Deep Learning Methods of Dynamic Sign Language Translation: A Rapid Umbrella Review

Countries globally are recognizing sign languages, which are used by more than 5% of national and global populations as official languages. In Africa, South Africa was the fourth country to officialize sign language in 2023. This trend combined with advances in computer vision and machine learning has resulted in an increase of scholarly literature on the translation between spoken and signed languages to bridge the language and communication divide. This has been accompanied by a high volume of published literature reviews on the subject, most of them distinguished by time and geographical contexts. Despite numerous published review papers on the subject, to date, no formal or umbrella reviews of these reviews have been performed to consolidate the similarities, contrasts, gaps, and opportunities reported in these reviews. In this paper, we conduct a first-of-its-kind, rapid systematic umbrella review of 11 literature reviews about machine translation and recognition of dynamic sign languages using machine learning. Three lenses namely, the evolution of machine learning and computer vision technologies and resulting applications to sign language translation; geographical contexts and associated applications; and the unique linguistic context of sign languages are used. We found that most reported literature has solved the sign language translation and recognition problem from a purely data-driven perspective, without due attention to specific linguistic complexities of the languages. Consequently, we found that most literature has not paid attention to proper corpora design, model design, and meaningful performance evaluation besides the generic and inherited measures of performance from spoken language processing and computer vision literature.