Multi-hop knowledge graph reasoning is an effective method for conducting complex knowledge question answering using natural language. Most existing researches adopt the iterative relation reasoning approach based on the overall semantics of the question to perform multi-hop knowledge graph reasoning. However, due to the overlooks of the local characteristics and fine-grained information of the relationships, there exist some issues such as large scales of candidate reasoning paths and unclear semantic relationships which need to be solved. In this paper, we propose a LLM(large language model) enhanced two-stage approach for multi-hop knowledge graph reasoning, aiming to enhance the efficiency and accuracy of complex knowledge question answering. Initially, our approach leverages the text understanding and generation capabilities of large language model, and employs structured knowledge template prompts to conduct coarse-grained relation filtering within a large set of candidate reasoning path relations, thus reducing the scale of the reasoning space. Subsequently, concerning the fine-grained evaluation of the semantic relations of candidate reasoning paths, a relation scoring mechanism based on SBERT+BiLSTM is designed, which is used to evaluate the relations from three perspectives: the similarity between the question and the relations, the relevance of the relations to the reasoning paths, and the degree of match between the question and candidate paths. Experimental results on standard datasets including QALD10, CWQ and GrailQA demonstrate that our approach achieves significant performance improvements in multi-hop KGQA tasks compared to existing approaches.