Uso de processamento de linguagem natural e aprendizagem de máquina para a extração de informação em editais de licitações não-estruturados.

Abstract

Bidding is the means adopted by the public administration to ensure equality of conditions for all who want to contract products or services with the government. It is also the role of the government to carry out the analysis and audit of documents derived from this process, in order to guarantee legal principles such as isonomy, legality, impersonality, morality, publicity, and administrative probity. Most documents related to bidding processes use the Portable Document Format (PDF). Such an unstructured format makes automated textual analysis more complex. The present work aims to develop an induction model, based on supervised classification, that is able to identify specific information contained in a bidding document, and thus add a layer of automation to the document audit process. For this, natural language processing techniques were used, and different machine learning models were analyzed to select the best model to be used in the classification task. The database used was extracted from the Portal of the Government of the State of Acre. At the end of the implementation, the model obtained great results and was able to identify the information of interest present in the documents in a satisfactory way.