Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 31·5 min read
Why are all LLMs Obsessed with Japanese Culture? On the Hidden Cultural and Regional Biases of LLMs
TL;DR
A new dataset reveals that large language models (LLMs) exhibit a cultural bias favoring Japan in their responses to culture-related questions. This bias is more pronounced in low-resource languages and emerges after supervised fine-tuning.
✦ Why It Matters
Consider diversifying training datasets to reduce cultural biases in AI applications.
Key Takeaways
- Evaluate LLMs using diverse cultural datasets.
- Monitor cultural biases during model training phases.
- Encourage multilingual training to enhance cultural representation.
- Investigate the impact of language on LLM outputs.
Full Summary
Large language models (LLMs) have been found to have limitations in cultural coverage and competence, often displaying biases towards specific cultures. This study introduces a dataset called Culture-Related Open Questions (CROQ), which includes questions in 24 languages to evaluate LLM responses.
The findings indicate that LLMs show a significant preference for Japan when answering culture-related queries, especially when prompted in low-resource languages. In contrast, responses in high-resource languages like English tend to be more diverse.
The research also identifies that cultural biases in LLMs become evident after the supervised fine-tuning phase, rather than during the initial pre-training. This insight highlights the need for more balanced training data to mitigate cultural biases in AI systems.