================
Workshop overview
================
Publication of negative results is difficult in most fields, but in NLP the problem is exacerbated by the near-universal focus on improvements in benchmarks. This situation implicitly discourages hypothesis-driven research, and it turns creation and fine-tuning of NLP models into art rather than science. Furthermore, it increases the time, effort, and carbon emissions spent on developing and tuning models, as the researchers have no opportunity to learn what has already been tried and failed.
This workshop on Insights from Negative Results in NLP invites both practical and theoretical unexpected or negative results that have important implications for future research, highlight methodological issues with existing approaches, and/or point out pervasive misunderstandings or bad practices. In particular, the most successful NLP models currently rely on different kinds of pretrained meaning representations (from word embeddings to Transformer-based models like BERT). To complement all the success stories, it would be insightful to see where and possibly why they fail. Any NLP tasks are welcome: sequence labeling, question answering, inference, dialogue, machine translation - you name it.
A successful negative results paper would contribute one of the following:
- experiments on (in)stability of the previously published results due to hardware, random initializations, etc.;
- ablation studies of components in previously proposed models, showing that their contributions are different from the initially reported;
- datasets or probing tasks showing that previous approaches do not generalize to other domains or language phenomena;
- extensions or annotations of existing datasets which show that prior successes are due to spurious statistical factors or annotation artifacts;
- the respective contributions of pre-training vs fine-tuning to the end result;
- cross-lingual studies showing that a technique X is only successful for a certain language or language family;
- broadly applicable recommendations for training/fine-tuning, especially if the X that didn’t work is something that many practitioners would think reasonable to try, and if the demonstration of X’s failure is accompanied by some explanation/hypothesis.
Some examples of insightful negative results papers are listed on the workshop website: https://insights-workshop.github.io/papers
Consider also the program of Insights 2020 (https://insights-workshop.github.io/2020/program/) and 2021 (https://insights-workshop.github.io/2021/program/).
=============
Important Dates
=============
- First CFP: December 20, 2021
- Second CFP: January 26, 2022
- Submission due: February 28, 2022
- Submission due for papers reviewed through ACL Rolling Review: March 21, 2022
- Notification of acceptance: March 26, 2022
- Camera-ready papers due: April 10, 2022
- Workshop: May 26 2022
All deadlines are 11.59 pm UTC -12h (“anywhere on Earth”).
==========
Submissions
==========
Submission is electronic, using the OpenReview conference management system. Submission link: https://openreview.net/group?id=aclweb.org/ACL/2022/Workshop/Insights
The workshop will accept short papers (up to 4 pages, excluding references), as well as 1-2 page non-archival abstract submissions for papers published elsewhere (e.g. in one of the main conferences or in non-NLP venues). The goal of this event is to stimulate a meaningful community-wide discussion of the deep issues in NLP methodology, and the authors of both types of submissions will be welcome to take part in our get-togethers.
The workshop will run its own review process, and papers can be submitted directly to the workshop by February 28, 2022. It is also possible to submit a paper accompanied with reviews from the ACL Rolling Review system by March 21, 2022. The submission deadline for ARR papers follows the ACL RR calendar [https://aclrollingreview.org/dates](https://aclrollingreview.org/dates).
Both research papers and abstracts must follow the ACL two-column format. Official style sheets:
- Overleaf template: https://www.overleaf.com/read/crtcwgxzjskr
- Latex/Word template download: https://github.com/acl-org/ACLPUB/tree/master/templates
Please do not modify these style files, nor should you use templates designed for other conferences. Submissions that do not conform to the required styles, including paper size, margin width, and font size restrictions, will be rejected without review.
=========
Authorship
=========
The author list for submissions should include all (and only) individuals who made substantial contributions to the work presented. No changes to the order or composition of authorship may be made after the paper submission deadline.
====================
Citation and Comparison
====================
You are expected to cite all refereed publications relevant to your submission, but you may be excused for not knowing about all unpublished work (especially work that has been recently posted and/or is not widely cited).
In cases where a preprint has been superseded by a refereed publication, the refereed publication should be cited instead of the preprint version.
Papers (whether refereed or not) appearing less than 3 months before the submission deadline are considered contemporaneous to your submission, and you are therefore not obliged to make detailed comparisons that require additional experimentation and/or in-depth analysis.
For more information, see the ACL Policies for Submission, Review, and Citation.
===========================
Multiple Submission Policy
===========================
The workshop cannot accept work for publication or presentation that will be (or has been) published elsewhere and that have been or will be submitted to other meetings or publications whose review periods overlap with that of Insights. Any questions regarding submissions can be sent to insights-workshop-organizers@googlegroups.com.
If the paper has been rejected from another venue, the authors will have the option to provide the original reviews and the author response. The new reviewers will not have access to this information, but the organizers will be able to take into account the fact that the paper has already been revised and improved.
==========
Ethics Policy
==========
*ACL workshops follow the conference guidelines for honouring the ACM Code of Ethics (https://www.acm.org/code-of-ethics). Per conference guidelines, a paper that may raise ethical issues needs to explicitly discuss them, and that discussion will be taken into account in the review process. We suggest considering the questions in the ARR responsible NLP checklist (https://aclrollingreview.org/responsibleNLPresearch/), in particular the questions related to the motivation for the choice of data to demonstrate the given negative result. We encourage realistic discussions of limitations. To help the reviewers quickly find which information is in your paper and where, you may choose to provide the filled-in checklist together with your submission.
Specific to the topic of negative results is the problem of revisiting published papers that cannot be reproduced. In most cases it comes down to general methodological problems, but if you have reason to believe the unreproducible result was deliberately fabricated, that should be discussed.
============
Reproducibility
============
Publishing negative results is not easy, partly because the author has the burden of proof that something truly does not work, rather than is caused by a bug.
We encourage the authors to link code repositories in the camera-ready versions, and consider the questions in the reproducibility section of the ARR responsible NLP checklist: https://aclrollingreview.org/responsibleNLPresearch/
At submission time, each submission can be accompanied by one PDF appendix for the paper, one PDF for prior reviews and author response, the optional checklist, one .tgz or .zip archive containing software, and one.tgz or .zip archive containing data (all fully anonymized). The appendix can document preprocessing decisions, model parameters, feature templates, lengthy proofs or derivations, pseudocode, sample system inputs/outputs, and other details that are necessary for the exact replication of the work. However, the paper submissions need to remain fully self-contained, as the supplementary materials are completely optional, and reviewers are not asked to review or download them. If you choose to provide the filled-in checklist together with the submission, it would help the reviewers and chairs to be able to quickly assess what information is provided and where.
===============
Anonymity Period
===============
We are not enforcing any anonymity period.
===========
Presentation
===========
All accepted papers must be presented at the workshop to appear in the proceedings. Authors of accepted papers must notify the program chairs by the camera-ready deadline if they wish to withdraw the paper. At least one author of each accepted paper must register for the workshop.
Previous presentations of the work (e.g. preprints on arXiv.org) should be noted in a footnote in the camera-ready version (but not in the anonymized version of the paper).
The workshop will take place on May 26 or 27 2022. The workshop will be hybrid with both in-person and virtual presentations.
==========
Contact info
==========
The third iteration of the workshop is organized by Shabnam Tafreshi, Joao Sedoc, Anna Rogers, Aleksandr Drozd and Anna Rumshisky. Any questions regarding the workshop can be sent to insights-workshop-organizers@googlegroups.com.