Research at Stanford may lead to computers that understand humans
June 6, 2012
After decades of trial and error, artificial intelligence applications that aim to understand human language are slowly starting to lose some of their brittleness. Now, a simple mathematical model developed by two psychologists at Stanford University could lead to further improvements, helping transform computers that display the mere veneer of intelligence into machines that truly understand what we are saying.
The Loebner Prize is a competition of the world's best "chatbots" - computer programs designed to simulate how a human interacts in a normal written conversation - that promises a grand prize of US$100,000 to the first program that can interact with another human in a natural way, undistinguishable from another human. The competition started in 1991, but the prize is still up for grabs and the transcripts from each year's winners tell us just how far we are (the answer: very) from ever reaching that goal.
However, there is hope yet. A new trend has emerged in the past few years and has led to the development of technologies like Siri, iPhone's "personal assistant." It entails using mathematical tools, namely probability and statistics, to try and model how people use language to communicate in social situations. The work at Stanford builds directly on this branch of research.
"A key part of this work was made possible by other research on Bayesian modeling, which uses principles from statistical reasoning to help us understand the structure of the mind," Assistant Professor Michael Frank told Gizmag. "There has recently been a lot of interest in using Bayesian models to understand how people think about other people - models of social cognition."
The researchers enlisted over 700 participants to take part in an online experiment in which they saw a set of objects and were asked to guess which one was being referred to by a particular word. By doing so, they modeled how a listener understands a speaker and how a speaker decides what to say. The researchers then created a mathematical equation to predict human behavior and determine the likelihood of someone referring to a particular object.
The mathematical model helps predict pragmatic reasoning and may lead to the manufacture of machines that can better understand inference, context and social rules, eliminating much of the communication breakdown that so often takes place in even the most advanced of today's natural language processing algorithms.
"We already use little bits of natural language to communicate with computers, like Siri on the iPhone, Google, or the phone answering systems that many companies now use. But these interfaces are still sometimes very frustrating to use. We hope that our work will help engineers build systems that can do a little better at guessing what human communicators mean," says Frank.
The researchers are currently applying their model to study hyperbole, sarcasm and other aspects of language. Their work is detailed in a paper published in the journal Science.