Model Tracing in MITutor: Knowledge Discovery by Topic Models

• In Event: Technology, Instruction, Cognition and Learning SIG. Assessment With Technology: Diverse Data, Collection, and Use

Abstract

This paper outlines a method to discover generalized behavioral examples in therapeutic dialogue that fail to adhere to treatment protocol. The proposed method of knowledge discovery by models relies on Latent Dirichlet Allocation (LDA), a type of statistical modeling for inferring the topics that occur in student responses to virtual therapist agents in MITutor, an intelligent tutoring system for training in motivational interviewing. We applied the method to 170 responses made by 21 students to virtual agents. Results suggest that an LDA model allows MITutor to not only check if a student response is incorrect, but also to gain insights into why it is incorrect and choose the most appropriate hint to deliver to students.

Authors

• Kent Ellsworth

• Eric G. Poitras, Dalhousie University

• Zac Imel, The University of Utah

• Derek Caperton

• Grin Lord

• Jake Van Epps, The University of Utah

• Michael Tanana, The University of Utah

• David Atkins, University of Washington