Speaker: Erin Olson (MIT)
Title: Loanwords and the Perceptual Map: A perspective from MaxEnt learning
Date/Time: Monday, 9/9 5:00-6:30pm
Location: 32-D831
Abstract:
The goal of phonological learning algorithms has largely been to arrive at a constraint ranking or weighting which is consistent with the data on which it is trained (Tesar & Smolensky, 2000, a.o.). However, phonological theorists have long understood that there should be additional conditions on such a ranking or weighting, whereby some constraints should always be ranked/weighted higher than others — e.g., the P(erceptual) Map of Steriade (2001). Two models which are successful in achieving both goals are those of Wilson (2006) and White (2013), who construct models which are capable of incorporating the PMap as an additional condition (bias) on learning. Both models are successful in capturing the behaviour of adult speakers when they are explicitly asked to learn a phonological process from data, and are successful in replicating results expected from the PMap. However, each model makes a different prediction when attempting to model the knowledge speakers have before being presented with data about a novel process. A real-world example of this situation comes from loanword phonology, where speakers are tasked with repairing phonologically marked structures without ever having encountered such structures — or the appropriate repair to such structures — in their native language. The model proposed by Wilson (2006) predicts that such speakers should behave as if they have no PMap-like bias, while that of White (2013) predicts that such a bias is indeed present.
In this talk, I will present the results of a variety of experiments which aim to test these hypotheses by modelling the phonological knowledge of a speaker of Cantonese, who must infer how to properly repair illicit phonological structures from English loanwords without ever having encountered the relevant structure in their native vocabulary. In Experiment 1, it will be established that, for a reasonable set of constraints, learning from only native language data is insufficient for arriving at a weighting that is also consistent with the attested loanword repairs of that language. Thus, a substantive, PMap-like bias will be imposed over the chosen constraint set in order to attempt to improve the model. Experiment 2 will show that encoding the bias in the manner proposed by Wilson (2006) is also insufficient for arriving at a weighting that predicts the loanword data. Experiment 3, in contrast, will show that encoding the bias in the manner proposed by White (2013) is successful, thus confirming the hypotheses outlined above.