RDF-based data integration is often hampered by the lack of methods to translate data locked in heterogeneous silos into RDF representations. In this paper, we tackle the challenge of bridging the gap between the Semantic Web and NoSQL worlds, by fostering the development of SPARQL interfaces to heterogeneous databases. To avoid defining yet another SPARQL translation method for each and every database, we propose a two-phase method. Firstly, a SPARQL query is translated into a pivot abstract query. This phase achieves as much of the translation process as possible regardless of the database. We show how optimizations at this abstract level can save subsequent work at the level of a target database query language. Secondly, the abstract query is translated into the query language of a target database, taking into account the specific database capabilities and constraints. We demonstrate the effectiveness of our method with the MongoDB NoSQL document store, such that arbitrary MongoDB documents can be aligned on existing domain ontologies and accessed with SPARQL. Finally, we draw on a real-world use case to report experimental results with respect to the effectiveness and performance of our approach.