AI-BASED DETECTION AND CORRECTION OF LINGUISTIC ERRORS IN MACHINE TRANSLATION SYSTEMS THEORETICAL FOUNDATIONS

Abstract

This article examines the theoretical foundations of detecting and correcting linguistic errors in AI-based machine translation (MT) systems, with a focus on morphologically rich and low-resource languages such as Uzbek. Although neural architectures have advanced, MT outputs still contain grammatical, lexical, semantic, syntactic, and cultural errors. The paper analyzes key error sources and modern correction strategies, including rule-based methods, neural models, Quality Estimation (QE), and automatic post-editing (APE). Special emphasis is placed on phraseological units—idioms, proverbs, and fixed expressions—which remain challenging due to their figurative and culture-dependent meanings. The study also discusses the limitations of BLEU- and METEOR-based evaluation and highlights the importance of hybrid human–AI workflows. Future research directions include corpus expansion, improved semantic modeling, and linguistically informed neural techniques.