DBGARE: Across-Within Dual Bipartite Graph Attention for Enhancing Distantly Supervised Relation Extraction

Hejian Gu, Hang Yu, Xiangfeng Luo

Abstract: Recently, a kind of method based on single bipartite graph combined with single attention mechanism was proposed to solve the challenges of noisy labels and long-tailed distributions in distantly supervised relation extraction, which achieved competitive results. However, we argue that single bipartite graph will lose interactive information within each bipartite graph partition and single attention will cause attention bias. In order to improve the robustness of this method for noisy labels and long-tailed distributions, we propose a novel framework DBGARE, whereby graphs across and within each bipartite graph partition are constructed to enrich unbalanced relations dependency, based on which, a novel technique dual attention mechanism is devised to avoid attention bias and promote information dissemination from data-riches to data-poors. Multiple experiments on various widely-used real-world datasets show the state-of-the-art effectiveness of our framework in both challenges above. Furthermore, almost previous benchmark datasets are in English, we provide CIPDS $$*$$ , a new Chinese dataset collected from Chinese Corpus in Chinese search engine, to expand distantly supervised relation extraction research field.