MVLP: Multi-perspective Vision-Language Pre-training Model for Ethically Aligned Meme Detection

Ethically aligned design aims to ensure that intelligent systems remain human-centric, serving humanity’s values and ethical principles. Moderating content for achieving ethically aligned design is a critical step toward the realization of trustworthy AI. However, current content moderation methods focus on the detection of harmful content and ignore ethical considerations. Likewise, multi-modal meme detection tasks lack designs for more complex ethical discriminations. Memes pose an interesting multi-modal fusion problem, i.e., their understanding requires a very specific combination of information from different modalities (the text and the image), and have become an influential medium for conveying ideas, socio-cultural, and political views. Therefore, ethically aligned meme detection is crucial to investigate modalities and ensure that they are aligned with our social values and goals. In this paper, we propose MVLP, a Multi-perspective Vision-Language Pre-training model for ethically aligned meme detection.