In 2011, my brother-in-law sold me a half case of 2007 Pahlmeyer Napa Valley Proprietary Red at cost — $480, significantly more than I had ever spent on wine. It was also my first Bordeaux-style red blend.
Pricey as it was for six bottles, I was floored by the incredible flavors that wine delivered. It remains my favorite wine, and I've (pardon the pun) fruitlessly searched for something that stacks up since.
The problem: choice
There are thousands of Bordeaux-style red blends to choose from, and my personal experience with wine recommendation engines has left much to be desired. So, short of poring over thousands of back issues of Wine Spectator, how can we:
Apply a novel approach to wine classification?
Quickly learn the distinctive features of a given vintage, by region?
Optimize wine selection, and pick a winner that might be under the radar?
There are currently over 5,000 distinct bottles of Bordeaux-style red blends available for purchase on Wine.com. Rather than segmenting these wines using traditional structured data — like price, vintage, winery, grape varietal — what if we could instead rely on the rich, expressive language used in the product description and expert reviews posted online?
Enter NLP (natural language processing).
A Quid network shows 3,380 nodes, where each node represents a description and review of a single bottle of Bordeaux-style red wine from Wine.com. Colors represent the themes of similar wines, based on language used in those descriptions.
Major themes in the visual above include both flavor (e.g., chocolate, explosive cherry) as well as style (e.g., subtlety, fullness) — which points to meaningful variation in how these wines are primarily described.
Several wines connect multiple themes (so-called "bridge nodes") — meaning these could be intriguing choices to try, as they exhibit dominant features from more than one segment. For example, you may be curious about a wine connecting the jam, blackcurrant and exotic spice segments.
While bold fruits like blackberry, blueberry, currant, and cherry are heavily represented, more unusual flavors seem to play a role as well, such as bay leaf, tobacco, sage, and even pencil lead.
The perfect pair
One cluster focused on how well certain wines pair with food, rather than the taste profile of the wine itself.
This cluster, focused on pairing wine with food, is quite dense, meaning the wines are commonly associated with more than one food pairing suggestion. In other words, many of the wines are alike and would go well with a variety of meals.
In this case, I'm most interested in unique wines on the periphery. Personally, I love ricotta and was pleased to see a funky subgroup focused on this cheese. Though I had never heard of it before, Poggio al Tesoro Bolgheri Sondraia 2013 has officially made my list of wines to try for 2017.
What language is used to characterize wines for each season? NLP helps us understand the most important keywords for a given vintage and region.
In the visual above, it's interesting to see that French blends are frequently described in abstract terms like "elegant" and "silk,” while American wines tend towards concrete flavors like "blackberry" and "cherry." I also find it curious that South American blends fairly consistently show vanilla flavor while other regions appear to have a bit more range. Italian blends also seem unique in that they alternate between flavor and style descriptors, showing distinctive tastes like "coffee" and "mint."
While looking across regions gives a broad view, examining the top descriptive term by itself doesn't get at the nuance I'm seeking. Let's dig deeper into just California red blends.
Here we can start to see the true power of NLP applied to wine. At a glance, I can learn that in Napa and Sonoma:
"Cassis", "Blackberry", and "Chocolate" are remarkably characteristic of red blends across most years.
On the other hand, "Elegant" and "Balanced" are probably reserved for French blends and are rarely used here.
2003 was a bizarre year, with several mentions of "tobacco" flavors relative to other vintages.
2011 was also unique, the only vintage with a high degree of "silk" and "structure" language — very uncharacteristic for this region
My favorite vintage (2007) is characterized by "Cassis," "Chocolate," "Layers," "Plum," and "Vanilla."
This shows one way to apply big data and machine learning techniques to wine and learn more about your own taste profile and why you enjoy what you enjoy, irrespective of expert opinion.
With NLP, we can directly identify the wines most similar to wines we already know are favorites. In other words, the more similarity in language used, the more likely two wines are to have connections to each other.
Existing wine recommendation approaches focus on structured data — i.e., how you've already rated certain vintages, regions, and varietals. The problem with this tactic is there's just too much variance across these dimensions, meaning we don't have data granular enough to consistently make an accurate prediction.
If this structured data were sufficient, the implication would be that all grapes within a given region for a given year are (more or less) created equal. Ask any sommelier if that premise holds and be prepared for a passionate lecture.
From the 11 red blends I've rated 4+ stars on the Vivino app, Quid was able to identify 73 associated wines, based entirely on similarities in language used.
Since this approach considers only language, Quid's algorithms recommended a variety of regions for me to try rather than just the three I input as training data.
A subset of these recommended wines (those with average price under $500) is shown below.
While anyone in the food and beverage industry can tell you that wine scoring is often wrong, let's assume it's a realistic proxy for quality. Using the output above would be one possible way to prioritize this list of 73 wines to try.
For example: I know I love the '07 Pahlmeyer, but there's a more affordable '12 Chateau Haut-Bailly available that also scored 96 points.
The '14 Realm the Bard was amazing, but I hadn't heard of the '09 Les Asteries which gained a higher score at a similar price point. Chimney Rock is certainly a household name and extremely representative of the Minerality, Subtlety, Crisp segment, but I'm interested in a 2004 d'Arenberg with a similar score that is more peripheral and possibly unique in its vintage.
What I appreciate most about this approach is that it's blind to anything other than descriptive language, meaning we're getting recommendations based on what the wine actually is rather than where it's from, the grapes used, or whether it was too rainy that year.
There are plenty of ways to leverage NLP in combination with traditional structured data to dramatically disrupt how we think about wine.
1. First, in consumer wine selection: Imagine the approach of automatically mining language similarity applied not just to product descriptions, but to the entire universe of Vivino reviews to make recommendations — incorporating language data could add serious horsepower to that platform. It would also be interesting to visualize a restaurant's entire wine list as a language network and use it to help diners explore the selection in a novel way before a meal.
And of course the approach can apply to more than wine. You can use this same methodology to see patterns in any kind of data, from surveys to product reviews.
2. Marketing insights: How are descriptions and flavors correlated to consumer purchase behaviors? How is this changing over time, and which messages are resonating? NLP could be a practical tool for food and beverage marketers when used in this way, especially wineries relying heavily on direct-to-consumer campaigns.
3. Guided wine tastings: What if we could add visual data to help people understand more about what they are tasting, and make it more likely for them to remember a given brand? Sales reps from distributors that hold tastings in wine shops and liquor stores could use tablets with pre-loaded data to help visually explain what people are tasting and how similar their wines are to other wines the customer may already be familiar with.
Looking at a single style of wine serves as a practical use case for this approach, but I'm excited by the possibility of extending this analysis across multiple varietals. Maybe then I could explain using data why 9 out of 10 blind tasters misidentify the 2014 Château de la Gardine as red instead of white (guilty).
In any event, I now have 73 new items on my to-do list. Salud.
Intelligence in your inbox
Sign up for a monthly look into how data and visualization are changing the way we view the world.