Deriving frequency effects from biases in learning
Maggie Baird · Proceedings of the Linguistic Society of America · 2021
This paper presents a phonological learner that derives frequency effects – the propensity of more frequent items undergo deletion and reduction processes at higher rates. The model is a bidirectional Maximum Entropy grammar which has two distinct learning steps, one mapping from UR to SR, and another mapping back from SR to UR using Bayesian inference. The model is tested on the case of t/d deletion in English and correctly derives the frequency-based pattern of deletion without access to surface patterns.