Relation extraction (RE) plays a crucial role in extracting information from Chinese texts but faces notable challenges, including difficulty in accurately recognizing overlapping triples and balancing the modeling of global and local semantic features. To address these issues, we propose a novel RE method that integrates a global–local perception mechanism and contrastive learning. The proposed mechanism comprises a local feature extraction module designed to capture fine-grained semantic distinctions between entity pairs and a global semantic association module that encodes sentence-level contextual information, thereby improving the overall feature representation. Moreover, the method introduces a contrastive optimization strategy based on contrastive learning, which enhances the model’s ability to cluster semantically similar relations while effectively distinguishing dissimilar ones. This strategy further improves the model’s discriminative ability to relation. Experimental results on two public Chinese datasets indicate that our model outperforms previous methods, confirming the effectiveness of our proposed approach in RE.