Connectionism is a modeling approach for understanding cognitive processes (for example, see Haykin, 1994). A connectionist model consists of nodes (mathematically modeled neuron-like elements) that are interconnected to each other through connection weights. A connectionist model learns to perform a cognitive task through processing of given input stimuli, and, in many models, feedback to the model's response. It can learn to process rule-like behaviors along with exceptions to rules. Thus, connectionist models are often used to demonstrate the importance of learning over innate knowledge, and to argue for nonmodular cognitive architecture (e.g. Seidenberg & McClelland, 1989).

Connectionism has been applied at all levels of Japanese sentence processing, from phonology (Ijuin et al., 1999), lexical-semantics (Tsuzuki, 1996, Ma et al., 2000), to syntax and sentential semantics (Negishi & Hanson, 2001; Tsuzuki et al., 1999; Motoki, Watanabe & Shimazu, 1998). A reasonable question to ask is what kind of unique contribution the connectionist research in Japanese psycholinguistics can make. Unlike English, both phonetic characters (katakana and hiragana) and ideographical characters (kanji) are used in Japanese. There are many Japanese words that have different kanji spellings but the same sound (and thus the same hiragana/katakana spelling). Syntactically, it is a head-final language and has case particles. In this chapter, connectionist models of Japanese sentence processing are examined, focusing on connectionist architecture, the claims made by connectionist models, and the use of linguistic features particular to Japanese.