I2C-UHU-Altair at EXIST2025: Multimodal Sexism Detection and Classification Using Advanced Vision-Language Models BLIP2 and Qwen, Large Language Models, and Learning with Disagreement Frameworks

Abstract

In this paper, we present the contributions of the I2C-UHU-Altair team to the EXIST2025 Lab at CLEF 2025, addressing the identification and classification of sexism in multimodal online content, particularly memes. Our system leverages recent advances in large language models (LLMs) and vision-language models (VLMs) to process both textual and visual information in a unified manner. We tackle three subtasks: binary classification of memes as sexist or non-sexist, classification of the author’s intent behind the meme, and multi-label categorization of sexist content. To enhance model robustness, we adopt the Learning with Disagreement framework, allowing the system to benefit from divergent annotations that reflect the inherent ambiguity and subjectivity in sociolinguistic tasks. We detail our multimodal architecture, preprocessing pipeline, and fine-tuning strategy. Our system demonstrated competitive performance in the shared task, achieving notable positions across all subtasks. Specifically, we ranked 5th in Subtask 2.1 (Soft-Soft), 3rd in Subtask 2.2 (Soft-Soft), and 3rd and 6th in Subtask 2.3 (Soft-Soft and Hard-Hard, respectively). Our findings highlight the potential of multimodal learning for detecting nuanced expressions of sexism in online environments and open avenues for future research in social media moderation and fairness-aware NLP.