Abstract
The paper addresses the lack of bilingual annotated resources for automatic fake news detection in the KazakhRussian media space. We introduce KazFakeCorpus, a balanced bilingual corpus annotated using a multi-level semantic scheme that captures message reliability, fake content type, disinformation technique, communicative intent, modality, and source characteristics. The corpus was constructed from verified Gov.kz materials for the REAL class and synthetically transformed texts for the FAKE class. After preprocessing and balancing, the final dataset contains 4,276 texts in Kazakh and Russian. Annotation was performed in Label Studio by two independent experts, while a pilot phase of 120 texts was used to refine annotation categories and guidelines. Inter-annotator agreement measured with Krippendorff’s Alpha ranged from 0.79 to 0.88, indicating substantial consistency. Corpus analysis revealed the prevalence of misattribution, clickbait, and emotional pressure techniques. The proposed framework models fake news as a structured semantic phenomenon and supports research in disinformation analysis, explainable NLP, and cross-lingual fake news detection for low-resource languages.