UNED-Annotate-Team at EXIST 2026: A Multimodal Multi-Task Approach to Sexism Identification in Memes

The harmful effects of sexism on social media change over time and can affect different groups of people in different ways. This is especially true for memes, where meaning is often conveyed through a combination of text and images, making sexist content harder to identify. As a result, people may interpret the same meme differently and disagree on whether it is harmful or sexist, a challenge that the Learning with Disagreement (LeWiDi) paradigm addresses precisely by preserving the variability in those annotations rather than forcing a single consensus label.

In this paper, we present the work of the UNED-Annotate-Team for Task 2.1 of the EXIST (sEXism Identification in Social neTworks) lab at CLEF 2026, which focuses on spotting sexism in Spanish and English memes. We propose a multimodal approach that integrates visual features with annotator demographic information (gender and age), exploring two distinct strategies: auxiliary multitask learning and contextual soft-token embeddings. We further enrich the textual representation with CLIP-derived visual semantic features and adopt a disagreement-aware annotation aggregation strategy aligned with the LeWiDi paradigm.

The results reveal the importance of prioritizing the proper generation of the visual semantic representation. We leveraged an efficient CLIP-based preprocessing strategy combined with early fusion, which significantly improved our detection of the minority (sexist) class. In the Soft assessment environment, we compared multitask learning with a novel soft token embedding strategy, demonstrating that integrating sociodemographic information as contextual tokens—rather than discrete auxiliary tasks—produces better alignment with the annotator’s subjectivity, achieving an F1 score of 72.36%.