Attention-Based Gated Convolutional Neural Networks for Distant Supervised Relation Extraction

Abstract

Distant supervision is an effective method to generate large-scale labeled data for relation extraction without expensive manual annotation, but it inevitably suffers from the wrong labeling problem, which would make the corpus much noisy. However, the existing research work mainly focuses on sentence-level noise filtering, without considering noisy words which widely exist inside sentences. In this paper, we propose an attention-based gated piecewise convolutional neural networks (AGPCNNs) for distant supervised relation extraction, which can effectively reduce word-level noise by selecting the inner-sentence features. On the one hand, we construct a piecewise convolutional neural network with gate mechanism to extract features that are related to relations. On the other hand, we employ a soft-label strategy to enable model to select important features automatically. Furthermore, we adopt an attention mechanism after the piecewise pooling layer to obtain high-level positive features for relation predicting. Experimental results show that our method can effectively filter word-level noise and outperforms all baseline systems significantly.