Asymmetric Generalisation of Harmony Triggers

Abstract

In vowel harmony systems, certain classes of segments may be preferred as triggers; in particular, Kaun (1995) notes that rounding harmony is preferentially triggered by non-high vowels. There is a plausible phonetic explanation for this: non-high vowels manifest F2 contrasts less prominently (Linker, 1982; Terbeek, 1977) and therefore benefit more from the boost in perceptual salience that harmony affords. In this paper, I present the results of an artificial grammar experiment suggesting, following Wilson (2006), that learners are systematically biased towards phonetically natural generalisations. Learners trained on a harmony system triggered by high vowels extended the generalisation to mid vowels, but learners trained on mid vowels varied in the breadth of their generalisation. I also show that a Maximum Entropy learner can successfully mimic the behaviour of human subjects if biases are included for both generality and phonetic naturalness.