Go to ACL ARR 2025 February homepagePurposefully Lost in Translation: Expanding The Stereotype Content Model for Cross-Cultural Stereotype Erasure
Abstract: Stereotype detection offers valuable insights for detecting implicit bias in language models. To mitigate such bias, stereotyping theories have been adopted in various NLP tasks. However, these implementations have primarily focused on English language models. As language models are increasingly applied across diverse languages and cultures, it is crucial to develop a model that addresses the range of stereotypes present in these languages and cultures. In this paper, we propose a framework for expanding the Stereotype Content Model (SCM) beyond the English language, demonstrated through the development and validation of our Korean SCM (KoSCM). We also present a translation framework designed to address the challenges related to data annotation, explore the cross-cultural validity of the SCM by evaluating the model against theory-grounded hypotheses, and introduce a novel method for stereotype erasure. To make the study of stereotyping more accessible to a broader range of researchers, we also present SCM prompting, a set of prompt engineering guidelines for LLMs aimed at stereotype detection. Our proposed CoT prompting improves the performance of LLMs by an average of 18.6\%. This study marks the first attempt to implement the SCM in a non-English language and with LLMs, paving the way for research on stereotypes across different languages and models.
Paper Type: Long
Research Area: Computational Social Science and Cultural Analytics
Research Area Keywords: language/cultural bias analysis, NLP tools for social analysis
Languages Studied: English, Korean
Submission Number: 5850
Loading