A Multi-grained Attention Network for Multi-labeled Distant Supervision Relation Extraction

Distant supervision is an efficient approach to generate labeled data automatically for relation extraction and it is by nature a multi-instance multi-label learning issue. However, previous methods mainly consider that each sentence explicitly express one relation and ignore the relation overlapping problem for which their models are incompetent to capture complete context features. The existing methods for multi-labeled distant supervision relation extraction incorporate with capsule network to handle the highly overlapping relations, but the computational complexity of the model is too high. To this end, we propose a novel multi-grained attention neural network for relation overlapping problem. Further, we devise a adaptive-class-balanced loss function to handle the class-imbalanced problem in distant supervision relation extraction. The experimental results on a widely used dataset show the effectiveness of proposed method to make multi-label predictions and to alleviate class-imbalanced problem.