A fine-tuned Arabic language model just placed third among all competitors at a specialized task: spotting exactly where machine translations go wrong.
The system comes from a team called TTLab, entered in the AlexandriaX-2026 shared task's error-detection subtask for Arabic. It treats the job as labeling individual words and characters in a translated sentence, marking precisely where an error starts and ends and what category it falls into. TTLab tested six different Arabic language models and found MARBERTv2 performed best, scoring 40.8 on the development set and 40.91 on the test set, good enough for third place overall. To handle the fact that some error types show up far less often than others in training data, the team used a technique called focal loss to force the model to pay more attention to rare categories, alongside decoding thresholds tuned separately for different Arabic dialects.
Automated error-span detection matters because Arabic has wide dialectal variation and comparatively little high-quality training data next to languages like English or Chinese, which makes manual review of machine-translated Arabic slow and expensive. A tool that flags not just that a translation is wrong, but exactly where and why, lets translation teams triage fixes instead of re-checking every sentence by hand.
The system found error locations reliably but still struggled to correctly classify rarer error types - a familiar bottleneck whenever training data is imbalanced, and a reminder that the fix here is more labeled examples of the categories nobody bothered to collect, not a fancier model.