ResFusionNet: A Residual Fusion Network for Multimodal Hate Speech Detection

Hate speech on social media has emerged as a significant challenge, driven by the increasing prevalence of multimodal content that often combines text, images, and other formats to convey harmful intent.This study introduces ResFusionNet, a novel residual fusion network tailored for multimodal hate speech detection.The proposed model integrates textual and visual features using state-ofthe-art representation techniques, particularly employing MPNet for text and EfficientNetV2 for images.By leveraging an early fusion approach augmented with residual blocks, ResFusionNet captures nuanced interdependencies between modalities, addressing the complexities posed by multimodal content.Experimental results on benchmark datasets, MMHS150K and Facebook Hateful Memes, demonstrate competitive performance to other methods in terms of accuracy, F1 scores.