Chinese Named Entity Recognition Method Based on RoFormer-GlobalPointer and Improved R-drop
Yujie Zhang, Jinhua Li, Shuaiwei Huan · 2022
Aiming at solving the problems of blurred entity boundary and poor generalization ability of traditional named entity recognition, a named entity recognition method based on RoFormer-GlobalPointer and R-drop is proposed. The word vector with long-term distance information and semantic information is obtained through the RoFormer pre-training model with rotational position encoding; the entity recognition is carried out by considering the head and tail of the entity as a whole GlobalPointer method, and the proposed R-drop method improves the generalization of the model. Experiments show that the model achieves 98.6% values on the Resume datasets. Compared with the baseline BERT-CRF model, the method is improved by 2.8%, which is higher than the current mainstream models.