Learning phonological underlying representations

One of the most important questions yet to be fully addressed in morphophonological learning is the simultaneous learning of underlying representation (URs) and the phonological grammar with alternating segments. For example, if provided with [kæt-s] ‘cats’ and [dɑg-z] ‘dogs’, how can a learner successfully learn the UR for plural is /z/ and acquire a grammar that enforces devoicing? This is a difficult learning task because it is a hidden structure problem. While URs are hidden and unobserved, the choice of URs and the learned phonology are mutually dependent: selecting /z/ as the UR leads us to a devoicing grammar while choosing /s/ as the UR requires a mapping from /s/ to [z]. Previous literature has approached the same issue with a variety of approaches, and it continues to be an active field of research.

My ongoing work in collaboration with Bruce Hayes adds to the existing literature by explicitly testing to what degree URs can abstract away from their surface realizations. We found a tentative support for a relatively concrete position indicating that URs might not be too distant from their observed forms, as more abstract URs could lead the learner to fail by getting stuck in a local optimum.

For more details, see below:

• AMP 2022 Poster

  Wang, Yang and Bruce Hayes. 2022. Learning underlying representations: Expectation-Maximization and the KK-Hierarchy. Poster presented at the Annual Meeting on Phonology. UCLA, Los Angeles, CA. October 2022.

• Manuscript

  Our manuscript is currently under revision. You can email me for a copy. I will attach a pdf soon.