Analyze feedback
Classify a single review, ticket or comment and see which words drove the result.
0 / 20,000
⌘↵ to analyze
Try an example
Each click picks a different sampleRecent
No analysis yet
Type or paste feedback on the left. With Live on, results update as you type.
Sentiment
Confidence
Low confidence. The model is unsure — the text may be mixed, sarcastic, off-topic or not English. Review it manually.
Probability
Key drivers
Pushing positive
Pushing negative
Annotated text
positive negativeBatch insights
Upload a CSV of reviews to see the sentiment mix, find uncertain cases and export labelled results.
Drop a CSV file here
Needs a header row. The column with the longest text is used — you can change it after upload. Up to 20 MB.
Sentiment mix
Confidence distribution
0%50%100%
| # | Review | Sentiment | Confidence |
|---|
Model
How the classifier was trained and how it performs on reviews it has never seen.
Per-class performance
test setConfusion matrix
rows = actual, columns = predictedModel selection
validation macro-F1Training setup
Known limitations
- Sarcasm is read literally — “Oh great, another two hour wait” scores positive.
- Some negations are missed — “no complaints whatsoever” lands on neutral.
- Mixed reviews are pulled toward whichever side has stronger words.
- Non-review text (questions, plain facts, gibberish) has no “not applicable” class; it usually gets low confidence.
- English only. Emoji and ratings like “10/10” are ignored by the tokenizer.