From National Curricula to Cultural Awareness: Constructing Open-Ended Culture-Specific Question Answering Dataset

TL;DR

The paper addresses the challenge of cultural and linguistic misalignment in large language models caused by English-centric data. It introduces CuCu, a multi-agent framework that leverages national curricula to automatically generate open-ended, culture-specific QA data, and applies it to the Korean Social Studies curriculum to produce KCaQA with 34,128 QA pairs across four languages. Quantitative and qualitative analyses show that KCaQA concentrates on curriculum-relevant Korean sociocultural topics and yields culturally grounded responses, validated by LLM-based judgments and human inspection. This approach offers a scalable path to culture-grounded supervision for sovereign or culture-adapted LLMs and can be extended to other nations with publicly available curricula. The work highlights practical benefits for improving cultural sensitivity and alignment in multilingual settings while acknowledging limitations related to curriculum scope and potential biases.

Abstract

Large language models (LLMs) achieve strong performance on many tasks, but their progress remains uneven across languages and cultures, often reflecting values latent in English-centric training data. To enable practical cultural alignment, we propose a scalable approach that leverages national social studies curricula as a foundation for culture-aware supervision. We introduce CuCu, an automated multi-agent LLM framework that transforms national textbook curricula into open-ended, culture-specific question-answer pairs. Applying CuCu to the Korean national social studies curriculum, we construct KCaQA, comprising 34.1k open-ended QA pairs. Our quantitative and qualitative analyses suggest that KCaQA covers culture-specific topics and produces responses grounded in local sociocultural contexts.

Figures (1)

• Figure 1: Overview of CuCu. CuCu constructs an open-ended culture-specific QA pairs using national curricula.