The proliferation of abusive content on social media across linguistically diverse regions presents a formidable challenge for automated content moderation systems. Multilingual Language Models (MLLMs), particularly transformer-based architectures, offer scalable solutions to this problem. However, evaluating their effectiveness across varied languages, dialects, and social platforms remains underexplored. This study investigates the performance, generalizability, and limitations of state-of-the-art MLLMs, including BERT, XLM-RoBERTa, mBERT, and their hybrid variants, for abusive content detection across multilingual and code-mixed datasets. We analyze over ten empirical studies covering between 3 and 13 languages, focusing on Indic and European languages, including Romanized and script-mixed forms. Our review highlights key architectural strategies such as transfer learning, lexicon integration, and ensemble modeling. Results indicate that models enhanced with social context or emoji embeddings outperform baseline transformers, achieving macro F1-scores as high as 0.91. We further assess platform-specific challenges and the scalability of these models in low-resource settings. This comparative study provides insights into model adaptability and effectiveness in real-world moderation systems and suggests directions for future research in multilingual content moderation.
Multilingual Language Models, Social Media Platforms, Abusive Content Detection, Transfer Learning, Ensemble Modeling
© IACyC is conducted under the administration of the CyberMACS Erasmus-Mundus Masters Programme. cybermacs.org