Este artigo trata da identificação dos domínios discursivos Jurídico, Entretenimento, Jornalístico, Fórum Virtual e Instrucional do português brasileiro no nível sentencial, amostrados do corpus Carolina. Avaliamos propriedades gramaticais, lexicais e semânticas. Demonstramos que os domínios são discerníveis e se organizam em uma escala consistente que associamos à distinção oral-envolvido vs. literato-informacional a partir da comparação com outros trabalhos. Treinamos classificadores Transformer em um novo dataset de sentenças para identificação de domínios, alcançando alta performance. Os padrões de erro dos modelos correlacionam-se com a escala identificada, indicando a captura desta dimensão de variação. Disponibilizamos publicamente os datasets e modelos produzidos.

This paper presents the challenges of building Carolina, a large open corpus of Brazilian Portuguese texts developed since 2020 using the Web as Corpus methodology enhanced with concerns about provenance and typology (WaC-wiPT). The corpus aims to serve both as a reliable source for research in Linguistics and as an important resource for Computer Science research on language models. Above all, this endeavor aims at removing Portuguese from the set of “low-resource languages”. This paper details the construction methodology of Carolina, with special attention to the issue of describing provenance and typology according to international standards, while briefly describing its relationship with other existing corpora, its current state of development, and its future directions. Este artigo apresenta os desafios da construção do Carolina, um grande corpus aberto de textos em português brasileiro em desenvolvimento desde 2020 que usa a metodologia ‘Web as Corpus’ aprimorada com preocupações de proveniência e tipologia (WaC-wiPT). O corpus pretende ser utilizado tanto como fonte confiável para pesquisas em Linguística quanto como importante recurso para pesquisas em Ciência da Computação sobre modelos de linguagem. Acima de tudo, este esforço visa retirar o português do conjunto das línguas de poucos recursos. Este artigo detalha a metodologia de construção do Carolina, com especial atenção ao problema da descrição de tipologia e proveniência segundo padrões internacionais; descrevemos também brevemente a sua relação com outros corpora existentes, seu estado atual de desenvolvimento e seus rumos futuros.

This paper presents the challenges of building Carolina, a large open corpus of Brazilian Portuguese texts developed since 2020 using the Web-as-Corpus methodology enhanced with provenance and typology concerns (WaC-wiPT). The corpus aims at being used both as a reliable source for research in Linguistics and as an important resource for Computer Science research on language models. Above all, this endeavor aims at removing Portuguese from the set of “low-resource languages”. This paper details Carolina's construction methodology, with special attention to the issue of describing provenance and typology according to international standards, while briefly describing its relationship with other existing corpora, its current state of development and its future directions.

This paper shows the results of an experimental approach on automatic reading of 16th-century Portuguese Inquisition manuscripts using two Handwritten Text Recognition (HTR) applications. Therefore, the experiment involved modeling the handwriting of one scribe, measuring the differences between the human-made and the machine transcriptions and analyzing those differences from a palaeographic perspective. Two independently developed HTR applications were selected for the experiment: Transkribus (by ReadCoop, an European Cooperative Society) and Lapelinc Transcriptor (by LaPeLinC, Corpus Linguistic Laboratory of the State University of Southern Bahia, Brazil). Both tools use machine learning technology. The results indicated that palaeography can contribute to the development of HTR technologies in a number of ways, apart from the evident contribution of helping in the deep understanding of the results. Finally, serving both pragmatic and scientific interests. So, with this article we express the desire to promote discussions about the state-of-the-art and the challenges ahead on the road to what many already call ‘digital palaeography’ - in particular, with regard to Early Modern Portuguese manuscripts.