Chinese Named Entity Recognition Based on Grid Tagging and Semantic Segmentation

Chinese Named Entity Recognition (CNER) is one of the fundamental tasks in Chinese Natural Language Processing and has been extensively researched. However, previous studies have primarily focused on specific types of CNER subtasks, which hindered the overall progress and practical implementation of CNER. Recently, grid-tagging-based methods have emerged as having significant advantages in information extraction tasks due to the flexible annotation framework and model architecture they provide. Building on this, we unify the CNER task as a character pair relation classification problem and solve it by predicting the character pair relation matrix, which is similar to the semantic segmentation task in computer vision. To do this, we utilize the U-shaped segmentation module to capture low-level and high-level features in the feature map of image style, enabling us to extract more semantic information. We also effectively model the boundary information of the entity, which in turn enhances the overall accuracy of entity recognition. To test the effectiveness of our method, we conduct experiments on two NER datasets containing flat and nested entities in the Chinese medical domain (CMeEEV2, MMC). Our results demonstrate that our approach improves the F1 score by 0.73% and 1.07% respectively compared to current state-of-the-art models, further validating its effectiveness.