Constraint programming is rapidly becoming the technology of choice for modeling and solving complex combinatorial problems. However, users of constraint programming technology need significant expertise in order to model their problem appropriately. The lack of availability of such expertise can be a significant bottleneck to the broader uptake of constraint technology in the real world. In this paper we are concerned with automating the formulation of constraint satisfaction problems from examples of solutions and non-solutions. We combine techniques from the fields of machine learning and constraint programming. In particular we present a portfolio of approaches to exploiting the semantics of the constraints that we acquire to improve the efficiency of the acquisition process. We demonstrate how inference and search can be used to extract useful information that would otherwise be hidden in the set of examples from which we learn the target constraint satisfaction problem. We demonstrate the utility of the approaches in a case-study domain.