Reducing the Emotional Distress of Content Moderators through LLM-based Target Substitution in Implicit and Explicit Hate-Speech (ACM)

Detecting hate speech is challenging due to its subtle, context-dependent nature—especially when it targets specific groups. This study explores how moderators’ backgrounds influence their emotional and cognitive responses, proposing a method to substitute real communities with fictional ones to reduce distress. Results show this approach lowers emotional impact while maintaining reasonable accuracy. Contextual familiarity boosts moderation performance, and although distress rises with prolonged exposure, substitution proves effective in easing emotional strain without compromising effectiveness.

https://dl.acm.org/doi/10.1145/3708319.3733712