Chinese named entity recognition based on multi-feature embedding
This paper proposes a Chinese named entity recognition method based on multi-feature embedding.Firstly,BERT model is used to extract the word vector containing rich context information,then BiLSTM is used to extract the features of the word embedding vector and the word root embedding vector,meanwhile,the iterative expansive convolutional neural network(IDCNN)is used to extract the features of the glpyh embedding vector,and then the three feature vectors are combined.Input multiple layers of self-attention mechanism,integrate features dynamically,and extract key features.Finally,conditional random field(CRF)is used for annotation and decoding.In order to further improve the performance,knowledge distillation is introduced and teacher model is used to guide student model training.F1 value of 96.51%was obtained on the Resume dataset,a significant improvement over the baseline model.
named entity recognitionmulti-feature embeddingmulti-head self-attention mechanismknowledge distillation
万方数据
