Advances in language modeling have led to remarkable accuracy on several NLP tasks, but most benchmarks used for evaluation are static, ignoring the practical setting under which training data from the past and present must be used for generalizing to future data. Consequently, training paradigms also ignore the time sensitivity of language and essentially treat all text as if it was written at a single point in time. Recent studies have shown that in a dynamic setting, where the test data is drawn from a different time period than the training data, the accuracy of such static models degrades as the gap between the two periods increases.
This workshop focuses on these time-related issues in NLP models and benchmarks. We invite researchers from both academia and industry to redesign experimental settings, benchmark datasets, and modeling by especially focusing on the “time” variable. We will welcome papers / work-in-progress on several topics including (but not limited to):
Dynamic Benchmarks: Evaluation of Model Degradation in Time
Measuring how NLP models age
Random splits vs time-based splits (past/future)
Latency (days vs years) at which models need to be updated for maintaining task accuracy
Time-sensitivity of different tasks and the type of knowledge which gets stale
Time-sensitivity of different domains (e.g., news vs scientific papers) and how domain shifts interact with time shifts
Sensitivity of different models and architectures to time shifts
Time-Aware Models
Incorporating time information into NLP models
Techniques for updating / replacing models which degrade with time
Learning strategies for improving temporal degradation
Trade-offs between updating a degraded model vs replacing it altogether
Mitigating catastrophic forgetting of old knowledge as we update models with new knowledge
Improving plasticity of models so that they can be easily updated
Retrieval based models for improving temporal generalization
Analysis of existing models / datasets
Characterizing whether degradation on a task is due to outdated facts or changes in language use
Effect of model scale on temporal degradation – do large models exhibit less degradation?
Efficiency / accuracy trade-offs when updating models
All accepted papers will be published in the workshop proceedings unless requested otherwise by the authors. Submissions can be made either via OpenReview where they will go through the standard double-blind process, or through ACL Rolling Review with existing reviews. See details below.
(Anywhere on Earth)
Submission deadline (for papers requiring review / non-archival): 10 October, 2022 12 October, 2022
Submission deadline (only for papers with ARR reviews): 25 October, 2022
Notification of acceptance: 31 October, 2022
Camera-ready paper deadline: 11 November, 2022
Workshop date: 7 December, 2022
We seek submissions of original work or work-in-progress. Submissions can be in the form of long/short papers (maximum of 4 pages for short papers and 8 pages for long papers, with unlimited space for references and appendix) and should follow EMNLP templates. Final versions of papers will be given one additional page of content so that reviewers’ comments can be taken into account.
Authors can choose to make their paper archival/non-archival. All accepted papers will be presented at the workshop.
Submission link: https://openreview.net/group?id=EMNLP/2022/Workshop/EvoNLP
For papers needing review click “EMNLP 2022 Workshop EvoNLP submission”
For papers from ARR click “EMNLP 2022 Workshop EvoNLP commitment Submission”
Non-archival track seeks recently accepted / published work (including EMNLP-Findings papers) as well as work-in-progress. It does not need to be anonymized and will not go through the review process. The submission should clearly indicate the original venue and will be accepted (and given a presentation/poster slot) if the organizers think the work will benefit from exposure to the audience of this workshop.
Submission: Please email your submission as a single PDF file to evonlp@googlegroups.com. Include “EvoNLP Non-Archival Submission” in the title and the author names and affiliation within the body of your email.
The workshop will feature a shared task on meaning shift detection in social media. Initial data already available! Winners of the shared task will also receive a cash prize. More details at the workshop website: https://sites.google.com/view/evonlp/shared-task.
Thanks to generous support from our sponsors Snap Inc., we will award the best paper award (with cash prize) to one of the submissions selected by our program committee and organizing committee. There will also be a cash prize for the best system demonstration paper in the shared task. The best papers will be given the opportunity for a talk to introduce their work.