Mind reading computers

Over the next couple weeks my posts may be a little bit lazy. I just bought a new condo (well new to me), and I have to put a whole bunch of work into it before I move in. But in anycase here's an interesting story from Forbes:

We've all heard the phrase before: It's not what you say, it's how you say it. Several years ago, Alex Pentland, a professor at the Massachusetts Institute of Technology's Media Lab, decided to find out whether or not this piece of conventional wisdom is true.

After studying the psychological literature, Pentland began to develop several simple computer-based systems to measure what scientists call "social signaling"--nonlinguistic behavior such as vocal intonation, facial movement and gesturing--during a conversation. He settled on four basic markers, which he designated activity, engagement, stress and mirroring.

Activity, or the fraction of time each person talks, is the simplest measurement. Not surprisingly, the more someone talks, the more interested in the conversation he or she is assumed to be. Engagement measures how the speakers influence one other. Are they talking in smooth succession, or are there long pauses between utterances? Does one speaker hesitate more often than another? Stress measures the variation in the pitch and volume of each speaker's voice to determine whether his or her voice betrays any discomfort or anxiety. Mirroring, finally, is a measurement of the speaker's empathy--how frequently he or she adopts the vocal intonations and inflections of the other, or repeats short phrases such as "uh huh" and "OK" if the other says them first.

Pentland then collaborated with other researchers in fields ranging from psychiatry to business in order to put his markers to the test. Could they be used to predict what would happen in various social situations, say, landing a date, or getting a job or a raise?

In most scenarios, the predictions that Pentland and his colleagues were able to make turned out to be shockingly accurate. Using nothing but these simple, nonlinguistic clues--and analyzing conversations that lasted between five minutes and just over an hour, depending on the experiment--the researchers were able to calculate the likelihood of a given outcome with an average accuracy rate of almost 90%.