A connectionist network with recurrent connections among units was designed to simulate human reader's performances in naming Japanese Kanji words. The network was trained to map the orthography of two-character Kanji words onto their pronunciations or phonology. After learning with the training corpus including approximately 4,000 two-character Kanji words, the network showed, in its naming latency, frequency and consistency effects and an interaction between these effects, largely comparable to these effects in naming latency of Japanese skilled readers. On the other hand, when naming nonwords, the network's accuracy was substantially worse than that of skilled readers. The results of the present study indicate that such a kind of network, although originally developed to simulate human readers' performance in alphabetic writing systems, can apply to Kanji as well, with caution that further elaboration is indispensable to cope with poor performance for nonwords.