A Generative Grammar of Cooking
Ganesh Bagler
arXiv:2211.09059
Read original publicationThe first formal generative grammar for cooking, treating recipes as combinatorial systems of culinary concepts (cubits) assembled by syntactic rules — making the creative act of cooking computable.
These are our reading notes and analysis. The original work belongs to its authors and publisher.
Core Thesis
Cooking, like language, follows a rule-based structure. A small set of culinary concepts — called cubits (culinary bits) — combined through three syntactic rules can generate an astronomical number of valid recipes, explaining creativity in cooking through "the infinite use of finite media."
Cubits: The Parts of Cooking
Analogous to parts of speech, cubits are the building blocks:
- I — Ingredient name (18,599 in RecipeDB)
- Q — Quantity-and-unit (71 terms)
- F — Form: chopped, sliced, diced (1,299 forms)
- P — Processing action: boil, fry, sauté (270 actions)
- T — Process descriptor: slowly, for 10 min (100+ descriptors)
- U — Utensil: pot, pan, skillet (69 utensils)
The Three Rules
- Ingredient Phrase (IP): Q + F + I → "100 gm of sliced potatoes"
- Processing Phrase (PP): P + IP(s) → "Boil the sliced potatoes"
- Culinary Sentence (S): PP + U → "Boil the sliced potatoes in a pot"
The magic lies in recursion: an ingredient phrase can itself be a culinary sentence ("Fry onion and boiled potato in a pan"), creating unlimited complexity from finite elements.
Data Foundation
Built on RecipeDB: 118,000+ recipes from 74 countries, 6 continents, annotated using Named Entity Recognition algorithms. The combinatorial space of valid culinary sentences is astronomical.
Why It Matters
This framework transforms cooking from art to computable system. Recipe generation models informed by this grammar can ensure structural validity, filter for dietary constraints, optimize carbon footprint, and respect culinary fingerprints of target cuisines. It is the theoretical foundation for making food knowledge computable.