NBCGP: A Global Pointer based Network with BiLSTM and Chinese Pretrained Model for Named Entity Recognition

The named entity recognition (NER) task aims to identify entities belonging to predefined semantic types (such as persons, locations, organizations, etc.) from a piece of text. Currently, both deep learning-based methods and pretrained model-based methods have achieved good results. However, for Chinese text, existing models have not shown ideal accuracy. This article proposes a novel NER model, based on the Chinese pretrained model NEZHA, bidirectional long short-term memory network (BiLSTM), and Global Pointer technology to build a neural network framework(NBCGP). The model uses NEZHA to capture contextual information for word embedding, further feature extraction is done by BiLSTM, then the outputs of NEZHA and BiLSTM are concatenated to obtain richer semantic information. The dimensionality reduction is performed through fully connected layers, and finally, labeling classification is done using Global Pointer. We tested this model on three common datasets and compared it with other classical models. Experimental results show that this model outperforms existing methods in terms of F1 score and speed, especially on datasets with nested entities. This demonstrates the effectiveness and superiority of this model.