Process models play an important role for specifying requirements of business-related software. However, the usefulness of process models is highly dependent on their quality. Recognizing this, researches have proposed various techniques for the automated quality assurance of process models. A considerable shortcoming of these techniques is the assumption that each activity label consistently refers to a single stream of action. If, however, activities textually describe control flow related aspects such as decisions or conditions, the analysis results of these tools are distorted. Due to the ambiguity that is associated with this misuse of natural language, also humans struggle with drawing valid conclusions from such inconsistently specified activities. In this paper, we therefore introduce the notion of canonicity to prevent the mixing of natural language and modeling language. We identify and formalize non-canonical patterns, which we then use to define automated techniques for detecting and refactoring activities that do not comply with it. We evaluated these techniques by the help of four process model collections from industry, which confirmed the applicability and accuracy of these techniques.