A Two-Stream Self-attention Multi-digraph Model for Chinese NER

As an important international language, Chinese has unique characteristics in many aspects, and Chinese Named Entity Recognition (CNER) is receiving increasing attention. However, traditional CNER methods face the challenge of balancing semantic disambiguation and the effective extraction of text sequence features. To tackle this issue, we propose a dynamic weighted fusion model based on a two-stream selfattention mechanism and multi-digraphs (DW-TSM). In this method, we first construct a graph structure by combining Chinese corpora with various entity dictionaries. We then embed the text that forms the graph using a gated graph neural network (GGNN). Subsequently, we input the shallow features obtained from the embedded text into an encoder containing a two-stream self-attention mechanism for feature extraction. We further fuse the output of each layer in the encoder through dynamic weighting and dimensionality reduction to enhance the model’s semantic recognition ability. Finally, we input the encoded feature representation into a standard BiLSTM-CRF layer to predict the globally optimal label sequence by considering the adjacent relationships between labels. Experimental results show that the proposed DW-TSM model outperforms traditional sequence labeling models and neural network models with an average F1 score improvement of approximately 1.7% on the People’s Daily, Weibo NER, and MSRA datasets, validating the superior performance of the proposed model in achieving effective named entity recognition in Chinese tasks.