Scalable and Culturally Specific Stereotype Dataset Construction via Human-LLM Collaboration

Short summary

Researchers introduce EspanStereo, a Spanish-language dataset for evaluating stereotypes in large language models across multiple Spanish-speaking regions, built through human-LLM collaborative annotation. The framework addresses the lack of non-English stereotype benchmarks and reveals significant regional variation in LLM bias. The methodology scales to other languages, offering a pathway toward comprehensive multilingual and cross-cultural bias evaluation.

• •New dataset (EspanStereo) enables stereotype evaluation for Spanish LLMs across multiple regions
• •Human-LLM collaborative framework reduces costs and identifies region-specific biases
• •Scalable methodology applicable to other languages and cultural contexts

Generated with AI, which can make mistakes.