Tuesday, December 9, 2008

Sentimental Software

Software generally isn't sentimental.

Ask a software program to find everything on the Web that talks about Barack Obama or Flat Panel TVs or Interocitors and it does pretty well. It doesn't matter how simple or obscure, how abundant or scarce, computers are very good at finding stuff.

However, ask a software program to find everything on the Web that says nice things about Obama, TVs or Interocitors and the program gets queasy and confused and apparently random in how it responds. Humans may understand nice, nasty and neutral in astonishing detail, and pick up on the slightest nuance in how things are expressed, but computers don't. We understand plain spoken feelings, we laugh at a joke, we grok sarcasm. Said another way, humans feel things, computers can't.

Lots of people have tried to make computers more sensitive creatures by writing elaborate programs that attempt to disentangle the tone of what humans write. They usually start by trying to get computers to understand written expression in the way we do, grammatically, syntactically and lexically. It's tricky. Stuff my three year-old understands in I Love You, Goodnight might be obvious to most software programs that try and detect tone, but my nine year-olds' Harry Potter would be a real stretch. And my New York Times would bamboozle most sentiment detection systems a lot of the time.

A lot of people are working on this problem, because there's a lot of commercial applications (Wikipedia has a nice summary of the technology, the business potential and some of the companies that are trying to solve this problem). It has great potential in automated trading applications, brand management, PR, politics... the list is long. Today, automated tone or sentiment detection is often built into media monitoring systems, but most companies selling these products acknowledge the results can be erratic and instead rely on human readers -- now there's a really, really interesting job . In my day job, I'd love an application that could accurately assess the nice-nasty-neutral tone of the stories my company gets. And I have to believe there's a great start-up opportunity in this space, especially serving some vertical market sectors.

If anyone knows of good sentiment detection software, let me know.