Free review sentiment analysis

Sentiment analysis is the automatic classification of text as positive, negative or neutral. It is used on customer reviews, survey answers, support tickets and social posts to turn thousands of free-text comments into counts you can track. Most systems also return a score — often between -1 and +1 — so you can rank the strongest complaints and the strongest praise rather than only counting labels.

There are two main approaches. Lexicon methods score each word against a dictionary of positive and negative terms, then adjust for negation (“not good”), intensifiers (“very good”) and punctuation, and add the results up. Machine learning methods instead learn from thousands of labelled examples, and modern ones use transformer language models that read the whole sentence in context. Lexicons are transparent and instant; models handle nuance better but are harder to explain.

Around 80-85% agreement with human labels is typical on short product reviews, and large language models push that into the low 90s on clean English text. Human annotators only agree with each other about 80-90% of the time, so that is roughly the ceiling. Accuracy falls on sarcasm, mixed reviews that praise one thing and criticise another, domain jargon, and comparisons such as “better than the last one”. Treat the aggregate trend as reliable and any single row as a suggestion.

For customer reviews, 70-80% positive is normal and healthy, because people who bother to review are skewed towards the happy and the furious. Anything below 60% positive usually signals a real product or delivery problem. The level matters less than the trend: a ten-point drop in positive share over a month is more informative than the absolute figure, and comparing your own categories against each other beats comparing against an industry benchmark.

NPS is a number the customer gives you; sentiment analysis is a number you infer from what they wrote. NPS asks how likely someone is to recommend you on a 0-10 scale and buckets them into promoters, passives and detractors. Sentiment analysis reads the free-text comment instead, which is where the reason lives. They are complementary: the score tells you how bad it is, the comments tell you why, and running sentiment analysis over NPS verbatims connects the two.

Export the reviews to one row per comment, classify each for sentiment, then group by something you can act on — product, month, channel, star rating — and look at the differences rather than the overall average. Pull out the words that appear far more often in negative reviews than positive ones; those are your themes. Then read a sample of the extremes by hand. Automated scoring finds where to look, but the fix almost always comes from reading the actual sentences. The tool on this page does the first pass.

Mostly not. “Great, another broken zip” contains only positive words, so lexicon methods score it positive and even trained models get it wrong more often than not — published accuracy on sarcasm benchmarks sits well below general sentiment accuracy. The practical defences are to read the extremes manually, to cross-check sentiment against the star rating where you have one, and to accept that a small share of mislabelled rows does not move an aggregate built from thousands.

Sentiment analysis places text on a single scale from negative to positive. Emotion detection assigns discrete feelings instead — anger, joy, sadness, fear, surprise, disgust — so two reviews that are equally negative can be separated into one that is furious and one that is disappointed. Emotion models are less accurate because the categories overlap and people label them inconsistently. Start with sentiment, add emotion only if the distinction would change what you do.