Legal topics are typically organized into trees of labels, where each branch, from the root (generic topics) to the leaf (specific topics), categorizes the vocabulary that describe lawsuits. In this article, we describe an approach, innovative in the Brazilian Justice System, of automatic hierarchical classification of a petition topic. The approach integrates methods based on transformers and complex networks to capture, in addition to the characteristics of the text, the relationship between the legal citations present in the petitions and the topic to which the petition refers. The validation of this approach is done through a benchmark that shows accuracy gains, as well as, a practical scenario with the implementation of a microservice on a National Justice Platform whose front-end implementation is already being used by a State Court to automatically suggest the classification of the petitions topic in the National Procedural System.