The English past tense: Analogy redux

The debate over how best to characterize inflectional morphology has been couched largely in terms of the "dual-mechanism" approach described in Pinker (1999) versus "single-mechanism" connectionist approaches derived from Rumelhart and McClelland (1986). There are, however, other single-mechanism approaches. The exemplar-based or analogical models of Daelemans et al. (2002) and Skousen (1989) also model inflectional usage accurately within a single-mechanism. The most striking theoretical claim peculiar to these purely analogical models is that they do not posit any resident linguistic generalizations for processing language. Instead they process new instances of usage by comparing them systematically to remembered instances of previous usage. Based on a comparison of their Minimal Generalization Model with an adaptation of Nosofsky's (1990) Generalized Context Model, Albright and Hayes (2003) argue that such purely analogical models are intrinsically inadequate for modeling the English past tense. This paper shows, however, that Skousen's (1989) Analogical Model performs as well as the Minimal Generalization Model. The implications of these results for cognitive linguistics are discussed.