Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings
Summary
This arXiv paper introduces Epicure, a family of skip-gram embeddings for food ingredients trained on a multilingual recipe corpus. It standardizes 4.14 million recipes across seven languages to 1,790 canonical ingredients, then builds a large ingredient-ingredient co-occurrence graph and a typed compound graph to explore relationships between ingredients and chemicals. The work compares three Metapath2Vec variants to map ingredients within a spectrum from chemistry to recipe context, enabling cross-lingual culinary AI research.