An Improved BERT-BiLSTM-CRF Model Integrating CNN and Transformer for Chinese Named Entity Recognition

This paper proposes a Chinese Named Entity Recognition (NER) method to address the gradient vanishing and exploding issues encountered by the Bidirectional Encoder Representation from Transformer (BERT)-Bidirectional Long Short-Term Memory (BiLSTM)–Conditional Random Field (CRF) model when processing long sequences. This approach incorporates Transformer to enhance the model's ability to capture long-range dependencies and contextual information, then further integrates Convolutional Neural Network (CNN) module to effectively capture local features, improve computational efficiency and enhance the model’s generalizability. The proposed model BERT-CNN-BiLSTMTransformer-CRF (Bert-CBTC) achieves efficient local feature extraction, robust global semantic modeling, accurate sequence labeling and complex contextual modeling. Extensive experiments on a Chinese NER dataset verify the model’s effectiveness, achieving an F1 score of 99.80, an accuracy of 99.97 on the dataset. Compared with other models, the proposed model improves Precision by 3.55-30.78, Recall by 6.17-43.3, and F1 score by 4.88-34.58, showing significant advantages. Beyond NER tasks, Bert-CBTC can be applied to other natural language processing (NLP) tasks such as relation extraction, providing a powerful solution for future text sequence modeling and showing excellent potential for adaptability and generalization.