Extended Abstract

Extending adaptor grammars to learn phonological alternations

---

Abstract

Recent advances in unsupervised learning of linguistic structure have demonstrated the feasibility of inferring latent morphological parses from an unannotated corpus given transparent underlying-to-surface mappings (ex., Adaptor Grammars), as well as in learning predictable phonological transformations from prespecified underlying morphemes to a range of surface allomorphs via a stochastic edit distance algorithm. In this paper we introduce a nonparametric Bayesian model which builds on the morpheme-segmentation success of AGs, and incorporates the ability to learn predictable phonological transformations of underlying forms to their surface allomorphs via the interaction of markedness and faithfulness principles, inspired by generative phonology. The unsupervised nature of this model (that is, no semantic information about the words being segmented is provided) is relevant not only computationally but also psychologically, as it mirrors developmental findings that young infants segment and cluster morphemes based solely on phonetic and distributional similarity. The model also incorporates many of the other cognitive restrictions infants during the initial period of morphophonological learning in an effort to make the model maximally realistic, and thus eventually useful in making quantitative predictions about the early stages of morphophonological acquisition that can be experimentally investigated. We evaluate the model on a novel dataset consisting of a complex system of allomorphy in Acehnese, an understudied Indonesian language.

Keywords: phonological acquisition, adaptor grammars, bayesian nonparametrics, morphophonology, Acehnese, allomorphy, morphological segmentation

How to Cite:

Breiss, C. & Wilson, C., (2020) “Extending adaptor grammars to learn phonological alternations”, Society for Computation in Linguistics 3(1), 480-483. doi: https://doi.org/10.7275/qnm4-fq42