Ensemble Approach for Religious Hate Speech Detection in Text-Based Images on Facebook

Hate speech detection in social media posts in religious affiliation is a challenging problem when it comes to text-based image contents. This study addresses this issue by utilizing a dataset consists of English text-based Facebook image posts and categorizing them on three categories: Catholic, Islam, and Hindu. Two models, TF-BERT with InceptionResNetV2 and TF-DistilBERT with DenseNet121 were implemented and enhanced using ensemble techniques, to improve accuracy and performance. Significant improvements are shown in results by ensemble models, outperforming standalone models. Ensemble models for the Catholic and Islam categories achieved accuracies of 0.8912 and 0.8351, with AUC-ROC scores of 0.9565 and 0.9311, respectively. Meanwhile, the Hindu category demonstrated the highest performance, achieving an accuracy of 0.9462 and an AUC-ROC of 0.9762. The proposed models surpassed the performance of previous studies that did not detect hate speech in religious affiliations through categorization. This study can be further extended to include additional religious affiliations.