Broadening input understanding in a language-based intelligent tutoring system
Martha W. Evens, Michael Glass · 1999
The CIRCSIM-Tutor intelligent tutoring system (ITS) tutors medical students about the baroreceptor reflex, a mechanism for controlling blood pressure. CIRCSIM-Tutor is predicated on the belief that students learn more when they are forced to use language to produce answers, thus the tutor processes free-text student input. The goal of this research was to replace the input understanding component in the current version of the ITS and suggest improvements which can be made for the next version. Even though the current version of CIRCSIM-Tutor asks closed questions which can be satisfied with short answers, student input can contain a variety of messy phenomena, including strange syntax, impromptu abbreviations, spelling errors, “near miss” answers to questions, and answers which are appropriate but unexpected, equations interpolated into the text, and hedged answers. Answers which don't match the category of what was being asked for can often be diagnostic of student difficulties which ought to be tutored. To address these issues I implemented an input understander based on finite-state transducers, similar to the approach used by many information extraction systems. I discuss how it is possible to add a category of more open questions to CIRCSIM-Tutor's tutorial repertoire, where the tutor tries to gather information from the student. These questions admit of a wide variety of possible answers, which I show can be processed by the technique of Latent Semantic Analysis. I argue that merely asking these questions, with a very imperfect understanding of the answers, should be enough for improving learning. The new input understander has been incorporated into CIRCSIM -Tutor where it is currently used by physiology classes at Rush Medical College.