This paper introduces an annotation proposal for the reflexive pronoun –se in Brazilian Portuguese with a view to classifying different strategies for impersonalization through the use of one supercategory. We carried out experiments on a gold standard treebank for Portuguese in the Universal Dependencies project and verified that the implementation of our proposal results in the training of a morphosyntactic annotation model that annotates syntactic dependencies 1.27 percentage point better in accuracy. Moreover, a detailed evaluation showed an increase of up to 6.34 accuracy in the annotation of verb arguments, one of the most important classes for carrying out various Natural Language Processing tasks, highlighting the importance of informed linguistic modeling decisions in practical NLP results.