💬 NLP

Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings

The paper presents Epicure, a family of multilingual ingredient embeddings trained on 4.14 million recipes that utilize distinct random-walk schemas to navigate a hybrid graph of ingredient co-occurrences and chemical flavor compounds, thereby positioning models at varying points along a spectrum between recipe context and chemical structure.

Original authors: Jakub Radzikowski, Josef Chen

Published 2026-05-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jakub Radzikowski, Josef Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a giant, invisible library where every single ingredient in the world is a book. The goal of this paper is to build a map of this library so that a computer can understand how ingredients relate to one another.

The authors, from a company called KAIKAKU.AI, created a new family of maps called Epicure. They built three slightly different versions of this map to see which way of organizing the library works best.

Here is the simple breakdown of what they did and what they found:

1. The Problem with the Old Map

Before this, there was a famous map called "FlavorGraph." It was great, but it had a few flaws:

It was mostly written in English, missing many global recipes.
It mixed two types of information (chemical smells and recipe habits) in a fixed way that couldn't be changed.
It had messy ingredient names (like "sourdough roll" vs. just "bread").

2. Building the New Library (The Data)

The authors gathered 4.14 million recipes from 11 different sources, covering seven languages (including English, Chinese, Russian, Spanish, and more).

The Cleanup: They used AI to clean up the messy ingredient lists. They turned 200,000 different ways of writing "onion" or "garlic" into a clean, standardized list of 1,790 ingredients.
The Chemical Link: They also connected these ingredients to a database of chemical compounds (the molecules that make food taste sweet, spicy, or citrusy).

3. The Three "Sibling" Maps

The authors trained three different models (maps) using the exact same data and computer settings. The only difference was how they walked through the library to learn the connections. Think of them as three different tour guides:

Guide A (Cooc): This guide only looks at recipes. If "tomato" and "basil" appear together in a recipe, they are friends. This guide learns based on cooking habits.
Guide B (Chem): This guide only looks at chemistry. If "tomato" and "basil" share the same chemical molecules, they are friends. This guide learns based on flavor profiles.
Guide C (Core): This guide is a mix. It looks at recipes and chemistry, but it walks through the recipe section more often. It tries to blend both worlds.

4. What the Maps Can Do

Once the maps were built, the authors tested them to see if they could answer culinary questions.

A. Finding Neighbors (The "What goes with this?" test)
If you ask the map, "What goes with Chicken?"

Guide A (Cooc) says: "Garlic, onion, and carrots." (Things you cook with chicken).
Guide B (Chem) says: "Beef and pork." (Things that taste similar to chicken chemically).
Guide C (Core) says: "Pork and beef, but also chicken broth." (A mix of both).

B. The "Magic Compass" (SLERP)
The most exciting part is a tool called SLERP. Imagine you have a compass needle pointing at "Chicken." You can rotate that needle toward a new direction, like "South Asian" or "Sweet Dessert."

If you rotate "Chicken" toward South Asian, the map instantly suggests: Curry leaf, fenugreek, and lentils.
If you rotate "Chocolate" toward Sweet Baking, it suggests: Vanilla, cocoa powder, and coffee.
The user can control how far to rotate. A small turn keeps you close to the original ingredient; a big turn takes you deep into a new flavor world.

5. The Big Discovery

The authors found that the "Chemistry" guide (Guide B) was actually the best at finding clear, distinct directions for things like "spicy," "fatty," or "Italian."

The Analogy: Imagine the library is a room.
- Guide A (Recipes) spreads the books out evenly across the whole room.
- Guide B (Chemistry) clusters the books tightly into specific corners.
- Surprisingly, the tightly clustered map (Guide B) made it easier to draw straight lines between concepts (like "Italian" vs. "Chinese"). It turned out that understanding the chemistry of food actually helped the computer understand the culture of food better than just looking at recipes alone.

Summary

The paper introduces Epicure, a new set of tools that helps computers understand food. It proves that by mixing recipe habits with chemical science, and by giving users a "knob" to control how much of each they want, we can build a digital kitchen assistant that can:

Suggest ingredients that go well together.
Help you explore new cuisines by rotating a "flavor dial."
Understand that "Chicken" can be a "Western dinner" or an "Indian curry" depending on which way you turn the dial.

The authors did not release the code yet, but they showed that this new way of mapping food works better than previous methods at finding clear, useful patterns.

Technical Summary: Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings

Problem Statement
Existing computational gastronomy approaches, such as FlavorGraph, have established comprehensive food embeddings by combining chemical data (FlavorDB) with recipe co-occurrence (Recipe1M+). However, prior work identified three critical limitations in these models: (1) they rely on single-language, English-centric corpora; (2) they fuse chemical and recipe-context signals into a fixed inductive bias, preventing control over the balance between chemistry and culinary context; and (3) they often utilize scattered vocabularies containing preparation details and non-food items. The authors posit that a useful ingredient embedding must not only encode culinary knowledge but also expose the underlying geometry as a controllable design axis, enabling tools that can navigate along supervised semantic directions or emergent cultural modes.

Methodology
The authors present Epicure, a family of three sibling skip-gram ingredient embeddings trained from scratch to address the aforementioned limitations. The methodology proceeds through five stages:

Corpus Aggregation: The authors aggregated 4.14 million recipes from 11 sources across seven languages (English, Chinese, Russian, Vietnamese, Spanish, Turkish, Indonesian, German, and Indian-English). Non-English terms were machine-translated to English using deterministic decoding.
Canonical Vocabulary Normalization: An LLM-augmented pipeline (using Claude Opus and Gemini Embeddings) normalized raw ingredient strings into a shared vocabulary of 1,790 canonical entries. This process filtered out non-food items and preparation modifiers, ensuring a clean node set.
Graph Construction: Three graph variants were constructed sharing the same 1,790 ingredient nodes and 203,508 NPMI-weighted ingredient-ingredient (I-I) co-occurrence edges.
- Cooc Graph: Contains only I-I edges.
- Core/Chem Graph: Adds 2,247 typed FlavorDB compound nodes connected via 80,019 typed ingredient-compound (I-C) edges. Compounds are tagged with 15 flavor categories (e.g., citrus, fatty, spicy).
Model Training: Three Metapath2Vec variants were trained with identical hyperparameters (300-dim embeddings, 20 epochs, negative sampling) but differing random-walk schemas:
- Epicure-Cooc: Walks only the pure I-I co-occurrence graph.
- Epicure-Core: Walks the typed-compound graph but injects pure I-I walks (10x repetition) to blend chemical and recipe-context signals.
- Epicure-Chem: Walks the typed-compound graph with no injected I-I walks, relying solely on compound-mediated paths.
Evaluation: The embeddings were evaluated using supervised direction probes (27 continuous sensory/nutrient directions and 8 cuisine macro-regions) and unsupervised factor analysis (FastICA decomposition followed by Gaussian Mixture Model partitioning).

Key Contributions

Controllable Inductive Bias: The paper introduces a design axis where the "chemistry-vs-recipe-context" spectrum is controlled at the walk schema level rather than being a fixed property of the model. This allows for three distinct embeddings (Cooc, Core, Chem) from a single experimental setup.
Multilingual, Canonical Embeddings: The work provides the first multilingual food embedding family trained on a normalized, LLM-curated vocabulary, moving beyond English-centric, unstructured ingredient lists.
Dual Operator Families: The authors demonstrate that a single 300-D space supports two complementary navigation operators:
1. Nearest-Neighbour Pairings: Retrieval of top-K neighbors and closest emergent modes.
2. SLERP Direction Arithmetic: Continuous rotation of a seed vector toward supervised pole vectors (e.g., "South Asian") or emergent factor-mode poles, controlled by an angle $\theta$ .

Results

Linear Recoverability: All three models linearly recover supervised probes. The Chem model consistently outperforms Cooc and Core in direction quality (mean Cohen's $d$ for cuisine separability: 3.07 for Chem vs. 2.43 for Cooc), suggesting that chemistry-mediated walks act as a structural prior that enhances the recoverability of concepts even those not directly encoded in the walk schema (e.g., basic tastes and USDA macronutrients).
Isotropy and Geometry: The models exhibit distinct geometric properties. Cooc and Chem are highly isotropic (Participation Ratio $\approx$ 174–183), while Core is more concentrated (PR = 94.2) due to the injected I-I walks. Despite this concentration, Core achieves the tightest emergent modes (coherence 0.833 vs. 0.611/0.703), indicating that the concentration is a deliberate design lever rather than a geometric defect.
Emergent Modes: Unsupervised decomposition recovered 20 interpretable factors per model, partitioned into 150–200 named culinary modes (e.g., "Sweet baking and dessert ingredients," "South Asian whole spice blends"). These modes show high coherence (5–6 $\times$ above random baselines) and represent tight, named neighborhoods.
Navigation Behavior:
- Pairings: Cooc tends to retrieve recipe companions (e.g., garlic for chicken), while Chem retrieves flavor-profile peers (e.g., beef for chicken).
- SLERP: Rotating seeds toward supervised or emergent poles yields predictable, label-aligned results. For instance, rotating "rice" toward "South Asian" retrieves curry leaf and lentils; rotating "chocolate" toward "sweet baking" retrieves baking powder and vanilla (or matcha in Chem, reflecting its chemistry bias).

Significance
The paper argues that the next step in computational gastronomy is not merely creating larger embeddings, but exposing the operators that act upon them. By treating the walk schema as a named, controllable axis, Epicure transforms a static vector space into a navigable interface. The three siblings allow users to choose between recipe-context neighbors (Cooc) and flavor-profile peers (Chem) on the same 300-D space. Furthermore, the combination of nearest-neighbor lookups and continuous SLERP rotation supports chef-facing tools that can "rotate, blend, or retrieve" along both supervised semantic directions and culturally coherent emergent modes. The authors conclude that this methodological shift—treating the walk schema as the experimental variable—applies to any future fusion of culinary signals, including nutrition, image, or sensory descriptors.