TRANSLATING THE MORPHOLOGICAL SYSTEM OF THE KAZAKH LANGUAGE USING ARTIFICIAL INTELLIGENCE TOOLS: EXPERIENCE AND TESTING

Abstract

The rapid proliferation of Large Language Models (LLMs) and Neural Machine Translation (NMT) architectures has generated considerable optimism regarding the automation of linguistic workflows. This optimism demands rigorous empirical scrutiny when applied to morphologically rich, agglutinative systems such as Kazakh. This study evaluates how ChatGPT (GPT-4), Google Translate and DeepL process the morphological density and stylistic nuances of literary Kazakh language while identifying the structural etiologies of their translation artifacts. The evaluation corpus comprises four textual excerpts systematically sampled from the canonical prose of Mukhtar Magauin and Sherkhan Murtaza, selected for their elevated literary registers, high morpheme-to-word ratios and culture-specific lexical depth. Each passage was ingested via zero-shot prompting to isolate baseline capabilities, and the resulting outputs were analyzed using a four-tiered philological error taxonomy encompassing morphological, semantic, syntactic and pragmatic-stylistic dimensions. The empirical findings establish a distinct hierarchy of architectural competence: ChatGPT demonstrates superior contextual modeling, exhibiting sensitivity to literary registers, long-distance syntactic dependencies and language-specific rhetorical devices such as vowel elongation and verbal reduplication; DeepL occupies an intermediate position, generating syntactically permissible output that consistently erodes cultural and idiomatic depth; Google Translate performs deficiently, manifesting critical semantic reversals, structural disruptions and the erroneous segmentation of transliterated proper nouns. The theoretical significance indicates that translation errors systematically cluster around specific typological features of Turkic languages that remain underserved by mainstream NLP architectures optimized for Indo-European structures. Practical significance deficits pose a concrete sociolinguistic risk to regional digitization initiatives, underscoring an immediate requirement for human-in-the-loop engineering. The article concludes that active and structural participation from trained philologists is indispensable to refine tokenization strategies, annotate specialized corpora and implement high-fidelity alignment protocols within automated translation systems.