Elamur Research⌕
☰
home / papers / 2026 / syuhhh@EEUCA 2026: A Three-Stage Progressive Training Framework for Fine-Grained Toxicity Detection in Online Gaming Communities

syuhhh@EEUCA 2026: A Three-Stage Progressive Training Framework for Fine-Grained Toxicity Detection in Online Gaming Communities

Paper · 2026 · Yuhao Shi, Yu Wang and Shengjie Zhao

Summary

The paper describes the winning entry in the GameTox shared task at the 9th EEUCA workshop. The task asks systems to assign one of six toxic-intent labels to in-game chat messages from World of Tanks. These messages are very short and full of slang, and the toxic classes are heavily outnumbered by non-toxic ones.

The authors build on XLM-RoBERTa-large and train it in three steps:

  1. masked-language-model adaptation on gaming chat corpora;
  2. transfer fine-tuning on multilingual Jigsaw toxicity data;
  3. task fine-tuning with a dual head that pairs class-weighted cross-entropy with supervised contrastive learning (SCL).

Around this core they add several extras. An LLM rewrites short messages with game context and generates synthetic examples for rare classes. Per-class decision thresholds are tuned with Nelder-Mead. For rare classes only, predictions are combined with ToxicBERT and LLM outputs.

The system reaches Macro F1 0.7041 and places first of 35 teams. This shows that domain adaptation and targeted handling of class imbalance matter more than the choice of backbone for moderating game chat.

Key points

  • Data: The GameTox dataset has 53,000 annotated World of Tanks utterances averaging about 12 tokens. Over 70% are Non-toxic, and the three rarest classes (Hate and Harassment, Threats, Extremism) together make up under 5%.
  • Results: The final Macro F1 of 0.7041 beats the best competing team (0.6725) by 3.16 pp and the vanilla XLM-RoBERTa-large baseline (0.6149) by 8.92 pp.
  • Ablation:
    • Domain MLM pre-training plus Jigsaw transfer gave the largest gain (+10.37 pp, as reported by the authors).
    • LLM short-text augmentation added +4.71 pp.
    • Long-tail synthesis added +1.90 pp.
    • SCL and the ensemble together added +2.31 pp.
  • Transfer data: The Stage 2 set draws on Jigsaw data in 5 languages (English, Russian, Turkish, Spanish, French) and keeps 238,465 samples after downsampling.
  • Remaining errors and code: Most errors come from confusing Insults with Hate and Harassment, and Non-toxic with Extremism. Code is released at github.com/oosyh/syuhhh-EEUCA2026.

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

🎓Tongji UniversityShanghai · China19 works · 2,162 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.Bias, fairness & safety · 517 works

Cite

Cite as Yuhao Shi, Yu Wang and Shengjie Zhao. 2026. syuhhh@EEUCA 2026: A Three-Stage Progressive Training Framework for Fine-Grained Toxicity Detection in Online Gaming Communities. Proceedings of the 9th Workshop on Event Extraction and Understanding: Challenges and Applications (EEUCA 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eeuca-1.17

@inproceedings{shi2026syuhhh,
  title = {{syuhhh@EEUCA 2026: A Three-Stage Progressive Training Framework for Fine-Grained Toxicity Detection in Online Gaming Communities}},
  author = {Yuhao Shi, Yu Wang and Shengjie Zhao},
  booktitle = {Proceedings of the 9th Workshop on Event Extraction and Understanding: Challenges and Applications (EEUCA 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.eeuca-1.17},
  url = {https://doi.org/10.18653/v1/2026.eeuca-1.17}
}

References

18 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Aashish Bhandari, Siddhant B Shah, Surendrabikram Thapa, Usman Naseem, and Mehwish Nasim. 2023. CrisisHateMM: Multimodal analysis of directed and undirected hate speech in text-embedded images from Russia-Ukraine conflict. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Reco… find
  2. Tommaso Caselli, Valerio Basile, Jelena Mitrović, and Michael Granitzer. 2020. HateBERT: Retraining BERT for abusive language detection in English. In Proceedings of the 5th Workshop on Online Abuse and Harms, pages 21–32.
  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 11th International AAAI Conference on Web and Social Media.
  4. 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: Human Language Technologies,…
  5. Ali Hürriyetoğlu, Surendrabikram Thapa, Hristo Tanev, Laxmi Thapa, and Surabhi Adhikari. 2026. Overview of the workshop on event extraction and understanding: Challenges and applications. In Proceedings of the 9th Workshop on Event Extraction and Understanding: Challenges and Applications (EEUCA). find
  6. Jigsaw/Conversation AI. 2018. Jigsaw toxic comment classification challenge. Kaggle Competition. find
  7. Jigsaw/Conversation AI. 2020. Jigsaw multilingual toxic comment classification. Kaggle Competition. find
  8. Hongli Liu, Yu Wang, and Shengjie Zhao. 2026. STAR: Semantic-temporal adaptive representation learning for few-shot action recognition. IEEE Transactions on Circuits and Systems for Video Technology. find
  9. Yu Wang, Shengjie Zhao, and Shiwei Chen. 2024. Action-semantic consistent knowledge for weakly-supervised action localization. IEEE Transactions on Multimedia, 26:10279–10289. find
  10. Usman Naseem, Shuvam Shiwakoti, Siddhant Bikram Shah, Surendrabikram Thapa, and Qi Zhang. 2025. GameTox: A comprehensive dataset and analysis for enhanced toxicity detection in online gaming communities. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association… find
  11. Yu Wang, Shengjie Zhao, and Shiwei Chen. 2025. SQL-Net: Semantic query learning for point-supervised temporal action localization. IEEE Transactions on Multimedia, 27:84–94. find
  12. Anil Singh Parihar, Surendrabikram Thapa, and Sushruti Mishra. 2021. Hate speech detection using natural language processing: Applications and challenges. In 2021 5th International Conference on Trends in Electronics and Informatics (ICOEI), pages 1302– 1308. IEEE. find
  13. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, and 1 others. 2019. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, volume 32.
  14. Shailesh Raman, Birju Patel, Sriram Srinivasan, Pnina Shachaf, and David Jurgens. 2021. Chat as currency: Linguistic features of toxicity in online gaming. In Proceedings of the 2021 ACM SIGCHI Conference on Human Factors in Computing Systems, pages 1– 14. find
  15. Surendrabikram Thapa, Shuvam Shiwakoti, Siddhant Bikram Shah, Surabhi Adhikari, Hariram Veeramani, Mehwish Nasim, and Usman Naseem. 2025. Large language models (LLM) in computational social science: Prospects, current state, and challenges. Social Network Analysis and Mining, 15(1):1–30. find
  16. Yu Wang, Shengjie Zhao, Jianyu Wang, and Xutao Chu. 2026. Learning local semantic signals and inter-class discrepancy for weakly supervised video anomaly detection. IEEE Transactions on Multimedia. find
  17. Yu Wang, Shengjie Zhao, Rongqing Zhang, Xiang Cheng, and Liuqing Yang. 2022. Multi-vehicle collaborative learning for trajectory prediction with spatio-temporal tensor fusion. IEEE Transactions on Intelligent Transportation Systems, 23(1):236–248. find
  18. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clément Delangue, Anthony Moi, Pier-ric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and 1 others. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Lan…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/shi-2026-syuhhh-eeuca-2026-three-stage/