Lang Xiong

Detecting mental manipulation is culturally dependent and a highly subjective task. To address this gap, we introduce CultureManip, a novel multilingual benchmark for mental manipulation detection. Using a taxonomy adapted from MentalManip, we tasked two native-speaking annotators per language with identifying manipulation presence and labeling specific techniques (e.g., Denial, Evasion, Feigning Innocence, Rationalization, Playing the Victim Role, Playing the Servant Role, Shaming or Belittlement, Intimidation, Brandishing Anger, Accusation, and Persuasion or Seduction) across four languages: English, Spanish, Chinese, and Tagalog. We then evaluated ChatGPT-3.5 Turbo’s zero-shot performance on the binary task of detecting manipulation.