DCM-CNER: A Dual-Channel Model for Clinical Named Entity Recognition Based on Embedded ConvNet and Gated Dilated CNN

As the volume of Chinese electronic medical records (EMRs) experiences an explosive growth, the application of clinical named entity recognition (CNER) technology becomes crucial for the effective utilization of EMR data and practical implementation of evidence-based medicine. While mainstream models excel in capturing global contextual information, their feature extraction mechanisms tend to be unidimensional, limiting their information retrieval capabilities. To address this issue, this paper proposes a Dual-Channel Model for CNER (DCM-CNER), which enhances feature extraction through the introduction of a customized convolutional stack, denoted as emConvNet (embedded ConvNet), for local feature extraction. The model further addresses the vanishing gradient issue by employing a newly designed Gated Dilated Convolutional Neural Network (GDCNN) module with a residual structure. Sequential and contextual information is extracted from the text using a Bidirectional Long Short-Term Memory (BiLSTM) layer. Additionally, a multi-head bilinear attention mechanism is introduced for parallel dynamic feature fusion. Experimental results demonstrate the superiority of the proposed DCM-CNER model, in comparison to the existing mainstream models and state-of-the-art models, achieving F1 scores of 94.15%, 85.26%, and 84.21% on the CCKS2017, CCKS2019, and CLUENER2020 datasets, respectively, thereby validating its effectiveness in performing the task of Chinese CNER in EMRs.