Elamur Research⌕
☰
home / papers / 2024 / CULTURE-GEN: Revealing Global Cultural Perception in Language Models through Natural Language Prompting

CULTURE-GEN: Revealing Global Cultural Perception in Language Models through Natural Language Prompting

Paper · 2024 · Huihan Li, Liwei Jiang, Jena D. Hwang et al.

Summary

The paper asks whether large language models know enough about the world's cultures, and whether they portray them fairly when generating free text. The authors built CULTURE-GEN, a dataset of model outputs for 110 countries and regions (the set covered by the World Values Survey) spanning 8 topics: preferred music, musical instruments, workout routines, TV shows or films, dishes, pictures, statues, and garments. They prompted gpt-4, llama2-13b and mistral-7b with short neighbor-themed templates, both with and without a stated nationality. They then linked the entities that appeared ("culture symbols") to specific cultures using an unsupervised ranking based on sentence probability, so no human labels or external knowledge bases were needed. The analysis shows that models "mark" non-default cultures with distinctive wording. Symbol diversity varies widely by region. Outputs that name no culture lean toward Western European, English-speaking and Nordic cultures. The framework can be applied to any model or culture, including low-resource settings, which supports research on cultural fairness and knowledge in LLMs.

Key points

  • Coverage: 110 countries and regions × 8 topics, with generations from gpt-4, llama2-13b and mistral-7b. Code and data are published at github.com/huihanlhh/Culture-Gen.
  • Cultural markedness: for Asian, Eastern European and African-Islamic cultures, models often put the word "traditional" before generated items. This happens in about 30% of llama2-13b outputs and nearly all gpt-4 outputs. Models also add explanations in parentheses, which implies readers would not recognize these items.
  • Uneven knowledge: symbol diversity is lower for some regions across every topic and model. It correlates with how often a culture name co-occurs with topic keywords in RedPajama (Kendall-τ up to 0.35). RedPajama is not the actual training data of these models, so this link is a conjecture.
  • Limitations: the aligned variants (llama2-13b-chat, mistral-7b-instruct) gave no usable answers because of their safeguards. The authors suggest OLMo, whose training data is fully open, for future attribution studies.
  • Proposed mitigation: broaden pretraining and instruction data to include more cultures, and pursue pluralistic alignment.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of WashingtonSeattle · United States353 works · 200,043 citations🔬Allen Institute for Artificial IntelligenceSeattle · United States275 works · 97,634 citations🎓University of Southern CaliforniaLos Angeles · United States162 works · 35,836 citations🔬Allen InstituteSeattle · United States7 works · 4,351 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksSafety Research · 543 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Huihan Li, Liwei Jiang, Jena D. Hwang, Hyunwoo Kim, Sebastin Santy, Taylor Sorensen, Bill Yuchen Lin, Nouha Dziri, Xiang Ren and Yejin Choi. 2024. CULTURE-GEN: Revealing Global Cultural Perception in Language Models through Natural Language Prompting. arXiv:2404.10199.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2404.10199 · arXiv 2404.10199

@misc{li2024culture,
  title = {{CULTURE-GEN: Revealing Global Cultural Perception in Language Models through Natural Language Prompting}},
  author = {Huihan Li, Liwei Jiang, Jena D. Hwang, Hyunwoo Kim, Sebastin Santy, Taylor Sorensen, Bill Yuchen Lin, Nouha Dziri, Xiang Ren and Yejin Choi},
  year = {2024},
  doi = {10.48550/arxiv.2404.10199},
  eprint = {2404.10199},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2404.10199}
}

References

36 references, 16 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Muhammad Farid Adilazuarda, Sagnik Mukherjee, Pradhyumna Lavania, Siddhant Singh, Ashutosh Dwivedi, Alham Fikri Aji, Jacki O’Neill, Ashutosh Modi, and Monojit Choudhury. Towards measuring and modeling” culture” in llms: A survey. arXiv preprint arXiv:2403.15412, 2024.
  2. Badr AlKhamissi, Muhammad ElNokrashy, Mai AlKhamissi, and Mona Diab. Investigating cultural alignment of large language models. arXiv preprint arXiv:2402.13231, 2024.
  3. Myra Cheng, Esin Durmus, and Dan Jurafsky. Marked personas: Using natural language prompts to measure stereotypes in language models. arXiv preprint arXiv:2305.18189, 2023.
  4. Together Computer. Redpajama: an open dataset for training large language models, October 2023. URL https://github.com/togethercomputer/RedPajama-Data. ↗
  5. Nicholas Deas, Jessica A. Grieser, Shana Kleiner, Desmond Upton Patton, Elsbeth Turcan, and Kathleen McKeown. Evaluation of african american language bias in natural language generation. ArXiv, abs/2305.14291, 2023. URL https://api.semanticscholar.org/CorpusID:258841278. ↗
  6. Esin Durmus, Karina Nyugen, Thomas Liao, Nicholas Schiefer, Amanda Askell, Anton Bakhtin, Carol Chen, Zac Hatfield-Dodds, Danny Hernandez, Nicholas Joseph, Liane Lovitt, Sam McCandlish, Orowa Sikder, Alex Tamkin, Janel Thamkul, Jared Kaplan, Jack Clark, and Deep Ganguli. Towards measuring the repres… — Towards Measuring the Representation of Subjective Global Opinions in Language Models
  7. Ashutosh Dwivedi, Pradhyumna Lavania, and Ashutosh Modi. Eticor: Corpus for analyzing llms for etiquettes. In Conference on Empirical Methods in Natural Language Processing, 2023. find
  8. Yanai Elazar, Akshita Bhagia, Ian Magnusson, Abhilasha Ravichander, Dustin Schwenk, Alane Suhr, Pete Walsh, Dirk Groeneveld, Luca Soldaini, Sameer Singh, et al. What’s in my big data? arXiv preprint arXiv:2310.20707, 2023. find
  9. Yi Ren Fung, Ruining Zhao, Jae Doo, Chenkai Sun, and Heng Ji. Massively multi-cultural knowledge acquisition & lm benchmarking. ArXiv, abs/2402.09369, 2024. URL https://api.semanticscholar.org/CorpusID:267657749. ↗
  10. Clifford Geertz. The interpretation of cultures, volume 5019. Basic books, 1973. find
  11. Dirk Groeneveld, Iz Beltagy, Pete Walsh, Akshita Bhagia, Rodney Kinney, Oyvind Tafjord, Ananya Harsh Jha, Hamish Ivison, Ian Magnusson, Yizhong Wang, et al. Olmo: Accelerating the science of language models. arXiv preprint arXiv:2402.00838, 2024. find
  12. Christian W Haerpfer and Kseniya Kizilova. The world values survey. The Wiley-Blackwell Encyclopedia of Globalization, pp. 1–5, 2012. find
  13. Jing Huang and Diyi Yang. Culturally aware natural language inference. In Findings of the Association for Computational Linguistics: EMNLP 2023, pp. 7591–7609, 2023.
  14. Amr Keleg and Walid Magdy. Dlama: A framework for curating culturally diverse facts for probing the knowledge of pretrained language models. arXiv preprint arXiv:2306.05076, 2023. find
  15. Khyati Khandelwal, Manuel Tonneau, Andrew M. Bean, Hannah Rose Kirk, and Scott A. Hale. Casteist but not racist? quantifying disparities in large language model bias between india and the west. ArXiv, abs/2309.08573, 2023. — [arXiv:2309.08573](https://arxiv.org/abs/2309.08573) find
  16. Cheng Li, Mengzhou Chen, Jindong Wang, Sunayana Sitaram, and Xing Xie. Culturellm: Incorporating cultural differences into large language models. ArXiv, abs/2402.10946, 2024. URL https://api.semanticscholar.org/CorpusID:267750997. ↗
  17. Weicheng Ma, Brian Chiang, Tong Wu, Lili Wang, and Soroush Vosoughi. Intersectional stereotypes in large language models: Dataset and analysis. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Findings of the Association for Computational Linguistics: EMNLP 2023, pp. 8589–8597, Singapore, Decemb… find
  18. Anjishnu Mukherjee, Chahat Raj, Ziwei Zhu, and Antonios Anastasopoulos. Global voices, local biases: Socio-cultural prejudices across languages. In Conference on Empirical Methods in Natural Language Processing, 2023. find
  19. Tarek Naous, Michael J Ryan, and Wei Xu. Having beer after prayer? measuring cultural bias in large language models. arXiv preprint arXiv:2305.14456, 2023.
  20. Tuan-Phong Nguyen, Simon Razniewski, Aparna Varde, and Gerhard Weikum. Extracting cultural commonsense knowledge at scale. In Proceedings of the ACM Web Conference 2023, WWW ’23. find
  21. ACM, April 2023a. doi: 10.1145/3543507.3583535. URL http://dx.doi.org/ 10.1145/3543507.3583535. find
  22. Tuan-Phong Nguyen, Simon Razniewski, Aparna Varde, and Gerhard Weikum. Extracting cultural commonsense knowledge at scale. In Proceedings of the ACM Web Conference 2023, pp. 1907–1917, 2023b. find
  23. Shramay Palta and Rachel Rudinger. Fork: A bite-sized test set for probing culinary cultural biases in commonsense reasoning models. In Findings of the Association for Computational Linguistics: ACL 2023, pp. 9952–9962, 2023.
  24. Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019.
  25. Aida Ramezani and Yang Xu. Knowledge of cultural moral norms in large language models. arXiv preprint arXiv:2306.01857, 2023.
  26. Michael J Ryan, William Held, and Diyi Yang. Unintended impacts of llm alignment on global representation. arXiv preprint arXiv:2402.15018, 2024.
  27. Patrick Schober, Christa Boer, and Lothar A Schwarte. Correlation coefficients: appropriate use and interpretation. Anesthesia & analgesia, 126(5):1763–1768, 2018. find
  28. Luca Soldaini, Rodney Kinney, Akshita Bhagia, Dustin Schwenk, David Atkinson, Russell Authur, Ben Bogin, Khyathi Chandu, Jennifer Dumas, Yanai Elazar, et al. Dolma: An open corpus of three trillion tokens for language model pretraining research. arXiv preprint arXiv:2402.00159, 2024. find
  29. Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Rytting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, et al. A roadmap to pluralistic alignment. arXiv preprint arXiv:2402.05070, 2024.
  30. Raphael Tang, Xinyu Crystina Zhang, Jimmy J. Lin, and Ferhan Ture. What do llamas really think? revealing preference biases in language model representations. ArXiv, abs/2311.18812, 2023. — [arXiv:2311.18812](https://arxiv.org/abs/2311.18812) find
  31. Yan Tao, Olga Viberg, Ryan S. Baker, and Rene F. Kizilcec. Auditing and mitigating cultural bias in llms, 2023. find
  32. Linda R Waugh. Marked and unmarked: A choice between unequals in semiotic structure. 1982. find
  33. Peter West, Chandra Bhagavatula, Jack Hessel, Jena Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. In Proceedings of the 2022 Conference of the North American Chapter of the Association f…
  34. Haryo Akbarianto Wibowo, Erland Hilman Fuadi, Made Nindyatama Nityasya, Radityo Eko Prasojo, and Alham Fikri Aji. Copal-id: Indonesian language reasoning with local culture and nuances. ArXiv, abs/2311.01012, 2023. URL https://api.semanticscholar.org/ CorpusID:264935209. ↗
  35. Da Yin, Hritik Bansal, Masoud Monajatipoor, Liunian Harold Li, and Kai-Wei Chang. Geomlama: Geo-diverse commonsense probing on multilingual pre-trained language models. arXiv preprint arXiv:2205.12247, 2022. find
  36. Caleb Ziems, Jane Dwivedi-Yu, Yi-Chia Wang, Alon Halevy, and Diyi Yang. Normbank: A knowledge bank of situational social norms. arXiv preprint arXiv:2305.17008, 2023. \#\# Limitations

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/li-2024-culture-gen-revealing-global-cultural/