How do you understand what thousands of customer reviews say without reading them one by one?

The question

Beyond a few hundred reviews, reading them by hand no longer works. The tool takes a review file (CSV, Excel, JSON, text, DOCX or PDF) and gives an overview: sentiment, themes, contradictions between rating and text, duplicates.

A multilingual model that runs locally

Sentiment is computed by Cardiff NLP’s XLM-T, a multilingual model converted to ONNX and quantised to int8. It runs on the computer, without sending reviews to an external service. Long reviews are split into 256-token segments, all of which are analysed.

When the best score is below 0.60, or too close to the second, the result is “Undetermined” rather than a forced label. A review with clearly positive and negative passages is labelled “Mixed”. These cases are offered for review.

An interface to explore and correct

Filters, charts and exports run in the browser. Themes rely on words and phrases that users can adapt to their domain. Every manual correction is exported along with the original sentiment.

What the check measures, and what it does not

On 24 constructed examples (8 French, 8 Portuguese, 8 English), the model gives 23 expected labels and one abstention. This check is not an independent benchmark: it does not support any claim about production accuracy. The model was trained on tweets, so a gap with retail reviews is possible, and irony or typos can still cause errors.