Abstract
Kazakh sign language is the focus of attention among scientists. In recent decades, increasing attention has been paid to the creation of technologies capable of removing barriers to communication and interaction between people with disabilities and society. The lack of a parallel corpus in sign language opens up the topic of text translation into gloss for urgent and in-depth research. Unfortunately, digital technologies are not accessible to people with hearing and speech impairments. Text-to-gloss translation is still relevant for Kazakh sign language. In particular, because one sign has several meanings, synthesizing it into gloss in a software environment takes a long time. This article discusses the translation of text into gloss using five different models based on a multilingual transformer. That is, it examines whether the models are faster or slower to read and whether they consume more or less resources. The mBART-large-50, M2M100-418M, mT5-base, MarianMT, and T5-base were selected for the experiment. All of these models were trained using the same parameters. The results were evaluated using widely used standard metrics (BLEU, ROUGE-L, and chrF++), and a qualitative analysis of the models was also performed. The study found that larger multilingual models provide more accurate translations. Although the mBART-large-50 and M2M100-418M models performed well, the other smaller multilingual models were faster and less resource-intensive. Translating text into a glossary is still relevant for Kazakh sign language. In particular, because one sign language has several meanings, synthesizing it into a glossary in a programming environment takes a significant amount of time.