Distant Supervision is a common technique for relation extraction from large amounts of free texts, but introduces wrong labeled sentences at the same time. Existing deep learning approaches mainly rely on CNN-based models. However, they fail to capture spatial patterns due to the inherent drawback of pooling operations and thus lead to suboptirnal performance. In this paper, we propose a novel framework based on Selective Capsule Network for distant supervision relation extraction. Compared with traditional CNN-based models, the involvement of capsule layers in the sentence encoder makes it more powerful in encoding spatial patterns, which is very important in determining the relation expressed in a sentence. To address the wrong labeling problem, we introduce a high-dimensional selection mechanism over multiple instances. It is one generalization of traditional selective attention mechanism and can be seamlessly integrated with the capsule network based encoder. Experimental results on a widely used dataset (NYT) show that our model significantly outperform all the state-of-the-art methods.
展开▼
