Application of Pre-trained Models Integrated with Prior Knowledge in Named Entity Recognition of Classical Chinese Medical Texts

Abstract This study constructs a high-quality BIO-annotated dataset for Named Entity Recognition (NER) tasks in traditional Chinese medicine (TCM), utilizing classical medical texts such as Shanghan Lun, Jingui Yaolüe, and Wenbing Tiaobian. The dataset encompasses six entity categories, including medicinal substances, diseases, and symptoms. To comprehensively evaluate model performance, we systematically compared eight mainstream methods: traditional models (HMM, CRF), deep learning models (BiLSTM, BiLSTM+CRF), pre-trained models (BERT, BERT+CRF, BERT+BiLSTM+CRF), and an improved model integrating domain-specific prior knowledge (BERT+BiLSTM+CRF+KNOW). Experimental results demonstrate that the BERT+BiLSTM+CRF+KNOW model with prior knowledge achieves state-of-the-art performance, with accuracy, recall, and F1-score reaching 90.03%, 87.74%, and 88.87%, respectively. This significantly outperforms other models (e.g., BERT+BiLSTM+CRF without prior knowledge yields an F1-score of 87.80%), validating the effectiveness of domain knowledge integration in enhancing semantic understanding of TCM texts.