What Matters in Artificial Learning, Sonority Hierarchy or Natural Classes?

Abstract

My research examines one proposed universal, the implicational nasal hierarchy scale, testing whether this scale is found with speakers of a language with no clear evidence for a nasal hierarchy.

Walker (2011) proposes a universal implicational nasalized segment scale based on evidence from typological frequency, Vowels > Glides > Liquids > Fricatives > Stops. She argues that if a more marked blocker class blocks harmony (vowels are least marked targets, so least likely to be blockers, and most likely to be targets), so do the less marked blocker classes (stops are most marked targets, so most likely to be blockers, and least likely to be targets). I address whether a pattern that is predicted by this implicational universal is easier to learn than one that is not. In particular, I investigate if it is easier to make a generalization when a more marked blocker (vowel)/target (stop) is presented during training and a less marked blocker (stop)/target (vowel) in testing rather than vice versa. 

In sum, individual and grouped results show evidence that both natural classes and a hierarchy play an important role in phonological artificial grammar learning.