Lexicographical corpus and computer vision

an AI-based methodology for large-scale semi-automatic collection and annotation of signs and the analysis of lexical variation (regionalisms) in Brazilian sign language

Abstract

The quantitative study of lexical variation (regionalisms) in Brazilian Sign Language (Libras) is methodologically hindered by the absence of large-scale, cherologically annotated lexicographical corpora. Traditional manual annotation methodologies, such as ELAN, are unfeasible for building massive corpora (tens of thousands of hours), and the praxis of conceptual glossing (translation) results in the loss of articulatory information (signifier), making the study of subtle cherological variants impossible. This article proposes an interdisciplinary methodological architecture that solves this dual bottleneck by articulating Corpus Linguistics, Computer Vision, and Computational Sociolinguistics. The objective is to detail a technically and ethically robust pipeline for large-scale semi-automatic collection and annotation, focused specifically on the discovery and analysis of sociolinguistic variation. The methodology begins with a systematic review (PRISMA) mapping the state-of-the-art (2018-2025), identifying the central gap: AI focuses almost exclusively on recognition (SLR) and translation (SLT), treating linguistic variation as noise rather than an object of study. The proposed architecture employs a two-phase pipeline. Phase 1 (Feature Extraction) uses pose estimation (optimized MediaPipe), following the optimization by dos Santos et al. (2025) , to convert videos (collected "in-the-wild") into vector representations (time series of landmarks), replacing glossing with a quantifiable "cherological transcription". Phase 2 (Semi-Automatic Annotation) utilizes Transformer models, trained on a seed Corpus, to generate lexical label suggestions. These suggestions are submitted to a human-in-the-loop validation interface, where Deaf linguists validate or correct the annotations, with the model being iteratively retrained. As a result, the methodology yields a massive relational database that links articulatory forms (vectors) to sociolinguistic metadata. This Corpus enables computational sociolinguistic analysis via unsupervised clustering techniques, allowing for pattern discovery and the identification of regional variants which are subsequently correlated with geographical data to map variation, enabling the creation of quantitative linguistic atlases for Libras. The conclusion is that this architecture overcomes the bottlenecks of traditional lexicography and offers a viable path for documenting variation. The article critically discusses the ethical foundations, positioning "Deaf-led Research" as a central methodological pillar to mitigate algorithmic biases (such as techno-ableism) and ensure the technology functions as a tool for documentation and empowerment, not for replacement or the erosion of linguistic rights.