{"id": "rc|human|rc0|review#1|14", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "I'm confused about the question() prompt. I thought these contained \"task-agnostic examples of how to transform statements to various questions\", but it seems like the format is not the same for all tasks? (Appendix H)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|17", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|19", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "ok I really found this paper hard to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|20", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "After reading the main paper the reader is left wondering as to what was actually proposed.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|21", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "The intro should really focus more on Figure 1 and highlight the overall algorithm better (generate the questions, generate the answers for these generated questions, aggregate the output).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|22", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "I feel that Figure 1+ Appendix F should form a good chunk of the main paper, showing what happens, why you believe that questions and answers that were generated by AMA are of good quality.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|23", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Additionally, apart from \"insights\" as to what makes good questions, authors don't provide enough details of what is being generated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|24", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "How many templates do you end up using? Is it task dependent? Were these templates selected for each dataset manually? Finally, the weak supervision should be also described in the main paper, as it is an integral part of the approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|26", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "after having (finally) understood the approach, I do find it witty.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|27", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Authors are generating a number of prompts that hopefully highlight different angles of the input, and combine the outputs of the model based on these prompts to come up with a final solution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|29", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "see my comments on clarity. As far as I understand authors open source their code (I have not seen where???) so this should help reproducibility, but overall just by reading the paper, I don't think I would have enough details to reproduce the approach.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|30", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Summary Of The Review:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|31", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Update: in light of additional comparison with prompt tuning, i am raising my score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|32", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "I appreciate all the additional experiments authors ran.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|33", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "I still encourage to improve the flow/the story/the presentation, but the method deserves to be seem", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|34", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Update: i am keeping my score of marginally above. I do think the narrative can be improved, but it is also an interesting peace of work and it is pretty witty to make the model generate the questions and answers to improve the prompting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|35", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Overall It is a very good idea of forcing the model to write the prompts that it is able to respond to.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|36", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "I do think the main paper should be reworked to include more examples/explanations of what is being proposed plus WS part.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|38", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Are all the models you tested are next token (perfix) trained, not span corruption trained correct?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|39", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "With respect to having less of improvement on closed book tasks - do you think it is because that the answers you show in the \"anwers\" part of the chain don't draw on closed book knowledge?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|40", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "I assume your \"answers\" templates are the same for various tasks so they don't include that closed book knowledge?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|42", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "You say in Table 1 that prompts can abstain from the predictions. What does it mean? Don't you get an output for each of your prompt chains?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|43", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Are questions prompts and answer prompts task specific? How did you create them How many chains do you create for each input during inference?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|44", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Building on this Isn't it the case that AMA performance will depend greatly on question() step? If the questions are not well done then when you create answer prompts, the model will not produce the result you are looking for", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#2|45", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "This seems to work well on simpler tasks (yes or no, question answering), but i am struggling to understand how the combining of the results will work for more generative tasks like summary tasks where variance of the predictions will be much higher", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|14", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "I found the story line a little confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|15", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "The aggregation based on weak supervision is touted as main accomplishment, but this actually only gives a small improvement in Table 1 (but this small push does give edge over GPT-3 sometimes).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|16", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "In many of the tasks the 6B model already beats the 175B without weak supervision based aggregation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|17", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Also this table is not really apples-to-apples comparison making it hard to interpret.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|18", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Should make it more clear what is the goal in this table --- is GPT-3 simply used more like a reference than a baseline?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|20", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "To explain more previous weakness, I was not clear when reading if it is enough to do prompt re-formatting or aggregation is also necessary.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|21", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "I think still I am a little not sure, also about whether re-formatting is necessary for aggregation to work well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|22", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "In general, it could help to tailor overall story as being primarily about prompt re-formating or aggregation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|23", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "It is well known that aggregating can be effective for prompting, and there are existing techniques like (self-consistency combined with chain of thought) that typically yield good improvements.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|24", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Perhaps worth using those as baselines, or mentioning why they weren't used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|25", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Maybe CoT is not effective for smaller models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|26", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|28", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "It would be helpful to have finetuning baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|29", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "For example Brown et al mention such baselines for BoolQ (91), CB (96.9), etc. in their Table 3.8. Others: DROP (89.1), NQ (44.5), WebQS (45.5).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|30", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Sometimes the improvement over GPT-3 is great but still far behind finetuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|32", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Self-consistency is a very crucial baseline / related work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|33", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Unlike the text says, it does not require additional training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|34", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "It is an insightful innovation that plurality vote is effective when you have different reasoning paths, which runs counterintuitive for how samples from a standard prompt behave.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|35", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Another related work with diverse prompting is Li et al. On the Advance of Making Language Models Better Reasoners. https://arxiv.org/abs/2206.02336", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|36", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "There is a lot of other works that describe strategies for exemplar selection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|37", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "One particular relevant example is Shi et al. (https://arxiv.org/pdf/2204.11454), which found aggregating 5 prompts (3 distinct exemplars each) is better than single prompt with 15 exemplars.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|38", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "There are other works that do sequential prompting like least-to-most and SeqZero.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|40", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "The analysis / diagnostic in Fig 5 is great, and builds confidence that sub-tasks are related to final performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|41", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "It was not always clear to me what is the source of weak supervision... Are you using the labels from the task, and if so, isn't this standard supervision? Are the prompts themselves considered the source of noise?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|42", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "It was not immediately clear to me what was meant by \"highly-correlated outputs\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|43", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "nit: Are you taking majority vote or plurality vote?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|44", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "nit: The framing of perfect vs. imperfect prompts is not the strongest.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#3|45", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Perhaps it is sufficient to reframe and say you are introducing an efficient and effective ensembling method?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|12", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "To apply AMA in real use cases, AMA needs 1000 training examples from the training set of the target task (specified in Appendix A.3.).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|13", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "This is an unrealistic setting because only a few training instances might be present depending on the task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|14", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "What if only a few training examples are present for a task (similar to few-shot setting)? Also, we need to train the probabilistic graphical model for each task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|15", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "This introduces additional latency for task adaptation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|16", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "For the AMA prompts on Appendix H, it seems that some tasks (DROP, SST, BoolQ, COPA) only do either prompt() or answer() stage, not both. How can we decide \"automatically\" when to do only prompt() or answer() or both depending on the target task?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|17", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "This paper is quite unclear on how the predictions are mapped to the output space. Are the predictions mapped to the output space using a predefined rule? (ex) no -> False for Figure 1)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|18", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Are there any results on a large-scale LM evaluation (175 OPT, GPT, BLOOM, etc,) comparing using AMA or not on the evaluation setting of Table 1 and 2?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|19", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "The structure of Section 3.2 and Section3.4 is divided into observation and solution (AMA).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|20", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "However, the observation experiments are not extensive enough. Is there a reason for selecting 3 tasks (CB, WSC, RTE)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|21", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Observing the findings on more tasks would make the findings more general and the motivation of AMA stronger.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|22", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "I think that StoryCloze should be included in Natural Language Understanding instead of NLI for Table 1 and 2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|23", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|24", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "The proposed method is novel in that it is effective and scalable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc0|review#4|25", "source": "rc", "split": "test", "paper_id": "rc0", "comment": "However, this paper contains some unclearness, especially on weak supervision and how the predictions are mapped to the output space. See Weakness part for details.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|22", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "1.It is claimed that these three tasks require understanding of both structure and content of the web-page.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|23", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "While it is easy to see that textual content plays a key role in each of the three tasks, the role played by the structure of the web-page is not clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|24", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "It can be argued that no significant HTML structure analysis or understanding is needed for these tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|25", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "For example, in Semantic Classification, what is most important for classifying HTML element 'input' into, say, 'username' is the value of its two attributes, 'type' and 'id'.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|26", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "As these attributes are in the close neighbourhood of 'input', parsing of HTML is not strictly necessary.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|27", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Therefore, it might a good idea to do some experiments that demonstrate unequivocally the need for HTML structure analysis or understanding in these tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|28", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "One such experiment could be to map all HTML tags in the web-page except the salient tags to the same token (say, ***) so that the input is now a sequence of salient tags, and ***.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|31", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The submission is well written and easy to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|32", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The three canonical tasks are described well and the adaptation of the various LLM for building models for these tasks are well explained.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|33", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The proposed solution is simple and appears to be effective for the tasks considered and the datasets chosen.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|35", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Experiments are designed well and should be easy to reproduce.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|36", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Datasets used in the experiments have been promised to be released.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|37", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The work should be interesting for practitioners.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|39", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "This work asks the question can off-the-shelf LLM trained on natural language text be used effectively for tasks that involve HTML pages.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|40", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "It proposes three tasks as canonical tasks in understanding HTML.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|41", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "It employs a variety of LLM to build models for the three tasks using a small amount of HTML data for fine tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|42", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "It shows that LLM does help these tasks significantly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#1|43", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "One key question not answered in this context is how much of HTML structure analysis and understanding is truly required for these questions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|9", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Models that are trained on task specific data to compare against is questionable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|10", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "I expect to see more task-specific models such as StructuralLM [1], MarkupLM [2] that respect the tree structure of HTML (or some other methods mentioned in the paper) to be evaluated against.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|11", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "However, the results seem to focus on LLMs only.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|12", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "I don't see the point of compare different variants of LLMs with or without pre-training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|13", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Basically, I need to see the following question answered: \"For specific web tasks, do I want to use a HTML specific model trained on the small dataset or just to fine-tune LLMs\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|14", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "[1] Li, C., Bi, B., Yan, M., Wang, W., Huang, S., Huang, F. and Si, L., 2021. StructuralLM: Structural Pre-training for Form Understanding. arXiv preprint arXiv:2105.11210.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|15", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "[1] Li, J., Xu, Y., Cui, L. and Wei, F., 2021. MarkupLM: Pre-training of text and markup language for visually-rich document understanding. arXiv preprint arXiv:2110.08518.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|17", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Clarity: The paper is easy to follow as the content of the paper is simple.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|18", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Quality: The experiments seem to be well-done.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|19", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Novelty: The novelty is limited in terms of methodology.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|20", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "It is mostly an empirical study for large LLMs on HTML texts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|21", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Reproducibility: Some details are missing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#2|22", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Either code or more detailed experimental setup should be provided in the supplementary material", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#3|10", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The novelty is primarily empirical, and the technical approach is essentially an application of known techniques (the standard pretrain-finetune paradigm of LLMs).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#3|11", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "I examined the supplementary material provided by the authors and only found the data files.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#3|12", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "It would be better if there were codes to prove its reproducibility.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#3|13", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#3|14", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Clarity: Most things are explained clearly either in the main paper or the appendix, except for the few points that are raised in the Weaknesses above.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#3|16", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Novelty: The novelty is primarily empirical, and the technical approach is essentially an application of known techniques (the standard pretrain-finetune paradigm of LLMs).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#3|17", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Reproducibility: The Supplementary Material only provides the data file.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#3|18", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "It is better to make the codes available for reproducing these results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#4|9", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "On the proposed Description Generation task, authors only spend two paragraphs presenting results, where WebD-T55-3B can obtain 90.8% accuracy (exact string match??).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#4|11", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "How to justify it from the existing semantic classification task, as in Figure 1(a) they look similar?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#4|14", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The modeling part lacks novelty, just basic pre-training + fine-tuning on LLMs, and seems there's no model design based on the features of HTML?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#4|15", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "On MiniWoB, it's not very surprising that using a larger model with pre-training, you can beat smaller models from scratch.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#4|16", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Instead, the fact that LLMs are worse than RL performances should be discussed more -- does it mean the task is bad, or RL is important (but it's hard to do RL for LLM)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#4|17", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "(Minor) In intro authors claim \"no processing of HTML\", but there is still some pre-processing?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#4|19", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Clarity: The paper is generally clear and easy to understand, some tables and figures in experiments are confusing or hard to read.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#4|20", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Novelty: The new task is somewhat interesting and novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#4|21", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The modeling part is not that novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc1|review#4|22", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Reproducibility: Authors claim to open-source the Description Generation dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#1|2", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "This is an interesting approach, however the experiments make me doubt the effectiveness of the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#1|3", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "Admittedly, the authors do point out that at the cost of fewer training parameters, the proposed approach attains the same performance as NLI encoders with MLP based or attention based classifiers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#1|4", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "However, the question becomes, how many fewer parameters are being learned to accept a performance that is in the same ball park but that does not exceed the SOTA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#2|1", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "I'd like to thank the authors for responding to my questions and for the additional experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#2|2", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "I think the new sentence embedding experiments make the paper quite a bit stronger - it would be interesting to scale them up to using SNLI + MNLI to see how much further they can go (right now they are still below Sentence-BERT which also was trained on SNLI in addition to MNLI).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#2|3", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "I think this paper is an interesting idea, and my only other main concern is the transfer results to other NLI datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#2|4", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "I think it would be a good idea to confirm that the difference between the two approaches is due to biases perhaps though an error analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#2|11", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "However, I do have some concerns about their claims that this helps in transfer learning to other NLI tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#2|12", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "Their main results show transfer performance comparing their approach and using an MLP, but it seems that overall, on all datasets, their approach transfers more poorly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#2|13", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "Two datasets that they do worse on they try to discount for either having terrible, below random performance (which is true) and the other for having the same biases as MNLI.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#2|14", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "However, if that was the case, I don't see why their model would perform worse since both theirs and the baseline would benefit from having these biases, but their model performs about 11 points less.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#2|15", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "So therefore, I don't find the transfer experiments convincing, though it is interesting how different the models do on some of these tasks - model performance is surprisingly task dependent", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#2|16", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "What I propose is for the authors to investigate if their sentence embeddings are in fact noticeably different than the ones trained in the more conventional matter.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#2|18", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "Im summary, I think this is an interesting experiment and it's nice to see that the MLP isn't doing a lot of heavy lifting (which also might be slightly counter to their hypothesis about the MLP containing a lot of entailment information).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#3|7", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "I especially appreciate the comparison to random matching heuristics, which already exhibits non-trivial performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#3|8", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "This is very reasonable because the neural network underlying random matching heuristics is still learnable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#3|12", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "1. One limitation of this paper is that the heuristic matching scorers are pretty ad hoc to the inference task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#3|14", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "Entailment, contradiction, and neutral scores are interesting, but hardly generalize to other sentence matching tasks (e.g., various IR applications).", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc2|review#3|17", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "As mentioned in the paper, NLI models do not transfer well to out-of-domain NLI samples, not to mention non-NLI tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#3|18", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "It would be interesting to see if the well-designed heuristic matching scores could ease the underlying model, so that it learns more generic sentence embeddings in general.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc2|review#3|20", "source": "rc", "split": "test", "paper_id": "rc2", "comment": "In references: Williams, Nagnia, Bowman: duplicate entry", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|7", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The paper should reduce the usage of color in the main text", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|8", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The paper should reduce the amount of branding and focus on the actual technique.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|9", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "I don't see how this has much to do with thinking fast and slow", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|10", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Figure 1 is confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|11", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The think step involves manipulating the inputs to get new sequences instead of just generating a set of relevant words?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|12", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Then, the second step is the actual summing over the sequences generated in the first step.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|13", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "There is a lack of technical details. In particular, it is not exactly clear how the sum step is performed. If this follows a probabilistic framework, please write down the steps formally.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|14", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "This paper first contrasts to chain of thought prompting, but it doesn't compare to that in experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|15", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Table 1 is confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|16", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "What's the difference between GPT-3 175B (davinci) and InstructGPT?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|19", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Without clear writing, it is hard to evaluate novelty.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#1|21", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The writing is not clear, making it hard to evaluate the contribution of this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#2|2", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The authors argued that this framework is good at performing many reasoning tasks with simple operations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#2|5", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "However, I am not sure if the improvement can be applied to other tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#2|6", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "For example, the task in Figure 1 requires evaluating every possible combinations of the choices.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#2|7", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "This is doable for 5 choices, but not so much for 100 choices or more (in SQuAD-type QA, the output could be any combinations of tokens) which is extremely expensive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#2|8", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Prompting LMs to score all possible candidates is also not accurate if the number of candidates is large, because the union of errors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#2|9", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The ThinkSum framework, however, is too general such that almost all existing methods fall into this framework, if you think scoring is the Think step, and argmax is the sum step.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#2|10", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Also, tasks experimented in this paper are classification tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#2|14", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Though, I may have missed some important parts of this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#2|16", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The scope of the proposed ThinkSum framework for LLMs is not well defined.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#2|17", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Please consider narrowing it down to a subset of tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#2|18", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The authors may also consider running some more challenging tasks to show the efficacy of the proposed framework.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#3|2", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The authors distinguish ThinkSum from chain-of-though like prompting methods by claiming that ThinkSum performs probabilistic inference instead of using LLMs to directly generate answers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#3|5", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Though the authors present ThinkSum as a general framework for LLM inference in Section 2, the definitions for \"Think\" and \"Sum\" are very vague and do not well relate to the example shown in Figure 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#3|6", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Then the rest of the paper is devoted to explain how ThinkSum can be used to tackle different tasks selected from BIG-bench in a case-by-case sense: each subsection covers one (class of) task, but it remains somewhat unclear how each example relates to the definition of ThinkSum presented in Section 2 and the connection between the individual tasks is unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#3|7", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Many new terminologies (in different colors) are introduced: some are completely unnecessary (e.g. product aggregation) and some do not have any clear explanations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#3|8", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The only baseline that the authors compare against to is the few-shot learning with GPT-3; the authors might want to include other prompt-based approaches as baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#3|10", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The paper is not well-written and hard to follow; even the definition of the ThinkSum paradigm is not clearly presented.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#3|12", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Despite the extensive demonstrations performed in this study, the authors need to present the ThinkSum paradigm in a more organized and clear way: even the core idea of ThinkSum is not clearly explained the examples throughout the paper do not help much explaining ThinkSum paradigm.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#3|13", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The experiments demonstrate the effectiveness of the approach in some sense but need comparisons against more baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#4|12", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Weaknesses One general issue may be how does the model scale up as the sets produced by the prompts to LLMs become large.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#4|13", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "In general, probabilistic inference may tend to become less reliable in this case so does that have a major influence on the SUM process.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#4|15", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The paper uses standard benchmark problems and is quite clear in terms of the description and setup and therefore to me it seems like the results should be reproducible,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#4|16", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "The novelty may not be as high since it mainly is an integration of methods but the impact of the approach seems significant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#4|19", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "Overall seems like an impactful paper with strong empirical results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc3|review#4|20", "source": "rc", "split": "test", "paper_id": "rc3", "comment": "They show that with their two-step approach, it is possible to achieve state-of-the-art results with much smaller models and thus should have an impact on scalability in future applications.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|6", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Collect a significant amount of training data for seq2seq tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|7", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Reconfirm empirically that using more data is always useful", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|9", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "The contribution of the paper is too thin in my opinion (i.e., showing that using more data is useful)", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|10", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "On Generaliability study: I don’t really understand this setup.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|11", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "The authors call it unseen tasks but then the model is fine-tuned using the annotated data of the task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|12", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "How is this different from taking any pretrained model and finetuning it on the same data?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|13", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "My understanding here is that the “unseen” part means that the tasks and data weren’t in the continued train step.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|14", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "But calling that generaliability seems to be an overstatement to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|15", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Is that what we all do with pretrained models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|16", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "There are no human evaluation results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|19", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Does the nature of informal text in the Personal Chat dataset have negative influences on formal text CNN/Daily Mail.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|20", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "I think there are many important questions about mixing random datasets in the context of language generation that have not been answered.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|21", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "The small improvement in automatic metrics do not fully justify the approach.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|24", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "How do you train (MVP+R) model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|25", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "the originality is marginal", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|27", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "In summary, the contribution of the paper is thin.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|28", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "The evaluation is not convincing (lack human evaluation).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#1|29", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "There are potential issues of concatenating multiple language generation datasets in the supervised pretraining step.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|3", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "However, the training setup is not designed in a way that allows us to compare unsupervised vs supervised pre-training since the backbone model is initialized with BART parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|4", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Thus the paper cannot answer the question which is posed, and instead shows that continued supervised pre-training using related tasks after starting with an unsupervised pre-trained model helps with downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|5", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "The model is initialized with BART and is further 'pre-trained' using supervised/multi-task learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|6", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "The main baseline is BART of the same architecture and some variants are compared.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|7", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Fine-tuning is done in two settings: full fine-tuning, and with prefix tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|12", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Improvements over BART baseline, which has the same architecture, in some tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|13", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "In some NLG cases achieves SoTA, although I've not verified every task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|17", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Clarity: It should be emphasized from the beginning (including abstract) that this is a study of continued multi-task supervised pre-training, and not a replacement for unsupervised pre-training.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|18", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Reproducibility: Missing details about how different tasks are mixed in supervised pre-training.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|19", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Novelty: continued pre-training with related tasks is not new, and is sometimes called intermediate fine-tuning. It is already known to improve performance in certain tasks.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|21", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "This paper should be re-written as a study of how to continue pre-training with multi-task supervised learning or properly study supervised pre-training by initializing the model from random parameters. At best, if re-written this way the paper shows how to achieve slightly better results on certain NLG tasks compared to BART with intermediate fine-tuning.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc4|review#2|22", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "The novelty is limited and the claims are incorrect.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|6", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "A collection of data mixtures over 77 NLG datasets, which are publicly released.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|7", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "A exploration effort on studying multi-task pretraining on the NLG area.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|8", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Code are released and could be easy to reproduce the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|10", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "The idea of multi-task pretraining is not novel, as it mostly follows Aribandi et al. (2022).", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|12", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "On Table 2, it seems the multi-task pretraining has minor gains over BART finetuning or Single; while on Table 9, the improvements are indeed significant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|13", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "The proposed variants of pretraining stages (e.g., MVP+S, MVP+R, MVP+M) don't seem effective as MVP (the results are on-par on most datasets from Table 2 & 3)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|15", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "On table 10, it seems there are a large improvement against ExT5 on the GEM benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|16", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Since the ExT5 paper doesn't release their model checkpoints, and the evaluation script of GEM benchmark can make a big difference on the results, it would be good to add BART finetune baseline on Table 10 too -- so that we can exclude the effect of evaluation difference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|18", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Overall, the paper makes a significant contribution on the collection of a large amount of NLG datasets and study the multi-task pretraining behaviors for NLG.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|19", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "Their code are released and the results can be reproduced.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|20", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "The idea presented in this paper is only marginally novel and mostly follow the prior work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc4|review#3|21", "source": "rc", "split": "test", "paper_id": "rc4", "comment": "The results of their proposed method also seems mixed in different datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|5", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "2. A relevant model architecture based method is Sodhani et al. 2018 (Towards Training Recurrent Neural Networks for Lifelong Learning) who use Net2Net to do zero-shot expansion of the model parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|6", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "3. Section 3.2 - you mention that any pseudo-example which does not have only one ANS token is discarded. Can you comment on how much discarding is needed to generate the required number of pseudo-samples?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|7", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "4. Why is it that every task was trained only for 9 epochs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|8", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "5. On page 5, you mention k=20. What is k? Where is this introduced?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|9", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "6. On page 5, you mention that MTL is used to determine whether forgetting is caused by a lack of model capacity. I am not sure if it is correct. Can you explain?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|10", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "7. Why not compare the approach with models like GEM? Keeping very few examples is ok. Even though you don’t beat GEM, it is good to see the comparison.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|11", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "8. Page 7: Is there any reason why you choose to go from large to small tasks? I feel like this is a favorable order. I would like to see how the model performs if you do the reverse order.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|12", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "9. Please remove the last line.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|13", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "10. I assume that the authors will release the code upon acceptance of the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|15", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "1. Page 2, 4th contribution: check the spelling for “pseudo-samples”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|16", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "2. Page 2, 5th last line: “After a completing a task” - fix it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|17", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "3. Table 1: I think the description is not correct. 1fEM is for wikiSQL, not WOZ. Also, it is better if you can describe these metrics in detail in the appendix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#1|20", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "I am happy with the authors' response and name change. I am increasing my score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#2|0", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "This paper studies the problem of lifelong language learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#2|5", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "1. Compare the proposed method with existing baselines. However, it seems that keep real data is missing from Table 3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#2|9", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "The weak points of this paper are as following:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#2|10", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "1. The assumption of modeling all tasks as QA might be strong;", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#2|11", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "2. The baseline from using real data is missing;", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#3|5", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "The motivation of the paper is clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#3|6", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "The comparison to old methods seems fair. The proposed method is clearly different from previous methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#3|7", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "One weakness of the paper is in the experimental results especially in section 5.4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#3|8", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "The statistical significance of the results in table 5 is missing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#3|9", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "As the authors have discovered, the performance is highly dependent on implementation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#3|10", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "In addition, the resulting performance might have a high variance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#3|11", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "From the data, it is hard to argue that LAMAL is better than MBPA++ by a significant margin.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc5|review#3|13", "source": "rc", "split": "test", "paper_id": "rc5", "comment": "Overall, the results are very interesting and worth a publication.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|20", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The main weakness comes from the experiment setup.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|21", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The main KILT results only have the original GENRE (using a different pre-trained model, BART) and CGR (even though the authors listed DPR and SEAL, the numbers are not in the same comparable settings).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|22", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The authors should provide a reproduced GENRE with the same pre-trained model and the same training recipe, as well as other baselines like BM25, a dense retrieval method (like DPR), and, if possible, SEAL.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|23", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "I have some other concerns in Table 2 and Table 3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|24", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "In Table 2, GENRE, DSI, and SEAL’s performance is so low on NQ-320K.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|25", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "I would appreciate it if the authors could clarify the setting (sorry if I missed it) or maybe retrain the GENRE on NQ-320K in the same setting as CGR (it shouldn’t be difficult since the code should be similar to CGR).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|27", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The proposed method is novel and intuitive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|28", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The authors could improve the clarity — e.g., the authors can make it clearer what text they put into the encoders for the contextual embeddings (in the paper they only vaguely mention that they are titles).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|29", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "Figure 1 is a bit confusing to me (especially Cape(1), Cape(2)).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|30", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The experiment setup can be significantly improved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|31", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "From the current experiments, it is hard to draw a clear conclusion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|33", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The proposed method -- using contextual embeddings instead of static token embeddings in decoder for autoregressive retrieval -- is novel, intuitive, and inspiring.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|34", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "However, the experiment has significant flaws (for the most direct and important baseline, the authors took numbers from the original implementation that uses even a different pre-trained model, which leads to unfair comparison).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|35", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "I am leaning toward rejection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#1|36", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "However, if the authors can update the results with the reproduced GENRE baseline, I am willing to raise my score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#2|9", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The proposed architecture is simple and elegant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#2|12", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The proposed model follows the framework of generative retrieval models but applies contextualized token embeddings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#2|13", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "Such contextualized embedding methods have been applied in KNN-LMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#2|18", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "This paper proposes a new generative retrieval model and the proposed model shows promising results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#3|8", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "Evaluation on datasets that are characterized by title, document ID or datasets for language tasks, and not on large-scale document retrieval collections where the specific content of the documents is important to satisfy the query, and where expressiveness and discriminative power of the document representations play an important role.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#3|12", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The proposed method makes sense but the contribution of adding a contextualized embedding space might be considered as light.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#3|15", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "It is not clear how the model guarantees a correct generation of key descriptors of a document, if the generative retrieval model and the contextual language model were not trained with it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#3|16", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "This relates to the core claim of the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#3|19", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "What if these representations were changed by using alternative technologies to build them?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#3|21", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "Could the proposed generative model be combined with older generative retrieval models that operate on bag-of-words or n-gram representations (e.g., Robertson-Sparck-Jones model, language retrieval model) in order to deal with specific content that should be generated?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#3|22", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "Overall references to older generative retrieval models could be added in related research.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#3|23", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The paper could be strengthened by analyzing the limitations of the work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#4|7", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "Unclear writing: Figure 1 and a verbal description in Section 3 is all we have to decipher what the paper does.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#4|8", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "I can guess what it does, but it's never fully clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#4|9", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The setup is questionable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#4|10", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "GENRE uses BART-large whereas this paper uses T5, so the focus on comparing directly with GENRE seems a bit misguided.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#4|11", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "Limited experiments (not training on all of KILT).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#4|12", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#4|13", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The paper is unclear on the model definition, making it hard to judge the quality.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc6|review#4|16", "source": "rc", "split": "test", "paper_id": "rc6", "comment": "The paper proposes a reasonable way to use external knowledge in generative retrievers, but the quality of innovation, experiments, and writing seems to fall below the standard of ICLR.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|3", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The paper introduces a dataset, AmbiBench, designed to measure how well models use the instrutions/few-shot examples to infer underlying task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|10", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "for uninformative instructions, both models & humans get 50% accuracy, but humans also get 100% on negation while models do no", "reviewer_type": "human", "label": "error", "error_type": "Inaccurate Summary", "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|11", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "for informative instructions, humans get perfect performance across all tasks but 1, with RLHF models (text-davinci-002) close behind.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|12", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Non-RLHF models do not do nearly as well: davinci does much worse than davinci instruct.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|13", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "They further find that ambiguously-specified examples provide a much bigger boost to generalization, when compared to finetuning on unambiguous examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|14", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "This provides evidence that finetuning works more robustly than few-shot prompting, and that explicitly training to handle ambiguity helps generalization performance.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|15", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Best model is not able to consistently verbalize the task it is trying to accomplish", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|16", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Main takeaways from paper: Scaled RLHF models perform best at disambiguating tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|17", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Finetuning helps over just few-shot prompting", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|27", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The tasks may be too simple / short / toy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|29", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Downside of using closed models (although no open models exist for this yet): we do not actually know the training procedure for the largest models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|32", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "page 2: what makes finetuning on in-context examples metalearning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|33", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "page 15: \"could could verbalize\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|35", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Potential improvement: run additional experiments on the FLAN-T5 models to see how instruction tuning helps smaller models (paper released after submission, so not required but would be useful additional evidence!)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|37", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "I am reading the updated version of the paper, which I find to be clearly written and thorough.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|40", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "It would be great to see more experiments with open RLHF models, but there unfortunately are not any available yet.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|41", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The synthetic dataset is novel, if simple.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|42", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "This is the first paper to do a clear and self-contained benchmark on task ambiguity (afaict).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#1|43", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "It evaluates a few known methods for getting the models to accomplish a task (few-shot prompting vs. finetuning) and confirms existing evidence that finetuning performs better.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|19", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "It is quite hard to draw many conclusions about the best performing systems because their training strategies are not public, and it is not clear what types of task ambiguity have been seen during training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|20", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The six synthetic tasks are quite simplistic and limited in scope.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|21", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "It is not clear how these findings would extend to more complex task types.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|23", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The choices of line color and hatching in Figure 3 are not very discriminative, and I struggled to untangle which line represents which model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|24", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "I suggest that more disctriminative line types are used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|27", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "This paper is well written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|28", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The experiments are well justified and explained.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|30", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The experiments are well designed but limited in scope.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|31", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The paper would be more valuable if it could could better describe how its findings may extend to broader types of task ambiguity, likely to be seen in downstream tasks of interest.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|33", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "This is the first paper that I've seen that properly investigates task ambiguity of this type in large language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|35", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The novel work in this paper should be reproducable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#2|36", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "However, the paper relies heavily on opaque models, and it is a little hard to draw meaningful conclusions about the best performing systems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#3|2", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Recent research has demonstrated the benefits of using natural language prompts to describe the task when fine-tuning a language model, but such prompts can be ambiguous in some cases.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc7|review#3|8", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "This paper only studies the ambiguity of a few toy tasks, which only require simple lexical understanding and the texts are synthetic and very short.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#3|9", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "It would be more interesting to see what the ambiguity is in real-world NLP tasks, and how they can be handled.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#3|11", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The \"uninformative instructions\" are very unnatural as they just replace the salient feature in a normal instruction with placeholder.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#3|12", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Nevertheless, the instructions still contain some description of the task, and it is easy to imagine that both human and model would make guesses based on the available, incomplete information.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#3|13", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The observation that the best language model achieve similar performance to human on such ambiguous task is not surprising, even though the setting is synthetic and simple.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc7|review#3|14", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "In my opinion, I don’t see many new insights from this paper that would help the community to better resolve task ambiguity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#3|16", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "This paper is clear, although the writing can be more fluent and concise.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#3|17", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The novelty is limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#3|19", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The reproducibility is unclear, as the authors do not mention whether the benchmark will be released.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|16", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Although the second paragraph of Section 4.3.1 (why text-davinci-002 outperforms humans in Figure 3) seems a bit hand wavy to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|19", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Please see my comments below.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|21", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Let me list my concerns and questions here, please answer and help me better understand the paper :)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|22", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "In Section 2.2, the author mentioned a few prior works discussing LLMs' unfavourable behaviours such as \"recency bias\" as proposed in [1].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|23", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Indeed, for most in-context learning or few-shot prompting work, such biases can have non-negligible effects on model performance (e.g., models tend to copy labels from the most recent examples as its prediction, due to such effects, model performance on a data point can sometimes vary from random to near perfect).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|24", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "The authors state that", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|25", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "In this work, we attempt to control for such factors by constructing a benchmark where multiple tasks are consistent with a single instruction or example, requiring the model to leverage multiple signals to disambiguate the intended task without relying on learned priors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|26", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Please help me understand this, I fail to understand how AmbiBench avoids or mitigates such effects?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|27", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "To my understanding, as long as few-shot prompting are provided as a sequence, such effects can exist.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|28", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "What are some reasons in Figure 2 and Figure 3 human annotator scores have a much bigger variance?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|29", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "I like how the authors include many LLM variants for comparison.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|30", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "However, I'm a bit concerned on, e.g., in Section 4.2.1, what's a good way to decouple models' 1) inability of disambiguate the task description from 2) inability of generate meaningful output via prompting (maybe a model can do disambiguation, but it's not big enough to perform well on prompting?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|31", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Given some recent discussions such as [2], is few-shot prompting the best way to investigate certain abilities of smaller models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|32", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Compared to, e.g., using a linear/non-linear probe to see if the LM representations contains certain properties we want.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|33", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "If the LM has trouble to even generate meaningful outputs (as reported by the authors, sometimes models generates tokens other than X and Y), it's a bit hard to tell how capable the model is on task disambiguation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|34", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "I'd like to hear more from the authors about the fine-tuning experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|35", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "I assume that fine-tuning could potentially make the disambiguation task much easier, even with only a small set of training data (68 per task).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|36", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "I fail to gain too much information from Section 4.4 more than fine-tuning on in-context data can outperform a general inference-only model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|37", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "As the authors mention, AmbiBench is a minimal complexity testbed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|38", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "I understand and agree that this is to make the benchmark more controllable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|39", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "However, for the same reason, I am less convinced about the generality of findings reported here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|40", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "What are some other types of task ambiguities in NLP tasks, maybe more subtle than simply occluding a category name?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|41", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "Maybe brought by natural language instructions rather than templatic language?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|42", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "What about in tasks other than sentence classification?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|43", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "It would be nice to have such discussion, which can give readers a better idea how the proposed method/benchmark can shed light on understanding task disambiguation in general.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|44", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "(less important:) Figure 1 is a bit misleading in some sense, I was honestly expecting to see how a LLM can explicitly guess the task description --- this might not be a crazy idea, given a sequence of examples (as many as 20 here), is there a way to model/measure such thing?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|45", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "(less important:) The two prompt templates introduced in Section 3.3 are quite arbitrary, especially the Q/A format barely makes sense.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|47", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "[1] Calibrate Before Use: Improving Few-Shot Performance of Language Models. Zhao et al., 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|48", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "[2] Emergent Abilities of Large Language Models. Wei et al., 2022.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|50", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "In general, I think the current version of the paper is around (maybe slightly above) the borderline.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc7|review#4|51", "source": "rc", "split": "test", "paper_id": "rc7", "comment": "I look forward to reading the authors responses to my questions so I can better understand the work and thus have a more precise evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|2", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "It confirms and extends some of the existing work on scaling laws, and as such it is a valuable contribution to the field.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|3", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "In particular, new questions addressed include: does the encoder-decoder architecture of NMT share the same scaling law as decoder-only LMs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|4", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "How does training data (backtranslated or not) affect scaling?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|7", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "I really enjoyed reading this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|8", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The research questions are clearly stated, and the results are fascinating to go through.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|10", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The number of experiments is impressive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|11", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "This adds weight to the papers conclusions, which are certainly thought-provoking.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|12", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "For example, the conclusion that scaling law should be bivariate, treating the number of encoder and decoder layers separately, makes me think we should not train models of symmetric number of layers anymore.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|14", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The footnote on comparison to Gordon et. al. in EMNLP2021 is too terse and should be expanded.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|15", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "That paper studied scaling laws for NMT as well, but in a different parameter regime.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|16", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "Do your conclusions affirm, reject, or extend theirs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|19", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The research questions are clear and the results are convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#1|20", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "I think it will be cited often in the future.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#2|8", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The proposed function of model capacity for encoder-decoder NMT model performance is useful for parameter allocation of NMT models training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#2|9", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "This paper provides experiments to explore how the composition of training/test data affects the test/training loss for scaling NMT models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#2|10", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "It's an important question to connect the other two questions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#2|11", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "It is important to investigate how generation quality evolves with the improvements in cross-entropy resulting from model scaling.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#2|12", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The problem is of great value for both NMT practice and research.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#2|14", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The authors have proposed three research questions, while the relationship among them is not clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#2|15", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The scaling property of the composition bias of the train/test sets has been less studied in previous work, and is most interesting to the reviewer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#2|16", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "This part can be more convincing if the authors can provide more experiments and analyses.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#2|19", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "However, the relationship among these research questions and findings is not clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#3|8", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "the research questions answered in this paper are interesting and help to better understand the mechanics of NMT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#3|9", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "As an expert in MT, I really found this paper interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#3|10", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "findings are very insightful", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#3|11", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "this paper will be impactful: it will guide NMT (or any other kind of encoder-decoder model such a BART) practictionners to build more efficient and effective models", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#3|12", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "another proof that target-original test sets are not suitable for NMT evaluation (Figure 8 should go on Twitter when you publish your paper!)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#3|13", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "evaluation is not only performed with BLEU (BLEURT is also used)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#3|15", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "almost the entire evaluation is not reproducible: undisclosed datasets, use of a non-standard BLEU implementation, pre-processing of datasets not detailed, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#3|16", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "the paper misses a related work section (more details than in the introduction).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#3|17", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The reader may not know related work on language models for instance", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#3|19", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "All the \"strengths\" pointed above justifies my recommendation for the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#4|8", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "They conduct extensive experiments on four language pairs and five different test domains.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#4|9", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "It is interesting that contrary to previous findings, this paper suggests that it is better to scale the decoder than the encoder, providing guidance for future research on deep NMT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#4|10", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The analysis on the translationese provides insights on the correlations between cross-entropy loss and generation quality, and its effect on the scaling laws.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#4|12", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The scaling laws have several test set-specific parameters, which makes it questionable if the laws have practical values.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#4|14", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "I appreciate that the authors make efforts to measure the outputs in both BLEU and BLEURT, but it would be better to include human evaluation results that can better reflect the generation quality of models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#4|15", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "The paper mainly focuses on the relationship between cross-entropy loss and the number of Transformer layers in large-scale settings, while there are a lot of other factors that can be considered.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc8|review#4|16", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "For example, it is unclear whether the scaling laws will hold if the architecture varies or if the hidden size changes.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc8|review#4|18", "source": "rc", "split": "test", "paper_id": "rc8", "comment": "While the paper has some minor issues, it conducts extensive experiments in large-scale settings and the results can provide insights for future research, thus I’m leaning towards an acceptance of the paper", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#1|8", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "While used components are effective, each of them is from other works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#1|10", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "There is no performance comparison with word-level pre-training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#1|11", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "Since the entity recognition task is word-oriented as mentioned in the paper, it would be better to perform pre-training on word-level rather than sentence-level.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#1|12", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "UniDoc does not improve the performance of document classification tasks over previous methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#1|13", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "Although the authors argue that it attributes to the difference of OCR engines, they should explain the magnitude of this difference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#1|14", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "They say the difference between EasyOCR and Tesseract, but should report performance metric and qualitative analysis on the RVL-CDIP dataset to compare real Microsoft OCR and EasyOCR.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|6", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "I feel that the methods/model architectures are largely putting together existing pieces of work from vision and language fields.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|7", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "In Sec 3.1, the proposed feature extraction, quantization, and gated cross-attention modules are all rather straightforward variations of existing work (i.e., the essential mechanisms are not new).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|9", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "In Sec 3.2, the three pre-training tasks also resemble well-known techniques, like masked language modeling in NLP and contrastive learning in vision.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|10", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "The alignment of vision and language is also not particularly surprising/insightful and provides little performance gain according to Table 2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|14", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "One thing that appears weird to me is the frozen sentence encoders initialized with the weights trained on NLI and STS datasets--if they have been fine-tuned on those tasks, they may capture too corpus-specific features to generalize to other domains (like the new pre-training documents).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|15", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "I don't think the computational constraints are valid excuses if the paper is proposing a new pre-training framework, which should be generic and not task-specific.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|21", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "However, I am not sure how significant the results are to the NeurIPS community, because none of the baselines compared, as far as I am aware of, are from ML conferences.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|22", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "I would like to see comparisons with stronger baselines (e.g., there are many better PLMs than BERT from the language field), and ideally how the framework performs on pure language applications (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|23", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "the GLUE & SQuAD benchmarks) since a lot of applications may not have images and texts simultaneously.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|24", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "If the UniDoc model cannot be effectively used for pure language applications, I am not convinced that this can be called a \"unified\" framework.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|32", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "There are concerns about the generalization ability of the model since it uses dataset-specific text encoders.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|33", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "The significance of the empirical contribution made is not clear to me (the baselines are not from the ML community and are not strong enough; it's not clear whether the model works for pure language applications as a \"unified\" framework)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#2|35", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "I think the authors have made reasonable responses regarding the limitations and potential negative social impact of their work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#3|6", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "Originality: Apart from some novelties of the pretraining objectives, the method reuses some of the existing architectures in a hierarchical way.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#3|12", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "To the best of my knowledge, the authors have identified and discussed the potential social impact of this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#4|8", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "The weaknesses of the paper: (1) The effects of some factors are not fully discussed (listed in the question part); (2) The result on RVL-CDIP is not SOTA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#4|9", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "The paper said that the potential reason could be using a weaker OCR, but not prove it directly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#4|12", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "Questions: (1) What is the effect of the gated mechanism used in cross-attention?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#4|13", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "What the results look like if the gate is removed?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#4|14", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "(2) Can we replace the gated cross attention layer with a normal Transformer (a single-stream model)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#4|15", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "(3) Is it possible to include Masked Language Models as well?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#4|16", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "If so, will the results become better with this typical pre-training task?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#4|17", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "(4) Are discrete representations indispensable?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#4|18", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "For example in Figure 1, can we use the RoI features from other documents in the same batch as negative examples for contrastive learning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc9|review#4|20", "source": "rc", "split": "test", "paper_id": "rc9", "comment": "As the paper said: \"...the document images used for pretraining may contain sensitive information and therefore the models trained on such data may inappropriately leak some private information...\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#1|12", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Originality: interesting work that attempts to define what true few-shot learning with language models should actually look like and proposes methods that fit this definition and benchmarks them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#1|13", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "As far as I'm aware, this type of work has not been pursued before.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#1|14", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Quality: the problem and methods are well-defined.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#1|15", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The experimental results seem to back up the hypothesis that true few-shot learning without using the information from held-out examples is still difficult for pre-trained language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#1|16", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Clarity: the paper is well-written.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#1|17", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The problem is well-motivated and the methods are clearly defined.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#1|18", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The experimental section is also very clear and the experiments measure interesting aspects of the discussed problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#1|19", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Significance: this paper could be impactful as it questions some of the assumptions of previous work showing few-shot learning in pre-trained language models and proposes best practices for future work for benchmarking few-shot learning in this setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#1|20", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Limitations And Societal Impact:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|7", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Overall, I think the paper highlights an important consideration for studying few-shot settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|8", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The fact that existing few-shot methods rely on large validation sets is unrealistic as authors point out, but often overlooked.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|9", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The experiments show that LMs indeed perform better with the held-out validation set than in the true few-shot learning setting across multiple models, two datasets, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|10", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The experiments nicely demonstrate that it’s important to consider the proposed setting to appropriately measure LM performance when there are only a few labeled examples available.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|11", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "One limitation of the experiment is the over-emphasis on the comparison with random selection of prompts and hyperparameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|12", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "This would be appropriate for the purposes of measuring the effectiveness of cross validation (or maximum description length), but not so much for the purposes of measuring the effect of not having a held-out validation set, which I thought was the main point of the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|13", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The authors include comparisons with tuning on a held-out validation set for most, but not all experiments. I think it’d be helpful to include those comparisons for all experiments and lean more on these comparisons over comparisons with random baselines. In particular, could you provide the results for Figure 3?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|14", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "My main comment is about the high-level claim that the existing literature “greatly overestimates” the few-shot capabilities of LMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|15", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "While it’s subjective what ‘greatly overestimates’ means, I think the claim perhaps comes off as a bit too strong given the empirical results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|16", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "In Figure 1, even though there are noticeable performance gaps between the true few-shot settings and the setting with a large held-out validation set, the models seem to perform reasonably well in the few-shot settings, regardless of prompt selection procedure.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|17", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The presence of a held-out validation set doesn’t make or break the model performance, and model size seems to affect the performance more than the prompt selection procedure.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|18", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "My suggestion would be to update the abstract and introduction to tone things down.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|20", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "One limitation of the experiments is that the paper has one dataset for each of prompt selection (with a fixed pool of prompts) and hyperparameter tuning.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|21", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "I think showing consistent results across datasets would make the results stronger.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|22", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The paper is missing a discussion of related work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|23", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "In particular, are there works in the few-shot literature that explores the “true few-shot” setting?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|24", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "I think the point raised in the paper are still good to highlight regardless, but it may be good to mention that some works already consider the “true few-shot” setting if that’s the case.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#2|25", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Limitations And Societal Impact:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|8", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The main finding of the paper is that both of these validation methods are only marginally better than random selection for both prompt tuning and model selection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|9", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "This is also significantly worse than selection based on the full validation data (with thousands of examples), which was the strategy used in “popular” few-shot learning papers like GPT-3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|10", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The results are indeed what one would expect but I think it is really important to make this point in order to avoid future research continuing to make such poor validation practices for few-shot learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|11", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The results in the paper cover many aspects of validation, albeit restricted to only these two simple choices, and yield analysis that can potentially inform experimentation for few-shot learning in the future.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|12", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "I believe the paper will have a stronger impact if they experimented with a few additional methods (see below) and also use some additional tasks for their prompt tuning experiments so that the analysis is more generally applicable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|14", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The choice of LAMA for prompt tuning: While I understand that the authors need a large collection of prompts to be able to reliably experiment with prompt selection, I am not sure how much inference we can draw from the experiments on this task for other few-shot tasks like QA, NLI, classification.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|15", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "As I understand it, the LAMA task is more a test of how much factual knowledge a pre-trained model retains as opposed to being a test for its few-shot abilities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|16", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "For instance, I don’t think there is a significant difference between zero-shot vs few-shot on LAMA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|17", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "It will be useful to see results for prompt tuning on other classification or QA tasks, for example from SuperGLUE, where the tasks require quite a different reasoning than LAMA.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|18", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Is leave-one-out the most appropriate cross-validation strategy here in few-shot experiments?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|20", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Another relevant approach to model selection is to use some additional, held-out, tasks to find model hyper-parameters and reuse them on all other tasks without any additional tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|21", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "I think this setting is termed as multi-distribution few-shot learning in section 2 and is also a reasonable strategy to select hyper-parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|22", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "This is a common approach in meta-learning, for instance where one has held out validation classes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|23", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "This was also previously discussed by [1] that proposed to use validation languages for evaluation of low-resource language transfer and [2] used a similar approach for few-shot classification using validation tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|24", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "While this may not always be possible for prompt-based methods, it is suitable for other methods like those in section 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|25", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "It will be useful to include this setting in the experiments in section 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|26", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Discussion of related works [1, 2] and their validation approach should be included.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|27", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "[1] Towards Realistic Practices In Low-Resource Natural Language Processing: The Development Set", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|28", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "[2] Self-Supervised Meta-Learning for Few-Shot Natural Language Classification Tasks", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|29", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Overall I really like the problem that the paper is bringing forward and I think this will be very useful in the future.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|30", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "My current score is mainly due to some of the above concerns which I hope the rebuttal will resolve.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|31", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "I understand doing all experiments can be difficult in the rebuttal timeframe but please respond to the best of your ability and I will be happy to revise my score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|33", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The authors' response resolved many of my concerns and I have updated the score accordingly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#3|34", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "I think it will be good to bring forth some of the additional experiments on SuperGlue in Appendix to the main paper and also include the results on multi-distribution validation provided in the rebuttal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|6", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Overall, this work shows that the few-shot ability of pretrained LMs is overestimated for true few shot learning, where model selection is fundamentally difficult and needs to be addressed in future work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|7", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Moreover, this work provides suggestions for fair comparison on few-shot learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|9", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "This paper criticizes recent advances made in prompt-based few-shot learning with pretrained LMs, arguing that the few-shot ability of pretrained LMs is overestimated for true few shot learning problems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|10", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "The paper proves this by experimental results on various tasks showing the few-shot learning performance drops significantly when no validation set is available.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|16", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "However, the contribution of this work is limited by its scope. The results for true few-shot learning are interesting but are not widely applicable. Because most previous work, including those being criticized, focus on the multi-distribution or tuned few-shot learning settings and make assumptions like a validation set is available or optimal hyperparameters for other tasks could be re-used for the new few-shot tasks. Indeed, future works should be explicit about their model selection method to make fair comparison. But this has been a well-known issue for few-shot learning under additional assumptions. For example, researchers in computer vision have been working on datasets with standard splits for fair comparison [1]. The true few-shot learning setup might not be entirely impactful.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|17", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "[1] Triantafillou, Eleni, et al. \"Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples.\" International Conference on Learning Representations. 2019.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|19", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "(1) The true few-shot learning setting is good for showing the few-shot learning ability of pretrained LM without any additional information except the few training examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|20", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "This is true since the few-shot performance of LMs is much better than random guess, although being suboptimal compared to SOTA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|21", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "As discussed above, much of the work criticized by the paper is not designed for true few-shot learning, but for more practical settings like multi-dimension or tuned few-shot learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|22", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "We do need a better comparison framework for multi-distribution and tuned few-shot learning, e.g., specifying which external tasks can be used for training/validation and how many validation data can be used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|23", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "For example, Tam et al. [1] assumes full access to development set while Schick and Schütze [2] didn’t make such assumption and thus the comparison between them is unfair.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|24", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "But this is a well-known issue for few shot learning or machine learning in general.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|25", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "(2) The paper also investigates whether the results for LAMA also hold for other classification tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|26", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "In the LAMA benchmark, the LMs are asked to predict the answer out of 20k candidates given a sentence.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|27", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "But for regular classification tasks, the number of candidates is often much smaller, i.e., only 2 for binary review classification (negative/positive).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|28", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "Also, when talking about 5-shot classification tasks, people often mean 5 examples per class.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|29", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "But in the Appendix B, only 5 examples in total are given.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|30", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "It is thus unsurprising that CV and MDL will fail in this case.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|32", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "[1] Derek Tam, Rakesh R Menon, Mohit Bansal, Shashank Srivastava, and Colin Raffel. Improving and simplifying pattern exploiting training. arxiv preprint arXiv:2103.11955, 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc10|review#4|33", "source": "rc", "split": "test", "paper_id": "rc10", "comment": "[2] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few-shot text classification and natural language inference. Computing Research Repository, arXiv:2001.07676, 2020.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|3", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "The paper also proposes an rASR metric that employs humans to evaluate how similar an adversarial example is to the vanilla sample.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|9", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "I am worried about the novelty.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|10", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "It seems that the proposed method just takes an average of embeddings from many different encoders.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|11", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "The motivation is not clear. The paper found that previous encoders fail to deal well with antonyms, but it seems that the proposed method did nothing special about antonyms either.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|12", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Some related word-level attacks should be discussed, e.g., [1], [2], [3].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|13", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Some designs need more justifications. E.g., what is the criterion when choosing the 7 datasets and why the threshold is just set as 2.5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|14", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "[1] Moustafa Alzantot, Yash Sharma, Ahmed Elgohary, Bo-Jhang Ho, Mani Srivastava, and Kai-Wei Chang. Generating natural language adversarial examples. In EMNLP, 2018.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|15", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "[2] Shuhuai Ren, Yihe Deng, Kun He, and Wanxiang Che. Generating natural language adversarial examples through probability weighted word saliency. In ACL, 2019.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|16", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "[3] Xinshuai Dong, Anh Tuan Luu, Rongrong Ji, and Hong Liu. Towards robustness against natural language word substitutions. In ICLR, 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|19", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "The novelty may not be enough.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|22", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Given the strength and weaknesses, I tend to reject.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|24", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Thank the authors for the effort in answering my questions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#1|25", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "After reading all the review comments and responses, I decided to keep my score unchanged.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|3", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Compared with previous methods such as USE, the proposed method SPE is simple and efficient.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|12", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "A. My first concern is though the experiments are properly designed, the comparison might not be fair.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|13", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "That is, as explored thoroughly in prior works such as Textfooler and BERT-Attack, the transferability of the adversarial samples is somewhat weak, that is, the adversarial samples are not universal to different classifiers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|14", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "While the proposed SPE is in fact trained with datasets including NLI, cola, rte, sst-2, yelp, etc., these datasets can be transferable, therefore, the SPE score can be viewed as another classifier in a way.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|15", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "I am concerned that the SPE similarity is in fact a reflection of classification, not as illustrated as ‘semantic-preserving’.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|16", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "If it is the case, then the advantage of low cost in the proposed scorer is less convincing since we can always train some simple classification models as a scorer to measure the adversarial sample quality.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|17", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "B. semantic preserving concept: As illustrated in Table 1, the authors intend to illustrate the concept of sentence semantic similarity when the sentence is perturbed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|18", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "These cases, which mostly are synonyms to antonyms change, are less observed in the similarity-based word-embedding used in the Textfooler but actually constantly happen in BERT-Attack generated substitutes, which cannot be properly recognized by sentence-level semantic scorers such as USE or BERT-Score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|19", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Therefore, the word-level and sentence-level semantic preserving performances can be different in measuring the quality of adversarial samples yet the paper seems to mix these two concepts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|20", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "C. The novelty of the proposed method: It is true that using USE or BERT-Score in evaluating the quality of the generated adversarial samples is costly and cannot obtain promising results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|22", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "How can the proposed SPE be used in more general scenarios as a semantic preserving sentence similarity checker is not discussed in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|30", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "The main concern is fairness in the evaluation process, the authors should provide some analysis of the transferability of attack methods and the connection to their methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|31", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "I have read the authors' responses and decided to keep my score unchanged.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#2|32", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "The technical quality of the paper would benefit from a major revision.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|9", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Several technical decisions are not justified.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|10", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Why use average of embeddings?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|11", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Why use 7 datasets?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|12", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Does some datasets more helpful then other datasets?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|13", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "For example, what if we remove the semantic classifiers from SPE when attacking Yelp dataset?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|14", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Why use 2.5 as the threshold for rASR?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|15", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "The sentences are just ``slightly similar'' if it is annotated as 3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|16", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "There are several issues with the experiments", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|17", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Why not verify the method on standard sentence similarity benchmarks such as SemEval?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|18", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "SPE can be considered as a sentence similarity metric.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|19", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "However, this paper neither verifies its performance on standard sentence similarity benchmarks nor justifies that this metric is only suitable for adversarial attack situations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|20", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Why use distilBERT on Rotten tomato dataset, while using RoBERTa on other datasets?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|21", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "There are many other adversarial attack methods, such as BERT-Attack, CLARE, SememePSO, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|22", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Does SPE also improves the quality for these methods?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|23", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "More insights is needed to demonstrate the improvement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|24", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "For example, certain type of changes are eliminated by applying SPE.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|26", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "I'm not sure if Eq. (1) is correct.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|27", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "Why does the complexity depend on the vocabulary size rather then the length of the sentence?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|28", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "What do you mean \"do not preserve the semantics and even the meaning of the text\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|29", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "What do you mean “eliminate the problem of vocabulary words”?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc11|review#3|30", "source": "rc", "split": "test", "paper_id": "rc11", "comment": "More philosophical question: Is SPE trying to improve the semantic similarity or to improve prediction consistency?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|20", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "The method is intuitive and the results are strong.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|21", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "It is not only a way to improve the efficiency of using in-context learning and instruction prompting, but the fact that the knowledge can be “internalized” is very interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|23", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "The experiment setup can be significantly improved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|24", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "First, the paper can be improved on how they demonstrate the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|25", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "There are very few tables/figures on the results in the main paper and most numbers scatter in the text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|26", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Second, the clarity of the experiment setup (what datasets they use for exp#1, how many pseudo input examples they generate, etc. sorry if I missed it) can be improved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|27", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "For Exp#2, it is very limited to only test context distillation on the SQL dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|28", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "It would be better to show other datasets too, like NLU, QA, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|30", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Even though the concept of “context distillation” is not new, the authors did a good job of conducting comprehensive research on three different perspectives.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|31", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "However, the presentation/clarity of the paper can be significantly improved, as suggested in the “weakness” section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|32", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Besides, I have several questions for the paper:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|33", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "(Sorry if I missed it) what model did you use to generate the input and how many inputs did you generate for each task?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|34", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "I would be curious to see some examples of generated input.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|35", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Is it possible to provide more results in exp#2 on datasets other than SQL?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|36", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Have you thought about/tried other optimization methods other than fine-tuning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|37", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "E.g., soft prompt tuning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|38", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "For the step-by-step reasoning experiment, have you tried it on models that are not fine-tuned on NatIns?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|39", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "E.g., directly using chain-of-thought on an LLM?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|41", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Overall the paper is well written and the conducted research is insightful and the results are strong.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|42", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "The paper should be accepted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#1|43", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "The quality of the presentation can be further improved per the weakness section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|12", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Still, I believe the authors do not tackle a really important point, critical to assess the full potential of the methodology: how the knowledge distilled in the network vanishes, or not, as it is trained on more and more tasks.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|13", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Therefore, it would have been interesting seeing results of this phenomena, as another hypothesis laid out on this premise.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|14", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "I believe such hypothesis would have been more relevant to evaluate than some of the ones considered by the authors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|15", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "The concern is that the knowledge distilled vanishes as the network is trained on more tasks, which could be indicated by the decrease in performance of those tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|16", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "If this is the case, the demanding fine-tune required would be less justifiable, and more ad-hoc networks or other fine-tuning methods might be a better alternative.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|17", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Besides, it would be valuable to have confidence intervals in the reported accuracies, in order to understand how the distillation process can be affected by the randomness in the set of provided inputs, as most experiments are just trained on a few epochs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|18", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Finally, I have two other questions:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|19", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "The explanation given in Section 2.4, for the simplification done on the soft-labels, is not really clear to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|20", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Could you further elaborate?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|21", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "In table 1, my intuition tells me that, when training on a single task, i.e. naive post-distill, the performance should be better than when mixing several.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|22", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Is not that right?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|23", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Or is it the case that the single task for naive training is the most complicated one, among all the ones used in the Mixed training?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|24", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "However, according to table 6, the performance of all tasks increases in the mixed version.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|25", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Could you provide some further intuition on this?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|27", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "The paper overall is clear, and the quality of writing is quite good.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|28", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Nevertheless, I consider the location of figures to be confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|29", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "For example, figure 1 is referred after figure 2, while actually both figures could be almost merged into a single one, that explains the idea of template, etc., through the example of number addition.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|30", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Then, while hypothesis 2 uses figure 5, this has been moved to the appendixes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|31", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Besides, figure 5 would benefit from square-axis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|32", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "For the sake of reproducibility and further supporting the statements, the authors provide further results in the appendix, as well as the code.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|33", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Finally, I found a couple typos:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|34", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Cognative: last line 1st page", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#2|35", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "By: in capital letters, in Table 4 caption", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#3|14", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "The authors do not clearly describe the learning objective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#3|15", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Section 2.4 says ‘we minimize the token-level KL divergence’, but there is no mathematical description of how they do this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#3|16", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "I couldn’t follow the motivation for ‘Sequential Distillation’ in section 2.3 and the problem addressed as well as the solution described are vague.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#3|17", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "The evaluation setting in 3.1 is vaguely described.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#3|18", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "‘We select 5 tasks where teacher can most significantly improve…’ - I don’t understand the motivation behind these choices.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#3|19", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "‘We plot the margins for each task in Figure 5’ - The discussion relies on a figure in the Appendix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#3|20", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "The main text should be self-contained and the reader should not have to rely on the appendix to understand the details.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|12", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Overclaiming: The paper made a strong emphasis that the proposed model is able to “internalize” the task, and made this claim many times overall in the paper. However, to my point of view, “internalizing the task” is exactly “fine-tuning”, and in fact, the proposed method is all about “better fine-tuning through distillation”. In this regard, I think the paper is overall overclaiming. Some details that indicate the proposed model is based on more assumptions than the baselines are not explicitly mentioned. For instance, based on my understanding, the proposed model is trained on the unlabeled portion of the test datasets, whereas baselines are not. Is it correct? (I am not super sure about this and asked this question in the “clarity” section.) If it is true, it is a fairly strong assumption to assum the unlabeled data of the test datasets, and should be mentioned very clearly. Not making this clear seems overclaiming.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|13", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Missing the most critical baseline I think the most important baseline is a student model that is instruction-tuned on a particular template (either teacher’s template or student’s template) with no distillation. To my understanding, this baseline is not provided, except for one experiment (Table 2). Instead, the paper claims that the proposed model effectively learns from the teacher model because it is better than the pre-distill student model. However, to my understanding, the pre-distill model is not instruct-tuned, thus it is unclear if this gain is coming from instruction tuning, or distillation from a teacher model. For this reason, I strongly believe that the paper should have mainly compared with the model instruction-tuned but without distillation, and should not claim the effectiveness of the distillation method without this baseline.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|14", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Side note: I think “Direct Gradient Descent” is basically the model that is instruction-tuned but without distillation. Is this correct? I think the name is not very intuitive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|15", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "To my understanding, it is using the template for the teacher model (long instruction, in-context examples) – and if it is correct, I think it’s also good to include the model using the template for the student model (concise instruction).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|16", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "This will support the assumption in the paper that longer instruction is more helpful than concise instruction, which this paper actually never explicitly showed (nor cited relevant prior work).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|17", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Experiments are not very consistent with no specific reason specified.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|19", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "In the experiments with Natural Instruction v2, the result is evaluated only on 5 datasets with the largest gains from instruction tuning and other 5 datasets with the smallest gains from instruction tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|20", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "It is not clear why the paper chooses to do so instead of reporting all datasets, especially because the paper did not report these two sets of 5 datasets separately and discuss how their results differ.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|24", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "The core idea of the paper is very clear and well-written.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|25", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Some experimental details are not super clear from the main paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|26", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "For instance, just to clarify:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|27", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Is it correct that the teacher model and the student model have the same size?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|28", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "Is it correct that the training tasks and test tasks are mutually exclusive when instruction-tuned, and all models (including all baselines and the proposed model) are only using the labeled data on training tasks but not the test tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|29", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "And is it correct that the proposed model performs distillation on the test dataset?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|30", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "In other words, the proposed model uses the unlabeled datasets of the test datasets, whereas the baselines do not?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc12|review#4|33", "source": "rc", "split": "test", "paper_id": "rc12", "comment": "However, I am concerned with overclaiming in the paper, missing critical baselines and inconsistency in evaluation of the model.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc13|review#1|15", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "The ensembling results appear to have ensembling take place over probabilities, rather than discrete predictions -- at least that appears to be the case given that only two models are being ensembled.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#1|16", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Is there any way to apply majority-vote style ensembling instead, to get around the issues with uncalibrated model confidence levels?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc13|review#1|17", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "The discussion on how to construct artificial long-term dependencies is very interesting, but it also is a little bit orthogonal to the remainder of the paper; at least with the current framing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#1|18", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "The paper would be stronger if it could provide a more cohesive narrative in terms of incoporating this section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#1|19", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "The use of BERT-base scale models naturally leaves open the question of whether anything changes at larger scale.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#1|20", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "While the fine-tuning paradigm becomes effective at these scales, recent research suggests that zero-shot capabilities tend to emerge at larger model sizes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#1|21", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Given that zero-shot capabilities are naturally connected to investigating the role of knowledge transfer in pre-training, there is a natural unresolved question of whether self-pretraining would fall behind if the experimental setup is upscaled by at least an order of magnitude.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#1|22", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "This is, of course, understandably hard to assess without enormous compute expenditures.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#1|27", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Overall the paper contains a thorough and comprehensive experimental evaluation of self-pretraining, and the result that self-pretraining can be as effective as off-the-shelf models can have important implications for our understanding of why the pre-training paradigm is as effective as it is.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|13", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "In the introduction the authors make the strong claim that “[...] the benefits from pretraining in this case cannot be attributed to knowledge transfer”, while in Section 2 the authors are more nuanced and allow for the possibility that “sufficient knowledge is [...] present in the downstream dataset”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|14", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "It seems possible that the models can acquire linguistic and factual knowledge from these datasets during pre-training, which cannot be learnt efficiently from only predicting the downstream task labels.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|15", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Since “knowledge transfer” arguments are concerned primarily with the type of knowledge transferred, rather than the source of knowledge, the authors would have to show that certain probing results cannot be replicated in their pre-trained models to support the claim from the introduction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|16", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "The observation that small task-specific pre-training corpora can bring strong benefits for a particular fine-tuning task was already made by Yao et al., 2022, which is not referenced in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|17", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "It is also relevant to Section 7, as Yao et al. construct examples by concatenating top-k BM25 retrieved documents, which is similar to using TF-IDF.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|18", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Another relevant reference in Section 7 is Levine et al., 2022, who propose a similar retrieval heuristic for constructing better pre-training examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|19", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "The name “self-pretraining” seems unfortunate since (a) it evokes parallels to self-training, where a model is used to generate pseudo-labels (see Amini et al,.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|20", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "2022), and (b) confounds general, task-agnostic pre-training with task-specific pre-training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|21", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Regardless of the RoBERTa-base experiments, it remains unclear how the results transfer to auto-regressive models, and how relevant they are to the web-scale pre-training of very large language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|22", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "While experiments would be unrealistic here, I would appreciate a fairer and more in-depth discussion in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|23", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Yao et al., NLP From Scratch Without Large-Scale Pretraining: A Simple and Efficient Framework, ICML 2022", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|24", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Levine et al., The Inductive Bias of In-Context Learning: Rethinking Pretraining Example Design, ICLR 2022", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|25", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Amini et al., Self-Training: A Survey, 2022", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#2|27", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "The paper is well written and presented, but limited in terms of novelty and originality.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|14", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "(1) It's not clear that the results answer the question of \"Our results suggest that in many scenarios, performance gains attributable to pretraining are driven primarily by the pretraining objective itself and are not always attributable to the incorporation of massive datasets\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|15", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "What happens if you sample upstream corpora to match the size of the downstream dataset, and then compare pre-training on downstream vs upstream with similar data sizes?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|16", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "(2) Related to (1), these datasets are all text classification datasets that are pretty similar, so it's somewhat unsurprising to me that pre-training on downstream corpora works well when fine-tuning on other datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|17", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Furthermore, for many of the benchmarks, the performance of RandomInit is already quite high and the gains from pre-training are pretty small already, so maybe it's not so surprising that pre-training on anything helps.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|18", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "I think the analysis of model size is quite helpful, since a natural question is whether or not these results are due to the ELECTRA small model being too small to benefit from pre-training on massive upstream datasets.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|19", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "But another concern in this vein that these downstream datasets are too simple to really see many gains from pre-training upstream datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|20", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "It'd be nice to see performance on other types of downstream datasets, like QA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|21", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "For instance, even looking like something like MNLi performance, it seems like performance is comparable to upstream datasets when pre-training on similar sentence-pair datasets (e.g., QQP or PAWS), but is far lower when pre-training on other downstream datasets that aren't as similar.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|23", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Why is the performance of PAWS on RandomInit 50%?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|24", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Is this a coincidence?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|25", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "I'm a bit surprised that it's so low.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|28", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "I do think the results are somewhat surprising and novel, although there's a bit of overclaiming (since I suspect that these results are largely due to the similarity of the downstream datasets).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|29", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "The work seems reproducible, with details in appendix A about the pre-training and fine-tuning process.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|32", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "However, I do think that this conclusion could be better-qualified (e.g., this is not a general conclusion, but one about text classification datasets).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|33", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Furthermore, it's not clear to me that this paper directly answers the question of whether pre-training data scale or pre-training objective is the key to better transfer performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#3|34", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Despite this, I think the work is thorough and interesting and would be of interest to the ICLR community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|11", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Despite the very impressive results, I think the elephant in the room, which I was surprised not to see any mention of in the paper, is how much do the results have to to do with the corpus being one of the downstream task corpus, or simply a smaller corpus.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|12", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "That is, a baseline of pretraining on a general pretrain corpus of similar size to the downstream corpora is missing and is very important to shed light on the results here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|13", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Is the conclusion from these results basically that you need far less pre-training in order to get good finetuning performance?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|14", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "I am aware that size doesn't explain it all, but I am also far from convinced that the corpora being downstream corpora makes an important difference here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|15", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "The error consistency experiments are interesting, but raise a few potential concerns.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|16", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "First, computing the prediction difference should be done on the validation set, not the test set (especially given the attempts to use this information to do smart ensembling, which could lead to overfitting the test set).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|17", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Second, there is no evidence presented in this paper that using the model confidence to select the correct model should lead to good ensemble performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|18", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Are models more confident in their correct predictions compared to their wrong predictions?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|19", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "This is a general desired property of models, but not one that is guaranteed and in particular not one that is explored in this paper (at least not reported), which makes it it not very surprising that these experiments failed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|20", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Finally, I am not sure I agree with the conclusion (stated from the conclusion section): \"The errors made by such self-pretrained models on the downstream tasks are significantly different from the ones made by the off-the-shelf models pretrained on upstream corpora.\" The error inconsistency numbers are indeed higher among the two approaches, but still overall quite low in 4/8 of the datasets, and even the higher ones are only around 10%.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|22", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Nonetheless, it does not seem related to long-term dependencies (as suggested by the title and the intro of that section), as even the authors do not argue that such dependencies exist in the reordered corpus.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|23", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Finally, more a suggestion than a weakness: what about zero/few-shot?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|25", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "This assumption stands in the heart the strong few-shot/zero-shot results observed in recent years, even with reasonably sized models (e.g., [1,2]).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|26", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "Did the authors consider running such experiments with their minimally pre-trained models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|27", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "I agree that this might be out of scope for this paper, but something the authors could consider.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc13|review#4|35", "source": "rc", "split": "test", "paper_id": "rc13", "comment": "However, many of the central claims in this paper are not validated, e.g., the importance of the downstream corpora, the significant difference in the error patterns between the different approaches, and the argument regarding long-term dependencies.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|1", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "This paper observes that existing models can only generate seen libraries and function calls at test time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|2", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "However, new functions and libraries are introduced all the time, and even a seen function call can have arguments that were not used in the training data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|3", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "To address such unseen functions in code generation, the authors propose a simple and effective approach for code generation by retrieving the relevant documentation, dubbed DocPrompting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|4", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The extensive experiments demonstrate that DocPrompting consistently improves NL->code models in two tasks, in two PLs, and across multiple strong base models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|5", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "DocPrompting improves strong base models such as CodeT5 by 2.85% in pass@1 (52% relative gain) in execution-based evaluation on the popular Python CoNaLa benchmark; on a new Bash dataset tldr, DocPrompting improves CodeT5 and GPT-Neo-1.3B by up to 6.9% exact match, and Codex by 6.78 charBLEU score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|8", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The idea of leveraging the newly added manuals and documentation to generate code containing unseen and unused functions and libraries is interesting and novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|9", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The work demonstrates an exciting endeavor direction to generate high-quality code.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|10", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The main results from Tables 1 & 3 demonstrate that the improvement of the proposed DocPrompting is significant across various pre-trained models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|11", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Conducting experiments on more than one benchmark and thorough ablations make the conclusion of this work convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|12", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The presentation of this work is excellent.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|13", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The rich use of diagrams and tables makes this paper easy to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|16", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "I examined the supplementary material provided by the authors and only found the data files.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|17", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "It would be better if there were codes to prove its reproducibility.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|19", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Please see the Strength And Weaknesses", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#1|22", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Considering the technical simplicity and reproducibility, I give a weak accept.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#2|3", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The authors experiment with using BM25 and training dense encoders for retrieving documentation, and find consistent improvements with DocPrompting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#2|4", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The authors focus on evaluation settings where functions in the test set are not seen during training, and also introduce the new code generation dataset “tldr”.", "reviewer_type": "human", "label": "error", "error_type": "Inaccurate Summary", "label_origin": "expert"}
{"id": "rc|human|rc14|review#2|12", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "I am unsure how fair the evaluation setting with unseen functions is for the baseline models, especially since the ExPrompting baseline cannot retrieve examples of unseen functions from the training set.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#2|18", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The paper is well executed and will likely lead to future work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|12", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "the technical approach is essentially an application of known techniques.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|13", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "(update: After reflection and discussion with the authors, I admit that there is novelty in the problem formulation as well, which goes beyond the setting considered by previous work.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|14", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "(update: The unsupported claims have been qualified in the new revision.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|15", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "(update: The paper has been reworded to make this more clear.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|16", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "It's possible there is some function-knowledge leakage due to the use of pretrained code models for retrieval (perhaps the authors can clarify this) (update: I am still curious whether this influences the results, but it is a bit orthogonal to the main point of the paper, since it would also affect baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|17", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The authors have added a brief discussion of this potential leakage to the appendix.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|20", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Overly-strong claims about existing models not being able to generalize (addressed by new revision)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|21", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The introduction makes very strong claims:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|22", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "all existing code generation models either learn directly from input-output pairs provided as training data ... or learn the mapping between input and output implicitly from naturally occuring corpora", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|23", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "all these works assume that all libraries and function calls were seen in the training data; and that at test time, the trained model will need to generate only seen libraries and function calls", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|24", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Thus, existing models inherently cannot generalize", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|25", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "These claims seem much too strong.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|26", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Although there has been a lot of work in the setting the authors discuss, there have also been a number of works that do not fit neatly into this categorization, and that allow models to generate code using new functions and libraries.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|28", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Hashimoto et al. (2018) and Lu et al. (2017) learn to retrieve from a dataset of examples at inference time, which could be distinct from those used at training time. Some of the other works the authors cite also do something similar (e.g. Parvez et al. (2021) and Pasupat et al. (2021))", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|29", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Shrivastava et al. (2021) show that code generation can be improved by conditioning on context extracted from other relevant files, which may include new functions", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|30", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Wu et al. (2022) use an external memory to allow transformers to memorize new data at training time and recall it later, and show that this allows it to predict newly-defined functions at test time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|32", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "I am not aware of prior work which specifically retrieves documentation to generate code from natural language, although there has been prior work on both retrieving relevant documentation (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|33", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "section 5.6 of this survey) and on retrieving existing code to generate code (e.g. Hashimoto et al. (2018)).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|34", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "As such, the empirical results provided by the authors do provide new evidence that documentation helps for code generation across a number of model architectures and tasks (as one might expect).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|35", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Additionally, the tldr task, the resplit of CoNaLa, and the documentation pool from DevDocs are new contributions of this work, and might be useful for evaluating improved documentation-retrieval models in the future.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|36", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Beyond this, the technique itself appears to be essentially the same as prior work: they use standard retrieval techniques based on existing algorithms and pre-trained models, and their generation models are also models that have been used for code or text generation before.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|38", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Overall, the paper is clearly written.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|39", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "One thing I thought could be improved: Section 3.1 starts from the assumption that we are given triplets $(n, c, \\mathcal{D}_n^*)$ where $n$ is a natural-language intent, $c$ is the target code, and $\\mathcal{D}_n^*$ is the \"ground truth docs\" for $n$.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|40", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "I didn't think this was very well explained, since it's not obvious why $n$ would have ground truth documentation associated with it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|41", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "From the rest of the paper and the appendix, it appears that $\\mathcal{D}_n^*$ is actually selected based on a heuristic that uses $c$, not $n$.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|42", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "I think it would be better to explicitly describe this process in section 3.1, e.g. by stating that we are given a pair $(n, c)$ and then we infer a set of reference docs $\\mathcal{D}_n^*$ from $c$.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|43", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "(I also wonder whether \"ground truth docs\" is the best terminology to use, since it's still being inferred heuristically and not manually observed or labeled.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|44", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "*(update: this has been addressed by the new revision.)*", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|45", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Other minor suggestions: *(update: addressed by the new revision.)*", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|46", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "- Section 4.2 states that ground-truth docs are constructed by \"matching the function name with its documentation\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|47", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "However, there may be multiple function names in an example, which makes this statement ambiguous.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|48", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "- I found page 7 hard to read, since it has a table at the top and a figure in the middle, with ordinary text interleaved around it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|49", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "- Figure 3 has overlapping text, and the axes are not labelled (I assume the x-axis is supposed to be $k$?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|51", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The dataset setup and experiments, which make up the bulk of the contribution, appear to be high quality overall.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|52", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The tldr dataset is an interesting source of documentation for bash code, and re-splitting the CoNaLa dataset based on use of new functions is a good idea for assessing generalization ability.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|53", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Experimentally, the authors do a good job of evaluating DocPrompting across a variety of metrics and for a variety of base models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|54", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "They also conduct a series of interesting ablations and extensions, including using oracle access to the appropriate documentation, comparing to example-based prompting, and investigating different retrievers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|55", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "One potential concern is that the strongest results on CoNaLa are given by CodeT5, but CodeT5 was pretrained on a dataset of code that might include some of the \"held-out\" Python functions described in 4.2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|56", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "This might be a source of knowledge leakage, since CodeT5-based retrievers might just pick documentation that includes functions that accomplish the task by simply memorizing the functions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|57", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "It would be helpful to analyze whether this leakage might have occurred, and if so, how much of an effect it might have on the results in Tables 3 and 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|58", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "(Am I right in assuming all of the results in Table 3 used a CodeT5-based retriever?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|59", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "*(update: Authors have added some discussion of this in the appendix.)*", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|60", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Additionally, I wasn't convinced by the argument in section 6.1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|61", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "If I understand correctly, this section compares n-gram overlap between the natural language $n$ and code $c$, and also n-gram overlap between the \"ground-truth\" docs $\\mathcal{D}_n^*$ (along with $n$) and the code $c$.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|62", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "But, from the appendix, $\\mathcal{D}_n^*$ is selected exactly by finding the documentation with the highest string overlap with $c$!", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|63", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "So it seems obvious that $\\mathcal{D}_n^*$ and $c$ would have high n-gram overlap, regardless of how much documentation does or doesn't \"ease the mapping between NL and code\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|64", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "(A better experiment might be to look at n-gram overlap between $c$ and the *retrieved* examples from the model, using only information from $n$ instead of the oracle.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|65", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "It would also be interesting to look at overlap between the NL description and the documentation to make sure the gap is actually being \"bridged\" on both sides.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|66", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "*(update: The authors have taken my suggestion here and presented results using the retrieved documents.)*", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|68", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The technique is fairly straightforward, and the authors release their new benchmarks and datasets, so it seems likely this work will be reproducible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#3|73", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "I have increased my score from 6 to 8.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|8", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "First paper to consider retrieving documentation for code generation", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|9", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Two curated datasets that can be useful for future work", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|10", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Extensive empirical experiments and ablation studies demonstrating the utility of their approach", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|18", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Adding retrieval will likely increase latency.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|19", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The authors do not discuss/report how this is affected.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|20", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "As the size of the document pool increases, it seems that model latency will also increase.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|21", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "While the reason for splitting paragraphs into individual documents is explained in the Appendix, it seems odd to treat them as independent documents.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|22", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "For example, in Figure 2, the two bottom documents cannot standalone without the top one.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|23", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "It is not clear how k is selected for the number of documents to retrieve.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|24", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Is it always 3?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|25", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The annotation process for writing the test cases is not clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|26", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "In Appendix B, it is written that the two first authors did this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|27", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "It is not clear whether each example was annotated by both of them and if so, what the agreement on that was.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|31", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Quality: Well-designed and experiments and well-written work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|33", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "However, in terms of technical novelty, already existing models and techniques are used for retrieval and generation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|34", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Reproducibility: Model training details are provided in the Appendix, which should be sufficient for reproducing these results, once the datasets are made available.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|36", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "While the technical novelty may be limited, there is novelty in terms of application and benchmark creation which will be useful for future work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|37", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "The experiments and empirical results are thorough.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc14|review#4|38", "source": "rc", "split": "test", "paper_id": "rc14", "comment": "Overall, I feel that this is a good paper contributing useful ideas and artifacts to the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|7", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Why only a subset of SentEval for the English experiments (3.2) but then the full SentEval in the multilingual XEval experiment (4.5.1)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|8", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Especially if you are trying to make a single-sentence encoder but then evaluating on SICK-E instead of SICK-R, which is arguably a more applicable eval set.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|9", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Why no XLM-R in the amazon reviews analysis (4.5.2)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|10", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Sec 1. Fig 1: box of chocolates", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|11", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Sec 2. --“conditional”, → “conditional,”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|12", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "--no quantitative comparison of using max vs mean pool vs CLS embedding", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|13", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "--first sentence is feed → first sentence is fed", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|15", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "--refer to using the same set of encoder weights for different inputs as siamese networks, as done in the sentence-bert paper https://en.wikipedia.org/wiki/Siamese_neural_network", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|16", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- v_d is used but not defined", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|17", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Sec 3. --Skip-thought originally used a sentence to predict the generation of both the preceding and succeeding sentences.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|18", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "This is functionally equivalent to your flip-flopping the order of the consecutive sequences.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|19", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "I would make the point that these steps are equivalent.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|20", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Note that this is also not necessarily making the task “more challenging” (and moreover I am not sure why “more challenging” equates with “better pretraining method for language understanding” -- and an ablation of this step is not included to show that it is in fact necessary and useful).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|21", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- similarly, no analysis of masking rate, nor explanation for why ‘more challenging’ is better.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|22", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- “We explore two transformer configurations, base and large, same as in the original BERT paper.” – fragmented", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|23", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- The number of projections N = 15.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|26", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- On the specific subset of SentEval tasks you’ve selected, the majority of the performance discrepancy is in the SICK-E task--otherwise, the overall #’s are rather interchangeable. How does this change if you add in the rest of the SentEval tasks, and why were they omitted? Analysis/exploration for why you get such a performance boost only on SICK-E would also be useful.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|27", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- “the length … set to be 256 tokens”: please clarify whether the “length” refers to the maximum length, or each sentence is a fixed-length chunk consisting of 256 tokens", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|28", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "--typo: “we also exploring”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|29", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Sec 4. --If your introduced bitext retrieval objective uses batch size, experiments comparing the effect of batch size is necessary.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|30", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- Please specify the value of the margin m being used in the experiments", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|31", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "--Choice of number of projections is also not motivated (and in fact contradicted by the ablation experiment finding that 15 is better)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|32", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- the motivation and contribution for XEVAL are great-- the explanation of the dataset is lacking.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|33", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "What translation API was used?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|34", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "How was the XEVAL score computed for each language?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|35", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Is it the full set of SentEval downstream tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|36", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- cite precedent for using the concatenation of u,v, u-v and u*v. (or show its effect via ablation)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|37", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- BR → CMLM+BR configuration not evaluated", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|40", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- typo: “has a significant upon mBERT”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|41", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Sec 5. --It is not clear that you can make the claim that the first PC only encodes language-identification information?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|42", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "--I assume Figure 3 is 2-dimensional TSNE (needs a cite), which comes with its own set of caveats as a visualization tool.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|43", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Quantitative clustering analysis such as silhouette coefficient might be more appropriate than a plot.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|44", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "If Figure 3 is not t-SNE, please specify the meaning of X and Y axes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|45", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- did not try higher than n=15 projections but claimed it was optimal", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|46", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- The description of the “skip” ablation is unclear: please clarify what is meant by “concatenated with the sequence outputs of s2”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#2|47", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "-- typos: “removing the first principal component … effectively eliminate”, “for both two models”, “representations … generally exhibits”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#3|0", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Although in this DNN dominating era, we cannot expect the explainability as we had before, I still believe that a solid work should be grounded on a reasonable basis, which could be in a high level, such as BERT and SBERT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#3|10", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "First, I think the explanations for the intuition of the proposed model can be clarified, especially in the introduction section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#3|12", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "The proposed model is the combination of Skip-Thought (Kiros et al., 2015) and BERT masked LM (Devlin et al., 2019).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#3|13", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Their experimental results show a detailed comparison of BERT but ignore much about the Skip-Thought.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#3|14", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Although the authors mentioned the results of the Skip-Thought model on the SentEval benchmark, the encoder used in the Skip-Thought (Kiros et al., 2015) is RNN while the author used the Transformer Encoder.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#3|15", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "I would appreciate a better and fair comparison of the Skip-Thought model by using same transformer encoder and same training corpus.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#3|16", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "I am not sure why you used the concatenation when you do the masked LM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#3|17", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Are there any other ways to do that?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#3|18", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "It can be more convincing to see some analysis or results here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|1", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "However, the paper continues to be centered around the SentEval benchmark results. While SentEval is a useful benchmark to evaluate sentence representations, it doesn't reflect well how these representations will be used in practice.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|19", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Model is largely based on prior work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|20", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "The main contribution is not clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|21", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "There are many settings considered in the paper and it is unclear if the proposed contributions are truly significant due to weak baselines are differences in data used for training different methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|22", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "There are several issues with the experimental setup.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|25", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Baselines: It is unfair to compare the proposed method against baselines like BERT which are not designed to produce fixed length encodings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|26", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Data used for pre-training: It is difficult to gauge how good the method is in comparison to other models due to differences in the data used for pre-training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|27", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Ideally, there should be a table comparing models that use the exact same resources.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|28", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "In Table 1, although BERT-base/large is trained on the same data, it is not a strong baseline since mean pooled representations from the encoder are treated as a sentence representation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|29", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Ideally, the model should be compared against a skip-thought baseline or an unsupervised sentence representation learning method that uses the same resources for training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|31", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "The authors need to make it clear exactly what resources are used for training each method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|32", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "Presentation can be improved, especially the organization of the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc15|review#4|33", "source": "rc", "split": "test", "paper_id": "rc15", "comment": "It is difficult to follow the paper and identify the main contributions in the current presentation.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|6", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "At a high level, the method proposed in this paper makes sense, but there is a critical problem in terms of novelty: the idea of using NLI to probe stereotypes is not new.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|7", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "In fact, nearly the same proposal outlined here is explored by Dev et al (2020), who additionally use the mechanism to probe for other kinds of stereotypes as well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|8", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "The hypothesis templates are interesting, but present a bit of a technical question.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|9", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "The hypothesis of the form \"This text talks about a female occupation\" refers to the text of the premise, rather than the events or entities in it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|10", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "In other words, it talks about the form of the premise, rather than its meaning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|11", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Of course, there's nothing wrong with this, but it breaks a crucial assumptions about how the NLI data (in particular the SNLI data) was sourced: the events and entities in the hypothesis refer to the events and entities in the premise as much as possible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|12", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "In contrast, the word \"text\" in the hypotheses constructed in this work refers to the entire text of the premise, and not its entities and events.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|13", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "It is not clear how this change affects model performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|14", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "One way to fix the issue is to change the hypothesis templates to use the same (or similar) words as the premises, and replace the occupation word with a gendered word.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|15", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "(But doing so would make the work even closer to that of Dev et al 2020.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|16", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "It is not clear why the neutral label is removed and the problem is converted into a binary problem of deciding whether the hypothesis is entailed or contradicted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|17", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "It seems that most of the hypotheses would actually be neutral, and a good model should allocate most of its probability mass to the neutral label.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|18", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Why do we have to force a choice between entail and contradict, when a stereotype-free model would actually predict neutral?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|19", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "The results in table 6 are interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|20", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "It seems that the models \"memorize\" the distributional correlations between gender and jobs differently for men and women.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|21", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Are there any conjectures about why this may be the case?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|22", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "It is not clear whether the bias that is being measured is in the representation (i.e. the *BERT embeddings) or the task (i.e., the NLI data).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|23", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "The experiments suggest that the problem is perhaps in both.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|24", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Previous work on stereotypes involving language has largely focused how they are encoded in the embeddings, and removing them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|25", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "This paper seems to argue that the provenance of the stereotypes is the training data for the task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|26", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "However, the final results suggest that this is not entirely the case, and the paper does say so in the section on debiasing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|27", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "It may be worth posing the question about the source of the biases early on in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|28", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Since the paper is talking about stereotypes in language technology, the authors should go over the work of Blodgett et al (2020) to better situate the motivations and outcomes of this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|29", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Indeed, there should be a discussion in the paper about the cultural context and assumptions that are implicit in the measurements.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|30", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "(For example, is the definition of B based on an American context? Would the measures transfer to a different country/cultural perspective?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|31", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Minor point: The plot in figure 1 should not be a line plot because the horizontal axis is categorical.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|32", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "A bar chart would be a better fit (and would convey the point more clearly).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|34", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Dev, Sunipa, Tao Li, Jeff M. Phillips, and Vivek Srikumar. \"On Measuring and Mitigating Biased Inferences of Word Embeddings.\" In AAAI, pp. 7659-7666. 2020.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#1|35", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Blodgett, Su Lin, Solon Barocas, Hal III Daumé, and Hanna Wallach. \"Language (technology) is power: The need to be explicit about NLP harms.\" In Proceedings of the Annual Meeting of the Association for Computational Lingustics (ACL). 2020.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#2|10", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "The ground truth for the proposed examples seems to be neutral (occupations are neither male nor female), so I would like to know how often evaluated systems actually predict this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#2|11", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "2.The bias measurement forces the models to predict either entailment or contradiction, where in fact the ground truth answer, in my opinion, for the proposed NLI examples, is neutral.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#2|12", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "(the occupation is neither male nor female).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#2|13", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "For all we know, the models are correctly predicting that with high probability, but the measurement is forcing a renormalization between entailment and contradiction examples (Section 3.2).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#2|14", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "This seems like it would be a problem for the \"delta P\" and \"B\" measurement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#2|16", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Zhao et al. had a turking process to make sure all entities in CONLL data were covered in the swaps.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#2|17", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Were any such measures taken here to deal with new NLI data?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#2|18", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "How was the list of swaps constructed in this case?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#2|19", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "The results are split between Table 5 and Table 6, so I was unsure where it helped, or in the unexplained Figure 2.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc16|review#2|20", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "From Figure 2, it seems like it hurt in some cases.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|5", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "However, I have some concern about this paper:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|6", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "When constructing the evaluation dataset, the authors replace the occupation word in the premise with other occupation words.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|7", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "However, this can lead to inconsistent semantics.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|8", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "For example, \"the doctor is operating\" becomes \"the teacher is operating\", which may not fit the realistic situation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|9", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "The constructed dataset contains only entailment pairs. How about analyzing the contradiction cases as well?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|10", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "When analyzing the results, the authors disregard the neutral case.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|11", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "I am wondering if they train the models in the same way.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|12", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "If not, it seems that there is a domain mismatch.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|13", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "This point is my primary concern.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|14", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "In the constructed dataset, the hypothesis is generated by templates and looks like the context is not very related to the premise.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|15", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "However, in most of NLI datasets, the premise and hypothesis are usually related.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|16", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "So the domain of training set and their constructed evaluation set are different.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|17", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "It can be reasonable for models to perform not well and have the bias on the evaluation set, since the domain changes a lot.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|18", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "Why not consider the existing hypothesis?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|19", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "For instance, replace \"he\" or \"his\" in the existing hypothesis with \"she\" or \"her\" so you can have female hypothesis and male hypothesis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|20", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "I don't quite understand the definition of B for evaluation. Is that probability for some predefined gender-specific occupations? If that is the case, how to define those words?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|21", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "What is the definition of \"bias\" in Figure 1?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|22", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "In Figure 6, how to calculate delta P for female and male respectively?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|23", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "In my understanding, delta P is the probability gap between female and male hypothesis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|24", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "I suggest that the authors use one table to show the difference before and after the data augmentation to compare the numbers more easily.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc16|review#3|26", "source": "rc", "split": "test", "paper_id": "rc16", "comment": "In Table 2, teacher => guard.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|5", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Along with the dataset, this paper presents a new model, ReGen.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|6", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "ReGen is composed of a retriever (based on SPLADE) and a reference encoder and target decoder (based on FiD).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|7", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "When trained on the WebBrain data, ReGen outperforms both large language model baselines and also a retrieval augmented model with a dense retriever (FiD + BART) according to a range of automatic metrics, and also annotator judgements of a small set of predictions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|8", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "I have some questions about some of these comparisions below, but it is clear that ReGen is capable of generating coherent and informative text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|9", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "It is also capable of providing references for this text, which is a significant advantage over methods that don't include retrieval.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|10", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "There are a number of nice ablations that show how the different systems' performances change with different numbers of retrievals and gold references.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|11", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "These comparisons are useful in understanding the tradeoffs between retrieving a lot of supporting evidence, with lower precision, vs fewer high quality references.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|14", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "This paper promises a massive new dataset for retrieval augmented text generation, that could be very useful to the NLP community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|15", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Having said that, I do have questions about the release with respect to copyright and privacy concerns (detailed below in the ethics section).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|16", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The end system can retrieve references, and link them to generated sentences, which is a very nice end product which could have real utility beyond unconstrained text generation, which cannot provide any form of attribution for its predictions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|17", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "There are a number of interesting modeling and training choices made, which lead to significant improvements over a very large language model (GPT3) and also similar retrieval agumented approaches with different choices of retriever and training procedure.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|19", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "There are a lot of separate contributions here that are not independently evaluated (data filtering / retrieval filtering / warmup strategy).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|20", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "It would be nice to see evaluations that validate these choices.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|21", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The reference passage filtering process selects the references included in the retrieval corpus references according to term overlap with the target Wikipedia article.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|22", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "If I'm understanding this process correctly, this means that the references stored in the corpus have been selected according to observations of the test time targets, so there is some leakage of information from the test targets into the model input.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|23", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The paper should discuss this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|24", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The prompt given to GPT3 \"Introduce [Page Title]\" does not mention that the target is Wikipedia-style text, and the example in Figure 1. does not look much like the start of a Wikipedia article.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|25", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Meanwhile, GPT3 is definitely able to generate Wikipedia style text if prompted to do so.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|26", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The comparision to GPT3 would be stronger if the prompt was better tied to the actual task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|29", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The paper is clearly written and all parts are pretty well described.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|30", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The authors commit to releasing the data, which will allow easy replication.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|32", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The dataset is similar in form to previous work, but extends that previous work to the scenario where references must be retrieved from a very large corpus.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|33", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Similarly, the model is a small modification of existing approaches but it appears to work better than well chosen baselines for this task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|35", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "There are a lot of details in the filtering procedures used to select the contents of the dataset, and assign references to sentences.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|36", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "These are all justified int the text, but are not supported by any sort of analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|37", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "If this is going to become a benchmark dataset, going forward, it would be good to see a discussion of how these choices affect the task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|38", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "In particular, see my question about test target leakage into the retrieval corpus above (under Weaknesses).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|41", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The new model is also quite different from previous work, which has generally relied on dense DPR-style retrievers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|42", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Overall, I think this paper is a nice contribution and I am currently leaning toward acceptance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#1|43", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "However, I also have some serious questions below about the data release strategy, and how the authors propose to handle concerns about releasing multiple terrabytes of data which may include copyrighted works or personal information.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|6", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The authors contribute a dataset \"WebBrain-Raw\", comprised of English wikipedia plus all crawlable references.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|7", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The dataset is cleaned with an interesting set of heuristics.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|8", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The authors additionally propose a baseline system called \"ReGen\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|9", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "ReGen combines several modeling choices:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|10", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "training a SPLADE retriever for this task with hard negatives are mined,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|11", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "tuning sparsity in document representations for retrieval,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|12", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "filtering retrieved documents trading off consistency and diversity,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|14", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "making generation more grounded in the references, by pre-training the generation component to generate individual sentences from a reference passage at time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|15", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The authors finally present analysis showing the usefulness of these modeling choices, compared to more vanilla choices.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|18", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The task examined in this paper (retrieval + multi-document summarization) is important.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|19", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The dataset released by this paper is much larger than comparable datasets and could be useful for furthering work on this topic.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|20", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The authors describe in detail many technical details of ReGen.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|21", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "These details can be useful to practitioners for reproducing ReGen’s results and in their own work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|23", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Limiting the generation to only the first passage of wikipedia pages is a pretty strong limitation, also making this work very close to existing multi-document summarization works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|24", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The authors do acknowledge this similarity though.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|25", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "In “Reference Passage Selection” and “Dataset Generation”, the authors select input passages for training with simple word overlap or BM25.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|26", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "It seems like this could be easily improved by using an entailment model, a dense retriever or even SPLADE.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|27", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The proposed \"WebBrain-Raw\" dataset would be more interesting if it was multilingual instead of English-only.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|29", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The paper is clear and good quality.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|30", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The system is described in detail and should be reproducible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|31", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The novelty of the approach is mainly in the size of the released corpus and in including retrieval as part of the task, together with multi-document summarization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|32", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The techniques proposed by the authors are very interesting, but they are only applied to the dataset they created, so it's not totally clear how their modeling choices would stack up against other methods on a more competitive benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|34", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "This looks like a useful dataset+baseline contribution for the important task of multi-document summarization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#2|35", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The techniques used by the authors seem solid.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|9", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The paper introduces WebBrain, which lets the model retrieve supporting evidence and generate factual articles given a factual query.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|10", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The proposed dataset is somewhat similar to the Wizard of Wikipedia (Dinan et al., 2018).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|11", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The newly proposed dataset is interesting and large-scale.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|12", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The authors crawled and cleaned Wikipedia.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|13", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The proposed task and corresponding dataset are very interesting and worthy of future research.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|14", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The paper proposes a new retrieval-augmented generation framework based on SPLADE and FiD.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|15", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The proposed methods achieve the best results over automatic and human evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|16", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The experiment section is very comprehensive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|17", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The authors conduct an ablation study with different retrieval models and show the impact of the different numbers of retrieved references.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|18", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The paper also checks the impact of a number of references.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|19", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Those results are clearly represented in tables or charts with detailed explanations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|20", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The paper shows the case study, human evaluation, and reference mark correction strategy in the appendix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|22", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The paper uses n-gram overlapping metrics for automatic evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|23", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The paper needs to include some newer metrics such as BERTscore (Zhang et al., 2019), and BARTScore (Yuan et al., 2021) which can check semantic similarity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|24", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Most of the experiment analyses are in quantitative way.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|25", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "I would like to see more qualitative analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|26", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Zhang, T., Kishore, V., Wu, F., Weinberger, K. Q., & Artzi, Y. (2019). Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|27", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Yuan, W., Neubig, G., & Liu, P. (2021). Bartscore: Evaluating generated text as text generation. Advances in Neural Information Processing Systems, 34, 27263-27277.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|28", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Dinan, E., Roller, S., Shuster, K., Fan, A., Auli, M., & Weston, J. (2018). Wizard of Wikipedia: Knowledge-powered conversational agents. arXiv preprint arXiv:1811.01241.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|30", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Some parts of the paper are not very clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|31", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The steps to create WebBrain-R and WebBrain-G is unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|32", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The paper attached the implementation details in the appendix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|33", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "It also provides examples from the dataset for readers to check.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|34", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "However, it does not provide any code for reproduction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|35", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "It shows the limitation and system demonstration in the Appendix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#3|38", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The experiment part is quite comprehensive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|6", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "This is an interesting and important task, and I think it's one that could be a fruitful combination of lot of existing work on both generation and understanding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|8", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Despite the grandiose motivations, the authors consider an extremely narrow instantiation of this task:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|9", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Only the first paragraph is used as the generation target, rather than entire articles.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|10", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "When I read the abstract and introduction I was excited to see how the authors would handle generation / understanding of long sequences, since this is an area where I think this benchmark would be particularly useful, but was disappointed to see that they largely punted on this problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|11", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Furthermore, the authors don't use the entirety of the reference articles, instead taking only a single paragraph per article that is specifically chosen to maximize informativeness, which I think gives a bit of an unfair advantage to the retrieval-based systems (since they're much more likely to retrieve something useful).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|12", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "In general, I feel like the paragraph \"Reference Passage Selection \" in 3.2 completely misses the mark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|14", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The original Wiki articles and reference articles tend to be very long.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|15", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "To adapt the capacity of most pre-trained language models (e.g., BERT has a 512 token limit), we use the first section of Wiki articles as the generation target, and we select the passage from the reference article with the highest PST value as the supporting passage.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|16", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "I'm not sure that I feel very good about our benchmarks having to be \"adapted\" to be consumable for existing models while trading off the essence of the task (synthesizing information from multiple long documents to generate a long document). We should set the benchmarks as goals, see how current models and techniques perform on them, and allow them to be targets for future work.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|17", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "I feel like the baselines are pretty weak, and the ReGen model is a pretty standard retrieve-and-generate baseline. I don't think there's too much modeling novelty here.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|18", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Furthermore, the query distribution seems pretty unnatural.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|19", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "For example, in the appendix, the query used is \"how google translate work\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|20", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "This is a query with a pretty clear information-seeking / question intent that can be answered by a paragraph (or paragraphs).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|21", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "However, the wikipedia page for \"Google Translate\" answers far more than just this question, it contains a variety of general information about the query \"google translate\" (which itself could have many different question intents).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|22", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "So, I feel like the source of supervision provides a different sort of information than what a query like this would actually want (let alone that in practice, only the first passage is used, which contains even less information because of the way that lead sections in wikipedia are structured to contain general high-level information: https://en.wikipedia.org/wiki/Wikipedia:Manual_of_Style/Lead_section).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|23", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The original motivation mentioned \"unseen factual queries\", but it's not really clear that generating an article is what you want for this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|24", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "If it's factual, you can probably answer it with a paragraph or two.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|25", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "I think this work would benefit from a clearer exposition of what these queries would actually be, and what the right source of supervision / data for effectively addressing these queries would look like.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|27", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "This paper was reasonably clear, and I thought that the setup was pretty solid.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|28", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "It's an entry in a long line of work on generating Wikipedia paragraphs from references or other external knowledge.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|30", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "While I like where this paper is going, I feel like the practical artifact and problem tackled here are a bit too small-scale and narrow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|31", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "We're at a point in NLP where things actually kind of (?) begin to work, and there have been a variety of impressive results on generating and understanding long documents.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|32", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "I think this work would really shine in pushing this direction, but instead, it compromises itself to fit the limitations of existing models, limiting its usefulness for future work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc17|review#4|33", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Thus, I'm not sure that I can recommend acceptance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|10", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Although it's encouraging to see that the zero-shot SuperGen outperforms the few-shot setting, I'm not fully convinced by the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|11", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "As been widely discussed, the data selection process under a few-shot setting can have a huge impact on evaluation results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|12", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Not to mention the impact brought by prompts, decoding strategies, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|13", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "I understand that some of these factors are difficult to control, so I suggest the following baselines for evaluation:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|14", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "PLMs fine-tuned on the original train set for each task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|15", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "While I understand that these results will outperform the results of SuperGen, I think it is still relevant and important to know what is the gap in performance between fine-tuning the PLMs and the zero-shot evaluations as shown in the Table.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|16", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "A knowledge distillation baseline: using the PLM to generate the output label when provided with an unlabeled text input (from some corpus), and then using these pseudo labels to train a discriminative model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|17", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "This would help shed light on what is the benefit of creating the synthetic data rather than directly transferring knowledge from the PLM to downstream models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|18", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "It would have been interesting to see if the generated synthetic data in a zero-shot manner could add to the labeled supervised data for the tasks via data augmentation, and help improve the performance of the fully supervised setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|20", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "\"we include the average and standard deviation over 5 different random seeds.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|21", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Are these 5 runs performed on 5 different few-shot train/dev sets or on the same pre-sampled set over 5 different runs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|23", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "In spite of excellent results on easy tasks such as simple text classification, the overall performance is still poor on tough tasks like CoLA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#1|24", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "This limitation has to be discussed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#2|9", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Calling this entire setup “zero-shot” is a bit of a stretch.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#2|10", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "The generator can be seen as zero-shot for generating task-specific examples, but the fine-tuned classifier not so much.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#2|11", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "For this reason, the truly zero-shot baseline seems weak and it is unsurprising that the classifier fine-tuned on synthetic data outperforms it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#2|12", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Perhaps an interesting comparison that was left out was fine-tuning on few-shot data vs. fine-tuning on a combination of few-shot and the synthetic data to quantify how much value is added by the synthetic data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#2|13", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "There’s nothing new and surprising about using synthetic data, it’s more the approach to generating the synthetic data that is the interesting contribution here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#2|14", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "The story (and title) could use some revision to clarify this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#2|16", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Have you considered generating synthetic data using the few-shot examples?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#2|17", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Do you think that would make it possible to generate higher quality data with a greater value-add?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#2|20", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "I think the limitations section doesn't really discuss potential societal impact. Perhaps the synthetic data generated by the model might contain unwanted biases that could be harmful.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc18|review#3|14", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Since prompt would be a key factor for the performance, it would be nice to have prompt ablation studies for each task, not just three tasks(MNLI, RTE, MRPC).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#3|15", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Even though SuperGen significantly outperforms zero-shot prompting and achieves a better result than all the few-shot methods, the performance cannot match to their SOTA results.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc18|review#3|18", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "In Table2, the performance of SuperGen without label smoothing for CoLA drops a lot.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#3|19", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Is there a specific reason for that?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#3|20", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Generator selection: Why did you not use large size PLMs (e.g., GPT3) from the beginning for the high-quality generation, if the model size matters for high-quality text generation like you mentioned in page 8, line 280?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc18|review#3|21", "source": "rc", "split": "test", "paper_id": "rc18", "comment": "Are there other regularization methods other than label smoothing and temporal ensembling that you might have considered?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#1|8", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "I think authors fail to properly state all the possible explanations to support the existence of transfer learning with scrambled or random input, If transfer learning were the explanation of why scrambled input keeps showing transfer learning capabilities an experiment training with the original data while tested on scrambled or vice versa is needed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#1|9", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "Why didn't the authors try that configuration?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#1|11", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "If results cannot be interpreted, authors should find other experiments to support (or refute) their claims.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#1|12", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "Despite the interesting results and large number of experiments, the inconclusive results make unclear the scientific contribution of this research piece.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#1|14", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "_Paper's title is somehow misleading \"Cross-Domain Knowledge Transfer for Pretrained Models\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#1|15", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "Authors use just BERT as pretrained model and a LSTM with GloVe embeddings as part of the experimentation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#1|16", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "_ The introduction and the Conclusions are inconsistent: \"we evaluate LSTMs using GloVe embeddings, BERT, and baseline models\" vs \"we take an English pretrained BERT off-the-shelf and fine-tune it with a scrambled English dataset\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#1|17", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "_table 4 is shown on page 6, but referenced on page 8.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#1|18", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "Please bring the table closer to its reference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|4", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "However, the authors fail to convince me why studying this distributional shift would bring new and relevant insights to the table.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|5", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "I think for two reasons:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|6", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "(a) The shift is not interesting in the sense that it is related to cross-domain or cross-lingual or cross-time shifts, or any other shifts that we observe in the wild.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|7", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "(b) The shift does not seem to add value over other synthetic shifts that have already been studied, e.g., word-level shuffling, character-level shuffling, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|8", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "You could have studied many other scrambling functions, so why this one?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|9", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "That said, I liked §7.3, which nicely baselines the scrambled model performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|10", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "§7.5 is also nice, except that it’s hard to see why any network would learn to reassociate words in the context of learning a particular task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|11", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "(Learning to reassociate words is not necessary to solve most tasks. Maybe in machine translation?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|13", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "The paper is overall well-written, but I found the use of the word ‘scrambling’ for a left-to-right in-order replacement of words, misleading.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|14", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "In related work, scrambling refers to scrambling of word order.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|15", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "I would also have liked to see a stronger motivation for focusing on frequency.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|18", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "I’m a little worried about community-wide overfitting of our intuitions about language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|19", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "There’s the popular alternatives (RoBERTa, GPT-2, t5, etc.), but also a lot of faster, fairer, more interpretable alternatives.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|20", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "(ii) The two sequential labelling tasks are very similar.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|21", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "POS is a give-away of NER.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|23", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "(iv) Using only a small subset of GLUE tasks looks like cherry-picking.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#2|24", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "(v) It is unclear how sensitive the results for each architecture are to hyper-parameter changes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#3|10", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "However, I want to see some measures of intrinsic evaluation related to the scrambled datasets as well (e.g., perplexity).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#3|13", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "Furthermore, could you every tried replacing word using ontology or a collection of semantically related words?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#3|14", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "Please add similar experiment to show that scrambling with synonyms and antonym have different effects (or positive vs negative replacements).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|9", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "Weakness: 2.I'm not convinced by the results about BERT in S7.3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|10", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "The confounder factor -- pre-training, is not considered.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|11", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "For instance, for the pre-trained BERT, the word frequencies information is already captured in the parameters (along with many other features that also hold for scrambled text).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|12", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "Such that when fine-tuned on scrambled data (word frequency preserved), it's not surprising that pre-trained BERT can clearly perform better.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|13", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "And the reason why LSTM is more robust is that the model size is very small such that it can relearn all weights.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|14", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "We can observe the same trend for 1-layer-BERT on classification tasks, the gaps are generally smaller than other model variants.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|15", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "Some claims in the paper can be further substantiated if the author can experiment with a BERT that is pre-trained on scrambled English.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|16", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "I also have some doubts about S6.2, in which the author claim that the bad performance is because sequence labeling tasks are more likely to rely on word identities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|17", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "However, as explored in Hewitt & Liang, those labeling supervisions (although in POS tagging) can be easily learned by fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|18", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "Can authors find a better explanation for the poor performance on labeling tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|19", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "The method proposed in this paper is widely used in many other studies, e.g., in adversarial attacking.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc19|review#4|22", "source": "rc", "split": "test", "paper_id": "rc19", "comment": "The title is about pre-trained models, but only BERT is discussed throughout the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|7", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The idea is well motivated, both wrt the novelty prompt generation and the contrastive confidence loss.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|8", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Empirical results are overall positive, with reasonable baselines and comparisons and some ablation studies.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|11", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "There are some details that are not very clear", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|12", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "S4.3 For contrastive baseline, I'm not sure how to interpret \"pairwise distant but internally compact\" in the context here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|13", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Are there novel classes and examples involved?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|14", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Might be good to clarify this question for the baselines here?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|15", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "S5.1 it was a bit difficult to follow the various comparisons made in ref to Table 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|16", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "E.g., I had a hard time figuring out what is being compared at \"This accuracy decrease is smaller\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|17", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "S5.2 \"To isolate the factors...\" what are the factors being isolated here and how is the isolation achieved by Table 2?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|18", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "S5.3 Generator size is concerned only with the example generator, while the novel-label generator is always GPT3?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|19", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "What happens if one generates novel labels with smaller models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|20", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "There's not much error analysis other than offhand comments with limited support.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|21", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "For example, when discussing vanilla outperforming OE+NP (\"We attribute this behavior to the noisiness of generation: some novelty prompted examples are similar to ID examples\"), although I intuitively agree with this sentiment, it would be great to have some empirical support by quantifying example similarity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|24", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Use of GPT3 might limit reproducibility but it won't be a permanent problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#1|27", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Publication would benefit the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|6", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Strength: Generating OOD examples with huge PLMs is an interesting idea.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|10", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Clarity: some important details are missing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|11", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Novelty: It brings some new ideas to OOD detection and classification.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|12", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Reproducibility: Others can reproduce the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|15", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The interesting part is novelty prompting in which examples that are semantically novel and bear some resemblance to ID examples are generated with prompts using huge language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|16", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "However, I have some concerns regarding to the presentation and the empirical results: (1) CCL+few shot data augmentation: the description is \"prompt with only an instruction and zero or more ID training examples\", then how many ID training examples are involved in CCL+few shot?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|17", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "What is the difference between the proposed method?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|18", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "(2) what is the classification model?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|19", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "(3) Looking at Table 1, I am curious why OE+novelty prompting performs such badly on TREC-10.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|20", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "If it is sensitive to noise, then why it renders descent results on the other three datasets?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|21", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "This is critical, as the baseline outperforms the proposed method on two of the four datasets but loses finally on the \"average score\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|22", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "(4) the improvement over CCL+few-shot is marginal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#2|23", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "That's why I am curious about the difference between the proposed method and CCL+few-shot.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|6", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "I appreciated the oracle experiments which frame how much potential room for improvement there may be in follow up work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|7", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The evaluation metrics made sense.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|8", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The authors report averages over a number of runs, yielding confidence intervals.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|9", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The biggest negative of the paper, in my view, is that the performance gains (particularly of the new loss function) are marginal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|10", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The new loss function outperforms OE mostly because OE seems to fail on TREC-10 (else, it's competitive).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|11", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The reason for OE failing is \"attributed to noisiness in generation\" of the OOD data --- this causes some concern, because while the new loss function seems to overcome this noise in the other dataset cases, I would be a bit nervous using it with any loss function given the massive performance drop on TREC-10: is it worth the risk of involving GPT-3 for a few AUC points?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|12", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The new data augmentation, while resulting in performance improvements over kFolden and Contrastive, also has more dependencies, i.e., access to GPT3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|13", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "In more detail, a few technical concerns:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|14", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "All of the out-of-label-distribution examples are synthetic, whereas, all of the in-label-distribution examples are (presumably) human written from the original corpus.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|15", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Are models implicitly just doing a \"human-authored\" vs. \"machine-authored\" classification task?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|16", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "This question is partly addressed --- given that the test sets are human-written, the method isn't only doing machine vs. human.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|17", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "But, I'd suggest to add an additional experiment that goes the other way: that is, an experiment where, at test time, the out-of-label-distribution examples are human-written (from the corpora), and the in-label-distribution examples are autogenerated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|18", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The authors propose a new loss function in eq.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|19", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "1 that goes beyond the uniform distribution loss of Hendrycks et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|21", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "While I can somewhat qualitatively see why this loss might be different than the uniform loss, the authors argue that it works better because it enables \"some novel examples can be more novel than others.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|22", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Can this claim be validated by examining the logits of the OE model vs. this one?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|23", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The uniform loss presumably (in a soft way) also allows some examples to be more novel than others.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|24", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The empirical impact of this loss function is also not entirely clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|25", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "OE seems to do badly on TREC-10 for some reason, but it generally seems to match (or sometimes outperform) the new proposed loss on the other corpora.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|26", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "I appreciated the oracle experiments --- but, somewhat inexplicably, OE seems to do much better with the gold data available vs. CCL.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|27", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The reason for this is not explored deeply --- why is that?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|28", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "All experiments are conducted with BERT-Base, which is a 4 year old model at this point.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|30", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The authors method can only work with text classification tasks, which is great, but --- some of the competing methods are more general, and the Wikitext + CCL method (which simulates the more general setting of not having GPT-3 but only an OOD set) only marginally outperforms prior solutions (kFolden, for example, doesn't even depend on the external set).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|32", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "see above for detail.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|34", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Eq 1 has a missing parenthesis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|35", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "Table 1's caption: \"Novlty\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|39", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "But, I have some doubts that the performance gains of the new method (a few AUC points) will entice folks to undertake the suggested protocol in practice, esp.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#3|40", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "given the GPT-3 dependence and also, the fact that choosing a slightly different loss function can cause catastrophic failure (e.g., on TREC-10).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#4|6", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The idea to use large pretrained generative model to generate OOD example is inspiring.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#4|7", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The experimental results are strong.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#4|9", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "While the results seems promising, I'm curious whether the choice of pretrained model (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#4|10", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "replace GPT-3 with GPT-2) impact the performance much.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#4|11", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "If the proposed method only works well on very large pretrained generative models, the computational cost could be extremely high, since the authors generate 100k OOD samples.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc20|review#4|12", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The proposed method seems hard to be extended to other tasks like sequence labeling and other domains like computer vision.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc20|review#4|14", "source": "rc", "split": "test", "paper_id": "rc20", "comment": "The paper is easy to follow, and the idea to use large pretrained generative model to generate OOD example is novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|4", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The paper is generally easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|5", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Designs of the pretraining objectives are reasonable and empirically effective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|6", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Experiments are solid and convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|8", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The paper demonstrates a good practice on how to pretrain a language model for a specific domain by injecting a variety of structured and unstructured domain knowledge.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|9", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Experiments verify that the knowledge-aware pretraining is indeed effective when you want to boost the performances of the tasks in specific domain.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|10", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The experiments are well designed and quite comprehensive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|12", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The paper is generally lack of novelty and inspiration, however.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|14", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "It is not surprising that when pretraining on the large amount of domain-specific data, the performance on the downstream tasks will improve.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|16", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "What are the principles of designing pretrianing objectives with regard to domain knowledge and is there any way to evaluate them without need of expensive pretraining and finetuning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|17", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The presentation of paper could be improved, especially for section 3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|18", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "For example, from the notation it is unclear which training objectives are defined on encoder and which are on both encoder and decoder.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|19", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Also in figure 1, there seems no clear distinction between the encoder and the decoder.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|21", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "How the five objectives are weighted in the pretraining?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|22", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "What are the “knowledge-tokens” and “non-knowledge” tokens in KMLM and KMS2S?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#1|23", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Are the results sensitive to them?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|11", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Integrating knowledge in transformer-based model is a topic of interest", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|12", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The experimental evaluation is solid.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|13", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The authors compare with previous works, and present sensible ablation results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|15", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Some details of the approach are under-specified/formulized.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|16", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The results may therefore not be reproducible. (see comments)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|17", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The paper is not self-contained (see comments)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|20", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The major difference from BERT is that our KMLM prioritizes knowledge tokens - what are the \"knowledge tokens\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|21", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "is it based on lexical match?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|22", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "does it only include \"content\" words (with no prepositions, for example)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|23", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "is ambiguity accounted for?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|24", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "is there a preliminary step of entity linking?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|25", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "What is a product \"aspect\" exactly?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|26", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "how are \"attributes\" and `\"features\" different?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|27", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The baseline method JAVE is missing a citation", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|28", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Table 1: why not report also precision and recall, in addition to F1?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|29", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "there seems to be enough space.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|30", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Table 2: note that the bold-faced results are not the best for all columns.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|31", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The evaluation tasks are very briefly described.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|32", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Perhaps examples can be given?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|33", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "with some explanation of how the proposed approach helps?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|34", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Some terms are under-specified: e.g., `\"intents\" in Sec. 4.2.3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|35", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "For the retrieval-based approach, we concatenate the dialogue context and use [SEP] token to separate context and response.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#2|36", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The [CLS] representation is fed into the output layer for classification.' - what is the context?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#3|5", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The authors present K-PLUG that pre-trained on Chinese e-commerce data, which contains approximately 25 million textual product descriptions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#3|6", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The authors evaluate three different downstream tasks, e-commerce KB completion, abstractive product summarization, and multi-turn dialogue, and show good results compared to its baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#3|8", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The paper is not easy-to-follow with too many abbreviations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#3|9", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The main concern of this work is: the comparison to existing works is too weak.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#3|10", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "As mentioned in Section 2.2, there are many related works also working on injecting knowledge or retrieving knowledge to improve language model pretraining.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#3|11", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "For example, KEPLER, SKEP, SenseBERT, and K-ADAPTER.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#3|12", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "However, the authors do not compare their training objectives/strategies to theirs, it is very hard for us to make a conclusion that which is better.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#3|13", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "It is not an apple-to-apple comparison.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#3|14", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The only baselines to compare within this paper is C-PLUG and E-PLUG.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#3|15", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The dataset used to pre-train the model contains expensive human annotation, which is not scalable. For example, the category labels (used for Product Entity Category Classification) and the summary (used for Product Entity Aspect Summary Generation). In other words, this kind of makes the pre-training process like a \"multi-task\" learning approach with human annotations, which in my opinion is hard to collect more data.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc21|review#3|16", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The ablation study results that are shown in Table 5 basically suggest that many of the used objective functions are not important and could be ignored as the difference with and without is really marginal (I am sure some of them are not significantly better and could be the same if the authors tune a bit the hyper-parameters).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|7", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The idea of domain-specific PLM is likely to be useful in practice but not widely explored.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|8", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Specifically, in the e-commerce field people often have heavy needs for automated text processing and also a large amount of data, making this line of work more useful in practice.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|9", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The pretraining objectives are properly designed, as they capture domain knowledge and have supervisions from the large-scale pretraining dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|10", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "The experiment results also justified the design.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|14", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Some downstream tasks are too similar to pretraining tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|15", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "For instance, Abstractive Product Summarization looks very similar to Product Entity Aspect Summary Generation (PEASG) in pretraining, given that the image modality is ignored.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|16", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "I understand that if pretraining tasks are similar to downstream tasks, the performance on downstream tasks will be better.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|17", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "However, such similarity limits the ability of the experiments on downstream tasks to show the usefulness of the pretrained model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|18", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "In other words, when the tasks are too similar, it becomes less clear whether the performance gain comes from the pretraining method or merely a larger training set.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|19", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Also, have you tried directly using the pretraining dataset in downstream tasks (it looks possible for E-commerce KB Completion and Abstractive Product Summarization)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|21", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "In Figure 1a, in the pretraining data aspect descriptions, the knowledge-base values and USPs are highlighted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|22", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "How are these labeled?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|23", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Are they directly available from the platform?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|24", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "For the pretraining dataset and downstream task datasets, each of them is said to be collected from \"a Chinese e-commerce platform\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|25", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Are they from the same or different platforms?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|26", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "It might be more clear to explicitly provide the platform names.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|27", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "In the formulation of KMLM and KMS2S (page 4) you mentioned \"knowledge tokens\", what are \"knowledge tokens\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|28", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Are they knowledge-base tokens and/or USP tokens?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|29", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "For multi-turn dialog, since it is highly user-oriented, it would be better to have human evaluations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc21|review#4|31", "source": "rc", "split": "test", "paper_id": "rc21", "comment": "Page 5, PEASG, Line 6: \"a aspect\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#1|8", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "It is not clear how this model helps in practice since it is achieving similar performance as other models like SpanBERT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#1|9", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "It is true that this model doesn’t introduce complexity while pre-training but pre-training is a one-time process and even if pre-training is slower, it does not affect downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#1|10", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "However, the proposed model introduces complexity in the downstream tasks and may affect training time in the downstream task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#1|11", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "The authors claim that the proposed model perform similar to models like SpanBERT but it is not clear if both models are improving on similar aspects.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#1|12", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "Some fine-grained analysis of results with these models might be insightful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#1|13", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "Another concern I have is that the proposed model can only work in the scenarios where sentence-level representation is required.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|5", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "The general idea is simple and easy to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|6", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "However, I have several concerns which are list as follows,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|7", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "Combining span-level information has been widely studied in the literature[1] (not cited).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|9", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "Besides, it is also worth discussing between Sec 5.2 and [1].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|10", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "The experiments are conducted on BERT-base and BERT-large models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|11", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "However, the two models are weak baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|13", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "As the paper focuses on training span representations, it is necessary to perform experiments on entity-based tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|14", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "The experiments setup may refer to the paper [2].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|15", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "The paper proposed to use the n-grams, it is also necessary to have a comparison with the pre-extracted spans using an off-the-shell chunker.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|16", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "I also have the following questions: If a pretrained chunker is utilized to extract chunks, what are the results like?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|17", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "In section 5.2, does the variant of n-grams affect the results?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|18", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "For these reasons, I do not recommend acceptance of this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#2|19", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "[1] Toshniwal et al. A Cross-Task Analysis of Text Span Representations, ACL RepL4NLP-2020 [2] Yamada et al. LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention, EMNLP 2020", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#3|9", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "Employing span-level information in PrLMs was proposed in previous works such as SpanBERT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#3|10", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "The paper presented a way that combines span-level information into the representations generated by PrLMs during the fine-tuning phase.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#3|11", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "However, as shown in Table 2, the proposed method is not better than SpanBERT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#3|12", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "So, what are the advantages of the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#3|13", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "As shown in Table 1, compared with the baselines, the proposed method does not bring remarkable promotion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#3|15", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "The subscript “j“ should be “1” in the second line of Formula (1).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#4|8", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "My concern is the significancy of the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#4|9", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "Since the improvement of the proposed method is marginal (~ 1% avg) and fine-tuning BERT on GLUE tasks might have high variance, authors should report the mean and std over multiple runs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc22|review#4|11", "source": "rc", "split": "test", "paper_id": "rc22", "comment": "If the model is pre-trained with span level information (e.g., spanBERT), will the proposed method outperform normal fine-tuning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|10", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The proposed NPPrompt method is simple and effective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|11", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "It only requires the label strings for doing zero-shot classification.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|12", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The empirical results show that it is on the same level as other methods that need more information about the labels.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|14", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The proposed method is pretty limited to the classification tasks, particularly for news classification, where the label names are informative and have many semantic relevant words in the vocabulary.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|15", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "It is also limited when the label names are multi-word expressions or phrases.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|16", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "It does not support the cases where the label names are not informative enough and additional label description is needed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|17", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The title and narrative can be misleading and an overclaim.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|18", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "As mentioned before, it is limited to masked LMs and the advantage is mainly for tasks like news classification.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|20", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "I suggest the authors extend the proposed method to multi-label classification settings where an example can have multiple acceptable news.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|21", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "A major limitation is the multi-word expression and out-of-vocabluary label names.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|22", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "How do you generalize the method to support that?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|23", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "Also, I suggest the authors try some multiple-choice QA problems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|25", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The paper is mostly clear and the novelty is incremental.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|26", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "I think the reproducibility should be okay since the method is pretty naive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|27", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "Can you show the kNN results for sentiment analysis tasks and NLI tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#1|28", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "I don't find them in the paper but they are quite important for qualitative analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#2|8", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "Proposed method is simple, intuitive and effective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#2|9", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "Conduct extensive experiments on a wide range of NLU tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#2|13", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "This paper is well-written.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#3|5", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The authors put significant effort on proving effectiveness of their method in a variety of NLP tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#3|6", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "However, I wanted to see significant test results to make sure that the improvements are not random.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#3|8", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The paper was well written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc23|review#3|9", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "I would require authors to add the github link for the code.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#1|9", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "The relative similarity of source language representations is interesting and worth exploring further.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#1|12", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "The paper is hard to follow and would benefit from a clearer focus rather than the broad range of topics covered here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#1|14", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "It is difficult to understand what the methods/terms (the information theoretic measure used, double descent) are - little time is spent explaining these.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#1|15", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "Double descent is discussed in the paper but it is still made not clear why this is relevant in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#1|16", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "Several portions of text are repeated - with some editing, space can be made to discuss concepts important to the paper", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#1|17", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "The authors make recommendations for modeling (Eg. using char level or byte level models for certain models - which have been extensively studied for this): this is not followed up with any concrete results on translation/downstream tasks or pointing out relevant work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|5", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "However, as much as I agree with some of the final conclusions, the soundness of the experiments appears to be in question.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|6", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "My main concerns are about the additional experiments with Chinese.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|7", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "The authors claimed that \"On the character level, target language ZH (ZHtrg) shows a different learning pattern throughout.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|8", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "There are two types of character-level representations used: Wubi and Pinyin.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|9", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "Wubi was originally invented for professional typesetters so that they can type fast.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|10", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "The segmentation may not be correlated with the meaning of the word at all, as claimed by the papers cited by the authors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|11", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "Pinyin, on the other hand, is highly ambiguous.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|12", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "One pinyin may representation dozens of words and the authors did not take tones into considerations at all.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|13", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "The author also mentioned that \"After filtering length to 300 characters maximum per line in parallel for the 6 languages, we made 3 subsets of the data with 1 million lines each\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|14", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "Each language carries meaning differently and the information density is drastically different.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|15", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "300 characters in Chinese carry much more information than 300 characters in English.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|16", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "This is an unfair comparison.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#2|17", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "It would make a lot more sense if the authors treat each language differently because of their orthographic differences.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|2", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "To do this, the authors perform controlled experiments by modeling text from parallel data from 6 typologically diverse languages.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|3", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "They pair the languages and perform experiments in 30 directions with Transformers, and compare 3 different unit representations: characters, bytes, and word-level (BPE).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|4", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "I appreciate the authors' effort for their systematic controlled experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|5", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "However, I'm leaning towards rejecting this paper since I think some of the claims made in the paper are too strong and not really backed up by their experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|7", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "The term \"Conditional-Language-Model\" can be misleading, since this paper model a target language conditioned on a source language, so more like in a machine translation setting rather than standard language modeling setting where you can also condition on the previous history.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|8", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "I'm also not sure if comparing perplexity by conditioning on another different language (source language) is correct.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|10", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "At the end of Section 2, the authors mention about \"generalizations\", but I couldn't really find any discussion about this in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|11", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "Maybe this can be clarified?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|12", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "I found that claiming script bias in character models is too strong, if the experiment only shows bias in ZH (and this is expected since its character-level has different notion with languages with Latin script).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|13", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "Byte-level: I found that there is still some \"erraticity\" in Figure 3(b) especially when the data size increases (which is more practical in real world application), so this is not entirely resolved?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|16", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "From characters to words to in between: Do we capture morphology? Clara Vania and Adam Lopez. ACL 2017.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#3|17", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "Multilingual Part-of-Speech Tagging with Bidirectional Long Short-Term Memory Models and Auxiliary Loss. Barbara Plank, Anders Søgaard and Yoav Goldberg. ACL 2016.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#4|3", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "This is a thorough paper with a large number of experiments and with interesting conclusions that nicely generalize the low level patterns observed in the experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#4|4", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "These conclusions are likely to be useful for the research community as part of its on going investigation of language transfer and the transformer model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#4|6", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "The language of the paper is often very complicated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#4|7", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "Just as a couple of examples: It was very hard for me to follow the abstract, the first paragraph, the (very long) sentence that start with \"in order to eliminate\" (1.1), item 3 in the list of contributions and this is just a partial list.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#4|8", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "I ask that if the paper is accepted the authors will try to improve this aspect.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#4|9", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "The writing is often over pedagogical and I often got the feeling that the authors try to educate their readers (but not in the positive sense of the word).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#4|10", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "I would try to avoid this style.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#4|11", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "This work seems highly relevant to the following paper:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#4|12", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "Towards Zero-shot Language Modeling. Edoardo Maria Ponti, Ivan Vulic,Ryan Cotterell, Roi Reichart and Anna Korhonen. EMNLP 2019", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc24|review#4|13", "source": "rc", "split": "test", "paper_id": "rc24", "comment": "I think the discussion parts can gain form comparing the conclusions of the two papers, when relevant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#1|10", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Regarding weakness one of the first comments is the lack of other approaches that treat the same problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#1|11", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The authors cite several related works in related cases (OOD detections, anomalies etc) but do not do any comparisons.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#1|12", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "I mean, the fact that they enhance few SOTA approaches is already good but I cannot assess how this pushes the current state-of-the-art methods in the specific problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#1|13", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Lack of statistical tests to evaluate significance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#1|14", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "In some cases improvements are quite small.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#1|15", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Maybe you could add them?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#1|16", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "I am curious if any other baseline like for example to consider the universum class as a side task would work here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#1|23", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Based on the presented results the proposed approach seems to help improving the models that it extends.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|13", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The first major concern is regarding the correctness of the motivation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|14", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "While the motivation for special treatments of the universum class is clear and intuitive, it's not convincing to me whether the limitations of conventional multi-class classification as pointed out in this paper actually matter that much.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|15", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Figure 1 illustrates that the decision boundaries learned by conventional classification models are \"open\", which makes it hard to distribute the universum class instances accurately.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|16", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "However, I doubt if this is really the case -- in the context of using deep neural nets (DNNs) as the feature encoder, the data representations  (i.e., $\\boldsymbol{h}$) and decision boundaries ($f(\\boldsymbol{h})$) are jointly learned, so even if the universum class instances have quite different semantics, it should not be hard for DNNs to map them into one compact cluster in the last layer's representation space.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|17", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "In other words, the limitations illustrated in Figure 1 (c) could be a problem for simple models like logistic regression or SVM, but do not seem to be a challenge for DNNs which can learn arbitrary non-linear functions for data representations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|18", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The second major concern is the additional parameters and hyperparameters introduced by the proposed algorithm.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|19", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Firstly, the GMM parameters need to be fit using target-class data -- when the amount of labeled data is small for some classes (e.g. few-shot learning or imbalanced learning), the GMM obtained may not be accurate/stable, and this limitation has not been discussed in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|20", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Secondly, a threshold value ($\\xi$) needs to be set for each class.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|21", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "It seems to be a hyperparameter that cannot be automatically learned, but the paper also didn't discuss how it can be appropriately set.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|22", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The authors also didn't show how sensitive the model is to these threshold values.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|23", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "This is quite concerning because it's impossible to manually set threshold values for each class, especially when the number of classes is large (e.g., in fine-grained NER/RE).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|24", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The paper is unclear in many technical details.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|25", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "For example, Section 3.3 mentions using N-pair loss for pretraining, but does not explain the motivation/reason.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|26", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Table 4 seems to show that the pretraining step is not useful, making the purpose for pretraining very confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|27", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Figure 3 studies the impact of last layer dimension and reports that the optimal value is 7.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|28", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "This is quite confusing because the common dimensionality of pretrained models are over 700; a dimension of 7 seems way too small for the last layer dimension.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|29", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Did you use another linear layer on top of the last BERT/RoBERTa encoder layer to convert the high-dimensional vector into lower-dimenisonal ones (e.g., 7-10 dimensions) before performing classification?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|30", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The performance improvements of the proposed method over conventional classification setting are quite marginal (in most cases less than 1 point).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|31", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The paper also didn't report error bars/stds or significant testing results to validate whether the improvements are meaningful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|34", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "I'd like to thank the authors for their provided responses and paper updates to address my concerns raised above.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|36", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "However, in the authors' rebuttal and updated paper version, many details have been further clarified, and I found them largely strengthened the contribution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|38", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The difficulty of learning with heterogeneous samples in the universum class has been clarified.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|39", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "I see that the miscellaneous patterns from the universum class that do not appear in training but in test time could be an issue for traditional classification models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|40", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "However, I feel that some concrete case studies would help (e.g., some embedding plots of the actually learned classification boundaries by the proposed method); Figure 1 does not seem to be the actual plots but looks more like conceptual illustrations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|41", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The learning process of GMM (especially how the threshold values are learned) and its limitations (e.g., hard to adapt to few-shot learning) have been better articulated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|42", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The motivation for pretraining looks clearer to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|43", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Glad to see the added significant tests and the new OECC baseline.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|45", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "However, I feel that the application scenarios of the method are somewhat narrow (i.e., there are many classification problems in NLP that do not have universum classes at all) and the method also has inherent limitations of not being applicable to low-data regimes (e.g., few-shot learning).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|46", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Although it does not appear to be a very strong and suitable paper for ICLR, I also don't want to gatekeep it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|47", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Hence, I'm recommending a weak accept.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|49", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Clarity: The motivation is clear, but many technical details are not explained sufficiently/clearly (see cons #3 above).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|50", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Quality: The major assumption made by the paper regarding why/how conventional treatments of the universum class could be an issue is not convincing (see cons #1 above)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|54", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "There are also concerns regarding the stability/generalization ability of the model due to the additionally introduced parameters and hyperparameters, and the lack of studies for these.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|55", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Many technical details have not been sufficiently discussed and thus remain confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#2|56", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The empirical advantage over standard classification baseline is quite marginal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|7", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "There is little analysis to show how the learned boundaries can better model class boundaries.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|8", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "A detailed analysis of precision/recall would be useful, and whether the approach improves multi-class decision boundaries, or only the boundary with the negative class.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|9", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The proposed approach emphasises on misclassified positive samples, which could have a similar effect on work that optimises F1-measures, see e.g., [1].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|11", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The authors have addressed some of my concerns mentioned above.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|12", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Hence I have increased the score to 6.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|13", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "[1] Ye, N., Chai, K. M., Lee, W. S., & Chieu, H. L. (2012). Optimizing F-measures: A tale of two approaches. In Proceedings of the 29th International Conference on Machine Learning (pp. 289-296). Omnipress.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|16", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "I am unable to follow the justification in section 3.5 that arrives at the proposed p(h_k,U) in Equation (4).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|17", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "In section 3.6, the boundary learning loss is optimised over wrongly classified samples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|18", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "Are these misclassified examples of the training set, or validation set?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|19", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "An analysis of how this loss function improves F1 while also optimising for cross-entropy could be useful to help understand the benefits of the approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|27", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "I would not recommend acceptance as I think the arguments made by the authors are not supported by sufficient analysis of the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc25|review#3|28", "source": "rc", "split": "test", "paper_id": "rc25", "comment": "The authors could also compare against approaches that directly optimises for f1-measures.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#1|7", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The motivation is clear and reasonable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#1|8", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The experiments support the claim and seem to be promising", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#1|10", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "To test the generalizability, I suggest to test on GLUE tasks as well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#1|11", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "It's not clear what size of models is using for the ablation study in Table 2", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#1|12", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "What will happen if we do not shuffle the sentences?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc26|review#1|14", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The proposed method is clearly described.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#1|16", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "They propose additional self-supervised loss to improve the language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|5", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Well-written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|7", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Section 4.1.1 states, \"To ensure fair comparisons, all models at the same scale had almost the same computational cost and number of parameters in both training and inference.\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|8", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "This is a detail people often forget, so I am happy to see that the paper takes computation into account.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|10", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The introduction argues that \"Without text structure, a text becomes a long continuous word sequence, which is hard to read and makes it difficult to identify key concepts and logical relationships.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|11", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "While this is probably, intuitively, true for humans, the paper uses the statement to argue against prior unsupervised sentence tasks and projects intuitive understanding of human attention and reading abilities onto machines (i.e., language models) without valid proof.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|12", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "This is a logical fallacy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|13", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "I do not believe that this is a major weakness for the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|14", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The paragraph just needs to be rewritten to avoid the incorrect argument.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|16", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Section 4.1.1 cites \"the sentence splitter of NLTK (Bird et al., 2009) for the use of TSP\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|17", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "NLTK's sentence splitter is Punkt (Kiss and Strunk, 2006), see also https://www.nltk.org/_modules/nltk/tokenize/punkt.html.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|18", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The paper should cite the Punkt paper together with the NLTK citation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|19", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Giving credit to the method's authors is not only fair, but it explicitly names the sentence splitter used, thus improving clarity for readers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|21", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Well-written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|23", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "See the point in Weaknesses and the Punkt sentece splitter.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#2|24", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The paper contains all the hyperparameters needed to reproduce the experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#3|6", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The paper is clearly stated and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#3|7", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The performance improves significantly in some of the SuperGLUE downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#3|9", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The hierarchical relationship labels are not convincing enough. It is even difficult for humans to predict whether two sentences are from the same document. Enforcing the model to learn the \"same document\" label might cause the overfitting of the sentence representations.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc26|review#3|10", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The ablation study also shows that order information is more important than hierarchical structure information.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#3|11", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The empirical results are not uniformly improved with the proposed TSP methods on the downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#3|13", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The clarity is good.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#3|14", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The quality and novelty are fair.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#3|15", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The experiments seem reproducible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|11", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The paper is clearly written and well-organized.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|12", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The proposed pretraining task TSP is shown to outperform the compared sentence-level pretraining objectives (NSP, SOP and SSO) on SuperGLUE.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|14", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The novelty of the method is quite limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|15", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The proposed objective TSP is essentially a combined (and slightly extended) version of NSP and SOP -- the \"hierarchical relation\" classification is extending NSP by further classifying the \"non-neighbor\" case into \"same-paragraph\" and \"same-document\"; the \"order relation\" classification is equivalent to SOP.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|16", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "In this sense, the proposed task is not new and should be compared with multi-task training that uses both NSP and SOP (this seems to be an important baseline that is missing from comparisons).", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|17", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The paper omitted discussions and comparisons with an important family of sentence-level pretraining tasks that use contrastive learning. For example, COCO-LM and CLEAR demonstrated that contrastive learning-based sequence-level tasks can effectively improve NLU task fine-tuning performance, and they can also be combined with word-level pretraining tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|18", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The empirical evaluation is quite problematic.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|19", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "My major concern comes from the fact that the paper does not follow the conventional pretraining/downstream task settings that have been extensively used by previous studies.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|20", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "(1) The pretraining corpus used in this paper is OpenWebText whereas previous pretraining methods use either Wikipedia + BookCorpus (in BERT) or further expanded corpora that include a wide variety of datasets besides OpenWebText (in RoBERTa).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|21", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The common pretraining corpora choice could be either of the two above.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|22", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Even if BookCorpus is not publicly available, Wikipedia should at least be included as a formal language pretraining source.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|23", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "If the pretraining experiments are conducted only on OpenWebText, it's very unclear whether the observed performance gain of TSP can be generalized to common/larger pretraining corpora.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|24", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "(2) Most pretrained language models (e.g., BERT, RoBERTa, ELECTRA) have been evaluated on GLUE whereas this paper only reports SuperGLUE results. This makes it hard to compare with previous pretrained models. While it is good to have SuperGLUE results, I believe it's important to conduct experiments on GLUE as well to facilitate easy comparisons with prior work. (3) The evaluation results only have the best performing results on the test set but not median results across all tasks (Table 6 only has averaged scores). This is problematic because some tasks have very small test sets and the scores could vary a lot. The common practice is to report median results over several random seeds on dev sets of all tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|26", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Meng, Yu et al. “COCO-LM: Correcting and Contrasting Text Sequences for Language Model Pretraining.” NeurIPS (2021).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|27", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Wu, Zhuofeng et al. “CLEAR: Contrastive Learning for Sentence Representation.” ArXiv abs/2012.15466 (2020).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|28", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "I thank the authors for providing responses to my raised concerns.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|29", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Unfortunately, I'm not convinced by the authors' arguments regarding my major concerns.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|31", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Regarding novelty: I never implied in my review that a method needs to be complicated to be novel; instead, simple methods can be novel as long as they provide new perspectives/insights.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|32", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "My critique regarding the novelty of the paper is mainly due to the fact that the proposed objective TSP is very similar to previously proposed sentence-level objectives (NSP and SOP) in formulation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|33", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The major difference between TSP and NSP+SOP, according to the authors' response, is that TSP considers paragraph-level relationships.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|34", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "However, the motivation for this idea is very vague -- how can one accurately/unambiguously define the real difference between a sentence and a paragraph, and the difference between a paragraph and a document?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|35", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "For example, in novels, it's very common that one or two sentences can be a paragraph; other times, a paragraph can also be an entire document.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|36", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "If the distinction between sentences, paragraphs and documents is even hard to tell for humans, then I don't see how the so-called \"paragraph-level relationships\" newly introduced in TSP can actually make a big difference compared to a simple combination of previous methods (NSP+SOP).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|37", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Regarding the comparisons to contrastive learning approaches: I'm not fully convinced by the argument that the comparisons between contrastive learning objectives and the proposed TSP objective can be omitted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|38", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "I understand that the two sets of methods are based on different principles, but they are indeed for the same goal -- improving the sequence-level representations of PLMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|39", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Therefore, unless TSP has other usages/benefits that contrastive learning fails to bring, I don't see why they cannot be fairly compared to each other.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|40", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Such comparisons do not exist in prior work probably because it is quite well known that NSP/SOP-style methods do not really work well (also shown in this paper), and many PLMs simply ended up not using any sequence-level objectives (e.g., RoBERTa, ELECTRA), so the contrastive learning papers didn't compare with NSP/SOP (of course, I do hope to see those results in the contrastive learning papers).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|41", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "The empirical gain of TSP does not seem convincing to me: Although TSP provides significant gains on GLUE/SuperGLUE average score, the gains are mainly from small/unstable tasks, like CoLA, MRPC, RTE and STS-B.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|42", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Fine-tuning on those tasks are notoriously unstable and is subject to high variance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|43", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "On large tasks like MNLI and SST-2 (note that they are more reliable than the previous tasks; RoBERTa paper uses these two tasks intead of the average GLUE score for ablation studies), the TSP objective barely provides any gain and even worsens the result.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|44", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Therefore, I suspect that the performance gain of TSP is simply due to randomness in fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|45", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Regarding the choice of pretraining corpus (OpenWebText): I'm indeed aware that OpenWebText is larger than Wikipedia, but Wikipedia is still a more standard choice for pretraining corpus if only one corpus is used (e.g., base model exploration in BERT, RoBERTa, ELECTRA all used Wikipedia + BookCorpus).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|46", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "This is probably because Wikipedia's text quality is arguably higher than OpenWebText which contains many short/casual/informal texts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|47", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Therefore, I'd suggest the authors follow the standard protocol for pretraining; otherwise, the findings and conclusions may not be reliable generalize well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|48", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "This is not my major concern about the paper though.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|50", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Clarity: The paper is clearly written and well-organized.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|51", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Quality: The contribution of the paper is not convincing due to the problematic empirical evaluation (see cons above).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc26|review#4|52", "source": "rc", "split": "test", "paper_id": "rc26", "comment": "Novelty: The novelty is quite limited. The proposed objective TSP is essentially a combined version of previous sentence-level tasks (see cons above).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|6", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "This is an interesting way to use large pre-trained language models such as GPT-3 to solve generative tasks like Machine Translation", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|7", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Combines multiple recipes from various sources (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|8", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "zero shot MT, few shot MT, back-translation) and shows how to apply them to a standard left-to-right pre-trained LM to achieve impressive translation results", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|10", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "While the results are interesting and illustrate the flexibility of a standard transformer language model, the key ingredient of this approach relies on having a language model that has seen a very large and diverse dataset that very likely includes some translation examples (for zero-shot translation to work).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|11", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Using such a language model to boot-strap self-training (called distillation in this paper) followed by back-translation is interesting, but should be grounded in techniques that have access to the same kind of large and diverse dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|12", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "One clear example is bi-text mining (e.g. see https://arxiv.org/abs/1911.04944) where one can simply mine a bunch of parallel text from a similar kind of dataset and then train a standard encoder-decoder machine translation model (with self-training / backtranslation if so desired).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|13", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "It is not clear what the advantage of the proposed method over such an approach would be except as a curious application of existing pre-trained models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|14", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Could the authors elaborate on this point and, if they agree, compare to the methods described above.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|16", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "If so, this would be great to discuss in related work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|18", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "While this work is an interesting way to explore the power of large pre-trained language models, I am not sure it is compared properly to methods (also starting with no parallel data) that may have access to the same kind of datasets used to train these language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|19", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "I am also not sure how this work compares to supervised and/or semi-supervised techniques using the same model sizes, or what is the state of the art for using transformer LMs with prompting for NMT (where parallel data is available).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#1|20", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Discussion of these points would be enlightening and may lead to updating my recommendation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#2|10", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Table 8 is also interesting: in this task we can only prepare monolingual corpora for each language/domain and only a few seed translations for prompts (if we didn't generate them from LMs).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#2|12", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Overall, the text is hard to read to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#2|13", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Many terms were introduced without definitions, the sentence frequently omitted important contexts, and syntax was often broken.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#2|16", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "The technique of model construction seems remarkable, and there is some concern about data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#2|17", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "The paper is somewhat hard to read due to syntactic/semantic issues, and I guess I possibly drops some context to assess the paper correctly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#2|18", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "I am not fully familiar with recent advances in so-called \"unsupervised translation\" and huge LMs, and I didn't judge if the proposed method is novel or not, and not provided comparison with existing studies.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|6", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "A new approach to construct UNMT with generative pre-trained language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|7", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "The proposed method achieves the new start-of-the-art unsupervised translation performance on the WMT14 English-French benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|9", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "This paper only verifies the effectiveness of the proposed method on the English-French language pair.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|10", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "It is better to conduct experiments on other language pairs, such as English-German, to make the result more reliable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|11", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Currently, the performance of UNMT model on distant language pair is very poor.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|12", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "It would be more exciting, if we also observe the significant improvement of the proposed method on some distant language pairs, such as English-Chinese and English-Japanese.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|14", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Overall, this paper is easy to follow and understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|15", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "I like this idea of constructing UNMT system with the GPT-X model and the proposed method is very interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|16", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Also, the experimental results on WMT14 English-French benchmark surprise me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|17", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "My main concern is the experiment settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|18", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "This paper merely verifies the effectiveness of the proposed method on the English-French language pair, making the experiment results not convincing enough.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|19", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "It is better to conduct experiments on other language pairs, such as English-German.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|20", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Previously, the performance of UNMT methods on distant language pair is very poor.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|21", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "As the proposed method replaces the bilingual dictionary with few-shot amplification, it seems to me that the proposed method may achieve better performance on distant language pairs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|22", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "I am looking forward to observing the unsupervised translation performance of the proposed method on some distant language pairs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|24", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "As the proposed method requires large-scale model to achieve the start-of-the-art performance, what is the training cost/time?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|27", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Unsupervised Neural Machine Translation with Weight Sharing. Yang et al., ACL 2018", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|28", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Phrase-Based & Neural Unsupervised Machine Translation. Lample et al., EMNLP 2018", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|29", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Unsupervised Neural Machine Translation with SMT as Posterior Regularization. Ren et al., AAAI 2019", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|30", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Extract and Edit: An Alternative to Back-Translation for Unsupervised Neural Machine Translation. Wu et al., NAACL 2019", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#3|31", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "An Effective Approach to Unsupervised Machine Translation. Artetxe et al., ACL 2019", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|7", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "State-of-the-art unsupervised translation results for both English-French and French-English.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|10", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Only one non-English language is considered.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|11", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "No low-resource language pairs considered, which is the realistic setting for unsupervised MT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|12", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "No in-depth discussion on the aspects which separate this work from other unsupervised MT works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|13", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "The authors claim that \"with sufficient scale, pre-training, and clever prompting, a single generative language model can implement the entire unsupervised neural machine translation pipeline, avoiding optimizations such as denoising autoencoding, auxiliary / adversarial losses in latent space, or ad-hoc bilingual dictionaries”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|14", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "While I can agree that removing auxiliary / adversarial losses or ad-hoc bilingual dictionaries are certainly steps in the right direction, current state-of-the-art unsupervised translation models have already taken these steps and do not rely on these tools.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|15", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "For example, MASS [6] pre-trains using only a denoising autoencoding objective and yet achieves strong results comparable to the ones found in this paper (accounting for model size) and additionally does not rely on clever prompting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|16", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "To be explicit, the MASS scores in Table 3 are comparable to the results of GPT-3 medium in Table 1, which brings into question if the gains found in this work are truly from the approach or simply from scaling.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|17", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "It thus appears that the main contributions of this work which separate it from MASS are 1) using a decoder-only model rather than the typical encoder-decoder architecture, 2) using language modeling as pre-training objective instead of a denoising autoencoding objective and 3) using an English-centric model rather than a model appropriately trained on both languages.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|18", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Of these, 1) is not particularly novel since GPT-3 has already demonstrated some initial translation performance for the language pairs in question.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|19", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Previous work [1] has also shown that the initial quality of the translation system for iterative back-translation has little effect on the final model's quality.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|20", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "These two findings make it unsurprising that the authors achieved strong unsupervised translation performance after back-translation, despite the relatively weak model that was used to generate the few-shot examples or to initialize the back-translation process.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|21", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Moreover, other contemporary work [7] has already shown that decoder-only models are sufficient to build strong translation models and provided a far more thorough analysis of this architectural bias than this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|23", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Models trained with masking strategies tend to struggle with zero-shot generation which hinders back-translation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|24", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "For example, mBART [4] relied on LangID filtering in the initial stages of their back-translation routine to overcome this issue.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|25", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "The authors additionally claim that \"[d]enoising autoencoding with a bilingual encoder can be viewed as a kind of latent bilingual lexicon induction, necessary for producing sufficiently aligned embeddings to kick-start back-translation.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|26", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "However, I see no reason why we can't view language modeling on multilingual data in the same light.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|27", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "The initial translation performance from GPT-3 already demonstrates that language model can serve as an acceptable substitute to denoising autoencoding for kick-starting back-translation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|28", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "With these thoughts in mind, we cannot consider this work’s successful removal of denoising autoencoding for training unsupervised translation models as a particularly difficult task nor a surprising finding but instead as more of a confirmation of an expected result.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|29", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "The English-centric nature of the model (point 3) is an interesting aspect, however it is understudied in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|30", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "The choice of language pair also hinders this study.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|31", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "For example, French is the second most populous language in GPT-3’s data [8].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|32", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Moreover, the original authors of GPT-3 have already demonstrated that GPT-3 can perform English < - - > French translation, further reducing the impressiveness of the result.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|33", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "It is also interesting to note that this English bias the authors comment on also affects traditional multilingual machine translation models (both supervised [2] and unsupervised [3]) due to the reliance on English-centric parallel data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|34", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "However, in this setting, this bias is even more pronounced since the English data does not arise from translation data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|35", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "It would be interesting to explore whether the overabundance of English monolingual data (or perhaps simply scale?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|36", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "is enough to make up for the scarcity of data available for some of the low-resource languages which appear in GPT-3’s corpus.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|37", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Similarly, it could be interesting to see if the scripts of the languages considered or their similarity with English is an important feature for this approach to succeed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|38", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "For example, do we expect GPT-3 to also excel at Hindi < - - > English translation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|39", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Or Gujarati < - - > English?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|40", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Both languages appear in GPT-3’s pre-training corpus, so one could test it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|41", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "The performance on these languages is crucial because the true value of unsupervised translation lies in low-resource language pairs, and their omission puts into question the validity of this method in realistic settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|42", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "The absence of true low-resource language pairs for evaluating unsupervised translation is unacceptable, especially when considering most of the recent literature in unsupervised translation includes such language pairs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|43", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Lastly, it would also be interesting to explore whether these models are able to generate more natural text, avoiding the pitfalls of translationese that traditionally affects models trained with supervised data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|45", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "What is [[TRANSLATE]] in Figure 1?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|46", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Is this just a stand-in for the prompt written in 4.1?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|47", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Many approaches in multilingual MT tend to discard the source language and use only the target language when biasing the model for translation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|48", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Yet it seems like you used the source language for both prompting GPT-3 and for the zero-shot format in the distillation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|49", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Leveraging the source language name could inhibit translation of code-mixed text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|50", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Did you experiment using only the target language instead in the prompt or attempted any investigation on whether including the source language name affected the translation performance?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|51", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Did you attempt this for non-English-centric language pairs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|53", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "[1] Artetxe et. al. Do all Roads Lead to Rome? Understanding the Role of Initialization in Iterative Back-Translation", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|54", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "[2] Arivazhagan et. al. Massively Multilingual Neural Machine Translation in the Wild: Findings and Challenges", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|55", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "[3] Garcia et.al. Harnessing Multilinguality in Unsupervised Machine Translation for Rare Languages", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|56", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "[4] Liu et.al.Multilingual Denoising Pre-training for Neural Machine Translation", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|57", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "[5] Siddhant et.al.Leveraging Monolingual Data with Self-Supervision for Multilingual Neural Machine Translation", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|58", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "[6] Song et.al.MASS: Masked Sequence to Sequence Pre-training for Language Generation", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|59", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "[7] Wang et.al.Language Models are Good Translators [8] https://github.com/openai/gpt-3/blob/master/dataset_statistics/languages_by_word_count.csv", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|62", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "The biggest flaw is that it is unclear whether anything presented in the work is truly novel or unknown to the machine translation community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc27|review#4|64", "source": "rc", "split": "test", "paper_id": "rc27", "comment": "Coupled with the limitations of only studying a single non-English language, no low-resource language pairs, as well no detailed discussion/experiments on the few aspects which distinguish it from other works (e.g.the overwhelming English-centric bias of GPT-3), I cannot recommend acceptance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|13", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Reveals problematic ingredients for likelihood-based answering.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|14", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Introduce the concept of MCSB and measure it by PPA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|15", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Concentrated results that significantly improves QA performance by using multiple-choice prompting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|19", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Mostly empirical analysis based on OpenAI playground.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|20", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Some major arguments are less supported.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|24", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "(Major concerns) How to make sure Codex model clearly outperforms Instruct model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|25", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "This is a critical question as the authors measures the main experiments (Table 2) that compare Multiple Chocie Prompting (MCP) and Cloze Prompting (CP) only with Codex model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|26", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "The capability to perform MCSB could be due to human feedback alignment by Reinforcement Learning rather than other points indicated by the authors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|27", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Are the PPA difference between Codex and Instruct (in Figure 2) statistically significant?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|28", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "While no statistical test has been provided, it seems not easy to decline null hypothesis that says the difference is a random effect.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|29", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Only Codex tested on OpenBookQA shows strong performance gain when using MCP, whereas Instruct outperforms Codex on the other two tasks in Table 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|30", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "More detailed experiments are necessary to convince how Codex achieve such higher accuracy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|32", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Any reason to choose OpenBookQA which also matters the performance of retriever?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|33", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Do you know how Codex model is exactly trained?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|34", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Codex model that you used could be first based on Instruct, then being further trained on code data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#1|35", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Equally likely, Codex model might perform it's own alignment similar to Instruct but based on the preference of generated codes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|6", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "This paper is well-written and easy to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|7", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "The authors propose a simple but effective prompting method, which outperforms previous CP methods, approaching or even surpassing SOTA performance.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|8", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Experimental results show that the MCQA ability of LLMs has been previously underestimated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|9", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "And there is a better way to prompt a single LLM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|10", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "The potential of multiple choices prompts can be further tapped.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|11", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Future work include prompt engineering is still promising.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|13", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "The novelty of this paper is limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|14", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Multiple choices prompting (MCP) has been used in other QA tasks, such as TruthfulQA and RACE.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|15", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Although the experimental results prove that the proposed multiple choices prompting (MCP) methods can outperform existing cloze prompting (CP) methods, the reasons behind it are still unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|16", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Since the authors have listed several problems within CP methods, I'm curious about whether these problems are all solved or avoided by their MCP methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|17", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "More analysis is needed to show this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|20", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "But it would be better to include more analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|22", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "A simple but effective prompting method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|23", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Novelty Novelty is limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|24", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Details can be found in the weakness part.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|26", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "The code is attached by the authors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|28", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Interesting paper but not good enough.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#2|29", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "It would be better to include more analysis on whether MCP can deal with these several problems that CP faces.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|4", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Strengths: The authors explain their approach well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|5", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "They also discuss the (somewhat surprising) variance between different models in their ability to separate the letter from the answer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|6", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "(They call this Multiple Choice Symbol Binding.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|7", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "The approach is evaluated on a wide range of (20) datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|8", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Weaknesses: The approach is not new, just discussed and evaluated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|9", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "The authors differentiate their suggested prompting from “prompt engineering”, which they seem to define as fine-tuning of prompts to increase model performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|10", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "However, I’m not convinced that these are fundamentally different, and would include research such as theirs in the general domain of prompt engineering.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|12", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "The paper is well written and I believe the experiments are verifiable with the given information, i.e.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|13", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "it should be possible to reproduce them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|14", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Regarding novelty, I am less convinced.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|15", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "The authors mention others having used the MCP approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|16", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "So the main addition here is the systematic discussion and wide range of experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc28|review#3|20", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "However, it is not novel: it is a fairly obvious way to prompt and has been tried before.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#1|8", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "I think it is a common belief that rehearsal-based is more robust against catastrophic forgetting, while regularization-method is more space-efficient in that it doesn't have to store examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#1|9", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "The fact that the last layer suffers from catastrophic forgetting is also not a surprising result, given that the lower layers are known to encode linguistic features and the upper layers encode task-specific features.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#2|8", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Moreover, the paper draws a conclusion that BERT is the most robust one and is a good option if a continual learning process is going to be conducted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#2|9", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Based on this, the authors provide a few analyses on BERT’s ‘secret’ for continual learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#2|10", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "However, compared with other pre-trained models, I don’t see that BERT is significantly better than others given the figures and tables.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#2|11", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "I feel from the figures and tables, BERT and other models look similar.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#2|12", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "The authors didn’t give a comprehensive explanation on how they read such information or a concrete quantitative comparison to support this claim.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#3|12", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "It would be really great to see how the insights after analysis can be used to improve performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#3|13", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "It's probably not absolutely required given the focus on probeing, but it would go a long way towards validating the insights.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#3|14", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Adjusting the numbers to achieve the 5/4/3/2 cuteness gets slightly in the way of understanding, unfortunately.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#3|15", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "The main reason for this is that the \"veins of CL methods\" has a different number over the course of the paper, which makes it hard to identify when a given list of N things is a list of the \"veins of CL methods\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#3|16", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Specifically, in the abstract and intro this number is 4, in section 2.3 this number is 3, in section 3.1 this number is 6, in Table 1 this number is 5 (two different sets of 5), in Figure 1 this number is 6, and in Figure 2 this number is 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#3|18", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Along the same lines, it would help to be consistent with the language around each set of N things.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#3|19", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "For example, the \"veins of CL methods\" are called at least \"veins\", \"schemes,\" and \"approaches\" at different points.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#3|21", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "The Figure here is too small to even attempt to parse.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#3|22", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "It would be very helpful to have a sentence that gives intuition about what's being measured with the accuracy metric.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#3|23", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "The definition is there, but it took me a second to realize that the intuitive idea is that it's measuring the accuracy of past tasks after the model has moved on to learning new tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|17", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Some plots seem to be a very small, and may be enlarged to use the 9th page.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|18", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Fig 2 plots should share a larger model legend (on the figure top or bottom), so the bars can become wider and easier to distinguish", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|19", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Fig 2 color could be made more distinguishable, especially since the plots are narrow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|20", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Tab 1 could underline the best non joint performance — makes it easier to glance", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|22", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "That Transformer layers, except the classification layer, do not adapt much during fine-tuning, is a known result (hence assumptions 1 and 2 in the paper), which should be cited — see BERTOLOGY Primer by Anna Rogers for references.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|23", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Here intermediate layers are shown to forget as well, so this is a nice new finding, that can be contrasted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|24", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Fig 3: I assume the figure color-scale is probe performance?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|25", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Also, the buffer size, e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|26", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "er-200 should be explained with an example.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|27", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Is it 200 er samples, 200 samples/ class?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|28", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "In section 5.1 it should be (re-)mentioned what the buffer size means.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|29", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Spelling errors (not all listed): using a TTS app/function makes it easy to find these", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|32", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "could be find ... can be found", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|33", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "tendency to forgetting ... to forget/ towards forgetting", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|34", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "reach the highest .. reaches", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|36", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "interferes ... with the PLMs ability to retrain _ representations", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|38", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Joint ... Joint multi-task training", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|39", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "from sketch ... from scratch", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|41", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "methods making sense ... make sense", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|42", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "The sentence \"(2) The classification layer clf is typically the most fragile of BERT, where continual learning learning methods making sense.\" is not understandable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|43", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Questions to the authors:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc29|review#4|54", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "I thus feel confident to recommend this paper for acceptance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#1|6", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "Does the system expects the gold document annotations to be available for training the retriever module?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#1|7", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "Section 2.2., talks about positive passages, this requires huge time and effort for annotation, so how does authors claim in abstract that the proposed technique does not require any language specific annotated training data?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc30|review#1|8", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "What happens in the scenario where a wikipedia page about a specific entity is not available in a particular language in which the question has been asked?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#1|9", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "Section 2.2 needs to have some more intuition and clarity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#1|10", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "What is intuition behind translating English answers to target language.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#1|11", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "Is there a mechanism to filter out spurious passages by the automatic passage labeling technique?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#1|12", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "How does the model learns cross lingual embeddings?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#1|13", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "do authors use a pre-trained model like m-bert?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#1|14", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "How does the proposed model results compare when compared with m-bert based methods?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#1|16", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "Limitations of the work has not been discussed by the authors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#1|17", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "Authors did promise to make the code publicly available to general public to enable further research in the direction ( societal impact)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#2|8", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "However, the overall idea is very similar to CRISS [1].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#2|9", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "The CRISS approach consists of a cross-lingual retrieval module and a seq2seq module, and they are optimized alternately.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#2|10", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "Similarly, the CORA approach in this paper also consists of DPR (Dense Passage Retriever, mbert structure) and GEN(Answer Generator, mt5 structure), and they are also optimized alternately.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#2|11", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "Although CRISS [1] doesn’t target at multilingual open QA, it has strong performance on cross-lingual sentence retrieval, which is related to multilingual open QA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#2|12", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "The authors should compare their work with CRISS [1].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#2|13", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "Authors claim that they introduce a new dense passage retrieval algorithm, but don’t mention any similar work about dense passage retrieval [2].", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc30|review#2|16", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "[1] Cross-lingual Retrieval for Iterative Self-Supervised Training", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#2|17", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "[2] Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#2|19", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "I don't see potential negative social impact of this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|11", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "The main difficulty of the proposed solution is how to train the MDPR and MGEN model components, given the limited multilingual raining data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|12", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "The initial model is trained with NQ (english) and TyDi QA and Xor-TiDi QA (multilingual data).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|13", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "To increase the amount of multilingual training data and language coverage, the authors introduce a data augmentation strategy where the initial model is used to harvest potential additional data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|14", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "This exploits the wikipedia “language links” resource.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|15", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "This iterative approach works by retrieving passages, looking up passages on the same topic in other languages using “language links”, and using the reader to see whether these can be used to generate the correct answer - if they can, they are added as additional training data for the next round.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|16", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "Another expansion mechanism exploits the fact that most answers are wikipedia entities, and can be translated using annotations available in the wikipedia “language links” resource.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|24", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "One limitation is that this won’t always work well if the language links aren’t high quality for some languages, or in other domains where language links aren’t available (either because wikipedia isn’t a good background corpus, or because answers are not wikipedia/wikidata entities).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|26", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "In terms of some constructive criticism, the MGen model could probably be made more powerful by using the Fusion-in-Decoder architecture, which would allow the model to scale to reading more passages per question, which has shown to greatly improve accuracy on open-domain QA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|27", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "I also wonder why the models’ performances are so different and sometimes inconsistent between MKQA and XOR-full (comparing tables 1 and 2) - for example, CORA scores 54.8 F1 for arabic on XOR-full, but loses a lot of performance arabic on MKQA (only 13.1 F1) .", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|28", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "Conversely, MT+DPR performs poorly for arabic on XOR-full (7.6 F1) but outperforms CORA on MKQA (16.0 F1).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|35", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "The authors provide a note on societal impacts which summarize the main impacts well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|36", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "This kind of work will likely have a positive impact on societal equality, unlocking the strong ODQA performance we see in high resource languages for lower resource languages.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|37", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "There are risks that CORA may provide wrong answers that might be harmful or offensive, but this is the case for monolingual systems too.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#3|39", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "The use of wikipedia language links is useful, but does limit the textual domains the data expansion technique can be used for to similar domains to wikipedia.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#4|12", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "But the novelty is limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#4|13", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "The most important part, I think, is the data expansion part described in Section 2.2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#4|14", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "From model perspective, there seems be have few innovation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc30|review#4|17", "source": "rc", "split": "test", "paper_id": "rc30", "comment": "no potential negative societal impact.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|1", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "This paper explores extreme parameter compression for pre-trained language model, especially BERT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|2", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "It introduces and compares several tensor decomposition methods and proposes to leverage Tucker decomposition as the final solution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|3", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The compressed BERT model achieves much smaller size with promising performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|15", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The authors do not include embedding layer and prediction layer size in experiments, while only report the Transformer encoder size.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|16", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "I know that this can make the size of compressed model really amazing (e.g., 1.8M) and the compression ratio amazing (e.g., 86M/12.3M=7) but is not fair as the whole model including the embedding layer are used when deploying.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|17", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "If the embedding layer is added, the model size will increase a lot, and the compression ratio will decrease, which make the experimental results less surprising.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|18", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "But this should be made clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|19", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The authors name a lot of related works, but compare only very few of them in the experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|20", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Some other method(s) are missing in the related works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|23", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Section 5.1, \"...are not exactly equal to the the raw weights...\", duplicate \"the\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|24", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Section 6.2, \"...outperforms ALBERT - the latter needs...while the latter does not...\", two \"latter\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|25", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "reference: [1] Xu, Jin, et al. \"NAS-BERT: Task-Agnostic and Adaptive-Size BERT Compression with Neural Architecture Search.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#1|29", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Though there exist some weaknesses, I think the paper is of good quality, if the authors could mitigate them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#2|10", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The paper proposed multiple potential ways of compressing weight matrices (matrix decomposition and tensor train decomposition) as some alternatives to the proposed Tucker decomposition.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#2|11", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "However, the author didn't compare with tensor train decomposition due to time constraints.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#2|12", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "I believe the paper will be more solid by adding ablation studies on tensor train decomposition.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#2|13", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The paper only performs experiments on BERT-base and TinyBERT models, but I believe that the compression method proposed in the paper should be more demanded by larger models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#2|14", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Some design choices in the paper seem arbitrary.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#2|16", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "I can understand the similarity of the attention query weight vectors across the layers, but I can’t understand why all the weights are merged.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc31|review#2|19", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "How does this method compare with previous works in terms of training/inference latency?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#2|21", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The paper would be better with more experiments on larger models and ablation studies.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|21", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The use of tensor decomposition for compressing neural networks has been explored extensively for CNNs, RNNs and Embeddings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|22", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The use here to compress transformers is a natural extension of the idea", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|23", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Decomposability and low-rank nature of FFN and MHA layers has been discussed previously in the literature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|24", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The authors themselves refer to these prior works", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|25", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Cordonnier et al 2021, further discusses the use of tucker decomposition to express MHA layers, albeit in a slightly different context.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|26", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "In order to improve the paper, I recommend providing more insights into the workings of the method and the bias that the fixed structure like tucker decomposition can lead to.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|27", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "I also recommend exploring the systems impact of running training and inference using tucker decomposed layers and why say 50x reduction in parameter count, does not lead to a 50x improvement in RPS (Table 4).", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|28", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Further, I would encourage the authors to explore and understand why finetuning with KD in Table 6 leads to such large accuracy drop.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|29", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "GD+TD should lead to better accuracy, but improving accuracy by 40% is an interesting data point.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|31", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "1. In section 4.2, the authors say \"During the inference phase, the terns that do not involve batch size b or seq length n could be calculated in an offline way...\". Could you expand on what you were referring to?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|32", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "2. Table 4 should have comparisons to sparsified BERT, esp for data points with 2-3x parameter compression. Both structured sparsity and random sparsity could achieve said compression. Eg - https://arxiv.org/abs/2109.04838", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|33", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "3. Tucker decomposition of matrices across layers forces common parameters between the matrices. A possible way this methodology could go wrong is if these matrices have different scales. Is the norm value across matrices similar for different layers? Have you looked at the problem from this point of view?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|35", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Overall, I think this paper is an incremental improvement to previous state-of-the-art.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|36", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Tucker decomposition has been used extensively in NN to compress RNNs, CNNs and Embeddings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|37", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Thus use of tucker decomposition and its ability to compress BERT MHA and FFN layers is incremental improvement over the previous results, especially given the fact that prior work has also shown that MHA and FFN layers can be decomposed in a low rank structure and talked about the redundancy in the parameters in those layers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#3|38", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "However, the results of the paper are interesting from an engineering point of view.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|1", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "After author rebuttal: thank you very much for the detailed response to both my and the other reviewers' comments. I have updated my score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|13", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The paper introduces many variants of decomposing the matrices (Table 2).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|14", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "However the results only seem to be based on one of the methods (i.e. IV).", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|15", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "I realize that some of variants have already been studied in the literature (e.g. Mao et al., Noach and Goldberg), but since the setup is not identical, it is crucial that the proposed approach is compared against both II and III.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|16", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Therefore, it is not clear whether the improvement in performance is coming from the actual proposed method, or something else.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|18", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The method requires a pipelined approach where one must first perform generalized distillation against the BERT model before doing task-specific distillation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|20", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "In GD, did you try distilling also the masked LM logits, in addition to (or instead of) the last layer hidden states and attention maps?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|21", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Have you checked the norms of the matrices?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|22", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Are they of similar scales?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|23", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "If not, do you obtain better performance by normalizing the matrices such that they are in the same scale?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|25", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "A new approach to model compression with strong-ish empirical results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|26", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Some reservations about the soundness of the method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc31|review#4|27", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "And it is furthermore not clear that the improvements are coming from the proposed approach as opposed to something else, since an ablation study across the different decomposition methods is missing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#1|8", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "Here the bound is formed by just dropping the \"+KL(p,q)\" term; I'll leave it up to the other reviewers to determine if this is the correct way to form this lower bound.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#1|9", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "Question for the authors: did you try other choices for the score function f_y (line 144)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#1|10", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "It would be nice to include some justification here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#1|14", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "In fact, they might have even spent more space making it extra clear than is needed -- Figure 1 helps a reader understand at a glance what is being optimized, then the terms in figure one are optimized in Figure 3 and we can see the contribution from the different terms.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#1|15", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "Most of the paper also goes into detail deriving and explaining these terms, and the paper could benefit (only a little) from some additional experiments taking some of that space.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#1|16", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "I would also recommend including more information about the hyperparameters used for the experiments; presumably default hparams were used, but it would be good for you to be clear about exactly what they were.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#1|17", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "Significance: If the experimental results hold on further experiments (e.g., with vision models) these results could be quite significant to the adversarial robustness community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#1|18", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "Most of the paper is focused on the derivation of the new approach, and the clarity of that part is also valuable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#1|19", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "Detailed comments: I would recommend including some vision experiments in the final version.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#1|22", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "I would recommend moving the related work section close to the end, but that's a style choice.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|8", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "One of the claimed contributions is to show the catastrophic forgetting phenomenon of adversarial finetuning of pretrained language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|9", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "I found this claim a bit not solid.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|10", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "From Figures 1 and 2, it is based on the relative L2 distance of the pretrained and finefuned models in the parameter space.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|11", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "However it does not necessarily mean it is really forgetting all the generic linguistic features.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|12", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "It is more straightforward to justify this by showing downstream tasks performance or language model performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|13", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "In addition, it is common to use a smaller learning rate or early stopping in fine tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|14", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "Hence, showing the enlarged discrepancy in parameters space in Figure 1 is not that informative.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|15", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "There is much more work in the image domain for adversarial training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|16", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "For example, TRADES [1], ensembling, early stopping [2], etc. I wonder whether the adversarial baseline is too simple and doesn’t incorporate few newer findings in the community.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|20", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "When mentioning pretrained language models, it is not clear to me in the beginning whether the proposed approach is for masked language models (e.g., BERT) or for autoregressive language models (e.g., gpt-2).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|22", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "The presented findings on finetuning on masked language models seems promising, but it is hard to justify as this is evaluated on new tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#2|23", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "In addition, it will be more insightful to see quantitative examples or more analysis for the actual adversarial attack (success, fail ones) beyond reporting accuracy -- this is because adversarial attacks could potentially be imperfect in the NLP domain, or there might be gradient obfuscation problems in claiming robust accuracy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#3|8", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "The improvements over the baseline models are relatively marginal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#3|9", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "The paper argues that previous adversarial fine-tuning would cost severe catastrophic forgetting issue, and I expect the improvement to be higher if the propose method can well tackle this issue.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#3|10", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "This paper needs analyses in terms of how the proposed method addresses the catastrophic forgetting issue compared to baseline models, since the main claim of the RIFT is to tackle this issue.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc32|review#3|12", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "Since this paper aims to cope with the classification-based tasks, it would be better to conduct experiments on more tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#3|13", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "For example, try on GLUE benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc32|review#3|14", "source": "rc", "split": "test", "paper_id": "rc32", "comment": "How many times did you run the experiments for each task and calculate the averaged performance? Since the improvements for many tasks are marginal and the variances looks not low enough. It would be better if you conduct a statistical significant test.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|16", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "What is the batch size at test time tuning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|17", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Is the added robust prefix the same for the entire test set, or the same within a batch but different across batches, or unique for every test example?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|18", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "This is an important point to assess whether the experiments are in an online setting where test data arrives in-stream or not.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|19", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Section 5 is not very convincing to me:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|20", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "(1) there are only several case studies without any quantitative results; I think it may appear in the appendix only or just uses a short paragraph in the main body (because the statements in Section 5 can only be good hypotheses which are not well-supported by quantitative evidence, also, attention is not a convincing proxy for the explanation, see point (2)), yet it takes more than one page in the current version;", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|21", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "(2) the interpretation from the perspective of attention weights bothers me a bit, attention as an explanatory tool is known to not be faithful [1] – a larger attention weight does not necessarily mean the final prediction depends on it more than others, and vice versa, thus maybe not overread it too much.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|22", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "As said above, it is ok to do it this way in the appendix, but using over one page of the main content for this is not convincing to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|23", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "In the attention visualization figures (e.g. Figure 4), why does a word attend itself as well instead of only attending to the contexts?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|24", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "For LM, I think that the diagonals in the attention figure should be zero, or am I missing something?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|25", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Also, the presentation could be improved here, in the figure caption you can explain what the rows and columns mean in the visualization to make it easier to read.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|26", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Section 6 is difficult to follow without reading the appendix, and it is disconnected from the rest of the paper in terms of the paper structure – having a theoretical interpretation section after experiments at the end of the paper is not a good presentation structure in my view.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|27", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "If you think Section 6 is important to have, move it before the experiment section and clarify more details to make it more self-contained; if it is not very important, you can just put it into the appendix while briefly mentioning it in the main body.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|28", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Besides the presentation issues above, there are some other minor presentation places which could be improved, e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|29", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "(1) Eq 4, an undefined variable X_C suddenly comes in without explanation", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|30", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "(2) Better to explain Eq (8) with more text given that this is an important equation for the proposed method", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|31", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "[1] Serrano et al. Is Attention Interpretable? ACL 2019", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#1|33", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "There are novel contributions of the paper both technically and empirically, however some major analysis sections are not convincing and a significant portion of the presentation needs to be improved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#2|4", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The authors study a novel problem in lightweight fine-tuning methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#2|5", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Most studies aim to match the performance of full model tuning via updating a subset of parameters but rarely study the robustness of lightweight fine-tuning methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#2|11", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The authors argue that robustness is important for lightweight tuning methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#2|12", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "But I still think it is better to provide a comparison between lightweight tuning methods will full tuning methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#2|13", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Provide a basic starting observation about whether lightweight methods bring more challenges on robustness or not/", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#2|15", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The studied problem is important and novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#2|16", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The proposed method is simple and clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#2|17", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The experiments justify the effectiveness of the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|8", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Review note: I am not familiar enough with Optimal Control Theory to evaluate the soundness of Section 6 / Appendix E. I will leave it outside the scope of my review and hope other reviewers can fill in the gap.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|19", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Inference Batch, batch size and its importance: The batch size at inference seems like an important variable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|20", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Indeed, my understanding of Section 3 and Eq.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|21", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "8 is that the specific batches used at inference will play an important role as they impact P_psi.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|22", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "This is my key concern with this paper as it brings up quite a few issues that should be mentioned.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|23", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Inference for a datapoint depends also on other datapoints in the batch.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|24", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "This is different enough from other ML setups that it should be highlighted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|25", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "It also causes reproducibility issues.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|26", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "For instance, batch norm avoids this by fixing batch statistics at inference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|27", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "It is not clear that the method works for low inference batch size (opt of Eq.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|29", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The inference batch size used is not mentioned anywhere.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|30", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "If the method requires batch size > some N (or if performance varies widely with batch size), this seems a strong assumption that should be made clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|31", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "You do not always get several samples of the attack on your system.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|32", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "It is not clear how well the method works when the inference batches are a mix of unperturbed samples and perturbed ones.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|33", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "This seems like a more realistic attack scenario.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|34", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "It is not clear how well the method works if there are different attacks in the inference batch, which also seems like a more realistic threat model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|35", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Overall, I feel like some answers to the above would diminish my concerns, most notably:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|36", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "What is the inference batch size used in the experiments?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|37", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "How does performance vary with inference batch size (one or two settings should be enough)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|38", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Does the method work when only x of N samples in the test batch are adversarial?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|39", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Does the example work when they are two types of attack", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|41", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "It would be helpful to remind the reader for each dataset how many classes they are / what the mode/random pred accuracy is to help interpret results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|42", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "For instance, on Table 1 under PWWS improves from 16.64 to 50 for SST-2 and 25 to 34 on SNLI but that is the same as the accuracy of a random predictor.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|44", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The writing could be improved substantially.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|45", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Some examples: p2 “remaining the pretrained models unmodified” -> “keeping”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|46", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Also on p3 “remaining its lightweightness”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|47", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "I suggest doing another pass as it does not align with the quality of the rest of the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|49", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "I feel like other methods for Parameter efficient transfer learning, such as Adapters, normal prompting should be quickly mentioned. Same goes for comparable approaches such as P-tuning.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|50", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "As mentioned earlier, I would caveat Section 5 with discussions of the validity of using attention for explanation, such as the Attention is Not Explanation paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|52", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Given the author's response, I am raising my score slightly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|53", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "There are still some concerns over the threat model but some of my questions on the impact of test batch size have been answered.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|56", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "One major caveat however is that it is not clear how flexible the inference setup is.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|57", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "This is critical as the predictions for a datapoint at inference depend on the other datapoints in the batch.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|58", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Currently, it seems like the inference is being done with a batch of all the same attack.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|59", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "These assumptions are simply not realistic as a threat model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|60", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Without understanding how this performs under a more realistic threat model, I cannot score this paper higher.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|61", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "I have highlighted experiments that would provide more realistic results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|62", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Despite my \"marginally below\" score, I do not think the paper can be accepted without an answer on this point.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|63", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Reject is too harsh for a paper that is otherwise promising.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#3|64", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The paper would also benefit from another writing pass, mentioning missing related work and clarifying some properties of the dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|16", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The experiments are OK but could be a bit more streamlined.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|17", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "I guess two things came up when I was looking at them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|19", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The other thing is that it is hard to contextualize how good the numbers in your framework are.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|20", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "For example, I have no idea how good 52% on VIPR for SST2 is.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|21", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "If previously proposed defenses are able to get 90% on the same task, then it is unlikely that the proposed method will gain traction, even if it is more computationally efficient with prefix tuning. So it would strengthen this paper a lot to have this comparison.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|23", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "I think the paper may benefit from spending just a little more time describing how susceptible prefix tuning is to adversarial attacks compared with regular finetuning.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|24", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "It seems like it is very susceptible, especially for easy tasks like text classification.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|25", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Appreciate the candidness in Figure 2 in showing that adversarial training takes longer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|26", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "It will be good to see if the proposed method not only improves performance against adversarial attacks, but also if it improves performance against different paraphrases of expressing the context and question, for example.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|28", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The grammar of the paper will benefit from having a native english speaker proof-read it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|29", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "E.g., \"as well as remaining the pretrained models unmodified\" can be re-written as \"without modifying the pretrained model parameters\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|30", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Since a lot of work was put into the paper, I will pay my respect and make a picky point: the bibliography could be cleaned up a bit.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|31", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "E.g., capitalize RoBERTa correctly, add URLs consistently, capitalize all conference names (e.g., IEEE transactions).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|33", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The paper studies the very timely topic of adversarial attacks against prefix tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|34", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The paper proposes a method that maintains the advantages of prefix-tuning against finetuning, a formidible problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|35", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The experiments are OK as of now, though expanding the scope to generation tasks could make the impact substantially larger.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|36", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "The characterization of the method could also be further improved by providing finetuning (with and without defenses) as baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc33|review#4|37", "source": "rc", "split": "test", "paper_id": "rc33", "comment": "Overall, I lean towards acceptance, though I am not an expert in adversarial attacks/defenses.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#1|9", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "This method can be seen as a variation of schedule sampling (Bengio et al., 2015) or a simplified version of UniLM (Dong et al., 2019). The novelty of this paper is limited.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc34|review#1|10", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "On the other hand, the authors should give more analysis about why this method can work in LLMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#1|11", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "The training set of pre-training models contains a lot of noise data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#1|13", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Bengio et al., 2015. Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#1|14", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Dong et al., 2019. Unified language model pre-training for natural language understanding and generation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#1|16", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "This clarity of this paper is clear, and the method mentioned in this paper is easy to reproduce.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#1|17", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "The novelty of this paper is limited.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc34|review#1|19", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "The novelty of this paper is limited. Methods similar to this paper have been widely used in previous work.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc34|review#1|20", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "On the other hand, the authors should give more analysis about why this method can work in LLMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|12", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.0) Kindly share the computational resources used, memory requirements and the training time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|14", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.1) In Table 1, along with the mean of the 3 random realizations, the spread should be provided as well, given that many of the performance numbers are within a single percentage point.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|15", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.2): The performance of FCM appears to be better than PaLM, but not significantly and consistently so.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|16", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "For significance, I’m using a criterion of at least 1% in absolute terms.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|17", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "It will be good if the authors can shed some light on the following:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|18", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.2.a) 0-shot: In 8/19 tasks, the performance improvement is < 1%.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|19", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.2.b) 1-shot: In 9/ 19 tasks, the performance improvement is < 1% and 2/19 times it’s worse.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|20", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.2.c) few-shot: In 4/ 19 tasks, the performance improvement is < 1% and 4/19 times it’s worse.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|21", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.2.d) Scaling trends: If we look at the consistency of performance improvement as the number of shots increase, there are very few tasks, on which FCM consistently stays better than PaLM by at least a percentage point.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|22", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "On the NLU tasks, FCM seems to lose its advantage as the number of shots increase, sometimes becoming much worse.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|23", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.3) Classification finetuning: results on SuperGLUE dev set only show significant improvement on 4 out of 8 tasks, with majority of the gains observed on WSC and CB.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|24", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "The authors should discuss the variability in performance gains.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|25", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.4) Scalability, 0-shot: In Table 4, FCM 1B results seem better than FCM 8B on CB, RTE, and WiC tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|27", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.5) Scalability, k-shot: In Table 5, the performance of both PaLM and FCM sometimes, and more often for FCM, degrades as the number of shots increases.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|29", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.6) Ablation of mask ratio: Irrespective of the average across all the tasks, FCM[0,0.15] is better only on 2 out of 8 tasks in the 0-shot setting and 3 of 8 tasks in the 1-shot setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|30", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "This shows that the choice of mask ratio is task dependent.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|32", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "(W.7) Comparison with Dropout: Contrary to the claim of dropout hurting FCM performance, it improves in 3 of 8 tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|33", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "It improves performance by 10.2 absolute % points for the MultiRC task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|34", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "It seems that the benefit of dropout is task dependent.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|36", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Clarity: The ideas are presented reasonably clearly and in a structured manner.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|37", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Quality: The quality of the paper is mostly good, with some minor issues pointed out below which are easily fixable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|39", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Appropriate use of capitalization in the ‘References’ section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|41", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Low on technical novelty.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|42", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "The simple, novel, scalable pre-training masking tactic seems to have a broad minor to moderate impact on a variety of tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|43", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "I expect this to become a part of the pretraining toolbox for causal LLMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#2|45", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Should be reproducible (subject to availability of computational resources).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#3|7", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Experimental details are also well explained.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#3|11", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "While the proposed method does not add any computational cost in theory, since previous tokens are randomly masked out, won't there be an engineering overhead since the previous hidden states cannot be cached which would have remained unchanged if the previous token weren't masked out otherwise?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#3|13", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "It would be great if the authors could elaborate more on the disadvantage of randomly making out previous tokens in terms of the actual engineering implementation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#3|14", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "It would be great if the authors could clarify if there were any misunderstanding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|5", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "The biggest weakness of this paper is ignoring related research papers and lack of comparison.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|6", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "[1] presents comprehensive study about different training objectives in zero-shot priming and finetuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|7", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "They compared 6 different training objectives in table 1 (https://arxiv.org/pdf/2205.11726.pdf): NXTUNI, NXTPRE, MSKUNI, MSKBI, HYBUNI, HYBPRE.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|8", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Casual language modeling such as GPT models belong to NXTUNI.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|9", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "They controlled hyper-parameters and training data to be same and studied these six different objectives.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|10", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "The most relevant method to this paper is HYBUNI (casual masking objective). [2] is the representative work of causal masking objective. This paper presents this new causal masking objective without comparing it with the most relevant casual masking method (CM3).", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|11", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "In addition, this paper should carry out apple-to-apple comparisons with other models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|12", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "For example, it needs to control model sizes (model parameters), batch size, hyper-parameters, training tokens, training devices, to be exactly the same.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|13", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "And control Total train compute (flops) to be similar.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|14", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Therefore, the qualified apple-to-apple comparison is only PaLM 1B v.s. FCM 1B (in table 3), PaLM 128M v.s. FCM 128M (in table 4).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|15", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "For others, PaLM 8B used different training data with FCM 8B, and T5 used different model parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|17", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "[1]. Artetxe, Mikel, et al. \"On the Role of Bidirectionality in Language Model Pre-Training.\" arXiv preprint arXiv:2205.11726 (2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|18", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "[2]. Aghajanyan, Armen, et al. \"Cm3: A causal masked multimodal model of the internet.\" arXiv preprint arXiv:2201.07520 (2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|21", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Due to lack of comparisons with relevant methods, I don’t think this paper has a high quality.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|23", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "The casual masking method proposed by this paper is quite similar to [2] and HYBUNI in [1]. [2] also adds some masks on the basis of casual language modeling.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|24", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "The difference is that [2] puts the masked tokens to the end of sentence, so the masked places can not only attend to previous tokens but also gather information from future tokens.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|26", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "But due to lack of comparison, I cannot see the performance differences.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|27", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "Reproducibility: easy to reproduce.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|29", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "[1]. Artetxe, Mikel, et al. \"On the Role of Bidirectionality in Language Model Pre-Training.\" arXiv preprint arXiv:2205.11726 (2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|30", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "[2]. Aghajanyan, Armen, et al. \"Cm3: A causal masked multimodal model of the internet.\" arXiv preprint arXiv:2201.07520 (2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|36", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "[1]. Artetxe, Mikel, et al. \"On the Role of Bidirectionality in Language Model Pre-Training.\" arXiv preprint arXiv:2205.11726 (2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc34|review#4|37", "source": "rc", "split": "test", "paper_id": "rc34", "comment": "[2]. Aghajanyan, Armen, et al. \"Cm3: A causal masked multimodal model of the internet.\" arXiv preprint arXiv:2201.07520 (2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|10", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The paper attempts to solve an important problem of LFLL in the context of the pre-trained Transformers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|11", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "With the rise of larger models, it is important to consider the scalability of these models for lifelong learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|13", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Empirical results showcase that the proposed approach is superior to existing methods like EWC and MAS when evaluated on the sequence of 2-4 domains/tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|15", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "For the LFLL problem, this paper proposes to use an existing prompt tuning-based transfer learning method where the underlying model is frozen and a few additional task/domain-specific parameters are learned.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|16", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Furthermore, it repurposes existing lifelong learning methods like deep generative replay [1], distillation-based methods [2], dynamically expandable network [3] as a unified framework when applied to prompt tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|18", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Another major limitation of this paper is that it completely ignores comparison with other parameter-efficient transfer learning methods like adapters [4].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|19", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Concretely, AdapterFusion [3] work proposes to learn a task-specific composition of adapters from previous tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|20", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "How does the prompt tuning-based method compare with adapter fusion when applied to the LFLL problem?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|21", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Is there a consensus that prompt tuning-based methods are better than adapters for few-shot learning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|22", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "By definition, the lifelong learning paradigm deals with a large number of tasks in sequence.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|23", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "However, the paper evaluates the proposed approach on the sequence of 2-4 tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|24", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "So it is unclear whether the gains persist as we increase the number of tasks to 10-15 (towards a more realistic evaluation setup).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|25", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "As we increase the number of tasks, does it enable positive forward/ backward transfer or increase the chance of negative transfer?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|26", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Overall, the paper is written in good style.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|27", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "However, there are key experimental details that are missing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|28", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "This paper defines prompts as a series of tunable tokens but then it is unclear how many tokens are defined per task prompt.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|29", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "This is important because as we scale the number of tasks, the input sequence length scales linearly, and with that self-attention scales quadratically.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|30", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "This issue limits the practical applicability of the prompt tuning-based method for lifelong learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|31", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "What does it mean to use validation sets for hyper-parameter selection?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|32", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Ideally, during lifelong learning, we do not know the sequence of tasks apriori.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|33", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "It is unfair to freeze a particular task order, then search for optimal hyper-parameters, and then report results on multiple runs with the same task order.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|34", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "At least, report the results on unseen task orders?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|35", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Or search on a subset of disjoint tasks [5].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|36", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The paper should consider updating their results with a fair hyper-parameter selection strategy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|38", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The paper trains a single prompt for both task solving and data generation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|39", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "It is unclear whether this is an optimal strategy when it comes to generating pseudo-samples?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|40", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Did the paper try learning a separate prompt for task solving and data generation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|41", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Does it help in addressing the quality of generated pseudo-samples?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|42", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Furthermore, how effective is the current pseudo-sample generation strategy?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|43", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "How many samples are discarded to get valid samples?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|44", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "There are a few hyper-parameters that are set without any justification: 16-shot learning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|45", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "How does the method perform with varying the number of training examples?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|46", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "How does the performance change with varying the number of pseudo-samples?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|47", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Does it hurt if we have more pseudo-samples (generating them is feasible and cheaper)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|48", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "In terms of the multi-task experiments (MT-PT), do we have separate prompts or a single prompt being trained with all data?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|49", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "If it is a single prompt then is it fair to compare it with LFPT5 with more capacity?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|50", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "If there are multiple prompts then how does one ensure that there are dedicated prompts for each task?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|51", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "[1] Sun, Fan-Keng, Cheng-Hao Ho, and Hung-Yi Lee. \"LAMOL: LAnguage MOdeling for Lifelong Language Learning.\" International Conference on Learning Representations. 2020.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|52", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "[2] Li, Zhizhong, and Derek Hoiem. \"Learning without forgetting.\" IEEE transactions on pattern analysis and machine intelligence 40.12 (2017): 2935-2947.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|53", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "[3] Pfeiffer, Jonas, et al. \"AdapterFusion: Non-Destructive Task Composition for Transfer Learning.\" Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|54", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "[4] Houlsby, Neil, et al. \"Parameter-efficient transfer learning for NLP.\" International Conference on Machine Learning. PMLR, 2019.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|55", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "[5] Chaudhry, Arslan, et al. \"Efficient Lifelong Learning with A-GEM.\" International Conference on Learning Representations. 2018.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|57", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Overall, I rate this paper marginally below the acceptance threshold.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|58", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "It is interesting to see the efficacy of prompt tuning for lifelong language learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|59", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "However, with the current experiments, it is unclear how they compare with adapters?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|60", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Also given this paper unifies existing approaches and adapts them for prompt tuning setup, it is marginally significant/ novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|61", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "There are several open questions (see weakness section + additional clarification questions) that need to be answered before understanding the overall benefits of the proposed framework.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#1|62", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "I am looking forward to the author's response in the rebuttal period and will consider updating my scores accordingly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|1", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "This paper proposes to leverage prompt tuning on large pre-trained language models to achieve Lifelong Few-shot Language Learning (LFLL).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|2", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Lifelong learning and few-shot learning have been considered different ML paradigms.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|3", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The authors argue that with the emergence of modern large-scale pre-trained language models, AI models now have the capability to achieve these 2 at the same time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|4", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "This work attempts to formally define the LFLL problem and benchmarked a strong pre-trained LM (T5) on 3 tasks (NER, text classification, and summarization) over 9 different datasets (defined as domains in LFLL).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|7", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Combining lifelong and few-shot learning is a new setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|8", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Experiments contain 3 different tasks and each has datasets from different domains.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|13", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The number of tasks and domains is minimal setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|14", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "To really become a benchmark to measure the progress of LFLL, more tasks/datasets will be needed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|16", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Question Answering and NLI are also very important NLP problems and many datasets exists from different domains.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|17", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Perhaps the authors can consider adding them to setup a more comprehensive benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|18", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Perhaps adding more pre-trained LMs such as GPT-2 and different sizes of T-5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|19", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The community will be interested how the model type / scale affect the LFLL capability.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|20", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "How will the pseudo data generation amount affect the learning / forgetting performance?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|21", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Since the data is generated by T-5, couldn't we generate as much as we want?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|22", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Maybe adding another ablation on this would be a good idea.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|24", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Overall I'm leaning positive about this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|25", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The proposed LFLL problem, in my opinion, is very important for the community to think about: how to adapt large-scale pre-trained LM to different tasks/domains with few samples?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|26", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "This paper sets up a strong baseline and benchmark for research along this direction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|27", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Although prompt tuning, KL regularization, and pseudo data generation are existing ideas, combining them and formulate this LFLL problem/benchmark is non-trivial contribution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#2|28", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "I think this work has potential, and if the authors can try to add more tasks/datasets/baselines, it may create more impact to the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|1", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The paper extended the use of prompt tuning to the scenario of life-long language learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|2", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "To prevent catastrophic forgetting, the paper proposes to generate pseudo samples as part of the training data when adapting to a new domain, and adding new set of prompt tokens when adapting to a new task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|3", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Using the T5 model, the paper demonstrated the effectiveness of the proposed methods using text classification, named entity recognition (NER), and summarization tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|4", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The method achieved better results for these tasks using the proposed life-long learning mechanisms, compared to fine-tuning, prompt tuning, and their combinations with regularization-based life-long learning methods such as elastic weight consolidation (EWC) and memory-aware synapses (MAS).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|7", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The idea of using prompt-tuning for life-long language learning is a promising direction and the methods proposed in the paper constitute a novel and effective way in bridging the gap between these 2 methodologies.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|8", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The discussion on few-shot learning also clearly demonstrated the difference of the proposed methods and a major collection of prior work using prompt-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|9", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The methods proposed in the paper are effective in utilizing the assumption of language modeling, such that tasks are formed as prompt-label pair sequences, and autoregressive sequence generation is used to anchor the model to previously learnt knowledge.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|11", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Comparisons to fine-tuning, prompt tuning, and their combinations with elastic weight consolidation (EWC) and memory-aware synapses (MAS) are solid choices to demonstrate the effectiveness of the proposed methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|12", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "In most of these comparisons, the proposed methods achieved better results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|13", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "There are a few weaknesses that I hope the paper can address to improve the scoring:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|14", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "In the paragraph below equation 4, the loss form is an addition of 3 values: 1) the task loss L^{task}; 2) the language modeling loss L^{lm}; 3) label consistency loss L^{KL}.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|15", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Both 1) and 2) are reasonable, but 3) is new in the paper and it is unclear why it is absolutely necessary. In particular, the paper should include an ablation study on a range of \\lambda_{KL}, not just a comparison between models with and without L^{KL}.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|16", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Some statistics that are useful for understanding the propose methodology are missing in the experimental section: 1) it is probably beneficial to include some numbers on the state-of-the-art results on methods for the original tasks and datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|17", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "They may not be fair statistics in a life-long scenario, but they are helpful for the readers to understand where the methods stand in a larger picture of machine learning methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|18", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "2) For the domain order experiments in section 5, it could be useful to also include accuracy for each of the datasets separately.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|19", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "This will provide the readers with additional knowledge on the relationship between task ordering and individual task performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|20", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Finally, some suggestions to the paper, but these will not affect the scoring:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|21", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Remove the quote in page 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|22", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The paper should include a brief introduction to the 2 regularization-based life-long learning methods that it compares with.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|23", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "These are elastic weight consolidation (EWC) and memory-aware synapses (MAS).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|24", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Try to reduce the use of abbreviations in tables.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|25", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "For example, by typing fine-tuning and prompt-tuning instead of FT and PT, it could make the table much easier to read without the need to refer to the paper text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|26", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Summary Of The Review:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|27", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Good novelty in extending prompt-tuning to life-long language learning using the generative properties of language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|28", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Good experiments demonstrating the effectiveness of the methods in preventing catastrophic forgetting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#3|29", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Some parts of the method are not sufficiently studied in the experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|1", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "This paper explores lifelong few-shot language learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|6", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Few-shot setting and lifelong learning are important settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|10", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "I did not understand why the proposed method is restricted to the few-shot setting. Do gains go away in the full data setting?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|11", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The experiments are OK but not probably not good enough to support the general claims about lifelong learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|12", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "E.g., it would be great to show more than 3 tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|13", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "I had a hard time understanding the paper, see below (and I will remove this comment from reasons to reject once it is clarified):", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|14", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The key part that I did not understand is why prompt tuning isn't already a sufficient life-long learning method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|15", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "My understanding of prompt tuning is that no model weights are modified, only the prompt is optimized, and optimized prompts are stored for each task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|16", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Then given a new task, a new prompt can be trained and stored.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|17", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "This works because the size of prompts is many orders of magnitude smaller than the size of the model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|18", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Maybe another way of asking this question is, how is the following different from regular prompt tuning: \"while adapting to a new task type, LFPT5 includes and tunes additional prompt embeddings for the new task\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|20", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Let's say we're up to 1000 tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|21", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "How much weight does the new task get in this mixture?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|22", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Won't this take a long time?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|24", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The introduction could make it more clear why the second two limitations of LLL are indeed limitations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|25", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Why NER in particular, instead of other tasks that haven't been explored thoroughly?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|26", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "The negative transfer point needs some more evidence as well, as I believe the decaNLP challenge showed positive transfer, right?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|28", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Is the finetuned BERT-Large a few-shot finetuned or with the full dataset?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|29", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "PT is not defined at first use in the abstract", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|30", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "All the acronyms are pretty confusing", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|31", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "Typo: review socre 5", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|33", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "This paper tackles an important task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|34", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "I didn't understand why their method specifically applies to the few-shot scenario.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|35", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "I also struggled to understand why the proposed method solves lifelong learning over regular prompt tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc35|review#4|36", "source": "rc", "split": "test", "paper_id": "rc35", "comment": "One sentence in the paper seems like an overclaim to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#1|7", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Optimizing fine-tuning for improved data efficiency (or more targeted task improvement) is an important research area.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#1|8", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Nice connections to retrieval augmented language models with corresponding retrieval experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#1|10", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The approach is extremely simple.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#1|11", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "This could also be a strength, but I wonder how much using only a single source (P3) of prompts is biasing results, since many templates used for creating that dataset are very similar across tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#1|12", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Expanding results across prompted datasets would be a good series of additional experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#1|13", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Given the hypotheses around the the role of negative interference in hurting performance (and the resulting benefit of querying more similar examples for fine-tuning), the paper would be much stronger if there were direct experiments exploring this", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#1|14", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Some of the model naming conventions are confusing/inconsistent (DEFT-T5-base vs DEFT-base).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#1|15", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "It wasn't clear to me what DEFT-base was defined as (vs XL which was the 3B param variants).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#1|17", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Some of the model naming conventions could be clearer", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#1|18", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The method (nearest neighbor augmented retrieval) isn't novel, though providing a detailed empirical assessment of its performance is a good contribution", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#1|19", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Uses public data for reproducibility, though unclear if code will be shared (though again, the methods are straightforward)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|9", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The problem is well posed, the method is clearly explained while being relatively novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|10", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Numbers are strong and experiments well conducted and results are interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|12", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "I have a few doubts about some of the stronger claims in the paper (in particular the negative transfer argument, see questions).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|13", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Few-shot fine-tuning has a long history in NLP.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|14", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "I’d like to see a fairer comparison to stronger baselines.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|15", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "IA3 is particular in the sense that it only fine-tunes a fraction of the parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|16", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Using a parameter efficient fine-tuning baseline needs a better justification (especially since DEFT-Few seems to be fully fine-tuned).", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|19", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "How did you handle imbalances in the P3 dataset?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|20", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "P3 has large subsets that can be 100x bigger than other subsets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|21", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "If these imbalances are not properly handled, DEFT could be overly sampling (random) or retrieving from the most represented subsets rather than the most relevant subsets (because the most relevant subsets are potentially small). In term, the “negative transfer” argument becomes less convincing.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|22", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "While I agree, it’s hard to re-train T0-3B on a well balanced P3, I’d like to see the random sampling baseline with a balanced P3 if it’s not already the case.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|23", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Introduction: “These findings suggest that prior work may have underestimated the cross-task generalization capabilities of large multitask models”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|24", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "I think this is a generous claim or at least imprecise (if prior works mean T0/FLAN): while previous works have focussed on the true zero-shot case (where there is no knowledge of the target task, nor through its prompt or through any labeled or unlabeled instances), the setup you are interested in requires access to target task (unlabeled or labeled).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|25", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "It seems that a given DEFT model optimized for target task A would perform well on a target task A, but worse on a target task B, whereas T0/FLAN would perform relatively worse than a “specialized” DEFT but better on average.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc36|review#2|26", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Could you showcase a few retrieved examples and their associated query?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|7", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The approach is simple and effective: The main idea is simple and effective - perform a \"task-attention\" on a set of training tasks to augment the test task few-shot data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|8", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "It is quite interesting to see that a naive example retrieval approach such as taking the union of the retrieved training examples can lead to significant gains (64.6 vs 60.4 on T0 test tasks for the few-shot setting).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|9", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The paper is clear and well-written", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|12", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The authors acknowledge this similarity in the related works, by stating that their approach differ from ReCross because ReCross \"make use of an additional re-ranker and multi-task trained model to retrieve useful instances.\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|13", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Per se, not using a modeling choice of a previously published paper is not a significant novelty: the novelty and contribution would come from showing that the modeling choice used in ReCross is not necessary or even harmful for achieving good performance in a different setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|14", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "These conclusions are currently missing from the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|15", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Given the good scores the paper reports on the T0 suite, I can imagine that the reranker might be only marginally useful, but this should be clearly stated and empirically supported.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|16", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The community goes fast and probably the authors started their work before having noticed the relevant related work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|17", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Nevertheless, at the moment, this paper presents the idea as totally new.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|18", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "It would be more suitable, for example, to acknowledge the similarity to Li et. al, 2022 directly in the intro and maybe make an extra effort to offer a simplification/ablation of some components of the aforementioned approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|19", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Some of the questions I can think of that would give a bit of novelty and require a small additional effort are:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|20", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Is the reranker used in ReCross useful?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|21", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "What happens if you fine-tune a multi-task trained model (T0-3B in your case, similar to ReCross), rather than a \"generalist\" model (T5-XL-LM)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|22", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Experimental setting can be stronger: Differently from ReCross, the authors only test on the T0 evaluation suite and build heavily on top of IA3 (Liu et al, 2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|23", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "I do believe that another benchmark is necessary.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|24", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "It would be interesting for example to explore the proposed method with Natural Instructions V2 dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|25", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "For example, this would give a better idea if the method works due to the nature of the prompts format in P3 (e.g . it retrieves examples with similar prompts) or due to other mechanisms.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|26", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Claims about task interference: The authors claim that their result show that there is negative task interference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|27", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "I would think that an important experiment to substantiate this claim is to report the results for 2) above, where the method is applied on top of a multi-task model trained on the multi-task training set.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|29", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "It would be nice to have Figure 3 and Figure 4 for the few-shot setting, and compare the performance to the T0-3B+IA3 when fine-tuned on the original test task data vs T5-XL-LM fine-tuned on original test task data + 2x, 3x, 4x of augmented data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|30", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The applicability of the method outside of NLP tasks might be discussed in the conclusion, e.g. in vision or multi-modal settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|33", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The reported results are convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|34", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Unfortunately, the overall contribution of this paper is diminished by the limited novelty with respect to ReCross (to appear at NeurIPS, 2022, and published on arXiv in April).", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|35", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The authors cite this paper in the related work, but I feel that the similarity is quite understated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|36", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "It is questionable whether a paper published in April should be considered related work or not and whether this should impact my review.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|37", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "In the current context, I would suggest the authors to do a small extra effort to 1) highlight similarities with ReCross and 2) differentiate their contribution w.r.t.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|38", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "ReCross, for example, by reporting ReCross performance and ablating some of the ReCross components (e.g. the reranker and the fact that ReCross uses the multi-task model as base model).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|39", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Right now, the paper appears to be written as if it was the first in proposing this retrieval idea.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#3|40", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "I didn't discover the similarity until I read the ReCross work in details.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|4", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "This paper proposes a simple yet effective idea on how to use existing resources (i.e., P3) to facilitate unseen tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|5", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The paper is well-written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|6", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The evaluation is good - it considers both 0-shot and few-shot settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|9", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Similar approaches exist like kNN-LM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|10", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Also, the proposed approach is also related to Mixture-of-Experts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|11", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "It is unknown if the task retrieval mechanism in the paper is optimal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|12", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "For example, what if we use BM25 for retrieval?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|13", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "Also is task embedding relevant here?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|14", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The proposed method has its drawback.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|15", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "It requires retrieval and training for each task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|16", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "In other words, for 3 tasks there are 3 models trained instead of the open-box T0.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|17", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "In Table 1, DEFT outperforms T0 but the improvement is somewhat limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|19", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "The paper is clear, but I doubt it can be easily reproduced by a third party given the scale of data involved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc36|review#4|20", "source": "rc", "split": "test", "paper_id": "rc36", "comment": "I recommend the authors release their code.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#1|16", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Table 1 shows that models with different percentages of redundant parameters have similar performance, implying that performance may not be proportional to \"well-trained\" parameters. It means that making more parameters \"well-trained\" does not necessarily improve the performance. This seems not totally in line with the paper's motivation.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc37|review#1|17", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Figure 5 is not straightforward to visualize the difference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#1|18", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "The curve is drawn only on the right subfigure.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc37|review#1|20", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "According to Figure 3, models trained with SAGE are susceptible to parameter pruning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#1|21", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "In general, we want efficient and compact models in deployment.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#1|22", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Will SAGE be harmful to developing efficient deployment models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|9", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "The use of Taylor approximation is only good close to the operating point for nonlinear functions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|10", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "However Theta_j,-j may be very large, and thus the approximation may be very bad.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|11", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "I do not agree that the memory and computational costs of SAGE are \"marginal\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|12", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "The EMA \"I-hat\" is a full copy of the model since we need one EMA per parameter, especially for large models this is substantial overhead.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|13", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "I would further appreciate timings to verify that indeed the computational overhead is \"marginal\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|14", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "It is not clear to me in what scale the quantities U and I-hat are, and hence, the multiplier to the learning-rate.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|15", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Because of this, it is not clear how one needs to change the original learning-rate \"eta\" when one turns on SAGE.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|16", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "I am confused by Fig5, it does not look like the two-moon dataset at all to me?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|17", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Please image-search \"two moon dataset\" and compare.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|18", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "minor nitpicks: 5) Fig1: Percent should be [0,100].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|19", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "6) I do not agree that one can conclude from the results that SAGE is more effective for small datasets than for large ones, since the \"points\" betwen the datasets/tasks do not live on the same scale.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|20", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "7) All experimental evaluation is about fine-tuning of already pre-trained transformer models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|21", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "This should be more accurately reflected in the title, for example by replacing the word \"training\" by \"fine-tuning\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|22", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "It would further be interesting to see how well it works when training from scratch; even if it does not work, it is still a valuable fine-tuning method, and stating this may save a lot of people a lot of resources and time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|23", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "8) In Figure 6, since we have per-parameter learning-rates, which one is shown for SAGE?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc37|review#2|24", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "9) Typo: \"smoothier\" -> \"smoother\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#3|13", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "The scaling formula seems somewhat ad hoc and hard to characterize. In particular, if the sensitivity of some parameter spikes on a given iteration, it will get a larger update than another parameter with the same moving-average sensitivity that has not spiked.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc37|review#3|14", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "There are some hints that the gains might be partly related to regularization (eg, better results on IWSLT than WMT).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#3|15", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "It would have been good to test the combination with dropout or something similar.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#3|16", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "The hyper-parameters for SAGE are exhaustively tuned, more so than for the baseline adaptive optimizers, so there’s a potential for bias.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#3|17", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Figure 7 counters this possibility, but only by showing heatmaps, so it's hard to gauge the actual numbers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#3|19", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Table 3 should show results for your implementation of Adamax that’s comparable to Adamax-SAGE, in addition to the Devlin et al numbers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#3|20", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Figure 2: Consider flipping these plots to the standard convention, and stacking them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#3|21", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Figure 4: Why do we care that local temporal variation is lower in SAGE than the baselines?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#4|13", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Architecture: The method is evaluated only with transformer architectures.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#4|14", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "In the context of image classification it would be interesting to evaluate it with CNN and others architecture than ViT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#4|15", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Optimisation: In image classification, the pre-training procedure used is quite sub-optimal since the paper of Dosovitsky et al.[1]", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#4|16", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "many improvements have been proposed such as the DeiT approach [2].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#4|17", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "It would be interesting to see if the proposed method still works when the model is trained with more regularisation and data-augmentation which may also lead to a better use of weights.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#4|18", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "Image Classification: Having only results of models pre-trained on ImageNet-21k and fine-tuned with the proposed method on downstream tasks is not very usual in image classification.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#4|19", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "It would be interesting to have results on ImageNet only where the SAGE method is used during the training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#4|20", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "[1] Dosovitskiy et al, An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, ICLR 2021", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc37|review#4|21", "source": "rc", "split": "test", "paper_id": "rc37", "comment": "[2] Touvron et al, Training data-efficient image transformers & distillation through attention, ICML 2021", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|14", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The content bias analysis is based on a synthetic dataset generated by GPT-2, so that the summary inference was performed on out-of-domain data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|15", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "If summarization models are sensitive to paraphrasing and structural features, then synthetic documents may also exhibit artifacts that influence the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|16", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "It would be important to perform a similar analysis on selected documents from CNN/DM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|17", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Alternatively, a human evaluation experiment using the synthetic data could provide more evidence that the bias is really introduced by the summarization models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|18", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The position bias analysis, the authors state that \"clearly amplified by MatchSum, PreSumm, and Azure, as shown in Fig. 4b.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|19", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "However, it is not clear if the observed pattern is caused by model preference or simply by truncation of inputs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|20", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "If we observe the figure in Appendix E, the distribution for PEGASUS and BART are more similar to the reference summaries distribution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|21", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "In contrast, MatchSum, Presumm, and Azure decrease the frequency sharply between 0.4 and 0.6.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|22", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Interestingly, Presumm and MatchSum use the same backbone model with a maximum input length of 512 tokens, whereas PEGASUS and BART support inputs of 1024 tokens, which might explain their capacity to select more content from the later positions of the articles.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|23", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "For reference, CNN/DM input documents have 766 words on average (\"Extractive Summarization as Text Matching\", Zhong et al., 2020).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|24", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Do the quantile calculation in the figures take into account this truncation effect?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|25", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "No details or citation is provided for the paraphrasing model. Just the link for its checkpoint at https://huggingface.co/tuner007/pegasus_paraphrase.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|26", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Authors mention that they \"conduct manual analysis on a randomly selected subset of the article\" but no further details are provided.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|27", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Experiments are based only on CNN/DM, which is well known for its lead bias for important content.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|28", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Experiments on additional datasets could make the claims stronger.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|30", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Typo in page 6: \"Overall, t hese results show a pattern of unpredictability...\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|31", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|32", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The work well-motivated and addresses a relevant problem in language generation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|33", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "However, the methods lack clarity in some respects (like the paraphrasing approach) and some conclusions are not strongly supported by empirical evidence (see weaknesses 1 and 2 above).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|34", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The authors mention the release of dataset and code, which helps in the reproducibility.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#1|38", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "For the lack of novelty and solid evidence for the claims, the recommendation is for the paper rejection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|11", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Only used single embedding word2vec (word embeddings can themselves be baised) and only considered one dataset CNN/Dailymail, without explaining any rationale behind the selection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|12", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "To add value, authors should have explored other text summarization datasets with different writing styles to really bring out the different style biases rather than just paraphrasing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|14", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The paper lacks any association/description between the metrics used for bias in this paper v/s different text summarisation metrics.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|15", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "It would have been nice to consider why and how these metrics are different to give a better understanding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|16", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Paper lacks any results about biases based on religion and race even though they claim to have done so in the initial section of the paper.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|17", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Only one incomplete figure about religion is present which is hard to comprehend as it lacks any legend.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|18", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The results and figures need to be explained further.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|19", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Explaining the ideal score for an automatic summarizer in the figures would have given a better idea.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|20", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "I did not really understand the dashed lines in the figure given as Baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|21", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "An explanation with an example will go a long way in making things much more clear and bringing out the results effectively.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|22", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "First it claims no work has been done on bias-ness in text summarization then it points out a previous work on structure biasness (Jung et al, 2019).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|23", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "In the paraphrasing of articles, the authors do not show how good the paraphrased sentences are.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|24", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "This raises the concern that if the paraphrased version of sentences itself were curated in a manner that they did not include proper information.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|25", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "A human evaluation of a few paraphrased sentences would have cleared this issue.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|26", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Also there are no examples from the synthetic dataset curated by the authors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|27", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "In the Fig.3, two of the extractive summarisers are not mentioned namely presumm and matchsum", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|29", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Clarity The paper is not very clear, there are a lot of loopholes that need to be fixed mainly concerning the explanation of the author’s ideas and implementation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|30", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Quality The quality needs improvement in terms of considering more experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|31", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Novelty The work is novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|32", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Reproducibility Yes it seems reproducible as they have provided the codes and datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|34", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The paper lacks comprehensive evaluations using different embeddings, models and datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#2|35", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Further, the paper can be hard to comprehend due to lack of explanations of figures and curated examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|14", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The exposition is confusing, several details are missing and assumptions are made without declaration.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|15", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "There is no qualitative analysis to help the reader better follow how the authors are changing structure or content, or what a synthetic summary looks like, or what does it mean to be an article about men only but still have sentences about women in them (Figure 3) or whether the authors' several perturbation strategies aren't altering valuable information?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|16", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The paper relies on the premise that a certain way of choosing a summary (say weighing what comes earlier in an article more), given an article is wrong.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|17", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "But it does not show what the alternative would be and whether that alternative is indeed more preferable to an end user.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|18", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "It is hard to say that given two different texts to an end user (since these are different as they've been reordered), they would somehow still write the same summary for both articles.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|19", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "There is also no end user analysis showing that these biased summaries from a model that rely more on what comes earlier in the article are offering less value to the end users than what a counterfactual unbiased summary would offer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|20", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Experiments in Figure 2 come close to this point but it does not come clearly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|21", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Summarization is not an objective task, so the gold standard is to show resulting summaries to humans and see whether they prefer one model or the other.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|22", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "But that's hard as the authors don't characterize what such an unbiased summary would be?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|23", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The paper appears rushed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|24", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Some examples: Figure 7 has no legend, Table 3's caption says some text is in red and other in blue, but the only color there is yellow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|25", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "On Page 3, its said that the generation details for the synthetic corpus are in Section 5, on Page 4 its said that they are in Appendix B, but it seems they are in neither.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|26", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Appendix B comes close to sharing some information but after reading it, I could still not tell how someone could replicate this synthetic data generation because details are sparse.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|27", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Also, a footnote in this paragraph says that more details are present in the Appendix?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|28", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Several claims have been presented as facts but no citations or data are offered to support them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|30", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Introduction: \"Readers, however, expect that summaries faithfully represent articles.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|31", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "First, what do you mean by faithful in this context?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|32", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Second, if this is the case, please add at least one behavioral science citation to this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|33", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Footnote 3: What notion of fairness does this relate to?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|34", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Section 3.3: \"While we could expect a human to generate a summary...\" What is the evidence to support this expectation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|35", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Is there any data, any pilot study you did, any previous work that looks at this?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|36", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Section 5: \"We conduct manual analysis...\" No details are shared on this analysis and no results are presented.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|37", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "I have read the experiments and results section several times, yet I can't tell whether I would be able to replicate the setup and reproduce these results given how sparse the details are.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|38", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "There are error bars in the plots.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|39", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "What do these represent?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|40", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|41", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The paper is not easy to understand, details are sparse and it is hard to judge the validity of the experiments without having a complete view of the setup.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|42", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Some of the insights are novel but I can't validate their correctness as there are not enough details in the paper to help me understand what is going on in each experiment.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|43", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Reproducibility is also difficult for the same reason.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#3|44", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "The Supplementary material is a zip file of zero bytes that does not expand so I can't run any code either.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|11", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "I'm not convinced whether the proxies/metrics for measuring content bias are accurate", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|12", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Lack of some human validation", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|13", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|14", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Overall, the paper is straightfoward to follow, the novelty is great and the analysis is interesting and reproducible with reasonable", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|16", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "I think the paper would be much stronger if the metrics were validated using some human analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|17", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "For example, can you select some gro", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|18", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "ups and show that humans find the representation score appropriate?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|19", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Moreover, can you select 10-100 samples that are both analyzed by", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|20", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "humans and the automated scores and show that the scores match the human evaluation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|21", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "There have been quite a bit of evidence that bias metrics at the embedding level do not follow extrinsic measures of bias and such hum", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|22", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "an validation would make the results more convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|23", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "On the Intrinsic and Extrinsic Fairness Evaluation Metrics for Contextualized Language Representations Intrinsic Bias Metrics Do Not Correlate with Application Bias", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|24", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Is there a reason why word2vec was used to create the representation scores (and not something newer, some LM-embeddings)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|25", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "I have a few suggestions for improvement:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|26", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "In the intro, when describing types of biases, show examples, much easier to follow with examples (and the reader is motivated to st", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|28", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Expand Table 1 with examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|29", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "List challenges for bias estimation in text summaries.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|30", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Include a validation for metrics used as explained above", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|31", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Fig 1: Do a and b differ in the order in which the groups appear in the text (i.e., a men first, women second and b women first, ben", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|33", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "I don't seem to find Footnote 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|34", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "Fig 7, no legend so not sure what the bar/colors/lines mean", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc38|review#4|35", "source": "rc", "split": "test", "paper_id": "rc38", "comment": "UPDATE: Thank you for your response, I will maintain my scores.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|3", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "One of the challenges is the max number of tokens that the model accepts, so they study multiple alternatives.", "reviewer_type": "human", "label": "error", "error_type": "Inaccurate Summary", "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|7", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The systematic study of the structure of the prompt, including influential related work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|8", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Focus on the number of digits to study the generalization", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|9", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Study of kind of errors in the proposed prompt strength the proposal", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|11", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The interference found is a negative result in an area poorly understood.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|12", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "It's entirely possible that a slight variation of the prompt managed to fix the issue.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|13", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "I'm not ready to accept strong interpretations of \"X prompt didn't work\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|14", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "I suggest lowering the tone of such findings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|15", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The jump from addition to combination with subtraction is relatively small at an abstract level: just four combinations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|16", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "This could become a strength if presented as a simple change that becomes challenging for the LM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|18", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The study of the composition of skills gets confusing around multiplication.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|22", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The work is as original as it can be given this is such a popular topic.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|23", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The paper didn't mention if the code would be distributed, but I assume so.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|24", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "End of the day, it's a set of templates.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|25", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "My main concern is presenting some blurry methods as contributions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|27", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "I'm not convinced that the experiments in GSM8k add clear insights beyond addition.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|29", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Perhaps there is a missing lesson on how the limitation of the number of tokens can be alleviated by composition.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|33", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Given how much space it took to increase the accuracy, we have to wonder if the max token is a solid ceiling for more precise reasoning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|34", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Page 6. \"Length 8\" in the caption of table 2 probably refers to multiplication.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#1|38", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "I recommend the acceptance of the paper based on the positive findings, while I recommend interpreting the results without assuming that LLMs would be able to achieve high performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|4", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The underlying idea of showing a language model how to break down “complex” data manipulation is quite interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|5", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "From the results, the model seems to benefit from this prompting technique.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|6", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "I found the paper somewhat hard to follow and the writing style rather confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|7", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "For instance, I did not understand what “defining algorithms as skills” means exactly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|8", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "This is mentioned several times in the manuscript but it looks to just boil down to using the aforementioned prompting technique.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|10", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The main drawback of the manuscript is that technical details of the experimentation are largely missing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|11", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The authors do not mention explicitly how many examples they provide in the prompt for each task, whether it is only one or multiple.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|12", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Details of the baseline are not provided either, such as which prompts are used, making the comparison between different ideas harder.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|13", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Section 3, evaluation metric paragraph -- “We measure [..] OOD performance [..] where the model sees shorter examples at training time and generalize” -- missing s, typo -- ”to longer problems at test time”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|14", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "This should be rephrased, there is no training/test time as long as a learning problem is not formulated and executed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|15", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Usage of space is not well-balanced, the authors dedicated a full subsection for explaining “two-number addition” (section 3.1) and only a few paragraphs for all the others.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|16", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "In particular, the algorithm for parity is not even mentioned but its results appear in table 2.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|17", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Section 4 shows that we can “teach” multiple problems at a time, such as subtraction and addition.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|18", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "This looks to perform worse than showing each problem separately, therefore I am not exactly sure what conclusions can be drawn from there.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|20", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The fact that this prompting fails on logical tasks (section 6) poses serious concerns on its applicability on harder problems other than simple mathematical operations and, therefore, its real-world utility.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|22", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Clarity can certainly be improved (see above).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|24", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "I would like to raise a concern about reproducibility of the experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|25", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "I wanted to reproduce some of the experiments but found out that OpenAI’s Codex code-davinci-002 model is not publicly available yet, which is somewhat awkward.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|26", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The best I could do was to try reproducing OOD results with the text-davinci-002 model, which exhibited similar reasoning processes but I got results (for 14 digits) that are much lower, i.e. 30%, than what is shown in Figure 3, i.e. approximately 97-98%.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|27", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "I tried with 10 samples where each addend has 14 digits, while providing the model with a single example of summation with 3-digit numbers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|28", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Perhaps the authors can explain why this happened?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|29", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Is that only due to a model difference or too few examples in the prompt?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|30", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Please find here a txt file showing the prompt I used (the same as in section G.2) and the response I got from the text-davinci-002 model: https://file.io/vqUWwuH8uZpq", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#2|32", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "While I find the prompting scheme interesting, I don’t find the overall contribution to be relevant for ICLR.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|8", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The method makes sense conceptually and demonstrates significant performance gain upon standard scratch pad/CoT in multiple aspects.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|9", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The performance improvement for length extrapolations is especially impressive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|13", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The algorithmic prompting method conceptually requires that we know how to break down the target operation (like parity) into recurrent steps and provide traces.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|14", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "With this level of specificity required, it is no longer necessary to use prompting to solve the tasks (we already solved it with our brain when providing the prompt and can totally write a program for it).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|15", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Some generalizations (other than length) from the demonstration to the actual task would make the method much more useful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|17", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The skill composition could also be very useful if the demonstration of combining skills A and B can be generalized to combine A and C or even B and C.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|18", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The length extrapolation intuitively only works because the recurrent step is the same for all lengths, so LM can follow the demonstration in a straightforward and local way.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|19", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Although this is arguably possible for all operations, it could lead to a very large state to carry over in each recurrent step.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|20", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "In fact, even for just addition, the copying of a large number of digits in context repetitively already seems very expensive given the limited context window.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|21", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The second-pass strategy seems a promising start to handle this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|22", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "Maybe the demonstration can just only demonstrate the recurrent step from f(n + 1) to f(n), so that each recurrent step can be generated in a separate context like a recursive call?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|23", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "For the final tool use demonstration, what is the advantage of <call_add_algo> over calling the injected calculator as designed in the original dataset?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|25", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The proposed method is built upon Chain of Thoughts and scratchpad line of ideas but the specific method is novel to the best of my knowledge.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|26", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "The claims are supported by thorough experiments and results under each setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc39|review#3|30", "source": "rc", "split": "test", "paper_id": "rc39", "comment": "But so far it is unclear whether this technique can be scaled to more complicated problems", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|15", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "No major weaknesses identified.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|17", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "It would be interesting to see the variance observed when applying parameter-efficient methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|18", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Previous work found large variance in downstream performance across different seeds when performing full fine-tuning [1,2,3].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|19", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "This is something you could add, for example, to Table 2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|21", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "In Figure 2, the performance of Adapter is 21.00, not 20.46", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|22", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "The \"v\" subscript is missing in Eq.5", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|23", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "[1] Dodge et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|24", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping.2021", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|25", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "[2] Mosbach et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|26", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "On the Stability of Fine-tuning BERT: Misconceptions, Explanations, and Strong Baselines. ICLR 2021", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|27", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "[3] Bugliarello et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#2|28", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Multimodal Pretraining Unmasked: A Meta-Analysis and a Unified Framework of Vision-and-Language BERTs. TACL 2021", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|24", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "There are no standard deviations for any results in the paper, which makes it hard to assess significance of many results, esp.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|25", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "since some of the performance gaps are small.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|26", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "For instance, in table 6, authors have a 0.4 BLEU discrepancy in their replication of full fine-tuning performance but draw conclusions on MAM being best based on a 0.2 BLEU gap.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|27", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "The gap is bigger when comparing MAM to methods not introduced by the authors, but still.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|28", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "The story makes sense but I am not 100% confident in the robustness of the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|29", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Using two decimal numbers for tasks (e.g: XSum) also gives a false impression of precision.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|30", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "I understand fine-tuning on some of these tasks (MT / XSUM) can be resource-intensive but having standard deviations in even a subset of the experiments would be useful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|31", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Interpretation of Figure 4 in Section 5.2 and effectiveness of MAM Adapters for encoder models", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|32", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "The authors highlight that while existing methods perform well on MNLI/SST, they are underwhelming on en-ro / XSum.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|33", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "They conclude that this means existing parameter-efficient transfer learning (PETL) approaches are not great for higher-resource / more challenging tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|34", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "However, I would point out that this data can also support a different conclusion: existing PETL approaches work well for encoder-only models but are not great for encoder-decoder models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|35", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Including the T5 datapoint is also not very relevant since in that case Superglue is treated as a single task (w/ one adapter).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|36", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Answering “For which tasks / architectures do PETL methods perform well” is an interesting question in itself and I feel like this section expedites this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|37", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "It is likely that both the architecture and #datapoints in the task matter.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|38", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "The author’s choice at the end of section 4.2 is to focus on XSum / en-ro MT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|39", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "It would be great to highlight that not only are those higher-resource tasks, but they are also generative ones, and thus all the conclusions of this paper might not apply widely to encoder models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|40", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Indeed, the results of MAM adapters in table 2 are quite mixed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|41", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Right now the paper is claiming more generality than deserved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|42", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Two possible fixes here: (1) Make claims less general and specific to enc-dec models, (2) Conduct more experiments on encoder-only models", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|44", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "The writing is subpar right now.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|45", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Example: section 4.4: “its counterpart at attention” -> “its attention counterpart”, “FFN can better utilize modification at larger capacities” -> modifications or “modification of the FFN is better at larger capacities”, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|46", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Minor: The citation style is often wrong for the sentence.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|47", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Use \\citet more often.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|48", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Minor: Replace “For classifications” with “for classification tasks” in the Appendix (twice)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|49", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Typos: Section 5 name Discussions -> Discussion ; Appendix A3 Learning -> Learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|51", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "For Figure 2, I feel like the full fine-tuning number should come from the original baseline, not the replication.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|52", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "The protocol in 4.5 for choosing the scaling parameter for Scaled PA is a bit surprising.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|53", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "I would just suggest that it should be an hyperparameter, the current description makes it seem like it is based on hyperparam tuning for LoRA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|54", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "It seems to me that 4.5 is about composition function and \\delta h functional form since LoRA changes both.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc40|review#3|55", "source": "rc", "split": "test", "paper_id": "rc40", "comment": "Revised score upwards after response, see below", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|8", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "(Unclear) How is the proposed GBST different from convolution, despite the scoring part?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|11", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "(Mixed) Charformer achieves worse predictive performance than subword models BERT and T5, which seems to suggest the proposed GBST gives little lift from character-level to subword-level.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|12", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "However, the Charformer models presented are smaller in size or faster in speed, or more efficient in FLOPS.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|13", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Thus, hopefully, it will provide people with more choices for their use cases.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|14", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "(Mixed) Experimental results do not show strong improvements on accuracy metrics.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|15", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Table 1-3 shows that Charformer is generally on par with Byte-level T5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|16", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Sometimes Charformer wins; other times Byte-level T5 wins.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|17", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Either with small differences.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|18", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "In the multilingual case (Table 4), we do see Charformer consistently outperforms Byte-level T5 with a small margin.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|19", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "In all cases, Charformer is significantly worse than subword models.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|20", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "The rescaled version Charformer_S is able to achieve comparable and sometimes better accuracy than subword models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|21", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "However, the re-scaling is not tied to Charformer and can be applied to other character-level models as well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|22", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "It is not clear if the gain is because of the proposed GBST or solely due to the re-scaling.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|23", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "(Con) It is not clear if the proposed GBST module learns meaningful subword tokenization.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|24", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Better qualitative evidence, e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|25", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "examples that highly scored subwords align with human intuition, or quantitative evidence, e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|26", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "how the learned tokenized subwords align with established methods, or how they are better at solving lexical tasks, would help.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|28", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "It would help if the authors can provide more and clearer technical details.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|29", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "If limited by pages, the author can point readers to appendices.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|30", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Make sure the main text still contains the necessary details when moving things to Appendices.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|31", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "(Con) No code provided.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|32", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Providing code will greatly help reproducibility and help clarify many technical details not fully described in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|34", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "How is the proposed GBST different from convolution, despite the scoring part?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|36", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "For example, what is the Transfomer stack used by Charformer? It seems like T5, but I don't find it mentioned anywhere.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|37", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "See other localized points below.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|38", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "All the tasks in the paper seem to be classification tasks.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|39", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Is Charformer encoder only? If so, how do the authors ensure a fair comparison with encoder-decoder models, e.g. T5? If not, is the decoder part of Charformer character-based, or it also uses some notion of subwords?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|41", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Eq (1). What is used as $F$ in the experiments?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|42", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "In \"Considering Offsets\" How is 1D convolution applied to X?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|43", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "How does it save computation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|44", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Figure 2. How to interpret this heatmap for block size > 1?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|45", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "How are they aligned?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#1|46", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "For example, we see a high score for block size = 3 at \"k\", is it for subword \"tok\" or subword \"ken\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#2|13", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "This way the representation captures information from nearby n-grams.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#2|16", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "My main concern is that the proposed method does not seem to be satisfying the criteria for a proper tokenization method.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc41|review#2|17", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "First, it does not learn to segment the input to different chunks to be fed in the contextualizer but rather downsamples the sequence in a fixed way.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#2|19", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Second, the soft selection at each position is performed over blocks of different size that correspond to that position and does not take into account global information e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#2|20", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "by computing scores of all possible segmentations to maximize for the most likely one.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#2|21", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "The current framing would suggest comparison to existing tokenization methods like BPE or recent optimized ones to the downstream task [1,2] in a controlled setting (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#2|23", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "In addition, there is no proper evaluation of the learned segmentations other than the visualization of the learned weights for a single example.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#2|26", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Studies that focus on tokenization like the ones cited below evaluate on tasks where tokenization is important for reaching state-of-the-art performance like machine translation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#2|27", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "It might be worth evaluating there.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#2|29", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "In section 8.2 about monolingual datasets, was the downsampling rate optimized only for the proposed model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#3|13", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "(minor) The technical novelty of this paper is only GBST, a dense layer over n-grams, while training setting follows ByT5, and the best SBase model is also an implementation of the deep encoder concept.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc41|review#3|14", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "The scalability of the model is not shown.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#3|15", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "All the experiment focusing on models (other than mT5) with a parameter scale of 200M.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#3|16", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "But the paper ByT5’s smallest model has 300M parameters and performs much better than with 200M parameters (comparing Table 4 of ByT5 and Table 4 of this paper).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#3|17", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "(minor) T5 seems much better than SBase under Zero-Shot settings, though it takes more computation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#3|18", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "From the last sentence of 3.1 Baselines, it seems that the ByT5 model used for comparison is unscaled, and I couldn’t find in the paper about structure of this ByT5 (sorry if I missed that).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#3|19", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "This is important since in the ByT5 paper all the model variants have much heavier weight in encoders, and there is also an ablation study to prove that heavier encoder can improve performance a lot.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#3|20", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "In this paper, the scaled Charformer-SBase is also much better than the unscaled Charformer-Base.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#3|21", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Thus it’s a bit strange that the best settings of both models are not used for comparison.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#5|13", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "It is not clear if the method works for other languages.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc41|review#5|15", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "The evaluations are based on base models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#5|16", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Since this model saves model size, would it achieve better performance compared with the same size (e.g., 200M like Byte-level T5 Base in Table 1)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#5|17", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Do you try large models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#5|18", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "Besides, it would be impactful if this paper showed that GBST module could actually generalize in other kinds of models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#5|20", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "1. I wonder if average pooling is good way to down sample though it is really fast, but how about MLP?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#5|21", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "2. In Eq.3, the indices should be 0 to M-1 or 1 to M (total number of block size is M), i.e. $P_i = \\softmax([p_{0, i}, p_{1, i}, \\dots, p_{M,i}]$ -> $P_i = \\softmax([p_{1, i}, p_{2, i}, \\dots, p_{M,i}]$", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#5|22", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "3. At Figure 1 (b), it seems the annotation `P5:8` in the fourth line is missed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#5|23", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "4. In section 2.1.2, block sizes 1 < b < M should be 1 <= b <= M", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc41|review#5|24", "source": "rc", "split": "test", "paper_id": "rc41", "comment": "5. In section 2.1.5, “latent subwords $\\hat{X} = [\\hat{X}_i, \\dots, \\hat{X}_M]$ should be “latent subwords $\\hat{X} = [\\hat{X}_i, \\dots, \\hat{X}_L]$", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|0", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "(This review is a collaboration between a senior and a junior reviewer. The junior wrote most of the review, and the senior modified the review. Both reviewers read the paper in detail.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|7", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "This paper is well written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|8", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "This work is timely because it shows that increasing the model size and pre-training data size is not the only way to achieve strong performance on language related tasks (which is the current trend).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|9", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Inductive biases like knowledge about entities can improve the performance of models to the point of achieving competitive results with bigger models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|10", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Some limitations of this work include the lack of experiments on more diverse architectures like encoder-decoders to make sure that the proposed technique not only works on autoregressive GPT-type models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|11", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Moreover, only one approach for entity integration is explored.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|13", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Interesting choice of the margin loss for entity prediction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|14", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "More motivation should be added into why this loss instead of the cross-entropy loss.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|15", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Optionally, an experiment comparing the performance of the model when trained with the cross entropy -vs- the margin loss would be interesting to have.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|16", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "At inference time, you are forced to put entity labels on the input text as a preprocessing step.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|17", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "How much does this impact the generalization capacity of the model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|18", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "How do you manage cases when an unseen entity is seen as input?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|19", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Are the experiments on QA really zero-shot?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|20", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "The examples presented in Appendix are not dummy as they reflect true questions and answers?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|22", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Have you tried your pre-training technique on an encoder-decoder architecture like T5?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|23", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "It would be interesting to see that this technique not only works with auto-regressive models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|24", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Have you tried to fine-tune your resulting pre-trained models on various tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|25", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "It would be interesting to see if your pretraining method also helps finetuning more efficiently on some tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|26", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "In addition, I would be curious to see if, while fine-tuning, the entity prediction accuracy and the LAMA probing performance of the model drops or remains more or less the same.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|27", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "[1] Sun, Yu, et al. \"Ernie: Enhanced representation through knowledge integration.\" arXiv preprint arXiv:1904.09223 (2019).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|28", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "[2] Ji, Yangfeng, et al. \"Dynamic Entity Representations in Neural Language Models.\" Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. 2017.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|30", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Thanks for the response!", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|31", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "After reading other reviews, I feel the novelty of this work is somewhat limited and it would be useful to highlight the differences with respect to previous work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|32", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Also a discussion on when your method is preferred over existing proposals.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#1|33", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "I am also decreasing my confidence score after reading other reviews.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|5", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Strengths: The idea of using entity information is certainly interesting and new.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|6", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Most details of the paper are clear from the writing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|7", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Weaknesses: The paper leaves open several questions that should really be explored and discussed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|8", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Given that the modeling contribution is simply adding an extra embedding layer, it would be good to have detailed analyses that provide more insights into the contribution itself. (see points 5,6,7 below)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|10", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "1. The details of the dictionary lookup are not entirely clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|11", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Do you lowercase all tokens?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|12", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Do you handle small misspellings in the tokens?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|13", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Do you perform fuzzy matches?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|14", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "2. Thanks for providing several implementation details.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|15", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "However, the values of hyperparameter $\\alpha$ are not discussed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|16", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Does this require a lot of tweaking or is the method stable to this choice?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|17", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "3. (minor) The ‘zero-shot’ experiments really seem to be ‘few-shot’ tests since you provide the model with 8 or so examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|18", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "4. The metrics used for the LAMA knowledge probing should be explained in the paper once (at least expand the abbreviations so readers can understand).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|19", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "5. Table 3: The parameter sizes indicated are misleading.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|20", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Large KALM has close to 700M params and not 300M as indicated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|21", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Similarly base KALM has ~550M params not 100M.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|22", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "6. Related to the above point, it seems like KALM base has more parameters than GPT-2 base (and even GPT-2 Large) and yet this aspect is not accounted for/investigated in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|23", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "In fact, the paper says “(KALM) more efficiently packs practical knowledge into its parameters … 12L KALM is very close … to 24L GPT-2 Large”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|24", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "This seems to be very misleading information.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|25", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Since Transformers are essentially a form of graph networks, I’m not sure entity embeddings can be treated separately from the model parameters since one can view those as edges part of the graph.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|26", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Concrete suggestion: Can you compare KALM vs GPT-2 (in all tables and Fig 1) while controlling for the #params?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|27", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "To me, the only closest comparison seems like KALM base vs GPT-2 Large where GPT-2 Large seems to be the better model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|28", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "7. I appreciate the authors efforts to test on a variety of benchmarks and I understand that training these huge models is not easy, but to me, the essence of the paper (i.e.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|29", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "addition of entity information) seems under-explored.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|30", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "For instance, could you perform ablations studies that only uses the entity embeddings in the input to see how much knowledge it provides?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|31", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Another ablation can be on the matching done in the dictionary lookup (in terms of frequency cutoffs, n=1,2,3,4 in the n-gram matching, etc.).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|32", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "One could also analyze the prediction accuracy of the model to see which types of entities (e.g. persons, organizations, etc.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|33", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "work better vs those that don’t help as much.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|35", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Update: Thanks to the authors for their response, which helped clarify several of my minor questions and I believe those can be revised with a writing pass.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|36", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "However, I still think this paper has two significant deficiencies:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|37", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "The parameter size comparison still seems flawed to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|38", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "The authors say that one can discount the entity embeddings, but can we really?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|39", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Aren’t they part of the model’s ‘representation’ even if the inference does not use all of them in a single forward pass?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|40", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Several neural net architectures exist (including large-scale LMs) that do not use all inference paths but still count them in the total params.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|41", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "The BERT vs GPT-2 example provided in the rebuttal is only a difference of 26k tokens (still significant!)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|42", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "but here we're talking about a few hundred million parameters!", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|43", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "At the very least, I think the true sizes of the models must be acknowledged and one can add the point on entity embeddings vs 'brain' parameters as a caveat, but it seems scientifically inaccurate to me to claim otherwise.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|44", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "To be clear, I don't think the size issue detracts much from the main contribution of adding entity embeddings here (i.e.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|45", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "this work may still be of interest even if the size of model is larger), but the current version of the paper has several claims about size savings that seem incorrect.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#2|46", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "The analysis of the proposed method (where it helps, where it fails, which hyperparams matter) is still lacking.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|4", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Since the main point of the paper is adding entity information to the models, the paper would be better to include more explanations about the entity information.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|5", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "For example, with the dictionary look-up, what is e_i exactly?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|6", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Why did the authors use such knowledge rather than other knowledge like POS or so?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|7", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Why do we need negative sampling?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|8", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "How many classes are there for entity?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|9", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "The models enjoy more information provided by entity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|10", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Then what about more data to provide more information?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|11", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "How much can we improve the performance by adding entity information rather than additional data samples?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|12", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "The authors said that the models were trained with entity information from scratch.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|13", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Can we use pretrained models, since we can simply add entity information to the input and output layers?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|14", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "It seems like that the additional information can fine-tune the pretrained models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|15", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "What is the role of the dummy QAs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#3|16", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "And how did the authors come up with the 8 dummy QAs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|6", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "The authors propose a simple (i.e. without changing model architecture) method to give knowledge-aware signal during pretraining.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|7", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Proposed method consistently outperforms GPT-2 on LAMA and zero-shot QA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|8", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Interesting to see that entities from fuzzy frequency-based matching can make LMs significantly better.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|10", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Need extra parameters to save entity embedding (~471M).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|11", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "However, as the authors mentioned, this can be viewed as an external memory module and embedding lookup doesn't need much computation.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|12", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "I would like to see how much KALM can transfer to downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|13", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "The authors already show zero-shot QA results, but it would be better to see fine-tuning results on knowledge-intensive tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|14", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "For instance, will KALM still outperform GPT-2 after finetuned on KILT tasks (https://arxiv.org/abs/2009.02252)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|15", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "(I know KILT released right before the due date of ICLR, but tasks in KILT are all already published)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|16", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Even though KALM outperforms GPT-2, it is still unclear for me the advantages of KALM compared to Entities as Experts (or EaE, https://arxiv.org/abs/2004.07202).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|17", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "For example, EaE outperforms KALM on the LAMA probing tasks, and EaE also outperforms T5-11B when finetuned.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|18", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "The authors mentioned fuzzy frequency-based entity matching is better than precise entity linking (e.g. EaE) on page 3, but EaE outperforms KALM on LAMA probing tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|19", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Based on the experimental results, precise entity linking seems more effective to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|20", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "It would be helpful if the authors can provide a more thorough comparison with EaE.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|21", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "- Typo? Page 3, \"United States\" at $w_{i:i+2}$ --> \"United Staes\" at $w_{i:i+1}$ ?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|24", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Thank you for preparing detailed responses, which helped me to clarify several questions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|25", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "However, one of my major questions is still unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|26", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Although 4 out of the 11 KILT tasks are already included in the main paper, most of them are LAMA knowledge probing tasks or zero-shot QA tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|27", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "It is still unclear how much and how robustly KALM can transfer to other downstream tasks with fine-tuning (e.g. Wizard-of-Wikipedia, FEVER, QA with fine-tuning).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|28", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "For instance, CorefBERT paper (which uses a similar idea but as you mentioned it use bidirecitonal attention) shows its transferability on QUOREF, six extractive QA benchmarks, DocRED, FEVER, five coreference resolution benchmarks and GLUE, which can convince me to choose CorefBERT over BERT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|29", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Experiments on the paper are promising but not diverse enough to make me choose KALM over GPT2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc42|review#4|30", "source": "rc", "split": "test", "paper_id": "rc42", "comment": "Thus, I would like to stick to my original score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#1|5", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "I believe this can be a useful technique but before considering such an approach as a general strategy, more theoretical insights should be provided.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#1|6", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "The authors report on ablation studies that demonstrate some empirical benefits.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#1|7", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "However, until I see more theoretical analysis on how the method improves convergence or lead to better losses by smoothing out the output of the objective function, I remain skeptical of the usefulness of this as a general training method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#2|3", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "The idea of dynamic temperature scaling has been tried in other works and tasks (e.g., attended temperature scaling).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#2|4", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "The paper parameterizes this mechanism with DNNs for the language model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#2|5", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "Though the idea looks interesting, it fails to explain why the scaling is better than other dynamic temperature scaling frameworks.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc43|review#2|6", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "The experiments are not solid.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#2|7", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "The baseline only includes Mos, which is not very strong.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#2|8", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "To validate whether this approach works with other LM of high-order attention or self-attention, a better baseline model is required (e.g., transformer, GPT).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#2|9", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "I would like to see this technique can help either NLU or NLG tasks, instead of just pure modeling.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#2|10", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "The case analysis section needs more examples instead of just cherry-picking few.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#3|4", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "Overall I found the paper well-motivated and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#3|6", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "The analysis is also interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#3|7", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "I vote for an acceptance, if the authors can polish the writing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#3|10", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "- Another view of the proposed method is that it learns a context-dependent weighting of the tokens in the vocabulary, such that \"important\" tokens (those with smaller \\tau) receive more gradient updates.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#3|11", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "Can the authors comment on this?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#3|13", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "- Adding onto above. [1] discusses the low-rank bottleneck of using a single softmax.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#3|14", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "Since elementwise matrix product can blow up the rank, how do the authors think the proposed method can serve as a more efficient way to deal with the softmax bottleneck?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#3|15", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "- Last but not least, the paper can be improved a lot if the authors can thoroughly polish the writing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc43|review#3|16", "source": "rc", "split": "test", "paper_id": "rc43", "comment": "[1] Breaking the Softmax Bottleneck: A High-Rank RNN Language Model. https://arxiv.org/abs/1711.03953", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|7", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "The approach is reasonably straight forward, although difficult to estimate the impact of a variety of hyperparameters and choices (some of which are varied and shown in later in the paper) at each stage due to the complexity of fine-tuning such a large model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|10", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "The method is only broadly applied to the 540B model that almost no one else has access to/can train or fine-tune.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|12", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "What happens if you don't have access to such a large model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|13", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Can you apply the entire method to a smaller model say < 10B?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|14", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "It'd be great to see if these approaches can be used with language models that are available to everyone and at scales that most of the community can use.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|15", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Replicating this is nearly impossible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|18", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Quality/Novelty: The authors integrate chain-of-thought prompting and self-consistency as means for self-improving large pretrained language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|19", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Technically this basically just takes tools from other work and integrates them, but the evaluation and the utilization is reasonably novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|20", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Originality: I think many people in the community over the years have approached a self-supervised, self-improvement goal for AI systems and NLP systems, but this specific approach is relatively original and to my knowledge has not been tried before.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|21", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Reproducibility: Access to the pretrained model is not available, and the sheer size of the models make reproducibility impossible for the vast majority of labs, universities, and companies alike.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|23", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "The paper is good.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|24", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "The method is relatively straight forward and empirically the method performs well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|25", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "The paper is clear and has additional experiments such as distillation that make it more compelling.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|26", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "There are limitations in that the models here are of squillions of parameters and not publicly available, so its nearly impossible to replicate.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|27", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "The method has not been tested on models that are of reasonable size or ones that are publicly available, which is a significant limitation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|28", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Given the compute resources at the authors' disposal, running a few additional experiments on relatively much smaller and publicly available models would strengthen the paper greatly and would be of significant interest to the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#1|29", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "In general this paper is of broad interest to the NLP subcommunity at ICLR, so I vote that it be accepted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|9", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Prompting is a popular technique for querying LLMs because of their size and training them is not efficient/possible in most scenarios.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|10", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Various prompting strategies have been developed in the past that have improved the performance of LLMs significantly, like Chain of thoughts, appending “let's do this step by step”, and so on.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|11", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "An obvious extension was to generate multiple times using the same/different strategies, and select the one with higher confidence or cluster them or do some majority voting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|12", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "This paper uses all the methods of the past and proposes an obvious extension that takes the confident samples and trains over them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|13", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "This is not novel at all with many researchers have thought about this but could not execute the idea because of a lack of resources, higher API costs, and non-availability of open-source LLMs like PaLM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|14", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Secondly, proposing an approach and claiming that “This is similar to how a human brain sometimes learns: given a question, think multiple times to derive different possible results, conclude on how the question should be solved, and then learn from or memorize its own solution” is just random with not theory or study to back this up.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|15", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "In other words, it is saying that humans can look at a problem, and think of a solution many times, and the one that humans conclude the most is the most likely output and use that to think and train themselves to memorize that solution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|16", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "I don't agree with this point at all (of course I am happy to listen to why authors thought about this).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|17", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Research where such an important claim is being made (LLMs are able to self-improve) needs to be shown on multiple models (small to very large) and just showing the results on a 540B model does not prove the point that large LMs can self-improve by training on the samples that they produced.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|18", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "When generating outputs with let's say a smaller model like T5-large, where the accuracy will not be that high, the model outputs cannot be used to train them as most of the time, the answers will be incorrect.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|19", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "So my guess is that this approach only works for PaLM or similar-sized models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|20", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Finally the limitation of the experiments on one model (and that too not open sourced) is not enough to justify the big claim of \"large language models can self improve\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|21", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "The reproducibility of the results is a concern to many.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|25", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "The paper is original in its implementation but not in ideation (very obvious extension that was not conducted by a lot of researchers due to lack of resources).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|26", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "However, the results cannot be reproduced due to the lack of open source policy for PaLM by google and there is no way to test if the results are accurate or not.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|29", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "This claim is very not supported by the experiments where the results are shown for just one LM and that too is not openly available.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|30", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "So the validity of the method cannot be tested/justified by the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#2|31", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Moreover, the approach is not novel in its ideation but only novel in its implementation (as the method was not implemented by others due to resource constraints) and hence I would suggest showing the novelty of the idea on more models and datasets to justify the claims.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|10", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "The paper presented an approach followed with empirical evidence, but did not provide why such an approach works in the first place (i.e., approach is not backed well enough).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|11", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Following are my concerns on the approach:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|12", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "The approach fine-tunes on the rationale-augmented reasoning paths using self-consistency, which means that self-consistency outputs without LMSI (fine-tuning approach) will be on par with CoT-Prompting with LMSI.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|13", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "This is more or less consistent with the observations in main Table-3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|14", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "As per my understanding, the model is not learning any new information, the proposed approach is acting like a filtering method where the results achieved by a self-consistency can be achieved by fine-tuned LMSI in CoT setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|15", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "I think the self-consistency results for LMSI can be achieved by using more samples for self-consistency without LMSI.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|16", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "These experiments are not there in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|17", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Like I said, there were a lot of ablations in the paper where some of them proved some points presented in previous work (like Table-7) and others are new (Figure-3).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|18", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "On the new ones, the paper failed to provide good discussion on the outcomes/observations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|19", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "For example, in Figure-3, both step-by-step and few-shot with step-by-step are converging for total samples at 80.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|20", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Why is that the case?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|22", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Please provide more discussion on whether self-improving a characteristic of large language models or it applies to any model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|24", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Any meaningful discussion would be helpful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|25", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Can you check if baseline self-consistency with more examples can match LMSI or not?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc44|review#3|27", "source": "rc", "split": "test", "paper_id": "rc44", "comment": "Overall, the paper has many interesting experiments but lacks proper evidence and discussion on why this approach works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|14", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "The promise of a better sentence scorer is potentially very impactful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|15", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "The authors chose to evaluate sentence re-ranking on multiple generated candidates for NMT and for ASR.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|16", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "I have several concerns about these experiments that made them not particularly convincing for me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|17", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "First, the experiments were, in my view, are a bit thin.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|18", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "For instance, the gain from SLM over BERT is not very large, and there is still a large gap between the oracle.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|20", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "As some examples, it could have been good to compare with T5, BART, XLNet, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|21", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "I am also not sure about the setup for the ablations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|22", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Why were the ablations not performed on all the datasets?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|23", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "I think this could have been addressed by putting a final row in Table 2, with SLM with one-third of the compute so that the compute is matched.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|24", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "It is fine if it does not win in all settings under the same compute, but in this case it would be good to be transparent about how much compute is needed to get the same performance, and whether there are performance gains in multiple different compute-matched settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|25", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Finally, I am not sure why NMT and ASR were chosen as the experimental settings in the first place.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|26", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "It could have been stronger if the paper had chosen a language understanding benchmark like GLUE, which could be a better fit for pretrained language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|27", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "I have voted for rejection primarily because the experiments are not particularly convincing, but if enough of the questions/concerns that I had are addressed in the rebuttal I will be happy to change my score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|29", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "The gains from SLM in the experiments required tuning a hyperparameter on a dev set and selecting the best value to use on the test set.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|30", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Was the same hyperparameter value usable across all datasets?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|31", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "In what hyperparameter range was the model performing better than the baselines?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|32", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "What happens if we don't tune this hyperparameter?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|33", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Why was NMT and ASR chosen instead of general language understanding, which how LMs such as BERT and GPT are mostly used today?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|34", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "If transcormer does better in those settings, then the impact of this work will be much greater.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|35", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "For instance, sentiment analysis can be formulated as scoring two sentences: \"the sentiment of <review> is positive\" and \"the sentiment of <review> is negative\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|36", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "This is called \"rank classification\", and this is how GPT-3 is evaluated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|37", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "While it is not necessary to re-train a GPT-3 size model, even on the small scale some experiments on language understanding tasks like these could improve the experimental validation of the paper immensely.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|38", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "I might have missed it but will the code be made publicly available to reproduce the experiments?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|39", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "This might be important to gain adoption for this method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|40", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Since there is a high-cost of re-training from scratch, is it possible for existing language models to be adapted for this new objective?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|41", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "This might enable more compute-efficient experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|43", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "The authors were up-front about the additional computational cost of their method, though additional experiments would be needed for the empirical gains in compute-matched settings to be convincing for me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#1|44", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "The societal impact part seems fine.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#2|7", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "The main idea seems like fairly straightforward variant of the XLNet architecture, not too dissimilar from other past proposals to modify its base architecture by changing the nature of the \"streams\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#2|14", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "It would be useful to have dataset and split sizes for the experiments of section 4 directly in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#2|15", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "IWSLT14 might be a small dataset, but I don't think German, Spanish, Italian, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#2|16", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "should be described as \"low-resource\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#2|18", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Could you clarify what the \"oracle\" line in Tabels 2 and 3 represents exactly?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#2|20", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "I do not see any potential negative societal impacts of this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#2|21", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Limitations are sufficiently addressed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|5", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Post-rebuttal: I raise soundness to 4/4 and my score from 4/10 to 6/10.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|6", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "The work is thorough and introduces an architectural innovation to more richly model $P(y_s|y_{\\backslash s})$ in one pass.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|7", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Unlike the paper originally claimed, it is not the first such model (T-TA and Electric); also, unsupervised sentence scoring remains niche (MLM scoring is more for inspection of pre-existing models, and for downstream tasks most works use discriminative ranking).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|8", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "However, the authors have done many experiments, it performs better than MLMs and previous models, and may have applications elsewhere (analogous to XLNet, also published at NeurIPS).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|9", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "There are still many grammar errors; please work with a copyeditor or native English speaker on any camera-ready.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|11", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "The authors succeed in proposing a scheme that models token probabilities conditioned on all other tokens (Eq.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|12", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "1), in a single pass, while preserving \"deep bidirectionality\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|13", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "The method appears sound, and while the mechanism is complicated, Figures 1 and 2 do well to explain it (though Figure 2b should have labels saying which axis is layer $l-1$ and which is layer $l$).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|14", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "However, I have concerns re: the clarity and significance of this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|15", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Firstly, there are two uncited works that have proposed one-pass solutions to token-wise probabilities:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|16", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "T-TA: J Shin, Y Lee, S Yoon, K Jung, “Fast and Accurate Deep Bidirectional Language Representations for Unsupervised Learning”, ACL 2020", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|17", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Electric: K Clark, M-T Luong, Q Le, C Manning, \"Pre-Training Transformers as Energy-Based Cloze Models\", EMNLP 2020", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|18", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "In particular, the first work's proposed T-TA has many similarities:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|19", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "a query initialized with position embeddings only", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|20", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "a query that is updated through to the end and is selectively blocked on the self position", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|21", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "while having some benefits (no 3x in cost?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|22", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "and disadvantages (the keys, values do not transform over layers).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|23", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "The authors should also acknowledge [18] in section 5.3, where an almost identical analysis is performed, even down to the choice of length and plot (Figure 3 here is largely [18]'s Figures 3+4).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|24", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "While the authors compare on some tasks (NMT, ASR) as prior work, it would be good to compare on the same datasets or even n-best lists.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|25", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "For example, [18], [20], and the above two works compare on a shared set of n-best LibriSpeech lists while this work creates new ones.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|26", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "It is also unclear e.g., what L20's \"quality of a sentence\" or what L23-24's \"precise sentence scoring\" mean.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|27", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Only in Section 2.1 (page 3) do we get two definitions:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|28", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "via CLM, which by the chain rule gives the log-likelihood score $\\log P(y_1, ..., y_{|y|})$", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|29", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "via MLM, which gives the pseudo-log-likelihood score  $\\sum_{i=1}^{|y|} \\log P(y_i \\mid y_{\\backslash y_i})$", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|30", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "SLM models the latter.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|31", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "It would be better to make this clear early, and then mention how/why these quantities are empirically (or even theoretically?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|32", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "good quantities to rescore with (so before in 26-38), with the second being better (which is why(?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|34", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Discriminative rescoring should also be discussed/contrasted early, or at least in the main text (not the appendix).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|35", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "I appreciate some of the analyses which complement past work, in particular A.3.1's discussion comparing sentence vs. fixed-length pre-training, and 5.3 showing SLM and MLM's similar scores.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|36", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Section 5.1 and 5.2 are unsatisfying though, as increasing the number of masked tokens would be expected to be poor given the mismatch with pretraining (in BERT, masking 15%).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|38", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "The word \"sliding\" evokes sliding windows (i.e., only a contiguous, fixed length of inputs are considered at any time), giving the wrong impression", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|39", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "L195-198: The structure of bi-LMs should be explained more", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|40", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "L163: should clarify that $p_i$ refer to position embeddings", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|41", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Writing should be proofread, e.g.,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|42", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "L18 and elsewhere: \"language modelings\" --> \"language modeling approaches\" or \"language modeling schemes\" or even \"language models\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|43", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "L85: \"in a chain-style rule\" --> \"via the chain rule\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|44", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Figure 2: \"reuse\" --> \"reuses\"; \"can not\" --> \"cannot\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|45", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Table 5, Appendix Table 5: \"test-clear\" --> \"test-clean\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|46", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "L233, 236, Appendix Table 5: \"LibrSpeech\" --> \"LibriSpeech\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|48", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Could you compare and contrast your approach with other one-pass models (T-TA, Electric)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|49", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Given only the sum of probabilities is ever used and since one must pretrain anyways, have you compared with the stochastic estimation or distillation methods (L96-L97)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|50", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Alternately, do you have an application where token-wise probabilities are essential?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|51", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Will you release your pretrained SLMs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|52", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "L199-200 (SLM for scoring, CLM for NLG, MLM for NLU) is not really explained -- are there experiments for this?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|54", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "since MLM and SLM model the same quantity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|55", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Have you considered ways of matching MLM and SLM costs that avoid MLM pretraining mismatch (and how do you formally define cost), e.g., a shallower MLM, lower attention dimensions?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|56", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Alternately, what if the MLM were pretrained with masked sets?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|58", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "While SLM gives token-wise probabilities out of the box, it does require pretraining from scratch and is rather specialized (whereas MLM and CLM have other uses).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc45|review#3|59", "source": "rc", "split": "test", "paper_id": "rc45", "comment": "Hopefully more uses can be explored to justify this cost; see works above for examples, like acceptability (BLiMP) or text similarity (STS).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|7", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "The method generally assumes that the decision of drop can be done independently.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|8", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "However, the positions of salient information would usually exhibit certain patterns such as appearing close to each other.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|9", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Do you have any idea to incorporate those characteristics into the method or learn those patterns?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|10", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "As the name of the method is SpanDrop, it would be much interesting if various selection/split methods of spans have been investigated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|11", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Of course, splitting by a sentence for passages is natural.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|13", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "I wonder about the additional training cost of SpanDrop compared to the standard training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|14", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "The expectation of length is the same for SpanDrop and Beta-SpanDrop.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|15", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "However, the expectation of length’s square would be different.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|16", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Therefore, I guess their computational overhead will be also different.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|17", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "The sampling cost would be almost negligible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|18", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Does it require additional training steps to converge due to its regularization effect?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|19", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Don’t it require modification of other regularization such as standard dropout?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#1|20", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Do you have any thoughts about whether SpanDrop will be generalized seamlessly to other tasks than NLP?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|9", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Several technical details are not discussed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|10", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "How to partition a long sequence into spans? For text sequences? For time series? Will the final results be sensitive to the partition of spans?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|11", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "If each word in NLP sequences is a span, the proposed method is the same as the word-level dropout baseline studied in previous works, e.g., [1].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|12", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "What's the tech novelty of this work?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|13", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "While the authors focus on the \"learning problems for long sequences where not all input elements contribute equally to the desired output\", the uniqueness of this kind of problems is not clear to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|14", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "In most (if not all) real-world sequence classification and prediction problems, not all input elements contribute equally to the desired output; otherwise, the problems will become much easier.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|15", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Experiments need to be improved and enhanced, e.g.,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|16", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "As reviewed in the second paragraph, there are many Transformer variants proposed for long sequences, but none of them is compared.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|17", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Although these approaches aim to approximate the original pairwise interaction with lower cost and are often interested in still capturing the interactions between every pair of input elements (e.g., the long sequence benchmark proposed by Tay et al., 2020), they can still be directly applied to long sequences where not all input elements contribute equally to the desired output.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|18", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "The four NLP datasets used in experiments are not representative tasks for long sequences.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|19", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "I suggest to test on the public benchmark datasets, e.g., [2].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|20", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "[1] Soft Contextual Data Augmentation for Neural Machine Translation, ACL 2019.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#2|21", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "[2] Long range arena: A benchmark for efficient transformers. In International Conference on Learning Representations, 2020.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#3|12", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "The procedure is fairly simple, and similar ideas have been used for regularizing models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#3|13", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "I find the solution similar to word dropout, and would be interesting to see comparison to that.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#3|14", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "The gains on real datasets are not high.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#3|17", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "It is not clear if similar benefits can be obtained by using word dropout at training time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#3|18", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "The experimental results on real datasets are weak.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|5", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "In experiments, consistent improvement is shown.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|8", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Effective data augmentation is a relevant topic.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|12", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "My major concern is the lack of comparison of other baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|13", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "In section5, the author discuss a bunch of related works on data augmentation, however, in natural language experiments these are not compared.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|14", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "I know that SCANDROP is easier to implement, but that does not mean you don't need to compare with other algorithms.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|15", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "It's not surprising that this data augmentation is effective, especially for the catfinding task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|16", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "While the improvement over electra-base is consistent, it's not large.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|18", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Sec1. I don't quite understand why when in data the contributing portion is small or sparse, then the problem is \"underspecified\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|19", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "It's still specified by the small contributing portion I think?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|20", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Are you suggesting the contributing part is so small, that for example, a classifier can not reach a decision?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|22", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "If you replace a span with a random token, it gives counterfactual sequence with the same length.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|24", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "Sec1. \"the data does not sufficiently define the goal for statistical models\", I don't understand what global is here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|28", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "(2) The lack of comparison of other baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|29", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "In section5, the author discuss a bunch of related works on data augmentation, however, in natural language experiments these are not compared.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc46|review#4|30", "source": "rc", "split": "test", "paper_id": "rc46", "comment": "I know that SCANDROP is easier to implement, but that does not mean you don't need to compare with other algorithms.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|5", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Overall, I like this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|6", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "The idea of using unit tests for transpilation is, although not formally guaranteed, a reasonable approach for transpilation in my opinion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|7", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Some of the language in the paper is a little off-putting as it seems to be disconnected (in my opinion) from deeply understanding the space of natural language translation (unstructured ambiguous languages) vs. programming language translation (structured unambiguous languages).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|8", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "It's my hope that if this paper is accepted, that the authors will polish the writing slightly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|9", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "For example: \"translation systems are not as effective for source code as for natural languages.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|10", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "This is a weird claim.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|11", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "The work that Alvin Cheung (Berkeley) and JRK (MIT) for Maaz Ahmad (UW Washington / Adobe Research) is wildly effective and has automatically transpiled ~300 functions with a geometric mean performance improvement of 3.36x.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|12", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Moreover, this approach is formally verified unlike the authors approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|13", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "In addition, it's already being used in production quality code (Adobe Photoshop v.21).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|14", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "There are many other such examples that seem to demonstrate the authors perhaps need to do a bit deeper literature review before making such bizarre claims.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|15", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "For example, the word \"transpilation\" never even appears in the paper, yet the whole paper is about transpilation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|16", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Do the authors not know about the field of \"transpilation\" and that they are working in this space?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|21", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "It's a reasonable (although not formal) way to perform transpilation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|22", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Using self-supervision is likely one of the most important aspects of the system because 99% of the code that we have today is currently unlabeled.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#1|23", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "To my knowledge, only a handful of labeled datasets exist (e.g., POJ-104, Google Code Jam, IBM/MIT's Project CodeNet, Microsoft's CodeXGLUE, etc.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|23", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "If ground truth parallel data were available for this task, then fine-tuning a model on this data would naturally result in gains in performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|24", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Since this parallel data is extremely scarce, the authors propose to utilize an automated test generation suite to build this parallel data from the model translations itself.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|25", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "I have 2 concerns with this proposal:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|26", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "It requires an automated unit test generation tool to be available for at least one of the languages translated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|27", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Unit tests generation is an active area of research and tools are available only for limited sets of languages and primarily for modern languages such as Java, Python, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|28", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "An important premise of automated code translation -- which the authors mention in their introduction and ethical concerns section -- is about modernizing legacy languages such as COBOL, where automated test generation systems might not be available at all.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|29", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "The proposed method therefore, in my opinion, is bottlenecked on the availability of automated unit test generation tools, and is not applicable when they are not available.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|30", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Automated unit test generation tools can themselves generate incorrect test cases -- test cases testing the incorrect method, incorrectly testing the method resulting in false positives or false negatives.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|31", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "AthenaTest [1], for instance, works better than EvoSuite and generates correct unit tests for the intended method only 16.21% of the time as per [1].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|32", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "26.71% of the times the test case generated by AthenaTest test for the wrong behavior.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|33", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Other factors such as dynamic typing (as in Python) [2] or the use of more complicated data structures (such as Generics in Java [3]) further affects the effectiveness of these test generation systems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|34", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "It would have been nice to get some insights into the quality of test cases generated and how this affects the performance of the translation model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|35", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "This effect, I believe, would be especially pronounced in cases where the code translated is a piece of more complicated codebase and not standalone functions that can be compiled in isolation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|36", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "The authors utilize EvoSuite more as a test data generation tool than a test case generation tool.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|37", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "EvoSuite generates test cases for Java, while this work trains models for Java, Python and C++.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|38", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Therefore, if I understand correctly, the way test cases are created for Python and C++ is by using the expected input/output pairs and assert statements generated for corresponding Java code, and putting these in a template for the other languages.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|39", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "I'm not completely convinced that this is an ideal strategy that can generalize to a broader set of languages or for more complicated pieces of code.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|40", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "For example: if the source code implements a Java class and the generated test operates on this class and their methods, how would this be translated to Python test without changing the procedure for Python test case generation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|41", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Or, how would this procedure work when there isn't a direct mapping of data types and constructs between the languages.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|42", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "This further reduces the applicability of this approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|43", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "If I understand correctly, the evaluation used in this work includes a reference implementation and a translated implementation, which are tested to generate the same output given the same input?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|44", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "This creates problems as authors note in Figure 4, where the gold translation is equivalent to the source code only on a small domain, while the model generated translation, though correct and better is incorrectly marked wrong.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|45", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "This, to me, indicates an issue with the utilized test set where gold implementations might not be correct translations, thereby affecting the scores.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|46", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Contemporary works in code generation and translation [4,5] have utilized expected input / output pairs to test the performance of the code generation system, and I wonder why the authors here chose to compare the performance against a reference implementation and not against input-output pairs directly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|47", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Additionally, we don't know how prevalent the problem of incorrect reference translations is in the utilized test set.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|49", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "The authors mention that they collected data from GitHub with permissive licenses.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|50", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Can the authors please share the list of licenses used for this data collection step?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|51", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Having this information in the appendix will provide a reference point for future work in this space on which licenses might be acceptable to use when working with GitHub data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|52", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "[1] Tufano, Michele, et al. \"Unit Test Case Generation with Transformers.\" arXiv preprint arXiv:2009.05617 (2020).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|53", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "[2] Lukasczyk, Stephan, Florian Kroiß, and Gordon Fraser.\"Automated Unit Test Generation for Python.\"International Symposium on Search Based Software Engineering.Springer, Cham, 2020.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|54", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "[3] Vogl, Sebastian, et al.\"EVOSUITE at the SBST 2021 Tool Competition.\"2021 IEEE/ACM 14th International Workshop on Search-Based Software Testing (SBST).IEEE, 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|55", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "[4] Hendrycks, Dan, et al.\"Measuring Coding Challenge Competence With APPS.\"arXiv preprint arXiv:2105.09938 (2021).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#2|56", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "[5] Chen, Mark, et al.\"Evaluating large language models trained on code.\"arXiv preprint arXiv:2107.03374 (2021).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#3|5", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "The paper is well written, easy to follow and presents an interesting and novel idea.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#3|6", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Technically, the work touches many parts - “static” text-to-text transformer, executing code, checking tests passing, observing and optimising test coverage.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#3|8", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "What would have been more interesting though is how well would it work with differences in the test generation/acceptance strategies.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#3|12", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "One of the biggest limitations of the approach seems to be around how unit tests are translated between the languages.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#3|13", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "The paper downplays this step in its expositions by simply saying that simple scripts can translate unit tests.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#3|14", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "While I agree that unit tests tend to be inputs and outputs for which translation should be easier, this is probably limiting the scope of what can possibly be translated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#3|15", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "The authors could include an appendix or discussion on how many of the tests could be translated successfully and what incorrect translations look like - e.g. tests that perform a sequence of APIs or expect exceptions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#3|16", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "A related question is if tests could be translated with some of the baseline models such as Transcoder/DOBF and if such an idea was considered and was/wasn’t successful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#3|17", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "It is not completely clear if the data filtering techniques (to only include functions for which tests can be generated) also play a role in the selection of the evaluation data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#3|18", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "I would be happy to get a confirmation of this from the authors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|9", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Using unit tests to improve quality of parallel datasets generated via backtranslation is not entirely new in the ML for software engineering domain.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|10", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "There were attempts to use it for bug patching applications, see for instance, DeepDebug (https://arxiv.org/abs/2105.09352) paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|11", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "It is also frequently used during inference, as a postprocessing step, and evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|15", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Authors claim that the unit tests suites can be used to test program semantics in any programming languages, provided there is a way to establish the mapping between types and parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|16", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "While this is true in most cases, have you observed and analyzed some corner cases, for instance, when an object needs to be instantiated in the body of the unit test method prior to assertions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|17", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Instantiating such variables may rely on imported classes or require translating extensive surrounding code context other than the unit test itself.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|18", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "How often does it happen, are these cases ignored?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|19", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Starting with 333k Java functions, authors obtain 103k high-quality test case suites with EvoSuites which are then iteratively used to generate parallel training corpus.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|20", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Have you analyzed the failed cases: • Does the model successfully generates data for small methods only (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|21", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "a single statement), and majority of failed cases are long methods?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|22", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "How does it handle test cases with multiple asserts?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|23", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "• When does EvoSuite fail to generate high-quality test suites?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|24", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "I wonder if EvoSuite scaffolding is an issue", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|25", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Maybe I am missing something, but why would generated test set mismatch to TransCoder tests set be an issue?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|26", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Are you generating tests sets for evaluation as well?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|27", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Does this change the baseline (published) TransCoder computational accuracy numbers?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|29", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "For multiple published papers the authors cite corresponding arXiv pre-prints.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|31", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Miltiadis Allamanis, Marc Brockschmidt, and M. Khademi. Learning to represent programs with graphs.ArXiv, abs/1711.00740, 2018 -> Published at ICLR 2018", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|32", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio.Neural machine translation by jointly learning to align and translate.-> published at ICLR 2015 … etc", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|33", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "maximize a criteria such as code coverage or mutation score -> \"maximize criteria such as code coverage or mutation score\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|34", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "“Otherwise, the mutant is said to be survived” -> “Otherwise, the mutant is said to have survived”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc47|review#4|35", "source": "rc", "split": "test", "paper_id": "rc47", "comment": "A1 caption (right), should perhaps be changed to “A generated unit test suite”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|14", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Weaknesses of the paper:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|15", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The main weakness in this work is one that relates to the overall goal of fact extraction from language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|16", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The “Oracle'' results from Table 1 are thought provoking: with perfect knowledge regarding the predicate/relation of test examples, and a subsequent 100% consistent response, the LLM is only able to obtain ~50% correct responses from T-Rex, which is an admittedly limited evaluation (41 “head” predicates, mostly of well known entities).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|17", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "While I understand that this work is clearly focused on the consistency issue, not necessarily correctness, it puts into question whether fact extraction from LMs is a worthwhile pursuit.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|18", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "I would have liked for the paper to dig a little deeper into this headroom question from the previous point.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|19", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Would it be possible to conduct a sampled qualitative evaluation of errors of the Oracle model in the ID cases?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|20", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Are the errors due to unseen triples during training time (e.g., not in Wikipedia), or maybe there are issues with model capacity (maybe a 10x version of the LM would be able to recall the prompted fact)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|21", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "In terms of writing, the most confusing section in the paper is Section 4.1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|22", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "After re-reading it twice, I was still not able to ascertain: (1) what data was used to train the models, and (2) what data was used to evaluate the models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|23", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The section makes reference to LAMA’s T-REX, LPAQA, ParaRel, as well as augmentations using BERT lexical replacements, as well as data from “Shin et al, 2020”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|24", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The section also talks about examples from these sources as well as templates (presumably filled in with WikiData triples?).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|25", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "I really think this section needs to be rewritten and the training, eval and test datasets should be much more precisely described.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|26", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "I would also encourage authors to release the exact datasets and splits to allow others to reproduce/improve on this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|27", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "But even with a data release, a precise description of how this data was constructed is very important.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|28", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "For the MoE and Oracle layers, the description in the paper is insufficient to determine the outputs presented to the first layer of the model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|29", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The depiction in Figure 2 hints that the entire sequence is rewritten using the fixed-length learned embeddings, and perhaps the subject or MASK embeddings are preserved?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|30", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "But actually sub-section 4.2 never formally describes how the embeddings are used to create the continuous prompts?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|31", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Are they prepended/appended to the original inputs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|32", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Or do they rewrite the original inputs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|33", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Do either the MASK or subject tokens get copied?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|34", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The LAMA benchmarks have one unfortunate characteristic: since it was constructed for BERT-style single token prediction, it has stripped down the original datasets (see the original version of T-Rex, which contains over 600 unique predicates vs. the 41 from LAMA: https://hadyelsahar.github.io/t-rex/ and https://aclanthology.org/L18-1544.pdf ).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|35", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "I wonder if a more comprehensive version of this would be to evaluate on a larger sequence-to-sequence model like BART https://arxiv.org/abs/1910.13461 or T5 https://arxiv.org/abs/1910.10683 (both available as HuggingFace models).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|36", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Given that this work leverages frozen LLMs, it seems that training and evaluation could be done relatively cheaply even for larger models with proper decoders.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|38", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "With respect to the MoE solution, the paper claims that the model does not use a weighted combination and opts to use the top-1 predicted relation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|39", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "I wonder if authors have tried using a weighted combination instead?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|40", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "If the relation classifier is trained with cross-entropy softmax loss, most of the weights will be close to one-hot (similar to top-1) except when the model is uncertain.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|41", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Therefore combining prompt embeddings may yield some benefit over top-1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|42", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Does this make sense?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|43", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Note sure this is a good idea, but: given that the LLM is frozen, it seems plausible that the continuous prompt embeddings learned in some of the models resemble existing embeddings from the original vocabulary.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|44", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "As such, would it make sense to attempt to “decode” the continuous prompt embeddings into the existing vocabulary?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|45", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "One could use a greedy decoding strategy of extracting the nearest neighbor (via dot product or cosine distance) from each continuous prompt embedding to the vocabulary input embedding table.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|46", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Have the authors tried inspecting the continuous prompts in this way?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|47", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "I wonder if the output is informative or whether these prompts are modeling purely latent variables.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|48", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Typo in Figure 1 “Canada si” -> “Canada is”,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#1|49", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Typo in page 6: “Cannonical” -> “canonical”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|11", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The paper does not draw a connection to prior work on robust optimization (for example, see references in [1]), which offers a more principled framework for formulating the objective of invariance to a class of input perturbations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|12", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "At the very least, it would be good to give a formal definition of robustness/consistency and cite this line of prior work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|13", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The paper does not provide a clear justification for why P-Adapters would be expected to improve consistency, and nothing in the training objective encourages the model to be consistent.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|14", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The argument might be more convincing if you could compare it to a training objective that does explicitly promote consistency--for example, by encouraging the hidden representations of different prompt paraphrases to be similar.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|15", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "As this paper notes, the factual probing benchmark has class imbalances and represents a very particular use case, so while P-Adapters appear to improve consistency here it’s unclear if these results will generalize to other settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|16", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Would it be possible to evaluate P-Adapters on a wider variety of tasks, following prior work on prompt optimization [2, 3]?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|17", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "For example, [4] provide 100 prompt templates for a wider variety of NLP tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|18", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The paper would be more compelling if you could show that P-Adapters improve consistency in other settings too.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|19", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Much of the motivation of this paper is based on assertions about user preferences, but there is no human evaluation to validate these claims.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|20", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "For example, would a typical user prefer to give a natural language input but get an inconsistent response?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|21", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Or would they rather pick a pre-defined relation type with better promise of consistency?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|22", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "In particular, it’s not clear why user preferences would place any restrictions on training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|23", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "MoE model: How accurate is the relation classifier?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|24", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "There is a big performance drop in the “OOD Prompts” setting, which leads me to wonder if the relation classifier was adequately trained.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|25", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Either way, relation classification results should be included and discussed in the main paper--a perfect relation classifier would lead to 100% consistency.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|26", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "All of the models suffer a performance drop on the “OOD Objects” setting, which indicates that the models have over-fit to the (imbalanced) distribution of entities in Wikidata.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|27", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "It’s unclear how to interpret consistency in this setting, because the models are consistently producing the incorrect response.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|28", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "It might be more informative to report something like the “Consistent-Acc.” measurement from [5].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|29", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Table 2 does not compare results to any prior work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|30", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "For example, can these results be compared to the “Consistency Improved PLM” results from [5]?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|32", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The metric called “precision@1” should be called “accuracy@1”", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|33", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Table 2 does not explain how the results are aggregated (prior work takes either the micro- or macro-average over relation types).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|34", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The description of the methods and the illustration (Figure 2) are somewhat unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|35", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Is there one MLP applied at each position?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|37", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The meaning of the different-colored tokens could also be explained in the caption.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|38", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The paper is missing some implementation details, such as the size of the BiLSTM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|39", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The results section (section 5) might be easier to read if it were divided into subsections or paragraphs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|40", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "It would be interesting to see the accuracy breakdown by relation type.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|41", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "[1] Sinha, Aman, Hongseok Namkoong, and John Duchi.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|42", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Certifying Some Distributional Robustness with Principled Adversarial Training. International Conference on Learning Representations. 2018.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|43", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "[2] Shin, Taylor, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. \"Eliciting Knowledge from Language Models Using Automatically Generated Prompts.\" EMNLP.2020.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|44", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "[3] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. GPT understands, too. arXiv preprint arXiv:2103.10385, 2021b.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|45", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "[4] Gao, Tianyu, Adam Fisch, and Danqi Chen. \"Making pre-trained language models better few-shot learners.\" ACL.2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#3|46", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "[5] Elazar, Yanai, Nora Kassner, Shauli Ravfogel, Abhilasha Ravichander, Eduard Hovy, Hinrich Schütze, and Yoav Goldberg. \"Measuring and improving consistency in pretrained language models.\" arXiv preprint arXiv:2102.01017. 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#4|12", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Although the proposed method involves experiments in four different settings including OOD keyword error, it might be valuable to investigate an OOD syntax error setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#4|13", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "That is, in a real scenario, users write the natural language prompt which may have a grammatical error.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#4|14", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Assessing the robustness of the proposed method in terms of grammatical errors can enhance the quality of the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#4|15", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "In Table 2, the performances of P-tuning in RoBERTa-large are better than the ones of Oracle.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#4|16", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Without any explanation, it is not convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#4|17", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Instead of listing the result in the tables, it would be clearer if the authors provide more explanations about them although they are in the appendix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|24", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "research questions are dispersed all over the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|25", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "What I mean by this:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|26", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Abstract: you say that you tackle the inconsistency problem and you want to use a lightweight approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|27", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Hence, I can see that there are two research questions: 1) inconsistency of LLM and 2) efficiency - how to tackle the inconsistency problem with a light weight model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|28", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Introduction: you say that focus is the continuous prompts and user-friendly prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|29", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "These are another two research questions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|30", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Discussion: you hypothesising that the success of P-Adapter is due to not updating the LLM parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|31", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Such that it is likely to maintain all of factual knowledge learned during pretraining.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|32", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Which to me is yet another research question - addressing the inconsistency of LLM with frozen weights vs fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|33", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "In Introduction you talk about the frozen weights of LLM but do not motivate why you decide to keep them frozen until the Discussion section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|34", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Can you please state and motivate clearly your (full) research question somewhere in Introduction?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|36", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "(Minor) Why didn’t you train BERT fully with the proposed natural language prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|37", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "You emphasise the importance of this experiment yourself in the Discussion section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|38", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "(Minor) Why didn’t you train (fine-tune) BERT with P-Adapters?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|39", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "This experiment will also allow you to answer your hypothesis in the Discussion section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|40", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "(Major) Did you try to freeze the parameters of LLM (or use an adapter) in the MoE model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|41", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "You claim that MoE performs better than P-Adapters because of the additional supervision information (in form of relations) the model gets during the training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|42", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "With the present experiments it is hard to tell.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc48|review#5|43", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Maybe MoE perform better because you also tune parameters of LLM?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|11", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "From these experiments, it is unclear whether models are actually \"learning to follow instructions\" or just learning a very large space of tasks from the fine-tuning procedure.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|12", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "In other words, even though prompt variance is reported at inference time, the models could potentially perform just as well with nonsense or missing prompts during fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|13", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "As far as I can tell, no experiments that rule out this possibility exist.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|14", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Although qualitatively useful, the analysis in 4.1 does not conclusively show that the number of instruction tuning clusters aids performance, or that this trend is likely to continue with more clusters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|15", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Most of the gain could be acquired by tasks which are most difficult, or most similar to the heldout task, and this analysis cannot disprove such an interpretation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|17", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "The paper is missing important details about hardware usage and training time", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|18", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Some possible issues which might be resolved by the additional questions below", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|20", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "For each dataset, we manually compose ten unique templates that use natural language instructions to describe the task for that dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|21", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Do you have unique prompts for each dataset or only for each dataset cluster?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|22", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Based on a cursory look at the supplementary material, I would assume the latter.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|23", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "I didn't fully understand the justification for the OPTIONS token.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|24", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Are the fine-tuned models successfully putting (almost) all of their probability mass on the corresponding options?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|25", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "How is the Base LM evaluated (if it's not fine-tuned, presumably it doesn't learn how to handle these options)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|26", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Figure 6A: why does the untuned model see worse performance with more parameters?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|28", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Figure 1 (Bottom) is possibly misleading, since AFAICT zero-shot FLAN underperforms few-shot GPT-3 on the majority of tasks", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|29", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Not clear what \"turning the task around\" means for some tasks, or why this is a useful type of prompt diversity", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#1|31", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "I give this paper a strong recommendation, in spite of some missing ablations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#2|14", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Cons: There are still a few questions that can be addressed to make the analysis comprehensive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#2|15", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Have authors try to use the FLAN prompts on GPT3 or BaseLM and how does the performance look like?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#2|16", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Since instruction tuning will adjust all the parameters in the original pre-trained language model, there is a question what about what is the potential impact of this tuning process?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#2|17", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Will it drops any knowledge of any tasks, which will be a disadvantage when the task's labeled data is available?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#2|18", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "In the Analysis C in the appendix, it will be good to have results for tasks other than classification such as summarization or question answering; and also to have a baseline where the BaseLM model is fine-tuned directly with the task labeled data (without prompt/soft-prompt).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#2|20", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Overall, the paper proposed an interesting idea and showed strong empirical results, hence I vote for accepting", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|4", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "For each dataset, we manually compose ten unique templates: Why not have templates per task cluster instead of per dataset?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|5", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "it is likely a relatively minor effect given the results from Appendix B but it seems like it could slightly prevent overfitting", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|7", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Nit: I would have tried to move the (datasets per cluster/templates per dataset) ablation to the main body as well and shortened Section 3", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|9", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "In figure 6A, why was performance not increasing for untuned models w.r.t model size?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|10", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "This seems to contradict findings from Brown et al where larger models did better on essentially all tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|11", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Were there perhaps some poor datasets that happened to be in the held-in split (since the held-out tasks don't seem to have the same trend)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|14", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "That more templates per dataset didn't help is particularly interesting and suggests some questions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|15", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "You hypothesize that more templates doesn't help because \"models at such scale do not easily overfit to a finetuning single task\" - but my intuition is for an opposite explanation -- that the models at such scale easily memorize a small number of templates!", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|16", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "One may even wonder if the instruction nature of the templates is helping at all.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|17", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "From what I can tell, Appendix C on prompt tuning (which is very interesting) is maybe the primary evidence the instructions are important.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|18", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "I think more could be done here, some ideas, probably there are better ways to test:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|19", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Have templates that leave out \"instructions\": I would guess it wouldn't affect held-in task performance much, but would affect held-out tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|20", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Consider HellaSwag/PiQA/etc, where FLAN underperformed few-shot and even zero-shot.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|21", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "One might hypothesize that if using a (subotimal) template that is less natural for language modeling, that zero-shot performance would suffer, but that FLAN performance wouldn't", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|22", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "One might hypothesize that the \"turn the task around\" templates help more than the other more straightforward templates that don't swap information between the prompt and response.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|23", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Easy but probably not great thing to try: held-out tasks with wrong/useless templates", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|24", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "A final thought: It's not obvious that using as many training examples per dataset as possible is optimal, given that the model could overfit to dataset-specific spurious correlations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|25", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "This could be another area to investigate", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|27", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "UnifiedQA seems potentially worth citing as prior work", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|29", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Overall, the paper's idea is powerful (but of somewhat limited novelty) and the results are good (but not great).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|30", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Its greatest strength IMO was the ablations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#3|31", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "My biggest complaint is that it's not completely clear the instructions themselves are important at all - I suggest a few more experiments, though they don't seem crucial.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|8", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Below are my concrete questions and comments:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|9", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Additional Tasks Results (3.4)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|10", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "In Appendix A.1, the paper mainly draws conclusions based on comparisons between GPT-3 and FLAN, which I do not think are fair: GPT-3 and FLAN differ in model size and pre-training data distribution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|11", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Instead, I think Base LM vs. FLAN might be a better comparison between “Off the shelf LM” and “instruction-tuned” model (though it won’t change the conclusion).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|12", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "It is also worth pointing out in the main paper that for most of the additional tasks, even though it does not lead to higher accuracy, the performance of FLAN is still at least comparable (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|13", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "<1% worse and difference generally negligible) to Base-LM 0-shot.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|14", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "The only outlier seems to be ReCorD where the performance drops significantly after instruction-tuning, and this probably deserves some discussion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|15", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Also I might have missed it - for the Base LM 137B zero-shot result, is it on the average template, or the best template?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|16", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Number of Task Clusters (section 4.1)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|17", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "For Figure 5 can you add the untuned model to the curve with the x-axis=0 (0-task cluster)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|18", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "This can help us understand how much even 1 cluster (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|20", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Explanation for scaling (section 4.2)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|21", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "It is an insightful empirical result that instruction tuning only works when model size reaches 68B.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|22", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "However, I am not entirely sure about the potential explanation of “model capacity”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|23", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "There might be two potential explanations to this phenomena: 1) “model capacity”: as the paper has mentioned, smaller pre-trained models do not have enough model capacity and underfit the instruction tuning data, and 2) “better OOD generalization”: better quality pre-trained models have higher OOD generalization ability (and OOD accuracy) and they are less likely to “overfit” to in-distribution data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|24", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "I personally find the second explanation more convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|25", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "For example, Sahn et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|26", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "(https://arxiv.org/abs/2110.08207) finds that even models with only 11B parameters can generalize to unseen tasks, using T5 (MLM) and a larger set of prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|27", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "The use of MLM objectives (might) improve the pre-training quality, while more prompts reduce the “overfitting to in-domain data” issue.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|28", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "I appreciate the fact that the author explicitly states the model capacity hypothesis more as a conjecture rather than a solid explanation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|29", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "It’d be great if the authors can support the explanation further with more empirical evidence.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|30", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "On the other hand, however, since the results from Sanh et al came out only 2 weeks ago, I would not change the score based on the response to this question.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|31", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "In-context Few-shot vs. Fine-tuned Few-shot (Section 4.3)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|32", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Can the authors compare “fine-tuning/prefix-tuning an instruction tuned model with 16 examples” (appendix C but with only 16 examples) with “in-context prompting” (in 4.3 of the main paper), similar to Chen et al. (https://arxiv.org/abs/2110.07814 )?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|33", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "This would further inform us how we should use the few-shot learning examples for larger language models: put it in-context, or fine-tune?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|34", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Again, since the comparison of Chen et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|35", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "came out only 2 weeks ago and the paper limit is 9 pages, I would not change the score based on the response to this question.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|37", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Results of Appendix C are interesting and potentially impactful - this might imply that instruction-tuned models will become the new “base” model for the pretraining-finetuning paradigm.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|38", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Is it possible to briefly mention it in the main paper as well (and redirect the readers to the appendix to see the full results)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|39", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "It might be too late to change the name, but “Finetuned LAnguage Net” (FLAN) is uninformative, since it does not capture any unique aspect of this method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|40", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "What does “LAnguage” mean here, \"natural language instruction\" or \"language model\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|41", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "If it is the former, then directly including the word “instruction” might be better; and hopefully it’s not the latter, since even fine-tuned BERT on SST-2 counts as a fine-tuned language model ...", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|42", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Typo Intro: Instruction tuning is “a” simple method that, as ….", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|43", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Conclusion: Moreover, our work supercedes recent work such “as”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|45", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "While the method is not new, the empirical results are strong and comprehensive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|46", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "Though I disagree on the interpretation of some empirical results, overall the additional analyses bring us further insights on what method works for very large language models (i.e.> 100B dense model).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc49|review#4|47", "source": "rc", "split": "test", "paper_id": "rc49", "comment": "I highly recommend the paper to be accepted to ICLR 2022.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|0", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Summary Of The Paper:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|7", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "The paper is well-written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|8", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "The paper is supported by many experiments to verify the claims that the authors proposed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|9", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Besides empirical results, the author also provided theoretical points of view to support their claim.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|11", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "The claim that flatness can decide the downstream performance is not well-supported.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|12", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "There are many factors that can affect the downstream performance, for example, the model size.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|13", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "One possible explanation is that the flatness could be a consequence of scaling up the model size, and the good performance is brought by the large model size, too.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|14", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "If that is the case, then flatness is not the reason for good downstream performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|15", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "The authors need to have more evidence to prove that flatness is the main reason that leads to good performance, otherwise, they may be both the consequence of another factor (like model size).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|16", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "When computing Hessian, I assume that you are using the loss of pretraining datasets, not downstream datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|17", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Then how can the flatness on the pretraining task reflect the situation on downstream tasks?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|18", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "This claim is not very intuitive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|19", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "It may need more explanations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|21", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Clarity&Quality: The paper is clear, well-written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|22", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Novelty: As far as I know, there is no existing paper discussing the same topic.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#1|23", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Reproducibility: The authors didn't provide the code. I think it's hard to reproduce the results by ourselves.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc50|review#2|6", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "This work seeks to better understand an important question in representation learning: how does pretraining performance correlate with downstream performance?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#2|7", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "This question is adequately answered in controlled settings where we observe the non-correlation between pretraining loss and downstream performance and the correlation between solution flatness and downstream performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#2|8", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "The investigation is both well-motivated and nicely executed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#2|9", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "In addition, a theoretical result is provided in support of the empirical observations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#2|10", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "It is, however, unclear what the practical implications of this work are.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#2|11", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "First of all, current large language models are not in the saturation regime, and it is hard to estimate when they will be as datasets grow with model size in tandem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#2|12", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Second, while this paper points that pretraining loss is not a reliable indicator of downstream performance, a simple remedy is to evaluate on downstream tasks during pretraining and compare models accordingly, which is likely already done in practice.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#2|13", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Finally, this paper does not demonstrate if the insight gleaned in this work can lead to additional “flatness regularization” that induces better downstream performance on real datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#2|14", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "It is understandable that large-scale experiments are expensive and are not expected, but given the rather empirical motivation of the paper, some validation on real data seems desirable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#2|16", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "The paper is well-written and presents novel results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#3|6", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "given the above observation, it becomes straightforward to infer that a pre-trained model with parameters at a flat minimum would lead to decent performance on the downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#3|8", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "the trace of a Hessian matrix in the formulation from this paper is very easy to obtain and one doesn't even need to materialise the matrix. The issue is that the Hessian matrix is evaluated around the global minimum of a model being trained on the pre-training data, and, without any knowledge of the pre-training data, it becomes very difficult to compute the Hessian matrix.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc50|review#3|9", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "The other issue is that it does require many samples to obtain a reasonable estimate of the trace.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#3|10", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "the trace of the Hessian matrix is only an overview of the magnitude of all eigenvalues.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#3|11", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "A smaller trace could mean that the minimum might be flatter than other minima with the same loss, but it could also mean that the minimum might be steep in only a few directions, and completely non-informative in others.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#3|12", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "The latter case wouldn't necessarily give us a strong transferable model, and yet the trace of the Hessian matrix wouldn't be able to tell.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#3|13", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "maybe I am missing the point and I am happy to be corrected, but the arguments made in this submission or the observations obtained are already well-known empirically, especially in the case of zero-shot and few-shot learning with very large pre-trained transformer models.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc50|review#3|14", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "In addition, the theorems mentioned in the submission don't seem necessary, or, in any case, support the arguments or the observations.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|7", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "In the controlled experiments, the authors empirically show the correlation between flatness and generalization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|8", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Even though language model achieves optimal pre-training loss, we can improve generalization performance of downstream task as we continue pre-training, which results in flatter minima.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|9", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Authors provide some theoretical analysis that connects flatter solution to the pre-training loss and generalization performance of downstream tasks in the limited settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|11", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "If they are properly addressed, I am willing to raise my score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|13", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "My expertise not learning theory, but in my humble opinion, what the authors claims is too bold.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|14", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "They claim that they \"prove that SGD with standard mini-batch noise implicitly prefers flatter minim in language models\", but it is limited to the special case --- Dyck language experiment which is far from practical scenario.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|15", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "It would be better to tone down the claim.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|16", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "I am a bit confused about the correlation between flatness and generalization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|17", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "For example, Dinh et al., [1] show that we can build equivalent models but with sharper minima, which might contradict to what the authors claim, specifically theorem 5.1.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|18", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "However, the authors did not properly tackle this issue even in the related work section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|19", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "It is not clear how the size of models and flat minima is related to each other.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|20", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "The authors claim that smaller transformer architecture is a subset of the larger transformer architecture family, but it does not make sense.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|21", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "If we increase the width of transformer or increase the vocabulary size, the smaller model cannot be a special case of larger model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|22", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "It would be better to explain why larger model converges to flatter minima than smaller one.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|23", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "In the experiments, it is unclear why the authors compare vanilla sgd training with adversarial training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|24", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "I think adversarial weight perturbation [2] is more relevant baseline since it explicitly regularize the flatness of weight loss landscape.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|25", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Minor typo: In page 4, (2) training for different number of steps -> (1) training for different number of steps", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|27", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Regarding the flat minima, what happens if we explicitly enforce the language model to converge flatter minima?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|28", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "For example, we can explicitly regularize the trace of Hessian to be small or use SAM [3] optimizer to enforce such regularization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|29", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Why is the cross-entropy loss is typically non-zero at the global minimizers?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|30", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Why do you assume that we only mask a single token in a sentence for masked language model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|31", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "In practice, we usually mask 15% of tokens of a sentence (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|32", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "BERT [4], RoBERTA [5]).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|34", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "[1] Dinh, Laurent, et al. \"Sharp minima can generalize for deep nets.\" International Conference on Machine Learning. PMLR, 2017.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|35", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "[2] Wu, Dongxian, Shu-Tao Xia, and Yisen Wang. \"Adversarial weight perturbation helps robust generalization.\" Advances in Neural Information Processing Systems 33 (2020): 2958-2969.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|36", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "[3] Foret, Pierre, et al. \"Sharpness-aware Minimization for Efficiently Improving Generalization.\" International Conference on Learning Representations. 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|37", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "[4] Devlin, Jacob, et al. \"BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.\" Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|38", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "[5] Liu, Yinhan, et al. \"Roberta: A robustly optimized bert pretraining approach.\" arXiv preprint arXiv:1907.11692 (2019).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|40", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Although this paper tackles interesting research question, it is unclear that flatter minima actually leads to better generalizations since it does not compare relevant baselines and lack of justification.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc50|review#4|41", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Please see detain in strength and weakness section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|13", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Given this challenge, I am not entirely sure whether this GPT-2 fine-tuning or GPT3 generation is very novel for this paper to be accepted as ICLR paper.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|14", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "As generation from fine-tuned large LM is the main contribution of this paper, I suggest the authors to compare with other possible ways of collecting this in-domain data (such as kNN or generation without fine-tuning) to contrast and highlight their scientific contribution.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|15", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Thinking about other possible ways to generate in-domain data,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|16", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "This paper points out that it is difficult to collect in-domain data such as pair-wise sentences and tabular data is hard to get.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|17", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Q) For tasks that simply require single sentence as an input, what happens if you just use NYT, wikipedia corpus or if you do kNN given training set?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|18", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Q) For the pair-wise input (x1, x2), what happens if you simply do nearest-neighbor with x1, x2 given popular unlabeled texts?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|19", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Q) For the tabular dataset, I am not sure if pre-training of GPT2 really adds value, what happens if you just train separate LM for the tabular data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|21", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Have you considered using adapters instead of finetuning whole GPT2 which seems too expensive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|22", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "What would be the result like if you don't really finetune GPT2 and do prompt-based generation? (not for few-shot)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|23", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "How about using g(x) notation in figure 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|24", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Do you require such a large pre-trained model such as GPT2 for something like tabular dataset?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|25", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "I am not sure how GPT2 pretraining helps the generation of feature-like datasets such as tabular data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|26", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Maybe inappropriate to say outperforming XGBoost on 2 out of 4 tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|27", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "(I think showing gains are fine over transformers, but besides connect-4 in Table 7, I am not sure the comparison to the XGBoost is that meaningful. The gaps between XGBoost and RoBERTa are remaining similar)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|28", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Table 2, \"We see consistent improvements with more GAL iterations,\", I am not sure if MNLI and RTE show that?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|29", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "What are the two metrics (metric1/metric2 format) for MNLI, MRPC, STS-B, QQP?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|30", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "I recommend that the authors denote metrics for all their tables.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|31", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Any reason that table 6 has part of GLUE only?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|32", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "How big is one iteration of GAL?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|33", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Does it correspond to running through all 40x unlabeled data?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|34", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Grammar If we had access to the oracle generative process, we were able to obtain the best KD and SSL results —> we could / we would be able to (?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|36", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "I want to first thank the authors for writing a clear paper and conducting extensive experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|37", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "The experimental results show performance gains, however, I am not convinced that this approach is novel as it has been shown in previous work that self-training can improve performance and as we know that GPT-2, 3 can generate quality examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|38", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "To appreciate the result more, I recommend authors to actually focus on discussing challenges in finetuning generator given training data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|39", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "If they could improve the quality of generation in a novel way, maybe the paper could become more appealing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|40", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "At the current state, I feel like GAL is very straightforward self-training with the unlabeled data generation with large-LM generation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#1|41", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "(Additional comments: In case I missed some contributions of this paper, If we remove the good quality of GPT2,3 what are the benefits that GAL is bringing as a framework? Could you elaborate on why generating in-domain data is a novel idea?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#2|14", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "-- (probable) high computation cost", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#2|16", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "The fact that GAL appears in the GLUE learderbord associated with the name XXX is a major issue if XXX is indeed the author of this article.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#2|17", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "As the GLUE leaderboard is mentioned a lot of time, this anonymization leak is important.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#2|21", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "On Fig 1., authors could mention explicitely and/or graphically that GAL correspond to the bottom path (and not to the refining of the discriminative path).", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc51|review#2|25", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "The fact that GAL appears in the GLUE learderbord associated with the name XXX [to be discussed privately] is a major issue if XXX is indeed the author of this article.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#2|26", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "As the GLUE leaderboard is mentioned a lot of time, this anonymization leak is important.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#2|27", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "The authors do not discuss the computation cost of combining GPT2 & RoBERTa-large modeling.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#2|31", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "I propose to accept this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|9", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "It is also surprising to see GPT-j could generate quality data with given few-shot labeled data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|11", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "First, the authors explicitly mention the small amount of labeled data in Figure 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|12", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Most of the experiments are conducted under full supervision except Table 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|13", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "It is not very true that only a small amount of labeled data is available considering MNLI includes 400k examples and RTE also has more than thousands of labeled examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|14", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "This claim matters since data augmentation is much more challenging and important for low-resource scenarios.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|16", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Considering that this idea is widely explored in the existing literature, I am more interested in the most distinguishing part: the authors use the unconditional autoregressive lanaguage model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|17", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "However, this claim is only investigated in Table 6.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|18", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "This still leaves many unanswered questions, like whether this observation is only validated for autoregressive models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|19", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "especially considering that the previous papers[1] show class-conditional LM also brings sizable gains.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|20", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Even though authors review many related works, they did not include other data augmentation works as baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|21", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Since this area is widely explored, it is important to understand how much benefits can be brought by this method contrasting other data augmentation options.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|22", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "The model with data augmentation and without augmentation is only ablation studies.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|23", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Considering a GPT-2 is incorporated additional for data generation, it is not surprising to observe improvements.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|24", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Comparison with other state-of-the-art data augmentation is needed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|25", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Besides the large language model as a data generator[1], many works like EDA[2], UDA[3] also explore this idea.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|26", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "I have several additional questions:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|27", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "In table 1, what data is used for knowledge distillation in DistiRoBERTa + KD?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|28", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Since KD is usually based on unlabeled data, I am not sure how the unlabeled data is achieved?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|29", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "If KD is based on trained labeled data, this seems not a fair comparison.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|30", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "This can only show that using additional LM for data augmentation is useful and such a claim is provided in many existing papers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|31", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "I will suggest adopting other simple ways to generate unlabeled data like UDA[3] and still KD loss on them", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|32", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "I am not sure about the claim on class-conditional generative models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|33", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Since class-conditional generative models still generate noisy data, how about using pseudo labeling on class-conditional generated data?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|34", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "The negative effect seems to be brought by noisy labels instead of conditional generated data quality since codntional-class generation is also shown to be effective in a recent paper [4] with a more challenging setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|35", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "[1] Jason W Wei and Kai Zou. 2019. Eda: Easy data augmentation techniques for boosting performance on text classification tasks", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|36", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "[2] Kumar, Varun and Choudhary, Ashutosh and Cho, Eunah Data Augmentation Using Pre-trained Transformer Models. worshop AACL 2020", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|37", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "[3] Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, Quoc V. Le Unsupervised Data Augmentation for Consistency Training. NeurIPS 2020", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#3|38", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "[4] Zirui Wang, Adams Wei Yu, Orhan Firat, Yuan Cao Zirui Wang, Adams Wei Yu, Orhan Firat, Yuan Cao", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|13", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Adding some of the things that could improve the publication below:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|14", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "1 - For figure 1, a NIT comment is that adding the unlabeled data to the same level as the models made it a bit confusing at first glance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|15", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Maybe moving the \"open domain unlabeled data\" cloud more to the left might make it more clear where the whole process actually starts (maybe something similar for the \"small amount of labeled data\" part.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|16", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "2- As you present the GLUE results on table 1, you split the scores with a \"/\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|17", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Since this is a paper that spans several areas of AI, it might be good to give more explanations on why this is the case and what are the metrics being considered in all your tables.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|18", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "3 - For table 1 and 2, some standard deviation for the multiple runs or significance testing might be useful to compare results that are similar to each other.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|19", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "4 - In section 5.3, a way to evaluate the impact of the quality for synthetic examples is to manually generate them by hand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|20", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "I do think that is an important component for the paper that was not covered as well as the rest.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|21", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "Understanding the limitation of current generative models and understanding \"how far we can go if we improve them\" based on curated high quality data might make your point stronger.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|22", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "5 - In table 5, you might want to add a baseline comparison without GAL so it is easier to compare the improvements.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|23", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "6 - For tabular tasks, I think that section was more confusing than helpful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|24", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "There was not a lot of analysis on why you decided to encode features the way you did and the impact it had on the two datasets where XGBoost outperforms your methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|25", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "It might be an interesting analysis to add or at least to discuss further.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc51|review#4|30", "source": "rc", "split": "test", "paper_id": "rc51", "comment": "My opinion is that this is a clear accept.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#1|9", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "The method is based on an intuition that a “mismatch between predictions of a model and labels can indicate candidates of noise specific to all held out labeling functions.” This intuition needs stronger support and explanation beyond a few examples mentioned in Figure 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#1|12", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "The paper is well written, clear and the method seems original.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|10", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "The [1] shares a similar motivation, which also aims to learn the weights of LFs to data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|11", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "The reported performance on TREC from [2] is 82.59, which is higher than that of the proposed method and its implementation in Table 2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|12", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "It is suggested to add more experiments on the some datasets from [2] to carefully validate the progress.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|13", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "The motivation is clear but corresponding empirical analysis are missing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|14", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "What are the refine results of the allocation of LFs to classes?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|15", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "[1] Giannis Karamanolakis, Subhabrata Mukherjee, Guoqing Zheng, and Ahmed Hassan Awadallah. Self-training with weak supervision. NAACL 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|16", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "[2] Yue Yu, Simiao Zuo, Haoming Jiang, Wendi Ren, Tuo Zhao, Chao Zhang. Fine-Tuning Pre-trained Language Model with Weak Supervision: A Contrastive-Regularized Self-Training Approach NAACL 2021", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|18", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Clarity, Quality: The paper is well written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|19", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Novelty: The paper has its merit but still some claims need to be further validated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|20", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Reproducibility: The authors provide psuedo code and implementation description.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|21", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "The code is included in the supplementary.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#2|24", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Some of concerns are included in the Strength And Weaknesses section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|10", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "The stated goal of the paper is to \"primary goal is to improve the LFs to classes allocation in order to correct the systematically biased label assignments\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|11", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "This targets fundamentally misspecified labeling functions, since a core assumption of vanilla DP is that LFs perform better than random chance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|12", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "These can happen in practice, but I'm not convinced flipping the label (essentially) is always beneficial especially in rare/infrequent classes settings.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|13", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "The paper suffers from clarity issues.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|14", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "For example, the feature-based and deep ULF methods need more methods elaboration and details", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|15", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "From the Table 2 results, which use the WRENCH benchmark implementations (https://arxiv.org/pdf/2109.11377.pdf) the reported performance numbers don't seem to match what was reported in the original WRENCH paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|16", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "See their Table 12 the best reported best results for RoBERTa + Cosine given various label model classes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|17", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "I've repeated the key parts below.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|18", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "This paper's implementation is underperforming and DeepULF does not outperform a tuned model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|19", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "My concern is that the perform benefits might simply wash out when using a tuned end model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|20", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Dataset/Metric        WRENCH/Label Model        This Paper        DeepULF\nYouTube/Acc          97.60 / MV                             96.4                    96.8\nSpouse/F1              40.50 / MV                            33.3                    36.9\nTREC/ACC             77.96 / MV                            65.8                    76.8\nSMS/F1                  96.67 / MV                            93.6                    96.1", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|21", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "There are stronger baseline label models than Flying Squid and the default Data Programming (Snorkel-DP) model See Mazzetto et al 2021. \"Semi-Supervised Aggregation of Dependent Weak Supervision Sources With Performance Guarantees\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|22", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Ideally this type of paper would be more systematic in exploring properties of the task themselves, including low frequency class problems, higher cardinalities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|23", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "It's hard to draw any strong insights from the experiments as presented.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|25", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "I have some concerns on the reported numbers (replicating WRENCH and reported above)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|26", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "I don't think the paper, overall, is particularly novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|27", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "There are virtues to a simple method for improving performance, say an deep empirical dive into k-fold estimation when designing labeling functions, but this manuscript doesn't provide enough details to be very illuminating.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|28", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Code will be released, but is not currently available", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#3|30", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Unfortunately given the clarity in presentation and the inconsistency with reported results (which would erase ULF's purported benefits), I don't think this paper is ready for acceptance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|4", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "This estimation is based on cross-validation-based confidence learning proposed by Northcutt et al (2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|10", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "However, using this confidence estimation is not straightforward (as pointed out in the experiments -- comparison against wscrossweigh and wscleanlab).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|12", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "The paper however lacks a theory to support the idea.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|13", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Incrementally updating the matrix T* and noise labels Y* sounds reasonable but can be dangerous as we don't know to where they will converge (or will they converge at all).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|14", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "In fact, there's even no objective for ULF to work on.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|15", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "(On the contrary, the noisy-channel matrix is part of the loss to be minimized.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|16", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Would it be possible to view T* and Y* as latent variables and put them in a theoretical framework (as in Snorkel or noisy-channel)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|17", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "When comparing ULF to data programming Snorkel, it is unclear whether ULF is able to deal with dependency between labelling functions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|18", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "The use of development sets in experiments for early stopping and hyper param tuning is typical. However, if these sets are clean (annotated labels are not noisy), then they don't seem to be used efficiently. For instance, why not using them to estimate the confidence directly? Or, at least they maybe more helpful than only to select p. Also, having clean dev. sets can be tricky as one can train a model directly on these sets in a supervised learning fashion.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|19", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "As estimating T* and Y* is the core of ULF, the paper should have some analyses on whether T* and Y* are estimated more correctly after each iteration.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|21", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Clarity: the paper is quite difficult to read as the cross-validation-based confidence estimation description is quite dense.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|24", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "They should be shorter to make room for analyses.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|29", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "It should be a welcomed contribution to the community whose work is on learning from noisy data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|30", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "Although I recommend to accept the paper, there are several points that the paper can be improved:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|31", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "the writing can be clearer (especially the description for cross-validation-based confidence estimation), some parts can be shorten for analyses.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|32", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "the paper will be stronger if ULF can be framed in a theoretical framework.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc52|review#4|33", "source": "rc", "split": "test", "paper_id": "rc52", "comment": "more analyses are needed to understand the quality of T* and Y*.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#1|10", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Even though there are experiments in section 7 trying to give an intuitive feel for the reason this method is working, there can be more discussion from a model-structure perspective on how the internal structure of the transformer is creating these embeddings for policy networks.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc53|review#1|12", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Q1: In Table 2, VirtualHome seems to be a very tough environment for any policy that isn't LID-ADG. I'd like to see comparison against other methods which have significance performance on this: even though they may not be using ADG, it would be nice to see how much can ADG push a method against a better performing model which might have more expert data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#1|13", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Q2: Can you provide more intuition/experiments on how different layers of GPT-2 in terms of self-attention are lighting up when a task is being performed? This question is aimed to poke at the structure of the transformer and how different layers are being used for interactive decision making", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc53|review#1|15", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Authors don't discuss the limitations enough of the model in the environment of VirtualHome and BabyAI. I'd like to see a discussion of failure modes of the model and how that traces back to the encoded embeddings.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|22", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "The proposed encoding method is templated-based (Line 155-156). Although the input encoding scheme (Section 7.1) may be a trivial problem, the encoding scheme may still affect the performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|23", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Searching for the optimal encoding scheme is an expensive process, which may bring a high cost of hand-crafted engineering.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|24", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Besides, the data gathering method also relies on hand-designed templates (Line 220).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|26", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "The related work of PLM is adequately cited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|27", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "But the authors should also introduce the background of policy learning so that the significance of this work can be highlighted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|29", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Compared to the work that uses traditional networks like DQN, the integration of PLM may affect the inference speed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|31", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Most parts of this paper are well written. However, there are some typos in the paper:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|32", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Line 53: pretrained LMs -> pre-trained LMs", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|33", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Line 104: language -> language. (missing full stop mark)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|34", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Some papers should be cited in a proper way: Line 108: [23], Line 109: [36], Line 285:[15], Line 287 [15]. For example, in Line 108, \"[23] show that\" needs to be rewritten as \"Frozen Pretrained Transformer (FPT) [23] show that\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|36", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "The authors provided the additional experiments for addressing my concern of scalability. The authors also revised the typos and added the related works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|38", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "1. Pre-trained language models limit the length of input text. For instance, the maximum input length of GPT-2 [1] is 1024 [2]. Will the length of encoded text in your experiments exceed this number? If not, can you figure out a solution for handling long-text input? Because the environment and policy description could be too long, we may not neglect this situation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|39", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "2. In the exploration stage (Section 4.2.2), what is the exact sampling method for the goal and initial state?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|40", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "3. The BabyAI 1.1 baselines [3] report the mean ± std in their papers. However, the evaluation of the BabyAI in this paper (Table 1 and 4) did not use different random runs. I did not find any justification in Section 5.2. Could you please offer an explanation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|41", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "[Rebuttal Updates] I confirmed the authors' answers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|43", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "[1] Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language models are unsupervised multitask learners..", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|45", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "[3] Hui, D. Y. T., Chevalier-Boisvert, M., Bahdanau, D., & Bengio, Y. (2020). BabyAI 1.1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|47", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Societal Impact: No potential negative societal impact. The authors provide a new perspective to aid policy learning with a pre-trained language model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|48", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Limitation: 1) Building text descriptions for each task still requires human labor. We do not know what textual format is optimal for policy learning. It varies from task to task, model to model. On the other hand, as I stated in Question 1, the long-text input could restrict the scalability of this framework.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#2|49", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "2) The proposed methods also need humans to design some templates/rules, as the authors mentioned in the conclusion part.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#3|11", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "One concern is about the reproducibility of the paper. The authors do not provide the detailed architecture and hyperparameters of the method. Although there is a short description in Appendix D.2. I don't think it is sufficient for reproduction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#3|12", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "It would be better if the authors can provide standard deviations in Tables 1 and 4. With a small number of demos, I believe that the standard deviation would be large.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#3|18", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Are all the tokens in brackets (such as [grab] and <apple>) treated as individual special tokens and added to the tokenizer or they are just treated as normal takes?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#3|19", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "I don't see the reason why the authors only use autoregressive language model. The next action prediction task doesn't seem to require an autoregressive model since it's not a sequence prediction task. I wonder how bidirectional Transformers such as BERT and RoBERTa perform.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc53|review#3|21", "source": "rc", "split": "test", "paper_id": "rc53", "comment": "Yes, the authors discuss them in Section 9.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|6", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "My main complaint with this paper is that it describes what's happening as a simplicity prior, or Occam's razor, but a more in-depth analysis of the pattern they observe might reveal that these systems have a prior towards specific types of rules or methods of compression, such as those that exploit hierarchical patterns or memorized chunks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|7", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Because there was a lack of analysis of the rule type, focusing only on an abstract idea of simplicity based on a particular method of compression, I found this unconvincing as a study of a general simplicity prior.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|12", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I wasn't satisfied with the attempt to explain what a concept was.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|13", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Within the paper, this is left very vague.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|14", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "What does it mean to say that an explanation is \"minimally acceptable\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|16", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "most representative AI systems - This is a really unclear claim", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|17", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I found the example about \"go stop go\" pretty confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|18", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I wonder if you could come up with a better example that doesn't involve iffy phrases like \"go stop go stop\" or even obscure English idioms like \"ready steady go\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|19", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I was confused as to the purpose of the paragraph on liens 114-123.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|20", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "It didn't seem to particularly relate to the rest of the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|21", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "GPT-3D-T0 is one of the most similar systems to humans - What does this mean?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|23", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "line 163 \"being w\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|25", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Because this paper has such a broad prior it is investigating, the results may be misleading if the true priors are more constrained than \"simplicity\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#1|26", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Otherwise, this seems to have a limited impact outside of understanding of the inductive bias of neural networks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|12", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "The paper mostly explores LM few-shot learning [1] under the machine teaching setup in [2].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|13", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Experiments appear to be designed carefully.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|14", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Multiple models are compared, both in terms of program complexity and number of examples given.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|15", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "The number of programs is low (only 8), which the authors justify as balancing between obtaining robust results and limiting the human effort needed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|16", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I would suggest keeping the current setup for the human comparisons, but greatly expanding the number of programs for the automated systems (much more than in the supplementary material).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|18", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "In addition to the mathematical definitions, I would appreciate if the authors could explain more thoroughly in plain text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|19", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "The findings appear reasonably significant for researchers involved in language modelling and machine teaching.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|20", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "As language models are commonly used, understanding their behavior, limitations and how they compare to humans seems important.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|22", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "L83: Is it really a failure?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|23", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Figure 1 is arguably overcrowded.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|24", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "It includes changes to both temperature and model size.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|25", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "For additional clarity, I think these should be presented separately.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|26", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "In Figure 1b, what is the data condition (WS, WS +AS I, WS + AS I + AS II)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|27", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "You group 2 programs under each complexity category.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|28", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Do you think that the paper would be more informative if the results of each program were presented on their own, or would it make the paper harder to read?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|29", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Does the order of examples matter?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|30", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "The matrices in Figure 4 are symmetric (along the bottom-left to top-right diagonal), so most information is duplicated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|31", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Why is the size of the additional examples not expressed in bits, as for the minimal witness sets?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|32", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Even with additional examples over the minimum witness set, performance of all systems and humans is still far from perfect.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|33", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I would be interested in knowing how many examples (or bits) are needed to reach certain performance thresholds (if they are at all reachable).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|34", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "If I understood correctly, the minimal witness set should allow to uniquely identify a program, under the assumption that shorter programs are preferred.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|35", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "On line 160, you mention \"In case of equal l, ties are broken lexicographically.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|36", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Does that mean that if a witness set is compatible with 2 programs of the same length, one program would be chosen arbitrarily?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|37", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "If that is the case, I would argue that the witness set is inappropriate to explain the chosen concept/program.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|38", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Does this phenomenon occur in practice?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|39", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "[1] Brown et al. (2020). Language models are few-shot learners. NeurIPS.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|40", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "[2] Telle et al. (2019). The teaching size: computable teachers and learners for universal languages. Machine Learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#2|43", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Limitations and societal impacts are discussed, albeit briefly, in the introduction and at the end of the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|11", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "In terms of presentation, I found the paper difficult to understand at points.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|12", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Even after reading section 3 it was unclear to me how witness sets could be so small and what exactly the authors meant by priors being aligned between learner and teacher, and I was only able to understand what the authors intended by reading the Telle et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|14", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "In particular, I think what I was missing was that witness sets only need to distinguish the shortest satisfying program from earlier, shorter ones.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|15", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Since much of section 3 is not used subsequently anyway, I would suggest that the authors instead use the space to emphasize this and how this leads to the experimental setup.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|16", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I would also recommend the authors put their GPT prompts and the way they attempted to elicit explanations from GPT in the main text, since it's very hard to interpret any of the results otherwise.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|17", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Methodologically, I'm a little concerned that the experiments don't fully address what the authors are after.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|18", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "In particular, it seems like what we'd really like to know is what concepts the learner has learned, but instead we measure accuracy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|19", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "This seems potentially problematic in situations where programs/concepts can have very similar outputs but differ drastically in their complexity, which is presumably the case here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|20", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "(For instance, do we have any way of telling from the experiments that GPT hasn't learned the exact desired program except on a list of 10,000 exception inputs where it behaves differently?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|21", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Furthermore, even if we were satisfied with just accuracy, why not get the accuracy on a much larger sample size than 15 per concept (at least for the non-human learners)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|22", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I think if we were able to get reliable explanations from the models we could resolve these questions to some extent, but as the authors note this is difficult to do.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|23", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I also wonder whether the machine teaching dimension of the paper changes the story/experiments much.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|24", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "For instance, if instead of feeding the learners the optimal witness sets the learners just received random examples, would the plots change?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|25", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "(This experiment might be worth doing!)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#3|26", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "If not, these experiments may tell us more about the priors/biases of these learners in general rather than about how these learners perform when being taught, with ramifications for questions Q1 and Q2 in section 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|7", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I would say upfront that I am not an expert in this topic, and have found many points in the paper unclear, perhaps due to this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|8", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "However, I wasn't fully convinced by the setup used by the authors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|9", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I am looking forward to the author response that might clarify some of these points.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|11", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "The main concern I have with this paper is that I am not sure how the concept of witness set as used in this paper is learnable by a machine or a human.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|12", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Looking at the examples in Table 1, there are numerous ways of generalizing the first WS, some of them of similar simplicity (e.g., copy the first character, copy the last character, copy the entire sequence, etc.).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|13", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Moreover, it is impossible to learn C4 from the single WS example.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|14", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "In fact, none of the examples contain the first rule (if input is '', print '00'), so how can any system be expected to learn this?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|15", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "This makes me unsure of the interpretation of the results in this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|17", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Can you provide a formal definition of a concept?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|18", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "The authors discuss it rather informally in the intro and related work (e.g., the concept of an odd number), but then treat it as a function in #150 and onwards.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|19", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Q1 asks \"Is a language model such as GPT able to capture algorithmic patterns from very few examples selected with a simplicity prior?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|20", "source": "rc", "split": "test", "paper_id": "rc54", "comment": ", but then Q2 asks Does accuracy increase from the theoretically minimal witness sets needed for teaching the concept with additional examples?\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|21", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "From the definition of \"capturing a concept\" in #154 we have \"iff for every S we have that p 􏰑 S iff c 􏰑 S.\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|22", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "In what sense does accuracy improve once you've captured a concept?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|23", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "#193: in which sense can a language relate to \"the way language models work internally (transformers)\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|24", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Please add random/majority baselines to the figures.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#4|25", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Minor: #41 story -> history?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|6", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Note: I am not an NLP researcher so I cannot comment on this work’s significance to that community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|15", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "My biggest concern is that nearly all of the results in the paper are based on the same 8 concepts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|16", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "It is very possible that some of the empirical findings are specific to quirks in these 8 concepts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|17", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "It makes sense that the human experiments needed a limited set, but it would greatly strengthen the paper if additional machine-only results were added with results averaged over hundreds or thousands of concepts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|18", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I am also concerned about the relative mismatch between the average accuracies of humans (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|19", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Figure 2b) versus their ability to provide correct explanations (Figure 3a).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|20", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Should we conclude that humans are sometimes guessing correctly?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|21", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Or that there are some quirks in the concepts, where the examples are suggestive of some other concept that is wrong, but happens to align with the test examples?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|22", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Also, the relatively low rate of explanation correctness across the board suggests that this task may just be too difficult for humans (it is certainly difficult for me, looking at the examples).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|23", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "For the “accuracy” results, if I were to guess 2 out of 5 tests correctly for a single concept, would that contribute 0.4 or 0.0 to the accuracy?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|24", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Please clarify in the text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|26", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I am wondering if the inclusion of the inductive programming methods (Louise and MH) are more distracting than informative, at least in the way that they are presented now, as first-class citizens alongside the GPT and human results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|27", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "My understanding is that the behavior of these methods is much less mysterious than that of GPT or humans -- if the background knowledge given to the inductive programming methods were perfectly aligned with P3, then we would see perfect agreement in the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|28", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "So really the results for these methods are just a function of the background knowledge.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|29", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I’m not sure what we are meant to take away from these results in particular.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|30", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "The link in footnote 1 (https://bit.ly/Code3357) cannot be accessed (permission error)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|31", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "For both humans and GPT, I would be curious to see if the results are robust to changes in the syntax.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|32", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "For example, instead of 0 and 1, what about a and b, or apple and banana, # and %, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|33", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "When presenting the non-human approaches with WS, then WS + AS I, then WS + AS I + AS II, are the methods “reset” after each phase?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|34", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "For example, with GPT-3, does it see the full history of phases, or is just given each phase independently?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|35", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "I am guessing the latter.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|36", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "If so, is it possible that the humans have an inherent advantage over the machines?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc54|review#5|37", "source": "rc", "split": "test", "paper_id": "rc54", "comment": "Intuitively, from the sequence, the humans could infer the relative complexity of the examples in each set, giving some additional hints about the target concept.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|11", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Some ad-hoc decisions (e.g., shortening the max sequence size) that are not well explained", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|15", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "I had a hard time processing the multitude of experimental results. In particular, it's not clear what the numbers represent in each table.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|16", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "I would clearly describe in the caption of each figure what the rows/columns represent and what metric is used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|17", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "In particular, the paper uses \"fairness\" without clearly explaining how fairness is measured.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|18", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Some of the design decisions are not well explained.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|19", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "In particular, I am curious about the maximum sequence size of 64 (which is quite limiting).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|20", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Is that to ensure that the style transfer and the GPT3 generation produce quality results?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|21", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Not clear what WR50 means.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|22", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "In Table 1, all columns except for BA represent some measure of fairness? If that's the case and if I understand the results WR (which I'm guessing stands for word replacement) seems quite performant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|23", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "How do you justify the complexity of adding the style transfer and GPT3 on top.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|24", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "What is the real advantage of the full C dataset?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|25", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "The fairness metric used seemed a bit ad hoc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|26", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "One of the more standard metrics could be use (see Fairness Definitions Explained and the following paper for an example of using equalized odds: Your Fairness May Vary: Pretrained Language Model Fairness in Toxic Text Classification - analysis of many LMs wrt fairness and model size/training size/random seed (in the context of toxic text prediction)).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|27", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Last but not least, I think \"fairness specifications\" is misleading and quite an overblown term for what the paper is about;", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#1|28", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "it made me think of some theoretical formalism/specification. Similar pairs of text across protected groups is a much more honest description of what the paper generates.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#2|17", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "For weaknesses, the authors could be clearer on what fairness metric is being used for tables such as Table D.3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#2|18", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Additionally, an ablation study could be added in terms of how changing the similarity classification threshold affects performance of the similarity model (and the downstream performance of downstream classifiers)?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc55|review#2|19", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Finally, I would have liked to see in the appendix table E1-4, the generations from the three methods on the same examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#2|20", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "There are also a few nits, for example, make sure to reference all variables in graphs (for example, D in Figure 1).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#2|21", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "As a note, the authors should reference this paper as well for automatic generation of evaluation data: https://arxiv.org/pdf/2202.03286.pdf.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#3|10", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "It seems like the approach would be really task dependent. I think it is good for authors to acknowledge this somehow and mention how the task can affect the outcomes.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc55|review#3|11", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Related to my above comment, authors only study toxicity problem.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc55|review#3|12", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "It would be good to study other tasks and compare the differences and provide more in depth discussion on what the differences are in terms of the filtering approach per task as well as human alignment results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#3|13", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "While the approach is effective in achieving more \"fair\" outcomes, the approach imposes loss on accuracy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#3|14", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Given that the task of toxicity detection is a subjective task and in many cases culturally dependent, what mechanisms authors take into account in their human studies other than recruiting more than one worker to label a hit and taking the maximum.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#3|15", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Please provide some discussion on this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|8", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Weaknesses (see more details in the section below):", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|9", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "The paper needs further improvement in clarity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|10", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Beyond individual fairness, a discussion of how the proposed dataset with individual fairness would affect group fairness notions is needed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|11", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "The broader impact of how relevant stakeholders would use the generated workflow is unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|13", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "In this section, I will provide more details of my comments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|17", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "However, I think the paper could be further improved in terms of clarity and discussion of multiple related notions (e.g., group fairness) and the potential broader impacts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|18", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "While the paper is fairly readable, there is substantial room for improvement in clarity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|19", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "The techniques involved in this paper span multiple domains, requiring the reader to have a good knowledge of algorithmic fairness, NLP, user study, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|21", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "However, it is hard to understand the experiment sections without referring to the appendix in detail.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|22", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "For example, in Table 2, the authors do not provide any description of the \"Metric/Method\" column in the main text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|23", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Similar problems happen in Table 3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|24", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "I also have a question about evaluating the diversity of generated pairs: the paper uses the fairness metric as the evaluation criterion for models trained on different constraint sets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|25", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "The word choice of \"diversity\" sounds strange, given the evaluation criterion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|26", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "A better word choice might be something like \"coverage\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|27", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Since the paper studies fairness.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|28", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "I am also interested in how the generated intuitive fairness specification affects group fairness notions such as demographic parity and equalized odds.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|29", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Besides, since the fairness specification is intuitively generated by humans, how do you provide more documentary specifications on this to the relevant stakeholders, such as ML/NLP practitioners and model users, to trust the overall workflow?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|30", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Lastly, I have some questions regarding reproducibility: I only see the codes in supplementary materials.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|31", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "Will the dataset be released?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc55|review#4|32", "source": "rc", "split": "test", "paper_id": "rc55", "comment": "If yes, how do you plan to release them in order to provide a better potential impact in the long run?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|8", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The paper reports state-of-the-art results for the HumanEval benchmark for code generation from natural language specifications.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|9", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The authors provide downloadable checkpoints for the trained models, which can be a useful resource for the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|10", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The paper introduces a multi-turn natural language to code dataset where the natural language instructions are interleaved with the intended generated code.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|12", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The model is evaluated only on one existing dataset, HumanEval, and not other similar datasets like APPS or MBPP.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|13", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "While the work represents a significant expenditure of resources, the amount of technical novelty is relatively low, and some of the new aspects in the construction of the model itself are not studied or described as much as they could be:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|14", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The \\textsc{BigPython} dataset seems responsible for much of the gains in the work, but the description of it in the paper is limited to \"We have compiled public, non-personal information from GitHub consisting of permissively licensed Python code in October 2021\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|15", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "There is significant space dedicated to describing the TPU-based training setup with JAX, but not much about the reasoning behind it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|16", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The paper doesn't have much experiments about hyperparameters and training data selection (e.g. the order in which the datasets were used for training), or further information/justification for the values used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|17", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "There is no release of training datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|18", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The deduplication methodology used during dataset preprocessing may be too weak.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|19", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "Only removing exact duplicate files may still leave many files which are very similar (for example, varying only in some comments at the top of the file).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|20", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "For more discussion on another code dataset, see https://twitter.com/miltos1/status/1497126435261083649", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|21", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "Considering that copies of the HumanEval dataset may have been posted to GitHub, the authors should investigate the possibility of training dataset contamination.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|24", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "I did not realize until quite late that the CodeGen-Multi and CodeGen-Mono models are trained on, respectively, The Pile plus the BigQuery dataset, and The Pile plus BigQuery plus BigPython. That fact is only mentioned in Section 4.3 and in the appendix, and not where the models are initially described (Section 2.1 and 2.2). This can be clearer in Section 2.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|27", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "A more detailed discussion about it would have been helpful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|30", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "It is unclear why Table 2 only has information about CodeGen-Mono and not the other models. More data here would be more convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|31", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "For the multi-turn dialogue data, the instructions for subsequent turns only make sense if the previous turns generated correct code (or at least code in line with the authors' intent).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#1|32", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "As such, it would be useful to see how well the model does when all the output for all previous turns are correct.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|7", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "(+) Training and releasing the models is in itself forms a substantial contribution to the field, enabling research groups with more limited resources to experiment on top of sound base models of significant sizes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|8", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "(+) The multi-turn dataset is a valuable contribution towards practical neural program synthesis applications.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|14", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "(Q1) Will the BigPy dataset be released?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|15", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "(Q2) Will MTPB be released?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|16", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "(Q3) MTPB: What is the longest intermediate piece of code required in a turn?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|17", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "(Q4) MTPB: Did you evaluate how realistic it is to assume that users can provide such close guidance, but not write the code themselves?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|19", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "page 3: \"models are called call the models as\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|20", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "page 3: \"While it is not our focus ...\" - this sentence is not grammatical, and I have no idea what point is being made here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|21", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "Fig 1 (nit): the solution produced by the model is wrong (email addresses are much complicated than this), but on par with how most humans would solve the problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|22", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "Sect. 4.3: this repeats earlier (more complete) descriptions of the used models - this is not necessary.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|23", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "page 4: \"performance on the MTPB improves as a function of the model size and data size\" - the latter claim is unsupported, as the additional data is also more specialized to the problem setting (i.e., more Python-specific). Please rephrase.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#2|26", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "I find the MTPB dataset to be largely uninteresting, as it relies on very fine-grained natural language supervision. It's not clear to me whether it adds value over the Django dataset of Oda et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|9", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "CodeGen model (single-turn) is only evaluated on a single dataset with only 164 problems. Why is the model not evaluated on other publicly available coding datasets such as MBPP and APPS?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|10", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "While the multi-turn dataset helps the models generate more correct code, I am not sure about its practical use cases.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|12", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "How does multi-turn task break down work for programs with control flow structures (for e.g. write code for binary search)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|13", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "From the examples, it appears that the MTPB dataset treats small loops and branches as a single sub-task, but how can a user specify sub-tasks within a loop/branch?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|15", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The paper is fairly well-written and reproducible with the models being publicly available.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|16", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "Novelty is limited since there are already several pre-training/dataset collection papers in this area.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|17", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "For the MTPB dataset, I don't fully understand the part about embedding the test case inputs within the prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|18", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "Why was this approach chosen over HumanEval's style of learning to complete given a function definition?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|19", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "How do you deal with multiple test cases? How does inference and evaluation work with multiple test cases?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|20", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "It will be useful to list Codex Davinci's performance in Table 1 (for completeness, even though it is a much bigger model).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|21", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "Similarly, it would be interesting to look at Codex's performance on the MTPB dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|23", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The contribution with regards to the CodeGen model is helpful for the community (since it is open-sourced and performs similar to Codex 12B model), but there is limited novelty there.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|24", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The MTPB dataset contribution needs more explanation regarding its practical use cases and its limited capability to handle more complex programs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#3|25", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "POST REBUTTAL: most of my questions are answered and I am increasing my score to a 8.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|8", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The evaluation result on the constructed dataset supports the proposed hypothesis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|9", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The performance difference in multi-turn and single-turn specification over levels of difficulty and models shows that the improvement is sizable except for easy problems with larger models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|13", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The motivation for multi-turn code generation is convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|14", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The supervision of comments and programs is helpful for multi-turn program synthesis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|16", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The idea of utilizing weak supervision of interleaved patterns is intuitive and convincing. This paper presents an effective way to make use of this idea.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|18", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "Codegen for program synthesis seems effective, especially when the user wants to generate pieces of code from input/output examples or natural language descriptions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|20", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The overall writing is clear. The authors utilize figures and tables well to illustrate the ideas. Figure 1 clearly shows the code generation process.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|22", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The authors clearly describe the related prior works from both the perspectives of program synthesis, large language models, and benchmarks for program synthesis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|24", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "The presentation of the experimental results is clear. Mainly, Table 4 provides understandable results showing that multi-turn specifications achieve better performance compared to single-turn specifications.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|26", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "I believe reproducing the results is possible given the clear description provided in the main paper and the appendix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|29", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "Some programmers like to write comments, while some are not. Also, different programmers write comments very differently.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|30", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "I suggest the authors investigate the effect of such differences.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|31", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "One very simple way could be to analyze the occurrence frequency of interleaved natural and programming patterns in the dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|32", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "While it is hard to formally define meaningful comments, it would be insightful to at least calculate the document frequency of interleaved natural and programming language.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc56|review#4|34", "source": "rc", "split": "test", "paper_id": "rc56", "comment": "In Table 3, the paper only compares the proposed method against GPT-Neo and GPT-J, which is not sufficient.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|4", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "My main concern with this approach is that the paper does not discuss how to apply this approach to more general problems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|5", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The problems considered here are very simple, and an algorithm to solve them needs to be known in order to generate the training data.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|7", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "In terms of the insight from a learning point-of-view, I also do not see anything surprising in this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|8", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "After decomposing the problem manually (i.e. by writing an algorithm to generate training data), the learning task is pretty simple (for addition, for example, all the model has to do is to extract the last digit from two numbers).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|9", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "So the high accuracy of the resulting method is not that surprising.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|11", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The writing style is a bit \"flashy\", which gives room for miscommunications.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|12", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "For example, the paper opens with the following statement:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|13", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Although neural networks have achieved amazing results on various domains, e.g., images, texts, audios, videos, games, etc., nearly all of them are classified as System 1 tasks (Kahneman, 2013), ...", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|14", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Afaik, this statement is not supported by the literature, and in particular not by the given reference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|15", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Following Kahneman's theory, tasks like playing the board game go would certainly require system 2 thinking.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|16", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "IMO, dropping the first paragraph would help the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|17", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The paper also makes claims such as \"the length of CoT can grow rapidly with the problem’s complexity\", without further explanation and I am not even sure what exactly this means.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|20", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "However, the technique only seems to work for problems for which we already have algorithms, and it is unclear to me if this could be extended to more general problems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#1|21", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The writing style should be improved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#2|4", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "On the negative side, the solution proposed is formulated as a fully supervised problem, meaning for each task the exact steps needed to solve the problem needs to be annotated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#2|5", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "This is not a scalable approach as for realistic real-world applications obtaining detailed annotations for reasoning problems is an expert task which is expensive to scale to large datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#2|6", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "As a result, outside of tasks such as arithmetic operations where the dataset can be created automatically (and one could argue that you wouldn't use a task specific LM as the paper doesn't deal with multiple tasks or out-of-distribution problems), the applicability of this approach is minimal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#2|9", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The novelty is minimal as the solution proposed is rather trivial as it doesn't try to solve the hard problem of reasoning without direct supervision, or other settings such as few-shot/zero-shot/multi-task and formulates the problem as a fully supervised task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#2|10", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Due to simplicity of approach, I think it's reproducible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#2|12", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Unfortunately, I don't see much impact from this work in practical terms.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#2|13", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The formulation doesn't scale to general reasoning problems, the tasks considered aren't tasks that you would train a specific LM to solve, and the algorithmic solution is very simple/trivial extension.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#2|14", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Hence, I'm not able to recommend this paper for publication.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|5", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "One important point missing from this work is how to systematically construct subproblems given any problem in order to teach the model learn the recursion of thoughts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|12", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "[-] The training of RoT requires (non-trivial) human inputs to design proper subproblems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|13", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "More specifically, I wonder", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|14", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "what is the limitation of the problems that RoT could solve: for example,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|15", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Does RoT have the ability to learn backtrack?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|16", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Some problems may require people to modify the earlier part of the answer based on the new observed information.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|17", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Does RoT have the ability to learn a problem where its subproblems have different structures (as recursion usually requires the same structure in the subproblems)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|18", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "What is the performance of RoT on (maybe small-scale) NP-hard algorithmic problems that may not have the divide-and-conquer structure such as TSP？", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|19", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "how to design proper subproblems in order to use RoT to train the models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|20", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "What are the criteria of the subproblems for RoT?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|21", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "What kinds of structures do they need to have？", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|22", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Are there systematic approaches to design subproblems for reasoning tasks in general?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|25", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The quality and originality of the work should meet or exceed the conference standard.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|26", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "My main concern is on how to properly train the RoT framework as it requires people to feed it with hand-designed subproblems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|27", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Therefore, it would be great if the authors could explain the limitation of the problems that RoT could solve and how to design proper subproblems in order to use RoT to train the models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc57|review#3|33", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "My main concern is on how to properly train the RoT framework as the training of RoT requires (non-trivial) human inputs to design proper subproblems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|11", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The work's most significant strength is its practical pipeline for automatically constructing knowledge graphs (KGs) from pre-trained language models (LMs).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|12", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Previous research on \"LMs as KGs\" or \"knowledge probing from LMs\" has primarily focused on prompt generation and does not cover the entire KG construction pipeline.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|15", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The paper also has a few minor flaws, and the writing could be improved, as listed below:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|17", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The multi-token issue in entity names.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|18", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The authors state in Sect.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|19", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "3.2 that \"Each entity consists of one or more tokens\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|20", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "How to resolve the multi-token issue in the likelihood computation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|21", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Or, for example, how to compute P_{LM}(t|p,h) if the name of entity t has more than one tokens?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|22", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "In Appx. A.5, the authors state that \"we enumerate the number of masks for every entity\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|23", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Is there a maximum value for such a number?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|24", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Is it a hyper-parameter?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|25", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Sect. 3.2 says that \"The use of the minimum individual log-likelihoods allows us to apply rich thresholding and pruning strategies (e.g., thresholding on log PLM(h|p) for proposing the head entities), as described more in the appendix.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|26", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "However, it seems that there is no related content in the appendix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|27", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Or, is it Appx.A.5?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|28", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "It would be preferable to include a subsection in Section 4.2 that simply introduces the baselines used in the experiment.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|29", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "In Sect. 4.2, the authors state that \"we restrict every entity to appear no more than 10 times to improve the diversity of generated knowledge\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|30", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "This, in my opinion, is not reasonable because it would result in a sparse KG.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|31", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Table 2 also demonstrates this problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|33", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Those systems, however, often involves a complex set of components -> Those systems, however, often involve a complex set of components.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|35", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Overall, the paper is clear, but the writing could be improved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|36", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The authors promise to release the code as well as the resulting KG.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#1|39", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "It can be further improved", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|7", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The problem of automatically constructing KG from pre-trained LMs is well-motivated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|8", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "It is interesting to combine the iterative KG construction into the current pre-trained LMs fashion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|11", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Some important technical details, e.g., how to generate multi-token entities and how to reduce the candidate set, are missing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|12", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The experiment is not comprehensive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|13", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "There is no objective evaluation.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|15", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Clarity: Some important technical details are missing, e.g., efficiently generating candidate entities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|16", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Please refer to the Summary of the review Q3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|17", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Quality: The model is only evaluated by crowdsourcing, with no objective evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|18", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The effectiveness is not significant compared with SOTA baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|19", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Novelty: The idea is generally interesting, the authors bring the old wisdom of iterative KG construction into the pre-trained LM fashion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|20", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "But the core technical contribution is just an LM-based reweighting. Methodological nvoelty is a shortage.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|21", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Reproducibility: Unable to evaluate, since the technical details are unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|24", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "My major concern is how to effectively handle the semantic drift, i.e., the semantics of new patterns and instance harvest may change drastically along with iteration or paraphrasing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|25", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "But authors did not discuss this problem in-depth or provide a method with a plausible bound.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|26", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "What we have in the model is the LM-based paraphrasing and weighting, however, the key point is, LM does not necessarily guarantee the semantics.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|27", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "For the diversity of paraphrase (second paragraph of Section 3.1), it is not clear (1) why to use a textual measurement (i.e., edit distance) to ensure diversity of paraphrases (second paragraph of Section 3.1).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|29", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "(2) how this ED constraint works. As ED works for a text pair, do all generated prompt pairs need to keep a certain distance?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|30", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Some important technical details are missing: (1) Efficient generation candidate entity pair generation of Section 3.2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|31", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Appendix A.5 has a few sentences talking about \"mask-filling\" but I cannot find the details of how to generate multi-token entities and how to use \"thresholding and pruning\" to reduce the candidate set.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|32", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "(2) For computing Eq.1, it is not clear (a) how to generate the query text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|33", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Is the query text formed by inserting a specific entity (e.g., insert h into p for P(h|p))?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|34", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "If so, what is the difference between Plm(t|p,h) and Plm(h,t|p)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|35", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "(b) how to compute the probability of Plm(h,t|p), Plm(h|p), and Plm(t|p,h).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|36", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Is the probability computed by the generative probability of the LM?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|37", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "In addition to the manual evaluation (crowdsourcing rating), the objective evaluation based on evaluation datasets should be given.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|38", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "For example, the authors can select datasets adopted by relevant tasks, e.g., knowledge graph completion, such as WRNN, FB15K-237, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#2|40", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The baseline methods should be introduced in a few sentences.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|4", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Strength: The proposed idea is simple and seems straightforward to implement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|5", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The proposed method needs minimal human input (only an initial prompt and 5 candidate entity pairs for each relation).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|6", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "It effectively leverages GPT3 to rephrase and generate better alternative prompts used for extracting new entity pairs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|7", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Overall, the paper is well written with detailed examples to illustrate the main idea.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|8", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Weaknesses: Although the model shows better results than the baseline (COMET), it still lacks more analysis in the evaluation to convince the effectiveness of the proposed model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|9", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "First, the result in Table 2 shows the precision of the prediction but not the recall.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|10", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "For example, the recall can be estimated on the known set of factual relations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|11", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The P-R curves in Figures 3 and 4 seem to partly answer this question.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|12", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "However, looking at the 0.8 level of precision, the recall is only less than 0.2, which raises another question about the overall effectiveness of the extraction method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|13", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The baseline WebChild in Table 2 seems to extract a much higher number of tuples at 4.6M, while the proposed model ROBERTANET only extract 6741 tuples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|14", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "It can be WebChild having much higher coverage, but this difference was not discussed in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|15", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Second, the quality of the extraction heavily depends on the compatibility scoring function (Eq.1).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|16", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "It would be better to evaluate more on the design choice of this formula and explain why the proposed formula is optional for this purpose.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|17", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Third, Looking at the extracted relations, it looks like some relations are more non-trivial than others.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|18", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "It would be useful to analyze more on the factual relations (e.g., the relations between named entities such as 'ceo_of', 'invent') because these relations will be more useful for downstream applications like QnA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|19", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "The LAMA dataset has more of these relations so authors can extract and analyze more case studies from this dataset.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|20", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "Another suggestion, in the GPT3 prompt: \"paraphrase: {sentence}\" used to extend the prompt set, would it be better to include example of all 5 seed entity pairs into the prompt to help eliminating the ambiguity?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc58|review#3|24", "source": "rc", "split": "test", "paper_id": "rc58", "comment": "However, the evaluation currently lacks analysis about the coverage and quality of the extracted relations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#1|22", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Selecting subsets of tasks and exploring fine-tuning (2.1, 2.2) takes up a lot of the paper and it did not perform as well as pre-training or impact the performance of pre-training much and so maybe less space could have been dedicated to these sections.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#1|24", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "This paper pushes forward work on large scale multi-task transfer learning, with a huge number and variety of tasks (107 for training) with many test suits and research questions explored.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#1|25", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "It is well written and convincing and I would like to see it published in ICLR.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|14", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Many of the results presented in motivating the work are over random subsets of pre-training tasks and random seeds for training. However, none of the results seem to give an idea of the variance of performance, nor any indication of how many different random subsets of tasks were used for the results.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|16", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The authors can analyze many interesting choices here to see how task scaling affects results. To give one instance, one can consider choices like selecting tasks proportional to their size (instead of random sampling) and evaluating how it affects performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|18", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Often the evaluation is in terms of “average” super-glue performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|19", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Since many improvements are actually marginal (less than 1%), it will be good to provide a spread of the results (at least in the appendix) so the readers know the results are not due to improvement on just a small subset of downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|20", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Many hyper-parameters and design choices for the pre-training are also difficult to find and not clearly specified.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|22", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Table 3: best effort has 48 tasks but random-mix is trained with 55, why?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|23", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "What is the variance of random-mix?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|24", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "What is the overlap between best-effort tasks and random tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|25", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "2.4: Is the amount of data and pre-training steps the same across the models being compared?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|26", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "If not, it will be good to make the amount of data and steps similar when comparing only C4 vs only C4 + ExMix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|27", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Moreover, it will actually be more useful to compare with respect to C4 performance at multiple data sizes, since pre-training data is more abundant than supervised multi-task data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|28", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Fig 3, adding multi-task over C4 only seems to help a little bit, looking at the spread of SuperGlue scores between the two settings will be more helpful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|29", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "These numbers (at R=2) seem different than what is reported in Fig 5, can you comment on why?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|30", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "2.5 and Fig 4: task scaling only helps at 512 batch-size, there doesn’t seem to be any consistent trend at batch-size of 128.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|31", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "This seems to indicate that batch-size is a very crucial axis for ablation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|32", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Did you try other batch-sizes, a plot of how performance varies with this parameter will be helpful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|33", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Moreover, batch-size here is confounded with the number of tasks that are sampled in a batch.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|34", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "What is the task sampling strategy used here and how does it affect this plot?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|35", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Fig. 4: What is the shaded area?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|36", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "if it is some statistic like standard deviation across random seeds, why is this not present at 0 and 107.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|37", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Fig. 4: What is the variance with respect to the random selection of tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|38", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Will a size-proportional selection of tasks lead to different results, for instance, to select 30 tasks pick the 30 with the largest data?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|39", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "In the super-glue results, improvements diminish at larger model sizes, to the extent that T5-XL performs similarly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|40", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "This seems like an interesting result and the trend indicates multi-task might become less useful at even larger model sizes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|41", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "It will be good to explore this further, if you have resources for such an experiment, as I consider this an important aspect of “scaling” which will fall within the scope of this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|42", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Can you comment on your choice of T5 1.1 instead of the original T5 models published with the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|43", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The original models also include downstream supervised tasks in their pre-training, just like this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|45", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "I am not convinced that the experiments fully support the claim that adding a random mix of more tasks overcomes negative transfer between tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|46", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The spread of SuperGlue results with the number of tasks is highly inconsistent and performance of some tasks often gets worse at more tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|47", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The large variance in Fig 4 at 55 and 80 actually indicates task selection is really important to get higher performance, though the best-effort approach, as presented in this paper, might not be the best to get that task mixture.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|48", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "This then requires a more thorough evaluation of the various choices for multi-task task selection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|49", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "I am also concerned with the large difference in trend at the two different batchsizes, which the authors didn't address.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|50", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Compared to previous work, I have to disagree with the authors that problems like task heads or loss scaling are mitigated by T5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|51", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "T5 still generates using a shared softmax classifier (over the whole vocabulary), instead confounding the problem with that of generation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|52", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Loss scaling is hidden in how you sample the instances/task for each batch.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|53", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The main point of difference that I see from the rebuttal is of using multiple seeds in Fig 3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|54", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The response also included new results using a larger T5 model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|55", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The result on the larger T5 model seems to confirm the trend that there are diminishing returns from multi-tasking at a larger scale.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|56", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "I am increasing my rating by one point (which in the new system makes it 5) as there are tidbits of interesting results in the paper, which if presented neutrally could have been interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|57", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "However, I am concerned that this will be used as a point of reference to say that randomly adding more tasks always helps, for which the paper doesn't provide sufficient evidence.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|59", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "In general, I think it is important to carefully study multi-task scaling in the post pre-training era.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|60", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "There are many papers in a similar vein recently and to be useful the paper should concentrate more on asking meaningful questions and rigorously analyzing the results to provide recommendations for future work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#2|61", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "I hope the authors take the criticism constructively to improve their work further, concentrating on more analysis rather than more downstream benchmark results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|11", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Given that SuperGLUE may represent one or several language understanding tasks, do the findings regards finetuning transfer, pre-finetuning hold still for other task families like generation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|12", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Or is that specifically tailored to SuperGLUE as mentioned in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|13", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Section 2.5 uses the same learning rate for different batch sizes, which might result in unfair comparison given there are previous works suggesting how batch size, learning rate could have compound effects on the downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|14", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "In section 2.4, will R beyond 10 still lead to consistent improvement of the SuperGLUE performance is unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|15", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Beyond the empirical observations, are there explanations (like task affinity beyond the accuracy correlation) that could explain more clearly the results of co-training by multi-task finetuning on top of pre-trained models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|16", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The significance of the results on different task/task families can be further validated by reporting std/mean results with multiple seeds as in (Section 2.5)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|17", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The zero-shot performance might also be a point to add given the multi-task prompting / finetuning seems to be helpful both for GPT-families and T5-families [1, 2].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|18", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "[1] Wei, Jason, et al. \"Finetuned language models are zero-shot learners.\" arXiv preprint arXiv:2109.01652 (2021).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|19", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "[2] Sanh, Victor, et al. \"Multitask Prompted Training Enables Zero-Shot Task Generalization.\" arXiv preprint arXiv:2110.08207 (2021).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|21", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The idea of leveraging multi-task learning for NLP has been introduced before as mentioned in the paper and not referred to as [3, 4].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|22", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "It is great to revisit and extensively test this idea at scale at the paradigm of pre-training methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|23", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "It also offers potentials for future works on how existing labeled datasets can be used to further improve NLP models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|24", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "[3] McCann, Bryan, et al. \"The natural language decathlon: Multitask learning as question answering.\" arXiv preprint arXiv:1806.08730 (2018).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#3|25", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "[4] Subramanian, Sandeep, et al. \"Learning general purpose distributed sentence representations via large scale multi-task learning.\" arXiv preprint arXiv:1804.00079 (2018).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|28", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The methodological novelty of the paper is fairly low.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|29", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "ExMix is a collection of pre-existing datasets and using multi-task signals during pre-training is not new either.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|30", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "The scientific value of this work predominantly comes from the scale of their multi-task setting, and their analyses.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|31", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "It is perhaps not surprising that a larger collection of supervised pre-training datasets would improve performance over these baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|32", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "All of section 2 is evaluated on average SuperGLUE performance only.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|33", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "While SuperGLUE is a strong baseline given the diverse set of tasks it encompasses, there is always the possibility that a subset of tasks within SuperGLUE are particularly well-suited for the ExT5 training setup and, as such, Section 2 risks overfitting to the SuperGLUE baseline.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|34", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "There are some details missing regarding Section 2.4, especially with respect to random seeds.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|35", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "In particular, Figure 3 shows some interesting behavior regarding the hyperparameter R - it is surprising to me that performance does not monotonically decrease after a certain point but rather seems to be trending upwards.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|36", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "What happens if R increases even more?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|37", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Or is the behavior at the tail end of R (R=5 or 10) due to random noise?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|38", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Section 2.5 compares different numbers of tasks trained on different batch-sizes.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|39", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "However, the models use the same R parameter across numbers of tasks, which seems unfair to low-task settings given that it was tuned on 107 tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|40", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Additionally, the models use the same learning rate regardless of batch-size, ostensibly a learning rate that was tuned for a batch-size of 512.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|41", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "It has been previously noted (e.g. by McCandlish et al., 2018) that there is a clear relationship between batch-size and learning rate; thus, if the same learning rates are being used across the 512 and 128 batch-size settings, it's likely that the results are heavily influenced by a poor learning rate for the 128 batch-size setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|43", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "Overall, the paper fairly convincingly argues that mixing supervised and self-supervised signals during pre-training is extremely beneficial to downstream performance and sample-efficiency.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc59|review#4|44", "source": "rc", "split": "test", "paper_id": "rc59", "comment": "However, there is limited methodological novelty and the benefits of scale and incorporating supervision may not be particularly surprising.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#1|13", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The lack of formal definitions, make it hard to understand how thrust is computed e.g., \"casting a set of instances ( the training data ) into the representation space \" The writing is mostly sloppy, e.g., c_0 is used before it is defined", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#1|14", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The experiment result also need more explanations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#1|15", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "For example in Table 3 it is good to see that using knowledge 25% of the time is as good as using it 75% of the time for many tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#1|16", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "However, one might wonder why only the result of UnifiedQA is shown?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc60|review#1|17", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "why not also compare to the case using knowledge 100% of the time?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc60|review#1|19", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "This paper is very unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#1|21", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "This paper is very unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#2|12", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Table 2, if my understanding is correct, with knowledge case is presented after the vertical bar (|).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#2|13", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "If that's true, please fix the description of the model which states the opposite.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#2|14", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Table 2, is the with knowledge same as using full knowledge?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc60|review#2|15", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Table 4, when comparing thrust with full knowledge, what are the exact performance numbers?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#2|16", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Please mention the evaluation times for both cases as well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#2|18", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "[1] Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., Driessche, G.V., Lespiau, J., Damoc, B., Clark, A., Casas, D.D., Guy, A., Menick, J., Ring, R., Hennigan, T.W., Huang, S., Maggiore, L., Jones, C., Cassirer, A., Brock, A., Paganini, M., Irving, G., Vinyals, O., Osindero, S., Simonyan, K., Rae, J.W., Elsen, E., & Sifre, L. (2022). Improving language models by retrieving from trillions of tokens. ICML.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#3|17", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "A main weakness is that the work is not reproducible, with respect to both the methods and datasets.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc60|review#3|22", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "A main weakness is that the work is not reproducible.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc60|review#3|23", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "It is not clear if the external knowledge used is published or not, and if and how it is possible to obtain it.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc60|review#3|24", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The methods (and experiments) are described informally, or obscurely, where the specific choices made are not motivated or compared to alternatives.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#3|26", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "This leap is hardly discussed or motivated in comparison to the literature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#3|27", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "While the motivation for identifying `hard cases' is 'limited bandwidth or budget to retrieve external knowledge', efficiency and computation time is not evaluated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#3|28", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Also, the choice or selecting some fixed portion of the queries (e.g., 25 percent) for expansion is not motivated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#3|32", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "There are some typos and style issues.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|12", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The motivation and suggested method are not well-aligned.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|13", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The authors claim that they simulate real-world cases, but the experimental setup is highly artificial, making the practicality of the algorithm dubious.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|14", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The experiments are insufficient to show the effectiveness of Thrust.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|16", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The authors claim that they try to simulate the real-world case where they have limited bandwidth to retrieve external knowledge from the perspective of cost-efficiency, but the claim is not persuasive because the way they choose the portion of the queries to use external knowledge is through relative comparison between queries after scoring all other queries, and not through absolute scoring on each independent query, e.g., by applying a score threshold.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|17", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "In other words, the suggested method seems impractical because each query would come to the system independently in the real world, not as a batch as in the experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|18", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Therefore, if cost-effectiveness is a concern, the authors should have proposed a method that works on each query independently to determine whether that query itself requires external knowledge or not.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|19", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Comparing the results in Table 2 and 3, on many of the tasks, using Thrust to partially utilize the external knowledge drops the performance compared to when all queries are answered with external knowledge (e.g., 70.2 → 56.8 (75% of queries) on e-SNLI UnifiedQA-base, 80 → 73.4% (75% of queries) for TriviaQA UnifiedQA-3b).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|20", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Therefore, the cost-effectiveness of Thrust, if it exists, comes with a sacrifice of accuracy for many cases, creating a trade-off.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|21", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Meanwhile, it is also unclear whether the method is actually cost-effective because the calculation of the Thrust score would also take up some inference time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|22", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The authors should compare the average inference speed between when all queries use the external knowledge as the input, and when part of the queries use the external knowledge while the Thrust score is calculated for all queries.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|23", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The authors only compare Thrust with random baseline, but in order to show the effectiveness of Thrust as a scoring metric, they should compare the algorithm with baselines simpler than Thrust but better than random, e.g., using BM25 scores.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|24", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "[Minor point] While the number of external knowledge instances to utilize and the method of utilizing knowledge might vary, the authors seem to test the effectiveness of Thrust with only one combination of them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|25", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The authors did not clearly indicate how many external knowledge instances are used for a query, but it seems like one by inferring from the context.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|26", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Also, they test only the case of using external knowledge as part of the input and did not explore the cross-attention-based method (e.g., FiD).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|28", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The misalignment between the motivation and the experiments makes the paper’s point unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|29", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Some experimental setup details are missing, making the results difficult to be reproduced for now.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|30", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "For example, the authors did not indicate the number of clusters they used in the experiments, and how many knowledge instances they used to augment a query (which seems to be one).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|31", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "In the caption of Table 2, “Performances with/without knowledge external knowledge are presented before/after the vertical bar, respectively,” it should be without/with based on the description of the results in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc60|review#4|32", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Such an important typo makes the paper difficult to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#1|9", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The paper does not achieve the expectations set by the title and abstract.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#1|10", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "I was expecting challenging tasks where humans struggle to find a proper prompt (like leading Dall-e to generate a horse riding an astronaut, which was cited as a limitation of Dall-e until someone found the proper prompt) but the proposed method can automatically find it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#1|11", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "Instead, most of the tasks can be solved with human-level accuracy using just the enumeration of a few input/output pairs as prompts (Fig. 4). These pairs would be available in any case since they are needed by the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#1|14", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The method and the experiments are sound, I appreciated the multiple runs to estimate the confidence of the accuracies.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#1|15", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The exaggerated hype is a smudge in a research work of good quality.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#1|16", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The broad topic of prompt engineering is very novel, while the specific idea of an automatic prompt engineer was first explored in the very recent Hononovic et al. (2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#1|17", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "I still consider the novelty of this work satisfactory, given the new method presented and the general novelty of the area of research.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#1|19", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "My recommendation is to accept this paper, the research is of interest and well performed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#1|23", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "This work made me think of https://arxiv.org/abs/2201.10222, the natural language program synthesis problem reminded me the Explanatory Learning problem, while the APE is similar to the CRN approach there, maybe you could find this paper interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|9", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The selection + filtering is a new addition to prior work done by [1] which improves the results significantly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|14", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "I think the tasks studied in this paper could be more challenging.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|15", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "Taking a look at Table1, most tasks require surface-level understanding of language syntax.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|16", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "Translation task is only done for single words. Semantic tasks from GLUE are at 90+% accuracy with simple transformer models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|17", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "Few tasks here require complex reasoning, or deeper language understanding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|18", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "I think those tasks require creating more nuanced instructions and seeing how this approach compares against humans would be very interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|19", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "Until then. it's unclear to me how this approach generalizes beyond simple tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|20", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "To build on my previous point, looking at figure 3, results on relatively harder tasks such as sentence similarity and membership show APE is doing worse than humans and vanilla generated prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|21", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The authors acknowledge this in the paper: \"Additionally, tasks such as Membership and Second Letter are intrinsically challenging for the model, and APE consistently performs worse than humans.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|22", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "I think the paper can significantly benefit from further studies on tasks such as summarization, intent classification and slot filling, etc. that are more challenging.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|23", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "[1] Or Honovich, Uri Shaham, Samuel R Bowman, and Omer Levy. Instruction induction: From few examples to natural language task descriptions. arXiv preprint arXiv:2205.10782, 2022.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|25", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "Please refer to Strengths.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|28", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "However, I'm not quite convinced if this approach can scale to more challenging NLP tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#2|29", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "I hope authors address this issue in the updated version of the paper. I can revisit my score given the new results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|7", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The proposed idea is very interesting and this line of works trying to find the best language to query large language models is very trendy these days.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|9", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The paper though is not very well written and there are a few parts that require some clarification.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|10", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The assumptions/intuitions in section 3.1 are not clear. For example why \"would like the model to predict the missing context before the demonstrations\"? What does it mean more versatile approach?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|11", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "I understand that for every set of training pair there will be a possible different instruction, do you consolidate them?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|12", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "Which one do you choose at inference time?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|13", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "Real examples in the the Figures and task descriptions would simplify the reading a lot.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|14", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "After the selection of instruction using different metrics, a set of final instruction are returned, how do these extraction are used? prompted all of them in an in-context learning scenario?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|15", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "In section 4.1, please explain why Honovich et al. (2022) is the right benchmark for the proposed task, redirecting the reader to the original paper is not good writing, since it is not a well established benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|16", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "In section 4.1 would be useful to have examples of input/output in the zero-shot and few-shot in context learning scenarios.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|17", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "Why is the model only tested on the dataset released by Honovich et al. (2022), and not on more standard benchmarks like GLUE or big bench?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|19", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The paper is not very clear, the idea is novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|20", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The authors do not mention to release any code, so I am not sure about reproducibility.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|22", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The idea is novel and interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|23", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "The paper is hard to read although the approach is quite straightforward.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|24", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "I suggest the authors to spend some time to make it more clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc61|review#3|25", "source": "rc", "split": "test", "paper_id": "rc61", "comment": "UPDATE: The authors made the paper more clear and added a set of experiments on more benchmarks that make the paper much stronger.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#1|10", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "In a way, we can see the multi tasks as a single giant task with many domains.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#1|11", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "If we view this as question answering, the paper is essentially training a system that learns to answer well given various domains of questions/tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#1|12", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "What would also be interesting is to see/what's missing in the paper, is to carefully study the difference in domains of the training tasks, versus zero-shot evaluation tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#1|14", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "However, I do understand that it would be an empirically expensive set of experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#1|15", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "The conclusion of this paper is that multi-task supervised learning makes the model generalize well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#1|17", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Questions Can authors provide intuition about the performance of T0 vs FLAN on the datasets that T0 does better, or worse?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#1|18", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Do it mainly have to do with auto-regressive style versus encoder-decoder (that T0 does not predict the input)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#1|20", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "This is a solid paper in terms of contribution to the language modeling line of work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#1|21", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "It will steer the community in the right direction of doing multi-task learning as a way to meta-learn new task or adapt to unseen tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#2|8", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "The idea of converting all the tasks to an unified format is not new, and has been explored in earlier works like DecaNLP, UFO-Entail, T5 and more recently in the CLUES few-shot benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#2|13", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "However, this analysis is missing for the training tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#2|14", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Since the objective of this work, is to evaluate \"true\" zero-shot performance of the models, it is necessary to perform this analysis to ensure that any text for the \"unseen\" tasks and datasets is not leaked from the training corpus.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#2|15", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "This is important since it is often difficult to explicitly partition tasks into disjoint sets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#2|16", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "For instance, the authors evaluate their models on held-out NLI tasks, but train their models on paraphrase identification tasks that bear some similarity with the entailment tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#2|17", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Can the authors report the multi-task model performance on NLI tasks without using the paraphrase identification tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#2|18", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "The authors compare the performance of their model and demonstrate the multi-task version (T0) to perform better than the single-task version (T5-LM) and outperform GPT-3 in primarily NLI tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#2|19", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "The performance in WiC seems a bit strange given that random choice should give an accuracy of 50%.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#2|24", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "The claim of \"true\" zero-shot generalization needs to be further supported by two more experiments as highlighted in the main review, namely, (i) analyzing the overlap in text from the MTL setup with that of the unseen tasks, and (ii) ablating paraphrase identification tasks to find the impact on NLI tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|9", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "There are some major comments which are presented here,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|10", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "1- why MNLI and QNLI are not considered in NLI tasks? 2- Why yellow-colored tasks in figure 2 are not considered as unseen tasks too? In other words, why different held-out sets are not considered for unseen tasks?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|11", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "3- The authors mentions that humans are not explicitly trained on NLI tasks, story completion, coreference resolution, or word sense disambiguation, and for that reason, these tasks are selected as held-out to test zero-shot generalization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|13", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "4- Section 5: the author mentioned training encoder-decoder model is computationally efficient since it is trained to only generate the target as opposed to GPT3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|14", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "However, isn't the model use more supervision to generate the input as well, which can benefit the zero-shot performance?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|15", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "later, the author mentioned that since T5 is pretrained with masked language modeling, which is different than conditional text generation, they used T5LM model, the language model adapted T5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|16", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Isn't this contradict their previous statement about benefits of target-only generation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|17", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "5- how about T5 performance with different sizes?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|18", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "6- it is mentioned that the best checkpoint was selected from step 12200 since it yields best validation metric.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|19", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "However, the training set of T0, T0+ and T0++ are different.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|20", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "are they using the same validation with same tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|21", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "7- Figure 4: why T0LM is not trained on the training mixture dataset?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|22", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "it is mostly underperforms T0", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|23", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "8- author mentions that they did not do prompt selection using best evaluation performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|24", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Since zero-shot tasks are different than training, don't they have a completely different prompt template?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|26", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "This way, the T0 can be evaluated on more BigBench tasks which has not in-vocabulary T5 tokens.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|27", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Moreover, using pretrained T5 model for T0 means using implicit multi-tasking trained during T5 pretraining.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|28", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "10- it would be interesting to see the T0 performance which trained with different number of mixed training tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|29", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "In other words, what is the impact of each training task, to zero-shot learning", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|30", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "11- T0 outperforms 7 out of 11 zero-shot datasets, but it outperforms on 2 out of 4 zero-shot tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|31", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "It underperforms on all datasets of story completion and coreference resolution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|32", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "what is the explanation for this?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|34", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "13- the contribution is not clear. if it is about the benefit of encoder-decoder model to decoder-only LM models for zero-shot learning, then a decoder-only LM model should be trained on the multi-tasked prompted training set. if the contribution is about the proposed multi-tasked prompted dataset", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|35", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "14- Figure 5: there is no constant improvement between T0, to T0+ and T0++.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|36", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "In other words, adding more dataset to multi-tasked trained does not always improve zero-shot learning on BigBench.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|37", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "what is the justification for this behaviour of encoder-decoder model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|38", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "if using more training data sometimes reduce performance, e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|39", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Known Unknown and Logical Deduction tasks, perhaps due to model-scale constraint, it would be helpful to evaluate with a larger T0.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|40", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "In other words, what is the relation between T0 model scale and training size to the zero-shot performance", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|41", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "15- Section 6.2: the author compared T0 with GPT3 on RTE tasks with different prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|42", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "what is the reason for just choosing RTE for this comparison only?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|43", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "shouldn't this comparison be done on all zero-shot tasks of Figure 4, to evaluate the robustness of GPT3 to prompts as well?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|44", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "16- The performance of T0 can also be compared with FLAN model in Figure 4 for each zero-shot task, despite different training set.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|46", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "18- Section 7: author mentioned that FLAN model with comparable size to T0 (8B) finds that increasing multi-task training reduce zero-shot performance, whereas they find the opposite.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|47", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "The only evaluation on the size of multi-task training dataset is presented on BigBench dataset in Figure 5, which is different than FLAN zero-shot evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|51", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "However, there are some shortcomings,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|52", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "1- The zero-shot evaluations are based on some selected tasks, including NLI, story completion, word sense disambiguation, and coreference resolution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|54", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "2- The author mentioned the benefits of encoder-decoder pretraining objective (target-only generation) to auto-regressive objective (input generation) as a key factor.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|55", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "However, there are no comparison with auto-regressive model with the same pretraining and zero-shot task evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|56", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "The only comparison is on the BigBench dataset, where Transformer-based language models are pretrained with different dataset, and also performance comparison on NLI, story completion, word sense disambiguation, and coreference resolution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#3|57", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "3- The paper is centered around understanding the multi-task pretraining, using explicit supervised learning of model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|18", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "I feel like your main baseline is the T5+LM model (T5 finetuned on the causal LMing task).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|19", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Therefore I think it would make the paper much stronger if you were to show how that baseline performs on the tasks in Figure 5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|20", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "I understand that the baselines that you use in Figure 5 were built and trained by an external third party, so I trust that these baselines aren’t too weak or unoptimzed somehow, but it would still be very important to show that you also improve over T5 in this setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|21", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "It would also be very interesting to see how well GPT-3 performs on those tasks, but I did not and will not penalize you for not having those numbers, as the GPT-3 API is not fully open and free.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|22", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "This paper seems like an extension of ideas discussed in iPET/PET (“It’s Not Just Size That Matters: Small Language Models Are Also Few-Shot Learners”, Schick & Schutze), and so to me it feels like it would be quite important to compare to that model in at least some of the experiments that you run, for example, if Figure 4 compared to iPET I feel the paper would be much stronger.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|23", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "What do you think?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|24", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Are T0 and iPET comparable?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|26", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Could you explain and/or cite relevant source “GPT-3 is behaviorally capable of zero-shot generalization to new tasks” (section 2) in the next draft?.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|27", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "I know that later on in the paper you do talk more about this, but I think it might also be important to explain this claim in section 2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|28", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "I think that many people believe that while GPT-3 wasn’t explicitly trained on tasks such as translation or adding numbers, GPT-3 can only do well on “zero-shot” tasks that it has implicitly observed during training", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|29", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "I’m using the chrome PDF reader and do not see an actual figure above the caption for figure 3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|30", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "“Unlike decoder-only language models such as GPT-3, is not trained to generate the input.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|31", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "This has computational benefits since in our prompted data, the target typically is much shorter than the input”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|32", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "You might not need to generate the input but you still need to encode it and so you only save the time taken on the softmax layer which is insignificant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|33", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "In addition, since you use an enc-dec model you need those extra encoder-attention sublayers so it’s not so clear that there’s an efficiency advantage here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|34", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "I would like to ask you to consider adding some actual experiments that show that this claim is correct, or cite a relevant paper, or just remove that sentence.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|35", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "“Baseline models are standard Transformer-based language models with varying size.” (Figure 5).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|36", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "What is a ‘standard transformer LM’?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|37", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "I don’t think such a thing really exists.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|38", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "I think it could be helpful in the paper if you would give more details, such as: position embedding, FF dimension, number of heads, head dimension, and number of layers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|39", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Make sure Figure 5 is explicitly referenced in the text in the relevant paragraph.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|40", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Consider making the following statement slightly weaker: “These results suggest that T0 is more robust to prompt formulation than GPT-3”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|41", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "It’s hard to draw such a strong conclusion from an experiment on one dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|42", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "In figure 6, what is the meaning of each dot?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc62|review#4|43", "source": "rc", "split": "test", "paper_id": "rc62", "comment": "Is it the performance for a given prompt? If so, that should be explained in the caption.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|5", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "That said, I have several concerns that make me question some of the claims of the paper:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|6", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The human evaluation shows very little performance difference between generation with data prefix (D-prefix) and model prefix (M-prefix), suggesting exposure bias is not a problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|7", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "But I would argue this is mostly an artifact of the experimental setup, as prefixing generation with D-string (of length L) only eliminates exposure bias up to position L, and then both evaluated systems (according to the generation process defined Section 4.1) use a standard auto-regressive LM generation process that makes them both subject to the exposure bias problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|8", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Since these generated strings are of length 30, there is plenty of room for exposure bias to crop in.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|9", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "(Indeed, Zhang et al., (2019) and Holtzman et al. (2019) have shown concrete examples of exposure bias artifacts appearing in much shorter sequences).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|10", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "So, the small performance difference between a preference M-prefixed and D-prefixed generation may very well be due to both setups being almost identical (same model, same auto-regressive inference algorithm, and the only difference is the prefix – which human raters are not even asked to judge directly).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|11", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "If exposure bias is indeed a problem, then it would affect both systems almost the same way, so this human evaluation can’t be used to either affirm or deny exposure bias is a significant problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|12", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "I also have concerns regarding the automatic evaluation based on BLEU.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|13", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "What the authors call “corpus-BLEU” is actually not the standard version of BLEU (see “Other Comments” below), as the version of the paper looks at n-gram (n=1 to 3) matches between the generated sentences and a large set of references that is not specific to a particular context.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|14", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "As the paper’s automatic evaluation is done in a completely context-agnostic way and relative to a large pool of references, it essentially only measures whether the model (with or w/o exposure bias) is able to generate plausible trigrams, but that sets the bar very low as we already have plenty of evidence showing neural language models are quite capable of generating reasonable trigrams (whether there is exposure bias or not), and in fact often much longer n-grams.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|15", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The authors claim that auto-regressive models appear to have “self-recovery ability” that mitigate any exposure bias, but I would say that it is hard to claim anything about self-recovery (at least in terms of automatic evaluation) when the evaluation metric operates over such a small window (<=3 words).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|16", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Abstract: “Although a lot of algorithms have been proposed to avoid teacher forcing and therefore alleviate exposure bias, there is little work showing how serious the exposure bias problem actually is.” I find this claim a bit misguided, as the numerous papers addressing exposure bias are empirical ones.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|17", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "They have shown improvements in various tasks such as machine translation, image captioning, and other generation tasks thanks to techniques aimed at reducing exposure bias.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|18", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Now, one could claim that significant improvements shown in these papers are due to reasons other than exposure bias, but if that is the case then the submission doesn’t do a much better job at isolating exposure bias from other factors (given my concerns in (1) and (2)).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|19", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Note that Zhang et al. (2019) actually does include a short section attempting to quantify the effect of exposure bias.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|20", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "For a paper that attempts to challenge the current understanding of a problem that has received very significant attention (i.e., exposure bias), it is quite thin in terms of related work (2 paragraphs).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|21", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The paper omitted important related work (e.g., Schmidt, 2019; Tan et al., 2019; Rennie et al., 2016), including an ACL best paper on the same topic and whose findings appear to be at odds with the current submission.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|24", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "That said, the paper makes rather strong claims (e.g., that “performance gain is minimal” if exposure bias is supposedly eliminated) that I find unfounded given my points in (1) and (2).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|25", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "In both evaluations, the setups evaluate a given model sequence W prefixed by a string sampled either from the data or the model, but that does not eliminate the fact that exposure bias is bound to appear within the generated sequence (length 30), which is the sequence that is evaluated in the end.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|27", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "“Corpus BLEU”: Note that BLEU, as originally defined in (Papineni et al., 2002), is in fact a corpus-level metric, as it aggregates n-gram statistics over an entire (test) corpus.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|28", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "So “corpus” in “corpus BLEU” is redundant and possibly confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|29", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The term “corpus BLEU” or “corpus-level BLEU” is generally used to contrast with various versions of “sentence-level BLEU.” Now it appears that “corpus BLEU” in this submission refers to the version of BLEU used in SeqGAN (Yu et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|30", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "), which is therein not called “corpus BLEU.” That distinction should be made clearer, considering that the use of “a large number of sentences from ground-truth data as references” is a significant departure from how BLEU was originally designed to work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|31", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Indeed, (corpus-level) BLEU (Papineni et al., 2002) doesn’t allow matching hypotheses and references across test instances, as opposed to the submission.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|32", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The authors justify their use of BLEU as it is a “well-established [metric] in the NLG literature,” but this is rather misleading as their specific version of BLEU is not well established and not what is commonly used in MT and NLG.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|33", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Table 1 doesn’t actually illustrate what the experimental setup of the paper does (i.e., Section 4.1), as the latter doesn’t include a shared prompt of 20 words that is supposed to make the two generated strings “more comparable”.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|34", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Since this prompt supposed to make the two strings more comparable is absent from the actual evaluation setup, I gather from the authors’ own words that they implicitly admit that string comparisons in their evaluation setup are not so comparable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|35", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Zhang et al., 2019: https://arxiv.org/pdf/1906.02448", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|37", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Rennie et al., 2016: https://arxiv.org/abs/1612.00563", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|38", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Tan et al., 2019: https://arxiv.org/abs/1811.09740", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|39", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Holtzman et al., 2019: https://arxiv.org/abs/1904.09751", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#1|40", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Chen and Cherry, 2014: https://www.aclweb.org/anthology/W14-3346/", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#2|5", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Concerns: The experiments only look at pure sampling from the model---i.e., they don't use temperature, nucleus sampling, greedy or beam search.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#2|6", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "These other generation settings are more important for applications, compared to pure generation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#2|7", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Also, they have been reported to exhibit more deviation from the corpus distribution (especially with respect to particular pathologies like repetition).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#2|8", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Evaluating generation in these other settings would increase the impact of the paper's conclusions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#2|9", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The human evaluation is helpful, but needs measures of inter-annotator agreement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#2|10", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Less significant: The qualitative experiment that starts the paper is very anecdotal, and I think the paper would be better off without it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#2|11", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The concrete examples in Table 1 are helpful for illustration, but I would not dedicate a section to them or call it an ‘experiment,’ as it is too limited in scope.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#2|12", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "It detracts from the stronger experiments later in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#2|13", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Likewise, the model in Example 1 seems too inaccurate to be illustrative, and the associated discussion in Appendix D is already well known I believe (that the “Teacher forcing” MLE objective just aims to choose parameters that maximize the likelihood of the corpus, using the chain rule).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|10", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "This contradicts the usual expectation about the EBP, namely that in the (M) case, the model would be \"lost\" in unknown territory and start to produce very poor text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|11", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The main conclusion of the authors is that: \"although the mismatch between the data and model prefix distribution exists, it is still in the model’s “comfortable zone”, and is not large enough induce drastic performance loss during generation\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|15", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The fact that the generation model does not move away from its \"comfort zone\" when generating prefixes could be related to two different dimensions: First, the prefixes that you generate are only moderately long, thus alleviating the EBP issue.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|16", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Second, you do not describe exactly how the generation is done.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|17", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "I gather that Transformer-XL uses a form of \"top-k\" sampling, that is, a generation mode more restricted than a standard (from the pure probabilistic viewpoint) \"ancestral sampling\". Such restricted sampling (as also beam-search) is known to improve the quality but limit the diversity of the generations. It would be interesting to see if you would obtain the same EBP conclusions with ancestral sampling.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|18", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The corpus-BLEU measure that you use is most of your experiments was designed for Machine Translation and appears to me to be a very weak measure of quality for open-ended generation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|19", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "It would actually be informative to compute a baseline for that measure in terms of continuations not from the model, but from the training data itself, giving a measure of the average \"quality\" of a gold-standard continuation relative to all the other gold-standard continutations from the training data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|20", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "How good would this \"gold corpus-BLEU\" be?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|21", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Probably not so good, and it would be interesting to compare the (D) and (M) continuations to these gold continuations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|23", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The JS divergence is not a distance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|24", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Please clarify exactly when \"prompts\" are used in your experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|25", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "While they are used in the human evaluations, it was not clear to me whether they were used in the EB-M experiments, for example.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|26", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The paper https://arxiv.org/abs/1906.05664 \"Calibration, Entropy Rates, and Memory in Language Models\" might be relevant as related work: in that paper the authors argue that neural language models tend to suffer from \"entropy drift\", namely the tendency to entropy of the next token prediction to be higher when conditioned on an (M) type prefix than on a (D) prefix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#3|27", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "However, if I am correct, they assume standard hierarchical sampling, and their empirical evaluations are not extensive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|13", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The main issue I have is that the formalization of exposure bias is not well-founded and I do not believe a convincing argument is made the claims of relatively little (or even no) exposure bias are actually supported by the experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|14", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "While the ways the authors try to quantify exposure bias are intuitive, there is really no evidence that these metrics measure what they claim to.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|15", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Corpus-BLEU has been used in many papers, though I don’t know of any showing actual correlation with a human metric, but we can at least say that it has precedent being used cumulatively.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|16", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "However, using the corpus-BLEU of individual samples in a fraction creates the potential for completely inaccurate estimates.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|17", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Even assuming corpus-BLEU works, there is no guarantee that it acts as an estimate of individual samples, in the exact same way that BLEU is known not to be accurate for individual sentences (Burch, 2006).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|18", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Furthermore the EB-C metric is basically proposed from thin air.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|19", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "It makes sense to compare the distributions, but how big of a difference should we expect?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|20", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "GANs do badly on EB-C, but they are known to be significantly worse even on normal metrics, so it’s not clear what this proves.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|21", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The qualitative evaluation is just the narrative of what the authors think.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|22", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "No intuition is given, simply “we didn’t find exposure bias when we looked at the data”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|23", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "It especially shouldn’t be described as failing “to show the significance of exposure bias”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|24", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The authors refer back to these thoughts as experiments, e.g. on page 5 with “This agrees with our observations in the qualitative experiments.” This seems inappropriate.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|25", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The discussion is extremely rushed and it is not very clear what we are supposed to conclude from it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|26", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "The authors show that their metric is not complete via a toy example, but talk about how MLE does not produce these kinds of solutions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|27", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "That’s fine and I believe that part of their argument, but I still feel that there is not much evidence that the metric would actually show significant differences if the quality of the writing went down.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|28", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Quality is a tricky thing to measure.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|29", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "To the authors’ credit they conduct a human study.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|30", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "However, it is unclear to me whether humans that are exposed to a generated prefix (which may already deviate somewhat from human language) would feel that the mistakes made due to exposure bias “match” the mistakes in the prefix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|31", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "If the generations were truly horrible, I agree we would see more deviation in this score, but the paper attempts to disprove the very existence of exposure bias.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|32", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "A more appropriate reframing would be to talk about the kinds of effects exposure bias couldn’t possibly be having, e.g. that exposure bias may only result in strange word choice but not total degeneration.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|33", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "This over-claim, and the lack of smaller claims to build-up to it that might have been more appropriate, make it difficult for me accept the described conclusions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|34", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Finally, previous works on open-ended text generation have used the portion of the corpus released by OpenAI.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|35", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "It’s hard to know how comparable these results are, and since both GPT-2 and a significant number of validation text examples from WebText are freely available this seems like a flaw.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|36", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Why was wiki103 chosen instead of WebText used alongside the pretrained GPT-2?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc63|review#4|37", "source": "rc", "split": "test", "paper_id": "rc63", "comment": "Callison-Burch, Chris, Miles Osborne, and Philipp Koehn. \"Re-evaluation the role of bleu in machine translation research.\" 11th Conference of the European Chapter of the Association for Computational Linguistics. 2006.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|10", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "I am bit unconvinced about claiming that this work requires no external architecture in Table 1.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|11", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "There are new parameters that are getting trained/fine-tuned before we can do inference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|12", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "The main difference I see between this and another approach like p-tuning is that the most of the existing architecture is getting reused vs in p-tuning there is explicitly an external LSTM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|13", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "It would've been nice to see what happens if we constrain the additional tokens to be a part of the same vocabulary.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|14", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "This is more or less the approach taken in AutoPrompt, except with the current training methodology.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|16", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "I don't think I quite understood Sec 4.3, based on this section, it looks like there is an explicit embedding of output classes into an embedding?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|17", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "How are these learnt without fine-tuning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|18", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "If they are learnt from scratch, then how is this different from regular fine-tuning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|19", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "While the work claims that it reduces the need to fine-tune, I am seeing this work as a point on the spectrum between effort-to-tune vs zero-shot.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|20", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "Engineering prompts is truly zero-shot, P-tuning is somewhere in middle that it requires an external LSTM, whereas this approach requires retraining the big model (modulo all parameters except T1, T2.. held constant).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|21", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "Is this fair to say?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|22", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "How computationally easy is it to actually fine-tune these unused embeddings vs learning an external model", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|23", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "How do we decide the number of unused tokens T1, T2.. to use?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|24", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "How much does this work translate to a completely novel task beyond classification?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|25", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "e.g. let's say constituency parsing or dependency parsing task, for which it is possible to engineer a novel prompt.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#1|28", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "I had some comments/questions on the key insights, but I don't think it should affect acceptance, unless I misunderstood.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#2|11", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "Missing information: Because the proposed Fluency Constraint Object is based on MLM, GPT-2-medium (in section 5.5 and figure 2-b) cannot use this objective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#2|13", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "In section 4.5, the authors said that their method requires no external parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#2|14", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "But this claim could be controversial since adding new tokens (for prompts and labels) on the pre-trained language models' token embedding seems more efficient than pre-training language models that already have unused or special tokens in their token embedding. I know many PLMs already have unused tokens in practice, but I would like to treat this claim theoretically.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc64|review#2|15", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "For GLUE tasks, I'd like to know the performance of DART with full training set.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#2|16", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "Comparing with Prefix-tuning and WARP in experiments would be great.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#2|17", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "But this is not mandatory.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#2|18", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "In figure 3 and 4, training iteration or epoch should be presented.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#2|23", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "Especially, focusing on differentiable label would be important observation for the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#3|12", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "From the results, only two datasets (SST-2 and MR) seem to have significant improvement over LM-BFF.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc64|review#3|13", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "It would be great to further discuss these results in detail.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#3|15", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "Overall, the approach is technically sound and is also very easy to apply to other pre-trained language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#3|16", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "The empirical results on the 10 popular datasets seem closer to LM-BFF and it’s hard to interpret if the approach empirically performs better than static/fixed prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#3|17", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "On the other hand, the proposed method performs much better on relation extraction and event extraction datasets, and it is not clearly discussed in the paper on these variations in the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#4|9", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "The notation/description of section 4 is not immediately intuitive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#4|10", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "I had to read it a couple of times before I could fully follow the method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#4|11", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "The authors could add more content to figure 1, which may resolve this issue.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#4|12", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "The choice in deciding how many template tokens are used is unclear.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc64|review#4|13", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "An additional ablation experiment trying different number of tokens to optimize could be illuminating (even for just 1 dataset).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#4|14", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "The ablation study would be better if specific numbers were provided.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#4|15", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "I know GPT3 access is hard to get, but I wish experiments with k=8 prompts could be compared against", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#4|16", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "Post rebuttal: I think another pass for clarity over the paper would be good for the final version, but otherwise I'm happy with the paper updates and I'm happy to see the ablation numbers aren't too sensitive to token count.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#4|17", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "I have updated my score to reflect this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc64|review#4|20", "source": "rc", "split": "test", "paper_id": "rc64", "comment": "While it has a couple minor issues, I think the broader community would find this paper interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|1", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The paper proposes a new way to improve the reasoning power of language models by decoding several sequences of CoT/Scratchpad reasoning and then using a voting-based scheme to choose an answer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|2", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "At the cost of having to sample multiple times from the model, the authors show a consistent improvement across tasks and models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|3", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Through thorough analysis, the authors show the utility of their method, and its advantages over other sampling techniques and highlight the best practices that might be useful in practical applications.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|5", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Clarity: The paper is very clearly written and is easy to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|6", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The takeaways from each section are clear and flow from one to the other.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|7", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The tasks, baseline sampling techniques, models and various tradeoffs while using different hyperparameters have been explored well!", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|8", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "They give a clear picture of what is useful in a practical setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|9", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Reproducibility: The method is easy to implement, but not all the models are publicly available or are accessible only through a paid API.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|10", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The method itself is quite general and can be used across several task types and models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|11", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Further Discussion: CoT/ Scratchpads are useful in understanding the output of a model; how would the authors suggest choosing a reasoning path to provide an explanation for the answer given by the model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|12", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Novelty: The paper is novel in its approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|13", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Quality: The paper, analysis and writing are of a very high quality", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|14", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|15", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Refer to main review.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#1|18", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "It has very limited weaknesses and will probably have a large impact on various practical applications.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|1", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "This paper investigates an improved version of Chain-of-thought prompting by leveraging self consistency and shows empirical evidence that their self consistency prompting indeed improve reasoning tasks performance compared with the original chain-of-thoguht prompting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|4", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "1.The self-consistency method is very well motivated, and authors did a great job presenting this motivation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|5", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The method is very principled.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|6", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "2.The empirical evidence of the superiority of the method is extensive and solid.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|7", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Weakness: authors’ notation of r_i is a bit unclear to me, my understanding is that they are just the generated text before the final answer (what authors call reasoning path).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|8", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "It is not clear to me what is the difference between the generated text before the final answer and what authors call the reasoning paths.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|9", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Did authors do something specific to exclude some potentially irrelevant text that were generated but did not contribute to reasoning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|10", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "If not, perhaps calling them reasoning paths is a bit confusing?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|11", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "It seems r_i and a_i always appear together, which gives me the impression that the only distinction is that a_i is the final answer and r_i is everything before that.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|12", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Another confusion is the presentation of the weighted sum, is it correct to understand the probability weight is just the logsumexp of the logprobs of the generation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|13", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "That makes sense to be, equations 1 is a bit misleading to me because it seems to place an emphasis on r_i as being the reasoning path when it’s just the probability of a specific generation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|14", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "I think it is okay to mention the generation is supposed to be reasoning paths, but it might be more clear to just note them as generated text in those formal expressions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|15", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|16", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "For clarity and quality see the previous section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|18", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "However the empirical results are nevertheless significant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|20", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "I’d recommend marginal acceptance, as I think the self consistency method is very principled and well motivated and empirical evidence is extensive and solid.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#2|21", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "However the notation and the presentation could be improved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|1", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "This paper proposes a decoding strategy to improve chain-of-thought (CoT) prompting in pre-trained language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|2", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The proposed strategy uses nucleus sampling to sample diverse outputs as opposed to a single one in CoT (based on greedy decoding) and selects the answer that is most consistent through marginalization.", "reviewer_type": "human", "label": "error", "error_type": "Inaccurate Summary", "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|3", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Evaluation performed on arithmetic and commonsense reasoning shows large improvements compared to CoT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|4", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Results are consistent across different language models and sizes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|7", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Proposes a simple and effective decoding strategy to improve LMs with CoT prompting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|8", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Contains extensive experiments with different LM types and sizes that demonstrate large gains compared to vanilla CoT without requiring additional training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|9", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Self-consistency strategy surfaces diverse outputs that correspond to rationales in CoT-prompted LMs and improves robustness to imperfect inputs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|10", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Also, it can provide uncertainty estimates for calibration.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|11", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Weaknesses The decoding strategy itself has limited novelty since sampling strategies (including nucleus sampling) and ranking according to model scores have been previously used for pre-trained LMs albeit not with CoT prompting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|12", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Even though it generally leads to improvements, the proposed decoding strategy creates a significant overhead proportional to the number of outputs sampled e.g 5x latency for 5 samples assuming that samples have equal lengths.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|13", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The marginalization process of self-consistency assumes that answers are coming from a fixed answer set, which decreases the scope of applicability.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|14", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|15", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The paper is well-written overall and its motivation is quite clear i.e.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|16", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "improving recently proposed CoT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|17", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The claims are supported by extensive experiments with a variety of model configurations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|18", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "I also do not have any major concerns about reproducibility since the authors provide sufficient descriptions and artifacts (code, prompts, model outputs, etc).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|19", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "In terms of originality, sampling different model outputs and ranking them according to a metric has been largely explored in the past (e.g. MT, text generation).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|20", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "In the context of pre-trained LMs, the GPT3 paper used beam search results for free-form completion, Adiwardana et al., (2020) used a very similar sample-and-rank method based on regular sampling, and more recently Jung et al. (2022) used nucleus sampling to sample diverse generated outputs to improve reasoning [1].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|21", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The last one is arguably parallel work but I am surprised it was not cited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|22", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Having said that, the approach itself is not that novel but it is a straightforward direction to explore for CoT LMs and the community can benefit from such a simple method and large-scale empirical findings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|24", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "It looks like that self-consistency is an instance of the sample-and-rank method that uses nucleus sampling instead of temperature sampling and majority vote (simple sum) instead of log probability.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|25", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Do the authors share this view?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|26", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "I'd suggest making the connection clear to avoid misinterpretations regarding novelty.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|27", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "In table 1, could the authors specify on which set the accuracy scores are computed?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|28", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "It's likely not very crucial since the majority vote was eventually selected, but it would be useful to know.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|31", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Overall, this paper conducts a large empirical study of a simple decoding strategy for pre-trained LMs that use CoT prompting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|32", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The decoding strategy itself has limited novelty but the empirical findings are quite interesting/useful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#3|33", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "The experiments show strong improvements over vanilla CoT prompting, which should be of interest to the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|1", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "This paper proposed a new decoding and answer generation method for LMs to achieve multi-step math problems in a “step-by-step” way.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|2", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Different from the standard “chain-of-thought” prompting which uses greedy decoding, this method samples several reasoning paths generated by LM, and then generates the final answer by finding the most frequent answer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|3", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Author conducted a series of experiments varying in LM type and LM size to assess the self-consistency method’s performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|8", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Weak: Self-consistency is an interesting point.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|9", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "However, my main consideration is that this paper does not go deep enough into the study of self-consistency.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|10", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "This lack of depth is mainly reflected in two perspectives (1) only the majority vote of multiple randomly generated answers is considered.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|11", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "There is a lack of in-depth observation on how to combine different answers; (2) different answers are obtained by controlling the temperature and k. For text generation, the effect of these two hyperparameters is unclear. In particular, how these two hyperparameters affect answer consistency is unclear.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|12", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Other weak points: 1 Experiment results suggest that LM cannot distinguish well between correct and incorrect answers (at least from the aspect of token probability), but why taking majority vote over answers generated from different reasoning paths works?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|13", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "More in-depth studies should be done.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|14", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "2 Since “diversity of reasoning paths is the key to better performance”, why other ensemble-based approaches (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|15", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "the multiple prompts setting) didn’t get larger gain for their diversity?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|16", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "3 Figure 2 is confusing: why greedy decode has different number of sampled paths?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|17", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "4 It would be better to show full sampled paths in table 4 and other examples if the purpose is to claim self-consistency can repair errored reasoning process.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|18", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|19", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Please refer to Strength And Weaknesses.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|21", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Overall, I suggest that the authors do a deeper study of the problem, e.g., for LMs, how does the problem of consistency arise and what is its fundamental nature?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|22", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "What exactly does improving self-consistency mean?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc65|review#4|23", "source": "rc", "split": "test", "paper_id": "rc65", "comment": "Why does majority vote improve the model effect?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|8", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The motivation for the study proposed in the paper is interesting for a number of reasons.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|9", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The volume of computation required by many modern transformer models has been prohibitively expensive and therefore out of reach for most researchers for quite a while.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|10", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "By studying the implications of constraining the computational resources on the ability of the model to perform well on certain tasks the authors could provide a way for researchers with limited budgets to participate and utilize these models in fundamentally new ways.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|11", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The trend in the paper to consider modifications that mainly reduce the gradient update cost without significantly impacting the total number of parameters in the model, based on the scaling laws, provides an interesting and unifying theme throughout.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|12", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The persistence of the scaling laws to influence the performance of the model on tasks is reinforced through empirical evidence throughout and yields interesting insights.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|13", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "Performance evaluation on a shoe-string budget of FLOPs compare to other prominent models is impressive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|15", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "Similar studies were conducted on a single node with 8 GPUs as noted by the authors. Though that setup had considerably more computational resources the total volume of computation was still a fraction of the amount used by many large research institutions. In light of that work, the scenario presented in this paper may seem somewhat derivative and only marginally interesting.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|16", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "It is not clear if or how the observations made in the cramming regime may be used to make more informed decisions regarding the training process in the normal training setting.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|18", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The writing is clear and the presentation of issues motivating the current work is adequately articulated in the text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|19", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "While I am not an expert in the transformer field I feel the authors did a good job explaining the connection between the scaling laws and the downstream performance of the models under consideration.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|20", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The novelty of the work pertains to the training strategies used to reduce computational costs without removing the total number of model parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#1|21", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "Although previous works looked at training with limited resources the author's study and extreme training scenario that is likely to be more pertinent and representative of the resources available to typical, non-institutional, researchers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|3", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "There are some interesting findings, such as per-gradient efficiency only depends on model size, several strategies to filter and sort the training data brought improvement, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|4", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "As a results, the authors were able to push the performance close to BERT base if excluding some outlier tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|7", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "This paper adds more insights on scaling-down, which is less understood as most concurrent efforts are around scaling-up.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|8", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The experiments have a good coverage in terms of testing various architectural changes from recent literature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|9", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The final performance on downstream tasks (GLUE) are impressive given the limited compute budget.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|11", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The biggest missing piece is an ablation study on the improvement from various architecture changes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|12", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "In the current version, the authors provided a comprehensive list of things they tried, what helped and what didn't.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|13", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "But it's unknown which change(s) brought the bigger improvement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|14", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "Related to this is the poor performance on CoLA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|15", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "Although the authors provided several hypotheses, some of them should be tested to verify whether they're actually related to any of the architecture or data changes, or mostly due to reduced model size.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|16", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "For example, one possible cause provided by the authors is that reasonable performance CoLA would need more training data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|17", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "But are models in these experiment trained on less data compared to BERT-base?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|18", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "Is it possible to see whether the performance gap can actually be closed if the models were trained longer than one day (i.e.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|22", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "For all the architectural modifications in Sec 4.2, does \"no improvement\" refer to pretraining loss or downstream tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|23", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "In Sec 4.3 batch size schedule, you found optimal performance from different batch size for pretraining (1536) and downstream tasks (4032).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|24", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "Why do you think pretraining loss benefit from smaller batch size?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|25", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "Similarly, could any of the architectural changes in Sec 4.2 have different effects on pretraining vs. downstream?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|26", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "For all the changes in Sec 4.2 and 4.3, when you test a specific modification, what were used for the rest of the architecture and training setup?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|27", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "How were they chosen?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#2|30", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "To make sure conclusions from this work would apply, some additional experiments with longer sequence length would be helpful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#3|4", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "In this paper, several modifications (architecture, training setup and datasets) are explored to check whether there is any improvement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#3|5", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "All of these aspects are important and interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#3|6", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "These can give good insights for the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#3|7", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "Some interesting conclusion are got, for example, training recipe and data setup lead to decent downstream performance on GLUE.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#3|9", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The investigation of modifications lack convincing experiments. For example, only one task performance (MNLI) is reported for when studying the impact of training hyper-parameters. Other tasks can have a different trend.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc66|review#3|10", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "And when exploring the effect of the architecture, only MLM loss is report.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc66|review#3|11", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The performance of downstream tasks can be also important.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc66|review#3|14", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "Clarity is clear, however, novelty is limit.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|4", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The authors investigate a wide range of design choices in terms of model architecture, training recipe, and data curation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|5", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "They also note that the final MLM loss is mostly correlated with the FLOPs spent, not the particular Transformer type and size.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|6", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "This motivates them to choose architectures that parallelize well on GPUs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|7", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The final result demonstrates that some combination of the variants proposed in the three years after BERT yields a model that is almost as performant as the original BERT base while using 1/50 of the FLOPs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|9", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "This work presents a thorough empirical investigation on a topic that is of interest to many researchers who do not have access to large compute clusters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|10", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The overall methodology appears to be sound and the final result promising.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|12", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "An additional concern is that it is not clear if the MLM loss is the best metric to track when comparing different model/training design choices.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|13", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "This paper (https://openreview.net/pdf?id=F5uYcwABMu), for example, clearly demonstrates that models with near-identical pretraining loss can perform very differently on downstream tasks due to implicit biases.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|14", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "This should be taken into consideration since many conclusions in this work are based on the MLM loss alone.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|15", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "I encourage the authors to also observe downstream performance and note if they agree or disagree with the MLM loss.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|16", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The connection between “cramming” and scaling laws can be clarified.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|17", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The scaling laws mentioned in this work are empirical observations that a model’s performance strongly correlates its size but not necessarily its shape.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|18", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The empirical results from this work show that this holds for the low-compute regime, which is somewhat surprising.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|19", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "However, these “laws” are merely empirical observations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|20", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "It is not clear what the authors mean by “we discuss scaling laws in the low compute regime and find that cramming is hard because the conclusions of Kaplan et al. (2020) so easy to reproduce” (emphasis mine) in the conclusion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|21", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "It would be better to simply state that the empirical observation from Kaplan et al. holds in the setups investigated, which motivated using architectures that parallelize well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|22", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "A related concern is that since the dominating factor for performance is the number of FLOPs we can squeeze out of a GPU within a given timeframe, this makes the conclusions of this work somewhat hardware-specific, e.g., they might not hold on TPUs or newer/older GPUs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|24", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "The paper is mostly well-written with just a few typos.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc66|review#4|25", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "I am not aware of similar investigations in the low-compute regime and believe that many in the community might find this work informative.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#1|13", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "The paper investigates temporal generalization in language model pretraining.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#1|15", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "Moreover, the considered online learning approach is a straightforward application of the existing dynamic evaluation method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#1|20", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "Clarification question: For closed-book QA, in Fig 4 the performance of the model when pre-trained on a dataset up to 2019 drops when evaluated on a test dataset from 2019.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#1|21", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "The paper attributes the performance drop to confusion around the names of presidents.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#1|22", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "However, the pre-trained model is fine-tuned on the QA dataset from 2011.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#1|23", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "Does the model undergo forgetting when fine-tuned on factual knowledge from 2011?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#1|24", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "Typos: Lines 234, 245, 361 should be Table 3 instead of Fig 3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#1|25", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "Line 368 should be Table 4 instead of Fig 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#2|18", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "L149: Fig. 2 -> you meant Table 2?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#2|19", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "L234: Fig. 3 -> Table 3", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#2|20", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "L245: Fig -> Table", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#2|24", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "==== Post Rebuttal ====", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#2|25", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "I've read all reviews and the authors' responses.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#2|26", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "Seems that there's no significant issue that I overlooked in my initial review, so I'm keeping my score and still recommending the acceptance of this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#3|16", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "This paper discussed \"limitations\" but it did not discuss possible negative societal impact of the work -- the paper could be improved by adding such a section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#3|17", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "Even though this benchmark is not likely to cause harm, it might be good to think about what might result from future work in this area (on both modeling and evaluation side, with respect to the field and society).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#3|18", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "You could also discuss the relevance of evaluation datasets like this one in terms of reducing harms of NLP (e.g. bias being amplified? or historical prejudices being baked in?).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#4|12", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "I'm actually surprised by how relatively ineffective dynamic evaluation is, especially for the news data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#4|13", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "It'd be great to know why.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#4|14", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "One theory, for example, is that catastrophic forgetting is occurring and the LM is being adapted to the news domain but becoming worse in overall performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc67|review#4|15", "source": "rc", "split": "test", "paper_id": "rc67", "comment": "A simple test for this would be to run the model on held-out data that is not temporally shifted and see how much perplexity changes before/after dynamic evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|6", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "This paper takes a closer look at how the model will perform at different layers & attention heads when presented with incorrect labels.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|7", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "The findings generally confirm that PLMs are still largely impacted by incorrect labels, which is a bit contradictory to the messages in Min et al., 2022* but aligns with Yoo et al., 2022*.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|8", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "The experiments look convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|9", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "It's also interesting that the method of \"logit lens\" can work with the hidden representation at different layers without further training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|10", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "After zeroing out the later layers, the final output layer can still work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|12", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "My biggest concern is on the underlying assumption of this paper---the model should not be impacted by the incorrect labels.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|13", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "It's controversial here because when we say in-context \"learning\", we actually want the model to be truthful to the provided demonstration examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|14", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "I am willing to hear the author's thoughts on this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|15", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Despite that this work might be the first to show that the copy of incorrect labels emerges in the later layers, this finding is not that surprising considering the early layers are mostly responsible for fusing information, and the later layers are mainly for generating the final output (see the references).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|16", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Regarding the proposed method, although removing the particular components can mitigate the negative impact of incorrect labels, it's also questionable how this will affect the generation capability.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|17", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "It would be helpful if the author can show how much this will change the model's performance when the labels are correct.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|19", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "* Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Rethinking the role of demonstrations: What makes in-context learning work?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|20", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "* Kang Min Yoo, Junyeob Kim, Hyuhng Joon Kim, Hyunsoo Cho, Hwiyeol Jo, Sang-Woo Lee, Sang-goo Lee, Taeuk Kim. Ground-Truth Labels Matter: A Deeper Look into Input-Label Demonstrations", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|21", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "* Nelson F. Liu, Matt Gardner, Yonatan Belinkov, Matthew E. Peters, Noah A. Smith. Linguistic Knowledge and Transferability of Contextual Representations", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|22", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "* Jesse Vig, Yonatan Belinkov. Analyzing the Structure of Attention in a Transformer Language Model", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|24", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "The paper is clearly written.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|25", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "The methods mostly follow the existing \"logit lens\" and \"induction heads\" approaches, but apply them into their analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|26", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "The findings from the analysis are generally useful but might not be novel enough.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#1|30", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Although it's interesting to see such an analysis of in-context learning, the findings are expected overall, and the proposed method cannot really compete with the performance of full language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|5", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "The findings that later layers contribute more to the errors under incorrect demonstrations are quite novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|6", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "The proposed method of zeroing-out later layers or certain heads seem to be effective in reducing the gaps between correct and incorrect demonstrations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|8", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "I think the overall analysis presents quite interesting findings but the robustness and generalizability of the analysis is questionable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|9", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "There are many arbitrary choices in performing the analysis/mitigation, e.g., to ablate the heads in Section 4, for the head choice (Appendix A.1 has the details), the authors \"first consider the 25 heads with the highest label-attending score, then select the 10 heads with the highest class-sensitivity score\"; further some bias was added \"towards heads in the later layers, like selecting the 5 heads of 25 that belong to layers 20 and later\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|10", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "How are those choices made?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|11", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "They look very arbitrary to me and I'm not sure if it can generalize to other models or other tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|12", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "the early-exit strategy in Figure 3, it's true that the overall trend shows later layers contribute more to the enlarged gaps, but the gap happens at different layers for different tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|13", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Zeroing out all layers after 16 might lead to an overall better performance, but for many tasks the performance under incorrect demos have already started degradation long before layer 16.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|14", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "In addition, this layer choice is based on the overall pattern analysis on these 8 datasets, would this conclusion generalize to a new task/dataset?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|15", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "how would people apply the method in practice?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|16", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "From Figure 3, the accuracy under correct demos are highly sensitive to the #layers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|17", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Would early-existing sometimes cause significant performance drop if not applied carefully after the critical layer?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|18", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "In addition, it seems like people need to perform extensive analysis for each model/task combination to identify the critical layers / false prefix-matching heads (as they all seem different across models/tasks), how can one apply the method in a scalable way for any new/unseen tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|19", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "In Table 1, any analysis to explain why the ablation doesn't work for natural language inference (SICK)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|20", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Are there more tasks in this category analyzed to show is it a single-task failure or the method doesn't work for the entire category?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|21", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "As an analysis paper, some of the key ablations are missing from the study, e.g., the false demos permute the label space for the entire demonstration.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|22", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "What if some of the labels remain correct and some of them are wrong?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|23", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Would that change the conclusion significantly?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|25", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Clarify: the paper is very clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|26", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Quality: the paper presents some interesting findings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|27", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Novelty: the finding that later layers contributed more to wrong predictions under incorrect demonstrations is quite novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|28", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Reproducibility: no code is provided, not sure if the analysis in the paper can be reproduced.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#2|32", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Thus I recommend weak rejection for now.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#3|6", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Weakness: technical novelty (see next section)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#3|8", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Clarity: This paper is written clearly, reading it was enjoyable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#3|11", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "This makes this paper somewhat novel but not significantly novel since there is little algorithmic innovation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#3|12", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "But given the significance of the problem itself and the insights obtained, I still think this paper is above the ICLR standard.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#3|14", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "I recommend accept as I think this paper provided significant empirical insights on a significant problem, and the presentation of this paper is clear, and the logical flow is coherent.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|8", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "The topic is interesting and timely.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|10", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "It includes detailed analysis on how attention heads contribute to in-context learning by attending to labels in the demonstrations that share the same true label as the test input.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|11", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "In fact, as far as I know, this is the first work that shows how in-context learning works with respect to attention heads (while Olsson et al identified such induction heads, they focus on language modeling rather than in-context learning with real NLP datasets).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|12", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "This analysis is also supported by experiments that measure the impact of zero-ing those attention heads.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|13", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "It is also quite clever to use synthetic dataset (called “unnatural dataset” in the paper) to find false prefix-matching heads and use them for NLP datasets that are real tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|15", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "The entire paper is based on the assumption that the model has to be robust to false demonstrations, which is not convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|16", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "First of all, the demonstrations that are called “false demonstrations” in the paper are hardly “false” in my opinion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|17", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "They still preserve the valid mapping between inputs and (an abstract notion of) labels, but the surface form of the labels are permuted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|18", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "For instance, false demonstrations in the binary sentiment classification (positive->negative, negative->positive) can be seen as true demonstrations whose defined task is “map a positive review to the word ‘negative’ and a negative review to the word ‘positive’”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|19", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "Therefore, it is not undesirable that the model predicts a permuted label instead of a true label (e.g., predicts ‘negative’ to a positive review), since this is the valid task defined by the demonstrations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|20", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "In fact, some analysis in the paper (e.g. Figure 2b) shows that the model performance against original true labels degrades only because the model accurately predicts the permuted true label.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|21", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "I would be more convinced by the idea if the false demonstrations were demonstrations with random labels, because then the task defined by the demonstrations is not the valid task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|24", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "“False demonstrations”: As I briefly mentioned earlier, I think the term “false demonstrations” might be too abstract and broad to describe what the paper actually does.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|25", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "It would be better to say “demonstrations with permuted labels”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|27", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "In summary, I think this paper has a set of very interesting findings about in-context learning and novel analysis on the connection between attention heads and in-context learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|28", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "It is well-written and is very easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|29", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "However, I have a critical concern in the most important underlying assumption in the paper – that the “false demonstrations” in the paper is not actually the false demonstrations but is rather a re-definition of the task, and it is unclear why the goal is to make the model predict the original, true label.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|30", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "If authors/other reviewers agree with this, this means the paper has to be almost completely rewritten:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|31", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "for instance, the authors can re-run the same experiments/analysis with random labels (instead of permuted labels) and likely the same findings will still hold;", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc68|review#4|32", "source": "rc", "split": "test", "paper_id": "rc68", "comment": "or alternatively, the paper can be re-written to be about how the model is able to perform the task when the task is redefined with permuted labels whose semantic meaning doesn’t match with pre-training, attribute this to a few attention heads, and prove it is true by showing zero-ing out such attention heads significantly degrades performance again redefined (permuted) true labels.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|5", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "The authors compared the proposed method with a few meta-learning methods on a diverse set of tasks, and show that their method outperforms other methods for most of the time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|6", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "The empirical analysis and ablations also provide insights for understanding the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|12", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "There are a few minor notes and suggestions below:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|13", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "I think in the problem definition, Line 98, it would be better to explicitly define the learning stages and a shared text classification model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|14", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "The meta-learning stage has the access to the source tasks, and the few-shot generalization stage does not.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|15", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "And there is a shared model after the meta-learning stage.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|16", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "In the preamble of Section 4, it would be better to illustrate more the motivation of the design choice and define the \"adaptation parameters\" clearly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|17", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "It is vague in the current presentation and makes people wonder about the underlying motivations of the design.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|18", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "It would be better to visualize the task representation in some way and show the task-to-task distances so that we can inspect if the closer tasks can locate near each other.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|19", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Line 202, BERT-base only allows 128 tokens? I don't think so.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|20", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "It should be 512 if I understand correctly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|21", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Please double-check it and correct me if I was wrong about this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|22", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "A few typos: Line 19: \"a intermediate\"; Line 201 BERT-BERT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#1|23", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Limitations And Societal Impact:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|3", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "The paper provides comprehensive evaluation on 10 tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|4", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Grad2Task outperforms strong baselines (Leopard, ProtoNet-BERT) in most settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|5", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Analysis with a toy dataset shows the task representations are informative.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|6", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Ablation study justified several design choices in Grad2Task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|17", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Notation are sometimes confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|18", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "(1) In Eq.3 the function d is overloaded.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|19", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "The function outputs the task embedding in the first part.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|20", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "It outputs RNN hidden representation in the second part.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|21", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "(2) In line 7-8 in Algorithm 1, e_j is a new variable that seems to be newly-introduced.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|22", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Please also provide hints for the reader to link e_j with p(y|x_j), e.g., refer the reader to Eq. 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|24", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Is there any specific reason that RNN is selected as the task embedding network?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|25", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "How long is the training process?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|26", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "How large is the final model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|27", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Some comparison in model size and training efficiency may be included.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|29", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Eq. 7. Is the first stage done on the support set or the query set or both?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|30", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Line 274: “w/ XY”, Table 2 “w/ X&Y”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|32", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "I believe the right side of Fig 1 is actually zooming in the part of [e_l^{task}] —> [Bottleneck Adapter].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|33", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Please consider highlighting this information to make the figure more accessible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|35", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "The authors provide detailed discussion on Grad2Task's limitations in Sec 7.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#2|36", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "I listed some of my concerns and suggestions in the main review.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|9", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Overall, I found the paper very hard to follow, there are numerous typos and the exposition made it really difficult to extract the key contribution, the motivations for the specific choices, and the key take-aways.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|10", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "The paper seems a lot like an amalgamation of many related concepts, which is not a bad thing but in the current version of the paper there is no clear motivation for the various choices.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|11", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "To provide just one instance, if the main motivation is to use task embeddings then there are many straight-forward ways to incorporate them in existing models, such as in the input to the encoder or the classifier, and carefully considering/evaluating such choices would have been useful in motivating the approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|12", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Another choice made by the authors is to use only the values in FIM corresponding to the adapters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|13", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Why is it sufficient to only use these, is there any analysis on the quality of the task embedding when restricting to only the adapters?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|14", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Similarly, the motivation for using task embeddings to generate Film adapters is not clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|15", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "For instance, one can use Film adapters with regular fine-tuning on the few-shot set to learn “task-specific” Film adapters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|16", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Note that this baseline is not present in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|17", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "So how is including the task embedding (which also uses gradient information) helpful over fine-tuning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|18", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Motivation as well as empirical results in this direction are lacking.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|19", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Moreover, there are several concerns regarding experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|20", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "First, as mentioned above, there are relevant baselines that need to be provided to understand the value of the proposed approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|21", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "These also include: ProtoNet with fine-tuning, ProtoNet with adapters fine-tuning, ProtoNet-BN with only Film fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|22", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Currently, the main baselines for the paper are ProtoNet and ProtoNet-BN, none of which use gradient information like the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|23", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "These baselines will help demonstrate the utility of task embeddings over fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|24", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "The authors use datasets from [1].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|25", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "However, I see there are 17 datasets used in that paper and the authors picked 7 of them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|26", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "I am not convinced that seq-length is an appropriate reason to drop 10 datasets, for instance, the authors can run the model with a larger sequence length during evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|27", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Moreover, there are decisions like only using a subset of classes in one of the datasets (Huffpost), and why this is done is also not clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|28", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "As such, due to the use of non-standard datasets, I am not sure how to interpret the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|29", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Ablations are carried on all the downstream tasks used for final evaluation, which makes me think the results are overfit to these tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|30", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "I didn’t find these particularly insightful either.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|31", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "For example, ablations on these choices would have been more helpful: using/not-using adapters or Film adapters, hidden-size of adapters, fine-tuning adapters vs using task embeddings, alternatives to incorporate task embeddings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|32", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Some hyper-parameter choices are missing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|33", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "For example in section 4.2, what are the hyper-parameters of the RNN d()?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|34", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Similarly, what is the architecture and hyper-parameters of the MLPs a(.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|36", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "There is no analysis on what the task embeddings are capturing and why they are useful here. This would have helped motivate the paper quite a bit.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|37", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "These concerns together make me hesitant about the value of the task-embedding approach presented here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|38", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "The authors should consider a thorough re-writing of the paper to provide motivation and to explain their approach clearly, and consider the suggested baselines and improvements to experiments explained above.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|39", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "[1] Learning to few-shot learn across diverse natural language classification tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|40", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "[2] Meta-dataset: A dataset of datasets for learning to learn from few examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|41", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "[3] Fast and Flexible Multi-Task Classification Using Conditional Neural Adaptive Processes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|42", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "[4] Film: Visual reasoning with a 382 general conditioning layer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|43", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "[5] Task2vec: Task embedding for meta-learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|45", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "The authors provided a lengthy response with additional experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|46", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "These partly address some concerns but raise a few more.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|47", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "When using more tasks and relevant baselines as reported here, I can see that the margins of improvements are significantly less than what is reported in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|48", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "For ex, FT-PN-BN is competitive with the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|49", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Some of the new results are also very puzzling.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|50", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Like why does prototypical networks after fine-tuning becomes much worse than without fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|51", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Another related problem I hadn't noticed in the first review was the difference between prototypical and other fine-tuning baselines, this difference seems too large, contrary to previous literature, and warrants more exploration.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|52", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Comparing some results in the rebuttal and in Table 1, it seems that most of the improvements for the proposed approach are coming from the 3 new introduced datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|53", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "These results are also very surprising.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|54", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "For instance, what is the property of SNIPS dataset on which Leopard and PN suddenly perform at half the accuracy of Grad2Task (21, 28 vs 52)!", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|55", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "I have still raised my score in favor weak accept since the rebuttal did address some concerns.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|56", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "Authors should incorporate these results and the presentation improvements in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc69|review#3|58", "source": "rc", "split": "test", "paper_id": "rc69", "comment": "There is an appropriate discussion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|12", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "I am unclear of some of the experimental settings and how much we can conclude from them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|13", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Particularly, when is bootstrapping+self-amplification+ensembling are used? Is it used only during translation?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|14", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "It also seems odd to use those techniques only for zero-shot and not adapt them (or explore possible adaptions) for 2-shot (given the explanation in 4.4)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|15", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "My main concern is, however, that if XGLM is not using bootstrapping+self-amplification etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|16", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "in Table 2 I am not sure how fair of a comparison it is. Perhaps with those techniques XGLM would exceed mT5 by a high margin.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|17", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "(In appendix G the additional techniques over SAP seems to make a huge difference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|18", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Appendix B does show that mT5 performs on par with a much larger GPT3 on translation when both are self-amplified but still would be curious about XGLM given its multilingual emphasis similar to mT5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|19", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "However, I acknolwedge the effectiveness of SAP compared to other variants of prompts to mT5)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|20", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "As the paper acknowledge, SAP requires running the whole encoder-decoder architecture in an autoregressive loop. This can be slow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|21", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Particularly critical is that caching isn't anymore possible (wherein caching is generally usable to speed up autoregressive decoding in causal decoders), because the encoder is bidirectional and in the autoregressive loop.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|22", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "I am increasing the score to 8.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|23", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "While the weaknesses are not completely resolved:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|24", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "re. 1 there are still enough evidence that mT5 can be made to perform on par with undirectional decoders as the authors have clarified.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|25", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Regarding 2, authors are upfront about it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|26", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "There are still some limitations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|27", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "The conclusions rely only on mT5 but authors are clear about their justification and there are not many other models at a similar large enough scale.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|28", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "In terms of impact, I am not entirely sure how meaningful the result will be - given other works on improving Seq2Seq pre-training is already underway independent of this work and SAP, as it is now, may not be as practical for adoption.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|29", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "However, I think the paper still brings a neat empirical insight, and sets up a bar with SAP that we can attempt to approximate or overcome.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|30", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "It can also open questions and investigation of why this sort of autoregression over the whole seq2seq setup works so well and if there can be some ways to make to computation easier.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|31", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "There is also a decent amount of experiments in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|32", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Overall, given these factors, I decided on raising the score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|33", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Some (optional) discussion topic the paper may (or may not) add:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|34", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Not related to prompt tuning but - [1] and [2] also uses a similar strategy of puting the bidirectional encoder within an autoregressive loop and shows better compositional generalization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|35", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "This could be worth some discussion and encourage future directions in more closely studying what this form of technique is bringing to the table.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|36", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "[1] Disentangled Sequence to Sequence Learning for Compositional Generalization - Zheng et al ACL 2022 [2] Recursive Decoding: A Situated Cognition Approach to Compositional Generation in Grounded Language Understanding -Setzler et al. Arxiv 2022", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|37", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|38", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Clarity: The text, motivation, and methodological descriptions are mostly clear. However, I am unclear on some experimental details as mentioned above.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|39", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Novelty/Originality: The method is novel to my awareness, but somewhat incrementally novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|40", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "However, although in hindsight, the approach is technically just a simple modification over normal prompting, the method still seems moderately effective and serve an interesting purpose (in bringing bidirectional models on par with unidirectional ones at least for the tasks considered), and this hasn't been done in this form before (to my knowledge).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|41", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Moreover, \"simplicity\" here can also count as a virtue. So given these factors, the contribution can still be considered significant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|42", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Reproducibility: Besides some unclarities, there seems to be enough explication of the resources and methods for it being reproducible although lack of code is not ideal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|43", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Summary Of The Review:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|44", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "SAP, a technique to effectively zero-shot/few-shot prompt bidirectional language models, is introduced.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|45", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Simple method; results look decent. Bidirectional models are shown to have the potential to be on par with or better than unidirectional model on instruction following and in-context learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#1|46", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "However, some experimental comparisons may not be completely fair and inference speed may take a big hit.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#2|13", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "The main question I have is: why? I don't think this is spelled out clearly enough in the main paper. Do the authors expect SAP-like approaches to eventually overtake supervised approaches? If so, the questions around computational complexity are central and should be discussed in much more detail in the main section of the paper, including perhaps some FLOPS statistics for the proposed approach and the current SOTA.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc70|review#2|14", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "The supervised translation baseline include in Table 2 should perhaps be replaced with a more recent and stronger multilingual model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#2|15", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "I would definitely challenge the classification of languages such as ko, ar, sw etc. as being low-resource.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#2|16", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "These are all languages for which at least 1M training sentences can be found (see e.g. Table 1 in https://arxiv.org/abs/2207.04672)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#2|17", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Minor: I believe the meaning of \"Supervised\" in Table 2 is only explained in the Appendix. I would recommend explaining it in the table caption.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#2|18", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#2|19", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "As stated above, I believe this paper to be of high quality, both from a scientific point of view as well as for its clarity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#2|20", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "The proposed approach is relatively simple but, I believe, original. The authors give enough detail to reproduce the experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#2|21", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Summary Of The Review:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#2|22", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "A strong paper proposing a conceptually simple yet effective prompt-based learning approach for models like mT5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#2|23", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Some questions around computational efficiency I believe should be given a bit more prominence.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|10", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "The technical novelty of the paper is somewhat limited, since the SAP prompt technique is just feeding the generated token back the prompt again so as to get the new representation of the encoder.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|11", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "The paper only verifies the effectiveness of the proposed SAP method on the mT5 models, however, claims that the performance improvements come from the Bidirectional LMs, which is not reliable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|12", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "This part of performance improvement may only come from SAP's effect on mT5, or SAP's effect on the encoder-decoder architecture model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|13", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Not only that, the underperforming attempt of running SAP on the T5 1.1 proves the unreliability of this conclusion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|14", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Lack of detail analysis on the computational efficiency.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|15", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|16", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "The quality is relatively good.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|17", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Summary Of The Review:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|18", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "This paper proposes an inference-only SAP prompting technique for mT5 model to achieve remarkable results on low-resource multilingual generative tasks, including MT, QA and Summarization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|19", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "The proposed method outperforms unidirectional models (GPT-3, XGLM) within fewer parameters and in-context training examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|20", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Even though the paper presents sufficient experiments and analysis, it does not give strong evidence to support the point that the low-resource performance boost comes from bidirectional pre-training objectives.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#3|21", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "What’s more, the technical novelty of the paper is kind of limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#4|15", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "This will make the inference more efficient as well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#4|16", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#4|17", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "The paper is well-written and understandable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#4|18", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "The procedure to adapt se2seq pretrained LMs for prompt-based learning is novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#4|19", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "The methodology for boostrapping prompts for unsupervised MT comes from existing working, but an extension in the form of English-centric bootstrapping method has been proposed. This is shown to be useful in some contexts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#4|21", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "Table 4/5: Why does T5+LM even lag behind vanilla mT5+SAP? Why does it also perform worse compared to vanilla mT5 on summarization?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#4|22", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "It is trained on a prefixLM kind of objective, so it does have bidirectional context as well as a causal decoder.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc70|review#4|23", "source": "rc", "split": "test", "paper_id": "rc70", "comment": "An analysis of the same would be useful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|7", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "But its execution can be significantly improved:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|9", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "LoRA only finetunes W_q and W_v. One missing baseline is a full-rank finetuning model, which only updates W_q and W_v.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|10", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "The memory/parameter efficiency of LoRA is exaggerated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|11", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Thorough proof reading is needed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|12", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Please see below for a long, but by no means exhaustive list of typos.-", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|13", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "One may argue that LoRA can take more training steps to converge due to low-rank updates.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|14", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "So it would be great to compare learning curves and training time overhead.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|16", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "What does VRAM stand for?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|17", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "If it is GPU memory, then the 2/3 saving of training memory is quite surprising to me—LoRA still needs to build the forward and backward computation graphs for the whole model.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|18", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Could the authors clarify?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|19", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "A (probably more general) variant of LoRA would be, instead of learning A and B, one forces the gradients to be of low-rank.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|20", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "So accumulating multiple low-rank steps could still yield a higher-rank update to the parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|21", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Have the authors considered this?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|23", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "“Reduce something by 10K/3 times” is strange.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|24", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Please consider changing it to reduce by a factor of 10K, or reduce by 2/3", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|25", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Please P(X), P(X = x), and p(x).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|26", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Whenever the paper talks about autoregressive language models, the math is never a language model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|27", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "A non-issue -> not an issue", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|28", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Less performant -> underperform", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|29", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "These layers have full-rank -> these layer have full ranks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#1|30", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "An inevitability -> Inevitable", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#2|14", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "One limitation of this method (also mentioned in the paper) is how the model works for multi-task serving setting, which is possibly the most common scenario in deploying large-scale pretrained models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#2|15", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Since LoRA is able to meet this requirement by not merging the pretrained weights (W_0) and low-rank weights (BA), it will be interesting if the author can report the comparison results of inference latency among (no-merging) LoRA and other methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#2|16", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Moreover, with regard to training efficiency, it is better to compare with adapters and prompt/prefix-tuning with the same amount of trainable parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#2|17", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Another line of work on mixture of experts (MoE) also shows promising results on mult-task/domain training (e.g., DEMix Layers: https://arxiv.org/abs/2108.05036), where we can selectively train expert modules for specific downstream task (while freezing the rest parameters).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#2|18", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "It will be interesting to see how RoLA compares with it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|11", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Missing comparison to the SOTA models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|12", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "The training efficiency remains unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|13", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "It would be useful if the authors could report the time during training and compare the latency to the baseline models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|16", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "I would say that the idea is relevant for achieving efficient pre-trained models that will be useful for many practical applications where latency is crucial.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|17", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "The paper is also well-written, with numerous experiments to support the claim.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|18", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Therefore, I think the claim is supported with empirical results; however, the paper's novelty is limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|19", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "The time efficiency still needs to be supported by training and inference time information.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|20", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "The approach is incremental, very similar to Adapters; adding small trainable parameters to the model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|23", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "References related to Matrix decomposition by design (please consider to add):", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|24", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Compressing Pre-trained Language Models by Matrix Decomposition", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|25", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Greenformer: Factorization toolkit for efficient deep neural networks", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|26", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Factorization Tricks For LSTM Networks", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|27", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "Lightweight and Efficient End-to-End Speech Recognition Using Low-Rank Transformer", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|29", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "What is the optimal rank for the factorized matrices? I was wondering to know if you have findings on choosing the optimal rank in the experiment.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc71|review#3|30", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "I would be happy to increase my score if the authors address my concerns.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#4|22", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "The training efficiency of LoRA is a bit unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#4|23", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "In \"Practical Benefits and Limitations.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#4|24", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "the authors mention a 2/3 reduction in VRAM by not storing the gradients of the previous steps in Adam, which is true but it is important to mention that the gradient of the original weights is still needed (although not stored) at the current optimization step to compute the gradient for the low-ranked weights (if I understood correctly your method).", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc71|review#4|26", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "I would encourage the authors to provide a discussion about the training efficiency.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#4|27", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "The authors did not report any baselines outside of the efficient fine-tuning setting.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc71|review#4|28", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "This is important for the reader to understand how far an efficient model is compared to the current SOTA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc71|review#4|29", "source": "rc", "split": "test", "paper_id": "rc71", "comment": "By no means, the reader, and me, expect new SOTA, but it is good to have these references in the table.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|6", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Memorization is a key problem to stop, and more empirical techniques are needed that do not sacrifice performance as differential privacy does.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|8", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "One of my big concerns is regarding the privacy evaluation of the method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|9", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The evaluation focuses specifically on preventing exact verbatim memorization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|10", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "A trivial baseline for stopping verbatim memorization is to just manually block the model from generating exact strings in the training set, potentially using a suffix tree like is done in Kandpal et al. 2022 and Katherine Lee et al. 2022.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|11", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Of course, this trivial baseline cannot prevent models from privacy risks such as (1) generating paraphrases of training texts, (2) leaking private information in a QA or dialogue setting (e.g., ask someone's SSN), and (3) cannot stop membership inference attacks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|12", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "To show the benefits of the method, it'd be great to see results in some of the above (or related) evaluations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|13", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "For example, run the membership inference methods from Kandpal et al. 2022, Carlini et al. 2020, or others.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|14", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Definitely checkout the privacy onion paper (Carlini et al. 2022) for another potential risk---unlearning some data can expose others.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|16", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "I have some concerns regarding the technical novelty.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|17", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The general idea of gradient ascent on some data to cause models to avoid learning is an idea that has been well-explored in the literature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|18", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "For example, the canonical Ganin et al. 2015 paper that proposes \"gradient reversal\" layers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|19", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "More recently, there is the large body of work on machine unlearning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|20", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Although the paper is correct that this is mainly focused on CV and basic regression tasks, there isn't a ton of new technical novelty required to apply the ideas to NLP.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|21", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "I find it weird to discuss de-duplication as a sort of baseline.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|22", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Deduplication should be almost always used as a first data processing stage to mitigate memorization (Kandpal et al. 2022; Carlini et al. 2022), and in that sense, it's quite orthogonal to the method proposed here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|23", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The quality and clarity is otherwise great, easy to read and follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#1|26", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "I find the method to be quite similar to past ideas and the privacy evaluation to be a bit limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|9", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The paper is very well written and follows easily, providing enough background to the reader wherever needed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|10", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The experimental setup is comprehensive with LM performance compared to a variety of tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|11", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "And benchmarks the results for future work to improve upon.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|12", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Many of the experimental decisions taken in this work were interesting: for example, using OPT method to replicate a deduplicated dataset, and using data from the Extraction dataset to measure unlearning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|13", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "While the LMs scale to larger sizes, it takes fewer epochs for the target sequences to be forgotten.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|14", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Sequential Unlearning is more Stable than Batch Unlearning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|15", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Do you have any intuition as to why this happens?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|17", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Thresholds: The thresholds should be set as mean + k* standard deviation and not just the mean.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|18", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "This current approach does not guarantee that with a high probability the examples are forgotten.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|19", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "How do the results change after making this modification?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|20", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "EL_n value: How was the value of n set to 10", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|21", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The Ablation/Understanding section of this paper is weak.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|22", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "In particular, the authors do ask all the right questions: \"Why Are Some Instances Harder to Forget?\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|23", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Why Are Some Instances Harder to Forget?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|24", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "but the study only provides empirical evidence to rehash the same phenomenon without much understanding gained for the reader.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|25", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The experimental study rather answers the question: \"Are Some Instances Harder to Forget?\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|28", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The authors use the term \"guarantees privacy\" for model extraction attacks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|29", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "I would not use terminology like guarantee for empirical evidence.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|31", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "While there may not be a lot of novelty in the methods designed in this work, I do believe that it has a significant contribution to the field of machine unlearning in large LMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#2|32", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The quality of writing is good, and the authors provide code for reproducibility.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|8", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The method is simple and intuitive and has nearly no overhead.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|9", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The problem is timely and relevant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|11", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "There are some experimental and design decisions that seem arbitrary and I don't really see the reasoning behind: a.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|12", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The choice of baseline: I don't really see why deduplication is chosen as a baseline to compare to?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|13", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "It seems the most irrelevant and also the most compute-intensive one.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|14", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The proposed method is not at all similar to deduplication, which is a pre-processing method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|15", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Also, deduplication itself is really not a valid 'privacy protection method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|16", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "I think the most appropriate baseline would actually be doing differential privacy, like DP-SGD, and then doing a comparison of memorization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|17", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "So for instance if you wanna delete sample a', you'd train a model on the full dataset, then unlearn it using your method, and then measure a's memorization before and after.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|18", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Then, you'd compare this with memorization of 'a' but under a model trained with DPSGD.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|19", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "DPSGD is the best point of comparison as by design, it is supposed to protect the membership, so it is like a pro-active approximate deletion method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|24", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Figure 1 is actually inaccurate in terms of DP, as you do not need to re-train with DP every time you get a deletion request, given how DP is actually defined.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|25", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "[1] Shokri, Reza, et al. \"Membership inference attacks against machine learning models.\" 2017 IEEE symposium on security and privacy (SP). IEEE, 2017.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|26", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "[2] Mireshghallah, Fatemehsadat, et al. \"Quantifying privacy risks of masked language models using membership inference attacks.\" arXiv preprint arXiv:2203.03929 (2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|27", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "[3] Carlini, Nicholas, et al. \"The secret sharer: Evaluating and testing unintended memorization in neural networks.\" 28th USENIX Security Symposium (USENIX Security 19). 2019.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|28", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "[4] Izzo, Zachary, et al. \"Approximate Data Deletion from Machine Learning Models.\" arXiv preprint arXiv:2002.10077 (2020).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|29", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "[5] Neel, Seth, Aaron Roth, and Saeed Sharifi-Malvajerdi. \"Descent-to-delete: Gradient-based methods for machine unlearning.\" Algorithmic Learning Theory. PMLR, 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|31", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The paper is well-written and well-structured.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|32", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The proposed method seems novel to me and the paper seems reproducible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|33", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "However it is lacking justification for the choice of baselines and metrics.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|35", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The paper's idea is interesting, simple, and effective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|36", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "However, I think a better baseline and better metrics are needed to be able to actually say that it outperforms other methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#3|37", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "I think a comparison with DP is needed, and also a comparison with other approximate deletion methods is also needed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|12", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Unlearning for large LMs is certainly a timely problem since large LMs have exploded in popularity, and are used as the foundation for almost all downstream NLP tasks, exacerbating any privacy concerns inherent in the pretrained LM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|13", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Thus, this work tackles an important problem of helping to mitigate some of the privacy issues that are likely to arise through the widespread use of large pretrained LMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|14", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The proposed approach is very simple yet effective and extremely efficient in comparison to retraining from scratch, which is generally intractable for large pretrained LMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|15", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "This makes the proposed approach very practical and likely to be used in the real world; it is also a much more appealing option than data preprocessing and expensive differential privacy methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|16", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Experiments using various LM model sizes, and evaluations on 9 different domain datasets provide decent evidence the proposed approach is a potentially viable unlearning method that warrants further study.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|18", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The forgetting guarantee requires extra validation data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|19", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "No results that vary the size of n, the hyperparameter used for the EL metric.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|20", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "This can give better insight into how much of the target examples are unlearned when the prompt varies in size.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|21", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Additionally, how much does the success of the EL metric vary depending on which n tokens are used as a prompt for this metric?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|22", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Only a small number of examples (32) are randomly selected to be unlearned.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|23", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Have the authors tried unlearning much larger portions of the training data and observing the effect on the resulting model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|24", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "It is still not quite clear to me why unlearning a larger batch of examples (128 vs 32) performs significantly worse than sequentially unlearning smaller batches of 32 examples at a time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|26", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "How is the \"hardness\" of forgetting defined.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|27", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Is it how much predictive performance changes when unlearning an example?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|28", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Or, is it defined as the number of epochs to successfully achieve forgetting for the given input?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|29", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The improved forgetting and utility performance of the larger GPT-NEO model compared to the smaller versions is interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|30", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "I would think these results suggest unlearning is \"easier\" for larger LMs with potentially more representational power.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|31", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Do the authors have any insight into the relation between a model that uses the proposed unlearning mechanism and a model retrained from scratch without the target training examples to be removed?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|33", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Table 3 is not color-blind friendly, consider showing prefixes as underlined text instead of blue-colored text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|34", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Footnote 6 (in the conclusion) should come after the period.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|36", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "The work appears novel, the writing is clear and is easy to read, and the approach is simple and straightforward enough that reproducing the results is likely feasible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|43", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "A more thorough investigation analyzing the predictive performance of the LLM as more instances are unlearned would greatly improve this paper and ultimately help readers and practitioners better understand the limitations of this approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc72|review#4|44", "source": "rc", "split": "test", "paper_id": "rc72", "comment": "Also, better understanding the difference between sequential and batch unlearning would also significantly improve this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|7", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I think this paper investigates a very interesting problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|10", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "However, I think the concerns that some readers might raise and complain include:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|12", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Is the top-3 accuracy meaningful for the task, especially for the spatial and cardinal problems where the concept space is very small, and for GPT-3 that knows to output in-domain words only?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|13", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Is the substring metric suitable for the color problem, especially in the “unseen concept” setup?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|14", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "For example, if “light blue” is a seen concept and “dark blue” is the test-time unseen concept, then answering the seen concept “light blue” in the unseen concept setup would result in a perfect accuracy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|15", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Would that defeat the purpose of testing generalization to unseen concepts, like in Table 2?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|18", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "However, is it possible that the model doesn’t understand the relationship between the concepts and the grounded representations, but instead utilizes a similarity between the test grounded representation and the grounded representations in the in-context prompts?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|19", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "For example, upon seeing the test representation (e.g., [0,1,0,0] in the spatial domain, or RGB (140, 0, 255) in the color domain), the model can use a simple strategy: copying the concept of a bunch of most similar representations in the in-line prompt examples (e.g., [0, 1, 0, 0, 0], or RGB (145, 0, 255)).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|20", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This strategy would not involve the concept of “left” or “pink”, and is robust to the rotation transformation (while not robust to the random transformation, if each point in the world was transformed independently).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|21", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This would align with the results in Table 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|22", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "To check whether this hypothesis is (partially) true, we can look at experiments like:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|23", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "A: Test on some real unseen concepts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|24", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This was done in the paper, like in spatial and cardinal columns in Table 2, Table 9, 10, 11 (in the appendix).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|25", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "But the performance is not very strong in these cases even for GPT-3 (top-1 accuracy).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|26", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "B: Test with fewer prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|27", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This is to prevent the model from memorizing similarity with the prompting examples too much.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|28", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This was also done in the paper (Figure 6 in the appendix).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|29", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Again, the performance is not strong, if the number of prompts goes below 20 or 60.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|30", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "C: Replace all of the concept names with concepts in an unrelated domain (e.g., substituting all “left” and “right” with “apple” and “orange”).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|31", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "If the performance is above baseline in this setup, should we conclude that the LM implicitly learns to map fruit concepts to the grounded spatial world?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|32", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This was not done in the paper and may be a good control experiment.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|33", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "(3) Though the paper investigates an interesting problem, the overall takeaway of this work is not very clear to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|34", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "How is the analysis useful for future work?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|35", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "(4) Some details in the paper should be checked.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|36", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "For example, in Section 2.1, the authors say that all models are pretrained on the 40GB OPENAI-WT dataset, but this is not true for GPT-3?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|37", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Also for the color experiments, it is not clear whether 60 or 6+57 or 70 (as mentioned in B.1.1 in the appendix) prompts were used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|39", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Overall I think this work investigates an interesting problem, but the main argument needs to be justified more carefully (as mentioned in (1), (2) above).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#1|40", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Also, the takeaway and impact of the work are not very clear to me, other than showing the somewhat inscrutable power of GPT-3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|4", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I think the modelling work presented is useful and interesting — and importantly: explained well and coherently.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|5", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "However, I find many statements not thought out as much as they deserve.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|6", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "For one, colours are as relative as left/right.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|7", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Something red looks dramatically different under white light than it does under other colours of light, and so on.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|8", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Also see visual illusions where the same shade of grey looks like different colours as a function of where it is in the image.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|9", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This is why categorical cognition plays a role in colours in humans.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|10", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The authors could at minimum note this, if they wish.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|11", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Another example is on the equivocation of \"text\" and \"language\" throughout the manuscript.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|12", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I also find it a bit strange to say \"north\" in humans is not grounded.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|13", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I assume it is always grounded to where the Sun is and what hemisphere we are in and so if somebody were to lie to us about where north is, we would eventually figure it out when we saw the movement of shadows, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|14", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "There is a lot of research on what humans indeed do here, so some literature review could be done if claims need to be made about actual human behaviour.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|15", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Overall, however, I find the modelling here very interesting and compelling.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|16", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I like that the authors explored a pre-trained model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|17", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Perhaps providing a coherent definition of \"grounded\" versus \"ungrounded\" might be useful to help formalise the distinction as used by the authors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#2|19", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The work is good quality and well-explained, but could benefit from slightly more details on the rationale behind some jargon and claims about human cognition.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#3|6", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Overall I really enjoyed this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#3|7", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The scope was fairly narrow and the experiments were not diverse enough to provide a sense of a thorough understanding of how LM's are forming this space, what possible confounds might be at play, or what other domains might have similar effects.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#3|9", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The topic is also very timely and important, and the implications of this work could help unify work in NLP, vision, and other grounded/situated learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#3|14", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "One thing I am considering is whether it's possible that huge parameter language models may be inducing mechanisms for performing spatial rotations on subspaces, though at some point it might become difficult to separate a model's ability to memorize the data and perform on-the-fly rotations, from something that might be called a non-trivial understanding of these concepts/spaces.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#3|15", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "That is mostly an aside as I can't present it as anything more than speculation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#3|16", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "One of the few questions about the color experiment results was to what extent the model was backing off to predict less specific color names, since it is something that is not well-captured in top-1 accuracy (though maybe some combination of this and distance evaluation starts to get at it).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#3|17", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "And why were 67 other colors inserted in the color prompts?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#3|19", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "How are grids incorporated into the model?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#3|20", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "In Appendix B2, it appears to be just as a string following \"World\", but is it a \"linearized\" string of one row after another?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#3|21", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Typos: Zellers reference \"within the prompt .\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#3|27", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The paper is well-written and references are good, although I may have missed an important reference or two since this is slightly outside of my research area, but barring the existence of similar uncited work, I think it's a paper worth publishing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|9", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The rotated baseline is also a great idea, and frankly I find the results on rotated spaces quite surprising, especially when it comes to color.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|10", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I would have expected a serious performance decrease.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|16", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I think there are some subtle but pretty important problems with the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|17", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "To fix them, it might be okay to just change the framing and back off from the \"isomorphism\" angle, but that feels like losing the whole point of the paper and leaving it kind of uninteresting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|18", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "On the other hand, I think maybe adding some tougher experiments and better baselines could patch up the holes and make the paper really nice.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|21", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "is mentioned as a data source, but I think deeper comparisons and more credit to that work should be provided.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|22", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This paper presents itself as providing clear evidence of learning \"isomorphic\" conceptual spaces from text alone, I think the Abdou et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|23", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "paper makes a case that is very similar (though meaningfully different, yes) and stronger than this one.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|24", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "At the very least, these results should be put in context of theirs (especially as their data is being used).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|25", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "In particular the Abdou et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|26", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "paper already provides pretty good evidence of conceptual grounding of colors using representational similarity analysis and linear mappings on color representations extracted from the LMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|27", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "These methods have some underlying assumptions but it seems to me that those assumptions are much weaker than the ones underlying the method used in this paper, which relies on the few-shot learning capabilities of the language model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|28", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "More on that in the next point.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|29", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I don't think the R-IV baseline is more fair than the R-ID one, as claimed in Sec. 2.5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|30", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This paper is about producing the correct term for the provided grounding, which is a different problem from producing an in-domain term.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|31", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Generating an in-domain term is a priori significantly easier (e.g., by hard-coding it).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|32", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "By including the requirement to generate an in-domain term, these experiments conflate the two problems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|33", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "So GPT-2, for example, does badly because in the few-shot format, it often can't decide on any answer with high enough confidence to make it to the top of the beam (see in Figure 3 how GPT-2 skips giving an answer at all and just moves on to the next prompt).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|34", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This is a classic example of \"boring\" sequence generation that happens for sequence models over high entropy output distributions (the literature on this focuses generally focuses on beam search, but I think the argument can reasonably apply in cases like this; see Holtzman et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|35", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "2019, Stahlberg and Byrne 2019, Eikema and Aziz 2020), which is an orthogonal issue to the grounding problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|36", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The issue here is a combination of that and the mere fact that the smaller GPT-2 models don't make sense of few-shot prompts in the effective way that GPT-3 does — again, arguably a separate issue from grounding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|37", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "In short, by relying on the few-shot learning paradigm to extract the conceptual structure of concern, we're limited to extracting conceptual structure that can surface through few-shot learning (just like, for example, linear probes can only extract linearly separable features).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|38", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This ends up being more a measure of the few-shot learning ability of the model than the conceptual grounding capabilities of the model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|39", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This is particularly problematic because it makes the baselines (ie, GPT-2 models) seem weaker than they probably should be, and it limits the baselines that can be used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|40", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I would point particularly to Abdou et al., who obtained alignments similar in quality to those from masked LMs (BERT, RoBERTa, ELECTRA) using just PMI statistics and fastText embeddings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|41", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "It's not clear to me if the large models are exhibiting conceptual generalization/grounding abilities much larger than those afforded by such baselines (once we move out of the few-shot learning format).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|42", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "To add to this issue, the headline experiment of generalizing to new concepts has at least one big issue that I find concerning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|43", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Including all n-1 concepts besides the target one, or many such concepts (eg, 60 colors) provides a very strong constraint in the input (in my view contradicting the \"only a small number of examples\" claim in the paper's abstract).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|44", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "While, yes, generalizing to new concepts does demonstrate that the model understands something about the domain, I think it's important that we understand this in graded terms.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|45", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The question this paper seems designed to address, based on the framing in the introduction and conclusion, is about whether language models can or do construct conceptual spaces that are roughly \"isomorphic\" to grounded ones.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|46", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "In order to test this, of course, we have to construct the isomorphism.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|47", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "But — and this is one of the classic problems with probing — if we are looking at it in terms of set isomorphism (i.e., without extra homomorphism structure — and we are, if we're measuring prediction accuracy), then it is always possible to reconstruct such an isomorphism if we have enough supervision (See Pimentel et al.).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|48", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The question is how much supervision we need versus how much is usefully obtainable from the pretrained model, hence approaches like Voita and Titov (2020)'s information-theoretic probing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|49", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "So it's not really a black and white question whether there is an \"isomorphic\" conceptual space inside the model, but a graded one.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|50", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "And when we talk about things in terms of an isomorphism, what we're really implying is that the vast majority of the information required to construct the isomorphism is easily accessible from the model representations, so less supervision is required.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|51", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This is the basic thinking behind the idea of a \"minimal grounding toehold\" which has been discussed as a prerequisite to grounding language models in conceptual spaces.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|52", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Some of this discussion in the NLP community subsequent to Bender and Koller's 2020 paper was catalogued in a blog post that goes into a lot more detail on the issue.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|53", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "That post proposed a color experiment extremely close to what is done in this paper, but an important difference in their proposal is that the input used by the LM to reconstruct the grounded lexicon is the minimal size that is required to span the conceptual space (i.e., constraining only the necessary degrees of freedom).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|54", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "One might use light/dark, red/green, and blue/yellow as the three basis dimensions to align with CIELAB, or even just to try red, green, and blue.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|55", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I think such a minimal experiment would be required to effectively address the isomorphism question.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|56", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This paper, by providing many more inputs, is doing something of a mix of probing (because of the few-shot/supervision aspect) and behavioral testing (because it's relying on inherent generalization behavior in the LM distribution).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|58", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The inclusion of a bunch of distinct concepts in the input can potentially also provide the model with a pragmatic cue that it will be presented with a new/unseen concept label if it receives a new/unseen kind of input, which adds another confound.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|59", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "This may not be much of an issue in the case of color since the space is fairly open and only 1/6 of the set of colors was covered anyway, but it still means a totally ungrounded baseline could do better than random.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|60", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "To sum up: to really address the question of interest in this paper, I think the unseen-concept experiments need to more strongly limit their inputs to a very small number (or even better, report curves showing performance dependent on number of inputs),", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|61", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "LM experiments which constrain the output space should be included as well, and predictive baselines outside of the few-shot paradigm should be provided.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|62", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The results of Abdou et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|63", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "seem to indicate (to me) that it should be possible to do quite a bit better than the GPT-2 results in this paper using some kind of simple model (for example with word embeddings or a probing model on LM representations).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|64", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "It's important to have these baselines so as not to misrepresent the performance trend with respect to model growth, which I suspect reflects more the ability of the model to do \"few-shot learning\" than to ground concepts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|65", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "(And to be clear: I don't think it would be a strike against this paper if few-shot learning for unseen concepts with very few colors didn't give good performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|66", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "That's still a cool experiment and an interesting result!)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|68", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I understand you would probably want to express the color in RGB for the purposes of LM testing, but Rotations in RGB seems like it isn't the best way of doing the isomorphism control: rotating in a space where distances are more perceptually meaningful (such as CIELAB) seems to me like it would have been a better approach, as while rotation preserves distances, the same distances in different parts of RGB space may correspond to different perceptual distances.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|69", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "On that note, I'm not sure if I totally understand the random world baseline.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|70", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "It sounds from the description in the text (Sec.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|71", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "2.4) that you ground words entirely randomly, but Figure 2 illustrates a \"random rotation\" which is much more constrained than this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|72", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "It suggests that a random rotation should destroy the underlying structure, but I don't see why this would be the case.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|73", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "As long as distances are preserved (under the assumption necessary for the control that distances are meaningful in the first place) the structure should be preserved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|74", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "So any rotation should work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|75", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "In what sense does a random rotation destroy the structure?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|76", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I don't buy it based just on the example shown in Figure 2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|77", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The string matching Top-1 accuracy metric seems quite permissive — are there examples, or is there a human verification step you can provide to show that it is reasonable in this case and not overly permissive?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|78", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "There seems to be some confusion about the training data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|79", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "When you first mention it, you use the term OPENAI-WT with no further description or reference (Sec2.1).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|80", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I don't think that is the dataset's name — I found it used as an abbreviation in one paper, Gehman et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|81", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "2020, but nowhere else.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|82", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The dataset is just called WebText.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|83", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Can you just call it WebText (or OpenAI WebText) and cite the GPT-2 paper directly when you name it (so it's clear where it's from)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|84", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Also, in Section 2.1 it says all models are trained on this data, but this is incorrect — GPT-3 was trained on a different mix of data including two corpora of books, Wikipedia, and Common Crawl.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|85", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Also, Appendix A calls the data \"OpenWT\", which sounds like \"OpenWebText\", which is an open dataset scraped using the same methods as WebText — but isn't the same dataset as was used to train GPT-2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|86", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Then, Section 2.4 does cite the GPT-2 paper for the data, but calls it CommonCrawl, which is incorrect — that's a different web scrape.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|88", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Sec. 2.2, fourth line: stray space after \"prompt\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|89", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "End of page 2: stray comma after \"detail\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|90", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The citation to Abdou et al.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|91", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "can be updated — the work is appearing in CoNLL 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|92", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Last sentence before Sec. 3: stray comma at \"R-IV, a fairer\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|94", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I think this is overall a nice paper with some nice ideas and interesting results, but it falls short of adequately addressing the research question that it is designed to.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|95", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The biggest flaws in my view are the inclusion of too much supervision in the unseen-concept experiments, which make the results less relevant to the \"isomorphism\" question inasmuch as it is interesting to the community, and the exclusive reliance on the few-shot learning paradigm, which adds a big confounding factor to the results that makes it hard to interpret any clear trends.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc73|review#4|96", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "I do think that this is really close to being a very nice paper as long as these issues are adequately addressed at least in the framing of the work if not the experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#1|11", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Their approach is an extension of Electra rather than something fundamentally new.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#1|14", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Related works section could be more elaborate.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#1|15", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "It would also be beneficial if the authors included a brief paragraph or two on how this work can be extended to the multilingual setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#1|16", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "All-Token MLM is just referred to in line 170.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#1|17", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Since this is critical for understanding the paper readers could be eased into it rather than assuming they have read ELECTRA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#1|18", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "In figure 3 the author shows training time in GPU hours.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#1|19", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "It is good to plot against examples seen or number of training steps (assuming all three used the same batch size).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#1|20", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "This would help clarify that there are no inefficiencies in the implementation of Roberta/ELECTRA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#1|21", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Line 249, authors mention that the benefit of stop gradients are for stability.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#1|22", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "It is not clear what they refer to.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#1|23", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Removing stop gradients seems to have a very positive effect on the average scores.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|7", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "In terms of originality, COCO-LM is built on ELECTRA, thus it lacks a novelty aspect.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|8", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "ELECTRA does not have a sequence-level objective, COCO-LM uses a contrasting objective (SCL) for sequences thus it performs better than ELECTRA on sequence level tasks (i.e.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|11", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Having said that, I think the contribution of the paper is quite marginal, however I think its strong point is the efficiency in training time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|13", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "What I find missing is that the paper hasn’t addressed the claim about missing language modelling benefit (line 107-113).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|14", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Although the authors presented word-similarity (table 3, line 292) results, I think that it’s rather a proxy evaluation but is not convincing enough about the benefit of language modeling capabilities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|15", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "As the prompt-based approach [1,2,3] is gaining popularity recently, I wonder if COCO-LM can be used in this setup where the model is required to generate (i.e., assessing language modeling capability).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|16", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "This is something MLM such as RoBERTA, BERT are doing quite well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|17", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "post rebuttal: As authors run additional experiments with prompt tuning and showed good results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|18", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "I increase my score to 7,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|19", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "[1] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pre-trained language models better few-shot learners.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|20", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "[2] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. GPT understands, too.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|21", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "[3] Guanghui Qin and Jason Eisner. Learning how to ask: Querying lms with mixtures of soft prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#2|23", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "The authors have adequately addressed the limitations and potential negative societal impact of their work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|16", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "1. While the paper claim the importance of language modeling capability of pre-trained models, the authors did not conduct experments on generation tasks that are more likely to require a well-performing language model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|17", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Experiments on word similarity and SquAD in section 5.3 cannot really reflect the capability of language modeling.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|18", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "The authors may consider to include tasks like language modeling, machine translation or text sumarization to strenghen this part, as this is one of the main motivations of COCO-LM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|19", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "2. Analysis of SCL in section 5.2 regarding few-shot abaility looks not convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|20", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "The paper claims that a more regularized representation space by SCL may result in better generalization ability in few-shot scenarios.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|21", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "However, results in Figure 7(c) and (d) do not meet our expectation such that COCO-LM achieves much more improvements with less labels and the improvements will gradually disappear with more labels.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|22", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Besides, the authors may check if COCO-LM brings benefits to sentence retrieval tasks with the learned anisotropy text representations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|23", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "3. The comparison with Megatron is a little overrated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|24", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "The performance of Megatron and COCO-LM is close to other approaches, for examples, RoBERTa, ELECTRA, and DeBERTa, which are with similar sizes as COCO-LM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|25", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "If the author claim that COCO-LM is parameter-efficient, the conclusion is also applicable to the above related works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|27", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "1. In experimental setup, why did the authors switch the types of BPE vocabulary, i.e., uncased and cased. Will the change of BPE cause the variance of performance?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|28", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "2. In Table 2, it looks like COCO-LM especially affects the performance on CoLA and RTE hence the final performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|29", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Can the authors provide some explanation on how the proposed pre-training tasks affect the two different GLEU tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|30", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "3. In section 5.1, the authors say that the benefits of the stop gradient operation are more on stability. What stability, the training process? If so, are there any learning curves of COCO-LM with and without stop gradient during pre-training to support this claim?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|31", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "4. In section 5.2, the term “Data Argumentation” seems wrong. Did the authors mean data augmentation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|33", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "1.Check the term “Argumentation” in line 164, 252, and 314.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|34", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "2.Line 283, “a unbalanced task”, should be “an unbalanced task”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|35", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "3.Line 326, “contrast pairs”, should be “contrastive pairs” to be consistent throughout the paper?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#3|37", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "There is no negative societal impact.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|5", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "The pretraining method achieves higher accuracy more efficiently than counterparts across various tasks (GLUE, SQuAD 2.0) and settings (base, base with more data and large).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|19", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "They experiment with three settings, varying model and data size.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|20", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Overall, the results are strong.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|28", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Overall, this paper makes a compelling case for this new pretraining method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|29", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "The results are strong, the ablation study comprehensive, the design choices are well motivated and their usefulness is supported by the analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|30", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Replicating some baselines also shows good scholarship.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|31", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "With general pretraining methods, it is easy to want more and ask for more experiments (e.g: SuperGLUE), as many tasks and setups are relevant to show the generality of the method, but I feel like the authors case is strong enough in this instance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|32", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "The biggest missing piece in the paper is the lack of hyperparameter details, because the appendix is not included.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|33", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "The sections do appear in my pdf’s reader table of content, but not the text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|34", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "There are also some result concerns, which I detail below.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|42", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "The appendix is currently missing and with it many of the hyperparameter details required for reproducibility.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|43", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Please share at least one run / set of results on the test set for each setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|44", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "This is necessary to ensure that the difference is not just in hyperparameter overfitting during fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|45", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "I understand that the claim “matches Megatron-3.9Bn, a model 10x bigger on MNLI” is compelling and makes for a great abstract.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|46", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "However, I think it is misleading and poor - scholarship.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|47", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Indeed, Megatron-3.9Bn is not a great baseline for a 1Bn+ model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|48", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "For instance, DeBERTA 1.5B+SIFT achieves 92.0/92.1 on MNLI (see Table 12 in their appendix, though not clear whether this is dev or test), quite a bit higher than Megatron 3.9Bn (and COCO-LM).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|49", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Minor: Some typos: Throughout the paper (3 times), please replace data argumentation with data augmentation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|50", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "“The analyses uses base setting” -> “The analysis uses the base setting” or “The analyses use the base setting”, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|51", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Minor: You can keep only one significant digit for GLUE/SQuAD results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|52", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Would make tables easier to read and I am not convinced a difference of <0.1 is significant on those tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|53", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Thanks for the response.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|54", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "Looking forward to the improvements to the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc74|review#4|57", "source": "rc", "split": "test", "paper_id": "rc74", "comment": "This is similar to BERT and has the same potential concerns.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#1|14", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "1.Structural locality inherent in datasets need to be captured by a set of custom locality feature functions which requires prior knowledge of the domain of the datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#1|15", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "2. Marginal improvements in results", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#1|16", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "3. No detailed discussion of learned parameters presented in Table 3. a. It appears that w doesn't matter so much for all non l_0 features. So it would be interesting to set w to 1 for these features and learn only b and for l_0 learn w.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#1|17", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "In Eqn 4. y should be w_t?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#2|18", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "While there are clear strengths, one weakness is that one may need to define structural properties in different types of datasets that one might use.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#2|19", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "For instance, it is clear that the model works for source codes and Wikipedia because associated structural information can be mined from the data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#2|20", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "It is unclear how does the method generalise across different tasks and datasets, i.e., beyond two datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#2|21", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "While the authors have addressed these limitations towards the end of the paper, the question is will the work be useful only to a small set of audience, or people from different domains can manually or automatically build such prior knowledge and incorporate it in this model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#2|22", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "The key advantages are clear from the paper, this seems to be the weakness that is hard to defend.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#2|23", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "One possible way to improve the argument so that we could obtain Wikidata-type structure for most datasets is to exploit entity detection and linking including automatically learning their relation (vector) information in a completely unsupervised way.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#2|24", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "The authors must note that I am simply giving ideas on how to strongly defend this weakness of the model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#2|25", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "In terms of experiments, these can be further improved by conducting some downstream application tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#2|26", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "Can the model be useful for document classification tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#2|27", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "Currently, it is very difficult to gauge the usefulness of the model through the limited experiments presented in the paper which mainly revolves around perplexity analysis and table 5 in the appendix has additional token prediction results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#3|9", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "The only weakness that I can identify is that the authors have used one dataset of each domain.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#3|10", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "Since there is minor improvement in the results, it is difficult to gain confidence that the results will indeed improve on other datasets as well", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#4|8", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "It seems that the structural locality and its value are known and not new.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#4|9", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "The paper's contribution then is adding the structural locality to the non-parametric language models.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc75|review#4|11", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "It seems to me that this contribution is not significantly novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#4|12", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "There is no comparison or discussion of other ways to add structural locality to LMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#4|13", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "It seems authors tried just a single approach that they present in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#4|14", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "Using locality info in authors' experiments leads to minor improvement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#4|15", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "This is not a significant weakness since there might be other tasks where locality will contribute more.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc75|review#4|16", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "However, it would have been interesting to see tasks or domains where improvement is more significant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|11", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The authors conduct experiments on multiple NLP tasks, including NLU and NLP tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|14", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Table 3 & 4 compare the proposed methods with the traditional usage (Full).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|15", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Full acts as a baseline method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|16", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "However, in Table 5 (NLG tasks), there's no baseline method for comparison.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|17", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "It will be better if the model can achieve good performnace when \\espilon is less than 5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|18", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Usually, \\espilon in a range of 0.1 ~ 5 can provide meaningful private protection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|19", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Also, the \\delta is large than it should be.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|20", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The value of \\delta on the order of 1/|D_train| are very dangerous according to [2].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|21", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "In that level of \\delta, they permit “preserving privacy” by publishing the complete records of a small number of database participants.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|22", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "(But in the experiments of E2E, the data size is 42000 and the \\delta is 1/50000.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|23", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Unsuitable \\epsilon and \\delta can ensure the utility (model performance) but the privacy protection becomes too weak.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|25", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "There's some hyper-parameters play an important role for model training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|26", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "But, tuning hyper-parameters requires additional information about the private information (accessing validation set or testing set), which leads to private leakage.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|27", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "So, how to count the information leakags of users?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|28", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "How to avoid it?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|29", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "This paper propose a meta-framework that handles several existing fine-tuning algorithms with DP-SGD.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|30", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The performance is amazing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|34", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The performance looks good but there are still some issues.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|35", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The technical novelty is limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|36", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The rebuttal solved most of my concerns.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#1|37", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Thanks for the author's efforts!", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#2|4", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "paper is very clearly written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#2|5", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Although one can say that the paper is mostly empirical and focused on showing that DP finetuning works when the curse of dimensionality for DPSGD is dealt with in one way or the other, I think it adds good value and a proof to the literature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#2|6", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "I do not have any major concerns with the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#2|8", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "I like the paper and think that it will add good value to the literature and the conference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#3|8", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The paper is well-written.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#3|9", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The presentation is clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#3|10", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "It is a great pleasure to read through the manuscript.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#3|11", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The experiments are well designed and executed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#3|12", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The results show improved privacy/utility trade-off over previous work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|7", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The paper is well written and organized.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|8", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The framework and the particular methods were explained clearly and the experiments were well executed, proving the claims in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|9", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The proposed framework could have a lot of impact for NLP practitioners when training DP.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|10", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Large models are notoriously difficult to train with DP from scratch.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|11", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Given there are huge amounts of public texts and pretrained models in the NLP community, fine-tuning with privacy as proposed in this paper could be standard for learning on sensitive data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|14", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Adapters, Compactors and LoRA are also not only parameter-efficient methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|15", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "How well might other sparsification techniques (e.g.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|16", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "lottery tickets [1], diff pruning [2]) compare to the proposed framework?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|17", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Baseline comparison is lacking, although the authors provided the non-private utility baseline, all non-private numbers are learned with full SGD.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|18", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "It would make more sense to provide non-private numbers for the corresponding parameter-efficient methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|19", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The comparison for RGP is also a bit unfair because it was originally proposed for training with privacy from scratch.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|20", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Is it likely that training large models from scratch with DP could be better with the parameter-efficient frameworks as well?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|22", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "[1] The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks. Jonathan Frankle, Michael Carbin", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|23", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "[2] Parameter-Efficient Transfer Learning with Diff Pruning. Demi Guo, Alexander M. Rush, Yoon Kim", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|25", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "Overall, I am leaning towards accepting this paper given the strong results and the hope for learning large models with differential privacy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc76|review#4|26", "source": "rc", "split": "test", "paper_id": "rc76", "comment": "The experiment section could be strengthened with better comparison to baseline methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#1|11", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "Lack of clarity at many places:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#1|12", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "a) It's clearer to demonstrate how you reformulate the input and output for the 3 tasks, as they seem very different compared to previous methods", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc77|review#1|13", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "b) It's better to add some description about the baseline systems", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#1|14", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "c) I'm especially interested in the decoding details on end-to-end entity linking, it's better to use a formular to demonstrate how you compute the log-probabilities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#1|16", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "In the entity disambiguation task, it seems that for each inference step, GENRE only generates one entity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#1|17", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "Is it possible to generate all entities simultaneously as what GENRE does in end-to-end entity linking?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#1|18", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "From table 5, an impressive result is that it seems that the pre-trained model is good at memorizing, even if the identifier is numeric.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#1|19", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "But from table 3, some of the baseline systems also benefit from pretrained models, while achieve much lower results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#1|20", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "What do you think causes this difference?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#2|7", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "One limitation of this paper is that they required the vocabulary of the entities to be the ones that have a Wikipedia page.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc77|review#2|8", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "And that their model relied on copying the surface form of entities (as suggested in the paper).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#2|9", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "From the experiments, the \"copying\" approach worked very well on Wikipedia entities, that are often common entities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#2|10", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "How can you improve the model to link rare entities not in the Wikipedia pages?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#2|12", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "Compared to the models with a large entity memory whose retrieval is performed with Maximum inner product search, how is the efficiency of your decoding strategy?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|13", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "There is no discussion or analysis on the performance of this new model on “tail entities” (entities that have few examples in the training set).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|14", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "A believe such a discussion would be interesting for two reasons:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|15", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "(1) if one wishes to use this type of model for much larger entity vocabularies, it is likely that a larger fraction of entities will have low number of examples, and", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|16", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "(2) one effect of contrastive learning (e.g., negative sampling) has on systems with explicit entity representations is some implicit training of all entities, which is lacking in the described autoregressive proposal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|17", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "Finally on this topic of tail entities: the “IDs” experiment, shown in Table 5, indicates that when entity mention and decode target are not related, performance suffers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|18", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "Similarly, for “tail entities” that are ambiguous with another “popular entity”, the model may be biased with the popular entity (results from “Cold Start” may support this hypothesis).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|19", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "Hard to say without some analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|20", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "While the constrained decoding is really interesting, it was not clear from the paper how crucial this was for good performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|21", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "Is this something absolutely critical for performance overall, or does it provide a modest performance improvement?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|22", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "While it may be clear to the authors, it would be highly informative to explicitly describe the performance impact on unconstrained beam decoding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|28", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "See comment above regarding analysis of “tail entities”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|29", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "Could any further insights on this topic be added to the paper?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|30", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "See comment above about ablating the constrained beam decoding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|31", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "This would not need to be a long or complex analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#3|32", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "Simply 1 or 2 data points for the reader to understand the magnitude of the importance of this decoding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#4|7", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "I have one big concern with the current format of the presentation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#4|8", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "In the current draft, It is not very clear whether the strong gain is coming from large scale pretr aining or the proposed method itself.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#4|9", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "To this end, the ablations shown in Table 7 and Table 8 should be reported in the main paper, with clear explanations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#4|10", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "As we all know, the model architecture (which is the focus of this paper) cannot be properly evaluated when the training set up is different (i.e., how much pretraining has been done, on what dataset?).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#4|11", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "Could you elaborate on this?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#4|12", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "In appendix, there's only result tables without in-line explanations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#4|14", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "If space allows, adding some more analysis on what types of entities do this model do better compared to other methods would be interesting (would lengthier names easier or harder?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#4|15", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "would it do better on popular entities or more long tail ones?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc77|review#4|16", "source": "rc", "split": "test", "paper_id": "rc77", "comment": "are these systems complementary to existing methods or mostly succeed and fail on the same set of examples?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|1", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "This paper attempts to find interpretable prompts for the few-shot tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|2", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "An iterative local search algorithm is proposed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|3", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The algorithm exploits the ability of LLMs to provide interpretable prompt.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|4", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The algorithm is evaluated on some simple datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|7", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "According to Table 4, the prompts found by iPrompt are interpretable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|8", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "According to Table 3, the prompts found by iPrompt have higher generalization ability than baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|10", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "the problem is not well-defined.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|11", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "In particular, how to assess the interpretability of prompts?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|12", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The metrics proposed in this paper only consider \"whether the generated explanation contains one of a set of problem-specific keywords\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|13", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "I would expect a more reasonable metric.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|14", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Whether the algorithm is reasonable is not explained.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|15", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "(1) The iteration does not always converge.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|16", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "(2) Why does the algorithm need multiple iterations?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|17", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "I do not understand the necessity of this iteration.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|18", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "(2) How to ensure that the model generates explanations for the relationship between input and output in the proposal step, e.g. in Fig 3, I don't understand why the LLMs generate \"combine the numbers\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|19", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "I have doubts about some experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|20", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "I think some baselines are designed incorrectly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|21", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "For example, (1) why is the accuracy of GPT-3 only 28.4 for the simple Inverse synthetic math task and 0.6% for the SST-2 task?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|22", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "I think the authors deliberately chose the setting that favors their method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|23", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "(2) AutoPrompt apparently does not have the ability of generalization across models, so it is not suitable as a baseline for Table 3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|24", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "(4) The task involved in the experiment is very simple.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|25", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "I think this means that the method proposed in this paper is not suitable for solving complex tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|26", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|27", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Please refer to Strength And Weaknesses", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|29", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Overall I feel that the problem is not clearly defined.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#1|30", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "This further led to confusion in the experimental design and inappropriate selection of baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|1", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "This paper proposes a method named iPrompt to automatically generate high-quality prompts using LLM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|2", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "iPrompt iteratively mutates top prompts by (1) generating explanations with an LLM (2) reranking them based on their performance and (3) exploring new candidates.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|3", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The authors verify the effectiveness of the proposed method on Math, Sentiment, NLU, and fMRI datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|6", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The paper is good writing and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|7", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The proposed method can search readable prompts without access to LLM parameters and gradients.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|9", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The proxy task(evaluation task) is important for the prompt search, however, the authors did not investigate more on the selection of the proxy task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|10", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The current proxy task is to measure the closeness and similarity to the \"ground-truth prompt\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|11", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "However, what is the definition of a \"ground-truth prompt\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|12", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The author did not elaborate more.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|13", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "If it is some existing description of the task label, it may not a good prompt to serve as a \"ground-truth\" prompt.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|14", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Did the author explore other evaluation methods to select good prompts?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|15", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Some results(e.g. Table5) are not convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|16", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "As the PLM zero-shot performance is sensitive to the prompt, the authors should report the average accuracy and std across several prompts when compared with other baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|17", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The Human-written performance of Rotten Tomatoes/SST-2/IMDB seems too low, as I have simply tried several manual prompts on GPT2-xl(1.3B) and achieve much higher accuracy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|18", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "(i.e., the IMDB accuracy of prompt ' The IMDb movie review in negative/positive sentiment is: \"<S1>\"' is 83.4)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|19", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The compared baseline is weak.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|20", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "More recent baselines should be considered, e.g., Making pre-trained language models better few-shot learners(https://arxiv.org/pdf/2012.15723.pdf)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|21", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|22", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Clarity and Quality: The paper is well-written and has a clear flow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|23", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Novelty: The idea that using the PLM model to generate a prompt by itself is not new.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|24", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "However, the proposed design of the prompt evolution algorithm is interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|25", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Reproducibility: Code is provided in the supplementary materails.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|27", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Overall, this paper investigates a practical and interesting direction that generates readable and high-quality prompts by PLM itself.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|28", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The designed algorithm is well-motivated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#2|29", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "However, the experiment part is relatively weak and cannot prove the empirical contribution of this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|1", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "This paper proposes a method called iPrompt to achieve interpretable autoprompting, that iteratively generates possible explanations of the input data, rerank them, and finally generate new explorations via truncating generated candidates and using them as prefixes to generate new candidates.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|2", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The authors show experiments over Synthetic Math and ANLI where the proposed approach outperforms AutoPrompt and Average Suffix decoding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|3", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "On sentiment classification datasets the authors show quite competitive performance of iPrompt compared to human-written prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|6", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The problem of auto generating prompts/explanations given a dataset is important to the research community on prompting/language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|7", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The idea of generating prompt candidates and then reranking/refining them as optimization is somewhat novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|9", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "The experiment results are relatively weak.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|10", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "the main results are only shown for two synthetic tasks, math and ANLI, and are only compared to two relatively weak baselines, AutoPrompt and suffix decoding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|11", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Why there is no human-written baselines?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|12", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "In addition, there are more recent methods that try to find better discrete prompts like:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|13", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "a) Gao et al. Making pre-trained language models better few-shot learners. ACL 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|14", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "b) Deng et al. RLprompt: Optimizing discrete text prompts with reinforcement learning. EMNLP 2022.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|15", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "c) Schick et al. Automatically identifying words that can serve as labels for few-shot text classification. COLING 2020.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|16", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "which should be cited and compared with if possible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|17", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Table 5, on the more realistic task sentiment classification, iPrompt was compared with human-written prompts which shows a relatively better performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|18", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "How are the human-written prompts determined?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|19", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Existing work has shown that instructions/prompts have a large effect on the final task performance, hence for a fair comparison the authors should show best/worst/avg results w.r.t.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|20", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "multiple possible human-written prompts, e.g., from PromptSource (https://github.com/bigscience-workshop/promptsource) where multiple prompts are crowd-sourced for various tasks including sentiment classification.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|21", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "In Table 3, for ANLI, iPrompt is on par with average suffix decoding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|22", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Is there any analysis to explain why it can't achieve a better performance on this task?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|23", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "All the experiment results are shown on relatively easy tasks (where the prompt can be easily defined/written).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|24", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Can the authors show results on more complex NLP tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|25", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "E.g., question-answering, natural language inference (the actual task instead of the synthetic one used in the paper).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|26", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "This could better demonstrate how the proposed method is able to generate more complex prompts and understand \"what each task is actually doing\" rather than just providing label choices.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|27", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Better interpretability of the generated prompts is a major claimed contribution of this paper, but it's not well supported and there is no associated study on this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|28", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Can the authors provide a human eval to show the generated prompts indeed have better interpretability (and how close it is to human-written prompts)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|29", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Based on Table 4, it's really hard to tell how interpretable the iPrompt-generated prompts are, since many of them do not make much sense to humans, e.g., \"Select currency code for a new\" (truncated randomly), \"What is the opposite of 1\" (what is \"1\"?).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|30", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Especially for the more realistic task sentiment classification, the prompts generated by iPrompt do not seem meaningful at all.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|31", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Some experimental details are not clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|32", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "What is $n$  used in experiments for each task?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|33", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Is there an ablation study on $n$ to show how many samples are needed for iPrompt to perform well?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|34", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|35", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Clarity and Quality: Overall the paper is well-written.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|36", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Some details are not clear, e.g., what is $n$ used in experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|37", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Novelty: The idea of generating interpretable prompts is somewhat novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|38", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Reproducibility: code is provided in supplementary materials.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|40", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Overall the problem studies is interesting and the method is somewhat novel, but the experimental results are fairly weak and the interpretability of the generated prompts are not well supported.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc78|review#3|41", "source": "rc", "split": "test", "paper_id": "rc78", "comment": "Thus I recommend rejection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#1|2", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "With empirical investigation, they show such apparently simple data augmentation closely matches the performance of the models trained on the manually annotated datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#1|4", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The proposed technique can be useful for legacy code translation where annotated data is scarce.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#1|5", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The empirical investigations are solid.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#1|6", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The paper is well-written and easy to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#1|7", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The authors did not provide any explanation/intuition/evidence why such a simple data augmentation technique works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#1|9", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The contrast between performance with noise is useful, but again that shows how sensitive this technique is with noise and for some legacy code (e.g., COBOL --- a language authors use to motivate the study) the amount of noise might be high.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#1|12", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "However, the results show the approach suffered for independently developed code (CodeNet), which is the most realistic case.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#1|13", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The approach is also susceptible to noise.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|2", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The authors first show that the document similarity methods can indeed align parallel documents and find that the word movers distance (WMD) is the most effective one.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|3", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Then, the authors show the high tolerance of models trained with noisy datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|4", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Based on the two findings, the authors finally apply the proposed method to a large, non-parallel code dataset, and observe a performance boost of using a noisy parallel dataset compared to randomly paired datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|13", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "My main doubt is the results presented in \"Section 5.3 RQ3: TRANSLATING BETWEEN A WIDER SET OF PROGRAMMING LANGUAGES\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|14", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The authors report CA@5 scores here but in the other place, BLEU/CodeBLEU/EM is used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|15", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "So, what is the intuition that uses different metrics?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|16", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "BLEU can also handle multi-reference evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|17", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "It would be nice if the authors could add these results in the author response.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|18", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The research is somewhat superficial.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|19", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The authors only show the models can tolerate certain noises but do not propose any simple heuristics to alleviate the impact of noises.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|20", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "For example, penalizing the noises (i.e., the targets not belonging to the source) during model training (fine-tuning). [1] might inspire the authors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|21", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "[1] Wu, Lijun, Jinhua Zhu, Di He, Fei Gao, Tao Qin, Jianhuang Lai, and Tie-Yan Liu.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|22", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Machine translation with weakly paired documents.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|23", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 4375-4384. 2019.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#2|25", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Section 3 Proposed Method: curated -> created", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc79|review#3|8", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Although I recognized that the main contribution of this paper is not proposing a sophisticated method to align samples, I thought the paper need more attention to design efficient strategy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#3|9", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The entire alignment algorithm takes O(D*f(D')) where f(D') is the complexity of GetSimilarDocuments().", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#3|10", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Unless any assumption this takes O(DD') meaning that it tends to be infeasible by increasing the data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#3|11", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Also, the algorithm seems a simple greedy method and may be heavily affected by the sampling order.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#3|12", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "According to Table 2 and 3, I observed that the pseudo-match accuracy introduced here does not reasonably reflect the real characteristics needed to be considered in code translation models since the BLEU of several systems compete with each other.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc79|review#3|15", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "As far as I saw the Figure 2 and 3, the accuracy is heavily reflected by the amount data available during training and some normalization should be needed to compare these metrics each other without biases. Figure 3 also show some remarkable characteristics in translating only into C++, but the paper did not mention about this point.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc79|review#3|16", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The motivation of the paper includes \"modernizing legacy applications,\" but the experiments does not reflect this perspective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#3|17", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "For example, the first section repeatedly referred to the COBOL applications, but we maybe not be able to reflect the experiments into this application due to large discrepancy between languages.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#3|19", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The alignment algorithm need to be improved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#3|20", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "This is not critical: this is still acceptable for the first observation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#4|4", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "In its current form, the problem statement and approach seem very trivial.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#4|5", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "First, the data sets seem rather unsophisticated and it is not clear what the pathway is from this method to say mining parallel corpora from github or some large-scale code repository websites.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#4|6", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "In its current form, it seems like this is preliminary work and with some more sophistication (in method + data sets) it is going to be a good future task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#4|7", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "I suggest that this be submitted to an appropriate workshop and a future attempt be made to submit this to ICLR.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#4|10", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The method and data set in this paper in its current form are rather trivial and more sophistication is required.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc79|review#4|11", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The techniques seem highly tuned for mining well-organized data sets from coding-problem websites.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|10", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The paper should definitely write more to motivate each of the re-ranker, currently, the writing of this part is too short, and especially this is one of the salient contributions of the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|14", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "However, the best performance of the models is far behind the state-of-the-art (which is surprising) and I thought the paper did a bad job of explaining this difference in performance and what could be done to fix it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|17", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The proposed approach of the paper is simple.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|18", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "I also agree with the conclusion that instead of racing towards building the largest model, we should carefully explore the few-shot capabilities of mid-sized models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|19", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Apart from the short motivation of the re-ranking methods, the paper is well-written and readable", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|21", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "My main concern about the paper is its very limited technical novelty.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|22", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "It uses Google search for retrieval and a set of static prompts (per dataset) to query the LLMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|23", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "It proposes three re-ranking methods (without much motivation).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|24", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Using Google search for retrieval and static prompting has been explored before.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|25", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "There also is a lot of work that explores re-ranking.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|26", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Although the paper claims that obtaining probabilities/scores from LMs have not been explored, there exist works such as Izacard and Grave 2021, ART (Sachan et al 2022) that use scores from LMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|27", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The paper also does not explore the easy-to-try extension of prompting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|28", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "For example, retrieving similar questions w.r.t a given question.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|29", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "This is easy to try and might work better and I would argue is a more practical prompting strategy than the current strategy of designing dataset-specific prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|31", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "One interesting result was Gopher 7B + retrieval was better than Gopher 280B without retrieval for NQ and HotpotQA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|32", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "However, this result does not transfer to reasoning benchmarks, so I am not sure if this result is widely applicable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|33", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Can you explain the product-of-experts and noisy-channel re-ranking in more detail?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|34", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "It would be also good to know the motivation behind applying this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|35", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Even though open-book models perform better than closed-book models the overall results when compared to SoTA performance is quite lacking.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|36", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "For example (55.9 v/s 38.4 in NQ, 65.2 v/s 30.3 in HotpotQA).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|38", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "I think the paper would benefit a lot if a section is added which discusses how can this gap be possibly closed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|39", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Table 1 shows the retrieval performance of Google search and the text in section 5.1 compares that to DPR and MDR.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|40", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Can you also report the retrieval performance wrt more recent (and better) retrievers such as contriever, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|41", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "[Minor] The abbreviation LSLM is first introduced in Section 3 and used in a few places later, but it is unclear what it stands for.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|42", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Large Semiparametric Language Models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|43", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "[Minor] The second paragraph of Sec A.4 seems to be out of place.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|45", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Clarity: I think the paper is easy to understand with the exception of an important section (3.3) where the re-ranking is discussed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|46", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "I think the paper could also benefit from a discussion on how to close the performance gap.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|47", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Quality: The paper does a good job in the experimental setup to show us that open-book models is better than closed-book models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|48", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Although useful, this is well-known in the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|49", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "I still think this is a high quality work with very limited novelty", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|50", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Novelty: The novelty of the paper is very lacking.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|51", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Using google search, static prompting are things that has been explored.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|52", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "I also think the paper could have tried other dynamic prompting approaches.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|53", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The re-ranking section was not motivated well and the re-rankers itself is not novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|54", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Reproducibility: Unfortunately, very few institutions have the resources to reproduce results on 280B LLMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#1|55", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "However, I appreciate the experiments on the open-source OPT-6B models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#2|8", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The paper is well-written and easy to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#2|9", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Plenty of empirical discussion and study.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#2|11", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The motivation of this work is unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#2|12", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "It reads like the most significant contribution is using Google search as the information retriever compared with previous work using other search engines on Wikipedia, etc.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc80|review#2|13", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The technical contribution is incremental, without significant novelty.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#2|14", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The 3 proposed methods can all be summarized as hyperparameter/model tuning on the retrieval-based LMs, i.e., tuning the search engine, input prompting, and ensemble strategy.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc80|review#2|15", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Although the authors have a lot of discussions on their empirical results, they don't finally deliver impressive insights or conclusions, besides using Google search and prompting performs better.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#2|17", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The paper has good clarity and easy to understand, but the motivation is unclear and the novelty is limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|6", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "It proposes 4 methods to rerank the generated answers through LM produced probability without fine-tuning.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|7", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The proposed methods achieve effective results on 4 downstream datasets.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|8", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "This paper also provides analysis on the model scale and inference time, and the ability to answer questions grounded in different dates.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|10", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The method is only tested on the Gopher model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|11", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "It would make this paper much stronger if it's also tested on other large-scale LMs such as OPT and GPT-3.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|12", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "More experiments should be done: How does the number of shots k affect the model performance?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|16", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Overall, the novelty and quality of this paper are good.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|17", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "But the clarity about different reranking methods is a concern.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|18", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "This paper needs to elaborate more on each reranking method and the difference between them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|19", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Currently there's just one-sentence description for each method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|20", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Especially for Product-of-Experts (PoE), which is the most effective one.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|21", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "How to combine those probabilities is not clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|22", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Moreover, the reproducibility is limited since it's only tested on the Gopher model, which seems not publicly available either through API or model weights.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc80|review#3|25", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Thus my recommendation is marginally above the acceptance threshold.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|8", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "This paper is well written and easy to follow the major idea and content.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|9", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The experiment results are solid, showing consistent improvements comparing to the huge closed book model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|10", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The proposed methods are simple but valuable to be explored, particularly presenting a new approach for the few-shot “in-context” learning in retrieval augmented language model under the autoregressive transformer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|12", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Only google search results are used for retrieval.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|13", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "It is not clear whether the improvement come from the high quality of google’s retrieved contents or from the retrieval-augmented method.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|14", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Other collection sources for retrieval such as Wikipedia or the training corpus of Gopher 280B also need to be used for evaluation, in addition to Google’s results on internet.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|15", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "In other words, the authors further need to whether the improvements over CB (in Table mainly come from by keeping or perfectly memorizing the corpus used for pretraining CB, or whether it comes from the use of new but well-matched corpus that is NOT seen when pretraining CB.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|16", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The proposed method is applied to the autoregressive transformer using the “in-context” learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|17", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "But, it is not clear how the method is extended to the encoder-decoder model like T5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|19", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "KILT or other knowledge-intensive datasets need to be further used for evaluation.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|20", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "What is the advantage comparing to other few-shot learning of retrieval-augmented models, such as: Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, Edouard Grave, Few-shot Learning with Retrieval Augmented Language Modelsm 2022:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|23", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The most parts of the paper are clear, and easy-to follow given its well-written style.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|25", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Also, the equation for direction inference needs to be checked: should the max be used instead of the summation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|26", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The proposed method of the few-shot “in-context” learning for internet-augmented model in the GPT3-style autoregressive language models is novel, which has not been explored yet.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|28", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Overall, this paper is well-written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|29", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The proposed idea of few-shot prompting for internet-augmented language models is novel, interesting, and timely proper in the literature of the autoregressive language model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|30", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "Experiment results are solid, showing improvements over the baseline huge Gopher 280B, and Gopher’s performance can be achieved by much smaller model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|31", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "The concluded direction of reconsidering towards huge pretrained models reasonably run from the experiments in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|33", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "In the right of Figure 3, CB 280B is only plotted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|34", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "But, it would be better or required to plot the results of other smaller CB models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|35", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "In Table 1, the results may rely on the performance of the retrieval.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc80|review#4|36", "source": "rc", "split": "test", "paper_id": "rc80", "comment": "In addition to using other resources like Wikipedia, can we control the retrieval performance by randomly putting noise irrelevant passages and check the degree of the performance from which the final QA performance start to decrease?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|7", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The problem of CLK itself is an important task and more realistic to downstream knowledge-intensive applications.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|8", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "I like the clear separation of the types of knowledge: time-invariant (to keep), outdated (to remove), new (to inject), as well as their collected datasets for reflecting the three types of knowledge update.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|9", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Different types of baseline methods are covered and compared with analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|11", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The formulation of the CKL problem is overly simplified. It basically only considers a single corpus (D_1) for updating the knowledge of previously learned LMs. A general setup should consider a streaming version of D_1 and make it a sequence of sub-corpus (D_1, D_2, ..., D_T) that arrive at different time steps. Also, the associated tasks in UpdatedLAMA and NewLAMA should reflect such a time series --- that is, the streaming version of the current probing tasks. The current formulation described in Section 3.1 only has a single time step. It's more like an offline learning problem with the forgetting constraints but not an (online) continual learning problem. The proposed setup is thus a bit far from the motivation of studying CKL --- maintaining an ever-changing LM.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|12", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The experiments are very limited to the LAMA probing which does not necessarily connect to real downstream applications of these LMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|13", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "It's also hard to justify whether such methods can maintain performance in general NLP tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|14", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The argument about the KILT experiments seems to only focus on testing the retention but not about the new/updated knowledge about D_1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|15", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The analysis of the CKL methods is not deep enough.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|16", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "How do these methods work and why do some outperform others?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|17", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "How do we know if an arbitrary new fact conflict with the existing knowledge or not on the fly? How do you define such conflicts properly? Say you have a sentence in D_0 \"Cristiano Ronaldo plays for XXX in 2010\", and you have another sentence in D_1 \"Cristiano Ronaldo plays for YYY now.\" In the current problem setup, how will such \"conflict\" be defined?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|19", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The presentation and the writing of the paper can be further improved with more illustrative examples and case studies for readers to qualitatively see the problem setup and the differences between these methods.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc81|review#1|23", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Please find more details above.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#2|4", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The paper is well organized and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#2|7", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The authors also conduct comprehensive experiments to verify the rationality of the proposed benchmark under the various settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#2|12", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The paper only performs experiments with an encoder-decoder model (T5). The experimental results will be more convincing if more pre-trained language models (such as GPT) are included. For example, we can explore the ability of different PLMs to avoid catastrophic forgetting and to acquire new knowledge while preserving invariant knowledge.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc81|review#2|13", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Consisting with the traditional setting of CL, the paper also creates a setting for multiple CKL phases.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#2|14", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "However, only a two-phase setting is considered.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#2|15", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "More experiments can be explored, such as five-phase or controlling the differences in the distribution of data at different phases.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#2|16", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The experimental findings in this paper are somewhat trivial.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#2|19", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The proposed benchmark and metric are technically sound.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#2|20", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "However, there are some concerns about experimental settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|4", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "They provided some findings on their continuous LM learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|7", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "They provide a new benchmark and metric to measure the retention of time-invariant knowledge, updated knowledge, and new knowledge.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|8", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Some interesting observations are provided, for example, 1) rehearsal methods do not work well in this setting (even though the reason is quite obvious because some knowledge is updated) and parameter-expansion methods achieve better results, 2) LMs are prone to more forgetting as they go through multiple CKL phases, 3) LMs should be pretrained with just a few epochs on less duplicating data for efficiency.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|10", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "In a real-world scenario, how can one know in advance that the new task is truly \"new\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|11", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "As mentioned on page 5, it is possible that the FUAR score is very large if \"no gain\" and \"preserve knowledge\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|12", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The authors did not show experiments or analysis in such a setting, where the new task has some \"knowledge overlapping\" with the learned tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|13", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "To make this concern more general, I feel the measurement of \"task similarity\" is missing in this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|14", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Can we also see perplexity as another dimension?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|15", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "EM scores cannot know \"how bad\" the prediction distribution is.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|16", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Do you think the conclusion of these LAMA tasks is the same as other NLP downstream tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|17", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Do you think others can easily replicate your numbers?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|18", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Do you run several splitting or seeds for multiple round CL settings?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|19", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Have you tried different sizes of T5 models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|20", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Maybe the GAP between vanilla and CKL methods will be smaller given larger models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|22", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "I have doubt on this sentence \"Moreover, the effectiveness of CKL methods is much reduced in multi-phase CKL, shown by the decrease of the gap between the FUAR of T5-Vanilla and the best performing CKL method in the scenario of two-phase and one phase, which is 0.92 and 0.46, respectively.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|23", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Isn't it only prove that T5-Vanilla can learn better FUAR scores SMALL-P1_x0019_--> SMALL-P2 than SMALL-P1_x0019_+ SMALL-P2 because you have similar FUAR scores for T5-Kadapters?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#3|26", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "If we can make sure the numbers from this paper are reproducible and comparable, then I think it could be a good testbed for future research in this direction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|7", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Strengths: In terms of novelty, the paper extended the definition of Continuous Learning to formulate Continuous Knowledge Learning that has unique challenges compared to the traditional CL.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|8", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "The paper also introduced a novel metric named FUAR to measure the trade-off between knowledge forgetting, update, or acquisition.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|9", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "This is a contribution to the field as this quantitative metric could facilitate direct comparison between models performing CKL tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|11", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Extensive experiments were conducted to benchmark the performance of various CL approaches (regularization, rehearsal, and parameter-expansion methods) on different aspects of the CKL task (retention of time-invariant knowledge, updating old knowledge, and acquiring new knowledge).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|12", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "In the Appendix, the authors also presented various ways to understand the model's learning process, such as the change of predicted outputs during the continued pretraining, as well as the failure analysis based on the type of probes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|13", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "These methods provide more insight into the model's learning process that went beyond the plain performance scores.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|14", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Weaknesses: Using T5, the authors showed that parameter-expansion methods have the most robust performance throughout all of the experimental settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|15", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "However, in the experiments with GPT-2 (a decoder-only model) in the Appendix, GPT2-MixReview (a rehearsal method) performs the best.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|16", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Although the authors mentioned that “We leave more exploration of applying CKL methods on decoder-only models such as GPT-2 architecture as future work”, they should still mention this critical discrepancy in the main body of the paper so that the readers are aware of the context of the findings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|17", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "After all, large language models take various forms and both T5 and GPT-2 are examples of large language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|18", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Questions: The authors demonstrated that parameter-expansion methods have the most robust performance throughout all of the experimental settings with T-5.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|19", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "Since the three methods (regularization, rehearsal, and parameter expansion) are not mutually exclusive, I was wondering if the authors have tried any combination of the approaches.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|20", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "For instance, the rehearsal approach could be combined with the parameter-expansion method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|21", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "I was wondering whether a combined approach would yield even higher performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc81|review#4|25", "source": "rc", "split": "test", "paper_id": "rc81", "comment": "This work is a big contribution to the community and would invite more research into this topic.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|0", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The paper proposes a method to improve the downstream performance of a pretrained Transformer on NLP tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|1", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The core idea is to not only use the output of the last Transformer layer for prediction, but let the model decide how to fuse the information from intermediate layers as well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|2", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "To dynamically decide which intermediate layers to use depending on the input example, the model uses a mechanism conceptually similar to self-attention, which yields a normalized importance score for each layer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|3", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The importance-weighted sum then yields a complementary representation to the last layer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|4", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "Lastly, another network produces a final, integrated representation from the output at the last layer and the complementary representation, which is then used for prediction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|5", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The model is evaluated on the GLUE benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|7", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The paper is well written and the experimental evaluation seems correct.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|8", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The paper has a nice ablation study which shows that the learned importance scores, the complementary representations, and the fusion network are needed to reach the model's full performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|10", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The main weakness is that the proposed extension to the baseline is relatively complex and rather heavy-weight in terms of new parameters (introducing ~25% more parameters compared to the baseline according to Table 2), yet only achieves a very marginal relative improvement of 0.2 percent over the baseline.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|11", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "It seems likely that this improvement could be achieved through much simpler means, e.g., additional self-attention layers on top of the last pretrained layer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|12", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "This is supported by the fact that the paper's analysis of the proposed importance score mechanism doesn't show comprehensible patterns (or the paper doesn't talk about it).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|13", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "I lean slightly towards rejection, because, although the proposed model is reasonable and could have potential, the experiments do not currently demonstrate that potential, and I hence expect it hard for the community to learn from this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|15", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "You cite several papers which demonstrate that different layers of pretrained Transformers encode different information, which is the motivation for your architecture.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|16", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "However, the cited sources use a feature extracting approach, i.e., they don't fine tune the encoder on their target task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|17", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "In the finetuning scenario (yours), can't we expect the model to learn to simply forward the relevant information from intermediate layers to the last layer?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|18", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "In Section 5 you analyze which intermediate layers are used for which task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|19", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "You state that we can see that the model learns to prefer different layers for each task, but don't go into detail.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|20", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "Can you relate the nature of the task to the layers it produces?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|21", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "For example, why does QNLI rely on layers 2-4 to some extent despite being a rather high-level understanding task?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|22", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "In Figure 2, right side you observe that different examples tend to different layers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|23", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "Did you qualitatively inspect the examples and try to find a pattern/link to their preferred intermediate layers?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|25", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "RoBERTa already achieves super-human results on the GLUE benchmark, meaning that it is probably already close to some upper limit.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|26", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "This makes it even more difficult for your model to substantially improve upon it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|27", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "I would consider evaluating it on the SuperGLUE benchmark instead, since there RoBERTa is not so close to human performance yet.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|28", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "Your model is designed to make use of representations at lower layers if a task requires it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|29", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "But all the tasks in your experiments are evaluated on high-level natural language understanding tasks, which typically require representations at higher layers.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|30", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "This makes it more likely that your model does not improve much over the baseline, because the last layer will arguably already contain much of the information needed for the tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|31", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "I think you should consider different tasks in your experiments to increase the chance of significant improvements.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|32", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "In the current state, it is unclear whether your improvements are coming from the specific model you chose or merely from the fact that you added a lot of parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#1|33", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "As a control, you could include an experiment where you simply add several self-attention layers on top of RoBERTa such that you match the number of parameters of your model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#2|0", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "This paper presents a new mechanism, called HIRE, to extract more information from the intermediate layers of pre-trained models, which will be further fused with the last layer of pre-trained models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#2|1", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The main contribution of this work is the newly proposed dynamic feature extractor HIRE and the fusion network.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#2|2", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "Experiments confirmed the effectiveness of the proposed method, and some interesting observations on the importance of different layers for different tasks were given (i.e. Figure 2).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#2|3", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "However, the proposed HIRE and Fusion modules can be viewed as the combination of some widely used deep learning mechanisms (e.g. Bi-GRU, softmax), using Bi-GRE to represent a sequence is a widely adopted choice, so the contribution in modeling is quite limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#2|4", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "Some existing works, like Self-Adaptive Hierarchical Sentence Model https://arxiv.org/pdf/1504.05070.pdf, have already proposed similar ideas to use more information from the intermediate layers of deep models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#2|5", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The paper is well-written and organized.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#2|7", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "All the models are evaluated on the GLUE dataset, experiments on more challenging tasks like QA (e.g. SQuAD 1.1/2.0) should be added.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#2|8", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "It would be interesting to visualize the importance scores of each layer for this challenging task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#2|9", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "As illustrated in Figure 2, different downstream task highly depends on different intermediate layers, the authors are suggested to conduct experiments under the multi-task setting, since the HIRE can adaptively select the intermediate layers, the proposed method should have a big advantage for this setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|1", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "In this paper, the authors propose a fine-tuning scheme for pre-trained language models, called Hidden Representation Extractor (HIRE).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|2", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The proposed method consists of two components.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|3", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "Hidden representation extractor (HIRE): a two-layer bi-GRU model summarizes the hidden representation of the encoder with a weighted-sum scheme.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|4", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "Fusion network: combine the weighed representation with original representation via the bi-GRU model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|5", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The authors apply the proposed approach on top of RoBERTa and evaluated on general benchmark GLUE.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|6", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The experimental results show that the proposed method could give moderate improvements over RoBERTa baseline system.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|7", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The authors' main contribution lies in two components: HIRE and fusion network.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|8", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "However, unfortunately, both components are not novel to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|9", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The idea of HIRE model is very similar to utilizing ELMo in down-stream tasks, where the hidden representations of each layer are weighted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|10", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The central part of the fusion network utilizes a concatenation scheme of [A;B;A+B;A*B], which is also a widely used approach, similar to Chen et al. (2017), Seo et al. (2017), Hu et al. (2019), etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|11", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "Besides the novelty of the proposed components, the experiments are also not well-designed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|12", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "I assume the proposed method could apply to various pre-trained language models (PLMs), as this is an approach that aims to apply on the fine-tuning stage.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|13", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "However, the authors only present the results on top of RoBERTa model, which cannot demonstrate that the proposed model can be generally applied to various PLMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|14", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "I would suggest the authors evaluate their approach in a more general way, such as build on top of various PLMs, or apply on PLM in another language.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|15", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "These would have helped to better demonstrate the effectiveness of the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|16", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "Considering the novelty and generalizability of the proposed method, I recommend rejection for this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|17", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "[1] Chen et al. Enhanced LSTM for Natural Language Inference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|18", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "[2] Seo et al. Bidirectional attention flow for machine comprehension.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|19", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "[3] Hu et al. Read+ verify: Machine reading comprehension with unanswerable questions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|21", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "A possible useful framework for fine-tuning down-stream tasks for pre-trained language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|23", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The design of the components is not novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|24", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The experiments are not well-designed, and the results only show marginal improvements over RoBERTa.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|25", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The generality of the proposed method is not well-studied, as the experiments are only performed on top of RoBERTa.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|27", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "Unlike Transformer-based models, RNN models (such as Bi-GRU) are not computationally efficient.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|28", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "How about your training time compared to the baselines?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|29", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "According to Table 2, using HIRE will increase the total parameter from 355 to 437, which will slow down the inference speed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|30", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "The authors did not cite ELECTRA (Clark et al., 2020) for comparison.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|31", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "After comparing the results, the proposed model is not competitive against ELECTRA, while it would cost more training and inference time, which will weaken the impact of the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|32", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "I strongly recommend the authors also carry out experiments on top of ELECTRA (or ALBERT-xxlarge, etc.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|33", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "to see if your approach generalizes well on various PLMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|35", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "I did not go through every detail of the writing and only list a few of the issues here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|36", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "page 1, section 1: CoNNL-2003 -> CoNLL-2003", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|37", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "page 1, section 1: (Peter et al., 2018b) points out ... -> Peter et al. (2018b) points out ...", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc82|review#3|38", "source": "rc", "split": "test", "paper_id": "rc82", "comment": "page 3, section 2.3: Tansformer -> Transformer; yeild -> yield; elementwise -> element-wise;", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#1|18", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "I believe this paper will be useful for the community. It should get accepted.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#2|13", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "(1) Novelty is limited as MD-based OOD metrics have been widely studied and used, though this paper used it in a different setting (non-classification).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#2|14", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "(2) It is not compared with other common OOD detection strategies, such as threshold based, energy based, contrastive based approaches. (3) It’s not clear how the proposed method generalizes to other tasks and model structures.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#2|16", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "(5) It is unclear how the background data is selected, and how well the background Gaussian is expected to generalize.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#2|18", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "More importantly, what if we know nothing about the OOD? It seems the background Gaussian would fail to work when there is no good definition of background data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#2|23", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "However, there is too much important material in the Appendix. Would suggest to improve the paper structure and format.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#2|26", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The proposed approaches are effective. However, MD-based OOD metrics have been extensively studied and used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#2|29", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Though the experiment settings and results are thoroughly explained, it is relatively challenging to reproduce all of them due to lack of necessary implementation details.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#2|32", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Novelty is a bit limited, lack of comparison with other common OOD approaches, and the generalization of the proposed approach is questionable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|8", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "There are some statements/claims that are either not well developed in the text (leaving to the reader to guess the intention of the authors when saying something) or that seem overeaching/inconsistent.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|9", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Please refer to my comments for detailed examples of these.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|10", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The paper can be hard to follow at times, and there are a lot of different experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|11", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Perhaps the authors would have a perceivingly stronger paper by making it less wide and more deep (i.e., cover less different points but go deeper into the points it does cover).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|14", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "I am not so positive about its reproducibility, since there are no statements about (a plan for open sourcing) code for reproducing the experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|15", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "One point I did not like very much is the fact the authors mention two external baselines at the end of their related work section (the ensemble methods proposed for e.g. machine translation), but do not compare their methods against those baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|18", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "My recommendation is a borderline accept.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|20", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The reson for \"borderline\" is detailed in my comments/questions below.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|21", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The examples of existing works on OOD detection could be greatly improved (2nd paragraph in the introduction).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|23", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "It is also an example in the vision domain, should you not include an example in the (conditional) language modelling domain?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|24", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Page 2, paragraph 1, last sentence (\"In Section 4, we show that while model perplexity is a reasonable choice for [performing selective generation with] in-domain examples, combining with our OOD score works much better when the input distribution is shifted\") -> Please make it clear that you are talking about selective generation, e.g., I added the text in-between brackets to do that as a suggestion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|25", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Page 2, section 2: In the first paragraph, it is not clear what is the (practical) relationship between maximum softmax probability (MSP) and perplexity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|26", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Please better elaborate your point, the paragraph reads like two more-or-less independent points glued together.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|27", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "A method that comes to mind is the mean over the MSP for a sequential output, for instance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|29", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Page 4, 2nd paragraph (\"While we use the ground-truth outputs to fit N (μw, Σw), we do not have ground-truth in general for the background examples (e.g. C4). We decode outputs from the trained CLMs and use those output embeddings to fit the background output Gaussian, N (μwδ , Σwδ ).\") -> This sentence is not clear. Please rephrase.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|30", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Your model requires input/output for the two tasks, i.e., summarisation and translation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|31", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "You have C4 for summarisation, and ParaCrawl for translation. Please elaborate that ParaCrawl does not have translation pairs, but paraphrases (right?), and that C4 does not include summaries (to the best of my knowledge).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|32", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Your descriptions seem incomplete and assume that the reader will make the leap about important points that should be made clear in your manuscript.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|33", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Page 4: \"Since we have implicitly defined two classes (...)\" -> There was no implicit definition, I would say this definition of the two classes was explicitly made.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|34", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "You are using AUROC as your metric, but I would like to see the plots rather than the summary metric (i.e., the actual curves), perhaps in the appendix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|35", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "As baselines, you are using the perplexity score but not the maximum softmax probability MSP.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|36", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Could you not use the mean MSP (over the sequence) as a proxy for OOD? Does it consistently underperform compared to using perplexity?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|37", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Page 6: \"Though RMD and Binary logits OOD scores both perform well at OOD detection, RMD OOD score is better at distinguishing near-OOD from far-OOD.\" -> Why does that matter? That is not what models were trained for.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|38", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "If you had trained models (RMD vs. Binary logits) to distinguish between near-OOD and far-OOD, perhaps the outcome would have been different. This split of near- vs. far-OOD is also a bit arbitrary, if you will make it central in your evaluation/discussion (like you do), perhaps it is a good idea to be more principled about it (e.g., using some objective/concrete measure to make one dataset \"more OOD than\" another).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|39", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "You hint at unigram overlap as a measure of OOD-ness later on, perhaps you should use the \"degree of OOD-ness\" as the percentage of unigram overlap between the two domains, for example.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|40", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Page 6: \"We observed that law has the highest unigram overlap rate (48.8%) and the second highest overall overlap with the in-domain data (Table A.7).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|41", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "This confirms that law is actually not OOD data and explains why no method can detect it.\" -> If your definition of OOD only depends on unigram overlap, you do not need anything other than computing the unigram overlap between domains to perform OOD detection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|42", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "I find this a very misleading comment, since the fact a dataset has high unigram overlap with another does not automatically rule out this dataset being OOD.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|43", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "You should include more nuance in these statements.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|44", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Page 7: \"Amazon Mechanical Turk workers were asked to evaluate summaries generated by the xsum model on a scale of 1-5 (bad-good) using 100 examples from xsum, cnn dailymail, reddit tifu, and samsum. We collected 3 ratings per example and took the median to reduce inter-rater noise\" -> This is a bad strategy to reduce inter-annotator agreement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|45", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "You are basically hiding the (possible) disagreement by only using the median, whereas in order to reduce the disagreement (and not mask it)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|46", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "you should instead 1) select high-quality annotators,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|47", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "2) make the annotation guidelines clear including concrete examples of how to annotate difficult cases,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|48", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "3) include control examples for annotators, etc. Please at least report the variance of the annotators, so the reader knows how noisy they are.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|49", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Page 8, Section 4.4: \"To evaluate that, we propose using the Quality vs Abstention Curve (QA), analogous to accuracy versus rejection curve used for selective prediction in the classification.\" -> Citation missing. You mention Malinin & Gales (2020); Xiao et al. (2020) later, but it is not clear how exactly they relate to QA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|50", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "If you are the one proposing this measure, please make it clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|51", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Page 8, Section 4.4: \"For translation, the OOD score is better than perplexity when abstention rate α > 0.5 and worse than perplexity when α < 0.5.\" -> From Figure 5, the cut-off point looks more like 0.6 or 0.65 rather than 0.5. Is that so?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|52", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Page 9, Section 5, last paragraph: \"In this work, we focus on developing scores that can be readily derived from the generative model itself, without much increase in computation.\" -> You did not compute how much time it takes to run your proposed models compared to the baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|53", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "I find it could be a limitation of your work, especially in light of the existing work using ensembles mentioned in your related work section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc83|review#3|54", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "You also did not include any such ensembles as your baselines. Why?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|2", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Towards this end, authors propose i) a discriminative loss to ensure generated texts produce enough diversity to differentiate between labels and ii) a meta-weighted generative loss to take into account the impact of different tokens towards a label.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|3", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "They also propose an ensembling classifier to add more robustness to deal with generated samples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|4", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Experimental results on various tasks from the GLUE benchmark shows that the proposed method outperforms other methods from the literature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|7", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The paper is enjoyable to read and the algorithm has been developed nicely throughout the narratives.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|8", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The authors clearly describe the problem they are intending to solve followed by motivating and describing the three major components of their algorithm - the discriminative loss, meta-weighted generative loss and the ensembled classifier.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|9", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Authors also provide the end-to-end algorithms which help the reader to understand and possibly implement the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|10", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Experimental results on the GLUE benchmark show clear and non-trivial improvement over existing baselines including the ones that rely on generating additional samples and those which do not.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|11", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Although the final proposed method is somewhat complicated, ablation studies on the importance of each of the component show relative improvement caused by individual components and the importance of all these pieces.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|12", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Overall, I find it to be a strong paper with a non-trivial contribution, clear narrative and strong empirical results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|14", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Qualitative analysis - Given that it is a generative paper, I'd like to see examples of how the generated texts look like for different labels. How different are those from the ones present in the training sample?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|15", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "On the same line, I'd like to see how well the meta-weight generation can learn the importance of various tokens. For example, authors provide an example that the token \"movie\" might not be relevant to generate texts for positive/negative movie reviews. Does such properties automatically emerge from the meta-learning procedure?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|16", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Susceptibility to hyper-parameter tuning - The optimization procedure is performing three interleaved gradient updates at every iteration.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|17", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "I'd imagine that such an optimization procedure will require sophisticated hyper-parameter tuning to obtain stable convergence.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|18", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "I'd like authors to comment on that along with the overall computational cost for this procedure.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|19", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The whole generator-discriminator paradigm is very similar to many works from the literature which are based on GANs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|20", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "I'd like the authors to clarify in the paper how their optimization process varies differently from the GAN.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|23", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Clarity The narrative of the manuscript is clear and pleasure to read - the problem is clearly described and every piece of the overall proposed method is well motivated and explained.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|24", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Quality The proposed method looks sound to me and it intuitively makes sense.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|26", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "As mentioned above, adding more qualitative evaluations will make the contribution stronger.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|27", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Novelty Although individual components might not be sufficiently novel, the overall contribution (i.e. the end-to-end algorithm) is novel and it cleverly combines and modifies existing method to solve the desired problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|28", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "I have no concerns regarding the novelty.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|29", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Reproducibility Authors provide sufficient details and pseudo-code corresponding to their methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|30", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "As mentioned above, some additional details on the stability and ease-of-training of the meta-learning procedure will be helpful (apologies if I have missed it).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#1|34", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "At this point, I am inclined to accept this paper based on my opinion alone.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|1", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The paper proposes a method to augment label-discriminative data in PLM few-shot learning setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|2", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The main idea is to improve the PLM-generator by introducing a label-discriminative loss in a meta-learning framework.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|3", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The improved PLM-generator is then used to generate augmentation data to finetune a classification PLM(Roberta-large).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|4", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The paper compares the proposed methods with several few-shot baselines on seven classification tasks of GLUE.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|7", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The overall framework to investigate the few-shot learning ability of PLM is somehow novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|8", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "It is interesting to design the $L_{disc}$ to encourage the PLM-generator to synthesize more label-discriminative augmentation examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|9", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The paper is well-written with clear logic.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|10", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Most part of the paper is easy to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|12", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The intuition that formulates label-discriminative text generation into a bi-level problem is not clear. What is the interpretation behind it? It seems that the bi-level structure is not necessary in this case, since there is no strict hierarchy of the inner and outer objectives ($L_\\text{gen}$ and $L_{disc}$): the order of inner and outer objectives can be simply swapped to achieve the same goal(i.e. using unweighted generation loss in the outer loop and weighted discriminative loss in the inner loop). The bi-level form here is more like to introduce “technical contribution”.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|13", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The two reason for not directly combining $L_\\text{disc}$ and $L_\\text{gen}$ is not well justified ( \"the generation-irrelevant loss  $L_\\text{disc}$ will unavoidably interfere with the language modeling process” and “A hyperparameter needs to be introduced to balance the weights of the two losses ”.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|14", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "I do not see why the bi-level formulation helps solve these issues.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|15", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Specifically, suppose the token weights are learned to minimize the discriminative loss, how can it guarantee such weights will not affect the language modeling process?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|16", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "In fact, even without direct intervention, the $L_\\text{disc}$ will affect $w$ and therefore affect the optimization of $L_\\text{gen}$.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|17", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Besides, it seems that introducing a hyperparameter to balance  $L_\\text{disc}$ and $L_\\text{gen}$ is acceptable, as the author introduces such a hyperparameter in Eqn. 5 to balance the two loss term for finetuning, which is actually in a similar situation.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|18", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "I remain skeptical about the performance gain of the proposed method. Firstly, the authors do not compare with the important baseline that directly uses few shots in the demonstration(as in-context examples) to generate augmentation data, and then use the labeled few shots and synthetic data to fine-tune a PLM classifier.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|20", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Besides, I am curious about to which extent more label-discriminative data can improve the moderate-size PLM classifier.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|21", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Table2 should include all the glue tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|22", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Complete results of the important baseline(w. $L_\\text{gen}$) can help justify which margin of improvement can the proposed method achieve.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|23", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The paper does not conduct quantitative and qualitative analysis to measure the quality of the generated text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|24", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Can the improved generator really generate more label-discriminative text?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|25", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Some automatic metrics/human evaluation should be introduced to measure the quality of the generated data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|26", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Usually, the bi-level formulations used in previous work present justifications of why such formulation is adopted, such as meta weight net, MAML, etc. However, no theoretical justification or intuitive explanation is given to explain the necessity and superiority of such a formulation in this paper.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|28", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Besides, it is well known that meta-learning methods are notoriously hard to tune and can be extremely sensitive to hyperparameter choices. It is hard to tune the framework on a large PLM with a small batch size.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|30", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The authors only conduct experiments on 2-class or 3-class NLU classification tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|31", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "It is questionable whether the proposed method is suitable for other tasks(e.g., multi-class classification and QA tasks).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|33", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Clarity, Quality, and Novelty: The paper is well-written and easy to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|34", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The intuition that formulates label-discriminative text generation into a meta-learning framework is not clear and a bit sudden.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|35", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Reproducibility: Code is provided in supplementary materials.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|37", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Overall, the paper is well-written with a clear flow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|38", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "However, (1) the intuition that formulates label-discriminative text generation into a bi-level problem is not clear, as there is no strict hierarchy of the two objectives. No convincing intuitive explanation or theoretical justification is given to explain the necessity and superiority of such a meta-framework in this paper.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc84|review#2|39", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "(2) Some important baselines are missing, and I remain skeptical about the quality improvement of generated data and performance gain of the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|2", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The core idea of the paper is to use a pre-trained language model (PLM) to generate texts which can be leveraged towards fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|3", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The authors propose two methods in the paper to achieve this: (i) a method to fine-tune the PLM to generate appropriate texts; (ii) a technique to fine-tune the classifier using the generated texts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|6", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "- The paper shows strong improvements over other baseline few-shot fine-tuning techniques (with / without use of data-augmentation).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|7", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The comparisons with data-augmentation techniques is more fair, as extra data is used during downstream fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|8", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "- The main technical novelty is in formulating the fine-tuning process of the generator with the meta-objective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|9", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Conditioning the generator on a weighted version of the input is a good way to generate label-coherent examples and the authors make the framework to work which is a strength.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|11", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "- Could the authors add the ablation where D_train and D_gen is together used for fine-tuning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|12", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "I could not find it in the paper and it would be beneficial to add this to truly understand the effectiveness of the proposed fine-tuning strategy (with temporal ensembles) for the classifier tuning stage.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|13", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "In Table. (2), I do find -temporal ensemble, however I am assuming it is a two step fine-tuning - first on D_train and then on D_gen without the regularizer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|14", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "- The compared baselines are not comprehensive.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|15", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The authors should add more baselines where data-augmentation is used for few-shot learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|16", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Some of the baseline techniques from https://aclanthology.org/2021.findings-acl.84.pdf can be investigated to strengthen the contribution of the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|17", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "While a comprehensive analysis of all the augmentations is out of scope for the paper, a few strongly performing augmentation techniques can be added as baselines to Table 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|18", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "I would like to see some quantitative analysis on what proportion of label noise is mitigated with the use of meta-objective in training the generator?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|19", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Downstream performance is definitely one indirect way to measure it, but having a quantitative analysis would strengthen the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|21", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The paper is clearly written and coherent to read.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|22", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The reweighting objective which is meta-learned is not novel, however the authors use it for generation task which is novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|24", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Overall, the paper is well-written, has satisfactory novelty and strong results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|25", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "However, I am leaning towards weak reject as I feel that the authors did not compare with enough baselines where augmentation is used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|26", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Given the strong results, I will be willing to increase my scores if the authors can :", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|27", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "(i) provide some motivation on why these 3 augmentation techniques were chosen to compare to or add some more baselines to Table 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#3|28", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "(ii) Some quantitative measurement of the mitigation of label noise on the meta-objective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|2", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Additionally, the generation of novel samples is guided by a meta-objective which uses a weighting mechanism to choose the most important tokens w.r.t a label and produce more discriminative text samples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|3", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The model achieves better results on the GLUE benchmark, which consists of natural language understanding tasks, compared to existing few-shot methods, both with and without data augmentation steps.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|6", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "+The idea of combining an autoregressive language model for data augmentation and another classification language model that performs the task, is simple and novel to some extent.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|7", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "+All used modules and components are reasonable and sound, including the pre-trained models in the framework and the considered datasets/tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|8", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "There is also some additional analysis and ablations included.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|9", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "+The empirical study shows that the proposed method FewGen outperforms other SOTA methods, with and without data augmentation on the GLUE benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|10", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "+The paper has a good flow, is well-written and understandable for the most part.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|12", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "-There is a lack of experiments without the data augmentation done by the autoregressive language model. To have a better understanding of how the data generation part helps, the authors should provide a baseline where the model is fine-tuned without augmentation. This will offer a fair comparison to the models trained w/o augmentation in Table 1, and also can highlight the contribution of the classification-only part of the method.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|13", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "-The paper states that “general idea of meta-learning is to formulate a meta objective to enable automatic learning of hyperparameters”, which is not entirely true, since that can be considered only as a subset of the meta-learning goals.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|14", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "In general, meta-learning is aiming to learn good representations and accrue meta-knowledge by observing batches of few-shot tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|15", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Therefore, it would be good if the authors can elaborate more how their method achieves this, i.e.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|16", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "to explain thoroughly the learning strategy, which is only briefly mentioned to be the online optimization strategy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|17", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "-The paper uses a learnable “w” as a feedforward net with 100-dimension hidden layer, but it is not explained why this is the only considered option.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|18", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "For instance, other options such as self-attention layers, do not seem to be considered.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|19", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Since this part is important for learning the token weights in the meta-weighted objective function, the structure of this part can be further explored, for instance by providing more ablation analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|20", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "-The paper is lacking more qualitative analysis and discussion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|21", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "It would be nice to provide more generated samples for given k-shots and for different tasks, and to discuss how they contribute to a better performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|22", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "-It is not entirely clear how the shots are defined in the downstream tasks. Are they predefined in the scope of the datasets? This part can be more elaborated.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|24", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Clarity and Quality: The paper is written in a clear and comprehensive manner for the most part", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|25", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Novelty: The method presented in this paper is mostly combining existing techniques in a novel manner.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|26", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "For example, they combine techniques from Meta-Weight-Net (which first proposed to learn weights as hyperparameters of the objective) and data augmentation to perform more accurate text classification.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|27", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Reproducibility: The paper offers sufficient information to reproduce the method, and also the code is available in the supplementary material.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|30", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "However, meta-learning comes a bit of a sudden and is not clearly motivated, nor the learning strategy is thoroughly explained.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|31", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "The authors should try to define the link between the proposed techniques, namely, data augmentation through text generation and meta-learning the weights, to have a more consistent storyline.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc84|review#4|32", "source": "rc", "split": "test", "paper_id": "rc84", "comment": "Also, it would also strengthen the paper if the authors consider the points and concerns raised in the “weakness” part.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|1", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The paper constructed framework where a language model is supplied with various types of external context.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|2", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The context is curated and retrieved, then concatenated with the language model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|3", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The retrieval part is frozen and the language model is initialized with LM adapted T5 and trained together with a dispatcher that selects which knowledge source to retrieve from.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|4", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The model shows improvement over non-context augmented, similarly finetuned models such as T0 on a wide range of tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|6", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Strength: Thorough evaluation on a wide range of tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|7", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Providing a broad picture of the effectiveness of the supplied context.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|9", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "1, The paper is weak on ablations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|10", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The author should include more baseline results, in particular the result where the same (initial) model is trained with the same hyperparameters with no non-trivial knowledge sources available.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|11", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "In their language this baseline should have the generalist only.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|12", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "They also did not analyze e.g. how retrieval noise/quality affect the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|13", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Ideally the author should look deeper into how are the retrieval results helping the model to make predictions that are otherwise difficult.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|14", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "2, A unified retrieval system that works across different knowledge sources and end-to-end trained with the language model would be more elegant and removes the expert dispatcher and make the method more freely generalizable to expanded knowledge sources without retraining.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|15", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "But this reviewer acknowledge the value of attempting to use out-of-the-box retrieval systems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|16", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "3, The paper claims to have experimental evidence that instance-adaptive selection is superior to task-adaptive selection of the context type.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|17", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "But this claim did not seem to be backed up by numbers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|18", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "4, The author should comment in the paper on the degradation on tasks such as OpenBookQA and PIQA when going from the no-context finetuned baseline of table 1 to their flagship model in table 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|19", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Some analysis on the win/loss patterns would be valuable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|20", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|21", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The work is original enough and provides good information on training general purpose language models that can utilize external knowledge sources.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|23", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "1, There is no listing of the 39 finetuning tasks. Making it difficult to evaluate the generalization ability of the resulting model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|24", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "2, The author should include a reference on the \"P3 task categorization framework\" and the identity of the \"P3 tasks\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|25", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "3, SwithTransformer -> SwitchTransformer,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|28", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The authors demonstrated incremental value of retrieved context from various datasources.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|29", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "It is light on ablations and analysis and rough on the edges in terms of presenting all necessary information and baselines.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#1|30", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "But overall it is a valuable paper. Improvements on the weaknesses of the paper could lead to better review ratings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|1", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "This paper proposes a novel semi-parametric language model architecture dubbed Knowledge-in-Context (KiC) which utilizes external knowledge memories of K (K=6 in the experiments) different structured knowledge types (dictionary, commonsense, Entity, event, script, causality) to make the prediction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|2", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "When a query is given, the knowledge selector classifies the query into (K + 1) classes, where the 0-th class represents that no external knowledge is required for the query.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|3", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Then, the knowledge memory that corresponds to the class with the highest predicted probability is selected, and the knowledge retrieved from the selected memory is concatenated with the query to serve as the input to the text-to-text model to generate the output.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|4", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "This can be viewed as a mixture-of-experts (MoE) architecture where the parameters of all experts are shared, and each expert has its own external knowledge base.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|5", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The authors show the effectiveness of the suggested method through extensive experiments on zero-shot setup, in-domain setup, and MMLU tasks, comparing the performance with the simple and state-of-the-art models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|6", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The authors also report that KiC seems to show emergent behaviors at the scale of 0.77B.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|9", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The paper proposes a novel semi-parametric method to use the external memory of heterogeneous types of structured knowledge, which demonstrates its effectiveness across many tasks in various setups with a relatively small number of parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|11", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The paper lacks ablation studies to investigate the effectiveness of each component of the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|12", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Several important details of the method are missing, making it less clear and the experimental results difficult to reproduce.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|13", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Suggested Ablation Studies & Analyses", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|14", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "In order to see how important is the knowledge-selecting mechanism, it would be interesting to see the performance of KiC without the knowledge selector where all knowledge memories are merged into one and then retrieved (however, it might be difficult if the retrieval mechanism is different for entity and script knowledge types, which is not clearly written in the paper).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|15", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "It would be interesting to compare the performance with the cases where unstructured knowledge source(s) is used to augment the input, in order to show the effect of using heterogeneous structured knowledge resources.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|16", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "It would be interesting to see the ratio of the knowledge source types selected by the knowledge selector to solve each task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|17", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "It would demonstrate the effectiveness of using the balancing loss and might provide the readers with further insights.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|18", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|19", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Several important details of the method are missing, making it less clear and the experimental results difficult to reproduce.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|20", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "It is not described in the paper what kind of model architecture the knowledge selector uses.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|21", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "For script data, it is not explained how the authors “retrieve the most relevant scenario” given the query (Is MPNet used here as well to encode the input query and each utterance in the script?).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|22", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "It is not described how a word is selected from the dictionary given the query, and how the authors preprocess the Wiktionary to create the dictionary knowledge that would serve as the input, especially when the word contains more than one meaning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|23", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "It is not clear how many knowledge instances are retrieved from the knowledge memory and attached to the query (It seems like one).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|24", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The actual storage footprint of each external knowledge memory is not reported in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|26", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "[Minor point] The motivation behind why the authors explored the use of only structured knowledge resources and did not include unstructured knowledge sources as a type of knowledge memory such as Wikipedia, remains unclear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|27", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "[Minor point] The formula for the balancing loss refers to the paper of SwitchTransformer, and is missing in this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|29", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The paper proposes a novel semi-parametric method to use the external memory of heterogeneous types of structured knowledge, which demonstrates its effectiveness across many tasks in various setups with a relatively small number of parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#2|30", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "However, many important details of the method and ablation studies are missing in the paper, making the need to update the manuscript essential to make the explanations clear and add several ablation studies and analyses which could provide more insights to the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#3|1", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "This paper tackles the problem of zero-shot task generalization to unseen tasks using a semi-parametric approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#3|2", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "In order to achieve this, they construct 6 different knowledge-rich external memory consisting of Dictionary, Commonsense, Entity, Event, Script, and Causality.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#3|3", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Like previous work, they perform multitask prompted fine-tuning on 40+ NLP tasks but while retrieving from external knowledge sources to perform the task.", "reviewer_type": "human", "label": "error", "error_type": "Unstated statement", "label_origin": "expert"}
{"id": "rc|human|rc85|review#3|4", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Since there are six knowledge sources, they train a MoE layer that dynamically routes to which source to retrieve the external knowledge from. This approach coined Knowledge-in-Context (KiC) enables a 770M LM to easily outperform LMs that are 4-39x larger by a significant margin.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#3|9", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The weakness is that it does not show any computational comparison compared to prior multitask prompted finetuning approaches.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#3|10", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Initial thought is that the proposed method might require much more computation (fine-tuning stage) since the MoE layer has to be trained to be able to dynamically select which knowledge source to route to AND also train the underlying LM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#3|11", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#3|12", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The paper is very easy to read will clear explantations in the technical parts such as the explanations regarding the MoE layer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#3|13", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "I think the proposed method is very novel in the sense that they utilized a MoE layer to retrieve from multiple external sources and that they utilized this to tackle the important task of generalizing to unseen tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#3|16", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Thus, I highly recommend this paper be accepted at this conference.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|1", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "This paper introduces Knowledge-in-Context (KiC), a semi-parametric LM consisting of a parametric text-to-text LM with a knowledge-rich external memory and a knowledge selector.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|2", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "During inference, the knowledge selector determines the sequence-to-expert assignment and a retriever retrieves the most relevant sequence from the expert.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|3", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The retrieved sequence is concatenated to the data instance and goes through the text-to-text LM. Results show that KiC outperforms LMs that are 4-39x larger by a significant margin.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|9", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Weakness: This paper lacks detailed ablation and analysis of the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|10", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The performance of retrieving from a mixture of 6 categories (treating a mixture of 6 categories as a single large corpus) is needed to show the effectiveness of MoE architecture for knowledge selection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|11", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "A naive mixture of 6 categories would not need any knowledge selection step.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|12", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Please also report the standard deviation across different evaluation prompts for each task in Table 2. Prompt sensitivity is also important in addition to median accuracy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|13", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "For each task, what is the occurrence of selection of each knowledge category for Table 2? This might answer the question of \"why does KiC works?\", if there are some patterns of the target task and selected expert.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|14", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Is there any ablation result showing the necessity of generalist (no external memory)? How does the performance vary if the generalist is absent?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|16", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Are there any results of KiC on a 3B (or larger) scale?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|17", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Is there any quantitative result of showing the benefit of using structured knowledge resources than plain text such as Wikipedia or C4 dataset?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|18", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "If the effect is similar, using plain text would be simpler because it does not require any knowledge selector.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|19", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|20", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "This paper is clear about the contribution and the proposed method. Although there are many lines of work related to semi-parametric zero-shot task adaptation, the method proposed in this paper is novel enough.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|22", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "This paper proposes KiC, an effective semi-parametric LM that outperforms much larger parametric LMs on various downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|23", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "The paper is well-written and the proposed method is novel enough.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc85|review#4|24", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "However, this paper lacks detailed analysis and ablation. Ablation studies on how each component of the proposed method benefits the target downstream task are needed, especially.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#1|10", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "W1 - I'm quite unsure about the set of prototypes. It seems not described in the main contents.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc86|review#1|11", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "I checked examples of prototypes in the appendix but the 1:1 label distribution (positive vs. negative) seems to me they are human-picked.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#1|12", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "W1-Q1 - If the set of prototypes pre-defined, do we really need to train the prototype?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#1|13", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "what if we just compute sentence similarity between input and the prototype?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#1|14", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "W2 - The main table is not clear, especially on faithfulness.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#1|15", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "What if we just consider the random prototype as a prototype?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#1|16", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "what is the faithfulness of that random prototype then?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#1|17", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Q2 - Just curious, in the introduction, why the post-hoc explanation is potentially reporting-bias?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#1|18", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Isn't this method also potentially reporting-bias since it only uses samples from train data as prototypes?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|6", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "This paper seems motivated by a prior NeurIPS 19 work \"This Looks Like That...\" in the sense that the architecture and loss designs are sourced from there.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|9", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Related works have limited coverage.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|15", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "However a couple confusing points still.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|16", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "There are many lines of explanation works, such as those use prompt engineering, information bottleneck, and purely generative approaches.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|17", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "This paper has limited coverage on these topics.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|18", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Touching different approaches is important here since the way ProtoTrex handle explanation might not easily extend to all other cases.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|19", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Even though this design was from the NeurIPS 19 paper, but in the task of NLP, how does the prototype embeddings compare against the label-wise weights in the final classification layer?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|20", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "This is to imagine that, without the use of the complicated loss in Eq 1, how does simply treating the label-wise embeddings as prototypes perform?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|21", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "The explanation generation (sec 3.5) needs elaboration.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|22", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "It seems this paper uses prototype embedding to find a training example as it nearest neighbor, and then use this data point as explanation to its prediction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|23", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "This design has certain limitations: a) not context/example dependent; b) this is hardly generation, instead, it is more in line with salience-based explanation works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|24", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Tab2 indeed shows some examples with explanation that partially depends on the input example.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|25", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "But no idea how they were generated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|26", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Tab 1b is confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|27", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "I don't understand what each number means.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|28", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "This paper very briefly went over some statements without getting into details.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|29", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "I don't see a solid explanation evaluation in this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|31", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "And not sure if rationale performances are based on token or sentence selection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|32", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Either way, this evaluation has nothing to do with generation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|33", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "And when it comes to generative explanations, ideally, there should be some human-based evaluation over a subset of testing data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#2|34", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "But no such thing in this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#3|11", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "According to the illustrations at the end of Section 3.4, there are many hyper-parameters should be set (along with lambda terms in training loss).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#3|12", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "How to determine these hyperparameters?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#3|13", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "If we move to a new dataset, we might lost in tuning these hyperparams.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#3|14", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "As the model incorporates additional modules (compared to pure classification model), it is questionable how efficient is Proto-Trex (both for parameter size and inference time).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#3|16", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "It is unclear where are the results from in Table 1b and 1c.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|12", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Hyperparameters: My initial impression of the method is that it involves far too many hyperparameters (various $\\lambda$ for combining losses, choosing similarity metrics etc).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|13", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "How hard was tuning hyperparameters?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|15", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "different models respond differently to the proposed approach.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|16", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "For instance, we see large performance gains on BERT but performance drops for SBERT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|17", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "It would be good to know why this happens.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|18", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "User interaction results are not convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|19", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "From Table-3, it seems like the interactions barely help, even though the prototype changes sometimes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|20", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Could the authors comment more about this?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|22", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Prototype quality: prototypes do not look very convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|23", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "I glanced at the prototypes in the Appendix and they seem to be quite redundant / of low diversity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|24", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "For instance, from Table-6 a lot of the prototypes either model “bad service and bad food” or “good service and good food” (P2/P4/P6/P8/P10).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|25", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "However yelp also has examples with “bad service, good food”, “good food, bad service”, examples not related to restaurants at all but the prototypes do not uncover these.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|26", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Faithfulness: How competitive are these numbers with post-hoc methods?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|27", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "I think for presentation purposes, it would be good to also compare this with SHAP/LIME/IG.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|28", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Overall, I think this paper is currently borderline.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc86|review#4|29", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "However, I’m happy to increase my score if the weaknesses can be quickly addressed —- the main ones are providing evidence that the learnt prototypes can be diverse and cover a lot more of the “modes” in the training distribution, providing a more comprehensive comparison with posthoc interpretability approaches, and confidence intervals.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|2", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Identify implicature understanding as a useful benchmark of LLM's communication ability.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|3", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Design a protocol for comparing implicature understanding of LLMs versus that of humans.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|4", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Roll-out the protocol on a host of existing LLMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|5", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Providing a deeper analysis of (a) how different properties of LLMs (e.g., instrucability or size) affect their performance, (b) which types of implicatures are harder for LLMs to understand, and (c) the effects of multi-shot learning on performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|6", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Broadly speaking, implicature refers to the act of saying one thing that implies another.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|7", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The authors observe that any entity (or model) that is claimed to \"understand language\" or \"be able to communicate\" must, in particular, understand implicature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|8", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Continuing the wide range of works aimed at understanding and evaluating LLMs, the authors design a protocol for evaluating an LLM's ability to understand implicature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|9", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The protocol entails prompting an LM (or human) with text of an implicature in the form of a question and answer; the task is to classify whether the answer means \"yes\" or \"no\" (logits of any token other than yes/no are masked, I think).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|10", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Performing better-than-random (50%) at this task necessitates implicature understanding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|11", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The authors use six prompt templates, some structured (like script) and some natural (like prose).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|12", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Using this protocol, the authors compare the performance of various LLMs and benchmark these against humans.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|13", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Overall, the authors find that the zero-shot performance of most LLMs is fairly close to random (around 60%), with the best models achieving around 72% accuracy -- compared to the mean 86% of humans.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|14", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "From this, the authors conclude that LLMs fail at zero-shot implicature understanding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|15", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Few-shot evaluation sees an increase to about 80%.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|16", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "In addition to these main findings, the authors perform several refined anlyses:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|17", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "breakdown by prompt type, showing that structured prompts are easier than natural prompts, but the overall ranking among models is not sensitive to prompt type", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|18", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "breakdown by implicatures type, showing that generalized implicature is easier than particularized implicature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|19", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "in k-shot evaluation, how performance scales with k, showing diminishing returns for $k \\geq 5$.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|20", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "how performance scales with model size (the conclusion here was unclear to me).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|23", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Rigorously evaluating implicature understanding is an interesting contribution to a timely and important area, namely, LLMs understanding of language and communication abilities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|24", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The protocol is extremely simple.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|25", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "This might seem like a Con, but often times identifying simple ideas and executing them correctly requires much heavy lifting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|26", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The authors put in the work to present their idea clearly: the paper is well-written and concepts are clearly defined and demarcated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|27", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Overall, the empirical evaluations are easy to understand and fairly complete (Cons 2-5, below, notwithstanding).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|28", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The in-depth breakdowns shed light on the general claim that LLMs fail at zero-shot implicature understanding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|29", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "I commend the authors for identifying a subset of their experiments that can be run without access to paid APIs or large compute.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|30", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The authors survey existing literature on implicature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|31", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Cons 6-8 aside, the findings of empirical evaluations are presented in a careful and complete manner.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|32", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The experiments seem reproducible (though I do not have API access or compute to attempt a full reproduction).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|33", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Cons & suggestions for improvement Items (6-8) refer to the code and reproducibility, and items (9-14) are minor comments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|37", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The perfomance gaps of LLMs and humans are about 14% in zero-shot evaluation, which decreases to 6% in few-shot evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|38", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Can you quantify what these gaps mean?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|39", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "How much better is 6% than 14%?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|40", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Given that humans perform imperfectly in this task (86%), I imagine there is a some x% threshold past which perfomance is \"essentially\" human.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|41", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "How close is this threshold to $x=86$?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|42", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "This likely requires a new set of experiments, but seems an important step forward that I would suggest as future work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|43", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "For example, a Turing test in which a distinguisher prompts a black box with an implicature (or sequence of these) and must distinguish an LLM from another human.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|44", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "In Table 1, the differences between Davinci-001 and Davinci-002 are within each others' standard deviation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|45", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Does this affect the statistical significance / validity of the claim that Davinci-001 outperforms Davinci-002 (page 6)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|46", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Not much changes in the paper if you say that the Davincis are indistinguishable, but I find this glaring and worth a brief discussion in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|47", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The identity of the asker/answerer among the pairs (Esther,Juan), (Karen,William) and (Bob,Alice) may affect the performance of the models: for example, a question asked by a Karen may be interpreted differently than a question asked by a William.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|48", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "I suggest adding a small ablation permuting the order of the pairs to test this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|49", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Why were human subjects only shown Prompt #2, rather than a mixture of all six prompts?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|50", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "As the authors find, structured prompts are easier for LLMs than natural ones.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|51", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "It is likely too late to test the effect of multiple vs. single prompt types on human performance, but I would expect to find an explicit discussion of this choice in the text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|53", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "No Python version is specified in the readme, and the code is incompatible with any version other than the intended one.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|54", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Which version should be used for reproducing the results?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|55", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "I managed to infer that it’s 3.9 using binary search:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|56", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Your use of the walrus operator (’:=’) implies Python≥3.8.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|57", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Your requirement of torch==1.10.2 implies Python<3.10.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|58", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Your use of union (’|’) operator implies Python≥3.9.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|59", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Even after finding the right Python version, the code does not run out of the box.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|60", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Specifically, I had to manually downgrade protobuf to 3.20 (pip install -Iv protobuf==3.20).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|61", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "You may want to add this to your requirements.txt.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|62", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "You should add a lightweight and easy-to-run sanity test.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|63", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "For example, an argument-less script that that trains a model on tiny data, validates on the same data, and asserts very high accuracy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|64", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "While the small experiment from the readme did reproduce for me, it took several minutes to run; it's good to have something that fails faster than that in case things aren't working.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|66", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "In the paragraph entitled \"On Prompting\" (page 8), the authors state that \"There is a narrative around large language models that if they fail a task, it might that the prompt was not the right one.\"", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|67", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "This statement should be supported by citations (just a few examples would suffice).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|68", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "A footnote next to a punctuation mark comes after the mark, rather than before it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|69", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "In page 5, move footnote 3 to come after the period.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|70", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Although Wittgenstein is cited in the appendix, I expected to find him cited alongside Grice and Huang in page 1, supporting the statement that meaning is determined by contexts, beliefs and social institutions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|71", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Not a big deal, I'll leave this decision up to the authors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|72", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Consider changing the title of Appendix B to \"Background on Implicatures\" or just \"Implicatures\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|73", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "On page 19, the authors write that Bach argues that \"conventional implicatures are [...] instances of something else\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|74", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Can you give an example of what this something might be?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|75", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "It would be nice to conclude Appendix B with a paragraph explaining in what way this work is informed by the literature mentioned in this section.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|76", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "What is the connection between conventional/conversational implicatures (in the appendix) and generalized/particularized implicatures?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|77", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "If conventional/conversational implicatures do not make an appearance in the rest of the work, why the long discussion of them?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|78", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|79", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "These are all covered in the sections above and below.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|81", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "This paper offers a significant contribution to an important area of research.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|82", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Overall, it is written well, its novelties are well-defined and explored, and the main claims are sufficiently supported by empirical findings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|83", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Do not take the many suggestions I made above as a serious negative sign; only a handful of them noticeably impact the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|84", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "I strongly suggest the authors to make the experiments more easily reproducible by fixing Cons 6-8.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#1|85", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "EDIT: The authors' rebuttal adequately addresses my concerns (primarily reproducibility and the title) and I have increased my score.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|1", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "This paper evaluates whether or not large language models understand implicature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|2", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "They create (automatically) a dataset that allows them to evaluate pre-trained models’ understanding of this phenomenon.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|3", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "They evaluate LLMs pre-trained on large text corpora and also ones that are fine-tuned with RL feedback to see the differences in models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|6", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "This paper is well-motivated and the evaluation is formulated clearly and concretely.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|7", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The authors spend time evaluating all models in different ways (e.g., relative accuracy, comparison across model sizes and classes, measuring the effect of in-context examples and so on) which is more comprehensive than most evaluations and helps understand model performance", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|8", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The authors report variance with the accuracy results which is good to see—however it looks like the variance is exceptionally high?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|9", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The authors also measure the effect of a larger number of samples (k), whether or not the wording of samples in the prompt affects performance, and they break down performance by example type which allows a nice and comprehensive analysis", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|15", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "There are several citations re: language models and implicature that are missing. They are:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|16", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "“Linguistic models for analyzing and detecting biased language. Recasens, M., Danescu-Niculescu-Mizil, C., and Jurafsky, D. (2013)”,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|17", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "“More than words: Syntactic packaging and implicit sentiment. Greene, S. and Resnik, P. (2009)”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|18", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "“Was it “said” or was it “claimed”? How linguistic bias affects generative language models. Patel and Pavlick (2021)”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|19", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|20", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "This paper is well written and seems reproducible (assuming the dataset is released).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|21", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Small clarification questions and typos are added below.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|23", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "1. Page 4. “..the order in which these examples are presented matters.” ← can the authors clarify why this is the case, since there is no explanation given.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|24", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "2. On naming “natural” vs. “structured” prompts. This seems somewhat non-intuitive–can the authors clarify this naming strategy?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|27", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Page 4. “.. change in protocol allows us to control for two two sources..”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#2|29", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "This paper studies an interesting problem however the framing of the paper is misleading.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|1", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "In this paper, the authors construct a test to examine LLM's ability to understand certain types of conversational implicatures by asking the models to convert non-direct answer to yes/no questions to yes/no.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|2", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "They found that intractable GPT-3 models perform much better than other baselines they compared but still a lot worse than human performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|3", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "They also showed that scaling doesn't help on this task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|4", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "In-context learning helps the task up to 5 examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|7", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "It is interesting to understand how LLM understand implicatures which is acquired through social learning for humans.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|8", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The contrastive task design is neat and effective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|9", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The experiments make sense and the results are informative.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|11", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Given the small size of the dataset, some of the results for smaller categories may not be significant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|12", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "It would be nice if the authors could calculate the uncertainty for Figure 3 & 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|17", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|18", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "This task and experiment design in this paper is very straightforward, and it is presented clearly.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|19", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Given access to the models used in this paper, I think the results should be reproducible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#3|21", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Good paper on analyzing implicature resolution of LLMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|1", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Authors evaluate pretrained large LMs on the binary implicature resolution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|2", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The results reveal that there is still a big gap between human performance and LM's on this task, indicating the challenge to interpret the language in context of the LMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|3", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "They also show that increasing the model size and prompt size (through few-shot prompting) is unlikely to close the gap, especially for the samples that require more context understanding.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|6", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The paper presents an interesting result of pretrained LMs performance on the binary implicature resolution task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|7", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "While we are seeing promising performance with the large LMs, especially with the newest ones like OpenAI's, there is still a gap with human performance on this implicature resolution task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|9", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The claim that large LMs are not zero-shot communicators is a little strong because it is only based on a single dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|10", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "There will be questions regarding whether this dataset is representative?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|11", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "the dataset size (600 samples) is enough or not?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|12", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "and even the quality of the samples/labels?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|13", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "For the samples that require context understanding (\"Particularised\"), authors can show more case studies and analyze the predictions on these samples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|14", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Also, I think it is better to plot the absolute accuracy scores in Figure 3 to illustrate whether the performance on this type of samples will increase with the model size.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|15", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The claim that \"increasing the model size and prompt size is unlikely to close the gap\" seems not true for the recent model from OpenAI.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|16", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Evidence in Figures 2 and 4 show that we are seeing performance gain when going to bigger model or having more few-shot examples in the prompt.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|17", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Furthermore, the current paper does not have enough analysis to understand why the current LM models fail and hence suggest directions for improvement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|18", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|19", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The idea is clearly presented, and the paper is easy to read.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|20", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The appendix section provides more details about the evaluation results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|21", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "The result seems reproduceable without much effort.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|23", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "Although this work reveals an interesting result when evaluating LMs for conversation context understanding, this evaluation is on a single test set, hence it is not strong enough to make a claim on the general effectiveness LMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc87|review#4|24", "source": "rc", "split": "test", "paper_id": "rc87", "comment": "It is also expected to see more analysis to understand why the current LMs model fail on this test and hence have some potential directions for improvements.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|5", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "I think the authors found the right critical bottlenecks for the MoE models, which are the trainability (model quality aspect) and the total number of parameter count (system performance aspect).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|6", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The proposed solution with the empirical results shows it's on the promising direction, although it's still far from completely addressing those foundational limitations of MoE,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|7", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The ablation and scaling laws section are very helpful to the research community to understand how to set the hyperparameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|9", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "How the speedup in table 1 is evaluated?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|10", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The gains in the downstream tasks are marginal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|11", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "It's better to report the variation of zero-shot results at nearby checkpoints as well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|12", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "I feel it's important to report the inference step time during autoregressive decoding to best demonstrate the gains from a smaller number of parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|13", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Because during decoding on accelerators, it's more often memory bound instead of compute bound.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|14", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "When decoding a single token given the prefix, the flops to compute each token is relatively small.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|15", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "However, the whole model parameters (in billions or even trillions) needs to send more HBM to the actual compute units during each decoded step.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|16", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "This HBM-cache communication is usually the dominant factor in the inference cost.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|17", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Double check the multi-rc results?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|18", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "It's a big jump from 1+ to 70+.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|20", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The flow of this paper is crystal clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|21", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The authors first identified the bottlenecks of an existing algorithm, found the root cause, proposed a solution, and finally demonstrated the effectiveness of the proposed solution with well prepared experimental evaluations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|22", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The text, tables, and figures are all of high quality.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|24", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "This paper worked on an important research topic: how to reduce the training and serving costs for large language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|25", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The proposed algorithm is only marginally novel but empirically significant.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|26", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The 5x reduction in the number of total parameters would improve the serving speed for MoE models by a lot.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|27", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "However, the gains in the downstream tasks are only marginal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|28", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "So it would be better if the authors clearly demonstrated why the large reduction in the parameter count matters using the metrics people care about: the serving latency, the step time, etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#1|29", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Alternatively, the authors can show the quality difference when matching the inference cost.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|6", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The proposed method is simple in design and implementation but achieves reasonably good results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|8", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The paper does not provide a fair comparison by fixing total parameters but ignore the computational cost (FLOPs) or activated parameters. In traditional MoE research, the general goal is to achieve better quality with a fixed computational cost (FLOPs), not with a fixed total parameters. The reviewer understand that this method provides a efficient way saving total parameters, but the reviewer suspects to achieve better quality, this method would also significantly increase the FLOPs compared to traditional top1 or top2 based routing used in switch transformer and GLaM.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|9", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Table 1 does not provide any details on computational cost.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|10", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "According to multiple prior works including T5 and the Chinchilla [3], there is always a tradeoff between model capacity and training tokens, the larger the model is, the lower training data/steps can be achieved within the same computational cost budget.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|11", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "According to Table 1, activated parameters are made fixed around 1B, however, it might not be clear about computational cost in this work's setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|13", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "GLaM adopts a top-2 based routing, that can yield much better results than top-1 based routing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|15", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "For example, Expert Choice [2] routing achieves heterogeneous experts such that different tokens can utilize a variable number of parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|20", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The paper is clearly written.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|21", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The proposed method is a minor improvement over traditional parameter sharing scheme like suggested in Universal Transformer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|22", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The paper can be reproduced but the results are not valid.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|25", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The paper's results are based on fixing total parameters but not on fixing computational cost (activated parameters and FLOPs), which can be unfair to many related works including GLaM.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|26", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The reviewer would not believe in any results that purely relying on parameter sharing, we could improve quality without introducing additional computational cost.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|27", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "For example, whether this method increases activated parameters (experts per token) is unclear and should be explained.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|28", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "For example, the paper can be increasing the number of layers or expert width or number of experts per token compared to GLaM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|29", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "All these increase activated parameters, thus inference time.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#2|30", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The paper should be also more proactive in explaining why quality gains can be achieved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|4", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "(1).This paper carried out analysis first and listed three challenges from analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|5", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "However, I did not know which MoE model does this paper study. In Figure 1, it shows “MoE” but I don’t know which MoE model is used to carry out experiments.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|6", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "There are plenty of MoE models such as Gshard (Lepikhin et al., 2020), Switch Transformer (Fedus et al., 2021), Base Layers (Lewis et al., 2021), HASH Layers (Roller et al., 2021), and etc.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|7", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Different MoE models may lead to different conclusions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|8", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The author needs to announce which model they used for analysis and add citations.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|9", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "(2).This article used only one MoE model to draw analysis conclusions, which I cannot agree with.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|10", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Because different MoE models may have different performance, analysis conclusions need to conduct experiments with at least two representative MoE models when talking about common challenges with MoE models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|11", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "(3).I am very suspicious about the expert pool method proposed in this article.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|12", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "How to choose the size of the expert pool. I speculate that the amount of experts required by a MoE model may be related to the diversity of the dataset.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|13", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Table 1 in BASE layers paper [1] shows similar words usually gathered to the same expert unit.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|14", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "However, this article only uses one dataset for pretraining, and does not use multiple datasets to test the required expert pool size.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|15", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "(4).Followed by the third problem, this paper selected the Pile dataset as the pre-training dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|16", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "However, the Pile dataset is full of duplicate documents (see [2] page 2), and this paper does not perform additional de-duplication processing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|17", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Because the dataset selected in the article has a lot of repetition and the tokens are not diverse, the size of the expert pool does not need to be large.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|18", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The conclusion is likely to change when changing to a different (diverse) pre-training dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|19", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "(5).As a MoE model, it is basically necessary to control the number of flos and compare it with the dense models and sparse models with the same number of flops, but this paper does not report total training flops number and total train computer (PF-days).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|20", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "In addition, this paper doesn’t compare it with a dense model with the same amount of flops in table 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|21", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "(6).I also have some questions about the experimental results of table 2.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|22", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "When we compared SaMoE (350M-128E) with dense model (350M), SaMoE should have more flops since it needs additional all2all communication cost.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|24", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "This SaMoE with more flops achieves a score 68.9.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|25", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "(7).Minor suggestion: usually we reported pretraining perplexity instead of validation loss in figure 3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|27", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "[1]. Lewis, Mike, et al. \"Base layers: Simplifying training of large, sparse models.\" International Conference on Machine Learning. PMLR, 2021.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|28", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "[2]. Zhang, Susan, et al. \"Opt: Open pre-trained transformer language models.\" arXiv preprint arXiv:2205.01068 (2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|31", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Due to the above weaknesses, this paper does not have a high quality.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|34", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "However, in the analysis part, it only carried out analysis experiments with one MoE model, which is hard to tell if findings applied to all MoE models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|35", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "In addition, this paper proposes to have a fixed number of global MoE layers, which is probably not suitable when a pre-training dataset has very diverse tokens.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc88|review#3|36", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "It happens that this paper selects the Pile as the pretraining dataset, and Pile is widely considered to contain many repeated sentences. (see [2] page 2).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|10", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "The analysis section on what works and why is a bit thin; Table 5 overall is quite confusing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|11", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "I believe answering the following questions may help add more clarity and strength to this part.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|12", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "What does Hits mean?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|14", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "What is Hits@1 (no explanation in text)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|15", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "What is (EM) following Hits@1?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|16", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "I'm guessing that's the subset of examples that the model answers correctly?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|17", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "If I'm guessing correctly, the definition of Not Recit, hits@20-recit., and hits@20-path follows a binary decision tree, on whether the ground truth answer appears in any recitation or any QA output.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|18", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "It is rather difficult to figure this out based on the text in 4.3.4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|19", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Maybe a logical tree of some sort can help improve readability here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|20", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Are there cases where, e.g., the ground-truth answer is not recited but the model still answers the question correctly?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|21", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "E.g., a more detailed characterization of Hits@1 would be interesting, unless reciting the gt answer is an empirical necessity – which by itself is interesting to point out if true.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|22", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "How are the 1024 examples selected?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|23", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "On a more minor note, the names for categories are very not self-explanatory and could be improved for better readability.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|24", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "There are some missing details that might slightly hurt readability", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|25", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Judging from Fig 2-(1), <Recitation N> is generated from – and directly related to – <Question N>, with other Q-R pairs as in-context examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|26", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Moving on to Fig 2-(2), it seems there are multiple QA pairs (Q0, A0, ...) inserted between <Recitation N> and <Question N>.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|27", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Are the questions in these QA pairs the same as those in the QR pairs in the top-left block of Fig 2?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|28", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Are they somehow related to <Question N>?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|29", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "If not, is there going to be some long-distance memory issues between <Recitation N> and <Question N>?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|30", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "In Fig 2-(3), are the multiple paths (containing Recitations-dots-Answers) repeating Fig 2-(2)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|31", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "In the multi-hop setting, do the numbered prompts like \"Recitation 1\" and \"Recitation 2\" naturally give rise to generating recitations of different topics, or does one have to train/update the LM to solicit such behaviour?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|32", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Since the multiple recited passages are generated in one-pass from the LLM decoding sequentially, ... How to use multiple numbered prompts to generate multiple recitations in one-pass sequentially?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|33", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "I'm personally curious about more details on the generated recitations, e.g., in NQ dataset, do the recitations resemble the \"ground truth\" passages in number of words/sentences/paragraphs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|34", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "What percentage of recitations contain the correct answers?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|35", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Minor points on motivations", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|36", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "One motivation mentioned for recitation is to match the form of causal LM pre-training objective (with the decimal of π example).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|37", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Personally, I tend to believe that recitation helps by making certain knowledge more explicit (value of π), and when primed on this explicit piece of information (in the form of recitation), the LM is then more likely to arrive at the correct answer (the 10th decimal of π).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|38", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "In other words, I'd argue that with or without recitation, the objective remains the same (i.e., masked or next-word prediction), so this motivation to me feels a bit mis-directed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|39", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Human's ability to recite relevant factoid knowledge before answering knowledge-intensive questions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|40", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Do humans have this ability?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|41", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Regardless of the answer, having this ability doesn't necessarily help humans answer knowledge-intensive questions more accurately, which I guess should be the true human-inspired motivation for the proposal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|42", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "In any case, citations are perhaps needed to make either claim.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|43", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|44", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Both the core idea and writing are both clear and comprehensible.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|45", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Missing details may well be compensated by open sourcing the code for reproducibility.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|46", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "The idea itself is well-motivated and novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|47", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Summary Of The Review:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|48", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "The paper is of overall good quality, with a sound idea and good execution by putting many moving parts together into a working system with empirical gains.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#1|49", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "The analysis part is a bit unclear and underwhelming.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|20", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(1) Major concerns that will be listed below.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|21", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(2) Unclear cost for running the proposed methods against the standard-prompting models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|22", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(3) Insufficient rationale to support gains and no gains", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|23", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|24", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Mostly clear and novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|25", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Summary Of The Review:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|26", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(1) As the passage-hint model is a corpus-specific treatment that additionally utilizes title hierarchies, the prompt-based model must have great added benefits.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|27", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "The core concern is that the model performs in-context learning based on the model-generated recitations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|28", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "No matter how you tune the temperature parameter, there must be a pathological trade-off between sacrificing diversity and sacrificing factuality.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|30", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(2) It will be great to have the performance graph (given two metrics) over increasing number of few-shot prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|31", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "If you could approximately measure the correctness of generated recitations, another figure that demonstrate the performance dependency with respect to the number of correct shots will benefit readers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|32", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(3) Another major concern is the computational trade-off from using many shots.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|33", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Table 1 shows that the proposed recite-and-answer model (with 5 shots) does not have significant impacts against standard-prompting (with more shots).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|34", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Can you address the claim that using the standard-models with more shots is more expensive than running your prompt-based model with less-shots?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|35", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(4) It is less clear that how you decide the number of shots you would use for standard-prompting model and your recite-and-answer model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|36", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "For example, UL2-20B is tested with 16 shots, whereas the in-house version used 64 shots.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|37", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Justify the rationale of picking the number of shots and the reason not to test on more shots for OPT-30B.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|38", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(5) Excluding UL2-20B (mostly encoder-decoder T5 structure except the mixture of denoiser variations with multiple special tokens), is your In-house model of decoder-only?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|39", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "If so, why In-house has significant performance gain on HotpotQA but no visible boost on TriviaQA, which is the entirely opposite to OPT-30B?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|40", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Any specific example-based explanation than a conjecture would be beneficial.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|41", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(6) The passage-hint model consists of significantly more steps.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|42", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "It will be useful to see overall time-cost trade-off between your model and standard models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|43", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(7) Is the passage-hint model flexible if the evidence documents do not have a clear and fine-grained title/subtitle structures?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|44", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "What if some of title components from this hierarchy is missing?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|45", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Should we expect performance degradation or performance boost because the passage hints are also generated being possibly incorrectly if the information structure is too granular.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|46", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(8) Overall, not much qualitative examples are provided.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#2|47", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Having some successful and failure cases by examples will be highly", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|17", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "The paper would be much stronger with experiments on GPT3, Instruct-GPT3 (davinci-002), and larger language models (larger in-house LMs?).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|18", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "It's not really clear from the paper whether recitation helps with larger scale, which I think is important for the generalizability of the method [1].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|19", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "This could work both ways --- I'm suspecting larger LMs will be better at both recitation and directly performing QA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|20", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "I think experiments on InstructGPT [4], T0 [3] or FLAN [2, 7] will be especially interesting, since it's been fine-tuned on instructions / examples / human preferences.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|21", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "A major advantage of retrieval augmented systems is their applicability on (1) tail distribution information; (2) generalization to information which was not present in the model's training set (like COVID for BERT).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|22", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "I think these are important limitations of the proposed method, and (1) is not really discussed (2 is just mentioned in the conclusion).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|23", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Are most of the correct recitations cases which were seen often during training?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|24", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Overall, the performance of closed-book models in this paper seems to significantly lag behind recent few-shot retrieval-augmented systems [5, 6].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|25", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "For instance, ATLAS [5] gets 42% on NQ with 64 examples and a smaller model, while the best number in this paper is 32% (5-10x larger model).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|26", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "While I agree that setting up retrieval is technically cumbersome, there are very good retrieval APIs available, which were utilized in [6] without any extra LM fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|27", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Note that I do think it's incredible that closed book LMs are doing so well, but practically (from a performance stand-point) it may be better to just retrieve some text from the web rather than ask an LM to generate it with few-shot examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|28", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Also, retrieval augmented LMs often have lesser parameters [5], so it's unclear which is a better method from an efficiency perspective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|30", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Perhaps the performance gains are because of the dedicated fine-tuning on Wikipedia data for the recitation LM model (which makes it overfit to Wikipedia)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|31", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Did you remove the passages from the test set questions while doing this fine-tuning?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|32", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Also I don't think enough experiments are done in the paper to justify its added complexity over vanilla LM-Recitation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|33", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "I would suggest moving it to the appendix, or performing experiments on all 3 datasets / models to show its benefit.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|42", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "This paper is relevant to https://arxiv.org/abs/2004.05483 and https://arxiv.org/pdf/2110.08387.pdf, it would be great to cite them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|43", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "In Table 4 (LM-Recitation_5), why is the number for different from Table 1 (Recite and answer)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|44", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "(16.34 in Table 1 vs 14.16 in Table 4)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|45", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|46", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Clarity - Very clear", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|47", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Quality - Very thorough experiments overall except for the experiments on passage hints.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|48", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "I would have liked other models being tested (weakness #1), but the experiments on the current set of tasks look good to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|49", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Novelty - Good novelty.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|50", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "The idea has similarity to chain-of-thought prompting, self-consistency prompting, and self-talk, but overall I think the idea is pretty new (especially in the context of large LMs and QA).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|51", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Reproducibility - Should be fully reproducible except the experiments on the in-house LM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|52", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Summary Of The Review:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|53", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "The paper has interesting ideas and surprising results, but I have two main concerns - (1) the paper does not evaluate the method on larger LMs which are available; (2) I don't think there's justification that this method is a replacement for retrieval in any way (weakness #2, #3).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|54", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "I am currently leaning reject, but will be happy to move to the accept range if weakness #1 is addressed via experiments on GPT3-170B and InstructGPT3-170B.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|55", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "After rebuttal: Thanks to the authors for the very detailed response!", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|56", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "I've decided to raise my score to 6 (accept range) due to the improvements shown on Codex.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#3|57", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "I would still suggest the authors to take a more balanced take in their conclusion, mentioning that while there are improvements over direct generation, there is still a gap behind retrieval-augmentation on NQ.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|9", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "As also mentioned is Limitation, the main drawback is the the updated of the model with new knowledge.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|10", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "This is even more evident when compared to simple DB retrieval in Table 5, where BM25 performs better that LM-Recitation 5-shot with 20-path (the best proposed model).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|12", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "What is the size of the index + docs (compressed) of BM25?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|13", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "is the size larger or smaller of the model parameter?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|14", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "if the size of the DB is larger of the model parameter, why not constrain it (e.g., randomly removing articles) and check if the performance in Table 5 changes?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|15", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "What is the inference speed of LM-Recitation K-shot with t-path?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|16", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "How's compared to BM25?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|18", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "what's the performance of larger LM (PaLM-500B or GPT-3/Instruct-GPT3) or smaller (GPT-J 6B)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|19", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Figure 4 shows the performance difference by sampling more self-consistency paths, but what append if the model samples more responses for the direct method (or chain-of-thoughts)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|20", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|21", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Clarity: Very clear Quality: Very high quality Novelty: somehow novel Reproducibility: 2 out of three LM are available for the research community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|22", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Summary Of The Review:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|23", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "The paper propose a two step approach to improve knowledge-intensive NLP tasks by using only LMs (no external retriever).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc89|review#4|24", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "However the model performance are a bit lagging compare to simple BM25 + LMs, plus the clear weakness of updated knowledge.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|2", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "The proposed benchmark was translated by native speakers of the target language with at least two years of professional experience.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|3", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "The experiments are well-designed, and the results are enough to support the claims.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|4", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "They find that on the MGSM benchmark, LLM-540B shows superior multilingual reasoning ability compared to GPT-3 on all languages with different settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|5", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Similarly, for the commonsense reasoning benchmark, LLM-540B also outperforms the SotA models: RoBERTa Large and XLM-R Large on two other benchmarks: XCOPA and XL-WiC, respectively.", "reviewer_type": "human", "label": "error", "error_type": "Inaccurate Summary", "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|8", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "The multilingual benchmark is translated by professional translators and verified by an additional one for a random subset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|9", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Experiments are well-designed with 4 prompting methods and 3 types of input exemplars.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|10", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Intensive analysis provides new insights for multilingual reasoning abilities of large language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|11", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "The paper is also well-written.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|14", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "The verification process is less discussed, and not convincing to guarantee the quality of the benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|15", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "The approach of translating a problem to English does not test the multilingual reasoning abilities of LLMs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|16", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|18", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "In the \"Manual translation process\" passage, who are the popular MT providers?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|19", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Questions to the authors", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|20", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Did you check the math background of translators?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|21", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "How do you measure the quality of translated examples compared to English examples?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|24", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Although the findings are interesting, the contributions are not significant enough for the multilingual research since the benchmark is small and the verification process is not well described (or implemented?), which may flip the current conclusions/claims if more multilingual examples with different types of problems are added.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#1|25", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "The big plus of this paper is the intensive experiments and detailed analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#2|1", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "This paper considers chain-of-thought (CoT) in multilingual setup and shows that despite a highly imbalanced pretraining dataset, LLM still learns strong CoT capability in non-English.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#2|2", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "It extends an existing dataset to multiple languages. It also compares different ways to construct exemplar and CoT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#2|7", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#2|8", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "This paper is well written and the experiment is well designed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#2|9", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "It considers CoT reasoning in a multilingual setup and introduces a new dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#2|10", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "However, portions of the experiment is impossible to replicate due to the proprietary nature of LLM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#2|11", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Is the dataset and CoT explanation publicly available?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc90|review#2|12", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Caption of table 6 has [CITE]?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|1", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "This paper introduced a multilingual arithmetic reasoning task with a corpus named MGSM, and used the few-shot chain-of-thought (Few-Shot CoT) prompting framework to evaluate the large language models’ few-shot learning capability.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|2", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Authors empirically found that the GPT-3 and LLM could provide reasonable performance on their constructed test set, and also provided results on two other multilingual reasoning tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|5", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "The paper shows interesting experimental results on the multilingual arithmetic reasoning task MGSM. Large-scale language models can perform well across languages, and even on underrepresented languages like Swahili.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|6", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "The paper provides extensive and convincing experiment results on other two multilingual tasks (common sense reasoning and word in-context semantic judgment).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|9", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Some experiments and additional clarification can be considered and added (see Questions).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|11", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "In Section 4, it would be better if the authors add the experimental results of adopting Zero-Shot CoT (Kojima, Takeshi, et al. 2022) [1], compared with the Few-Shot CoT model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|12", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "One major question I have in mind is, is the current sample size of MGSM sufficient as a benchmark corpus? It would be better if the test sample size can be larger than the current 250, as the original GSM8K contains 1,000 test samples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|13", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "It would be better if authors can add some insight or feature-based analysis on the emergent ability of cross-lingual reasoning of the tested language models GPT-3 and LLM.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|15", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "I was also wondering whether the performance on MGSM (shown in Table 3) will be affected by adding some text perturbation (here the perturbation should not affect the numeral text), or adding some spans with code-switching.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|16", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "In Table 7, why are the number of few-shot exemplars in different language settings not the same? For instance, only 1 exemplar is provided for ru, th, te, and bn, which is smaller than that of en, de, fr, and es (6 exemplars).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|18", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "[1] Kojima, Takeshi, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. \"Large Language Models are Zero-Shot Reasoners.\" arXiv preprint arXiv:2205.11916 (2022).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|19", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|20", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Clarity: Very clear, the paper is well-written.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#3|22", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Novelty: This work is a multilingual version extension of the few-shot chain-of-thought reasoner. They empirically evaluated the multilingual reasoning capability of two large scale language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|1", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "This paper extends the prevalent GSM8K benchmark to a multilignual verison covering ten languages, and accordingly verify a recently proposed technique Chain-of-though upon it, resulting a best performance of 55 with the most competent LLM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|2", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "The authors also verify such capability on established benchmark XCOPA, achieving a new state-of-the-art.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|3", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "The paper is well presented and very clear to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|6", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "1. The paper is well presented and easy to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|7", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "2. a multilingual version of the popular benchmark GSM8K is proposed and released, which might be of interest to the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|8", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "3. new state-of-the-art on XCOPA and XLWiC is set.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|11", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "2. The empirical investigation is not sufficient for a solid benchmarking of the new datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|12", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Specifically, I believe several recently-proposed methods like ZeroShotCoT[1] and SelfConsistencyCoT[2] are also necessary and not difficult to implement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|14", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "And the empirical verification of CoT on more datasets are also not sufficient for another new paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|15", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "[1] Large Language Models are Zero-Shot Reasoners", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|16", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "[2] Self-Consistency Improves Chain of Thought Reasoning in Language Models", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|17", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Clarity, Quality, Novelty And Reproducibility:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|18", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Clarity: The claim is clear and backed by experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|19", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Quality: The paper is well written. The quality is ok but not extremely solid. Experimental investigations do not cover necessary scope.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|21", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Reproducibility: The dataset is released and the code is available.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|22", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Although I assume most people do not have access to GPT3 or the refered 540B LLM, which are the only keys to obtain such SoTA results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|25", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "But the technique is almost identical to CoT, and the results can be well expected, leaving it only a very incremental contribution.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|26", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Besides, such empirical verification is also not sufficient as important baselines are missing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc90|review#4|27", "source": "rc", "split": "test", "paper_id": "rc90", "comment": "Therefore, I vote for borderline reject, but I'm not absolutely sure and are open for discussion or rebuttal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|10", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The method is not compared with previous models in the experiments, such as Transformer-XL, Compressive Transformers, and other efficient Transformers for LM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|11", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "There is no ablation study in terms of modeling.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|13", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Is the application restricted to sequence decoding?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|14", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Can we finetune the proposed model as BERT on downstream tasks, and how about the performance?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|15", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The above question is related to generalization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|16", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Such kind of kNN method seems to suffer from too much memorization instead of generalization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|17", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "If the model cannot generalize well, it would not be good at zero- or few-shot learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|18", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "I agree that language modeling is a meaningful task, but it is more like a testbed for modeling techniques that improve generalization.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|19", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Otherwise, the shortcut of memorizing context would hide the real challenge.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|20", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The code seems unavailable.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|21", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "It is unclear how many engineering efforts are required to reproduce the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|22", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The paper writing ignores many details that are helpful for reproducing the method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|23", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Even the code would be released, the standalone paper should be informative enough.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|24", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The kNN layer is densely triggered for each position, which seems quite heavy.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|25", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "How about using a gate to determine whether we need to perform kNN?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|26", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Is there any evaluation about the performance of the learned kNN search? For example, we could conduct some human evaluations to see whether the retrieved knowledge makes sense for prediction.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|27", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The model is only evaluated on the language modeling task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|28", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "More tasks would be helpful to indicate whether the proposed method is general enough in terms of data distribution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|30", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "I am okay to accept the submission.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#1|31", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "But it can be improved as mentioned in the above review (e.g., including ablation studies, more tasks, comparisons with previous models, and analysis).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|22", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Section 3 could see more formalization and better description on what are the input for searching the kNN, how the model is trained (maybe, an algorithm figure).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|23", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Experimentation could be greatly improved: no hyperparameter search or reasoning behind the selection, no baselines with other memory architectures showing differences either performance or specific use-cases (beyond plain transformers).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|24", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Memory sizes are too big for the data selected.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|25", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "It is unclear why such big memories are still improving the model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|26", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The results on PG-19 doesn't seem to replicate previous work (original paper shows 33.6 for Compressive transformer, whereas this work shows ~19 for plain transformers).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|27", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Results are very close for each memory size and there are no error measurements (e.g., std dev).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|28", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "It is hard to conclude anything about memory sizes and their effectiveness.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|29", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Hard to replicate: several proposed datasets are downloaded from the web, it will be hard to download the same files.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|30", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "No details on the specific hardware and software used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|32", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "What do you refer with \"16 attention heads of size 64\"?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|33", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Describe the content of the figures in the caption (e.g., what is the architecture about?)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|40", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "My recommendation is to reject this work given the poor quality of the experimental section (see above and below).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|41", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Also, there will be some challenges to reproduce this work with respect to the datasets that should be addressed as well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|43", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Please, could you clarify how did you choose the hyperparameters and justify them?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|44", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "What is being used as search in k-NN?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|45", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The output of the layer connected to the memory?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|46", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Are the keys used in the search or both key/values?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|47", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Why did you choose k=128 in k-NN?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|48", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "What would happen if these value changes?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|49", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Why do you use one layer of memory, if two improve the results?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|50", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Could you add std dev.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|51", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "values to all your experiments?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|52", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Please explain why does the arXiv dataset have such a low perplexity?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#2|53", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "How many sequences are used for training, validation and testing?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#3|9", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "there are some benchmarks that require models to be able to handle long inputs, such as document-level QA datasets and document-level retrieval.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|4", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "This paper shows a novel, simple and intuitive method for extending the number of tokens a transformer can access at every timestep.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|5", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The authors augment the model with a kNN datastore, so that in the top layer, the model attends to the local context using normal attention but then also attends to up to 262k previous tokens by using an approximate nearest-neighbors datastore.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|6", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "This is basically equivalent to approximately attending over those previous tokens.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|7", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "No gradients are passed to the representations in the datastore in any way, making this easy to implement and relatively efficient to train.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|8", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The authors present results on multiple language modeling datasets that have long-range dependencies.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|9", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "On most of the datasets the gains from using the proposed method are quite small (<=.1 perplexity), and it’s not clear that they are statistically significant, but on the C4 dataset the model reduces perplexity by 2 which I believe to be a strong result.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|10", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "I believe that this idea is promising and that giving it a stage at ICLR will push the community to try further improving upon it, hopefully leading to larger gains.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|20", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The authors do not compare their model to any previously published model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|21", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Throughout the paper the authors only compare their model to baselines that they designed and trained.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|22", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Doing this for some of the datasets is fine, but not even having one dataset in which you compare to strong, previously published models makes it harder to convince the reader that you have found something interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|23", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "You did show results for PG19, which has been used before in LMing papers, so it’s not clear why you haven’t compared to previously published models on that dataset, such as the compressive transformer?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|24", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The numbers in the compressive transformer are very different from the numbers your models obtain, so I’m assuming that you preprocessed that dataset differently and so have a different vocabulary, making those results incomparable?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|25", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Could you re-train your model using the same settings as the compressive transformer so that we could compare the results of your method and theirs?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|26", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The baseline has just 6 layers, which is significantly smaller than the current transformer LMs that are being discussed by the community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|27", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "I’m definitely not expecting to see GPT-3 sized models in every research paper (or even close to that) but I think 16 layers would make more sense for such an exploration.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|28", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "I’m afraid the gains shown here would not transfer to larger models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|29", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "On the other hand, larger models might produce better representations resulting in even bigger gains.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|30", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The authors do not show standard deviation numbers for any of the baselines on the various datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|31", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "This makes understanding the results on each dataset, for the different models, harder.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|32", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "For example, on the arxiv dataset the authors talk about a drop from 2.927 to 2.877 perplexity between two different models, but is a 0.05 perplexity difference really statistically significant here?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|34", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Of course, it’s totally fine if this model runs slower or uses more memory than the baseline- I expect such preliminary investigations to potentially be nonoptimal, with later work making it an actually viable model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|35", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "I would not penalize this paper for having slow runtime or using too much memory.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|36", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "But I think it's important to discuss these things.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|38", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Although the perplexity improvements on many of the datasets are quite small (<=.1 perplexity) this could be a non-issue, and more of a problem with the perplexity metric than with the model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|39", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The perplexity metric, and most of the data that we train our LMs on, heavily incentivizes models to get short term dependencies right.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|40", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Getting the model to look back into the past is both hard and barely incentivized, and so these small initial gains that are attained in this paper might be improved when we find a better loss function that incentivizes looking back more, or when we find data that has more long-term dependencies and less short-term ones.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|41", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The way I see it, there are two different ways to augment the transformer with a memory module.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|42", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The memory module could be ‘global’ or ‘local’.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|43", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Local memory transformers have a memory module that can only access the recent past.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|44", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "So models like the compressive transformer or Transformer-XL are local memory transformers.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|45", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "On the other hand, models like the kNN-LM or REALM are ‘global’ memory models, since they can access the entire datastore for every prediction, without any kind of limitation that is dependent on the current timestep and the recent previous timesteps.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|46", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "The model proposed here is clearly a ‘local’ memory transformer, since it performs approximate attention over the recent past.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|47", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "I believe that that is part of what makes this model work, and I think the paper might benefit from a discussion of this and maybe even an empirical test.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|48", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Does this model improve performance on WikiText-103, where all the articles are in a shuffled ordering?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|49", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "I don’t think it would, and that’s totally fine, but I think readers would benefit from a discussion of this property of the model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|50", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "I think it could be helpful if the authors had more comparisons where their model and the baseline can access the same amount of tokens for every prediction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|51", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "For example, if you train one of the memory models with 2k tokens of memory and 512 tokens of local context it would be interesting to compare it to the baseline model trained on 2.5k tokens.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|52", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "Of course, since your model performs approximate attention we do expect to see worse perplexity values, but it would be very insightful to see precisely how much using this approximate attention method degrades performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|53", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "For section 3.2, consider referencing [CrossBatch] which discusses this distributional shift in section 3.2 of their paper, in the vision domain.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|55", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "[Compressive Transformer] Jack W. Rae, Anna Potapenko, Siddhant M. Jayakumar, Timothy P. Lillicrap. “Compressive Transformers for Long-Range Sequence Modelling”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|56", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "[REALM] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, Ming-Wei Chang. “REALM: Retrieval-Augmented Language Model Pre-Training”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc91|review#4|57", "source": "rc", "split": "test", "paper_id": "rc91", "comment": "[CrossBatch] Xun Wang, Haozhi Zhang, Weilin Huang, Matthew R. Scott. “Cross-Batch Memory for Embedding Learning”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|2", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "While there are no new tasks, the new benhmark adds value by combining complimentary lines of work in a well designed evaluation framework.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|5", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The tasks are chosen to cover a range of domain, task, and learning transfer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|6", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "There is a detailed investigation of the trade-off between large test sets with few test episodes, and smaller test sets with many test episodes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|7", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The test setting is chosen to maximise precision of the benchmark given a moderate compute budget.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|8", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The paper presents a new prompt-based approach to few-shot learning that converts each of the classification tasks into a multi-choice QA task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|9", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The details of this conversion are not given in the paper and I would like to see them listed.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|10", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Comparisons are also made to previous models that were applied to some subset of the tasks included in Fleet.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|11", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "While this paper does not introduce any new tasks or evaluation methodologies, it does add value by combining several lines of complimentary work on few-shot learning into a single evaluation framework that has been well designed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|12", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The prompt based UniFew baseline also seems both simple and effective, although I would like to see more details of the prompt template design in the main paper.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|14", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The authors point out that Fleet is limited by it's restriction to text-classification tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|15", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "It is clearly not a defnitive benchmark, but still valuable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#1|16", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The paper hints at the design of templates for the UniFew baseline but does not provide enough details for replication.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|1", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The authors introduce FLEET a new benchmark for evaluating few-shot learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|2", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "It's focused on classification tasks in NLP.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|3", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The features of this new benchmark are:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|4", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "evaluation of 4 different types of transfer", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|5", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "textual labels for zero-shot classification", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|6", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "experimentally determined sample size", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|7", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "In addition, they propose a new strong baseline called UniFew which is prompt-based model built on top of UnifiedQA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|8", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "They show that this method is competitive with other few-shot systems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|9", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Finally, they release the code and framework for further extension.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|11", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Overall, this work is going to be useful for the few-shot learning community.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|18", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The benchmark only includes classification tasks and in English only.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|20", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "More details on how the benchmark will look and how it's been designed would have been helpful. Screenshots of the baselines and the metrics reported would be good.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|22", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Can existing benchmarks like GLUE or SuperGLUE be converted to few-shot benchmarks by creating few-shot versions of them?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|23", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Wouldn't that be better since we will know both few-shot and fully supervised performance on the same set of tasks?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|24", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "It's a bit suboptimal for everyone to create a different set of tasks as a benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|26", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Include references about multilingual benchmarks: XTREME, XGLUE, XTREME-R", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|27", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Number of shots and classes: the numbers chosen here are a bit arbitrary and more concrete experiments could have been better.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|28", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The list of tasks is in the Appendix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|29", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "This is important and should have been included in the main paper", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|31", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Yes they have with their short statement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|32", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "They should have included considerations that go into selecting the \"few\" examples that the models fine-tune on.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#2|33", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "These need to be free of any bias.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#3|3", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The authors also provide a simple yet effective prompt-based learning baseline along with the FLEET benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#3|12", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Currently, FLEET only contains one entity typing dataset (CoNLL).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#3|13", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "This dataset is only used for task & pre-train transfer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#3|14", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "One could a) include more entity typing datasets b) and offer class transfer or domain transfer on the entity typing task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#3|16", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "The authors provide a section addressing the limitations in the provided benchmark.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#4|1", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "This paper proposes FLEET, a benchmark for few-shot learning in NLP.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#4|6", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "First, the benchmark contains a lot of different settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#4|8", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "On the other hand, it practically makes understanding, reporting, and using the benchmark a bit messy. For example, people will report very specific settings such as FLEET with {4-shot, with meta learning, task transfer}.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc92|review#4|10", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Second, a common criticism of GLUE and other benchmarks which applies to FLEET is that the benchmarks (1) simply repackage and resell other datasets as a new benchmark without adding substantial new value and (2) contain a relatively ad-hoc selection of individual datasets. FLEET is similar in that it sort of throws together a collection of tasks (sentiment, NLI, etc.) mainly guided by low-level reasons (e.g., the class imbalances) rather than higher-level ones (e.g., based on practical few-shot scenarios, different types of reasoning, tests of spurious correlations, etc. ).", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc92|review#4|11", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Third, a lower-level point is that I think few-shot benchmarks should contain multiple prompts / label descriptions for each dataset.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#4|12", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "There can be high variance across different prompts (Zhao et al. ICML 2021, Gao et al. ACL 2021, etc.)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#4|13", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "and so I think the reaction to this should be to report accuracy across multiple different prompts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#4|14", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "Finally, the proposed method is extremely similar to this concurrent work https://arxiv.org/abs/2104.04670.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc92|review#4|15", "source": "rc", "split": "test", "paper_id": "rc92", "comment": "This is not a negative or a positive, I am simply pointing this out as an FYI to the authors, as they should probably cite/discuss this in a future version of their draft.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|7", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "The writing is generally clear.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|8", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "The paper is addressing an important phenomenon that parameter-efficient tuning which has deficient ID performance, while full fine-tuning has worse OOD performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|10", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "First of all, the presented methodology is not well motivated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|11", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Parameter-efficient tuning aims to reduce number of parameters that need to save for each task by only fine-tuning a small number of additional parameters.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|12", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "However, the proposed method produces an ensemble model that fine-tunes all parameters, losing the parameter-efficiency.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|13", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "If the goal of this paper is to improve the OOD performance of full fine-tuning, it should compare with methods that improve OOD performance which has a vast amount of related work. And this literature review is also missing from the paper. On the other hand, if the goal of this paper is to improve the ID performance of parameter-efficient tuning, then it should propose a method for improving parameter-efficient tuning methods rather than combine it with full fine-tuning.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|14", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Second, although it's unfair to mention recent work after submission time, there is recent work on parameter-tuning that shows proper improvements over parameter-efficient tuning methods could match the performance of full fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|15", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "This makes the motivation of this paper less meaningful.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|16", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Third, the experiments do not reflect how and why the hyperparameters (length of prefix vectors, bottleneck dimension) of parameter-efficient tuning are selected, which is important for the discussions and conclusions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|17", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "For example, if with tuned hyperparameters, the ID results of parameter-efficient tuning methods could be close to full fine-tuning while preserving good OOD performance, then the proposed method makes no sense any more.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|19", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Prefix-tuning prepends tunable vectors to the projected key and value vectors instead of the input.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|20", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Another question is, the reported ROUGE-2 score on XSUM is quite low (21.2) on ID compared to the original reported result in BART paper (22.27) on the full test set.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|21", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "I would guess the ID performance is even higher or at lease to this number. Why is it?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|23", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "The motivation of this paper is not well-defined.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#1|24", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "The goal of this paper is to improve the OOD performance of a full-finetuning model, however, it didn't compare with any method on OOD generalization / domain adaptation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#2|8", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "This paper addressed interesting research topic and the proposed method is interesting and promising.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#2|9", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "This paper is clear and therefore it is easy to follow and to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#2|10", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "The analyses are interesting and provide insightful information.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#2|14", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "The experimental results are mixed, i.e. it is difficult to draw strong conclusions from the results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#2|15", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Furthermore, it is not clear where the results are significant different among systems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#2|16", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "The definitions of ID and OOD are not well described in this paper and need to be improved.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc93|review#2|17", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Furthermore, it is not clear how many examples of ID and OOD were used in the datasets.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc93|review#2|19", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Overall, I rate this paper as marginally above the acceptance threshold, mainly because the idea is interesting and somewhat novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#2|20", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "However, there are some weaknesses esp. in the experimental setup and results that make this paper a borderline paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|6", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "The current paper looks perfect up to section 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|7", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "It properly discusses different fine-tuning schema as well as the trade-off between OOD performance and ID performance with respect to the use of a fune-tuning scheme.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|8", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "It properly defines the problem (i.e., the trade-off) and proposes a simple yet effective way to overcome it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|9", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "It tests its method on three different NLG tasks: Table2Text Generation, Summarisation, and QA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|10", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "However, starting from section 5, it makes me hard to recommend an acceptance for this paper in the current format.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|11", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "This is because of two major concerns.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|12", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "One is that the interpretations of experimental results are problematic, weight might make the rest discussions and conclusions do not stand anymore.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|13", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "In section 5, regarding Table 2, the authors say \"when evaluating ID, full fine-tuning achieves the stronger results, whereas lightweight fine-tuning achieves only 81% of the skyline on average\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|14", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "This statement is only true \"on average\" and overlooks the huge differences between different tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|15", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "For example, for ID data of the Data2Text task, full fine-tuning achieves a score of 63.25 while lightweight fine-tuning achieve 63.18.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|16", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "There is no significant difference between the performance on ID of these two schema.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|17", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Meanwhile, in a similar vein, for OOD samples of the summarisation task, I also found no significant difference between the two schema.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|18", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "The situation described by the author only exists when doing question answering.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|19", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "This suggests that such a trade-off is task-dependent to a large extent and it seems to me that if a task relies on the inputs in a more directed way (e.g., QA is often accomplished by simply copying text from inputs), then the trade-off is more significant and the proposed model is more useful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|20", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Whereas, if a task relies on the inputs in a more indirect way (e.g., Data2Text requires the generator to plan what to produce in the first place), then the trade-off is less significant and improvement made by the proposed model is less.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|21", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Anyway, such a phenomenon should not be overlooked and covered by only saying \"on average\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|25", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "I generally like the idea of this paper and the paper is perfect up to the place where the experiments are introduced.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#3|26", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Discussions of this paper overlook major phenomena in the results, making the discussions and, probably, the conclusions are, in part, wrong.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#4|5", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "I am not convinced and even surprised by the authors that \"ensembling is one of the methods that we propose as well\".", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#4|6", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "I think the ensemble method is existing and NOT proposed by authors, and authors only apply it to the setting lightweight fine-tuning and full fine-tuning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#4|8", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "The idea of combining adapter-finetuning and full-finetuning to handle both in-domain data and out-of-domain data is interesting and clearly motivated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#4|9", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "This paper is well written and presents their motivation clearly and finally also provides theoretical analysis on multi-class logistic regression to help understand why it can work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#4|12", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Although cocktail fine-tuning only does inference once, it's not free and needs a knowledge distillation process before training, which requires inference over whole-training data using the adapter-finetuning model and can be time-consuming.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc93|review#4|13", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "The authors should talk about this limitation in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#4|17", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "It's also well written and clearly motivated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#4|18", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "Further, it provides a theoretical analysis of multi-class logistic regression to explain why it works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#4|19", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "However, I think the authors overclaimed their contribution.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#4|21", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "In addition, the authors mention that \"ensembling is one of the methods that we propose as well\", which I don't agree with.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc93|review#4|22", "source": "rc", "split": "test", "paper_id": "rc93", "comment": "I would recommend declining this paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|6", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "While we have seen a number of interesting papers focusing on debiasing static word embeddings, debiasing (even measuring) contextualized word embeddings is underexplored.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|7", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "As pretrained text encoders become more and more powerful, this paper is approaching a very challenging and important problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|8", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The way that the proposed method to correct the attention scores is smart.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|9", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "More specifically, it attaches a sequence of words from different social groups to the original training sentence, which makes it very controllable to adjust the attention scores assigned to specific words.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|10", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The proposed method also has an advantage of reducing multiple types of biases (e.g. gender, race, age, ...) simultaneously and there can be more than 2 groups for certain biases (although in Table 1, examples are all binary).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|12", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The experimental results, especially in Table 4, are not very convincing.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|13", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The proposed method seems only work well on reducing race bias.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|14", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "Gender bias seems to be the hardest one to mitigate across all methods and Kaneko is doing better on reducing religion bias.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|15", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The paper also need to provide more details on hyper-parameter tuning, e.g. lambda in Eq (3).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|16", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "And also more details and ablation studies on negative sampling and layer selection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|17", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "I am not fully convinced about the motivation (intuition) in Sec. 3.1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|18", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "Figure1 definitely presents some good examples of bias in encoders but it is hard to conclude that bias is mostly from the encoder.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|19", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The paper needs more quantitative analysis on bias in encoders and also decoders to support this claim.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|20", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The paper mentioned that bias measurements like WEAT have been questioned.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|21", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "However, Stereo-set and Crows-Pairs have also been criticized [1].", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|22", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "5.I am wondering that the positions of words in s_g may also affect the attention scores.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|23", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "I may have missed something here but has any analysis on the relative positions has been done? E.g. \"man\" comes first or \"woman\" comes first.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|24", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "If not, it would be interesting to see if we need to randomly change the order of words in the tuples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#1|25", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "[1] Stereotyping Norwegian Salmon: An Inventory of Pitfalls in Fairness Benchmark Datasets", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#2|5", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The debiasing based on attention is a new method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#2|6", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The empirical results seem good for likelihood based and inference based evaluation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#2|8", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "While the work has its merit in decreasing the biases in some metrics, It is bad using those dataset as leaderboard without looking into them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#2|9", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "[1] elaborates on the few problems with CrowS-Pairs and StereoSet.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#2|10", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "In [2], it criticizes all the work focusing on word level debiasing and those are really not removing them, but hiding them.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#2|11", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "To me, this paper is no different from prior work in this aspect.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc94|review#2|12", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "Also given most of the differences are quite small, I am not sure whether the paper is making meaningful progress.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#2|14", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "How does it handle multiple token words, different synonyms?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#2|15", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "Also, as described in section 5, it is not scalable for words that need not be debiased.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#2|16", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The proposed approach is also not very novel.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc94|review#2|17", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "It is an extension of counterfactural data augmentation in the attention with regularization with original attention.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|10", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "For evaluation, the paper shows the performance of several transformer-based models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|11", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The evaluation of the method includes three different benchmarks: two benchmarks that measure bias intrinsically (Crows-Pairs and StereoSet) and an NLI task designed for measuring bias.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|12", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "To make sure that the semantic strengths of the model are not lost, the models are also evaluated using the GLUE tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|13", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The model obtains competitive performance for the GLUE tasks, while reducing bias compared to original models and a couple of related works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|15", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The method for bias mitigation introduced by the paper is intuitive and relatively simple.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|16", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "In particular, I liked the idea of sample augmentation as opposed to trying to identify parts of the original text that refers to social groups and trying to alter the weights attending to those tokens, which is much higher to achieve.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|18", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "The method requires a teacher, unaltered model to be trained in parallel, which can be computationally involved.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|19", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "I would have liked to understand how much the performance of the model (perhaps on the GLUE tasks) is changed if such a teacher model is not used.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|20", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "I am concerned about the evaluation of bias.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|21", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "As recent studies showed, intrinsic measures of bias do not necessarily correlate with bias measures on a concrete, down-stream task that can be used in an application.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|22", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "https://arxiv.org/abs/2012.15859 The paper includes the NLI task, however, I perceive that task very similar to the intrinsic tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|23", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "It doesn't provide insights in what it would happen wrt bias in a down-stream task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|24", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "I would have liked the evaluation of the model to include a task for which group annotations exist and the bias for the original and modified model are presented.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|25", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "For example, the Jigsaw dataset could be used for such an evaluation (https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/data), or, for a smaller dataset, HateXplain could be used (https://github.com/hate-alert/HateXplain).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|26", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "Without such an evaluation for an application oriented task, it is hard to understand whether the few points gained by the new model will lead to significant bias improvement in a downstream task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|27", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "You may want to take a look at this paper that lists some of the problems with benchmarks such as StereoSet: https://www.microsoft.com/en-us/research/uploads/prod/2021/06/The_Salmon_paper.pdf", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|28", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "I am not sure if I understood the need to separate the original input from the artificially constructed one.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc94|review#3|29", "source": "rc", "split": "test", "paper_id": "rc94", "comment": "Do you think that the model will get confused if no separated is used and, thus, its semantic performance would decrease?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|11", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "In-context learning is a hot, rapidly evolving area of research and of high interest to the ICLR and NLP communities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|18", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "The idea of automatically constructing graphs of unlabeled examples using similarity measures is relatively old in the label propagation literature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|21", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "The main weakness of the paper is the lack of a thorough description and analysis (including ablations) of the vote-k algorithm and its relationship to existing clustering algorithms.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|22", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "I would have hoped for a more complete discussion in Section 2.1, including the contents of Appendix G which seem important for readers to fully understand the proposed vote-k algorithm.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|23", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "One particularly confusing part of vote-k: it was not immediately clear to me how this differs or is better than simple k-means clustering to discover centroids?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|24", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Once a weighted graph is constructed (as described in Section 2.1, using cosine similarities of sentence-level representations), it should be possible to run any graph-based algorithm for finding centroids.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|25", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "I think a naive baseline (that does not use model prediction confidence scores) would be to use k-means.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|26", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Related to the point above, I would have liked to see more ablations on the vote-k selection process. How important is it to use the model confidence “stratification” strategy in the algorithm? This is particularly useful to know since, given a very large set of unlabeled examples |U|, it can be costly to run all u in |U| model predictions to collect confidence scores.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|27", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "One source of bias that was not addressed in the paper relates to the choice of unlabeled examples (|U|).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|28", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "In this study, the unlabeled corpus for all 10 datasets has been extracted from an existing labeled/supervised corpus (that has been previously annotated).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|29", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "This may not seem like it, but I believe there is huge amount of supervision and work simply choosing which examples to label in the first place.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|30", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "In practice, however, for a new task/dataset, this will not be the case if we want to do only “selective annotation”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|31", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "I was quite unsure (maybe a little skeptical) that the proposed procedure would work for a “random” or uncurated corpus of unlabeled examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|32", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "There is also no discussion on label distribution in the vote-k process.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|33", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Again, I think the current experimental evaluation is taking advantage of curated training datasets where labels tend to be nicely distributed/represented.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|34", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "So focusing on performing input similarity and model confidence may be enough.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|35", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "For example, for some tasks, a fairly random unlabeled corpus may contain a huge number of “negative” examples (say you want to automatically find pairs of sentences for entailment - most random pairs of sentences will be NEUTRAL).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|36", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Measuring similarities of inputs alone does not seem to be sufficient to match the expected posterior distribution of labels?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|37", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Can you please elaborate on this point?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|38", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "With respect to the experimental setup, it is a bit jarring the number of combinations of (dataset, model) pairs (Table 1).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|39", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "I understand that it is expensive/infeasible to run all datasets on all models, and I don’t have a good suggestion to improve on this.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|40", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "But I fear that the end result is that subsequent/future work that builds on vote-k may need to re-evaluate your implementation on a partial set of dataset/tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|42", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Question about potential clarification in initialization of Algorithm 1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|43", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Maybe I did not understand this part of the algorithm, but when running initially |U| contains all examples, and |L| is empty.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|44", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "My understanding is that the score in line 4 of Algorithm 1 (also equation in Section 2.1) generates score of 0 for every node since no u in |U| are yet connected to a node in |L|.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|45", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Is the argmax using some stable ordering when all scores are 0 to guarantee determinism?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|46", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Does this choice not influence the final set of examples in |L| ?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|55", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "The novelty of this work is mixed: the idea of using retrieval for in-context learning prompts is not new (this work cites Liu et al., 2022 and Rubin et al., 2022)..", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|56", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "The idea of constructing graphs of unlabeled examples is also not new (common in label propagation literature).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|57", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "But novelty here is (1) in combining these ideas to make a practical system, and (2) the experimental work demonstrating some insights comparing this with fine-tuning and other selective annotation schemes.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|59", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Given that the code for vote-k be released, there should be no impediments to reproducing the results presented in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|63", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "I do think there are some unanswered questions that I list above, including (1) whether this scheme would be robust to a truly unlabeled or uncurated corpus, and (2) the relationship of vote-k with simpler k-means and other clustering algorithms.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#1|64", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Given that these issues do not seem overly problematic, I’m recommending the acceptance of this work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|14", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Weaknesses: Some of the design choices need to be elaborated on further and additional analysis-based experiments would also be useful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|15", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "I elaborate on these further below for the authors to address.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|16", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "For the classification tasks, what do the label distributions look like over the labeled subset L?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|17", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Since the selective annotation is based entirely on similarities derived from sentence embeddings, there is nothing explicit ensuring that the label distribution over the selected subset is not skewed.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|18", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Or, is the label distribution on the annotated subsets derived via vote-k indeed skewed for some tasks and the performance improvements are mainly coming from improvements on the labels that are well-represented?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|19", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Some discussion of this would be useful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|20", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "It seems like the optimal value of k in vote-k would depend on the number of instances in the unlabeled set that changes with the tasks. A single value of k=150 was chosen for experiments across all tasks. Could the authors justify this choice further?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|21", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "From the diversity and representativeness measures shown in Table 10 in Appendix F, the difference between Random and vote-k does not appear to be very large.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|22", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "In the least, it's unclear how to assess the differences shown in this table.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|23", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "A qualitative analysis might be more revealing here.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|24", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "It might be interesting to see some examples, especially when the annotation budget is 18, of the kinds of instances that get selected depending on the task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|25", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "t-SNE plots of these selected examples in the larger context of unlabeled instances might also be a good visualization to show.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|29", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "The proposed technique has been empirically validated well and code for this work has been made available to encourage reproducibility.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|31", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "I have rated this submission as a marginal accept for now.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#2|32", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Look forward to the authors' responses to my questions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#3|12", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Do the authors take the cost of encoding every unlabeled training instance using Sentence-BERT into account when comparing methods?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#3|13", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "I believe this was not mentioned in the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#3|14", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "In section 2.1, It is not very clear how the first M/10 examples are labeled. It is mentioned that “the current labeled L has M/10 samples” which is used to label other instances. Are they manually labeled?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc95|review#3|15", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "It is stated that the vote-k method encourages diversity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#3|16", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Since all the labels are removed first, does this method encourage or keep inclusivity of all the classes?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#3|17", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Have the authors looked at whether their method favors examples from certain classes?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#3|18", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "In section 2.1, “we conduct experiments three times and report the average score”. What/where is the randomness? Examples selected or model prediction?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc95|review#3|20", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "In section 4.4, I think some more detail is needed as to why the combination of vote-k and similarity retrieval works, but vote-k alone with random selection of supporting examples has almost the same score as random selection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#3|25", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "The authors motivated this approach well, and a detailed analysis is carried out.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#3|26", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Some parts about the proposed method are not clear, and there are a few missing and potentially misleading pieces in the paper.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc95|review#4|8", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "Weaknesses: Some important alternatives and baselines are not involved in the comparison.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#4|13", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "I do not think the proposed method is novel, though it is effective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#4|14", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "It is an ensemble of existing techniques.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#4|17", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "The method is simple yet effective, and the insights from the experiments are interesting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#4|18", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "However, I have a few concerns: Regarding to comparison: (1) can you compare vote-k with the following alternative?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#4|19", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "(a) replace vote-k with random selection; (b) run PLM and obtain confident scores; and (c) select annotations according to the confident scores.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc95|review#4|22", "source": "rc", "split": "test", "paper_id": "rc95", "comment": "From the description in Page 3, it seems that the remaining 9M/10 examples are selected according to the confident scores, but Line 11-16 in Algorithm 1 has nothing to do with the confident scores.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|1", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "This paper provides an argument for reconsidering how we construct a single training example during LM (pre)training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|2", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "It gives a formal proof that dependencies between two texts are vanishingly weak when the texts are presented in separate training examples (but not when they are in the same training example).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|3", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Intuitively, in order to make cross-textual inferences between texts in different examples, a model must be able to access information stored in its weights, but commonly used learning rates and model architectures ensure that the updates to the weights are too small to recover this information.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|4", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The paper suggests remedying this problem by changing the way in which we construct training examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|5", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Rather than obtaining examples by separating texts into contiguous chunks, we can construct examples from multiple semantically related chunks found by computing the similarity of text embeddings in the training corpus.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|6", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Empirical results are presented which show that task-adaptive (i.e.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|7", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "post-pretraining) training on such data leads to greatly improved performance on zero-shot semantic similarity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|8", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Further results show that doing pretraining with these kinds of examples improves zero-shot QA performance by a small margin (from an extremely low baseline).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|10", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The main results of this paper are of broad interest.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|11", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Questions about how the nature of the input to LMs affects learning are important and poorly understood.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|12", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The claim that the within- and between-example is crucial is novel and thought provoking.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|13", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The arguments are highly technical and require a lot of expertise, (including familiarity in particular with Levine, 2020, on which the proof builds substantially) to understand.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|14", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "This is not necessarily a strength or a weakness---there is great value to developing techniques for making this kind of formal argument.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|15", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "However, I find it difficult to follow even the thread of the argument at an intuitive level, which is a shame because it's interesting to me.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|16", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I suspect I'm not the only potential reader who will have this problem.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|17", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The paper could reach a much wider audience (which I think would be an improvement) if it provided more intuitive paraphrases of the main results in Section 2.3.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|18", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "There are a few unaddressed issues with the authors' main suggestion: that we can improve LMs’ abilities to make cross-example connections by constructing examples out of interleaved texts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|19", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Assuming a fixed example length, this means portions of text from the same source that otherwise would have been in the same example in the standard chunking scheme would now be separated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|20", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "However, it seems that, in general, dependencies within a text are more important than dependencies between random semantically related sentences from different texts.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|22", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I also have some doubts about the empirical results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|23", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The results about kNN-task adaptive pretraining seem to have a confound: The method for retrieving data for this intermediate training step involves finding sentences with high similarity, but the task on which this method is shown to lead to improvements is itself a semantic similarity task. The claim is that the method should be helpful for many kinds of tasks, but my intuition is that this \"coincidental\" alignment of the data collection and the evaluation task makes the result less generalizable.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|24", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The empirical results about zero-shot QA don't help much because the task is clearly too difficult.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|25", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The improvement on the models trained with the new method is from F1 of ~0.001 to ~0.01.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|26", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Do larger LMs succeed at this task?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|27", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "If so, then it seems that models tested in this paper are just too small or ineffective for the task.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|28", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Is there an easier task where this method leads to strong performance but ordinary pretraining does not?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|31", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "These kinds of arguments are valuable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|32", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "However, the argument is difficult to follow for a non-specialist audience.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#1|33", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Also, the empirical results the authors report have some confounds due to questionable task selection.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|1", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The paper addresses the bias induced by the inconsistency between the pretraining examples and downstream examples.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|2", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Specifically, the text is continuous during pretraining while could be non-neighboring in downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|3", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The authors proposed two methods: kNN-TAPT and kNN-pre-training in Task Adaptive Pre-training (TAPT) step or general Pre-training step respectively, where each training sample is composed of a sentence and its semantically closed neighbors (by kNN search).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|4", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The method is effective in sentence similarity tasks (in kNN-TAMP scenario) and closed book open domain QA tasks (in kNN-Pre-training scenario).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|5", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The paper also gives a theoretical analysis of the so-called in-context bias, by quantifying the NLM’s ability to model dependencies between two sentences that appear in the same training example (the in-context representation) and in different training examples (the sequential representation).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|12", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "1.The introduction of the motivation (the concept of in-context bias) is not easy to understand at the very beginning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|13", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The paper said: “the pretrained NLM can model much stronger dependencies between text segments that appeared in the same training example, than it can between text segments that appeared in different training examples.” Acutally it seems quite natural for me and I did not realize it is a problem until I saw more explanations in section 1.1.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|14", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "2.The theory is a bit complicated and not easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|15", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "3.The experiments are limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|16", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The authors only conduct the evaluation on sentence similarity tasks and open domain QA tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|17", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "However, there are many other tasks that involve sentence pairs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|18", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "For example, sentence inference tasks such as MNLI and RTE are common tasks in NLP field.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|19", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The authors should conduct experiments on more types of sentence pair tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|21", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The paper addresses the inductive bias of in-context learning.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|22", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The studied problem is meaningful and could be a potential breakthrough for pretrained language model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#2|23", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "However, the written of the paper is a bit hard to follow and the experiments is somewhat limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|1", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The authors show empirically and theoretically that transformer based language models cannot effectively exploit dependencies between sentences in different inputs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|2", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "For the theory part, they introduce a new measure called \\eps-sep rank.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|3", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "A function is said to have low \\eps-sep rank if it can be approximated up to error \\eps by a function with low separation rank.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|4", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Bounding the measure shows that if learning rate is low, the information from different training instances cannot be well integrated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|5", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Empirically, they present experiments where the task specific pre training was altered to include non-task specific similar sentences in the context of the LM.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|6", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "They show that augmenting context of pre training inputs with related similar sentences from Wikipedia helped improve end performance on similarity and zero shot open domain QA tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|12", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "1. I think the presentation of the paper needs to be improved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|13", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "For the theoretical analysis, it was not clear to me what is the contribution of the current analysis compared to the Levine 2020 paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|14", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I felt similar conclusions can be drawn from the results of that paper as well.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|15", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "It will be good to rewrite highlighting the contributions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|16", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I also noticed a lot of repeated text in section 2 from Levine 2020 paper; will be good to modify.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|17", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "2. The empirical part of the paper shows improved performance of adding similar sentences to the context of LM training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|18", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I felt this was quite separate from the theoretical analysis. It does not follow from the theoretical results that adding similar sentences will be a good thing. As a result, having these in the same paper looked incoherent to me.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|19", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "3. The experimental results are weak.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|20", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The gains are not super high.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|21", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I will be more convinced if evaluation is done on a wider range of tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|23", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I was not clear about the impact of the theoretical analysis presented, compared to past work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|24", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "It will be good to rewrite the section clarifying the contributions and novel conclusions from the analysis.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|25", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The empirical section sounded disparate from the theoretical part of the paper, and the results are weak.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#3|26", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I suggest evaluating on a wider variety of tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|1", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The paper presents a theoretical analysis of the strengths of including similar input examples S1 and S2 into the same actual transformer input.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|2", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The analysis contrasts using examples of concatenated S1;S2 with running the two in separate inputs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|3", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The paper then proves empirically that including related examples in the same input (via a method the article calls KNN-pretraining) allows transformers to learn much faster in cases when cross-document dependency is relevant.", "reviewer_type": "human", "label": "error", "error_type": "Inaccurate Summary", "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|7", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The theoretical contributions will have a high impact on training transformer-based models", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|8", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The theoretical analysis is supplemented by experimental analysis", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|12", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "correct on less than 50 questions out of 20, 000 in the evaluation set – Separating thousands by commas (American system) is ok.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|13", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Separating them by space is also ok.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|14", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Doing both is not.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|15", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "correct on roughly than 250 questions in the evaluation set – \"than\" shouldn't be there", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|16", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Equation 4 would have been easier to read if the right-hand side was expanded to S1 and S2's {w_2^j} notation from the previous line.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|18", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Despite its theoretical focus, the paper is easy to read and to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#4|19", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The theoretical analysis is supplemented by empirical results.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|1", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "This paper formalizes the notion of an in-context bias---the pre-trained neural language models are better at modeling dependencies that appear within one contiguous pre-training text chunk, rather than those that appear in different pre-training text chunks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|2", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The authors perform a theoretical analysis of this phenomenon and link the inability to model dependencies between different examples with the models' low-pretraining learning rate.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|3", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The authors then propose two methods for improving pre-training example design by adding sentences that are similar (KNN), but from different documents.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|4", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "They empirically show that this works better than reasonable baselines on sentence similarity tasks (e.g., STS and SICK), as well as NaturalQuestions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|6", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I thought that this paper was very thought-provoking, and I appreciated the attempts to better understand what is going on with pre-trained language models, why they work well, and what might we be able to improve from theses insights.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|8", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Thorough theoretical analysis that reveals the connection between (practically-necessary) small learning rates and inability to use dependencies across text chunks", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|9", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Useful framing and discussion of the \"in-context bias\", where models are more likely to learn dependencies within text chunks seen during pre-training.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|10", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "reasonable initial experimental results demonstrating some ways to help models better use cross-text-chunk dependencies (put them into a contiguous text chunk), providing some hope that these results could make models better.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|12", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I felt like the empirical validation could have been stronger.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|13", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "The only two tasks examined are sentence similarity tasks (which seem a bit more like a sanity check), and NaturalQuestions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|14", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "It's particularly surprising to me that this works so well on NQ, and I wish the authors had dug a bit deeper into this, but I also recognize that page limits exist.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|15", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Where else do you think the in-context bias could be useful?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|16", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "It wasn't intuitive for me that it'd be useful for NQ.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|17", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "Do you have any initial experiments on the \"self-improving\" aspect of this technique?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|18", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "This is mentioned several times, but there are no initial results or anything suggesting that it might be a promising direction to pursue.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|20", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I thought this paper was interesting and brings a new perspective to practical design decisions used in pre-training language models.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc96|review#5|21", "source": "rc", "split": "test", "paper_id": "rc96", "comment": "I think this paper is relevant to the ICLR audience, and that said audience would be excited to know about it.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|9", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "Here are a number of questions for the authors:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|10", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "Your TopOnly setting is supposed to mimic how the encoder-decoder model uses the topmost-layer source encodings for translation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|11", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "This setting indeed shows improvements and achieves lower perplexity on almost all language pairs.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|12", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "One can interpret that two fundamentally different architectures are still bound by similar features.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|13", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "Did you investigate why the topmost-layer is the most informative way (at least in the tried approaches) of transferring information from the source to the target generation?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|14", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "Following on that, looks like that the difference between the Deep and TopOnlyDeep variations is very small.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|15", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "What is your understanding on the difference of information captured by deeper and multiple layers versus deeper and topmost layer?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|16", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "Did you compare Deep and TopOnly setups?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|17", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "This question is mostly about the PrefixLM model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|19", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "You conclude in one of the sections that \"When there are fewer parameters, the model with inductive biases favoring translation will achieve better quality\" and follow it with four observations two of which I want to discuss here. It is not entirely clear to me why \"deeper LMs (Deep) rather than wider\" and \"training LMs without source-side language modeling loss (TgtOnly)\" are both considered as high inductive biases. For the former, how is deeper models have more inductive biases than wider models? For the later, wouldn't the other way around be correct?", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|20", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "You observe that the PrefixLM model performs surprisingly well on zero-shot directions.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|21", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "Do you have any intuitions or have you looked into finding why that's the case?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|22", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "According to your experiments, it looks like CausalLM (unidirectional) model is not performing well, or at least as good as PrefixLM or EncoderDecoder models, in almost any settings.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|23", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "This is intuitive because the model has access to less information at each step.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#1|24", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "I'm wondering if you can think of a scenario where such a model will be more effective.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#2|5", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "This is an OK paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#2|7", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "However, after a careful reading of the paper, I still could hardly find enough insights from the paper.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#2|8", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "How could the findings benefit the community? The design principle of EncDec-based NMT is the sub-optimal performance/efficiency of LM-based NMT, and this paper just says: \"yes, the intuition is true\", which is not exciting at all.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc97|review#2|10", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "It would be nice if the authors could discuss the applicability of the findings in the author response and the future version.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#2|11", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "And better yet, directly utilizing the findings to enhance existing NMT systems.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#2|14", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "This is an OK paper but the current version is not exciting, which could hardly attract attention from the community.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc97|review#2|15", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "The authors should answer the question: how could the paper findings benefit the community?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#3|9", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "Despite good empirical efforts, the motivation of this paper seems somewhat unclear.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc97|review#3|11", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "I feel like the authors didn't convey an incentive for this.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc97|review#3|12", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "The paper conducted extensive experiments to show how scaling affects LMs for MT, however, few suggestions based on the findings are given for future development of MT.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#3|15", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "Section 5 seems to mainly examine p and L_{\\inf} In equation (7), whereas \\alpha remains undetermined. How was the value of \\alpha determined for diagrams in Figure 3? Get fitted from Figure 2 I guess?", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc97|review#3|20", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "I tend to consider it as a borderline paper but are willing to change my mind if I did miss something upon the author response.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#4|12", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "Weaknesses: Some experiments require more careful design to make findings more reliable.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#4|13", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "Such as, in Figure 2(d), the Test Bleu of EndDec, CasualLM-Deep, and Prefix-TopOnly-Deep decrease as the parameters increase in the interval of #Params 0.19-0.43E8.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#4|14", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "This phenomenon is not in line with the general trend, also the intuition.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#4|15", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "I suspect that the experimental settings could be improved.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#4|16", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "Othervise, the authors need to give a reasonable explanation.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#4|17", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "There is another question I would like to promote a discussion.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#4|18", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "That is how to measure the transferring capabilities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#4|19", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "The conclusions drawn from Figure 6 is that LMs do not really facilitate the transfer to low-resource languages, because LM-based models do not outperform EncDec model in all #params scales.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#4|20", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "However, we cannot see the improvement each model gain through transferring over the baseline.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc97|review#4|21", "source": "rc", "split": "test", "paper_id": "rc97", "comment": "These relative improvements over the baseline (not the absolute scores) may be more appropriate to measure the transferring capabilities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#1|10", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "The idea of using entity abstraction is straightforward and not very novel as doing POS tagging for input is widely used in many NLP models, especially in traditional NLP pipelines.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc98|review#1|11", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "While the performance on synthetic datasets is promising, on realistic datasets the proposed method fails to improve the performance.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc98|review#1|12", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "Therefore the effectiveness of the method is not fully supported by the experiments.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc98|review#1|17", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "Therefore I believe this paper is below the bar.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#2|9", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "The paper only shows end-to-end performance on the three tasks without deeper analysis on why entity abstraction does not always improve the performance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#2|10", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "Like the authors pointed out, the current conclusion could be due to the noisy entity type predictions or the fact that the tasks do not require entity abstraction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#2|11", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "Without the deeper analysis, readers cannot generalize the conclusion beyond the three datasets mentioned in this paper to other unexplored downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#2|12", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "Only one pretrained transformer model t5-small is tested, therefore the empirical findings of this paper cannot be applied to other models, which usually contain much more parameters and perform better on various downstream tasks.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#2|15", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "As the related work section discussed, there are a lot of other related works also incorporating knowledge from named entities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#2|16", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "To me, using only the entity type is marginally novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#2|17", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "Also, the empirical findings lack deeper analysis and were merely conducted on a single and toy-sized pretrained language model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#2|18", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "For empirical results, the only significant improvement is on a synthetic dataset while the gain on real-world datasets is also marginal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#2|19", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "Therefore, the significance of the paper's contribution is limited.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#3|8", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "The experiment part lacks analysis on the model which could provide more intuitions on how the model works.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#3|9", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "More analytic experiments will bring more interesting findings of the proposed model.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#3|10", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "For instance, we could dive deep into why the model variants perform differently between CLUTRR and HotpotQA & CoQA.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#3|13", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "It will be of great help for revealing the intuition of the model if introducing more analytic experiments.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|11", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "My major concern is about the experiments. Specifically,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|12", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "There is very minor boost in performance on QA datasets.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|15", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "The performance of baselines are far below the current top performers.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|16", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "I understand that the paper is not interested in achieving new state-of-the-art, and the test set used might be different.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|17", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "However, in that case you can not prove that the proposed methods are still beneficial on a model that achieves way better baseline results.", "reviewer_type": "human", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|18", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "An analogy here is that something working under low-resource setting is not necessarily effective when there is adequate access to data.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|19", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "The method of incorporating abstraction is not universal.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|21", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "If there is a new reasoning dataset, which architecture in Table 1 should people use?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|24", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "In Section 4,1, it is mentioned that “the inverse relation of “father” can be “son” or “daughter”, we accept both”. What does “accept both” mean here? “Ann is the daughter of Brett” and “Ann is the son of Brett” can not be both correct. Right?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|25", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "Also in Section 4.1, “We generated 390,000 examples that were roughly split 70/30 between training and testing”.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|26", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "But in Section 4.1.1, the size of test set becomes 10,000.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|27", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "Which one is correct?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|28", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "The lvl.2 output of CLUTTR in Appendix B does not use the answer template specified in Section 4.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|29", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "TACRED has been mentioned multiple times in the Appendix.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|30", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "How is TACRED used in this work?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc98|review#4|33", "source": "rc", "split": "test", "paper_id": "rc98", "comment": "I do not recommend acceptance of the paper in its current form.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|7", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The concept of the paper is nicely presented for the next AI abundant society where demand for human judgments is more increased.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|8", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "They proposed the novel similarity approximation method with a smaller number of judgements.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|9", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "They provide the evaluation data in the supplemental material, which is beneficial.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|13", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The title In particular, the linkage between title and abstract seems not clear at the first glance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|14", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Please consider to update the title and abstract so that readers can easily understand the topic theme.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|15", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Is English best? There are many languages.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|16", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "In my opinion, I view that authors selected English language to represent the basis of all things in nature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|17", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Why English(e.g., language system such as polysemous, grammar, dataset availability…)?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|18", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Justification of English employment would make the paper better.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|19", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Related to above topic, the paper report “N = 1,492 US participants for the new behavioral experiments reported in this paper.”", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|20", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Are they all Americans and share the same contexts? (e.g., cultures…)", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|21", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Crowd scouring How long does it take to STEP-Tag? It is good that measure actual time that participants consumed. Comparing with that with the time used for writing captions may strengthen your claim.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|24", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Quality Good. Issues for human judgment cost have been solved with the proposed idea.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|25", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Clarity Good. In the experiment, model they used is sufficient.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|26", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Originality Good. The idea of methods themselves are not very novel.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#1|27", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "However, I believe the exhaustive amount of examination is valuable, which increase the originality, thus it should be highly evaluated.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#2|1", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The paper proposes a new class of similarity approximation methods based on language.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#2|2", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "To collect the language data required by these new methods, the authors also developed and validated a new adaptive tag collection pipeline, which is significantly cheaper than the classical methods.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#2|3", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Finally, the authors also develop ‘stacked’ methods that combine language embeddings with DNN embeddings, and find that these consistently provide the best approximations for human similarity across all the three modalities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#2|6", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The paper is overall well-written and easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#2|7", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Multiple datasets are evaluated for the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#2|8", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Multiple modalities are explored and evaluated on the proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#2|11", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "No baselines are compared to the proposed method in the experiment section.", "reviewer_type": "human", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|human|rc99|review#2|13", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The paper is generally well-written. The authors consider multiple modalities and datasets to evaluate their proposed method.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|1", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The authors make the observation that representations learned by DL models produce proximity scores very different from human evaluations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|2", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "They introduce a simple and scalable technique to make the human and model produced similarity scores closer.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|3", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Essentially, text descriptions or tags for various input data points (across modalities) are passed through LLMs (or word-frequency methods) and the resulting representations used for proximity scores.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|4", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Stacked representations combining the existing model and the text description representation is shown to consistently match human similarity metrics better.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|7", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "I think this is a well-motivated problem. Learned representations are often used as features in the small-data regimen or sometimes directly for getting proximity scores in an AI setting.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|8", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "This paper address the human interpretability of these representations by (i) confirming that human similarities and proximity scores from models can vary a lot,", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|9", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "(ii) text-descriptions or tags can be leveraged and stacking these representations with the model-learned representations can help.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|10", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "I also appreciate that the technique is scalable and in many cases not that much of an overhead to implement.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|11", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "I appreciate the arguments in the related text that leverage cognitive science literature.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|12", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "In addition, the paper is easy to follow.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|15", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "I was wondering if we could get some numbers on if the stacked representations help in additional downstream tasks like say classification (i.e. does the performance on imagenet improve if you use imagenet + text). However, I understand that this can be significant undertaking and do not want to base my review on this experiment but it is a potential future direction.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|17", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The paper presents a straightforward but well-motivated idea.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|18", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Having representations match human similarity judgement is indeed useful.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|19", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Implementing this technique would require (i) ability to acquire free-text or tag values which are domain-dependent but not prohibitively expensive, (ii) ability to obtain LLM representations which are straightforward since high quality implementations and libraries exist for these now. Reproducibility is not an issue.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|22", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Clearly described technique that is scalable, easy to implement", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|23", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Techniques like these that are easy to implement and help with interpretability are of great use in the small-data regimen (where the bulk of us are).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#3|24", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "I would like to see this paper at ICLR.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|1", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The authors consider the task of approximating human didactic similarity judgments over N pairs of (images/audio/text).", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|2", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "While high quality human judgments over all pairs is ideal, (N choose 2) is too big for large N.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|3", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The authors consider gathering N descriptions of the objects (tags or captions), extracting text representations of those descriptions, and then computing text-text similarity.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|4", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Experiments across a large number of models and datasets reveal the surprising effectiveness of computing text-text similarities.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|6", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Reconstructing pairwise human similarity judgments across N objects is an interesting task, and the author's idea of gathering textual descriptions is clever.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|7", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The performance of the stacked models (which depend on domain-specific pretrained models and text-text similarities) and text-only models are surprisingly strong across a variety of domains --- this finding has implications for the relative surprising representational strength of text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|8", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The annotation process they describe for their tag corpus is interesting --- multiple rounds of annotations are undertaken with crowdworkers fixing each-other's errors.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|9", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The authors consider a large number of models across many domains --- the benchmark experiments are extensive, and the \"size vs. accuracy\" plots in the end are cool.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|10", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Figure 5 is amazing --- few works offer such clear practical guidance.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|11", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Finally, I liked the hopeful message in the end --- that a combined multimodal approach seems to work best, so don't completely discard the domain-specific representations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|12", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "It should be noted that the cited work where some of the data is from, Marjieh et al. 2022, is quite similar in the sense that it also proposes to estimate human similarity judgments. This work appears to extend that work by exploring more types of models and their combinations.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|13", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "My main concerns are that this set of experiments suggests some clear next setups that I don't nesc. think are out of scope for this work. Specifically:", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|14", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "If the goal in practice is to reconstruct human judgments --- I would have liked the authors to compare against a supervised setup.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|15", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "How much of the remaining misalignments can be fixed by gathering n << N^2 additional pairwise judgments and then training a supervised model ontop of the models?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|16", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "I would have liked to have seen more experiments with even less supervision --- is it possible to gather n < N captions, and train a model on those very few pairs to map to text, and then use the resulting modality --> text model as input to the models shown here?", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|17", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "I would have appreciated more discussion of tags vs. captions for the images.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|18", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "In figure 3A, I assume that the text only methods use captions unless specified otherwise. But --- because the tags were not fed to, e.g., BERT/RoBERTa, I can't tell if the performance gains are due to the full captions being used, or if the gains are due to the more expressive models. A trivial linearization of, e.g., \"A photo of a tag1, tag2, and tag3\" handed to a LLM would have been a nice comparison to see.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|19", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Finally, I would have been interested in a fully LLM setup where instead of pairwise cosine similarities, both text pairs are fed to a very large LLM (e.g., OPT or GPT3) and the model outputs a likert rating as text.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|21", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "As I mentioned, the novelty is somewhat limited because of the prior human judgements work the authors get their data from.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|22", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "But, the extensive set of experiments is interesting/new.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|30", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "The authors seemingly ran all of the experiments I suggested and have incorporated them into the paper.Amazing :-)", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|human|rc99|review#4|31", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Furthermore, the authors clarified the difference with prior work.", "reviewer_type": "human", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|9", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "- Applying LLMs to HTML understanding is a novel and important problem with practical applications in web automation and retrieval.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|10", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "This is one of the first works to systematically study this.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|11", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "- The three proposed benchmark tasks capture key aspects of HTML understanding.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|12", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The tasks are well-motivated and the datasets used are suitable.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|13", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "- Experimenting with a wide range of state-of-the-art LLM architectures provides useful insights into which models work best for these tasks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|14", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "- Demonstrating strong transfer from LLMs to HTML understanding tasks, without HTML-specific architectures or training, is an important finding that can spur more work in this direction.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|15", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "- Open-sourcing a large automatically labeled HTML dataset from CommonCrawl is a valuable contribution to the community.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|18", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The bidirectional attention hypothesis is stated but not investigated in depth.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|19", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "- The snippet extraction procedure for handling large HTML pages has some ad-hoc aspects.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|20", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "A more principled approach to processing long HTML sequences would be good to explore.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|21", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "- More discussion of the potential limitations of repurposing LLMs for HTML understanding and ideas for future improvements would have strengthened the paper.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|24", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The task formulations, model architectures, datasets, and experimental methodology are all described in sufficient detail.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|25", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "- The technical quality of the work is high.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|26", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The experiments are comprehensive and the results are significant.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|28", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "- The datasets and model details are provided, facilitating reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|29", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Open-sourcing the CommonCrawl dataset is a big plus.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|31", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "This paper makes a valuable contribution by demonstrating the effectiveness of LLMs for HTML understanding tasks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|32", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The strong transfer results, without requiring specialized architectures, can serve as a new baseline and open up avenues for future work.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|claude_opus|34", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The new dataset is also a useful resource for the community.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gpt4|2", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Despite the significant capabilities of LLMs in various natural language tasks, their potential in parsing raw HTML has been relatively unexplored.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gpt4|9", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The paper addresses an under-explored yet critical area of leveraging LLMs for HTML understanding, relevant for web-based automation and information retrieval.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gpt4|10", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Demonstrates significant improvements in accuracy and sample efficiency over existing models by fine-tuning pre-trained LLMs on HTML understanding tasks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gpt4|11", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The comprehensive experiments across different models, tasks, and a novel dataset contribute valuable insights to the field.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gpt4|12", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Open-sourcing the large-scale HTML dataset distilled from CommonCrawl is a significant contribution that will likely spur further research.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gpt4|15", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "A comparison with state-of-the-art models in related fields, such as document understanding or web page segmentation, might have provided a broader context.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gpt4|19", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Quality: The research is of high quality, with rigorous experimentation and a significant dataset contribution.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gpt4|20", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Novelty: The approach of fine-tuning pre-trained LLMs for HTML understanding tasks is novel, especially the application across diverse tasks like semantic classification, description generation, and web navigation.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gpt4|21", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Reproducibility: With the detailed methodology, open-sourced dataset, and clear descriptions of the fine-tuning process, the study appears highly reproducible.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gpt4|24", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The research is significant for its potential applications in web automation and information retrieval.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gpt4|25", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The open-source dataset further contributes to the field's advancement.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|8", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Novel approach: This work is among the first to explore LLMs for understanding raw HTML without pre-parsing, offering a new perspective on web automation and understanding.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|9", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Strong performance: LLMs, especially T5 models, demonstrate remarkable performance across all tasks, achieving state-of-the-art results on the MiniWoB benchmark with significantly less training data.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|10", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Sample efficiency: Pre-trained LLMs require significantly less labeled data compared to models trained from scratch, reducing data collection costs and time.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|11", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Open-source dataset: The paper contributes a new large-scale HTML dataset for description generation, promoting further research in this area.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|13", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Limited context window: The paper highlights the context window length as a bottleneck for LLMs in handling real-world websites, which are often much larger than the pages used in the benchmarks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Snippet extraction impact: While necessary for longer pages, snippet extraction can lead to information loss and performance degradation.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|18", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Quality: The research is of high quality, employing rigorous evaluation methods and comparing against relevant baselines.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|19", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Novelty: While the individual components (LLMs, HTML understanding tasks) are not entirely new, the combination and the focus on raw HTML processing offer significant novelty.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|20", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "Reproducibility: The paper provides details about the datasets and models, and with the planned release of the code and dataset, the results should be reproducible.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|22", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "This paper presents a compelling case for using LLMs in HTML understanding tasks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|23", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "The strong performance, sample efficiency, and minimal pre-processing requirements offer a promising direction for web automation and related applications.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc1|gemini_pro_1.5|24", "source": "rc", "split": "test", "paper_id": "rc1", "comment": "However, the limitations of context window size and the impact of snippet extraction require further investigation.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|14", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "1.While the WebBrain-Raw dataset is very large, the paper lacks details on how the queries and articles are split/sampled to create the WebBrain-R and WebBrain-G datasets used for experiments.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|15", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "More dataset statistics and analysis would be helpful.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|16", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "2.The generator in ReGen uses a simple reference mark correction strategy in decoding, but a more principled approach to improve the accuracy of citing evidence during generation is needed.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|17", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The authors do discuss this as a limitation and future work.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|18", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "3.Generating complete Wikipedia-style articles is extremely challenging.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|19", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The current work focuses on generating the first section, which is an important first step but still quite far from the ultimate goal.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|20", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "More discussion on the gaps and potential roadmap would strengthen the motivation and impact.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|21", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "4.While the human evaluation results are encouraging, details are lacking on the specific criteria, annotation process, number of samples, and inter-annotator agreement.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|22", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "A larger-scale and more rigorous human evaluation would make the results more convincing.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|26", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The overall quality of the work is high, with the main contributions being the new large-scale WebBrain-Raw dataset and the novel factuality-enhanced ReGen framework.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|28", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Assuming the code and data are released, the experiments should be reproducible.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|31", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "While there is still a big gap to generating full Wikipedia articles, this work represents an important step in that direction.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|32", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The paper could be strengthened with more dataset details, deeper analysis on remaining challenges, and more rigorous human evaluation.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|claude_opus|33", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "But overall, this is a high-quality submission with notable contributions to the field.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|gpt4|12", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "The research quality is high, with thorough experimentation and a clear presentation of results.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|gpt4|18", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "While the paper is strong in most aspects, further discussion on its limitations and broader applicability would strengthen its impact.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|gemini_pro_1.5|12", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Retrieval quality: While ReGen shows improvement, the reliance on retrieved evidence still leads to performance below that achieved with gold references, highlighting the need for further advancements in retrieval quality.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|gemini_pro_1.5|13", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Reference mark accuracy: The current model struggles to accurately identify the source of information for generated sentences, impacting factualness and readability.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Limited article length: The current focus is on generating short articles, whereas full Wikipedia pages contain much more information.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|gemini_pro_1.5|15", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Generating longer, comprehensive articles remains a challenge.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc17|gemini_pro_1.5|24", "source": "rc", "split": "test", "paper_id": "rc17", "comment": "Future work should focus on addressing these limitations and expanding the task to generate longer, more comprehensive articles.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|claude_opus|15", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "- For label names without semantic meanings, NPPrompt still requires some keywords to work well.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|claude_opus|16", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "How to extend it to arbitrary labels is unclear.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|claude_opus|18", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "How to incorporate few-shot examples when available to further improve performance is not explored.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|claude_opus|20", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "Whether the technique generalizes to other languages is unknown.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|claude_opus|22", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The paper is clearly written and easy to follow.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|claude_opus|23", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The proposed NPPrompt method is simple and intuitive.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|claude_opus|24", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The extensive experiments are good quality, covering a diverse set of important NLP tasks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|claude_opus|25", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The key novelty is enabling an effective fully zero-shot learning approach for PLMs without relying on any labeled data, unlabeled corpus or human annotations.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|claude_opus|31", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "However, the code is not provided with the submission, so reproducing the full results may still take some effort.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gpt4|3", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The method outperforms existing zero-shot approaches by significant margins on standard benchmarks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gpt4|8", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "1.Reproducibility: While the paper mentions experiments and results, detailed settings for reproducibility (e.g., hyperparameters, computational resources) could be better emphasized.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gpt4|9", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "2.Comparison with Baselines: The paper could benefit from a deeper analysis of why NPPrompt outperforms baselines, including insights into cases where it might not perform as expected.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gpt4|10", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "3.Generalization to other languages or domains: The paper primarily focuses on English; extending and evaluating the approach for other languages or specific domains could provide further insights.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gpt4|14", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "Reproducibility could be enhanced with more detailed experimental settings and access to code or datasets.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gpt4|18", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "However, improvements in documenting reproducibility and extending analyses could further strengthen the paper.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gemini_pro_1.5|3", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "The paper evaluates NPPrompt on tasks like text classification, text entailment, and paraphrase identification, demonstrating significant performance improvements over previous zero-shot methods.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gemini_pro_1.5|8", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "Limited analysis for certain scenarios: The paper acknowledges that NPPrompt may require additional keywords for label names lacking semantic meaning.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gemini_pro_1.5|9", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "Focus on zero-shot setting: The study primarily focuses on the zero-shot setting, leaving the exploration of few-shot scenarios for future work.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gemini_pro_1.5|10", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "Limited task coverage: The evaluation primarily focuses on text classification and NLU tasks from the GLUE benchmark.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gemini_pro_1.5|16", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "Reproducibility: While the paper lacks code, it provides detailed descriptions of the methodology and experimental setup, facilitating potential replication by other researchers.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc23|gemini_pro_1.5|20", "source": "rc", "split": "test", "paper_id": "rc23", "comment": "However, further investigation is needed to address its limitations in handling labels with limited semantic meaning and to explore its applicability to other NLP tasks and few-shot learning scenarios.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc28|claude_opus|22", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "- Dataset leakage is a potential concern that could have been analyzed in a bit more depth, though the authors do provide reasonable arguments for why it's unlikely to be a major factor.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc28|claude_opus|27", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "The overall quality of research seems high, with extensive experiments supporting the key claims.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc28|claude_opus|33", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "In summary, this is a high-quality paper that makes a valuable contribution to the field of using LLMs for MCQA.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc28|claude_opus|37", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Overall, this paper has the potential to shift how LLMs are applied to MCQA tasks going forward.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc28|gpt4|11", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "- The paper could benefit from a more in-depth discussion of the limitations of the proposed MCSB approach, including potential biases introduced by symbol binding.", "reviewer_type": "ai", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|llm|rc28|gpt4|12", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "- Comparison with state-of-the-art methods focuses primarily on performance metrics, with less emphasis on computational efficiency and scalability.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc28|gpt4|19", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Summary of the Review", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc28|gpt4|22", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "While the paper is of high quality and novelty, further exploration of its limitations and a broader comparison with existing methods could strengthen its impact.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc28|gemini_pro_1.5|0", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Summary of the Paper:", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc28|gemini_pro_1.5|26", "source": "rc", "split": "test", "paper_id": "rc28", "comment": "Further investigation into the factors influencing MCSB ability and exploration of prompt engineering for MCP are promising directions for future work.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc99|claude_opus|18", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "- Discussion of potential risks (e.g.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc99|claude_opus|19", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "encoding human biases) is quite brief", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc99|gpt4|13", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Additionally, the performance gap noted between human judgments and DNNs might be model-specific rather than a general limitation of DNNs.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc99|gemini_pro_1.5|12", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Limited interpretability analysis: While interpretability is claimed as a key advantage of the proposed methods, the paper lacks a dedicated analysis to support this claim. Further investigation is needed to assess how interpretable the generated prompts and representations actually are.", "reviewer_type": "ai", "label": "error", "error_type": "Unstated statement", "label_origin": "expert"}
{"id": "rc|llm|rc99|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc99", "comment": "Lack of comparison with recent prompting methods: The paper does not compare with some recent methods for finding better discrete prompts, which would be helpful to gauge the relative performance of the proposed methods.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc29|claude_opus|12", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "- While the paper experiments with an extensive combination of models and methods, the justification for choosing those particular PLMs, CL methods and datasets could be elaborated more.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc29|claude_opus|13", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "- The layer-wise probing reveals useful insights, but the underlying reasons why different layers in a PLM demonstrate different remembering/forgetting behavior remains unclear.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc29|claude_opus|17", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "- The paper is overall well-motivated, clearly organized, and easy to follow.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc29|gemini_pro_1.5|7", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Detailed analysis: The layer-wise probing analysis offers insights into the forgetting behavior within PLMs and the effectiveness of different CL methods on different layers.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc29|gemini_pro_1.5|8", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Identifies open research questions: The paper highlights several important research directions for future work in PLM-based continual learning.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc29|gemini_pro_1.5|9", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Open-source resources: The authors commit to releasing code and datasets, which will benefit the research community.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc29|gemini_pro_1.5|16", "source": "rc", "split": "test", "paper_id": "rc29", "comment": "Missing comparisons: The paper does not compare with contrastive learning-based sentence-level pretraining tasks, which have shown promising results in CL.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|6", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The proposed method is shown to be orthogonal and complementary to other compression techniques like knowledge distillation.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|9", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "- Comprehensively compares different tensor decomposition methods for compressing PLMs, providing useful insights", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|10", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "- Achieves very high compression ratios (e.g.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|11", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "1/48 parameters) while maintaining most of the performance, demonstrating the effectiveness of the Tucker decomposition approach", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|12", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "- Proposes practical decomposition and reconstruction protocols that enable efficient compression and inference", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|13", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "- Shows the method is complementary to knowledge distillation, indicating potential for combining with other compression techniques", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|16", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "- Theoretical analysis of why Tucker decomposition works well for compressing Transformers is somewhat limited", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|17", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "- Does not explore the limits of compression - even smaller models may be possible with some performance trade-off", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|20", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The paper is clearly written and easy to follow.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|21", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The methodology is explained in detail and the experimental setup is reproducible.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|23", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The overall quality of research and presentation is high.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|27", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The comparison between decomposition methods is rigorous and novel.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|28", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "A few areas that could be further explored are evaluating on an even wider range of NLP tasks, providing more theoretical justification, analyzing impacts on pre-training, and pushing compression to the limit.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|claude_opus|30", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Overall, this is a high-quality paper that advances the important problem of making large language models more efficient.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|0", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Summary of the Paper", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|4", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The compressed BERT models, with significantly reduced parameters, achieve comparable or slightly better performance than the original BERT on the GLUE benchmark.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|5", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Additionally, a tiny version of BERT, with less than 2M parameters (excluding the embedding layer) and 2.7 times faster inference, achieves 96.7% of BERT-base's performance.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|6", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The paper also examines the synergy of the proposed method with knowledge distillation, indicating its orthogonality to existing compression techniques.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|9", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The paper tackles the significant issue of PLM size, making a compelling case for the need for more efficient, smaller models without compromising performance.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|10", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "It introduces an innovative application of Tucker decomposition for PLM compression, which is both novel and technically sound.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|11", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The empirical results are impressive, especially the achievement of a tiny BERT variant that retains most of the performance of the full model while being significantly smaller and faster.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|16", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Clarity, Quality, Novelty, and Reproducibility", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|17", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The paper is well-written and clear, making it accessible to readers familiar with PLMs and tensor decomposition techniques.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|18", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The research is of high quality, employing rigorous experimental methodologies and providing a comprehensive comparison with baseline models.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|19", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The novelty lies in applying Tucker decomposition to compress PLMs efficiently, which is a relatively unexplored area.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|20", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The experiments appear reproducible, with sufficient details on the implementation and evaluation settings.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|21", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Summary of the Review", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|22", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The paper presents a significant contribution to the field of NLP by proposing an effective method for compressing PLMs using Tucker decomposition.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gpt4|23", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Its approach not only preserves the performance of the original models on benchmark tasks but also improves inference speed, addressing a critical need for more efficient NLP models.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|3", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The authors argue that existing compression methods like knowledge distillation and quantization have limitations in achieving high compression ratios.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|9", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Comprehensive framework: The paper provides a well-structured framework for discussing and comparing different matrix/tensor decomposition methods for PLM compression.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|10", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "High compression ratios: Tucker decomposition enables significant parameter reduction, achieving up to 48x compression in Transformer layers with minimal performance loss.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|11", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Faster inference: The reconstruction protocol leads to faster inference speeds even with similar FLOPS compared to the original model.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|12", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Orthogonal to existing methods: The approach is complementary to existing compression methods like knowledge distillation, allowing for further compression of distilled models.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|18", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The paper is clearly written and well-organized.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|19", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The methodology is sound and follows established practices in tensor decomposition and knowledge distillation.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|20", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The experimental results are convincing and demonstrate the effectiveness of the proposed approach.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|22", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The paper provides sufficient details for reproducibility, although the specific implementation details of the knowledge distillation process could be further elaborated.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|25", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "The method achieves impressive compression ratios and even improves inference speed while maintaining good performance on the GLUE benchmark.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|26", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "However, the reliance on knowledge distillation and the limited novelty of the findings are potential drawbacks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|27", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Further investigation into the impact on generation capabilities and exploration of alternative methods to address the decomposition discrepancy would be valuable.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|29", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "It would be helpful to discuss the limitations of the proposed method and potential future research directions to address them.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|30", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Exploring the impact of the compression on different downstream tasks, including generation tasks, would provide a more comprehensive understanding of the method's impact.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|31", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Investigating alternative approaches to compensate for the decomposition discrepancy, such as fine-tuning strategies or regularization techniques, could be beneficial.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc31|gemini_pro_1.5|32", "source": "rc", "split": "test", "paper_id": "rc31", "comment": "Overall, this paper presents a valuable contribution to the field of PLM compression and is worthy of acceptance with minor revisions and clarifications.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc89|claude_opus|14", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "- Evaluation is limited to CBQA datasets; broader applicability to other knowledge-intensive NLP tasks is not explored.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc89|claude_opus|16", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "- Some key implementation details are missing, such as specifics of the synthetic data generation process.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc89|claude_opus|20", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Key details are provided to understand the methodology.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc89|claude_opus|21", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "However, some specifics are lacking that would be needed to fully reproduce the results, such as details on the synthetic data generation process for fine-tuning.", "reviewer_type": "ai", "label": "error", "error_type": "Unstated statement", "label_origin": "expert"}
{"id": "rc|llm|rc89|claude_opus|28", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Overall, this work makes a valuable contribution to improving CBQA with language models, though further work is needed to explore broader applicability and knowledge updating.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc89|gpt4|12", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "RECITE represents a significant advancement in leveraging LLMs for knowledge-intensive tasks, offering a promising direction for enhancing factual accuracy without external data reliance.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc89|gpt4|13", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "The approach is innovative, well-executed, and backed by compelling experimental evidence.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc89|gpt4|14", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Despite potential limitations, the paper's contributions to the field are substantial.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc89|gemini_pro_1.5|15", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "Computational cost: Sampling multiple recitations and employing self-consistency can be computationally expensive, especially for large LLMs.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc89|gemini_pro_1.5|19", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "The RECITE framework and its components are novel and represent a significant contribution to the field of knowledge-intensive NLP.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc89|gemini_pro_1.5|24", "source": "rc", "split": "test", "paper_id": "rc89", "comment": "While potential limitations exist, the paper's strengths outweigh them, making it a valuable contribution to the NLP community.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|6", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Ablation studies demonstrate that preserving access to the LLM's original embeddings of the subject in the prompt is important for P-Adapter performance.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|9", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "- Addresses an important problem of extracting factual knowledge consistently from LLMs", "reviewer_type": "ai", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|10", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "- P-Adapters are lightweight and require no additional prompt annotations at inference time", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|11", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "- Substantial improvements in precision and consistency over the baseline", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|12", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "- Insightful ablation studies on what makes P-Adapters effective", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|14", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "- Overall precision is still modest, with the best models only correctly predicting 50% of facts", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|19", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The paper is clearly written and the proposed P-Adapter approach is novel in aiming to adapt diverse natural language prompts into effective continuous prompts for factual extraction.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|20", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The experiments cover multiple out-of-distribution settings to test robustness.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|21", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Key implementation details, hyperparameters, and datasets are specified.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|22", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Code is provided to aid reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|25", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "P-Adapters demonstrate substantial improvements over baselines while requiring no additional annotations at inference time.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|26", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Insightful ablations reveal the importance of preserving access to the original subject embeddings.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|27", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "While overall precision remains modest, this work is a promising step towards practical use of LLMs as knowledge bases.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|claude_opus|28", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The paper is clear and technically sound.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gpt4|7", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "- The methodology presents a significant innovation in handling variability in natural language prompts for querying LLMs, offering a practical solution to improve the robustness of factual information extraction without additional annotation costs.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gpt4|8", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "- Experimental results show clear benefits in terms of precision and consistency, with P-Adapters outperforming baseline approaches by a notable margin across different settings, including out-of-distribution scenarios.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gpt4|9", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "- The paper is well-organized, articulating the problem, proposed solution, and results in a coherent and comprehensible manner, contributing positively to its clarity and overall quality of research.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gpt4|14", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The paper is clearly written, detailing the methodology, experimental setup, and results with sufficient clarity to enable reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gpt4|15", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The approach is novel, addressing the inconsistency problem in querying LLMs with natural language prompts in a unique and effective way.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gpt4|16", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The authors provide a comprehensive comparison with existing methods, demonstrating the quality and novelty of their contributions.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gpt4|17", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Additionally, they mention that the code and data used for experiments will be made available, supporting the paper's reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gpt4|19", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The paper introduces a promising approach to enhance the consistency and precision of factual information extraction from LLMs using P-Adapters.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gpt4|20", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "By innovatively transforming natural language prompts into continuous prompts, the authors address a significant challenge in NLP, making a notable contribution to the field.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|8", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Addresses an important problem: P-Adapters tackle the issue of inconsistency in factual information extraction from LLMs, which is crucial for reliable knowledge base applications.", "reviewer_type": "ai", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|9", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Lightweight and efficient: P-Adapters require minimal additional parameters and training data compared to other approaches like MoE models.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|10", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Improved robustness and consistency: P-Adapters demonstrate significant improvements in both precision and consistency across various evaluation settings, including out-of-distribution scenarios.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|11", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "User-friendly: P-Adapters only require natural language prompts at inference time, aligning with current user expectations for IR systems.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|13", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Modest overall performance: While P-Adapters improve upon the baseline, the absolute precision achieved by even the best models remains relatively low, particularly in out-of-distribution settings.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Limited interpretability analysis: The paper claims improved interpretability of generated prompts but lacks a dedicated study or human evaluation to support this claim.", "reviewer_type": "ai", "label": "error", "error_type": "Unstated statement", "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|15", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Unclear experimental details: Some experimental details, such as the number of samples needed for training, are not fully elaborated.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|16", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "The authors should consider conducting a human evaluation to assess the interpretability of the generated prompts.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|18", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Providing more detailed information about the experimental setup and hyperparameter choices would enhance the paper's reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|19", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Overall, this paper presents a valuable contribution to the field of factual information extraction from LLMs.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|20", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "With further investigation and refinement, P-Adapters have the potential to significantly improve the robustness and consistency of LLM-based knowledge bases.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|22", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Clarity and Quality: The paper is generally well-written and easy to follow, although some experimental details could be clearer.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|23", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Novelty: The P-Adapter approach is novel and offers a promising direction for improving factual information extraction from LLMs.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|24", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "Reproducibility: The authors provide code and data resources to facilitate the reproduction of their results.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|27", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "They demonstrate promising improvements in both precision and consistency, even in out-of-distribution settings.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc48|gemini_pro_1.5|28", "source": "rc", "split": "test", "paper_id": "rc48", "comment": "However, the overall performance remains modest, and further analysis is needed to support claims of improved interpretability.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc50|claude_opus|15", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "- The main limitation is that the empirical evaluation focuses on simplified datasets due to computational constraints.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc50|claude_opus|16", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "It's unclear how well the results generalize to more complex real-world datasets and larger models.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc50|claude_opus|18", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "- The paper lacks experiments comparing flatness measures other than trace of Hessian.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc50|claude_opus|24", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "However, the code for the experiments is not provided which somewhat limits reproducibility.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc50|gpt4|13", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "While the paper introduces the concept of flatness, it does not provide concrete guidelines on how to achieve or measure flatness in practical settings.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc50|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Limited downstream tasks: The experiments primarily focus on a few specific downstream tasks, which might not be representative of the diverse range of NLP applications.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc50|gemini_pro_1.5|15", "source": "rc", "split": "test", "paper_id": "rc50", "comment": "Computational feasibility: Reaching the saturation regime, where the pre-training loss is near-optimal, might be computationally challenging for large-scale models and datasets.", "reviewer_type": "ai", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|llm|rc85|claude_opus|14", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "- More details can be provided on the prompting and knowledge selection/retrieval process with concrete examples to improve clarity.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc85|claude_opus|15", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "- Ablation studies on the impact of different knowledge types and the knowledge selector would strengthen the paper.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc85|claude_opus|20", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Key implementation details are provided, though some more specifics on the prompting and knowledge selection would further aid reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc85|claude_opus|26", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Some added details on the methodology and discussion on limitations would further strengthen the paper.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc85|claude_opus|27", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Overall, this is a high-quality paper that makes valuable contributions to the field.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc85|gpt4|14", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Limitations in Knowledge Coverage: While KiC covers a broad range of knowledge types, the scalability of adding new knowledge types or updating existing ones remains unclear.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc85|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Limited analysis of knowledge quality: The paper lacks quantitative and qualitative analysis to assess the quality of the retrieved knowledge pieces.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc85|gemini_pro_1.5|16", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Applicability to large PLMs: The current implementation of KiC may not be readily scalable to large PLMs due to computational constraints.", "reviewer_type": "ai", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|llm|rc85|gemini_pro_1.5|24", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Can you provide more details on the quantitative and qualitative analysis of the retrieved knowledge pieces?", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc85|gemini_pro_1.5|25", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Have you considered comparing KiC with a baseline that uses few-shot examples for data augmentation?", "reviewer_type": "ai", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|llm|rc85|gemini_pro_1.5|26", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "How can KiC be scaled to larger PLMs, addressing the computational constraints?", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc85|gemini_pro_1.5|27", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Do you plan to explore the performance of KiC on other NLP tasks beyond classification?", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc85|gemini_pro_1.5|29", "source": "rc", "split": "test", "paper_id": "rc85", "comment": "Addressing the weaknesses and exploring the additional questions will further strengthen the work and broaden its impact.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|7", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "- RoT provides an effective and scalable approach to enable language models to perform multi-step reasoning on problems that would otherwise exceed the maximum context size.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|8", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "- The idea of recursively dividing problems and solving subproblems in separate contexts is well-motivated and aligns with how humans approach complex problems through abstraction and recursion.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|9", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "- The paper demonstrates the generality and effectiveness of RoT on a diverse set of arithmetic and algorithmic reasoning tasks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|10", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "- RoT leads to dramatic improvements in reasoning capabilities, not just for large pre-trained models like GPT-3, but also for small randomly initialized Transformers and LSTMs.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|12", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "- RoT currently requires supervision in the form of ground truth intermediate reasoning steps for each task.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|13", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Reducing this expensive supervision could make RoT more widely applicable.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|15", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "- Some additional analysis on the limitations of RoT and failure cases could provide further insights.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|17", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "- The paper is generally well-written and the key ideas of RoT are clearly explained.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|18", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "- The extensive experiments on multiple reasoning tasks demonstrate the effectiveness and versatility of the proposed approach.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|19", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The results convincingly show that RoT leads to significant improvements in reasoning capabilities.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|20", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "- Recursively combining language models with multiple contexts for reasoning is a novel paradigm that addresses a key limitation of existing approaches constrained by a single context length.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|21", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "- The authors plan to release source code which would facilitate reproducibility.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|27", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Overall, this is a high-quality paper that proposes a novel and impactful approach to language model reasoning.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|claude_opus|28", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The key limitation is the need for expensive supervision which could be a useful direction for future work.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gpt4|6", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "RoT addresses a critical limitation in LMs regarding context size and demonstrates a significant improvement in solving complex reasoning tasks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gpt4|7", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The framework is model-agnostic, allowing its application across various models, including Transformers and LSTMs.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gpt4|8", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The paper provides extensive experiments and comparisons, showing that RoT can enable even small models to perform complex reasoning at a level that may be challenging even for humans.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gpt4|10", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The approach requires supervised learning for each task, potentially limiting its applicability to problems where such detailed supervision is feasible.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gpt4|12", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Despite its potential, the approach's dependence on recursive problem-solving may not be suitable for all types of reasoning tasks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gpt4|16", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The novelty of the approach lies in addressing the context size limitation by leveraging recursive problem-solving, a significant step forward in the field.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gpt4|17", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "The inclusion of source code for reproducing experiments enhances the paper's reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gpt4|21", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "While there are limitations, such as the need for task-specific supervision, the strengths and contributions of this work significantly outweigh its weaknesses.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|7", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Novelty: RoT presents a novel approach to reasoning with LMs, addressing the critical issue of context limitations.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|9", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Divide-and-conquer: The recursive approach allows for efficient and scalable reasoning, enabling LMs to tackle problems requiring extensive intermediate steps.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|10", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Performance: RoT demonstrates significant performance improvements on various tasks, achieving near-perfect accuracy even for small LMs.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|11", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Reproducibility: The paper provides the source code for full experiment reproduction.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|13", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Supervision: RoT currently requires supervised training data with ground truth intermediate steps, which can be expensive to obtain.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|15", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Clarity: Some aspects of the paper, particularly the technical details of the recursive procedures, could benefit from improved clarity and elaboration. Providing more detailed explanations and examples of the recursive procedures would improve clarity.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|16", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Exploring methods for reducing the reliance on supervised training data would be valuable for future work.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|18", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Clarity: The paper is generally well-written, but some sections describing the recursive procedures and training data distribution could be made clearer.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|19", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Quality: The research is well-motivated, presents a significant contribution, and is supported by extensive experiments.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|20", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Novelty: RoT is a highly novel approach to reasoning with LMs, offering a unique solution to the context size limitation.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|21", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "Reproducibility: The paper provides the source code, allowing for full experiment reproduction and facilitating further research.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc57|gemini_pro_1.5|26", "source": "rc", "split": "test", "paper_id": "rc57", "comment": "While some aspects could benefit from improved clarity and the current reliance on supervised training data is a limitation, the overall contribution is highly significant and warrants acceptance.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|claude_opus|3", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Thrust leverages the distribution of training instances in the PTLM's representation space.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|claude_opus|12", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "- Extensive experiments convincingly demonstrate the effectiveness of the IAPEK approach using Thrust, leading to performance gains while being more cost-efficient.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|claude_opus|15", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "- While the Thrust metric is evaluated on various tasks and datasets, it would be good to see even more comprehensive experiments, e.g.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|claude_opus|17", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "- Some details on Thrust could be explained further, e.g.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|claude_opus|18", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "how hyperparameters like number of clusters are chosen, runtime cost of computing Thrust.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|claude_opus|19", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "- Qualitative examples illustrating when Thrust decides to seek external knowledge would provide additional insights.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|claude_opus|22", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The paper is clearly written and the proposed method is well-explained.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|claude_opus|25", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Key details are provided to allow reproducibility, though the code is not open-sourced yet (but the authors intend to release it).", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|claude_opus|30", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Some additional details, analysis and even more comprehensive evaluation could further strengthen this already solid work.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gpt4|10", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Reproducibility: The promise to release code and data after acceptance encourages reproducibility and further research.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gpt4|15", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The paper is well-written, presenting a novel approach to a significant problem in NLP.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gpt4|17", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "However, the novelty is somewhat diminished by the lack of comparison with existing methods.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gpt4|18", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The commitment to releasing code and data post-acceptance is commendable and will aid in reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gpt4|22", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "However, a more comprehensive comparison with existing methods could have further highlighted its novelty and superiority.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gpt4|23", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The commitment to open-source principles enhances the paper's value to the research community.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|12", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Overclaiming: The paper's claim that the proposed model \"internalizes\" the task through distillation might be overstated, as it essentially boils down to improved fine-tuning.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|13", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Missing Baseline: The evaluation lacks a crucial baseline - a student model instruction-tuned without distillation.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "This makes it difficult to isolate the gains specifically attributable to the distillation process.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|15", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Inconsistency in Evaluation: The selection of datasets for evaluation in some experiments appears inconsistent and lacks clear justification.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|16", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Limited Analysis of Failure Cases: While the paper acknowledges the failure case of ARC-C, a deeper analysis of why THRUST fails in such scenarios would be beneficial.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|17", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Including the missing baseline of an instruction-tuned student model without distillation is crucial to accurately assess the effectiveness of the proposed distillation method.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|18", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Providing a more detailed explanation for the selection of datasets in certain experiments would improve the clarity and consistency of the evaluation.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|19", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "A deeper analysis of the failure case of ARC-C would be valuable in understanding the limitations of THRUST and potentially guiding future improvements.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|20", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Clarifying the details regarding model size, training/test data usage, and distillation on the test dataset would enhance the paper's clarity.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|22", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Addressing the outlined concerns and incorporating the suggested improvements would further strengthen the paper and increase its impact on the community.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|24", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The paper is generally well-written and presents the core idea clearly.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|25", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "However, some experimental details could be elaborated upon for better clarity.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|27", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "The authors mention releasing code upon acceptance, which will enable reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|31", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "However, certain aspects like overclaiming, missing baselines, and inconsistent evaluation require further attention.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc60|gemini_pro_1.5|32", "source": "rc", "split": "test", "paper_id": "rc60", "comment": "Addressing these concerns and providing a deeper analysis of failure cases would strengthen the paper significantly.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc66|claude_opus|14", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "- The poor performance on the CoLA task is not sufficiently explained and requires more analysis to understand the limitations of the crammed models.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc66|claude_opus|17", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "- More discussion on how the findings can guide future research on efficient training of large language models would be useful.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc66|gpt4|15", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "However, the complexity of some descriptions might require prior knowledge in the field for full comprehension.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc66|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "CoLA performance: The significant drop in performance on CoLA suggests potential limitations of the Crammed LM, either in terms of hyperparameter sensitivity or the ability to capture linguistic acceptability with limited training data.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc66|gemini_pro_1.5|23", "source": "rc", "split": "test", "paper_id": "rc66", "comment": "However, further investigations are needed to address the weaknesses identified and explore the full potential of cramming language models.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|claude_opus|8", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "1) The largest model tested (GPT-3 with 175B parameters) can learn to ground words to conceptual spaces to a significant degree, while smaller models struggle.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|claude_opus|9", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "2) The models generalize to unseen instances in the original and rotated spaces, but not in random unstructured spaces, suggesting they are not merely memorizing.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|claude_opus|10", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "3) To some extent, the models can generalize to unseen but related concepts (e.g.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|claude_opus|11", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "grounding \"blue\" after being shown how to ground \"red\").", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|claude_opus|12", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "4) Analysis of model errors shows the large model tends to predict incorrect but semantically related concepts.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|claude_opus|21", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "It's unclear if the findings would generalize to more complex grounded concepts.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|claude_opus|23", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "- More insight into what enables the largest model to succeed (e.g.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|claude_opus|24", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "analyzing its representations) would strengthen the explanatory power of the work.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|claude_opus|26", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The paper is clearly written and the methodology is described in enough detail to enable reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|claude_opus|27", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The research question is novel and the results are significant, shedding light on the grounding capabilities of LLMs.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|claude_opus|28", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "The experiments are carefully designed and analyzed.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|gpt4|0", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Summary of the Paper", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|gpt4|13", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Clarity, Quality, Novelty, and Reproducibility", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|gpt4|14", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Clarity:The paper is well-written, with clear explanations of the methodology, experiments, and findings.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|gpt4|15", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Quality and Rigor:The research is of high quality, with rigorous experimental design and thorough analysis of results.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|gpt4|16", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Novelty:The paper contributes novel insights into the possibility of grounding LMs in conceptual spaces using only textual information.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|gpt4|17", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Reproducibility:Detailed descriptions of the experimental setup and access to model details enhance the reproducibility of the results.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|gpt4|18", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Summary of the Review", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|gemini_pro_1.5|12", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Performance of smaller models: Smaller LLMs perform poorly on the tasks, suggesting that massive scale might be necessary for this approach to work effectively.", "reviewer_type": "ai", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|llm|rc73|gemini_pro_1.5|15", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Clarity: The paper is clearly written and well-organized, with concrete examples and detailed explanations of the methodology.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|gemini_pro_1.5|17", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Novelty: The idea of grounding LLMs by leveraging their existing conceptual structure is novel and potentially impactful.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc73|gemini_pro_1.5|18", "source": "rc", "split": "test", "paper_id": "rc73", "comment": "Reproducibility: The paper provides sufficient details about the data generation process and model configurations to allow for reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc75|claude_opus|9", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "The gains from adding locality features, while consistent, are relatively modest in terms of perplexity and accuracy.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc75|claude_opus|10", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "More discussion on the significance of the improvements would be helpful.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc75|claude_opus|11", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "The choice of locality features seems somewhat ad-hoc based on domain knowledge.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc75|claude_opus|12", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "More principled methods for defining the features could be explored.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc75|claude_opus|13", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "Limitations of the approach when locality information is not available or already captured by the model are not discussed in depth.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc75|gpt4|11", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "Limited domain analysis: While the paper provides insights into Java and Wikipedia datasets, the exploration of other domains where structural locality could play a crucial role is missing.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc75|gpt4|17", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "The inclusion of source code and detailed experiment setups enhances the paper's reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc75|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "Exploring automated feature extraction and extending the approach to other domains would broaden its impact.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc75|gemini_pro_1.5|15", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "Datastore size: The datastore size for Wikipedia is quite large, raising potential concerns about computational efficiency and scalability.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc75|gemini_pro_1.5|25", "source": "rc", "split": "test", "paper_id": "rc75", "comment": "Further exploration of its applicability to other domains and addressing the potential limitations would be valuable future directions.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc79|claude_opus|2", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The authors explore the effectiveness of various document similarity methods in creating parallel code datasets, study the tolerance of code translation models to varying degrees of noise in the training data, and demonstrate the practical utility of their approach by creating datasets and training models for translating between programming language pairs not explored in prior work.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc79|claude_opus|9", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "This is an important finding for the practicality of the proposed approach.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc79|claude_opus|12", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "- While the paper shows promising results, the datasets created are still noisy by design.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc79|claude_opus|13", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The impact of this noise on downstream applications of the code translation models is not fully explored.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc79|claude_opus|26", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Overall, this is a strong paper that makes a valuable contribution to the code translation field.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc79|claude_opus|27", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Some weaknesses exist in evaluating downstream impact and scalability.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc79|gpt4|13", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "The reproducibility of the results may be challenging due to the complexity of the setup and potential variability in the document similarity measures.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc79|gpt4|18", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "However, reproducibility might require access to specific datasets and clarity on the document similarity measures used.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc79|gemini_pro_1.5|11", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Noise concerns: While the authors analyze noise tolerance, the inherent noise in automatically generated datasets could still impact translation quality and modelgeneralizability.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc79|gemini_pro_1.5|12", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Limited analysis of failure cases: The paper lacks a deeper analysis of why the proposed method performs poorly on certain tasks or language pairs.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc79|gemini_pro_1.5|13", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Scalability concerns: Identifying critical layers and false prefix-matching heads for each model/task combination might not be scalable for practical applications.", "reviewer_type": "ai", "label": "error", "error_type": "Unstated statement", "label_origin": "expert"}
{"id": "rc|llm|rc79|gemini_pro_1.5|16", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "A deeper analysis of failure cases and the reasons behind them would be valuable to understand the limitations of the proposed method.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc79|gemini_pro_1.5|17", "source": "rc", "split": "test", "paper_id": "rc79", "comment": "Investigating the scalability of the approach for practical applications and exploring ways to address potential bottlenecks is crucial for future work.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|16", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Analysis of selective generation focuses on abstention rate, but it would be interesting to see the coverage vs quality trade-off, i.e.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|17", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "what % of examples can be generated while maintaining a specified quality bar.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|21", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The paper is clearly written and easy to follow.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|22", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The methodology is explained in detail, including algorithms for fitting the Gaussians and computing OOD scores.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|23", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Assumptions made are stated upfront.", "reviewer_type": "ai", "label": "error", "error_type": "Unstated statement", "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|25", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Results are extensive and include both automatic and human evaluation.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|26", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Key strengths and weaknesses of the method are analyzed.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|27", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The approach of using the CLM's own embeddings for OOD detection without a separate model is quite novel, as is the extension of selective prediction to language generation under distribution shift.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|28", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The authors plan to release code which would aid reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|30", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Details on hardware requirements and runtimes could further improve reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|32", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "This paper makes valuable contributions to OOD detection and selective generation for conditional language models under distributional shift.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|33", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The proposed OOD scores are highly effective while being simple and lightweight.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|35", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The key weakness is the lack of discussion on computational efficiency.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|claude_opus|36", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "But overall, this is a strong paper that would be of interest to the NLP community and spur further work in this important direction.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|gpt4|11", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The approach's dependency on specific embeddings and the potential for its performance to vary across different models or tasks is not extensively discussed.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|gpt4|14", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The paper is well-written, presenting a clear and logical progression of ideas.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|gpt4|15", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The research is of high quality, leveraging both theoretical insights and empirical evaluations.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|gpt4|16", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The novelty of the approach lies in its effective combination of perplexity scores and OOD detection for improving CLM outputs.", "reviewer_type": "ai", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|llm|rc83|gpt4|17", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The provided details and evaluation framework contribute to the paper's reproducibility, although additional information on the implementation and potential limitations would further enhance it.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|gpt4|21", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Despite some areas for further exploration, the paper's strengths in novelty, methodological clarity, and potential impact make a compelling case for its acceptance.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "Clarity: Some aspects of the methodology, particularly the feature-based and deep ULF methods, could benefit from further elaboration and details.", "reviewer_type": "ai", "label": "error", "error_type": "Unstated statement", "label_origin": "expert"}
{"id": "rc|llm|rc83|gemini_pro_1.5|16", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The paper is generally well-written and easy to follow.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|gemini_pro_1.5|17", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The research is of high quality, with thorough experiments and analysis.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|gemini_pro_1.5|18", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The proposed OOD detection scores and selective generation framework are novel and significant contributions.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc83|gemini_pro_1.5|19", "source": "rc", "split": "test", "paper_id": "rc83", "comment": "The authors mention that code will be released, which will enhance reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|claude_opus|12", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "The paper is generally well-written and easy to follow.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|claude_opus|13", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "The proposed Proto-Trex architecture is novel in generating prototypical explanations for transformer language models.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|claude_opus|14", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "The interactive learning component is also an interesting extension.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|claude_opus|16", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "However, some details on reproducibility, such as hyperparameters and model selection, are missing.", "reviewer_type": "ai", "label": "error", "error_type": "Neglect", "label_origin": "expert"}
{"id": "rc|llm|rc86|gpt4|11", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "The complexity of the Proto-Trex architecture might pose challenges for widespread adoption, especially in resource-constrained environments.", "reviewer_type": "ai", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|llm|rc86|gpt4|12", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "While promising, the evaluation of user interactions and their impact on model performance and interpretability could benefit from further in-depth studies involving a broader range of participants.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gpt4|14", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "The paper is well-written, presenting a clear and concise explanation of the Proto-Trex and iProto-Trex frameworks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gpt4|15", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "The research is of high quality, contributing novel insights into interpretability in NLP.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gpt4|16", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "The proposed methods are novel, offering a practical approach to improving the transparency of transformer LMs.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gpt4|17", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "The experiments are well-designed, and the paper provides sufficient detail for reproducibility, though the complexity of the models might pose some challenges.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gemini_pro_1.5|12", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Trade-off with accuracy: While generally competitive, Proto-Trex sometimes shows a slight decrease in accuracy compared to non-interpretable baselines.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gemini_pro_1.5|13", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Limited scope: The evaluation focuses on sentiment classification tasks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Further research is needed to assess thegeneralizability of the approach to other NLP tasks.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gemini_pro_1.5|15", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Word-level explanations: While the paper explores word-level prototypes, the explanations generated seem less coherent and interpretable than sentence-level ones.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gemini_pro_1.5|17", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "It would be interesting to see how Proto-Trex performs on other NLP tasks beyond sentiment classification.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gemini_pro_1.5|18", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Further research could explore alternative methods for selecting and representing word-level prototypes to improve their interpretability.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gemini_pro_1.5|19", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Investigating the impact of different user types and expertise levels on iProto-Trex could be valuable.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gemini_pro_1.5|22", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Clarity: The paper is generally well-written and easy to follow, although some technical details might require further clarification. Quality: The research is well-conducted, with a solid methodology and comprehensive experiments. Novelty: The use of prototype networks for explaining Transformer LMs is novel and offers a promising direction for interpretable NLP.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc86|gemini_pro_1.5|23", "source": "rc", "split": "test", "paper_id": "rc86", "comment": "Reproducibility: The paper provides access to code and trained models, facilitating reproducibility and further research.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|claude_opus|10", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "This is an important result.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|claude_opus|13", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "- Well-written paper that clearly describes the limitations of prior MoE approaches and the key aspects of the SaMoE architecture.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|claude_opus|15", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "- While the rationale for the self-adaptive combination is clear, it would be useful to provide some additional intuition or analysis on why this particular formulation is effective.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|claude_opus|19", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "- The work is high quality, with strong empirical results supporting the claimed contributions.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|claude_opus|20", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Ablations and comparisons provide additional evidence for the importance of the proposed components.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|claude_opus|21", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "- The self-adaptive combination scheme is a novel contribution that hasn't been explored for MoE models before to my knowledge.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|claude_opus|22", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Decoupling expert parameters from model depth is a key insight.", "reviewer_type": "ai", "label": "error", "error_type": "Misunderstanding", "label_origin": "expert"}
{"id": "rc|llm|rc88|claude_opus|23", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "- Sufficient implementation details (e.g.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|claude_opus|24", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "appendix, code release) are provided to enable reproducibility.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|0", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Summary of the Paper", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|7", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Clarity and Methodological Rigor: The paper is well-structured, presenting a clear problem statement, detailed methodology, and extensive experimental results.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|8", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The proposed solution is innovative, addressing a significant issue in the scaling of MoE models.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|10", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "This advancement could make large-scale language models more accessible and cost-effective.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|14", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Limited Discussion on Limitations and Broader Impact: While the paper extensively discusses the technical aspects and benefits of SaMoE, it lacks a detailed discussion on potential limitations, broader impacts, and ethical considerations of deploying such models.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|16", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Clarity, Quality, Novelty, and Reproducibility", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|17", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Clarity: The paper is well-written, with a logical flow that facilitates understanding of complex concepts.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|18", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The use of figures and tables effectively illustrates the methodology and results.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|19", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Quality: The research is of high quality, demonstrating a deep understanding of MoE models and presenting a novel approach to improving their parameter efficiency.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|20", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Novelty: SaMoE introduces a unique method for combining expert layers, distinguishing it from previous works on MoE models.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|21", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "The self-adaptive expert combination mechanism is a novel contribution to the field.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gpt4|24", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Summary of the Review", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gemini_pro_1.5|13", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Limited exploration of proxy tasks: The paper relies on a proxy task to measure the closeness to a \"ground-truth prompt\" during prompt search, but the definition and selection of this proxy task are not thoroughly investigated.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gemini_pro_1.5|14", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Some unconvincing results: Certain results, such as the reported human-written performance on some datasets, seem lower than expected and require further clarification.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gemini_pro_1.5|17", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Clarity and Quality: The paper is well-written and easy to follow, with a clear explanation of the proposed method and analysis of the challenges addressed.", "reviewer_type": "ai", "label": "error", "error_type": "Unstated statement", "label_origin": "expert"}
{"id": "rc|llm|rc88|gemini_pro_1.5|18", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Novelty: While the general idea of expert sharing is not entirely new, the specific design of SaMoE's self-adaptive expert combination is novel and shows significant improvements.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}
{"id": "rc|llm|rc88|gemini_pro_1.5|19", "source": "rc", "split": "test", "paper_id": "rc88", "comment": "Reproducibility: The authors' commitment to open-sourcing the code strengthens the potential for reproducibility and further research on SaMoE.", "reviewer_type": "ai", "label": "correct", "error_type": null, "label_origin": "expert"}