European and Italian legal systems are densely interconnected, making their management, interpretation, and analysis an increasingly complex challenge. Traditional document-based approaches often failed to capture the dynamic structure and the semantic dependencies that arise between national and supranational norms. This thesis addresses these challenges by adopting a strategy based on the property graph model, enabling a structured representation of legislative information as a network of entities (nodes) and relationships (edges). The main contribution of this work is the development of a complete ETL (Extract-Transform-Load) pipeline that automatically collects, parses, and normalizes legal data from the official repositories of the Publications Office of the European Union. The extracted information is transformed into a property graph that captures both the internal composition and the interrelations among European legal acts. It is then integrated with the pre-existing Italian legal graph, forming a unified multilevel legislative knowledge graph that models the relationships between the two systems. To demonstrate the potential of this integrated representation, an experimental link prediction task is carried out employing Graph Neural Networks (GNNs). The model leverages both structural and semantic features to predict indirect influences of European legislation on Italian laws. The experiments show high discriminative performance in transductive settings and robust generalization under hybrid transductive-inductive conditions, confirming the viability of graph-based and machine-learning approaches for legislative knowledge discovery. This work contributes to the advancement of graph-based legislative information systems, demonstrating once again their potentialities to support AI-assisted legislative research.
La fitta interconnessione tra i sistemi giuridici europeo e italiano rende la loro gestione, l'interpretazione e analisi una sfida sempre più complessa. I tradizionali approcci basati sui documenti non sono in grado di cogliere appieno la struttura dinamica e le dipendenze semantiche che emergono tra le norme nazionali e sovranazionali. Questa tesi affronta tali problematiche adottando una strategia basata su property graph, che consentono una rappresentazione strutturata delle informazioni legislative come una rete di entità (nodi) e relazioni (archi). Il principale contributo di questo lavoro consiste nello sviluppo di una pipeline ETL (Extract-Transform-Load) in grado di raccogliere e normalizzare automaticamente i dati legislativi provenienti dall'archivio ufficiale dell’Ufficio delle Pubblicazioni dell’Unione Europea. Le informazioni estratte vengono trasformate in un property graph che cattura sia la composizione interna che le relazioni tra le norme europee. Esso viene poi integrato con il grafo legislativo italiano preesistente, dando origine a un grafo di conoscenza legale multilivello che rappresenta interamente le relazioni tra i due sistemi giuridici. Per dimostrare le potenzialità di questa rappresentazione integrata, è stato condotto un esperimento di link prediction basato su Graph Neural Networks (GNNs). Il modello sfrutta caratteristiche strutturali e semantiche per prevedere le influenze indirette della legislazione europea sulle norme italiane. I risultati mostrano elevate prestazioni nel contesto transduttivo e una solida capacità di generalizzazione in condizioni ibride transduttivo-induttive, confermando la validità degli approcci basati su grafi e machine learning per l'espansione della conoscenza giuridica. Questo lavoro contribuisce all’avanzamento dei sistemi di informazione giuridica basati su grafi, dimostrando ancora una volta le loro potenzialità nel supportare la ricerca giuridica assistita dall’intelligenza artificiale.
Integration of italian and european legislative graphs for cross-level link prediction
BALDESSARI, GUIDO
2024/2025
Abstract
European and Italian legal systems are densely interconnected, making their management, interpretation, and analysis an increasingly complex challenge. Traditional document-based approaches often failed to capture the dynamic structure and the semantic dependencies that arise between national and supranational norms. This thesis addresses these challenges by adopting a strategy based on the property graph model, enabling a structured representation of legislative information as a network of entities (nodes) and relationships (edges). The main contribution of this work is the development of a complete ETL (Extract-Transform-Load) pipeline that automatically collects, parses, and normalizes legal data from the official repositories of the Publications Office of the European Union. The extracted information is transformed into a property graph that captures both the internal composition and the interrelations among European legal acts. It is then integrated with the pre-existing Italian legal graph, forming a unified multilevel legislative knowledge graph that models the relationships between the two systems. To demonstrate the potential of this integrated representation, an experimental link prediction task is carried out employing Graph Neural Networks (GNNs). The model leverages both structural and semantic features to predict indirect influences of European legislation on Italian laws. The experiments show high discriminative performance in transductive settings and robust generalization under hybrid transductive-inductive conditions, confirming the viability of graph-based and machine-learning approaches for legislative knowledge discovery. This work contributes to the advancement of graph-based legislative information systems, demonstrating once again their potentialities to support AI-assisted legislative research.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Baldessari.pdf
solo utenti autorizzati a partire dal 18/11/2026
Descrizione: Tesi di laurea magistrale
Dimensione
2.35 MB
Formato
Adobe PDF
|
2.35 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/246334