资源描述:
《abstract (doc)摘要(doc)》由会员上传分享,免费在线阅读,更多相关内容在应用文档-天天文库。
AbstractInthisdissertationIinvestigatetheneuralmechanismsunderlyingthehumanabilitytolearn,storeandmakeuseofgrammaticalstructure,so-calledsyntax,inlanguage.IndoingsoIincorporateinsightsfromlinguistics,cognitivepsychologyandneuro-biology.Fromlinguisticresearchitisknownthatthestructureofnearlyalllanguagesexhibitstwoessentialcharacteristics:languageisproductive--fromalimitednumberofwordsandrulesonecanproduceandunderstandanunlimitednumberofnovelsentences.Further,languageishierarchical--sentencesareconstructedfromphrases,whichinturncanbeconstructedfromotherphrases,etc.Thesetwostructuralpropertiesoflanguageprovideminimumrequirementsthatasystemoflanguageprocessing,suchasthebrain,mustsatisfy.Afirstcontributionofthisdissertationisthatitattemptstoformulatetheserequirementsasconciselyaspossible,allowingforastrictevaluationofexistingmodelsofneuralprocessinginthebrain(so-calledneuralnetworks).Fromthisevaluationitisconcludedthatconventionaltypesofneuralnetworks(inparticularso-calledrecurrent,fullydistributednetworks)areunsuitedformodelinglanguage,duetocertainoversimplifyingassumptions.IntheremainderofthisthesisIthereforedevelopanoveltypeofneuralnetwork,basedonaneuraltheoryofsyntaxthatdoestakeintoaccountthehierarchicalstructureandproductivityoflanguage.ItisinspiredbyJeffHawkins'sMemoryPredictionFramework(MPF),whichisatheoryofinformationprocessinginthebrainthatstates,amongotherthings,thatthemainfunctionoftheneocortexistopredict,inordertoanticipatenovelsituations.AccordingtoHawkins,tothisendtheneocortexstoresallprocessedinformationastemporalsequencesofpatterns,inahierarchicalfashion.Cellularcolumnsthatarepositionedhigherinthecorticalhierarchyrepresentmoreabstractconcepts,andspanlongertimesbyvirtueoftemporalcompression.WhereasHawkinsapplieshistheoryprimarilytotheareaofvisualperception,inmydissertationIemphasizetheanalogiesbetweenvisualprocessingandlanguageprocessing:temporalcompressionisatypicalfeatureofsyntacticcategories(astheyencodesequencesofwords);wheneverthesecategoriesarerecognizedinanearlystageofthesentence,theycanbeexpandedtopredictthesubsequentcourseofthesentence.Iproposethereforethatsyntacticcategories,likevisualcategories,arerepresentedlocallyinthebrainwithincorticalcolumns,andmoreoverthatthehierarchicalandtopologicalorganizationofsuch`syntactic'columnsconstitutesagrammar.Asecondsourceofinspirationformyresearchistheroleofmemoryinlanguageprocessingandacquisition.Animportantquestionthataneuraltheoryoflanguagehastoaddressconcernsthenatureofthesmallestproductiveunitsoflanguagethatarestoredinmemory.Whenproducinganovelsentenceitseemsthatlanguageusersoftenreuseentirememorizedsentencefragments,whosemeaningsarenotpredictablefromtheconstituentwords.Examplesofsuchmulti-wordconstructionsare`Howdoyoudo?'or`kickthebucket',buttherearealsoproductiveconstructionswithoneormoreopen`slots',suchas`themoreyouthinkaboutX,thelessyouunderstand',orcompletelyabstractand unlexicalizedconstructions.Accordingtocertainlinguistictheorieseverysentenceinalanguagecanbeformedbycombiningconstructionsofvaryingdegreesofcomplexityandabstractness.InordertoanswerthequestionaboutthestorageofconstructionsIproposethatinlinguistics,asincognitivepsychology,onemustdistinguishbetweentwokindsofmemorysystems:amemorysystemforabstract,relationalknowledge,so-called`semantic'memory,andamemorysystemforpersonallyexperiencedeventsor`episodes'(forinstancethememoryofabirthdayparty),embeddedinatemporalandspatialcontext,so-called`episodic'memory.Icontendthat,whileabstractrulesandsyntacticcategoriesofalanguagearepartofasemanticmemoryforlanguage,anepisodicmemoryisresponsibleforstoringsentencefragments,andevenentiresentences.Episodicmemoryalsoplaysanimportantroleinlanguageacquisition,assumingthatourlinguisticknowledgeisnotinnate,butoriginatesfromtheassimilationofmanyindividuallinguisticexperiences.Animportantclaimofthisthesisisthatlanguageacquisition,likeknowledgeacquisitioninothercognitivedomains,canbeunderstoodasagradualtransformationalprocessofconcreteepisodicexperiencesintoasystemofabstract,semanticmemories.Startingfromtheassumptionthatuniversalmechanismsofmemoryprocessinginthebrainalsogovernlanguageproductionandacquisition,Iformulateanexplicittheoryabouttheinteractionbetweenanepisodicandasemanticmemoryforlanguage,calledthe``HierarchicalPredictionNetwork''(HPN),thatisappliedtosentenceprocessingandacquisition.HPNfurtherincorporatestheideasoftheMPF,withsomeimportantmodifications.ThesemanticmemoryforlanguageisconceivedofinHPNasaneuralnetwork,inwhichthenodes(correspondingtosyntacticandlexicalcorticalcolumns)derivetheirfunctionfromtheirtopologicalarrangementinthenetwork.Thismeansthattwonodesthatfulfillasimilarfunctionwithinthesyntacticanalysisofasentencearepositionedwithineachother'svicinityinsomehigh-dimensionalspace.(Thisismotivatedbythetopologicalorganizationof,forinstance,theorientationcolumnsinareaV1inthevisualcortex,whereneighboringcolumnsaretunedtosimilarorientationsoflinesegments.)Asyntacticanalysis(parse)ofasentenceinHPNconsistsofatrajectorythroughthenetwork,that(dynamically)bindsasetofnodes,astheyexchangetheirtopologicaladdressesviaacentralhub.(Thisisinspiredbyresearchonhowprimitivevisualcategoriesareboundintocomplexcontoursorshapes.)Byvirtueofflexiblebindingsbetweenthenodes(asopposedtothestaticbindingsinconventionalneuralnetworks)HPNcanaccountfortheproductivityoflanguage.InHPN,theepisodicmemoryforlanguageisembeddedwithinthesemanticmemory,intheformofpermanentmemorytraces,whichareleftbehindinthenetworknodesthatwereinvolvedinprocessingasentence.Thisway,thenetworkanalysisofaprocessedsentencecanalwaysbereconstructedatalatertimebymeansofthememorytraces. Moreover,novelsentencescanbeconstructedbycombiningpartialtracesofpreviouslyprocessedsentences.Thisthesisisorganizedasfollows:Chapter1introducesthegoalsofmyresearch,andmotivatesthechosenapproach.Chapter2introducestheMemoryPredictionFramework,andprovidestheneuro-biologicalbackgroundfortheneuraltheoryofsyntax.Chapter3coverssomebasicconceptsfromthefieldof(computational)linguistics,withaspecialfocusonparsingtechniquesthatwillbeusedintheHPNmodel.Chapter4containsacriticalreviewoftheliteratureonneuralnetworksoflanguageprocessing,withinthecontextofthedebateonthefundamentalcharacteristicsofstructureinlanguage:productivityandhierarchy.InChapters5to8Idevelop,inmultiplestages,theHPNmodel.Inordertoquantitativelyevaluatethepredictionsoftheneuraltheoryofsyntax,IdescribeacomputerimplementationofHPN,thatallowstorunsimulationsbasedontensofthousandsofsentences.Chapter5startsbyintroducingthebasicHPNmodelwithoutanepisodicmemory,whichshowsHPN'sabilitytolearnasyntactictopologyfromsimple,artificiallygeneratedsentences.Subsequently,inChapters6and7Idiscussanextendedmodel(andcomputerimplementation)thatintegratesanepisodicmemorywithasemanticmemoryforlanguage,yetforsimplicitylacksatopology.Ievaluatethismodelonalargenumberofrealisticsentenceswithrespecttoitsperformanceonsyntacticsentenceanalysis.Finally,inChapter8allthecomponentsofHPNareintegratedwithinasingleimplementation,whichdemonstrateshowanabstractgrammarintheformofanetworktopologyisconstructedoutofepisodiclinguisticexperiences.InthischapterIemphasizetheparallelsbetweenlanguageacquisitionandtheprocessofmemoryconsolidation--thetransformationbythebrainofinformationconsistingofconcreteepisodesintoanetworkofabstract,semanticknowledge.InChapter9Ipresentageneraldiscussionandmanyideasforfutureresearch.Themainconclusionofmydissertationisthatitisbothpossibleandworthwiletocoupleinsightsfrom(computational)linguisticstoneuro-biologicalinsights,andviceversa.Ontheonehand,fromthetoughfunctionaldemandsthatlanguageposesoninformationprocessingbythebrainonecaninferanumberofnon-trivialconclusionsregardingneuralconnectivityandstorageinthebrain;ontheotherhand,thephysiologicallimitationsofthebrain'shardwarepresentsomeunexpectedchallengesfortheoriesofsyntax,forinstanceconcerningtheuseoftopology.