Aggregating Class Interactions for Hierarchical Attention Relation Extraction

Abstract

Distantly supervised relation extraction is a powerful learning method to recognize relations of entity pairs. However, wrong label problem is inevitable among large-scale training data. In this work we propose a hierarchical attention neural network to effectively alleviate the impact of noise instances. Moreover under distantly supervised scenario, connections and dependencies widely appear among relation classes, which we call class interactions. Previous end-to-end methods that considered the relations as independent failed to make use of these interactions. To better utilize these important interactions, we propose a soft target as training objective to learn class relationships jointly. Experiments show that our model outperforms state-of-the-art methods.