Reproduce “YouTube-SL-25: A Large-Scale, Open-Domain Multilingual Sign Language Parallel Corpus” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Garrett Tanzer, Biao Zhang
Researchers at Google and Google DeepMind introduce YouTube-SL-25, the largest open-domain multilingual parallel corpus of sign language videos with aligned captions, comprising over 3,200 hours across more than 25 sign languages. The dataset and associated baseline models advance sign-to-text translation and sign language identification, demonstrating performance improvements through multilingual pretraining and transfer learning.
Key insights
- A scalable corpus construction methodology, relying on coarse manual triaging rather than labor-intensive per-video native signer annotation, can efficiently build large, multilingual sign language datasets.
- Multilingual pretraining on a diverse corpus like YouTube-SL-25 provides substantial benefits for both high-resource and low-resource sign languages, improving translation and identification performance.
- The open-domain and diverse nature of the corpus fosters the development of more generalizable sign language AI models, moving beyond specialized or closed-domain datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.