Relation extraction is a core task in natural language processing, focusing on predicting the relation labels between given entities in a text. However, existing relation extraction models face several challenges, including insufficient logical reasoning, inadequate semantic information in relation labels, and being prone to misclassification. To address these issues, we propose the Updating Relation Label Word Representations Prompt Contrastive Learning(UPCL) Framework. The framework (1) designs a novel template that provides explicit reasoning steps and can improve the ability of the model to perform complex reasoning. Within this framework, (2) the representation of relation label word is updated by using sentence information in the training set, (3) furthermore the representation is trained using a contrastive learning strategy. Experimental results show that our model has demonstrated improved performance on three relation extraction datasets, proving the effectiveness of our model. To verify the model’s generalization capability, we also design multiple experiments for different scenarios, and experiments demonstrate that UPCL significantly outperforms baselines in various datasets.