Protótipo de um sistema de geração automatizada de pareceres de auditoria baseado em Aprendizado de Máquina

Abstract

This work presents the development of a prototype artificial intelligence system for automating the generation of audit reports in municipal management, developed in partnership with the Comptroller and Audit Office of the Municipality of Sobral, Ceará, Brazil. The system uses natural language processing and machine learning techniques to analyze responses from audited agencies and generate automated reports based on historical data. The methodology involved collecting and analyzing audit data from 2017 to 2023, including information on audited items, agency responses, and the corresponding issued reports. The system was developed using technologies such as Python, machine learning libraries (scikit-learn), natural language processing (sentence-transformers, NLTK), and an interactive web interface with Gradio. It implements semantic similarity algorithms to identify patterns in historical responses and generates reports tailored to the specific context of each new response, incorporating contextual compatibility checks and the removal of inappropriate specific information. System validation demonstrated its ability to produce consistent and contextually relevant reports, with varying confidence levels based on similarity to historical cases. The results indicate that the system can significantly contribute to the modernization and efficiency of public audit processes by reducing the time required for report preparation and standardizing the quality of analyses carried out by the municipal comptroller.