Reproduce “LLMs as Cultural Archives: Cultural Commonsense Knowledge Graph Extraction” 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

Junior Cedric Tonga, Chen Cecilia Liu, Iryna Gurevych, Fajri Koto

Researchers developed an iterative, prompt-based framework to extract multi-step, inferential cultural commonsense knowledge from LLMs, structuring it into a Cultural Commonsense Knowledge Graph (CCKG). This CCKG consistently enhanced the cultural reasoning and generation abilities of smaller language models, improving performance on benchmarks like IndoCulture and ArabCulture.

Key insights

  • LLMs encode multi-step cultural commonsense knowledge, but extracting coherent inferential chains is more effective when prompting in English, even for non-English cultures.
  • Explicitly structuring cultural knowledge into a CCKG provides an effective method to augment and improve the cultural reasoning and generation abilities of smaller LLMs.
  • Cultural commonsense reasoning is less amenable to Chain-of-Thought (CoT) prompting, benefiting more from direct, in-context cultural knowledge examples derived from the CCKG.

Summary by alphaXiv.

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