Recently, automated short-answer grading (ASAG) methods based on deep neural networks (DNN) have attained high scoring accuracy. However, the accuracy requires further improvement especially for large-scale and high-stakes tests because a slight scoring error will strongly influence many examinees. To improve the accuracy, this study proposes a new DNN-based ASAG method that utilizes examinees’ abilities which are estimated using an item response theory model from their true-false responses for objective exam questions offering with a target short-answer question.
