Improving Long-tail Relation Extraction with Knowledge-aware Hierarchical Attention

Abstract

Relation Extraction (RE) is a crucial step to complete Knowledge Graph (KG) by recognizing relations between entity pairs. However, it usually suffers from the long-tail issue, especially when using distantly supervision algorithm. In this paper, inspired by the rich semantic correlations between head relations and tail relations, we proposed a knowledge-aware hierarchical attention (KA-HATT) relation extraction model. According to relational hierarchy, the multiple layers of attention were established, which take advantage of the knowledge from data-rich classes to boost the performance of data-poor classes at the tail. We have conducted extensive experiments on available dataset New York Times (NYT). Experimental results show that, compared with baseline models, our model achieves significant improvements on relation extraction, especially for long-tail relations.