SemEval-2022: International Workshop on Semantic Evaluation
https://semeval.github.io/
We invite proposals for tasks to be run as part of SemEval-2022. SemEval (the International Workshop on Semantic Evaluation) is an ongoing series of evaluations of computational semantics systems, organized under the umbrella of SIGLEX, the Special Interest Group on the Lexicon of the Association for Computational Linguistics.
SemEval tasks explore the nature of meaning in natural languages: how to characterize meaning and how to compute it. This is achieved in practical terms, using shared datasets and standardized evaluation metrics to quantify the strengths and weaknesses and possible solutions. SemEval tasks encompass a broad range of semantic topics from the lexical level to the discourse level, including word sense identification, semantic parsing, coreference resolution, and sentiment analysis, among others.
For SemEval-2022, we welcome any task that can test an automatic system for semantic analysis of text, which could be an intrinsic semantic evaluation, or an application-oriented evaluation. We especially encourage tasks for languages other than English, cross-lingual tasks, and tasks that develop novel applications of computational semantics. See the websites of previous editions of SemEval to get an idea about the range of tasks explored, e.g. SemEval-2019: http://alt.qcri.org/semeval2019/ and SemEval-2020: http://alt.qcri.org/semeval2020/
We strongly encourage proposals based on pilot studies that have already generated initial data, as this can provide concrete examples and can help to foresee the challenges of preparing the full task. In the event of receiving many proposals, preference will be given to proposals that have already run a pilot study.
In case you are not sure whether a task is suitable for SemEval, please feel free to get in touch with the SemEval organizers at semeval-organizers [at] googlegroups.com to discuss your idea.
* Task Selection *
Task proposals will be reviewed by experts, and reviews will serve as the basis for acceptance decisions. Everything else being equal, more innovative new tasks will be given preference over task reruns. Task proposals will be evaluated on:
- Novelty: Is the task on a compelling new problem that has not been explored much in the community? Is the task a rerun, but covering substantially new ground (new sub-tasks, new types of data, new languages, etc.)?
- Interest: Is the proposed task likely to attract a sufficient number of participants?
- Data: Are the plans for collecting data convincing? Will the resulting data be of high quality? Will annotations have meaningfully high inter-annotator agreements? Have all appropriate licenses for use and re-use of the data after the evaluation been secured? Have all international privacy concerns been addressed? Will the data annotation be ready on time?
- Evaluation: Is the methodology for evaluation sound? Is the necessary infrastructure available or can it be built in time for the shared task? Will research inspired by this task be able to evaluate in the same manner and on the same data after the initial task?
- Impact: What is the expected impact of the data in this task on future research beyond the SemEval Workshop?
* New Tasks vs. Task Reruns *
We welcome both new tasks and task reruns. For a new task, the proposal should address whether the task would be able to attract participants. Preference will be given to novel tasks that have not received much attention yet.
For reruns of previous shared tasks (whether or not the previous task was part of SemEval), the proposal should address the need for another iteration of the task. Valid reasons include: a new form of evaluation (e.g. a new evaluation metric, a new application-oriented scenario), new genres or domains (e.g. social media, domain-specific corpora), or a significant expansion in scale. We further discourage carrying over a previous task and just adding new subtasks, as this can lead to the accumulation of too many subtasks. Evaluating on a different dataset with the same task formulation typically should not be considered a separate subtask.
* Task Organization *
We welcome people who have never organized a SemEval task before, as well as those who have. Apart from providing trial, training, and test data, task organizers are expected to:
- Verify the data annotations have sufficient inter-annotator agreement
- Verify licenses for the data allow its use in the competition and afterwards. In particular, text that is publicly available online is not necessarily in the public domain; unless a license has been provided, the author retains all rights associated with their work, including copying, sharing and publishing. For more information, see: https://creativecommons.org/faq/#what-is-copyright-and-why-does-it-matter
- Resolve any potential security, privacy, or ethical concerns about the data
- Make the data available in a long-term repository under an appropriate license, preferably using Zenodo: https://zenodo.org/communities/semeval/
- Provide task participants with format checkers and standard scorers.
- Provide task participants with baseline systems to use as a starting point (in order to lower the obstacles to participation). A baseline system typically contains code that reads the data, creates a baseline response (e.g. random guessing, majority class prediction), and outputs the evaluation results. Whenever possible, baseline systems should be written in widely used programming languages and/or should be implemented as a component for standard NLP pipelines such as UIMA or GATE.
- Create a mailing list and website for the task and post all relevant information there.
- Create a CodaLab or other similar competition for the task and upload the evaluation script.
- Manage submissions on CodaLab or the similar competition site.
- Write a task description paper to be included in SemEval proceedings, and present it at the workshop.
- Manage participants’ system description submissions to the task, and possibly shepherd papers that need additional help in improving the writing.
- Review one or two other task description papers.
* Important dates *
Task proposals due March 22, 2021
Task selection notification May 14, 2021
* Preliminary timetable *
Trial data ready July 30, 2021
Training data ready September 3, 2021
Evaluation data ready December 3, 2021
Evaluation start January 10, 2022
Evaluation end by January 31, 2022 (latest date; task organizers may choose an earlier date)
Paper submission due February 23, 2022
Notification to authors March 31, 2022
Camera ready due April 30, 2022
SemEval workshop Summer 2022 (co-located with a major NLP conference)
Tasks that fail to keep up with crucial deadlines such as the dates for having the task and CodaLab website up and dates for uploading trial, training, and evaluation data may be cancelled at the discretion of SemEval organizers. While consideration will be given to extenuating circumstances, our goal is to provide sufficient time for the participants to develop strong and well-thought-out systems. Cancelled tasks will be encouraged to submit proposals for the subsequent year’s SemEval. To reduce the risk of tasks failing to meet the deadlines, we are unlikely to accept multiple tasks with overlap in the task organizers.
* Submission Details *
The task proposal should be a self-contained document of no longer than 3 pages (plus additional pages for references). All submissions must be in PDF format, following the ACL 2021 template:
LaTeX and Microsoft Word: https://2021.aclweb.org/downloads/acl-ijcnlp2021-templates.zip
Overleaf: https://www.overleaf.com/latex/templates/instructions-for-acl-ijcnlp-202...
Each proposal should contain the following:
- Overview
- Summary of the task
- Why this task is needed and which communities would be interested in participating
- Expected impact of the task
- Data & Resources
- How the training/testing data will be produced. Please discuss whether existing corpora will be re-used.
- Details of copyright, so that the data can be used by the research community both during the SemEval evaluation and afterwards
- How much data will be produced
- How data quality will be ensured and evaluated
- An example of what the data would look like
- Resources required to produce the data and prepare the task for participants (annotation cost, annotation time, computation time, etc.)
- Assessment of any concerns with respect to ethics, privacy, or security (e.g. personally identifiable information of private individuals; potential for systems to cause harm)
- Pilot Task (strongly recommended)
- Details of the pilot task
- What lessons were learned and how these will impact the task design
- Evaluation
- The evaluation methodology to be used, including clear evaluation criteria
- For Task Reruns
- Justification for why a new iteration of the task is needed (see criteria above)
- What will differ from the previous iteration
- Expected impact of the rerun compared with the previous iteration
- Task organizers
- Names, affiliations, brief description of research interests and relevant experience, email addresses.
- (if applicable) years and task numbers, of any SemEval tasks you have run in the past
Proposals will be reviewed by an independent group of area experts who may not have familiarity with recent SemEval tasks, and therefore all proposals should be written in a self-explanatory manner and contain sufficient examples.
The submission webpage is: https://www.softconf.com/acl2021/w13_SemEval2022
Chairs:
Nathan Schneider, Georgetown University
Alexis Palmer, University of North Texas
Guy Emerson, University of Cambridge
Natalie Schluter, IT University of Copenhagen
Contact: semeval-organizers [at] googlegroups.com