Elamur Research⌕
☰
home / papers / 2026 / Aatman at SemEval-2026 Task 9: Transfer Learning for Multilingual Polarization Detection

Aatman at SemEval-2026 Task 9: Transfer Learning for Multilingual Polarization Detection

Paper · 2026 · Aatman Vaidya

Summary

This system paper addresses Subtask 1 of SemEval-2026 Task 9 (POLAR), which asks whether an online post contains polarized content. It is a binary classification problem covering 22 languages. The POLAR data aggregates posts from news sites, Reddit, Bluesky, blogs and regional forums on topics such as elections, conflicts, migration and gender rights, with about 3,000–5,000 labeled items per language.

The author compares three strategies:

  • fine-tuning seven multilingual encoder models;
  • zero- and few-shot prompting of open-weight LLMs, including the safety classifier LlamaGuard;
  • sequential transfer learning, in which mDeBERTa is first trained on pooled English and Hindi harmful-language datasets and then on the POLAR training split.

Fine-tuned encoders clearly outperform prompted LLMs, and mDeBERTa is the strongest baseline. The extra transfer stage raises it to the best Macro-F1 of 0.781. The findings suggest that compact supervised encoders, enriched with related hate and abuse data, remain a cheap and strong choice for multilingual moderation. General-purpose and safety-tuned LLMs transfer poorly to implicit, context-dependent polarization.

Key points

  • Data split: Subtask 1 has 73,681 training, 3,687 development and 33,288 test instances. The 22 languages include Russian, Polish and German but not Ukrainian.
  • Encoder baselines (Macro-F1): mDeBERTa 0.759, mmBERT 0.708, XLM-RoBERTa 0.693, Glot500 0.670, GTE-multilingual-MLM 0.664, mBERT 0.660, HateBERT 0.639.
  • LLM prompting: The best zero-shot score was 0.600. The results table attributes it to Qwen-2.5-14B, while the text mentions 8B models only. Few-shot prompting used 15 polarized and 15 non-polarized examples and scored 0.694 for Qwen2.5-8B, 0.579 for Llama-3.1-8B and 0.557 for LlamaGuard-3-8B.
  • Transfer learning: The auxiliary corpora were HateCheckHIn, HateXplain, Implicit Hate, MACD, OLID, ToxiGen and Uli. They are limited to English and Hindi, the languages the author speaks. Sequential fine-tuning of mDeBERTa reached 0.781 Macro-F1, a gain of +0.022 over the baseline.

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

🎓University of TübingenTübingen · Germany75 works · 41,081 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.Multilingual & cross-lingual NLP · 828 works

Cite

Cite as Aatman Vaidya. 2026. Aatman at SemEval-2026 Task 9: Transfer Learning for Multilingual Polarization Detection. 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-1.370

@inproceedings{vaidya2026aatman,
  title = {{Aatman at SemEval-2026 Task 9: Transfer Learning for Multilingual Polarization Detection}},
  author = {Aatman Vaidya},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.370},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.370}
}

References

18 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. arXiv preprint arXiv:2111.09543.
  2. Ayyoob Imani, Peiqin Lin, Amir Hossein Kargaran, Silvia Severini, Masoud Jalili Sabet, Nora Kassner, Chunlan Ma, Helmut Schmid, André FT Martins, François Yvon, and 1 others. 2023. Glot500: Scaling multilingual corpora and language models to 500 languages. arXiv preprint arXiv:2305.12182. find
  3. Deepak Kumar, Yousef Anees AbuHashem, and Zakir Durumeric. 2024. Watch your language: Investigating content moderation with large language models. In Proceedings of the International AAAI Conference on Web and Social Media, volume 18, pages 865– 878. find
  4. Marc Marone, Orion Weller, William Fleshman, Eugene Yang, Dawn Lawrie, and Benjamin Van Durme. 2025. mmbert: A modern multilingual encoder with annealed language learning. arXiv preprint arXiv:2509.06888.
  5. Binny Mathew, Punyajoy Saha, Seid Muhie Yimam, Chris Biemann, Pawan Goyal, and Animesh Mukherjee. 2021. Hatexplain: A benchmark dataset for explainable hate speech detection. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 14867–14875. find
  6. Matteo Melis, Gabriella Lapesa, and Dennis Assenmacher. 2025. A modular taxonomy for hate speech definitions and its impact on zero-shot llm classification performance. arXiv preprint arXiv:2506.18576. find
  7. Marzieh Mozafari, Reza Farahbakhsh, and Noel Crespi. 2019. A bert-based transfer learning approach for hate speech detection in online social media. In International conference on complex networks and their applications, pages 928–940. Springer. find
  8. Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Garrido Veliz, P Sam Sahil, Yiran Zhang, Marco Antonio Stranisci, Idris Abdulmumin, "Ozge Alacam, Cengiz Acar"urk, Aisha Jabr, Saba Anwar, Abinew Ali Ayele, Elena Tutubalina, Aung Kyaw Htet, Xintong Wang, Surendrabikram Thapa, Tanmoy Chak… — SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization
  9. Gabriel Nicholas and Aliya Bhatia. 2023. Lost in translation: Large language models in non-english content analysis. arXiv preprint arXiv:2306.07377.
  10. Flor Miriam Plaza-del Arco, Debora Nozza, and Dirk Hovy. 2023. Respectful or toxic? using zero-shot learning with language models to detect hate speech. In The 7th workshop on online abuse and harms (woah), pages 60–68. find
  11. Mohamed Ibrahim Ragab, Ensaf Hussein Mohamed, and Walaa Medhat. 2025. Multilingual propaganda detection: Exploring transformer-based models mbert, xlm-roberta, and mt5. In Proceedings of the first International Workshop on Nakba Narratives as Language Resources, pages 75–82. find
  12. Rishabh Ranjan, Likhith Ayinala, Mayank Vatsa, and Richa Singh. 2025. Multimodal zero-shot framework for deepfake hate speech detection in low-resource languages. arXiv preprint arXiv:2506.08372. — [arXiv:2506.08372](https://arxiv.org/abs/2506.08372) find
  13. Punyajoy Saha, Divyanshu Sheth, Kushal Kedia, Binny Mathew, and Animesh Mukherjee. 2022. Rationale-guided few-shot classification to detect abusive language. arXiv preprint arXiv:2211.17046. — [arXiv:2211.17046](https://arxiv.org/abs/2211.17046) find
  14. Orion Weller, Kathryn Ricci, Marc Marone, Antoine Chaffin, Dawn Lawrie, and Benjamin Van Durme. 2025. Seq vs seq: An open suite of paired encoders and decoders. arXiv preprint arXiv:2507.11412. — [arXiv:2507.11412](https://arxiv.org/abs/2507.11412) find
  15. Zecharias Zelalem and Peter Guest. 2021. Why facebook keeps failing in ethiopia. Rest of World. find
  16. Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, and 1 others. 2024. mgte: Generalized long-context text representation and reranking models for multilingual text retrieval. arXiv preprint arXiv:2407.19669.
  17. A 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… — POLAR: A Benchmark for Multilingual, Multicultural, and Multi-Event Online Polarization
  18. Language Specific Performance Zero Shot Encoder-only Language Few shot Safety xlm-roberta-base mbert-base mdeberta-base mgte-multi lingual-mlm-base glot-500 -base mmbert -base hatebert qwen2.5 -14B llama3.1 -8B qwen2.5 -8B llama guard3 -8B Amharic 0.5356 0.4196 0.6927 0.4448 0.482 0.6082 0.4613 0.48… find

Updated

↑ back to top

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