Using machine learning like this https://arxiv.org/pdf/1506.05869.pdf would be fun for this and avoid(?) canned responses...