The Weekly Newsletter of MIT Linguistics

Phonology Circle 9/11 - Roni Katzir (TAU & MIT) and Ezer Rasin (MIT)

Speaker: Roni Katzir (Tel Aviv University & MIT) and Ezer Rasin (MIT)
Title: Learning opacity, optionality, and abstract URs using Minimum Description Length
Date and time: Monday, September 11, 5:00-6:30pm
Location: 32-D831

We discuss two posters that we will present at the Annual Meeting on Phonology 2017:

1. Acquiring opaque phonological interactions using Minimum Description Length

Opacity poses an obvious challenge for the child learning the phonology of their ambient language: a process to be acquired loses support because of environments in which it was supposed to apply but didn’t (e.g., counterfeeding) or in which it wasn’t supposed to apply but did (e.g., counterbleeding). Not surprisingly, no learners in the literature can handle opacity distributionally, from unanalyzed input data alone. Children, however, do manage to acquire opacity in a variety of languages (Dell 1981, McCarthy 2007). This talk shows how opaque interactions (including optionality, counterfeeding, and counterbleeding) can be acquired using the principle of Minimum Description Length (MDL; Solomonoff 1964, Rissanen 1978). Specifically, we use an adaptation of Rasin & Katzir 2016’s MDL learner (originally used for OT phonology) to rule-based phonology and show how it applies to various cases of opacity. We then show simulations on artificial-language data illustrating the mechanization of the idea.

2. Minimum Description Length subsumes Free Ride effects in UR learning

Human learners have been argued to infer URs that are sometimes different from their corresponding surface representation (SR) even without being forced to do so by an alternation. McCarthy (2005) proposes the Free Ride Principle (FRP), which allows the learner to extend non- identical mappings in alternating forms to non-alternating forms. Recently, Rasin & Katzir (2016) proposed Minimum Description Length (MDL) as an evaluation metric for OT. The present work shows how MDL supports the induction of nonidentical URs both in cases that have motivated the FRP and in cases that have been argued to involve nonidentical URs in the absence of supporting alternations (and are thus outside the scope of the FRP). As a result, the FRP is redundant.