A Graph-Enhanced LLM-Based Question Answering System for the AI Act

The rapid evolution of Artificial Intelligence (AI) has brought both new opportunities for innovation ethical challenges across various sectors of our society. To address these challenges, the European Union (EU) approved the AI Act in 2024, with the goal of establishing a regulatory framework for the creation of AI systems in a trustworthy and ethically balanced manner. However, given the unstructured nature of such documents and their complex language, ensuring compliance with such regulations remain a complex task. This thesis proposes a graph-enhanced, LLM-based Question Answering (QA) system designed to automatically retrieve requested insights from the AI Act. The proposed approach combines the interpretability of structured graph representations with the retrieval capabilities of Large Language Models (LLMs), addressing the issue of providing accountable and traceable answers to user's questions. Inspired by the GraphReader framework, but adapted for the legal domain, this model structures the AI Act into Chunks, AtomicFacts, and KeyElements, creating a graph-based representation that helps efficient and transparent information retrieval. Unlike traditional Knowledge Graph (KG)-based approaches, which require manually defined ontologies and explicit entity linking, this system leverages LLMs for document structuring, storage, and retrieval, ensuring scalability and adaptability to evolving regulations.