Inteligência artificial na análise de patologias corruptivas: delimitação jurisprudencial nas decisões do TCU do conceito aberto de cláusula restritiva ao caráter competitivo em editais de licitação

Nenhuma Miniatura disponível

Data

2020-12-22

Título da Revista

ISSN da Revista

Título de Volume

Editor

Universidade Federal de Goiás

Resumo

In Brazil, the innovations brought by the new Civil Procedure Code (Law No. 13105/2015), formalized the importance of using precedents at all levels in the Brazilian legal system. This research arises from the need to congregate all decisions that deal with the same theme, in contrast to simple keyword searches used these days, in order to provide greater legal certainty. In administrative law, the Brazilian Federal Audit Court’s (TCU) decisions serves as a guide for public actions, as they establish practical criteria that allow public managers to conclude for a possible solution in a specific case. In this context, this empirical research had as main objectives: (i) to create a replicable method of delineating jurisprudence and identifying precedents with the use of Artificial Intelligence (AI) applicable, above all, to TCU decisions that deal with clauses restricting the competitiveness in bidding notices; and (ii) to contribute to the administrative precedents culture sedimentation and to the jurisprudential delimitation of indeterminate legal concepts, especially those related to corrupt practices. The area chosen for jurisprudential extraction application is found at the heart of frauds in public bidding, that one of its corruption pathologies is the existence of clauses in bidding notices that restrict the competitiveness in public purchases. Thus, this research tested AI techniques (text mining for document classification) in the TCU decisions (between 1992 and 2018), with the purpose of systematize the jurisprudential delimitation and consequent unambiguous administrative precedents identification, which give concreteness to the concept of competitiveness restrictive clauses. More than 300 judgments, previously labeled in 11 different classes related to bidding clauses considered restrictive to the competitive nature of the public purchase, were used to train machine learning and deep learning models for multilabel classification, in order to verify whether the machine would be able to point out which others TCU decisions were related to any of these classes. The results using convolutional neural networks for the training and test phases proved to be reasonable, as they presented an evaluation metric of 82.69%. However, in the supervised assessment stage, the trained deep learning algorithm was inconclusive for the desired jurisprudential delimitation. Despite the unsatisfactory results this research was successful, at least partially, in reaching its general objectives, since the mapping of the computational state of art applied to the research problem and the operational details described provide know-how transferring for futures research that envisages the use of AI for jurisprudential systematization. It was also demonstrated the importance of systematically tracking administrative precedents so that there is greater legal security not only for public managers in their administrative actions, but also in improving the algorithms used by AI applications that search for possible irregularities, such as Alice the robot from TCU. Still, this study presents the construction of the jurisprudential line with indication of the probable administrative precedents and their rationes decidendi for the cases of clauses in bidding notices, considered restrictive because they demand technical capacity from bidders in an irregular manner.

Descrição

Citação

SILVA, S. T. T. Inteligência artificial na análise de patologias corruptivas: delimitação jurisprudencial nas decisões do TCU do conceito aberto de cláusula restritiva ao caráter competitivo em editais de licitação. 2020. 197 f. Dissertação (Mestrado em Direito e Políticas Públicas) - Universidade Federal de Goiás, Goiânia, 2020.