The purpose of relation extraction is to discern semantic relationships between entities within sentences, positioning it as a classification task. In recent years, the integration of graph neural networks (GNNs) with large language models (LLMs) has shown promise in addressing relation extraction challenges. However, prevalent graph-centric approaches still grapple with notable challenges: firstly, they often fail to harness dependency relation information comprehensively; secondly, a standardized protocol for pruning dependency trees in the context of these models remains elusive. To mitigate these concerns, we propose a two-phase graph convolutional network (GCN) tailored for relation extraction, leveraging the contextual understanding provided by a large language model. In the initial phase of our model, we integrate node representations derived from the LLM, dependency relation type representations, and dependency type weights to collectively derive novel node embeddings, thereby exhaustively leveraging dependency relation information. Transitioning to the second phase, we leverage the adjacency matrix extracted from the dependency tree to execute graph convolution operations, guided by insights from the LLM. This dual-phase approach enables our model to dynamically and autonomously prune the dependency tree while benefiting from the contextual richness of the large language model. We evaluated our methodology on two publicly accessible datasets, and the outcomes demonstrate that our model surpasses prior research in terms of the F1 score, attaining state-of-the-art performance. Furthermore, supplementary ablation experiments underscore the efficacy of each constituent within our proposed model, validating the contribution of both the GCN and the large language model.