LDRC: Long-tail Distantly Supervised Relation Extraction via Contrastive Learning

Abstract

Long-tail problem is one of the major challenges in distantly supervised relation extraction. Some recent works on the long-tail problem attempt to transfer knowledge from data-rich and semantically similar head classes to data-poor tail classes using a relation hierarchical tree. These methods, however, are based on the assumption that long-tail and head relations have a strong correlation, which does not always hold true, and the model’s ability to learn long-tail relations is essentially not improved. In this paper, a novel joint learning framework that combines relation extraction and contrastive learning is proposed, allowing the model to directly learn the subtle differences between different categories to improve long-tail relation extraction. Experimental results show that our proposed model outperforms the current state-of-the-art (SOTA) model on various mainstream datasets.