Context-Aware Toxicity Detection in Game Chat: Domain-Adaptive Pretraining with Match Metadata under Limited Labels

Toxic chat is a persistent problem in competitive games, spurring ongoing moderation efforts. But these efforts face challenges: toxicity is context-dependent, may span multiple messages, and is shaped by player and team dynamics. We adapt RoBERTa for game moderation by modeling text and structured match metadata (player and team identifiers, with only ~100 labeled matches. We introduce Contextual Domain-Adaptive Pretraining (C-DAP), a DAP variant that injects player/team indicators as custom tokens during pretraining and finetuning to learn gaming-specific contextual embeddings. Using datasets from Defense of the Ancients 2 ( DOTA 2 ) and Call of Duty®: Modern Warfare®III ( MWIII ), we compare sources of context (metadata, prior interactions, conversational history), identify when they help most, and show how to leverage them. C-DAP consistently outperforms strong baselines while remaining simple and reusable. Our results highlight the value of context-aware, domain-specific modeling for proactive moderation and offer a practical recipe for training robust, game-tailored text toxicity detectors with low annotation effort.