The framework developed in this paper can deal with scenarios where selected sub-ontologies of a large ontology are offered as views to users, based on contexts like the access rights of a user, the trust level required by the application, or the level of detail requested by the user. Instead of materializing a large number of different sub-ontologies, we propose to keep just one ontology, but equip each axiom with a label from an appropriate context lattice. The different contexts of this ontology are then also expressed by elements of this lattice. For large-scale ontologies, certain consequences (like the subsumption hierarchy) are often pre-computed. Instead of pre-computing these consequences for every context, our approach computes just one label (called a boundary) for each consequence such that a comparison of the user label with the consequence label determines whether the consequence follows from the sub-ontology determined by the context. We describe different black-box approaches for computing boundaries, and present first experimental results that compare the efficiency of these approaches on large real-world ontologies. Black-box means that, rather than requiring modifications of existing reasoning procedures, these approaches can use such procedures directly as sub-procedures, which allows us to employ existing highly-optimized reasoners. Similar to designing ontologies, the process of assigning axiom labels is error-prone. For this reason, we also address the problem of how to repair the labelling of an ontology in case the knowledge engineer notices that the computed boundary of a consequence does not coincide with her intuition regarding in which context the consequence should or should not be visible.