This paper presents the development of Retrato_Cantado, a dataset of sentences extracted from Brazilian song lyrics and manually annotated to identify and categorize predicative constructions that describe individuals. The corpus findings validate the effectiveness of lexical-syntactic patterns for identifying predicative sentences, confirming their suitability for large-scale linguistic annotation tasks. The dataset also serves as a valuable resource for the analysis of textual discourse and the representation of social groups in Brazilian culture. We additionally trained a person-characterization classifier to illustrate the applicability of the dataset to the automatic detection of predicative descriptions, which achieved high accuracy and highlights the potential for creating more specialized models capable of detecting physical and sociocognitive categories, as well as performing sentiment polarity analysis.