Assessment of Engagement for Intelligent Educational Agents: a Pilot Study with Middle School Students

Adaptive learning is an educational method that utilizes computers as an interactive teaching device. Intelligent tutoring systems, or educational agents, use adaptive learning techniques to adapt to each student’s needs and learning styles in order to individualize learning. Effective educational agents should accomplish two essential goals during the learning process – 1) monitor engagement of the student during the interaction and 2) apply behavioral strategies to maintain the student’s attention when engagement decreases. In this paper, we focus on the first objective of monitoring student engagement. Most educational agents do not monitor engagement explicitly, but rather assume engagement and adapt their interaction based on the student’s responses to questions and tasks. A few advanced methods have begun to incorporate models of engagement through vision-based algorithms that assess behavioral cues such as eye gaze, head pose, gestures, and facial expressions. Unfortunately, these methods typically require a heavy computation load, memory/storage constraints, and high power consumption. In addition, these behavioral cues do not correlate well with achievement of highlevel cognitive tasks. As an alternative, our proposed model of engagement uses physical events, such as keyboard and mouse events. This approach requires fewer resources and lower power consumption, which is also ideally suited for mobile educational agents such as handheld tablets and robotic platforms. In this paper, we discuss our engagement model which uses techniques that determine behavioral user state and correlate these findings to mouse and keyboard events. In particular, we observe three event processes: total time required to answer a question; accuracy of responses; and proper function executions. We evaluate the correctness of our model based on an investigation involving a middle-school after-school program in which a 15-question math exam that varies in cognitive difficulty is used for assessment. Eye gaze and head pose techniques are referenced for the baseline metric of engagement. We conclude the investigation with a survey to gather the subject’s perspective of their mental state after the exam. We found that our model of engagement is comparable to the eye gaze and head pose techniques for low-level cognitive tasks. When high-level cognitive thinking is required, our model is more accurate than the eye gaze and head pose techniques due to the students’ nonfocused gazes during questions requiring deep thought or use of outside variables for assistance such as their fingers to count. The large time delay associated with the lack of eye contact between the student and the computer screen causes the aforementioned algorithms to incorrectly declare the subjects as being disengaged. Furthermore, speed and validity of responses can help to determine how well the student understands the material, and this is confirmed through the survey responses and video observations. This information will be used later to integrate instructional scaffolding and adaptation with the educational agent.