| Resumen |
This paper addresses the EXIST 2025 lab with a comprehensive approach to classifying sexism across diverse digital media formats, including tweets, memes, and TikTok videos. The study explores tailored methodologies for each modality, reflecting their distinct characteristics. For Task 1 (binary sexism identification), we compare two approaches: the transformer-based HateBERT model and the generative Claude 3.7 model. HateBERT, optimized through tweet preprocessing, regularized training, and multitask learning, demonstrates robustness in textual analysis. In contrast, Claude 3.7 leverages advanced multimodal capabilities, integrating visual and textual cues for flexible and effective content interpretation. For Tasks 2 and 3—focused on classifying the intent behind sexist content and identifying its specific type—Claude 3.7 is used exclusively. It effectively incorporates multimodal inputs, including visual frames from memes and videos, enabling nuanced distinctions such as direct sexist expressions versus judgmental critiques. Our results reveal substantial performance differences across tasks and modalities, with Claude 3.7 achieving first place in Task 2 in 8 out of the 9 evaluation metrics reported by the organizers. © 2024 Copyright for this paper by its authors. |
