Enhancing Post-Editing of Kazakh Translations Using Fine-Tuned Large Language Models

Akbayan Bekarystankyzy, Diana Rakhimova, Aliya Zhiger, Assel Sakatay, Nazym Zhumakhan, Aigerim Yerimbetova, Дина Оралбекова, Mussa Turdalyuly

2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.3390/a19030199
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Machine translation for low-resource languages such as Kazakh remains a complex task due to the scarcity of training data, intricate morphological structures, and culturally specific linguistic characteristics. This study presents the first extensive exploration of fine-tuning large language models for automated post-editing of Kazakh translations. We introduce KazPE, a carefully curated and annotated dataset that includes 10,008 training sentences and 311 test sentences spanning six domains: the medical, scientific, journalistic, oral, fiction, and legal. The dataset features detailed error classifications across 9 linguistic categories. Our method fine-tunes GPT-4.1 mini using supervised learning to enhance translation quality by systematically correcting targeted errors. According to human evaluations, conducted on a continuous 0–1 scale, the fine-tuned model achieves an average quality score of 0.84, surpassing the baseline score of 0.80, corresponding to a 5% relative improvement. The greatest improvements are observed in handling morphological and lexical errors, as well as in domain-specific texts—particularly in legal (+17%) and medical (+22%) domains. In addition, the translations were evaluated using the automatic metrics: BLEU, TER and METEOR. The fine-tuned model shows improvements across all automatic metrics (BLEU, TER, METEOR), which confirms better n-gram overlap with reference texts, fewer edits needed, and enhanced lexical and semantic alignment with the reference texts. Comprehensive error analysis shows that the fine-tuning process effectively mitigates challenges related to Kazakh’s agglutinative morphology and specialized terminology, while preserving accuracy on already correct sentences. This research establishes the first structured evaluation framework for Kazakh translation post-editing and offers valuable guidance for enhancing machine translation in morphologically rich, low-resource languages. To facilitate further progress in Turkic language processing, we publicly release the KazPE dataset, trained models, and evaluation framework.