Chinese Medical Named Entity Recognition Using Pre-Trained Language Models and an Efficient Global Pointer Mechanism

Abstract

Medical Named Entity Recognition (MNER) is a critical task in medical text mining, serving as a foundation for intelligent diagnosis, disease prediction, and related tasks. However, Chinese medical texts present unique challenges, such as diverse entity classifications, varying entity lengths, and complex entity nesting, which hinder NER performance. To address these issues, we propose a Chinese MNER method leveraging pre-trained models and the efficient global pointer (EGP). Our approach incorporates three key innovations: (1) data augmentation techniques, including easy data augmentation and remote supervision with a medical entity dictionary, to address imbalanced entity types distribution and varying entity lengths; (2) a character-word fusion strategy that integrates RoBERTa-generated character vectors and Word2Vec-generated word vectors, enhancing semantic representation; and (3) an improved decoding layer based on EGP, enabling efficientrecognition of both nested and flat entities while reducing computational costs. Experimentals show that our method achieves F1 scores of 75.87% and 92.77% on the CMeEE-V2 and CCKS2020 datasets, respectively, outperforming the RoBERTa-BiLSTM-CRF baseline by 3.06% and 4.38%, respectively.