أحمد أبو حسن

Misinformation classifiers for Arabic are overwhelmingly trained on Modern Standard Arabic, yet most viral claims circulate in dialect. We introduce a benchmark of 38,000 fact-checked claims across nine dialect groups and show that dialect-aware pre-training reduces false negatives by 31% relative to strong multilingual baselines.

We describe the design and first release of LAHJA, a 12-million-word corpus of dialectal Arabic contributed by 4,100 volunteers with explicit, revocable consent. We detail our collection protocol, annotation scheme, and baseline results for dialect identification and sentiment analysis.

تقيس هذه الدراسة أداء خمسة نماذج لغوية كبيرة على مهام الفهم والتوليد بخمس لهجات عربية، وتُظهر فجوة في الأداء تصل إلى 24 نقطة بين الفصحى واللهجات المغاربية. نقترح إطارًا للتقييم العادل بين اللهجات ونناقش أثره على الخدمات الرقمية الحكومية.

Using 60,000 paired posts, we measure how automated moderation systems treat equivalent content written in English, Modern Standard Arabic, and Levantine Arabic. Dialectal posts were removed at 2.3× the rate of their English equivalents, with disproportionate impact on news and advocacy accounts.