Abstract
Recognizing medical named entities is a crucial aspect of applying deep learning in the medical domain. Automated methods for identifying specific entities from medical literature or other texts can enhance the efficiency and accuracy of information processing, elevate medical service quality, and aid clinical decision-making. Nonetheless, current methods exhibit limitations in contextual awareness and insufficient consideration of contextual relevance and interactions between entities. In this study, we initially encode medical text inputs using the Chinese pre-trained RoBERTa-wwm-ext model to extract comprehensive contextual features and semantic information. Subsequently, we employ recurrent neural networks in conjunction with the multi-head attention mechanism as the primary gating structure for parallel processing and capturing inter-entity dependencies. Finally, we leverage conditional random fields in combination with the cross-entropy loss function to enhance entity recognition accuracy and ensure label sequence consistency. Extensive experiments conducted on datasets including MCSCSet and CMeEE demonstrate that the proposed model attains F1 scores of 91.90% and 64.36% on the respective datasets, outperforming other related models. These findings confirm the efficacy of our method for recognizing named entities in Chinese medical texts.