Elamur Research⌕
☰
home / papers / 2026 / GHAD NLP at AdabEval2026: Transformer-Based Approach for Arabic Politeness and Pragmatic Category Classification

GHAD NLP at AdabEval2026: Transformer-Based Approach for Arabic Politeness and Pragmatic Category Classification

Summary

The paper describes the GHAD NLP team's system for the AdabEval 2026 shared task. The task is automatic politeness detection in Arabic social media posts and has two subtasks. In the first, a text must be assigned to one of three classes: polite, neutral, or impolite. In the second, one or more of nine pragmatic functions must be identified, such as criticism, insult, prayer, greeting, or expression of respect. The data include both Modern Standard Arabic and dialects from four platforms, including YouTube and X. For the first subtask, the authors separately fine-tuned several Arabic and multilingual transformers: MARBERTv2, CAMeLBERT, XLM-R, XLM-R Large, and others. They combined their probabilities by weighted soft voting, with the weights selected by exhaustive search on the validation set. For the second subtask, they built a multi-label model based on MARBERT. The text is normalized before it, and a separate threshold for each category is applied after it. The system ranked first in both subtasks. However, the result for the second subtask is noticeably lower, so recognizing fine-grained pragmatic functions remains difficult.

Key points

  • Dataset (Codabench): 7,928 examples in total. Subtask A has 4,895 examples for training, 693 for validation, and 1,406 for testing. The classes are highly imbalanced: about 75% neutral texts and only 8.5% impolite ones. Subtask B has 2,049 examples for training, 291 for validation, and 588 for testing. Each text can have up to four categories, and each is accompanied by keywords as lexical justification.
  • Ensemble for A: the authors searched over weights from 0 to 1 in steps of 0.1, subject to their summing to 1. This yielded 3,003 combinations, and macro-F1 on validation was computed for each. XLM-R Large received the largest weight (0.40). GATE AraBERT v1 received a weight of 0, so it was not included in the final ensemble.
  • Subtask B: diacritics and tatweel are removed from the text, and orthographic variants are unified. The threshold for each label is chosen by grid search on validation. Training settings: learning rate 2×10⁻⁵, maximum length of 192 tokens, up to 50 epochs with early stopping.
  • Results: in subtask A, macro-F1 0.89 and accuracy 0.93; in subtask B, macro-F1 0.58. The team ranked first in both subtasks.
  • Context: in prior work on 10,000 annotated ADAB examples, the best model was MARBERT (accuracy 0.9119, macro-F1 0.8582).

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 10.63317/4k9uztg6j9jr

@inproceedings{alfattni2026ghad,
  title = {{GHAD NLP at AdabEval2026: Transformer-Based Approach for Arabic Politeness and Pragmatic Category Classification}},
  author = {Alfattni, Ghada and Kurdi, Ghader},
  booktitle = {The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks},
  year = {2026},
  doi = {10.63317/4k9uztg6j9jr},
  url = {https://doi.org/10.63317/4k9uztg6j9jr}
}

References

18 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Muhammad Abdul-Mageed, AbdelRahim Elmadany, et al. 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, pages 7088–7105. ↗ our card
  2. Hend Al-Khalifa, Nadia Ghezaiel, and Maria Bounnit. 2024. Analyzing politeness in arabic tweets: A preliminary study. In Proceedings of the 7th International Conference on Natural Language and Speech Processing (ICNLSP 2024), pages 352–359. find our card
  3. Hend Al-Khalifa, Nadia Ghezaiel, Maria Bounnit, Hend Hamed Alhazmi, Noof Abdullah Alfear, Reem Fahad Alqifari, Ameera Masoud Almasoud, and Sharefah Ahmed Al-Ghamdi. 2026a. Adab: Arabic dataset for automated politeness benchmarking – a large-scale resource for computational sociopragmatics. In Proceedings of the 2026 International Conference on Language Resources and Evaluation (LREC 2026), Palma, Mallorca, Spain. ↗
  4. Hend Al-Khalifa, Nadia Ghezaiel, Maria Bounnit, Hend Hamed Alhazmi, Noof Abdullah Alfear, Reem Fahad Alqifari, Ameera Masoud Almasoud, and Sharefah Ahmed Al-Ghamdi. 2026b. Adab: Arabic dataset for automated politeness benchmarking–a large-scale resource for computational sociopragmatics. arXiv preprint arXiv:2602.13870. ↗
  5. Reem K Al Motairi and Mohammed Hadwan. 2024. Sentiment analysis methods for arabic content on social media: A systematic review. Ingénierie Des systèmes D’information, 29(1). find
  6. Amal Albalawi and WM Yafooz. 2025. Arabic offensive text classification using emojis: including emoji data in arabic natural language processing. International Journal of Electrical and Computer Engineering, 15(3):3332–3345. ↗
  7. Ali Alhazmi, Rohana Mahmud, Norisma Idris, Mohamed Elhag Mohamed Abo, and Christopher Eke. 2024. A systematic literature review of hate speech identification on arabic twitter data: research challenges and future directions. PeerJ Computer Science, 10:e1966. ↗
  8. Reem Alqifari, Hend Al-Khalifa, Nadia Ghezaiel Hammouda, Maria Bounnit, Hend Al-Hazmi, Ameera Almasoud, Sharefah AlGhamdi, and Noof Alfear. 2026. The adabeval 2026 shared task on arabic politeness detection. In Proceedings of The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks (OSACT 2026), Palma, Mallorca, Spain. Co-located with the 2026 International Conference on Language Resources and Evaluation (LREC 2026). 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. ↗ our card
  10. Cristian Danescu-Niculescu-Mizil, Moritz Sudhof, Dan Jurafsky, Jure Leskovec, and Christopher Potts. 2013. A computational approach to politeness with application to social factors. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 250–259. ↗ our card
  11. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL. ↗
  12. Go Inoue, Bashar Alhafni, Nurpeiis Baimukan, Houda Bouamor, and Nizar Habash. 2021. The interplay of variant, size, and task type in Arabic pre-trained language models. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, Kyiv, Ukraine (Online). Association for Computational Linguistics. find our card
  13. Zachary C. Lipton, Charles Elkan, and Balakrishnan Naryanaswamy. 2014. Thresholding classifiers to maximize f1 score. arXiv preprint arXiv:1402.1892. ↗
  14. Hamdy Mubarak, Ammar Rashed, Kareem Darwish, Younes Samih, and Ahmed Abdelali. 2021. Arabic offensive language on twitter: Analysis and experiments. In Proceedings of the sixth arabic natural language processing workshop, pages 126–135. ↗
  15. Priyanshu Priya, Mauajama Firdaus, and Asif Ekbal. 2024. Computational politeness in natural language processing: A survey. ACM Computing Surveys, 56(9):1–42. ↗
  16. Marina Sokolova and Guy Lapalme. 2009. A systematic analysis of performance measures for classification tasks. Information Processing & Management, 45(4):427–437. ↗
  17. Anirudh Srinivasan and Eunsol Choi. 2022. Tydip: A dataset for politeness classification in nine typologically diverse languages. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5723–5738. ↗ our card
  18. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38–45. 9. Appendix 9.1. Hyperparameter Configuration Details Parameter Max sequence length Tokeniser value 256 AutoTokeniser from pre-trained model DataCollatorWithPadding 8 16 2 16 3 × 10 −5 Up to 15 (early stopping) AdamW 0.01 0.06 Cosine FP16 if CUDA available Argmax (no threshold tuning) Data collator Batch size (train) Batch size (eval) Gradient accumulation Effective batch size Learning rate Epochs Optimiser Weight decay Warmup ratio LR scheduler Mixed precision Threshold optimisation Table 5: Configuration for subtask A. Parameter Max sequence length Tokeniser Value 192 AutoTokeniser from pre-trained model DataCollatorWithPadding 10 10 1 10 2 × 10 −5 Up to 50 (early stopping) AdamW 0.01 0.06 Linear (default Hugging-Face) FP16 if CUDA available Validation-based grid search Data collator Batch size (train) Batch size (eval) Gradient accumulation Effective batch size Learning rate Epochs Optimiser Weight decay Warmup ratio LR scheduler Mixed precision Threshold optimisation Table 6: Configuration for subtask B. find our card

↑ back to top