Enriching the Student Model in an Intelligent Tutoring System

An intelligent tutoring system is a computer-based self-learning system which provides personalized learning content to students based on their needs and preferences. The importance of a students’ affective component in learning has motivated adaptive ITS to include learners’ affective states in their student models. Learner-centered emotions such as frustration, boredom, and confusion are considered in computer learning environments like ITS instead of other basic emotions such as happiness, sadness and fear. In our research we detect and respond to students’ frustration while they interact with an ITS. The existing approaches used to identify affective states include human observation, self-reporting, modeling affective states, face-based emotion recognition systems, and analyzing data from physical and physiological sensors. Among these, data-mining approaches and affective state modeling are feasible for the large scale deployment of ITS. Systems using data-mining approaches to detect frustration have reported high accuracy, while systems that detect frustration by modeling affective states not only detect a student’s affective state but also the reason for that state. In our approach we combine these approaches. We begin with the theoretical definition of frustration, and operationalize it as a linear regression model by selecting and appropriately combining features from log file data. We respond to students’ frustration by displaying messages which motivate students to continue the session instead of getting more frustrated. These messages were created to praise the student’s effort, attribute the results to external factors, to show sympathy for failure and to get feedback from the students. The messages were displayed based on the reasons for frustration. We have implemented our research in Mindspark, which is a mathematics ITS with a large scale deployment, developed by Educational Initiatives, India. The facial observations of students were collected using human observers, in order to build a ground truth database for training and validating the frustration model. We used 932 facial observations data from 27 students to create and validate our frustration model. Our approach shows comparable results to existing data-mining approaches and also with approaches that model the reasons for the students’ frustration. Our approach to responding to frustration was implemented in three schools in India. Data from 188 students from the three schools, collected across two weeks was used for our analysis. The number of frustration instances per session after implementing our approach were analyzed. Our approach to responding to frustration reduced the frustration instances statistically significantly–(p < 0.05)–in Mindspark sessions. We then generalized our theory-driven approach to detect