Traducción automática de lengua de señas en español utilizando Deep Learning

Pedro Dal Bianco

PhD thesis, Facultad de Informática, Universidad Nacional de La Plata

Full title: Traducción automática de lengua de señas en español utilizando Deep Learning - Estrategias en contextos de pocos datos (in English: Automatic sign language translation in Spanish using Deep Learning: strategies for low-data settings). Advisors: Franco Ronchetti and Facundo Quiroga. Defended on 30 March 2026.

Summary

The thesis translates Argentinian Sign Language (LSA) into written Spanish when little data is available. It follows two lines:

• Resources: the LSA-T corpus, over 20 hours of video from the CN Sordos channel with Spanish text and pose keypoints, and Seni.ar, a web app to validate it.
• Models: a compact, gloss-free, pose-based Transformer, extended in two ways against data scarcity: multilingual training on a unified set of sign language corpora, and paraphrases of the target sentences generated with large language models.

An interpretability study shows that the encoder attends to clusters of frames that match single signs, and that during decoding the model shifts from the visual input to the linguistic context.

Citation

@phdthesis{dalbianco2026thesis,
  title={Traducci{\'o}n autom{\'a}tica de lengua de se{\~n}as en espa{\~n}ol utilizando Deep Learning},
  author={Dal Bianco, Pedro Alejandro},
  school={Facultad de Inform{\'a}tica, Universidad Nacional de La Plata},
  year={2026},
  doi={10.35537/10915/196569}
}