2019Q3 Reports: Program Chairs

Program Committee

Organising Committee

General Chair

Jill Burstein, Educational Testing Service, USA

Program Co-Chairs

Christy Doran, Interactions LLC, USA
Thamar Solorio, University of Houston, USA

Industry Track Co-chairs

Rohit Kumar
Anastassia Loukina, Educational Testing Service, USA
Michelle Morales, IBM, USA

Workshop Co-Chairs

Smaranda Muresan, Columbia University, USA
Swapna Somasundaran, Educational Testing Service, USA
Elena Volodina, University of Gothenburg, Sweden

Tutorial Co-Chairs

Anoop Sarkar, Simon Fraser University, Canada
Michael Strube, Heidelberg Institute for Theoretical Studies, Germany

System Demonstration Co-Chairs

Waleed Ammar, Allen Institute for AI, USA
Annie Louis, University of Edinburgh, Scotland
Nasrin Mostafazadeh, Elemental Cognition, USA

Publication Co-Chairs

Stephanie Lukin, U.S. Army Research Laboratory
Alla Roskovskaya, City University of New York, USA

Handbook Chair

Steve DeNeefe, SDL, USA

Student Research Workshop Co-Chairs & Faculty Advisors

Sudipta Kar, University of Houston, USA
Farah Nadeem, University of Washington, USA
Laura Wendlandt, University of Michigan, USA
Greg Durrett, University of Texas at Austin, USA
Na-Rae Han, University of Pittsburgh, USA

Diversity & Inclusion Co-Chairs

Jason Eisner, Johns Hopkins University, USA
Natalie Schluter, IT University, Copenhagen, Denmark

Publicity & Social Media Co-Chairs

Yuval Pinter, Georgia Institute of Technology, USA
Rachael Tatman, Kaggle, USA

Website & Conference App Chair

Nitin Madnani, Educational Testing Service, USA

Student Volunteer Coordinator

Lu Wang, Northeastern University, USA

Video Chair

Spencer Whitehead, Rensselaer Polytechnic Institute, USA

Remote Presentation Co-Chairs

Meg Mitchell, Google, USA
Abhinav Misra, Educational Testing Service, USA

Local Sponsorship Co-Chairs

Chris Callison-Burch, University of Pennsylvania, USA
Tonya Custis, Thomson Reuters, USA

Local Organization

Priscilla Rasmussen, ACL

Area Chairs

FORMATTING TBD

Biomedical NLP & Clinical Text Processing Bridget McInnes, Virginia Commonwealth University, USA Byron C. Wallace, Northeastern University, USA Cognitive Modeling – Psycholinguistics Serguei Pakhomov, University of Minnesota, USA Emily Prud’hommeaux, Boston College, USA Dialog and Interactive systems Nobuhiro Kaji, Yahoo Japan Corporation, Japan Zornitsa Kozareva, Google, USA Sujith Ravi, Google, USA Michael White, Ohio State University, USA Discourse and Pragmatics Ruihong Huang, Texas A&M University, USA Vincent Ng, University of Texas at Dallas, USA Ethics, Bias and Fairness Saif Mohammad, National Research Council Canada, Canada Mark Yatskar, University of Washington, USA Generation He He, Amazon Web Services, USA Wei Xu, Ohio State University, USA Yue Zhang, Westlake University, China Information Extraction Heng Ji, Rensselaer Polytechnic Institute, USA David McClosky, Google, USA Gerard de Melo, Rutgers University, USA Timothy Miller, Boston Children’s Hospital, USA Mo Yu, IBM Research, USA Information Retrieval Sumit Bhatia, IBM’s India Research Laboratory, India Dina Demner-Fushman, US National Library of Medicine, USA Machine Learning for NLP Ryan Cotterell, Johns Hopkins University, USA Daichi Mochihashi, The Institute of Statistical Mathematics, Japan Marie-Francine Moens, KU Leuven, Belgium Vikram Ramanarayanan, Educational Testing Service, USA Anna Rumshisky, University of Massachusetts Lowell, USA Natalie Schluter, IT University of Copenhagen, Denmark Machine Translation Rafael E. Banchs, HLT Institute for Infocomm Research A*Star, Singapore Daniel Cer, Google Research, USA Haitao Mi, Ant Financial US, USA Preslav Nakov, Qatar Computing Research Institute, Qatar Zhaopeng Tu, Tencent, China Mixed Topics Ion Androutsopoulos, Athens Univ. of Economics and Business, Greece Steven Bethard, University of Arizona, USA Multilingualism, Cross lingual resources Željko Agić, IT University of Copenhagen, Denmark Ekaterina Shutova, University of Amsterdam, Netherlands Yulia Tsvetkov, Carnegie Mellon University, USA Ivan Vulic, Cambridge University, UK NLP Applications T. J. Hazen, Microsoft, USA Alessandro Moschitti, Amazon, USA Shimei Pan, University of Maryland Baltimore County, USA Wenpeng Yin, University of Pennsylvania, USA Su-Youn Yoon, Educational Testing Service, USA Phonology, Morphology and Word Segmentation Ramy Eskander, Columbia University, USA Grzegorz Kondrak, University of Alberta, Canada Question Answering Eduardo Blanco, University of North Texas, USA Christos Christodoulopoulos, Amazon, USA Asif Ekbal, Indian Institute of Technology Patna, India Yansong Feng, Peking University, China Tim Rocktäschel, Facebook, USA Avi Sil, IBM Research, USA Resources and Evaluation Torsten Zesch, University of Duisburg-Essen, Germany Tristan Miller, Technische Universität Darmstadt, Germany Semantics Ebrahim Bagheri, Ryerson University, Canada Samuel Bowman, New York University, USA Matt Gardner, Allen Institute for Artificial Intelligence, USA Kevin Gimpel, Toyota Technological Institute at Chicago, USA Daisuke Kawahara, Kyoto University, Japan Carlos Ramisch, Aix Marseille University, France Sentiment Analysis Isabelle Augenstein, University of Copenhagen, Denmark Wai Lam, The Chinese University of Hong Kong, Hong Kong Soujanya Poria, Nanyang Technological University, Singapore Ivan Vladimir Meza Ruiz, UNAM, Mexico Social Media Dan Goldwasser, Purdue University, USA Michael J. Paul, University of Colorado Boulder, USA Sara Rosenthal, IBM Research, USA Paolo Rosso, Universitat Politècnica de València, Spain Chenhao Tan, University of Colorado Boulder, USA Xiaodan Zhu, Queen’s University, Canada Speech Keelan Evanini, Educational Testing Service, USA Yang Liu, LAIX Inc, USA Style Beata Beigman Klebanov, Educational Testing Service, USA Manuel Montes, Instituto Nacional de Astrofísica, Óptica y Electrónica, Mexico Joel Tetreault, Grammarly, USA Summarization Mohit Bansal, University of North Carolina Chapel Hill, USA Fei Liu, University of Central Florida, USA Ani Nenkova, University of Pennsylvania, USA Tagging, Chunking, Syntax and Parsing Adam Lopez, University of Edinburgh, Scotland Roi Reichart, Technion – Israel Institute of Technology, Israel Agata Savary, University of Tours, France Guillaume Wisniewski, Université Paris Sud, France Text Mining Kai-Wei Chang, University of California Los Angeles, USA Anna Feldman, Montclair State University, USA Shervin Malmasi, Harvard Medical School, USA Verónica Pérez-Rosas, University of Michigan, USA Kevin Small, Amazon, USA Diyi Yang, Carnegie Mellon University, USA Theory and Formalisms Valia Kordoni, Humboldt University Berlin, Germany Andreas Maletti, University of Stuttgart, Germany Vision, Robotics and other grounding Francis Ferraro, University of Maryland Baltimore County, USA Vicente Ordóñez, University of Virginia, USA William Yang Wang, University of California Santa Barbara, USA

Main Innovations

• Conference theme

The CFP made a special request for papers addressing the tension between data privacy and model bias in NLP, including: using NLP for surveillance and profiling, balancing the need for broadly representative data sets with protections for individuals, understanding and addressing model bias, and where bias correction becomes censorship. The three invited speakers were all selected to tie into the theme, and a Best Thematic Paper was selected.

• Land Acknowledgement

Similar to what has been done in recent *CL conferences, the opening session included a land acknowledgement to recognize and honor Indigeneous Peoples.

• Video Poster Highlights

This year included one minute slides with pre recorded audio that showcase the posters to be presented that day. The goal was to provide more visibility to posters. These were shown during the welcome reception, breakfast and breaks.

• Remote Presentations

Remote presentations were supported for both talks and posters, via an application form to the committee.

• Diversity & Inclusion Organization

The new Diversity & Inclusion team piloted a number of new initiatives including:

• • additional questions on the registration form to identify any accommodations
  • preferred pronouns (optionally) added to badges
  • I’m hiring/I’m looking for a job/I’m new badge stickers

<bunch of others, pull from their report>?

• Submissions

This year we followed a two-stage submission process, in which abstracts were due one week before full papers. Our goal was to get a head start on assigning papers to areas, and recruiting additional area chairs where submissions exceeded our predicted volume.

• • Pro: early response to areas with larger than predicted number of papers
  • Con: too much overhead for PCs, as authors repeatedly contacted chairs to request that papers be moved between long and short, or asked about changes to authorship, titles and abstracts.

• Full papers available for bidding: reviewers loved it, authors did not

An overview of statistics

Authors were permitted to switch format (long/short) when they submitted the full papers, so the total in the chart below uses 2271 as the total number of submissions, discounting the 103 that never submitted a full paper in the second phase. Seventy nine papers were desk-rejected due to anonymity, formatting, or dual-submission violations; 456 papers withdrawn prior to acceptance decisions being sent, although some were withdrawn part way through the review process; and an additional 11 papers were withdrawn after acceptance notifications had been sent. Keeping the acceptance rate consistent with past years meant 5 parallel tracks were needed to fit more papers into 3 days--as the conference grows, decisions will have to be made about continuing to add more tracks, adding more days to the main conference, or lowering the acceptance rate. The overall technical program consisted of 423 main conference papers, plus 9 TACL papers, 23 SRW papers, 28 Industry papers, and 24 demos. The TACL and SRW papers were integrated into the program, and marked SRW or TACL accordingly.

Acceptance break-down: \begin{table}[h] \centering \begin{tabular}{|l|l|l|l|l|} \hline

&\textbf{Long}& \textbf{Short} &\textbf{Total} & \textbf{TACL}\\ \hline

Reviewed & 1067 & 666 & 1733 & \\ Accepted as talk & 140 & 72 & 212 & 4\\ Accepted as poster & 141 & 70 & 211 & 5\\ Total Accepted & 281 (26.3\%) & 142 (21.3\%) & 423 (24.4\%) & 9\\ \hline \end{tabular} \end{table}

== Detailed statistics by area

Area Long (%) Short (%) Area Long (%) Short (%) Bio and clinical NLP 7 (57) 28 (17) Question Answering 73 (36) 41 (17) Cognitive modeling 24 (29) 14 (14) Resources and Evaluation 33 (27) 20 (20) Dialog and Interactive systems 64 (20) 18 (27) Semantics 80 (13) 42 (11) Discourse and Pragmatics 38 (21)

      11 (36)

Sentiment Analysis 32 (28) 40 (20) Ethics, Bias and Fairness 16 (25) 12 (50) Social Media 44 (18) 41 (36) Generation 46 (14) 19 (23) Speech 19 (31) 9 (33) Information Extraction 46 (28) 16 (12) Style 24 ( (25) 16 (25) Information Retrieval 22 (22) 13 (30) Summarization 22 (27) 28 (28) Machine Learning for NLP 100 (29) 22 (22) Syntax 36 (52) 54 (13) Machine Translation 49 (30) 53 (18) Text Mining 101 (18) 29 (24) Multilingual NLP 43 (25) 28 (10) Theory and Formalisms 12 (58) 12 (16) NLP Applications 60 (30) 41 (17) Vision & Robotics 41 (12) 22 (36) Phonology 24 (33)

      24 (25)

== Conference tracks The Industry Track, in its second year, had 28 accepted papers (10 oral and 18 posters, acceptance rate: ~28%), and ran a lunchtime Careers in Industry panel which was very well attended. Panelists were Judith Klavans, Yunyao Li, Owen Rambow, and Joel Tetreault and the moderator was Phil Resnik.

The Student Research Workshop had 23 accepted papers, distributed throughout the conference, and 19 submissions received pre-submission mentoring. For the first time, both archival and non-archival submissions were offered, meaning that authors who opted for the non-archival version will not have a paper available in the archive and are free to publish elsewhere.

There were 25 accepted Demos, which were spread across several of the poster sessions.

Review Process

Issued a wide call for volunteers for Area Chairs (ACs) and reviewers. Volunteers were scanned by PCs and assigned ACs/reviewer roles. PCs created 25 specific areas + one for “Mixed Topics” and assigned at least 2 ACs per topic area. After abstract deadline we added more ACs to teams with larger than predicted submissions

We used a hybrid reviewing form, combining elements of the EMNLP 2018, NAACL-HLT 2018 and ACL 2018, with a 6-point overall rating scale so there was no “easy out” mid-point, distinct sections of summary, strengths and weaknesses to make easy to scan and compare relevant sections, and the minimum length feature of START enabled to elicit more consistently substantive content for the authors.

Authors were blind to Area Chairs Review assignment Criteria: Fairness, Expertise, Interest Method: area chair expertise + Toronto Paper Matching System (TPMS) + reviewer bids Many reviewers did not have TPMS profiles Goal was no more than 5 papers per reviewer, some reviewers agreed to handle more First-round accept/reject suggestions were made by area chairs Final decisions were made by the program chairs

No author response: due to time constraints and finding from NAACL 2018 that it had little impact. Authors were unhappy about this, they really want to be able to respond to reviews. Video Poster highlights: instead of 1-minute madness, A/V failures have made it hard to assess effectiveness. SRW papers integrated into sessions: positive feedback from participants, better experience for students Did not repeat Test of Time awards from 2018--should this happen every N years to allow for sliding window?

Recruiting area chairs (ACs) and reviewers

Response Area Chair Reviewer Female 24.4 25.2 Male 73 71.7 Prefer not to answer 2.6 3.1

Assigning papers to areas and reviewers

Assignment to areas was based on keywords and manual inspection of the paper. Assignment of papers to reviewers followed a combination of TPMS, reviewer bidding, and manual tweaking.

Deciding on the reject-without-review papers

Our process for identifying desk rejects has been very similar to what other PCs have done in the past. First, the area chairs check their batch of assigned papers and report any issues to us. As the reviewing begins, reviewers may also identify issues that were not caught by ACs, which they flag up to ACs or directly to PCs. We then review each of these issues and make a final decision, to ensure that papers are handled consistently. This means each paper is reviewed for non-content issues by at least three people. The major categories for desk rejects are: Violations to the dual submission policy specified in the call for papers Violations to the anonymity policy as specified in the call for papers “Format cheating” submissions not following the clearly stated format and style guidelines either in LaTeX or Word (thanks to Emily and Leon for introducing the concept). As of February 7th, out of 2378 submissions, there were 44 rejections for format issues, 24 for anonymity violations, and 11 for dual submissions. This means that a total of 3% of the submissions were desk-rejected.

A large pool of reviewers

Similar to what other PCs have done in the past, we distributed a wide call for volunteers to recruit the Area Chairs and Reviewers--we seeded the areas with volunteers who responded, and then Area Chairs filled out the remainder of their respective committees. Our goal was to ensure greater diversity by including in each area some participants who may not have been previously involved, and therefore would not have been invited if the committees were built from lists of previous reviewers. 390 of 1321 reviewers were reviewing for NAACL for the first time.

Structured review form

We used a hybrid reviewing form, combining elements of the EMNLP 2018, NAACL-HLT 2018 and ACL 2018, with a 6-point overall rating scale so there was no “easy out” mid-point, distinct sections of summary, strengths and weaknesses to make easy to scan and compare relevant sections, and the minimum length feature of START enabled to elicit more consistently substantive content for the authors. This received excellent feedback from authors but which some reviewers complained about and others outright circumvented via html tags or repeated filler content.

Abstract Submissions

This year we followed a two-stage submission process, in which abstracts were due one week before full papers. Our goal was to get a head start on assigning papers to areas, and recruiting additional area chairs where submissions exceeded our predicted volume. Relative to the projected numbers from NAACL-HLT 2018, several areas received a higher-than-predicted number of submissions: Biomedical/Clinical, Dialogue and Vision. Text Mining ended up with the overall largest number of submissions.

Review process

Authors were permitted to switch format (long/short) when they submitted the full papers, so the total in the chart below uses 2271 as the total number of submissions, discounting the 103 that never submitted a full paper in the second phase. Seventy nine papers were desk-rejected due to anonymity, formatting, or dual-submission violations; 456 papers withdrawn prior to acceptance decisions being sent, although some were withdrawn part way through the review process; and an additional 11 papers were withdrawn after acceptance notifications had been sent. Keeping the acceptance rate consistent with past years meant we needed 5 parallel tracks to fit more papers into 3 days--as the conference grows, decisions will have to be made about continuing to add more tracks, adding more days to the main conference, or lowering the acceptance rate. The overall technical program consists of 423 main conference papers, plus 9 TACL papers, 23 SRW papers, 28 Industry papers, and 24 demos. The TACL and SRW papers are integrated into the program, and are marked SRW or TACL accordingly.

To be added : X reviews were received by the end of the review period, Y others within the next week.; Importance of double blind reviewing

Best paper awards

• Best Thematic Paper:

What’s in a Name? Reducing Bias in Bios Without Access to Protected Attributes

Alexey Romanov, Maria De-Arteaga, Hanna Wallach, Jennifer Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Geyik, Krishnaram Kenthapadi, Anna Rumshisky and Adam Kalai

• Best Explainable NLP Paper:

CNM: An Interpretable Complex-valued Network for Matching

Qiuchi Li, Benyou Wang and Massimo Melucci

• Best Long Paper

BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova

• Best Short Paper

Probing the Need for Visual Context in Multimodal Machine Translation

Ozan Caglayan, Pranava Madhyastha, Lucia Specia and Loïc Barrault

• Best Resource Paper

CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge

Alon Talmor, Jonathan Herzig, Nicholas Lourie and Jonathan Berant

Presentations

• Long-paper presentations: 22 sessions in total (4 sessions in parallel), duration: 15 minutes for talk + 3 minutes for questions + 2 dedicated Industry Track sessions
• Short-paper presentations: 12 sessions in total (4 sessions in parallel), duration: 12 minutes for talk + 3 minutes for questions
• Best-paper presentation: 1 session at the end of the last day
• Posters: 8 sessions in total (1 session in parallel with every non-plenary talk session) + 1 dedicated Industry Poster session

Issues and recommendations

TBD