Baby Talk
Learning language, researchers are finding, is an astonishing act of brain computation--and it's performed by people too young to tie their shoes
Learning words is one thing; learning the abstract rules of grammar is another. When Noam Chomsky first voiced his idea that language is hard-wired in the brain, he didn't have the benefit of the current revolution in cognitive science, which has begun to pry open the human mind with sophisticated psychological experiments and new computer models. Until recently, linguists could only parse languages and marvel at how quickly children master their abstract rules, which give every human being who can speak (or sign) the power to express an infinite number of ideas from a finite number of words.
There also are a finite number of ways that languages construct sentences. As Chomsky once put it, from a Martian's-eye view, everybody on Earth speaks a single tongue that has thousands of mutually unintelligible dialects. For instance, all people make sentences from noun phrases, like "The quick brown fox," and verb phrases, like "jumped over the fence." And virtually all of the world's 6,000 or so languages allow phrases to be moved around in a sentence to form questions, relative clauses, and passive constructions.
Statistical wizards. Chomsky posited that children were born knowing these and a handful of other basic laws of language and that they learn their parents' native tongue with the help of a "language acquisition device," preprogrammed circuits in the brain. Findings like Newport's are suggesting to some researchers that perhaps children can use statistical regularities to extract not only individual words from what they hear but also the rules for cobbling words together into sentences.
This idea is shared by computational linguists, who have designed computer models called artificial neural networks that are very simplified versions of the brain and that can "learn" some aspects of language. Artificial neural networks mimic the way that nerve cells, or neurons, inside a brain are hooked up. The result is a device that shares some basic properties with the brain and that can accomplish some linguistic feats that real children perform. For example, a neural network can make general categories out of a jumble of words coming in, just as a child learns that certain kinds of words refer to objects while others refer to actions. Nobody has to teach kids that words like dog and telephone are nouns, while go and jump are verbs; the way they use such words in sentences demonstrates that they know the difference. Neural networks also can learn some aspects of the meaning of words, and they can infer some rules of syntax, or word order. Therefore, a computer that was fed English sentences would be able to produce a phrase like "Johnny ate fish," rather than "Johnny fish ate," which is correct in Japanese. These computer models even make some of the same mistakes that real children do, says Mark Seidenberg, a computational linguist at the University of Southern California. A neural network designed by a student of Seidenberg's to learn to conjugate verbs sometimes issued sentences like "He jumped me the ball," which any parent will recognize as the kind of error that could have come from the mouths of babes.
advertisement
