Research on Chinese named Entity Recognition based on RoBERTa and word fusion

Abstract

With the rapid progress of Internet medicine, a good deal of data is generated every day, which is of great significance to clinical decision-making system and medical entity research. For electronic medical records in Chinese(CEMR) named entity recognition(NER) task of long entity, the entity chaos, border demarcation difficulties and other issues, this paper proposes a fusion based on RoBERTa, and words of Chinese named entity recognition method. This method uses the joint feature representation of characters and entity words depended on the pre-training model RoBERTa and the medical field lexicon created by ourselves, which can precisely spilt entity boundaries, so as to solve the influence of unregistered words. The experimental results show that our model has preferable performance and significant improvement compared with the baseline model. Specifically, the F1 value of CCKS2020 CEMR dataset reached 88.71%, 1.16% higher than all the baseline models.