The connectionist approach to reading aloud has been a serious challenge to the traditional dual-route theory, but a critical question concerning the theoretical distinction of the connectionist approach from the dual-route theory remains unresolved. That is, through what kind of internal structure a single-route connectionist model represents the two seemingly distinct kinds of ability to process regularities and exceptions without relying on dual-route structure? By taking a model from Plaut et al. (1996) and examining it closely, the present study attempts to answer this question. Various forms of network analysis demonstrate that the representational system in hidden unit space is structured in the same way regardless of learning regularities or exceptions. Further analyses about the effect of the reading network's exception learning upon its nonword reading reveal a proper viewpoint on the connectionist mechanism for a quasiregular task. Unlike the dual-route assumption, exception learning in connectionist modeling of reading aloud does affect the model's nonword reading performance. This is analogous to the phenomenon that "noise capturing" or "overfitting" in statistical modeling affects the model's generalization performance. In reality, however, the severity of "ordinary exceptions" in normal word reading happens to be not high enough to ruin the network's nonword reading, as "noise" does in statistical modeling. |