Detecting Transitional Provisions in EU Legislation with Hybrid AI for Better Regulation

Maintaining an accessible and up-to-date legislative framework is crucial for ensuring transparent governance and enabling effective citizen participation in democratic processes. This paper examines methodological approaches for the automated detection and modeling of textual modifications within EU Regulations and Directives, with particular attention to transitional provisions and temporal concepts (retroactivity, postponement, relative dates), as well as legislative ones (entry into force, application, sunset clause). We present a comparative analysis of four distinct approaches: (1) a rule-based pipeline utilizing regular expressions to capture standardized legal expressions, (2) a SpaCy-based pipeline leveraging advanced NLP capabilities for entity recognition and pattern matching, (3) a Large Language Model (LLM)-based pipeline using the OpenRouter API for contextual extraction, and (4) RDF/Cellar extraction serving as the benchmark. Each methodology is designed to generate standardized Akoma Ntoso (AKN) metadata annotations, facilitating point-in-time understanding of legislative modifications and temporal validity. The rule-based and SpaCy-based approaches offer computational efficiency and procedural transparency, while the LLM-based method demonstrates potential for handling complex linguistic variations and contextual dependencies. RDF extraction provides authoritative temporal information for evaluation purposes. This comparative framework provides insights into the relative strengths and limitations of each approach, contributing to the ongoing development of advanced legal informatics tools for legislative drafting and eParticipation initiatives.