Abstract

Language acquisition algorithms tend to cluster around the two logical end points of a data-fitting scale: conservative models such as those employing the Subset Principle (Berwick 1985) and the Size Principle (Tenenbaum and Griffiths 2001) maintain that learners cling as closely as possible to positive data exemplars, whereas aggressive models such as the one assumed in the Sound Pattern of English (Chomsky and Halle 1968) hold that learners postulate the simplest and broadest-scoped generalizations compatible with the observed data.

In this talk we demonstrate using phonological examples that these two classes of learning algorithm make testably different predictions about the possible range of historical changes, and we compare these predictions to the empirical record. The available diachronic facts are shown to primarily involve scope expansion, which is more consistent with the predictions of the aggressive algorithms.