Baby Statisticians Learn Language

By Joanne Kenen

WASHINGTON (Reuter) - Infants are surprisingly adeptlearners, able to use simple statistical skills to analyzesounds and recognize which ones form words, researchers saidThursday.

The findings run counter to much conventional wisdom abouthow babies acquire language and suggest that they are far moresophisticated learners than previously believed.

``We've underestimated what babies are capable of,'' saidJenny Saffran, a University of Rochester graduate student whodesigned the experiment involved. A report she co-wrote with herprofessors appears Friday in the journal Science.

Linguists agree that babies use a variety of cues to learnlanguage, including pitch, stress, pauses and rhythm.

Saffran's experiment filtered out all those cues and simplystudied if babies could apply elementary statistical principlesto sounds. She found that, by computing which combinations theyheard most often, the infants were able to figure out which oneswere words.

The researchers do not suggest the babies understood wordsthe way adults do, as sound symbols attached to meaning.

But they said the experiment establishes that afteronly twominutes infants can remember familier sequences of sounds,distinguish them from unfamiliar ones, and make thesedistinctions based on statistical recurrence without any othernormal speech clues.

``It's possible, although it's a question for futureresearch, that babies' learning abilities can actually accountfor far more complicated aspects of language acquisition thanpeople had imagined,'' Saffran said in an interview.

For instance, her findings suggest that babies can usestatistics to figure out that ``pretty'' and ``baby'' are words,but that the syllables ``ty-ba'' -- the end of pretty and thebeginning of baby -- are not.

Linguists and psychologists have long debated howmuch of ababy's language abilities were innate and how much was based onexperience, with the consensus in recent years that much of itwas innate. The Science report provides evidence for thelearning end of the spectrum.

``The nature of this learning is surprising; a purelyinductive statistically driven process, based on only twominutes of incidental input with no reward or punishment,''University of California at San Diego psychologists ElizabethBates and Jeffrey Elman wrote in a Science commentary.

``It contradicts the widespread belief that humans cannotand do not use generalized statistical procedures to acquirelanguage,'' a theory that has gained widespread acceptance sincelinguist Noam Chomsky began arguing it 40 years ago.

Saffran did two studies, each with 24 eight-month-oldinfants who heard a voice synthesizer in a flat monotone speakfour words of a nonsense language for two minutes with no pausesbetween syllables. It sounded like ``bidakupadotigolabubidaku.''

To test their recognition of the words later,Saffran's teamrelied on the ``novelty preference'' effect. Babies pay moreattention to something new than to something familiar.

In one experiment, the babies were played those ``words''mixed in with ``nonwords'' containing some of the same syllablesin different order. In the other, the infants heard a mix ofthe ``words'' and word fragments.

In both groups, the babies paid more attention to the newmix of sounds than to the familiar ones like ``bidaku.''

``No one believed that young infants could be such rapidlearners. Now we have evidence that months before infants beginto produce words, they can very rapidly learn which sounds arelikely to go together to form words,'' said professor RichardAslin, a co-author of Saffran's article.

Article from Mercury Online service,12:37 PM ET 12/12/96