Named Entity Recognition on COVID-19 Scientific Papers

Text mining techniques, especially named entity recognition (NER), play a vital role in supporting researchers for keeping track of hundred thousand of papers on COVID-19 related literature. Although a few research has been performed NER on COVID-19 scientific papers, very little is currently known concerning the behaviors of current entity recognition models in this new domain. Therefore, this ongoing study attempts to analyze current NER models’ performance and limitations on the CORD-19 dataset. By examining three NER models, this study showed that NER performance is improved with the similarity between the testing and pretraining data. When there are little manually annotated resources for COVID-19 NER exist, our analysis suggested that for training purposes, enhancing the dictionary for seed annotation is effective (not necessarily requiring costly human annotation).