1. Two scores on a drink page
Next to the name of a drink you see the score out of 10. It is the average of every rating we have for that drink: ratings from our own community and ratings calculated from public tasting notes. Every rating counts once.
Below it is the community score: the average of the ratings written on Tasting Intelligence itself, by people with an account. When every rating of a drink comes from our community, both numbers are the same. The page then shows the score once, "based on N community reviews".
Score out of 10
All ratings: from the community and calculated from public tasting notes. Broad, and available early.
Community score
Only ratings written here, by people who tasted the drink and chose the number themselves.
- The colour of the ring follows the score: green from 8, gold from 6, orange below 6.
- Latest review is the date of the newest rating that counts in the score.
The brand average
A brand page shows one brand average: the average of every rating of every bottle of that brand, from our community and calculated from public tasting notes. Every rating counts once, so a bottle with many ratings weighs more than a bottle with a few. The line below it says how many of those ratings come from our community, and shows their average once there are enough of them, the same threshold as on a drink page. A brand page never gives search engines stars.
2. Where ratings come from
Community ratings
Someone tasted the drink and gave it a rating from 0 to 10 with their tasting note. We use that number exactly as given.
Calculated ratings
Many drinks are described in public tasting notes elsewhere on the web. We do not copy the score of the original site. A language model reads the text and estimates how positive it is.
How a rating is calculated
- The tasting note is translated into English, the language the analysis runs in.
- RoBERTa, a language model from the BERT family, reads the text and gives three probabilities: negative, neutral and positive.
- Those probabilities become one number from 1 to 10: an entirely negative text gives 1, a neutral one 5.5, an entirely positive one 10.
A calculated rating therefore measures the tone of a text, not a number the reviewer chose.
3. Calculated or community: pros and cons
Calculated ratings
Pros
- Breadth: a drink gets a profile and a score long before many people have rated it here.
- Consistency: every text is read by the same method.
Cons
- It reads tone, not taste: a dry, factual description lands near the middle, whatever the reviewer thought of the drink.
- Irony and mixed notes are hard to read, and very high or very low verdicts are flattened.
- Nobody chose the number: it is our estimate of what the text says.
Community ratings
Pros
- Someone tasted the drink and chose the number themselves.
- Written on this site, under a profile you can open.
Cons
- Still scarce: it takes time before a drink has enough community ratings.
- One palate per rating: personal taste weighs in, which is why we wait for several.
Why community ratings count for more
A rating someone gave after tasting says more than an estimate from a text. That is why we show the community score separately, and why it is the only score we share with search engines. In the score out of 10 every rating still counts once: calculated ratings make sure a drink has a score while the community grows.
4. Reading the community score
The community score appears from 5 community ratings. With one or two, a single opinion can move the average by whole points; with more, it starts to say something about the drink.
Example: still collecting
Example: live
- The bars stand for the ratings needed. Every filled bar is one community rating; the numbers in brackets say the same.
- Once every bar is filled, the bars make way for the average and the number of community reviews it is based on.
- No community reviews yet means nobody has rated the drink here so far. Rate it to fill the first bar.
5. What search engines see
Search engines such as Google can show stars next to a page in their results. We only give them the community score, never calculated ratings, because those stars should reflect reviews written here. A drink without a live community score has no stars in search results.
6. Under the Hood
The models and software behind this page, and when they run.
-
fast-langdetect Google Cloud Translation
Language detection & translation: A review in another language is translated to English first, because the analysis models read English. The original text stays on the site.
When: Every new review, written here or collected from other sites.
-
RoBERTa
cardiffnlp/twitter-roberta-base-sentiment-latestRating 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.