Source-linked AI summary
MudawanSn: A Gold-Standard Wolof-Arabic Parallel Corpus for Machine Translation
Mouhamed Mbaye, Thierno Diop
TL;DR
Wolof–MSA lacks a publicly available, dedicated gold-standard parallel corpus, limiting direct resources for this low-resource language pair. MudawanSn provides 1,271 manually translated sentence pairs, a reproducible construction workflow, and benchmarks showing improved translation after fine-tuning. AfriNLLB-12 achieves the strongest reported scores in both directions.
Problem
No publicly available manually translated gold-standard parallel corpus had been developed specifically for Wolof–MSA, despite both languages appearing in broader multilingual resources.
Method
MudawanSn constructs 1,271 sentence-aligned Wolof-to-MSA pairs from Senegalese news and documents alignment, semantic verification, and quality-control procedures.
Results
Fine-tuning improves translation quality in both directions, with AfriNLLB-12 reaching 7.76 BLEU and 30.72 chrF++ for Wolof-to-Arabic and 8.75 BLEU and 33.08 chrF++ for Arabic-to-Wolof.
Takeaways & Limitations
The released corpus, splits, and code provide a foundation for machine translation and broader linguistic and NLP research on Wolof–Arabic and related low-resource African language pairs.
Takeaways & Limitations
The corpus covers Latin-orthography Wolof and Senegalese online news, while the Arabic side was produced by a single translator without independent inter-translator agreement.
Abstract
from arXiv · showhide
We present MudawanSn, a gold-standard resource of 1,271 sentence-aligned pairs manually translated from Wolof into Modern Standard Arabic (MSA). The source texts are drawn from the MasakhaNER corpus and cover politics, society, religion, and sports in Senegalese news discourse. Although multilingual resources such as FLORES-200 and NTREX include both Wolof and Arabic, no publicly available parallel corpus is specifically designed for the Wolof-Modern Standard Arabic language pair. We describe the corpus construction protocol, sentence alignment procedure, and quality-control workflow. We benchmark four machine translation systems spanning three architectural families: NLLB-200 (600M), mT5-base, and two AfriNLLB variants, showing that fine-tuning on MudawanSn yields substantial improvements in both translation directions. The best-performing model, AfriNLLB-12, achieves 7.76 BLEU and 30.72 chrF++ for Wolof-to-Arabic, and 8.75 BLEU and 33.08 chrF++ for Arabic-to-Wolof. The corpus is released under the CC BY-NC license and is publicly available on Hugging Face and GitHub.
1 Introduction
Wolof–MSA lacks a publicly available, manually curated parallel corpus despite strong linguistic and cultural ties and existing multilingual coverage. MudawanSn addresses this gap with 1,271 aligned pairs, a reproducible construction protocol, and machine-translation benchmarks showing gains from fine-tuning.
- Existing resources pair Wolof mainly with French, while multilingual collections containing Wolof and Arabic were not designed as dedicated, manually curated Wolof–MSA corpora.
- MudawanSn releases 1,271 manually translated and sentence-aligned Wolof–MSA pairs with documentation and reproduction resources.
- The corpus construction protocol combines sentence extraction, translation, multilingual alignment, LASER3 semantic verification, and rigorous quality control.
- Fine-tuning substantially improves translation quality in both directions across four evaluated architectures, with AfriNLLB-12 achieving the best overall performance.
2 Related Work
Existing Wolof resources are concentrated in French- or English-centered collections, while Arabic–African pairs remain comparatively scarce and lack dedicated gold-standard data. MudawanSn fills this gap as a public, human-curated Wolof–MSA news corpus with provenance and reproducible baselines.
- 2.1 Parallel Corpora Involving Wolof: FLORES-200 and NTREX include Wolof and Arabic but are evaluation benchmarks rather than large-scale training corpora.
- 2.1 Parallel Corpora Involving Wolof: No publicly available, manually translated, gold-standard corpus had previously targeted Wolof–MSA specifically; MudawanSn addresses this gap with baselines and a reproducible protocol.
- 2.2 Parallel Corpora Involving Arabic and African Languages: Arabic–African language pairs remain relatively scarce in dedicated resources, and African-focused initiatives have not yet supplied curated, sentence-aligned gold-standard corpora for specific pairs.
- 2.3 Positioning of Our Contribution: MudawanSn is presented as the first publicly available human-curated sentence-level Wolof–MSA resource for the news domain.
- 2.3 Positioning of Our Contribution: Compared with prior resources, MudawanSn is manually verified, culturally situated in Senegalese news, and released with provenance tracing each pair to its upstream source.
3 Corpus Construction and Quality Control
MudawanSn was constructed through a reproducible pipeline that cleans, enriches, aligns, and semantically verifies manually translated Wolof–Arabic sentence pairs. Manual inspection and LASER3-based checks support the corpus’s alignment quality.
- Corpus construction: 1,271 Wolof–Arabic pairs were manually translated, sentence-aligned, and processed through an end-to-end cleaning, enrichment, alignment, and semantic-verification pipeline.The Wolof sources are Senegalese news covering politics, society, religion, and sports.
- Multilingual source matching: 98.19% of pairs were enriched with original French sources and metadata through exact, substring, and fuzzy token-overlap matching.The remaining 1.81% retain MasakhaNER fallback records.
- Sentence alignment and quality control: Each pair underwent manual inspection, while LASER3 similarity provided quantitative alignment validation.The workflow targeted one-to-one semantic correspondence and checked for alignment shifts, omissions, and spelling variations.
- Sentence alignment and quality control: 0.7508 mean and 0.7649 median LASER3 similarity scores were obtained across the 1,271 pairs.96.06% of pairs scored above 0.60, and all pairs below that threshold received manual review.
- Corpus statistics: The corpus contains 25,543 Wolof words and 17,972 Arabic words, with source texts predominantly drawn from MAFAND.Statistics use whitespace tokenization to avoid overfragmentation from multilingual subword tokenizers.
4 Baseline Experiments
The experiments compare four systems across three architectural families under zero-shot and bidirectional fine-tuning settings. Fine-tuning improves the multilingual models, with AfriNLLB-12 achieving the strongest lexical scores in both translation directions, while metric rankings differ for AfriCOMET.
- Experimental setup: The benchmark covers NLLB-200, mT5-base, and full and pruned AfriNLLB models evaluated zero-shot and after ten epochs of bidirectional fine-tuning.Metrics include BLEU, chrF++, and AfriCOMET.
- Results: 7.76 BLEU and 30.72 chrF++ made AfriNLLB-12 the strongest lexical system for Wolof-to-Arabic, while 8.75 BLEU and 33.08 chrF++ led Arabic-to-Wolof.NLLB-200 also improved substantially with fine-tuning, and AfriNLLB-8 remained within 0.7 BLEU of AfriNLLB-12 in Wolof-to-Arabic.
- Results: Fine-tuning improved NLLB-200 by 5.35 BLEU and 13.89 chrF++ in Wolof-to-Arabic, with consistent gains across multilingual systems.The evaluation uses 85% training, 7.5% development, and 7.5% test data.
- Results: mT5-base gained slightly when Arabic was the target but showed no gain when Wolof was the target.The paper relates this asymmetry to mC4’s substantial Arabic coverage and virtually absent Wolof coverage.
- Results: NLLB-200 fine-tuned achieved the highest AfriCOMET scores, although AfriNLLB-12 led BLEU and chrF++ in both directions.Because AfriCOMET was not trained on Wolof–Arabic data, lexical metrics are used as the primary comparison and human evaluation is left for future work.
5 Conclusion
MudawanSn contributes a publicly available gold-standard Wolof–Arabic corpus of 1,271 manually translated and aligned sentence pairs, together with baseline experiments. Its open release is positioned as a foundation for broader linguistic and low-resource African NLP research.
- Conclusion: MudawanSn is a publicly available gold-standard corpus of 1,271 manually translated and aligned Wolof–Arabic sentence pairs from Senegalese news.The paper also reports baseline machine translation experiments spanning three architectural families.
- Conclusion: The corpus is presented as a foundation for contrastive Wolof–Arabic studies, Wolof-aware tokenizers and morphological analyzers, and broader low-resource African NLP research.The corpus, splits, and code are released openly under a CC BY-NC license.
Limitations
MudawanSn’s coverage is constrained by its Latin-only script, modest size, single-translator design, selective treatment of idioms, and Senegalese online-news domain focus.
- Script coverage: The corpus excludes Wolofal, limiting coverage to Latin-orthography Wolof and leaving Arabic-derived script effects for future study.The authors identify scarce Ajami-Wolof resources and limited Latin–Ajami transliteration tools as barriers to extension.
- Corpus size: At 1,271 sentence-aligned pairs, the corpus remains modest in size, with scaling and broader domain coverage left for future work.The authors attribute this constraint to the manual effort and multi-stage quality control required to build the resource from scratch.
- Single-translator design: A single bilingual translator produced the Arabic side, so translation choices may reflect individual stylistic and lexical preferences without inter-translator agreement measurement.The workflow included dictionary verification, native-speaker consultation, and human review of flagged semantic-verification pairs, but no independent second translation.
- Coverage of idiomatic expressions: The translation process avoided highly localized idioms, improving mapping and alignment robustness but underrepresenting figurative Senegalese news Wolof.This trade-off prioritizes unambiguous parallel quality for machine-translation training over full idiomatic coverage.
- Domain coverage: Because all texts come from Senegalese online news, politics and civic discourse are over-represented while conversational, literary, religious, and technical registers are absent.Systems fine-tuned only on this corpus should therefore transfer imperfectly to other registers.