Letter-to-sound rules do not seem to provide a more basic level for the child to bootstrap into reading, whereas Seymour's logographic route does. There is evidence that phonetic reading happens sometime in the course of development, this is likely to result from the child's induction of a procedure which encodes the letter-sound mapping in some form. This could well be a PDP network or a set of rules. I'm still not sure what the mature model should look like, but as Coltheart (1991) points out in his commentary, it does not look like Seidenberg and McClelland's PDP. Neither does it look like a set of letter-to-sound rules.
2. Skoyles proposes that children start reading using an abstract letter-to-sound procedure. The use of such a procedure is intended to answer the criticism that a neural network model, such as that of Seidenberg and McClelland (1989), requires extensive error correcting feedback in order to learn the orthography to phonology mapping. All that is achieved in this proposal is that we now have to decide where the letter-to-sound rules come from. The proposal would have some merit if it could be argued that the letter-to-sound mapping were more basic or more easily learned than the "non-phonetic" procedure it is used to train. It is not clear that this is the case.
3. If letter-to-sound rules are to be effective for word recognition, they need to have a reasonable degree of accuracy. Rule sets exist that can translate around 80% of words correctly and the rest almost correctly (Elovitz et al., 1976; Venezky, 1970), however, they are the result of careful analysis of a large number of print words and their phonological forms. In order to learn a set of rules of even moderate accuracy, the child would need to be trained on a significant number of print-phonology pairs -- just as the PDP network is, in fact. The alternative is that the child is taught these rules explicitly, for instance, via phonics instruction. This would imply that children not taught phonetically would be at a severe disadvantage, which is not the case (for instance see Johnston and Thompson 1987).
4. My own experiments with algorithms that learn letter-to-sound rules from examples have resulted in rule-sets capable of slightly better performance than Seidenberg and McClelland's PDP network, although not up to the standard cited by Besner et al. (1990) for mature readers. Coltheart et al. (1991) have achieved even better performance with a similar algorithm using a different set of assumptions; their rulesets are capable of naming 85% of the non-word stimuli taken from Besner et al. (1990) correctly.
5. So it seems that the child does not face an easier task learning letter-to-sound rules than the one used in training a PDP network. What does come first then? Seymour and Elder (1986) provide evidence for a preliminary stage where the child recognises words as visual stimuli -- without any reference to the phonological information that they encode. This logographic procedure is certainly more basic than any procedure which uses learned letter-to-sound associations and could provide the source of training examples that Skoyles seeks. Unfortunately, however, the logographic route is definitely non-phonetic. Stuart and Coltheart (1988) suggest that this logographic stage might be bypassed when the child has well developed phonological skills. The mechanism they suggest by which phonological knowledge could be producing these errors relies on the pre-establishment of something akin to Seymour's logographic route. My own work has tried to make the logographic process more concrete in a computational model and provides an account of the Stuart and Coltheart data based on this route alone.
6. In summary, letter-to-sound rules do not seem to provide a more basic level for the child to bootstrap into reading, whereas Seymour's logographic route does. There is evidence that phonetic reading happens sometime in the course of development, this is likely to result from the child's induction of a procedure which encodes the letter-sound mapping in some form. This could well be a PDP network or a set of rules. I'm still not sure what the mature model should look like, but as Coltheart (1991) points out in his commentary, it does not look like Seidenberg and McClelland's PDP. Neither does it look like a set of letter-to-sound rules.
Skoyles J. (1991) Connectionism, Reading and the Limits of Cognition. PSYCOLOQUY 2.8.4.
Coltheart M. (1991) Connectionist modeling of human language processing: The case of reading (Commentary on Skoyles 1991). PSYCOLOQUY 184.108.40.206
Coltheart, M., Curtis, B., Atkins, P. and Schreter, Z. (1991) Computational modelling of reading aloud: connectionist and nonconnectionist approaches. Psychonomic Society meeting, San Francisco, November 1991.
Besner, D., Twilley, L., McCann, R.S. and Seergobin, K. (1990) On the association between connectionism and data: are a few words necessary? Psychological Review, 97, 432-446.
Elovitz, H.S. and Johnson, R. and McHugh, A. and Shore, J.E. (1976) Letter-to-sound rules for automatic translation of English text to phonetics IEEE Transactions on Acoustics Speech and Signal Processing
Venezky, R.L. (1970) The structure of English Orthography, Mouton, The Hague.
Seidenberg, M. S. and McClelland, J. l. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96, 523-568.
Seymour, P.H.K. and Elder L. (1986) Beginning reading without phonology. Cognitive Neuropsychology, 3, 1, 1-36
Stuart, M. and Coltheart, M (1988) Does reading develop in a sequence of stages? Cognition, 30, 139-181
Johnston, R.S. and Thompson, G.B. (1987) Is dependence on phonological information an invariant stage in reading development? Unpublished Manuscript, Department of Education, Victoria University of Wellington, New Zealand.