Answers are grounded in this paper and cite their sources.
Public administration continuously issues official acts through official gazettes; however, their publication in unstructured formats hinders automated querying and computational reuse. This article investigates the ability of small language models (SLMs) to extract structured information from administrative acts into a unified JSON schema, comparing zero-shot and one-shot prompting regimes. A two-stage benchmark protocol is proposed and applied to acts from the Official Gazette of the State of Ceará: an exploratory evaluation of various models using 50 instances to assess structural compliance, followed by an evaluation involving 2,300 acts—measuring structural compliance, average inference time, and informational agreement with extractions produced by GPT-4o (used as an automated reference). The results reveal performance variations among the models across the evaluated dimensions. Qwen2.5-1.5B demonstrated the highest structural compliance, reaching 96.18% in zero-shot and 95.44% in one-shot settings, with average inference times of 33.04 and 36.91 seconds, respectively. Conversely, Qwen3-4B achieved the highest average informational agreement (0.465 in one-shot), performing particularly well in the similarity of summaries (0.788) and the comparison of complex lists, despite lower structural compliance (approximately 77%). It should be noted that the one-shot regime did not consistently improve model performance, and that structural compliance does not necessarily imply higher informational agreement. The study concludes that SLMs show potential for specific administrative information extraction tasks, provided that their selection process separately considers structural reliability and agreement with an automated reference.
No takes yet. Share an insight, caveat, or question.
Silva et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: