Detecting hateful content reliably is essential for healthy online communities, yet performance varies widely across languages. This work addresses a practical question: can training data from one language support hate‑speech detection in another with comparable accuracy? To investigate, we conduct a large‑scale study covering 14 languages. We observe consistent transfer patterns: linguistic similarity emerges as the strongest predictor, while temporal overlap and dataset‑size differences play smaller roles. Although these measurable factors explain most of the trends, they do not fully account for all cases. A complementary qualitative analysis suggests that shared cultural cues and communicative norms may explain the remaining patterns, especially within certain language clusters; this hypothesis is based on inspection rather than causal measurement. The study also includes low‑resource languages, offering practical insights into when cross‑lingual transfer can aid them and where its limits lie. Overall, the findings inform the design of more reliable multilingual hate‑detection systems.
Review