Chapter 2. The Simulations
connectionist visual word recognition, namely that a single system can learn to pro- nounce both exception and nonwords at a level comparable to human subjects. If the networks were not able to pronounce nonwords, this would validate the claim of the Dual-Route model whereby there have to be two separate processes, one rule-based one that can pronounce nonwords but not exception words and one lexicon-based one that can pronounce exception words but not nonwords (Coltheart et al., 1993; Weekes and Coltheart, 1996). Thus for example the poor nonword performance of the Seiden- berg and McClelland (1989) network was one of the major criticisms for that model because the network was far below the average performance of human subjects. This model was subsequently improved and achieved levels of nonword reading compara- ble to humans (Plaut et al., 1996; Harm and Seidenberg, 1999), showing that it was not a problem inherent of the architecture of this connectionist model. One of the prob- lems of the earlier model which contributed to the poor nonword reading was the use of Wichelfeatures as input and output representations. Since, for both the current net- works, a slot-based representation was used, it is expected that the networks will not suffer from the same problem regarding the nonword reading.
As testing corpus, Glushko’s nonwords were used (Glushko, 1979) as seen in Plaut et al. (1996). All nonwords used with their acceptable pronunciations can be seen in appendix D. The non-words are divided into two types: consistent and inconsistent nonwords. A consistent word is a word whose orthography-to-phonology correspon- dence is consistent with the one of its orthographic neighbours. If this is not the case, then the word is inconsistent. Note that in the present study this difference has not been made for the words in the original training corpus. However, since the nonwords were already subdivided in this way in Plaut et al. (1996), this division was kept for the nonwords. For testing, a nonword was deemed to be correct if the output of the network matched one of the possible pronunciations of the nonword.