Phonology, reading acquisition, and dyslexia: Insights from connectionist models.

Abstract

The development of reading skill and bases of developmental dyslexia were explored using connectionist models. Four issues were examined: the acquisition of phonological knowledge prior to reading, how this knowledge facilitates learning to read, phonological and nonphonological bases of dyslexia, and effects of literacy on phonological representation. Compared with simple feedforward networks, representing phonological knowledge in an attractor network yielded improved learning and generalization. Phonological and surface forms of developmental dyslexia, which are usually attributed to impairments in distinct lexical and nonlexical processing "routes," were derived from different types of damage to the network. The results provide a computationally explicit account of many aspects of reading acquisition using connectionist principles.

Literacy effects on language and vision: Emergent effects from an amodal shared resource (ASR) computational model

“…This they argue is consistent with evidence indicating that literacy leads to increased sensitivity to smaller phonological units. Harm and Seidenberg's (1999) model suggests that increased componentiality may be a consequence of literacy, consistent with the psycholinguistic grain-size theory and the restructuring hypothesis. However, the model represented phonology in terms of individual phonemes, thus increasing the chances that the model will discover phoneme-level representations.…”

“…We also suggest that alternative methods of implementing cognitive efficiency within neural networks would also not simulate the pattern of the low-literate participants. For instance, reduction in the resources available for forming mappings between representations (Harm & Seidenberg, 1999) would impede the model's ability to learn tasks that both high and low literacy groups are capable of performing, such as vision to semantics, or phonological to semantics mappings. A further alternative implementation by reducing the overall levels of activation passing between layers within the model (Monaghan & Shillcock, 2004), or by increasing the threshold such that more activation is required before activating a response, is also unlikely to simulate the focused distinctions between high and low literate populations.…”

“…Harm and Seidenberg (1999) compared behaviour of a computational model trained to generate stable phonological representations of monosyllabic words, where the phonological representation was input as a set of phoneme features representing each phoneme in the word, to a model that in addition mapped orthographic representations onto the phonological representations, and assessed the effect of this literacy on the model's performance on a range of single-word phonological processing tasks. They found that when one phoneme within a word was affected by noise, the literate model was better able to restore the phoneme.…”

“…The model provides as an output a dynamic measure of the location of fixation across multi-object scenes over time, which is dependent on the integrated processing of information across all three modalities. The model also differs from previous models of effects on phonological processing where phonology has been used as both an input and output measure (Harm & Seidenberg, 1999). In contrast the chosen model has a different dependent variable 'eye gaze' therefore phonological manipulations are only indirectly related to behaviour.…”

“…There are several ways to simulate variation in cognitive efficiency in neural network models, including addition of noise to activations of units, reductions in processing resources to form mappings between representations, or reduction in overall levels of activation within the network (Harm & Seidenberg, 1999;Monaghan & Shillcock, 2004). We chose to simulate cognitive efficiency in the current models in terms of addition of noise.…”