Adaptive Question Generation for Student Modeling in Probabilistic Domains

Abstract

Problem solving behavior remains to be the most trustable source for modeling student knowledge in intelligent tutoring systems. In this work we focus on diagnostic problem solving, as an essential question type associated with probabilistic domains. Student answer for such questions indicates the knowledge discrepancies between the student and his/her stored model. In this paper we introduce an algorithm that adaptively generates different appropriate follow-up questions to accurately determine the knowledge discrepancies in the student model. Answers to these follow-up questions are used to update the student model. Verification is conducted on the updated model based on the matching between student and generated model answers to the presented questions. Results suggest that tracking the student knowledge discrepancies using the generated follow-up questions improves the prediction accuracy of the student answers by 20% compared to relying only on the diagnostic questions alone. In addition, approximation of the student model enhanced by 40% relative to that obtained using the diagnostic questions alone.