Multilingual clickbait detection with BERT: An empirical evaluation of training strategies

Clickbait detection is an important component of automated content analysis systems used in large-scale web and social media platforms. While prior work has primarily focused on monolingual settings, practical deployments often require robust performance across multiple languages, including low-resource scenarios. This paper presents a controlled empirical study of transformer-based clickbait detection across nine typologically diverse languages. Using publicly available datasets normalized to a common binary label scheme, we evaluate three training strategies under a unified protocol: (1) monolingual fine-tuning with native BERT models, (2) cross-lingual transfer using with multilingual BERT models, and (3) joint multilingual training with multilingual BERT models. We complement these supervised settings with zero-shot and few-shot prompting using Gemini. Beyond the strategy ranking, we report three error-analysis findings that explain why differences arise. First, false negatives are consistently associated with longer headlines across all languages and training strategies, indicating a universal structural bottleneck in headline-only classification. Second, languages with very high inter-class semantic similarity (cosine >0.95 for Bengali and Urdu) impose a performance ceiling that no training strategy fully overcomes. Third, error clusters are language-specific, which helps explain the strong asymmetry observed in cross-lingual transfer. Together, these findings provide practical guidance for deploying clickbait detection in multilingual environments and demonstrate that error-aware evaluation reveals substantially more than accuracy comparisons alone.