Learning Patient Representations from Text