Interacting with Political Narratives Through LLMs: An Approach Based on Ontologies and Graph Embeddings

Narratives are essential tools through which politicians and public figures construct shared meanings and shape public perception, both locally and globally. This paper introduces a computational approach for systematically identifying and analyzing narrative structures in political speeches, aiming to enhance our understanding of how politicians try to convey their messages. A novel ontology, OntoNarr (Ontology for Narrative Representation), is defined and used to identify narrative schemas within the full text of the speeches. The core contribution is a more interpretable and conceptually coherent method for comparing political speeches based on their underlying narrative structures. This is achieved by converting ontology-based representations into graph embeddings and visualizing them using scatterplots. Unlike traditional NLP pipelines that rely primarily on lexical and syntactic features, this method incorporates a formal semantic structure, addressing key limitations in conventional analysis. A case study involving speeches from four politicians demonstrates how historical context influences the choice of narrative schema while also revealing some cross-temporal and cross-ideological similarities. Lastly, a method from granular computing is used to quantitatively evaluate the ontology-based approach.