We present two dialogue systems for lan- guage learning which both restrict the di- alog to a specific domain thereby pro- moting robustness and the learning of a given vocabulary. The systems vary in how much they constrain the learner's answer : one system places no other constrain on the learner than that provided by the re- stricted domain and the dialog context ; the other provides the learner with an exercise whose solution is the expected answer. The first system uses supervised learning for simulating a human tutor whilst the second one uses natural language gener- ation techniques to produce grammar ex- ercises which guide the learner toward the expected answer.