A connectionist approach to naming disorders of Japanese in dyslexic patients

The triangle model is a computational model for lexical processing which computes word orthography, phonology and semantics using the architecture of a parallel distributed processing network. The computation takes the form of interactions among neuron-like processing units. In the present research, a Japanese triangle model computed phonology directly from orthography for both Kanji and Kana strings, with additional input to phonology from a component representing putative word semantics. This model successfully simulated certain effects seen in the performance of Japanese skilled readers. Moreover, different types of damage to the model reproduced data on both the surface and phonological forms of acquired dyslexia. After damage to the semantic component, the model’s reading performance remained good for Kana words, Kana nonwords, and Kanji words with consistent character-sound correspondences, but was significantly impaired on Kanji words with atypical correspondences: this simulates surface dyslexia. After damage to the phonological component of the model, the network’s performance remained good for both Kanji and Kana words but was impaired on Kana nonwords: this simulates phonological dyslexia. These results are basically comparable to those of previous models developed for English, and thus demonstrate that the same computational principles of the triangle model can be applied to alphabetic and non-alphabetic writing systems. Mechanisms and properties of the model for Japanese are discussed.