Arabic language models aligned with country cultural preferences
Aligning Cultural Behaviors in Arabic Language Models via Preferences
Computation and LanguageArtificial Intelligence
Summary
Arabic language models have gotten better, but they often miss the unique cultural behaviors of different Arab countries. The authors created ArabPref, a large dataset that shows preferences based on country-specific cultural traits. They tested and improved how well language models reflect these cultural differences using this data, making the models behave in ways more aligned with local culture. They also built a multiple-choice test to check cultural alignment, validated by native speakers to ensure accuracy.
What this means in practice
- •For natural language processing teams: Improve Arabic language models to generate responses that reflect country-specific cultural behaviors accurately.
- •For multilingual chatbot developers: Create chatbots that adapt their behavior and responses according to cultural preferences of different Arab nations.
Authors
Junhong Liang, Daniil Orel, Saeed Almheiri, Bilal Elbouardi, Younes Samih, Abed Alhakim Freihat, Zhuohan Xie, Momina Ahsan, Muhammad Dehan Al Kautsar, Sarfraz Ahmad, Omar El Herraoui, Mohamed Anwar, Preslav Nakov, Fajri Koto
Abstract
Arabic LLMs have improved rapidly, but aligning them with country-specific cultural behavioral preferences remains underexplored. Previous work has focused primarily on instruction-tuning for factual knowledge, overlooking nuanced cultural behaviors. We propose ArabPref, which to our knowledge is the first large-scale pan-Arab preference data set for the alignment of cultural behavior conditioned on the country. Our contribution includes two resources in English and Modern Standard Arabic: a culturally grounded preference dataset for training and evaluation, and a multiple-choice benchmark designed to assess culturally aligned behavior across nations. All test data are validated by native speakers to ensure authenticity. To optimize cultural alignment, we experiment with post-training methods across multilingual and Arabic-centric language models, evaluating performance on both generation tasks and our cultural multiple-choice benchmark. By releasing an evaluation benchmark, we provide a foundation for systematically studying and improving culturally aligned behavior in language models.