Towards learning user-adaptive state models in a conversational recommender system