A CRFs-Based Approach Empowered with Word Representation Features to Learning Biomedical Named Entities from Medical Text

Targeting at identifying specific types of entities, biomedical named entity recognition is a fundamental task of biomedical text processing. This paper presents a CRFs-based approach to learning disease entities by identifying their boundaries in texts. Two types of word representation features are proposed and used including word embedding features and cluster-based features. In addition, an external disease dictionary feature is also explored in the learning process. Based on a publically available NCBI disease corpus, we evaluate the performance of the CRFs-based model with the combination of these word representation features. The results show that using these features can significantly improve BNER performance with an increase of 24.7% on F1 measure, demonstrating the effectiveness of the proposed features and the feature-empowered approach.