Similarity, feature-based generalization and bias in novel onset clusters

The past decade has seen significant advances in modeling gradient preferences of English speakers using grammars with probabilistic rules, constraint rankings, or weights inferred from the statistical properties of the lexicon (Coleman and Pierrehumbert 1997; Treiman, Kessler, Knewasser, Tincoff, and Bowman 2000; Frisch, Large, and Pisoni 2000; Bailey and Hahn 2001; Hammond 2004; Hayes and Wilson, in press; and many others). In study after study, where lexical counts have shown probabilistic differences between different cluster types, experiments with native speakers have shown corresponding behavior preferences. At the same, studies investigating substantive biases in phonology have begun to focus on cases in which speakers’ preferences do not appear to follow straightforwardly from lexical statistics (Moreton 2002, 2007; Wilson 2003, 2006; Zhang and Lai 2006; Berent et al, in press). For example, Berent et al. (in press) show that English speakers prefer initial #bn over #bd, even though few or no English words begin with either cluster. These results show that speaker preferences cannot be guided solely by obvious lexical statistics such as literal biphone frequency. However, they leave open the question of whether such preferences might be learned by some less direct means, perhaps involving similarity, phonological features, or some other mechanism for generalizing from attested to unattested sequeces. In this talk, I report on some attempts to take on this question, exploring the capacity of similarity-based and natural class-based models to predict generalization to novel clusters.