Classificação do teor das intimações recebidas do sistema processo judicial eletrônico: uma análise comparativa dos algoritmos de redes neurais recorrentes

Abstract

This study aimed to develop a Machine Learning model for analysis and classification of the genre of the movement, based on the content of subpoenas received from the Electronic Judicial Process (PJe) system, within the scope of the 1st instance, with the purpose to speed up the work of Public Prosecutors. For that, the effectiveness of different Network architectures was explored Recurrent Neural Networks (RNNs) in classifying the content of judicial subpoenas received by the Public Ministry of State of Bahia (MPBA). Three variants were used of RNNs - Simple RNN, LSTM and BLSTM - to classify texts. The models were trained and evaluated with based on test accuracy, precision and recall, using the confusion matrix. The prediction results showed that the LSTM and BLSTM models presented greater overall accuracy, while the Simple RNN models, although less accurate, required less time of training. The overall accuracy metrics, precision for the classification of the genre of the "Science" movement and the revocation to "Non-Science" of the chosen models of each architecture were, respectively: Simple RNN - 86%, 76% and 83%; LSTM - 88%, 80% and 87%; BLSTM - 88%, 79% and 86%.