Wine Talk

Blind tasting is one of the favorite games among wine enthusiasts. A mysterious bottle of wine wrapped in opaque papers or pouches, poured into a delicate hand-blown crystal glass, revealing its clear deep crimson color. Rich and opulent, pure and high-toned bouquet jumped out of the glass. Beautiful wild red cherries, vigor, fresh, and reverberant. A hint of roses and violets crackling with a touch of white pepper. Stony satiny tannins perfectly balanced with its tension and energy. What is it, you wonder in ecstasy, swirling the elixir gently to take in all its intricate aromatics. Familiar memories conjured up in your mind. You were wondering within a deep forest, thick with pristine primeval growths. As the humid scent of life wafts from the moss-covered trees. You walked towards the heart of forest in search of solace. Suddenly you noticed a ray of light. You smelt flowers and red fruits that seem out of place in these woods. Unexpectedly the forest opened to a small clear spring, pouring forth like a miracle, like an oasis in the desert. The restoring water appears out of nowhere, and the surface glitters like so many jewels lit by the heavens. Drawn by the beauty, you quietly approached the spring. In that moment, the breeze rippling across the water, delivers to your nostrils the smell of sweet flowers and wild red fruit, so sweet and exalted. Up in the air, a pair of violet butterflies tangling in flight! (Drops of God, Volume 4. The First Apostle, a.k.a. George Roumier Les Amoureuses. 2001. In the words of Yukata Kanzaki, the fictional world-known wine critic.) There is no other lieu-dit in the world that could lead you to a virgin forest like this than Les Amoureuses, the premier cru in Chambolle Musigny, in Cote de Nuits of Cote d'Or, Burgundy. This is George Roumier Les Amoureuses. Vintage... let's say, 1999? You announce to your wine lover friend with whom the bottle is shared, with a somwhat confident smile.

But seriously, blind tasting is one of the essential skills of many wine professionals. For an importer or retailer, to be able to pick out the best quality wine (or the wine most likely to sell) at the most reasonable price contributes directly to the profitability or survival rate of the business. For a wine writer or critic, the capability of correctly judging wine's quality and ageability is very much hinged upon his or her own reputation. For a sommelier, correctly identifying wines blindly not only creates the wow factor for the restaurant (Like the tales around well-known sommeliers Raj Parr, Fred Dame --- exposed in New York Times articles on scandals though, Larry Stone, and the likes), but also helps tremendously with building a best wine program given a limited budget.

Therefore, it is no wonder that most rigorous wine study programs include a section on blind tasting in examinations. In The Court of Master Sommelier's tasting exams required to earn the title Master Sommelier, for instance, candidates have to pass an oral example of 25 minutes to identify six wines --- three white, three red --- correctly in terms of grape variety, region of production, and vintage, by first describing them, one by one, from colors in sight, to aromas on the nose, to flavors or other elements in the mouth, and then concluding with deduced identities. In The Institute of Master of Wine's tasting exams required to obtain the diploma of Master of Wine, as another example, candidates have to sit a written exam of three hours while tasting 12 wines, answering in the form of essays different winemaking techniques or climatic characteristics exemplified in the color, aromas, and tastes of wines, with attempts to identify either the vintage, region, or grape variety, possibly funneling (For instance, if at one point you think the closest you could get with a wine is that it is an Italian red wine due to perhaps its volatile acidity, drying tannins, prominent herbal characters, and an acidic spine. But no clue if it is a Brunello di Montalcino, a Barolo, or an Etna Rosso, you could potentially funnel by putting down that you think it could be all three with a slight inclination towards Etna Rosso due to its volanic characteristics.) when uncertainty arises.

There have been quite a few different schools of thought regarding how to blind taste, what makes an excellent blind taster, what to look for to improve blind tasting skills, and so forth. One of the most widely accepted approach is deductive tasting, possibly popularized by The Court of Master Sommelier and Wine and Spirits Education Trust, which essentially separates the process of blind tasting into two steps: first, describe the wine in terms of color, aroma, and flavor, and structure, as precisely and objectively as possible; second, given the resulting descriptors, logically deduce the identity of the wine without referring back to the liquid in the glass. The first step requires a palate tuned to accurately identify a wide range of aromas and flavors in different forms and levels of doses, from exotic fruits like lychee or tamarind to esoteric flowers like marigold or azalea, from Asian five spices to Comte cheese and Herbes de Provence, from potting soil after an early summer rain, to pencil shavings and graphite, let alone cat urine, dirty socks, wet dogs, barnyard funk, and leather belts. And that is why ``licking rocks and eating dirt`` are not uncommon perhaps not only among geologists, but also sommeliers --- at least those serious about improving palate sensitivity, I guess. It is only when one can objectively identify all the elements in a wine precisely in a consistent manner, can the second step --- logical deduction --- really shine.

In this section, I will focus on this second step, the logical deduction. Various advice and toolkits for tuning your palate for the first step have been passed down: constant training with wine aroma kits, the sniff and scratch book series by Richard Betts, roasting plain popcorns with different spices --- a tip by Jill Zimorski, cooking with a wide range of ingredients and condiments, paying attention to not only the flavors but also the structural elements, the textures, and types and shapes of acidity, to name just a few. Let's assume for now --- don't worry, we have solutions to be discussed later for when this assumption is hardly met --- that we all have reached the point when we have the perfect palates able to capture the whole spectrum of aromas, flavors, and sensations in a glass of wine.

To logically deduce the identity of any given wine, is to compare the wine in question to stereotypes of wines with known identities in our memory, and find the identity of the most similar stereotype. Therefore, the second step of deductive tasting --- logical deduction --- can be further divided into two parts.

A first and major part of training for logical deduction, is to build up a robust and comprehensive database of stereotypes of wines of different origins, varieties, vintages, and producers, etc. with objective and subjective descriptors. How would you describe a stereotypical 10-year-old Condrieu from the 2010 vintage? What are the characteristics of a 2013 Aglianico del Vulture in 2020? Luckily, such summmarization tasks are not unique to wine tastings and there is indeed a lot to borrow from the field of natural language processing (or more generally, machine learning to accomplish this task in a data-driven manner with much greater precision and capacity than human memorization and manual work. We describe several such frameworks in Summarization.

Secondly, comparing the descriptors of the given wine to those of stereotypes in our collected database in the first step. Humans are notoriously bad in such tasks. For example, in the case of blind tasting, one taster decided to narrow down to only Savennieres, the most ``cerebral`` wine producing region, after detecting both Botrytis --- honey, marmalade, saffron --- and oxidative --- almond paste, bruised apple, cheese rind --- aromas. However, wouldn't an aged Montrachet from certain vintages and producers also best exemplify both Botrytis and oxidative notes? One might further confirm or refute the choice of Savennieres with the level, shape, and structure of acidity, as well as unique aromatics since Chenin supposedly is of searingly high and crescendo acidity according to Nick Jackson's excellent blind tasting book Beyond Flavors, and radiant of fragrances like chamomile, jasmine, honeysuckle, wasabi, and dried stone and tree fruits sometime a touch of tropical too. However, more often than not, one starts to hallucinate certain aromas signature of Savennieres with such an objective in mind, falling victim to the confirmation bias. Blame it on the subjectivity of wine tasting!

In another example, one taster might have quickly eliminated varieties like Nebbiolo, Grenache, Pinot Noir, Gamay(, and more indigenous varieties such as Freisa, Ruche, Prie Rouge, Nerello Mascalese, Baga, etc.) simply based on the deep purple color in the glass. However a Hubert Lignier Charmes Chambertin, a Ch√Ęteau de Beaucastel Ch√Ęteauneuf-du-Pape, or a Yvon Metras Moulin-a-Vent would easily defeat that assumption, in which case the taster would have simply bypassed the correct identities at the initial judgment. The advice of funneling, or enlisting all the possible ``grape laterals`` --- easily confused or similar grape varieties --- has been circulated among some wine study circles. But a Pinot Noir might be similar to Gamay, which might be similar to Nebbiolo in some capacity, which then is similar to Sangiovese (ever mistook a Brunello for Barolo?) or Nerello Mascalese or Xinomavro, and the chain never stops....

Such begs the question, is there an optimal or systematic way to move the deduction process consistently towards the correct answer as much as possible? In what steps and based on what characteristics should one eliminate or funnel? For example, Abigail might start with color, then aromatics on the nose, then flavors and finally structural components on the palate, therefore deduce by eliminating most varieties by color, draw initial conclusions based aromatics on the nose and palate, and narrow down to or confirm the final conclusion with the structural components. But Bob perhaps might argue one should use the structural components to come up with a list of initial conclusions and drill down to a few based on aromatics and flavors, and finally confirm with color or quality indicators. Yet another pro Claire might instead use fruit categories and conditions (crunchy tree fruit or jammy stone fruit?) on the nose versus on the palate (if fruit went tart on the palate compared to the nose it might be indicative of certain regions) for initial conclusions, and structural components for final conclusions. Or if Claire is not good at judging the level or type of acidity, she might choose to not reply on structural components as much and use them sparingly. Whose strategies might most consistently lead to the most correct answers in blind tasting sessions? What is the optimal strategy based on one's strengths and weaknesses? For instance, if Claire is confident in her ability to detect spices but lacking in acidity calls, whereas Bob can never detect Rotundone (the chemical compound supposedly responsible for smells of black pepper) but is excellent in accessing fruit aromas and flavors. How should their blind tasting strategies differ to accommodate these strengths and weaknesses? What if we were blind tasting for vintage alone, or variety alone, or country? How would the optimal deduction strategy change according to the target? Intuitively, there might be a much smaller set of characteristics we watch out for if we are trying to decide on the country alone, compared to vintage or variety. Such are exactly what we dive into in Decision Tree and Multi-task Learning.

AI Talk


Decision Tree

Multi-task Learning

Back to Top