AUTOMATICALLY UNDERSTANDING HANDWRITTEN SELF-EXPLANATIONS

Research has demonstrated that self-explanation hones student's metacognitive skills and increases their performance. We have found however, that not all self-explanation is substantive. Our goal is to develop computational techniques capable of determining whether a student's explanation is relevant or not. This will then enable us, for example, to create an interactive tutoring system capable of prompting students to continue their explanations when necessary. This is a tractable task as self-explanations typically contain a small number of possible concepts. The language used to express these concepts can vary greatly, but our task is only to identify the existence of the concepts, not to perform general machine interpretation. In this paper, we present early work on the automatic understanding of students' handwritten self-explanation of their solutions to homework problems in an engineering statics course. We employ an open information extraction technique popularly used to identify relations present in broadcast news transcripts. In our study, this technique achieved up to 97% accuracy at identifying when the content of a student's self-explanation did not match the concepts used by experts in explaining their own work on the same problem.