Abstract Named entity recognition (NER) is crucial for automating the extraction of key legal entities from unstructured legal texts. However, Chinese legal documents present significant and unique challenges, particularly the coexistence of fine-grained local terminology patterns and complex, long-range contextual dependencies. This paper proposes a novel hybrid architecture, bidirectional encoder representations from transformers (BERT)-convolutional neural network (CNN)-Transformer-conditional random field, to robustly enhance NER performance in Chinese legal documents. This synergy is strategically designed to address this dual challenge: the CNN module efficiently captures local features and precise entity boundaries, while the Transformer model effectively models long-distance dependencies and global context, thereby enabling accurate recognition of complex and interrelated legal entities. Evaluations on the CAIL2018 dataset demonstrate that the proposed model achieves a strong F1 score of 91.85%, significantly outperforming several strong baseline methods. Ablation studies confirm the critical, synergistic contribution of both CNN and Transformer in the overall performance, validating the necessity of combining local feature extraction and global dependency modeling. This work provides a robust and highly effective solution for legal entity recognition in Chinese legal texts, with potential applications in intelligent legal document processing systems. All code and data are publicly available at https://github.com/Hero-Legend/CNN-Transformer-for-Legal-NER.
Zhou et al. (Sun,) studied this question.