MudawanSn: A Gold-Standard Wolof-Arabic Parallel Corpus for Machine Translation
cs.CL
Quick verdict
First public Wolof–MSA corpus (1,271 human pairs); transparent protocol, real fine-tuning gains, but tiny n=96 test and one translator weaken 'gold-standard'.
Summary
The paper presents MudawanSn, a manually translated, sentence-aligned corpus of 1,271 Wolof–Modern Standard Arabic pairs, sourced from Senegalese news in MasakhaNER and positioned as the first public gold-standard bitext for this pair. Construction involved a single bilingual translator (dictionary-aided, no MT in the loop), NFC normalization, a three-stage French-source matching pipeline against galsenai/french-wolof-translation (exact 53.89%, substring 34.30%, >=85% token-overlap 9.99%; 98.19% enriched), manual 1:1 alignment, LASER3 similarity screening (mean 0.7508; pairs <0.60 human-reviewed), and a TSV release under CC BY-NC with fixed 85/7.5/7.5 splits. Baselines compare NLLB-200-600M, mT5-base, and two AfriNLLB variants, zero-shot vs. fine-tuned (10 epochs, joint bidirectional training), scored with sacreBLEU BLEU, chrF++, and AfriCOMET. Fine-tuning yields large gains for all multilingual models; AfriNLLB-12 is best lexically (7.76 BLEU / 30.72 chrF++ wo->ar; 8.75 / 33.08 ar->wo), while fine-tuned NLLB-200 scores highest on AfriCOMET, a metric the authors deem unreliable here. Strengths: a genuinely new resource, transparent provenance, candid limitations. Main weaknesses: a 96-pair test set with no significance testing or confidence intervals, a single-translator 'gold-standard' claim without inter-annotator agreement, unspecified sentence-selection criteria, and possible train/test leakage from sentence-level rather than document-level splitting.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- MudawanSn is the first publicly available, manually translated, gold-standard parallel corpus specifically designed for the Wolof-Modern Standard Arabic language pair.
- The multi-stage construction protocol (manual translation, French-source matching, 1:1 alignment, LASER3 screening with human review) yields high parallel fidelity, with 96.06% of pairs scoring above 0.60 LASER3 similarity.
- Fine-tuning on MudawanSn substantially improves all tested multilingual models in both directions, e.g., NLLB-200 improves by +5.35 BLEU and +13.89 chrF++ for Wolof-to-Arabic.
- AfriNLLB-12 achieves the best lexical scores (7.76 BLEU / 30.72 chrF++ wo->ar; 8.75 BLEU / 33.08 chrF++ ar->wo), though NLLB-200 fine-tuned scores highest on AfriCOMET, which the authors attribute to the pair being outside AfriCOMET's training coverage.
Strengths
Fills a genuine, well-argued resource gap
Section 2 and Table 1 carefully situate MudawanSn against FLORES-200, NTREX-128, MAFAND-MT, SMOL, and FineTranslations; the qualifier 'specifically designed' for Wolof-MSA is defensible, and the Senegalese-context motivation (Wolofal/Ajami tradition, Ngom 2017) is substantive rather than decorative.
Transparent provenance and enrichment pipeline
The three-stage French-source matching (Section 3.4) is fully quantified: 53.89% exact normalized, 34.30% substring, 9.99% fuzzy token-overlap matches, reaching 98.19% enrichment with french_source and MAFAND metadata, so every pair is traceable to upstream outlets.
Multi-layer quality control with no MT circularity
Informative baseline study with a honest metric discussion
Candid limitations section
Weaknesses
Tiny evaluation set with no uncertainty quantification
'Gold-standard' rests on a single translator
Unspecified sentence-selection criteria
Sentence-level split risks train/test leakage
Interpretation baselines missing for LASER3 threshold
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MudawanSn?
MudawanSn is a publicly released parallel corpus for machine translation between Wolof (written in Latin script) and Modern Standard Arabic. It contains 1,271 sentence pairs manually translated by a human translator from Senegalese news articles, quality-checked with manual alignment review and LASER3 semantic-similarity screening, and released under the CC BY-NC license on Hugging Face and GitHub with fixed train/dev/test splits.
How well do current models translate between Wolof and Arabic?
According to the paper's benchmarks, zero-shot multilingual models are near-unusable (about 1-5 BLEU). After fine-tuning on MudawanSn, the best model (AfriNLLB-12) reaches 7.76 BLEU / 30.72 chrF++ for Wolof-to-Arabic and 8.75 BLEU / 33.08 chrF++ for Arabic-to-Wolof on a 96-sentence test set. chrF++ is much higher than BLEU because the languages share little vocabulary, so character-level overlap is a fairer yardstick. These scores indicate usable but still far-from-fluent translation.
Can I use MudawanSn for commercial products?
No. The corpus is released under CC BY-NC, a non-commercial license (inherited from the MasakhaNER source data). It is suited for research, evaluation, and fine-tuning experiments; commercial deployment would require separate licensing of the underlying data.
Does the corpus include Wolofal (Wolof written in Arabic script)?
No. MudawanSn covers only Latin-script Wolof paired with Arabic-script Modern Standard Arabic. The authors list Wolofal (Ajami) coverage as future work, noting that transliteration tools and public Ajami-Wolof resources are currently scarce.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1 (pipeline overview) is available only as extracted text; its visual content was not independently analyzable.
- The Hugging Face and GitHub artifacts could not be inspected from the text, so actual data quality, license files, and code completeness are taken on the paper's word.
- Tables were extracted inline; minor formatting ambiguities (e.g., AfriCOMET column alignment in Table 3) were resolved from surrounding prose.
- The arXiv identifier (2609.17539) and the stamped submission date (14 Jul 2026) are inconsistent in the extracted text; versioning could not be verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.