Connectionist modelling of Chinese character pronunciation

Introduction Psychologists and cognitive scientists aim to understand the universals involved when the brain deals with written language. Seidenberg and McClelland’s “triangle model” of the reading of monosyllabic English words has been substantially developed (e.g. Harm & Seidenberg, 1999), but with little application to languages other than English. Here we report our initial application of this approach to the reading of Chinese, a radically different orthography from English, illuminating processing universals in reading. The granularity of Chinese orthography is a fundamental issue in modelling its pronunciation. The character is the unit of pronunciation. Several formats exist, but a central paradigm exists in which a phonetic radical, on the right, specifies pronunciation information, and a semantic radical, on the left, specifies information about meaning (Figure 1). Chen et al. (1996) argue that these radicals (plus any remaining sub-character material) are the functional processing units of Chinese reading. Therefore, the typical granularity is substantially coarser than that found in English orthography. Is there thus a principled architecture for modelling Chinese pronunciation that uses distributed representations in a non-trivial way? We resolve the issue by referring to the anatomy of the visual pathways.