Decoupled Attention and Character–Word Mask for Chinese Nested Named Entity Recognition

Abstract

Vocabulary integration is an effective approach for improving named entity recognition performance. However, existing methods exhibit insufficient decoupling capability in modeling the position and content between characters and words, and under static matching mechanisms, models tend to memorize fixed character–word co-occurrence patterns, resulting in limited generalization. To address these issues, a Chinese nested NER method based on attention decoupling and character–word association mask is proposed. This method designs a decoupled mechanism that decomposes traditional character–word interactions into three independent matrices, content–content, content–position, and position–content, thereby achieving decoupled representation of position and content and enabling character position embeddings to actively perceive lexical word content. Meanwhile, a character–word association masking mechanism is designed to randomly mask character–word associations during training, preventing over-reliance on lexicons and mitigating overfitting. Experiments conducted on four datasets—Weibo, OntoNotes 4.0, CMeEE-V2, and CNERTA—demonstrate that the proposed method achieves F1-scores of 73.06%, 83.22%, 75.25%, and 81.68%, respectively, outperforming mainstream baseline methods and verifying the effectiveness and generalization capability of the model.