UNDERSTANDING NORMAL AND IMPAIRED WORD READING - COMPUTATIONAL PRINCIPLES IN QUASI-REGULAR DOMAINS

A connectionist approach to processing in quasi-regular domains, as ex emplified by English word reading, is developed. Networks using approp riately structured orthographic and phonological representations were trained to read both regular and exception words, and yet were also ab le to read pronounceable nonwords as well as skilled readers. A mathem atical analysis of a simplified system clarifies the close relationshi p of word frequency and spelling-sound consistency in influencing nami ng latencies. These insights were verified in subsequent simulations, including an attractor network that accounted for latency data directl y in its time to settle on a response. Further analyses of the ability of networks to reproduce data on acquired surface dyslexia support a view of the reading system that incorporates a graded division of labo r between semantic and phonological processes, and contrasts in import ant ways with the standard dual-route account.