Avaliação de Modelos Generativos para Extração de Informações em Processos da Pró-Reitoria de Desenvolvimento Institucional e Recursos Humanos do Instituto Federal de Goiás

Abstract

Efficiency in administrative management represents a growing challenge in the public sector, particularly given the volume of unstructured documents requiring manual analysis. At the Federal Institute of Goiás (IFG), processes such as Training Leave require staff to dedicate significant time locating data scattered across attachments and dispatches. This work proposes and evaluates a solution based on Large Language Models (LLMs) for the automatic Information Extraction (IE) from these proceedings. The methodology adopted a two-stage architecture (Generation and Validation) and investigated the impact of four text selection strategies (Full Body, Requirement-only, Sentence Chunking, and Token Chunking) on the performance of both commercial (GPT family and DeepSeek) and open-source models (GPT-OSS, DeepSeek). Experiments demonstrated that the text selection strategy exerts a greater influence on extraction quality than the model choice itself. The "Full Body" approach proved to be the most robust, allowing models to capture complex semantic relationships that are lost in chunking strategies. Results indicate that while commercial models like GPT-4o-mini offer greater overall consistency, free models, when paired with a broad context strategy, achieve competitive similarity indices (exceeding 0.90 in specific scenarios). This study delivers a validated analysis systematic for this domain, demonstrating the technical feasibility of intelligent automation in personnel management.