We proposes an enhanced zero-shot relation extraction method based on a self-feedback mechanism for generating relation descriptions (SFDG-RE). By leveraging a self-feedback optimization mechanism and the powerful prior knowledge of large language models (LLMs), this method generates high-quality relation description texts and performs semantic matching, addressing the challenges of data sparsity and generalization in dynamic environments for zero-shot relation extraction. Unlike previous methods that rely on extensive labeled data or manual tuning, our approach decomposes the relation extraction task into description generation and semantic matching tasks, introducing a self-feedback mechanism. SFDG-RE iteratively optimizes relation descriptions during generation and employs a multi-task learning model along with a self-attention mechanism, significantly improving the model’s matching performance and generalization capability. Compared to traditional methods, SFDG-RE not only innovatively integrates self-feedback with relation description generation but also demonstrates substantial performance improvements in experiments. The results indicate that our method outperforms existing state-of-the-art methods across multiple benchmark datasets and exhibits better robustness and flexibility as the number of relations increases compared to baseline methods.