Buryat historical sources: digital infrastructure of machine translation

The study is dedicated to a vast yet still underexplored corpus of Buryat historical sources in the Old Written Mongolian language, preserved in academic and archival institutions in Russia. The introduction of these documents into scientific and public circulation, reflecting all aspects of Buryat society from the 18th to the early 20th centuries, requires the application of modern digital methods. The most promising approach to addressing this task is a comprehensive strategy employing digital humanities tools, which includes digitizing sources, full-text input of texts using romanized transliteration, creating a text Mongolian-Russian corpus with metatextual, structural, and morphological markup according to TEI standards, translating sources, and forming a balanced parallel Mongolian-Russian corpus. Special attention was paid to the genre, chronological, and territorial representativeness of the corpus data. The prepared digital materials serve as a foundation for applying machine learning methods to tackle tasks in optical text recognition for Mongolian script and machine translation. To assess the prospects of machine translation for Buryat historical sources, computational experiments were conducted with transformer models trained "from scratch" and large pre-trained multilingual models (mBART-50, mT5). The scientific novelty of the work is defined by the creation of the first methodologically grounded digital platform for studying Buryat historical sources, adapted to the characteristics of the local rendition of Old Written Mongolian. Important competencies have been gained for organizing the complete cycle of digital processing of historical sources from the digitization of archival documents to the formation of a balanced parallel corpus. For the first time, an online corpus of unique historical texts has been assembled and published, equipped with analytical tools, as well as specialized datasets for training AI models. It has been established that for low-resource language pairs, the most effective strategy is fine-tuning pre-trained multilingual models rather than modifying neural network architectures. The research lays the groundwork for creating comprehensive tools for digitizing Buryat written heritage, which will open new perspectives for historical, linguistic, and cultural studies.