Improving biomedical named entity recognition through transfer learning and asymmetric tri-training

Abstract

Today, electronic health records have turned into prime sources of information for physicians looking after their patients. EHRs and computerized patient data resources have expedited the accelerated discovery of formerly obscure biomedical and clinical information. Because of the lengthy, error-prone, non-scalable, and expensive manual abstraction process, natural language processing (NLP) procedures are being wielded more and more in biomedical and clinical fields. One of the building blocks of all NLP systems, Named Entity Recognition (NER) is considered a sub-activity of information retrieval. To extract biomedical knowledge from electronic health records, a prerequisite is the efficient recognition of biomedical entities. Deep-learning techniques have gained more and more consideration recently for the above-mentioned task. Notwithstanding, these methods are based on high-caliber, high-cost, labeled data. In this work, a biomedical-named entity recognition model based on transfer learning and asymmetric tri-training is proposed to diminish the limited annotated data problem in the biomedical-named entity recognition domain. The proposed model showed a significant improvement of more than 9% over the baseline BiLSTM-CRF model in the exact F1 scores on four different datasets considered in this work.