Intersectional Bias in Japanese Large Language Models from a Contextualized Perspective

Hitomi Yanaka, Xinqi He, Jie Lu, Namgi Han, Sun-Jin Oh, Ryoma Kumon, Yuki Matsuoka, Kenji Watabe, Yuko Itatsu2025 · DOI

Introduction

As Large Language Models (LLMs) are increasingly integrated into social systems, evaluating their fairness becomes a critical safety concern. However, most existing bias research focuses on single attributes in isolation—such as gender, age, or nationality. In reality, discrimination often occurs at the "intersection" of these identities, where the experience of a young foreign woman differs significantly from that of an older local man. Furthermore, in the Japanese context, bias often manifests through subtle microaggressions rather than overt hate speech.

This paper introduces inter-JBBQ, a new benchmark designed to evaluate intersectional bias in Japanese LLMs. By focusing on how multiple social attributes overlap within specific social contexts, the researchers demonstrate that bias is not merely additive but is deeply tied to the situational context of the prompt.

The inter-JBBQ Benchmark

The primary contribution of this work is the inter-JBBQ dataset, which builds upon the Japanese Bias Benchmark for Question Answering (JBBQ). The dataset consists of 1,400 question pairs covering four sensitive topics in Japanese society: house renting, marriage markets, academic research, and social etiquette.

Methodology

The benchmark utilizes a multiple-choice Question Answering (QA) framework with two distinct context types:

  1. Ambiguous Context: Scenarios where no factual information is provided to answer the question (e.g., "Person A and Person B participated in an event"). In these cases, the only correct answer is "Unknown."
  2. Disambiguated Context: Scenarios where factual information is provided to allow for a single correct answer.

By comparing model responses across these contexts, the researchers can determine if a model relies on harmful stereotypes (negative questions) or remains neutral when information is missing. The study evaluated several models, including the Swallow series (Japanese-optimized models based on Llama-3.1) and GPT-4o, using both standard prompts and "debias" prompts that explicitly instruct the model to avoid social bias.

Key Insights

Bias is Context-Dependent

The research reveals that the "direction" of bias changes depending on the topic. For instance, in the "marriage market" category, models consistently favored "female" personas over "male" or "transgender" personas, reflecting a specific social commodification. However, in "academic research," the bias was less consistent and shifted unpredictably as more attributes—such as income or nationality—were introduced. This suggests that a model might appear "fair" in one scenario while exhibiting significant bias in another.

The "Unknown" Trap and Model Scale

A critical measure of bias is whether a model correctly identifies ambiguity. GPT-4o demonstrated high accuracy (~100%) in choosing "Unknown" when facts were missing, likely due to extensive safety alignment. In contrast, smaller models (8B) often struggled, falling below 50% accuracy and instead "hallucinating" a choice based on underlying stereotypes. Larger models (70B) and those with instruction tuning (the "+i" versions) were significantly more robust at avoiding these biased leaps in logic.

Intersectionality is Non-Additive

One of the most significant findings is that intersectionality is not a simple addition of biases. Adding a second or third attribute does not just increase a bias score; it can completely flip the model's preference. For example, a model's preference for a specific gender might vanish or reverse when a specific nationality or employment status is introduced. This highlights the necessity of evaluating models using complex, multi-attribute scenarios rather than single-variable tests.

Results and Limitations

Performance Summary

  • GPT-4o: Showed the highest resistance to bias in ambiguous settings, though it still exhibited stereotypical alignments in specific social contexts.
  • Swallow-70B-Instruct: Performed well (85-91% accuracy on ambiguity), outperforming its smaller counterparts.
  • Prompting Effects: While "debias" prompts improved GPT-4o's performance, they had inconsistent effects on the Swallow models, sometimes even degrading accuracy on factual questions.

Future Directions

The authors acknowledge several limitations, including the current scope of only four social topics and the use of template-based questions which may lack linguistic variety. Future work will focus on using LLMs to generate more natural-sounding templates and expanding the benchmark to include a wider array of social attributes and more recent models like Gemini or DeepSeek.

Conclusion

The inter-JBBQ benchmark provides a vital tool for understanding the nuanced ways LLMs process social identities in Japan. By demonstrating that bias is highly contextual and non-linear, this research underscores the need for more sophisticated, intersectional evaluation frameworks to ensure the development of fair and safe AI systems.

Summary updated

BibTeX
@misc{yanaka2025intersectional,
  author = {Hitomi Yanaka and Xinqi He and Jie Lu and Namgi Han and Sun-Jin Oh and Ryoma Kumon and Yuki Matsuoka and Kenji Watabe and Yuko Itatsu},
  title = {{Intersectional Bias in Japanese Large Language Models from a Contextualized Perspective}},
  year = {2025},
  doi = {10.18653/v1/2025.gebnlp-1.2},
}
Figure