Speaker: Levi Driscoll (MIT)
Title: P-Map Biases Enable Joint Learning of Allomorphy and Phonology
Time: Monday, October 21 5pm-7pm
Location: 32-D831
Abstract: The objective of this study is to build an unsupervised learner that can deduce a set of stems and affixes and a phonological grammar when presented with a corpus of surface forms to which some phonological processes have applied. The central question under investigation is whether introducing a learning bias informed by the P-Map (Steriade 2008) enables the learner to outperform one that considers an unrestricted hypothesis space.
This model is unique in that it is tasked with segmenting a set of surface forms, determining whether the proposed morphemes are related, and generating a constraint-based grammar with no prior information about meanings associated with forms or phonological processes that may have applied to the data. Other models are often primarily concerned with segmentation (Goldsmith 2000 et seq.), or learn phonology given perfectly segmented input (Boersma & Pater 2016), or learn phonological rules rather than constraint weights or rankings (Albright & Hayes 2003, Calamaro & Jarosz 2015). The present model seeks to simulate learning both morphology and phonology with no endowment beyond knowledge of natural classes and a set of unweighted constraints.