Processamento de documentos jurídicos longos: comparação e avaliação de métodos baseados em Modelos de Linguagem

Abstract

The Brazilian legal system faces a structural scenario of accumulating lawsuits, which compromises its efficiency and adherence to the constitutional principle of a reasonable trial duration. As of August 2024, the number of pending cases exceeded 80 million, according to data from the Judiciary's National Database. This substantial volume of litigation directly impacts the speed and effectiveness of judicial service delivery, necessitating the development of technological tools capable of supporting the management, screening, and comprehension of legal documents—documents whose length and complexity challenge traditional human and automated analysis workflows. In this context, understanding how advanced text processing techniques can contribute to streamlining judicial activity lies at the heart of this dissertation. Existing applications in the field, such as BumbaBERT, show promising results for optimizing procedural workflows but remain constrained by the structural limitations of their underlying Transformer architecture, particularly due to high computational complexity. To address this issue, this dissertation project proposes and evaluates a set of strategies aimed at the efficient processing of long documents, using initial complaints linked to Incidents for the Resolution of Repetitive Demands (IRDR) as a case study. Building on gaps identified in the literature and the practical motivation stemming from the UEMA-TJMA technical cooperation agreement—specifically regarding the difficulty of adapting language models to extensive legal texts—the methodological process was guided by the Data Science Trajectories (DST) framework. This approach provided a foundation for understanding the domain and planning solutions, as well as for identifying a taxonomy of methods capable of organizing the field of automatic long-document classification into three categories: truncation methods derived from baselines (e.g., BumbaBERT, LegalBERT-PT); ...decomposition-recomposition (e.g., ToBERT) and content synthesis based on sentence selection strategies (e.g., TextRank, LexRank, SBERT, LlaMa). Based on this framework, empirical experimentation and statistical validation were conducted. Consequently, the study involved implementing and comparing eight architectures based on fine-tuning BumbaBERT—totaling 40 experiments—that considered performance metrics such as accuracy, F1-score, precision, and recall; computational efficiency indicators such as time, inference speed, and memory usage; statistical significance tests; and practical implementation feasibility. The results demonstrated that hierarchical architectures outperform content synthesis-based approaches, achieving a better balance between precision and stability, albeit at a higher computational cost. This finding reinforces the importance of preserving the integral argumentative structure of legal texts to ensure interpretive consistency and the reliability of automated classifications. Thus, the work contributes scientifically to the advancement of natural language processing in the legal domain by demonstrating how established strategies can be reinterpreted and adapted to address the linguistic and structural specificities of Brazilian legal texts. From a technological and institutional perspective, the study offers a reproducible artifact capable of integration into the TJMA’s automation system, thereby contributing to reduced case processing times and the strengthening of digital transformation policies in the public sector. Finally, regarding the social dimension, the study reaffirms the role of digital transformation as a tool for democratizing access to justice, fostering innovation that combines technical precision, ethical responsibility, and a commitment to the public interest.