Intelligent systems for teaching and learning assistance are missing from most applications, even though recent improvements in NLP allow us to imagine innovative solutions. The creation of a question-answer system based on academic sources would accelerate, improve and motivate student learning. In this context, we are interested in the generation of questions through neural approaches. With the recent annotation of a corpus of questions and answers in French, we have the resources to evaluate and develop such approaches. Nevertheless, several obstacles must be considered: the amount of qualitative data is not sufficient to train generative approaches; in the context of a stand-alone application we do not explicitly have the support for generation. In this study we propose different methods for extracting these supports comparing and analyzing the results on our corpus and those of the literature.