The winner takes it all — almost: cumulativity in grammatical variation

Abstract

In this article we show that cumulativity is necessary to account for probabilistic variation found in actual language use, and we compare the accuracy of the predictions that different versions of stochastic OT make. We distinguish two versions of cumulativity, namely ganging-up cumulativity and counting cumulativity . We will compare how Paul Boersma's (1998) version of stochastic OT on the one hand and Maximum entropy models on the other hand deal with cumulativity. The second part of the article reports empirical data on English genitive variation. It turns out that both versions of cumulativity obtain in the empirical data. In the last part of the article, we compare the predictions of the two theories with respect to this empirical domain. The maximum entropy model proves to be clearly superior, both with respect to the accuracy of its predictions and to its learnability properties.