Aligning Language Models to User Opinions