AlexandriaX ArabicNLP 2026 Shared Task

Participant Support

FAQ

What is AlexandriaX?

AlexandriaX-2026 is an ArabicNLP 2026 shared task on context-aware dialectal Arabic MT and MT evaluation, cross-dialect Arabic MT, and span-level MT error detection and classification.

Can a team participate in only one subtask?

Yes. Teams may participate in Subtask 1, Subtask 2, Subtask 3, or any combination of them. Subtask 1 also has constrained and unconstrained tracks (participants can participate in either or both).

What is allowed in the constrained MT track?

Participants may use only the provided Alexandria training and development data. Submitted systems must not exceed 5B total parameters.

What is allowed in the unconstrained MT track?

External data, pretrained models, and large-scale resources are allowed. There is no model-parameter limit in the unconstrained track.

Which dialects are covered in Subtask 1?

Jordanian, Lebanese, Palestinian, Syrian, Saudi, Omani, Yemeni, Egyptian, Sudanese, Libyan, Moroccan, Mauritanian, and Tunisian.

Which dialects are covered in Subtask 2?

Modern Standard Arabic, Palestinian, Moroccan, Tunisian, Egyptian, Algerian, Lebanese, Yemeni, Omani, Saudi, and Libyan.

Which dialects are covered in Subtask 3?

Egyptian, Emirati, Mauritanian, Moroccan, and Palestinian.

How is Subtask 1 ranked?

Systems are evaluated by the overall average across all countries. spBLEU is the primary metric for the leaderboard.

How is Subtask 2 ranked?

Systems are evaluated by the overall average across all countries. spBLEU is the primary metric for the leaderboard.

How is Subtask 3 ranked?

Systems are evaluated for span detection and error classification, with the main ranking determined by Labeled Span F1.

When will the data be available?

Training and development data, baseline code, and evaluation scripts are scheduled for June 1, 2026. Blind test data will be released on July 20, 2026.