Speech and Sentiment

It's interesting to see the growth in the speech analytics industry.  Call center agents are using it regularly now.  

"In fact, the business of analyzing words and their sentiments, called speech analytics, is a $214 million market, according to estimates from DMG Consulting, and used in finance, insurance, health, travel, retailing and telecommunications."


And what some companies, like Beyond Verbal, claim is possible.

What Language Reveals

Part of my vision of Refleta was to bring in disparate data streams and use them to form a picture of the consumer. Once I realized that device data was difficult to get at (mainly because most of middle America still doesn't own many devices) I turned to social media data. Turns out there's a lot you can glean from a Facebook post or Tweet. To start, my team built a sentiment analysis engine and ran it against the post and tweets data we were granted access to by alpha testers. We were able to discern when testers were feeling sad, happy, angry, etc. even without the specific use of those words.

But that wasn't a novel idea. In fact, there is plenty of research out there that shows the language you use says a lot about you.

And now there's even more interesting work in the space. Check it out.

Deep Learning

Google recently introduced a potentially very interesting open source tool called word2vec. It is software designed to understand the relationship between words without human intervention.  How? Through the use of "deep learning" - basically neural network models on steroids.  Essentially these models understand the features of an input (in this case words) and how those inputs relate to each other.


You don't have to understand the ins and outs to see how potentially useful software that can do this is.  One potential application is better understanding of tweets.  Typical sentiment analysis, for example, has a difficult time with tweets due to their short nature and the fact that they are often infused with symbols and sarcasm.  With deep learning techniques it may be possible to decipher the sentiment of tweets more accurately.  Can you imagine?

Big Data and Recruiting

Recruiting is using big data to uncover potential talent. But right now it's unclear if all the data currently aggregated can predict whether a candidate will be a cultural fit or work out in the long run.

Whether these algorithms work in the long run (where it's defined as the employee remaining employed by the employer for X years), only time will tell and it should be fairly simple to collect the retention data on employees.

When it comes to cultural fit understanding a candidate's personality is important. I suspect many of these algorithms may not be factoring in data around personality and soft skills. The ethical/moral issues aside, it's possible that sentiment analysis on the prospect's social media communication might give an employer a sense over time of a candidate's emotional stability. Or even an analysis of pronoun usage - studies have shown pronoun usage can highlight depression. But is it fair to do this?

Still, if you break down soft skills to key components like effective communication, leadership, conflict style, etc., I have to imagine there are ways to determine it with data. We reveal so much about ourselves - the independent variables are out there.

Sentiment Analysis

The bane of sentiment analysis is sarcasm. It's very difficult for a computer to know if "that was way fun" is actually someone exclaiming enjoyment or being sarcastic. Heck, it can even be tough for humans to tell the difference. A USC Annenberg lab out of Los Angeles is trying to solve the problem. They built their model using human annotation.

Only time will tell if they can crack the code. Some folks on Twitter simply use (*S) to indicate sarcasm, but it hasn't caught on universally. And that's a shame (*S).