Context: Social control is an act of exercising citizenship, which contributes to the strengthening of popular sovereignty. The Audit Courts analyze hiring data from official diaries. For there to be transparency in public administration and social control can be exercised, this data needs to be classified and presented in a friendly way. Problem: The large number of published contracts makes it difficult to process these data, and makes manual classification of the objects of these contracts almost impossible, resulting in damage to social control and consequently to the effectiveness of the public service. Solution: This paper explores automatic classification models for public procurement objects aimed at dealing with the COVID-19 pandemic, using classical and deep Machine Learning approaches. The models were trained with a set of data extracted from the manual classification of hiring published in the official diaries. IS Theory: The work was conceived following the Information Processing Theory, in particular, on the concept that compares information processing with the human learning model. Method: The research has a predictive character, and its evaluation was carried out through proof of concept. The analysis of the results was performed using a quantitative approach. Summary of Results: The obtained results achieved an accuracy of 96% using the BERTimbau model, which is a pre-trained BERT model for the Portuguese language. Additionally, the model that used deep learning outperformed the model with document embeddings by 5% and by more than 10% the models using the classical approaches. Contributions and Impact in the IS Area: The main contribution of the article is to make possible a model for automatic classification of public expenses to increase transparency and improve monitoring by Courts of Auditors and society in general.
