New Equations May Help Robots Grasp Common Talk
|Psychologists Noah Goodman (right) and Frank Michael (left) wrote mathematical equations that quantify how people use context to understand descriptions. The equations could help computer scientists build context into robots.|
When people walk by a sign that says, "Baby sale — all stock 50 percent off," they know the store is selling cribs, strollers or snap-crotch playsuits, not actual babies. They're working off several pieces of context and common knowledge: They know that stores usually sell baby supplies rather than babies, that any operation selling babies wouldn't advertise this way, and that "baby sale" is common shorthand for "baby supplies sale."
Now, how can such a sense of context be added into the code of a future robot that reads and listens?
Two psychologists recently invented mathematical equations for the probability that someone is referring to a particular object and the probability that a speaker will choose a certain word to describe an object. The equations make it much easier to put context into computers. The work represents a step toward designing computer systems that better understand how people talk. Such research helps improve robotic systems that listen to people, such as automated customer service. It also may help develop therapies for people with language disorders.
The context the researchers studied was much simpler than what's needed to understand "baby sale," however. To test their equations, they had 745 study volunteers look at a series of three pictures. One three-picture set might contain a blue square, a blue circle and a green square, for example.
One group of volunteers was asked to bet on whether another person, while trying to describe the blue circle in one word, was more likely to say "blue" or "circle." The researchers told another group of volunteers that someone had said "blue," then asked the volunteers to guess which picture the speaker meant.
"We modeled how a listener understands a speaker and how a speaker decides what to say," said one of the study authors, Noah Goodman of Stanford University.
Goodman and colleague Michael Frank, also at Stanford, found that volunteers' bets corresponded with the equations Goodman and Frank had written.
The researchers are now using the equations to study hyperbole and sarcasm, which are also context-dependent and difficult for a computer to understand.
"It will take years of work, but the dream is of a computer that really is thinking about what you want and what you mean rather than just what you said," Frank said.
Frank and Goodman published their equations May 25 in the journal Science.