Context-Aware Toxicity Detection in Multiplayer Games: Integrating Domain-Adaptive Pretraining and Match Metadata

TL;DR

This work tackles proactive toxicity moderation in multiplayer games by emphasizing context beyond single messages. It proposes a framework using Domain-Adaptive Pretraining (DAP) with MLM and metadata tokens on a DistilRoBERTa-base backbone to capture game-specific slang and cross-message context. Through DOTA 2 and MWIII datasets, it shows that toxicity is often context-dependent, that metadata and long-term player behavior can boost predictive performance, and that gains from DAP are dataset-dependent. These results highlight the importance of context-aware, domain-specific moderation for scalable, proactive game-community management.

Abstract

The detrimental effects of toxicity in competitive online video games are widely acknowledged, prompting publishers to monitor player chat conversations. This is challenging due to the context-dependent nature of toxicity, often spread across multiple messages or informed by non-textual interactions. Traditional toxicity detectors focus on isolated messages, missing the broader context needed for accurate moderation. This is especially problematic in video games, where interactions involve specialized slang, abbreviations, and typos, making it difficult for standard models to detect toxicity, especially given its rarity. We adapted RoBERTa LLM to support moderation tailored to video games, integrating both textual and non-textual context. By enhancing pretrained embeddings with metadata and addressing the unique slang and language quirks through domain adaptive pretraining, our method better captures the nuances of player interactions. Using two gaming datasets - from Defense of the Ancients 2 (DOTA 2) and Call of Duty$^\circledR$: Modern Warfare$^\circledR$III (MWIII) we demonstrate which sources of context (metadata, prior interactions...) are most useful, how to best leverage them to boost performance, and the conditions conducive to doing so. This work underscores the importance of context-aware and domain-specific approaches for proactive moderation.

Figures (4)

• Figure 1: Context levels within a match. Message 6 is evaluated for toxicity, 0 is the start of the match. In addition to the 'toxic' label, the annotator records whether or not the message is toxic only because of previous messages ('contextual' label).
• Figure 2: Differing performance boost of DAP between datasets (subplot a.), performance boost from using metadata only in pretraining (subplot b.), more context is not always better (subplot c.). Subplots a. and c. are averaged across experiments, the error bar being standard deviation between experiments. Subplot b. is averaged across repeated runs of single experiments, the error bar being standard deviation between runs.
• Figure 3: Experiment pipeline. In part A) we experiment with including metadata in domain adaptive pretraining. Part B) involves different context levels as shown in Figure \ref{['context_levels']}.
• Figure 4: Example of epoch selection for the experiment using distilRoBERTa-base and messages from current player with learning rate 5e-6. Metrics from all other experiments can be viewed in the same format at https://github.com/AdrienSF/proactive-contextful-toxicity-detection.