---
title: "Low-Resource Machine Translation for Moroccan Arabic (2025)"
description: "The authors train systems in both directions on the Darija Open Dataset (DODa), using a random 80/15/5 split. 26.79 BLEU. 22 references."
authors: ["Alexei Rosca", "Abderrahmane Issam", "Gerasimos Spanakis"]
year: 2025
venue: "Proceedings of the First Workshop on Advancing NLP for Low-Resource Languages"
source: "https://aclanthology.org/2025.lowresnlp-1.5/"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/rosca-2025-low-resource-machine-translation-moroccan/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/rosca-2025-low-resource-machine-translation-moroccan/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/18/) / Low-Resource Machine Translation for Moroccan Arabic

# Low-Resource Machine Translation for Moroccan Arabic

**Authors:** Alexei Rosca, [Abderrahmane Issam](https://md.research.elamur.ai/free/en/authors/abderrahmane-issam/) and [Gerasimos Spanakis](https://md.research.elamur.ai/free/en/authors/gerasimos-spanakis/) · **Year:** 2025 · **Venue:** [Proceedings of the First Workshop on Advancing NLP for Low-Resource Languages](https://md.research.elamur.ai/free/en/venues/lowresnlp/) ([Workshop on Advancing NLP for Low-Resource Languages 2025](https://md.research.elamur.ai/free/en/venues/lowresnlp/2025/)) · **Publisher:** [INCOMA Ltd.](https://md.research.elamur.ai/free/en/publishers/incoma-ltd/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2025.lowresnlp-1.5/) · [PDF](https://aclanthology.org/2025.lowresnlp-1.5.pdf) · [OpenAlex](https://openalex.org/W7133186254)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
The paper asks which standard low-resource NMT techniques actually help for Moroccan Arabic (Darija), a dialect with very little parallel data. The authors train systems in both directions on the Darija Open Dataset (DODa), using a random 80/15/5 split. They compare three strategies: back-translation, paraphrase augmentation with a BART-based paraphraser, and transfer learning. For transfer learning they fine-tune three models: the multilingual NLLB, Arabic-oriented OPUS-MT models, and BERT-OPUS, a hybrid built from a DarijaBERT encoder and an OPUS-MT Arabic→English decoder. Pretrained models combined with back-translated data work best, reaching 26.79 BLEU for Darija→English and 9.98 for English→Darija. Two weaknesses stand out. Gains seen on in-domain DODa data often shrink or even reverse on the BIBLE and MADAR test sets. Translating into Darija also remains far harder than translating out of it. The study gives a realistic picture of what current recipes deliver for dialects and where they still fail.

## Key points
- Training data is DODa, with more than 45,000 Darija–English pairs. Out-of-domain testing uses BIBLE (500 New Testament pairs) and MADAR (5,500 pairs).
- BERT-OPUS (about 200M parameters) gives the top Darija→English score of 26.79 BLEU with back-translation. It edges out NLLB-200-distilled-600M, which scores 26.65 with back-translation.
- Fine-tuning NLLB raises its Darija→English score on DODa from 8.66 to 26.50 BLEU. On BIBLE, however, the same fine-tuning drops it from 20.43 to 13.53.
- For English→Darija, NLLB reaches 9.98 BLEU with back-translation. OPUS-MT-En-Ar reaches only 2.58 after plain fine-tuning, which the authors attribute to its decoder being trained to produce MSA. Across models, the best result in one direction is more than 16 BLEU points above the best in the other.
- Paraphrasing helps when applied to the source side (English→Darija, 9.68 vs 8.10 BLEU for NLLB). It hurts when applied to the target side (Darija→English, 20.57 vs 26.50 BLEU).

## Interesting ideas
- When augmenting data, rewrite only the input side and keep the gold output untouched; paraphrasing the targets added noise to what the model had to imitate.
- A model fine-tuned on one dataset should always be scored on test sets from other domains, because in-domain gains here sometimes turned into losses on BIBLE and MADAR.
- An encoder-decoder can be stitched together from separately pretrained parts: a language-specific encoder plus a decoder from a related high-resource pair, then fine-tuned jointly. This smaller hybrid matched a larger multilingual model.
- Back-translation works by running a reverse-direction model over plentiful target-side text to create synthetic training pairs, which gave the most reliable gains in this study.
- Generating into the poorly resourced side proved much harder than analysing it, so each direction deserves its own separate measurement.

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (learn): when expanding training sets for the readers, rephrase the questions or story text while leaving the gold answers and frames fixed, so augmentation never corrupts the supervision signal.
- **I2** (learn): every retraining round could report scores on held-out genres alongside the in-domain set, so that a "win" that only overfits one source is caught early.
- **I3** (tree): gparse could warm-start its encoder from a domain- or register-specific pretrained model rather than a generic one, reusing a separately trained head for the dependency output.
- **I4** (learn): the graph→text generator could turn unlabeled meaning graphs into synthetic sentences, giving the parser extra aligned text–graph pairs that can be checked against witness parsers.
- **I5** (back): evaluate the text → graph → text compressor separately for parsing and for generation, since a strong analysis score may hide a much weaker surface realizer.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [Maastricht University](https://md.research.elamur.ai/free/en/universities/maastricht-university/) — University, Maastricht, Netherlands · 7 works · 605 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works

🧭 Topics (3): [🔁 Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) · 1,105 works · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 3,066 works · [🌱 Low-resource languages](https://md.research.elamur.ai/free/en/topics/low-resource-languages/) · 287 works

## Cite

Cite as: Alexei Rosca, Abderrahmane Issam and Gerasimos Spanakis. 2025. Low-Resource Machine Translation for Moroccan Arabic. Proceedings of the First Workshop on Advancing NLP for Low-Resource Languages.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{rosca2025low,
  title = {{Low-Resource Machine Translation for Moroccan Arabic}},
  author = {Alexei Rosca, Abderrahmane Issam and Gerasimos Spanakis},
  booktitle = {Proceedings of the First Workshop on Advancing NLP for Low-Resource Languages},
  year = {2025},
  doi = {10.26615/978-954-452-100-4-005},
  url = {https://doi.org/10.26615/978-954-452-100-4-005}
}
```

- [← previous Loanwords in the Language of Archaic Biblical Hebrew Poetry: A Case for Akkadianisms](https://md.research.elamur.ai/free/en/papers/notarius-2025-loanwords-language-archaic-biblical-hebrew/)
- [next → Lucas, Peter J. 2024. Old English Poetry from Manuscript to Message. Utrecht Studies in Medieval Literacy 58. Turnhout: Brepols. Pp. xviii + 398. ISBN 9782503600314.](https://md.research.elamur.ai/free/en/papers/griffith-2025-lucas-peter-j-2024-old/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/18/)

## References

22 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Francis Bond and Ryan Foster. 2013. [Linking and extending an open multilingual Wordnet](https://md.research.elamur.ai/free/en/papers/bond-2013-linking-extending-open-multilingual-wordnet/). In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1352–1362, Sofia, Bulgaria. Association for Computational Linguistics.
2. Jeremy Howard and Sebastian Ruder. 2018. Universal language model fine-tuning for text classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 328–339, Melbourne, Australia. Association for Computational Linguistics. [↗](https://doi.org/10.18653/v1/p18-1031)
3. Houda Bouamor, Nizar Habash, Mohammad Salameh, Wajdi Zaghouani, Owen Rambow, Dana Abdulrahim, Ossama Obeid, Salam Khalifa, Fadhl Eryani, Alexander Erdmann, and 1 others. 2018. The madar arabic dialect corpus and lexicon. In Proceedings of the eleventh international conference on language resources and evaluation (LREC 2018). [find](https://scholar.google.com/scholar?q=Houda+Bouamor%2C+Nizar+Habash%2C+Mohammad+Salameh%2C+Wajdi+Zaghouani%2C+Owen+Rambow%2C+Dana+Abdulrahim%2C+Ossama+Obeid%2C+Salam+Khalifa%2C+Fadhl+Eryani%2C+Alexander+Erdmann%2C+and+1+others%2E+2018%2E+The+madar+arabic+dialect)
4. Abderrahmane Issam and Khalil Mrini. 2021. Goud.ma: a news article dataset for summarization in moroccan darija. In 3rd Workshop on African Natural Language Processing. [find](https://scholar.google.com/scholar?q=Abderrahmane+Issam+and+Khalil+Mrini%2E+2021%2E+Goud%2Ema%3A+a+news+article+dataset+for+summarization+in+moroccan+darija%2E+In+3rd+Workshop+on+African+Natural+Language+Processing%2E)
5. ElMehdi Boujou, Hamza Chataoui, Abdellah el Mekki, Saad Benjelloun, Ikram Chairi, and Ismail Berrada. 2021. [An open access nlp dataset for arabic dialects : Data collection, labeling, and model construction](https://md.research.elamur.ai/free/en/papers/boujou-2021-open-access-nlp-dataset-arabic/).
6. Melvin Johnson, Mike Schuster, Quoc V Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, and 1 others. 2017. [Google’s multilingual neural machine translation system: Enabling zero-shot translation](https://md.research.elamur.ai/free/en/papers/johnson-2017-google-s-multilingual-neural-machine/). Transactions of the Association for Computational Linguistics, 5:339–351.
7. Chris Callison-Burch, Philipp Koehn, and Miles Osborne. 2006. Improved statistical machine translation using paraphrases. In Human Language Technology Surafel M Lakew, Stephan Gouws, Yulia Tsvetkov, Vukosi Marivate, and Stefan Weber. 2020. Low-resource neural machine translation: A benchmark for five african languages. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2602–2611. The University of Edinburgh’s neural MT systems for WMT17. In Proceedings of the Second Conference on Machine Translation, pages 389–399, Copenhagen, Denmark. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Chris+Callison%2DBurch%2C+Philipp+Koehn%2C+and+Miles+Osborne%2E+2006%2E+Improved+statistical+machine+translation+using+paraphrases%2E+In+Human+Language+Technology+Surafel+M+Lakew%2C+Stephan+Gouws%2C+Yulia+Tsvetkov%2C+V)
8. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 86–96, Berlin, Germany. Association for Computational Linguistics. [↗](https://doi.org/10.18653/v1/p16-1009)
9. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. [BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension](https://md.research.elamur.ai/free/en/papers/lewis-2020-bart-denoising-sequence-sequence-pre/). In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online. Association for Computational Linguistics. [↗](https://doi.org/10.18653/v1/2020.acl-main.703)
10. Ridouane Tachicart, Karim Bouzoubaa, and Hamid Jaafar. 2014. Building a moroccan dialect electronic dictionary (mded). [find](https://scholar.google.com/scholar?q=Ridouane+Tachicart%2C+Karim+Bouzoubaa%2C+and+Hamid+Jaafar%2E+2014%2E+Building+a+moroccan+dialect+electronic+dictionary+%28mded%29%2E)
11. Jonathan Mallinson, Rico Sennrich, and Maria Lapata. 2017. [Paraphrasing revisited with neural machine translation](https://md.research.elamur.ai/free/en/papers/mallinson-2017-paraphrasing-revisited-neural-machine-translation/). In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 881–893. Association for Computational Linguistics (ACL). The 15th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2017 ; Conference date: 03-04-2017 Through 07-04-2017.
12. NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, and 20 others. 2022. No language left behind: Scaling human-centered machine translation. Preprint, arXiv:2207.04672. [↗](https://arxiv.org/abs/2207.04672)
13. Jörg Tiedemann and Santhosh Thottingal. 2020. OPUS-MT – building open translation services for the world. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation, pages 479–480, Lisboa, Portugal. European Association for Machine Translation. [find](https://scholar.google.com/scholar?q=J%C3%B6rg+Tiedemann+and+Santhosh+Thottingal%2E+2020%2E+OPUS%2DMT+%E2%80%93+building+open+translation+services+for+the+world%2E+In+Proceedings+of+the+22nd+Annual+Conference+of+the+European+Association+for+Machine+Translati)
14. Hanane Nour Moussa and Asmaa Mourhir. 2023. Darnercorp: An annotated named entity recognition dataset in the moroccan dialect. Data in Brief, 48:109234. [find](https://scholar.google.com/scholar?q=Hanane+Nour+Moussa+and+Asmaa+Mourhir%2E+2023%2E+Darnercorp%3A+An+annotated+named+entity+recognition+dataset+in+the+moroccan+dialect%2E+Data+in+Brief%2C+48%3A109234%2E)
15. Khalil Mrini and Francis Bond. 2017. Building the moroccan darija wordnet (mdw) using bilingual resources. [find](https://scholar.google.com/scholar?q=Khalil+Mrini+and+Francis+Bond%2E+2017%2E+Building+the+moroccan+darija+wordnet+%28mdw%29+using+bilingual+resources%2E)
16. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. [Attention is all you need](https://md.research.elamur.ai/free/en/papers/vaswani-2017-attention-all-need/). In Advances in neural information processing systems, pages 5998–6008.
17. Aissam Outchakoucht and Hamza Es-Samaali. 2021. [Moroccan dialect -darija- open dataset](https://md.research.elamur.ai/free/en/papers/outchakoucht-2021-moroccan-dialect-darija-open-dataset/). Preprint, arXiv:2103.09687. [↗](https://arxiv.org/abs/2103.09687)
18. Aissam Outchakoucht and Hamza Es-Samaali. 2024. [The evolution of darija open dataset: Introducing version 2](https://md.research.elamur.ai/free/en/papers/outchakoucht-2024-evolution-darija-open-dataset-introducing/). Preprint, arXiv:2405.13016. [↗](https://arxiv.org/abs/2405.13016)
19. Pidong Wang, Preslav Nakov, and Hwee Tou Ng. 2016. Source language adaptation approaches for resource-poor machine translation. Computational Linguistics, 42(2):277–306. [find](https://scholar.google.com/scholar?q=Pidong+Wang%2C+Preslav+Nakov%2C+and+Hwee+Tou+Ng%2E+2016%2E+Source+language+adaptation+approaches+for+resource%2Dpoor+machine+translation%2E+Computational+Linguistics%2C+42%282%29%3A277%E2%80%93306%2E)
20. Sascha Rothe, Shashi Narayan, and Aliaksei Severyn. 2020. [Leveraging pre-trained checkpoints for sequence generation tasks](https://md.research.elamur.ai/free/en/papers/rothe-2019-leveraging-pre-trained-checkpoints-sequence/). Transactions of the Association for Computational Linguistics, 8:264–280.
21. Barret Zoph, Deniz Yuret, Jonathan May, and Kevin Knight. 2016. Transfer learning for low-resource neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1568–1575, Austin, Texas. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Barret+Zoph%2C+Deniz+Yuret%2C+Jonathan+May%2C+and+Kevin+Knight%2E+2016%2E+Transfer+learning+for+low%2Dresource+neural+machine+translation%2E+In+Proceedings+of+the+2016+Conference+on+Empirical+Methods+in+Natural+Lan)
22. Hassan Sajjad, Ahmed Abdelali, Nadir Durrani, and Fahim Dalvi. 2020. Arabench: Benchmarking dialectal arabic-english machine translation. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5094–5107. [find](https://scholar.google.com/scholar?q=Hassan+Sajjad%2C+Ahmed+Abdelali%2C+Nadir+Durrani%2C+and+Fahim+Dalvi%2E+2020%2E+Arabench%3A+Benchmarking+dialectal+arabic%2Denglish+machine+translation%2E+In+Proceedings+of+the+28th+International+Conference+on+Computa)

*Updated 2 November 2026*