Libras and Artificial Intelligence

lexicographic challenges

Abstract

This article discusses the theoretical and methodological challenges of the lexicographic treatment of Brazilian Sign Language (Libras) in the age of artificial intelligence, with attention to the risks and possibilities opened by technologies for recognition, translation, annotation, and automated organization of multimodal data. The discussion is based on the understanding of Libras as a visual-spatial language, endowed with its own grammatical organization, linguistic variation, and specific modes of lexical and terminological constitution. In this sense, the article problematizes the tendency to reduce signs to Portuguese equivalents or glosses, especially in digital and computational contexts in which the signed language is converted into processable data. Methodologically, the study is characterized as a theoretical-critical reflection grounded in research on Libras lexicography and terminology, digital dictionaries and glossaries, the production of term-signs, computational processing of sign languages, datasets, annotation, algorithmic biases, and Deaf participation in artificial intelligence projects. The discussion shows that dictionaries, glossaries, and digital platforms in Libras play a relevant role in documenting both general and specialized lexicon, but still face problems such as dispersion of records, heterogeneity of formats, fragile access, insufficient metadata, gaps in validation, and dependence on Portuguese as the main search axis. It is argued that artificial intelligence may contribute to the organization, indexing, retrieval, and comparison of signs, provided that it does not replace linguistic analysis, human curation, and community validation. The article also emphasizes that poorly representative datasets, gloss-centered models, and decontextualized metrics may reinforce erasures, undue standardization, and epistemic inequalities. As a way forward, the article proposes a multimodal, ethical, and participatory lexicography of Libras, guided by the linguistic description of the signed language, the documentation of variation, the ethical governance of data, and the centrality of the Deaf community in the production, validation, and circulation of lexicographic resources. It concludes that the use of artificial intelligence in this field should be understood as situated sociotechnical support, rather than as an autonomous solution to the lexicographic challenges of Libras.