Community Dynamics

1. The Controversy Verdict

The Controversy Score reveals the "Social Risk" of a bottle. By calculating the standard deviation σ of all ratings, we determine if the community is in agreement or deeply divided.

2. Score Drivers (Flavors & Character)

Score Drivers show which flavors and which character traits go together with a higher or a lower score for one drink. They answer why people rate it the way they do, not just how high.

How to read them

• The bar is the difference between the average score of the reviews that mention the term and the average of the reviews that mention any flavor (or any trait) of this drink (on the axis), weighted by the number of reviews. Reviews that name no flavor at all are mostly short or neutral and score lower; counting them would make almost every term look positive.
• The label is the number of reviews. A term needs at least 3, and every review counts once, however often it repeats the word.
• Weighting: a term from 3 reviews keeps 3/8 of its effect, one from 45 reviews 9/10. A signal from many reviewers outranks a lucky outlier.
• Small effects (less than 0.3 points after weighting) are left out. When nothing lowers the score noticeably, the chart says so: that is a finding too.
• No overlap: a flavor only appears under Flavor Drivers, and orange blossom is not counted as blossom as well.
• No verdict words: delicious or awful are left out. A review that says awful scores low by definition, so it tells you that people dislike a drink, not why.

Example: in the charts above, Medicinal and Harsh pull the score down while Vanilla and Smooth lift it. If you dislike medicinal notes, weigh that more heavily than the average score.

3. Under the Hood

The models and software behind this page, and when they run.

• RoBERTa cardiffnlp/twitter-roberta-base-sentiment-latest

  Rating from text: A review collected from another site has no rating we may copy, so a BERT-family language model calculates one from the text (1 to 10). Reviews from our community keep the rating their author gave.

  When: Only reviews collected from other sites.

• spaCy en_core_web_lg

  Word analysis: Picks the adjectives and nouns from each review, in their base form and without filler words, to link words to opinions.

  When: Every review.

Deep Dives & Documentation