Elamur Research⌕
☰
home / papers / 2026 / DigiS-FBK at SemEval-2026 Task 9: Multi-task Learning for Multilingual and Cross-cultural Polarization Classification

DigiS-FBK at SemEval-2026 Task 9: Multi-task Learning for Multilingual and Cross-cultural Polarization Classification

Paper · 2026 · Veronica Orsanigo, Alan Ramponi and Elisa Leonardelli

Summary

The paper describes the DigiS-FBK team's systems for SemEval-2026 Task 9 (POLAR), which treats online polarization as a text classification problem across many languages, cultures and events. POLAR has three linked subtasks: deciding whether a text is polarized (S1), labelling its target type (S2, multi-label), and labelling the rhetorical way it is expressed (S3, multi-label). The subtasks depend on each other; for example, a non-polarized text cannot have a type or manifestation. The authors therefore train them jointly with multi-task learning. Each model pairs a shared encoder-only transformer with a separate linear head per subtask, and they compare these joint models against single-task baselines. They use XLM-RoBERTa for most languages and monolingual encoders for English, Italian, Spanish and Polish. Multi-task training beat the single-task setups on every subtask, although scores varied considerably between languages. The work matters because it shows that compact encoder models can handle several polarization subtasks at once. This makes the approach a practical choice when compute or time is too limited for LLMs.

Key points

  • Data: The POLAR data covers 22 languages from seven language families for S1 and S2, and 18 languages for S3 (Burmese, Italian, Polish and Russian have no S3 labels). Each language has roughly 1,700–7,000 training instances. In total there are about 73.7k original train, 3.7k dev and 33.3k test instances, and the authors re-split train/dev into 68,513 and 8,853 instances.
  • Setup: Models were built with MaChAmp v0.4.2 using default hyperparameters. The encoders were XLM-RoBERTa (multilingual), RoBERTa (English), AlBERTo (Italian), BETO (Spanish) and Polbert (Polish). The tested configurations were [S1], [S2] and [S3] alone, plus the joint combinations [S1, S2], [S1, S3] and [S1, S2, S3].
  • Best configurations: Training on all three subtasks gave the best results for S1 and S2. For S3, the pair [S1, S3] performed best.
  • Monolingual vs. multilingual: Monolingual encoders worked well for English and Spanish. For Italian and Polish, they did worse on S2 than XLM-RoBERTa, and the Italian model never predicted the Political label on the test set.
  • Approaches that did not help:
    • Sequential fine-tuning starting from S1 gave no gains on dev.
    • Forcing S2 and S3 labels to zero whenever S1 was predicted 0 left F1 unchanged.
    • Most remaining errors are false negatives, which the authors attribute to strong class imbalance in S2 and S3.

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 TrentoTrento · Italy85 works · 15,524 citations🔬Fondazione Bruno KesslerTrento · Italy35 works · 43,945 citations

 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 Veronica Orsanigo, Alan Ramponi and Elisa Leonardelli. 2026. DigiS-FBK at SemEval-2026 Task 9: Multi-task Learning for Multilingual and Cross-cultural Polarization 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-1.357

@inproceedings{orsanigo2026digis,
  title = {{DigiS-FBK at SemEval-2026 Task 9: Multi-task Learning for Multilingual and Cross-cultural Polarization Classification}},
  author = {Veronica Orsanigo, Alan Ramponi and Elisa Leonardelli},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.357},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.357}
}

References

13 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Rie Kubota Ando and Tong Zhang. 2005. A framework for learning predictive structures from multiple tasks and unlabeled data. Journal of Machine Learning Research, 6:1817–1853. find
  2. Galen Andrew and Jianfeng Gao. 2007. Scalable training of L1-regularized log-linear models. In Proceedings of the 24th International Conference on Machine Learning, pages 33–40. find
  3. José Cañete, Gabriel Chaperon, Rodrigo Fuentes, Jou-Hui Ho, Hojin Kang, and Jorge Pérez. 2023. Spanish pre-trained BERT model and evaluation data. arXiv preprint arXiv:2308.02976. find
  4. Rich Caruana. 1997. Multitask learning. Machine learning, 28(1):41–75. find
  5. 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
  6. Dariusz Kłeczek. 2020. Polbert: Attacking polish NLP tasks with transformers. In Proceedings of the Pol-Eval 2020 Workshop, Warsaw, Poland. Institute of Computer Science, Polish Academy of Sciences. find
  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. Zikun Liu, Chen Luo, and Jia Lu. 2024. Hate speech in the internet context: unpacking the roles of internet penetration, online legal regulation, and online opinion polarization from a transnational perspective. Information Development, 40(4):533–549. find
  9. Raquel Martínez-España, Julio Fernández-Pedauye, José Giner-Pérez de Lucía, Jose Miguel Rojo-Martínez, Kaoutar Bakdid-Albane, and Juan José García-Escribano. 2024. Methodology for measuring individual affective polarization using sentiment analysis in social networks. IEEE Access, 12:102035– 102049. find
  10. Saif Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and Colin Cherry. 2016. find
  11. SemEval-2016 task 6: Detecting stance in tweets. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016), pages 31– 41, San Diego, California. Association for Computational Linguistics.
  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, Elena Tutubalina, Aung Kyaw Htet, Xintong Wang, Surendrabikram Thapa, Tanmoy Chakr… — SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization
  13. Michela Del Vicario, Walter Quattrociocchi, Antonio Scala, and Fabiana Zollo. 2019. Polarization and fake news: Early warning of potential misinformation targets. ACM Transactions on the Web (TWEB), 13(2):1–22. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/orsanigo-2026-digis-fbk-semeval-2026-task/