It’s one of the biggest challenges in computing — getting a machine to think like a human. This long-standing computational problem is one that captivates public interest, as evidenced by the much-hyped 1997 chess match between IBM’s Deep Blue supercomputer and world champion Garry Kasparov. The machine won a six-game match by two wins to one with three draws, but more importantly, the game brought about an international love affair with supercomputing. It’s actually been a long-time since supercomputing has so moved the world. Arguably, not even the 2008 accomplishment of breaking the petaflop barrier created such an intense international stir. There’s something about a machine being able to do something so seemingly human, like playing a centuries old game of strategy, that touches hearts and minds more than achieving some remote-sounding number of computations per minute.
But it turns out that playing chess is actually not such a great predictor of “human-ness” for a machine. It’s actually pretty easy for computers to beat humans at well-defined tasks such as playing rule-oriented games or predicting weather changes. What’s not so easy is for machines to understand language — indeed semantics is one area where humans still have the clear edge.
In a recent New York Times article, author Steve Lohr covers current advances in the field of computational semantics being undertaken by a group of researchers at Carnegie Mellon University.
Team leader Tom M. Mitchell, a computer scientist and chairman of the machine learning department, outlines the nature of the challenge: “For all the advances in computer science, we still don’t have a computer that can learn as humans do, cumulatively, over the long term.”
The researchers are working on a project, called the Never-Ending Language Learning system, or NELL. NELL is fed facts, which are grouped into semantic categories, such as cities, companies, sports teams, actors, universities, plants and 274 others. Examples of category facts are “San Francisco is a city” and “sunflower is a plant.” NELL has been able to glean 390,000 facts by scanning hundreds of millions of Web pages. The larger the pool of facts, the more refined the system will get.
So much of language understanding is predicated on an underlying knowledge base and that’s what NELL is developing. In the sentence: “The girl caught the butterfly with the spots,” a human reader innately understands that “spots” refers to the butterfly because the human knows that butterflies are likely to be spotted whereas girls are not. Such “basic” knowledge that we take for granted confounds the computer. This general knowledge can only be learned, and that’s why NELL was programmed to learn so many facts.
There have been similar attempted artificial learning programs, but NELL is different in that the system is being taught to learn on its own with little assistance from researchers, although if they notice NELL has gotten something blatantly wrong, like classifying an Internet cookie as a baked good, they will correct those errors.