This paper focuses on question selection methods for conversational recommender systems. We consider a scenario, where given an initial user query, the recommender system may ask the user to provide additional features describing the searched products. The objective is to generate questions/features that a user would likely reply, and if replied, would effectively reduce the result size of the initial query. Classical entropy-based feature selection methods are effective in term of result size reduction, but they select questions uncorrelated with user needs and therefore unlikely to be replied. We propose two feature-selection methods that combine feature entropy with an appropriate measure of feature relevance. We evaluated these methods in a set of simulated interactions where a probabilistic model of user behavior is exploited. The results show that these methods outperform entropy-based feature selection.