BioSentinel at EXIST 2026: Soft-Label Optimization with XLM-RoBERTa for Sexism Intent Classification in Memes
2026-07-27 • Computation and Language
Computation and Language
AI summaryⓘ
The authors took part in a challenge to classify memes based on their communicative intent, such as whether they are direct, judgemental, or not sexist. They used a text-focused model called xlm-roberta-base and trained it to predict both hard labels and soft probability distributions due to differing opinions in the data. Their system performed moderately well, ranking about mid-range among many submissions. They found that combining two types of loss functions helped improve predictions, and they studied how annotator disagreement influenced model design in subjective tasks. They also experimented with model settings to better understand performance.
memescommunicative intentLearning with Disagreement (Le-Wi-Di)xlm-roberta-baseKL divergencecross-entropy losssoft labelshard labelsannotator disagreementsubjective NLP tasks
Authors
Chandru Munisamy, Karthikeya Raguveer, Alapan Kuila
Abstract
This paper describes the BioSentinel team's participation in EXIST 2026 Task 2.2: Source Intention in Memes, part of the CLEF 2026 evaluation campaign. The task requires classifying the communicative intent behind memes as direct, judgemental, or no (non-sexist), under a Learning with Disagreement (Le-Wi-Di) paradigm that mandates both hard-label and soft-label (probability distribution) predictions. We present a text-centric approach built on xlm-roberta-base (270M parameters) trained with a composite loss function combining KL divergence on soft annotator distributions and weighted cross-entropy on hard labels. On the official test set, the system achieved an ICM-Soft-Norm of 0.3229 and ICM-Norm of 0.3778, with a hard F1-score of 0.4236, ranking 40th (out of 118 submissions) in the soft-soft evaluation and 49th (out of 187 submissions) in the hard-hard evaluation. We provide an analysis of the dataset characteristics, exploratory larger-architecture runs, and the role of annotator disagreement in shaping model design for subjective NLP tasks. Ablation results show that KL loss improves soft-label metrics, while CE loss improves hard-label accuracy. We also report a separate validation-set temperature analysis.