Incorporating Instance Correlations in Distantly Supervised Relation Extraction

Luhao Zhang, Linmei Hu, Chuan Shi

Abstract: Distantly-supervised relation extraction has proven to be effective to find relational facts from texts. However, the existing approaches treat the instances in the same bag independently and ignore the semantic structural information. In this paper, we propose a graph convolution network (GCN) model with an attention mechanism to improve relation extraction. For each bag, the model first builds a graph through the dependency tree of each instance in this bag. In this way, the correlations between instances are built through their common words. The learned node (word) embeddings which encode the bag information are then fed into the sentence encoder, i.e., text CNN to obtain better representations of sentences. Besides, an instance-level attention mechanism is introduced to select valid instances and learn the textual relation embedding. Finally, the learned embedding is used to train our relation classifier. Experiments on two benchmark datasets demonstrate that our model significantly outperforms the compared baselines.