Distant Supervision for Relation Extraction with Hierarchical Attention and Entity Descriptions

Distant supervision for relation extraction is an effective method to find novel relational facts from plain text. However, distant supervision inevitably suffers from wrong label problem, and most existing methods generally focus on direct sentences containing entity pairs, but ignore massive information from background knowledge. To tackle these problems, we propose a novel hierarchical attention model to select valid instances and capture vital semantic information in them. Furthermore, we incorporate entity descriptions extracted from Wikipedia into the hierarchical attention model to provide supplementary background knowledge. The proposed architecture can not only combat the noise introduced by distant supervision, but also adequately extract latent and helpful background information. The experimental results on both Chinese and English datasets show that the proposed approach consistently achieves significant improvements on relation extraction as compared with strong baselines.