Towardsatheoryoflexicals教学讲解课件

TowardsatheoryoflexicalsemanticsCogs3000Week7GoalsofatheoryoflexicalsemanticsAtheoryoflexicalsemanticsneedsnotonlytoaccountforthemeaningofindividualwordsbuthowwordmeaningschangeincontextwithotherwords.Considerthemeaningofgoodinthephrasegoodfriend(<loyal,reliable>)Nowconsiderthemeaningofthesamewordinthephrasegoodloverorgoodmeal.Thereisachameleon-likequalitytothemeaningofsomewords,whichneedstobeaccountedfor.LexicalsemanticsandPragmaticknowledgeThemeaningsofwordsaresaidtobestoredinthementaldictionary,

alongwiththeirphonologicalandmorpho-syntacticfeatures.Lexicalknowledgeclearlyvariesfromspeakertospeaker.LanguagesalsovaryfromoneanotherinhowtheylexicalizeourknowledgeandperceptionoftheworldHence,lexicalsemanticsthoughintimatelyconnectedwithpragmatics(knowledgeoftheworld)isdistinctfromit.Lexicalknowledgeisconfinedtoknowledgeofwordmeaningsandwordusage.TheproblemofpolysemyDictionarieshavemultiplemeaningslistedforanygivenwordandthemorecommontheword,themoremeaningsitwilltypicallyhavelistedagainstit.Meaningsof‘show’thatarefoundinthepocketOxford.Thecontextshavebeenscrambled.Semanticanomalydetection…Anotherwaywemighttesttheadequacyofatheoryoflexicalsemantics.Itisthedetectionofanomalyintheactualexpressionwhichisofinteresttoushere,ratherthanhowtheanomalousutterancewasproduced.Anothertest…Abilityofthesystemtogenerateacceptableparaphrasesforindividualwordsorphrasalexpressions.Toexpresscommonalitiesanddifferencesofmeaningbetweenverbalexpressionsthataresemanticallyrelated.abottleofwineversusajugofwhiskeyAfundamentalunresolvedproblemonhowtoapproachtheanalysisofwordandphrasemeaningsWhethersemantictheoryshouldbegroundedinlogicorinpsychology; insomeformalizedpropositionalcalculusoftruthconditionsforpossibleworlds, orinperceptualexperienceanditsinternalcognitiverepresentation.Theanswertothislongstandingcontroversymaylieinthelevelofanalysisadopted.Apropositionalcalculusisclearlymoreapplicabletosentence-levelsemanticanalysis.Thewayweperceiveobjectsintheworldisclearlyofcriticalinterestforthesemanticpropertiesofwords.SemanticNetworksOneoftheearliestattemptstoconstructacomputationalmodeloflexicalsemanticswasundertakenbyRossQuillian(1968)Asystemambitiouslydubbed‘Theteachablelanguagecomprehender’(TLC)Aprogramdesignedtobecapableofbeingtaughtto‘comprehend’Englishtext.TheTLCenvisagedthelexiconasasemanticnetworkencodingbothworldknowledgeandthemeaningsofwords.TheTLCQuillian’ssemanticnetworkisasymbolic,notaconnectionistnetwork.Thenodesofthenetworkrepresentconceptsorwords, thatarelinkedbyarcs,whichrepresentasmallnumberofrelationaltypes.Themeaningofawordmaybeexpressed,firstbyaccessingitsrootnodeinthenetworkandtraversingthenetworkelementstowhichthisnodeislinkedinaseriesofsteps.Aword’smeaningisdefinedbyotherwords,whosemeaningsaredefined,inturn,byotherwords.HowtheTLCrepresentsthepolysemouswordplantThesethreemeaningsofplantaredisjunctiveintheirusage.Thebasicmeaningofplant(PLANT1)andpointerstoitsalternativemeanings(PLANT2andPLANT3)areexpressedina‘planeofwordmemory’.SemanticnetworkforplantSemanticrepresentationsinTLCFourbasictypesofarclinkingnodesinasemanticnetworkInadditiontothebasiclogicalrelationsofdisjunction(OR)andconjunction(AND)Quillian’ssystemrecognizedasubordinate-superordinaterelation,whichothershavelabelled‘theISArelation’andwhichdefinesarelationshipofclassmembership(e.g.,dog-animal).Athirdtypeoflinkage,‘propertyattribution’linksonenodedirectlytoanother,declaringonenodeisapropertyofanothernode.Afourthtypeofcomplexlinkagebetweenthreenodesisusedtocapture‘thematicroles’: thethree-wayrelationshipthataverboraprepositiontypicallycontractswithnounsthattheyhead(e.g.,FROM3(foodgoal

airsource)).

Quillian’sclassificationofarctypesissomewhatidiosyncratic.Heoverlookscertainkindsofsemanticrelationwidelyrecognizedinlexicalsemanticanalysis.Apart-wholerelationship,‘theHASrelation’,isusefulforrepresentingfunctionalcomponentsofmanyobjects(e.g.,cat:HAS(AND(fur,whiskers,eyes,etc.))Anothersemanticrelation,confinedtoverbs,butusefulforrelatingbasicverbmeaningstosynonymsthatelaborateamannerofexecutingthebasicverbisreferredtoastroponymy. walk:stroll,wander,march,etc.TheproblemofsemanticdecompositionWhatrelationalmeanings(setofarctypes)shouldberegardedasprimitive?Quillian’sapproachtothisdifficulttheoreticalproblemispragmatic.Howwelldoesthesystemachieveitstaskofsimulatinghumanlexicalsemanticprocessinginrelationtotheeffortrequiredtobuildthenetwork?

Verificationthroughtesting(engineeringapproach)TestingyoursemanticprocessorFindthesemanticlinkagesbetweeneachwordpair.Expresseachlinkageinashortsentence.TestingtheTLCHowTLCgeneratesitsresponsesGototherootnodeofeachwordinagivenpair.Traversethearcsofitsmemoryplane,lookingfornodesthatmatchtheothermemberofthepair.e.g.,GotoPLANT1;searchplaneofPLANT1forinstancesofLIVE.DothesameforLIVE1,lookingforinstancesofPLANT,keepingarecordofany‘intersections’thatarefound.Extendsearchforintersectionsfromtheprimarymeaningplaneofeachwordtoplanesatonearcremoved,alternatingbetweeneachmemberofthewordpairandkeepingarecordofthesearchpathforanylinkagesthatwereencountered.ThissearchofsemanticspaceQuilliancharacterizedasaformof‘spreadingactivation’HowTLCgeneratesitsresponsesThepathassociatedwitheach‘intersection’issavedintemporarymemoryuntilthesearchfinished.Finally,aseparateprogramwasusedto‘translate’eachintersectionpathintoa‘kindof,MeTarzan.YouJane’verbalexpression.Butwhatdoesthisdemonstrate?

AvariantoftheTuringtest.FromwordtosentencemeaningsToaccountforpolysemy,requirestheconceptualmachineryneededforexpressingsentencemeanings.Thepivotalroleoftheverb.Atheoryofthematicroleassignment.Johngavethelibraryabook<GIVE(Ag

John)(Thbook)(Golibrary)>Inmanystudiesofaphasiaandlanguageacquisition,theabilitytoassignthematicrolescorrectlyhasbeentakenasanoperationalcriterionofsentencecomprehension.TheproblemofsemanticparaphraseTheydidn’tletJohngooutoftheroom.TheypreventedJohnfromleavingtheroom.Postulateasmallsetofsemanticprimitives; andaparsingmechanism,orsetofinferentialrules.CharlesFillmore(1968)kill=<CAUSE(Ag__BECOME(NOT(ALIVE(Th__))))>Whatapproachescouldbetakentotheproblemofsemanticdecompositionofverbmeanings?ConceptualdependencytheoryRogerSchank(1975)Verbmeaningsareexpressibleascompositionalfunctionsofjust11semanticprimitivesortypesofACT.PROPEL,MOVE,INGEST,EXPEL,GRASP,PTRANS,ATRANS,MTRANS,MBUILD,SPEAK,INTENDConceptualDependencydiagramfor‘Johnateafrog.’Theconceptualdependencygraphisderivedfromsemanticinformationinlexicalitemsandfromsomegeneralrulesofconceptualstructureformation.eatisanACToftypeINGEST,whichinvolvestheACTofMOVingsomeobjectYviaoneshandtoone’smouth.AugmentedConceptualDependencydiagramfor‘Johnateafrog.’

Thesemanticprimitivesofverbmeaningaregroundedinaconcretespatio-temporalworldofphysicalobjectsthatcollidewithoneanotherorareactivelypropelledthroughchangesofstateorlocationbywillfulbeings.Primitives,ATRANS,MTRANS,andMBUILDaremetaphoricalextensionsfromthephysicaltothementalworldofconceptsotherwiseapplicabletoatangibleworldofconcreteobjects.

Cognitionasmetaphor

MTRANSmeans‘totransferinformation’.HencetotellXaboutYistotransferknowledgefromonemetaphoricallocationinX’sdomaintothatofY.MBUILDmeans‘tocreateorcombinethoughts’.CognitiveGrammarhaselaboratedthesenotionsofmind-as-body-projection,throughtheconceptof

imageschemasandtheirroleinexplainingpatternsofpolysemy(seeLee,2002).Evaluationofsymbolicmodelsoflexicalsemantics.

Whatshouldaperformancemodelofwordmeaningshouldbeabletoaccomplish?Tohandlepolysemy(variationofwordmeaningwithphrasalcontext).Tocomputeobviousandnon-obviousrelationshipsofmeaningbetweenwordsandphrases.Toperformsemanticparaphrase.Thesegoalsweremoreexplicitlyandcomprehensivelyaddressedbythepreviousgenerationofsymbolicexpertsystemstylemodels(Quillian,1968,Schank,1975).Symbolicnetworksprovideadatastructuretosupportlexicalinference,whichenablestheconstructionofTuringtestsformodelsofwordcomprehension.ProblemswithsymbolicnetworkmodelsoflexicalsemanticsFragilityofperformanceastheygrowmorecomplex;theirinflexibility.Symbolicsemanticnetworkshavenocapacitytolearnfromlanguageexperience.Experiencewithexpertsystemmodelsofsemanticmemoryalsodemonstratedthatafullyfledgedmodelofword-to-sentencelevelsemanticprocessingisinfeasibleatthepresenttime.Theartofthepossible…InvestigatingsemanticstructuresWhendoesphrasalcontextstarttoexerciseaninfluenceonthesense-meaningofaword?Aresomesemanticrelationsmoreaccessibleormorerapidlycomputedon-linethanothers?Issemanticmemoryorganizedintorelativelydiscretemodulesorisitdistributedinahomogenoussystem?

Towhatextentarethererelativelyindependent,modalityspecific,semanticsystemsthatservedifferentdomainslexicalusage?Doesthefactthatsomelexicalitemsrepresentclustersofhighlycorrelatedpropertiesorfeaturesrendertheirretrievalmorerobustthanotheritemswhosedistinguishingsemanticfeaturesaremoresparseandessentiallyuncorrelated?

Theroleofcontextinword-sensedisambiguation.Behaviouralstudiesofon-linelexicalretrieval,usingcross-modalsemanticprimingparadigm,Swinney(1979).Bothmeaningsofanambiguouswordsuchasbug(<insect>or<covertlisteningdevice>)areinitiallyactivated,regardlessofastronglybiassingdiscoursecontext: ‘Rumourhaditthatforyearsthebuildinghadbeenplaguedwithproblems.Themanwasnotsurprisedwhenhefoundseveralspiders,roaches,andotherbugs

inthecornerofhisroom’.Swinney’sfindingsAlexicaldecisiontothecontextuallyinappropriatebutrelatedword‘spy’’wasfacilitatedjustasmuchasthecontextuallyappropriaterelatedword‘ant’.Whenadelaywasintroducedbypresentingtheprobewordthreesyllablesaftersubjectsheardtheambiguousprimeword,onlythecontextuallyappropriateprobe(ant)receivedasignificantprimingeffect.Whatdotheseresultssignify?Lexicalmeaningretrievalisaprocesswhichisinitiallyautonomousofcontexteffects.Thesemanticprimingparadigmasawindowtoon-linewordmeaningretrieval.Somemethodologicalissues:thespecifictypeofprimingparadigmused,thepotentialimpactofnon-automaticorstrategicinfluencesuponprimingeffects,thedistinctionbetweenassociativeandsemanticeffectsinpriming.SemanticrelationtypesandassociativestrengthbetweenprimeandprobewordsMossetal.,1995,1997.Fourkindsofsemanticrelationwerecrossedwithatwolevelfactorofassociativelinkage(associatedvs.non-associated).Categorycoordinaterelationsarethemostfrequentlyusedinstudiesofsemanticpriming.SemanticrelationsCategorycoordinationmaybeamorereadilyidentifiablesemanticrelationforpairsofnaturalobjectsthanartifactpairs… becausetaxonomicrelationsbasedonperceptualpropertiesaremoreapplicabletonaturalobjectsthantoartifacts.Functionalpropertiesmaybemoresalientforartifacts,whichtendtobeidentifiedbytheirusages.SemanticrelationsCategoryspecificsemanticdeficitsforsubclassesofnaturalobjects(suchasfruitandvegetables)havebeenfoundinsomeaphasics(Warrington,1975) whichithasbeenarguedreflectmodularityofsemanticcategoriesfornaturalobjects.Mossetal.(1995)distinguishedbetweeninstrumentalandscriptedfunctionalrelations.Results:associativeandsemanticprimingandtheeffectofprimetypeTherewasanappreciable‘boost’totheprimingeffectifprimeandprobewordswerealsoassociativelylinked.Inallbutthecaseofthe‘functionalscript’typeofsemanticrelation,therewasasignificantindependenteffectforsemanticallyrelatedprimes.Semanticprimingcannotbeattributedsolelytonormativeassociationstrength.However,Thestrengthofprimingeffectsassociatedwithdifferentkindsofprime-probesemanticrelationshipalsovariedwiththemodeofstimuluspresentation.Difficulttoarguethatanyhierarchyofaccessibilityorrapidityofactivationfordifferenttypesofsemanticrelationexists.Someconclusions:SemanticPriming.Theimportanceofsemanticoverform-based(phonologicalororthographicsimilarity)primingeffectsappearstobepromotedbyuseofthecross-modalprimingparadigm.Welackabodyoffindingsonthetimecourseofactivationofsemanticpropertiestoindicatethattheparadigmreallyconstitutesawindowonlexicalsemanticprocessing.Semanticprimingeffectsrepresentmorethanactivationonthebasisofnormativeassociativestrengthbetweenprimeandprobe.BrainimagingstudiesoflexicalsemanticactivationThequestionsposedintheseinvestigationsunderstandablyhaveamorelocalist(localizationist)flavour.Studiesofnormalbrainactivityinlexicalprocessinghavebeeninfluencedbyrare,butstrikingcasereportsofcategory-specificsemanticdeficitsinaphasia.Neuro-imagingstudiesofsemanticorganizationinthenormalbrainarealsoinformedbyneurologicalmodelsofthedivisionofmentalorcomputationallabourinthebrain.Thedivisionofmentallabourinthebrainbetweensensoryandmotorfunctions,episodicversusnon-episodicmemory,ordeclarativeversusproceduralknowledge(Ullman,2001)Modularorlocalistnotionsofbrainorganizationforperception,action,andmemoryaretakenasthestartingpointforneuro-linguisticmodelsoflexicalsemanticorganization.Neuro-linguisticmodelsoflexicalsemanticorganizationNounsofhighimagability(‘pictureproducers’)maybeexpectedtohavemoreofaposterior(temporal-occipital)locusofrepresentationinthebrain.Verbs(‘actionwords’),moreofananterior(frontal)representation.Conflictingfindingshavebereportedintheimagingliteratureonthequestionofcerebrallocalizationfordifferentsemanticandgrammaticalcategoriesofwords.Tyler,Stamatakis,Dick,Bright,FletcherandMossH(2003)UsedfMRItoexaminetheneuralactivationassociatedwithconceptualprocessingofnounsreferringtoanimalsandtoolsandforverbsreferringtotool-associatedactions(e.g.,drilling,painting)andbiologicalactions(e.g.,walking,jumping).Theyfoundthatobjectnamesandtheirassociatedactionsactivatedthesameareasofthetemporallobe,suggesting?...Namesforobjectsandactionsactivateasingledistributedsemanticnetwork.

Noevidenceofcategoryspecificityforeitherobjectsoractions,andthesamepatternofactivationwasevokedfortheappearanceandmotionattributesofbothlivingandnonlivingcategories.Anon-differentiated,distributednetworkforconceptualknowledgerepresentation?Tworecentstudiessuggestthatitmaybenecessarytoqualifytheconclusionwhenattentionispaidtosemanticandmorpho-syntacticattributesofthestimuliandthesubject’stask.AnfMRIstudyofsemanticjudgementswithinflectednounsandverbs(Tyler,Bright,Fletcher,&Stamatakis,2004).Subjectspresentedwithasemanticjudgementtaskinvolvingnounsandverbs,whereallitemswereexplicitlyinflectedforthatcategory(withregularplural{-s}orprogressive{-ing}marking).Thesemanticjudgementtaskinvolveddecidingwhetherthethirdmemberoftripletbelongedordidnotbelongtoasemanticclassdeterminedbytheprecedingpair.Abaselineconditionwasincludedinwhichsubjectsjudgedwhetheratargetlettersequence(incapitals)matchedordidnotmatchtheprecedingletterpair.AnalysisThebaselinetaskfMRIactivationpatternsweresubtractedfromtheactivationpatternsobtainedfromthenounandverbjudgements,inordertoisolatethesemanticprocessingcomponentofthejudgementtaskfromtheorthographicprocessingcomponents.ThesubtractionofnounandverbactivationpatternsResults:theword–baselinesubtractionSeveralpeakactivationareaswereidentified.Theyclusteredintothreeorfourregions,predominantlyin