This paper presents an innovative solution for automated classification of urgent messages in online courses, using domain adaptation with DIET-RASA (Dual Intent and Entity Transformer) and XLNet (Extra-Long Transformer). Domain adaptation in this context is even more challenging because, in addition to the shift between distributions, the supports of the domain conditional distributions are also not coincident. We tackle this problem via generalization, mapping intent and lexicon from the source domain to the target domain using pre-trained models and embedding sequences of varying length. We provide in detail the main steps of the proposed pipeline and we also demonstrate the effectiveness of our approach in different scenarios, including cross-domain and low-resource environments. Our experiments show that our approach outperforms other state-of-the-art methods, being able to achieve an accuracy of up to 92% in some scenarios with a real-world dataset. Our approach can be used to offer personalized and agile support to students, regardless of the specificities of each domain.