Abstract.Currently, one of the pressing issues in the Republic of Kazakhstan is the accurate translation of legal texts from Kazakh into Russian and English, as well as from these languages into Kazakh. This scientific work analyzes translation errors using widely known machine translation systems such as Yandex and Google, based on legal texts sourced from specialized legal databases in the Kazakh–Russian language pair. The main goal of the study is to achieve precise and semantically accurate translation of sentences and terminology specific to the legal field. To this end, a corpus of 96,555 sentences and phrases was compiled using a specialized program, collecting data from legal documents, court decisions, and official websites. This corpus was used to train the MarianMT neural machine translation system within the Kazakh–Russian language pair. To further improve translation quality, the KazRobert transformer model was applied. The study provides a comprehensive explanation of the KazRobert model’s architecture and its mathematical foundations. Translation quality was evaluated using internationally recognized metrics such as BLEU, TER, and METEOR. The study presents a comparative analysis of two outcomes: the initial results from the MarianMT model alone, and the improved results from the same model fine-tuned with KazRobert. The findings indicate that the proposed approach outperforms previous models, including the OpenNMT-based system developed by the same authors.
The experiments demonstrated that increasing the corpus size and the number of legal terms positively impacts translation quality. Furthermore, the research suggests that this method can be effectively adapted for other Turkic languages that share structural similarities with Kazakh.

Keywords: neural machine translation, MarianMT machine translation, KazRobert model, transformer model, legal domain corpus, BLEU translation metric, TER translation metric, METEOR translation metric.