Cross-corpus named entity recognition on Chinese electronic medical records based on label sharing
Xuelei Wang, Degen Huang · International Conference on Artificial Intelligence and Intelligent Information Processing (AIIIP 2022) · 2022
Named entity recognition (NER) on electronic medical records (EMRs) is a prerequisite task of medical information extraction. However, due to the high data sensitivity and the labeling difficulty of EMRs, there are few available data resources, making it difficult for NER to reach a practical level on EMRs. To alleviate the problem of low-resource, we propose a label sharing-based cross-corpus NER (LSCC-NER) model, which consists of the shared and private encoders divided by the cross-corpus label similarity. We design a category-wise multi-head self-attention unit for each encoder and introduce the entity category prediction task (ECP) to realize the division. Besides, considering the nested entities and data noise in EMRs, we utilize span-based decoding methods and adversarial training to further improve the robustness of our model. The experiments on two evaluation datasets of EMRs show that our proposed LSCC-NER model can achieve higher recognition performance compared with common transfer learning methods.