Named Entity Recognition Method Based on Fusion of Chinese Character and Word Features Using BERT Model

To address the issues of low accuracy and insufficient utilization of text information in the compliance recognition of text data of power marketing online cooperative channel business, this paper proposes a Chinese Named Entity Recognition (CNER) method based on word fusion Bidirectional Encoder Representation from Transformers (BERT) language model in support of the construction of a compliance recognition system. The method designs a structure of word feature fusion to effectively fuse text characters and related word features inside the pretrained model BERT, utilizing the text word boundary feature information. It then introduces a Bidirectional Gated Recurrent Unit (BiGRU) after the BERT, which is used to further learn the global context features of the text, and finally outputs the global optimal results of entity labeling through Conditional Random Field (CRF) decoding layer. The model constructed based on this method are validated on three public datasets Weibo, Ontonotes4, and MSRA, and the F1-scores of the models reach 71.33 %, 81.56 %, and 95.40 %,, which are all improved compared with the mainstream models.