Uma metodologia para agrupamento de processos judiciais baseada em aprendizagem profunda aplicada à justiça trabalhista brasileira

Abstract

The Brazilian legal system establishes measures to ensure the speed of resolution of lawsuits, such as the principle of reasonable duration of lawsuits, the principle of speed, procedural economy and due legal flow, with a view to optimizing the judicial process. However, the Judicial Courts are experiencing a period of major budget cuts and a reduction in the number of magistrates and civil collaborators. In this current situation, Information Technology, more specifically Artificial Intelligence (AI), has been successful in facing many complex problems in the area of Natural Language Processing (NLP), which can help to alleviate the difficulties faced by these restrictions. Among the needs that can be met with AI, the automation of document classification is of great interest, because, when correctly classified, they can indicate paths to be followed by the judiciary to ensure the speedy resolution of legal proceedings. In this context, this work seeks to detect the degree of similarity between court documents that can be achieved in the inferred groups through the use of unsupervised learning. Then, through the application of nine Natural Language Processing techniques, which are (i) Frequency of Terms - Inverse Frequency of Documents (TF-IDF); (ii) Word2Vec with CBoW specialized in the Brazilian language corpus; (iii) Word2Vec with Skip-gram specialized in the Brazilian language corpus; (iv) BERT (Bidirectional Encoder Representations from Transformers) trained for general purposes for Brazilian Portuguese; (v) BERT specialized with the corpus of the Brazilian labor judiciary; (vi) General purpose trained GPT-2 for Brazilian Portuguese; (vii) GPT-2 specialized in the corpus of the Brazilian labor judiciary; (viii) Roberta trained for general purposes in Brazilian Portuguese; and (ix) Roberta, specialized in the corpus of the Brazilian labor judiciary, developed a model for grouping lawsuits, which is calculated based on the cosine of the distance between the elements of the group to its centroid. The Ordinary Appeal (acronym in Portuguese for “Recurso Ordinário” - RO) was chosen as the type of reference document because it is the document that triggers the process to go to the higher court and because there is currently a large contingent of lawsuits awaiting judgment in the 2nd instance. After the steps of data extraction, pre-processing and cleaning, the documents had their content transformed into a vector representation, using the nine NLP techniques mentioned above. For specialized BERT, specialized GPT-2 and specialized RoBERTa, a previous step of generating dense vectors of information representation - word embeddings was necessary. Preliminarily, through the presented study, it can be seen that the use of specialized word embedding models is an important advance in the quality of models that use document content as resources, especially when it comes to NLP models based on Transformers architecture.