Elamur Research⌕
☰
home / papers / 2026 / Gradient Descenders at SemEval-2026 Task 9: Data-Centric Counterfactual Augmentation for Multi-Label Hate Speech Detection

Gradient Descenders at SemEval-2026 Task 9: Data-Centric Counterfactual Augmentation for Multi-Label Hate Speech Detection

Paper · 2026 · Tran Phuoc Thanh Nhan and Dang Van Thin

Summary

The paper describes the Gradient Descenders system for SemEval-2026 Task 9 Subtask 2. The task is multi-label hate speech detection over five categories: political, racial/ethnic, religious, gender/sexual and other. The authors used only the English data, which has two problems. Political content strongly dominates, and identity-based hate almost never appears without political framing. Because of this, a classifier may treat political vocabulary as the signal for toxicity.

To counter this, the team generated counterfactual training examples with an LLM using three prompting strategies:

  • Additive label-flipping: adds identity-based attacks to texts that are already toxic.
  • Context decoupling: removes political framing so that only identity hate remains.
  • Cross-domain identity substitution: swaps one targeted group for another.

They then fine-tuned a Twitter-adapted RoBERTa-base model with an MLP classification head on the expanded corpus. The final system scored 44.15% Macro-F1 on the official test set, above the organizer baseline. The work shows that targeted synthetic data, rather than changes to the model architecture, can improve recall on rare classes in multi-label settings with heavy imbalance and overlapping labels.

Key points

  • Original data: The English training set has 3,222 samples. Political labels make up over 54.95% of labels. Purely racial samples account for only 0.25%, and pure religious or gender samples are nearly absent.
  • Effect of augmentation on label counts:
    • Political: 1,150 → 3,615
    • Racial: 281 → 1,689
    • Religious: 112 → 1,125
    • Gender: 72 → 1,108
    • Other: 126 → 827
    • The 2,047 "no label" samples were left unchanged.
  • Sampling targets: Two-label intersections were upsampled to 500 examples each. Three-label combinations and the newly created "pure" classes were capped at 300 each.
  • Model and training: The backbone is a RoBERTa-base model pretrained on 154M tweets (Loureiro et al., 2023). Preprocessing used ekphrasis. Training used BCE loss, a 0.5 decision threshold, lr 2e-5, batch size 16, 10 epochs and AdamW, on Kaggle T4 GPUs.
  • Result: The system reached a Macro-F1 of 44.15% on the official test set, outperforming the organizer baseline.

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 2Research subjects and languages the work is about.Sentiment & emotion · 972 worksBias, fairness & safety · 517 works

Cite

Cite as Tran Phuoc Thanh Nhan and Dang Van Thin. 2026. Gradient Descenders at SemEval-2026 Task 9: Data-Centric Counterfactual Augmentation for Multi-Label Hate Speech 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-1.113

@inproceedings{nhan2026gradient,
  title = {{Gradient Descenders at SemEval-2026 Task 9: Data-Centric Counterfactual Augmentation for Multi-Label Hate Speech Detection}},
  author = {Tran Phuoc Thanh Nhan and Dang {Van Thin}},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.113},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.113}
}

References

11 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Christos Baziotis, Nikos Pelekis, and Christos Doulk-eridis. 2017. Datastories at semeval-2017 task 4: Daniel Loureiro, Kiamehr Rezaee, Talayeh Riahi, Francesco Barbieri, Leonardo Neves, Luis Espinosa Anke, and José Camacho-Collados. 2023. Tweet insights: A visualization platform to extract temporal… — [arXiv:2308.02142](https://arxiv.org/abs/2308.02142) find
  2. 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
  3. Political Political ∩ Racial Political Political ∩ Racial +350 +115 +350 +126 Orig. Final Growth 14.3× 20.8× Standard Augmentation (2-Label Intersections) Political ∩ Other Injection Political Political ∩ Racial Injection Political +423 +315 77 185 500 500 6.5× 2.7× Complex Intersections (3-Labels)… — POLAR: A Benchmark for Multilingual, Multicultural, and Multi-Event Online Polarization
  4. Van Bach Nguyen, Christin Seifert, and Jörg Schlötterer. 2025. Guiding llms to generate high-fidelity and high-quality counterfactual explanations for text classification. ArXiv, abs/2503.04463. — [arXiv:2503.04463](https://arxiv.org/abs/2503.04463) find
  5. Someen Park, Jaehoon Kim, Seungwan Jin, Sohyun Park, and Kyungsik Han. 2024. PREDICT: Multi-agent-based debate simulation for generalized hate speech detection. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 20963–20987, Miami, Florida, USA. Associat… find
  6. Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 41–58, Online. Association for Computational Linguistics. find
  7. Qianli Wang, Nils Feldhus, Simon Ostermann, Luis-Felipe Villa-Arenas, Sebastian Moller, and Vera Schmitt. 2025. Fitcf: A framework for automatic feature importance-guided counterfactual example generation. ArXiv, abs/2501.00777. find
  8. 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, San Diego, California. Association for Computational Linguistics.
  9. Letian Peng, Yuwei Zhang, and Jingbo Shang. 2023. Controllable data augmentation for few-shot text mining with chain-of-thought attribute manipulation. In Annual Meeting of the Association for Computational Linguistics. find
  10. Paul Röttger, Bertie Vidgen, Dong Nguyen, Zeerak Waseem, Helen Margetts, and Janet Pierrehumbert. 2021. HateCheck: Functional tests for hate speech detection models. In Proceedings of the 59th Annual Meeting of the Association for Computational *Table 4: Performance comparison with competition basel…
  11. Micro-Average P R F1 Pol. Class-wise F1-Score Rac. Rel. Gen. Oth. Organizer Baselines mBERT XLM-RoBERTa RemBERT twitter-roberta-hate LaBSE 31.29 27.33 43.17 36.99 34.68 Our Experiments Internal Baseline Ours (Augmented) 29.22 44.15 67.48 52.73 64.50 57.94 Improvement (∆) +14.93 -2.98 +5.21 +1.84 +1.… find

Updated

↑ back to top

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