thaulab@EEUCA 2026: Who Said What to Whom? A Targeting-Aware Neural-Symbolic Pipeline for Gaming Toxicity Detection

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

Summary

The thaulab@EEUCA 2026 system addresses toxicity classification in gaming chat through a three-stage neural-symbolic pipeline. It combines an ensemble of two compact transformers, DeBERTa-v3-base (184M) and XLM-RoBERTa-base (278M), with a Linguistically-Informed Mediator (LIM). The LIM resolves inter-model disagreements using corpus-backed lexical normalization, class-conditional unigram scoring, multilingual profanity detection, and agentive targeting analysis. This specifically targets safety-critical minority classes like Hate Harassment, Threats, and Extremism, which face an extreme 1,450:1 Non-toxic to Extremism ratio. A two-stage data augmentation strategy, using only provided training data, tackles this imbalance. The system achieved a Macro F1 of 0.6441 (3rd rank) and an accuracy of 0.9062 (1st rank) on the official test set. The pipeline is domain-portable, requiring only a game-specific entity lexicon substitution for adaptation.

Key takeaway

For Machine Learning Engineers building content moderation systems, this pipeline demonstrates how combining neural models with a linguistically-informed mediator can significantly improve performance on critical, imbalanced toxicity classes. You should consider integrating targeted linguistic analysis and disagreement resolution into your moderation pipeline, especially when dealing with low-resource, high-impact categories like hate speech or threats. This approach offers a robust and portable solution for real-world gaming environments.

Key insights

A neural-symbolic pipeline enhances gaming toxicity detection by combining transformer ensembles with a linguistically-informed mediator.

Principles

• Hybrid neural-symbolic approaches improve classification robustness.
• Targeting minority classes is crucial for safety-critical moderation.
• Data augmentation can mitigate extreme class imbalance effectively.

Method

The system uses a three-stage pipeline: transformer ensemble, a Linguistically-Informed Mediator (LIM) for disagreement resolution (lexical normalization, unigram scoring, profanity detection, agentive targeting), and two-stage data augmentation.

In practice

• Employ DeBERTa-v3-base and XLM-RoBERTa-base for toxicity tasks.
• Integrate linguistic mediators for critical minority class resolution.
• Apply two-stage data augmentation for severe class imbalance.

Topics

• Gaming Toxicity Detection
• Neural-Symbolic AI
• Transformer Models
• DeBERTa-v3-base
• XLM-RoBERTa-base
• Class Imbalance
• Data Augmentation

Code references

• Anmol2059/thaulab_EEUCA

Best for: AI Engineer, 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.
• Reasoners or Translators? Contamination-aware Evaluation and Neuro-Symbolic Robustness in Tax Law — LLM legal reasoning performance is inflated by data contamination; neuro-symbolic systems offer superior robustness and generalization.
• From Prompt Risk to Response Risk: Paired Analysis of Safety Behavior of Large Language Model — LLM safety analysis benefits from paired prompt-to-response severity transitions and joint relevance assessment.
• Holy $#!t: Are popular toxicity models simply profanity detectors? — Toxicity models often misclassify positive profanity due to inadequate training data and contextual understanding.
• 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.
• 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.