Representation learning on Text-Attributed Graphs (TAGs) is crucial for deeply analyzing complex network data in real-world scenarios. Current approaches which utilize Pretrained Language Model (PLMs) and Graph Neural Networks (GNNs) mainly analyze the intrinsic correlation between semantic and structural information at a single granularity. However, semantic and structural features inherently exhibit multi-granularity characteristics, which are essential for effectively modeling TAGs. In this paper, we propose a multi-granularity representation learning method for Text-Attributed Graphs (MugraTAG) to achieve effective integration of semantic and structural information at various granularities. Firstly, we construct a multi-granularity TAG via a graph coarsening strategy, leveraging Large Language Models (LLMs) to integrate and inherit structural and semantic attributes across granularities. Secondly, we design the embedding method to explicitly capture intra-granularity structural and semantic features. Finally, we perform inter-granularity fusion by integrating node representations across granularities, yielding unified node embeddings enriched with hierarchical structural and semantic insights. Experimental results on five publicly available TAG datasets demonstrate that our proposed multi-granularity integration framework achieves better performance compared to various baselines, particularly achieving a 10.35% average improvement on PubMed dataset.