Elamur Research⌕
☰
home / papers / 2026 / PolAR Bears at SemEval-2026 Task 9: Parameter-Efficient Fine-Tuning and Cross-Lingual Augmentation for Multilingual Polarization Detection

PolAR Bears at SemEval-2026 Task 9: Parameter-Efficient Fine-Tuning and Cross-Lingual Augmentation for Multilingual Polarization Detection

Paper · 2026 · Vinay Ulli and Jyoti Kumari

Summary

The authors address online polarization detection in four under-resourced Indian languages (Hindi, Bengali, Telugu and Odia) within SemEval-2026 Task 9. The task has three parts: binary polarization detection, multi-label classification of target groups, and multi-label identification of rhetorical manifestations. Labelled data is scarce, so they machine-translated every training item with IndicTrans2 into the other three languages and kept the original labels, roughly quadrupling the training set. One framework handles all three subtasks. Qwen3-4B-Instruct receives ChatML prompts that list the relevant categories. Instead of generating answers, the model's final-layer hidden states are mean-pooled and passed to a linear classifier, and the model is adapted cheaply with QLoRA. Binary detection worked well, but the multi-label subtasks performed strongly only for Hindi. The authors attribute this to three causes: Qwen's Hindi-heavy pre-training, translation flattening rhetorical subtlety, and a fixed 0.5 decision threshold. They ran no baselines, so the augmentation's own contribution remains unmeasured. The work offers an inexpensive single-GPU recipe and exposes its limits for fine-grained tasks in lower-resource languages.

Key points

  • Data augmentation: IndicTrans2 generated 12 translation directions among the four languages, growing the training corpus from about 11,350 to 43,244 samples. Translation quality was checked only automatically, via bge-m3 cosine similarity, with no human review.
  • Model setup: Qwen3-4B-Instruct was fine-tuned with QLoRA (4-bit NF4, LoRA r=16, α=32, dropout 0.05 on all attention and MLP projections).
    • Only about 33M parameters (0.81%) were trained, on a single A100 40GB GPU.
    • Training used AdamW with a learning rate of 3e-4 and an effective batch size of 32.
  • Subtask 1 (binary detection): average Macro F1 of 0.813, ranging from 0.7589 (Odia) to a best of 0.8668 (Telugu).
  • Subtasks 2 and 3 (multi-label): average Macro F1 was only 0.354 and 0.303. Hindi was the exception, at 0.7008 and 0.7248, while Telugu dropped to 0.1299 in Subtask 2 and Odia to 0.1196 in Subtask 3.
  • Proposed future work: class-aware threshold tuning and curation of native (non-translated) datasets. Comparisons with mBERT, IndicBERT or an unaugmented baseline are still missing.

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

🎓Banaras Hindu UniversityVaranasi · India6 works · 260 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 Vinay Ulli and Jyoti Kumari. 2026. PolAR Bears at SemEval-2026 Task 9: Parameter-Efficient Fine-Tuning and Cross-Lingual Augmentation 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-1.279

@inproceedings{ulli2026polar,
  title = {{PolAR Bears at SemEval-2026 Task 9: Parameter-Efficient Fine-Tuning and Cross-Lingual Augmentation for Multilingual Polarization Detection}},
  author = {Vinay Ulli and Jyoti Kumari},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.279},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.279}
}

References

14 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216, 4(5).
  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. Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. 2017. Automated hate speech detection and the problem of offensive language. In Proceedings of the International AAAI Conference on Web and Social Media, volume 11, pages 512– 515.
  4. Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2024. QLoRA: Efficient finetuning of quantized llms. In Advances in Neural Information Processing Systems, volume 36.
  5. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics, pages 4171–4186.
  6. Jay Gala, Diptesh Kanojia, Sudhanshu Bhatia, Hitesh Doshi, Sumanth Doddapaneni, and 1 others. 2023. IndicTrans2: Towards high-quality and accessible machine translation models for all 22 scheduled indian languages. Transactions of the Association for Computational Linguistics, 12.
  7. Edward J Hu, Yelong Shen, Phil 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.
  8. Divyanshu Kakwani, Anoop Kunchukuttan, Satish Golla, Gokul N.C., Pushpak Bhattacharyya, Mitesh M. Khapra, and Pratyush Kumar. 2020. IndicNLPSuite: Monolingual corpora, evaluation benchmarks and pre-trained multilingual language models for indian languages. In Findings of the Association for Computat…
  9. Julia Mendelsohn, Vinodkumar Prabhakaran, and Dan Jurafsky. 2021. A framework for the computational linguistic analysis of dehumanization. In Proceedings of Frontiers in NLP. find
  10. Niklas Muennighoff. 2022. Sgpt: Gpt sentence embeddings for semantic search. arXiv preprint arXiv:2202.08904.
  11. 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
  12. 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
  13. Zeerak Waseem and Dirk Hovy. 2016. Hateful symbols or hateful people? predictive features for hate speech detection on Twitter. In Proceedings of the NAACL Student Research Workshop, pages 88–93.
  14. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ulli-2026-polar-bears-semeval-2026-task/