TOWARDS BUILDING PEDAGOGICAL AGENTS
BASED ON EXPERIMENTS
A Preliminary Result
Hanju Lee and Kazuo Hiraki
Graduate School of Arts and Sciences, The University of Tokyo, Tokyo, Japan
Keywords: Pedagogical Agent, Natural Pedagogy, Educational Applications.
Abstract: Pedagogical agents are computer generated characters that supports learning. Pedagogical agents gained
increasing interest in past decades, but research on pedagogical agents have produced mixed results on
learning outcomes. We plan to adopt trial-and-error approach, based on knowledge gathered from field of
cognitive science. While the main purpose of this paper is to discuss future direction of our work, it
describes an experiment carried out as a preliminary step. Five participants were presented with word
learning task, with and without physical image of pedagogical agent. The result suggests that pedagogical
agent itself does not harm learning, even when it is irrelevant to learning material.
1 INTRODUCTION
Pedagogical agents are characters presented on
computer screen that promote user’s learning
experience. Pedagogical agents are required to be
presented as physical image, but other functions
such as movement, voice are considered
supplemental.
Although research concerning pedagogical
agents has started promisingly in early 1990s, since
then it has produced mixed results, often producing
no or little educational outcome. Some authors argue
pedagogical agents have little direct impact on
learning (Baylor, 2011), but the conditions of use of
pedagogical agents (i.e. education contents,
pedagogy) and its design varies greatly among
studies, thus whether pedagogical agents can or
cannot produce positive educational outcome
remains unanswered (Heidig and Clarebout, 2011).
Research on pedagogical agents is highly
complex, requiring convergence of several fields of
studies, such as computer science, cognitive science,
and pedagogy. Among these, pedagogy is arguably
of utmost importance when designing pedagogical
agent, as it defines the goal of the system. In present
research, we follow natural pedagogy proposed by
Csibra and Gergely (2009). Natural pedagogy theory
claims that human communication is specifically
adapted to allow the transmission of generic
knowledge, and ability to receive such signals is
naturally given. Our goal is to assess subtle
movements triggering natural pedagogy and
implement them in pedagogical agents, hence
fostering effective and fast knowledge transfer. We
also follow teaching-as-transfer view over teaching-
as-communication view, as the latter has been
rejected both empirically and theoretically
(Kirschner et al., 2006), and as we believe
knowledge transfer is necessary for any other
educational methods.
The purpose of this paper is to describe an
experiment carried out as preliminary step of our
research, and to discuss future work.
2 EXPERIMENT DESIGN
Before working on ways to design effective
pedagogical agents, we felt the need to confirm that
fundamental elements of pedagogical agents do not
hinder user’s learning. Since irrelevant adjunct
information is said to hurt learning (Harp and
Mayer, 1998), elements of pedagogical agents such
as physical image and movement may actually act as
interference. Although numerous studies have
already shown pedagogical agents do not cause such
damage (Moreno et al., 2001, Moundridou and
Virvou, 2002), their interaction model may have
376
Lee H. and Hiraki K..
TOWARDS BUILDING PEDAGOGICAL AGENTS BASED ON EXPERIMENTS - A Preliminary Result.
DOI: 10.5220/0003957203760379
In Proceedings of the 4th International Conference on Computer Supported Education (CSEDU-2012), pages 376-379
ISBN: 978-989-8565-06-8
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
canceled out the original effect. To answer this
question, we created an agent that has minimal
interaction function and is irrelevant to learning
material. A simple foreign language word retention
task was used. The participants were presented with
read out words in their native language (Japanese)
and foreign language (Korean), and were asked to
remember the latter. A photo was presented for each
word, to clarify the meaning.
The design included one within-subject
variables: pedagogical agent’s visibility. During
learning phase, two conditions were provided: with
physical image of pedagogical agent (Ag condition),
and without physical image of pedagogical agent
(no-Ag condition). Each participant learned under
both conditions, and was tested immediately.
3 METHOD
3.1 Pedagogical Agent
While limiting its functions, we aimed to design
agent’s appearance to be humanlike, to open up the
potential to equip it with subtle human behaviours.
The agent was developed as three-dimensional
character using Face Robot©, and photorealistic
textures were used to enhance graphical quality (see
figure 1). Although the agent is capable of
behaviours such as blinking, eye movement, and
complex facial expression, only lip movement was
applied for this experiment. The agent’s lip
movement was synchronized to pre-recorded voice
using viseme method, and was done semi-
automatically by Face Robot©.
Figure 1: A close-up view of the pedagogical agent.
3.4 Participants
At current stage, we were able to gather 5 college
students (3 females). Every participant was native
Japanese speakers, and had minimal pre-experience
with Korean, the foreign language used in this
experiment.
3.5 Procedure
After arriving at the laboratory, participants were
told to seat in front of a computer screen, and were
instructed to memorize Korean words. They were
told there would be a test afterwards. After learning
phase, participants were tested after 1-min break.
The experiment was done in soundproof
environment. Whole process took around ten
minutes.
3.5.1 Learning Phase
For learning phase, Ag condition and no-Ag
condition were provided (see figure2). After learning
under one condition, each participant learned under
another condition without taking any break. To
assess order effect on which condition comes first,
order of conditions was counter balanced. Fifteen
words were presented for each condition, total of
thirty words. A photo was displayed, and a word
describing the photo was presented by voice output,
first Japanese then Korean. Each word was read out
twice. Every word was a noun, consisting of less
than five syllables (for both Japanese and Korean).
Words were presented in random order.
Figure 2: The participants learned from two conditions. In
Ag condition (left) pedagogical agent was visible, and in
no-Ag condition (right) it was not.
3.5.2 Test Phase
After 1-min break, participants were tested whether
they remembered the words. Four photos were
presented and word describing one of them was read
out twice, and participants had to pick the right
answer (see figure3). All thirty words were tested.
Figure 3: A sample of computer screen during test phase.
TOWARDSBUILDINGPEDAGOGICALAGENTSBASEDONEXPERIMENTS-APreliminaryResult
377
3.5.3 Voice
Voice was pre-recorded by native Korean speaker,
who also speaks near-perfect Japanese. Text-to-
speech was not used for several reasons:
Does not support multiple language in same
voice;
Lacks modulation, which may be crucial
for natural pedagogy
Human voice have been reported not only
to be preferable over synthesized voice, but
to have advantage on learning
outcomes(Atkinson et al., 2005)
4 RESULTS
Because the number of participants enrolled in the
experiment was limited, the data were not analyzed
statistically. However, the test scores (see Table 1)
showed no significant difference between A and no-
Ag condition. Four participants remembered words
better when they were taught under Ag condition,
and while one participant remembered less, the
margin was minimal (one word). The order effect of
conditions did not appear consistently, on either
direction. All participants scored above chance level,
which was 25%.
Table 1: Number of errors of each participants (P1~P5).
ORDER OF CONDITIONS (FIRST)
Ag no-Ag
P1 P2 P3 P4 P5
Total 17 7 2 5 3
no-Ag 11 5 2 2 2
Ag 6 2 0 3 1
5 CONCLUSIONS AND FUTURE
WORK
In this experiment, the physical image of
pedagogical agent did not damage learning
outcomes, even when it was irrelevant to learning
material and lacked interaction. Although more
experiment is required regarding participant group
and more complex learning materials, one can
carefully argue that pedagogical agent itself is not
harmful to learning.
In future work, we plan on bracing pedagogical
agent with more sophisticated behaviours that could
trigger more smooth information transfer. Attention
guiding system using eye movement, gesture and
facial expression is being implanted, using
technologies such as motion tracker, camera and
eye-tracker, and we want to improve it to point
where it could actually produce positive learning
outcomes. We are also working on automatic voice
modulation system, in belief that voice modulation
plays great role during information transfer process.
We are planning to gather participants from
much younger groups (i.e. preschoolers), to find the
natural way in which human receives information.
They can provide new and important insight, since
they are not yet trained by traditional ways of
learning, and they do not posses ability to overcome
inefficiencies that may reside on our way of teaching,
unlike adults.
As for learning materials, while adding more
complexity, they will be carefully selected from
subjects that learning outcomes could be tested
numerically and presented clearly. We will also
attempt to compare our system with other teaching
methods, and identify the cause of gaps between.
While this study focuses on efficient information
transfer, we do not seek to bring older way of
teaching back which was criticized of putting too
much emphasis on memorizing and strong guidance.
Instead, our ultimate goal is to make the transfer
process as fast and not painful as possible to let
more time to be spent on creative, meaningful
learning.
Finally, we will be working with revision process
in mind, which differs from previous attempts made
by other groups. Experiments and updating policy
will be based on extensive knowledge gathered from
field of cognitive science.
ACKNOWLEDGEMENTS
This study was supported by JST, CREST.
REFERENCES
Atkinson, R. K., Mayer, R. E. & Merrill, M. M. 2005.
Fostering social agency in multimedia learning:
Examining the impact of an animated agent's voice.
Contemporary Educational Psychology, 30, 117-139.
Baylor, A. L. 2011. The design of motivational agents and
avatars. Etr&D-Educational Technology Research and
Development, 59, 291-300.
Csibra, G. & Gergely, G. 2009. Natural pedagogy. Trends
in Cognitive Sciences, 13, 148-153.
Harp, S. F. & Mayer, R. E. 1998. How seductive details do
their damage: A theory of cognitive interest in science
learning. Journal of Educational Psychology, 90, 414-
434.
CSEDU2012-4thInternationalConferenceonComputerSupportedEducation
378
Heidig, S. & Clarebout, G. 2011. Do pedagogical agents
make a difference to student motivation and learning?
Educational Research Review, 6, 27-54.
Kirschner, P. A., Sweller, J. & Clark, R. E. 2006. Why
minimal guidance during instruction does not work:
An analysis of the failure of constructivist, discovery,
problem-based, experiential, and inquiry-based
teaching. Educational Psychologist, 41, 75-86.
Moreno, R., Mayer, R. E., Spires, H. A. & Lester, J. C.
2001. The case for social agency in computer-based
teaching: Do students learn more deeply when they
interact with animated pedagogical agents? Cognition
and Instruction, 19, 177-213.
Moundridou, M. & Virvou, M. 2002. Evaluating the
persona effect of an interface agent in a tutoring
system. Journal of Computer Assisted Learning, 18,
253-261.
TOWARDSBUILDINGPEDAGOGICALAGENTSBASEDONEXPERIMENTS-APreliminaryResult
379
