Elamur Research⌕
☰
home / papers / 2026 / BARRAC: Adaptation of an English Aspect-based Sentiment Analysis Approach for Classification Tasks in Arabic Dialects

BARRAC: Adaptation of an English Aspect-based Sentiment Analysis Approach for Classification Tasks in Arabic Dialects

Summary

Arabic NLP has many dedicated models and datasets, but labelled data for dialectal tasks such as sarcasm detection, dialect identification and cross-dialect sentiment is still scarce. The authors test whether a method built for English can be adapted to fill this gap, rather than building yet another resource. They take DS²-ABSA, a framework that synthesises training data for aspect-based sentiment analysis of consumer reviews, and adapt it as BARRAC:

  • Attributes: product-centred aspect categories and aspect terms become task-specific Arabic linguistic devices and linguistic markers.
  • Domains: the two review domains are widened to ten public-discourse domains.
  • Training: noisy self-training is replaced by two stages. The model first trains on GPT-4o-generated pseudo-labelled data, then fine-tunes on just 100 gold labels.

Across five dialect datasets with an ARBERTv2 backbone, BARRAC reaches a mean macro-F1 of 63.93. This beats the strongest few-label method by about 3 points and is on par with GPT-4o on average, using far fewer parameters. The work suggests that porting task-level methods across languages is a viable complement to porting models and benchmarks.

Key points

  • The evaluation covers five public datasets and three tasks: Ar-Sentiment, Ar-Sarcasm, Ar-Dialects (Abu Farha and Magdy, 2020), Sa'7r and DART. Every model gets 100 real labels, and BARRAC is compared against 14 baselines.

  • BARRAC scores 63.93 mean macro-F1. That is +8.48 over the best supervised baseline and roughly +3 over few-label SOTA (Cluster&Tune, IDoFew). GPT-4o averages 63.98, and BARRAC beats it on four of five datasets. GPT-4o stays clearly ahead on Ar-Sentiment (72.47 vs 65.61).

  • Intermediate training (synthetic first, then gold) beats noisy self-training (best 60.92) by +3.01. The largest gains are on Ar-Dialects (+4.82) and Ar-Sarcasm (+2.63).

  • The ablation shows how much each part contributes:

    Removed componentChange in mean F1
    Instance-driven stream−4.11
    Key-point-driven stream−1.61
    Label normalisation−2.02
    Synthetic stage−8.47
    Gold fine-tuning stage−27.26

    Without the gold stage, Ar-Sentiment falls to 16.06 and Ar-Dialects to 13.47.

  • A manual error analysis of 20 misclassified examples per dataset groups errors into four kinds: insufficient context, linguistic overlap, label ambiguity and annotation ambiguity. Low-support dialects (Gulf, Levantine, Maghrebi) remain weak spots.

Interesting ideas

  • I1. Generate synthetic examples from a two-level attribute hierarchy: first a coarse mechanism (e.g. verbal_irony or a dialect region), then fine surface cues tied to it, and ask the LLM to write sentences satisfying both, so the generated data covers phenomena deliberately rather than at random. — For Mova (learn): LLM teacher tiers could generate sentences on demand for rare constructions by pairing a phenomenon (e.g. raising verbs) with specific trigger words, which would fill gaps in silver data.
  • I2. Use noisy synthetic labels only as a starting point for the model's weights, and let a small set of clean gold labels make the final adjustments, instead of mixing the two throughout training; this keeps label noise from persisting into the final model. — For Mova (tree): gparse could pretrain on large silver treebanks and then do a final fine-tuning pass solely on gold UD trees, rather than training on silver and gold mixed together.
  • I3. Run an ablation that removes the gold stage entirely to measure how far synthetic data alone can go; it showed that synthetic data handles coarse binary distinctions but cannot set fine-grained class boundaries. — For Mova (learn): this works as a negative control: train on silver only, compare against silver+gold, and see which relations or labels silver data fails to calibrate.
  • I4. In error analysis, keep a separate category for cases where the gold label itself looks debatable, apart from genuine model errors, which helps estimate how much error is not learnable at all. — For Mova (answer): failures on school tests and fable QA could be tagged as "questionable key" or "insufficient context" before anyone writes repair rules for them.
  • I5. Adapt a method to a new domain by swapping only the task objective in a fixed prompt template, keeping the rest of the prompt unchanged, so the differences between tasks stay controlled and comparable. — For Mova (frames): one fixed teacher-prompt template for labelling semantic roles, speech acts or irony, varying only the objective, would make annotations from different tasks consistent and easier to audit.

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

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2609.38820 · arXiv 2609.38820

@misc{almutairi2026barrac,
  title = {{BARRAC: Adaptation of an English Aspect-based Sentiment Analysis Approach for Classification Tasks in Arabic Dialects}},
  author = {Almutairi, Ali and Mohammadi, Gelareh and Razzak, Imran and Joshi, Aditya},
  year = {2026},
  doi = {10.48550/arxiv.2609.38820},
  eprint = {2609.38820},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2609.38820}
}

References

25 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sura F. Abdhood, Nazlia Omar, and Sabrina Tiun. 2025. Data augmentation for arabic text classification: A review of current methods, challenges and prospective directions. PeerJ Computer Science, 11:e2685. find
  2. Muhammad Abdul-Mageed, AbdelRahim Elmadany, and El Moatez Billah Nagoudi. 2021. 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, Online. Association for Computational Linguistics. ↗ our card
  3. Muhammad Abdul-Mageed, AbdelRahim Elmadany, Chiyu Zhang, El Moatez Billah Nagoudi, Houda Bouamor, and Nizar Habash. 2023. NADI 2023: The fourth nuanced Arabic dialect identification shared task. In Proceedings of ArabicNLP 2023, pages 600– 613, Singapore (Hybrid). Association for Computational Linguistics. find
  4. Muhammad Abdul-Mageed, Amr Keleg, Abdel-Rahim Elmadany, Chiyu Zhang, Injy Hamed, Walid Magdy, Houda Bouamor, and Nizar Habash. 2024. NADI 2024: The fifth nuanced Arabic dialect identification shared task. In Proceedings of the Second Arabic Natural Language Processing Conference, pages 709–728, Bangkok, Thailand. Association for Computational Linguistics. find
  5. Ibrahim Abu Farha and Walid Magdy. 2020. From Arabic sentiment analysis to sarcasm detection: The ArSarcasm dataset. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Offensive Language Detection, pages 32–39, Marseille, France. find our card
  6. Halah AlMazrua, Najla AlHazzani, Amaal AlDa-wod, Lama AlAwlaqi, Noura AlReshoudi, Hend Al-Khalifa, and Luluh AlDhubayi. 2022. Sa‘7r: A saudi dialect irony dataset. In Proceedinsg of the 5th Workshop on Open-Source Arabic Corpora and Processing Tools with Shared Tasks on Qur’an QA and Fine-Grained Hate Speech Detection, pages 60–70, Marseille, France. European Language Resources Association. find
  7. Israa Alsarsour, Esraa Mohamed, Reem Suwaileh, and Tamer Elsayed. 2018. DART: A large dataset of dialectal Arabic tweets. In LREC 2018. find
  8. Abdullah Alsuhaibani, Hamad Zogan, Imran Razzak, Shoaib Jameel, and Guandong Xu. 2024. Idofew: Intermediate training using dual-clustering in language models for few labels text classification. In WSDM, page 18–27. find
  9. 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. find our card
  10. M Saiful Bari, Yazeed Alnumay, Norah A. Alzahrani, Nouf M. Alotaibi, Hisham Abdullah Alyahya, Sultan AlRashed, Faisal Abdulrahman Mirza, Shaykhah Z. Alsubaie, Hassan A. Alahmed, Ghadah Alabduljabbar, Raghad Alkhathran, Yousef Almushayqih, Raneem Alnajim, Salman Alsubaihi, Maryam Al Mansour, Saad Amin Hassan, Dr. Majed Alrubaian, Ali Alammari, Zaki Alawami, and 7 others. 2025. ALLam: Large language models for arabic and english. In The Thirteenth International Conference on Learning Representations. find our card
  11. Kenneth R. Beesley. 1996. Arabic finite-state morphological analysis and generation. In COLING 1996 Volume 1: The 16th International Conference on Computational Linguistics. find
  12. Yonatan Belinkov, Alexander Magidow, Alberto Barrón-Cedeño, Avi Shmidman, and Maxim Studying the history of Romanov. 2018. the arabic language: Language technology and a large-scale historical corpus. Preprint, arXiv:1809.03891. ↗
  13. Naaima Boudad, Rdouan Faizi, Rachid Oulad Haj Thami, and Raddouane Chiheb. 2018. Sentiment analysis in arabic: A review of the literature. Ain Shams Engineering Journal, 9(4):2479–2490. find
  14. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, pages 4171–4186. find
  15. Meta. 2024. Introducing meta llama 3: The most capable openly available llm to date. find
  16. OpenAI. 2025. GPT-Family (April 14 version). Large language model. find
  17. Ali Safaya, Moutasem Abdullatif, and Deniz Yuret. 2020. KUISAIL at SemEval-2020 task 12: BERT-CNN for offensive speech identification in social media. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 2054–2059, Barcelona (online). find our card
  18. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2020. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. Preprint, arXiv:1910.01108. ↗
  19. Haitham Seelawi, Ibraheem Tuffaha, Mahmoud Gzawi, Wael Farhan, Bashar Talafha, Riham Badawi, Zyad Sober, Oday Al-Dweik, Abed Alhakim Freihat, and Hussein Al-Natsheh. 2021. ALUE: Arabic language understanding evaluation. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 173– 184, Kyiv, Ukraine (Virtual). Association for Computational Linguistics. find
  20. Eyal Shnarch, Ariel Gera, Alon Halfon, Lena Dankin, Leshem Choshen, Ranit Aharonov, and Noam Slonim. 2022. Cluster & tune: Boost cold start performance in text classification. In ACL, pages 7639–7653. find
  21. Bashar Talafha, Hawau Olamide Toyin, Peter Sullivan, AbdelRahim A. Elmadany, Abdurrahman Juma, Amirbek Djanibekov, Chiyu Zhang, Hamad Alshehhi, Hanan Aldarmaki, Mustafa Jarrar, Nizar Habash, and Muhammad Abdul-Mageed. 2025. NADI 2025: The first multidialectal Arabic speech processing shared task. In Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks, pages 720–733, Suzhou, China. Association for Computational Linguistics. find
  22. Qwen Team. 2024. Qwen2.5: A party of foundation models. find
  23. Hongling Xu, Yice Zhang, Qianlong Wang, and Ruifeng Xu. 2025. DS 2 -ABSA: Dual-stream data synthesis with label refinement for few-shot aspect-based sentiment analysis. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15460–15478, Vienna, Austria. Association for Computational Linguistics. find
  24. Jianqing Zhu, Huang Huang, Zhihang Lin, Juhao Liang, Zhengyang Tang, Khalid Almubarak, Abdulmohsen Alharthik, Bang An, Juncai He, Xiangbo Wu, Fei Yu, Junying Chen, Zhuoheng Ma, Yuhao Du, He Zhang, Emad A. Alghamdi, Lian Zhang, Ruoyu Sun, Haizhou Li, and 2 others. 2025. Second language (arabic) acquisition of llms via progressive vocabulary expansion. Preprint, arXiv:2412.12310. ↗
  25. Henry Peng Zou and Cornelia Caragea. 2023. JointMatch: A unified approach for diverse and collaborative pseudo-labeling to semi-supervised text classification. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7290– 7301. Association for Computational Linguistics. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/almutairi-2026-barrac-adaptation-english-aspect-based/