PSN: Persian Social Norms Dataset for Cross-Cultural AI

2024-06-13

arXiv.org

AI Summary

This work addresses the performance bottlenecks of large language models (LLMs) in understanding and classifying social norms within Iranian society—challenges arising from cultural specificity and the linguistic complexity of Persian. To this end, we introduce PSN (Persian Social Norms), the first high-quality, multidimensional Persian-language social norm dataset, comprising over 1,700 normative statements annotated with contextual scenarios, situational context, and culturally grounded labels, accompanied by expert English translations. Methodologically, we propose a novel pipeline integrating LLM-assisted generation, culturally informed prompt engineering, native-speaker collaborative validation, bilingual alignment, and structured semantic annotation—ensuring scalability, cultural authenticity, and ethical compliance. PSN is the first publicly available, reusable Persian social norm resource, substantially bridging a critical gap in culturally aware AI research. It serves as foundational infrastructure for evaluating, training, and modeling normative adaptation in cross-cultural AI systems.

Technology Category

• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Social Cognition And Interaction
• Application Domains: Humanities & Computational Social Science

Application Category

• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

Datasets capturing cultural norms are essential for developing globally aware AI systems. We present Persian Social Norms (PSN) a novel dataset of over 1.7k Persian social norms, including environments, contexts, and cultural labels, alongside English translations. Leveraging large language models and prompt-engineering techniques, we generated potential norms that were reviewed by native speakers for quality and ethical compliance. As the first Persian dataset of its kind, this resource enables computational modeling of norm adaptation, a crucial challenge for cross-cultural AI informed by diverse cultural perspectives.

Problem

• Evaluates LLMs on classifying Iranian social norms.
• Highlights performance gaps in understanding cultural contexts.
• Introduces ISN dataset for cross-cultural AI analysis.

Innovation

• Introduced Iranian Social Norms (ISN) dataset
• Evaluated 6 LLMs on Iranian norm classification
• Highlighted linguistic challenges in Farsi prompts