I2C-UHU at EXIST2024: Learning from Divergence and Perspectivism for Sexism Identification and Source Intent Classification

Abstract

In this paper, we present the contributions of the I2C-UHU team to the EXIST2024 Lab at CLEF 2024, focusing on the identification of sexism and the classification of source intent in social media texts. State-of-the-art transformer models are employed to address the complex and nuanced nature of sexist language. We adopt a two-fold approach: firstly, classifying tweets as sexist or non-sexist, and secondly, categorizing sexist tweets based on intent. Our innovative approach, employing Learning with Disagreement, incorporates diverse perspectives from multiple annotators, enhancing the robustness and accuracy of our models. We detail our data preprocessing, augmentation techniques, and hyperparameter optimization strategies. Our results in the competition demonstrated effectiveness, with our entries achieving positive rankings in the two tasks in which we participated. In Task 1, we secured the 10th position out of 70 participants on the hard labels leaderboard and the 13th position out of 40 for soft labels. In Task 2, we achieved the 11th position out of 46 participants for hard labels and the 17th position out of 35 in the best run for soft labels. Our findings provide a foundation for future research and practical applications in social media moderation and policy-making.