Despite their semantic-rich nature, online communities have, to date, largely been analysed through examining longitudinal changes in social networks, community uptake, or simple term-usage and language adoption. As a result, the evolution of communities on a semantic level, i.e. how concepts emerge, and how these concepts relate to previously discussed concepts, has largely been ignored. In this paper we present a graph-based exploration of the semantic evolution of online communities, thereby capturing dynamics of online communities on a conceptual level. We first examine how semantic graphs (concept graphs and entity graphs) of communities evolve, and then characterise such evolution using logistic population growth models. We demonstrate the value of such models by analysing how sample communities evolve and use our results to predict churn rates in community forums.