Peter Kleiweg, John Nerbonne (1999)

`An FGREP investigation into phonotactics'

Abstract

We discuss experiments with neural networks being trained in a phonotactic processing task. A recurrent network not only learns to predict the next letter given a partial processed word, but also learns to represent the letters in a manner meaningful to the processing task. To this end, we use Miikkulainen's [1993] FGREP, augmented with an algorithm we call dispersion, to improve distinctness among the set of letter representations.

Our goal is to create a more realistic model of how humans might process natural language.