Over the past few years, graph structures have proven their effectiveness to represent interaction between textual information on many natural language tasks. The new GAT architecture has significantly improved the results in node classification tasks thanks to their attention mechanism based on the features of the vertices. In parallel, recent publications have shown that taking into account the topological aspect of the graph can attenuate some problems such as over-smoothing and the bottleneck phenomenon. In this paper we propose a way to improve GAT by considering the graph curvature in the attention mechanism. Experiments conducted on various datasets show that our method improves the original method GAT and outperforms recent GNNs specialized in node classification.