Abstract: Relation extraction is an important task in the field of information extraction. In order to better solve the noise problem and deep semantic representation of sentences in the dataset, a distant supervision relation extraction model combining reinforcement learning and densely connected convolutional networks (DenseNet) is proposed, which is divided into sentence selector and relation classifier. In the sentence selector, the reinforcement learning method could search for noisy sentences and effectively improve the quality of input data. In the relation classifier, DenseNet could realize feature reuse and learn richer semantic features. The experimental results on the NYT dataset prove that the sentence selector can effectively filter noise, and the relation extraction performance of the model is better than the baseline model.