Método não supervisionado de sumarização extrativa de textos jurídicos com alinhamento de grafos semânticos guiados por atenção

Abstract

The massive volume and technical complexity of legal documents in Brazil impose a major challenge to the efficiency of the judicial system. Automatic summarization emerges as an alternative to mitigate this overload and assist the work of judges and lawyers. However, the application of deep learning models in Law faces critical obstacles: algorithmic opacity (“black-box”), the unacceptable risk of factual hallucinations in generative models, and the severe scarcity of labeled data for training. Thus, the development of solutions that unite factual fidelity and interpretability is essential. In this context, this work proposes an unsupervised extractive summarization method focused on the legal domain, structured on the modeling of semantic graphs guided by attention mechanisms. The method extracts self-attention weights from an expert language model (Legal-BERTimbau) and filters noisy connections via dynamic binarization using Otsu’s method. The text is converted into a directed graph, thematically partitioned by the Hierarchical Infomap algorithm to isolate the argumentative axes. Topic alignment is performed in a dense vector space (Sentence-BERT), and sentences are ranked by the Maximum Attention heuristic, respecting a strict compression limit of 10%. In the evaluation using the RulingBR dataset, the proposed model outperformed classical unsupervised algorithms in the ROUGE-1 (36.61%) and ROUGE-L (20.74%) metrics. Additional experiments with an Extractive Oracle demarcated the upper bound of the task at a ROUGE-1 of 65.21% and ROUGE-L of 47.37%, while a hybrid extractive approach guided by an LLM (GPT-5 mini) achieved a ROUGE-L of 21.31%. Thus, the developed method proves promising by ensuring the integrity of the original text, free from procedural hallucinations, additionally offering a visual explainability interface that makes sentence selection fully auditable.