Bayesian Learning over Conflicting Data: Predictions for Language Change

This paper is an analysis of the claim that a universal ban on certain ('anti-markedness') grammars is necessary in order to explain their non-occurrence in the languages of the world. To assess the validity of this hypothesis I examine the implications of one sound change (a > ə) for learning in a specific phonological domain (stress assignment), making explicit assumptions about the type of data that results, and the learning function that computes over that data. The preliminary conclusion is that restrictions on possible end-point languages are unneeded, and that the most likely outcome of change is a lexicon that is inconsistent with respect to a single generating rule.