Dual-stream feature fusion encoding for Chinese named entity recognition

Abstract

Recently, the attention mechanism and convolutional operation have been widely applied in Chinese Named Entity Recognition (NER) owing to their parallelization ability. However, in Chinese NER, the attention mechanism tends to model for global information and almost ignores local feature information between characters. On the contrary, convolutional operation can capture the local information while inability to solve patterns with discontinuous characters. In this paper, we propose a novel dual stream feature fusion encoding method. Specifically, we design a Dual-stream Temporal Network (DSTN), in which we use the advantages of both convolutional operation and self-attention mechanism while alleviate their respective drawbacks. DSTN can effectively capture the local and global feature information by encoding the characters. Besides, we also present a loss calculation method, namely Multi-loss, which can prevent the model from over-fitting. The experiment results on two NER datasets showed that our method has excellent performance and efficiency than most mainstream methods.