Semantic matching refers to the process of measuring the similarity between texts. It is fundamental for various NLP tasks. However, challenges such as contextual fragmentation and insufficient global modeling limit its applications. This paper proposes a hybrid framework that integrates BERT and LLMs. First, a dynamic segmentation strategy is designed based on LDA topic modeling and keywords extraction to partition text segments and prevent semantic fragmentation. Second, BERT is employed to extract local features, while LLMs generate cross-paragraph global semantic supplements to improve contextual coherence. Furthermore, a weighted fusion mechanism and a bidirectional matching network are introduced to comprehensively compute similarity by integrating both local and global information. Experimental results demonstrate that the proposed model significantly outperforms existing baselines. This method effectively balances fine-grained local features and global semantic associations.