Modelling the Acquisition of Reading Skills

Existing direct route connectionist models of reading aloud (i.e. text to phoneme conversion) can now learn to perform on their training data and unseen words with accuracy comparable to that of humans. They also exhibit a number of developmental, reaction time and brain damage effects that are observed experimentally. However, various deficiencies (such as their failure to perform reliable lexical decision and to show the pseudohomophone effect) indicate the importance of incorporating some form of lexical/ semantic system into these models. In this paper we present a preliminary investigation of this problem. A general framework is outlined that models network activation flow between orthography, phonology and semantics. Explicit simulations of the learning process in small scale networks of this type show how this approach can account for many aspects of reading (and related tasks) that are not possible without the influence of the semantic system.