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

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Gaming & Interactive Media · Depth: Advanced, short

Summary

The syuhhh@EEUCA 2026 system secured 1st place among 35 teams in the Shared Task on Fine-Grained Toxicity Detection in Online Gaming (GameTox) at the 9th EEUCA Workshop, co-located with ACL 2026. This system addresses 6-class fine-grained toxic intent classification using the official GameTox dataset, which comprises 53,000 real-world World of Tanks chat utterances. The core of the system is a three-stage progressive training framework built upon XLM-RoBERTa-large. This framework includes gaming domain adaptive Masked Language Model (MLM) pre-training, multilingual toxicity transfer fine-tuning, and supervised contrastive learning (SCL)-enhanced target task tuning. Additionally, the system incorporates LLM-driven data augmentation and long-tailed class synthesis to improve performance. It achieved a Macro F1 score of 0.7041, with ablation studies confirming each module's contribution. The code has been released to support further research.

Key takeaway

For NLP Engineers developing content moderation systems in specialized domains like online gaming, this framework offers a proven approach to achieve high accuracy. You should consider implementing a multi-stage training pipeline, starting with domain-adaptive pre-training and incorporating supervised contrastive learning. Utilizing LLM-driven data augmentation can also significantly address challenges with long-tailed classes, improving your system's ability to detect nuanced toxic intent effectively.

Key insights

A multi-stage training framework with domain adaptation and contrastive learning excels at fine-grained toxicity detection in gaming.

Principles

• Domain-specific pre-training improves performance.
• Multilingual transfer fine-tuning enhances robustness.
• Supervised contrastive learning refines classification.

Method

A three-stage progressive training framework: (1) gaming domain adaptive MLM pre-training, (2) multilingual toxicity transfer fine-tuning, and (3) SCL-enhanced target task tuning. Includes LLM data augmentation.

In practice

• Apply multi-stage training for domain-specific NLP.
• Use LLM augmentation for scarce data classes.
• Implement SCL for fine-grained text classification.

Topics

• Fine-Grained Toxicity Detection
• Online Gaming Communities
• XLM-RoBERTa-large
• Supervised Contrastive Learning
• LLM Data Augmentation
• Natural Language Processing

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Holy $#!t: Are popular toxicity models simply profanity detectors? — Toxicity models often misclassify positive profanity due to inadequate training data and contextual understanding.
• 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.
• 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.
• Do We Still Need Humans in the Loop? Comparing Human and LLM Annotation in Active Learning for Hostility Detection — LLM labels can match human-level F1 scores for hostility detection but exhibit distinct error profiles.
• Multi-Stage Training for Abusive Comment Detection in Indic Languages — Multi-stage training with language preprocessing and model ensembles improves abusive comment detection while minimizing false positives.
• mdok-style at SemEval-2026 Task 10: Finetuning LLMs for Conspiracy Detection — Data augmentation and self-training can effectively finetune LLMs for specialized text classification tasks with limited data.

Open in AIssential →

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