Multi-granularity Semantic Extraction and Multi-task Fusion for Chinese Medical Entity Normalization

Medical Entity Normalization (MEN) aims to map informal clinical mentions to formal medical concepts in certain terminologies. Despite the rapid development of MEN research in English, Chinese MEN presents unique challenges, including (1) semantic confusion arising from missing word delimiters and compound terms, (2) limited annotated data with context-deficient texts, and (3) the multi-implication issue where a single mention may correspond to multiple formal concepts. In this study, we propose a novel Chinese MEN framework that integrates multi-granularity semantic extraction and multi-task fusion to tackle these challenges. First, we introduce a word-lattice structure to capture rich word-level semantics of Chinese entities while mitigating segmentation errors. Second, we employ a pre-trained medical language model to encode character-level semantics with reduced data reliance, enhanced by adversarial training for robust few-shot fine-tuning. Third, we utilize a multi-task fusion framework to jointly model implication number prediction and mention-concept matching, effectively addressing the multi-implication issue. By incorporating multi-similarity loss to guide online hard negative mining, the multi-task training process effectively captures discriminative features of ambiguous concepts. Our model is evaluated on clinical entity normalization datasets from the CHIP 2019 and 2020, achieving high accuracy of 91.54% for procedures and 68.2% for diagnoses, significantly surpassing other baseline methods. The suggested model is expected to disambiguate Chinese medical terms, providing a solid foundation for downstream clinical applications.