� Many of the intelligent tutoring systems that havebeen developed during the last 20 years haveproven to be quite successful, particularly in thedomains of mathematics, science, and technology.They produce significant learning gains beyondclassroom environments. They are capable ofengaging most students’ attention and interest forhours. We have been working on a new generationof intelligent tutoring systems that hold mixed-initiative conversational dialogues with the learn-er. The tutoring systems present challenging prob-lems and questions to the learner, the learner typesin answers in English, and there is a lengthy mul-titurn dialogue as complete solutions or answersevolve. This article presents the tutoring systemsthat we have been developing. AUTOTUTOR is a con-versational agent, with a talking head, that helpscollege students learn about computer literacy.ANDES, ATLAS, AND WHY2 help adults learn aboutphysics. Instead of being mere information-deliv-ery systems, our systems help students activelyconstruct knowledge through conversations.

Intelligent tutoring systems (ITSs) are clearlyone of the successful enterprises in AI.There is a long list of ITSs that have been

tested on humans and have proven to facilitatelearning. There are well-tested tutors of alge-bra, geometry, and computer languages (suchas PACT [Koedinger et al. 1997]); physics (suchas ANDES [Gertner and VanLehn 2000; VanLehn1996]); and electronics (such as SHERLOCK [Les-gold et al. 1992]). These ITSs use a variety ofcomputational modules that are familiar tothose of us in the world of AI: production sys-tems, Bayesian networks, schema templates,theorem proving, and explanatory reasoning.According to the current estimates, the arsenalof sophisticated computational modules inher-ited from AI produce learning gains of approx-

imately .3 to 1.0 standard deviation units com-pared with students learning the same contentin a classroom (Corbett et al. 1999).

The next generation of ITSs is expected to goone step further by adopting conversationalinterfaces. The tutor will speak to the studentwith an agent that has synthesized speech,facial expressions, and gestures, in addition tothe normal business of having the computerdisplay text, graphics, and animation. Animat-ed conversational agents have now been devel-oped to the point that they can be integratedwith ITSs (Cassell and Thorisson, 1999; John-son, Rickel, and Lester 2000; Lester et al. 1999).Learners will be able to type in their responsesin English in addition to the conventionalpoint and click. Recent developments in com-putational linguistics (Jurafsky and Martin2000) have made it a realistic goal to havecomputers comprehend language, at least toan extent where the ITS can respond withsomething relevant and useful. Speech recog-nition would be highly desirable, of course, aslong as it is also reliable.

At this point, we are uncertain whether theconversational interfaces will produce incre-mental gains in learning over and above theexisting ITSs (Corbett et al. 1999). However,there are reasons for being optimistic. One rea-son is that human tutors produce impressivelearning gains (between .4 and 2.3 standarddeviation units over classroom teachers), eventhough the vast majority of tutors in a school’ssystem have modest domain knowledge, haveno training in pedagogical techniques, andrarely use the sophisticated tutoring strategiesof ITSs (Cohen, Kulik, and Kulik 1982; Graess-er, Person, and Magliano, 1995).

A second reason is that there are at least twosuccess cases, namely, the AUTOTUTOR and

Articles

WINTER 2001 39

Intelligent TutoringSystems with Conver-

sational DialogueArthur C. Graesser, Kurt VanLehn, Carolyn P. Rosé,

Pamela W. Jordan, and Derek Harter

Copyright © 2001, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2001 / $2.00

AI Magazine Volume 22 Number 4 (2001) (© AAAI)

This article describes some of the tutoringsystems that we are developing to simulateconversational dialogue. We begin with AUTO-TUTOR. Then we describe a series of physicstutors that vary from conventional ITS systems(the ANDES tutor) to agents that attempt to com-prehend natural language and plan dialoguemoves (ATLAS and WHY2).

AUTOTUTOR

The Tutoring Research Group (TRG) at the Uni-versity of Memphis developed AUTOTUTOR tosimulate the dialogue patterns of typicalhuman tutors (Graesser et al. 1999; Person etal. 2001). AUTOTUTOR tries to comprehend stu-dent contributions and simulate dialoguemoves of either normal (unskilled) tutors orsophisticated tutors. AUTOTUTOR is currentlybeing developed for college students who aretaking an introductory course in computer lit-eracy. These students learn the fundamentals

ATLAS systems that we discuss in this article.AUTOTUTOR (Graesser et al. 1999) is a fully auto-mated computer tutor that has tutoredapproximately 200 college students in anintroductory course in computer literacy. Anearly version of AUTOTUTOR improved learningby .5 standard deviation units (that is, about ahalf a letter grade) when compared to a con-trol condition where students reread yokedchapters in the book. ATLAS (VanLehn et al.2000) is a computer tutor for college physicsthat focuses on improving students’ conceptu-al knowledge. In a recent pilot evaluation, stu-dents who used ATLAS scored .9 standard devi-ation units higher than students who used asimilar tutoring system that did not use natur-al language dialogues. Thus, it appears thatthere is something about conversational dia-logue that plays an important role in learning.We believe that the most effective tutoring sys-tems of the future will be a hybrid betweennormal conversational patterns and the idealpedagogical strategies in the ITS enterprise.

Articles

40 AI MAGAZINE

Figure 1. A Screen Shot of AUTOTUTOR.

of computer hardware, the operating system,and the internet.

Figure 1 is a screen shot that illustrates theinterface of AUTOTUTOR. The left window has atalking head that acts as a dialogue partnerwith the learner. The talking head deliversAUTOTUTOR’S dialogue moves with synthesizedspeech, intonation, facial expressions, nods,and gestures. The major question (or problem)that the learner is working on is both spokenby AUTOTUTOR and is printed at the top of thescreen. The major questions are generated sys-tematically from a curriculum script, a modulethat we discuss later. AUTOTUTOR’S major ques-tions are not the fill-in-the-blank, true-false, ormultiple-choice questions that are so popularin the U.S. educational system. Instead, thequestions invite lengthy explanations anddeep reasoning (such as why, how, and what-ifquestions). The goal is to encourage students toarticulate lengthier answers that exhibit deepreasoning rather than deliver short snippets ofshallow knowledge. There is a continuous mul-titurn tutorial dialogue between AUTOTUTOR

and the learner during the course of answeringa deep-reasoning question. When consideringboth the learner and AUTOTUTOR, it typicallytakes 10 to 30 turns during the tutorial dia-logue to answer a single question from the cur-riculum script. The learner types in his/hercontributions during the exchange by key-board, as reflected in the bottom window. Forsome topics, as in figure 1, there are graphicdisplays and animation, with components thatAUTOTUTOR points to. AUTOTUTOR was designedto be a good conversational partner that com-prehends, speaks, points, and displays emo-tions, all in a coordinated fashion.

An Example AUTOTUTOR-Learner DialogueFigure 2 shows a dialogue between a college stu-dent and AUTOTUTOR. Prior to this question, thestudent had been asked and attempted toanswer 6 previous questions about the internet.The internet was the macrotopic, and studentswere tutored by answering several deep-reason-ing questions about the internet. It should benoted that this is not a fabricated toy conversa-tion. It is a bona fide dialogue from our corpusof approximately 200 AUTOTUTOR-student dia-logues in a computer literacy course.

AUTOTUTOR begins this exchange by asking ahow question in turn 1: What hardware do youneed to take photos and send them over theinternet? However, AUTOTUTOR doesn’t merelypop the question out of the blue. It first pre-sents a discourse marker that signals a changein topic (All right, let’s go on.), presents a con-

text to frame the question (You want to takephotos and send them over the internet.), andthen presents a discourse marker that signalsthe questions (Consider this problem.). There-fore, AUTOTUTOR monitors different levels of dis-course structure and functions of dialoguemoves. AUTOTUTOR inserts appropriate discoursemarkers that clarify these levels and functionsto the learner. Without these discourse markers,learners are confused about what AUTOTUTOR isdoing and what they are supposed to do next.A dialogue advancer network (DAN) has beendesigned to manage the conversational dia-logue (Person et al. 2001). The DAN is a finite-

Articles

WINTER 2001 41

TUTOR-1: Alright, let's go on. You want to take photos and send them over the Internet. Consider this problem. What hardware do you need to take photos and send them over the Internet? STUDENT-1: First of all you need to have access to the internet.TUTOR-2: What else?STUDENT-2: a camera TUTOR-3: Yeah. For what type of connection do you need a network card?STUDENT-3: network cardTUTOR-4: Okay. Now, if you have access to the Internet through a network card, then your connection is… STUDENT-4: I don't know.TUTOR-5: Is direct. Okay. A network card is needed if you have a direct connection to the Internet. Okay. How does the user get hooked up to the internet?STUDENT-5: by a plug.TUTOR-6: Once again, a network card is needed if you have a direct connection to the Internet. Okay, what does a modem do?STUDENT-6: connects network card and internetTUTOR-7: Okay. Try this. A modem is required when there is… STUDENT-7: a connectionTUTOR-8: a dial up connection....

TUTOR-30: Let’s review. To send your photos on the Internet, you need either a digital camera or a regular camera to take the photos. If you use a regular camera, you need a scanner to scan them onto a computer disk. If you have a direct connection to the Internet, then you need a network card. A modem is needed if you have a dial up connection.

Figure 2. An Example Tutorial Dialogue between AUTOTUTOR and a Learner.

if you have a direct connection to theinternet.

Expectation 4: A modem is needed if youhave a dial-up connection.

AUTOTUTOR decides which expectation tohandle next and then selects dialogue movesthat flesh out the expectation. The dialoguemoves vary in directness and information con-tent. The most indirect dialogue moves arehints, the most direct are assertions, andprompts are in between. Hints are often articu-lated in the form of questions, designed to leadthe learner to construct the expected informa-tion. Assertions directly articulate the expectedinformation. Prompts try to get the learner toproduce a single word in the expectation. Forexample, the tutor turns 3, 4, 5, and 6 in figure2 are all trying to get the learner to articulateexpectation 3. Hints are in the tutor-3 turn (Forwhat type of connection do you need a net-work card?) and the tutor-5 turn (How does theuser get hooked up to the internet?). Promptsare in tutor-4 (If you have access to the internetthrough a network card, then your connectionis…, with a hand gesture encouraging thelearner to type in information). Assertions arein tutor-5 and tutor-6 (A network card is need-ed if you have a direct connection to the inter-net.). AUTOTUTOR attempts to get the learner toarticulate any given expectation E by goingthrough two cycles of hint-prompt-assertion.Most students manage to articulate the expec-tation within the six dialogue moves (hint-prompt-assertion-hint-prompt-assertion).AUTOTUTOR exits the six-move cycle as soon asthe student has articulated the expectedanswer. Interestingly, sometimes students areunable to articulate an expectation even afterAUTOTUTOR spoke it in the previous turn. Afterexpectation E is fleshed out, AUTOTUTOR selectsanother expectation.

How Does AUTOTUTOR Know Whether a Student Has Covered an Expectation?AUTOTUTOR does a surprisingly good job evalu-ating the quality of the answers that learnerstype in. AUTOTUTOR attempts to “comprehend”the student input by segmenting the contribu-tions into speech acts and matching the stu-dent’s speech acts to the expectations. Latentsemantic analysis (LSA) is used to computethese matches (Landauer, Foltz, and Laham1998). When the tutor’s expectation E is com-pared with the learner’s speech act A, a cosinematch score is computed that varies from 0 (nomatch) to 1.0 (perfect match). AUTOTUTOR con-siders each combination of speech acts that thelearner makes during the evolution of an

state automaton that can handle different class-es of information that learners type in. TheDAN is augmented by production rules that aresensitive to learner ability and several parame-ters of the dialogue history.

How Does AUTOTUTOR Handle the Stu-dent’s Initial Answer to the Question?After AUTOTUTOR asks the question in the tutor-1 turn, the student gives an initial answer inthe student-1 turn. The answer is very incom-plete. A complete answer would include all thepoints in the summary at the final turn (tutor-30). What does AUTOTUTOR do with this incom-plete student contribution? AUTOTUTOR doesn’tsimply grade the answer (for example, good,bad, incomplete, a quantitative score) as manyconventional tutoring systems do. AUTOTUTOR

also stimulates a multiturn conversation that isdesigned to extract more information from thestudent and get the student to articulate piecesof the answer. Thus, instead of being an infor-mation-delivery system that bombards the stu-dent with a large volume of information, AUTO-TUTOR is a discourse prosthesis that attempts toget the student to do the talking and exploreswhat the student knows. AUTOTUTOR adopts theeducational philosophy that students learn byactively constructing explanations and elabo-rations of the material (Chi et al. 1994; Conatiand VanLehn 1999).

How Does AUTOTUTOR Get the Learner to Do the Talking?AUTOTUTOR has a number of dialogue moves toget the learner to do the talking. For starters,there are open-ended pumps that encouragethe student to say more, such as What else? inthe tutor-2 turn. Pumps are frequent dialoguemoves after the student gives an initial answer,just as is the case with human tutors. The tutorpumps the learner for what the learner knowsbefore drilling down to specific pieces of ananswer. After the student is pumped for infor-mation, AUTOTUTOR selects a piece of informa-tion to focus on. Both human tutors and AUTO-TUTOR have a set of expectations about whatshould be included in the answer. What theydo is manage the multiturn dialogue to coverthese expected answers. A complete answer tothe example question in figure 2 would havefour expectations, as listed here:

Expectation 1: You need a digital cameraor regular camera to take the photos.

Expectation 2: If you use a regular cam-era, you need to scan the pictures onto thecomputer disk with a scanner.

Expectation 3: A network card is needed

AUTOTUTOR isa discourseprosthesis

that attemptsto get the

student to dothe talking

and exploreswhat the

studentknows.

Articles

42 AI MAGAZINE

answer to a major question; the value of thehighest cosine match is used when computingwhether the student covers expectation E. LSAis a statistical, corpus-based method of repre-senting knowledge. LSA provides the founda-tion for grading essays, even essays that are notwell formed grammatically, semantically, andrhetorically. LSA-based essay graders can assigngrades to essays as reliably as experts in compo-sition (Landauer et al. 1998). Our research hasrevealed that AUTOTUTOR is almost as good asan expert in computer literacy in evaluatingthe quality of student answers in the tutorialdialogue (Graesser et al. 2000).

How Does AUTOTUTOR Select the Next Expectation to Cover?AUTOTUTOR uses LSA in conjunction with vari-ous criteria when deciding which expectationto cover next. After each student turn, AUTOTU-TOR updates the LSA score for each of the fourexpectations listed earlier. An expectation isconsidered covered if it meets or exceeds somethreshold value (for example, .70 in our cur-rent tutor). One selection criterion uses thezone of proximal development to select thenext expectation, which is the highest LSAscore that is below threshold. A second criteri-on uses coherence, the expectation that hasthe highest LSA overlap with the previousexpectation that was covered. Other criteriathat are currently being implemented are pre-conditions and pivotal expectations. Ideally,AUTOTUTOR will decide to cover a new expecta-tion in a fashion that both blends into the con-versation and that advances the agenda in anoptimal way. AUTOTUTOR generates a summaryafter all the expectations are covered (for exam-ple, the tutor-30 turn).

How Does AUTOTUTORGive Feedback to the Student?There are three levels of feedback. First, there isbackchannel feedback that acknowledges thelearner’s input. AUTOTUTOR periodically nodsand says uh-huh after learners type in impor-tant nouns but is not differentially sensitive tothe correctness of the student’s nouns. Thebackchannel feedback occurs online as thelearner types in the words of the turn. Learnersfeel that they have an impact on AUTOTUTOR

when they get feedback at this fine-grain level.Second, AUTOTUTOR gives evaluative pedagogicalfeedback on the learner’s previous turn based onthe LSA values of the learner’s speech acts. Thefacial expressions and intonation convey differ-ent levels of feedback, such as negative (forexample, not really while head shakes), neutralnegative (okay with a skeptical look), neutral

positive (okay at a moderate nod rate), and pos-itive (right with a fast head nod). Third, there iscorrective feedback that repairs bugs and miscon-ceptions that learners articulate. Of course,these bugs and their corrections need to beanticipated ahead of time in AUTOTUTOR’S cur-riculum script. This anticipation of contentmimics human tutors. Most human tutorsanticipate that learners will have a variety ofparticular bugs and misconceptions when theycover particular topics. An expert tutor oftenhas canned routines for handling the particularerrors that students make. AUTOTUTOR currentlysplices in correct information after these errorsoccur, as in turn tutor-8. Sometimes studenterrors are ignored, as in tutor-4 and tutor-7.These errors are ignored because AUTOTUTOR hasnot anticipated them by virtue of the contentin the curriculum script. AUTOTUTOR evaluatesstudent input by matching it to what it knowsin the curriculum script, not by constructing anovel interpretation from whole cloth.

How Does AUTOTUTOR Handle Mixed-Initiative Dialogue?We know from research on human tutoringthat it is the tutor who controls the lion’s shareof the tutoring agenda (Graesser, Person, andMagliano 1995). Students rarely ask informa-tion-seeking questions and introduce new top-ics. However, when learners do take the initia-tive, AUTOTUTOR needs to be ready to handlethese contributions. AUTOTUTOR does a moder-ately good job in managing mixed-initiativedialogue. AUTOTUTOR classifies the learner’sspeech acts into the following categories:

Assertion (RAM is a type of primary mem-ory.)

WH-question (What does bus mean andother questions that begin with who,what, when, where, why, how, and so on.)

YES-NO question (Is the floppy disk work-ing?)

Metacognitive comment (I don’t under-stand.)

Metacommunicative act (Could you re-peat that?)

Short response (okay, yes)

Obviously, AUTOTUTOR’S dialogue moves onturn N + 1 need to be sensitive to the speechacts expressed by the learner in turn N. Whenthe student asks a What does X mean? ques-tion, the tutor answers the question by givinga definition from a glossary. When the learnermakes an assertion, the tutor evaluates thequality of the assertion and gives short evalua-tive feedback. When the learner asks, What did

Our researchhas revealedthatAUTOTUTOR isalmost asgood as anexpert incomputerliteracy inevaluating thequality ofstudentanswers in thetutorialdialogue ….

Articles

WINTER 2001 43

riculum script with deep-reasoning questionsand problems. The developer then computesLSA vectors on the content of the curriculumscripts. A glossary of important terms and theirdefinitions is also prepared. After that, thebuilt-in modules of AUTOTUTOR do all the rest.AUTOTUTOR is currently implemented in JAVA forPENTIUM computers, so there are no barriers towidespread use.

ANDES: A Physics Tutoring System That Does Not Use Natural Language

The goal of the second project is to use naturallanguage–processing technology to improvean already successful intelligent tutoring sys-tem named ANDES (Gertner and VanLehn2000; VanLehn 1996). ANDES is intended to beused as an adjunct to college and high-schoolphysics courses to help students do theirhomework problems.

Figure 3 shows the ANDES screen. A physicsproblem is presented in the upper-left window.Students draw vectors below it, define variables

you say? AUTOTUTOR repeats what it said in thelast turn. The DAN manages the mixed-initia-tive dialogue.

The Curriculum ScriptAUTOTUTOR has a curriculum script that orga-nizes the content of the topics covered in thetutorial dialogue. There are 36 topics, one foreach major question or problem that requiresdeep reasoning. Associated with each topic area set of expectations, a set of hints and promptsfor each expectation, a set of anticipated bugs-misconceptions and their corrections, and(optionally) pictures or animations. It is veryeasy for a lesson planner to create the contentfor these topics because they are Englishdescriptions rather than structured code. Ofcourse, pictures and animations would requireappropriate media files. We are currently devel-oping an authoring tool that makes it easy tocreate the curriculum scripts. Our ultimate goalis to make it very easy to create an AUTOTUTOR

for a new knowledge domain. First, the devel-oper creates an LSA space after identifying acorpus of electronic documents on the domainknowledge. The lesson planner creates a cur-

Articles

44 AI MAGAZINE

Figure 3. The ANDES Tutoring System.

in the upper-right window, and enter equationsin the lower-right window. When studentsenter a vector, variable, or equation, ANDES willcolor the entry green if it is correct and red if itis incorrect. This approach is called immediatefeedback and is known to enhance learningfrom problem solving (Anderson et al. 1995).

To give immediate feedback, ANDES mustunderstand the student’s entries no matterhow the student tries to solve the problem.ANDES uses a rule-based expert system to solvethe problem in all correct ways. It gives nega-tive feedback if the student’s entry does notmatch one of the steps of one of the solutionsfrom the expert model. For this reason, ANDES

and similar tutoring systems are known asmodel-tracing tutors. They follow the student’sreasoning by comparing it to a trace of themodel’s reasoning.

How Does ANDES Hint and Give Help?Students can ask ANDES for help by either click-ing on the menu item What do I do next? or byselecting a red entry and clicking on the menuitem What’s wrong with that? ANDES uses aBayesian network to help it determine whichstep in the expert’s solution to give the studenthelp on (Gertner, Conati, and VanLehn 1998).It prints in the lower-left window a short mes-sage, such as the one shown in figure 3. Themessage is only a hint about what is wrong orwhat to do next. Often a mere hint suffices, andthe students are able to correct their difficultyand move on. However, if the hint fails, thenthe student can ask for help again. ANDES gener-ates a second hint that is more specific than thefirst. If the student continues to ask for help,ANDES’s last hint will essentially tell the studentwhat to do next. This technique of giving helpis based on human-authored hint sequences.Each hint is represented as a template. It is filledin with text that is specific to the situationwhere help was requested. Such hint sequencesare often used in intelligent tutoring systemsand are known to enhance learning from prob-lem solving (McKendree 1990).

During evaluations in the fall of 2000 at theU.S. Naval Academy, students using ANDES

scored about a letter grade (0.92 standard devi-ation units) higher on the midterm exam thanstudents in a control group (Shelby et al. 2002).Log file data indicate that students are usingthe help and hint facilities as expected. Ques-tionaire data indicate that many of them preferdoing their homework on ANDES to doing itwith paper and pencil.

Other intelligent tutoring systems use similarmodel tracing, immediate feedback, and hintsequences techniques, and many have been

shown to be effective (for example, Anderson etal. [1995]; McKendree, Radlinski, and Atwood[1992]; Reiser et al. [2002]). A new company,Carnegie Learning,1 is producing such tutors foruse in high-school mathematics classes. As offall 2000, approximately 10 percent of the alge-bra I classes in the United States will be usingone of the Carnegie Learning tutors. Clearly,this AI technology is rapidly maturing.

Criticisms of ANDES and Other SimilarTutoring SystemsThe pedagogy of immediate feedback and hintsequences has sometimes been criticized forfailing to encourage deep learning. The follow-ing four criticisms are occasionally raised bycolleagues:

First, if students don’t reflect on the tutor’shints but merely keep guessing until they findan action that gets positive feedback, they canlearn to do the right thing for the wrong rea-sons, and the tutor will never detect the shallowlearning (Aleven, Koedinger, and Cross 1999).

Second, the tutor does not ask students toexplain their actions, so students might notlearn the domain’s language. Educators haverecently advocated that students learn to “talkscience.” Talking science is allegedly part of adeep understanding of the science. It also facil-itates writing scientifically, working collabora-tively in groups, and participating in the cul-ture of science.

Third, to understand the students’ thinking,the user interface of such systems requires stu-dents to display many of the details of their rea-soning. This design doesn’t promote steppingback to see the “basic approach” one has usedto solve a problem. Even students who havereceived high grades in a physics course can sel-dom describe their basic approaches to solvinga problem (Chi, Feltovich, and Glaser 1981).

Fourth, when students learn quantitativeskills, such as algebra or physics problem solv-ing, they are usually not encouraged to see theirwork from a qualitative, semantic perspective.As a consequence, they fail to induce versionsof the skills that can be used to solve qualitativeproblems and check quantitative ones for rea-sonableness. Even physics students with highgrades often score poorly on tests of qualitativephysics (Halloun and Hestenes 1985).

Many of these objections can be made to justabout any form of instruction. Even experttutors and teachers have difficulty getting stu-dents to learn deeply. Therefore, these criticismsof intelligent tutoring systems should onlyencourage us to improve them, not reject them.

There are two common themes in this list offour criticisms. First, all four involve integrat-

Talkingscience isallegedly partof a deepunderstandingof the science.

Articles

WINTER 2001 45

ing technology that was originally developedfor CIRCSIM tutor (Freedman and Evens 1996),the BASIC ELECTRICITY AND ELECTRONICS (BEE) tutor(Rosé, Di Eugenio, and Moore 1999), and theCOCONUT model of collaborative dialogue (DiEugenio et al. 2000). A number of natural lan-guage–understanding authoring tools havebeen developed, including the LC-FLEX parser(Rosé and Lavie 2001).

Currently, ATLAS plays only a small role in thestudent’s total problem-solving process. Mostof the time, the students interact with ANDES

just as they ordinarily do. However, if ATLAS

notices an opportunity to promote deep learn-ing, it takes control of the interaction andbegins a natural language dialogue. AlthoughATLAS can ask students to make ANDES actions aspart of the dialogue (for example, it might havethe student draw a single vector), most of thedialogue is conducted in a scrolling text win-dow, which replaces the hint window shown inthe lower left of figure 3. When ATLAS has fin-ished leading the student through a line of rea-soning, it signs off and lets the student returnto solving the problem with ANDES.

The ATLAS dialogues are called knowledge con-struction dialogues (KCDs) because they aredesigned to encourage students to infer or con-struct the target knowledge. For example, ANDES

might simply tell the student that when anobject moving in a straight line is slowingdown, its acceleration is in the opposite direc-tion to it velocity. ATLAS will instead try to drawthe knowledge out of the student with a dia-logue such as the one shown in figure 4, wherethe student derived the target principle from adeeper one. KCDs are intended to provide deep-er knowledge by connecting principles, relatingthem to commonsense knowledge, and givingthe student practice in talking about them.

Knowledge Construction Dialogues to Teach PrinciplesTo date, ATLAS conducts just one kind of KCD,namely, those that teach domain principles.Currently, we are concentrating on only asmall portion of physics, so only 55 principlesare covered. Even so, building so many KCDswas daunting enough that we built tools tohelp us. With these tools, we were able to buildknowledge sources for our KCDs in only threeperson-months.

The primary design concept is to representKCDs as recursive finite-state networks. Statescorrespond to tutor utterances (usually ques-tions), and arcs correspond to student respons-es. A few arcs are special in that they either calla subdialogue or return from one. Such recur-sive finite- state networks are often used in spo-

ing problem-solving knowledge with otherknowledge, namely (1) principles or rationales;(2) domain language; (3) abstract, basicapproaches; and (4) qualitative rules of infer-ence. Second, the kinds of instructional activi-ty that are currently used to tap these otherkinds of knowledge make critical use of naturallanguage. Although one can invent graphic orformal notations to teach these kinds of knowl-edge on a computer, they might be more con-fusing to the students and instructors than theknowledge that they are trying to convey.Moreover, students and instructors are likely toresist learning a new formalism, even a graphicone, if they will only use it temporarily.

ATLAS: A Natural LanguageEnhancement for

Model-Tracing Tutors We believe that tutoring systems must use nat-ural language if they are to become more effec-tive at encouraging deep learning. Therefore,we have begun building ATLAS, a module thatcan be added to ANDES or other similar model-tracing tutoring systems to conduct naturallanguage dialogues and thereby promote deeplearning. ATLAS uses natural language-process-

Articles

46 AI MAGAZINE

1. Tutor: As you know, you can think of an acceleration vector as changing a velocity vector by pulling on its tip. In this case, the elevator’s velocity vector points straight down and you drew its acceleration vector pointing down too. What does your acceleration vector do to the velocity vector?

2. Student: It stretches it downward.

3. Tutor: Does that make it longer?

4. Student: Yes.

5. Tutor: Which means the elevator is speeding up, right?

6. Student: Yeah, but its not really speeding up. It's slowing down.

7. Tutor: Right. So maybe your acceleration vector’s direction is wrong. What direction should it be?

8. Student: Up.

9. Tutor: Right. So what have you learned about the direction of acceleration when an object is moving in a straight line and slowing down?

10. Student: It’s in the opposite direction from the velocity.

11. Tutor: Yes! That’s exactly right. Good job.

Figure 4. A Hypothetical Dialogue between ATLAS and a Student.

ken language dialogue systems (Jurafsky andMartin 2000), so it makes sense to start withthem and see where they break down.

Our primary tool is the KCD editor (figure 5).In the upper-left window, the author selects atopic, which is deceleration in this case. Thisselection causes a shorthand form of the recipes(discourse plans) to appear in the upper-rightwindow. Selecting a tutor-student interactionbrings up windows for seeing the tutor’s contri-bution (as with the lower-right window) andthe student’s expected answers (middle win-dows). The left-middle window is for correctanswers. As in AUTOTUTOR, the student’s expect-ed answer is represented as a set of expectations(left and opposite in this case). The right-mid-dle window is for incorrect answers. When oneof these is selected, a subdialogue for handlingit is displayed in the lower-left window. Noticethat the author enters natural language text forthe tutor contribution, the expectations, andalmost everything else.

In a limited sense, the KCDs are intended tobe better than naturally occurring dialogues.Just as most text expresses its ideas more clearlythan informal oral expositions, the KCD isintended to express its ideas more clearly thanthe oral tutorial dialogues that human tutorsgenerate. Thus, we need a way for expert physi-cists, tutors, and educators to critique the KCDsand suggest improvements. Because the under-lying finite-state network can be complex, it isnot useful to merely print it out and let expertspencil in comments. The second tool facilitatescritiquing KCDs by allowing expert physicistsand psychologists to navigate around the net-work and enter comments on individual statesand arcs (figure 6). It presents a dialogue in theleft column and allows the user to enter com-ments in the right column. Because there aremany expected responses for each tutorial con-tribution, the user can select a response from apull-down menu, causing the whole dialogueto adjust, opening up new boxes for the user’s

Articles

WINTER 2001 47

Figure 5. The Knowledge Construction Dialogue Editor.

should have ATLAS say line 7 instead of line 3.Although the authors only see and edit nat-

ural language text, we cannot expect studentsto type in exactly the responses that theauthors enter. The compiler uses CARMEL (Rosé2000) to translate the expected studentresponses into semantic structures. Thus, itshould recognize the expected responses evenif they are not expressed with the same wordsand syntax as the author-provided versions.

Current Status and Future DirectionsThe initial version of ATLAS was pilot tested inthe fall of 2000. Five students used ATLAS, andfive students used ANDES without its hintsequences instead of the ATLAS KCDs. Despitethe small number of subjects, the ATLAS stu-dents scored significantly higher than theANDES students on a conceptual posttest. Sur-prisingly, the effect was large: The ATLAS stu-dents gained about .9 standard deviation unitsmore than the ANDES students. Moreover, theyscored about the same as the ANDES students ona quantitative posttest, suggesting that theimprovements were limited to the materialtaught by ATLAS, as expected.

Five issues dominate ATLAS’s future develop-ment. First, writing an effective KCD anddebugging it with real students is an inherentlylabor-intensive process. We will continuebuilding tools to expedite the process. Second,the conventional wisdom is that a recursivefinite-state network does not provide sufficientflexibility for managing complex dialogues.Although a reactive planner interprets ATLAS’snetworks, we do not currently make use of allits power. Thus, a second important direction isto determine how much of this additionalpower is necessary for conducting more effec-tive tutorial dialogues with students. Third, thecurrent version of ATLAS does not make use ofthe full power offered by the CARMEL core-understanding component. Thus, anotherrelated direction is determining how sophisti-cated an analysis of student input is necessaryfor the system to determine how best to pro-ceed with its dialogue with the student. Fourth,we have deliberately left the ANDES system’s twomajor knowledge sources alone, so ANDES is stillresponsible for solving physics problems anddeciding which hint sequence is appropriate.Thus, KCDs are used mostly to replace hintsequences. We are not sure if this simple designwill allow pedagogically useful dialogues orwhether we will need to port some of ANDES’sknowledge to ATLAS. Fifth, we plan to extendATLAS’s capabilities to additional types of knowl-edge-construction dialogues, such as goal-scaf-folding dialogues.

comments. This tool runs in a web browser, soexperts can use it remotely.

At any point in its development, a KCD canbe compiled into executable code. The code isinterpreted by a reactive planning engine(Freedman 1999). The engine does not simplyfollow the finite-state network. Instead, it hasrudimentary (but growing) capabilities fortreating the network as a plan for the conver-sation that it will adapt as necessary. Forexample, in the conversation in figure 4, sup-pose the student said at line 2, “The accelera-tion makes the velocity vector longer, so theelevator should be going faster.” The reactiveplanner should recognize that the student hasskipped ahead in the conversation plan, so it

Articles

48 AI MAGAZINE

Figure 6. The Knowledge Construction Dialogue Commenter.

WHY2: Tutoring Qualitative Explanations

All tutoring systems have students perform atask, and they help the students do it. Sometutoring systems, such as ANDES and ATLAS, havethe student solve problems. Other tutoring sys-tems, such as AUTOTUTOR, ask the student deepquestions and help them formulate a correct,complete answer. Recent work with humantutors (for example, Chi et al. [2001]) suggeststhat a good activity for teaching is to have stu-dents explain physical systems qualitatively.Although it is possible to have them expresstheir explanations in formal or graphic lan-guages (for example, CYCLEPAD [Forbus et al.1998]), we believe that they will learn more ifthey can express their explanations in naturallanguage. Thus, the goal of the WHY2 project isto coach students as they explain physical sys-tems in natural language.

WHY2 is intended to be a successor of one ofthe first intelligent tutoring systems in the lit-erature, the WHY system. WHY was envisionedand partially implemented by Albert Stevensand Alan Collins (Stevens and Collins 1977).They studied experts helping students articu-late such explanations and tried to embed theirtutorial strategies in the WHY system. Stevensand Collins discovered that students had agreat many misconceptions about nature.These misconceptions would only surfacewhen students expressed their ideas qualita-tively because they could solve textbook quan-titative problems correctly (Halloun andHestenes 1985). Since this time, considerableeffort has been expended by physics educatorsto discover, catalog, and invent remedies forstudent misconceptions. The remedies are usu-ally intended for classroom or laboratories andhave had only moderate success (Hake 1998).By adapting them to the tutorial setting andembedding the tutorial strategies uncovered byCollins, Stevens, and others, WHY2 might bemuch more successful.

The basic idea of WHY2 is to ask the studentto type in an explanation for a simple physicalsituation, such as the battery-bulb circuitshown in figure 7. WHY2 analyzes the student’sexplanation (line 1 in Figure 7) to see if the stu-dent has any misconceptions. If it detects amisconception, it invokes a knowledge con-struction dialogue (KCD), such as the oneshown in lines 2 through 9. During this dia-logue, further misunderstandings may arise,which can cause another KCD to be selectedand applied (see lines 10 onward).

WHY2 is a joint project involving both theAUTOTUTOR and ATLAS groups. It began recently

and is still in the design stages. A corpus ofexplanations from students has been collectedand is being analyzed to see what kinds of mis-conceptions and language the students areusing. Our plan is to use a combination of theLSA technology from AUTOTUTOR and thesemantic composition technology from ATLAS.

Articles

WINTER 2001 49

Question: A battery is connected to a bulb by two wires. The bulb lights. Why?

1. Student: There are two currents which come out, one current positive and one negative. They combine in the bulb to produce light and heat.

2. Tutor: I see. Which current is composed of moving electrons?

3. Student: I’m not sure. Both?

4. Tutor: Right, both are. An electrical current is just moving electrons. Which direction do the electrons move in your two currents?

5. Student: I think the electrons move out from the battery to the bulb. The two currents meet there and produce energy.

6. Tutor: Are the electrons actually destroyed in the bulb?

7. Student: I’m not sure.

8. Tutor: Actually, electrons are not destroyed or lost in an electrical current. Instead, they go out one side of the battery, go through the bulb and go back in the other side of the battery. Thus, your two currents are actually the same flow, only the electrons are leaving the battery for one and returning to the battery for the other. Given everything you know about batteries and bulbs, does this make sense?

9. Student: But I thought the battery dies when it runs out of electrons.

10. Tutor: Batteries die when they run out of the chemical energy that they use to pump electrons around the circuit. The battery is like a water pump that produces a powerful current, and the bulb is like a paddle wheel that is driven by the water current. When the water pump runs out of energy, the current slows to a stop. As this happens, does the number of electrons in the battery change?

11. Student: Maybe a little. But it stays about the same.

12. Tutor: Exactly right! I think you understand electrical current better now. See if you can write a better explanation.

Figure 7. A Hypothetical Dialogue between a Student and WHY2.

Generation Computer Tutors: Learn fromor Ignore Human Tutors? In Proceedings ofthe 1999 Conference of Computer-HumanInteraction, 85–86. New York: Associationof Computing Machinery.

Di Eugenio, B.; Jordan, P. W.; Thomason, R.H.; and Moore, J. D. 2000. The AgreementProcess: An Empirical Investigation ofHuman-Human Computer-Mediated Dia-logues. International Journal of Human-Com-puter Studies 53(6): 1017–1076.

Forbus, K. D.; Everett, J. O.; Ureel, L.;Brokowski, M.; Baher, J.; and Kuehne, S. E.1998. Distributed Coaching for an Intelli-gent Learning Environment. Paper present-ed at the AAAWorkshop on QuantitativeReasoning, 26–29 May, Cape Cod, Massa-chusetts.

Freedman, R. 1999. ATLAS: A Plan Managerfor Mixed-Initiative, Multimodal Dialogue.Paper presented at the 1999 AAAI Work-shop on Mixed-Initiative Intelligence, 19July, Orlando, Florida.

Freedman, R., and Evens, M. W. 1996. Gen-erating and Revising Hierarchical Multi-Turn Text Plans in an ITS. In IntelligentTutoring Systems: Proceedings of the 1996Conference, eds. C. Frasson, G. Gauthier,and A. Lesgold, 632–640. Berlin: Springer.

Gertner, A.; Conati, C.; and VanLehn, K.1998. Procedural Help in ANDES: GeneratingHints Using a Bayesian Network StudentModel. In Proceedings of the FifteenthNational Conference on Artificial Intelli-gence, 106–111. Menlo Park, Calif.: Ameri-can Association for Artificial Intelligence.

Gertner, A. S., and VanLehn, K. 2000. ANDES:A Coached Problem-Solving Environmentfor Physics. In Intelligent Tutoring Systems:Fifth International Conference, ITS 2000, eds.G. Gautheier, C. Frasson, and K. VanLehn,133–142. New York: Springer.

Graesser, A. C.; Person, N. K.; andMagliano, J. P. 1995. Collaborative Dia-logue Patterns in Naturalistic One-on-OneTutoring. Applied Cognitive Psychology 9(4):495–522.

Graesser, A. C.; Wiemer-Hastings, K.;Wiemer-Hastings, P.; Kreuz, R.; and theTutoring Research Group 1999. AUTOTUTOR:A Simulation of a Human Tutor. Journal ofCognitive Systems Research 1(1): 35–51.

Graesser, A. C.; Wiemer-Hastings, P.;Wiemer-Hastings, K.; Harter, D.; Person, N.;and the Tutoring Research Group. 2000.Using Latent Semantic Analysis to Evaluatethe Contributions of Students in AUTOTU-TOR. Interactive Learning Environments 8(2):129–148.

Hake, R. R. 1998. Interactive-Engagementversus Traditional Methods: A Six-Thou-sand Student Survey of Mechanics TestData for Introductory Physics Students.

AcknowledgmentsThe AUTOTUTOR research was support-ed by grants from the National ScienceFoundation (SBR 9720314) and theOffice of Naval Research (N00014-00-1-0600). The ANDES research was sup-ported by grant N00014-96-1-0260from the Cognitive Sciences Divisionof the Office of Naval Research. TheATLAS research is supported by grant9720359 from the LIS program of theNational Science Foundation. TheWHY2 research is supported by grantN00014-00-1-0600 from the CognitiveSciences Division of the Office ofNaval Research.

Note1. www.carnegielearning.com.

ReferencesAleven, V.; Koedinger, K. R.; and Cross, K.1999. Tutoring Answer Explanation FostersLearning with Understanding. In ArtificialIntelligence in Education, eds. S. P. Lajoie andM. Vivet, 199–206. Amsterdam: IOS.

Anderson, J. R.; Corbett, A. T.; Koedinger, K.R.; and Pelletier, R. 1995. Cognitive Tutors:Lessons Learned. The Journal of the LearningSciences 4(2): 167–207.

Cassell, J., and Thorisson, K. R. 1999. ThePower of a Nod and a Glance: Envelope ver-sus Emotional Feedback in Animated Con-versational Agents. Applied Artificial Intelli-gence 13(3): 519–538.

Chi, M. T. H.; Feltovich, P.; and Glaser, R.1981. Categorization and Representation ofPhysics Problems by Experts and Novices.Cognitive Science 5(2): 121–152.

Chi, M. T. H.; de Leeuw, N.; Chiu, M.; andLaVancher, C. 1994. Eliciting Self-Explana-tions Improves Understanding. CognitiveScience 18(3): 439–477.

Chi, M. T. H.; Siler, S.; Jeong, H.; Yamauchi,T.; and Hausmann, R. G. 2001. Learningfrom Tutoring: A Student-Centered versus aTutor-Centered Approach. Cognitive Science.Forthcoming.

Cohen, P. A.; Kulik, J. A.; and Kulik, C. C.1982. Educational Outcomes of Tutoring: AMeta-Analysis of Findings. American Educa-tional Research Journal 19(2): 237–248.

Conati, C., and VanLehn, K. 1999. Teach-ing Metacognitive Skills: Implementationand Evaluation of a Tutoring System toGuide Self-Explanation While Learningfrom Examples. In Artificial Intelligence inEducation, eds. S. P. Lajoie and M. Vivet,297–304. Amsterdam: IOS.

Corbett, A.; Anderson, J.; Graesser, A.;Koedinger, K.; and VanLehn, K. 1999. Third

The KCDs of ATLAS will be generalizedto incorporate elements of the DANsof AUTOTUTOR.

Our dialogue technology can bestressed by the complexity of the lan-guage and discourse we anticipatefrom the students. However, if we canmake it work, the pedagogical payoffswill be enormous. Repairing the quali-tative misconceptions of physics is adifficult and fundamentally importantproblem.

ConclusionsWe discussed three projects that haveseveral similarities. AUTOTUTOR, ATLAS,and WHY2 all endorse the idea that stu-dents learn best if they constructknowledge themselves. Thus, their dia-logues try to elicit knowledge from thestudent by asking leading questions.They only tell the student the knowl-edge as a last resort. All three projectsmanage dialogues by using finite-statenetworks. Because we anticipate build-ing hundreds of such networks, theprojects are building tools to letdomain authors enter these dialoguesin natural language. All three projectsuse robust natural language–under-standing techniques—LSA for AUTOTU-TOR, CARMEL for ATLAS, and a combina-tion of the two for WHY2. All threeprojects began by analyzing data fromhuman tutors and are using evalua-tions with human students through-out their design cycle.

Although the three tutoring systemshave the common objective of helpingstudents perform activities, the specifictasks and knowledge domains arerather different. AUTOTUTOR’S studentsare answering deep questions aboutcomputer technology, ATLAS’s studentsare solving quantitative problems, andWHY2’s students are explaining physi-cal systems qualitatively. We mightultimately discover that the conversa-tional patterns need to be different forthese different domains and tasks. Thatis, dialogue styles might need to be dis-tinctively tailored to particular classesof knowledge domains. A generic dia-logue style might prove to be unsatis-factory. Whatever discoveries emerge,we suspect they will support one basicclaim: Conversational dialogue sub-stantially improves learning.

Articles

50 AI MAGAZINE

American Journal of Physics 66(4): 64–74.

Halloun, I. A., and Hestenes, D. 1985. Com-mon Sense Concepts about Motion. Ameri-can Journal of Physics 53(11): 1056–1065.

Hestenes, D.; Wells, M.; and Swackhamer,G. 1992. Force Concept Inventory. ThePhysics Teacher 30(3): 141–158.

Johnson, W. L.; Rickel, J. W.; and Lester, J.C. 2000. Animated Pedagogical Agents:Face-to-Face Interaction in InteractiveLearning Environments. International Jour-nal of Artificial Intelligence in Education11(1): 47–78.

Jurafsky, D., and Martin, J. E. 2000. Speechand Language Processing: An Introduction toNatural Language Processing, ComputationalLinguistics, and Speech Recognition. UpperSaddle River, N.J.: Prentice Hall.

Koedinger, K. R.; Anderson, J. R.; Hadley, W.H.; and Mark, M. A. 1997. Intelligent Tutor-ing Goes to School in the Big City. Journalof Artificial Intelligence in Education 8(1):30–43.

Kulik, J. A., and Kulik, C. L. C. 1988. Timingof Feedback and Verbal Learning. Review ofEducational Research 58(1): 79–97.

Landauer, T. K.; Foltz, P. W.; and Laham, D.1998. An Introduction to Latent SemanticAnalysis. Discourse Processes 25(2–3):259–284.

Lesgold, A.; Lajoie, S.; Bunzo, M.; andEggan, G. 1992. SHERLOCK: A Coached Prac-tice Environment for an Electronics Trou-bleshooting Job. In Computer-AssistedInstruction and Intelligent Tutoring Systems,eds. J. H. Larkin and R. W. Chabay,201–238. Hillsdale, N.J.: Lawrence Erlbaum.

Lester, J. C.; Voerman, J. L.; Townes, S. G.;and Callaway, C. B. 1999. Deictic Believ-ability: Coordinating Gesture, Locomotion,and Speech in Life-Like Pedagogical Agents.Applied Artificial Intelligence 13(4–5):383–414.

McKendree, J. 1990. Effective FeedbackContent for Tutoring Complex Skills.Human-Computer Interaction 5:381–413.

McKendree, J.; Radlinski, B.; and Atwood,M. E. 1992. The GRACE Tutor: A QualifiedSuccess. In Intelligent Tutoring Systems: Sec-ond International Conference, eds. C. Frasson,G. Gautheir, and G. I. McCalla, 677–684.Berlin: Springer-Verlag.

Person, N. K.; Graesser, A. C.; Kreuz, R. J.;Pomeroy, V.; and the Tutoring ResearchGroup. 2001. Simulating Human Tutor Dia-logue Moves in AUTOTUTOR. InternationalJournal of Artificial Intelligence in Education.Forthcoming.

Reiser, B. J.; Copen, W. A.; Ranney, M.;Hamid, A.; and Kimberg, D. Y. 2002. Cogni-tive and Motivational Consequences ofTutoring and Discovery Learning. Cognitionand Instruction. Forthcoming.

Rosé, C. P. 2000. A Framework for RobustSemantic Interpretation. In Proceedings ofthe First Meeting of the North American Chap-ter of the Association for ComputationalLinguistics, 311–318. San Francisco, Calif.:Morgan Kaufmann.

Rosé, C. P., and Lavie, A. 2001. BalancingRobustness and Efficiency in Unification-Augmented Context-Free Parsers for LargePractical Applications. In Robustness in Lan-guage and Speech Technology, eds. J. C. Jun-qua and G. V. Noord, 239–269. Amsterdam:Kluwer Academic.

Rosé, C. P.; DiEugenio, B.; and Moore, J.1999. A Dialogue-Based Tutoring Systemfor Basic Electricity and Electronics. In Arti-ficial Intelligence in Education, eds. S. P.Lajoie and M. Vivet, 759–761. Amsterdam:IOS.

Shelby, R. N.; Schulze, K. G.; Treacy, D. J.;Wintersgill, M. C.; VanLehn, K.; and Wein-stein, A. 2001. The Assessment of ANDES

Tutor. Forthcoming.

Stevens, A., and Collins, A. 1977. The GoalStructure of a Socratic Tutor. In Proceedingsof the National ACM Conference, 256–263.New York: Association of ComputingMachinery.

VanLehn, K. 1996. Conceptual and Met-alearning during Coached Problem Solv-ing. In Proceedings of the Third IntelligentTutoring Systems Conference, eds. C. Frasson,G. Gauthier, and A. Lesgold, 29–47. Berlin:Springer-Verlag.

VanLehn, K.; Freedman, R.; Jordan, P.; Mur-ray, C.; Osan, R.; Ringenberg, M.; Rosé, C.P.; Schulze, K.; Shelby, R.; Treacy, D.; Wein-stein, A.; and Wintersgill, M. 2000. Fadingand Deepening: The Next Steps for ANDES

and Other Model-Tracing Tutors. In Intelli-gent Tutoring Systems: Fifth InternationalConference, ITS 2000, eds. G. Gauthier, C.Frasson, and K. VanLehn, 474–483. Berlin:Springer-Verlag.

Arthur Graesser is a pro-fessor of psychology andcomputer science at theUniversity of Memphis,codirector of the Insti-tute for Intelligent Sys-tems, and director of theCenter for Applied Psy-chological Research. He

has conducted research on tutorial dia-logue in intelligent tutoring systems and iscurrent editor of the journal DiscourseProcesses. His e-mail address is a-graesser@memphis.edu.

Kurt VanLehn is a pro-fessor of computer sci-ence and intelligent sys-tems at the University ofPittsburgh, director ofthe Center for Interdisci-plinary Research onConstructive LearningEnvironments, and a

senior scientist at the Learning Researchand Development Center. His main inter-ests are applications of AI to tutoring andassessment. He is a senior editor for thejournal Cognitive Science. His e-mail addressis vanlehn@cs.pitt.edu.

Carolyn Rosé is aresearch associate at theLearning Research andDevelopment Center atthe University of Pitts-burgh. Her main re-search focus is on devel-oping robust languageunderstanding technolo-

gy and authoring tools to facilitate therapid development of dialogue interfacesfor tutoring systems. Her e-mail address isrosecp@pitt.edu.

Pamela Jordan is aresearch associate at theLearning Research andDevelopment Center atthe University of Pitts-burgh. Her main interestsare in computer-mediat-ed natural language dia-logue, the analysis of

these dialogues to identify effective commu-nication strategies, and the creation of dia-logue agents to test strategies. Her e-mailaddress is pjordan@pitt.edu.

Derek Harter is a Ph.D.candidate in computerscience at the Universityof Memphis. He holds aB.S. in computer sciencefrom Purdue Universityand an M.S. in computerscience and AI fromJohns Hopkins Universi-

ty. His main interests are in dynamic andembodied models of cognition andneurologically inspired models of actionselection for autonomous agents. His e-mail address is dharter@memphism.edu.

Articles

WINTER 2001 51

Articles

52 AI MAGAZINE

please join us!

� member s electronic library � national conference on artificial intel-ligence � discounts on ai books � innovative applications of artificialintelligence � discounts on journals � robot exhibition � spring sym-posium series � aaai press � fall symposium series � botball tourna-

ment � ai magazine � classic paper award � aaai fellows � allennewell award � distinguished service award � mobile robot competi-tion � ai topics website � technical reports � grants for workshops,

conferences, and symposia � electronic directory

american association for artificial intelligence445 burgess drive � menlo park, california 94025 usa

www.aaai.org � membership@aaai.org650-321-4457 � 650-321-4457 f

ax