Distant supervision is an effective method to generate large-scale la-beled data for relation extraction without expensive manual annotation,but it inevitably suffers from the wrong labeling problem,which would make the cor-pus much noisy.However,the existing research work mainly focuses on sen-tence-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 ex-traction,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 rela-tions.On the other hand,we employ a soft-label strategy to enable model to se-lect important features automatically.Furthermore,we adopt an attention mech-anism 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.