A Bootstrapping Approach With CRF and Deep Learning Models for Improving the Biomedical Named Entity Recognition in Multi-Domains

Abstract

Biomedical named entity recognition (biomedical NER) is a core component to build biomedical text processing systems, such as biomedical information retrieval and question answering systems. Recently, many studies based on machine learning have been developed for a biomedical NER. The machine learning-based approaches generally require significant amounts of annotated corpora to achieve high performance. However, it is expensive to manually create a large number of high-quality corpora due to the demand for biomedical experts. In addition, most existing corpora have focused on several specific sub-domains, such as disease, protein, and species. It is difficult for a biomedical NER system trained with these corpora to provide much information for biomedical text processing systems. In this paper, we propose a method for automatically generating the machine-labeled biomedical NER corpus that covers various sub-domains by using proper categories from the semantic groups of a unified medical language system (UMLS). We use a bootstrapping approach with a small amount of manually annotated corpus to automatically generate a significant amount of corpus and then construct a biomedical NER system trained with the machine-labeled corpus. At last, we train two machine learning-based classifiers, conditional random fields (CRFs) and long short-term memory (LSTM), with the machine-labeled data to improve performance. The experimental results show that the proposed method is effective to improve performance. As a result, the proposed one obtains higher performance in 23.69% than the model that trained only a small amount of manually annotated corpus in F1-score.

Keywords

• Biomedical named entity recognition
• bootstrapping
• information extraction
• semi-supervised learning