How to Model a Likely Behavior of a Pedagogical Agent from a Real Situation

The aim of this work is to model the behavior verbal and nonverbal behavior of a Pedagogical Agent (PA) can be integrated into an Intelligent Tutoring System. The following research questions were posed: what is the nonverbal component of an educational communication? How to study this component for a computational model of plausible behavior of a virtual agent? What correlations between educational actions and the direction of gaze of a human agent? To carry out exploratory work, a methodological approach based on multi-modal video corpus study was adopted. Within a multidisciplinary team consisting of computer scientists and didactics of mathematics, an educational situation in which a virtual pedagogical agent is likely was developed. Dyadic interactions between teachers and learners late second early third (15-16 years) in a skills assessment interview in mathematics following the resolution of exercises by students with a mathematics software was filmed. A multi-level annotation scheme to annotate the observed behavior was proposed. The multidisciplinary research subject (ITS, Human-Machine Interfaces, Educational sciences, educational, linguistic etc.) due to the development of the coding scheme a delicate but important work given the wealth of knowledge from different disciplines. After a portion of the collected work annotation corpus statistical measures derived from annotations carried out suggest different strategies for teachers in terms of gaze direction depending on the learner profile and pedagogical actions. These measures have enabled to extract rules to control the nonverbal behavior of a PA. Keywords—pedagogical agent; nonverbal communication; behavior; corpus analysis