Traditional and context-specific spam detection in low resource settings

Social media data has a mix of high and low-quality content. One form commonly studied content is spam. Most studies assume that spam context-neutral. We show on different Twitter sets context-specific exists identifiable. then compare multiple traditional machine learning models neural network model uses pre-trained BERT language to capture contextual features for identifying spam, both context-specific, using only content-based features. The outperforms the with an F1 score 0.91. Because training are notoriously imbalanced, we also investigate impact this imbalance simple Bag-of-Words best extreme imbalance, but fine-tunes from other domains significantly improves score, not levels domain-specific models. This suggests strategy employed may vary depending upon level in set, amount available low resource setting, prevalence vs. Finally, make our use by research community.