Abstract In previous research on Chinese named entity recognition (CNER), most studies have focused mainly on using lexical information to enhance character-level representations. Such methods often miss task-relevant information from the data, and remain vulnerable to noise, which hurts performance, ultimately leading to suboptimal model performance. In this paper, we propose a vocabulary-enhanced CNER method based on information interaction and optimization. Initially, character feature vectors and lexical feature vectors are obtained, respectively, and the information is interactively integrated through graph attention networks. Additionally, we introduce a novel highway feature refinement learning network, designed to process and optimize the integrated information for complex relationships, thereby obtaining information useful for downstream tasks. Our model achieves F1 values of 71.62%, 95.88%, and 82.06% on the Weibo, Resume, and OntoNotes datasets, respectively, showing its superior performance in lexical enhancement and deep-level useful information acquisition compared to other models. Particularly on the Weibo dataset, which is characterized by a significant amount of noisy data, our model still achieves a higher F1 score, significantly outperforming other baseline models and exhibiting stronger generalization capability and flexibility.
Ke et al. (2026) studied this question.