Automatic Chinese Reading Comprehension Grading by LSTM with Knowledge Adaptation

Owing to the subjectivity of graders and the complexity of assessment standard, grading is a tough problem in the field of education. This paper presents an algorithm for automatic grading of open-ended Chinese reading comprehension questions. Due to the high complexity of feature engineering and the lack of consideration for word order in frequency based word embedding models, we utilize long-short term memory recurrent neural network to extract semantic feature in student answers automatically. In addition, we also try to impose the knowledge adaptation from web corpus to student answers, and represent the students' responses to vectors which are fed into the memory network. Along this line, the workload of teacher and the subjectivity in reading comprehension grading can both be reduced obviously. What's more, the automatic grading methods for Chinese reading comprehension will be more thorough. The experimental results on five Chinese and two English data sets demonstrate the superior performance over compared baselines.