Researchers have developed C$^2$TU, a novel method to identify and neutralize cloaked toxicity in Chinese social media content. This approach uses a homophone graph and a toxic lexicon to detect disguised toxic words, then filters out non-toxic candidates and corrects cloaked terms to their original toxic meanings. Two variants of the method exist: one based on BERT and another utilizing Large Language Models (LLMs) that addresses autoregressive limitations to leverage full semantic context for improved detection. Experiments show C$^2$TU significantly outperforms existing methods, achieving up to a 71% increase in F1 score and a 35% increase in accuracy on Chinese toxic datasets. AI
IMPACT Enhances content moderation capabilities for platforms dealing with sophisticated forms of online abuse.
RANK_REASON Academic paper detailing a new method for toxicity detection. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
