Learning opacity in Stratal Maximum Entropy Grammar

(Kiparsky 1971, 1973) draws attention to cases of historical change that suggest that at least some opaque phonological process interactions are difficult to learn. In his work on Stratal Optimality Theory (Stratal OT), he further claims that independent evidence about the stratal affiliation of a process is helpful in learning an opaque interaction (Kiparsky 2000). In this paper, we develop a computationally implemented learner for a weighted constraint version of Stratal OT, and show that for an initial set of test cases, opaque patterns are indeed generally more difficult than their transparent counterparts, and that information about stratal affiliation does make learning of an opaque pattern easier. Our results support the viability of Stratal OT as a theory of opaque interactions (Bermúdez-Otero 1999, 2003, Kiparsky 2000; see below). However, it is not clear whether a stratal setup is the only source of opaque interactions – in fact, some cases of opacity may stem from other factors (McCarthy 2007). A broader assessment of alternative theories of opacity will not be attempted in this paper; see Jarosz (2016) for work on learning opacity in one of these alternative frameworks (Jarosz 2014). In the Stratal OT approach to opacity (see also especially Bermúdez-Otero 1999, 2003), grammars are chained together to produce outputs that would not be possible within a single level. For example, the classic case of the opaque interaction between diphthong raising and flapping in Canadian English illustrated in (1), mitre can be produced by raising in the first (word level) grammar, and then flapping in the second (phrase level) one. As we will see shortly, the final candidate cannot be produced by a single OT grammar, with the standard set of constraints assumed for this case (though see Pater 2014).