Elamur Research⌕
☰
home / papers / 2026 / Phatthachdau at SemEval-2026 Task 9: A Multi-Stage Augment-Judge-Train Pipeline for Multilingual Online Polarization Detection

Phatthachdau at SemEval-2026 Task 9: A Multi-Stage Augment-Judge-Train Pipeline for Multilingual Online Polarization Detection

Paper · 2026 · Phan Phat

Summary

This system paper tackles binary polarization detection (Subtask 1, POLARDETECT) in SemEval-2026 Task 9, which uses the multilingual POLAR dataset. The authors work on two languages with opposite difficulties. Hausa is low-resource and heavily skewed, with only about 11% of examples polarized. English is high-resource, but much of its hostility is implicit. They propose a three-stage Augment-Judge-Train (AJT) pipeline:

  • Augment: Gemini 2.0 Flash, prompted with a taxonomy of polarization types (political, ethnic, religious, gender), creates synthetic data. For Hausa this means colloquial polarized examples. For English it means hard negatives and texts that imitate the partisan style of the seed data.
  • Judge: Gemini 2.5 Pro scores each synthetic sample, and weak ones are discarded.
  • Train: A weighted soft-voting ensemble combines fine-tuned encoders with LoRA-adapted Llama models, each trained on different mixes of original and synthetic data.

The system ranked first for Hausa and in the top 10 for English. The results suggest that culture-aware, quality-filtered LLM augmentation is a practical fix for severe class imbalance in under-resourced languages.

Key points

  • Results: Hausa reached Macro-F1 0.8336, ranking 1st (runner-up: 0.8324). English reached 0.8092, ranking 10th (winner: 0.8252).
  • Quality filter: The judge (Gemini 2.5 Pro, temperature 0.30) rated samples on a 0–1 scale and kept those scoring ≥ 0.70. Generation used temperature 0.75 for English and 0.85 for Hausa.
  • Augmentation volume: Hausa received 2,481 synthetic polarized samples, added on top of 392 original polarized ones. English received 1,827 non-polarized and 1,096 polarized samples.
  • Models: Hausa used xlm-roberta-base-finetuned-hausa. English used RoBERTa-base plus Meta-Llama-3.1-8B with LoRA (r=16, α=32, 4-bit NF4). Everything was trained on a single 16GB P100 GPU.
  • Ensemble: The ensemble has fixed weights, for example a 0.40 top weight on one English LLM-LoRA model, and a 0.5 decision threshold. Code and augmented data are on GitHub (kelvin2250/SemEval2026_Task9).

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, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksSocial Sciences · 1,772 worksSociology and Political Science · 577 works

🧭 Topics 1Research subjects and languages the work is about.Multilingual & cross-lingual NLP · 828 works

Cite

Cite as Phan Phat. 2026. Phatthachdau at SemEval-2026 Task 9: A Multi-Stage Augment-Judge-Train Pipeline for Multilingual Online 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-1.208

@inproceedings{phat2026phatthachdau,
  title = {{Phatthachdau at SemEval-2026 Task 9: A Multi-Stage Augment-Judge-Train Pipeline for Multilingual Online Polarization Detection}},
  author = {Phan Phat},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.208},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.208}
}

References

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

  1. Build Up. 2025. The polarization footprint in europe. Technical report, Build Up. find
  2. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association fo… find
  3. Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. QLoRA: Efficient finetuning of quantized LLMs. In Advances in Neural Information Processing Systems, volume 36.
  4. Thomas G Dietterich. 2000. Ensemble methods in machine learning. Multiple classifier systems, pages 1–15. find
  5. Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, and 1 others. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805.
  6. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.
  7. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
  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. 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 Ky… — [arXiv:2505.20624](https://arxiv.org/abs/2505.20624) find
  10. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Hao, Zhanghao Wu, Sylvain Ba, Eugene Zhuang, Zihao Lin, Zhuohan Li, Eric Xing, and 1 others. 2023. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Advances in Neural Information Processing Systems, volume 36.

Updated

↑ back to top

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