Paper about leveraging machine-generated rationales to facilitate generalization in conversations accepted to ACL.