A Contrastive Learning Based Approach to Detect Sexism in Memes

The widespread use of social media has created a unique challenge in detecting and mitigating sexism in online content. In this paper, we present our approach for detecting sexism in memes, developed for Task 4 of the EXIST 2024 challenge. The task was based on binary classification problem to detect whether or not a meme is sexist, within the context of a learning with disagreement paradigm. In our approach, We used ResNet50 and m-BERT models finetuned on EXIST 2024 dataset to get image and text embeddings. These embeddings, along with the annotators' data, were subsequently used to train a model using contrastive learning. The results on the test data demonstrate the effectiveness of contrastive learning techniques in addressing multimodal tasks.