---
title: "FNLP412@EEUCA 2026 — paper by Radulescu (2026)"
description: "This shared-task paper (EEUCA 2026) addresses fine-grained toxicity classification of in-game chat. 16 references."
authors: ["Mihai Radu Radulescu"]
year: 2026
venue: "Proceedings of the 9th Workshop on Event Extraction and Understanding: Challenges and Applications (EEUCA 2026)"
source: "https://aclanthology.org/2026.eeuca-1.10/"
doi: "10.18653/v1/2026.eeuca-1.10"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/radulescu-2026-fnlp412-eeuca-2026-understanding-toxic/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/radulescu-2026-fnlp412-eeuca-2026-understanding-toxic/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/30/) / FNLP412@EEUCA 2026: Understanding Toxic Behavioral Intent in Gaming Chat Logs using Transfer Learning and Synthetic Data Augmentation

# FNLP412@EEUCA 2026: Understanding Toxic Behavioral Intent in Gaming Chat Logs using Transfer Learning and Synthetic Data Augmentation

**Authors:** Mihai Radu Radulescu · **Year:** 2026 · **Venue:** [Proceedings of the 9th Workshop on Event Extraction and Understanding: Challenges and Applications (EEUCA 2026)](https://md.research.elamur.ai/free/en/venues/eeuca/) ([Workshop on Event Extraction and Understanding: Challenges and Applications 2026](https://md.research.elamur.ai/free/en/venues/eeuca/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.eeuca-1.10/) · [PDF](https://aclanthology.org/2026.eeuca-1.10.pdf) · [DOI 10.18653/v1/2026.eeuca-1.10](https://doi.org/10.18653/v1/2026.eeuca-1.10) · [OpenAlex](https://openalex.org/W7166836336)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
This shared-task paper (EEUCA 2026) addresses fine-grained toxicity classification of in-game chat. It uses the GameTox corpus of World of Tanks messages, labelled on a six-level scale from non-toxic to extremism. The data are hard to model for three reasons: utterances are extremely short, they mix English with Russian, German, Polish, French and other languages, and the label distribution is heavily skewed.

The author cleans the text with URL and user-mention tags, lowercasing, collapsing of repeated characters and a hand-built slang/obfuscation dictionary. Gemini-generated synthetic examples are added for the two rarest classes. Starting from a TF-IDF Logistic Regression baseline, the paper compares several multilingual transformers. It also tests ensembling, teacher–student pseudo-labelling and per-class threshold tuning with a severity-ordered decision rule.

The best system is an MDeBERTa-V3 model trained first on the Jigsaw multilingual toxicity data for one epoch, then fine-tuned on GameTox for five. It reaches 0.6725 Macro-F1, second place on the leaderboard. The same two-stage recipe actually hurt XLM-RoBERTa. The results suggest that intermediate training on generic toxicity data can help moderation of noisy, multilingual gaming chat.

## Key points
- **GameTox class balance:** 81% of messages are non-toxic, and class 5 (extremism) has only 27 samples (0.06%).
- **GameTox message length:** messages average 13.84 characters, with a 75th percentile of 18 characters.
- **Augmentation:** 50 synthetic samples were generated with Gemini 3 Pro for classes 4–5 (threats, extremism).
- **Test Macro-F1 results:**
  - LR baseline: 0.4665
  - Fine-tuned XLM-RoBERTa-base: 0.5752
  - XLMR with Jigsaw pre-training: 0.4755
  - MDeBERTa-V3 with Jigsaw: 0.6725 (accuracy 0.8992)
  - The DeBERTa run used a max length of 256 during the Jigsaw stage, while XLMR used 64, which the author later lists as a limitation.
- **Other variants:** none beat the MDeBERTa-V3 model.
  - An ensemble of M3 and M4 scored 0.6165.
  - Pseudo-labelled DeBERTa-V3-base students scored about 0.61–0.63; the text and Table 1 disagree on which variant got which score.
- **Constraints and future work:** training ran on a MacBook M4 Pro at roughly one hour per epoch, and per-language performance was not analysed. Suggested future work includes Focal Loss, Adversarial Weight Perturbation, backtranslation and CONDA as a pre-training source.

*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

- [University of Bucharest](https://md.research.elamur.ai/free/en/universities/university-of-bucharest/) — University, Bucharest, Romania · 21 works · 516 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works

🧭 Topics (1): [🛡️ Bias, fairness & safety](https://md.research.elamur.ai/free/en/topics/bias-fairness-safety/) · 517 works

## Cite

Cite as: Mihai Radu Radulescu. 2026. FNLP412@EEUCA 2026: Understanding Toxic Behavioral Intent in Gaming Chat Logs using Transfer Learning and Synthetic Data Augmentation. 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.

```bibtex
@inproceedings{radulescu2026fnlp412,
  title = {{FNLP412@EEUCA 2026: Understanding Toxic Behavioral Intent in Gaming Chat Logs using Transfer Learning and Synthetic Data Augmentation}},
  author = {Mihai Radu Radulescu},
  booktitle = {Proceedings of the 9th Workshop on Event Extraction and Understanding: Challenges and Applications (EEUCA 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.eeuca-1.10},
  url = {https://doi.org/10.18653/v1/2026.eeuca-1.10}
}
```

- [← previous FlowHN: Adaptive Token Routing for Efficient Parallel Hybrid Networks](https://md.research.elamur.ai/free/en/papers/moradi-2026-flowhn-adaptive-token-routing-efficient/)
- [next → Focus particles and scalar inferences across humans and language models](https://md.research.elamur.ai/free/en/papers/brousse-2026-focus-particles-scalar-inferences-across/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/30/)

## References

16 references, 2 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](https://scholar.google.com/scholar?q=Aashish+Bhandari%2C+Siddhant+B+Shah%2C+Surendrabikram+Thapa%2C+Usman+Naseem%2C+and+Mehwish+Nasim%2E+2023%2E+Crisishatemm%3A+Multimodal+analysis+of+directed+and+undirected+hate+speech+in+text%2Dembedded+images+from+ru)
2. 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](https://scholar.google.com/scholar?q=Usman+Naseem%2C+Shuvam+Shiwakoti%2C+Siddhant+Bikram+Shah%2C+Surendrabikram+Thapa%2C+and+Qi+Zhang%2E+2025%2E+Gametox%3A+A+comprehensive+dataset+and+analysis+for+enhanced+toxicity+detection+in+online+gaming+communiti)
3. Qian Chen, Zhu Zhuo, and Wen Wang. 2019. Bert for joint intent classification and slot filling. Preprint, arXiv:1902.10909. [find](https://scholar.google.com/scholar?q=Qian+Chen%2C+Zhu+Zhuo%2C+and+Wen+Wang%2E+2019%2E+Bert+for+joint+intent+classification+and+slot+filling%2E+Preprint%2C+arXiv%3A1902%2E10909%2E)
4. Karen Spärck Jones. 1972. A statistical interpretation of term specificity and its application in retrieval. Journal of Documentation, 28(1):11–21. [find](https://scholar.google.com/scholar?q=Karen+Sp%C3%A4rck+Jones%2E+1972%2E+A+statistical+interpretation+of+term+specificity+and+its+application+in+retrieval%2E+Journal+of+Documentation%2C+28%281%29%3A11%E2%80%9321%2E)
5. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. CoRR, abs/1911.02116. [find](https://scholar.google.com/scholar?q=Alexis+Conneau%2C+Kartikay+Khandelwal%2C+Naman+Goyal%2C+Vishrav+Chaudhary%2C+Guillaume+Wenzek%2C+Francisco+Guzm%C3%A1n%2C+Edouard+Grave%2C+Myle+Ott%2C+Luke+Zettlemoyer%2C+and+Veselin+Stoyanov%2E+2019%2E+Unsupervised+cross%2Dlingu)
6. Amane Sugiyama and Naoki Yoshinaga. 2019. Data augmentation using back-translation for context-aware neural machine translation. In Proceedings of the Fourth Workshop on Discourse in Machine Translation (DiscoMT 2019), pages 35–44, Hong Kong, China. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Amane+Sugiyama+and+Naoki+Yoshinaga%2E+2019%2E+Data+augmentation+using+back%2Dtranslation+for+context%2Daware+neural+machine+translation%2E+In+Proceedings+of+the+Fourth+Workshop+on+Discourse+in+Machine+Translati)
7. Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. [Deberta: Decoding-enhanced bert with disentangled attention](https://md.research.elamur.ai/free/en/papers/he-2020-deberta-decoding-enhanced-bert-disentangled/). In International Conference on Learning Representations.
8. 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. Preprint, arXiv:2403.16685. [find](https://scholar.google.com/scholar?q=Nhat+M%2E+Hoang%2C+Xuan+Long+Do%2C+Duc+Anh+Do%2C+Duc+Anh+Vu%2C+and+Luu+Anh+Tuan%2E+2024%2E+Toxcl%3A+A+unified+framework+for+toxic+speech+detection+and+explanation%2E+Preprint%2C+arXiv%3A2403%2E16685%2E)
9. 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](https://scholar.google.com/scholar?q=Ali+H%C3%BCrriyetog%CC%86lu%2C+Surendrabikram+Thapa%2C+Hristo+Tanev%2C+Laxmi+Thapa%2C+and+Surabhi+Adhikari%2E+2026%2E+Overview+of+the+workshop+on+event+extraction+and+understanding%3A+Challenges+and+applications%2E+In+Proceedi)
10. Yuanzhe Jia, Weixuan Wu, Feiqi Cao, and Soyeon Caren Han. 2024. In-game toxic language detection: Shared task and attention residuals (student abstract). Proceedings of the AAAI Conference on Artificial Intelligence, 37(13):16238–16239. [find](https://scholar.google.com/scholar?q=Yuanzhe+Jia%2C+Weixuan+Wu%2C+Feiqi+Cao%2C+and+Soyeon+Caren+Han%2E+2024%2E+In%2Dgame+toxic+language+detection%3A+Shared+task+and+attention+residuals+%28student+abstract%29%2E+Proceedings+of+the+AAAI+Conference+on+Artifici)
11. Ian Kivlichan, Jeffrey Sorensen, Julia Elliott, Lucy Vasserman, Martin Görner, and Phil Culliton. 2020. Jigsaw multilingual toxic comment classification. [https://kaggle.com/competitions/jigsaw-m](https://kaggle.com/competitions/jigsaw-m) ultilingual-toxic-comment-classification. Kaggle. [↗](https://kaggle.com/competitions/jigsaw-m)
12. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2018. [Focal loss for dense object detection](https://md.research.elamur.ai/free/en/papers/lin-2017-focal-loss-dense-object-detection/). Preprint, arXiv:1708.02002.
13. Surendrabikram Thapa, Shuvam Shiwakoti, Siddhant Bikram Shah, Kritesh Rauniyar, Laxmi Thapa, Surabhi Adhikari, Kristina T. Johnson, Ali Hür-riyetoğlu, Hristo Tanev, and Usman Naseem. 2026. Understanding toxic behavior in gaming communities using ai to promote healthier digital spaces. In Proceeding… [find](https://scholar.google.com/scholar?q=Surendrabikram+Thapa%2C+Shuvam+Shiwakoti%2C+Siddhant+Bikram+Shah%2C+Kritesh+Rauniyar%2C+Laxmi+Thapa%2C+Surabhi+Adhikari%2C+Kristina+T%2E+Johnson%2C+Ali+H%C3%BCr%2Driyetog%CC%86lu%2C+Hristo+Tanev%2C+and+Usman+Naseem%2E+2026%2E+Understand)
14. Henry Weld, Guanghao Huang, Jean Lee, Tongshu Zhang, Kunze Wang, Xinghong Guo, Siqu Long, Josiah Poon, and Soyeon Caren Han. 2021. Conda: a contextual dual-annotated dataset for in-game toxicity understanding and detection. Preprint, arXiv:2106.06213. [find](https://scholar.google.com/scholar?q=Henry+Weld%2C+Guanghao+Huang%2C+Jean+Lee%2C+Tongshu+Zhang%2C+Kunze+Wang%2C+Xinghong+Guo%2C+Siqu+Long%2C+Josiah+Poon%2C+and+Soyeon+Caren+Han%2E+2021%2E+Conda%3A+a+contextual+dual%2Dannotated+dataset+for+in%2Dgame+toxicity+under)
15. Dongxian Wu, Shu tao Xia, and Yisen Wang. 2020. Adversarial weight perturbation helps robust generalization. Preprint, arXiv:2004.05884. [find](https://scholar.google.com/scholar?q=Dongxian+Wu%2C+Shu+tao+Xia%2C+and+Yisen+Wang%2E+2020%2E+Adversarial+weight+perturbation+helps+robust+generalization%2E+Preprint%2C+arXiv%3A2004%2E05884%2E)
16. Zachary Yang, Nicolas Grenon-Godbout, and Reihaneh Rabbany. 2023. Towards detecting contextual real-time toxicity for in-game chat. In Findings of the Association for Computational Linguistics: EMNLP 2023. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Zachary+Yang%2C+Nicolas+Grenon%2DGodbout%2C+and+Reihaneh+Rabbany%2E+2023%2E+Towards+detecting+contextual+real%2Dtime+toxicity+for+in%2Dgame+chat%2E+In+Findings+of+the+Association+for+Computational+Linguistics%3A+EMNLP+)

*Updated 1 November 2026*