2000 Volume 15 Issue 6 Pages 996_2
Student modeling, or more specifically, the diagnosis of student errors, is the key to intelligent individualized educational systems. One of the major bottlenecks in the development of student modeling systems lies in the difficulty of automatically acquiring the background knowledge, particularly the bug library, which is the collection of common errors made by students in a particular domain. This dissertation addresses the problem of automatically discovering and detecting novice programmer errors at the knowledge as well as behavior levels. Knowkedge errors are incorrect or missing pieces of knowledge, which, in turn, result in errors in behavior, e.g., erroneous programs. A multistrategy error detection and discovery system, MEDD, is presented that solves this problem in two steps. The first step involves intentionbased diagnosis by a subsystem called MDD, in which the programmer's intention is discerned, and the discrepancies between the student's program and the intended program are computed and then further examined whether they occur in an erroneous program. The second step involves multistrategy conceptual clustering, in which the behavior-level errors computed by MDD are clustered by another subsystem called MMD into an error hierarchy, whose main subtrees form intensional definitions of error classes that denote knowledge-level errors. Results show that MEED/MDD is capable of correctly determining the behavior-level errors of incorrect novice programs, and that MEDD/MMD can discover the knowledge-level errors in novice programs even when several knowledge-level errors occur in a single program, and when the programs are fed into MEDD in a different order. The novelty of MEDD as a system for student modeling is three-fold. First MEDD automatically constructs and extends bug libraries for novice (Prolog) programming. Second, the bug library is constructed/extended at the same time as a student model in inferred. Third, a student model can be constructed from a single behavior, with the help of knowledge learned from past multiple behaviors. The novelty of MEDD as a machine learning system is two-fold. First, MEDD learns first order classifiers for first order objects. Second, similarity and causality are more tightly coupled in MEDD than in any other extant system for unsupervised concept formation.
