Learning Phonological Underlying Representations: The Role of Abstractness

Finished · Phonological Learning

Summary

This paper produces a learning model that is quite impressive. When confronted with observed forms from a paradigm, it is capable of parsing morpheme boundaries (in concatenative systems), generating UR candidates and SR candidates, and uses EM learning to discover optimal URs for each morpheme along with constraint weights. Their main interest is at what level of UR abstraction do patterns become unlearnable? They go through many phonological patterns and show that URs featurally composite URs included in the set allow for successful learning, but higher levels of UR abstraction, like covert URs, precludes successful learning because the learner gets stuck in a local optimum.

They also suggest that because reanalysis has a occurred in many languages that exhibit some form of composite URs, this suggests that learners do not have access to these representations and this is why they reanalyze. But, this argument seems to fall short for me because there HAS to be access to these URs at some point in the language, otherwise the phonological patterns would never have arisen. The fact that, at one point, these patterns occurred (and still do productively in some languages) suggests that they can be arrived at by learners. However, I agree that at some level, high degrees of abstraction are dispreferred.