Sentiment analysis
Sentiment analysis uses natural language processing to identify the opinion or emotion expressed in text, most often classifying it as positive, negative or neutral. It lets teams read the tone of thousands of reviews, comments or transcripts at once.
How it works
Early systems counted words from sentiment dictionaries: "love" scores positive, "broken" negative. Machine-learning models learn from labeled examples instead, and modern language models read whole sentences in context. Beyond overall polarity, aspect-based sentiment analysis attaches sentiment to specific topics, so "great screen, terrible battery" scores positive for screen and negative for battery.
Sentiment models struggle most with sarcasm, irony, mixed feelings and domain language, so teams validate a model on a hand-labeled sample of their own data before trusting it. Results are reported as the share of positive, negative and neutral mentions, overall and for each aspect, and tracked over time or compared across products.
A worked example
Across 40,000 reviews of a hotel group, aspect-based sentiment analysis finds overall sentiment 71 percent positive, but only 38 percent for "check-in". Reading 200 of the negative check-in reviews traces most of them to rooms not ready at the advertised 3 p.m.
Related
Automate this in Pulse
Sarcasm such as 'great, another update' is where sentiment models go wrong, and reading the Pulse transcript around a flagged remark shows the analyst which way it was meant.