Learning Non-Local Phonological Alternations via Automatic Creation of Tiers

Anonymous

Keywords: phonological alternations, tiers, non-local, learning

TL;DR: We present a computational model that learns non-local alternations by automatically constructing alternation-relevant tiers, and accurately matches the behavior of humans on prior artificial language experiments.

Abstract: Phonological alternations often involve dependencies between adjacent segments. Despite the apparent long-distance nature of alternations such as consonant and vowel harmony, even these can be reduced to dependencies between adjacent segments by projecting a subset of segments onto a new representation, often called a tier. Tiers are known to simplify learning non-local dependencies and are consistent with human behavior in artificial language experiments. However, little is known about the mechanism by which learners may construct such a representation. In this work, we propose a computational model that learns non-local alternations by automatically constructing alternation-relevant tiers. The model is sensitive to adjacent dependencies and---when adjacency fails---uses this same sensitivity to construct a tier that reduces the relevant non-local dependencies to local ones, falling within its purview. The model accurately matches the behavior of humans on prior artificial language experiments. This submission describes preliminary work and is intended as an extended abstract.