A named entity recognition model embedding rehabilitation medicine Chinese character knowledge graph

Background Named entity recognition (NER) is pivotal for medical information extraction and clinical decision support. However, most studies on Chinese medicine NER account for single-feature and multi-feature embeddings of Chinese characters, neglecting the multi-feature correlations inherent in Chinese characters. This limitation is exacerbated in rehabilitation medicine due to sparse annotated data, complex terminology, and significant character-level polysemy, which traditional models struggle to address effectively. Methods To bridge this gap, this article constructs a novel NER framework integrating a rehabilitation medicine chinese character knowledge graph (RMCCKG). The RMCCKG not only encompasses Chinese character features, including character, radical, pinyin, stroke, structure, part of speech, and morphology, but also embraces interrelationships between these features. Expanding on this foundation, we propose the RMCCKG+BERT-BiLSTM-CRF NER model. We employ RMCCKG embeddings and Bidirectional Encoder Representations from Transformers (BERT) embeddings as joint input representations, where the RMCCKG embeddings serve as prior knowledge to assist the model in better understanding the text content. This model enables character-level semantic enhancement This model enables character-level semantic enhancement through a hybrid architecture combining bidirectional long short-term memory (BiLSTM) and conditional random field (CRF) layers. Results Experiments on our self-developed Rehab dataset and the public CMeEE dataset demonstrate that our model outperforms baseline methods, achieving a 3.96% F1-score improvement. Further, our studies further reveal that low-dimensional fusion of RMCCKG embeddings yields optimal performance, with significant gains in low-frequency entity recognition.