The Gradient Lexicon and Phonology Learner


During acquisition, the learner must memorize all the words of her language, as well as phonological generalizations over them. These tasks interact: the contents of the lexicon form the empirical basis for the acquisition of phonological grammar, and the contents of the phonological grammar can constrain how words are represented in the lexicon (e.g. Mehler, Dupoux, and Segui (1990)). This interaction becomes especially complex for patterns with exceptions. Fidelholz (1979) points out that patterns with just a few exceptions may be represented differently than patterns with many. I present a computational model in which features of a lexical item (like stress) can be represented with continuously-valued weights. These weights simulate the memory strength associated with each feature of a word. When the features of lexical items are learned concurrently with the phonological grammar, different degrees of exceptionality are represented differently, both in the grammar and in the lexical entries themselves.


Slides from the 2017 LSA annual meeting, 6 January, 2017

The learner used for the talk

Input files (tab-delimited text): no exceptions , 5 exceptions , 100 exceptions

Python script for generating new input files

probability plot

    Probability of encoding stress on a new word depends on the number of stress exceptions in the language.