Author response

The value of the author response mechanism is frequently debated in our field and can be a source of stress for authors. On the one hand, when our work is being reviewed by others, it can feel helpless to not have the opportunity to respond to those reviews. On the other hand, there is the perennial question about whether author responses ever “help” (in the sense of taking a paper over the line to “accept” from “reject”). (On that point, see this very thoughtful analysis by Hal Daumé III for the process for NAACL 2013.) And finally there is the issue that author responses must be turned around in a short time and can be tricky to write: How to strike the right tone (firm, polite, confident; not pleading or angry) especially when we might still be feeling the sting of negative reviews. As reviewers, we have seen both very effective author responses (expressing gratitude for feedback and pointing out sources of misunderstanding) and very ineffective ones (pure vitriol, or long lists of promises of what will be accomplished before the camera-ready version).

In light of all of this, what we settled on for COLING 2018 is an optional author response to be seen by the area chairs only – and not the reviewers. Thus we are providing authors with the opportunity to flag reviewer misunderstandings for area chairs and to answer questions raised by reviews. The latter should only be done when the information is already available and can be indicated in a short statement (e.g. “Indeed, we did set the random seed and will include this information in the camera ready” but not “That is an interesting idea for a further experiment, we will run that one and include the numbers in the camera ready”). We also note that author response is optional and area chairs will not read anything into the lack of an author response.

Author response will run from 20-24 April.

Why this route? Well, the quantitative evidence is that pointing out reviewer mistakes rarely leads to a change in scores. The folk knowledge has been for some time that responses are really used by ACs to detect misaligned reviews. So rather than encourage an intrinsically difficult communication that has had little to no effect in the past, we instead divert the replies to go to the authoritative party they are relevant to. This gives a little extra work for ACs, but as they’re acting in pairs and areas are roughly the same compact size, our hope is that time can be spent more on working out the dialog around a paper and less on administering a huge set of authors and reviewers.

Lessons Learned

The role of PC chair is interesting in many ways. It provides a perhaps unparalleled opportunity to influence the way in which research is approached and presented in our field. For COLING 2018, we have been taking this responsibility very seriously and working hard, through both decisions for the review process and the publicization of those ideas in this blog, the push the field in directions that we believe will be fruitful, including stronger interdisciplinarity and more reproducibility.

On the flip side, the role of PC chair comes with some serious downsides. One is the heart rending process of deciding on and then informing authors of desk rejects. We did our utmost to do this as fairly as humanly possible, starting with publicizing our desk reject policy. We hoped that that move would reduce the number of desk rejects, and it may have, but there were still a handful of papers rejected without review under the policy.

The most common reason for a desk reject by a long way was the paper’s length (ie. documents were submitted with more than 9 content pages). Papers in the completely incorrect template were also desk rejected, as were those with squashed line spacing, reduced font size, removed author boxes, and so on. Other reasons for desk rejection were bad anonymisation; some papers, for example, linked to the author’s private github repository. This is the sort of thing that can really wait until camera ready. All papers sent in other templates were desk rejected (we saw e.g. NAACL, ACL, NIPS formats). One paper was rejected for breaking the arXiv embargo period, having been published there fewer than 30 days before the COLING deadline. No edits were allowed after the deadline had passed. This was a very unpleasant process overall and we can only make a plea to authors to follow the guidelines so that work gets the attention it needs, instead of rejection without feedback. That way there don’t have to be any desk rejects at all. They are often desperately unpleasant to send, and probably even worse to receive.

In this blog post, we wanted to briefly reflect on what we have learned about the kind of practices that put people in the corners that lead to the kind of mistakes that result in desk rejects. In general, we see that there is a culture of last-minutism in our field. Deadlines can inspire people to get things done that otherwise seem impossible, but doing things in a rush also has downsides. Here are some DOs and DON’Ts of paper submission that we hope will spare people some pain in the future:

  • Do access the submission system early, so you know what awaits.
  • Do read the CFP carefully. Such documents can be intimidating, especially for first-time submitters, but the information there all has a purpose, and it’s easier to make use of if you get it early.
  • Don’t leave submitting your final paper until the absolute last minute. If something goes wrong (e.g. submitting the wrong pdf, losing your internet connection), you’ll have missed the deadline. This happens regularly and is wasteful. Sometimes you might not find out it was the wrong PDF until after the deadline, or might be so rushed that the paper spills over the page limit unnoticed. This means the hard work has to wait for another conference.

And finally a couple of thoughts on interacting with PC chairs, especially in large conferences:

  • Please don’t ask the PC chairs to upload a PDF for you after the deadline. The deadline is a deadline. Asking for it to be bent is asking the PC chairs to not apply policies evenly and fairly.
  • Do be aware that the PC chairs in a conference this size are communicating with ~1000 authors and ~1000 reviewers, and keep that in mind as you make requests.

COLING 2018 Submissions Overview

We’ve had a successful COLING so far, with over a thousand papers submitted, covering a variety of areas. In total, 1017 papers were submitted to the main conference, all full-length.

Each submitted paper had a distinct type assigned by the authors, that affects how it is reviewed. These were developed based on our earlier blog post on paper types. The “NLP Engineering Experiment paper” was unsurprisingly the dominant type, though only made up for 65% of all papers. We were very happy to receive 25 survey papers, 31 position papers, and 35 reproduction papers—as well as a solid 106 resource papers and a strong showing of 163 computationally-aided linguistic analysis papers, the second largest contingent.

Some papers were withdrawn or desk rejected before review began in earnest. Between ACs and PC co-chairs, in total, 32 papers were rejected without review. Excluding desk rejects, so far 41 papers have been withdrawn from consideration by the authors.

Allocating papers to areas gave each area a mean and median of 27 papers. The largest area has 31 papers and the smallest 19. We interpret this as indicating that area chairs will not be overloaded, leading to better review quality and interpretation.

Author survey results

Shortly after the submission deadline, we sent out a survey to our authors, with the goal of better understanding how our outreach was working.

Respondents

We sent the notification of the survey via START to all corresponding authors (so roughly 1000 people) and asked them to share it with co-authors. The survey recorded 434 total responses, which is a pretty satisfying response rate!

Of those 434, 302 (69.6%) indicated that they were submitting to COLING for the first time, and 101 (23.3%) to a major NLP conference for the first time.

Outreach

We asked how people first found out about COLING 2018. The most popular response was “Web search” (44.2%), followed by “Call for Papers sent over email (e.g. corpora mailing list, ACL mailing list)” (35.9%), then “Other” (12.4%) and “Social media” (7.4%).  The “Other” answers included word-of-mouth, knowing to expect COLING to come around in 2018, and websites that aggregate CFPs.

Paper types

We wanted to find out if people were aware of the paper types (since this is relatively unusual in our field) before submitting their papers, and if so, how they found out. Most—349 (80.4%)—were aware of the paper types ahead of time.  Of these, the vast majority (93.4%) found out about the paper types via the Call for Papers. Otherwise, people found out because someone else told them (7.4%), via our Twitter or Facebook feeds (6.0%), or via our blog (3.7%).

We also asked if it was clear to authors which paper type was appropriate for their paper and if they think paper types are a good idea. The answers in both cases were pretty strongly positive: 78.8% said it was clear and 91.0% said it was a good idea. (Interestingly, 74 people who said it wasn’t clear which paper type was a good fit for theirs nonetheless said it was a good idea, and 21 people who thought it was clear which paper type fit nonetheless said it wasn’t.)

Writing mentoring program

We wanted to know if our authors were aware of the writing mentoring program, and for those who were but didn’t take advantage of it, why not. 277 respondents (63.8%) said they were aware of it. The most common reason chosen for not taking advantage of it was “I didn’t/couldn’t have a draft ready in time.” (150 respondents), followed by “I have good mentoring available to me in my local institution” (97 respondents). The other two options available in that check-all-that-apply question were “I have a lot of practice writing papers already” (74 respondents) and “Other” (10). Alas, a few people indicated that they only discovered it too late.

Other channels

We have been putting significant effort into getting information out about our process, but still worry that the channels we’re using aren’t reaching everyone. We asked “What other channels would you like to see information like this publicized on?” referring specifically to the paper types. Most people did not respond, or indicated that what we’re doing is enough. Other responses included: LINGUIST List, LinkedIn, Instagram, ResearchGate, and email. Ideas for email include creating a conference-specific mailing list that people can subscribe to and sending out messages to all email addresses registered in START.

We include these ideas here for posterity (and the benefit of future people filling this role). We have used LinkedIn and Weibo in a limited capacity and are using Twitter and Facebook. Adding additional social media (Instagram) sounds plausible, but is not in our plans for this year. An email list that people could opt into for updates makes a lot of sense, though there’s still the problem of getting the word out about that list. Perhaps a good way to do that would be to include that info in the CFP (starting from the first CFP). Emailing everyone through START may not be feasible (depending on START’s email privacy policy) and at any rate wouldn’t help reach those who have never submitted to a compling/NLP conference before.

Blog readership

Of course we wanted to know if our authors are reading this blog.  44.7% of respondents weren’t aware of the blog (prior to being asked that question!), 15.0% had found it only recently, 24.9% had been aware of it for at least a month but less than 6, and 15.4% indicated that they’ve been aware of it for at least 6 months. 9.2% of respondents read (almost) everything we post, 32.0% read it sometimes, and the remainder don’t read it or read it only rarely.

We also wanted to know if the PC blog helped our authors to understand our submission process or shape their submissions to COLING 2018. 22.8% indicated “Yes, a lot!” and 28.1% “Yes, a little”. On the no side, 22.1% chose “No, not really” and 27.0% “No, not at all”. “Yes, a lot!” people, we’re doing this for you 🙂

 

Outstanding Mentors

The COLING 2018 writing mentoring program went extremely well—we are grateful to all of the mentors who volunteered their time to provide thoughtful comments to the authors who participated. Furthermore, the prompts we used in the writing mentoring form (listed in the description of the program) were effective in eliciting useful feedback for authors.

There is great willingness in our field to participate from the mentoring side.  Over 100 mentors signed up, which means we could have provided mentoring for even more papers than we did. It seems that the biggest hurdle to success for such a program is getting the word out to those who would most likely benefit from it. (We’ve got another blog post in the works about outreach & responses to our author survey.)

Reviewing the work of the mentors to find those to recognize as outstanding mentors was inspiring—and the task of choosing difficult—because so many did such a great job. Even if the mentored papers aren’t ultimate accepted to COLING, the authors who received mentoring will have benefited from thoughtful, constructive feedback on their work which we hope will inform both future writings on the same topic and perhaps even their approach to writing on other topics.

Against that background, the following mentors distinguished themselves as particularly outstanding:

  • Kevin Cohen
  • Carla Parra Escartín
  • David Mimno
  • Emily Morgan
  • Irina Temnikova
  • Jennifer Williams

Thank you to all of our mentors!

Reviewing in an interdisciplinary field

The process of peer review, when functioning at its best, ensures that work that is published in archival venues is carefully vetted, such that the results likely to be reliable and the presentation of those results interpretable by scholars in the field at large. But what does “peer review” mean for an interdisciplinary field? Who are the relevant “peers”? We believe that for both goals—vetting reliability of results and vetting readability of presentation—reviewing in an interdisciplinary field ideally involves reviewers coming from different perspectives.

In our particular context, we had the added (near-)novelty of our keyword-based area assignment system. (Near-novelty, because this was pioneered by NAACL 2016.) This means our areas are not named, but rather emerge from the clusters of both reviewer interests and paper topics. On the upside, this means that really popular topics (“semantics”, or “MT”) can be spread across areas, such that we don’t have area chairs scrambling for large numbers of additional reviewers. On the downside, the areas can’t be named until the dust has settled, so reviewers don’t necessarily have a clear sense of which area (in the traditional sense) they are assigned to. In addition, interests that were very popular and therefore not highly discriminative (e.g. “MT”) weren’t given very much weight alone by the clustering algorithm.

During the bidding process, we had a handful of requests from reviewers to change areas (really a small number, considering the overall size of the reviewing pool!). These requests came in three types. First, there were a few who just said: There’s relatively little in this area that I feel qualified to review, could I try a different area? In all cases, we were able to find another area that was a better match.

Second, there were reviewers who said “My research interest is X, and I don’t see any papers on X in my area. I only want to review papers on X.” We found these remarks a little surprising, as our understanding of the field is that it is not a collection of independent areas that don’t inform each other, but rather a collection of ways of looking at the same very general and very pervasive phenomenon: human language and the ways in which it can be processed by computers. Indeed, we structured the keywords into multiple groups—targets, tasks, approaches, languages and genres—to increase intersection on at least a few areas of any given reviewer’s expertise. We very much hope that the majority of researchers in our field read outside their specific subfield and are open to influence from other subfields on their own.

The third type was reviewers, typically of a more linguistic than computational orientation, who expressed concern that because they aren’t familiar with the details of the models being used, they wouldn’t be able to review effectively. To these reviewers, we pointed out that it is equally important to look critically at the evaluation (what data is being used and how) and the relationship of the work to the linguistic concepts it is drawing on. Having reviewers with deep linguistic expertise is critical and both COLING and the authors very much benefit from it.

To create the best cross-field review, then, it helps to examine each of one’s strengths and compare these with the multiple facets presented by any paper. No single reviewer is likely to be expert in every area and aspect of a manuscript; but, there’s a good chance that, as long as some care has been applied to matching, there will be some crossover expertise. Be bold with that expertise. And indeed, the coverage of knowledge that multiple reviewers have is often complementary. As a reviewer, you can bring some knowledge to reviewing at least one aspect of a paper—more so than others sharing the workload—even if that is not the aspect initially expected.