Elamur Research⌕
☰
home / papers / 2026 / Stochastic Gradient Descenders at SemEval-2026 Task 9: Few-Shot LLM Prompting for Polarization Type Classification

Stochastic Gradient Descenders at SemEval-2026 Task 9: Few-Shot LLM Prompting for Polarization Type Classification

Paper · 2026 · Huynh Phu and Dang Thin

Summary

The paper describes a system for SemEval-2026 Task 9 (POLAR), Subtask 2, English track. The task is multi-label classification of social media posts into five polarization types: political/ideological, racial/ethnic, religious, gender/sexual and other. The main difficulty is strong label imbalance. Some categories are very rare, and many posts carry no label at all, which hurts macro-averaged F1. The authors compared three levels of supervision:

  • mDeBERTa-v3-base with domain-adaptive pre-training on the multilingual POLAR corpus and Asymmetric Loss;
  • Qwen2.5-32B-Instruct adapted with LoRA;
  • training-free few-shot prompting of Llama-3.3-70B-Instruct.

Both fine-tuned families failed to find any correct gender/sexual instances on the development set. The prompted Llama model produced non-zero F1 for every category and therefore had the best macro score, so it became the submission. It placed second among 29 English entries. The work shows that under macro-F1, covering rare labels matters more than peak accuracy on frequent ones. Large instruction-tuned models handle this well without parameter updates.

Key points

  • Data: the English split has 3,222 training, 160 development and 1,452 test instances. Gender/sexual has only 72 training positives (2.2%) and 3 development positives.
  • Prompt design: each prompt asks the model to act as a cautious classifier, gives a definition for each category and includes 15 demonstrations. The demonstrations are one positive per type plus ten non-polarized texts, shuffled. Temperature is 0.1 and the output is JSON.
  • Development macro F1:
    • Llama few-shot: 0.378
    • Qwen-LoRA (detailed prompt): 0.372
    • Best mDeBERTa run: 0.354
  • Official test results: macro F1 of 0.5157, ranked 2nd of 29, against 0.5322 for the winner and 0.3333 for the baseline. Gender/sexual reached F1 0.531, while "Other" reached only 0.225.
  • Error analysis:
    • 68.3% of test items received an exactly correct label set.
    • 188 of 919 non-polarized texts were wrongly tagged political.
    • 51.7% of multi-label items were only partially predicted.
    • 81.7% of "Other" predictions were false positives.

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

🎓Vietnam National University Ho Chi Minh CityHo Chi Minh City · Vietnam21 works · 14 citations🎓University of Information TechnologyHo Chi Minh City · Vietnam13 works · 7 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.Language models & pretraining · 3,320 works

Cite

Cite as Huynh Phu and Dang Thin. 2026. Stochastic Gradient Descenders at SemEval-2026 Task 9: Few-Shot LLM Prompting for Polarization Type Classification. 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-1.129

@inproceedings{phu2026stochastic,
  title = {{Stochastic Gradient Descenders at SemEval-2026 Task 9: Few-Shot LLM Prompting for Polarization Type Classification}},
  author = {Huynh Phu and Dang Thin},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.129},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.129}
}

References

13 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Francesco Barbieri, Jose Camacho-Collados, Luis Espinosa-Anke, and Leonardo Neves. 2020. TweetEval: Unified benchmark and comparative evaluation for tweet classification. Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1644–1650.
  2. Emanuel Ben-Baruch, Tal Ridnik, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. 2021. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 82–91. find
  3. Tom Brown and 1 others. 2020. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33:1877–1901.
  4. Aaron Grattafiori and 1 others. 2024. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  5. Suchin Gururangan, Ana Marasović, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. 2020. Don’t stop pretraining: Adapt language models to domains and tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8342–8360. Associati… find
  6. Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2023. DeBERTaV3: Improving DeBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing. In Proceedings of the 11th International Conference on Learning Representations (ICLR).
  7. 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 Proceedings of the 10th International Conference on Learning Representations (ICLR).
  8. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2980– 2988.
  9. System message: You are a neutral and objective classifier for detecting polarization. Your task is to determine whether a given text clearly incites division, intolerance, or conflict between groups. Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Garrido Veliz, P Sam Sahil, Yiran Zha… — SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization
  10. 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… — POLAR: A Benchmark for Multilingual, Multicultural, and Multi-Event Online Polarization
  11. Jason Wei and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824–24837.
  12. An Yang and 1 others. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. find
  13. A Classification guidelines – Assign 1 only if the text clearly promotes division, intolerance, or conflict. – Neutral reporting, questions, or balanced discussions should receive 0 for all categories. – When the text is ambiguous or merely mentions a sensitive topic without a polarizing stance, def… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/phu-2026-stochastic-gradient-descenders-semeval-2026/