The hidden dangers of computer modelling: remarks on Sokolik and Smith's connectionist learning model of French gender

Sokolik and Smith (1992) conducted three computer modelling experiments involving the learning of French gender attribution using a connectionist architecture. Their device learns to classify nouns quickly, given only letter-plus-position information about word spellings and feedback about the match between the target and the device's output. In particular, the device is not programmed with rules nor does it acquire any. Sokolik and Smith conclude that their model of gender attribution and the acquisition thereof has relevance for SLA. In these remarks, I show why this conclusion is wrong.