|Abstract:||This paper examines a case of parasitic tone harmony from Dioula d'Odienné, in which there is a ganging effect of local and long-distance segmental features that increases the likelihood of tone harmony. I use the Dioula data to revisit the recent claim put forth by e.g., Potts et al. 2010 that ganging effects can be captured solely by cumulativity of weighted constraints in a log-linear model of phonology, and that constraint conjunction--the previous method of obtaining ganging--is no longer necessary given additive constraint weights (cf. Smolensky 2006). In the present analysis, I argue that the simple addition of constraint weights is insufficient to adequately model the probabilistic ganging effects of similarity found in the Dioula corpus. Weighted constraint conjunction, implemented as an interaction term in Maximum Entropy Harmony Grammar (e.g., Goldwater & Johnson 2003), provides a significantly improved model of ganging. To justify weighted conjunctions, an information-theoretic method for model selection and comparison in probabilistic Optimality Theory is presented, for use when addressing quantitative natural language data.
|Area/Keywords:||Phonology, Harmonic Grammar, Maximum Entropy, surface correspondence, similarity, constraint conjunction, cumulativity, superadditivity, model selection