Predicting object and scene descriptions with an information-theoretic model of pragmatics