Study on Chinese Named Entity Recognition Based on Dynamic Fusion and Adversarial Training

This paper aims at the Chinese Named Entity Recognition task, uses NEZHA Chinese pre-trained language model as the word embedding layer, then adopts the BiLSTM network architecture to encode it, and finally connects the CRF layer to optimize the output sequence. In order to enhance the fusion of semantic features of the NEZHA model in the upper, middle, and lower layers, an attention mechanism has been adopted to integrate the NEZHA coding layers. At first, weight was given to each representation generated by its 12 Transformer coding layers. Secondly, the weight value was dynamically adjusted through supervised training, and then the generated layer representation was weighted average to get the final word embedded representation. Finally, some noise was introduced to the input data, which is used for adversarial training to improve the generalization and robustness of the model. The results show that the F1 Score of the proposed model on Chinese Clinical Named Entity Recognition Dataset and people’s daily corpus are respectively 98.52% and 96.84%, which are 2.36% and 4.21% respectively higher than the benchmark model Bert BiLSTM CRF.