LINUS@EEUCA 2026: Fine-grained Toxicity Detection in Gaming Chat using Multilingual Transformers

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Gaming & Interactive Media · Depth: Expert, quick

Summary

The LINUS@EEUCA 2026 team developed a fine-grained toxicity detection system for online gaming chat, addressing the challenge of subtle, context-dependent language. Their approach utilized the GameTox dataset, comprising approximately 53,000 World of Tanks chat utterances annotated across six categories, including Insults and Flaming, Hate and Harassment, and Threats. The best performing model was a multilingual BERT variant, mmBERT-base, fine-tuned with class-weighted cross-entropy loss. This model achieved a Macro F1 score of 0.5882 during validation and an official test Macro F1 of 0.5104 on the shared task leaderboard. An internal evaluation on a development split yielded 0.4282, prompting analysis into distributional sensitivity to gaming slang and class imbalance.

Key takeaway

For NLP engineers developing moderation systems for online gaming platforms, you should consider fine-tuning multilingual BERT models like mmBERT-base. Implementing class-weighted cross-entropy loss is critical to effectively handle the inherent class imbalance and context-specific slang found in gaming chat. This approach can improve the detection of subtle toxic behaviors across various categories, enhancing player safety and community health.

Key insights

Fine-grained toxicity detection in gaming chat is challenging but achievable using multilingual transformers and weighted loss.

Principles

• Context-dependent language complicates toxicity detection.
• Class-weighted loss is crucial for imbalanced datasets.

Method

Fine-tune mmBERT-base with class-weighted cross-entropy loss on a domain-specific dataset like GameTox for multi-label toxicity classification.

In practice

• Apply mmBERT-base for nuanced gaming chat moderation.
• Implement class-weighted loss to handle imbalanced toxicity categories.

Topics

• Toxicity Detection
• Gaming Chat
• Multilingual Transformers
• mmBERT-base
• Class Imbalance
• GameTox Dataset

Code references

• sunilRegmi-ai/eeuca-toxicity-detection

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Anatomy of Data Repositories for the Analysis and Detection of Toxicity in Portuguese — Lexical diversity and annotation practices significantly impact cross-domain generalization in hate speech detection datasets.
• LinguIUTics at PsyDefDetect: Iterative Imbalance-Aware Fine-tuning of Qwen3-8B for Psychological Defense Mechanism Classification — Iterative imbalance-aware fine-tuning of LLMs, combined with augmentation and post-processing, significantly improves rare-class NLP performance.
• Muted: Multilingual Targeted Offensive Speech Identification and Visualization — Muted identifies targeted offensive speech using transformer models and spaCy, requiring no fine-tuning for toxic span detection.
• On GATE, Text and Social Media Analysis, and Detecting Misinformation Online — A multi-layer analytical pipeline effectively extracts and categorizes digital identity aspects of agrifood products from online discourse.
• Harder to Defend: Towards Chinese Toxicity Attacks via Implicit Enhancement and Obfuscation Rewriting — Chinese implicit toxicity, combining indirectness and obfuscation, presents a major challenge for LLM detectors, with CITA revealing high missed-detection rates.
• A Multitask Transformer for Offensive Language Detection and Target Identification in HateBR — Multitask learning improves offensive language detection consistency but may reduce fine-grained target attribution.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.