2017Q3 Reports: Program Chairs
Innovations
As compared to previous ACL conferences, this year's main innovations were:
- In addition to best paper awards, a larger number of outstanding papers were selected. The original plan was to identify roughly 1-2% of submissions; ultimately 11 papers out of 1290 submissions (0.85%) were identified as outstanding by the awards committee.
- Instead of 20-minute talks, long papers were presented in 15-minute talks, plus 5 minutes for questions and changing speakers. Short papers were given 12 minutes plus 4 for questions and transition.
- Authors were allowed to submit an appendix of any length with their papers. Reviewers were not required to read appendices.
- Area chairs provided meta-reviews for all papers that were not obvious rejections.
- The deadline for short papers was ahead of the deadline for long papers.
- We introduced a new area labeled “Other” to handle papers that did not fit into traditional areas, as well as a few COI papers.
Rationale
Outstanding papers were added because the community is growing and so is the number of papers at each conference, including the number of papers that deserve recognition as being of particular importance and quality.
Appendices were added to help address the replication problem. The relevant passage in the Call for Papers reads as follows:
- ACL 2016 also encourages the submission of supplementary material to report preprocessing decisions, model parameters, and other details necessary for the replication of the experiments reported in the paper. Seemingly small preprocessing decisions can sometimes make a large difference in performance, so it is crucial to record such decisions to precisely characterize state-of-the-art methods.
- Nonetheless, supplementary material should be supplementary (rather than central) to the paper. It may include explanations or details of proofs or derivations that do not fit into the paper, lists of features or feature templates, sample inputs and outputs for a system, pseudo-code or source code, and data. The paper should not rely on the supplementary material: while the paper may refer to and cite the supplementary material and the supplementary material will be available to reviewers, they will not be asked to review or even download the supplementary material. Authors should refer to the contents of the supplementary material in the paper submission, so that reviewers interested in these supplementary details will know where to look.
Meta-reviews were strongly recommended by the NAACL 2015 program co-chairs.
The deadline for short papers was moved ahead of the deadline for long papers to coordinate deadlines with NAACL, such that rejected NAACL submissions could be reworked into ACL submissions.
The “Other” area was introduced to give papers in non-traditional research areas a better chance at fair review. Rather than conventional bidding, the “Other” area chairs were given access to the reviewer pool directly (while bidding was taking place for papers in the rest of the areas), so that these papers were given the first chance at the most appropriate reviewers.
Submissions and Presentations
ACL 2016 received a total of 1288 valid submissions, of which 825 were long papers and 463 were short papers. 21 long papers and 9 short papers were rejected without review due to non-anonymity or formatting issues. The remaining submissions were each assigned to one of 19 areas, and managed by a program committee of 38 area chairs and 884 reviewers (including secondary reviewers indicated on the review forms). 231 (28%) of the 825 qualifying long papers and 97 (21%) of the 463 qualifying short papers were selected for presentation at the conference. Of the accepted long papers, 116 were selected for oral presentation, and 115 for poster presentation. Of the accepted short papers, 49 have oral and 48 have poster presentations. The oral versus poster decision was made based on the recommendations of reviewers, which we took as a noisy signal of the intended audience’s preference of format for each paper.
In addition, 25 TACL papers will be presented at ACL – 24 as talks and one as a poster. Including TACL papers, there will be 189 oral and 163 poster presentations at the main ACL conference. The table below shows the number of reviewed submissions in each area for long and short papers, as well as the number of papers accepted in each area. Approximately 59 short and 52 long papers were withdrawn before review was completed; these are not included in the table.
| area | long reviewed | long accepted | short reviewed | short accepted | total submissions | percentage of total submissions | total accepted | percentage of total accepted | area acceptance rate | outstanding papers |
| Semantics | 114 | 43 | 66 | 13 | 180 | 14.0% | 56 | 17.1% | 31.1% | 3 |
| Information Extraction, Question Answering, and Text Mining | 122 | 27 | 48 | 9 | 170 | 13.2% | 36 | 11.0% | 21.2% | 1 |
| Sentiment Analysis and Opinion Mining | 77 | 9 | 28 | 3 | 105 | 8.2% | 12 | 3.7% | 11.4% | |
| Document Analysis | 53 | 13 | 49 | 8 | 102 | 7.9% | 21 | 6.4% | 20.6% | |
| Machine Translation | 58 | 15 | 36 | 9 | 94 | 7.3% | 24 | 7.3% | 25.5% | 1 |
| Tagging, Chunking, Syntax, and Parsing | 48 | 18 | 33 | 9 | 81 | 6.3% | 27 | 8.2% | 33.3% | 2 |
| Social Media | 39 | 6 | 30 | 8 | 69 | 5.4% | 14 | 4.3% | 20.3% | |
| Machine Learning | 46 | 15 | 22 | 6 | 68 | 5.3% | 21 | 6.4% | 30.9% | 1 |
| Resources and Evaluation | 44 | 17 | 21 | 6 | 65 | 5.0% | 23 | 7.0% | 35.4% | |
| Other | 34 | 10 | 27 | 5 | 61 | 4.7% | 15 | 4.6% | 24.6% | |
| Discourse and Pragmatics | 42 | 15 | 18 | 4 | 60 | 4.7% | 19 | 5.8% | 31.7% | |
| Summarization | 29 | 5 | 19 | 2 | 48 | 3.7% | 7 | 2.1% | 14.6% | |
| Multilinguality | 19 | 6 | 24 | 5 | 43 | 3.3% | 11 | 3.4% | 25.6% | |
| Phonology, Morphology, and Word Segmentation | 23 | 6 | 10 | 4 | 33 | 2.6% | 10 | 3.0% | 30.3% | 1 |
| Generation | 20 | 8 | 9 | 3 | 29 | 2.3% | 11 | 3.4% | 37.9% | |
| Dialog and Interactive Systems | 20 | 7 | 8 | 0 | 28 | 2.2% | 7 | 2.1% | 25.0% | 1 |
| Cognitive Modeling and Psycholinguistics | 17 | 7 | 8 | 1 | 25 | 1.9% | 8 | 2.4% | 32.0% | 1 |
| Vision, Robots, and Grounding | 14 | 3 | 6 | 2 | 20 | 1.6% | 5 | 1.5% | 25.0% | |
| Speech | 6 | 1 | 1 | 0 | 7 | 0.5% | 1 | 0.3% | 14.3% | |
| total | 825 | 231 | 463 | 97 | 1288 | 100.0% | 328 | 100.0% | 25.5% | 11 |
The top five areas with the highest number of submissions this year were Semantics; Information Extraction, Question Answering, and Text Mining; Sentiment Analysis and Opinion Mining; Document Analysis (including text categorization, topic models, and retrieval), and Machine Translation.
Review Process
The page limit was 8 pages for long paper submissions and 4 pages for short paper submissions (each with unlimited pages for references). Camera-ready versions were given one additional page: 9 plus ∞ for long papers, 5 plus ∞ for short papers. As mentioned above, authors were additionally allowed to submit an appendix of unlimited length.
We changed the review forms slightly. We divided the SOUNDNESS category into one category for theoretical soundness and one for empirical soundness. We also pre-structured the comment box in the review form with three headings, “strengths,” “weaknesses,” and “discussion.” Reviewers were free to delete these headings when authoring their reviews.
Reviewer load balancing continues to be a challenge for this conference. Reviewers were invited directly by area chairs, then asked later to fill out a survey indicating which areas they were willing to review for. After papers were submitted and routed to areas, we used a tool provided by Mark Dredze (and used in previous ACL conferences) to improve the load balancing across areas. This worked well for short papers, but not as well for the much larger set of long papers, and even with considerable manual corrections, there was considerable disparity in the reviewing load across areas. We suspect the greedy algorithm implemented in this tool is partly to blame, but cannot rule out the possibility that we just needed more reviewers or more flexible reviewers to achieve balance. We expect that reviewer recruiting and assignment will continue to be a challenge as ACL grows.
There was a large number of papers for which area chairs had conflicts of interest, about 5-10% of submissions, including some papers where one of the co-PCs had a conflict. While the ACL policy at that time stated that “The identity of the reviewers of a paper shall be withheld from all people who have a conflict of interest in that paper,” we found that this was not doable for this large number of papers. Together with the ACL 2016 coordinating committee, we decided on the following procedure for ACL 2016:
- As the START tool allows ACL PC co-chairs to see all papers and all reviewers, PC co-chairs did not submit to the conference, but they do handle most aspects of the process for all papers that were submitted to the conference. When one PC co-chair had a COI with a paper, the other co-chair made final decisions about acceptance and presentation format.
- Papers co-authored by area chairs are assigned to areas other than the one that they chair.
- If an area chair has a COI with a paper (but is not an author on it), the other area chairs of the same area handle the paper, but the paper remains in the area. START supports this by making such papers invisible to the area chair with the COI. (Note though that there were some glitches where papers were not assigned meta-reviews because of visibility issues. These were handled informally when discovered.)
- If all area chairs of an area have a COI with the same paper, the paper is re-assigned to a different area. The “Other” area was a reasonable last resort.
As mentioned above, area chairs were asked to submit a meta-review (a functionality that is supported by START) for all papers that were not clear rejects. We found this to be extremely helpful in deciding on borderline papers.
Outstanding and Best Papers
The area chairs nominated 18 papers from 13 areas for recognition as outstanding papers. Out of these, the paper awards committee selected 9 outstanding long papers, of which one was awarded “best paper” and another “best student paper.” For short papers, 4 papers were nominated and 2 selected as outstanding. The paper awards committee was formed before the paper deadlines, and papers were given to the committee in their originally submitted, anonymized form. Reviews and meta-reviews of the nominated papers were also provided to the committee on request.
Presentations
The oral presentations are arranged in up to seven parallel sessions. There are two large poster sessions including dinner on two evenings of the conference. We manually grouped the papers into sessions, relying more on the topic of each paper than its area. In the end, only 10 out of 44 oral sessions were made up of papers all reviewed in the same area.
Based on recommendations of past program chairs, we aimed to have at least 11 m2 available for every poster presented in the poster sessions, to make the space comfortable and easy to move in. Additional space was booked at the poster session venue to allow for this; prior to the conference we estimated that more than 12 m2 would be available for every poster at the busier Monday night poster session.
Timeline
In anticipation of a larger pool of submissions, we intentionally scheduled the short and long paper deadlines as a joint, single date deadline..
The complete timeline after submission is given below:
- Feb 9-12 Paper bidding
- Feb 13 ACs assign papers to reviewers
- Feb 13-27 Paper reviewing
- Feb 28 AC check that all reviews have been received
- Mar 13-15 Author response period
- Mar 16-20 Reviewers update reviews after reading author response
- Mar 25 ACs send preliminary ranking to PCs
- Mar 28 ACs produce Meta reviews for borderline papers; ACs produce final rankings and Accept / Reject decisions
- Mar 30 Notification of acceptance
- April 22 Camera ready deadline
We had one exception to the schedule of completing transmission of the acceptance decisions, being about 12 hours late, due to operations difficulty.
Also, recruitment of the invited speakers started later than we had initially envisioned, which may have led to problems recruiting an appropriate external-to-the-field speaker.
Recommendations
We also detail our actions with respect to the outcomes and recommendations from ACL 2016, on two of their relevant points; see [2016 Program Chairs' report].
> 2. Many reviews were late. At the time that author response started, one third of the papers had at least one review missing, and some papers had all three reviews missing. We recommend leaving a few extra days between the end of reviewing and the start of author response, and starting some way of passing information about delinquent reviewers forward from conference to conference.
We mitigated this somewhat by having a shorter initial review cycle. While certain reviewers were late at this stage, we had a lengthened dialogue period that made it much easier to control for delays in reviews coming in. We recommend also setting an outstanding reviewer recognition award to a somewhat large proportion of reviewers (perhaps 5%) to spur on-time review and the necessary service time to do a good job of reviewing.
> 3. As discussed above, the reviewer load balancing task needs a more principled solution so that enough reviewers are recruited in advance of the deadlines and so that load balancing is handled smoothly with a good outcome.
This was mitigated somewhat by the solution of re-using the previous NAACL and ACL roles for reviewers. However, this has the potential problem (noted by Michael Strube) that personal information of reviewers is circulated to new chairs without explicit permission by the reviewer. Post-conference, we will try to solicit reviewers' explicit permission to have ACL store their personal profile for subsequent program committees.