Representações Embeddings Orientadas à Linguagem jurídica Brasileira

Abstract

The automatic processing of legal texts arranged in natural language makes it possible to build a range of applications based on artificial intelligence, such as classification and grouping of processes by subject, document summarization, and translation into citizen language. In this sense, the Brazilian judiciary launched the Justice 4.0 program, looking to encourage the development of solutions that offer speed in procedural activities. Notably, technical language is predominant in this application domain, requiring specialized models for the segment. Bearing in mind this context, this work aims to build models embeddings oriented to the legal sphere with a view to feeding applications in the area. In this sense, approximately 5.3 million documents were extracted from Brazilian justice institutions from the most varied spheres, such as civil, criminal, and labor. The models were evaluated by classifying initial petitions, and the results obtained were promising when compared to generalist models of the Portuguese language. Such research findings demonstrate that models trained with legal documents better understand the segment’s language’s specificities and can potentially promote new applications for the sector