📄 The Anthropology of Food: How NLP can Help us Unravel the Food cultures of the World

Summary

This position paper connects food anthropology with NLP. Anthropology treats food as a carrier of identity, memory, power and social norms. Its main method, ethnography, gives rich "thick" descriptions but is slow and hard to compare across many societies. Earlier NLP work on food has stayed mostly at a "thin" level: recipes, sentiment, trivia. The authors propose a two-part framing: food as an artifact (the object studied) and food as a lens (a way to study class, identity and historical change).

They survey 21 relevant studies and propose interdisciplinary research questions. In a case study, they test what LLMs internally encode about daily eating habits worldwide. The models group countries in ways that echo colonial and linguistic history, such as shared habits among former colonies. They largely miss globalization and migration effects, which real food-delivery data shows clearly. The paper argues that cultural benchmarks for LLMs should go beyond factual recall. It also warns about stereotyping and over-claiming what models know.

Key points

• Survey scope: The authors searched ACL, EMNLP, EACL and NAACL, plus Google Scholar and Semantic Scholar for interdisciplinary work. After excluding nutrition- and agriculture-only papers, 21 studies remained, most published in 2023–2025.
• Existing work: Prior resources such as FORK (US-centric culinary commonsense QA) and MAPO (multilingual ingredient prediction) use food mainly as a diagnostic or factual test rather than a lens on lived practice.
• Case study result: Country groupings derived from LLM representations show colonial-era cohesion. Links between clusters are sparse, so transnational food flows are underrepresented. The authors suggest these patterns may partly reflect the training data.
• Real-world comparison: Order data from DoorDash, Talabat and Zomato (2021–2024) spanned seven countries: the USA, Canada, the UAE, Oman, India, Australia and New Zealand. It shows pizza as a global staple and Indian dishes (naan, butter chicken, biryani) as widely ordered across countries.
• Positioning: The authors present NLP as a complement to anthropologists, not a replacement. LLM-derived patterns serve as testable hypotheses to check against migration, trade or ethnographic evidence. They call for reflexivity, cultural relativism and participatory design.