Developing a Kazakh question-answering model: standing on the shoulders of multilingual giants

At present, there is a proliferation of AI-powered systems being developed to enable natural language interactions with humans. However, these systems are primarily designed for a limited set of approximately 40 dominant languages, while the majority of the world’s over 7,000 living languages, including Kazakh, are categorized as under-resourced languages that lack cutting-edge AI technologies like Siri, GPT-3, etc. The solution is to leverage state-of-the-art multilingual models, which have recently emerged as vital resources for lowresource languages, facilitating the rapid development of native language models with acceptable quality. This study focuses on the comprehensive evaluation of various multilingual models, including mT5, XLM-RoBERTa, and mDeBERTa with the aim to determine their suitability as donor models for Kazakh question answering. Furthermore, the study addresses the research question of whether large multilingual models genuinely outperform native language models in the task of question answering. Experimental results confirm this superiority and lead to the conclusion that there is no need to train native models from scratch, considering the associated economic and environmental costs.