arxivcs.CLcs.AI2026-07-17
Conditional Reliability of Toxicity Signals for Multilingual and Code-Mixed Abuse Detection
Indraveni Chebolu, Rohan Singh, Arnab Mallick, Harmesh Rana
Moderation systems increasingly rely on external toxicity tools, but those tools are unreliable under code-mixing, transliteration, slang, and language mismatch. We study the \emph{conditional reliability} of toxicity priors in Indian multilingual and code-mixed short text: Engli…