Extração de esquemas em documentos legais não estruturados utilizando LLMs

Abstract

This work explores the use of Large Language Models (LLMs) for schema extraction from unstructured legal documents. The study proposes an approach that integrates data collection techniques, text preprocessing, prompt engineering, and embedding-based evaluation to examine the applicability of LLMs in identifying and organizing legal information efficiently. For evaluation, the Mistral 7B Instruct v0.2 model was applied to 471 Price Registration Records (ARPs), segmented into 2,000-token blocks. Results from the per-document analysis indicate strong performance in data type classification (Type Accuracy: 0.946; Type Precision: 0.972), but moderate semantic performance (Semantic Accuracy: 0.412; Semantic Coverage: 0.714), revealing consistent typing but limitations in semantic correspondence. A second evaluation stage examined a unified JSON constructed from all extracted schemas, in which semantic metrics reached maximum values, showing that the model successfully recovers all expected fields at least once, although with increased noise. The findings suggest that LLMs can serve as auxiliary tools for legal schema extraction, reducing manual effort, but still require additional semantic verification and structural consolidation strategies for practical deployment.