Connectionist approaches have been providing new views for understanding both normal and disordered reading processes (e.g., Seidenberg & McClelland, 1989; Plaut, McClelland, Seidenberg, & Patterson, 1996). The present research was designed to simulate normal naming processes of Japanese Kanji words using a connectionist network. The network was trained to map the orthography of two-character Kanji words onto their pronunciations or phonology. Each Kanji character was represented by a 16 x 16 grid pattern on the input layer, and the word's phonology at the output layer consisted of phonological codes for the two component Kanji characters. The training corpus included 4,136 two-character Kanji-words with frequencies higher than four occurrences per million. After 900 training epochs, the network could correctly name 99.8 ~ of the 4,136 words, including those with inconsistent or atypical character-sound correspondences. In terms of efficiency of word naming, the network showed frequency and consistency effects and an interaction between these variables, largely comparable to these effects in the naming latencies of Japanese skilled readers (Fushimi, Ijuin, Patterson, & Tatsumi, 1999). On the other hand, when naming nonwords, the network's performance was substantially worse than that of skilled readers. Properties of the connectionist approach for Japanese Kanji word naming are discussed.

Key words: connectionist model, Japanese kanji, naming, consistency effect