COMPARATIVE ANALYSIS OF UZBEK RUSSIAN MACHINE TRANSLATION MODELS

Abstract

Uzbek-Russian machine translation has moved from good enough for gist to a tool that increasingly shapes academic writing, administrative workflows, media consumption, and bilingual education in Uzbekistan and the broader post-Soviet space. Yet, for this language pair, the operational question is no longer whether neural MT works, but which model family works better under which linguistic pressures: agglutinative Uzbek morphology, flexible Russian word order, divergent evidential and modality markers, and the dense presence of culture-bound realia and proper names. This article compares six widely used or strategically important model lines for Uzbek↔Russian: Google Translate, Yandex Translate, Microsoft Translator, DeepL, Meta NLLB-200, and Meta M2M-100. The comparison combines documented language coverage and platform affordances, an error-oriented linguistic test set of thirty diagnostic items spanning official, academic, and literary registers, and a synthesis of Russian-language scholarship on MT typology, evaluation, and error analysis. The results show a stable split between convenience-optimized cloud systems and controllable open multilingual models: the former dominate in speed, integration, and user experience, while the latter provide stronger levers for domain adaptation, reproducibility, and privacy-preserving deployment. The paper formulates a forward-looking research agenda for Uzbek-Russian MT centered on balanced parallel corpora, script-robust normalization, hybrid evaluation protocols, and modular pipelines that couple translation with quality estimation and post-editing.