Laying a foundation for bidialectalism: necessary biases for algorithmic learning of two dialects of Estonian

This paper investigates how the vowel patterns of two closely-related dialects of Estonian can be described using as much shared representation as possible, as well as what parameters and biases are necessary for the Gradual Learning Algorithm (Boersma & Hayes, 2001) to be able to learn restrictive grammars for both dialects. Standard Estonian and the minority Kihnu dialect share the same vowel inventory but differ in their distribution of those vowels, with Standard Estonian being subject to positional restrictions and Kihnu Estonian demonstrating front-back vowel harmony. I extend the constraint set that Kiparsky and Pajusalu (2003) propose to account for the vowel harmony typology in Balto-Finnic languages, and show via Low-Faithfulness Constraint Demotion (Hayes, 2004) that there exist ranking of these constraints that account for the patterns in both dialects. I also process the Estonian Dialect Corpus (Lindström, 2013) and use the contents as learning data for runs of the Gradual Learning Algorithm implemented by both OTSoft (Hayes et al., 2013) and the author. Convergence on grammars for both dialects is contingent on three biases being employed during the learning process: low initial faithfulness, specific over general faithfulness, and the Magri (2012) update rule. This suggests that even for relatively simple phonological patterns, the learning environment must be delicately balanced in order to account for the behaviour of two different grammars grounded in the same shared framework.

This paper is my second Qualifying Paper, one of the requirements for candidacy in the UBC Department of Linguistics. Download manuscript here.