Named entity recognition based on fusion of global and local semantic features for Chinese electronic medical records

Named entity recognition for Chinese electronic medical records often suffers from insufficient feature representation due to inadequate integration of global contextual information and local semantic patterns. To address this issue, this paper proposes Named Entity Recognition based on the fusion of global and local semantic features for the Chinese electronic medical record(NERF). NERF employs bidirectional gated recurrent units to capture contextual semantics, a self-attention mechanism to model global dependencies, and a multilayer perceptron to perform nonlinear feature transformation on token-level representations. The global features and the local features are fused via addition and decoded using a conditional random field. In addition, adversarial training is used to improve robustness and generalization. The comparison experimental results on the CCKS2017 dataset shows that NERF outperforms all compared methods. Its precision, recall, and F1-score are 90.89%, 94.49%, and 92.65%, respectively.