The Problem: "Campfire" vs. "Smoke"
Human language is messy and creative. If you drink a peaty whisky, you might say it tastes like "smoke." Another person might describe the exact same glass as tasting like a "campfire," "ash," or even a "medicinal bandage."
To a standard computer search, "Campfire" and "Smoke" are completely different words. This is the challenge of flavor analysis. A simple keyword search misses the nuance of the experience.
Our goal was to build a system that understands meaning, not just spelling.
We wanted our software to recognize that a user mentioning "campfire" is actually contributing to the "Smoky" flavor profile, even without an exact keyword match.
The Process
To transform thousands of fragmented reviews into a structured Tasting Wheel, our engine follows four key steps:
Collection
Ethically harvesting reviews and tasting notes from diverse digital sources.
Cleaning
Removing "noise" (like shipping feedback) to isolate purely sensory descriptors.
Vectorization
Converting descriptors into mathematical coordinates using Natural Language Processing.
Visualization
Mapping these coordinates onto the Flavor Wheel to generate a visual fingerprint.
The Engine: Word Vectors
How does the AI understand that "Lemon" and "Lime" are related? We utilize Word Embeddings, powered by the spaCy NLP guide.
Imagine a vast 3D map where every word has a specific GPS coordinate.
- Words with similar meanings, like "Apple" and "Pear," live in the same neighborhood.
- Unrelated concepts, like "Apple" and "Leather," are located in entirely different cities.
By measuring the Cosine Similarity between these coordinates, we can mathematically prove that "Toffee" is a 90% match for "Caramel." This ensures accuracy even when the specific category name is never mentioned.
Visualizing the AI's Logic
The heatmap below illustrates how our algorithm clusters flavors. It visualizes the statistical correlations within the Tasting Intelligence Master Profile.
Adjust the Similarity Threshold:
The dense clusters in the heatmap confirm that the algorithm effectively identifies the sensory "fingerprint" of each flavor category.