Improving Text Matching with Semantic Dependency Graph via Message Passing Neural Network

摘要

Text matching is a core natural language processing research problem. Deep semantic alignment and comparison between two text sequences lie in the core of text matching. While the attention-based model achieves high accuracy through word-level or char-lever alignment, they ignore the deep semantic relations between words and have poor generalization performance. This paper presents a neural approach to leveraging the Chinese Semantic Dependency Graph for text matching. This model uses Message Passing neural network to encode the semantic relation between word and use these semantic associations to assist semantic alignment and comparison. Experimental results demonstrate that our method substantially achieves state-of-the-art performance compare to the strong baseline model. The further discussion shows that our model can improve the text alignment process and have better robustness and comprehensibility.