Definition Response Scoring with Probabilistic Ordinal Regression

Word knowledge is often partial, rather than all-or-none. In this paper, we describe a method for estimating partial word knowledge on a trial-by-trial basis. Users generate a free-form synonym for a newly learned word. We then apply a probabilistic regression model that combines features based on Latent Semantic Analysis (LSA) with features derived from a large-scale, multi-relation word graph model to estimate the similarity of the user response to the actual meaning. This method allows us to predict multiple levels of accuracy, i.e., responses that precisely capture a word's meaning versus those that are partially correct or incorrect. We train and evaluate our approach using a new gold-standard corpus of expert responses, and find consistently superior performance compared to a state-of-the-art multi-class logistic regression baseline. These findings are a promising step toward a new kind of adaptive tutoring system that provides fine-grained, continuous feedback as learners acquire richer, more complete knowledge of words.