An FGREP Model of Phonotactics

Peter Kleiweg, John Nerbonne and Ivelin Stoianov
Alfa-informatica, BCN, U. Groningen
{kleiweg,nerbonne,stoianov}@let.rug.nl

Mikkulainen introduced FGREP architectures as a means of finding optimal data representations in neural networks. In FGREP models, backpropagation of error and weight adjustment is continued past hidden layers into the layer of input data. These too are adjusted to minimize error. The current paper applies FGREP to the problem of Dutch phonotactics, in particular, the problem of distinguishing allowed from disallowed sequences of phonemes, explored in earlier CLIN meetings by Tjong Kim Sang and by the present authors. Earlier work aimed at learning all Dutch monosyllables as these are found in CELEX, a somewhat unnatural level of generality if compared to human ability. We therefore also experiment with a simplification of the task, to make it more comparable to human performance.