Abstract: Existing joint extraction tasks of entities and relationships introduce distant supervision strategies to automatically generate large-scale training data, leading to severe problems of noisy data during data processing. To address the issue of noisy data, this paper proposes an entity relation joint extraction model with reinforcement learning integration. The model consists of two components: reinforcement learning and joint extraction model. The joint extraction model is composed of a graph convolutional network and a multi-head self-attention mechanism. Firstly, reinforcement learning is utilized to eliminate noisy sentences from the original dataset, and the denoised high-quality sentences are input into the joint extraction model. Secondly, the joint extraction model is employed to predict and extract entities and relationships from the input sentences, and provide feedback rewards to the reinforcement learning component to guide it in selecting high-quality sentences. Finally, the reinforcement learning and joint extraction models are jointly trained and iteratively optimized. The experiments demonstrating that the proposed model can effectively address the issue of data noise and outperform baseline methods in entity relationship extraction.