Elamur Research⌕
☰
home / papers / 2026 / TAGA@EEUCA 2026: Token-Attribution Guided Attention for Fine-Grained Toxic Behaviour Classification in Online Gaming Communities

TAGA@EEUCA 2026: Token-Attribution Guided Attention for Fine-Grained Toxic Behaviour Classification in Online Gaming Communities

Paper · 2026 · Akshyat Shah, Shashi Sah, Aryan Gupta et al.

Summary

The paper presents TAGA, a system for the EEUCA 2026 Shared Task on toxic behaviour in gaming chat. The task is to assign each utterance to one of six intent classes in the GameTox dataset, about 53K English World of Tanks messages. The data is heavily skewed: Threats and Extremism make up well under one percent. Gaming chat is also hard to classify because of slang, leetspeak and censored profanity.

The pipeline has four main parts:

  • Preprocessing: leetspeak decoding, expansion of gaming abbreviations, and regex-based uncensoring.
  • Encoder: a DeBERTa-v3-base backbone.
  • Token attributions: the Detoxify scorer measures how much the toxicity score drops when each token is deleted (leave-one-out).
  • Attribution-guided pooling: these attribution scores become bias terms in an attention pooling layer.

Focal loss with label smoothing, sentence-level features from two Detoxify models, and oversampling complete the system. A five-phase ablation raises test macro-F1 from 0.498 to 0.618. Unlike the earlier configurations, the final model predicts the rarest class, Extremism, at all. This shows that signals from pretrained toxicity scorers can replace manual token-level annotation as an inductive bias.

Key points

  • Data: GameTox (Naseem et al., 2025), split into 42,959 train, 5,367 validation and 5,375 test utterances. About 81% of messages are Non-Toxic, Threats 0.14% and Extremism 0.06%.
  • Attribution method: leave-one-out scoring with the Detoxify unbiased model on 4 channels (toxicity, threat, insult, identity attack). It is applied only to utterances whose baseline toxicity exceeds τ = 0.15.
  • Ablation results (test macro-F1):
    • E0 baseline: 0.498
    • E1 preprocessing: 0.508
    • E2 loss redesign: 0.511
    • E3 architecture: 0.495 (a drop)
    • E4 Dual TAGA: 0.590
    • E5 full TAGA: 0.618
  • Detoxify features need support: adding them without oversampling or the auxiliary loss (E4a) lowered test macro-F1 to 0.448. With both added, performance jumped substantially.
  • Per-class results and limitations:
    • Extremism F1 rose from 0 to 0.667, but Hate stays low at 0.469, partly due to confusion with Extremism.
    • Leave-one-out attribution costs O(n) Detoxify passes per toxic utterance.
    • The system was evaluated on only one game.

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

🎓Delhi Technological UniversityNew Delhi · India4 works · 196 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksPsychology · 1,552 worksSocial Psychology · 306 works

🧭 Topics 1Research subjects and languages the work is about.Bias, fairness & safety · 517 works

Cite

Cite as Akshyat Shah, Shashi Sah, Aryan Gupta and Kavinder Singh. 2026. TAGA@EEUCA 2026: Token-Attribution Guided Attention for Fine-Grained Toxic Behaviour Classification 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-1.22

@inproceedings{shah2026taga,
  title = {{TAGA@EEUCA 2026: Token-Attribution Guided Attention for Fine-Grained Toxic Behaviour Classification in Online Gaming Communities}},
  author = {Akshyat Shah, Shashi Sah, Aryan Gupta and Kavinder Singh},
  booktitle = {Proceedings of the 9th Workshop on Event Extraction and Understanding: Challenges and Applications (EEUCA 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.eeuca-1.22},
  url = {https://doi.org/10.18653/v1/2026.eeuca-1.22}
}

References

23 references, 3 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. Jeremy Blackburn and Haewoon Kwak. 2014. Stfu noob! predicting crowdsourced decisions on toxic behavior in online games. In Proceedings of the 23rd International Conference on World Wide Web, pages 877–888. find
  3. Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2023. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. In International Conference on Learning Representations.
  4. Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations.
  5. Nhat M Hoang, Xuan Long Do, Duc Anh Do, Duc Anh Vu, and Luu Anh Tuan. 2024. Toxcl: A unified framework for toxic speech detection and explanation. arXiv preprint arXiv:2403.16685. find
  6. 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
  7. Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, and 1 others. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825. find
  8. Qian Chen, Zhu Zhuo, and Wen Wang. 2019. Bert for joint intent classification and slot filling. arXiv preprint arXiv:1902.10909. find
  9. Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. 2016. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174. find
  10. Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, and 1 others. 2024. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70):1–53. find
  11. Bruno Mendes da Silva, Mirian Tavares, Filipa Cerol, Susana Mendes da Silva, Paulo Falcão, and Beatriz Isca Alves. 2020. Playing against hate speech – how teens see hate speech in video games and online gaming communities. volume 3, pages 34–52. find
  12. Rosa Falotico and Piero Quatto. 2015. Fleiss’ kappa statistic without paradoxes. Quality & Quantity, 49:463–470. find
  13. Antigoni Founta, Constantinos Djouvas, Despoina Chatzakou, Ilias Leontiadis, Jeremy Blackburn, Gianluca Stringhini, Athena Vakali, Michael Sirivianos, and Nicolas Kourtellis. 2018. Large scale crowdsourcing and characterization of twitter abusive behavior. In Proceedings of the International AAAI Co… find
  14. Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, and 1 others. 2024. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295. find
  15. Laura Hanu and Unitary team. 2020. Detoxify. https://github.com/unitaryai/detoxify. ↗
  16. Jigsaw/Conversation AI. 2018. Toxic comment classification challenge. https://www.kaggle.com/c/j igsaw-toxic-comment-classification-chall enge. ↗
  17. Haewoon Kwak, Jeremy Blackburn, and Seungyeop Han. 2015. Exploring cyberbullying and other toxic behavior in team competition online games. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, pages 3739– 3748. find
  18. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980–2988.
  19. Ilya Loshchilov and Frank Hutter. 2017. SGDR: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations. find
  20. Marcus Märtens, Siqi Shen, Alexandru Iosup, and Fernando Kuipers. 2015. Toxicity detection in multiplayer online games. In 2015 International Workshop on Network and Systems Support for Games (NetGames), pages 1–6. find
  21. Binny Mathew, Punyajoy Saha, Seid Muhie Yimam, Chris Biemann, Pawan Goyal, and Animesh Mukherjee. 2021. Hatexplain: A benchmark dataset for explainable hate speech detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14867–14875. find
  22. Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao Wu. 2018. Mixed precision training. In International Conference on Learning Representations. 2019. Predicting the type and target… find
  23. 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

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/shah-2026-taga-eeuca-2026-token-attribution/