Wineinformatics: The Evaluation of the Computational Wine Wheel in the Past Decades

    


    


    Figure 1: 2009 Kosta Browne Pinot Noir Sonoma Coast’s review on both chemical and sensory analysis, which received 95 points from Wine Spectator.

    

In consumers’ perspective, using wine reviews is considered more approachable to learn about the quality and the characteristics of a wine than using physicochemical laboratory data. Laboratory data is often difficult to obtain due to its associated cost and in general not for consumers. Wine reviews, on the other hand, are much more available. There are hundreds of wine review magazines and websites, such as Wine Spectator, Robert Parker, or wine.com. Each of these sources has hundreds of thousands of reviews by wine experts for a wide range of wines.

However, structured data, such as physicochemical laboratory records, can be easily interpreted and analyzed by computers. In contrast, unstructured data, like wine reviews showed in figure, require natural language processing techniques to enable computers to understand human language-based reviews. With a vast repository of millions of wine reviews sourced from diverse outlets, often available at minimal expense, there lies immense potential to uncover valuable insights beneficial to a wide audience. Therefore, the Computational Wine Wheel (CWW) was developed started from 2014 to focus on processing wine reviews into computer understandable format so that data mining techniques, such as classification, clustering, association rules and regression, can be applied to the data collected from the reviews various wine related knowledge discovery [8-14]. In this review, the evolution of Computational Wine Wheel in the past decade is described and discussed, transitioning from single-source to multi-source evaluation.

The evolution of the Computational Wine Wheel (CWW)

In order to use words in wine reviews to classify wine scores, the very first Computational Wine Wheel (CWW) has developed in [15]. It is essentially a filter, or a sieve, that collects important words, or attributes, from a review. These attributes, then, are used in classification algorithms to determine the quality of the wine. For example, below is the review of Dow’s Vintage Port 2011 from Wine Spectator:

Powerful, refined and luscious, with a surplus of dark plum, kirsch and cassis flavors that are unctuous and long. Shows plenty of grip, presenting a long, full finish, filled with Asian spice and raspberry tart accents. Rich and chocolaty. One for the ages. Best from 2030 through 2060.

The bolded words are the attributes that can be extracted from the review. These attributes can be of different types: savory (chocolaty, tart), body (long), or adjective (powerful, refined). In order to create the Computational Wine Wheel, a technique in natural language processing called “Bag of Words” is used.

Bag of Words

From a collection of reviews, “Bag of Words” extracts the attributes by tokenizing, removing the stop-words, normalizing the tokens, and creating the master dictionary. The first step, tokenizing, is done by breaking the sentences into unique words and phrases. Then, the stop-words, such as articles (a, an, the) or prepositions (to, for), are removed from the list of words. Token normalization condenses words with similar meaning into a single representation, such as “red” and “reddish” into “red.” Finally, all the words that are still in the list at the end form a master dictionary, or a sieve, to be used to extract attributes from reviews.

Note that attributes can contain more than one word. For example, “raspberry tart” is considered to be one attribute. Another note is that similar words are not always necessarily normalized into a single attribute. For example, “apple” and “fresh apple” are condensed into “apple,” but “green apple” is considered a separate attribute, because in wine, green apple gives a distinct flavor. This is where domain knowledge is very important when it comes to natural language processing.

Wine Aroma Wheel

One of the earlier attempts to create a master dictionary for wine reviews is the Wine Aroma Wheel, developed by Ann. C. Nobel [16]. It contains words that describe fragrance and flavors and consists of 12 categories, each with subcategories that maps to different taste, scent and aromatic qualities of red and white wines. While the Wine Aroma Wheel is useful to study wine, one of its limitations is that it does not include adjective and wine body attributes. For example, if the Wine Aroma Wheel is applied to the review of Dow’s Vintage Port 2011, the only attributes extracted are dark plum, kirsch, cassis, Asian spice, raspberry tart, and chocolaty, which are all savory attributes.

Wine Reviews

Quality data with minimal noise is essential for successful data science research. Therefore, the cornerstone of this research lies in high-quality wine reviews. According to the Wine School of Philadelphia, there exist five prominent wine review platforms offering extensive databases of professional wine critiques.

Wine Spectator [23], a renowned wine magazine since 1979, meticulously evaluates over 15,000 wines worldwide. It stands as a pinnacle in the field and a primary source for previous Wineinformatics research. Wine Enthusiast [24], established in 1988, annually reviews a staggering 24,000 wines. Its reviews are accessible for free with a simple email registration. Beyond wine critiques, the magazine encompasses a broad-spectrum including wine accessories, storage solutions, education, food pairing, and lifestyle features.

Antonio Galloni's Vinous [25], established in 2012, boasts a team of top-tier wine critics offering online wine evaluations. It enjoys high esteem in the wine trading realm, emerging as a trailblazer in contemporary wine literature. Robert Parker’s Wine Advocate [26], arguably the most eminent wine critic, revolutionized the industry with his Wine Advocate magazine in 1978. Introducing the widely adopted 100-point scale, his influence has been profound, shaping the wine trade landscape significantly. Decanter [27], founded in London in 1975, not only offers comprehensive wine reviews but also delves into wine producers and regions. While their critiques are detailed and focused, they occasionally diverge from the perspectives of other aforementioned magazines.

Among various prestigious wine magazines, Wine Spectator can be considered as an easier data source to start aggregating wine reviews because of their strong on-line wine review search database and consistent wine reviews. These reviews are mostly comprised of specific tasting notes and observations while avoiding superfluous anecdotes and non-related information [3].

Computational Wine Wheel

In order to overcome the Wine Aroma Wheel’s limitations, [17] introduced the Computational Wine Wheel (CWW) for the first time in 2014. In that research, top 100 wines from Wine Spectator in 2011 was studied and utilized with bag of words technique to create the master dictionary. Merging the concept of wine aroma wheel and natural language processing methods, not only were flavor attributes included, other physical attributes, such as acidity or tannins, were also present. The end result was the first version of the CWW, with 12 categories and 28 subcategories that contained 547 original tokens and 376 normalized tokens. Both hierarchical clustering and association classification algorithms were applied on the dataset contained 1000 wine reviews processed by the first CWW and received satisfactory results [17].

Later on, [18] improved the original CWW by using Wine Spectator’s reviews of the top 100 wines in a 10-year period, from 2003 to 2013. The newer version, CWW2.0, was extended to include 14 categories with 34 subcategories. There are 1932 original tokens and 986 normalized tokens for a more comprehensive list of attributes. From 2016~2022, numerous dataset were generated and processed by the CWW2.0 to discover different types of information related to wine: In [13, 20-21], a large dataset contains more than 100,000 Wine Spectator wine reviews with vintage 2006-2015 were collected and processed thorough CWW2.0 to study the ranking of wine reviewers [20], regression on wine price and grade [21] as well as multi-label and multi-target methods in Wineinformatics [13]; a smaller dataset targeted on Bordeaux with more than 14,000 wine reviews and an even smaller focused on elite Bordeaux with 1359 wine reviews were proposed and studied to know more about 21st century Bordeaux wines [12]. The large Bordeaux dataset is currently publically available through IEEE data port [22].

In 2022, Robert Parker’s reviews were included in the Wineinforamtics researches and performed a direct comparison with Wine Spectator’s review [11]. In 2023, the Computational Wine Wheel 2.0 was revised into 3.0 by including reviews from Robert Parker’s Wine Advocate and adopts Neural Networks into Wineinformatics researches [19]; 513 of Robert Parker’s elite Bordeaux reviews were included in the creation of the new Computational Wine Wheel. More attributes were added to the new version CWW3.0 to increase the number of original tokens to 2589 that were condensed to 1191 normalized tokens. Below is a table that compares all 3 versions of the Computational Wine Wheel. The subcategories with no word count in the CWW columns were not included in the original wheel. Since this review is focus on the evolution of the Computational Wine Wheel, how to utilize the CWW is clearly described in [3 and 19]. Generally speaking, the more tokens stored in the Original (3rd column in table1), the more words used by human language can be picked up by the Computational Wine Wheel; meanwhile the more tokens used by Normalized (4th column in table1), the more attributes can be provided to the machine learning algorithms. Therefore, table1 clearly suggests that the evolution of the Computational Wine Wheel has made it far more robust compared to its state a decade ago.

  


    
Category 

    
Subcategory 

    
Original 

    
Normalized 

  

  


    
CWW1.0 

    
CWW2.0 

    
CWW3.0 

    
CWW1.0 

    
CWW2.0 

    
CWW3.0 

  

  


    
Caramel 

    
Caramel 

    
9 

    
71 

    
97 

    
7 

    
40 

    
56 

  

  


    
Chemical 

    
Petroleum 

    
3 

    
9 

    
11 

    
1 

    
5 

    
6 

  

  


    
Sulfur

    


    
11 

    
11 

    


    
10 

    
10 

  

  


    
Pungent

    


    
4 

    
4 

    


    
3 

    
4 

  

  


    
Earthy 

    
Earthy

    
18 

    
72 

    
128 

    
2 

    
31 

    
47 

  

  


    
Moldy

    


    
2 

    
2 

    


    
2 

    
2 

  

  


    
Floral 

    
Floral

    
15 

    
61 

    
87 

    
15 

    
39 

    
45 

  

  


    
Fruity 

    
Berry

    
18 

    
49 

    
84 

    
15 

    
28 

    
39 

  

  


    
Citrus

    
11 

    
37 

    
56 

    
11 

    
23 

    
35 

  

  


    
Dried Fruit

    
21 

    
67 

    
76 

    
21 

    
60 

    
65 

  

  


    
Fruit

    
5 

    
22 

    
42 

    
4 

    
9 

    
16 

  

  


    
Other

    
7 

    
25 

    
22 

    
7 

    
18 

    
9 

  

  


    
Tree Fruit

    
12 

    
39 

    
55 

    
9 

    
31 

    
40 

  

  


    
Tropical Fruit

    
15 

    
48 

    
67 

    
11 

    
27 

    
36 

  

  


    
Fresh 

    
Fresh

    
15 

    
41 

    
75 

    
12 

    
29 

    
44 

  

  


    
Dried

    
6 

    
25 

    
50 

    
6 

    
21 

    
39 

  

  


    
Canned/Cooked

    
7 

    
16 

    
18 

    
7 

    
15 

    
17 

  

  


    
Meat 

    
Meat

    
1 

    
25 

    
36 

    
1 

    
13 

    
21 

  

  


    
Microbiological 

    
Yeasty

    
3 

    
5 

    
5 

    
3 

    
4 

    
4 

  

  


    
Lactic

    
3 

    
14 

    
14 

    
2 

    
6 

    
6 

  

  


    
Nutty 

    
Nutty

    
3 

    
25 

    
27 

    
3 

    
15 

    
20 

  

  


    
Overall 

    
Tannins

    
24 

    
90 

    
124 

    
3 

    
4 

    
6 

  

  


    
Body

    
17 

    
50 

    
61 

    
10 

    
23 

    
17 

  

  


    
Structure

    
9 

    
40 

    
51 

    
2 

    
2 

    
2 

  

  


    
Acidity

    
14 

    
40 

    
61 

    
3 

    
3 

    
4 

  

  


    
Finish

    
50 

    
184 

    
233 

    
6 

    
5 

    
13 

  

  


    
Flavor/Descriptors

    
217 

    
649 

    
889 

    
179 

    
432 

    
467 

  

  


    
Oxidized 

    
Oxidized

    


    
1 

    
2 

    


    
1 

    
2 

  

  


    
Pungent 

    
Hot

    


    
3 

    
3 

    


    
2 

    
2 

  

  


    
Cold

    


    
1 

    
1 

    


    
1 

    
1 

  

  


    
Spicy 

    
Spice

    
26 

    
83 

    
85 

    
21 

    
44 

    
53 

  

  


    
Wood 

    
Resinous

    
6 

    
24 

    
31 

    
6 

    
9 

    
12 

  

  


    
Phenolic

    
1 

    
6 

    
6 

    
1 

    
4 

    
5 

  

  


    
Burned

    
11 

    
47 

    
51 

    
8 

    
26 

    
28 

  

Table 1: Number of tokens in each subcategory comparison between three versions of the Computational Wine Wheel.