عنوان

A Multilinear Approach to the Unsupervised Learning of Morphology

TL50942

انگلیسی

A Multilinear Approach to the Unsupervised Learning of Morphology

[Thesis]

Meyer, Anthony

Dickinson, Markus

Indiana University

2019

187 p.

Ph.D.

Indiana University

2019

This dissertation presents a multilinear approach to the unsupervised learning of morphology (ULM), where multilinear refers to a multi-tiered architecture that allows for the handling of both concatenative and nonconcatenative phenomena in a general, unified way, as in autosegmental morphology. This dissertation reformulates autosegmental theory in graph-theoretic terms. That is, it identifies the essential properties that make autosegmental theory so conducive to modeling nonconcatenative morphology and shows that these properties are equivalent to the mathematical properties of a bipartite graph. This observation makes it possible to recast the autosegmental formalism as a graphical machine-learning model, namely the Multiple Cause Mixture Model (MCMM), a bipartite graphical model related to the Restricted Boltzmann Machine. The dissertation's experimental component consists of the development and evaluation of Multimorph, an MCMM-driven ULM system. The evaluation method takes a "dual-paradigm" approach, comprising both intrinsic and extrinsic components. The latter evaluates the system as a component of a larger chain of processes. This in line with lexeme-based theories of morphology, i.e., theories that regard morphology as a distinct but mediating layer of linguistic organization situated between phonology and syntax/semantics. The results of the experiments demonstrate the soundness and promise of a multilinear approach to the unsupervised learning of morphology.

Computer science

Linguistics

Meyer, Anthony

Dickinson, Markus

Indiana University

p

[Thesis]

276903

a

Y