<p>Multi-hop knowledge graph question answering represents a complex task that requires multi-step reasoning and relies on multiple intermediate triplets for answer inference. While existing methods are dedicated to enhancing logical reasoning capabilities, they exhibit deficiencies in handling spurious path problems. Traditional path retrieval methods often generate numerous candidate paths due to redundant expansion, making it difficult to effectively filter spurious reasoning paths, thereby reducing the interpretability of results. Meanwhile, shallow modeling of question semantics tends to overlook crucial contextual clues, which subsequently affects reasoning accuracy. To address these challenges, this paper proposes the LLM-based Semantic Enhancement and Context-Guided GNNs (LSECG) framework that integrates GNNs with large language model semantic parsing, achieving context-sensitive semantic representation through large model semantic enhancement and feature extraction modules. Furthermore, we design a multi-level node representation module and an adaptive structure-aware pooling subgraph context aggregation method that effectively integrates local and global information while dynamically adjusting node aggregation information through adaptive weighting. Experimental results on the WebQuestionsSP (WebQSP) and ComplexWebQuestions (CWQ) benchmark datasets validate the effectiveness of the proposed method in terms of reasoning accuracy and result interpretability. Ablation experiments further demonstrate the expressiveness and robustness of each module.</p>