By Dennis Baron
A computer at Carnegie Mellon University is reading the internet and learning from it in much the same way that humans learn language and acquire knowledge, by soaking it all up and figuring it out in our heads.
People’s brains work better some days than others, and eventually we will all run out of steam, but the creators of NELL, the Never Ending Language Learner, want it to run forever, getting better every day in every way, until it becomes the largest repository imaginable of all that’s e’er been thought or writ.
Since the first “electronic brains” began to appear in the late 1940s, it has been the goal of computer engineers and the occasional mad scientist to fashion machines that think and learn like people do. Or at least machines that perform functions analogous to some aspects of human thought, and which also self-correct by analyzing their mistakes and doing better next time around.
Setting out to create an infinite and immortal database is a big task: there’s a lot for NELL to learn in cyberspace, and a whole lot more that has yet to be digitized. But since NELL was activated a few months ago it has learned over 440,000 separate things with an accuracy of 74% which, to put it in terms that any Carnegie Mellon undergraduate can understand, is a C. In contrast, I have no idea how to count what I’ve learned since my own brain went on line, and no idea how many of the things that I know are actually correct, which suggests that all I’ve got on my cerebral transcript is an Incomplete.
NELL’s programmers seeded it with some facts and relations so that it had something to start with, then set it loose on the internet to look for more. NELL sorts what it finds into categories like mountains, scientists, writers, reptiles, universities, web sites, or sports teams, and relations like “teamPlaysSport, bookWriter, companyProducesProduct.”
NELL also judges the facts it finds, promoting some of them to the higher category of “beliefs” if they come from a single trusted source, or if they come from multiple sources that are less reliable. According to the researchers, “More than half of the beliefs were promoted based on evidence from multiple [i.e., less reliable] sources,” making NELL more of a rumor mill than a trusted source. And once NELL promotes a fact to a belief, it stays a belief: “In our current implementation, once a candidate fact is promoted as a belief, it is never demoted,” a process that sounds more like religion than science.
Sometimes NELL makes mistakes: the computer incorrectly labeled “right posterior” as a body part. NELL proved smart enough to call ketchup a condiment, not a vegetable, a mislabeling that we owe to the “great communicator,” Ronald Regan. But its human handlers had to tell NELL that Klingon is not an ethnic group, despite the fact that many earthlings think it is. Alex Trebek would be happy to know that, unlike Sean Connery, NELL has no trouble classifying therapists as a “profession,” but the computer trips up on the rapists, which it thinks could possibly be “awardtrophytournament” (confidence level, 50%).
NELL knows that cookies are a “baked good,” but that caused the computer to assume that persistent cookies and internet cookies are also baked goods. But that’s not surprising, since it still hasn’t learned what metaphors are—NELL is only 87.5% confident that metaphors are “tools” (plus, according to NELL, there’s a 50-50 chance that metaphors are actually “book writers”).
Told by its programmers that Risk is a board game, NELL predicts w