Parsing Arabic Dialects Revisited: New Benchmarks, Models, and Insights

Abstract

Parsing dialectal Arabic remains underexplored, with limited progress over the past two decades. Existing Modern Standard Arabic (MSA) parsers perform poorly on dialectal data, motivating the need for dialect-specific approaches. We revisit this task using modern neural models and present new results on Egyptian and Gulf Arabic dependency parsing. We demonstrate that even small amounts of dialectal training data yield substantial improvements in parsing accuracy. Our contributions include: (1) introducing a new annotated dataset for Gulf Arabic, (2) releasing a state-of-the-art multi-variety Arabic parser, and (3) employing dialect identification as a diagnostic tool to better understand how training data affects parsing performance across dialects and test sets.

Resources

Details