemainmodelshavebeenproposed.Thesearethesalienceim- balancemodel(Ortony,1979,1993),thestructuralmappingmodel(Gentner,1983; Gentner&Clements,1988),andtheclassinclusionmodel(Glucksberg&Keysar, 1990,1993).Thesalienceimbalancemodelproposesthatmetaphorsaresimilarity statementswithtwotermsthatshareattributes.However,thesalienceoftheseattrib- utesismuchhigherinthesecondtermthanthefirst.Thecomparisonservestoem- phasizetheseattributesinthefirstterm.Thestructuralmappingmodelsuggeststhat topicandvehiclecanbematchedinthreeways:intermsoftheirrelationalstructure (i.e.,inthehierarchicalorganizationoftheirpropertiesandattributes),intermsof thosepropertiesthemselves,orintermsofbothrelationalstructureandproperties. Peopletendtoshowapreferenceforrelationalmappingsinmetaphors.Theclassin- clusionmodelproposesthatmetaphorsareunderstoodascategoricalassertions.Ina metaphor“AisB,”AisassignedtoacategorydenotedbyB(i.e.,Richardfallsinto theclassofbrave,aggressivethings,ofwhichalionisaprototypicalmember).Only thosecategoriesofwhichBisamemberthatcouldalsoplausiblycontainAarecon- sideredastheintendedmeaningofthecategoricalassertion. Theviewofmetaphorasaformofcategorizationseemsperhapsmostconsis- tent with the claim that metaphor comprehension requires no special processes overandaboveliteralcomprehension.Boththesalienceimbalancemodelandthe structural mapping model imply a property-matching procedure that is engaged 8 THOMAS AND MARESCHAL fornonliteralcomparisons(Glucksberg,McGlone,&Manfredi,1997).Moreover, Glucksbergetal.(1997)arguedthattheclassinclusiontheoryisempiricallydis- tinguishablefromtheseproperty-matchingmodels.Althoughliteralcomparisons areasymmetric(inthatthesimilarityoftwotermscanberateddifferentlydepend- ingontheorderofpresentation;e.g.,Tversky&Gati,1982),classinclusionstate- mentsshouldbemorethanasymmetric;theyshouldbenonreversible.“Thelionis Richard”shouldmakeverymuchlesssensethan“Richardisalion,”unlessRich- ardhappenstobeaprototypicalmemberofacategoryofwhichlioncouldalsobea member.Second,Glucksbergetal.claimedthatthetopicandvehicleshouldmake verydifferent(althoughinteractive)contributionstothemetaphor’smeaning,and that these contributions are predictable. The vehicle provides the properties that maybeattributedtothetopic,butthelistener’sfamiliaritywiththetopicconstrains thosepropertiesthatmaybeattributedtoit.Glucksbergetal.primedcomprehen- sionofmetaphoricalcomparisonsbypreexposuretoeithertopicorvehicle.They predictedthatonlycomparisonsinvolvingtopicswithfewpotentiallyrelevantat- tributes,orvehicleswithfewpropertiesavailableascandidateattributes,should benefitfrompreexposure.Intheirview,neitherproperty-matchingmodelshould predict the nonreversibility or specific interactivity effects. Nevertheless, Glucksberg et al. found empirical support for both of their predictions. TheclassinclusionmodelcontrastswithLakoffandcolleagues’theorythatmeta- phorsrelyonestablishedmappingsbetweenpairsofdomainsinlong-termmemory (Lakoff,1987,1990,1993;Lakoff&Johnson,1980;Lakoff&Turner,1989).Thus, comprehensionofthemetaphor“thisrelationshipisgoingnowhere”proceedsviaa preexistingsystemofcorrespondencesbetweentheconceptualdomainsofloveand journey. The class inclusion theory, on the other hand, posits no such preexisting metaphoricalstructures.Inacomparisonoftheclassinclusionandconceptualmeta- phortheories,McGlone(1996)determinedthatitwasnotyetpossibletofindcon- clusiveevidenceforeithertheory.McGlonepresentedfourexperiments,employing metaphorparaphrasing,comparison,andcuedrecall,theresultsofwhichhetookto supporttheclassinclusiontheoryovertheconceptualmetaphortheory.However,he admittedthattheuseoftheseofflinemeasuresmaynothavetappedtheuseofcon- ceptual metaphors during online interpretation. Evidence for the class inclusion modelcomesfromtheirreversibilityofmetaphorsandrelateddiscoursephenomena (Glucksberg,1991),whereastheprimaryevidencefortheconceptualmetaphorthe- orycomesfromtheobservedsystematicityofidiomaticexpressionsincertainse- mantic domains. Lakoff (1993) criticized the class inclusion model for its use of metaphoricalattributivecategoriestomediatemetaphorcomprehension.Thus,the metaphor “my job is a jail” must be understood via appeal to the category of re- strainingthings(ofwhichjailisaprototypicalmember).However,theapplicationof the term restraining to the concept job is itself metaphorical. Yet Lakoff’s (1993) owntheoryincursthesameprobleminhisuseoftheinvarianceprinciple,bywhich domainsarelinkedinlong-termmemory.Thus,thedomainsofcontainersandcate- CONNECTIONIST MODEL OF METAPHOR AS CATEGORIZATION 9 gories,forinstance,arelinkedinaparticularwaysuchthat“sourcedomaininteriors correspondtotargetdomaininteriors”(p.215).However,thenotionoftheinterior ofacontainercanonlybemetaphoricallyappliedtotheconceptcategory.Insum,it isprematuretorejecteitherofthesetheoriesatthecurrenttime.Inwhatfollows,we concentrateontheclassinclusiontheory. Inthisarticle,ouraimistoproposeacomputationalmodelofmetaphorcom- prehensionbasedonacategorizationdevice,asopposedtotheproperty-matching device that would have to lie at the heart of a salience imbalance or a structural mappingmodel.Becauseourmodelisbasedonapreviouslyproposedmechanism ofsemanticmemory,itexemplifiestheideathatmetaphorcomprehensionisnota specialfunctionofthelanguageprocessingsystem.Indeed,wesuggestthatwithin this mechanism, literal and metaphorical comparisons are distinguished only quantitatively,notqualitatively.Theimplementedmodeldemonstratesinconcrete termshowtopicandvehicleinteractinmetaphorcomprehension,addressingsome ofthevaguenessintheinteractionposition.Finally,weshowhowthemodelac- countsforbothoftheempiricalfindingsdemonstratedbyGlucksbergetal.(1997) and how it generates new predictions. First,however,welayouttheassumptionsofthemetaphorbypatterncomple- tion (MPC) model. ASSUMPTIONS OF THE MODEL The model builds on the following assumptions: 1. Theaimofcomprehensionistheongoingdevelopmentofasemanticrepre- sentation, and that representation is feature based. 2. The ongoing semantic representation is continually monitored against ex- pectationsbasedonacommongroundbetweenlistenerandspeaker.Specifically with regard to metaphor comprehension, the ongoing semantic representation is monitoredfordegreeofexpectedmeaningchange.(Itwillbemonitoredinother ways for other nonliteral communication.) 3. Comparisonsoftheform“AisB”areclassinclusionstatementswherethe intendedmeaningisthatAisamemberofcategoryBandsoshouldinherititsat- tributes (Glucksberg & Keysar, 1990, 1993). 4. Themeaningproducedbyametaphoristheresultofusingacategorizationmech- anismtotransferattributesfromBtoAwhenAisnotinfactamemberofB.However, membershipofBisnotallornothing,butdependsondegreeoffeaturaloverlap. 5. Thecategorizationmechanismisanautoassociativeneuralnetwork.Cate- gorymembershipisestablishedbytheaccuracyofreproductionofanovelinputA toanetworktrainedtoreproduceexemplarsofcategoryB.Theoutputofsucha network is a version of A transformed to make it more consistent with B.