On the analogical modelling of the English past-tense: A critical assessment

Abstract

The formation of English past tenses has provided the arena in which alternative computational models of psycholinguistic processes have been most intensely tested. The debate has mainly revolved around accounts based on a single, analogising mechanism as exemplified by connectionist models and the dual mechanism theory of Pinker (1999) which combines elements of a pattern associator memory with a rule-based system. Prasada and Pinker (1993) provide evidence against the connectionist approach by showing that their networks failed to reproduce the behaviour of their English speaking subjects when forming the past tense of a subset of nonce verbs. Eddington (2000a) provides an alternative single mechanism account of this data using exemplar-based, analogising models which are claimed to produce behaviour comparable to Prasada and Pinker's subjects. On the basis of these results he argues that not only is a general rejection of single mechanism, analogising accounts premature but also that exemplar-based models ⿿may prove to have advantages over their connectionist counterparts⿿.This paper tests these theoretically important claims. It is shown that the contrast between the particular exemplar-based theory, analogical modelling (AM), and connectionist model which forms the basis of Eddington's arguments is, in fact, a function of differences in the mappings the two simulations compute and that once these are removed, the AM advantage disappears. A detailed analysis of the predictions of the particular AM simulation presented further shows that, although the output bears superficial similarities with those produced by Prasada and Pinker's subjects, other properties are more problematic.