ieeexplore.ieee.org
Integrating Large Language Models and Ontologies for Legal Text Impact Extraction and Classification: an Application to Regulatory Impact Assessment | IEEE Conference Publication | IEEE Xplore
Regulatory Impact Assessment is an essential process for evaluating policy feasibility and balancing stakeholder interests prior to enactment. However, manual analysis is often time-consuming and error-prone due to the unstructured nature of legal documents and complex cross-referencing networks. To address this challenge, we propose a framework that automates legal information extraction and multi-label impact classification by integrating a Large Language Model (LLM) with ontologybased reasoning. Initially, an LLM utilizes few-shot prompt engineering to extract structured entities (agents, actions, states) and primary legal signals (deontic words, impacted targets) from raw texts. Subsequently, this extracted data is mapped into a domain-specific ontology, where a HermiT reasoner applies formal Description Logic axioms to automatically classify impacts into qualitative or quantitative benefits, costs, and constraints. Experimental evaluations on a curated dataset demonstrate the framework's effectiveness, achieving an F1-Score of 84.30% in impact identification, and a Macro-F1 of 84.71% with an Exact Match of 75.15% in complex multi-label classification. By delegating reasoning logic to the ontology, our approach effectively mitigates LLM hallucination, delivering precise, consistent, and explainable legal analysis.
