Abstract
In this paper we explore the linguistic components of toxic behavior by using crowdsourced data from over 590 thousand cases of accused toxic players in a popular match-based competition game, League of Legends. We perform a series of linguistic analyses to gain a deeper understanding of the role communication plays in the expression of toxic behavior. We characterize linguistic behavior of toxic players and compare it with that of typical players in an online competition game. We also find empirical support describing how a player transitions from typical to toxic behavior. Our findings can be helpful to automatically detect and warn players who may become toxic and thus insulate potential victims from toxic playing in advance.
Introduction
Toxic behavior—ranging from griefing to verbal abuse—is a significant challenge in multiplayer online games. While players often recognize toxicity intuitively, defining and detecting it automatically is difficult because it frequently overlaps with competitive "trash talk." This research aims to move beyond simple keyword filtering by analyzing the linguistic evolution of toxic players during a match. By identifying specific patterns and the timing of toxic speech, we can build a foundation for systems that predict and mitigate toxicity before it degrades the player experience.
Methodology and Dataset
The study utilizes a massive dataset from the League of Legends (LoL) Tribunal, a legacy crowdsourced system where players voted on whether reported peers were actually toxic. The dataset includes 590,311 Tribunal cases representing over 2.1 million matches, containing roughly 24 million messages from "toxic" players (those convicted by the Tribunal) and 33 million messages from "typical" players.
Researchers employed three primary analytical techniques:
- N-gram Analysis: Extracting uni-grams and bi-grams to identify "discriminative" terms used almost exclusively by toxic players.
- Temporal Density Plotting: Mapping chat volume across "normalized time" (0% to 100% of a match duration).
- Phase Transition Analysis: Tracking the "last time of use" for strategic and social words to identify the point where behavior shifts from typical to toxic.
Linguistic Signatures of Toxicity
Analysis reveals distinct linguistic profiles for typical versus toxic players. Toxic players send 38% fewer messages than typical players, but their messages are longer on average (3.139 words vs. 2.732 words). While common profanity like "fucking" is used by both groups, toxic players are uniquely identified by specific abusive bi-grams.
Typical players focus on coordination and social niceties (e.g., "group push," "sorry," ":D"), whereas toxic players utilize language aimed at demeaning others (e.g., "fucking retard," "report noob").
Temporal Evolution of Toxicity
Toxicity is often a reactive "phase transition." Many toxic players begin a match behaving like typical players, using strategic terms. However, as the game progresses unfavorably, they hit a breaking point. A key predictor of this shift is the absence of helpfulness; toxic players stop using teamwork-related words (e.g., "ward," "ult") and social cues long before the match ends.
As shown in the temporal density analysis, different types of toxicity peak at different stages. Early-game toxicity often involves threats of "intentional feeding," mid-game frustration focuses on lane performance (e.g., "fucking bot"), and late-game toxicity—accounting for 82.94% of discriminative toxic bi-grams—is dominated by externalizing blame (e.g., "report noob").
Conclusion
This study demonstrates that toxicity in online games is not just about vocabulary but about a temporal shift in communication. The transition from strategic cooperation to abusive blaming provides a measurable signal for automated detection. Future work should focus on interactional context—how one player's comments trigger another's toxicity—to further distinguish between harmless competitive banter and truly harmful behavior.
Summary updated