Development and Evaluation of a Small Kazakh Language Corpus to Improve the Efficiency of Multilingual NLP Systems in Low-Resource Environments

Arailym Tleubayeva, Sultan Aubakirov, Aisultan Tabuldin, Adai Shomanov

2025

Closed access Metadata (CSL-JSON)
DOI 10.1109/sist61657.2025.11139363
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

This study tackles NLP challenges in low-resource settings by developing the Small Kazakh Language Corpus—a high-quality, annotated collection of Kazakh texts sourced from news, scientific publications, and Wikipedia. The corpus was used to fine-tune two models for masked language modeling: the multilingual XLM-RoBERTa-base and the Kazakh-specific nur-dev/roberta-kaz-large. Fine-tuning notably improved XLM-RoBERTa-base's accuracy from 48.84% to 68.85% and F1-score from 41.96% to 68.62%, while nur-dev/roberta-kaz-large showed more modest gains. These findings demonstrate the critical role of language-specific resources in enhancing multilingual NLP systems and provide a solid foundation for further research in applications such as machine translation, sentiment analysis, and question answering.