Large margin estimation of n-gram language models for speech recognition via linear programming