Structuration argumentative des débats parlementaires par IA : vers une modélisation interprétable des dynamiques discursives

This paper proposes an interpretable computational framework for structuring parliamentary debates through argumentative annotation, semantic clustering, and relational graph modeling. Drawing on open data from the French National Assembly, we first review existing work in argument mining, computational political analysis, stylometry, and document structuring in the context of digital humanities. We then describe a processing pipeline based on large language models for sentence-level argumentative categorization, complemented by sentiment analysis and semantic embeddings. The resulting structured representation enables the construction of an argument-centered knowledge graph, onto which parliamentary actors are projected according to their discursive practices. The analysis highlights both the internal argumentative coherence of individual interventions and the relative rarity of explicit argumentative relations between interventions, reflecting the institutional nature of parliamentary discourse. Beyond these empirical findings, the paper demonstrates how such a structuring approach enhances the interpretability, explorability, and downstream usability of political corpora, particularly in the context of Retrieval-Augmented Generation systems. More broadly, it argues that document structuring should be considered not merely as a technical preprocessing step, but as an interpretative infrastructure for computational hermeneutics in the humanities and social sciences.