Language models encode linguistic proprieties and are used as input for more specific models. Using their word representations as-is for specialised and low-resource domains might be less efficient. Methods of adapting them exist, but these models often overlook global information about how words, terms, and concepts relate to each other in a corpus due to their strong reliance on attention. We consider that global information can influence the results of the downstream tasks, and combination with contextual information is performed using graph convolution networks or GCN built on vocabulary graphs. By outperforming baselines, we show that this architecture is profitable for domain-specific tasks.