Going deeper: Automatic short-answer grading by combining student and question models

Abstract

As various educational technologies have rapidly become more powerful and more prevalent, especially from the 2010s onward, the demand of automated grading natural language responses has become a major area of research. In this work, we leverage the classic student and domain/question models that are widely used in the field of intelligent tutoring systems to the task of automatic short-answer grading (ASAG). ASAG is the process of applying natural language processing techniques to assess student-authored short answers, and conventional ASAG systems often mainly focus upon student answers, referred as answer-based. In recent years, various deep learning models have gained great popularity in a wide range of domains. While classic machine learning methods have been widely employed to ASAG, as far as we know, deep learning models have not been applied to it probably because the lexical features from short answers provide limited information. In this work, we explore the effectiveness of a deep learning model, deep belief networks (DBN), to the task of ASAG. Overall, our results on a real-world corpus demonstrate that 1) leveraging student and question models to the conventional answer-based approach can greatly enhance the performance of ASAG, and 2) deep learning models such as DBN can be productively applied to the task of ASAG.