I started reading Pinker's Words and Rules a few months ago (I didn't actually do more than start then). He makes this point - apparently there are two distinct language systems running in parallel - one that simply regurgitates words (which accounts for irregulars, and the fact that very common words are more likely to be irregular) and a rule system that is more flexible but that has higher processing overhead. (Thanks for reminding me, I really need to get back and finish it.)