Sentiment analysis
Sentiment analysis is the use of language-processing methods to classify text as positive, negative or neutral, so that opinion in large volumes of content can be summarized.
Last updated September 28, 2026
How it works
Early systems scored text using word lists, so "excellent" added points and "terrible" subtracted them. Current systems use trained language models that classify a passage in context. Output is usually a label (positive, negative, neutral) or a numeric score, sometimes with finer categories such as emotions.
The hard part is what the label refers to. A news article about a lawsuit contains negative language without expressing an opinion of the defendant. A sentence can be positive about one company and negative about another. Sarcasm, quotation and domain vocabulary also cause errors. The quality question for any classifier is whether it measures the tone of the text or the stance toward the specific subject, and the two can give different numbers.
For that reason, teams should validate sentiment output on a manually reviewed sample from their own data before relying on aggregate scores.
Common misconceptions
Sentiment is not reputation, and a positive share is not a satisfaction score. Aggregates also hide important detail, such as a small but influential group of critics. Read the underlying items before drawing conclusions. See reputation management and social listening for the wider context.
See how these signals show up in your own coverage on the PeakMetrics platform or run the free AI Perceptions check. Back to the glossary.