Previous Edition
StanceEval-2024
The first Arabic stance detection shared task, held at ArabicNLP 2024 in Bangkok, attracted ~30 competing teams and laid the foundation for StanceEval-2026.
Visit StanceEval-2024 ↗
StanceEval-2026 is a shared task on stance detection in Arabic social media. Given a tweet and a target, systems must determine whether the writer's stance is Favor, Against, or None — across both seen and unseen targets, using the multi-label Mawqif-v2 dataset.
Stance detection identifies whether a text expresses a position for, against, or neutral toward a specific target — distinct from sentiment analysis in that the target may be abstract or not explicitly mentioned in the text.
Given an Arabic post and a target, participants classify the expressed stance. Sentiment and sarcasm labels are available as optional auxiliary features.
Models are trained and evaluated on targets observed during training — testing in-domain classification performance.
Models are evaluated on entirely new topics — testing zero-shot generalization beyond targets seen during training.
Training & dev data and evaluation scripts released publicly.
Teams must be registered. Blind test data made available for submission.
Leaderboard finalized and winning teams announced.
Camera-ready system description papers submitted by participating teams.
Budapest, Hungary — shared task presentations and prize ceremony.
Registration opens May 16, 2026 alongside the data release. Join teams from around the world advancing Arabic NLP.
An extended version of the original Mawqif corpus collected from X.com. Each tweet is manually annotated with stance, sentiment, and sarcasm. New unseen-target tweets were labeled by three annotators, with a fourth resolving disagreements.
📂 View Dataset on GitHub ↗| Split | Target | Type | # Tweets | % Favor | % Against | % None |
|---|---|---|---|---|---|---|
| Train | COVID-19 Vaccine | Seen | 1,167 | 43.62 | 43.53 | 12.85 |
| Digital Transformation | Seen | 1,145 | 76.77 | 12.40 | 10.83 | |
| Women Empowerment | Seen | 1,190 | 63.87 | 31.18 | 4.96 | |
| Dev | COVID-19 Vaccine | Seen | 206 | 43.69 | 43.69 | 12.62 |
| Digital Transformation | Seen | 203 | 76.85 | 12.32 | 10.84 | |
| Women Empowerment | Seen | 210 | 63.81 | 30.95 | 5.24 | |
| Test | Target 1 | Unseen | 312 | 22.12 | 72.44 | 5.45 |
| Target 2 | Unseen | 332 | 43.67 | 35.24 | 21.08 | |
| Target 3 | Unseen | 352 | 44.89 | 45.45 | 9.66 | |
| Total | — | 5,117 | — | — | — | |
Systems are evaluated per topic. Favg2 is the primary ranking metric, aligning with prior Arabic stance detection literature. Favg3 includes the None class for a fuller picture.
The evaluation script is available below alongside the baseline code to help participants build on the existing implementation.
$1,600 in total prizes across both tracks — $800 per track — distributed to the top three winners.
PhD student, KFUPM · Teaching assistant, Tabuk University.
Assistant Professor, KFUPM & Research Fellow, SDAIA-KFUPM JRC for AI.
Associate Professor, KFUPM & Fellow, SDAIA-KFUPM JRC.
Assistant Professor, KFUPM. Associate Editor, NLP Journal.
Assistant Professor, KFUPM.
PhD Candidate, KFUPM.
MSc student, KFUPM.
AI Researcher & Engineer, JRCAI.
Senior Principal AI Researcher, HUMAIN.
Have questions or need support? Reach us by email or join our Discord for real-time discussion and announcements.
The first Arabic stance detection shared task, held at ArabicNLP 2024 in Bangkok, attracted ~30 competing teams and laid the foundation for StanceEval-2026.
Visit StanceEval-2024 ↗