How Sentiment Analysis Can Avoid Becoming Snake Oil
We’ve already covered a lot of the problems with Sentiment Analysis, but we thought we’d try being positive for once.
Sentiment is obviously important, but it’s not the only important thing in language. Much more important is isolating topics and themes within text corpora.
You’ll discover much, much more than you would have done with a “sentiment first” approach.
Are Girls Rude?
In our last major retail project, we discovered an overwhelmingly strong association between the terms “girl”, “rude” and “unhelpful”. Sentiment analysis gives us just enough information to reach the wrong answer: People are unhappy with female staff because they’re rude and unhelpful.
However, when we looked at the text sources in detail, the real problem became clear. It wasn’t just that “girls” were frequently rude: When female staff were called rude or unhelpful, they were almost exclusively referred to as “girls”.
It’s all Downton Abbey downtown at your local supermarket.
The problem isn’t about staff. It’s about the customer’s prejudices and the store’s lack of provision to combat them. We call this the “Downton Abbey” problem.
Shop assistants are at the bottom of the social pecking order. And whether or not we like to admit it, our society devalues female ability. We’re therefore working with two kinds of inferiority broadcasting. Hence “girl”, and its in-built senses of “frivolous”, “inexperienced”, “junior”.
Since we expect our social inferiors to be deferential, any display of non-deferential behaviour feels rude. And any female member of staff not being suitably deferent is dismissed as a rude, unhelpful girl.
It’s a real problem, but who in your Customer Experience team would talk about it without fear of being branded a class warrior?
What’s our solution? As usual, it’s all about messaging. Make it clear to the customer just how much experience each staff member has, so they know exactly what to expect from them. Acknowledge the underlying assumptions while deftly nullifying them.
And you can’t get that far with a “negative” flag. Sentiment was only one part of the problem here, and you can’t solve a problem when you’re being told you have a different one.
Think you’ve got a Downton Abbey problem? Maybe something a bit more serious, like a Game of Thrones problem (in which your staff are unable to help customers because they’ve been violently murdered)? If so, email Chris to discuss how text analytics can improve your marketing and customer relationship management.