Modelling human teaching tactics and strategies for

Modelling human teaching tactics and strategies for tutoring systems

Article (Unspecified)

http://sro.sussex.ac.uk

du Boulay, Benedict and Luckin, Rosemary (2001) Modelling human teaching tactics and strategies for tutoring systems. International Journal of Artificial Intelligence in Education, 12 (3). pp. 235-256. ISSN 1560-4292

This version is available from Sussex Research Online: http://sro.sussex.ac.uk/id/eprint/1330/

This document is made available in accordance with publisher policies and may differ from the published version or from the version of record. If you wish to cite this item you are advised to consult the publisher’s version. Please see the URL above for details on accessing the published version.

Copyright and reuse: Sussex Research Online is a digital repository of the research output of the University.

Copyright and all moral rights to the version of the paper presented here belong to the individual author(s) and/or other copyright owners. To the extent reasonable and practicable, the material made available in SRO has been checked for eligibility before being made available.

Copies of full text items generally can be reproduced, displayed or performed and given to third parties in any format or medium for personal research or study, educational, or not-for-profit purposes without prior permission or charge, provided that the authors, title and full bibliographic details are credited, a hyperlink and/or URL is given for the original metadata page and the content is not changed in any way.

http://sro.sussex.ac.uk/

MODELLING HUMAN TEACHING TACTICS AND

STRATEGIES FOR TUTORING SYSTEMS

Benedict du Boulay and Rosemary Luckin , HumanCentred Technology Research Group,Schoolof CognitiveandComputingSciences,University of Sussex, BN19QH,[email protected],[email protected]

Abstract. One of the promisesof ITSs and ILEs is that they will teachand assistlearningin an intelligent manner. Historically this hastendedto meanconcentratingon the interface,on the representationof thedomainandon the representationof the student’s knowledge. Sosystemshave attemptedto provide studentswith reificationsbothof what is to be learnedandof thelearningprocess,aswell asoptimally sequencingandadjustingactivities, problemsandfeedbackto besthelpthemlearnthatdomain.

We now have embodied(anddisembodied)teachingagentsandcomputer-basedpeers,andthe field demonstratesa muchgreaterinterestin metacognitionandin collaborative activitiesandtools to supportthat collaboration.Neverthelessthe issueof the teachingcompetenceofITSsandILEs is still important,aswell asthemorespecificquestionasto whethersystemscanandshouldmimic humanteachers.Indeedincreasinginterestin embodiedagentshasthrownthespotlightbackonhow suchagentsshouldbehave with respectto learners.

In the mid 1980sOhlssonand othersoffered critiquesof ITSs and ILEs in termsof thelimited rangeandadaptabilityof their teachingactionsascomparedto thewealthof tacticsandstrategiesemployed by humanexpert teachers.So arewe in any betterpositionin modellingteachingthanwewerein the80s?Are thesecriticismsstill asvalid todayasthey werethen?

This paperreviews progressin understandingcertainaspectsof humanexpert teachingandin developingtutoring systemsthat implementthosehumanteachingstrategiesandtactics. Itconcentratesparticularlyon how systemshave dealtwith studentanswersandhow they havedealtwith motivationalissues,referringparticularlyto work carriedoutatSussex: for example,on respondingeffectively to thestudent’s motivationalstate,on contingentandVygotskianin-spiredteachingstrategiesandon theplausibilityproblem.This latteris concernedwith whethertacticsthat are effectively appliedby humanteacherscan be as effective when embodiedinmachineteachers.

INTR ODUCTION

IntelligentTeachingSystemsandIntelligentLearningEnvironmentsinitially evolvedin aratherlop-sidedway. First, mucheffort wasput into developinghighly detailedmodelsof particulardomains.So, for example,SOPHIEin its variousversionsembodieda highly detailedrepre-sentationof an electroniccircuit at variouslevels of granularity, from a singledevice suchasa resistor, via a functionalsub-structureup to a completepower supply (Brown andBurton,1975). It alsomodelleddiagnostictacticsandstrategy andcouldreactsensiblywhenstudentsexhibited lessthan optimal trouble-shootingbehaviour. It also had, for its time, good natu-ral languagecapabilitiesandcould respondto a wide rangeof domainspecificquestionsandcommands(BurtonandBrown, 1977).Othersystemswereableto exploit their domainknowl-edge,including knowledgeof misunderstandingsof thedomainin orderto make fine-graineddiagnosticjudgementsaboutstudents.For example,Debuggy (Burton,1982)andmorerecentsystemslike it werecapableof building ahighly detailed(student)modelof anindividual’s sub-tractionbehaviour, but left it to humanteachersto embarkon appropriateremediation.Indeedwhethertherewasany valuein undertakingsuchfine-graineddiagnosiswasitself questioned,

asreteachingthewholeprocedureratherthanjust theincorrectlyunderstoodportionseemsjustaseffective (Sleemanetal., 1989).

Second,teachinginvolves a wide variety of communicative activities suchasexplaining,persuading,arguing,demonstrating,describingandsoon,andtheseareskills thatarealsousedin otherthaneducationalcontexts. Onecouldimaginea teachingsystemthatimplementedthismoregeneralcommunicative competenceandthenspecialisedit asneededfor the particulareducationalcontext at the time. The theoriesof teachingthat were implementedin machineteachers1 werenotgroundedin suchgeneralcommunicativecompetence(becauseit wasbeyondthestateof theart) but necessarilytreated“teaching”asan isolatedandlargely self-containedskill.

An exampleof one of theseisolatedand self-containedskills was SocraticTutoring, amethodof teachingbasedon askingthestudenta seriesof carefullyconstructedquestionsthatwould leadstudentsto recognizeandfix gapsandinconsistenciesin what they know of a do-main (Collins et al., 1975). Anotherexampleis provided by thevarioussystemsproducedbyAndersonandhiscolleagueswhichmonitoredstudents’problem-solvingin afine-grainedman-nerandhadthecapabilityof reactingimmediatelyif thestudentdepartedfrom thepaththatanideal studentwould have followed (seee.g.,AndersonandReiser,1985). This endowed ma-chineteachersof that erawith a certaincommunicative brittlenessthat could underminetheirotherskills.

Of course,therewereattemptsto build tutorialsystemswith moreversatileeducationalcom-municative competence.OnesuchsystemwasGUIDON (Clancey, 1982)which incorporatedrulesfor “selectingdiscoursepatterns”,for “choosingdomainknowledge” andfor “maintain-ing the communicationmodel”. Within the category of discoursepatterntherewererulesforrespondingto a studenthypothesis,which resembledSocraticTutoring, as well as rules fordealingwith otheraspectsof theinteraction.

AnothersuchsystemwastheMeno-tutorwhichincorporatedaDiscourseManagementNet-work (Woolf, 1988). Indeed,therearevarioussimilaritiesbetweenthis network andOhlsson’staxonomy(seeFigure1). This systemcouldmake useof thecurrentdiscoursecontext to dis-tinguishandexecutea rangeof differentkindsof tutorial tactics(for example,briefly acknowl-edginga student’s incorrectanswer)andstrategic rules(for example,undertakinga seriesofshallow questionsaboutavarietyof topics).

Sooverall therewasratherunevenprogressin thefollowing areas,with mostsystemshavingrathera restrictedrepertoireof teachingactions,andwork concentratingon (ii) below:

i. Thedevelopmentof avariedrepertoireof teachingactions.

ii. Thedevelopmentof effective strategic andtacticalmeans-endsrulesfor thedeploymentof theteachingactions.

iii. Thedevelopmentof suchbasic,communicative skills andcompetenceasexplainingar-guing,convincing, cajoling,detectingmisunderstandings,dealingwith interruptionsandsideissuesetc.

iv. The developmentof theoriesof motivation and affect that would enablethe judiciouschangeof topic,useof a joke, impositionof a threat,offer of praiseandsoon.

Thekind of criticism thatwaslevelledat machineteacherswasthesameasthatoften lev-elledat AI in general,namelythatthey tendedto concentrateon toy worlds(albeitoftenhighlydetailedtoy worlds)andthatthey tendedto degradebadlywhenmovedoutsidetheirown sphere

1Theterm“machineteachers”is usedasa generaldescriptionof systemsthatadjustthemselvesto theneedsoftheirstudents.Thismaymeanposingaproblem,evaluatinganansweror adjustingthelevel of help,but it couldalsoincludesadjustingsomeaspectof (say)asimulationin anintelligentlearningenvironment,to increasethelikelihoodthatproductive explorationsareundertaken.

2

of competence(seee.g.,Dreyfus, 1979). This meantthat the teachingstyle of mostmachineteacherswas gearedtowardsa rather“convergent”, “syllabus bound” teachingand diagnos-tic style (seee.g.,Ohlsson,1987). By contrast,a humanteacheris able to integrate topicsacrosswidely differingdomains,changestyleandapproachastheoccasiondemands,appealtocommon-senseknowledgeandreasoninganduseall thecommunicationandsocialskills at hisor herdisposal.A machineteacheroftenappearsploddingandrelentless,dominatedby its owndomain-specificknowledgeandunableto deploy any of thatchangeof paceandperspective thatmakesgoodteachingwhatit is.

Anothercriticism of machineteachersis thatthey tendedto embodyamodelof teachinginwhich the teacherknows best. It wasnot that thesystemsof the time werewholly concernedwith “transmitting” knowledgeandmaintainingagency in interactionswith their users:someexcellentsystemswerearrangedaslearningenvironmentswhich reactedintelligently to movesinstigatedby users.But they werenot thekindsof systemsthatcarriedout themorefacilitatingrolesof teachers,suchashelpingstudentsto work moreeffectively together, helpingstudentsreflect on what they had learnedand doneor guiding studentsin open-endedproject work.Thereweregoodreasonsfor this,namelythatmodellingsuchfacilitatingskills neededartificialintelligenceabilitiesthatwere(andlargely still are)beyondthestateof theart. Soit’s not thatthedesignersof suchsystemshadanimpoverishedview of education;it wasmuchmoreacaseof doingwhatwaspossible.

This paperis divided into four moresectionsafter this Introduction. The next section(2)expandsthe critique above by giving a brief accountof Ohlsson’s andothers’analysesof therestrictedteachingcapabilityandversatilityof systemsbuilt upto aboutthemid-eighties.It citesfurtherexamplesfrom thatperiod.Section3 thenexaminesthreemethodologiesfor developingteachingcapabilityandversatility. Themostimportantof thesemethodologiesis theobservationof humanteachers,andexamplesof suchwork aredescribed,concentratingin particularonhowteachersdealwith studenterrorsandhow teachersmotivatestudents.Teachingin all its variedformscoversmany morefactorsthanjustthesetwo,but thesehavebeenchosenasrepresentativeof themodellingeffort of artificial intelligencein education.

In order to seewhat progresshasbeenmadesincethe mid-eighties,Section4 draws thethreadstogetherfrom thehistoricalanalysisandexaminesa numberof contemporarysystemsthatattemptto embodyclever teachingtacticsor strategies,includingonethatattemptsto dealwith motivationalissues.Theconcludingsectionofferssomethoughtsonthedegreeof progressmade.

We mustoffer two immediatedisclaimers.Thepaperdoesnot attemptto beanexhaustivereview of what is known aboutteaching,or of thosepartsof this knowledgethat have beenincorporatedinto systems,thoughit doesprovide a numberof pointersto this large literature.Rather, it attemptsto highlight key issuesandsystems,drawn largely from work atSussex, thatexemplify thecomplexity of this task.

Theseconddisclaimerconcernstheparticularfocusof thepaperonkey aspectsof acertainkind of teaching.We have concentratedlargely on systemsthatembodydomainknowledgeorskills to be learned,ratherthanon moreopen-endedsystems,e.g. that facilitatedialoguebe-tweenstudents.This appearsmore“teacher-centred”than“learner-centred”but is not intendedto expressany valuejudgementbetweenthesetwo differentwaysto conceptualiseeducationalinteractions.

RESTRICTED REPERTOIRE OF TEACHING ACTIONS

Versatilehumanteachershave an enormousrepertoireof teachingactionsat their disposal.Theserangefrom caseswherethe teachergetsthestudentto do almostall thework (“explainthis to me...”, “solve this problem...”, “write anessaycomparing...”, “chooseaprojectto ...”) tocaseswheretheteacherexercisesmoredirectagency (“this is how it is done...”, “think of it this

3

way...”, “if I wereyou,I would...”), via all kindsof intermediateandindirectcases,for examplewherethe teacherorganisesan educationalsettingwhich facilitatesseveral studentsworkingtogetherin an effective way. An educationalencounter, viewed asa kind of ordinaryhumancommunication,canexploit all therichnessof context, modality, interactionandcontentof or-dinaryhumancommunicationaswe experienceit both in day to dayconversations(includingfeaturessuchastonesof voice,irony, humour, glances,silencesandsoon)andin lessinteractiveformssuchastelevision, theatre,books,newspapersetc. Evenin a distanceeducationcontext,the teacherasauthorof thematerialswill be mindful of the learningsituationof thestudents.They will probablybe isolatedandso materials(suchasvideo-clips)and activities (suchasself-assessmentquestions)will beincorporatedto keepthelearnermotivatedandself-reflective.

Ohlsson(1987)providedanexcellentcritiqueof Artificial Intelligencein Education(AIED)in termsof its historicallynarrow focusonmodellinganddiagnosisattheexpenseof (theharder)remedialactionsandteaching.He offeredananalysisof someof themany teachingoperationsthatmight beassociatedwith teachinga procedureof somekind, seeFigure1. Notethat thesewerespecificallyconcernedwith teachinga procedure, thoughsomeof theseoperationsmightalsoapply in teachinga principle or a concept. Theseoperationsincludeteachingactionsas-sociatedwith settingthe sceneaswell aswith indicatingthe nutsandbolts of the procedure.Settingthe scenecaninvolve clarificationof goalsaswell as justificationof individual stepsor pointing out similaritiesto similar proceduresalreadywell understoodby thestudent.It isimportantto reiteratethatteachingcanbemuchmorethanassistingthemasteryof procedures,principlesandconcepts.

PRESENTING TARGET PROCEDURE

Define terms,describe procedureprompt recall

Demonstrateinteractivepromptedannotatedapplied

Practiceguidedannotatedcorrectedhintsdrill

PRESENTING PRECURSORS

PrimingReviewingMarking familiar & unfamil-iar steps

PRESENTING PURPOSES

Giving a goalCriticise precursorsGeneralisation or replace-ment of precursors

PRESENTING JUSTIFICATIONS

AnnotatingTransparent casesEquivalent procedureVerification

alternativeinverseempirical test

DEALING WITH ERRORS

RevealExplainMark

DEALING WITH SOLUTIONS

FeedbackPrompt self-checkPrompt self-reviewPrompt self-annotation

Figure1: TeachingaProcedure:SomePrinciplesof IntelligentTutoring(Adaptedfrom Ohlsson(1987))

4

Perhapsthemostinterestingitemson Ohlsson’s list arethoseassociatedwith reactionstothe studentgettinghis or her answercorrect: actionsdesignedto get the studentto checktherobustness,applicability, assumptionsetcof thesolution.Will it work for all cases?Supposingthe initial conditionshadbeenslightly different? Is it expressedin themostgeneralform? Isit similar to other solutions? Are theresimilarities in the way the solution was constructedcomparedto otherproblem-solvingepisodes?Is thesolutionoptimal?Wastheproblem-solvingoptimal? And so on. The difficulty of achieving this with a machineteacherarisesnot fromany difficulty of posingtheright kindsof questionto thestudent,but of beingableto undertakeany but the mostcursoryanalysisof the student’s answer. An associatedproblemhereis therestrictedmodality (typically text anddiagramson a screen)that wasthenavailable. Thereissomethingratherspecialaboutvoiceandgesturethatcancut throughcomplex materialto revealtheessentialpoint. Thelatestgenerationof embodiedagentsexploit this,aswe seelater.

While it may be possibleto constrainthe languagewith which the studentrefersto theknowledgedomainitself, it is muchharderto constrainthemeta-languagein which thestudentmakesevaluative statementsaboutsolutions.Thestandardtricks at thedomainlevel of usingsemanticgrammars,menusor otherdevicesto restrictinputaremuchlesseasyto applyto meta-languagebecausethedomainhasbeenbroadenedandthestudentis askedto makecomparisonsacrossthe domainor indeedbetweendomains.With notableexceptions,suchasCollins andBrown (1988),thedesignersof systemstargetedby Ohlssonweremuchlessinterestthannowin thewholeissueof metacognitionandtoolsto supportplanningandreflection(say)werelesscommon.

Therealsotendedto beabasicasymmetrybetweenstudentandmachineteacherin thatthecomplexity of the textual or diagrammaticoutputfrom themachineteacherwasusuallyfar inexcessof thecomplexity of typedinput from thestudentthat could beunderstoodby thema-chineteacher. Sometimessystemsallowedstudentsto typein freeform text, but typically thiswaseitherignoredor only partially comprehendedby the system.Of course,a studentcouldbemonitoredinteractively while building up a complex object(e.g. suchasa Lisp programintheLisp tutor: AndersonandReiser,1985)or be provided with a posthocanalysisof a com-plex objectthatheor shehasbuilt (suchasa Pascalprogramin Proust:JohnsonandSoloway,1987),but in generalmachineteachersmademuchhigherdemandson thelanguageandimageunderstandingof their studentsthantheir studentsmadeof them. If thestudentthoughtaboutthedomainin a differentway from themachineteacher, or solvedproblemsin anidiosyncraticmannertherewasusuallynowaythatthestudentcouldtell themachineteacherthatthiswasthecaseor any wayfor themachineteacherto beableto makeanevaluative,comparative commentaboutthestudent’s view or methodin comparisonwith its own.

AlthoughwehaveconcentratedonOhlsson’s (1987)critique,hewasnot theonly researcherto have calledinto questionthe teachingcapabilityof thatgenerationof machineteachers.Inanothercritical analysisof Intelligent ComputerAssistedInstruction(ICAI), Ridgway (1988)summedup theactivities missingfrom machineteacher’s repertoireas:

1. Encouragingexplanationsfrom pupilsto eachotherandto theteacher.

2. Groupwork thatfacilitatesmetacognitive activities andis itself inherentlyvaluable,

3. Metacognitive skills suchasself-explanationandself-evaluation.

Although concernedwith a wider classof systems,namelyadvice-giving expert systems,Carroll andMcKendree(1987)2 criticisedmany approachesthenusedfor machineteachersaslackingin empiricaljustificationor generalisabilityacrossdomains.For example,with respectto SocraticTutoring,they wrote:

2Many of thequestionsraisedin thiswide-rangingsurvey of thefield arestill relevanttoday.

5

“It is likely thatthis is aneffective stylefor interactive tutoringin many situations,but it is alsosignificantthatno evidenceis offeredto supportthisassumption.Thepossibility exists that the Socraticstyle is often adoptedfor tutorially irrelevantreasons.Giving the systemcontrol of the dialogueallows a simplequestion-listknowledgestructure.”

(Carroll andMcKendree,1987,page15)

In summary, thecriticismsabovecanbereducedto two majorissues.First,tutoringsystemshave focusedon too narrow a rangeof typesof educationalinteraction,i.e. taking ratherateacher-centredview of the enterpriseand not attemptinga more learner-centredfacilitatingrole. Second,even within a teacher-centredframework, mostsystemshave adoptedratheranarrow rangeof teachingtacticsandstrategies.

So in principle how could the issuesraisedin thesecriticisms, especiallythe second,bedealtwith?

DEVELOPMENT OF TEACHING STRATEGIES

Thereare threeprincipledmethodologiesfor developing the teachingexpertisein AIED sys-tems.First is theobservationof humanteachersfollowedby anencodingof effective examplesof theseteacher’s expertise,typically in theform of rules.Thesecondis basedon learningthe-ory andderivesa teachingtheoryfrom that. Thethird is basedon observationsof realstudentsor onarunnablesimulationmodelof thestudentandderivesateachingtheoryfrom experimentswith suchstudentsor modelsof students.

Derived from expert human teachers

An influential earlyexampleof themethodologyof learningfrom expert humanteacherswasSocraticTutoring(Collins et al., 1975).SocraticTutoringprovidesa numberof detailedteach-ing tacticsfor eliciting from andthenconfrontinga learnerwith her misconceptionsin somedomain.A generalisationof thisapproach,“Inquiry Teaching”is offeredby CollinsandStevens(1991).A morerecentexampleof thegeneralmethodologyis providedby Lepperet al. (1993)who analysedthe methodsthat humanteachersuseto maintainstudentsin a positive motiva-tionalstatewith respectto their learning.

Thereis an issuein usingexpert teachersasa source.Typically they will have beeninflu-encedto agreateror lesserextentby thetheoriesof teachingandlearningthey wereexposedtoduring training,so thereis a dangerthatonemight besimply observingthesetheoriesfilteredthroughtheir applicationby thechosenexperts. Anotherissueis that therearemany differentstylesandphilosophiesof teaching,further fragmentedby individual personalitydifferencesanddomainnorms.Finally whatteacherssaythey doandwhatthey actuallydomaybeatoddswith eachother, seeBlisset al. (1996),describedlater.

Therearemany waysto examinethe issueof what constitutesexpert humanteachingbe-haviour. Oneway is to have regard for the literatureon expertiseandanalysethe ways thatexpertteachersaresimilar to expertsin otherfields.For example,Sternberg andHorvath(1995)reviewedwork onthestructureof expertteacherknowledge.Typically, this is foundto bemuchmorehighly structuredthanthat of novice, or even experiencedteachers.They looked at the“efficiency” with whichthesekindsof expertsolvedproblemsin theirdomain.They foundthat,aswith otherexperts,automaticitylendsspeed.They alsolooked at “insight” issuesandcon-cludedthatexpertteachershaveabetterinsightinto thedeepratherthanthesurfacestructureoflearningandteachingsituations.

Althoughnot concernedto assistthe implementationof machineteachers,theanalysisbySchoenfeld(1998)of how teachers’context-specificbeliefs,goalsandknowledgeareactivated

6

andinteractin classroomsettingsis couchedin the languageof cognitive science.Theauthoranalysesa numberof mathematicslessonsin enoughdetailto demonstratehow far we yethaveto go in orderto duplicateskilled humanteachingandto offer hints towardsa framework forsuchanenterprise.

Another way is to look at the fine structureof how humanteachersdeal with particularissues,suchasmaintainingstudents’motivation(e.g.,LepperandChabay,1988),offering cor-rective feedback(e.g.,Fox, 1991),or detectingandrepairingdialoguefailures(Douglas,1991).In the latter case,Douglasfound in herstudyof teachersthat “Expert andnovice tutorsmadeaboutthesamenumberof [communication]failures,but theexpertwasmarkedly betterat de-tectingandrepairingthem.” — an observation that supportsSternberg andHorvath’s (1995)analysisof teachingexpertise,above.

Amongthemany studiescitedby Sternberg andHorvath(1995)wasanempiricalanalysisof expertandnoviceteachersin theareaof mathematics(LeinhardtandGreeno,1991).EchoingOhlsson’s (1987)critique,they foundthat

“The expertteachershad,with theclass,a largerepertoireof routines,usuallywithseveralformsof eachone.In somecases,weobservedteachersapparentlyteachingnew routinesto their classes.Themainfeaturesof thesemutuallyknown routineswerethat (a) they werevery flexible, (b) ordercould be shiftedandpiecestakenfrom onesegmentandappliedto another, (c) little or no monitoringof executionwasrequired,and(d) little or no explanationwasrequiredfor carryingthemout.Theseroutineshadsimple,transparentobjectives: to increasetheamountof timethat studentsweredirectly engagedin learningor practicingmathematicsandre-ceiving feedback,to reducethe cognitive load for the teacher, and to establishaframethatpermittedeasytransmissionof informationin mutuallyknown andrec-ognizedsettings.”

(LeinhardtandGreeno,1991,page265)

More recently, in a seriesof studiesGraesseret al. (2000) studiedboth expert teachersandnon-expert humantutors3. They found that even untrainedhumantutorswereextremelyeffectiveandthattheirmethodsdid notseemto correspondto any of thestandardmethodologiessuchasSocraticTutoring(Collins et al., 1975),erroridentificationandcorrection(CorbettandAnderson,1992)or sophisticatedmotivationaltechniques(Lepperet al., 1993). However theyconcludedthat

“Tutorsclearly needto be trainedhow to usethe sophisticatedtutoring skills be-causethey donotroutinelyemergein naturalistictutoringwith untrainedtutors.Webelieve thatthemosteffective computertutor will bea hybrid betweennaturalistictutorial dialogandidealpedagogicalstrategies”

(Graesseretal., 2000,page51)

Theway thatexperthumanteachersinterweave a wide rangeof actionsthatdealwith cog-nitive, metacognitive andaffective issuesis exemplifiedby Lajoie et al.’s (2000)observationof anexpertmedicalinstructor. They observedhow theinstructorassignedrolesto eachof theparticipantsin thesmallgroupof first yearmedicalstudentsdiscussingpatients.

“One studentwasasked to presentan actualpatientcaseto thegroup,describingthepatient’s relevant medicalhistoryandcurrentsituation. A secondstudentwas

3Thedistinctionbetween“teaching”and“tutoring” is opento debate.In referringto humanteachers,wewill usetheterm“tutor” to imply someonewith lessformal trainingin pedagogythana“teacher”.Theliteratureof ArtificialIntelligencein Educationusesthe terms“tutor” and“teacher” interchangeablyirrespective of the expertiseof thesystemsodesignated.

7

then asked to summarizethe samecasebasedon the verbal accountof the firststudent.Next, a third studentwasasked to producea problemlist for the patientat the blackboardwith the assistanceof the otherstudentsin the group. Finally,the fourth studentwasasked to leadthe groupin developinga list of differentialdiagnosesfor thecase.”

(Lajoie etal., 2000,page59)

Throughtheorganizationof thestudent’s rolesandthroughbothspecificandgeneralfeed-backthe instructorwasableto dealwith a numberof issuessimultaneously. At thecognitivelevel, boththedivisionof labourandhis feedbackprovidedscaffolding for thecomplex, cogni-tive taskof arriving at a diagnosis.At themetacognitive level theinteractionbetweenstudentsandtheir feedbackprovidedmodelsfor reflective self-criticalexaminationof how datawasbe-ing usedandhow decisionswerebeingarrived at. At the affective level, encouragementwasbeingprovidedwhenneeded.And finally, at the“communityof practice”level, studentswerebeingapprenticedinto medicallyacceptedwaysof behaving.

It is unrealisticto expectthatan all-embracing,prescriptive theoryof teachingwill easilyemerge given the complex, socialnatureof the enterprise.It would be like expectinga pre-scriptive theoryof “beinga politician” or “beinganactor”. Of coursein eachof theseactivitiesthereareguidelineswhich thenovice teacher(or politician or actor)canmake useof andsometheoriesandpracticaltips on, say, how to beconvincing, how to explain effectively or how toreflecton performance.For example,in the latter case,it canbe very useful for a teachertoseea videotapeof a lessonhe or shehastaughtandto discussthe performancewith a moreexperiencedperson.

But thesetheoriescannever be entirely prescriptive in that the activities do not occur ina vacuumbut often dependfor their effectivenesson the personalitiesof the participants.Anauthoritarian,disciplinedteachermaybejust aseffective (alongcertaindimensions)asa moreeasy-goingpersonwith a laissez-faireapproach.It dependson how well theindividual teachercanexploit his or herown personalitytraitsfor thejob in hand(Rutteret al., 1979).Theaboveshouldnotbetakento meanthatthetheoreticalstudyof teachinginteractionsis misguided,onlythatsocialphenomenaareenormouslysubtle.

Within the overall spaceof possibleteachingsituations,let us examinetwo in particular:dealingwith motivationalissuesanddealingwith errors.Thesetwo have beenchosenbecausethey interactwith eachotherandalsobecausethey nicely illustratea polarisationof emphasisbetweensystemsthat Ohlsson(1987) criticised and what hasbeenachieved sincethen. Ofcourse,thevery notionof “dealingwith errors”betraysanexpectationof correctandincorrectanswers,andthiswill applyonly in a limited kind of learningsituation.

In orderto maintainthesenseof historiccontinuity, theissueof dealingwith errorsis takenfirst. This issuehasbeencentralin thedevelopmentof machineteachers.

Dealingwith errors

Many machineteachershave addressedtheissueof dealingwith studenterrors,so it is naturalthatdesignersof suchsystemshave lookedto humanexpert teachersfor insight. An importantteachingcontroversy aroundsuchsystemsis the issueof the tactical vs. the strategic valueof providing immediatehelp whenerrorsaredetected.Tactically the studentis helpedat thepoint wherethey have thebestchanceof understandingandexploiting thathelp. Strategically,however, they maybebeingdeniedopportunitiesto figureout for themselveswhatwentwrongandwhatto donext.

A review of theliteratureonthecomparisonbetweenexperthumanteachersandthesekindsof intelligent tutoring systemsis provided by Merrill et al. (1992b). They notedthatdifferentexpertsadopteddifferentstylesof tutorial feedback:

8

“Fox (1991)andLepperet al. (1991)arguedthat tutorsusevery subtlefeedbackuponerrorsorobstaclestomaximisestudents’problemsolvingsuccess.TheMcArthuret al. (1990) resultsalsosuggestthat tutors follow students’solutionsvery care-fully, but indicatethat this feedbackcanbevery directive. McArthur et al. arguedthat tutorsgive explicit feedback,sometimeseven telling studentshow to solve aproblem,andcarefullystructurestudents’tasksby remindingthemof theproblemgoals.Littmanetal. (1990)andMerrill etal. (1992a)arguedthatthecontext of theerroris critical in determiningfeedback”

(Merrill etal., 1992b,page283)

Thereare clearly differing stylesof teachingthat position themselves at different pointsin the trade-off betweenproviding too muchhelp (thuspotentiallyinhibiting thedevelopmentof problem-solvingstrategies), and providing too little help (thus risking the possibility thatstudentsbecomelost anddemotivated). For example,in thecontext of industrialtraining, theRecovery Boiler Tutorofferedthetraineeanumberof “precautionarymessages. . .whena full-scaledisasteris imminent” ratherthanspecificallynegative feedbackstatements.Thesemightredirectthestudent:Haveyouconsidered. . . ; or, draw theirattentionto anunobservedrelation-ship: Did younoticethe relationshipbetween. . . ; or, confirmthoseactionswhich arehelpful(Woolf, 1988).

An importantfacetof the issueof how muchspecifichelp to offer is the degreeto whichstudentsare in fact sufficiently self-aware to know when they do in fact needhelp (seee.g.,AlevenandKoedinger,2000).

Merrill et al. (1992b)provide a critical comparisonof a rangeof model-basedtutoringsys-temsin relationto theabove findingsfrom humantutors. An interestingissuethat is exploredaspartof this discussionis theway that theinterfaceto thesystemcanbea crucialelementinhelpingthestudentunderstandthenatureof theproblem-domain,overandbeyondany feedbackthata tutormight additionallyprovide (Reiseretal., 1992).

In comparinghumanandcomputer-basedtutors’ waysof dealingwith errors,Merrill et al.(1992b)notesomesimilarities,e.g. bothhelp“studentsdetectandrepairerrorsandovercomeimpasses”.However, they alsonotea numberof differences.Humantutorsoffer less“explicitverbalizationsof thestudent’s misconception”thanmachineteachers;they aremoreflexible inthetiming andthenatureof their feedback;andthey aremoresubtlein indicatingto thestudentthatanerrorhasoccurred,e.g.via slight pausesor intakesof breath.

Motivatingstudents

An importantaspectof humanteachingexpertisethatfiguresonly weakly in Ohlsson’s (1987)criticismcentresaroundskills employedby teachersto build andmaintainstudents’engagementwith thetaskandtheirmotivationto learn.

Accordingto Lepperetal. (1993)thefocusfor many expertteachersis not justoncognitiveissues,suchaswhattaskto teachnext or whaterrorthestudenthascommitted.They alsofocusstronglyon affective issues.How canthestudentbestimulatedandchallenged?How canthestudent’s confidenceandsenseof controlover thelearningsituationbemaintainedor improved.In pursuit of theseaffective goals,humanteachersareoften indirect in their way of helpingstudentsto setgoalsor in reactingto theirerrors:“Now tell mehowyou got that6?”.

In thetricky issueof detectingthestudent’s currentmotivationalstateLepperet al. (1993)suggestthatexperiencedteachersmake useof “the student’s facialexpressions,bodylanguage,intonation,andotherparalinguisticcues”.

Thereis anotherway of gatheringevidenceaboutthestudentsmotivationalstate.Studenteffort, ratherthanperformance,is alsoa reasonablyreliableindicationof intrinsic motivation(Keller,1983).Learnerswhodisplayahighlevel of effort (detectedthroughtheiractivities,sug-gestions,responses)maydeserve praiseevenwhentheir performanceis non-optimal.Thereis

9

awide literatureon therelationbetweenextrinsic rewards(suchaspraise)andintrinsicmotiva-tion. Eisenberger’s (1992)work suggestsrewardingeffort is effective over thelong-termandanextensive meta-analysisof the literatureby CameronandPierce(1994)suggeststhat (contraryto receivedwisdom)extrinsic rewardssuchaspraisedo notdecreaseintrinsicmotivation.

Observation of humanteachersthus lendsforce to Ohlsson’s (1987) criticisms. Indeed,they go beyond it by pointingout theimportanceof affective issuesin determininghow expertteachersbehave. Before turning to examinesystemsthat post-dateOhlsson’s criticisms, webriefly examinethetwo othermethodologiesfor deriving thebehaviour of amachineteacher.

Derived from learning theory

We have alreadynotedthat thereare threemethodologiesfor deriving a teachingtheory fora machineteacher. The secondof the threemethodologiesstartsfrom a learningtheoryandderivesappropriateteachingtacticsandstrategiesfrom thattheory.

Beforecontinuingwith our discussionsabouthow learningtheorieshave informedthede-signof computersystemsthatteachwe needto saya few wordsaboutthenatureof knowledgeitself. Thenatureof knowledgehaslongbeenanareaof active researchanddiscussionamongstphilosophers,psychologistsandeducationalistsaswell asthoseinvolvedin thedevelopmentofartificial intelligencesystems.Is knowledgeabsoluteor relative? Doesit exist asan externalobjectthatcanbeknown or is it boundup with eachindividual’s environmentandexperience?What are the implicationsof what we belief aboutthe natureof knowledgefor the way weperceive our own knowledgeandtheprocessesby which we acquirethatknowledge?In otherwordshow do our beliefsaboutknowledgeinteractwith theway we understandtheprocessoflearning?Thereis no roomfor usto do justiceto a discussionof thenatureof epistemologicalbeliefsandtheir role in learninghere. We thereforeraisethequestionin the reader’s mind asonethat they would needto considerfor a fuller explorationof thenatureof learningtheories.However, in this currentpaperwe now placeourselvesonestepremoved from this discussionandconcentrateon thenatureof thelearningtheoriesthathave influencedsystemdesignratherthan the epistemologyupon which they are founded. For examplesof work in this areathereaderis referredto von Glasersfeld(1984,1987);Wilson andCole(1991);Wilson (1997).

Let usnow look at somespecificexamplesof learningtheoriesandtheir interactionswithteachingtheories. ConversationTheory (PaskandScott,1975)and its reification in variousteachingsystemsis an exampleof this approach. As with SocraticTutoring, ConversationTheoryis concernedessentiallywith epistemologyratherthanwith affectiveaspectsof teachingandlearning. It is basedon a view of learningconsistingof two interactingprocesses.Oneoperatesat the domainlevel, for instanceaddinglinks, facts,rulesandprinciples. The otherworksat themeta-level, notinggapsandinconsistenciesin what is known at thedomainlevel.Thesetwo processescanoperateinsideanindividual learneror they canbedistributedasrolesbetweenmorethanoneparticipant.Overall thetheorysetsconditionsto ensurethatthelearnerconstructsa multifacetedunderstandingof a domainthat allows her to describe(to herselforto others)the inter-relationshipsbetweenconcepts.In somewaysthis is echoedby the “self-explanation” view of effective learning(Chi et al., 1989). A further exampleof the secondmethodology, whichalsopartiallyaddressessomeof theaffectiveissues,is ContingentTeaching(Wood andMiddleton,1975). Herethe ideais to maintainthe learner’s agency in a learninginteractionby providing only sufficientassistanceatany point to enableherto makeprogressonthetask.Theevaluationof this strategy in thehandsof non-teacherswhohadbeendeliberatelytaughtit shows that it is effective. However it was found sometimesto go againstthe grainfor experiencedteacherswho oftenwish to provide morehelpat variouspointsthanthetheorypermits(Woodetal.,1978).Successfullyhelpingteachersto applya teachingstrategy basedonscaffolding (Woodet al., 1976)canbedifficult, asthestudyby Bliss et al. (1996)shows. Theyfoundthat,evenafterteachers’reflective observationof their own andtheir colleagueslessons,

10

focusingon opportunitiesandmethodsof scaffolding pupils’ learning:

“. . . they professedimproved practiceanddemonstratedgreaterconfidencein dis-cussingscaffolding, but therewasnosignificantincreasein thenumberof instancesthatcouldbedescribedasscaffolding. Whenscaffolds wereusedthesewereusu-ally onaone-to-onebasis.It wasduringthisphasethatwerealisedthatourteacherscould ‘talk scaffolding’ but appearedto implementit only marginally. Their focuswason teachingratherthanon pupils’ learning.”

(Bliss etal., 1996,pages44-45)

The rangeof tutors developedby Andersonand his colleagueshasprovided an influen-tial, andcontroversial,modelof teachingin the areaof dealingwith errors,seefor example,(Andersonand Reiser,1985; Andersonet al., 1985; Corbettand Anderson,1992;Andersonet al., 1995),andmorerecently(Koedingeret al., 1997,2000).Theform of teachinghasbeencharacterisedasone-to-onewith fine-graineddiagnosisandremediationfor multi-stepproblem-solving in a varietyof formal domainssuchasgeometry, programmingandalgebra.It appliesonly to problemswith tightly constrainedsolutionmethodsandthis is clearly limiting in termsof the kinds of educationalinteractionbetweenmachineteacherandstudentthat canbe sup-ported.However, recentwork by Koedingeretal. (1997)usingthePAT systemto teachalgebrahasshown how someof theselimitations canbe circumventedby payingspecialattentiontocontextual factors(e.g.asemphasisedby Schoenfeld,1998).In particular, greatcarewastakento involve theschoolsandtheteacherswho wouldbeusingthesystemandcarefulthoughtwasgiven to the useof the Tutor within theclassroom.The systemwasusednot on a one-to-onebasisbut by teamsof studentswho werealsoexpectedto carryout activities relatedto theuseof thetutor, but not involving thetutor, suchasmakingpresentationsto their peers.This useofexplanationbetweenlearnershelpsto counteroneof Ridgway’s (1988)criticisms,mentionedearlier.

Thedesignof thesesystemsis derived from a theoryof learningthat,at base,providesanaccountof the developmentof expertisewhich explainshow declarative knowledgeis trans-formedinto proceduralknowledgeandhow this transformationcanbe supportedby learningenvironments(Anderson,1990).A consequenceof thetheoryis thatattentionis paidto ensur-ing that learnersarekeptawareof thegoalandsub-goalhierarchyof theproblemsolvingtheyhave embarkedon.

The intelligenceof thesesystemsis deployed in several ways. Model Tracing,basedonrepresentingknowledgeof how to dothetaskin termsof production-rules,is usedto keepclosetrackof all thestudent’s actionsastheproblemis solvedandflag errorsasthey occur, suchasmisplottinga point or enteringa value in an incorrectcell in the spreadsheet.It alsoadjuststhe help feedbackaccordingto the specificproblem-solvingcontext in which it is requested.KnowledgeTracingis usedto choosethenext appropriateproblemsoasto move thestudentsin a timely but effective mannerthroughthecurriculum.

Learningtheoriesarestill beingusedto inform systemdesign:for example,Constructivism(AkhrasandSelf, 2000)andReciprocalTeaching(ChanandChou,1997). In addition,Grand-bastien(1999)stressestheneedfor effective methodsto access,organiseandusetheexpertiseof the teacheror trainer. Startingfrom a modelof learning,Winne (1997)suggestshow stu-dentsmight behelpedto developbetter“self-regulated”learningcapability(i.e. improve theirmetacognitive skills).

Derived from studiesof students

Thethird methodologyfor deriving a teachingtheoryis basedon observationof students.Onemethodologyobserveshow studentsof differenttypesrespondto a particularteachingmethod,for exampleassessinghow studentsof differing ability farewith a particularmachineteacher.

11

Anothermethodologycomparesdiffering methodsacrossstudents,and a third methodologyobservesinteractioneffectsbetweenstudentcharacteristicsandteachingmethods.

Thesemethodologiescomein two forms. Therecanbeempiricalobservationsof real stu-dentsor therecanbeanalysesof thereactionsof simulatedstudents.

Studiesof RealStudents

Within theeducationalliteratureasa wholethereis a hugeliteratureon how studentsof differ-ing characteristicsrespondto differing teachingmethods.Therangeof characteristicsincludegender, ability, learningstyle,backgroundknowledge,ageandsoon. For a review of this hugeareathe readeris referredto CronbachandSnow (1977). Within the artificial intelligenceineducationcommunity, many studieshave lookedat how studentsof differing ability andback-groundrespondto particularsystems.As asinglerecentexampleof thisstyleof work weselectArroyo et al. (2000).They categoriseda cohortof studentsby genderandby level of cognitivedevelopment. They wantedto establishhow variationsin the style of hints in the context ofanarithmeticprograminteractedwith genderandwith cognitive development.Hints variedontwo dimensions:degreeof interactivity andthenatureof thesymbolismused.They looked atthereductionin thenumberof mistakeson a problemfollowing a hint asoneof thedependentvariables.They founda numberof interactioneffects(e.g. that“high cognitive ability studentsdo betterwith highly symbolichintswhile low cognitive ability studentsdo worsewith highlysymbolichints”). Theseandrelatedresultsshouldenabletheprogramto make“macroadaptive”(Shute,1995)changesto its teachingstrategy to suit particularsub-groupsof students.

Despitethehugewealthof work, it is difficult to derive generalguidelinesaboutthediffer-entialeffectof students’characteristicsof sufficientprecisionandreliably to supportthedesignof machineteachers.

In termsof looking at the effectsof variationsof teachingmethod,an importantindirectinfluenceonprogresshasbeenthework of Bloom(1984)andhiscolleagues.They investigatedhow variousfactors,suchascuesandexplanations,reinforcementandfeedback,affect studentlearning,taking conventionalclassroomteachingas the baseline.They found that highly in-dividualisedexpert teaching,shifts thedistribution of achievementscoresof studentsby abouttwo standarddeviationscomparedto the moreusualsituationwhereoneteacherdealswith aclassroomof students.They alsofoundthattherangeof individual differencesreduced.

This two standarddeviation improvement,or Two Sigmashift, hasbecomea goalat whichdesignersof machineteachersaim. A standardmethodof evaluationof sucha systemis tocompareit with conventionalnon-computer-basedinteractionteachingthesametopic, thoughtherehave beensomecomparisonsof “smart” and“dumb” versionsof thesamesoftware. Formoreon evaluationof AIED systems,seedu Boulay(2000);Self (1993).

Studiesof SimulatedStudents

The secondform of the methodologythat is basedon observation of studentsusessimulatedratherthanreal students.The exampleswe know of this methodologytypically comparedif-ferent teachingmethodsacrossidenticalsimulatedstudentsratherthanmodellingstudentsofdiffering characteristicsandobservinghow a particularteachingmethodaffectsthemdifferen-tially. This methodologybuilds a computationalmodelof thelearneror of thelearningprocessandderivesa teachingstrategy or constraintson teachingbehaviour by observingthemodel’sresponseto differentteachingactions.For example,VanLehnetal. (1994)comparedtwo strate-giesfor teachingsubtractionto a productionrulemodelof a subtractionlearnerandconcluded,on the basisof the amountof processingengagedin by the model,that the “equal additions”strategy was more effective than the more widely taught “decomposition”strategy. With asimilar generalmethodologyVanLehn(1987)derived “felicity conditions”for thestructureoftutorial examples,for instancethatthey shouldonly containonenew subprocedure.

12

This methodologyoffersan interestingavenuefor researchbut in termsof makingpredic-tions abouthow real studentswill react,it dependscrucially on the fidelity of the underlyingsimulations.

EXAMPLES OF RECENT PROGRESS

This sectiondoesnot aim to bea comprehensive review of thecurrentstateof theart in mod-elling teachingfor tutoring systems.It offers a numberof examplesof recentandrelativelyrecentsystemsthatattemptto go beyondtherestrictionsoutlinedby Ohlsson(1987)andRidg-way(1988).Two of theexamplesderivedirectly from theanalysisin theprevioussection.Thusthe first subsectionexploits the literatureon expert teachersto tackle the centralissueof theaffectivedimensionin teachinganddescribesasystemconcernedwith modellingteachers’mo-tivationalexpertise.Thesecondsubsectionshows how ContingentTeachinganda Vygotskianlearningtheorycanbeexploitedanddealswith anaspectthathasbeenof centralinterestfromthestartof AIED, namelyadjustingthekind of activities andthehelp provided to studentstosucceedon thoseactivities. Thesystemcandynamicallyadjustthe terminologyit usesto de-scribeits domainto thestudentaswell asmakeadjustmentsto boththecomplexity of thetasksit setsaswell asthehelpit provides.

Becauseof their increasingvisibility, thethird subsectionlooksat severalexamplesof ped-agogicalagentsandexamineshow they areaffecting thedebateaboutmodellingteachers’be-haviour, includingtheissueof theperceivedplausibilityof suchsystems.

Note that one way to track the changesin the teachingability of modernsystemsis toexaminetools for building machineteachers,namelyauthoringsystems.This is not theplaceto survey suchsystemsandan excellent recentsurvey is provided by Murray (1999). He isupbeatabouttheir capabilityto representtutorial strategiesandtacticsbeyondsimpleissuesofcurriculumsequencingandplanning:

“Instructionaldecisionsat themicro level includewhenandhow to give explana-tions, summaries,examples,andanalogies;what type of hinting andfeedbacktogive; andwhat typeof questionsandexercisesto offer thestudent.. . .Also char-acteristicto systemsin this category is the ability to representmultiple tutoringstrategiesand“meta-strategies” that selectthe appropriatetutoring strategy for agivensituation.”

(Murray, 1999,page102)

Amongthemany systemssurveyedby Murray (1999),REDEEMstandsasagoodexampleof an authoringsystemwith the capabilityof specifyinga wide variety of teachingstrategies(Major etal., 1997).Thissystemprovidestoolsfor authors(usuallyteachers)to reuseandreor-ganizeexistingnon-adaptive pagesof tutorial materialinto a responsive andadaptive system.

Adding motivational competence

As we have seen,theoriesof instructionalmotivationelaboratetheinfluenceof issueslike con-fidence,challenge,controlandcuriosityin thelearningprocess(Keller,1983;MaloneandLep-per, 1987)andsuggestinstructionaltacticsto keepthestudentin anoptimal learningstateandprovide appealingandeffective interactions.Of course,it is an openquestionas to how farin practiceoneneedsa separatetheoryof motivation. It could be arguedthat if onegetsthecognitive,metacognitive andcontextual issuesright, thenall will bewell. Eachof theseis itselfa complex issue,andso for thepurposesof progressit seemssensibleto clarify themeansbywhichstudentscanbemotivated.

Work on motivational issuesis proceedingalong two fronts. The first reasonsabouttheaffective stateof theteacher. With therisein interestin creatingpedagogicalagents,increasing

13

attentionis beingpaidto equippingthemwith affective competence.For example,Lesteret al.(1999)describethetechniquesunderpinningCOSMO,a pedagogicalagentthatcanadaptbothits facialexpression,its toneof voice,its gesturesandthestructureof its utterancesto indicateits own affective stateandto addaffective forceandfocusto its interactionswith learners.Bypresentingits reactionsto thestudentin amorevaried,human-like waythehopeis thatstudentswill be bettermotivatedto learnandbetterable to judgewhat is importantin what is beingadvised.Theseissuesareelaboratedin a latersectionof thepaper.

Onanotherfront work is progressingonasystem,MORE,to reasonabouttheaffectivestateof the student(del Soldato,1994;del Soldatoanddu Boulay,1995). To this endimplement-ing motivationaltechniquesdemandsshapingthesystem,includingdomainrepresentationandstudentmodel,in severalaspects.In whatfollows theassumptionis thatthestudentis workingon topicsassignedaspart of a curriculumratherthanworking on topicspurely of their ownchoosing,in which casethemotivation issuesarelikely to bevery different. In particular, thesystemmust:

1. Detectthestudent’s motivationalstate;

2. Reactwith thepurposeof motivatingdistracted,lessconfidentor discontentedstudents,or sustainingthedispositionof alreadymotivatedstudents.

The notionof a system’s reaction— triggeringparticularmotivationaltactics— suggeststhatacomprehensive instructionalplanshouldconsistof a “traditional” instructionalplancom-binedwith a motivationalplan. Wasson(1990)proposedthedivision of instructionalplanninginto two streams:contentplanning(“which topic to teachnext”), followedby delivery planning(“how to teachthenext topic”). At first sightthemotivationalplanseemsto becompletelyem-beddedin thedelivery plan. However, motivationaltacticsdo not alwayssimply completethetraditionalcontentplanning:sometimesthey competewith it aswell. A typicalexampleof sucha conflict is thenecessityfor lessconfidentstudentsto build their confidenceby accumulatingexperienceof success,in which casethesystemcouldprovide problemslikely to bepositivelyanswered— basedon topics that the studentalreadyknows. This is relatedto the needforpracticein learning,an issuenot well exploredin AIED systems.Furthermore,while the de-tectionof a learner’s motivationalprofile shapesthestudentmodel,thesystem’s reaction(e.g.suggestinganeasierproblem,askingapuzzlingquestion)dependson theresourcesavailableinthedomainrepresentation.

Typicaldomain-basedplannersselectactionsaccordingto whetherthelearnerknowsatopicor hasmastereda skill. Themethodologyis twofold: detectingthecurrentstateof thelearner’sknowledgeand skill (studentmodelling) and reactingappropriatelyin order to increasethisknowledgeand skill (teachingexpertise). To take accountof motivational factors,we haveextendedthe twin activities of “detectingthe state”and“reactingappropriately”by addingamotivationalstateandmotivationalplanningto thetraditionalITS architecture.Sometimestheadviceofferedby themotivationalplannerdisagreeswith thedomain-basedplan,while in othercasesbothplanscomplementeachother(delSoldatoandduBoulay,1995).

Let us consider, as an example,the situationin which the studentsucceedsin solving aproblem,in this casefinding the bug in a program. A typical domain-basedplannerwouldacknowledgethe right answerandsuggest(or directly provide) a harderproblem,thusmak-ing surethe studentis traversingthedomainin a progressive manner. Suchbehaviour is wellexemplifiedby Peachey andMcCalla’s (1986)instructionalplanner:whenthe learnermastersan instructionalgoal, the plannerfocusesnext on goalsthat requirethe topic just masteredaspre-requisites,traversingthe domainin the directionof a specificultimategoal. In this case,knowingor not knowingthe topic, or exhibiting or not exhibiting the relevantskill, is theonlyissuein thestudentmodelthatdrivestheselectionof suitableactions.

14

Motivationalplanningtakesinto accountothervariablesin the studentmodelandwidensthe tutor’s spaceof possiblereactions.Justby consideringbinary statesof effort (little/large)andconfidence(low/ok) resultsin four differentsituations,eachonerequiringa suitablesetofactionsfrom thetutor, thebinarystatesof effort andconfidence.

When the student’s confidenceis diagnosedasbeing low, the major goal for the planneris to help the learnerregain a reasonablelevel of confidence,and one of the tacticsfor im-proving confidenceis to increasethe student’s experienceof success. The tutor shouldthenselecta task likely to be solved successfullyagain(e.g. a similar taskto the onethe studenthasjustaccomplished).This is aclearexampleof disagreementbetweenthedomain-basedandthemotivationalplanner, sincesimply traversingthedomainto thenext hardertopic hasbeendeliberatelyavoided.

Ontheotherhand,if providing aright answerrequireslittle effort from thestudent(evenaninsecureone)the tutor shouldmove to hardertasks.The tutor shouldmake thedifficulty-levelpromotionvery clear, bothby praisingthesuccessesobtainedsofarandwarningaboutthenewdifficultieswhich arelikely to beencounteredat theharderlevel. Thestudentthenis preparedto copewith new failureswithout feelingtoode-motivated.

Let usnow considerthecaseof ataskthatdoesnotrequireverymucheffort from anormallyconfidentlearner. For a typical domain-basedplannersucha situationwould beideal,whereasfrom a motivationalperspective thetaskcouldbeperceived asbeingirrelevantor “boring”, orin otherwords,de-motivating. Thetutor shouldthenincreasethedegreeof challengeprovidedby theinteraction,by adjustingthedifficulty level to a harderonewherethestudentwould notalways(easily)performthetask,andsomeeffort wouldhave to bespentto achieve success.

Similar analyseshave beenmadefor caseswherethestudentfails at a taskor givesup ona taskandthesetacticswereimplementedasproductionrulesin themotivationalplanner. Theissuehereis not so muchwhethertheseparticulartacticsarecorrect,but the fact that tacticssuchasthis canbemodelledexplicitly within thesystem.If thesearenot thebestrules,thenotherscanreplacethemwithout having to redesignthesystemfrom scratch.

Theneedfor flexibility in theway thatmotivationaltacticsareimplementis underlinedbytheevaluationof MORE.Oneof theissuesthatemergedwasthereluctanceof somestudentstoacceptcertainteachingtacticsfrom a machineasopposedto a human— therefusalto providehelpwhenaskedor therefusalto allow thestudentto give up on anunsolvedproblem.This isanexampleof the“plausibility problem”,whichwe discusslater.

Judging task difficulty and degreeof assistance

Assumingthat a learneris in a reasonablestateof motivation, the teachercan thenfocusonwhatthelearnershoulddo andhow they shouldbehelped.

Wenow discusstheissueof adjustingthecomplexity of thelearningenvironment,thecon-tentanddifficulty of theactivities, thelanguagein which they areexpressed,andthequality ofhintsandsuggestionsin interactive learningenvironments(ILEs). In particular, wedescribetheeducationalphilosophyunderpinningtheEcolabsoftware.

TheEcolabis anInteractive LearningEnvironment(ILE) whichaimsto helpchildrenaged10–11yearslearnaboutfood chainsandwebs. The Ecolabprovides a flexible environmentwhichcanbeviewedfrom differentperspectivesandrun in differentmodesandin increasinglycomplex phases.In additionto providing thechild with thefacilities to build, activateandob-serveasimulatedecologicalcommunity, theEcolabalsoprovidesthechild with smallactivitiesof differenttypes,suchasfindingoutwhatanimalseat,whicharepredatorsandwhichareprey,establishingtheenergy changesassociatedwith feeding,andsettingup a self-sustainingsmallecosystem.Theactivitiesaredesignedto structurethechild’s interactionswith thesystem.Theyprovide a goal towardswhich thechild’s actionscanbedirectedandvary in thecomplexity oftherelationshipswhich thechild is requiredto investigate.

15

This systemexplorestheway thatVygotsky’s Zoneof ProximalDevelopmentcanbeusedin the designof learnermodels(Luckin, 1998;Luckin anddu Boulay,1999). The theoreticalfoundationrequiresthesystemto adoptthe role of a moreableassistantfor a learner. It mustprovideappropriatelychallengingactivitiesandtheright quantityandqualityof assistance.Thelearnermodelmusttrackboththelearner’s capabilityandherpotentialin orderto maintaintheappropriatedegreeof collaborative assistance.

Oneof thelinks betweenthework on theZPD andthework on motivation is thenotionofeffort. For themotivationalplannertheamountof effort expendedby thestudentis a measureof hermotivationalstate.For Vygotsky, anappropriatedegreeof mentaleffort is apre-requisitefor learning.

TheZoneof ProximalDevelopment(ZPD) (Vygotsky, 1978)is createdwhentwo or morepeopleform a collaborative learningpartnershipin which the moreablemembersenablethelessable membersto achieve their goal. In order for a collaboratorto be successfulin therole of a moreablelearningpartnershemustconstructa sharedsituationdefinition (Wertsch,1984)whereall membershave somecommonknowledgeaboutthecurrentproblem. This in-tersubjectivity canonly beachieved if theteacher/collaborator hasa dynamicrepresentationofthelearner’s currentknowledgeandunderstanding.TheZPD alsohasa spatialanalogywhichquantifiesa learner’s potential(Vygotsky, 1986). It is the fertile areabetweenwhat shecanachieve independentlyandwhat shecanachieve with assistancefrom another. In essencetheZPDrequirescollaborationor assistancefor alearnerfrom anothermoreablepartner. Theactiv-ities which form a partof thechild’s effective educationmustbe(just) beyondtherangeof herindependentability. Thelearningpartnermustprovide appropriatelychallengingactivities andtheright quantityandqualityof assistance.In theEcolabthelearningpartnerrole is adoptedbythesystem,andso the learnermodelmusttrackboth the learner’s capabilityandherpotentialin orderto maintaintheappropriatedegreeof collaborative assistance.

The strongfocuson adaptingto the userby adjustingthe amountof help that is initiallyofferedis similar to the adaptive mechanismsin the SHERLOCKtutors(seee.g.,Katz et al.,1992;Lesgoldet al., 1992). A differencefrom SHERLOCKis that thereis alsoadjustmentboth to the natureof the activities undertaken by usersand to the language in which theseactivities areexpressed.Theworking assumptionis thatmoreabstractlanguageis harderandlearnersmovefrom theconcretetowardtheabstract.An alternativeview mightoffer theabstractterminologyearlierasan aid to generalisation.The emphasiswhich the Ecolabplacesuponextendingthelearnerbeyondwhatshecanachievealoneandthenproviding sufficientassistanceto ensurethatshedoesnot fail alsosetsit apartfrom othersystem’s suchasthatof Becket al.(1997),whichgenerateproblemsof controlleddifficulty andaimto tailor thehintsandhelpthesystemoffersto theindividual’s particularneeds.TheEcolabextendsthework donewith othersystemswhich have usedtheZPD conceptin relationto thelearnermodellingtasksuchasthesystemof Gegg-Harrison(1992)which offers the learnerguidedproblem-solvingsessionsinwhich they aregivenassistancein solvingdifficult Prologproblems.

The Ecolabcanassistthechild in several ways. First, it canoffer 5 levels of gradedhelpspecificto theparticularsituation;second,the difficulty level of theactivity itself canalsobeadjusted(activity differentiation).Finally, thedefinitionof thedomainitself allows topicsto beaddressedby thelearneratvaryinglevelsof complexity and(independently)usingterminologyof varying levels of abstractness.So, for example,activities caninvolve simplebilateralrela-tionshipsbetweensay“rabbits” and“grass”, or the samesimplerelationshipdescribedin themoreabstractterms“herbivore” and“primary consumer”.In addition,morecomplex relation-ships(suchasbetweendistantmembersof thesamefood web)canalsobedescribedeitherinsimpleor moreabstractlanguage.

The Ecolabusesa bayesianbelief network modelof the difficulty of transitionsbetweennodesandthehistoryof successandhelprequiredat previousnodesto decide:

16

� Which nodein the curriculum will be tacklednext — which level of complexity andwhich level of terminologyabstraction.

� Whatlevel of helpwill beoffered.

� How muchactivity differentiationwill beofferedto thechild.

Thesystemmaintainsa modelof theecologycurriculumbasedon a BayesianBelief Net-work. Eachnodein the curriculumrepresentsa rule to be learned. The rulesare linked viapre-requisitesthatimposeapartialorderon therulesin thecurriculum.Thereis astartingnodeandalsoanotionalfinishingnode,namelythemostcomplex ruleexploredvia themostabstractterminology.

Eachnodeis associatedwith aprobabilityvaluethatindicatesthelikelihoodthatthelearnercancomplete,unassisted,activities associatedwith thatnode.Thesystemusesthesevaluestodistinguishnodesthatareeithertoo easy(outsidetheZPD), just too hard(within theZPD) ormuchtoohard(outsidetheZPD) for that learnerto completeunassisted.

Thenext nodefor the learneris chosenfrom thosethatarejust too hard. Thesystemusesdataaboutthe learner’s progresswith previousnodesto setboththedegreeof difficulty of theactivity chosenaswell asthequalityof theinitial helpwhich is offeredif needed.

Onceanactivity is completedtheactualamountof helpthatthelearnerusedis noted.Thismay be more than was expected,if the activity turnedout to be harder, or can be lessthanexpectedif in fact the learnerdid not needany help. The amountof help actuallyprovidedis usedby the systemto updatethe probability valueof masteryat that node. This value isthenpropagatedthroughthenetwork to updatetheprobabilityvaluesat all othernodeslinkedto it via pre-requisites(Luckin, 1998). A nodeis againchosenthat is just too hard. This mayinvolveeitheraprogressionthroughthecurriculum(i.e. to amorecomplex ruleor towardsmoreabstractterminology)or stayingat thecurrentnodeandtacklingadifferentactivity.

Student’s input to theEcolabwaslargely unambiguous,e.g. button pressesto choosedif-ferentaspectsof the interfaceor to chooseanimalsand their actionsto assembleinto smallprograms.Whereinappropriateactionswerechosenthesystemgeneratedhelpmessagesat theappropriatelevel of specificity, dependingon its view of the learner’s degreeof masteryof thetopic. Whenthe learnermadesucha mistake the Ecolabdid not try to reasonaboutwhat thelearnermight have hadin mind. Its modelof the learnerwasan overlay of standardecologyknowledgeexpressedasasetof probabilitiesthatthelearnerhadmasteredeachof thetopics.

An issuethatemergedfrom thework with Ecolab(which is thesubjectof currentresearch)concernsthepupils’ degreeof insightinto theirown stateof learningandpossibleneedfor help,seealso(AlevenandKoedinger,2000)mentionedearlier.

Making the teachermanifestand believable

Thegenerationof systemscritiquedby Ohlsson(1987)realisedtheir teachingexpertisethroughtheir textual interactionswith studentsor throughchangesto the interfacesonto the domainsbeingstudied.With therapidimprovementin graphicalandaudiotechnologymany new possi-bilities for animatedpedagogicalagentspresentthemselves.Suchsystemsstill have to addressthesamerangeof teachingproblemsasbefore,but they cannow bring a wider rangeof tacticsto bear(e.g.achangeof facialexpression,or achangeof verbalemphasis).

Johnsonet al. (2000)describea numberof pedagogicalagentsfor differentdomains,in-cludingSteve(for teachingaboutoperatingmachinery),Adele(for teachingmedicine),HermantheBug (for biology) andCOSMO(for adviceaboutinternetprotocols).They arguethatsuchsystemsbring extrapossibilitiesin thefollowing areas:

17

� Interactive demonstrations:wheretheagentcando thetask,point to itemsin asimulatedenvironmentaswell ashandover the taskto the learnersandcommenton their perfor-mance.

� Navigationalguidance:wheretheagentcanassistthelearnersto find theirwayroundandestablishtheirbearingswithin acomplex VR world.

� Useof gazeandgesture:wheretheagentcanexploit its own gazeandgestureto indicateits currentfocusof attentionin anon-verbalmanner.

� Useof non-verbalfeedbackandconversationalsignals:wheretheagentcanindicatethatthestudenthascompleteda taskcorrectlyor incorrectlyby differentkindsof nodof theheador by changingits facialexpression,or exploit eye contact,or adjusttoneof voiceandemphasis.Thesekindsof capabilitygo someway towardsproviding boththesubtleandnon-intrusive feedbackemployedby experthumanteachers,describedearlier.

� Conveying andeliciting emotion: wherethe agentcan indicatesurprise,pleasure,dis-pleasure,puzzlementandotheremotionsappropriateto thecurrentstateof the learninginteraction.

� Virtual teammates:wheretheagentcanplay therole of oneor moreteammatesin taskswherethelearnerneedsto learnhow to coordinateactionsfor arolewith theactionstakenby otherroleplayers.

Embodiedpedagogicalagentsoffer the possibility that someof the subtletechniquesem-ployed by expert teachers(Merrill et al., 1992b)cannow beappliedby machineteachers.Ofcourse,theseextra possibilitiesbring extra complexity: for example,not just a matterof decid-ing what to sayandwhento sayit, but alsoa matterof exactly how to sayit. So oneof thecentralandlong-standingproblemsof thefield hasre-emergedwith new force.

In additionto any problemsof educationaleffectivenessin practice,machineteachersarevulnerableto whatLepperet al. (1993)call the“Plausibility Problem”:

“Evenif thecomputercouldaccuratelydiagnosethestudent’saffectivestateandevenif thecomputercouldrespondto thatstate(in combinationwith its diagnosisof thelearner’s cognitive state)exactly asa humantutor would, thereremainsonefinal potentialdifficulty: theplausibility, or perhapstheacceptability, problem.Theissuehereis whetherthesameactionsandthesamestatementsthathumantutorsusewill have thesameeffect if deliveredinsteadby a computer, evena computerwith avirtually humanvoice.” (Lepperet al., 1993,page102)

In otherwords,will humanteachingtacticsandstrategies,or tacticsderivedfrom learningtheoriesor learningsystemswork effectively for a machineteacher?We alreadynotedhowstudentsfoundcertainactionsof theMOREmachineteacherunacceptable.For amoreextendeddiscussionof this issueseeLepperet al. (1993);du Boulayetal. (1999).

CONCLUSIONS

How far hasthe situationimproved from that describedby Ohlsson?Sincethe mid-eightiestherehave beentwo very usefuldevelopments.Partly asa resultof thedesireto improve thecapabilitiesof suchsystems,therehasbeenanincreasingamountof researchinto humanexpertteachingpractice.Of course,teachinghasbeenstudiedfor millenia, but themorerecentworkhasstudiedit at a level of granularityandwith thepossibilitythat thetacticsandstrategiesob-servedmightbeimplementable.Thishasleadto agradualfilling in of thejigsaw of capabilities,takingawider rangeof issuesinto accountsuchasmotivationandindividual differences.

18

Second,theadventof pedagogicalagentshasagainthrown thespotlightbackontothewholeissueof teachingexpertiseand the subtletyof learner-teacherinteractions. Thereare someencouragingevaluationsof suchsystems(seee.g.,Lesteret al., 1997)but they alsoraisemanyinterestingand,asyet,unresolvedissuessuchastheir plausibilityandtheiracceptabilityacrossa rangeof educationalcontexts.

TheInternethasbeenahugeinfluenceonthedevelopmentof systemsfor education.Overallthis hasfavouredmore learner-centredapproachesthan thosein which teachingtacticsandstrategiesareto thefore, seeCollins et al. (2000);McCalla(2000)for analysesof thesetrends.Even within a web-basedlearned-centredparadigm,the systemcan make variousautomaticadjustments,e.g. to which pagesaremadeaccessibleto a particularlearner, seeBrusilovskyet al. (1998)for anexample.Theintroductionof networked technologieswhich allow learnersto interactacrosswidely distributedgeographicallocationsenablesinteractionsbetweenhumanlearnersandteacherswhichwerepreviouslyunavailable.Are theissueswhichwerepertinenttotraditionalface-to-facehumanteachingandlearningstill pertinentor shouldwebeexploringthechangesin humanteachingwithin this paradigmin orderto inform our designsfor intelligentsystemsto supportthis learning?

We shouldnot losesightof thestrengthsof machineteachers,despitetheir failings. In ad-dition to beingableto reify thelearningdomainandthelearningor problem-solvingprocessinwaysnoteasilyopento humanteachers,machineteachershave theability to actin apatientandconsistentmanner. Thisconsistency canbebothin termsof theirknowledgeandstrategy aswellasin termsof theiremotionalreactions.As humanteacherswearewell awareof theoccasionalemotionalintensityof certaineducationalinteractionsandwe have alreadycited the studybyBliss et al. (1996) that observed a disparitybetweenwhat teacherssaid they weredoing andwhatthey actuallydid in theclassroom.Machineteachersdonotneedto beprey to theseprob-lems— unless,of course,our theoryof educationsuggestssuchintensityor unpredictabilityneedsto playa role!

In future, asmachineteachersevolve, no doubtwe will seethe emergenceof personalitytypesamongstthem,with somebeingjokey, alert andquick-fire while othersaremorewell-manneredandpedestrian.Eachkind may suit sometypesof studenton someoccasions.Assystemsbecomemoreversatilewemayseetheemergenceof thepossibilityof somenegotiationoverwhatis to belearned:thiswouldbelikely to helpthemotivationissuesmentionedearlier.

We arealsostartingto seetheemergenceof systemsthatmonitor theinteractionsamongststudentswhile they learn in order to ensurethat all partiesplay an effective role. A certainamountcanbe achieved herewithout the needfor complex naturallanguageprocessingtech-niques(seee.g.,Soller,2001).

Acknowledgements

Wethankthereviewersof thefirst draft of thispaperfor their many excellentandhelpful com-ments.

References

Akhras,F. N. andSelf,J.A. (2000).Systemintelligencein constructivist learning.InternationalJournalof Artificial Intelligencein Education, 11. to appear.

Aleven,V. andKoedinger, K. R. (2000). Limitations of studentcontrol: Do studentsknow whentheyneedhelp?In IntelligentTutoringSystems:5thInternationalConference, ITS2000,Montreal, number1839in LectureNotesin ComputerScience,pages292–303.Springer, Berlin.

Anderson,J.R. (1990). TheAdaptiveCharacterof Thought. LawrenceErlbaumAssociates,Hillsdale,New Jersey.

19

Anderson,J.R.,Boyle, C. F., andYost,G. (1985).Thegeometrytutor. In Proceedingsof IJCAI’85, LosAngeles,California, pages1–7.

Anderson,J. R., Corbett,A. T., Koedinger, K. R., andPelletier, R. (1995). Cognitive tutors: Lessonslearned.TheJournalof theLearningSciences, 4(2):167–207.

Anderson,J.R. andReiser, B. J.(1985).TheLISP tutor. BYTE, 10(4):159–175.

Arroyo, I., Beck,J. E., Woolf, B. P., Beal,C. R., andSchultz,K. (2000). Macroadaptinganimalwatchto genderandcognitive differenceswith respectto hint interactivity andsymbolism. In IntelligentTutoring Systems:5th InternationalConference, ITS2000,Montreal, number1839in LectureNotesin ComputerScience,pages574–583.Springer, Berlin.

Beck, J., Stern,M., andWoolf, B. P. (1997). Using the studentmodel to control problemdifficulty.In Jameson,A., Paris,C., andTasso,C., editors,UserModeling: Proceedingsof SixthInternationalConference, UM97, pages278–288,New York. SpringerWien.

Bliss, J., Askew, M., andMacrae,S. (1996). Effective teachingand learning: Scaffolding revisited.Oxford Review of Education, 22(1):37–61.

Bloom,B. S. (1984). The2 sigmaproblem:Thesearchfor methodsof groupinstructionaseffective asone-to-onetutoring. EducationalResearcher, 13(6):4–16.

Brown, J.S.andBurton,R. R. (1975). Multiple representationsof knowledgefor tutorial reasoning.InBobrow, D. G.andCollins,A., editors,RepresentationandUnderstanding, pages311–349.AcademicPress,New York.

Brusilovsky, P., Kobsa,A., andVassileva,J.,editors(1998).AdaptiveHypertextandHypermedia. KluwerAcademicPublishers,Dordrecht.

Burton,R. R. (1982). Diagnosingbugsin a simpleproceduralskill. In Sleeman,D. andBrown, J. S.,editors,IntelligentTutoring Systems, pages157–183.AcademicPress.

Burton,R.R.andBrown,J.S.(1977).Semanticgrammar:A techniquefor constructingnaturallanguageinterfacesto instructionalsystems.TechnicalReport3587,Bolt BeranekandNewmanInc.

Cameron,J.andPierce,W. D. (1994).Reinforcement,reward,andintrinsicmotivation:A meta-analysis.Review of EducationalResearch, 64(3):363–423.

Carroll,J.andMcKendree,J. (1987). Interfacedesignissuesfor advice-giving expertsystems.Commu-nicationsof theACM, 30(1):14–31.

Chan,T.-W. andChou,C.-Y. (1997). Exploringthedesignof computersupportsfor reciprocaltutoring.InternationalJournalof Artificial Intelligencein Education, 8(1):1–29.

Chi, M., Bassok,M., Lewis, M., Reimann,P., andGlaser, R. (1989). Self-explanations:How studentsstudyanduseexamplesin learningto solveproblems.CognitiveScience, 13:145–182.

Clancey, W. J. (1982). Tutoring rulesfor guidinga casemethoddialogue. In Sleeman,D. andBrown,J.S.,editors,IntelligentTutoringSystems, pages201–225.AcademicPress.

Collins, A. andBrown, J. S. (1988). Thecomputerasa tool for learningthroughreflection. In Mandl,H. andLesgold,A., editors,LearningIssuesfor IntelligentTutoring Systems, pages1–18.Springer-Verlag,New York.

Collins,A., Neville, P., andBielaczyc,K. (2000).Theroleof differentmediain designingenvironments.InternationalJournalof Artificial Intelligencein Education, 11(2):144–162.

Collins, A. andStevens,A. L. (1991). A cognitive theoryof inquiry teaching. In Goodyear, P., ed-itor, Teaching Knowledge and Intelligent Tutoring, pages203–230.Ablex PublishingCorporation,Norwood,New Jersey.

20

Collins, A., Warnock,E. H., Aiello, N., andMiller, M. L. (1975). Reasoningfrom incompleteknowl-edge.In Bobrow, D. G. andCollins, A., editors,RepresentationandUnderstanding, pages383–415.AcademicPress,New York.

Corbett,A. T. andAnderson,J.R. (1992).LISPintelligenttutoringsystem:Researchin skill acquisition.In Larkin, J. H. andChabay, R. W., editors,Computer-AssistedInstructionand Intelligent TutoringSystems:SharedGoalsandComplementaryApproaches, pages73–109.LawrenceErlbaum.

Cronbach,L. J.andSnow, R. E. (1977).AptitudesandInstructionalMethods:A Handbookfor Researchon Interactions. Irvington,New York.

delSoldato,T. (1994).Motivationin tutoringsystems.TechnicalReportCSRP303,Schoolof CognitiveandComputingSciences,Universityof Sussex.

del Soldato,T. anddu Boulay, B. (1995). Implementationof motivationaltacticsin tutoring systems.Journalof Artificial Intelligencein Education, 6(4):337–378.

Douglas,S.A. (1991).Tutoringasinteraction:Detectingandrepairingtutoringfailures.In Goodyear, P.,editor, Teaching Knowledge andIntelligentTutoring, pages123–147.Ablex PublishingCorporation,Norwood,New Jersey.

Dreyfus, H. (1979). WhatComputers Can’t Do: TheLimits of Artificial Intelligence. HarperandRow,New York, secondedition.

du Boulay, B. (2000). Canwe learnfrom ITSs? In Gauthier, G., Frasson,C., andVanLehn,K., edi-tors,IntelligentTutoring Systems:Proceedingsof 5th InternationalConference, ITS2000,Montreal,number1839in LecturesNotesin ComputerScience,pages9–17.Springer-Verlag.

du Boulay, B., Luckin, R., anddelSoldato,T. (1999).Theplausibilityproblem:Humanteachingtacticsin the‘hands’of amachine.In Lajoie,S.P. andVivet,M., editors,Artificial Intelligencein Education:OpenLearningEnvironments:New ComputationalTechnologies to SupportLearning, Explorationand Collaboration. Proceedingsof the International Conferenceof the AI-ED Societyon ArtificialIntelligenceandEducation,LeMansFrance, pages225–232.IOSPress.

Eisenberger, R. (1992).Learnedindustriousness.Psychological Review, 99(2):248–267.

Fox, B. A. (1991). Cognitive andinteractionalaspectsof correctionin tutoring. In Goodyear, P., ed-itor, Teaching Knowledge and Intelligent Tutoring, pages149–172.Ablex PublishingCorporation,Norwood,New Jersey.

Gegg-Harrison,T. S. (1992). Adapting instructionto the studentscapabilities. Journal of ArtificialIntelligencein Education, 3(2):169–181.

Graesser, A. C., Person,N., Harter, D., andthe Tutoring ResearchGroup(2000). Teachingtacticsinautotutor. In ModellingHumanTeachingTacticsandStrategies:WorkshopW1at ITS’2000,Montreal.

Grandbastien,M. (1999). Teachingexpertiseis at the coreof ITS research.InternationalJournal ofArtificial Intelligencein Education, 10(3–4):335–349.

Johnson,W. L., Rickel, J. W., and Lester, J. C. (2000). Animatedpedagogicalagents:Face-to-faceinteractionin interactive learningenvironments. International Journal of Artificial IntelligenceinEducation, 11(1):47–78.

Johnson,W. L. andSoloway, E. (1987). Proust:anautomaticdebuggerfor pascalprograms.In Kears-ley, G. P., editor, Artificial Intelligence& Instruction: applicationsand methods. Addison-WesleyPublishing,Reading,Massachusetts.

Katz,S.,Lesgold,A., Eggan,G., andGordin,M. (1992). Modelling thestudentin SherlockII. Journalof Artificial Intelligencein Education, 3(4):495–518.

Keller, J.M. (1983).Motivationaldesignof instruction.In Reigeluth,C. M., editor, Instructional-designTheoriesandModels:AnOverview of their CurrentStatus. LawrenceErlbaum.

21

Koedinger, K. R., Anderson,J.R., Hadley, W. H., andMark, M. A. (1997). Intelligent tutoringgoestoschoolin thebig city. InternationalJournalof Artificial Intelligencein Education, 8(1):30–43.

Koedinger, K. R., Corbett,A. T., Ritter, S., andShapiro,L. J. (2000). Carnegie learning’s cognitivetutor: Summaryresearchresults. Summaryof researchresultsavailablefrom CARNEGIElearning,Pittsburgh,www.carnegielearning.com.

Lajoie,S.P., Wiseman,J.,andFaremo,S.(2000).Tutoringstrategiesfor effective instructionin internalmedicine.In ModellingHumanTeachingTacticsandStrategies:WorkshopW1at ITS’2000,Montreal.

Leinhardt,G. andGreeno,J.G. (1991).Thecognitiveskill of teaching.In Goodyear, P., editor, TeachingKnowledge and IntelligentTutoring, pages233–268.Ablex PublishingCorporation,Norwood,NewJersey.

Lepper, M., Aspinwall, L., Mumme,D., andChabay, R. (1991). Self-perceptionandsocialperceptionprocessesin tutoring: Subtlesocialcontrol strategiesof expert tutors. In Olson,J. andZanna,M.,editors,SelfInferenceProcesses:TheSixthOntarioSymposiumin SocialPsychology, pages217–237.LawrenceErlbaumAssociates,Hillsdale,New Jersey.

Lepper, M. R. andChabay, R. (1988). Socializingthe intelligent tutor: Bringing empathyto computertutors. In Mandl,H. andLesgold,A., editors,LearningIssuesfor IntelligentTutoring Systems, pages242–257.Springer-Verlag,New York.

Lepper, M. R., Woolverton,M., Mumme,D. L., andGurtner, J.-L. (1993). Motivationaltechniquesofexperthumantutors:Lessonsfor thedesignof computer-basedtutors.In Lajoie,S.P. andDerry, S.J.,editors,ComputersasCognitiveTools, pages75–105.LawrenceErlbaum,Hillsdale,New Jersey.

Lesgold,A., Lajoie, S., Bunzo,M., andEggan,G. (1992). Sherlock:A coachedpracticeenvironmentfor anelectronicstroubleshootingjob. In Larkin, J.H. andChabay, R. W., editors,Computer-AssistedInstructionandIntelligentTutoringSystems, pages289–317.LawrenceErlbaumAssociates,Hillsdale,New Jersey.

Lester, J. C., Converse,S. A., Stone,B. A., Kahler, S. A., andBarlow, S. T. (1997). Animatedpeda-gogicalagentsandproblem-solvingeffectiveness:A large-scaleempiricalevaluation. In du Boulay,B. andMizoguchi,R., editors,Artificial Intelligencein Education:Knowledge andMedia in Learn-ing Systems.Proceedingsof the AI-ED 97 World Conferenceon Artificial Intelligencein Education,volume50 of Frontiers in Artificial Intelligence, pages23–30,Kobe,Japan.IOSPress.

Lester, J.C.,Towns,S.G.,andFitzgerald,P. J.(1999).Achieving affectiveimpact:Visualemotivecom-municationin lifelik epedagogicalagents.InternationalJournalof Artificial Intelligencein Education,10(3–4):278–291.

Littman,D., Pinto,J.,andSoloway, E. (1990).Theknowledgerequiredfor tutorial planning:An empir-ical analysis.InteractiveLearningEnvironments, 1(2):124–151.

Luckin, R. (1998). ‘ECOLAB’: Explorationsin theZoneof ProximalDevelopment.TechnicalReportCSRP386,Schoolof CognitiveandComputingSciences,Universityof Sussex.

Luckin, R. anddu Boulay, B. (1999). Ecolab:Thedevelopmentandevaluationof a Vygotskiandesignframework. InternationalJournalof Artificial Intelligencein Education, 10(2):198–220.

Major, N., Ainsworth, S., andWood,D. (1997). REDEEM:Exploiting symbiosisbetweenpsychologyandauthoringenvironments.InternationalJournalof Artificial Intelligencein Education, 8(3–4):317–340.

Malone,T. andLepper, M. R. (1987). Making learningfun. In Snow, R. andFarr, M., editors,Aptitude,LearningandInstruction: ConativeandAffectiveProcessAnalyses. LawrenceErlbaum.

McArthur, D., Stasz,C., andZmuidzinas,M. (1990). Tutoring techniquesin algebra. Cognition andInstruction, 7:197–244.

22

McCalla, G. (2000). Life and learningin the electronicvillage: The importanceof localizationforthe designof environmentsto supportlearning. In Intelligent Tutoring Systems:5th InternationalConference, ITS2000,Montreal, number1839in LectureNotesin ComputerScience,pages31–32.Springer, Berlin.

Merrill, D. C.,Reiser, B. J.,Beekelaar, R.,andHamid,A. (1992a).Makingprocessesvisible: Scaffoldinglearningwith reasoning-congruentrepresentations.In Frasson,C., Gauthier, G., andMcCalla,G. I.,editors,IntelligentTutoringSystems:SecondInternationalConference, ITS’92,Montreal, pages103–110,New York. SpringerVerlag.

Merrill, D. C., Reiser, B. J., Ranney, M., andTrafton, J. G. (1992b). Effective tutoring techniques:Acomparisonof humantutorsandintelligenttutoringsystems.TheJournal Of TheLearningSciences,2(3):277–305.

Murray, T. (1999).Authoringintelligenttutoringsystems:An analysisof thestateof theart. InternationalJournalof Artificial Intelligencein Education, 10(1):98–129.

Ohlsson,S. (1987). Someprinciplesof intelligent tutoring. In Lawler, R. W. andYazdani,M., edi-tors, LearningEnvironmentsand Tutoring Systems:LearningEnvironmentsand Tutoring Systems,volume1, pages203–237.Ablex Publishing,Norwood,New Jersey.

Pask,G. andScott,B. (1975).CASTE:A systemfor exhibiting learningstrategiesandregulatinguncer-tainties.InternationalJournalof Man-MachineStudies, 5:17–52.

Peachey, D. andMcCalla,G. (1986).Usingplanningtechniquesin intelligenttutoringsystems.Interna-tional Journalof Man-machineStudies, 24:77–98.

Reiser, B. J., Kimberg, D. Y., Lovett, M. C., andRanney, M. (1992). Knowledgerepresentationandexplanationin GIL, anintelligenttutor for programming.In Larkin, J.H. andChabay, R. W., editors,Computer-AssistedInstructionand IntelligentTutoring Systems, pages112–149.LawrenceErlbaumAssociates,Hillsdale,New Jersey.

Ridgway, J. (1988). Of courseICAI is impossible. . .worsethough, it might be seditious. In Self,J., editor, Artificial IntelligenceandHumanLearning, pages28–48.ChapmanandHall Computing,London.

Rutter, M., Maughan,B., Mortimore, P., andOuston,J. (1979). FifteenThousandHours: SecondarySchoolsandTheirEffectsonChildren. PaulChapmanPublishing.

Schoenfeld,A. H. (1998). Towardsa theoryof teaching-in-context. Issuesin Education, 4(1):1–94.http://www-gre.berkeley.edu/faculty/aschoenfeld.

Self,J. (1993).Specialissueonevaluation.Journalof Artificial Intelligencein Education, 4(2/3).

Shute,V. J. (1995). SMART: Studentmodellingapproachfor responsive tutoring. UserModellingandUser-AdaptedInteraction, 5(1):1–44.

Sleeman,D., Kelly, A., Martinak,R.,Ward,R., andMoore,J.(1989).Studiesof diagnosisandremedia-tion with highschoolalgebrastudents.CognitiveScience, 13(4):551–568.

Soller, A. L. (2001). Supportingsocialinteractionin anintelligentcollaborative learningsystem.Inter-nationalJournalof Artificial Intelligencein Education, 12(1):40–62.

Sternberg,R.J.andHorvath,J.A. (1995).A prototypeview of expertteaching.EducationalResearcher,24(6):9–17.

VanLehn,K. (1987).Learningonesubprocedureperlesson.Artificial Intelligence, 31(1):1–40.

VanLehn,K., Ohlsson,S.,andNason,R. (1994).Applicationsof simulatedstudents.Journalof ArtificialIntelligencein Education, 5(2):135–175.

von Glasersfeld,E. (1984). An introductionto radicalconstructivism. In Watzlawick, P., editor, TheInventedReality, pages17–40.W.W. Norton& Company, New York.

23

von Glasersfeld,E. (1987). Learningasa constructive activity. In Janvier, C., editor, Problemsof rep-resentationin theteaching andlearningof mathematics, pages3–17.LawrenceErlbaumAssociates,New Jersey.

Vygotsky, L. S. (1978). Mind in Society:theDevelopmentof Higher Psychological Processes. HarvardUniversitypress,Cambridge,MA.

Vygotsky, L. S.(1986).ThoughtandLanguage. TheM.I.T. Press,Cambridge,MA.

Wasson,(Brecht),B. (1990). Determiningthe Focusof Instruction: ContentPlanning for IntelligentTutoring Systems. PhD thesis,Departmentof ComputationalScience,University of Saskatchewan,Canada.

Wertsch,J. (1984). The zoneof proximal development:Someconceptualissues. In Rogoff, B. andWertsch,J.,editors,Children’s Learningin the“Zone of ProximalDevelopment”, volume23, pages7–19.Jossey-Bass,SanFrancisco.

Wilson, B. (1997). Thepostmodernparadigm.In Dills, C. andRomiszowski, A., editors,Instructionaldevelopmentparadigms. EducationalTechnologyPublications,EnglewoodCliffs NJ.

Wilson,B. andCole,P. (1991).A review of cognitiveteachingmodels.EducationalTechnologyResearchandDevelopment, 39(4):47–64.

Winne,P. H. (1997). Experimentingto bootstrapself-regulatedlearning. Journal of EducationalPsy-chology, 89(3):397–410.

Wood,D. J.,Bruner, J.S.,andRoss,G. (1976).Theroleof tutoringin problemsolving.Journalof ChildPsychologyandPsychiatry, 17:89–100.

Wood, D. J. and Middleton, D. J. (1975). A study of assistedproblemsolving. British Journal ofPsychology, 66:181–191.

Wood,D. J., Wood,H. A., andMiddleton,D. (1978). An experimentalevaluationof four face-to-faceteachingstrategies. InternationalJournalof Behavioural Development, 1:131–147.

Woolf, B. P. (1988).Representingcomplex knowledgein anintelligentmachinetutor. In Self,J.,editor,Artificial IntelligenceandHumanLearning, pages3–27.ChapmanandHall Computing,London.

24

Documents

Modelling human teaching tactics and strategies for