Institute of Formal and Applied Linguistics

Court verdicts contain rich information about criminal behavior, but their unstructured format limits systematic analysis. We present an LLM-assisted pipeline for constructing structured datasets from factual statements contained in the verdict texts, combining attribute exploration, span annotation, and value clustering. Attributes are partly defined by legal experts and partly induced from the data, enabling a comparison between expert-defined and data-driven schema design. We validate our results on the Slovak dataset through careful manual evaluation, highlighting both the strengths and current limitations of the proposed pipeline and their impact on downstream legal analytics.