New evidence shows that babies learning to understand language rely more heavily than previously thought on patterns in the language they hear, according to results presented by a panel of scientists at the annual meeting of the American Association for the Advancement of Science.
The new data alter a long-established trend that had swung the emphasis in language learning research onto pre-programmed mechanisms built into the human brain.
“The thinking for years had been that something as complex as language could not be acquired with fairly simple learning mechanisms that rely on word frequency, associations between adjacent words, and other statistical phenomena,” says Rebecca Gomez, a presenter at the panel and an assistant professor of psychology in The Krieger School of Arts and Sciences at The Johns Hopkins University. “But we now have new methods for assessing learning in infants, and we’ve been able to demonstrate that this type of learning is in fact very strong.”
Gomez uses exposure to artificial languages to study language learning in infants and in adults.
“I’m interested in finding out how far sensitivity to statistical information in language can actually take an infant in the language learning problem,” Gomez says. “Artificial languages allow me to eliminate any interference from prior language learning and focus more closely on what the learner is responding to.”
Artificial language words have no meaning and include such examples as “pel,” “wadim,” and “jic.” But Gomez organizes them to mimic grammatical patterns in natural language.
“I’ll find a regularity in natural language that I think might be contributing to learning in an important way, and I’ll extract that and put it into an artificial language in a very simplified form,” Gomez explains.
Gomez investigated the conditions under which infants switch the focus of their learning from patterns involving adjacent words to patterns involving words separated by other words in a sentence. The two learning mechanisms, known respectively as adjacent dependencies and long- distance dependencies, are both important in language learning.
“It’s been shown in prior research that adjacent dependencies are very useful in learning, and might even be a kind of default mode for language learners,” says Gomez. “We set out to test whether we could switch infants out of that default mode by making long-distance learning more useful.”
Gomez exposed 18-month-old infants to three-word strings of an artificial language. The experiments took place in a booth with speakers on the right and left of a central chair, where a parent sat with the infant. In the initial, training period of exposure, the first and third words in the sequence were linked and limited in variety to two pairs of words. If the first word was “pel,” the third word was always “jic”; if the first word was “vot”, the third word was always “rud.” The second word varied, and was taken either from a small set of words or from a very large word set.
After the training period, the infants were then re-exposed to a mixture of sentences that were either “grammatical,” agreeing with the rules established in the training period that linked the first and third words, or “ungrammatical,” conflicting with the rules. Infants who learned the distinction listened longer to an “ungrammatical” burst of speech, as if the “grammatical” stimuli were no longer interesting.
“What we actually found is that there’s no discrimination between ‘grammatical’ and ‘ungrammatical’ when the middle element only draws on a small set of words,” Gomez says, noting that when the middle element word set is small, infants can easily use the first word to predict the second, or the second word to predict the third, both of which are adjacent dependencies.
“With a large enough set size, actually 24, all of a sudden infants and adults will learn the dependency between the first and third words. Twelve-month-olds do not show this ability, suggesting it develops with age.
“This is an important demonstration, because it shows that in addition to having access to multiple learning mechanisms, infants can switch to a new mechanism when the original mechanism is no longer optimal,” says Gomez.
Gomez and other infant researchers are optimistic that they are seeing the first signs of a shift in language learning research.
“We’ve always known that some aspects of language had to be learned, rather than pre-programmed, because there are differences in languages,” she says.
Scientists have recently discovered an increasing number of statistical regularities in language that infants can use to help learn language, and a more even-handed view of the balance between associative learning and preprogrammed learning mechanisms built into the brain is starting to emerge.
Gomez notes that she doesn’t think the debate between the two sides should be aimed at eliminating one viewpoint, and that she feels both sides have valid contributions to make to science’s understanding of how infants learn language.