id stringlengths 17 29 | source stringclasses 1
value | split stringclasses 1
value | paper_id stringclasses 100
values | comment stringlengths 15 932 | reviewer_type stringclasses 2
values | label stringclasses 2
values | error_type stringclasses 4
values | label_origin stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|
rc|human|rc0|review#1|14 | rc | test | rc0 | 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) | human | correct | null | expert |
rc|human|rc0|review#2|17 | rc | test | rc0 | Clarity, Quality, Novelty And Reproducibility: | human | correct | null | expert |
rc|human|rc0|review#2|19 | rc | test | rc0 | ok I really found this paper hard to follow. | human | correct | null | expert |
rc|human|rc0|review#2|20 | rc | test | rc0 | After reading the main paper the reader is left wondering as to what was actually proposed. | human | error | Neglect | expert |
rc|human|rc0|review#2|21 | rc | test | rc0 | 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). | human | correct | null | expert |
rc|human|rc0|review#2|22 | rc | test | rc0 | 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. | human | correct | null | expert |
rc|human|rc0|review#2|23 | rc | test | rc0 | Additionally, apart from "insights" as to what makes good questions, authors don't provide enough details of what is being generated. | human | correct | null | expert |
rc|human|rc0|review#2|24 | rc | test | rc0 | 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. | human | correct | null | expert |
rc|human|rc0|review#2|26 | rc | test | rc0 | after having (finally) understood the approach, I do find it witty. | human | correct | null | expert |
rc|human|rc0|review#2|27 | rc | test | rc0 | 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. | human | correct | null | expert |
rc|human|rc0|review#2|29 | rc | test | rc0 | 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. | human | error | Neglect | expert |
rc|human|rc0|review#2|30 | rc | test | rc0 | Summary Of The Review: | human | correct | null | expert |
rc|human|rc0|review#2|31 | rc | test | rc0 | Update: in light of additional comparison with prompt tuning, i am raising my score. | human | correct | null | expert |
rc|human|rc0|review#2|32 | rc | test | rc0 | I appreciate all the additional experiments authors ran. | human | correct | null | expert |
rc|human|rc0|review#2|33 | rc | test | rc0 | I still encourage to improve the flow/the story/the presentation, but the method deserves to be seem | human | correct | null | expert |
rc|human|rc0|review#2|34 | rc | test | rc0 | 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. | human | correct | null | expert |
rc|human|rc0|review#2|35 | rc | test | rc0 | Overall It is a very good idea of forcing the model to write the prompts that it is able to respond to. | human | correct | null | expert |
rc|human|rc0|review#2|36 | rc | test | rc0 | I do think the main paper should be reworked to include more examples/explanations of what is being proposed plus WS part. | human | correct | null | expert |
rc|human|rc0|review#2|38 | rc | test | rc0 | Are all the models you tested are next token (perfix) trained, not span corruption trained correct? | human | correct | null | expert |
rc|human|rc0|review#2|39 | rc | test | rc0 | 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? | human | correct | null | expert |
rc|human|rc0|review#2|40 | rc | test | rc0 | I assume your "answers" templates are the same for various tasks so they don't include that closed book knowledge? | human | correct | null | expert |
rc|human|rc0|review#2|42 | rc | test | rc0 | 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? | human | correct | null | expert |
rc|human|rc0|review#2|43 | rc | test | rc0 | Are questions prompts and answer prompts task specific? How did you create them How many chains do you create for each input during inference? | human | correct | null | expert |
rc|human|rc0|review#2|44 | rc | test | rc0 | 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 | human | correct | null | expert |
rc|human|rc0|review#2|45 | rc | test | rc0 | 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 | human | correct | null | expert |
rc|human|rc0|review#3|14 | rc | test | rc0 | I found the story line a little confusing. | human | correct | null | expert |
rc|human|rc0|review#3|15 | rc | test | rc0 | 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). | human | correct | null | expert |
rc|human|rc0|review#3|16 | rc | test | rc0 | In many of the tasks the 6B model already beats the 175B without weak supervision based aggregation. | human | correct | null | expert |
rc|human|rc0|review#3|17 | rc | test | rc0 | Also this table is not really apples-to-apples comparison making it hard to interpret. | human | correct | null | expert |
rc|human|rc0|review#3|18 | rc | test | rc0 | Should make it more clear what is the goal in this table --- is GPT-3 simply used more like a reference than a baseline? | human | correct | null | expert |
rc|human|rc0|review#3|20 | rc | test | rc0 | 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. | human | correct | null | expert |
rc|human|rc0|review#3|21 | rc | test | rc0 | I think still I am a little not sure, also about whether re-formatting is necessary for aggregation to work well. | human | correct | null | expert |
rc|human|rc0|review#3|22 | rc | test | rc0 | In general, it could help to tailor overall story as being primarily about prompt re-formating or aggregation. | human | correct | null | expert |
rc|human|rc0|review#3|23 | rc | test | rc0 | 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. | human | correct | null | expert |
rc|human|rc0|review#3|24 | rc | test | rc0 | Perhaps worth using those as baselines, or mentioning why they weren't used. | human | correct | null | expert |
rc|human|rc0|review#3|25 | rc | test | rc0 | Maybe CoT is not effective for smaller models? | human | correct | null | expert |
rc|human|rc0|review#3|26 | rc | test | rc0 | Clarity, Quality, Novelty And Reproducibility: | human | correct | null | expert |
rc|human|rc0|review#3|28 | rc | test | rc0 | It would be helpful to have finetuning baselines. | human | correct | null | expert |
rc|human|rc0|review#3|29 | rc | test | rc0 | 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). | human | correct | null | expert |
rc|human|rc0|review#3|30 | rc | test | rc0 | Sometimes the improvement over GPT-3 is great but still far behind finetuning. | human | correct | null | expert |
rc|human|rc0|review#3|32 | rc | test | rc0 | Self-consistency is a very crucial baseline / related work. | human | correct | null | expert |
rc|human|rc0|review#3|33 | rc | test | rc0 | Unlike the text says, it does not require additional training. | human | correct | null | expert |
rc|human|rc0|review#3|34 | rc | test | rc0 | 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. | human | correct | null | expert |
rc|human|rc0|review#3|35 | rc | test | rc0 | 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 | human | correct | null | expert |
rc|human|rc0|review#3|36 | rc | test | rc0 | There is a lot of other works that describe strategies for exemplar selection. | human | correct | null | expert |
rc|human|rc0|review#3|37 | rc | test | rc0 | 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. | human | correct | null | expert |
rc|human|rc0|review#3|38 | rc | test | rc0 | There are other works that do sequential prompting like least-to-most and SeqZero. | human | correct | null | expert |
rc|human|rc0|review#3|40 | rc | test | rc0 | The analysis / diagnostic in Fig 5 is great, and builds confidence that sub-tasks are related to final performance. | human | correct | null | expert |
rc|human|rc0|review#3|41 | rc | test | rc0 | 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? | human | correct | null | expert |
rc|human|rc0|review#3|42 | rc | test | rc0 | It was not immediately clear to me what was meant by "highly-correlated outputs". | human | correct | null | expert |
rc|human|rc0|review#3|43 | rc | test | rc0 | nit: Are you taking majority vote or plurality vote? | human | correct | null | expert |
rc|human|rc0|review#3|44 | rc | test | rc0 | nit: The framing of perfect vs. imperfect prompts is not the strongest. | human | correct | null | expert |
rc|human|rc0|review#3|45 | rc | test | rc0 | Perhaps it is sufficient to reframe and say you are introducing an efficient and effective ensembling method? | human | correct | null | expert |
rc|human|rc0|review#4|12 | rc | test | rc0 | 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.). | human | correct | null | expert |
rc|human|rc0|review#4|13 | rc | test | rc0 | This is an unrealistic setting because only a few training instances might be present depending on the task. | human | correct | null | expert |
rc|human|rc0|review#4|14 | rc | test | rc0 | 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. | human | correct | null | expert |
rc|human|rc0|review#4|15 | rc | test | rc0 | This introduces additional latency for task adaptation. | human | correct | null | expert |
rc|human|rc0|review#4|16 | rc | test | rc0 | 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? | human | error | Misunderstanding | expert |
rc|human|rc0|review#4|17 | rc | test | rc0 | 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) | human | correct | null | expert |
rc|human|rc0|review#4|18 | rc | test | rc0 | 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? | human | correct | null | expert |
rc|human|rc0|review#4|19 | rc | test | rc0 | The structure of Section 3.2 and Section3.4 is divided into observation and solution (AMA). | human | correct | null | expert |
rc|human|rc0|review#4|20 | rc | test | rc0 | However, the observation experiments are not extensive enough. Is there a reason for selecting 3 tasks (CB, WSC, RTE)? | human | correct | null | expert |
rc|human|rc0|review#4|21 | rc | test | rc0 | Observing the findings on more tasks would make the findings more general and the motivation of AMA stronger. | human | correct | null | expert |
rc|human|rc0|review#4|22 | rc | test | rc0 | I think that StoryCloze should be included in Natural Language Understanding instead of NLI for Table 1 and 2. | human | correct | null | expert |
rc|human|rc0|review#4|23 | rc | test | rc0 | Clarity, Quality, Novelty And Reproducibility: | human | correct | null | expert |
rc|human|rc0|review#4|24 | rc | test | rc0 | The proposed method is novel in that it is effective and scalable. | human | correct | null | expert |
rc|human|rc0|review#4|25 | rc | test | rc0 | 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. | human | correct | null | expert |
rc|human|rc1|review#1|22 | rc | test | rc1 | 1.It is claimed that these three tasks require understanding of both structure and content of the web-page. | human | correct | null | expert |
rc|human|rc1|review#1|23 | rc | test | rc1 | 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. | human | correct | null | expert |
rc|human|rc1|review#1|24 | rc | test | rc1 | It can be argued that no significant HTML structure analysis or understanding is needed for these tasks. | human | correct | null | expert |
rc|human|rc1|review#1|25 | rc | test | rc1 | 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'. | human | correct | null | expert |
rc|human|rc1|review#1|26 | rc | test | rc1 | As these attributes are in the close neighbourhood of 'input', parsing of HTML is not strictly necessary. | human | correct | null | expert |
rc|human|rc1|review#1|27 | rc | test | rc1 | Therefore, it might a good idea to do some experiments that demonstrate unequivocally the need for HTML structure analysis or understanding in these tasks. | human | correct | null | expert |
rc|human|rc1|review#1|28 | rc | test | rc1 | 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 ***. | human | correct | null | expert |
rc|human|rc1|review#1|31 | rc | test | rc1 | The submission is well written and easy to understand. | human | correct | null | expert |
rc|human|rc1|review#1|32 | rc | test | rc1 | The three canonical tasks are described well and the adaptation of the various LLM for building models for these tasks are well explained. | human | correct | null | expert |
rc|human|rc1|review#1|33 | rc | test | rc1 | The proposed solution is simple and appears to be effective for the tasks considered and the datasets chosen. | human | correct | null | expert |
rc|human|rc1|review#1|35 | rc | test | rc1 | Experiments are designed well and should be easy to reproduce. | human | correct | null | expert |
rc|human|rc1|review#1|36 | rc | test | rc1 | Datasets used in the experiments have been promised to be released. | human | correct | null | expert |
rc|human|rc1|review#1|37 | rc | test | rc1 | The work should be interesting for practitioners. | human | correct | null | expert |
rc|human|rc1|review#1|39 | rc | test | rc1 | This work asks the question can off-the-shelf LLM trained on natural language text be used effectively for tasks that involve HTML pages. | human | correct | null | expert |
rc|human|rc1|review#1|40 | rc | test | rc1 | It proposes three tasks as canonical tasks in understanding HTML. | human | correct | null | expert |
rc|human|rc1|review#1|41 | rc | test | rc1 | It employs a variety of LLM to build models for the three tasks using a small amount of HTML data for fine tuning. | human | correct | null | expert |
rc|human|rc1|review#1|42 | rc | test | rc1 | It shows that LLM does help these tasks significantly. | human | correct | null | expert |
rc|human|rc1|review#1|43 | rc | test | rc1 | One key question not answered in this context is how much of HTML structure analysis and understanding is truly required for these questions. | human | correct | null | expert |
rc|human|rc1|review#2|9 | rc | test | rc1 | Models that are trained on task specific data to compare against is questionable. | human | correct | null | expert |
rc|human|rc1|review#2|10 | rc | test | rc1 | 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. | human | correct | null | expert |
rc|human|rc1|review#2|11 | rc | test | rc1 | However, the results seem to focus on LLMs only. | human | correct | null | expert |
rc|human|rc1|review#2|12 | rc | test | rc1 | I don't see the point of compare different variants of LLMs with or without pre-training. | human | correct | null | expert |
rc|human|rc1|review#2|13 | rc | test | rc1 | 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"? | human | correct | null | expert |
rc|human|rc1|review#2|14 | rc | test | rc1 | [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. | human | correct | null | expert |
rc|human|rc1|review#2|15 | rc | test | rc1 | [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. | human | correct | null | expert |
rc|human|rc1|review#2|17 | rc | test | rc1 | Clarity: The paper is easy to follow as the content of the paper is simple. | human | correct | null | expert |
rc|human|rc1|review#2|18 | rc | test | rc1 | Quality: The experiments seem to be well-done. | human | correct | null | expert |
rc|human|rc1|review#2|19 | rc | test | rc1 | Novelty: The novelty is limited in terms of methodology. | human | correct | null | expert |
rc|human|rc1|review#2|20 | rc | test | rc1 | It is mostly an empirical study for large LLMs on HTML texts. | human | correct | null | expert |
rc|human|rc1|review#2|21 | rc | test | rc1 | Reproducibility: Some details are missing. | human | correct | null | expert |
rc|human|rc1|review#2|22 | rc | test | rc1 | Either code or more detailed experimental setup should be provided in the supplementary material | human | correct | null | expert |
rc|human|rc1|review#3|10 | rc | test | rc1 | The novelty is primarily empirical, and the technical approach is essentially an application of known techniques (the standard pretrain-finetune paradigm of LLMs). | human | correct | null | expert |
rc|human|rc1|review#3|11 | rc | test | rc1 | I examined the supplementary material provided by the authors and only found the data files. | human | correct | null | expert |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.