A comment can have one or more mentions
The legacy module scored the whole comment. Sentiment breaks it apart: every distinct thing the guest said becomes a mention, with its own polarity and its own verbatim quote.
-
A single comment can yield one mention or ten — that is why the header shows mentions and guests as two different numbers.
-
Nothing is translated first. The engine reads Spanish, English and Portuguese in their original language and returns the same taxonomy.
Everything downstream — indexes, rankings, trends — is counted over mentions, never over comments.
Three axes: topic, attribute and area
Every mention is classified on two axes: the topic it talks about and the attribute being judged. The area is not picked by the model: it is resolved from a curated topic × attribute matrix.
-
Same combination, same area, always. That is what makes the area ranking actionable rather than arguable.
- The attribute is the cross-cutting axis: Cleanliness is read at once across bathroom, room, bed, towels, shower, pool and restaurant.
The topic answers what was talked about, the attribute what was judged, and the area who owns it.
How the Sentiment Index is calculated
Positives divided by positives plus negatives, times 100. Neutral mentions and suggestions are counted and can be read, but stay out of the denominator: they neither reward nor penalise.
-
The same formula applies at any cut: property, area, topic, attribute or a single topic × attribute cell.
-
Reading scale: 70 or above is good, 40 to 69 needs attention, below 40 is critical.
The index is always shown, at any sample size — there is no minimum that hides it.
Comparing periods without noise
Each period is compared against the immediately preceding window of the same length — 30 days against the previous 30 — and recalculated with whatever filters you have applied.
-
If either window has fewer than 5 polar mentions, the delta is hidden instead of shown.
- A 40-point swing built on 3 mentions is noise, not a trend. The index is still shown; only the comparison waits for enough evidence.
No delta is hand-written — every arrow you see was computed from dated mentions.
What you gain
The screen is built so that any number can be traced back to the sentence the guest actually wrote.
-
From the score to the why: pick a topic, see its dimensions, read the quotes — three columns, two clicks.
- Cross-cutting reading: you can ask how Cleanliness is doing across the whole property, not just inside one topic.
- One metric fewer: the Perfection Index is retired — it moved almost identically to the sentiment index (r = 0.92–0.98).
The same guests, the same comments, many more answers.
