Análise textual das decisões proferidas em colegiado do Tribunal de Justiça do Tocantins

Abstract

The growing demand for speed and efficiency in the Judiciary has driven the adoption of solu- tions based on Artificial Intelligence. In this context, this dissertation proposes a methodology grounded in Natural Language Processing (NLP) and unsupervised learning for clustering ju- dicial decisions, with a focus on screening appeals submitted to the office of court presidencies. The underlying hypothesis is that automated organization of judicial decisions based on seman- tic similarity can contribute to identifying repetitive or settled themes, thereby avoiding the submission of unnecessary appeals to higher courts. The proposed methodology comprises sev- eral stages, including text preprocessing, vectorization using SBERT family models with em- phasis on the legal-domain model stjiris/bert-large-portuguese-cased-legal-mlm-sts-v1.0 clus- tering with the K-Means algorithm, dimensionality reduction via t-SNE, and evaluation of clus- tering quality through internal metrics such as Silhouette Score, Davies-Bouldin Index, and Calinski-Harabasz Index. Qualitative analysis was conducted using topic modeling with Latent Dirichlet Allocation (LDA), interactive visualizations with PyLDAvis, and word cloud gener- ation, allowing a deeper interpretation of the resulting clusters. The results showed the for- mation of thematically consistent groups, with emphasis on categories such as Tax Law, Pro- bate Law, Administrative Law, and Civil Liability. These findings demonstrate that the pro- posed approach is effective in organizing judicial decisions by semantic similarity, offering practical potential to support the screening and organization of judicial processes especially in contexts involving large document volumes and lack of labeled data.