Artificial Intelligence and Language Acquisition - Modelling An Artificial Language Acquisition Mechanism for Representing Inflections of Words Within the Context of Artificial Mind System and Kernel Memory

Tetsuya Hoya and Noriko A. Toyoda
Saitama Institute of Technology and Meikai University

The approaches in terms of artificial intelligence are also beneficial to modelling the language faculty in general. A careful investigation of not only the results obtained utilising artificial devices but also the underlying theory can still provide other insightful accounts in the general study of language. In the talk, I would like to focus upon a specific approach of it, namely, how we can develop an artificial language acquisition mechanism that can explain the Berko’s WUG-test, i.e. how children acquire the plural form of an unknown word like WUG, within a uniform context of artificial mind system (AMS) based upon a new connectionist model called kernel memory (KM), both of which were summarised in the recent monograph (Hoya, 2005).
Pinker (Pinker 2000) challenged the Rumelhart and McClleland's pattern association model using the so-called multi-layered perceptron neural network (MLP-NN) by enumerating several points where the model failed to simulate the human language faculty. In contrast, from an engineering point of view, many researchers are now aware that the practical implementation of MLP-NN based approaches is quite often hindered due mainly to their numerical instability.
AMS is a holistic model of artificial brain proposed from an engineer’s view, with weaving the ideas from various psychological/cognitive studies, and the approach essentially follows the modularity principle of mind (Fodor, 1983 and Hobson 1999); the functionality of mind can be subdivided into i) input, ii) assimilating, and iii) output modules, the scope of which is also modelling-affinitive. In the AMS, each module, as well as their interactive processes in between, is then modelled by means of KM. Then, I would like to emphasise that the AMS&KM approach can not only get rid of various problems that are fundamental to conventional connectionist models but also be extended to the study in general language acquisition such as the WUG-Test. Although at first sight the study of WUG-test seems to deal with a small fraction of language acquisition (i.e. inflection of a noun), it can hopefully reveal one of the core principles of how humans acquire language. I would therefore like to take up this example to show how the approach by AMS&KM can be applied to practical development of artificial language acquisition devices, whilst demonstrating some preliminary simulation results obtained so far within the engineering context.

Fodor, J. (1983): The Modularity of Mind. MIT Press.
Hobson, A. (1999): Consciousness and Brain. New York: W. H. Freeman and Company.
Hoya, T. (2005): Artificial Mind System - Kernel Memory Approach. Studies in Computational Intelligence (SCI): Vol. 1. Heidelberg: Springer-Verlag.
Pinker S. (2000): Words and Rules - The Ingredients of Language. Perennial