In this paper we examine how a mechanism that learns word classes from distributional information can contribute to the simulation of child language. Using a novel measure of noun richness, it is shown that the ratio of nouns to verbs in young children’s speech is considerably higher than in adult speech. Simulations with MOSAIC show that this effect can be partially (but not completely) explained by an utterance-final bias in learning. The remainder of the effect is explained by the early emergence of a productive noun category, which can be learned through distributional analysis.