Glossary · Reviews & analysis
What is review sentiment analysis?
Sentiment analysis labels each review by tone. SellerSort uses four classes — positive, neutral, negative and mixed — and keeps the probability for each, so a review that praises the product and complains about delivery is counted as mixed rather than forced into one side.
What this comes down to
- Four classes: positive, neutral, negative, mixed.
- “Mixed” exists because many useful reviews contain both praise and a complaint.
- Every probability is kept, so thresholds can move without re-analysing.
- Sentiment says how a review feels; topics say what it is about.
Why four classes, not two
“Arrived three days late and the box was crushed, but the mug itself is lovely” is neither a complaint nor praise alone. A two-way split would bury the delivery problem under a positive label. Counting it as mixed keeps it visible where problems are counted.
The Review Analyzer orders topics by negative and mixed mentions together for this reason.
Sentiment is half the answer
Knowing 17% of reviews are negative does not tell you what to fix. Sentiment becomes useful combined with topic labels — shipping, sizing, quality — and with severity, which separates an annoyance from a product that broke. The summary counts all three from per-review judgments, in code, so the same reviews always give the same numbers.
How each class is defined
| Class | Definition |
|---|---|
| Positive | The reviewer is satisfied overall. Small caveats are fine as long as the verdict is favourable. |
| Neutral | The reviewer states facts or describes the item without praise or complaint, or is plainly indifferent. |
| Negative | The reviewer is dissatisfied overall, even if they name one thing they liked. |
| Mixed | The reviewer is genuinely split — they praise some parts and condemn others, and neither side wins. Use this only when a single overall verdict would misrepresent them. |
These are the exact definitions the model is given. The question asks for the reviewer’s overall verdict, taking the whole review together, so one sharp sentence in an otherwise happy review does not flip it to negative.
Questions
Is star rating the same as sentiment?
No. A four-star review can describe a real problem and a two-star one can be about delivery rather than the product. Ratings are kept when the source has them; sentiment is read from the text.
Can I see how sure the analyzer was?
Yes. The CSV export carries a sentiment_confidence column next to the sentiment, so you can filter out the calls it was least sure about and read those yourself.
Read next
- Review severityA score from 1 to 5 of how serious the problem in a review is; SellerSort counts reviews at severity 4 or above as urgent.
- Actionable reviewA review that needs a response from the seller — an unresolved problem, a direct question, or something owed to the customer.
- How to analyze thousands of customer reviews without reading them allTurning a pile of free-text reviews into counts you can act on: topics, severity, and the difference between a loud complaint and a common one.
- How to prioritise negative reviews by severityNot every one-star review is urgent. Rank complaints by severity and how often they recur, so the fix list starts with what actually hurts.