Shared lexical representations: evidence from first-language acquisition

There is a striking inconsistency in how similarity is treated in local connectionist models: phonological and semantic similarity are taken into account, whereas lexical similarity is plainly ignored. In this article, lexical similarity, in particular the ordering of segments within words, is shown to influence linguistic behaviour. A detailed analysis of one child’s acquisition of German reveals that the correct pronunciation of /k/ in the same phonological environment (e.g. Klammer ‘clothes peg’ and klein ‘little’) is learned in different words at about the same time. This is interpreted to mean that the same horizontal link has been adjusted in the learning process. To account for identical links between segment tokens from different words, lexical items are claimed to have shared representations. More generally, memory representations are assumed to be superpositional in nature. The shared-representation hypothesis argues against local connectionism in favour of a model which endorses a one-to-many relationship between processing and linguistic units. In the final analysis, fully distributed representations may well be necessary to account for all of the data.*