Wine chemistry and language bottleneck hypothesis: insights from large language models

Area of Science:

• Sensory science
• Computational linguistics
• Food chemistry

Background:

• Human taste perception is challenging to model computationally, especially for complex products like wine.
• Wine's elaborate terminology offers a unique case study for taste perception research.
• Large language models (LLMs) are increasingly used to analyze human-generated text, including wine descriptions.

Purpose of the Study:

• To investigate wine taste perception across chemical composition, taste receptor interactions, and linguistic terminology.
• To explore the relationship between wine's chemical properties and its descriptive language.
• To propose a conceptual framework for understanding the mediation of taste perception.

Main Methods:

• Analysis of wine chemistry (approx. 100 parameters).
• Modeling of ligand-taste receptor interactions (fewer than 10 parameters).
• Analysis of wine language terminology (approx. 25 axes) using binary oppositions and canonical correlation analysis (CCA).
• Comparison of LLM handling of wine terminology with human perception.

Main Results:

• Canonical correlation analysis revealed no statistically significant correlation between wine chemistry and wine terminology (p ≈ 0.5).
• The study introduces a conceptual framework suggesting a low-dimensional taste space (less than 10 dimensions) mediates between chemical and linguistic spaces.
• This framework provides a potential explanation for the observed poor correspondence between wine chemistry and terminology.

Conclusions:

• A low-dimensional taste space likely mediates between the high-dimensional chemical sensometabolome and the wine language space.
• The findings suggest that direct correlations between chemical composition and linguistic descriptions of wine are limited.
• Further validation with paired chemical-sensory datasets and nonlinear methods is required.