Elamur Research⌕
☰
home / papers / 2022 / Arabic Dialect Identification with a Few Labeled Examples Using Generative Adversarial Networks

Arabic Dialect Identification with a Few Labeled Examples Using Generative Adversarial Networks

Paper · 2022 · Mahmoud Yusuf, Marwan Torki and Nagwa El-Makky

Summary

The paper addresses Arabic dialect identification when only a handful of labeled examples exist for each dialect. Labeled data for this task is scarce, unevenly distributed across countries, and often noisy, because some corpora assign labels from a user's location. The authors attach a Semi-Supervised GAN (SS-GAN), following the GAN-BERT design, to the Arabic pretrained encoders ARBERT and MARBERT. A generator produces fake sentence vectors from noise. A discriminator then has to separate real from fake while also predicting the dialect class, and it learns from both labeled and unlabeled examples. They also try a 2-stage schedule: first train with the GAN attached, then remove it and keep fine-tuning the plain encoder. Experiments use tiny random subsets (0.01%–10%) of four corpora: QADI, NADI 2021, ArSarcasm and AOC. In the smallest-data regimes, macro-F1 often improves two- to four-fold over plain fine-tuning, and training converges faster. As the labeled data grows, the gap disappears. The result matters for any low-resource variety where raw text is cheap but annotation is expensive.

Key points

  • Base models are ARBERT and MARBERT. The SS-GAN extension is modeled on GAN-BERT (Croce et al., 2020), and the authors report it as the first use of this setup for Arabic dialect identification.
  • The four corpora are QADI (367,353 train / 3,304 test, 18 country dialects, clean and fairly balanced), NADI 2021 (imbalanced, with 4 classes covering over half the data), ArSarcasm (mostly MSA/Egyptian) and AOC (86,542 train / 10,812 test, 4 classes).
  • Selected GAN-MARBERT vs MARBERT macro-F1 results:
    • NADI at 1% (210 examples): about 8 vs 2.8.
    • AOC at 0.02% (17 examples): over 26 vs 17.
    • AOC at 0.05% (184 examples): 47 vs 18.
    • QADI at 0.01–0.02% (37–74 examples): more than four times the baseline.
  • The 2-stage setup uses 5 epochs with the GAN and 5 without, learning rate 2e-5 and sequence length 40. Its results versus plain fine-tuning:
    • ArSarcasm at 1%: 32 vs 20.5 (ARBERT).
    • NADI at 1%: 10.9 vs 7.2.
    • QADI at 0.01%: 7.8 vs 2.2.
    • AOC at 0.05%: 43.9 vs 18.5.
  • Gains shrink or vanish with more labels. For example, ArSarcasm at 5% gives 45.7 for the 2-stage setup vs 47 for ARBERT.

Interesting ideas

  • I1. The discriminator gets one extra "fake" class on top of the real labels, so unlabeled real text still trains it by teaching what real input looks like.
  • I2. The adversarial part is temporary scaffolding: it helps early training and is then removed, so inference costs nothing extra.
  • I3. The authors explain why standard augmentation fails here: word swaps lose meaning, dialect synonym lists barely exist, and back-translation drifts toward the standard variety (MSA) and erases the very feature being classified.
  • I4. The evaluation sweeps learning curves down to a few dozen labeled examples, which shows where an extension helps and where it stops mattering.
  • I5. The paper warns that labels taken from metadata (user location) introduce systematic noise that limits generalization.

For Mova

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

  • I1 (learn): A parser or tagger head could use this to draw signal from large raw English corpora while the gold set stays tiny, as a complement to silver data from witness parsers.
  • I2 (tree): gparse could warm up with an auxiliary adversarial or unlabeled-data module for a few epochs, then drop it and ship the unchanged parser.
  • I3 (back): When the text → graph → text compressor or translation is used to create training data, check that it keeps register and dialect markers and does not normalize them away.
  • I4 (learn): Report each new component at several training-set fractions, with the plain baseline as a control, instead of a single full-data score.
  • I5 (learn): Treat any label taken from document source or metadata as a weak witness that must agree with others before it enters silver data.

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

🎓Alexandria UniversityAlexandria · Egypt39 works · 287 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 2,908 works

Cite

Cite as Mahmoud Yusuf, Marwan Torki and Nagwa El-Makky. 2022. Arabic Dialect Identification with a Few Labeled Examples Using Generative Adversarial Networks. Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2022.aacl-main.16

@inproceedings{yusuf2022arabic,
  title = {{Arabic Dialect Identification with a Few Labeled Examples Using Generative Adversarial Networks}},
  author = {Mahmoud Yusuf, Marwan Torki and Nagwa El-Makky},
  booktitle = {Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)},
  year = {2022},
  doi = {10.18653/v1/2022.aacl-main.16},
  url = {https://doi.org/10.18653/v1/2022.aacl-main.16}
}

References

18 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ahmed Abdelali, Hamdy Mubarak, Younes Samih, Sabit Hassan, and Kareem Darwish. 2021. QADI: Arabic dialect identification in the wild. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 1–10, Kyiv, Ukraine (Virtual). Association for Computational Linguistics.
  2. Muhammad Abdul-Mageed, AbdelRahim Elmadany, et al. 2021a. Arbert & marbert: Deep bidirectional transformers for arabic. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 7088–7105. ↗
  3. Muhammad Abdul-Mageed, Chiyu Zhang, Houda Bouamor, and Nizar Habash. 2020. Nadi 2020: The first nuanced arabic dialect identification shared task. In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pages 97–110. ↗
  4. Muhammad Abdul-Mageed, Chiyu Zhang, AbdelRahim Elmadany, Houda Bouamor, and Nizar Habash. 2021b. Nadi 2021: The second nuanced arabic dialect identification shared task. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 244–259. ↗
  5. Badr AlKhamissi, Mohamed Gabr, Muhammad El-Nokrashy, and Khaled Essam. 2021. Adapting marbert for improved arabic dialect identification: Submission to the nadi 2021 shared task. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 260–264. ↗
  6. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. Arabert: Transformer-based model for arabic language understanding. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Offensive Language Detection, pages 9–15. ↗
  7. Laila Bashmal and Daliyah AlZeer. 2021. Arsarcasm shared task: An ensemble bert model for sarcas-mdetection in arabic tweets. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 323–328. find
  8. Houda Bouamor, Sabit Hassan, and Nizar Habash. 2019. The madar shared task on arabic fine-grained dialect identification. In Proceedings of the Fourth Arabic Natural Language Processing Workshop, pages 199– 207.
  9. Danilo Croce, Giuseppe Castellucci, and Roberto Basili. 2020. Gan-bert: Generative adversarial learning for robust text classification with a bunch of labeled examples. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2114–2119. find
  10. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171– 4186. ↗
  11. Siddhant Garg and Goutham Ramakrishnan. 2020. Bae: Bert-based adversarial examples for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6174–6181.
  12. Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets. Advances in neural information processing systems, 27. find
  13. Moataz Mansour, Moustafa Tohamy, Zeyad Ezzat, and Marwan Torki. 2020. Arabic dialect identification using BERT fine-tuning. In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pages 308–312, Barcelona, Spain (Online). Association for Computational Linguistics.
  14. Mahmoud Nabil, Mohamed Aly, and Amir Atiya. 2015. Astd: Arabic sentiment tweets dataset. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 2515–2519. ↗
  15. Sara Rosenthal, Noura Farra, and Preslav Nakov. 2017. Semeval-2017 task 4: Sentiment analysis in twitter. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 502– 518.
  16. Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. 2016. Improved techniques for training gans. Advances in neural information processing systems, 29:2234– 2242. ↗
  17. Toshiyuki Takezawa, Genichiro Kikui, Masahide Mizushima, and Eiichiro Sumita. 2007. Multilingual spoken language corpus development for communication research. In International Journal of Computational Linguistics & Chinese Language Processing, Volume 12, Number 3, September 2007: Special Issue on Invited Papers from ISCSLP 2006, pages 303–324. find
  18. Omar Zaidan and Chris Callison-Burch. 2011. The arabic online commentary dataset: an annotated dataset of informal arabic with high dialectal content. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 37–41.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yusuf-2022-arabic-dialect-identification-few-labeled/