AlexandriaX 2026: The First Shared Task on Dialectal Arabic Machine Translation
Authors: Abdellah El Mekki, AbdelRahim A. Elmadany, Samar M. Magdy, Saad Ezzini, Mo El-Haj, Mustafa Jarrar, Zaid Alyafeai, Bernard Ghanem, Muhammad Abdul-Mageed
Organizations: The University of British Columbia · King Fahd University of Petroleum and Minerals · Lancaster University · VinUniversity · Hamad Bin Khalifa University · King Abdullah University of Science and Technology · Canada Research Chair in NLP and ML
Dialectal Arabic machine translation (MT) remains challenging despite recent progress in Arabic language technologies, particularly because effective translation requires modeling not only semantic content but also dialectal variation, conversational context, speaker and addressee characteristics, and sociolinguistic appropriateness. Moreover, conventional MT metrics provide limited insight into the linguistic errors produced by dialectal systems. We present the AlexandriaX 2026 Shared Task on Dialectal Arabic MT, which addresses these challenges through three complementary subtasks: (1) context-aware English-to-Dialectal Arabic dialogue translation across 13 Arabic varieties, (2) cross-dialect Arabic translation in the financial domain covering six Arabic dialects, and (3) span-level MT error detection and classification using linguistically motivated error categories across five Arabic varieties. The shared task attracted 38 registrations for Subtask 1, 33 for Subtask 2, and 35 for Subtask 3. Twelve unique teams submitted their system description papers, all of which we accepted for publication. The best system on Subtask 1 achieved 30.42 spBLEU in the constrained track and 33.49 spBLEU in the unconstrained track. On Subtask 2, the top system obtained 28.40 spBLEU. On Subtask 3, the best system achieved an overall score of 49.82, outperforming the 24.29 baseline. Taken together, the results of the leading systems across the three subtasks highlight the benefits of explicit dialect modeling, context-aware generation, retrieval and reranking, and specialized approaches to interpretable MT error analysis. All the resources of AlexandriaX 2026 shared task are publicly available, including data, baselines, and evaluation code on our project page: https://alexandriax.dlnlp.ai.
Figures & tables
Team Name
Affiliation
Subtask 1
Subtask 2
Subtask 3
C
U
ALEXIS Ibrahim et al. (2026)
Alexandria University
✓
Alkhder Alkhder et al. (2026)
Sakarya University
✓
✓
ArabicMTDiagnostics Yassine (2026)
Lebanese University
✓
Axiom Hossain (2026)
CUET
✓
CUNI Jon and Bojar (2026)
CUNI
✓
Table 1: AlexandriaX 2026 test phase participants by subtask and track. ✓ denotes submitted predictions; uncited teams did not submit a system paper. For Subtask 1, C and U indicate constrained and unconstrained tracks.
Rank
Team
Average spBLEU
Average chrF++
1
Thakaa
33.49
48.13
2
CUNI
32.50
47.05
3
LahjaMT
28.54
44.02
4
NAMAA Community
27.41
42.58
5
Rosetta
25.09
41.02
–
Baseline
21.17
37.45
Table 2: Official results for the unconstrained track of Subtask 1. Systems are ranked by mean spBLEU across the 13 target Arabic varieties; mean chrF++ is reported as a secondary metric. Alkhder submitted the same constrained system to both tracks, hence its identical scores in Table 3 .
Rank
Team
Average spBLEU
Average chrF++
1
Thakaa
30.42
45.42
2
LahjaMT
28.30
43.81
3
FCDS
26.48
42.30
4
Rosetta
26.10
41.79
5
NAMAA Community
23.26
39.03
–
Baseline
21.17
37.45
Table 3: Official results for the constrained track of Subtask 1. Systems are ranked by mean spBLEU across the 13 target Arabic varieties; mean chrF++ is reported as a secondary metric.
Rank
Team
Global spBLEU
Global chrF++
1
Thakaa
28.40
42.15
2
ALEXIS
25.84
39.02
3
NAMAA Community
24.53
39.04
4
L3IA
14.07
29.75
–
Baseline (Zero-Shot Gemma-4-31B)
24.69
38.44
–
Baseline (Leave-as-is)
12.81
27.31
Table 4: Official results for Subtask 2 (global corpus-level spBLEU and chrF++ evaluated across all 3,574 blind test instances on CodaBench).
Rank
Team
Overall Score
Exact Match F 1
Overlap F 1
Error Class Macro-F 1
1
Thakaa
49.82
23.19
54.26
45.38
2
NAMAA Community
46.41
22.39
54.48
38.34
3
TTLab
40.91
19.40
46.00
35.82
4
REGLAT
40.21
0.41
42.72
37.70
5
Axiom
39.26
16.64
48.13
30.40
6
ArabicMTDiagnostics
34.84
12.22
38.18
31.51
Table 5: Official results for Subtask 3. Systems are ranked by the Overall Score, computed as the mean of Overlap F1 and Error Class Macro-F1. Exact Match F1 is reported as an additional measure of span-localization accuracy.
Team
EG
JO
LB
LY
MA
MR
OM
PS
SA
SD
SY
TN
YE
Avg.
Thakaa
37.54
38.80
35.75
29.45
27.57
20.18
38.10
36.04
36.95
30.32
42.92
33.16
28.57
33.49
CUNI
38.29
40.10
34.69
30.85
28.11
18.79
29.31
35.36
36.05
32.38
38.68
32.79
27.12
32.50
LahjaMT
32.59
34.93
31.81
24.70
24.29
18.00
28.61
31.69
31.39
25.53
37.02
29.62
20.85
28.54
NAMAA Community
32.52
33.92
29.53
23.75
22.83
15.89
26.18
30.26
30.80
25.66
35.27
27.25
22.51
27.41
Rosetta
30.41
32.26
28.09
20.14
21.84
10.83
24.78
29.56
29.58
20.49
33.64
24.50
20.11
25.09
Baseline
26.22
27.61
24.34
18.18
14.97
9.83
20.33
24.27
25.22
18.47
29.34
19.65
16.73
21.17
Table 6: Subtask 1 spBLEU scores by target Arabic variety for the unconstrained track. Avg. is the arithmetic mean across the 13 varieties. Alkhder’s scores match Table 7 because the same system was submitted to both tracks.
Team
EG
JO
LB
LY
MA
MR
OM
PS
SA
SD
SY
TN
YE
Avg.
Thakaa
33.07
36.17
32.44
26.72
24.14
17.67
35.51
33.22
32.86
28.17
39.75
29.41
26.37
30.42
LahjaMT
32.59
34.93
31.81
23.32
24.29
18.00
28.61
31.69
31.39
23.81
37.02
29.62
20.85
28.30
FCDS
29.89
33.30
30.22
21.31
20.64
17.27
26.95
30.44
29.97
19.19
35.34
27.63
22.12
26.48
Rosetta
30.85
33.39
29.70
19.82
21.42
12.13
26.07
30.55
30.56
21.59
34.73
26.12
22.36
26.10
NAMAA Community
26.82
29.93
26.28
15.62
18.46
15.46
23.12
26.27
26.14
15.66
33.97
24.71
19.90
23.26
Baseline
26.22
27.61
24.34
18.18
14.97
9.83
20.33
24.27
25.22
18.47
29.34
19.65
16.73
21.17
Table 7: Subtask 1 spBLEU scores by target Arabic variety for the constrained track. Avg. is the arithmetic mean across the 13 varieties.
Team
Egyptian
Lebanese
Moroccan
Palestinian
Saudi
Tunisian
Avg.
Thakaa
36.47
34.98
24.32
20.48
26.55
16.44
26.54
ALEXIS
13.53
14.68
33.59
29.35
37.55
27.23
25.99
NAMAA Community
33.18
29.05
18.57
17.76
25.67
14.63
23.14
L3IA
15.10
14.58
16.28
11.51
13.13
7.19
12.97
Baseline (Leave-as-is)
13.96
17.32
8.90
10.94
12.34
5.11
11.43
Table 8: Subtask 2 spBLEU scores grouped by target dialect. Each score aggregates results from the five corresponding source dialects, and Avg. is the unweighted macro-average (arithmetic mean) of the six displayed target-dialect scores, which differs from the global corpus-level score in Table 4 due to dialect sample size disparity.
Team
EGY
MAU
MOR
PAL
UAE
Avg.
Thakaa
35.19 / 62.79
17.48 / 37.78
33.63 / 57.79
12.90 / 53.21
19.05 / 59.64
23.65 / 54.24
NAMAA Community
29.82 / 64.86
17.89 / 35.91
24.24 / 56.23
19.51 / 55.45
21.25 / 58.46
22.54 / 54.18
TTLab
23.93 / 58.63
17.78 / 30.32
16.16 / 46.48
18.00 / 38.46
20.00 / 51.64
19.17 / 45.11
REGLAT
2.02 / 43.03
0.00 / 36.45
0.00 / 44.02
0.00 / 44.69
0.00 / 45.27
0.40 / 42.69
Axiom
24.19 / 45.50
12.50 / 38.88
21.88 / 50.90
13.91 / 49.80
10.96 / 52.05
16.69 / 47.43
ArabicMTDiagnostics
10.69 / 32.98
16.16 / 33.75
6.61 / 34.87
8.93 / 42.69
17.61 / 44.09
12.00 / 37.68
Table 9: Subtask 3 span-localization performance by target dialect, reported as Exact Match F 1 / Overlap F 1 . Avg. is the arithmetic mean of the five dialect-specific scores. The best result for each metric and dialect is shown in bold.
Team
EGY
MAU
MOR
PAL
UAE
Avg.
Thakaa
50.00
40.78
42.48
43.55
48.98
45.16
NAMAA Community
40.35
29.27
36.36
29.27
52.50
37.55
TTLab
37.61
26.67
36.36
30.00
44.62
35.05
REGLAT
36.36
59.52
16.28
24.74
49.18
37.22
Axiom
32.26
23.21
31.25
19.13
42.47
29.66
ArabicMTDiagnostics
30.53
34.34
21.49
30.36
38.99
31.14
Table 10: Subtask 3 error-classification performance by target dialect, measured using Error Class Macro-F 1 . Avg. is the unweighted mean of the five dialect-specific scores. The best result for each dialect is shown in bold.
Current Machine Translation (MT) systems for Arabic often struggle to account for dialectal diversity, frequently homogenizing dialectal inputs into Modern Standard Arabic (MSA) and offering limited user control over the target vernacular. In this work, we propose a context-aware and steerable framework for dialectal Arabic MT that explicitly models regional and sociolinguistic variation. Our primary technical contribution is a Rule-Based Data Augmentation (RBDA) pipeline that expands a 3,000-sentence seed corpus into a balanced 57,000-sentence parallel dataset, covering eight regional varieties eg., Egyptian, Levantine, Gulf, etc. By fine-tuning an mT5-base model conditioned on lightweight metadata tags, our approach enables controllable generation across dialects and social registers in the translation output. Through a combination of automatic evaluation and qualitative analysis, we observe an apparent accuracy-fidelity trade-off: high-resource baselines such as NLLB (No Language Left Behind) achieve higher aggregate BLEU scores (13.75) by defaulting toward the MSA mean, while exhibiting limited dialectal specificity. In contrast, our model achieves lower BLEU scores (8.19) but produces outputs that align more closely with the intended regional varieties. Supporting qualitative evaluation, including an LLM-assisted cultural authenticity analysis, suggests improved dialectal alignment compared to baseline systems (4.80/5 vs. 1.0/5). These findings highlight the limitations of standard MT metrics for dialect-sensitive tasks and motivate the need for evaluation practices that better reflect linguistic diversity in Arabic MT.
This paper describes team Alkhder's submission to Subtask 1 of the AlexandriaX-2026 shared task, which asks systems to translate English dialogue turns into thirteen Arabic dialects. We fine-tune NileChat-3B-Base with QLoRA on the Alexandria parallel dialogue corpus, training one set of adapter weights for all dialects. Each turn is translated conditioned on the target country and dialect, the domain, the participant roles, the speaker-addressee gender direction, and the preceding dialogue context. At inference, conversations are translated autoregressively and the model is fed its own previous outputs back as context without using any gold target-side test information. Under the constrained track, using only Alexandria data, it trains in about one hour and translates the private test set in under two hours. The system reaches a country-macro-averaged spBLEU of 20.36 and chrF++ of 36.64 on both tracks.
Hasan Alkhder, Maher Pasa, Maria Hanini +2
Sakarya University, Türkiye · Altınbaş University, Türkiye · Imperial College London +1
We present TTLab's submission to the AlexandriaX-2026 Subtask~3 on Arabic MT error span detection and classification. Our system frames the task as token-level classification over surface forms, preserving character offsets to ensure exact alignment with the evaluation metric. To handle severe label imbalance, we employ a focal loss with class weighting and dialect-specific decoding thresholds. Among six Arabic pre-trained encoders, MARBERTv2 achieves the best overall performance of 40.8 and 40.91 on the development and test set, respectively, ranking \nth3 out of all participating teams. While our system localizes error spans effectively, classification of rare error types remains challenging, highlighting the need for data augmentation for tail categories. The code is available at {[\faGithub~ TTLab at AlexandriaX-2026](https://github.com/ENTAILab/arabic-dialectal-mt-error-span-detection)
Ali Abusaleh, Bhuvanesh Verma, Alexander Mehler
Text Technology Lab (TTLab), Goethe University Frankfurt