Legal Domain Adaptation in Portuguese Language Models - Developing and Evaluating RoBERTa-based Models on Legal Corpora

• Ano de defesa: 2024
• Autor(a) principal: Garcia, Eduardo Augusto Santos
• Orientador(a): Silva, Nádia Félix Felipe da
• Banca de defesa: Silva, Nádia Félix Felipe da, Lima, Eliomar Araújo de, Soares, Anderson da Silva, Placca, José Avelino
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Programa de Pós-graduação em Ciência da Computação (INF)
• Departamento: Instituto de Informática - INF (RMG)
• País: Brasil
• Palavras-chave em Português: Processamento de linguagem natural; Modelo de linguagem; Domínio legal; Benchmark Jurídico
• Palavras-chave em Inglês: Natural language processing; Language model,; Legal Domain; Legal Benchmark
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tede/13781

This research investigates the application of Natural Language Processing (NLP) within the legal domain for the Portuguese language, emphasizing the importance of domain adaptation for pre-trained language models, such as RoBERTa, using specialized legal corpora. We compiled and pre-processed a Portuguese legal corpus, named LegalPT, addressing the challenges of high near-duplicate document rates in legal corpora and conducting a comparison with generic web-scraped corpora. Experiments with these corpora revealed that pre-training on a combined dataset of legal and general data resulted in a more effective model for legal tasks. Our model, called RoBERTaLexPT, outperformed larger models trained solely on generic corpora, such as BERTimbau and Albertina-PT-*, and other legal models from similar works. For evaluating the performance of these models, we propose in this Master’s dissertation a legal benchmark composed of several datasets, including LeNER-Br, RRI, FGV, UlyssesNER-Br, CEIAEntidades, and CEIA-Frases. This study contributes to the improvement of NLP solutions in the Brazilian legal context by openly providing enhanced models, a specialized corpus, and a rigorous benchmark suite.