Scaling Cultural Resources for Improving Generative Models

2025-10-29

AI Summary

Generative models exhibit significant performance degradation in multilingual and cross-cultural settings, primarily due to the scarcity of high-quality, culturally salient, and globally representative multilingual data resources. Method: We propose a reproducible and scalable multi-path data construction framework that integrates multilingual web harvesting, automated cultural salience filtering, and community-driven data contribution. This framework systematically expands culturally aware training and evaluation datasets. Technically, we design novel cultural bias detection metrics and a multidimensional evaluation protocol—enabling, for the first time, quantitative assessment of generative models’ cross-cultural applicability. Results: Experiments demonstrate that our framework effectively uncovers cultural biases across critical dimensions—including value expression, regional commonsense knowledge, and social norms. It establishes an extensible data infrastructure and standardized benchmark for fairness-aware model optimization and global deployment.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Large Multimodal Models (LMMs)
• Philosophy and Ethics of AI: Bias, Fairness & Equity

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Generative models are known to have reduced performance in different global cultural contexts and languages. While continual data updates have been commonly conducted to improve overall model performance, bolstering and evaluating this cross-cultural competence of generative AI models requires data resources to be intentionally expanded to include global contexts and languages. In this work, we construct a repeatable, scalable, multi-pronged pipeline to collect and contribute culturally salient, multilingual data. We posit that such data can assess the state of the global applicability of our models and thus, in turn, help identify and improve upon cross-cultural gaps.

Problem

• Improving generative models' performance across global cultures
• Addressing reduced model efficacy in diverse cultural contexts
• Developing scalable data pipelines for cross-cultural AI evaluation

Innovation

• Constructed scalable pipeline for cultural data collection
• Collected multilingual culturally salient data resources
• Assessed and improved model cross-cultural applicability gaps