Modeling reading development: Cumulative, incremental learning in a computational model of word naming

Abstract

Natural reading development gradually builds up to the adult vocabulary over a period of years. This has an effect on lexical processing: early-acquired words are processed more quickly and more accurately than later-acquired words. We present a connectionist model of reading, learning to map orthography onto phonology to simulate this natural reading development. The model learned early words more robustly than late words, and also showed interactions between age of acquisition and spelling-sound consistency that have been reported for skilled adult readers. In additional simulations, we demonstrated that age of acquisition effects are a consequence of incremental exposure to words in concert with changes in plasticity as learning proceeds, and are not due to uncontrolled differences in ease of reading between early and late-acquired words. Models which do not learn through cumulative training are unable to explain age of acquisition and related effects.

Introduction

Computational models of reading map orthographic input representations (letters or letter sequences) onto phonological (sound-based) and/or semantic (meaning-based) representations. In the model of Harm and Seidenberg (1999), for example, the input units represent the letters in written words while the output units represent phonological features of phonemes (individual speech sounds) in spoken words. The model is trained over time to generate the appropriate phonological output for each of the different inputs that correspond to the written words it learns to “read”. As Harm and Seidenberg (1999) observed, their model built on several previous computational models of reading (Plaut et al., 1996, Seidenberg and McClelland, 1989) and has since been extended by others (Harm and Seidenberg, 2004, Perry et al., 2007). A feature of all of these attempts to simulate visual word recognition is that the words that the model is required to learn are all entered into training together from the outset. But as Harm and Seidenberg (1999) also noted, natural reading development is not like that: young children embarking upon the task of learning to read will typically begin with a first reading book that contains 5 or 10 words, repeated several times. The next book will repeat those words and add a few more, and so on. For a reader of English it can take 10–15 years to acquire a full, adult reading vocabulary (Nation, 2009). This process of gradual, cumulative, incremental training that starts with a few words and builds a vocabulary over time is quite unlike the training implemented in existing computational models of visual word recognition.

The present study aimed to explore the consequences of training a computational model of reading in a more natural way. Our goal was to discover whether the final state of a developmentally-trained model would show an influence of the point at which words were entered into training and, if it did, whether that influence would be greater for words with irregular or exceptional spelling-sound correspondences than for words with regular, consistent correspondences. The reasons for investigating these possibilities are embedded in the literature on effects of age of acquisition (AoA) in adult reading. The next section will summarise the relevant findings.