Parallel Corpora and Neural Models for the Uzbek Language: Current State and Development Paths

This article examines the role of parallel corpora in developing neural machine translation for the Uzbek language. The concept of a parallel corpus is analysed, existing resources for Uzbek are reviewed, and methodological principles of corpus construction are formulated: representativeness, alignment quality, metadata annotation, and licensing. The article analyses how corpus characteristics affect neural model behaviour — error types observed in English–Uzbek MT are correlated with corpus gaps. Development paths are proposed: national corpus programs, domain-specific corpora, community-driven collection, and post-editing data feedback loops. The conclusion argues that corpus construction is the single most effective investment for improving Uzbek MT quality, and that a coordinated national effort, combining academic, governmental, and community resources, is needed to move Uzbek from low-resource to medium-resource status.