Printer Friendly

Neural networks for learning verbs.

Children show a remarkable ability to learn language, whatever its idiosyncrasies. In English, one well-studied example of this process involves the effort required to learn the past tense of verbs. Although expressing the past tense usually involves simply adding the ending "-ed" to the verb's present tense, English has a number of important exceptions, especially among frequently used verbs.

Psychologists have theorized that children learn such forms by going through a set of stages. At first, they merely memorize a few important, common verbs, remembering that "go" and "went" go together, "make" and "made:' "play" and "played," and so on. As their vocabulary expands, they presumably learn a rule: Given an unfamiliar verb, add the ending "-ed." However, in this second stage, children often apply the rule indiscriminately - even to such verbs as "go," saying "goed" or "wxnted." In the third stage, they realize they shouldn't use the rule in every case, and they proceed to learn the correct form of each of the exceptions.

Psychologist David Rumelhart and his co-workers at Stanford University have now developed a neural-network computer model that challenges the view that such learning occurs in stages by the adoption of new rules. Their simple model stumbles into the same kinds of mistakes that children make as they learn their verbs, yet it doesn't rely on the addition of new rules to achieve the desired results.

A neural network consists of a number of "elements"roughly corresponding to neurons in the human brain connected together in various ways. Learning involves changes in the strengths of these connections as the network responds to various types of inputs. Rumelhart and his colleagues used a simple type of neural network. They offered it examples of verbs in both their present and past forms, which strengthened connections activated by frequently occurring verbs. The network then attempted to predict the past tense of new verbs.

Following the same pattern as young children, the network initially tended to add "-ed" to new verbs. As it learned additional examples, it began to organize its knowledge of exceptions into clusters- for example, "ring" and "rang," "sing" and "sang" would belong to the same cluster. Eventually, the network learned how to handle a wide range of exceptions.

"To a first approximation, a very simple learning system recapitulated data collected from children:' Rumelhart concludes.

The findings suggest that children may learn verb forms not by remembering and invoking rules they have been taught, but simply by listening and generalizing from one example to another. The learning process itself yields the observed behavior, Rumelhart notes. "That's all you need to explain this kind of result."

"We don't expect our model to be perfect:' he adds. "It provides a hypothesis rather than an answer and suggests what else might be tried or tested."
COPYRIGHT 1993 Science Service, Inc.
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1993, Gale Group. All rights reserved. Gale Group is a Thomson Corporation Company.

Article Details
Printer friendly Cite/link Email Feedback
Title Annotation:computer model suggests how children learn verb forms
Author:Peterson, Ivars
Publication:Science News
Article Type:Brief Article
Date:Feb 27, 1993
Previous Article:Using the sounds of hearing.
Next Article:Atoms in buckyball cages.

Related Articles
Neural networks set sights on visual processing in brain.
High society on the brain.
Backing up 'back prop.' (back propagation in neural networks)
The brain in the machine: biologically inspired computer models renew debates over the nature of thought.
Possible applications of neurocomputing in defense.
A neural network - could it work for you?
The AI factory; how artificial intelligence will create 'smart plants.' (Cover Story)
Neural-net neighbors learn from each other.
Decision support software for tax.
Thinking computers.

Terms of use | Copyright © 2017 Farlex, Inc. | Feedback | For webmasters