Generative Models vs Discriminative Models: Which Performs Better in Detecting Cyberbullying in Memes?

Accessibility to the internet has led to a massive increase in the usage of social networking and online communication apps over the past decade. With so much content available online, these platforms suffer from governance and monitoring problems making the users susceptible to online cyberbullying and trolling. Recently, this trolling and hate comes majorly from memes that combine text and image modalities. Many studies show how cyberbullying can harm the mental well-being of the affected individuals. Previous studies also tried to study the role of sentiment, emotions, and sarcasm in identifying hateful memes setting up the meme detection task as a multi-modal, multitask problem. In the past, meme detection models were discriminative, but more recently generative models have been used to solve non-generation tasks such as aspect-based sentiment analysis and span detection. Motivated by this, in this work, we propose a unified Multimodal Generative framework, MGex by reframing the multitasking problem of detecting cyberbullying, sentiment, emotion, and sarcasm as a multimodal text-to-text generation problem. For this purpose, we use MultiBully dataset which provides annotation for all these labels. Here, we evaluate and contrast our proposed generative framework with several multitasking baselines and state-of-the-art models. Disclaimer: The article contains offensive text and profanity. This is owing to the nature of the work and does not reflect any opinion or stand of the authors.