Arabic dialect dependency treebank enables better parsing across cities
A Parallel Multi-Dialect Arabic Dependency Treebank
Computation and Language
Summary
Arabic language tools have mostly focused on the formal Modern Standard Arabic, leaving city dialects less studied. The authors created a new data collection called MADARTB that has sentences from five Arabic city dialects, all annotated to show how each word connects grammatically. They found that while the overall sentence structures are similar, local details differ notably between dialects. They also tested language parsers on this data and showed traditional supervised methods still slightly outperform large language models in understanding these dialects. The dataset and models are publicly available for others to use.
What this means in practice
- •For natural language processing teams: Improve Arabic dialect parsing accuracy by training models on a large set of parallel annotated sentences from multiple city dialects.
- •For machine translation developers: Enhance translation systems by incorporating syntactic analysis from diverse Arabic dialects using this comprehensive parallel treebank.
Authors
Noor Abo Mokh, Mohamed Adel, Jamila El-Gizuli, Go Inoue, Ahmed Elshabrawy, Nizar Habash
Abstract
Dependency parsing resource development for Arabic has largely focused on Modern Standard Arabic, leaving dialectal Arabic comparatively under-resourced. We introduce MADARTB, the first large-scale parallel dependency treebank spanning five city dialects: Beirut, Cairo, Doha, Rabat and Tunis, and extending an existing parallel MSA version. In total MADARTB consists of 5,000 sentences annotated using the CATiB dependency scheme. We analyze syntactic similarity across all city dialects and MSA. Although global construction profiles exhibit high similarity across varieties, local syntactic structures frequently diverge, revealing substantial structural variation across dialects. We further establish a cross-dialect dependency parsing comparison using a supervised parser and four LLMs. The supervised parser achieves 87.5% LAS, while the best LLM setup trails at 85.7% LAS across all six varieties on the test set. We release our annotations and models publicly.