Enhanced Chinese Named Entity Recognition with Transformer-Based Multi-feature Fusion

In the field of Named Entity Recognition (NER), the integration of semantic features is crucial for enhancing model performance. Particularly in Chinese NER tasks, the inclusion of lexical information into character-based models is essential due to the lack of clear word boundaries. Additionally, effectively fusing multiple semantic features is of utmost importance for further performance enhancement. To address these challenges, this paper proposes a novel Chinese Named Entity Recognition model, termed DFL-NER. The model adopts an improved Transformer feature fusion method, beginning with the extraction of word vectors using the pre-trained DeBERTa model. Subsequently, lexical features are integrated into character vectors through an enhanced SoftLexicon algorithm, and ultimately, dynamic word vectors, lexical features, and glyph features are combined through a feature fusion module to enrich semantic representation. Experimental results demonstrate that compared to the baseline model, DFL-NER achieves F1-score improvements of 7.14%, 0.37%, and 1.81% on the Weibo, Resume, and MSRA datasets, respectively, showcasing outstanding performance.