Detecting implicit sexism in digital social networks via contrastive learning-based adaptive network

Abstract

The pervasive reach of social media has accelerated the spread of hateful content, including sexism. Sexism is a subtle and context-dependent form of bias often missed by conventional detection systems. We propose ASCEND (Adaptive Supervised Contrastive lEarning for implicit sexism detectioN), a novel framework for implicit sexism detection through a threshold-based supervised contrastive learning mechanism. ASCEND employs a learnable similarity threshold to identify positive pairs. This ensures that only semantically aligned samples with cosine similarity above this threshold are pulled together in the embedding space. This adaptive filtering controls false positives and false negatives. To further enhance representation learning, ASCEND integrates word-level attention to focus on contextually salient terms. Also, it augments textual embeddings with sentiment, emotion, and toxicity features for richer semantic understanding. The model jointly optimizes contrastive and cross-entropy losses, enabling robust fine-grained classification. Extensive evaluations on EXIST2021 and MLSC show ASCEND outperforming competitive baselines and large language models by up to 36.14% macro-F1, with consistent gains across tasks. Ablation studies confirm the complementary roles of thresholding, attention, and auxiliary features. These results demonstrate ASCEND’s effectiveness in capturing the linguistic signals of implicit sexism and its adaptability to diverse social media contexts. © 2026 Elsevier Ltd.