Abstract: Asking questions during a lecture is a central part of the traditional classroom setting which benefits both students and instructors in many ways. However, no previous work has studied the task of automatically generating student questions based on explicit lecture context. We study the feasibility of automatically generating student questions given the lecture transcript windows where the questions were asked. First, we create a data set of student questions and their corresponding lecture transcript windows. Using this data set, we investigate variants of T5, a sequence-to-sequence generative language model, for a preliminary exploration of this task. Specifically, we compare the effects of training with continuous prefix tuning and pre-training with search engine queries. Question generation evaluation results on two MOOCs show that that pre-training on search engine queries tends to make the generation model more precise whereas continuous prefix tuning offers mixed results.