You could extract quoted speech from the data (especially in Q&A format) and treat that as "chat" that the model should learn from.