SinaLab

📢 Announcement: Shared task at ArabicNLP at EMNLP 2026:

مسابقة بحثية - قدرة الذكاء الاصطناعي على التعميم مع البيانات غير المرئية مسبقا
𝗔𝗿𝗮𝗚𝗲𝗻𝗿𝗲: 𝗔 𝗛𝗶𝗲𝗿𝗮𝗿𝗰𝗵𝗶𝗰𝗮𝗹 𝗗𝗲𝗳𝗶𝗻𝗶𝘁𝗶𝗼𝗻-𝗚𝘂𝗶𝗱𝗲𝗱 𝗔𝗿𝗮𝗯𝗶𝗰 𝗚𝗲𝗻𝗿𝗲 𝗖𝗹𝗮𝘀𝘀𝗶𝗳𝗶𝗰𝗮𝘁𝗶𝗼𝗻

Participate: https://wp.lancs.ac.uk/aragenre/

A challenging task designed to test how NLP systems generalise to the unseen:
•⁠ ⁠Few-shot training and development data
•⁠ ⁠Cross-lingual genre definitions guiding Arabic classification
•⁠ ⁠A large hidden test set of naturally occurring Arabic texts
•⁠ ⁠Unseen genre-definition combinations

While AraGenre focuses on Arabic, the task addresses a wider challenge in NLP: building methods that can generalise from limited data and support languages and varieties that remain underrepresented in current NLP research.

Organisers:
Mo El-Haj
Saad Ezzini
Mustafa Jarrar
Shadi Abudalfa

#𝗔𝗿𝗮𝗚𝗲𝗻𝗿𝗲

Join us at to learn 𝐇𝐨𝐰 𝐭𝐨 𝐮𝐬𝐞 𝐨𝐧𝐭𝐨𝐥𝐨𝐠𝐢𝐞𝐬 𝐢𝐧 𝐋𝐋𝐌 𝐩𝐫𝐨𝐦𝐩𝐭𝐬?
We are excited to release our 𝐖𝐨𝐣𝐨𝐨𝐝𝐎𝐧𝐭𝐨𝐥𝐨𝐠𝐲 for 𝐍𝐚𝐦𝐞𝐝 𝐄𝐧𝐭𝐢𝐭𝐲 𝐑𝐞𝐜𝐨𝐠𝐧𝐢𝐭𝐢𝐨𝐧 and 𝐑𝐞𝐥𝐚𝐭𝐢𝐨𝐧 𝐄𝐱𝐭𝐫𝐚𝐜𝐭𝐢𝐨𝐧 at at .

يسعدنا الإعلان عن "أنطولوجيا وجود"، لتوحيد طرق استخراج أسماء الأعلام والعلاقات بينها، وكذلك استخدام الانطولوجيا أثناء استخدام النماذج اللغوية الضخمة.

- Paper: https://lnkd.in/dyVydKvN
- Download: https://lnkd.in/dYDka4sw
- YouTube: https://lnkd.in/dsN2mTrs
- Slides: https://lnkd.in/dQJm55S4

𝐒𝐡𝐨𝐫𝐭 𝐬𝐮𝐦𝐦𝐚𝐫𝐲:
Information Extraction tasks such as Named Entity Recognition and Relation Extraction often rely on diverse tagsets and annotation guidelines, creating challenges for model generalization, cross-dataset evaluation, and interoperability. To overcome these issues, WojoodOntology provides a unified semantic framework that aligns heterogeneous tagsets and guidelines.
It supports two ontology-based mapping strategies: rule-based unidirectional alignment and ontology-based prompting, where ontology concepts are embedded directly into large language model prompts to enable more effective bidirectional mappings. Experiments demonstrate that ontology-based prompting improves out-of-domain mapping accuracy by 15% compared to rule-based methods. Additionally, WojoodOntology is aligned with Schema.org and Wikidata, ensuring compatibility with knowledge graphs and facilitating wider industry adoption.

#𝐒𝐡𝐚𝐫𝐞𝐝𝐓𝐚𝐬𝐤
Alaa Aljabari Nagham Hamad Mohammed Khalilia Mustafa Jarrar Birzeit University

Article: 𝐫𝐞𝐥𝐚𝐭𝐢𝐨𝐧 𝐞𝐱𝐭𝐫𝐚𝐜𝐭𝐢𝐨𝐧 𝐜𝐨𝐫𝐩𝐮𝐬. ~550K tokens, 40 relations.
نشر ورقة في مؤتمر مرموق للطالبة ألاء الجعبري - برنامج الدكتوراة بعلم الحاسوب. الورقة تقدم منهجية مدونة جديدة لاستخراج العلاقات بين أسماء الأعلام في النصوص العربية.

Article:https://www.jarrar.info/publications/JKJ25.pdf
Download: https://sina.birzeit.edu/relations/
Demo: https://sina.birzeit.edu/relations/
Slides:https://www.jarrar.info/Talks/wojood_relations_paper.pdf

𝐬𝐮𝐦𝐦𝐚𝐫𝐲: We introduce WojoodRelations, the largest Arabic relation extraction corpus, containing 33K sentences and 15K annotated relation triples across 40 distinct relation types. We also propose new methods for supervised and LLM-based RE, with our models achieving up to 92.9% F1, setting strong baselines for future research.




SinaLab
الدراسات العليا والأبحاث في جامعة بيرزيت
كلية الهندسة -جامعة بيرزيت
Mustafa Jarrar
Hamad Bin Khalifa University
Birzeit University | جامعة بيرزيت
PhD in Computer Science / دكتوراة علم حاسوب