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

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 TAGA (Token-Attribution Guided Attention) system was developed for the EEUCA 2026 Shared Task. Its goal is fine-grained toxic behavior classification in online gaming communities. Online gaming chat presents unique moderation challenges. These include domain-specific slang, deliberate obfuscation, and informal language. Extreme data imbalance for categories like threats also complicates detection. TAGA employs a leave-one-out attribution method using the Detoxify toxicity scorer. This computes per-token attribution scores, which are then projected into learned attention biases. This steers the model towards toxicity-indicative tokens. An ablation study demonstrated performance gains. Key components were domain-specific preprocessing, focal loss with label smoothing, attribution-guided attention pooling, and dual-model Detoxify features with strategic oversampling. The final system achieved a test macro-F1 score of 0.618. Importantly, it produced non-zero predictions despite the dataset's extreme data imbalance.

Key takeaway

For Machine Learning Engineers developing content moderation systems for specialized domains like online gaming, TAGA's approach offers a robust solution. You should consider integrating token-attribution guided attention, using tools like Detoxify for per-token scoring, and applying domain-specific preprocessing. This method effectively addresses challenges posed by slang, obfuscation, and extreme data imbalance. It achieved a macro-F1 score of 0.618. This ensures non-zero predictions for rare toxic categories.

Key insights

TAGA uses token attribution from Detoxify to guide attention, improving fine-grained toxic behavior classification in online gaming chats.

Principles

• Attribution-guided attention improves token relevance.
• Domain-specific preprocessing is vital for slang.
• Strategic oversampling handles data imbalance.

Method

TAGA computes per-token attribution scores using a leave-one-out method with Detoxify, then projects these into learned attention biases to steer the model towards toxicity-indicative tokens.

In practice

• Implement attribution-guided attention for NLP.
• Use Detoxify for token-level toxicity scoring.
• Apply focal loss with label smoothing.

Topics

• Token-Attribution Guided Attention
• Toxic Behavior Classification
• Online Gaming
• Detoxify
• Data Imbalance
• DeBERTa-v3-base

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

Related on AIssential

• Measuring and Mitigating Toxicity in Large Language Models: A Comprehensive Replication Study — Inference-time toxicity mitigation is effective for explicit toxicity but struggles with implicit hate speech and incurs significant latency.
• BetXplain: An Explanation-Annotated Dataset for Detecting Manipulative Betting Advertisements on Social Media — BetXplain provides an explanation-annotated dataset to advance automated detection of manipulative social media betting ads and their mental health impacts.
• CausalDetox: Causal Head Selection and Intervention for Language Model Detoxification — Causal head selection and intervention effectively detoxify LLMs while preserving fluency and accelerating the process.
• Token-weighted Direct Preference Optimization with Attention — Using an LLM's intrinsic attention to weight tokens during preference optimization significantly improves alignment and performance.
• Detoxification for LLM: From Dataset Itself — Detoxifying pretraining datasets directly and semantically reduces LLM toxicity more effectively than post-training methods.
• TantraTagger: A Benchmark Dataset for Tantrayukti-Based Discourse Structure Labelling in Sanskrit Śāstra Texts — TantraTagger is a benchmark dataset for discourse structure labelling in Sanskrit Śāstra Texts.

Open in AIssential →

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