Abstract: This article addresses the rhetorical labeling of judicial decisions using hierarchical neural networks, focusing on decisions from the Indian judicial system. To this end, four datasets containing judicial documents annotated with different rhetorical labels are used. The study uses the Hierarchical Sequential Labeling Network model as a basis, which combines representations extracted from BERT at the token level, a Bi-LSTM layer, and an attention pooling layer, allowing the generation of sentence-level representations and facilitating the identification of specific rhetorical roles.
Data: The work is based on four datasets: (i) Build, composed of 184 judicial documents from the Indian judicial system, with rhetorical labels in criminal and tax cases; (ii) Paheli, which includes 50 decisions from the Supreme Court of India categorized into five branches of law; (iii) M-CL, focused on 50 competition law cases; and (iv) M-IT, with 50 decisions on tax law. Each document is segmented into sentences annotated with a fixed set of rhetorical roles (7 or 13 labels, depending on the case), divided into training (80%), validation (10%), and test (10%) partitions.
Base model: All experiments are built on the Hierarchical Sequential Labeling Network model. This model encodes each sentence independently, using BERT to obtain token-level representations, which then pass through a Bi-LSTM network to capture dependencies within the text. Then, an attention layer obtains the final representation of the sentence, used to predict its rhetorical role.
This type of solution is especially relevant for legal technology and natural language processing companies like Q2BSTUDIO, a company specialized in development and technological services. At Q2BSTUDIO, we work on integrating complex NLP models for tasks such as automated legal analysis, semantic legal search systems, and structuring unstructured information. Research like the one described here serves as a reference framework for building tools that optimize legal decision-making through AI and advanced linguistic analysis.
At Q2BSTUDIO, we believe in the fusion of technological innovation and the needs of the legal world, and we continuously adopt cutting-edge models like those presented in this article to develop practical, efficient solutions aligned with the ethical and technical regulations necessary for their implementation in real-world environments.
