Researchers have developed MADARTB, the first large-scale parallel dependency treebank for Arabic, encompassing five city dialects (Beirut, Cairo, Doha, Rabat, and Tunis) and Modern Standard Arabic. The dataset contains 5,000 sentences annotated with the CATiB dependency scheme, revealing significant syntactic variation across dialects despite overall construction similarities. A comparison using a supervised parser and four LLMs showed the supervised parser achieving 87.5% LAS, while the best LLM reached 85.7% LAS on the test set. The annotations and models are being released publicly. AI
IMPACT Provides a valuable resource for developing and evaluating Arabic NLP models, particularly for dialectal variations.
RANK_REASON Publication of a new academic dataset and research paper. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
