A THEORETICAL FRAMEWORK FOR COMPARING NEURAL MACHINE TRANSLATION AND LARGE LANGUAGE MODELS IN THE UZBEK-RUSSIAN LANGUAGE PAIR

Abstract

Institutions in Uzbekistan now choose between dedicated neural machine translation engines and general-purpose large language models for Uzbek-to-Russian work, and they choose without evidence: no published study measures the two families against each other on this pair. This article does not supply the measurement. It supplies the theoretical framework the measurement would require, and argues that the framework cannot be borrowed unchanged from the existing literature. Three arguments are developed. The comparative evidence accumulated in shared-task evaluation establishes a high-resource advantage and a low-resource deficit for general-purpose models, but the Uzbek-to-Russian configuration, a low-resource agglutinative source with a high-resource fusional target, is not among the configurations tested, and the direction of the combined effect is underdetermined. The two families of system are optimised over different objects and therefore fail in different ways, so a comparison expressed only as a quality score cannot distinguish them in the respect that matters to a post-editor. The typological distance between Uzbek and Russian, in suffix chaining, in the absence of grammatical gender and in constituent order, makes the pair a distinct theoretical case rather than another instance of a known one. The article closes with what evaluation theory permits to be claimed and with the design constraints that follow.