Neural networks and child language development: a simulation using a modular neural network architecture

Abstract

A modular architecture comprising neural networks that employ both supervised and unsupervised learning strategies to learn complex patterns is presented for simulating child language development during infancy, particularly 9-24 months. It is argued that in order to simulate aspects of human learning, particularly the uniquely human language learning, it is important to use a modular architecture: an architecture that can realistically simulate the effects of the environment, through supervised learning networks, and the effects of self-learning or 'innate development', through unsupervised networks. We present ACCLAIM-a modular neural network architecture which synthesises two Kohonen maps, two Hebbian connection based networks and two backpropagation networks. ACCLAIM was used to simulate the learning of concepts, words, conceptual relations, semantic relations and simple word-order rules, mimicking two crucial phases of a child's language development, that is one-word and two-word language. ACCLAIM can produce child-like one-word and two-word sentences. The training data used for the simulation was taken from two major longitudinal studies of child language development.