Measuring Discursive Illiberalism: A Lexicon-Based and Machine Learning Pipeline for Latin American Presidential Rhetoric 2000-2024

Research on democratic erosion has increasingly turned from abrupt authoritarian ruptures to the gradual, discursive normalization of illiberal politics inside electorally legitimate regimes. Yet existing measures, built largely on expert surveys and regime-level coding, do not capture the everyday rhetorical legitimation through which executives erode liberal democratic norms. This article develops and validates a scalable, replicable method for measuring discursive illiberalism at the sentence level. We introduce a two-stage pipeline: a theory-anchored lexicon of red-flag terms, paired with a synthetic filter that discards negated or third party-attributed matches, first identifies illiberal sentences and a multinomial classifier then assigns each identified sentence to one of seven theoretically grounded domains. The lexicon was built iteratively, refined over nine rounds of error analysis against real presidential Discourse, and validated against known-groups discourse and an independent Large Language Model reader. Applied as a proof of concept to Latin American presidential rhetoric (2000-2024), the instrument yields a usable, auditable ranking of presidents by illiberal intensity and composition; its precision, recall, and discriminant validity with respect to populism are still being established with the rigor this measurement problem requires. The method offers a transferable framework for the large-scale, longitudinal study of democratic backsliding in Latin America. Keywords: Illiberalism; Democratic backsliding; Political discourse; Latin America; Computational Social Science