Evaluating Student Learning Gains in Two Versions of AutoTutor

Evaluating Student Learning Gains in Two
Versions of AutoTutor
Natalie K. Person 1 , Laura Bautista 2 , Arthur C. Graesser 2 , Eric Mathews 1 ,
and The Tutoring Research Group 2
1 Department of Psychology, Rhodes College, 2000 North Parkway, Memphis, TN 38112,
2 Department of Psychology, The University of Memphis, Memphis, TN 38152-6400
Abstract: The pedagogical effectiveness of two versions of AutoTutor was assessed in a student learning outcome
study. Sixty students enrolled in a computer literacy course received tutoring from one of the versions of AutoTutor
on one of the following topics: Hardware, Operating Systems, and the Internet. Students were also required to reread
material on two of the previously mentioned topics. All participants then received a comprehension test on all three
topics. A within-subjects design enabled the following conditions to be compared: AutoTutor versus a reread
condition versus a control condition. Results indicated that AutoTutor was an effective pedagogical tool compared to
the other learning controls. Both versions of AutoTutor provided an effect size increment of approximately .5
standard deviation units when compared to the reread and control condition.
1 Background
AutoTutor is an animated pedagogical agent that participates in a conversation with the learner
while simulating the dialog moves that are frequently used by typical human tutors. AutoTutor is
currently designed to help college students learn about topics that are typically covered in an
introductory computer literacy course (e.g., hardware, operating systems, and the Internet).
Elaborated descriptions of AutoTutor’s architecture have been discussed in previous
publications, and therefore, will only receive brief mention in this paper [8, 11, 15, 16, 20, 25,
28, 34, 35].
AutoTutor’s discourse patterns and pedagogical strategies are based on a previous project that
dissected 100 hours of naturalistic tutoring sessions [12, 13, 31]. Instead of merely being an
information delivery system that bombards the student with a large volume of information,
AutoTutor serves as a discourse facilitator or collaborative scaffold that assists the student in
actively constructing knowledge. Hence, a central educational philosophy behind AutoTutor is
that effective learning occurs when students actively do the following: (1) construct subjective
explanations and elaborations of the material, (2) ask and answer questions, and (3) solve
problems that require deep reasoning [2, 4, 7, 23].
We currently have two versions of AutoTutor. AutoTutor 1.1 simulates the dialog moves of
normal, untrained human tutors, whereas AutoTutor 2.0 simulates dialog moves that are
motivated by more sophisticated, ideal tutoring strategies. Our analyses of human tutoring
sessions revealed that typical human tutors do not use most of the ideal tutoring strategies that
have been identified in education and the intelligent tutoring system enterprise. These strategies
include the Socratic method [5], modeling-scaffolding-fading [6], reciprocal training [29],
anchored situated learning [1], error identification and correction [1, 23, 38], building on

prerequisites [10], and sophisticated motivational techniques [22]. Detailed discourse analyses
have been performed on small samples of accomplished tutors in an attempt to identify
sophisticated tutoring strategies [9, 17, 26, 27, 36]. However, we discovered that the vast
majority of these sophisticated tutoring strategies were virtually nonexistent in the untrained
tutoring sessions that we videotaped and analyzed. Tutors clearly need to be trained how to use
the sophisticated tutoring skills because they do not routinely emerge in naturalistic tutoring with
untrained tutors. In this paper we report a study that assessed the impact of two versions of
AutoTutor on student learning gains. We begin with a brief overview of AutoTutor and then
report the results of the empirical study.
2 Brief Overview of AutoTutor
AutoTutor works by having a conversation with the learner. AutoTutor appears as an animated
agent that acts as a dialog partner with the learner. The animated agent delivers AutoTutor’s
dialog moves with synthesized speech, intonation, facial expressions, and gestures. The major
question or problem that is being worked on is both spoken by AutoTutor and is printed at the
top of the screen. The major questions/problems are generated systematically from a curriculum
script, a module discussed below. AutoTutor’s major questions and problems are not the fill-inthe
blank, true/false, or multiple-choice questions that are so popular in the US educational
system. Instead, the questions and problems invite lengthy explanations and deep reasoning
(e.g., answers to why, how, what-if questions). The goal is to encourage students to articulate
lengthier answers that exhibit deep reasoning, rather than to recite small bits of shallow
knowledge. There is a continuous multi-turn tutorial dialog between AutoTutor and the learner
during the course of answering a question (or solving a problem). When considering the turns of
both the learner and AutoTutor, it typically takes 10 to 20 conversational turns to answer a single
question or solve a problem from the curriculum script. The learner types in her/his
contributions during the exchange on the keyboard. For some topics, there are graphical displays
and animation, with components that AutoTutor refers to. The ultimate goal is to have
AutoTutor be a good conversational partner that comprehends, speaks, points, and displays
emotions, all in a coordinated fashion.
3 AutoTutor’s Architechure
3.1 Curriculum Script
A curriculum script is a loosely ordered set of skills, concepts, example problems, and questionanswer
units. Each topic in the curriculum script is represented as a structured set of words,
sentences, or paragraphs in a free text format. Associated with each topic (problem or question)
is a focal question, a set of basic noun-like concepts, a set of ideal good answer aspects (each
being roughly a sentence of 10 - 20 words), different forms of expressing or eliciting each ideal
answer aspect (i.e., a hint, prompt, versus assertion), a set of anticipated bad answers (i.e., bugs,
misconceptions), a correction for each bad answer, a summary of the answer or solution, and a
set of anticipated topic-related questions and answers.

3.2 Natural language extraction and speech act classification
AutoTutor must be able to classify the speech acts of student contributions in order to flexibly
respond to what the student types in. First, AutoTutor segments the string of words and
punctuation marks within a learner’s turn into speech act units, relying on punctuation to perform
this segmentation. Then each speech act is assigned to one of the following speech act categories:
Assertion, WH-question, YES/NO question, Metacognitive comment (e.g., I don’t understand),
Metacommunicative act (e.g., Could you repeat that?), and Short Response.
3.3 Latent Semantic Analysis
The fact that world knowledge is inextricably bound to the process of comprehending language
and discourse is widely acknowledged, but researchers in computational linguistics and artificial
intelligence have not had a satisfactory approach to handling the deep abyss of world knowledge.
Recently, latent semantic analysis (LSA) has been proposed as a statistical representation of a
large body of world knowledge [20, 21]. LSA capitalizes on the fact that particular words appear
in particular texts (called “documents”). Each word, sentence, or text ends up being a weighted
vector on the K dimensions. The “match” (i.e., similarity in meaning, conceptual relatedness)
between two words, sentences, or texts is computed as a geometric cosine (or dot product)
between the two vectors, with values ranging from 0 to 1. AutoTutor has successfully used LSA
as the backbone for assessing the quality of student assertions, based on matches to good answers
and anticipated bad answers in the curriculum script [15].
3.4 Dialog Move Generator
AutoTutor currently generates the following dialog moves: main questions, short feedback (i.e.,
positive, neutral, negative), pumps (“uh huh”, “tell me more”), prompts ("The primary memories
of the CPU are ROM and _____"), hints, assertions, corrections, and summaries. As mentioned
earlier, we currently have two versions of AutoTutor, AutoTutor 1.1 and AutoTutor 2.0.
AutoTutor 1.1 simulates the dialog moves of untrained (yet effective) human tutors, whereas
AutoTutor 2.0 is a hybrid between naturalistic tutorial dialog and ideal pedagogical strategies.
The two versions primarily differ in terms of the mechanisms that control the particular dialog
moves that are generated after a student contribution. The dialog move mechanisms for both
AutoTutor versions are discussed later.
3.5 Dialog Advancer Network
The Dialog Advancer Network (DAN) is a mechanism that manages the conversation that occurs
between a student and AutoTutor [30, 32, 33, 34]. The DAN is comprised of a set of customized
pathways that are tailored to particular student speech act categories (e.g., Assertion,
Metacognitive comments). The DAN enables AutoTutor to micro-adapt each tutor-generated
dialog move to the preceding student turn. For example, if a student wants AutoTutor to repeat
the last dialog move, the DAN contains a Metacommunicative pathway that allows AutoTutor to
adapt to the student’s request and respond appropriately. A DAN pathway may include one or a

combination of the following components: (1) discourse markers (e.g., “Okay” or “Moving on”),
(2) AutoTutor dialog moves (e.g., Positive Feedback, Pump, or Assertion), (3) answers to WH-
or Yes/No questions, or (4) canned expressions (e.g., “That’s a good question, but I can’t answer
that right now”).
3.6 Animated Agent
The persona for AutoTutor was created in MetaCreations Poser 3 and is controlled by Microsoft
Agent. AutoTutor is a three-dimensional embodied agent who remains on the screen throughout
the entire tutoring session. AutoTutor communicates with the learner via synthesized speech,
facial expressions, and simple hand gestures. Each of these communication parameters can be
adjusted to maximize AutoTutor’s overall effectiveness as a tutor and conversational partner.
Although a great deal more could be said about the workings of the animated agent, these
mechanisms have been described elsewhere [25, 35] and are simply beyond the scope of this
paper.
4 Two Versions of AutoTutor
4.1 AutoTutor 1.1
The dialog moves in AutoTutor 1.1 are generated by 15 fuzzy production rules [19] that
primarily exploit data provided by the LSA module [15, 34]. AutoTutor1.1’s production rules are
tuned to the following LSA parameters: (a) Student Assertion Quality, (b) Student Ability Level,
and (c) Topic Coverage. Each production rule specifies the LSA parameter values for which a
particular dialog move should be generated. For example, consider the following dialog move
rules:
(1) IF [Student Assertion match with good answer text = HIGH or VERY HIGH]
THEN [select POSITIVE FEEDBACK dialog move]
(2) IF [Student Ability = MEDIUM or HIGH & Student Assertion match with
good answer text = LOW] THEN [select HINT dialog move]
In Rule (1) AutoTutor will provide Positive Feedback (e.g., “Right”) in response to a high quality
student Assertion, whereas in Rule (2) AutoTutor will generate a Hint to bring the relatively high
ability student back on track (e.g., “What about the size of the programs you need to run?”). The
dialog move generator currently controls 12 dialog moves: Pump, Hint, Splice, Prompt, Prompt
Response, Elaboration, Summary, and five forms of immediate short-feedback (positive,
positive-neutral, neutral, negative-neutral, and negative).
During the tutorial conversation for each tutoring topic, AutoTutor must keep track of which
good answer aspects have been covered along with which dialog moves have been previously
generated. AutoTutor 1.1 uses the LSA Topic Coverage metric to track the extent to which each
good answer aspect (Ai) for a topic has been covered in the tutorial conversation. That is, LSA
computes the extent to which the various tutor and student turns cover the good answer aspects
associated with a particular topic. The Topic Coverage metric varies from 0 to 1 and gets updated
for each good answer aspect with each tutor and student turn. If some threshold (t) is met or

exceeded, then the Ai is considered covered. AutoTutor also must decide which good answer
aspect to cover next. In AutoTutor 1.1, the selection of the next good answer aspect to cover is
determined by the zone of proximal development. AutoTutor 1.1 decides on the next aspect to
cover by selecting the aspect that has the highest subthreshold coverage score. Therefore,
AutoTutor 1.1 builds on the fringes of what the student knows or what has occurred in the
discourse history of the tutorial conversation. A topic is finished when all of the aspects have
coverage values that meet or exceed the threshold t.
4.2 AutoTutor 2.0
We believe that the most effective computer tutor will be a hybrid between naturalistic tutorial
dialog and ideal pedagogical strategies. AutoTutor 2.0 incorporates tutoring tactics that attempt
to get the student to articulate the good answer aspect that is selected. AutoTutor 1.1 considers Ai
as covered if it is articulated by either the student or the tutor, whereas AutoTutor 2.0 counts only
what the student says when evaluating coverage. Therefore, if Ai is not articulated by the student,
it is not considered as covered. This forces the student to articulate the explanations in their
entirety, an extreme form of constructivism. In order to flesh out a particular Ai; AutoTutor 2.0
uses discourse patterns that organize dialog moves in terms of their progressive specificity. Hints
are less specific than Prompts, and Prompts are less specific than Elaborations. Thus, AutoTutor
2.0 cycles through a Hint-Prompt-Elaboration pattern until the student articulates the Ai. The
other dialog moves (e.g., short feedbacks and summaries) are controlled by the fuzzy production
rules that were described for AutoTutor 1.1.
AutoTutor 2.0 has two additional features for selecting the next Ai to be covered. First,
AutoTutor 2.0 enhances discourse coherence by selecting the next Ai that is most similar to the
previous aspect that was covered. Second, AutoTutor 2.0 selects pivotal aspects that have a high
family resemblance to the remaining uncovered aspects; that is, AutoTutor 2.0 attempts to select
an aspect that has the greatest content overlap with the remaining aspects to be covered.
Whereas AutoTutor 1.1 capitalizes on the zone of proximal development exclusively, AutoTutor
2.0 also considers conversational coherence and pivotal aspects when selecting the next good
answer aspect to cover.
5 Evaluation of Student Learning Outcomes
5.1 Methods
The methodologies for testing the two versions of AutoTutor (i.e., versions 1.1 and 2.0) were
identical. The participants were 60 students in a computer literacy course at the University of
Memphis. Thirty-six students participated in the AutoTutor 1.1 testing, 24 in the AutoTutor 2.0
testing. The students received extra credit in the computer literacy course for participating in the
experiment. There were three experimental conditions: AutoTutor (student interacted with
AutoTutor to learn about one of the three computer literacy topics, Hardware, Operating systems,
or Internet), Reread (student reread material in the course textbook about one of the three topics),
and no-read Control (student does not re-read or interact with AutoTutor for one of the three
topics). It should be noted that the students were rereading the material that they had previously
covered in the computer literacy course, not learning it for the first time. That is, students had

eceived lectures on the material, had been assigned relevant chapters to read, and had been
tested on the topics by the course instructor. A repeated-measures design ensured that all students
participated in each of the three conditions. The assignment of the three conditions to the three
computer literacy topics was counterbalanced across subjects to control for possible order effects.
All conditions occurred sequentially with minimal time elapsing between conditions. The time
spent rereading the material and interacting with AutoTutor was restricted. For AutoTutor 1.1,
students were given 45 minutes to reread the material and 45 minutes to interact with AutoTutor.
These times were extended to 55 minutes for the AutoTutor 2.0 sessions because AutoTutor 2.0
interactions are (by design) longer.
5.2 Outcome measures
There were 3 outcome measures. We selected a sample of 18 multiple-choice questions from the
test-bank that accompanies the textbook used in the computer literacy course. An equal number
of questions was selected for each of the three topics. We discovered that all of the test-bank
questions were shallow according to Bloom’s taxonomy. A computer literacy expert constructed
a sample of 12 deep multiple-choice questions, four questions for each of the three topics that
tapped causal inferences and reasoning. And finally, there was a cloze test that had 4 critical
words deleted from the ideal answers of each topic; the students filled in a total of 72 blanks with
answers. The three measures were combined into a composite score for each student. The
proportion of correct responses of the composite score served as the metric of student learning
gains. Students were given unlimited time to complete the tests.
5.2 Results
A 2 (AutoTutor Version) x 3 (Experimental Condition) repeated-measures ANOVA was
performed to determine whether the composite score means differed in the various conditions.
The results of this analysis indicated that there were significant differences among the three
experimental conditions, with means of .43, .37, and .35 in the AutoTutor, Reread, and Control
conditions, respectively, F(2, 70) = 6.10, p< .05. Planned comparisons showed the following
pattern: AutoTutor > Reread = Control. The effect size of AutoTutor over Control was .50
standard deviations. This is encouraging given that students spent the same amount of time in the
AutoTutor (50.6 minutes) and Reread (49 minutes) conditions. Surprisingly, there was no main
effect for AutoTutor Version or any significant interactions.
A repeated-measures ANOVA was performed that crossed the three conditions with the three
types of tests (Shallow, Deep, and Cloze). There was a significant main effect of condition, F(2,
70) = 48.03, p< .05, MSe = .038, a significant main effect of test, F(2, 70) = 3.06, p< .05, MSe =
.037, and no significant interaction. The effect size advantages of AutoTutor over Control were
.15 for the shallow test questions, .28 for the deep questions, and .64 for the cloze test.
6 Conclusions
The results support the conclusion that AutoTutor has a significant impact on student learning
gains compared to the other learning and control conditions. We are encouraged by these findings

for two reasons. First, AutoTutor is (to our knowledge) the first animated conversational
computer tutor to produce such learning outcomes in students. Second, students and educators
alike should be pleased that sessions with AutoTutor do not require time commitments beyond
those that students would normally make studying the material.
We anticipated that the more sophisticated strategies of AutoTutor 2.0 would lead to more
positive learning outcomes than the rule-based generation in AutoTutor 1.1. One possible reason
for this non-difference between the AutoTutor versions is that AutoTutor 2.0 sessions were
approximately twice as long as the version 1.1 sessions, 160.58 turns versus 88.35 turns,
respectively. We reported above that there was no difference in the amount of time students spent
in the AutoTutor versus the Reread condition; however, there were significant differences in the
average amounts of time students spent interacting with AutoTutor 1.1 versus 2.0. On average,
students spent 38.4 minutes interacting with AutoTutor 1.1 and 69.0 minutes with the 2.0
version. Hence, it may be the case that AutoTutor 2.0 is a better overall tutor; however, students
experienced fatigue in the considerably lengthier sessions possibly masking the effects of the 2.0
version.
References
[1] Anderson, J. R., Corbett, A. T., Koedinger, K. R., & Pelletier, R. (1995). Cognitive tutors: Lessons learned.
The Journal of the Learning Sciences, 4, 167-207.
[2] Bransford, J. D., Goldman, S. R., & Vye, N. J. (1991). Making a difference in people’s ability to think:
Reflections on a decade of work and some hopes for the future. In R. J. Sternberg & L. Okagaki (Eds.),
Influences on children (pp. 147-180). Hillsdale, NJ: Erlbaum.
[3] Cassell, J., & Thorisson, K.R. (1999). The power of a nod and a glance: Envelope vs. emotional feedback in
animated conversational agents. Applied Artificial Intelligence, 13, 519-538.
[4] Chi, M. T. H., de Leeuw, N., Chiu, M., & LaVancher, C. (1994). Eliciting self-explanations improves
understanding. Cognitive Science, 18, 439-477.
[5] Collins, A. (1985). Teaching reasoning skills. In S.F. Chipman, J.W. Segal, & R. Glaser (Eds), Thinking and
learning skills (vol. 2, pp 579-586). Hillsdale, NJ: Erlbaum.
[6] Collins, A., Brown, J. S., & Newman, S. E. (1989). Cognitive apprenticeship: Teaching the craft of reading,
writing, and mathematics. In L. B. Resnick (Ed.), Knowing, learning, and instruction: Essays in honor of
Robert Glaser (pp. 453-494). Hillsdale, NJ: Erlbaum.
[7] Conati, C., & VanLehn, K. (1999). Teaching metacognitive skills: Implementation and evaluation of a tutoring
system to guide self-explanation while learning from examples. In S.P. Lajoie and M. Vivet, Artificial
Intelligence in Education (pp. 297-304). Amsterdam: IOS Press.
[8] Foltz, P.W. (1996). Latent semantic analysis for text-based research. Behavior Research Methods,
Instruments, and Computers, 28, 197-202.
[9] Fox, B. (1993). The human tutorial dialog project. Hillsdale, NJ: Erlbaum
[10] Gagné, R. M. (1977). The conditions of learning (3rd ed.). New York: Holdt, Rinehart, & Winston.
[11] Graesser, A.C., Franklin, S., & Wiemer-Hastings, P. & the Tutoring Research Group (1998). Simulating
smooth tutorial dialog with pedagogical value. Proceedings of the American Association for Artificial
Intelligence (pp. 163-167). Menlo Park, CA: AAAI Press.
[12] Graesser, A.C., & Person, N.K. (1994). Question asking during tutoring. American Educational Research
Journal, 31, 104 -137.
[13] Graesser, A.C., Person, N.K., & Magliano, J.P. (1995). Collaborative dialog patterns in naturalistic one-on-
one tutoring. Applied Cognitive Psychology, 9, 359-387.
[14] Graesser, A.C., Wiemer-Hastings, K., Wiemer-Hastings, P., Kreuz, R., & TRG (1999). AutoTutor: A
simulation of a human tutor. Journal of Cognitive Systems Research, 1, 35-51.
[15] Graesser, A.C., Wiemer-Hastings, P., Wiemer-Hastings, K., Harter, D., Person, N., & TRG (2000). Using
latent semantic analysis to evaluate the contributions of students in AutoTutor. Interactive Learning
Environments.

[16] Hu, X., Graesser, A. C., & the Tutoring Research Group (1998). Using WordNet and latent semantic
analysis to evaluate the conversational contributions of learners in the tutorial dialog. Proceedings of
the International Conference on Computers in Education, Vol. 2, (pp. 337-341). Beijing, China:
Springer
[17] Hume, G. D., Michael, J.A., Rovick, A., & Evens, M. W. (1996). Hinting as a tactic in one-on-one tutoring.
The Journal of the Learning Sciences, 5, 23-47.
[18] Johnson, W. L., & Rickel, J. W., & Lester, J.C. (in press). Animated pedagogical agents: Face-to-face
interaction in interactive learning environments. International Journal of Artificial Intelligence in Education.
[19] Kosko, B. (1992). Neural networks and fuzzy systems. New York: Prentice Hall.
[20] Landauer, T.K., & Dumais, S.T. (1997). A solution to Plato’s problem: The latent semantic analysis theory of
acquisition, induction, and representation of knowledge. Psychological Review.
[21] Landauer, T.K., Foltz, P.W., Laham, D. (1998). An introduction to latent semantic analysis. Discourse
Processes, 25, 259-284.
[22] Lepper, M. R., Woolverton, M., Mumme, D.L., & Gurtner, J.L. (1991). Motivational techniques of expert
human tutors: Lessons for the design of computer-based tutors. In S.P. Lajoie & S.J. Derry (Eds.), Computers
as cognitive tools (pp. 75-105). Hillsdale, NJ: Erbaum.
[23] Lesgold, A., Lajoie, S., Bunzo, M., & Eggan, G. (1992). SHERLOCK: A coached practice environment for an
electronics troubleshooting job. In J. H. Larkin & R. W. Chabay (Eds.), Computer-assisted instruction and
intelligent tutoring systems (pp. 201-238). Hillsdale, NJ: Erlbaum.
[24] Mayer, R. E., & Moreno, R. (1998). A split attention effect in multimedia learning: Evidence for dual
processing systems in working memory, Journal of Educational Psychology, 90, 312-320.
[25] McCauley, L., Gholson, B., Hu, X., Graesser, A. C., & the Tutoring Research Group (1998). Delivering
smooth tutorial dialog using a talking head. Proceedings of the Workshop on Embodied Conversation
Characters (pp. 31-38). Tahoe City, CA: AAAI and ACM.
[26] Merrill, D. C., Reiser, B. J., Ranney, M., & Trafton, J. G. (1992). Effective tutoring techniques: A comparison
of human tutors and intelligent tutoring systems. The Journal of the Learning Sciences, 2, 277-305.
[27] Moore, J.D. (1995). Participating in explanatory dialogues. Cambridge, MA: MIT Press.
[28] Olde, B. A., Hoeffner, J., Chipman, P., Graesser, A. C., & the Tutoring Research Group (1999). A
connectionist model for part of speech tagging. Proceedings of the American Association for Artificial
Intelligence (pp. 172-176). Menlo Park, CA: AAAI Press.
[29] Palinscar, A. S., & Brown, A. (1984). Reciprocal teaching of comprehension-fostering and comprehension-
monitoring activities. Cognition & Instruction, 1, 117-175.
[30] Person, N. K., Bautista, L., Kreuz, R. J., Graesser, A. C. & the Tutoring Research Group (2000). The dialog
advancer network: A conversation manager for AutoTutor. ITS 2000 Proceedings of the Workshop on
Modeling Human Teaching Tactics and Strategies. Montreal, Canada.
[31] Person, N. K., & Graesser, A. C. (1999). Evolution of discourse in cross-age tutoring. In A.M. O’Donnell
and A. King (Eds.), Cognitive perspectives on peer learning (pp. 69-86). Mahwah, NJ: Erlbaum.
[32] Person, N. K., Graesser, A. C., & the Tutoring Research Group (2000). Designing AutoTutor to be an
Effective Conversational Partner. Proceedings for the 4 th International Conference of the Learning Sciences.
Ann Arbor, MI.
[33] Person, N. K., Graesser, A. C., Harter, D., Mathews, E. C., & the Tutoring Research Group (2000). Dialog
Move Generation and Conversation Management in AutoTutor. Proceedings for the AAAI Fall Symposium
Series: Building Dialogue Systems for Tutorial Applications. Falmouth, Massachusetts.
[34] Person, N. K., Graesser, A. C., Kreuz, R. J., Pomeroy, V., & the Tutoring Research Group (2000). Simulating
human tutor dialog moves in AutoTutor. International Journal of Artificial Intelligence in Education.
[35] Person, N. K., Klettke, B., Link, K., Kreuz, R. J., & the Tutoring Research Group (1999). The integration of
affective responses into AutoTutor. Proceeding of the International Workshop on Affect in Interactions (pp.
167-178). Siena, Italy.
[36] Putnam, R. T. (1987). Structuring and adjusting content for students: A study of live and simulated tutoring
of addition. American Educational Research Journal, 24, 13-48.
[37] Soller, A., Linton, F., Goodman, B., & Lesgold, A. (1999). Toward intelligent analysis and support of
collaborative learning interaction. In S.P. Lajoie & M. Vivet (Eds.), Artificial Intelligence in Education (pp.
75-82). Amsterdam: IOS Press.
[38] van Lehn, K. (1990). Mind bugs: The origins of procedural misconceptions. Cambridge, MA: MIT Press.