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

Despite the development of novel tools for automatic food quality control, human taste perception is difficult to represent on computers in textual or conceptual space forms. Wine is a particularly interesting object of taste studies because of its elaborate terminology. We investigate taste perception at three different levels: the chemical composition of wine (∼100 parameters), the interaction of ligands with taste receptors (<10 parameters), and wine language terminology (∼25 axes). Wine terminology is studied via the analysis of binary oppositions. Given that wine terminology is often perceived as subjective, it is particularly intriguing to study how large language models (LLMs) handle it in comparison with humans. While the internal semantic space of LLMs may involve thousands of dimensions, it is trained on human wine descriptions. Canonical correlation analysis (CCA) showed no statistically significant correlation between wine chemistry and wine terminology (p ≈ 0.5). While this null result does not rule out the bottleneck hypothesis, it provides no confirmatory evidence. The primary contribution of this work is the conceptual framework proposing that a low-dimensional taste space (less than 10 dimensions) mediates between the high-dimensional chemical sensometabolome (∼100 dimensions) and the wine language space (∼25 dimensions). This framework offers a plausible explanation for why wine chemistry and wine terminology show poor correspondence, but rigorous validation requires future studies with paired chemical-sensory datasets and nonlinear methods.