Chinese medical entity recognition based on the dual-branch TENER model.

Abstract

Named Entity Recognition (NER) is a long-standing fundamental problem in various research fields of Natural Language Processing (NLP) and has been practiced in many application scenarios. However, the application results of NER methods in Chinese electronic medical records (EMRs) are not satisfactory, mainly due to the following two problems: (1) Existing methods do not take into account the impact of medical terminology on model recognition performance, resulting in poor model performance. (2) Existing methods do not fully utilize the Chinese language features contained in EMR, resulting in poor model robustness. Therefore, it is imminent to solve these two problems regarding the performance of the NER model for EMRs. In this paper, a TENER-based radical feature and entity augmentation model for NER in Chinese EMRs is proposed. The TENER model is first used in the pre-training stage to extract deep semantic information from each layer of the feature extractor. In the decoder part, the recognition of medical entity boundary and entity category are divided into two branch tasks. We compare the overall performance of the proposed model with existing models on different datasets using the computed F1 score evaluation metric. The experimental results show that our model achieves the best F1 score of 82.67%, 74.37%, 70.16% on the CCKS2019, ERTCMM, and CEMR data sets. Meanwhile, in the CMeEE challenge, our model surpassed the top-3 with the F1 score of 68.39%. Our proposed model is the first to divide the NER task into a two-branch tasks, entity boundary and types recognition. Firstly, the medical entity dictionary information is integrated into TENER to obtain the feature information of professional terms in Chinese EMRs. Secondly, the features of Chinese radicals in Chinese EMRs extracted by CNN are added to the entity category recognition task. Finally, the effectiveness of the model is validated on four datasets and competitive results are achieved.