Authors - Ussen Marassulov, Orken Mamyrbaev, Gulnur Kazbekova, Aigerim Yerimbetova, Madina Sambetbayeva, Duman Telman Abstract - This paper evaluates fake news detection in Kazakh and Russian media by comparing sparse TF-IDF baselines with multilingual transformer models. The task is formulated as binary text classification with Real (0) and Fake (1) labels. The protocol combines duplicate-aware cleaning, split-leakage verification, label-consistency checks, non-neural baselines, transformer finetuning, and bidirectional cross-lingual testing. After exact title + text deduplication, the corpus contained 38,013 Kazakh records, 37,181 Russian records, and 75,194 bilingual records. Five regimes were evaluated: KZ_only, RU_only, MIX_only, KZ->RU, and RU->KZ. TF-IDF remained highly competitive in in-domain testing, reaching Macro-F1 = 0.9979 on Kazakh, 0.9984 on Russian, and 0.9989 on mixed data. In cross-lingual testing, multilingual transformers showed clearer advantages: mBERT reached Macro-F1 = 0.9833 in KZ->RU, while XLM-R achieved 0.9947 in RU->KZ. The results indicate that lexical baselines are strong when train and test data share the same language setting, whereas multilingual transformers are more reliable when the target language changes.
