Linguistic Analysis of Toxic Behavior in an Online Video Game

Haewoon Kwak, Jeremy Blackburn

TL;DR

Researchers analyzed over 2 million League of Legends matches to decode the linguistic DNA of "toxic" players. They discovered that toxicity isn't just about using bad words; it’s a temporal "phase transition" where players stop communicating strategically and start lashing out as a match progresses.

What Problem Does This Paper Address?

Toxic behavior—ranging from griefing to verbal abuse—is a plague in multiplayer online games. While most people "know it when they see it," toxicity is notoriously difficult to define and detect automatically because it is subjective and often overlaps with competitive "trash talk."

The authors aim to move beyond simple keyword filtering. They seek to understand the linguistic evolution of a toxic player during a match. By identifying the specific patterns and timing of toxic speech, they hope to provide the foundation for systems that can predict and mitigate toxicity before it ruins the experience for others.

Methods

The researchers employed a multi-step linguistic analysis:

  • N-gram Analysis: They extracted uni-grams (single words) and bi-grams (two-word phrases) to identify "discriminative" terms—words used almost exclusively by toxic players versus typical players.
  • Temporal Density Plotting: They mapped chat volume across "normalized time" (0% to 100% of a match duration) to see when different groups talk the most.
  • Phase Transition Analysis: By tracking the "last time of use" for common strategic and social words (like "sorry" or "ward"), they identified the specific point in a match where a player’s behavior shifts from typical to toxic.

Datasets

The study utilizes a massive, real-world dataset from the League of Legends (LoL) Tribunal, a now-legacy crowdsourced system where players voted on whether reported peers were actually toxic.

  • Scale: 590,311 Tribunal cases representing over 2.1 million individual matches.
  • Content: The data includes roughly 24 million messages from "toxic" players (those convicted by the Tribunal) and 33 million messages from "typical" players (opponents who were not reported).

Key Insights and Ideas

  • The "Phase Transition": Toxicity is often reactive. Many toxic players start the game behaving exactly like typical players—using strategic terms and social cues. However, as the game "plays out contrary to their desires," they hit a breaking point where they stop using positive/strategic language and transition into purely abusive behavior.
  • The Silence of Strategy: One of the strongest predictors of a "turning" player isn't just the presence of slurs, but the absence of helpfulness. Toxic players stop using words related to teamwork (e.g., "ward," "ult," "blue") and social bonding (e.g., "sorry," ":D," "gj") long before the match ends.
  • Chat Volume is Bi-modal: Both typical and toxic players chat most at the very beginning (strategy/ice-breaking) and the very end (GGs or blaming) of a match. However, typical players are significantly more active in the "ice-breaking" phase, while toxic players dominate the mid-to-late game chat volume.
  • The "F-Word" Ambiguity: Interestingly, the word "fucking" was not a discriminative uni-gram. It is used frequently by both typical and toxic players, highlighting that profanity alone doesn't define toxicity in a competitive subculture—context and intent do.

Key Results

  • Message Length: Toxic players send 38% fewer messages than typical players, but their messages are longer on average (3.139 words vs. 2.732 words).
  • Late-Game Toxicity: 82.94% of discriminative toxic bi-grams (like "report noob") occur in the late stage of the game, suggesting that toxicity is frequently a response to an impending loss.
  • Linguistic Signatures: The researchers identified 252 bi-grams exclusive to toxic players. Early-game toxicity is often "intentional feeding" (e.g., "ill feed"), mid-game is often lane-specific frustration (e.g., "fucking bot"), and late-game is focused on externalizing blame (e.g., "report noob").

Limitations and Future Work

  • Lack of Interactional Context: The study treats messages as individual units. The authors acknowledge that they don't yet analyze the "flow" of conversation—how one player's comment triggers another's toxicity.
  • Trash Talk vs. Toxicity: While they identified discriminative words, building a real-time warning system requires even finer nuance to avoid punishing harmless competitive banter (trash talk) that is part of the game's culture.
  • Future Direction: The authors suggest that future systems should focus on "interaction analysis" to reveal the narratives of how players influence each other's behavior.

Summary updated

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.
BibTeX
@inproceedings{kwak2014linguistic,
  author = {Haewoon Kwak and Jeremy Blackburn},
  title = {{Linguistic Analysis of Toxic Behavior in an Online Video Game}},
  booktitle = {SocInfo Workshops},
  year = {2014},
  url = {https://openreview.net/forum?id=oIh4ASjEK6rb},
}

Loading related content...

Figure