JurisBERT: Transformer-based model for embedding legal texts

• Ano de defesa: 2022
• Autor(a) principal: Charles Felipe Oliveira Viegas
• Orientador(a): Renato Porfirio Ishii
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Fundação Universidade Federal de Mato Grosso do Sul
• País: Brasil
• Palavras-chave em Português: Retrieving Legal Precedents, Semantic Textual Similarity, Sentence Embedding, BERT
• Link de acesso: https://repositorio.ufms.br/handle/123456789/5119

We propose in this paper a new extension of BERT (Bidirectional Encoder Representations from Transformers), called JurisBERT. It is applied in Semantic Textual Similarity (STS) and there is a considered improvement in fastness, in precision and it requires less computational resources than other approaches. JurisBERT was trained from scratch with specific domain texts to deal with laws, treatises, and precedents, and has better precision compared to other BERT models, which was our main finding of this work. Furthermore, our approach considers the concept of sublanguage, i.e., a model pre-trained in a language (Brazilian Portuguese) passes through refining (fine-tuning) to better attend to a specific domain, in our case, the legal field. JurisBERT includes 24,000 pairs of ementas with degrees of similarity varying from 0 to 3. We extract these ementas from search mechanisms available on the courts' websites, in order to validate the approach with real data. Our experiments showed JurisBERT is better than other models in four scenarios: multilingual BERT and BERTimbau without fine-tuning in around 22% and 12% precision (F1), respectively; and with fine-tuning in around 20% and 4%. Moreover, our approach reduced 5 times the training steps, besides using accessible hardware, i.e., low-cost GPGPU architectures. This result demonstrates that not always pre-trained models, such as BERT Multilingual and BERTimbau, which are heavy, and require specialized and expensive hardware, are the best solution. So, we have proven that training the BERT from scratch with domain-specific texts has greater accuracy and shorter training time than large and general pre-trained models. The source code is available at https://github.com/juridics/brazilian-legal-text-dataset.