Millions of people use Twitter, Facebook, and other social networks. For these big internet companies— Twitter, Facebook, Amazon , and Google —this behavior is a treasure trove, a vast collection of personal information that can help them better understand who you are and, ultimately, show you what you dream about and want to buy. But that’s easier said than done. Their ability to discover this data depends on how good their algorithms are, and by “good algorithms,” we mean algorithms that can understand how a human thinks. As we know, machines aren’t very good at this.
But a new algorithm developed at Stanford University could help companies change that reality by giving computers the ability to reliably interpret our data. It's called Neural Sentiment Analysis, or NaSent for short. The algorithm seeks to improve current methods of written language analysis by drawing inspiration from the human brain.
NaSent is part of computer science known as deep learning, a new field that aims to develop programs that can process data in the same way the human brain does. The movement started in the academic world, but since then it has spread to web giants such as Google and Facebook.
“We’re trying to push deep learning emotional understanding closer to human levels of ability — given that previous models have plateaued in terms of performance,” says Richard Socher, the Stanford University graduate student who developed NaSent with artificial intelligence researchers Chris Manning and Andrew Ng.
The goal, Socher says, is to develop algorithms that can operate without constant human intervention. In the past, sentiment analysis has focused mostly on models that ignore word order or human experience. While this works in very simple cases, it will never reach human level of understanding.
Of course, despite promising early tests, the algorithm needs improvements. For example, if it encounters phrases, it seems to struggle to recognize words it hasn’t encountered before. To strengthen the system, Socher and his team have started pulling more data from Twitter and other databases. They have also created a live demo where someone can type in their own phrases. The demo creates a tree structure and assigns a polarity label to each word. If users think NaSent is misinterpreting a particular word or phrase, they can give it a different label, as Wired.
