System Description Paper
A paper documenting the methods, experiments, findings, and lessons learned from participation in ArGuard subtasks, with up to four pages of main content, following ACL formatting guidelines.
NLP/CORE Unranked
ArabicNLP2026-ArGuard 2026
ArabicNLP 2026 Shared Task on ArGuard: Harmful Content Detection in Arabic Memes and LLM Prompts
ArGuard 2026 is a shared task at ArabicNLP 2026 focused on detecting harmful content in Arabic multimodal memes and LLM prompts. It features two tracks—multimodal hateful meme detection and textual harmful prompt detection—each with binary classification and fine-grained categorization subtasks, emphasizing Arabic-specific challenges like dialects, code-switching, sarcasm, and visual-textual interactions.
Paper fit
Contribution paths
A strong submission should clearly identify its contribution and evaluate it appropriately.
Research areas in scope
01
Task A: Multimodal Hateful Meme Detection
A1: Hateful vs. Not Hateful classificationA2: Multi-label fine-grained category prediction (mocking, incitement, dehumanization, slurs, contempt, inferiority, exclusion, stereotyping, extremism, threat, insults, historical, humor, sarcasm, and other categories)
02
Task B: Textual Harmful Prompt Detection
B1: Safe vs. Unsafe prompt detectionB2: Harm category classification (self-harm, harm to others, harassment, adult content, bullying, hate speech, fraud or illegal activities)
Policies worth checking twice
- Papers must use the official ACL two-column format with unmodified style files.
- Main content is limited to four pages; references and appendices may extend beyond.
- Limitations and Ethical Considerations may appear on a fifth page before references.
- Papers are not anonymized; author names, affiliations, and contact information must be included.
- Submissions must cite the official ArGuard overview and dataset papers.
- All system description papers must be submitted via the OpenReview portal.
- Camera-ready papers must pass ACL PubCheck for formatting compliance.
- Teams may submit a single paper covering multiple subtasks, though not required.
Official sources
Compiled from the official call for papers. The organizers’ pages remain authoritative.
Last verified September 9, 2026