Gradable adjectives, vagueness, and optimal language use: A speaker-oriented model