Event-B provides a flexible framework for stepwise system development via refinement. The framework supports steps for (a) refining events (one-by-one), (b) splitting events (one-by-many), and (c) introducing new events. In each of the steps events can moreover possibly be anticipated or convergent. All such steps are accompanied with precise proof obligations. Still, it remains unclear what the exact relationship - in terms of a behaviour-oriented semantics - between an Event-B machine and its refinement is. In this paper, we give a CSP account of Event-B refinement, with a treatment for the first time of splitting events and of anticipated events. To this end, we define a CSP semantics for Event-B and show how the different forms of Event-B refinement can be captured as CSP refinement.