AI Sentiment Analyzer

There are two families of method. The older scores each word against a dictionary marked positive or negative, adds the result up and applies rules for negation and intensifiers — fast, cheap, inspectable, and easily fooled by idiom. The newer hands the whole passage to a language model trained on enormous amounts of writing and asks it to judge, which copes far better with context and phrasing while leaving you no arithmetic to check. This page is the second kind: the passage goes to a language model under an instruction to return a label, a confidence estimate, the dominant emotions and a line of reasoning.

Sarcasm contradicts itself on purpose — the words approve and the meaning does not, and whatever flips it usually sits outside the sentence, in tone of voice, shared history or a running joke. 'Brilliant, another update that breaks the login' holds a positive word and a bad event, and which wins depends on knowing a broken login is bad news for this writer. Models handle it better than dictionaries because they weigh the passage whole, but dry sarcasm, understatement and anything resting on the previous message in a thread still slip past. If a verdict on a sarcastic line looks wrong, it is.

Not in any statistical sense. Nothing counts outcomes or holds a validation set, so the number is written alongside the label rather than derived from it: an impression of how clear-cut the passage was, expressed as a figure because a figure was asked for. It works as a rough flag, since a low one usually does mean the text is genuinely ambiguous and worth a human read. It does not work averaged, plotted over months, or wired to a threshold that triggers something. If you need a calibrated score, hand-label fifty of your own texts and compare — that is a different piece of work.

Neutral means no evident feeling in either direction, which is more common in real feedback than people expect: delivery updates, factual bug reports, questions, one-word answers. Mixed means both directions are present at once and neither clearly wins. The distinction is what makes labels useful for triage — a mixed result almost always deserves a human read, a neutral one usually does not. Short text drifts neutral simply because there is nothing to go on; 'fine', 'ok' and 'as described' carry plenty of unexpressed opinion, and no method recovers what the writer did not type.

It leaves your device. There is no local model and no offline mode: the passage is sent to this site's /api/ai endpoint and on to Anthropic's API, so with the network off nothing happens at all. Nothing in that path writes what you typed into a database here — the server logs record failures, not input — and beyond that point retention is governed by the AI provider's policy rather than by anything on this page. One related caution: instructions buried inside text you are analysing are meant to be treated as ordinary content rather than obeyed, and that is a mitigation, not a guarantee.

Not here. There is one box and one button, and each run carries a single passage with no memory of the one before, so a hundred rows means a hundred separate runs whose labels were never held to a common standard. One submission is capped at 40,000 characters, which covers a long thread but not an export, and requests are limited to six a minute from one address — comfortable for a person reading answers, useless for a loop. For a real corpus, script it against an API with your own fixed category list, or use something built for tabular data.

Because they measure different things. A rating is a summary judgement made afterwards and anchored to habit — some people never award five, others award five for anything that arrives intact — while the text describes a specific experience as it is remembered. A four-star review can be a list of grievances from a generous writer; a two-star one can praise the product and blame the courier. Where the two disagree, the text is usually the more informative and the rating is the one that gets counted. That gap is the reason to read sentiment at all.

Aspect-based analysis splits a passage by subject — delivery, price, packaging, support — and returns a verdict for each, instead of one verdict for the lot. You need it the moment the question becomes what people are unhappy about rather than whether they are unhappy, and it is the only way a review praising the product and damning the shipping turns into something anyone can act on. This page does not do it: one passage in, one overall label out, with any split described in the sentence of reasoning rather than broken into fields. A structured version means naming your aspects in the prompt yourself.