A State of the art in Computational Linguistics

AStateoftheartinComputationalLinguisticsGiacomoFerrariDepartmentofHumanStudies-UniversitàdelPiemonteOrientaleA.AvogadrotoappearintheProceedingsofthe17thInternationalCongressofLinguists-Prague1.ThepositionofComputationalLinguisticsComputationalLinguisticshasalonghistory,datingbacktotheFifties,duringwhichitdevelopedawholesetofcomputationalmodelsandimplementations,theories,methodologiesandapplications.Itisdifficulttogiveasensibleaccountofitspresentstatewithoutgoingbackalittletothemainstepsthroughwhichthisdisciplineevolvedtowardsitspresentstate.Sinceitsorigins,ComputationalLinguisticshasbeeninanintermediatepositionbetweenComputerScienceandArtificialIntelligence,LinguisticsandCognitiveScience,andEngineering.ComputerScienceitselfsharesitsrootswithComputationalLinguistics;parsing,whichiscentralforthedesignofcompilersforprogramminglanguages(AhoandUllmann1977:6),isalsothebuildingblockofanynaturallanguageprocessingengine,andbotharetherealizationsofthechomskiantheoryofformallanguages(Chomsky1957).Thesametheory,togetherwiththecorrespondingcomputationalmodel,hasgivenacontributiontothegeneralhypothesisofArtificialIntelligence,thathumanbehavioursusuallyjudgedintelligentcouldbesimulatedinacomputerinaprincipledway.Oversimplifying,ArtificialIntelligenceaimsatmodellinganumberofbehavioursthroughthreeverygeneralparadigms,theoremproving,problemsolvingandplanning,andlanguageunderstandingandproduction.Thehistoryofbothdisciplinesisrichinintersections,especiallybetweenlanguageprocessingandplanning,asinSHRDLU(Winograd1971)or,morerecently,inARGOT(Allenetal.1982,Allen1983),withallitspracticalandtheoreticalfollow-ups;moderndialoguesystemsinalltheirformsandapplicationsarederivedfromthedialoguemodelJ.AllendesignedforARGOT.Thecommitmentto“simulationofbehaviour”,sharedbyArtificialIntelligenceandandarelevantpartofComputationalLinguistics,makesthemalsosharetheeffortfor“cognitivemodelling”ofdifferenthumanbehaviours,includingtheuseoflanguage.ThisisprobablyoneofthereasonswhyLinguisticsappearsinthesetofsciencesoriginallyinterestedinthearisingofthenewdisciplinecalledCognitiveScience().SincetheSeventies,whenlanguagetechnologyreachedastateofmaturitysuchastoallowtherealizationofsomeapplications,Engineeringhasbeeninterestedinsomeofthelanguageprocessingtechniques,anditappearedsoonthattheapproachintroducedbyengineerswascertainlylesstheoreticallyandcognitivelyinteresting,butmoreeffectiveinmanyways.Bynow,wecansaythatwhileComputationalLinguistswere,andare,moreinterestedinthecorrectnessandplausibilityoftheirmodels,Engineerswere,andare,moreinterestedintheusabilityoftoolsandtechniques,evenatthecostofsome“dirty”solutions.ThehistoryofComputationalLinguisticsinthelastdecadesismuchthehistoryoftheevolvingrelationsbetweenalltheseconjuringapproaches.2.MainAchievementsIthasbeenforalongtimeacommonplacethatComputationalLinguistics,asArtificialIntelligence,hascreatedanumberofmodels,programmes,andprototypeswhichclaimedalotbutdidnothingatall.Thisisanextremeviewthatpointsouttheabsolutelytheoreticalcharacterofwhathasbeendone,despitethestrongclaimsofpracticalutility,oftenduetoconstraintsimposedbysponsorsandfundingagencies.Nevertheless,itisjustthankstotheseattempts,oftenimpractical,thoughtheoreticallyclean,thatsomeofthefieldsofComputationalLinguisticsreachedmaturityandhasbeenabletoproducestabletechnologies,asapreparatorysteptoengineering.Thefollowingparagraphswillanalysethemainareasinwhichsomeachievementshavebeenreached.2.1ParsingandParsingTechnologyParsinghasbeenprobablythefieldofComputationalLinguisticswheresomeresultshavebeenachievedsinceanearlystage.ThefirstparserswhichboresomerelationtogrammaticaltheorygobacktotheearlySeventies.ThecommonbackgroundknowledgethatmadepossiblesuchadevelopmentisGenerativeandTransformationalGrammar,butitscomputationalinterpretationgivesrisetoenumberofdifferentmodels,withdifferentbothtechnicalandtheoreticalimpacts.Thekeyproblemtosolveistoreachalogical(deep)structureofthesentence,suchastosatisfytheconstraintofbeingmappedontosomesortofsemantic(executable)representation.Transformationalgrammardoesnotofferadirectsolutiontothisproblem,leavingafreespacetoseveralinterpretations.Themostorthodoxoftheseinterpretations(Friedman1971)proposedaparserthat,afterhavingproducedasurfacetree,bymeansofacontext-freegrammar,triedtoyieldsuchadeepstructureusingthesamesetoftransformationsproposedbyChomsky,invertinginputandoutput.Onthecontrary,themostsuccessfulsolutionwastheoneofferedbyATN(Woods1970),asnotonlyitwasagoodanswertothequestionofdealinginasingleprocessthemappingofasurfacesentence/structureontothecorrespondinglogicalform,butgavealsoorigintoanumberofgrammatical(Kaplan1973,1975;KaplanandBresnan1982)andpsycholinguistic(Kaplan1972,WannerandMaratsos1978)debates.Thiswasalsotheprototypeofanentirefamilyofnaturallanguagegrammarsforparsing,calledAugmentedPhraseStructureGrammars(APSGs,seeHeidorn1975);ingeneralAPSGsareordinarycontext-freegrammars,augmentedwithinstructionstostorethealreadyreached(sub)treesandtocarrylinguistictests.Fromaplaincom