Exploring Cultural Bias in Language Models Through Word Grouping Games
Abstract: Large language models (LLMs) can exhibit cultural bias, overlooking and misrepresenting cultural nuances. Models who unequally represent global cultures can reinforce harmful stereotypes. Evaluating the extent of cultural bias in an LLM, then, is crucial to equitable model development. Most previous works focus on question-answering (QA) tasks~\cite{palta-rudinger-2023-fork}.
QA tasks focus on one correct answer given the cultural context, despite in many cases, there being a group of correct answers with shared characteristics for a given question.
We proposed a task focusing on word groups, Word Grouping Game (WGG) that implicitly evaluates the model’s cultural knowledge and norms. In WGG, LLMs are given a pool of words, where they must separate the words into groups of four words tied under a common topic. In order to perform well in the game, the model also needs to perform culture-related reasoning. We evaluated the game with two cultures, Latinx/Hispanic and Chinese, in both the native language and an English translation for comparison. Through experimentation, we find biases towards Chinese culture-based groupings, as well as disparities in performance between open- and closed-source models based on the language used for a given game.
Paper Type: Short
Research Area: Ethics, Bias, and Fairness
Research Area Keywords: LLM evaluation, multi-lingual, game, benchmark
Contribution Types: Model analysis & interpretability, Data resources
Languages Studied: English, Spanish, Chinese
Submission Number: 2426
Loading