UCLA/USC Joint Phonology Seminar
I describe an implemented system for learning the allomorphs of morphemes. It inputs glossed paradigms and outputs the allomorphs of each morpheme. The system takes the form of a maxent grammar, with constraints preferring minimal alternation and segmental contiguity. Learning the allomorphs first greatly simplifies the learning of underlying forms: string alignment of allomorphs reveals the full set of alternations, which can be used to construct a highly informative candidate set from which the correct UR’s and constraint rankings are readily discovered. To illustrate, I show how the system learns UR’s and rankings for several well-known phonology problems.