Business processes can be captured in various forms, ranging from unstructured textual descriptions to formalized process models. This increases the risk of having deviating representations of what should describe the same process. To address this issue, this paper presents an approach for identifying deviations between process descriptions and process models and subsequently repairing these deviations in the process descriptions. Thus, it complements existing research that by now mainly focused on repairing process models. The proposed approach employs process structure trees, mapping elements from the textual description to tree nodes, and subsequently discovering deviations. Afterwards, new process fragments for the deviating parts of the description are generated by means of sentence templates and a Large Language Model. To repair the identified deviations, those fragments are then used to replace the deviating ones resulting in a description more closely aligned with the process model. The approach is empirically evaluated using an existing dataset wherein deviations are introduced. The findings show that our method can discover and repair deviations reliably and efficiently.