Software requirements are commonly written in natural language, making them prone to ambiguity, incompleteness and inconsistency. By converting requirements to formal semantic representations, emerging problems can be detected at an early stage of the development process, thus reducing the number of ensuing errors and the development costs. In this paper, we treat the mapping from requirements to formal representations as a semantic parsing task. We describe a novel data set for this task that involves two contributions: first, we establish an ontology for formally representing requirements; and second, we introduce an iterative annotation scheme, in which formal representations are derived through step-wise refinements.