Measuring rhetorical styles to understand political ideology and economic discourse in the Brazilian Senate

Abstract

This dissertation uses the Evidence–Minus–Intuition (EMI) measure to analyze rhetorical style in speeches delivered in the Brazilian Federal Senate between 1995 and 2022. EMI measures the relative proximity of political speeches to evidence-oriented and intuition-oriented language. The analysis combines word embeddings, dictionary-based rhetorical measurement, topic modeling, TF–IDF lexical representations, and supervised learning models to examine how rhetorical style varies over time, across ideological classes, and across substantive macro-themes. The results show a substantial decline in EMI over time, suggesting a movement away from evidence-oriented language and toward more intuition-oriented rhetoric. However, EMI distributions overlap substantially across left, center, and right-wing speeches. The supervised learning exercises show that binary EMI orientation and topic indicators do not improve the TF–IDF lexical baseline in the simple ablation design. Extended specifications with senator-level and electoral covariates improve random-split performance, but robustness checks indicate weaker generalization. Overall, EMI is best interpreted as a complementary measure of rhetorical style rather than as an independent classifier of party-based ideology