Customer review sentiment analysis
Sentiment analysis usually means labelling a review positive or negative. That works for short, single-topic reviews; it does not work for real ones. What is useful is a sentiment per topic.
What aspect-level sentiment is
A review is represented not by one label but by topic–sentiment pairs. Each pair carries its own severity score and evidence quote, so the same review can count as positive on one topic and negative on another.
Uncertainty is not hidden
The model is not obliged to produce a judgement for every review. Without a clear enough signal the result is marked `unknown` and excluded from aggregation. Turning a low-confidence guess into a firm percentage makes that percentage look more trustworthy than it is.
Confidence and sample size are read together
A high-confidence result on a tiny sample and a large-sample result with low confidence are each insufficient on their own. The dashboard shows both on every insight and does not publish findings below the minimum sample rule.
Review text is not an instruction
Reviews are untrusted user content. Even if one contains something addressed to the model, it is not acted on; the text is processed only as data to be analysed. Model output is likewise never rendered directly as HTML in the interface.
Method
Every number on this page comes from the analysis pipeline nalyzed runs over publicly visible Google Maps reviews: normalisation, language detection, personal-data redaction, business-type topic classification and aspect-level sentiment. Results are stored with their evidence quote, sample size and confidence score; a result whose evidence does not appear verbatim in the source review is not published. When collection does not finish, coverage is shown as `partial` and no screen claims that every review was analysed.