Reproduce “Same Lesson, Different Story: Cross-Lingual Reconstruction of Cultural Narratives in Large Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jory Alshaalan, Haya Albaker, Abeer Aldayel, Aljawharah Alabdullatif, Rehab Alahmadi

Researchers at King Saud University investigated how Large Language Models (LLMs) reconstruct cultural narratives across languages using proverb prompts, finding that while abstract moral meaning is largely preserved, narratives undergo systematic reconstruction in terms of agency and power dynamics. The study also revealed strong inter-model semantic convergence among LLMs and an asymmetrical redistribution of social positioning under cross-lingual conditions.

Key insights

  • LLMs primarily encode proverbs as abstract moral lessons, with cross-lingual prompts acting as a mechanism for narrative reconstruction around these shared abstract concepts, rather than preserving culturally specific narrative templates.
  • The study extended the 'Artificial Hivemind effect' to multilingual cultural narratives, observing strong inter-model semantic convergence among different LLMs, suggesting they arrive at similar semantic interpretations of proverb meanings.
  • Current multilingual narrative evaluations relying solely on semantic similarity metrics may overestimate cultural preservation, as these metrics can overlook significant, culturally meaningful variations in narrative expression, agency, and power relations despite preserved core meaning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.