Glossary · S

Sentiment analysis

Silhouette of a person at a laptop by a window

Definition

Sentiment analysis is a method that automatically estimates the tone of a conversation or message, for example as positive, neutral or tense.

Whether a conversation is going well is often clear after a few sentences. Sentiment analysis tries to make that call automatically, not for one conversation but for many at once.

#What is measured

For text, the analysis looks at wording and phrasing. "Again", "for weeks" or "finally" are signals that add up to a picture. For calls, it usually works on the transcript produced by speech analytics.

The result is a classification, such as positive, neutral or tense, often as a curve across the conversation. The single value matters less than the change: does the mood shift after a particular piece of information?

Sentiment over the course of a conversation: a hint for the agent, not a rating of her as a person. Sample.

#What it is good for

Trends. Which topics most often come with frustration?
Causes. After which product or process changes does sentiment drop?
Support. Team leads spot tense conversations and can offer help.
Quality. Conspicuous conversations are selected for quality review.

#Limits

Irony, dialect, jargon and cultural differences easily lead to misjudgements. For a single conversation the analysis is therefore unreliable. It only becomes robust across many contacts.

It is not suited to judging the performance of individual agents. A customer's mood says more about her request than about the person handling it.

#Co-determination and law

In Germany, technical systems capable of monitoring employees' behaviour or performance are subject to the works council's co-determination (Section 87(1) no. 6 BetrVG). Purpose, level of analysis and access rights belong on the table early.