June 05, 2004

What do wine tasting notes communicate?

Following up on our earlier discussion of winetalk, I did a PsycInfo search on "wine". I didn't come up with much, to my surprise, but there were a few interesting things. One was Frédéric Brochet and Denis Dubourdieu, "Wine Descriptive Language Supports Cognitive Specificity of Chemical Senses", Brain and Language, Volume 77, Issue 2 , May 2001, Pages 187-196. The authors analyzed several large corpora of tasting notes, in both French and English, using a method that factored the vocabulary into "classes" based on co-occurrence patterns, and concluded that

(1) Class number and organization are different among experts so that each expert has his own discourse strategy. (2) Wine language is based on prototypes and not on detailed analytical description. (3) Prototypes include not only sensory but also idealistic and hedonistic information.

Most striking for me was how strongly Brochet and Dubourdieu claim a sort of "indeterminacy of radical translation" for winetalk, based not only on the idea that different individuals connect words to experience differently, but also on the idea that the underlying sensory apparatus is not really shared:

...the number of classes and their nature are broadly different among subjects. Lawless (1984) demonstrated that experts were not significantly able to recognize wines based on a description given by others, even when they were experts. Given the small number of terms common to the several authors studied here, it seems clear that wine descriptions are deeply individual and that they make sense mainly to the taster him- or herself. The results confirm that a consensual language for the description of wine does not exist and that only ‘‘individual’’ languages appear in published works. The analysis of the compiled corpus showed only convergence through color. Category divergence was confirmed by Berglund (1973), who demonstrated with basic odorants that flavor categories do not exist at an interindividual level but that they were accurate for individuals. These differences in language used to describe taste sensations may arise from genetic differences among individuals (Buck, 1993). Olfactory receptors may be encoded by a very large multigene family, so the probability that two individuals will possess the same receptors is very low. This diversity is enhanced by the diversity of learning associated to chemical senses. Individuals do not learn to designate odors in the same way so that a same sensation, a same signal, will be categorized differently, which will lead to different denomination, i.e., different languages. This shows that communication of wine sensory properties is not accurate (Lehrer, 1975).

It's obvious to start with that tasting notes don't tell us anything about the wine that was tasted, except insofar as we can make inferences from the linguistic reactions of the person that did the tasting. Brochet and Dubourdieu are arguing that such inferences must be very indirect ones at best, and that in fact the descriptions convey almost nothing about the wine, but only something about the taster's descriptions of his or her reactions. These in turn are doubly decoupled from our own, first by the different choice and use of terminology, and second by a different underlying biology of sensory receptors. Unless, perhaps, we allow our own intrinsic perceptions to be overshadowed by the influence of what we've read.

I'm reluctant to accept this argument. For one thing, I suspect that you could use the same techniques of linguistic analysis to make a similar argument with respect to nearly any sort of linguistic description of human experience, perceptual or otherwise. It's well known that descriptions are reliably inconsistent across individuals, even experts from the same subculture (see e.g. Furnas et al., "Statistical semantics: Analysis of the potential performance of key-word information systems", Bell System Technical Journal. Vol 62(6, Pt 3), Jul-Aug 1983, pp. 1753-1806). This doesn't necessarily mean that the underlying perceptions and categorizations are incommensurable or even that the descriptions are not communicatively effective.

Also, the idea of a strict logical distinction between sensation and expectation probably makes even less sense in the case of food and drink than it does elsewhere. From what little I know of the psychophysics, it seems likely that the integration of perception and expectation takes place at all levels of processing, and at time scales ranging from recent individual experience to evolutionary development across generations. On this view, we might value winetalk for the same reason that we value a lot of other writing: not because the writer's perceptions are the same as our own, but because they're different.

It's still a rather bleak view of communication: either the text we read is just pumping some energy into resonances of our own neuronal system, in a way that is only accidentally connected to the pattern originally expressed; or the text overshadows our own reactions, affecting them only by partly replacing them with an echo of the expert's views, disconnected from our own sensory experience. It's nicer to think of a message being composed, sent, received, understood, evaluated, and acted on.

With respect to the author's hypothesis that "[o]lfactory receptors may be encoded by a very large multigene family, so the probability that two individuals will possess the same receptors is very low", I haven't been able to find any concrete information one way or the other. However, this research finds that "humans have accumulated mutations in odor receptor genes four times faster than have the other primates", and that "[o]lfactory receptors are the largest 'superfamily' of genes in mammals, with over 1,000 different genes. But in humans, more than 60 percent of these genes no longer work." It's not clear (to me at least) how much population variation in this superfamily exists among the current human population. However, from what I can see, the situation might well be as the authors suggest: there are a thousand or so featural dimensions that are available in principle; there is a high mutation rate; there is presumably relatively little selective pressure on the details of the system, as long as adequate overall identification and discrimination can be maintained; and apparently in recent times there has been little selective pressure in the hominid line even to maintain the acuity of the system as a whole.

Another cross-cultural note: Brochet and Dubourdieu use the Max Reinert's Alceste software package, which performs a sort of factor analysis of term-by-document matrices, and seems to be fairly widely used by researchers in France. Rienert describes his approach here, in a way that is somehow very French, despite the pervasive influence of Anglophones such as Harris, Wittgenstein and Peirce. Alceste seems to have a lot in common with Latent Semantic Analysis, originally proposed here, and widely used in the Anglophone world -- but neither tradition seems to be aware of the existence of the other one. This is not strictly a matter of country and language, since there is (for instance) an LSA site at Grenoble, but there is still an interesting cultural divergence here.

Posted by Mark Liberman at June 5, 2004 07:42 AM