Mercado J.S., Ma Q., de Kinderen S., Winter K., Cabot J.
Data and Knowledge Engineering, vol. 164, art. no. 102578, 2026
Large Language Models (LLMs) have the potential to support the transformation of natural language legal text into a regulatory model, a task conventionally known to be time consuming and error prone when done manually. In this paper, we introduce CLERK: a Companion LLM Expert for modeling Regulatory Knowledge existing in natural language legal texts. CLERK captures regulatory knowledge in the format of Legal Goal Requirements Language (GRL) models. CLERK offers three key contributions, utilizing established prompting techniques: (1) Adopting the Tree-of-Thought (ToT) prompting framework, CLERK streamlines the regulatory modeling process by breaking down complex steps into manageable tasks and focusing on those essential for constructing a Legal GRL model only. (2) The ToT framework enables self-evaluation of intermediate outputs. (3) CLERK enhances consistency and clarity, by leveraging additional in-context learning prompting techniques, such as few-shot prompting and output formatting with an explicit syntax definition. Experiments with eight regulatory articles from two domains (healthcare and energy communities) display a notable improvement brought about by CLERK compared to previous approaches. This improvement pertains to identifying relevant actors, goals and their deontic modalities, as well as the relationships among goals.
