Segmentação de textos jurídicos: estudo de caso em petições

Abstract

The growing volume of legal cases highlighted in the 2023 report of the Conselho Nacional de Justiça highlights the urgency of effective strategies for screening and document analysis. With 81.4 million cases pending at the end of 2022, the Judiciary faced a major challenge in terms of management and decision-making regarding this vast set of documents. These documents, mostly textual, can be seen as a sequence of semantically coherent segments, designed to facilitate the transition between different subtopics within a single document. In this way, dividing a document into similar segments can help improve and/or speed up downstream applications. In this sense, the segmentation of initial petitions into descriptions named fact, thesis and request (FTP) emerges as a promising approach. Therefore, this study proposes the construction of an initial petition segmenter, using neural networks together with incorporations from the Bidirection Encoder Representations from Transformers model in a database of petitions to be annotated manually, discovered by Cross Industry Standard Process for Data Mining methodology. The objective lies in performing automated identification of the FTP triplet in order to provide the most relevant parts of the petition in the legal context. Furthermore, the proposed approach can be adapted to other natural language processing tasks, such as summarization and textual search. By focusing on the essential elements for analysis and response, it is expected to reduce the volume of information to be processed, thus optimizing operational efficiency. Furthermore, the division of the petition in particular facilitates the quick and accurate identification of key elements, accelerating the understanding process, especially when sent in conjunction with automated systems and solutions from the Justice 4.0 program