With the development of UAV cluster technology, the effective management of its network structure, especially the understanding of the nature of the edge between UAVs, has become a key research topic. In the UAV cluster network, under some complex conditions, the links and nodes in the network will carry more or less text information. In this paper, we propose a novel edge representation learning framework (Learn Edge Representation by combining GNN and LM models, LERGL), which combines the structure understanding capability of graph neural networks with the text processing capability of pre-trained language models. Specifically, in this framework, the graph neural network is responsible for encoding the graph structure data, while the language model is responsible for parsing the text attributes on the side. In each iteration, the edge features are extracted from the language model and then injected into the graph neural network for information exchange, and the feature expressions of the two are effectively fused through the attention mechanism. By conducting experiments on three real-world datasets, we verify the effectiveness of the LERGL framework for edge classification tasks.