The Weekly Newsletter of MIT Linguistics

Phonology Circle 11/7 - Gaja Jarosz & Aleksei Nazarov

Presenter: Gaja Jarosz (UMass Amherst) & Aleksei Nazarov (University of Toronto) 
Title: Domain-general learning of stress parameters: A typological investigation
Date/Time: Wednesday, November 7, 5:00-6:30pm
Location: 32-D831

Existing learners for stress in OT (e.g., Tesar and Smolensky 2000, Boersma and Pater 2013/2016, Jarosz 2013, 2015) are all so-called domain-general learners (need not be specified in UG). However, Pearl (2007, 2011) argues that Principles & Parameters models of stress (e.g., Dresher and Kaye 1990) do necessitate domain-specific learning mechanisms (specified in UG). Nazarov and Jarosz (2017) proposed a domain-general learning with more inference power than Yang’s (2002) learner (which Pearl 2011 showed didn’t work for stress parameters), arguing that domain-general learners may be sufficient for both OT and parametric models.

However, Nazarov and Jarosz (2017) only tested their learner on a small set of stress systems with a subset of Dresher and Kaye’s parameters. To validate these initial results, we have now tested our learner, in addition to 3 versions of Yang’s (2002) learner, on all 280 stress systems that are possible in Dresher and Kaye’s (1990) framework. Similarly to Nazarov and Jarosz (2017), it is found that the proposed learner converges on the correct stress system about 95% of the time, while Yang’s learner does so no more than 10% of the time. To further compare the learners, we have extracted a few statistics on the stress systems themselves that predict their performance under each learner. Finally, we found that the stress systems never learned by Nazarov and Jarosz’s learner are unattested (save for two cases that warrant further investigation), suggesting that the only function of domain-specific learning mechanisms might be to enable the learning of unattested languages.