Named Entity Recognition of Chinese Diabetes Based on Multi-Feature Fusion and Multi-Head Attention Mechanism

Named Entity Recognition (NER) for diabetes in China is the foundation for processing Chinese diabetic medical data. However, the structure of Chinese diabetes medical data is complex, with issues such as confusion between entity types and ambiguous entity boundaries, which pose difficulties and challenges for named entity recognition in Chinese diabetes data. Current mainstream named entity recognition models typically use single-feature representations of text, which do not fully use the features of the text, resulting in poor entity recognition performance. Therefore, this paper proposes a Chinese diabetes named entity recognition method that uses multi-feature fusion and a multi-head attention mechanism. It uses the RoBERTa-wwm pre-trained model to obtain character vectors from Chinese diabetes medical texts and enhances their semantic features by incorporating Chinese radical-level structural feature vectors. The obtained vectors are processed through a BiGRU module to extract global context features and a CNN module to capture multi-scale local features. After fusing these features, the most relevant feature information for entities is obtained by integrating the multi-head attention mechanism. Finally, the conditional random field is used for entity recognition to complete extraction of the entities. Experimental results show that the proposed method exhibits superior performance and achieves better results on the Chinese diabetes annotation dataset.