Natural classes are not enough: Biased generalization in novel onset clusters

It has long been recognized that a powerful source of information about the phonological knowledge that speakers have comes from the fact that they can generalize this knowledge to novel items. A classic example of this, discussed by Chomsky and Halle (1965) and recited to countless undergraduate classes and strangers on airplanes ever since, is the fact that native speakers typically judge blick [blIk] to be a possible word of English, but tend to agree that bnick [bnIk] would not be possible. Although *blick is perhaps the most famous example of generalization to a novel string, it is unfortunately not all that revealing about the form that the relevant knowledge takes or how speakers acquire it, since all that would be required in order to reject *bnick and favor blick is a quick look at the lexicon: there are no English words starting with #bn. A slightly more telling fact is that speakers show preferences for some attested sequences over others (blick [blIk] ?thwick [TwIk]). This particular preference could follow straightforwardly from the fact that there are relatively few English words starting with [Tw] (thwart, thwack), but at least it shows that speakers have more refined knowledge than simply whether there are any existing words that start the same way.