Abstract

The legal domain presents several challenges for Natural Language Processing (NLP), particularly due to its linguistic complexity and lack of public datasets. Named Entity Recognition (NER), a subarea of NLP, has been successfully used to extract useful knowledge from legal texts. Its widespread use is limited by the lack of legal text corpora. This paper introduces UlyssesLegalNER-Br, a comprehensive corpus of Brazilian legal documents for NER, covering bills, case laws and laws, including the first NER corpus based exclusively on Brazilian laws. This research expand the UlyssesNER-Br corpus, previously focused only on the Brazilian legislative domain. The proposed corpus has 560 public documents annotated using a hybrid approach, organized in 9 categories and 23 fine-grained types, experimentally evaluated with the CRF, BiLSTM, and BERTimbau architectures. The corpus was experimentally evaluated regarding predictive performance, computational cost and label-level results. The best micro F1 96.18% was achieved by BERTimbau on the unified corpus, providing a strong baseline for Brazilian legal NER. At the label level, six categories and seven types presented a F1-score above 95%, while the lowest were distributed in the interval 71-82%.