What is sentiment analysis?

The technique comes out of natural language processing research from the early 2000s, where it was called opinion mining. Early systems counted words against sentiment lexicons; supervised machine learning replaced them; large language models now handle most of it, with meaningfully better grasp of context. It became a standard feature of social listening tools because tone at scale is something no human team can read manually.

In practice teams track net sentiment, the positive share minus the negative, and watch for deltas around launches, campaigns, and incidents. Sentiment also feeds composite brand health scores. Used this way, as a trendline read consistently over time, it is genuinely useful: a sharp sentiment move is a reliable prompt to go look at what happened, even when the absolute number is soft.

The accuracy limits are real and worth internalizing. Sarcasm, irony, slang, and mixed feelings in a single post defeat classifiers regularly; sentiment often attaches to the wrong target, positive about the ad but negative about the brand; and performance drops outside English. Human annotators only agree with each other about eight times in ten, which puts a ceiling on what any tool can honestly claim. And because most posts are neutral, headline sentiment often swings on a small slice of the data. The discipline that saves you: read the verbatims behind any shift before reporting it.

How this shows up in Waldo

Waldo returns analysis with the receipts attached: every output links to sources, and the raw layer of posts and verbatims is one call away from any aggregate. When a sentiment number moves, checking it against the actual language people used is a single step, not an export and a spreadsheet.

Related terms and reading