I2C-UHU at EXIST 2024: Transformer-Based Detection of Sexism and Source Intention in Memes Using a Learning with Disagreement Approach

Abstract

In this paper, the I2C-UHU Group addresses the Exist-2024 challenges of Sexism Identification and Source Intention in Memes. We developed an ensemble of classifiers based on Transformer technology and adopted a Learning with Disagreement (LeWiDi) approach to analyze data from multiple annotators’ perspectives. Techniques for constructing datasets and optimizing hyperparameters were explored, enhancing model performance through varied combinations. The optimal models were refined by weighting according to prediction accuracy. Our submissions for Task 4 achieved ranks of 4th with ICM-Hard and ICM-Soft scores of 0.5668 and 0.4476, respectively. For Task 5, we secured 2nd and 10th places with ICM-Hard and ICM-Soft scores of 0.4119 and 0.2023, respectively.