CPMFA: A Character Pair-Based Method for Chinese Nested Named Entity Recognition.

Chinese Nested Named Entity Recognition (CNNER) faces several challenges due to the language diversity phenomena, the complexity of the language, and the imbalanced distribution of entity types in Chinese text. To address these challenges in CNNER, we propose a new method called CPMFA (Character Pair-based method with Multi-feature representation and Attention mechanism). The CPMFA method predicts the predefined relations of character pairs in a sentence, and identifies nested named entities based on these relations. First, our method utilizes the pre-trained language model LERT (Linguistically-motivated Bidirectional Encoder Representation from Transformer), and Bidirectional Long Short-Term Memory (BiLSTM) to generate comprehensive and precise character representations. Second, our method uses multi-feature representation to capture complex semantic information within the text, and employs the Pyramid Squeeze Attention (PSA) module to emphasize key features. Finally, to overcome the challenge of the imbalanced distribution of entity types, PolyLoss function is integrated into our model training process. Results of experiments show that the proposed CPMFA method achieves an F1 score of 83.79%. Compared to other mainstream span-based methods, the proposed CPMFA method has excellent performance in CNNER.