POPOLARE: A Populism and Polarization Classification Framework for Italian Texts

The influence of political discourse in forming public opinion has intensified the need for tools able of capturing complex ideological patterns. This require innovative, data-driven approaches to analyze and interpret political language with both precision and transparency. This thesis presents POPOLARE, a two-fold populism and political polarization framework for political Italian speeches based on Natural Language Processing and Machine Learning. Individual transcripts are transformed into textual document representations and then aggregated to derive representations for each speaker. Based on these representations, populism and polarization detection tasks are performed. A key novelty lies in the use of Generative AI for data annotation, and explainability techniques for model interpretation. Results show that simple models combined with lexical representations perform best, and that interpretable features enhance both accuracy and transparency. POPOLARE provides a replicable approach for ideological analysis, with future directions including multilingual extension and deeper use of explainable AI.