2021Q3 Reports: EMNLP 2020

Overview

EMNLP 2020 was held 15-20 November 2020, with the main conference running 16-19 November, and tutorials held before the conference, on 15 November and workshops afterwards, on 19-20 November.

Regarding the format of EMNLP 2020, by late spring 2020, with the world in the midst of the COVID pandemic and no sense as to when (or how) it might end, the EMNLP 2020 Committee decided that EMNLP 2020 should be held virtually. It made such a request to the ACL Manager (Priscilla Rasmussen) and the ACL Treasurer (David Yarowsky), who were able to comply by renegotiating ACL's contract with the Barcelo Bavaro Grand Resort (Dominican Republic) to host EMNLP there in November 2021, rather than 2020. While this seemed a good solution because no site had yet been arranged for EMNLP 2021, we do not yet know its full consequences, as travel in 2021 is still unlikely to be completely unconstrained.

In previous *ACL and EMNLP conferences, the unsung heroes of the Organizing Committee were the members of the Publications Committee, who are responsible for producing typographically error-free copies of all the papers for the conference publication. While they are still critical to the conference (and still unsung), in 2020, with the forced adoption of a virtual format for EMNLP, the circle of unsung heroes has expanded to include those committees tasked with designing and running the virtual infra-structure of the conference. Here, we were indebted to the Virtual Infra-structure Chairs of EMNLP 2020 (Yang Feng, Yansong Feng, Jan-Christoph Klie, Zhongyu Wong, Eduardo Blanco) and their advisors (Hao Fang and Sudha Rao). I would caution that, going forward, we cannot continue to rely on students and post-docs to provide this service: It is just too much work for people with other responsibilities during the three months prior to the conference, and that is most of us. I suspect this will also be true if and when we are able to run conferences in "hybrid" format.

3359 papers were submitted to EMNLP, of which 752 were accepted. As with other recent *ACL and EMNLP conferences, processing and reviewing them all involved a five-level hierarchy of 3 Programme Chairs, 33 Senior Area Chairs, 196 Area Chairs, 2633 primary Reviewers assisted by 523 secondary Reviewers. Also involved in this process for the first time was a Ethics panel who had to deal with some difficult decisions both for the main conference and at least one of the workshops. Going forward, I regret that this will continue to be an important part of the reviewing process.

For more information, there is a detailed description of the process by which papers were selected and new innovations in publication decisions in the Programme Chairs' Preface (https://aclanthology.org/2020.emnlp-main.0.pdf, pp. vii-x).

EMNLP PROGRAMME

Since EMNLP had presenters and participants from 21 of 24 time zones around the world, the conference, its workshops and tutorials were distributed over three blocks that covered most of the day, with a short 3-hour gap which was an awkward time for almost everyone. As a result, decisions had to be made vis-a-vis plenary sessions, as there was no slot during which all participants were likely to be awake (or to be awake but unlikely to disturb other family members, since essentially everyone was attending from home). Thus participants were unlikely to have attended all the live plenary sessions when they were held.

Authors were encouraged to record a 12-minute presentation of their papers, for which automated captions were provided. We tried to get this finished a week before the conference, as one complaint we took on-board was that at ACL-2020, talks were only posted at the start of the conference, so people didn't have time to hear them. We also had a group of volunteers willing to help authors to correct the automated captions to their papers, but very few authors took us up on this. It was suggested that a Language Model be built based on the ACL Anthology (weighting more recent publications more heavily) which could improve the accuracy of automated captions.

EMNLP WORKSHOPS

The biggest problem for the EMNLP Organizing Committee were the Workshops, since we didn't realize what a problem it would be to allow each Workshop (which vary significantly in size) to chose its own format -- that is, what parts of the Workshop would be live in GatherTown, what parts would be live on Zoom, what would be recorded, how and where posters would be shown, and how would questions and discussion be handled. I don't even want to be reminded of what a mess it was, even though our Workshop Co-Chairs -- Lonneke van der Plas and Jackie Chi Kit Cheung -- handled all the problems with admirable patience and resourcefulness. Our recommendation was to avoid giving Workshop Coordinators so much freedom with their format. A single format can and should be used for all workshops.

WHERE EMNLP 2020 EXCELLED

Two areas in which EMNLP 2020 excelled were (1) Diversity and Inclusion activities, organized and run by its D&I Co-Chairs (Isabelle Augenstein and Chris Brew) and its D&I Student Chairs (Prathyusha Jwalapuram and Murathan Kurfali), and Publicity, run by Anna Rogers and Ruifeng Xu. D&I activities included Affinity Social Group sessions, large-group and small-group Mentoring sessions, Birds of Feather sessions, two undergraduate panels, and Gather.Town sessions for socializing. The D&I Committee also handled awards of broadband usage and child care, which made it possible for more people to participate from around the world. Our Publicity Chairs both broadcast important information in real-time and monitored what people were saying about the conference. I believe they caught and allowed us to solve several problems before they blew up in our face. Again, I recommend having an active and attentive Publicity Chairs.

SPONSORS

Finally, we should thank the sponsors of EMNLP 2020, for each of whom we built a virtual "booth" where they could distribute material and meet with people interested in the possibility of jobs or internships: Bloomberg Engineering, Google Research, Apple, Amazon Science, Baidu, Megagon Labs, Facebook, DeepMind, Grammarly, ByteDance, Zeta Alpha, Babelscape, Naver, Adobe, Hitachi, Salesforce, and USC Viterbi Lab.