Optimizing Language Models for Dialogue