Representation and Reasoning for Deeper Natural Language Understanding in a Physics Tutoring System

Students ’ natural language (NL) explanations in the domain of qualitative mechanics lie in-between unre-stricted NL and the constrained NL of “proper ” do-main statements. Analyzing such input and providing appropriate tutorial feedback requires extracting infor-mation relevant to the physics domain and diagnosing this information for possible errors and gaps in reason-ing. In this paper we will describe two approaches to solving the diagnosis problem: weighted abductive rea-soning and assumption-based truth maintenance system (ATMS). We also outline the features of knowledge rep-resentation (KR) designed to capture relevant semantics and to facilitate computational feasibility