A connectionist model of alexia: Covert recognition and case mixing effects

A connectionist model was developed to simulate the production of phonological, semantic and orthographic lexical representations from an orthographic input code. When lesioned, the network displayed many of the characteristics of the neurological disorder, pure alexia. These included: greater accuracy in the recognition of letters compared to words, and (in some cases) spared semantic categorization and lexical decision ability relative to naming. Performance was superior for high relative to low frequency words and (for some lesions) for ‘high imageability’ words with more semantic referents. Similar to alexic patients, errors in the model tended to be visually rather than semantically related to target words and disrupted input had a greater effect on naming than on lexical decision performance. There were also tendencies apparent in the model which have not so far been reported for alexic patients but are found in other neurological patients, including category specificity effects and superior performance on superordinate relative to subordinate semantic categorization. Noise was added to the input units of the unlesioned and lesioned model in an attempt to simulate case mixing effects on both normal and alexic reading. The unlesioned model demonstrated similar effects to normal readers with mixed case stimuli (Besner & McCann, 1987; Mayall & Humphreys, 1996). Further, effects on the lesioned model were comparable to those found with alexic patients (Bub & Arguin, 1995; Mayall & Humphreys, submitted). The similarity of the performance characteristics of the model to those of pure alexic patients suggests that the architectures of the modelled and the human reading system encompass common properties, and that covert recognition in pure alexia can be attributed to the architecture of the word processing system.