A Major Finding: Outlook Negative for Sentiment Analysis

A study published by the Zurich University of Applied Sciences tested nine state-of-the-art sentiment analysis APIs against a set of 30000 verbatims pre-scored by humans. It found that the average accuracy is only 50%. You can get that by flipping a coin.

 Screen Shot 2014-01-22 at 14.27.33

The paper suggests we should combine tools to overcome their individual weaknesses – and raise accuracy by as much as 9%. Really, why bother?

Why would you settle for knowing what the consumer might feel when you can get more reliable insights by using text analytics to study themes. After all, people talk about what matters to them, and text analytics can tell you what people are talking about. Then, if you add on a layer of linguistic interpretation, and you’ll find out why they’re talking about it.