Mapping Global Cuisine: Ingredient Embeddings from Millions of Recipes

This research demonstrates how to encode culinary knowledge into high-dimensional embeddings that preserve both recipe context and chemical flavor compounds, enabling better understanding of ingredients and cooking relationships across languages. The 2-megabyte model efficiently captures patterns from 4.14 million recipes across seven languages, offering potential applications in recipe recommendation, food innovation, and cross-cultural culinary analysis. As large language models increasingly handle real-world problem domains, this work shows how to meaningfully represent specialized knowledge in compact, interpretable form.