Sentiment analysis

Sentiment analysis is software that reads a piece of text or speech, such as a post, a product review or a customer call, and decides whether the person sounds pleased, angry or neutral. It lets a company read thousands of comments a day, which no team of people could read by hand.

Here is how it works in general. The software collects the text: posts, reviews, survey comments, emails or chat messages. For a phone call, speech recognition first turns the speech into text. The software then gives each piece a label, positive, negative or neutral, often with a score, for example from −1 to +1. Better systems go further. They label each topic within the text separately. So "the delivery was fast but the shoes are tight" counts as pleased about the delivery and unhappy about the fit. This is called aspect-based sentiment analysis. Finally, the software adds up the scores by day, product or branch, and alerts a person when the share of negative comments suddenly rises.

The method has changed over the years. Early tools used lists of words: "excellent" scored as positive, "worst" as negative. These lists failed on sarcasm, such as "Great, my order is late again", and on Hinglish. Later tools used machine-learning models trained on examples that people had labelled. Today, large AI models read the whole sentence and its context. They understand sarcasm and mixed languages much better, and they can also explain why a comment is negative.

Take, for example, a phone brand after a new launch. A social listening tool reads tens of thousands of posts about the phone, and shows that complaints about its battery tripled overnight.

Consider an insurer's customer care centre. Conversation intelligence software scores every recorded call, so a supervisor can listen first to the calls in which the customer grew angry.

Imagine a company that runs a staff survey every quarter. Its employee engagement survey tool reads 5,000 anonymous comments, and reports which teams sound most frustrated, and about what.

How is sentiment analysis different from topic analysis? Topic analysis finds what people are talking about, such as the delivery, the price or the app crashing. Sentiment analysis finds how they feel about it. Customer experience management software uses both together: it shows which topics make customers the angriest, so the company knows what to fix first.