A Method of Chinese NER Based on BERT Model and Coarse-Grained Features
Yafei Wang, Hua Li, Peilong Zhang, Xianrong Wang, Yulin Liu · 2022
Named entity recognition (NER) is a fundamental task and an important aspect in the fields of information extraction, natural language understanding and retrieval systems. Currently, NER based on deep learning is better than traditional feature and kernel function-based approaches in the field of feature extraction depth and modeling accuracy. The traditional word-based feature approach tends to ignore the coarse-grained word features, and the recognition of named entities in Chinese requires a comprehensive consideration of word-level and word-level features. To address the above issues, the combination of the BERT pre-training model and the Lattice network structure integrates coarse-grained features and word features, and CRF is used to decode sequence labels. The BERT-Lattice-CRF model is proposed in this paper, we use the public data set Resume for testing. Then we can conclude that the F1 value of the model on the data set is significantly improved.