Chinese NER Based on Adversarial Training and Interactive Attention
Sihao Wu, Xiaoling Wu, Rui Xie, Wentao Zhan · 2024
In recent years, the Chinese named entity recognition model based on BERT-BiLSTM-CRF has become one of the mainstreams. BiLSTM model is better at extracting global features, but is not good at extracting local features and assigning feature information weights adequately. In order to improve the recognition effect of named entities, a model based on adversarial training and interactive attention (AT-IA) is proposed. Firstly, RoBERTa is used to preprocess the text in the embedding layer. At the same time, slight disturbances are added to the obtained vector to generate adversarial samples, and the word vector and adversarial samples are input into the feature layer to enhance the recognition ability of fuzzy entity boundaries. BiLSTM and TextCNN are used in the feature extraction layer to solve the problem of insufficient feature extraction ability of single neural network. Then the extracted features are processed through the proposed interactive attention layer, and finally the two feature vectors are fused and input to CRF layer for decoding. The model proposed in this paper is tested on MSRA and Resume Chinese datasets, and the F1 scores of 95.01% and 96.73% are obtained respectively. The proposed model outperforms other state-of-art models, which proves its effectiveness in Chinese named entity recognition.