A Cross-Lingual Hybrid Neural Network with Interaction Enhancement for Grading Short-Answer Texts

Automatic Short-Answer Grading (ASAG) is an application for recognizing textual entailment in smart education. With the continuous expansion of the application scope of artificial neural networks, many deep learning models have been applied to grading short-answer texts. However, the coding structures and interaction forms of existing models are still too simple to meet the requirements of the ASAG task, resulting in low scoring accuracy. To address these challenges, we propose a cross-lingual hybrid neural network with interaction enhancement for ASAG. First, we sequentially use a convolutional neural network and bidirectional Long Short-Term Memory (LSTM) network to encode the answer text. Then, we introduce an interaction enhancement layer consisting of reference-answer-to-student-answer and student-answer-to-reference-answer attentions, and we combine the attentions and their inputs to form enhanced representations of answer texts. Finally, we introduce two Siamese Bi-LSTM networks to fuse the enhanced representations of answer texts and combine their multiple pooled vectors for grade classification on a multi-linear prediction layer. The experimental results show that our model significantly improves the performance of various simple models for Chinese and English ASAG tasks. The code is available online at https://github.com/wuhan-1222/DL_ASAG.