A Survey on Multi-Modal Hate Speech Detection

Hate speech is a speech that expresses an attack on a person based on their ethnicity, race, sexual identity, belief, age, or unfitness. The integral complexity of this activity makes it necessary to differentiate hate speech from other types of social molestation. Online content commonly considers two kinds of data, i.e., text and images. Memes can also be used to share views and ideas through social media. Primarily memes are generated only for amusement, while some memes could be hate-related, generally a collection of text and images. Identification of offensive memes can help to identify and mitigate harmful effects on social media. This task is challenging because the image and text in such memes might be irrelevant. To make this model, one should deeply understand the images and texts. Almost all the models are based on past neural networks, and someworks are also related to identifying deceptive content for image and text data. Evenson, present-day techniques are competitive but not reaching optimum performance due to a significant gap in connecting images and text data. With a few papers, a model that detects hate speech from text and images is considered a vision language task. Most authors are motivated by Visual Questions answering the problem that identifies answers with image and text inputs.