Transfer Learning for Multilingual Abusive Meme Detection
Mithun Das, Animesh Mukherjee · 2023
The exponential growth of social media platforms has permitted people to connect worldwide. However, it has also fueled the elevation of several harmful and abusive content on the Internet. Repeated exposure to abusive content may lead to psychological effects on the target users. Thus it is necessary to detect such abusive content in all forms to keep these platforms safe and healthy. So far, several works have been done for abusive speech detection; however, most of these are text-based. Yet, social media contents are often multimodal, comprising text, images, videos, etc. Internet memes have recently emerged as a predominant mode of content shared on social media and are used to express vitriol or harm toward others. Hence it is essential to detect such abusive memes. Although several works have been done for abusive/harmful meme detection, most of these are in English with only a very few extending to non-English datasets. Therefore, one of the immediate solutions is to detect abusive memes in one language and transfer them to other languages. This work explores several model transfer techniques to bridge the gap by creating various baseline models.