其他
浅谈学术论文rebuttal
学术论文是发布自己或团队最新研究进展正式且最快捷的途径,也是和同行交流想法最方便、高效的方式。当同行评议(Peer review)作为学术成果正式发布的必经之路已运行200余年[1]时,用正确的姿势进行review rebuttal便成为提高论文录用机率甚至扭转乾坤的最后一搏。本文从“What is peer review?”、“How to rebuttal?”及“Does rebuttal matter?”三部分来聊聊学术论文(主要针对人工智能领域会议和期刊)rebuttal的那些事。
What is peer review?
How to rebuttal?
提交PDF文档作为rebuttal(有页数和模版限制,方便提交新表格和插图,如CVPR/ICCV/ECCV); 提交一段文本作为rebuttal(有字符数限制,不方便提交新表格和插图,如AAAI/IJCAI); 提交PDF文档作为rebuttal,同时提交修改后的论文(重提交的文件有类似的政策限制,如WACV)。
Rebuttal时,针对不同review类型: 喷「novelty」:这类最为麻烦但也是作者们遇到可能最多的意见,作此类rebuttal时不妨重新梳理和强调文章的重要贡献,然后澄清并不是trivial的简单combine,再强调一下motivation和intuition,用另一种方式将文章亮点表达出来。同时,可以尝试“围魏救赵”,即:若审稿人针对方法的某个部件提出novelty不足,可强调其他部件或整个方法的范式是前所未有的;或claim说方法简单有效,思路全然不同; 喷「factual error」:审稿人一旦找出文章的事实性错误,作者不妨大方承认,并表示感谢,同时表示会在final version中更正错误;另一种情况是,可能就是因为作者自己没写明白,才使得审稿人错误理解,如此,也可大方承认,说“我们已经修改了这部分描述,实际上是这样做的,并不是你理解的那样,blabla”; 喷「涨点不足」:此类一般有两种,一则无证据的裸喷涨点不足;二则有证据(提供了reference)喷涨点不足或喷没有对比reference中结果。针对一,可找些证据(如列reference)论证自己方法的涨点幅度和其他state-of-the-art的涨点幅度是可比的,“你看,别人发在顶会的结果相比baseline也是涨这么多”;针对二,可试着找出这些“证据”方法和自己方法的不同之处或实验细节的不公平之处,比如图像分辨率不同、backbone不同等; 喷「实验不足」:有条件做实验的,rebuttal中补上即可;若实验规模太大,rebuttal期间无条件做出,可在rebuttal中承诺final version中补上(这样力度会相对较弱);而对于要求不合理的实验意见,可实事求是的说明为何无需做实验; Rebuttal时不要漏点,要逐点回应做到有问必答。若因篇幅有限,可将类似的意见合成一点,万不可因篇幅有限擅自删除一些要点或遗漏要点,以免造成含糊不清、浑水摸鱼之嫌,一旦被审稿人发现会在paper discussion阶段当作硬伤来“置于死地”;此外,除非基本上全是positive评价,要充分利用rebuttal file的空间,充分表达自己观点,力图将所有评分均拉到positive区间确保十拿九稳; Rebuttal时需要揣摩审稿人倾向,“一切可以团结的力量都要团结,不中立的可以争取为中立,反动的也可以分化和利用”。有的审稿人会在意见中明确表示,“如果解决了xxx,我就会提升评分”,对此一定要充分争取;对于某些审稿人提出的不足(如novelty),可能刚巧是另一位审稿人提出的优点(“This paper is interesting and novel”),一定要为我所用,让两位审稿人在paper discussion中“短兵相接”;对于borderline的审稿人,一定要充分“拉拢腐蚀”;对于初审给了positive分数的审稿人,一定要巩固基础;对于初审给了negative分数的审稿人,一定要放绝大多数的精力和rebuttal篇幅来解释澄清,争取“冰释前嫌”; Rebuttal是“一盘棋”,整篇rebuttal需要统筹协调,与正文、review配合的相得益彰,同时还需注意rebuttal篇幅资源的分配和优化。哪位审稿人应多分配笔墨、哪个问题应多着力回应都需要根据整体审稿意见情况深入思考、统筹安排; Rebuttal中能缩写的尽量缩写,如约定俗成的PCA、CNN等不必展开,从而节省空间,将资源留给更需要的回应; Rebuttal时若发现审稿人的factual error,如ta提出的某个观点有显然错误、提出需要对比的数据集显然不是该领域常用的数据等,作者可在rebuttal回应此人时首先指出其错误,先下一城,赢得主动。要知道rebuttal除了该审稿人之外,其他审稿人以及AC都会看到。此外,这一问题还可以在AC message(见下文)中指出,降低该审稿人意见在AC心中的置信度; ……
Please note that Assigned Reviewer #id has made some statements that are either against the common-sense in our field or self-contradictory (ironically his/her own confidence rating is "very confident"). blabla We want to bring to your attention the very flawed review #id. This reviewer is self-contradictory, cf. Comment #id1, Comment #id2, and Response #id. blabla We would like to raise attention to AC that unfortunately Reviewer #id holds a very biased view towards the contributions of our paper. blabla
开头 Thank you for your suggestion. Thank you for the positive/detailed/constructive comments. We sincerely thank all reviewers and ACs for their time and efforts. Below please find the responses to some specific comments. We thank the reviewers for their useful comments. The common questions are first answered, then we clarify questions from every individual review. We thank the useful suggestions from the reviewers. Some important or common questions are first addressed, followed by answers to individual reviews. 表达同意 We thank the reviewer for pointing out this issue. We agree with you and have incorporated this suggestion throughout our paper. We have reflected this comment by … We can/will add/compare/revise/correct ... in our revised manuscript/our final version. Due to the rebuttal policy, “authors should not include new experimental results in the rebuttal”, additional results may not be included. However, we will add these mentioned experiments and discussions in our final version. Thank you for the constructive comment.(对于CVPR/ICCV/ECCV rebuttal不能提供新结果的政策) 表达不同意 We respectfully disagree with Reviewer #id that ... The reviewer might have overlooked Table #id ... We can compare ... but it is not quite related to our work ... We have to emphasize that ... The reviewer raises an interesting concern. However, our work ... Thank you for the comment, but we cannot fully agree with the comment. As stated/emphasized ... You have raised an important point; however, we believe that ... would be outside the scope of our paper because … This is a valid assessment of …; however, we believe that ... would be more appropriate because ... 解释澄清 We have indeed stated/included/discussed/compared/reported/clarified/elaborated ... in our original paper ... (cf. Line #id). As we stated in Line #id, ... We have rewritten ... to be more in line with your comments. We hope that the edited section clarifies … 额外信息与解释 We have included a new figure/table (cf. Figure/Table #id) to further illustrate… We have supplemented the xxx section with explanations of ... Thank you for the comment. We will explore this in future work.
Does rebuttal matter?
「逆天改命」:首轮Strong Reject改为positive评分的仅占所有review的0.8%; 「强势逆转」:首轮Weak Reject改为positive评分的占所有review的6.9%; 「力挽狂澜」:首轮Borderline改为positive评分的有33.6%; 「错失良机」:首轮Borderline改为negative评分的约56%; 「画蛇添足」:首轮Weak Accept改为negative评分的有15.5%; 「弄巧成拙」:首轮Strong Accept改为negative评分的有5.2%;
后记
参考
^Aliaksandr Birukou, Joseph Wakeling, Claudio Bartolini, Fabio Casati, Maurizio Marchese, Katsiaryna Mirylenka, Nardine Osman, Azzurra Ragone, Carles Sierra, and Aalam Wassef. Alternatives to peer review: Novel approaches for research evaluation. Frontiers in Computational Neuroscience, 2011, 5:56. ^CVPR/ICCV/ECCV等会议在使用Toronto Paper Matching System作为准则之一进行论文分配:http://www.cs.toronto.edu/~zemel/documents/tpms.pdf ^指审稿人或AC在投稿池子中申请审稿,按程度不同可大致分为:eager、willing、in a pinch、not willing等。AAAI/IJCAI等会议在沿用此方式。 ^http://cvpr2019.thecvf.com/files/CVPR%202019%20-%20Welcome%20Slides%20Final.pdf ^Yang Gao, Steffen Eger, Ilia Kuznetsov, Iryna Gurevych, Yusuke Miyao. Does My Rebuttal Matter? Insights from a Major NLP Conference. NAACL-HLT 2019
历史文章推荐
人体姿态估计的过去,现在,未来
给研究新生的建议,光看论文是学不好的,一定要看书,看书,看书!
So Young Sohn:信用评级与专利保护中的AI技术概览
Shai Ben-David:无监督学习中的鲜花与荆棘
不是我们喜新厌旧,而是RAdam确实是好用,新的State of the Art优化器RAdam
机器学习中的评价指标
CVPR2019 |《胶囊网络(Capsule Networks)综述》,附93页PPT下载
AiLearning:一个 GitHub万星的中文机器学习资源
你正在看吗?👇