Previous supervised methods for dialogue topic segmentation have primarily employed Transformer-based models for this classification task. However, they often suffered from issues such as a shallow understanding of topic structures and an overreliance on local surface features. To address these challenges, we propose GTLC, a novel hierarchical framework that integrates a global topic module and a local coreference module, drawing inspiration from window-attention theory. Our approach simulates the two stages of human semantic comprehension: first, a broad perspective to grasp general semantics, followed by a more focused view to identify key details. Specifically, we leverage large language models (LLMs) to extract global topic information at the sentence level while capturing local coreference relations of key entities at the word level. The distilled capabilities of the model are then employed to segment dialogues based on the combined information from both levels. Experimental results on three public datasets demonstrate the effectiveness of GTLC, highlighting its ability to enhance topic segmentation through multi-granularity linguistic cues.