Artificial Intelligence and ComputerAssisted Language Instruction.docx

Artificial Intelligence and ComputerAssisted Language Instruction.docx

《Artificial Intelligence and ComputerAssisted Language Instruction.docx》由会员分享，可在线阅读，更多相关《Artificial Intelligence and ComputerAssisted Language Instruction.docx（17页珍藏版）》请在冰豆网上搜索。

ArtificialIntelligenceandComputerAssistedLanguageInstruction

ArtificialIntelligenceandComputer-AssistedLanguageInstruction:

APerspective

AlanBailin

UniversityofWesternOntario

Abstract:

ThearticleattemptstooutlinethemajorcomponentsofCALI-AI（computer-assistedlanguageinstructionincorporatingartificialintelligencetechniques）.ThearticlebeginsbydiscussingbrieflythecentralassumptiononwhichCALI-AIworkisbased,thatis,thathumancognitiveabilitiescanbereproducedbymechanicalmeans.ItthenproceedstoexaminethefollowingcomponentsofCALI-AI:

（1）naturallanguageprocessing,problemsolving,（3）languagelearning,and（4）modelingteacherbehavior.Thearticleconcludeswithadiscussionofthewaysinwhichlanguageteacherscanparticipateinthedevelopmentofthefield.

KEYWORDS:

artificialintelligence,computer-assistedlanguageinstruction,naturallanguageprocessing,languageteaching.

Introduction

ThisarticleattemptstooutlinethemajorcomponentsofCALI-Al（computer-assistedlanguageinstructionincorporatingartificialintelligencetechniques!

）fromtheperspectiveofhowthesecomponentsareusedatpresentandhowtheymightbeusedinthefuture.Insodoing,ittriestoshowthatCALI-AIhasfeaturesthatdistinguishitfromotherAIapplicationsandthatcanallowittomakeitsowndistinctcontributionsbothtoCALIandtoAI.

ThearticlebeginsbydiscussingbrieflythecentralassumptiononwhichCALI-AIworkisbased,thatis,thathumancognitiveabilitiescanbereproducedbymechanicalmeans.ItthenproceedstoexaminethefollowingcomponentsofCALI-AI:

（1）naturallanguageprocessing,

（2）problemsolving,（3）languagelearning,and（4）modelingteacherbehavior.Thearticleendswithadiscussionofthewaysinwhichlanguageteacherscanparticipateinthedevelopmentofthefield.

HumanCognitionandComputability

TheultimategoalofCALI-AIistomodelinarobustwaythecognitivebehaviorofhumansinaparticularsocialrole:

thatoflanguageteacher.Atleast

25

inthisregard,CALI-AIisnotsimplyanattemptatsophisticatedprogramming.Itis,aboveall,anattempttoachieveatrueAIgoal:

thereplicationbymachineofsignificantaspectsofhumancognitiveabilities.Totestwhetherornotamachinecouldreplicatehumancognitivebehavior,AlanTuringsuggestedthatahumanshouldinteractwithitwithoutanyknowledgeaboutwhetherornots/hewastalkingtoamachine.Ifthehumanbelieveds/hewastalkingtoanotherhuman,themachinecouldbeconsidered"truly"intelligent.ThistestisknownastheTuringTest.

ThetestwouldhavenovalidityinrelationtomanyAIsystemsdesignedforpurelymilitaryorindustrialpurposesbecausethesesystemsdonotreallyaimatsimulatinghumanbehavior.Rather,theyareintendedsimplytoaidinthemakingofpurelytechnicaldecisions（thecorrectmixtureofingredientsinaSoupmix,orthecorrecttechnicalresponsetoincomingmissiles—seeBuchnanan1985forexamples）.Ontheotherhand,ifCALI-AIultimatelydoesachieveitsgoal,itshouldbeabletopasstheTuringtest,becauseitwillhavesuccessfullyreplicatedasignificantaspectofhumanbehavior—thatofalanguageteacher.Atrulysuccessfulsystemwouldbehaveinwaysindistinguishablefromthatofahumanperformingthesameteachingfunction.

Wearefarfromachievingthisgoal.CurrentCALI-AIprojectscannotand,intheauthor'sopinion,shouldnotbeusedinplaceofateacher.TheyaretrulyapartofCAI—computer-assistedinstruction.Nevertheless,evenatthisstage,agreatdealofwhatateacherdoescanbereplicatedbymachine.CALI-AIcancheckthesyntaxofastudent'swrittenwork,createenvironmentsinwhichstudentsuselanguageinpedagogicallybeneficialways,andprovidesophisticatedfeedbacktostudentsengagedindrill-and-practiceexercises.

UnderlyingbothCALT-AI'ssmallbutsignificantactualcontributionsanditspotentialcontributionsisthequestionofwhataspectsofhumancognitivebehavioracomputercanreplicate.Inthemostgeneralterms,theansweristhatamachinecanreplicateanyaspectofhumanbehaviorwhichcanberepresentedorsimulatedbycomputationalmeans.Itmustbestressedthat"computational"heredoesnotreallymeaninvolvingnumbers（notatleastinanarithmeticsense）.Rather,itreferstoanythingwhichcanbedescribedintermsofa"Turingmachine,"

ATuringmachineisnota"real"machine,butratheranautomaton,thatis,anidealizedabstractmodelofamachine.itisspecificallyintendednotfornumericaloperations（althoughitcanbeusedtocomputethem）,butratherforthegeneralmanipulationofsymbols.ATuringmachine（seefigure1）consistsof

26

（1）afinitenumberofstates,

（2）atapeofinfinitelength,（3）afinitenumberoftapesymbols（includingtheblank（B）symbol）,and（4）atapehead.Eachtapecellcontainsonlyonetapesymbol,andthetapeheadscansonlyonecellatatime.Themachinecanperformthefollowingkindsofoperations:

（1）itcaneraseatapesymbolonthecellwhichthetapeheadisscanningandreplaceitwithanon-blanktapesymbol,

（2）itcanmovethetapeheadonecelltoeithertheleftorright,（3）itcanchange4）itcan"halt"（thatis,stop）completely（HopcraftandUllman1969,80ffandPartee1978,162ff）.

Figure1:

TheTuringMachine

ThebasicassumptionofAIingeneral,andCALI-AIinparticular,isthathumancognition—oratleastasignificantportionofit—canbereplicatedbymeansoftheploddingstep-by-stepmovesofaTuringmachine.UnderlyingCALI-AIthenisnotsomemagician'shocus-pocus,butratherthecareful,preciseanalysisofwhatlanguageteachinginvolves.Whetherornotthebasicassumptionprovestenable,theattempttodevelopCALI-AIshouldleadustoabetterunderstandingofwhatitmeanstoteach.AsweexaminethecomponentsofCALI-AI,thisshouldbekeptinmind.

NaturalLanguageProcessing

Thefieldofnaturallanguageprocessingcanbedividedintothefollowingareas:

syntax,semantic/pragmatics,morphology,speechprocessing,andlanguagegeneration.Below,eachwillbeexaminedinturn,

Syntax

Thetwobasicareasinwhichsyntaxisimportantinnaturallanguageprocessingareparsingandlanguagegeneration.Thissectionconcernsonly

27

parsingbecause,unlikelanguagegeneration,itisanareaofapplicationwherethesyntaxoperatestoalargedegreeindependentlyofothernaturallanguageprocessingcomponents（semantics,morphology,etc.）.

Theconsiderableworkthathasbeendoneinnaturallanguageprocessinghasledtoavarietyofapproachesand,asaconsequence,anumberofdifferentwaysofcategorizingparsers（seetable1）.Thesecategorizationsprovideuswithawayofexploringthepropertiesofparsers.

HowTheyParse:

Top-Down,Bottom-Up,Wait-and-SeeParser（WASP）

HowTheyExplore:

Backtracking,ParallelParsing

FormalGrammar:

Type0,contextSensitive,ContextGree,RegularGrammar

LinguisticGrammar:

Government-Binding（GB）,Lexical-FunctionalGrammar（LFG）,StructureGrammar（GPSG）

Table1

Onewayofclassifyingparsersisintermsofhowtheparsingprocedureoperates.Atop-downparserbeginswiththemajorsyntacticunitsofasentence,thentriestofindtheimmediateconstituentsofeachofthese,andsoonuntilthewordunitsarereached.Abottom-upparser,ontheotherhand,triestobuildthestructuresfromthewordleveluptothesentencelevel（seeGrishman1986,27andWinograd1983,90-91）.Await-and-see-parser（WASP）doesnottrytobuildamajorcategoryfromthestart,but,asthenameimplies,waitsuntilithastheconstituentsnecessaryformakinganidentification.Inotherwords,forthemostpart,ittriestotaketheguessworkoutofparsing（seeWinston1984,309ffandMarcus1980）.

However,evenWASPsmustguessoccasionally,andlikeotherparsers,needtoexploremorethanonesyntacticanalysisbeforedecidingonanappropriateparse.Parserscanbeclassifiedintermsofhowtheyexplorethesepossibilities.Aparserissaidtobacktrackifitexploresonepossiblesyntacticstructureafteranotheruntilitfindstheonewhichisrequired.Parallelparsing,ontheotherhand,meansthattheparserexploresallthealternativesatthesametime（seeGrishman1986,27fandWinograd1983,368-369）.

Parserscanalsobeclassifiedintermsoftheformalgrammartypewithwhichtheycanbeidentified（i.e.,canbeconsideredmathematicallyequivalentto）.Formalgrammarsaredescribedusing"productions."Theseareruleswhichtakethefollowingform:

A-->B

AruleofthisformisunderstoodtomeanthatAconsistsofB.

28

Classificationintermsofformalgrammartypesrelatestorestrictionsonwhatsymbolscangoontheleftandrightsidesofthearrow.Context-sensitivegrammarsaregrammarsinwhichtherightsideoftherule（theB）mustcontainatleastasmanysymbolsastheleftside（AB-->BCandAB-->BCD,butnotAB-->B）,Incontext-freegrammars,theleftsidemustcontainonlyonesymbolandtherightcanhaveanynumberaslongasitisnotsolelycomprisedofthesymbolfor"theemptysentence"（A-->BorA-->BCD,butnotA-->{empty}orAB-->CD）.Regulargrammarsareevenmorerestricted.Therecanbeonlyonesymbolontheleftandatmosttwosymbolsontheright,atleastoneofwhichmustbea"terminal"symbol,thatisasymbolwhichcannotappearontheleft.Inaddition,inaregulargrammartheterminalsymbolmustalwaysbeontherightorontheleft（A-->aBorA-->BaandA-->b,butnotA-->aBandA-->Ba,wherelower-caselettersdenoteterminalsymbols）.Almostallmodernparsersarebasedoncontext-freeorregulargrammars（althoughtheyareoften"augmented"withadditi