Mining knowledge components from many untagged questions

Abstract

An ongoing study is being run to ensure that the McGraw-Hill Education LearnSmart platform teaches students as efficiently as possible. The first step in doing so is to identify what Knowledge Components (KCs) exist in the content; while the content is tagged by experts, these tags need to be re-calibrated periodically. LearnSmart courses are organized into chapters corresponding to those found in a textbook; each chapter can have anywhere from about a hundred to a few thousand questions. The KC extraction algorithms proposed by Barnes [1] and Desmarais et al [3] are applied on a chapter-by-chapter basis. To assess the ability of each mined q matrix to describe the observed learning, the PFA model of Pavlik et al [4] is fitted to it and a cross-validated AUC is calculated. The models are assessed based on whether PFA's predictions of student correctness are accurate. Early results show that both algorithms do a reasonable job of describing student progress, but q matrices with very different numbers of KCs fit observed data similarly well. Consequently, further consideration is required before automated extraction is practical in this context.