Online abuse targeting women is a pervasive challenge across digital platforms, directly impacting mental health, limiting freedom of expression, and creating hostile environments. While early computational approaches treated harmful content identification as a broad, flat binary classification task (e.g., separating hate speech from neutral text), recent research has shifted toward a more nuanced objective: explainable and fine-grained online sexism detection.
This research direction focuses on breaking down sexism into structured taxonomies, capturing implicit or subtle biases, and providing explanations for model decisions. Achieving this requires moving beyond simple keyword matching toward architectures capable of hierarchical reasoning, task-adaptive representation learning, and robust handling of severe data imbalance.
The Challenge of Nuance and Implicitness
Why does standard text classification fail when applied to online sexism? The primary reason is that sexist language frequently avoids explicit slurs, relying instead on subtle framing, cultural tropes, or implicit assertions of inferiority.
Traditional machine learning algorithms struggle to capture these dynamics. For instance, Judith Jeyafreeda Andrew demonstrates in JudithJeyafreeda at SemEval-2023 Task 10: Machine Learning for Explainable Detection of Online Sexism that while basic classifiers can achieve reasonable performance on coarse binary detection (sexist vs. non-sexist), they fail significantly when tasked with fine-grained categorization. Similarly, Jayant Panwar et al. show in PanwarJayant at SemEval-2023 Task 10: Exploring the Effectiveness of Conventional Machine Learning Techniques for Online Sexism Detection that conventional models lack the semantic capacity to parse the complex, multi-class nature of sexist discourse.
To foster safer online environments, models must be sensitive to these underlying structures. As noted by Fareen Tasneem et al. in KingsmanTrio at SemEval-2023 Task 10: Analyzing the Effectiveness of Transfer Learning Models for Explainable Online Sexism Detection, transfer learning models are essential precisely because they can map the implicit contextual relationships that define subtle prejudice.
Deconstructing Sexism via Hierarchical Taxonomies
To make detection explainable, researchers decompose "sexism" into explicit behavioral categories. Rather than outputting a single toxic score, systems are designed to categorize the type of harm being committed.
In a large-scale study of digital gender-based violence, Nabamallika Dehingia et al. developed and validated a multi-tier taxonomy in Violence against women on Twitter in India: Testing a taxonomy for online misogyny and measuring its prevalence during COVID-19. Their framework breaks online misogyny into distinct operational categories:
- Sexist abuse and direct harassment
- Sexual objectification
- Threats of harm and violence
- Asserting inferiority or enforcing rigid gender roles
- Justifying violence or dismissing systemic inequality
This structured framing underpins major benchmark evaluations, such as the SemEval-2023 Task 10 on the Explainable Detection of Online Sexism (EDOS). As detailed by Amit Das et al. in AU_NLP at SemEval-2023 Task 10: Explainable Detection of Online Sexism Using Fine-tuned RoBERTa, the task is organized into three progressive layers of granularity: Subtask A (binary detection), Subtask B (broad category identification), and Subtask C (fine-grained vector classification).
[Social Media Post]
│
▼
┌────────────────────────────────────────┐
│ Subtask A: Is it sexist? │ ──► (Binary Filter)
└──────────────────┬─────────────────────┘
│ Yes
▼
┌────────────────────────────────────────┐
│ Subtask B: What is the category? │ ──► (e.g., Sexual Objectification)
└──────────────────┬─────────────────────┘
│
▼
┌────────────────────────────────────────┐
│ Subtask C: Fine-grained Vector │ ──► (Specific behavioral sub-type)
└────────────────────────────────────────┘
Operating across these granularities allows models to build internal representations that mirror human sociological frameworks. Mengyuan Zhou leverages this structure in PingAnLifeInsurance at SemEval-2023 Task 10: Using Multi-Task Learning to Better Detect Online Sexism by implementing a hierarchical, multi-task learning framework. By optimizing the model simultaneously across all three subtasks, the shared encoder learns that coarse-grained features (is it harmful?) should inform and constrain fine-grained decisions (how is it harmful?), leading to more robust and explainable predictions.
Architectural Strategies: Ensembles and Adaptation
To capture the linguistic variation of online platforms, researchers heavily rely on deep pre-trained transformer architectures like BERT and RoBERTa. However, a single model often suffers from blind spots due to the highly varied manifestations of online harassment.
To mitigate this, teams frequently deploy model ensembles. C. Jerin Mahibha et al. (Brainstormers_msec at SemEval-2023 Task 10: Detection of sexism related comments in social media using deep learning) and Shruti Sriram et al. (PoSh at SemEval-2023 Task 10: Explainable Detection of Online Sexism) demonstrate that combining diverse architectures—such as BERT, DistilBERT, RoBERTa, and XLNet—smooths out individual model variances, yielding higher macro-F1 scores on nuanced text. Similarly, Lavinia Felicia Fudulu et al. show in I2C-Huelva at SemEval-2023 Task 10: Ensembling Transformers Models for the Detection of Online Sexism that careful hyperparameter tuning within an ensemble framework is critical for stabilizing multi-class classification performance.
Beyond standard fine-tuning, Task-Adaptive Pretraining (TAPT) has emerged as a key mechanism for bridging the gap between generic language models and social media discourse. Hadiseh Mahmoudi demonstrates in IUST_NLP at SemEval-2023 Task 10: Explainable Detecting Sexism with Transformers and Task-adaptive Pretraining that running an intermediate masked language modeling phase on domain-specific, unlabeled social media text allows the transformer's attention heads to adapt to internet slang, abbreviations, and informal syntax before tackling the target sexism tasks.
Furthermore, the choice of model family introduces trade-offs in interpretability. Kanishk Verma et al. conduct a comparative analysis between encoder-only and decoder-only models in DCU at SemEval-2023 Task 10: A Comparative Analysis of Encoder-only and Decoder-only Language Models with Insights into Interpretability. Their work reveals that while generative decoder models offer flexible explanations, deep interpretability analysis exposes significant ambiguities in existing ground-truth annotations, highlighting the ongoing difficulty of establishing gold-standard baselines for subjective human behavior.
Mitigating Data Scarcity and Class Imbalance
A foundational barrier to training deep sexism detectors is data distribution. In any given social media stream, explicit or fine-grained sexist comments are heavily outnumbered by benign text, creating a severe class imbalance. Furthermore, annotating fine-grained data requires extensive human labor, leading to a scarcity of labeled examples.
To prevent models from simply predicting the majority non-sexist class, Doaa Obeidat et al. utilize selective ensembles paired with focal loss in JUST_ONE at SemEval-2023 Task 10: Explainable Detection of Online Sexism (EDOS). Focal loss dynamically down-weights the loss assigned to easy-to-classify, benign examples, forcing the network to focus its gradient updates on hard, ambiguous, or rare sexist instances.
Other approaches expand the training signal through semi-supervised learning and data augmentation:
- Consistency Training and Synonym Replacement: Yehui Xu and Haiyan Ding (Tsingriver at SemEval-2023 Task 10: Labeled Data Augmentation in Consistency Training) introduce a method that augments labeled data using TF-IDF-based synonym replacement. By training the model to output identical predictions for both the original text and its perturbed counterpart, the network learns semantic invariance, preventing it from overfitting to specific keywords.
- Unsupervised Data Augmentation (UDA): To leverage vast pools of unlabeled social media text, Wiebke Petersen et al. (hhuEDOS at SemEval-2023 Task 10: Explainable Detection of Online Sexism (EDOS) Binary Sexism Detection (Subtask A)) and Bingjie Yu et al. (DUTIR at SemEval-2023 Task 10: Semi-supervised Learning for Sexism Detection in English) employ semi-supervised frameworks. UDA enforces consistency constraints across large unlabeled corpora, ensuring that the model's latent space remains stable even when exposed to unannotated linguistic variations.
- In-Context LLM Augmentation: Weston Feely et al. demonstrate in QCon at SemEval-2023 Task 10: Data Augmentation and Model Ensembling for Detection of Online Sexism that combining transformer ensembles with in-context learning via large language models like GPT-3 can generate synthetic training examples or provide zero-shot validation for highly complex, fine-grained categories.
Extending to Low-Resource Contexts
The data scarcity problem is magnified when moving outside of English. The broader field of online threat mitigation emphasizes that identifying the specific groups targeted by offline harm is as critical as predicting the harm itself, a concept underscored in the Proceedings of the Fourth Workshop on Threat, Aggression & Cyberbullying @ LREC-COLING-2024.
When applied to low-resource languages, general-purpose multilingual models often fail to capture localized cultural nuances. To address this, Farjana Alam Tofa et al. focus on abusive comments targeting women in Malayalam text in CUET_Novice@DravidianLangTech 2025: Abusive Comment Detection in Malayalam Text Targeting Women on Social Media Using Transformer-Based Models. By fine-tuning a language-specific encoder (Malayalam-BERT), they show that localized pre-training allows models to successfully capture gender-based harassment within specific regional digital ecosystems, proving that the principles of fine-grained detection can scale effectively across diverse languages and cultures.
Key Foundational Work
DCU at SemEval-2023 Task 10: A Comparative Analysis of Encoder-only and Decoder-only Language Models with Insights into Interpretability (2023) — contributed 1 concept elements to this direction.
CNLP-NITS at SemEval-2023 Task 10: Online sexism prediction, PREDHATE! (2023) — contributed 1 concept elements to this direction.
hhuEDOS at SemEval-2023 Task 10: Explainable Detection of Online Sexism (EDOS) Binary Sexism Detection (Subtask A) (2023) — contributed 1 concept elements to this direction.