The exploitation of distributional information in syllable processing

There is now growing evidence that people are sensitive to the statistical regularities embedded into linguistic utterances, but the exact nature of the distributional information to which human performance is sensitive is an issue that has been surprisingly neglected as yet. In order to address this issue, we first propose an overview of some basic measures of association, going from the simple co-occurrence frequency to the normative measure of contingency, rw: We then report an experiment collecting judgments of word-likeness as a function of the relationship between the phonemes composing the rimes (VC). The contingency between Vs and Cs, as assessed by rw; was the best predictor of children and adult judgments. Surprisingly, the forward transitional probability ðPðC=VÞ; which is the main measure considered by language researchers, was a poor predictor of performance, whereas the backward transitional probability ðPðV=CÞ) made a sizeable contribution. We then analyze the ability of computational models to account for these results, successively considering a connectionist model based on the automatic computation of statistical regularities (SRN) [Cogn. Sci. 14 (1990) 179] and a model in which the sensitivity to statistical regularities emerges as a by-product of the attentional processing of the incoming information (Parser) [J. Memory Language 39 (1998) 246]. Somewhat ironically, Parser, which implements no specific mechanisms for statistical computations, proves to be a better predictor of performance than the SRN. The generality of these results, and their implications for the issue of automaticity, are discussed. q 2003 Elsevier Ltd. All rights reserved.