FlavorGraph embeddings reveal culinary tacit knowledge

The recent computer science study titled 'Epicure: Multidimensional Flavor Structure in Food Ingredient Embeddings' reveals that FlavorGraph's 300-dimensional ingredient embeddings capture chefs' implicit understanding of flavor, texture, and cultural identity. Utilizing a pipeline enhanced by a large language model, the researchers refined 6,653 raw ingredients into 1,032 standardized entries, thereby enhancing the recoverable structure. They discovered a minimum of fifteen distinct dimensions that can be classified independently, covering aspects such as taste, texture, geography, food processing, and culture.