Researchers create first Korean toxicity dataset tackling obfuscated harmful language

Language models deployed in online settings increasingly face the challenge of toxic content, but many systems struggle when users intentionally hide harmful language. A research team has addressed this gap by developing KOTOX, a specialized dataset for identifying and remediating disguised toxic expressions in Korean.

The challenge of hidden toxicity#

Korean presents particular vulnerabilities to obfuscation due to its linguistic properties. The language's agglutinative morphology — the ability to combine multiple meaningful units into single words — and the flexibility of Hangeul orthography make it easy for users to mask toxic expressions in ways that traditional detection systems miss. Despite this risk, obfuscation techniques in Korean have received minimal research attention until now.

A linguistically grounded approach#

The researchers categorized Korean obfuscation patterns into linguistically meaningful categories and developed transformation rules based on real-world examples. They then packaged these rules into an open-source framework, enabling systematic generation of obfuscated text variants. The resulting KOTOX dataset pairs neutral and toxic sentences with their obfuscated versions, creating a comprehensive resource for training robust models.

Why it matters#

This is the first dataset to simultaneously address both deobfuscation and detoxification for Korean language models. Systems trained on the dataset perform better at handling disguised toxic content without degrading their ability to process ordinary text. By providing both linguistic insight and practical tools, the work aims to improve how language models deployed for Korean speakers can identify and mitigate harmful content, even when intentionally concealed.

The code and dataset have been made publicly available to support further research in this area.