Reviewers agree that this paper contains valuable ideas and evaluations. Authors response on adding sanity check and others (R1's comments) has increased our confidence. Please carefully review remaining comments (e.g., R2's comment on benefits of non-feature based interpretability methods) for your final version.