Hardness-guided domain adaptation to recognize biomedical named entities under low-resource scenarios

Abstract

Domain adaptation is a promising approach to address data scarcity in low-resource scenarios. However, applying it to token-level tasks like biomedical Named Entity Recognition (bioNER) poses challenges due to the unique linguistic characteristics of clinical narratives, often resulting in unsatisfactory performance.

In this paper, we introduce a hardness-guided domain adaptation (HGDA) framework specifically designed for bioNER tasks. This framework effectively utilizes domain hardness information to enhance the adaptability of the learned model in low-resource situations.

Experimental results on various biomedical datasets demonstrate that our model significantly outperforms the recently published state-of-the-art (SOTA) MetaNER model, highlighting the effectiveness of our approach in improving performance in challenging contexts.