Elamur Research⌕
☰
home / papers / 2026 / REGLAT at SemEval-2026 Task 9: Enhancing Arabic Online Polarization Detection Using AraBERT and Synonym Replacement Augmentation

REGLAT at SemEval-2026 Task 9: Enhancing Arabic Online Polarization Detection Using AraBERT and Synonym Replacement Augmentation

Paper · 2026 · Ahmed M. Fetouh, Rahmath Mohammed, Omer Dawood et al.

Summary

The paper describes REGLAT's entry to SemEval-2026 Task 9, Subtask 1. The task is binary detection of polarized posts in Arabic social media, a setting made hard by dialect diversity, rich morphology and limited labeled data.

The team used the Arabic part of the POLAR benchmark and fine-tuned a single model, AraBERTv2 (aubmindlab/bert-base-arabertv2). Before training, they cleaned every post with the model's own ArabertPreprocessor. They then enlarged the training data with synonym replacement drawn from a small hand-built dictionary of Arabic sentiment words, which roughly doubled its size.

On the blind test set the system reached macro F1 0.831 and accuracy 0.833. This placed it 11th of 45 teams, 0.018 below the winner. The development score was lower, at macro F1 0.790.

The authors argue that one well-tuned transformer with cheap lexical augmentation is a strong, reproducible baseline for low-resource Arabic tasks. They do not, however, report an ablation that isolates what the augmentation contributes.

Key points

  • Data: the Arabic training split of POLAR has about 3,380 posts. Synonym augmentation expanded it to 6,032 examples, with labels kept unchanged.
  • Augmentation recipe: each post received 1–3 augmented copies, and each copy had 1–3 words swapped for dictionary synonyms. Copies identical to the original were discarded.
  • Training setup: AraBERTv2, maximum length 256 tokens, 6 epochs, learning rate 2e-5, batch size 32 for training and 16 for evaluation. The checkpoint was chosen by macro F1 using the Hugging Face Trainer.
  • Validation vs. development scores: per-epoch validation macro F1 peaked near 0.939, far above the 0.790 on the official dev set. The paper does not explain this gap.
  • Error analysis:
    • Dialectal spelling, sarcasm and implicit polarization caused many mistakes.
    • Emotionally intense but neutral posts produced false positives.
    • Subtle polarization produced false negatives.
    • Augmentation sometimes introduced semantic drift.

Interesting ideas

  • I1. Synonym replacement from a small curated dictionary, discarding unchanged copies, is a cheap way to multiply scarce labeled data without a generator model.
  • I2. Running the exact preprocessor that the pretrained model saw during pretraining keeps inputs aligned with the model's assumptions.
  • I3. The error analysis separates emotional intensity from polarization, since strong feeling alone triggered false positives.
  • I4. The authors admit augmentation can cause semantic drift that leaves noisy, mislabeled copies.
  • I5. The large gap between in-training validation and the official dev or blind scores shows why held-out sets must be kept independent of augmented data.

For Mova

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

  • I1 (learn): silver training sentences could be multiplied by swapping lemmas within a VerbNet class or WordNet synset, keeping only copies whose UD tree and frame keep the same shape.
  • I2 (tokens): normalisation in crate en should match the conventions of the UD treebanks the parsers were trained on, so tokenizer output never drifts from what the tagger expects.
  • I3 (frames): prag could keep affect intensity and stance or attitude toward a target as distinct annotations, so irony and speech-act judgments do not mistake loud language for a position.
  • I4 (learn): every augmented example could face a negative-control check, such as witness parsers or an LLM teacher confirming the meaning or label survived, before it enters training.
  • I5 (learn): evaluation splits should be fixed before any augmentation, so that augmented paraphrases of test items cannot inflate the scores.

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

🎓Prince Sattam Bin Abdulaziz UniversityAl Kharj · Saudi Arabia17 works · 489 citations🎓Benha UniversityBanhā · Egypt14 works · 89 citations🎓ElSewedy University of TechnologyCairo · Egypt5 works · 5 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 Ahmed M. Fetouh, Rahmath Mohammed, Omer Dawood, Mariam Labib, Nsrin Ashraf and Hamada Nayel. 2026. REGLAT at SemEval-2026 Task 9: Enhancing Arabic Online Polarization Detection Using AraBERT and Synonym Replacement Augmentation. Proceedings of the 20th International Workshop on Semantic Evaluation (2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.semeval-1.226

@inproceedings{fetouh2026reglat,
  title = {{REGLAT at SemEval-2026 Task 9: Enhancing Arabic Online Polarization Detection Using AraBERT and Synonym Replacement Augmentation}},
  author = {Ahmed M. Fetouh, Rahmath Mohammed, Omer Dawood, Mariam Labib, Nsrin Ashraf and Hamada Nayel},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.226},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.226}
}

References

10 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ahmed H. AbuElAtta, Mahmoud Sobhy, Ahmed A. El-Sawy, and Hamada Nayel. 2023. Arabic regional dialect identification (ardi) using pair of continuous bag-of-words and data augmentation. International Journal of Advanced Computer Science and Applications, 14(11). find
  2. Reema Khaled AlRowais and Duaa H. Alsaeed. 2023. Arabic stance detection of covid-19 vaccination using transformer-based approaches: A comparison study. Arab Gulf Journal of Scientific Research, 42(1):89– 107. ↗
  3. 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, Marseille, France. European Language Resources Association. ↗
  4. Javier Borge-Holthoefer, Walid Magdy, Kareem Darwish, and Ingmar Weber. 2015. Content and network dynamics behind egyptian political polarization on twitter. In Proceedings of the 18th ACM Conference on Computer Supported Cooperative Work & Social Computing (CSCW), pages 700–711. ACM. find
  5. Amey Hengle, Aditya Kshirsagar, Tanmoy Chakrabarty, and Kathleen McKeown. 2021. Combining context-free and contextualized representations for arabic *Table 3: Performance of our system on Arabic data for Subtask 1 (Polarization Detection). All values are reported for the Arabic language only.* Set Accuracy Precision Recall F1 (Binary) F1 (Macro) F1 (Micro) Development Blind Test 0.7929 0.8330 0.7703 0.8230 0.7600 0.7988 0.7651 0.8107 0.7900 0.831 0.7929 0.833 sarcasm detection and sentiment identification. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 207–214, Online. Association for Computational Linguistics. find
  6. Lana Kazkaz. 2022. Role of new opinion leaders on social media in political and religious polarization (jordanian case study). Lebanese Science Journal, 21(2):233–251. find
  7. Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Garrido Veliz, P Sam Sahil, Yiran Zhang, Marco Antonio Stranisci, Idris Abdulmumin, Özge Alacam, Cengiz Acarürk, Aisha Jabr, Saba Anwar, Abinew Ali Ayele, Elena Tutubalina, Aung Kyaw Htet, Xintong Wang, Surendrabikram Thapa, Tanmoy Chakraborty, and 15 others. 2026a. SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization. In Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval-2026), San Diego, CA, USA. Association for Computational Linguistics.
  8. Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Garrido Veliz, P Sam Sahil, Yiran Zhang, Marco Antonio Stranisci, Idris Abdulmumin, Özge Alacam, Cengiz Acartürk, Aisha Jabr, Saba Anwar, Abinew Ali Ayele, Simona Frenda, Alessandra Teresa Cignarella, Elena Tutubalina, Oleg Rogov, Aung Kyaw Htet, and 24 others. 2026b. Polar: A benchmark for multilingual, multicultural, and multi-event online polarization. Preprint, arXiv:2505.20624. ↗
  9. Ingmar Weber, Venkata Rama Kiran Garimella, and Alaa Batayneh. 2013. Secular vs. islamist polarization in egypt on twitter. In Proceedings of the 2013 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 290–297. ACM. find
  10. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, and 3 others. 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, Online. Association for Computational Linguistics.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/fetouh-2026-reglat-semeval-2026-task-9/