In most current industrial software engineering projects, the majority of requirements documents are written almost entirely in natural language. However, specifying the requirements in natural language has one major drawback, namely the inherent imprecision, i.e., ambiguity, incompleteness, and inaccuracy, of natural language. Since the requirements document forms the basis of the whole development process, such defects can have severe consequences for the whole project. Therefore, it is important to deal with these defects in a requirements specification right from the start. This paper presents an approach for reducing the problem of imprecision in natural language requirements specifications with the use of natural language patterns, which allow formulating requirements sentences in a less ambiguous, more complete, and more accurate way. To ensure the applicability of our approach we based our patterns on a metamodel for requirements statements for embedded systems. With this metamodel, we ensure that all forms of requirements statements are described with the patterns. We validated the effectiveness of the patterns by using them to rewrite a substantial, previously written, requirements specification to eliminate its imprecisions.