Morphological Complexity and Neural Machine Translation Quality in Kyrgyz: a Genre-Stratified Evaluation Using chrF
Year: 2026
Type: conference-paper
Abstract: Kyrgyz is an agglutinative language, within the Turkic family. Its word content is formed in a process of adding various suffixes to a single root in a chain, each suffix corresponding to case, number, tense, mood, or possession [1]. This is useful because speakers can express far more in a single word, but it makes machine translation (MT) relatively hard as a single Kyrgyz word is equivalent to a phrase in English. Even misinterpreting one suffix can result in complete semantic distortion. Nevertheless, with the latest volume of literature on Turkic MT, the influence of Kyrgyz inflectional morphology on the quality of translation remains relatively under researched. The current study aims to address this by taking MT quality in literary and analytical text as a source of statistical analysis for the context of morphological complexity. This paper investigates this gap by asking: (1) What is the extent of morphological complexity in relation to MT quality in Kyrgyz? (2) To what extent do the different types of text (literary-analytical) affect accuracy of translation? (3) What error types occur due to insufficient morphological processing? Contributions. The study offers three major contributions: (1) the first systematic research of morphological effects on Kyrgyz MT quality; (2) evaluation based on genre stratification with a $29 \% \mathrm{chrF}$ score degradation in complex texts; (3) specific errors determined by insufficient morphological processing.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7167258701