Multimodal Deep Learning with Discriminant Descriptors for Offensive Memes Detection

A meme is a visual representation that illustrates a thought or concept. Memes are spreading steadily among people in this era of rapidly expanding social media platforms, and they are becoming increasingly popular forms of expression. In the domain of meme and emotion analysis, the detection of offensives is a crucial task. However, it can be difficult to identify and comprehend the underlying emotion of a meme because its content is multimodal. Additionally, there is a lack of memes datasets that address how offensive a meme is, and the existing ones in this context have a bias towards the dominant labels or categories, leading to an imbalanced training set. In this article, we present a descriptive balanced dataset to help detect the offensive nature of the meme's content using a proposed multimodal deep learning model. Two deep semantic models, baseline BERT and hateXplain-BERT, are systematically combined with several deep ResNet architectures to estimate the severity of the offensive memes. This process is based on the Meme-Merge collection that we construct from two publicly available datasets. The experimental results demonstrate the model's effectiveness in classifying offensive memes, achieving F1 scores of 0.7315 and 0.7140 for the baseline datasets and Meme-Merge, respectively. The proposed multimodal deep learning approach also outperformed the baseline model in three meme tasks: metaphor understanding, sentiment understanding, and intention detection.