GChat-Detect: Domain-Adaptive Toxicity Detection for Online Gaming Chats Using DistilBERT

Abstract

Online multiplayer gaming platforms host millions of concurrent users, making them fertile ground for toxic communication, hate speech, and harassment. Existing content moderation systems struggle with gaming-specific slang, abbrevi-ations, and context-dependent expressions that are absent from general-purpose natural language processing (NLP) corpora. This study, proposes GChat-Detect, a fine-tuned DistilBERT model enhanced with a game-domain adaptive vocabulary layer trained on the publicly available Jigsaw Toxic Comment Classification Challenge dataset. This paper introduces a GameAdapt pre-processing pipeline that handles gaming abbreviations, emote sequences, and code-mixed chat language. This work deploys the model that achieves 93.4% accuracy, 92.8% macro F1-score, and an AUC-ROC of 0.974, outperforming baseline approaches including Logistic Regression, SVM, BiLSTM, and BERT-base. An ablation study confirms that domain-adaptive fine-tuning contributes a 2.2% improvement in F1-score over generic fine-tuning. The proposed framework offers a real-time deployable solution for toxicity detection in online gaming environments relevant to emerging digital economies.