I recently ran across a feature I’d never seen before on the Yahoo news site: real-time sentiment analysis.
Sentiment analysis can be a really valuable tool, such as when a company is evaluating social media conversation about it. Some high end programs also provide purchase and churn intent and predictive analytics about individual behavior. Those programs can start at few thousand dollars a month on up.
But what’s it mean for this story? In this case you can see that of the 7200+ comments made on this story about Trump’s lawyers threatening a 17 year old girl who created a website where kittens can punch him (don’t get me started…), about 59% were rated neutral, 35% negative and 6% positive. In aggregate the conversation was considered moderately negative.
Which means what? Context is everything. In this case is a “negative” comment negative about the girl, the site, or Trump? Those two comments on the left illustrate the problem perfectly: I would call the first one “negative” and the second “positive”, but they’re both anti-Trump. How did the software score them?
And since online sentiment is often 80-90% neutral, a story that is 59% neutral and 35% negative may technically be “mostly neutral” but by Internet standards it’s definitely slanting negative.
In the high end systems you train them by reviewing how they rated individual comments and override the original rating. And better make sure you have some multi-lingual people around: the high end systems can also monitor dozens of languages.
But I think Yahoo still has a little work to do to provide a valuable version of sentiment analysis for its stories.