Elamur Research⌕
☰
home / papers / 2026 / DeepSemantics at SemEval-2026 Task 9: Label-Wise Optimization with Adaptive Focal Loss for Polarization Manifestation Identification

DeepSemantics at SemEval-2026 Task 9: Label-Wise Optimization with Adaptive Focal Loss for Polarization Manifestation Identification

Paper · 2026 · Eliasse Tiao, Josue R. Edou and Mahugnon A. L. Gohouede

Summary

The paper describes the DeepSemantics entry for Subtask 3 of SemEval-2026 Task 9 (POLAR 2026). In this subtask, systems must flag social media posts for six polarization categories: Dehumanization, Vilification and Stereotype, along with Invalidation, Extreme Language and Lack of Empathy. The authors cover English and Hausa. Both datasets are very sparse and skewed, and Hausa is also a low-resource language.

Instead of predicting all labels jointly, they train a separate binary classifier per label (One-vs-Rest). The English models use RoBERTa-base with Adaptive Focal Loss. The Hausa models use Afro-XLM-R-small, because focal loss made training unstable there. Hausa training instead uses weighted binary cross-entropy plus capped oversampling of rare positives with EDA augmentation. Each label also gets its own tuned decision threshold.

On English validation data, One-vs-Rest clearly beat a joint multi-label baseline. On the official test sets, the system placed 5th in Hausa and 14th in English. The work shows that label-wise modeling and imbalance-aware training help with rare categories in a low-resource language.

Key points

  • Test results: official macro-F1 was 0.464 for English (14th place) and 0.192 for Hausa (5th place). The strongest English labels were Vilification (0.625) and Extreme Language (0.564).
  • One-vs-Rest vs. joint model: on English validation data, One-vs-Rest reached 0.50 macro-F1 against 0.33 for the joint multi-label baseline. Hausa used One-vs-Rest only and scored 0.20 on validation.
  • Oversampling rule: oversampling activates only when negatives outnumber positives by more than 5:1. Positives are then duplicated so the ratio stays at or below 5:1, followed by EDA (synonym replacement, random insertion).
  • Threshold tuning: each label's threshold comes from a grid search over [0.01, 0.99] and is averaged across stratified 5-fold splits.
  • Label dependence and limitations:
    • Pairwise χ² tests found dependence between almost all English label pairs, but only some Hausa pairs.
    • In Hausa, 92.5% of polarized texts carry a single label.
    • All results come from a single run (seed 42) on free Google Colab.

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

🎓African Institute for Mathematical SciencesCape Town · South Africa8 works · 797 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

Cite

Cite as Eliasse Tiao, Josue R. Edou and Mahugnon A. L. Gohouede. 2026. DeepSemantics at SemEval-2026 Task 9: Label-Wise Optimization with Adaptive Focal Loss for Polarization Manifestation Identification. 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-1.210

@inproceedings{tiao2026deepsemantics,
  title = {{DeepSemantics at SemEval-2026 Task 9: Label-Wise Optimization with Adaptive Focal Loss for Polarization Manifestation Identification}},
  author = {Eliasse Tiao, Josue R. Edou and Mahugnon A. L. Gohouede},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.210},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.210}
}

References

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

  1. Jesujoba O. Alabi, David Ifeoluwa Adelani, Marius Mosbach, and Dietrich Klakow. 2022. Adapting pre-trained language models to African languages via multilingual adaptive fine-tuning. In Proceedings of the 29th International Conference on Computational Linguistics, pages 4336–4349, Gyeongju, Republic…
  2. Pablo Barberá. 2020. Social media, echo chambers, and political polarization. Social Media + Society, 6(3):1–12. find
  3. 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
  4. Alexis Conneau. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of ACL, pages 8440–8451. find
  5. Jesse Read, Bernhard Pfahringer, Geoff Holmes, and Eibe Frank. 2011. Classifier chains for multi-label classification. Machine Learning, 85(3):333–359. find
  6. Michael Conover, Jacob Ratkiewicz, Matthew Francisco, Bruno Gonçalves, Filippo Menczer, and Alessandro Flammini. 2011. Political polarization on twitter. find
  7. Grigorios Tsoumakas and Ioannis Manousos Katakis. 2007. Multi-label classification: An overview. Int. J. Data Warehous. Min., 3:1–13. find
  8. Jacob Devlin. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186.
  9. Jason Wei and Kai Zou. 2019a. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. In EMNLP-IJCNLP, pages 6382–6388.
  10. Steven Y. Feng, Varun Gangal, Jason Wei, Sarath Chandar, Soroush Vosoughi, Teruko Mitamura, and Eduard Hovy. 2021. A survey of data augmentation approaches for nlp. In Findings of ACL-IJCNLP, pages 968–988.
  11. Jason Wei and Kai Zou. 2019b. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. Preprint, arXiv:1901.11196.

Updated

↑ back to top

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