TLDR: This research paper introduces novel machine learning models, Speech Concept Bottleneck Models (SCBM and SCBMT), designed for detecting and categorizing sexism in social media text as part of the EXIST 2025 challenge. While a fine-tuned XLM-RoBERTa model proved highly effective for binary sexism identification, the SCBMT model demonstrated superior performance in more complex multi-label sexism categorization and source intention tasks. A key advantage of SCBMT is its valuable interpretability, achieved through adjective-based explanations, and its strong language robustness across English and Spanish content.
Sexism has become a pervasive issue across social media platforms and online discussions, significantly impacting the well-being of individuals and potentially leading to real-world harm. To combat this growing problem, the Sexism Identification in Social Networks (EXIST) challenge is held annually, with the 2025 edition at CLEF 2025 focusing on advancing automatic tools for identifying sexism in multimedia content.
Researchers from St. Pölten University of Applied Sciences have contributed to this effort by focusing on Task 1 of the EXIST benchmark: identifying and classifying sexism in textual social media posts. Their work addresses three specific subtasks: Sexism Identification in Tweets (binary classification: sexist or non-sexist), Source Intention in Tweets (multiclass classification: direct, reported, or judgmental intention), and Sexism Categorization in Tweets (multi-label classification: identifying specific types like ideological inequality, stereotyping, objectification, sexual violence, and misogyny).
Innovative Models for Sexism Detection
The team implemented three distinct machine learning models to tackle these subtasks, each offering unique strengths. The first is the Speech Concept Bottleneck Model (SCBM), a human-interpretable model that uses large language models (LLMs) to transform input text into an adjective-based representation. This means the model identifies how well specific descriptive adjectives (like “abusive,” “misogynistic,” or “factual”) describe the text, providing clear, human-understandable reasons for its predictions. This approach offers both local (instance-level) and global (class-level) explanations for its decisions.
The second model, the Speech Concept Bottleneck Model with Transformer (SCBMT), extends SCBM by combining these adjective-based representations with contextual embeddings generated by a transformer model, specifically XLM-RoBERTa-large. This fusion aims to balance the interpretability of SCBM with the powerful semantic understanding of transformers, capturing a broader range of textual features. Unlike traditional transformer models, the transformer component in SCBMT is not fine-tuned, offering a more lightweight alternative.
As a baseline for comparison, the researchers also fine-tuned a multilingual XLM-RoBERTa transformer model. This model was trained on the EXIST2025 dataset, augmented with data from previous EXIST benchmarks (EXIST2022 and EXIST2024) to enhance its robustness and generalizability.
Experimental Approach and Key Findings
The experiments involved training these models on the EXIST2025 dataset, which contains over 10,000 English and Spanish text posts from X/Twitter, annotated by multiple individuals. The researchers also explored the impact of incorporating annotators’ demographic information and strategies to mitigate class imbalance within the datasets.
The results showed varied performance across the subtasks. For Subtask 1.1 (Sexism Identification), the fine-tuned XLM-RoBERTa model performed exceptionally well, achieving a high rank in the overall benchmark. This suggests that for straightforward binary classification, a robust, fine-tuned transformer is highly effective, especially when augmented with additional data.
However, for the more nuanced and fine-grained classification tasks of Subtask 1.2 (Source Intention) and Subtask 1.3 (Sexism Categorization), the SCBMT model demonstrated superior performance. It achieved stronger results, particularly for Spanish content, outperforming both the standalone SCBM and the fine-tuned XLM-RoBERTa. This highlights SCBMT’s ability to capture the complementary semantic information needed for complex multi-label and multi-class problems.
Also Read:
- LOTUS: A New Framework for Evaluating Advanced Image Captioning
- New Method Extends AI Safety from Text to Images
Interpretability and Future Directions
A significant advantage of the SCBM and SCBMT models is their inherent interpretability. By using descriptive adjectives as “bottleneck concepts,” these models can explain their predictions in human-understandable terms. For instance, in identifying direct sexist messages, adjectives like “invasive” and “violence-inciting” were highly relevant, while “chauvinistic” and “victim-blaming” were associated with judgmental content. This level of transparency is crucial for understanding how AI systems make decisions in sensitive areas like sexism detection.
The study also noted the surprising language robustness of SCBMT, performing comparably well on Spanish data even though the adjectives and prompts were initially designed for English. This suggests the generalizability of the semantic cues captured by the adjective-based concepts. Future work aims to refine prompting techniques for adjective generation and expand the lexicon with input from domain experts to further improve coverage and interpretability.
This research provides valuable insights into developing effective and transparent AI tools for combating online sexism. For more detailed information, you can read the full research paper here.
