Skip to main content
Back

IGC-BERT Catches League of Legends Toxic Players With 12-Point Recall Boost

Researchers built L2DTnH, a fine-grained League of Legends chat dataset, and fine-tuned BERT into IGC-BERT to catch in-match toxicity offline.

Pro

  • New ESORICS 2026 paper releases L2DTnH, a hand-labeled dataset of 15.2k League of Legends chat messages.
  • Eight veteran LoL players annotated toxic vs non-toxic messages drawn from the Tribunal dataset.
  • Fine-tuned IGC-BERT outperforms general-purpose toxicity APIs like Perspective and Detoxify on in-match chat.
  • Aggregating messages from the same player boosts recall by 12 points versus single-message classification.
  • Companion browser extension flags toxic web content locally, without sending data to third-party AI servers.
  • Kids-show false positives show why domain-specific toxicity detection needs domain-specific training data.

Toxicity in competitive multiplayer games is one of those problems everyone acknowledges but few actually try to fix at the source, meaning the chat window during a live match. A new paper accepted to ESORICS 2026 takes a swing at that gap with a fine-grained dataset of League of Legends chat messages, a fine-tuned BERT classifier called IGC-BERT, and a browser extension that runs entirely on your machine.

The setup looks simple, but the interesting parts sit in the methodology and the limitations the authors are honest about.

The gap nobody was filling

Few works have proposed countermeasures developed and tested on textual messages sent during a match, meaning when the harassment actually occurs. The authors back this up with a systematic literature review: they screened 1,039 papers and found that only 15 proposed ML/NLP-based mechanisms to counter toxicity or harassment, all of them relying on datasets with a limited scope in openness, source, or labeling granularity.

General-purpose toxicity classifiers like Perspective API or Detoxify are trained on comment sections from news sites and social platforms. Gaming chat has its own vocabulary, its own rhythm (short bursts, mid-fight abbreviations, in-game slang like "bot lane noob"), and its own norms about what counts as banter versus harassment. A classifier trained on Reddit comments does not know that.

Building L2DTnH from the Tribunal archives

L2DTnH draws from the well-known Tribunal dataset, which reports the chatlogs of League of Legends matches that included some verified harassment, but without any fine-grained label, making it hardly usable for toxicity detection. The raw material existed, but nobody had labeled it at the message level.

To fix that, the authors partnered with 8 expert League of Legends players and created a fine-grained labelled dataset, L2DTnH, containing 1.4k toxic and 13.8k non-toxic messages exchanged during LoL matches. Using veteran players rather than crowdworkers matters here: they can tell the difference between an insult, a strategic callout, and sarcasm inside a specific match context.