Fake News Detection in the Kazakh Language: A Review of Corpus Annotation, Multilingual Modeling, and Low-Resource Challenges

Abstract

This focused narrative review examines fake news detection for Kazakh as a low-resource natural language processing problem. It synthesizes work on corpus construction, binary and multi-level annotation, traditional classifiers, multilingual transformers, evidence-aware verification, and social and multimodal context. The available Kazakh evidence is promising but narrow: public labeled data remain scarce, reported results are difficult to compare, and very high scores from small single-split datasets may not generalize across sources, topics, or time. Kazakh morphology, code-switching, and cultural context also limit direct transfer through machine translation. We therefore argue for public benchmarks with source- and time-aware splits, stronger annotation quality reporting, retrieval-grounded and explainable models, and human oversight. Large language models may support claim extraction and fact-checking assistance, but require Kazakh evidence retrieval, calibration, and careful evaluation rather than unsupported generation.