Intersectional Bias in Japanese Large Language Models from a Contextualized Perspective

TL;DR

This paper introduces a Japanese benchmark to evaluate intersectional bias in large language models, revealing that bias varies with context even when social attributes are held constant.

Contribution

It presents the inter-JBBQ benchmark for assessing intersectional bias in Japanese LLMs within a question-answering framework, highlighting contextual variations.

Findings

Bias output varies with context despite equal social attribute combinations

GPT-4o and Swallow exhibit intersectional bias in Japanese LLMs

Benchmark enables nuanced analysis of social biases

Abstract

An increasing number of studies have examined the social bias of rapidly developed large language models (LLMs). Although most of these studies have focused on bias occurring in a single social attribute, research in social science has shown that social bias often occurs in the form of intersectionality -- the constitutive and contextualized perspective on bias aroused by social attributes. In this study, we construct the Japanese benchmark inter-JBBQ, designed to evaluate the intersectional bias in LLMs on the question-answering setting. Using inter-JBBQ to analyze GPT-4o and Swallow, we find that biased output varies according to its contexts even with the equal combination of social attributes.