Loanwords and the perceptual map : a perspective from MaxEnt Learning

Abstract

This dissertation examines the predictions of two computational models of grammar within the domain of loanword phonology. These models, formulated within a Maximum Entropy (MaxEnt) framework, have been shown to be successful when simulating the effects that a substantive bias such as the Perceptual Map (PMap) hypothesis of Steriade (2001) may have on a phonological learner. While previous studies have focused primarily on modelling data taken from artificial grammar learning experiments (Wilson, 2006; White, 2013), this dissertation will instead model loanword adaptation. Loanword adaptation was chosen as a useful test domain as speakers will often choose to repair phonotactically-illicit loanwords in ways that are not attested in their native grammar. It thus provides a wealth of data about how speakers structure their grammar in the absence of overt phonological evidence. To this end, a case study of English loanword adaptation in Cantonese is undertaken.

It will be shown that the patterns of consonant deletion and vowel epenthesis used by speakers of Cantonese to adapt English words are compatible with the PMap, and can be modelled through the MaxEnt learners mentioned above. It will also be shown through a series of computational simulations that Wilson's (2006) learner fails to acquire the grammar necessary to account for the patterns of loanword adaptation, while White's (2013) learner succeeds. This is a result of the way in which the PMap is encoded within these learners. While both encode the PMap as a series of asymmetrical Gaussian distributions on the weights of constraints, Wilson (2006) encodes this asymmetry through the variances, or plasticities, of the distributions, while White (2013) encodes it through the means, or target weights. A grammar which encodes the PMap through asymmetrical plasticities must encounter evidence from the phonology of the language in order to alter the weights of constraints.

However, the loanword phonology of Cantonese crucially lacks such phonological evidence, and Wilson's (2006) model cannot make use of it when establishing constraint asymmetries. White's (2013) model, however, allows constraint asymmetries to be maintained in the absence of overt evidence, and results in more accurate grammars of Cantonese loanword adaptation.