Abstract
Recent advancements in Artificial Intelligence have yielded promising results in addressing complex challenges within Natural Language Processing (NLP), serving as a vital tool for expediting judicial proceedings in the legal domain. This study focuses on the detection of similarity among judicial documents within an inference group, employing eight NLP techniques grounded in transformer architecture, specifically applied to a case study of legal proceedings in the Brazilian judicial system. The transformer-based models utilised - BERT, GPT-2, RoBERTa, and LlaMA - were pre-trained on general-purpose corpora of Brazilian Portuguese and subsequently fine-tuned for the legal sector using a dataset of 210,000 legal cases. Vector representations of each legal document were generated based on their embeddings, facilitating the clustering of lawsuits and enabling an evaluation of each model's performance through the cosine distance between group elements and their centroid. The results demonstrated that transformer-based models outperformed traditional NLP techniques, with the LlaMA model, specifically fine-tuned for the Brazilian legal domain, achieving the highest accuracy. This research presents a methodology employed in a real case involving substantial documentary content that can be adapted for various applications. It conducts a comparative analysis of existing techniques focused on a non-English language to quantitatively explain the results obtained with various NLP transformers-based models. This approach advances the current state of the art in NLP applications within the legal sector and contributes to the achievement of Sustainable Development Goals.