Automatic Chinese Short Answer Grading with Deep Autoencoder

Xi Yang, Yuwei Huang, Fuzhen Zhuang, Lishan Zhang, Shengquan Yu

Abstract: Short answer question is a common assessment type of teaching and learning. Automatic short answer grading is the task of automatically scoring short natural language responses. Most previous auto-graders mainly rely on target answers given by teachers. However, target answers are not always available. In this paper, a deep autoencoder based algorithm for automatic short answer grading is presented. The proposed algorithm can be built without expressly defining target answers, and learn the lower-dimensional representation of student responses. For the sake of reducing the influence of data imbalance, we introduce the expectation regularization term of label ratio into the model. The experimental results demonstrate the effectiveness of our proposed method.