Identifying Patterns For Short Answer Scoring Using Graph-based Lexico-Semantic Text Matching

Short answer scoring systems typically use regular expressions, templates or logic expressions to detect the presence of specific terms or concepts among student responses. Previous work has shown that manually developed regular expressions can provide effective scoring, however manual development can be quite time consuming. In this work we present a new approach that uses word-order graphs to identify important patterns from humanprovided rubric texts and top-scoring student answers. The approach also uses semantic metrics to determine groups of related words, which can represent alternative answers. We evaluate our approach on two datasets: (1) the Kaggle Short Answer dataset (ASAP-SAS, 2012), and ( We show that our automated approach performs better than the best performing Kaggle entry and generalizes as a method to the Mohler dataset.