YouTube-SL-25: A Large-Scale, Open-Domain Multilingual Sign Language Parallel Corpus
This work addresses the data scarcity problem for Deaf/Hard of Hearing communities by providing a large-scale, multilingual sign language dataset, enabling more inclusive machine learning research.
The authors tackled the data bottleneck in sign language machine learning by creating YouTube-SL-25, a large-scale, open-domain multilingual corpus with over 3000 hours of videos across more than 25 sign languages, which is over 3 times the size of YouTube-ASL and the largest parallel sign language dataset to date. They demonstrated that multilingual transfer benefits both higher- and lower-resource sign languages within the corpus, as shown by baseline results on sign-to-text tasks using a T5-based model across 4 sign languages.
Even for better-studied sign languages like American Sign Language (ASL), data is the bottleneck for machine learning research. The situation is worse yet for the many other sign languages used by Deaf/Hard of Hearing communities around the world. In this paper, we present YouTube-SL-25, a large-scale, open-domain multilingual corpus of sign language videos with seemingly well-aligned captions drawn from YouTube. With >3000 hours of videos across >25 sign languages, YouTube-SL-25 is a) >3x the size of YouTube-ASL, b) the largest parallel sign language dataset to date, and c) the first or largest parallel dataset for many of its component languages. We provide baselines for sign-to-text tasks using a unified multilingual multitask model based on T5 and report scores on benchmarks across 4 sign languages. The results demonstrate that multilingual transfer benefits both higher- and lower-resource sign languages within YouTube-SL-25.