Modelisation des emotions de l'apprenant et interventions implicites pour les Systemes Tutoriels Intelligents.

摘要

Modeling the user's experience within Human-Computer Interaction is an important challenge for the design and development of intelligent adaptive systems. In this context, a particular attention is given to the user's emotional reactions, as they decisively influence his cognitive abilities, such as perception and decision-making. Emotion modeling is particularly relevant for Emotionally Intelligent Tutoring Systems (EITS). These systems seek to identify the learner's emotions during tutoring sessions, and to optimize his interaction experience using a variety of intervention strategies.;This thesis aims to improve current methods on emotion modeling, as well as the emotional strategies that are presently used within EITS to influence the learner's emotions. More precisely, our first objective was to propose a new method to recognize the learner's emotional state, using different sources of information that allow to measure emotions accurately, whilst taking account of individual characteristics that can have an impact on the manifestation of emotions. To that end, we have developed a multimodal approach combining several physiological measures (brain activity, galvanic responses and heart rate) with individual variables, to detect a specific emotion, which is frequently observed within computer tutoring, namely : uncertainty. First, we have identified the key physiological indicators that are associated to this state, and the individual characteristics that contribute to its manifestation. Then, we have developed predictive models to automatically detect this state from the analyzed variables, trough machine learning algorithm training.;Our second objective was to propose a unified approach to simultaneously recognize a combination of several emotions, and to explicitly evaluate the impact of these emotions on the learner's interaction experience. For this purpose, we have developed a hierarchical, probabilistic and dynamic framework, which allows one to track the learner's emotional changes over time, and to automatically infer the trend that characterizes his interaction experience namely : flow, stuck or off-task. Flow is an optimal experience : a state in which the learner is completely focused and involved within the learning activity. The state of stuck is a non-optimal trend of the interaction where the learner has difficulty to maintain focused attention. Finally, the off-task behavior is an extremely unfavorable state where the learner is not involved anymore within the learning session. The proposed framework integrates three-modality diagnostic variables that sense the learner's experience including : physiology, behavior and performance, in conjunction with predictive variables that represent the current context of the interaction and the learner's personal characteristics. A human-subject study was conducted to validate our approach through an experimental protocol designed to deliberately elicit the three targeted trends during the learners' interaction with different learning environments.;Finally, our third objective was to propose new strategies to positively influence the learner's emotional state, without interrupting the dynamics of the learning session. To this end, we have introduced the concept of implicit emotional strategies : a novel approach to subtly impact the learner's emotions, in order to improve his learning experience. These strategies use the subliminal perception, and more precisely a technique known as affective priming. This technique aims to unconsciously solicit the learner's emotions, through the projection of primes charged with specific affective connotations. We have implemented an implicit emotional strategy using a particular form of affective priming namely : the evaluative conditioning, which is designed to unconsciously enhance self-esteem. An experimental study was conducted in order to evaluate the impact of this strategy on the learners' emotional reactions and performance.