Sentiment Analyser

About the Sentiment Analyser

Sentiment analysis has a reputation problem it has earned. Presented as a percentage it looks like a measurement, and people treat it as one — sixty-eight per cent positive goes into a slide, gets compared against last quarter, and becomes a number a decision rests on. It is not a measurement. It is a reading, and readings of tone are wrong often enough that the decimal point is a lie.

This page gives you the reading and refuses the false precision. You get a verdict with the reasoning attached, the themes people keep returning to, and a verbatim quotation for each finding so you can check the interpretation against the words that produced it. Those quotations are the important part. A theme without an example is an assertion; a theme with three examples is something you can argue with.

Two modes cover the two situations. Whole-text analysis reads a single piece and describes the shape of the feeling in it, including whether it is uniform or split — which is a materially different finding from an average. Line-by-line treats each line as its own item, which is what you want when you have pasted a column of reviews or survey responses out of a spreadsheet, and it ends with a tally rather than a percentage.

The instruction that matters most is the one about ambiguity. Sarcasm reads as praise to any system that scores words. Politeness masking complaint is the standard register of British and Indian customer feedback both, and reads as neutral or mildly positive when it means nothing of the sort. Where those are suspected, this page is told to say so rather than force the item into a category, and an honest "this could be either" is worth more than a confident classification that is backwards.

What it cannot do is tell you why. That takes talking to the people who wrote it.

How to use it

1. 1Paste the feedback — reviews, survey answers, support messages, or a single long piece.
2. 2Choose line-by-line when each line is a separate response, or whole-text for one continuous document.
3. 3Keep the themes on. The recurring subjects are usually more actionable than the overall verdict.
4. 4Keep the examples on and read them. They are how you check the interpretation rather than trust it.
5. 5Treat the tally as an indication. Where it disagrees with your instinct, read the source before believing either.

Questions

Why is there no percentage score?

Because a percentage implies a precision that does not exist here. Sarcasm, understatement and cultural register all defeat automated tone reading, and the errors are not evenly distributed — they cluster in exactly the responses that matter most. A tally of items with the reasoning visible is honest about what it is. A decimal is not.

Does it handle sarcasm?

Better than a word-scoring system, and imperfectly. It is instructed to flag suspected sarcasm as ambiguous rather than to resolve it, on the grounds that a marked uncertainty you can check beats a confident classification you cannot. Expect to catch some yourself.

How much feedback can I paste at once?

Twenty-four thousand characters, which is a few hundred short reviews. For larger sets, run it in batches and compare the themes between them — themes that hold across batches are the real ones, and themes that appear in only one are usually noise.

Can I use this to score individual customers or staff?

Please do not. It is not accurate enough at the level of a single item to carry a consequence for a named person, and the errors fall hardest on people who write indirectly or in a second language. It is a tool for finding patterns in a pile, not for judging one item in it.