Representation, Learning, Generalization and Damage in Neural Network Models of Reading Aloud

We present a new class of neural network models of reading aloud based on Sejnowski & Rosenberg’s NETtalk. Unlike previous models, they are not restricted to mono-syllabic words, require no complicated inputoutput representations such as Wickelfeatures and require no preprocessing to align the letters and phonemes in the training data. The best cases are able to achieve perfect performance on the Seidenberg & McClelland training corpus (which includes many irregular words) and in excess of 95% on a standard set of pronounceable non-words. Evidence is presented that relate the output activation error scores in the model to naming latencies in humans. Several possible accounts of developmental surface dyslexia are identified and on various forms of damage the models exhibit symptoms similar to acquired surface dyslexia. However, their inability to account for lexical decision, the pseudohomophone effect and phonological dyslexia indicate that we will still need to introduce an additional lexical/semantic route before we have a complete model of reading aloud. Nevertheless, the models’ simplicity, performance and room for improvement make them a promising basis for the graphemephoneme conversion route of a realistic dual route model of reading. Edinburgh University Technical Report 94/1 – November 1994