Elamur Research⌕
☰
home / papers / 2026 / PolarizedTeam at SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization

PolarizedTeam at SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization

Paper · 2026 · Maria Nestor, Maroan Al Shrafat, Ioana Pește et al.

Summary

The paper describes PolarizedTeam's entry to SemEval-2026 Task 9 (POLAR), which asks systems to find polarized online content in 22 languages. Subtask 1 is a binary polarized/non-polarized decision. Subtask 2 assigns any of five polarization types: political, racial/ethnic, religious, gender/sexual and other. The texts are short, noisy and unevenly labelled, and some languages have almost no positives for certain categories.

The authors trained a separate model for each language. For Subtask 1 they compared mBERT, XLM-RoBERTa-Large and a hybrid that feeds mBERT token embeddings into a unidirectional LSTM. The hybrid scored best, with an average F1 of 0.76. For Subtask 2 they built up their system in three steps:

  • an mBERT baseline using mean pooling and class-weighted BCE;
  • XLM-RoBERTa-Base combining [CLS] and mean-pooled vectors, trained with Focal Loss;
  • a final model in which every label is handled by a dedicated attention head paired with a separate classifier.

Macro-F1 rose with each step, reaching 0.507 for the final model. The work shows that label-aware modelling and imbalance handling help with rare categories and low-resource languages. Very short texts and the lack of cross-lingual transfer remain weak points.

Key points

  • Data: The POLAR dataset covers 22 languages, including Amharic, Hausa, Khmer, Odia, Russian and Polish, with roughly 1,700–7,000 snippets per language. For Subtask 2, all-zero (non-polarized) examples were downsampled to half the number of polarized ones.
  • Subtask 1 averages (F1 / accuracy):
    • mBERT: 0.61 / 0.70
    • XLM-RoBERTa-Large: 0.67 / 0.71
    • mBERT+LSTM: 0.76 / 0.80
    • The best per-language hybrid scores were on Chinese (0.897) and Nepali (0.880).
  • Subtask 2 average macro-F1:
    • mBERT baseline: 0.453
    • XLM-RoBERTa-Base with Focal Loss: 0.495
    • Label-wise attention: 0.507
    • The final model's gains were noted for skewed languages such as Amharic, Hausa and Telugu.
  • Training setup: PyTorch with HuggingFace Transformers; AdamW at a learning rate of 1e-5 with linear decay; batch size 32–64; at most 5 epochs with early stopping; maximum length of 256 tokens; no lemmatization.
  • Code: The code is released at github.com/FreaksMind/POLAR-SEMEVAL. The authors list remaining limitations as rare labels, very short texts, and per-language training that blocks cross-lingual transfer.

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

🎓Alexandru Ioan Cuza UniversityIași · Romania8 works · 402 citations🔬Romanian AcademyBucharest · Romania8 works · 841 citations🔬Academia Română, Filiala IașiIași · Romania3 works · 3 citations🔬Academia Oamenilor de Știință din RomâniaBucharest · Romania3 works · 1 citation

 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 Maria Nestor, Maroan Al Shrafat, Ioana Pește, Daniela Gifu and Diana Trandabăț. 2026. PolarizedTeam at SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization. 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-1.180

@inproceedings{nestor2026polarizedteam,
  title = {{PolarizedTeam at SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization}},
  author = {Maria Nestor, Maroan {Al Shrafat}, Ioana Pește, Daniela Gifu and Diana Trandabăț},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.180},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.180}
}

References

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

  1. 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. Preprint, arXiv:1911.02116. find
  2. Maksym Del and Mark Fishel. 2022. Similarity of sentence representations in multilingual lms: Resolving conflicting literature and case study of baltic languages. Preprint, arXiv:2109.01207. — [arXiv:2109.01207](https://arxiv.org/abs/2109.01207) find
  3. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. Preprint, arXiv:1810.04805.
  4. Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2023. DeBERTav3: Improving deBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing. In The Eleventh International Conference on Learning Representations.
  5. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural Computation, 9(8):1735– 1780. find
  6. Lo, Soda Marem and Stranisci, Marco A. and Cignarella, Alessandra Teresa and Frenda, Simona and Basile, Valerio and Bosco, Cristina and Jezek, Elisabetta and Patti, Viviana. 2025. Subjectivity in stereotypes against migrants in Italian : an experimental annotation procedure. In Proceedings of the El… find
  7. Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Garrido Veliz, P Sam Sahil, Yiran Zhang, Marco Antonio Stranisci, Idris Abdulmumin, Özge Alacam, Cengiz Acarürk, Aisha Jabr, Saba Anwar, Abinew Ali Ayele, Elena Tutubalina, Aung Kyaw Htet, Xintong Wang, Surendrabikram Thapa, Tanmoy Chakra… find
  8. Average F1 Score 0.45 0.49 0.50 *Table 6: Average F1 scores across all languages.* and Multievent Online Polarization. In Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval-2026), San Diego, CA, USA. Association for Computational Linguistics. find
  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… — POLAR: A Benchmark for Multilingual, Multicultural, and Multi-Event Online Polarization
  10. Dana Warmsley. 2017. On the detection of hate speech, hate speakers and polarized groups in online social media. find
  11. Wentao Xu, Wenlu Fan, Shiqian Lu, Tenghao Li, and Bin Wang. 2025. Polarized patterns of language toxicity and sentiment of debunking posts on social media. Preprint, arXiv:2501.06274. — [arXiv:2501.06274](https://arxiv.org/abs/2501.06274) find

Updated

↑ back to top

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