CLTL at EXIST 2025: Identifying Sexist Memes Using an Ensemble of Shallow and Transformer Models

We present the CLTL system developed for identifying and categorizing sexist memes (Task 2) in English at the EXIST 2025 Shared Task at CLEF 2025. The task consisted of three subtasks: (Task 2.1) sexism identification, where memes were classified as sexist or not-sexist; (Task 2.2) source intention classification, where sexist memes were further classified as either direct or judgemental; and (Task 2.3) sexism categorization, where memes were classified into one or more overlapping fine-grained classes: ideological and inequality, stereotyping and dominance, objectification, sexual violence, and misogyny and non-sexual violence. Our submissions were based on a hard majority voting ensemble strategy, where the component models included a multimodal model that combined the representations of Swin Transformer V2 and a pre-trained language model (RoBERTa or BERT), and the text-only models that used meme text and image captions as input. The text-only approaches included pre-trained transformer models (RoBERTa, BERT, and a BERTweet model fine-tuned for sexism detection) and a conventional machine learning approach, namely an SVM with stylometric and emotion-based features. Our experiments demonstrated that an ensemble that incorporates deep learning and conventional machine learning approaches is efficient for the sexist meme detection task. Our best runs with an ICM-Hard score of 0.2850 for Task 2.1, -0.0645 for Task 2.2, and -0.4214 for Task 2.3, were ranked 6th out of 18 runs, 4th out of 15 runs, and 5th out of 14 runs, on the English leaderboard, respectively.