Ablations Studies in Pose-Based Sign Language Translation

Abstract

Sign Language Translation (SLT) task is rapidly evolving from isolated recognition approachesto a complex continuous gloss-free translation. Recent models leverage multimodallearning and large-scale language models to directly translate visual input into text. However,gloss-free SLT still faces persistent problems such as signer variability and mismatch with spokenlanguage. This work investigates how pose-based pre-processing techniques can improvethe performance of gloss-free SLT in a T5-based transformer framework.