Abstract: Docket files, also known as plumitifs, are legal text documents describing judicial cases. They are present in most jurisdictions and are meant to provide a window on legal systems. They contain information of a judicial case such as parties’ identities, accusations’ provisions, decisions, and pleas. However, this information is cryptic, using abbreviations, and making references to the criminal code. In this paper, we explore the use of neural text generators to improve the legal accuracy of the docket file verbalization regarding the accusations, decisions, and pleas sections. We introduce a legal accuracy evaluation scale used by jurists to manually assess the performance of three architectures with different levels of prior knowledge injection. We also study the correlation of our human evaluation methodology with automatic metrics.