A Dynamic Character and Word Information Fusion Method for Chinese Named Entity Recognition

Named Entity Recognition(NER) is a fundamental task in natural language processing and plays a crucial role in many downstream tasks. Existing research on Chinese Named Entity Recognition mainly relies on character information and supplementary static word information, which still cannot well solve the problem of multiple meanings of one word without obvious separators in Chinese Named Entity Recognition. To address this problem, a dynamic character and word information fusion method for Chinese Named Entity Recognition is proposed based on a pre-trained language model. The dynamic word information is supplemented by using a BERT model based on word training, a Cross-Attention mechanism is introduced in the encoding stage to obtain the character and word potential information, neural networks with different structures are used to extract the implicit features for character and word information respectively, and a Gate mechanism is designed to filter and fuse the implicit features, and finally the labels are obtained by decoding through a Conditional Random Field. Through experiments on five Chinese datasets, the F1 values are 80.84% for Ontonotes 4.0, 95.20% for MSRA, 95.82% for Resume, 71.72% for Weibo and 79.62% for CLUENER2020. The highest enhancement of the proposed method is 3.52%, and the experimental results show that the proposed dynamic character and word information fusion method can effectively enhance Chinese Named Entity Recognition.