A Self-Organizing Neural Network Model of Lexical and Morphological Acquisition

University Of Richmond, Richmond VA

Investigators

Abstract

A crucial aspect of human language learning is the learner's ability to generalize existing patterns to novel instances. The issue of generalization is a focal point of current debates on mechanisms of language acquisition. In the last ten years, many researchers have used the acquisition of the English past tense as an example to debate whether language acquisition should be viewed as a symbolic, rule-based learning process or as a connectionist, statistical learning process. However, most of this debate has revolved around a specific cluster of connectionist models, the back-propagation network as a model of language acquisition. The back-propagation algorithm, especially in the context of language acquisition, has several limitations now. In this study, we explore self-organizing neural networks, in particular, the self-organizing feature maps as models of language acquisition. In contrast to back-propagation, the self-organizing network uses unsupervised learning that requires no presence of a supervisor or an explicit teacher; learning is achieved entirely by the system's self-organization in response to the input environment. Moreover, multiple self-organizing networks can be connected via Hebbian learning, a biologically motivated co-occurrence learning mechanism. Our project involves first the development of a connectionist model of language acquisition based on principles of self-organization and Hebbian learning. It further involves the modeling of the acquisition of the lexicon and morphology, with respect to the emergence of structured lexical representations and the relationship between generalization and representation. These studies should allow us to determine (1) whether and how structured lexical representations can emerge from self-organizing processes of learning, rather than being available innately; (2) the extent to which generalization is a function of the new representation; and (3) whether and how self-organizing processes lead to the recovery from generalization errors. We propose that self-organization and Hebbian learning provide the necessary computational and psycholinguistic mechanisms for lexical representation, morphological generalization, and recovery in language acquisition. Our project will integrate previous empirical and modeling results in a new light, and offer a new theoretical perspective as well as a methodological tool for the study of the acquisition of the lexicon and the morphology.