The paper proposes a two-stage model for sentence-level paraphrase generation, trained end-to-end. The first stage is content planning (specifically predicting a 'latent' bag of keywords). The second one is the surface realization stage (forming a sentence relying on the keywords). The model is interesting and novel. The evaluation is sufficiently convincing (the author response, I believe, addressed initial concerns of the reviewer 1). There is a consensus between the reviewers that the work should be accepted. I find the approach creative and think that it may have an impact on other text generation tasks, beyond paraphrasing. I encourage the authors to address the reviewers' comments and incorporate information presented in the author response (e.g., see R2's comment regarding the presentation and details about the baselines). [This meta-review was reviewed and revised by the Program Chairs]