Using Contextualized Representations For Biomedical Entity Recognition

Abstract

Distributed representations are usually used as input features in text mining tasks. Previous works have shown its potential in encoding semantics. Generally, there existing two representation methods namely static and dynamic, which means they are context-free and context-dependent respectively. Many works have demonstrated that context based representations significantly improved performance in natural language processing field. Therefore, in this paper, we utilize contextualized representations to recognize biomedical entities and evaluate the results at entity-level on BC2GM and BC5CDR-disease datasets. Results show that we obtain a F1-score of 75.16% and 75.97%, which improving 2.54% and 3.96% respectively compared with context-free representations. It indicates that the method based on contextualized representations is promising for entity recognition tasks.