Connectionist Modelling of Word Recognition

Connectionist models offer concrete mechanisms for cognitive processes. When these models mimic the performance of human subjects they can offer insights into the computations which might underlie human cognition. We illustrate this with the performance of a recurrent connectionist network which produces the meaning of words in response to their spelling pattern. It mimics a paradoxical pattern of errors produced by people trying to read degraded words. The reason why the network produces the surprising error pattern lies in the nature of the attractors which it develops as it learns to map spelling patterns to semantics. The key role of attractor structure in the successful simulation suggests that the normal adult semantic reading route may involve attractor dynamics, and thus the paradoxical error pattern is explained.