Similaridade Semântica entre Processos Judiciais: Uma Abordagem para Identificação de Litigância Predatória

Abstract

According to the National Council of Justice, the Brazilian Judiciary receives more than 30 million court cases every year. This excessive volume of data highlights the need for automated solutions to increase the efficiency and productivity of the judicial system. In this scenario, one of the challenges faced is predatory litigation, characterized by the filing of lawsuits in bad faith, in which the same parties (plaintiff, defendant and lawyer) submit multiple claims on the same issue, generating exhaustion of judicial resources. In 2020, around 30% of civil and consumer law cases were classified as predatory litigation, generating a minimum cost of R$ 10.7 billion to the judiciary. In addition to the financial losses, this practice overloads judges and civil servants, increasing the time it takes to process legitimate cases and compromising the efficiency of the system. This paper investigates the use of embeddings based on Transformers, such as Bidirectional Encoder Representations from Transformers, to analyze predatory litigation patterns by means of semantic similarity. These models are capable of capturing the semantic meaning of words and phrases, considering the context in which they are inserted. Information Retrieval techniques were used to construct and evaluate the research, which allows the effectiveness of the models to be measured in identifying semantically similar cases. Four language models were evaluated (BERTikal, BERTimbau, BumbaBERT Small and RoBERTa ptBR), of which BERTimbau had the best performance (Recall@k of 69.23% and MAP of 35.58%), followed by BumbaBERT Small (Recall@k of 61.54% and MAP of 31.13%). Other models tested performed less well. These results demonstrate the viability of the proposal, contributing to the automation of tasks in the Judiciary and advances in Natural Language Processing techniques