Cultura-Arabica: Probing and Enhancing Arabic Cultural Awareness in Large Language Models via LoRA
Year: 2025
Type: conference-paper
Abstract: Large Language Models (LLMs) have demonstrated impressive multilingual capabilities; however, their reasoning often reflects Englishcentric perspectives, which can limit accuracy in culture-specific contexts.Arabic, with its diverse dialects, rich historical heritage, and complex socio-cultural norms, presents a particularly challenging setting for such evaluation.To address this gap, we participated in the PalmX 2025 shared task, which benchmarks cultural reasoning in Arabic through multiple-choice questions covering traditions, social norms, history, geography, arts, and dialectal expressions.By applying parameterefficient adaptation and culturally informed prompt formatting, we aligned model outputs with both linguistic correctness and cultural relevance.Our approach achieved an accuracy of 71.65%, securing second place overall and closely matching the top system.These results demonstrate that targeted adaptation can significantly enhance cultural reasoning in LLMs, paving the way for more culturally aware Artificial Intelligence.
Authors: Pulkit Chatwal, Santosh Kumar Mishra
Institution: Rajiv Gandhi Institute of Petroleum Technology
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037787