Detecção de empresas potencialmente não confiáveis por meio de extratos de compras governamentais: uma aplicação com modelos de linguagem natural.

Abstract

In the context of government procurement in Brazil, efficiency and continuous monitoring of expenditure represent significant challenges for public management. In 2023, the Brazilian Federal Government issued a total of 1,761,910 invoices for various types of acquisitions, resulting in an amount of R$ 76.62 billion in negotiations with private entities (TRANSPARÊNCIA, 2024). These government procurements cover a broad spectrum of the public sector, including procurement for the purchase of materials used in the construction of highways, maintenance of schools and hospitals, use of goods by public servants, among other purposes. The acquisition of these inputs is distributed in various locations throughout the country, generating a growing and diverse volume of information, present in contracts and invoices for products and services.

However, these government procurements are often a fertile ground for the occurrence of collusion and fraud (OECD, 2007), such as overpricing of products, supplier monopolies, bribes to public officials, etc.

The objective of this work is to compare the performance of Natural Language Processing (NLP) models in the task of detecting, based on extracts of government purchases, companies that have already been punished by government agencies, such as General Comptroller of the Union. The data used are public and periodically updated through the Federal Government's Open Data portal.

The results of this work show that it is possible to use natural language models as a pre-stage of investigation of suspicious purchases, providing a classification of potentially problematic purchases that can later be evaluated by an expert, thus reducing the human workload by reducing the purchase list to a smaller and more focused amount.