Automatically Discovering How Misogyny is Framed on Social Media.
| Authors | Rakshitha Rao Ailneni & Sanda M. Harabagiu |
| Year | 2025 |
| Venue | NAACL 2025 |
| Paper | View on DBLP |
Abstract
Abstract not yet available in this stub. Read the full paper →
Engineering Breakdown
Plain English
This paper addresses the problem of automatically identifying and categorizing how misogyny is framed and expressed on social media platforms. The authors developed a method to discover the distinct linguistic patterns and rhetorical frames that characterize misogynistic content, moving beyond simple binary detection to understand the nuanced ways that gendered hostility manifests online. This work is important because social media platforms need more sophisticated tools than keyword matching to enforce policies against gender-based harassment, and understanding these frames is essential for content moderation at scale. The paper likely combines NLP techniques with frame analysis to extract recurring patterns from social media datasets and map them to underlying misogynistic ideologies.
Core Technical Contribution
The core innovation is applying frame semantics and discourse analysis at scale to automatically discover and categorize how misogyny is rhetorically constructed on social media, rather than treating it as a monolithic phenomenon. Unlike prior work that focuses on binary misogyny classification (is this content misogynistic: yes/no), this approach extracts the underlying frames—the conceptual structures and narrative patterns—that enable misogynistic expression. The technical contribution likely involves unsupervised or semi-supervised methods to cluster similar linguistic patterns and then interpret those clusters as coherent frames, potentially using language models or dependency parsing to surface the semantic relationships. This shift from classification to frame discovery represents a meaningful advance because it provides explainability and enables targeted interventions at the level of how ideas are communicated, not just whether flagged content exists.
How It Works
The system likely begins by collecting a corpus of misogynistic content from social media platforms, preprocessed for language understanding and noise removal. Next, the authors probably extract linguistic and semantic features using techniques like dependency parsing, semantic role labeling, or embeddings to represent the conceptual structure of each text. An unsupervised clustering algorithm (possibly K-means, hierarchical clustering, or a neural clustering approach) groups similar texts based on shared frames, with the number of clusters potentially determined by silhouette analysis or other intrinsic evaluation metrics. The discovered clusters are then interpreted and labeled by examining prototypical examples and shared linguistic patterns within each cluster to identify the overarching frame or narrative. Finally, the system validates these frames by checking their coherence and coverage against the original corpus, and may apply a language model to generate natural language descriptions of each frame for interpretability. The output is a taxonomy of misogynistic frames with examples, allowing platforms to understand not just whether content is harmful, but why and how it constructs that harm.
Production Impact
For engineers building content moderation systems, this approach transforms how platforms can scale policy enforcement and explainability around gender-based harassment. Instead of maintaining lists of keywords or training binary classifiers that provide little insight into why content violates policy, teams can implement frame-based moderation that catches variations of the same harmful rhetoric under different linguistic surface forms. This has immediate benefits: reduced false negatives (catching paraphrased misogyny), improved appeals handling (explaining to users exactly which frame their content violates and why), and better training data for annotation (using frame definitions as guidelines for labelers). However, the trade-offs are real: frame discovery requires substantial labeled data upfront and NLP expertise, the computational cost of parsing and clustering large datasets is non-trivial (potentially 5-10x the cost of simple keyword matching), and frames discovered in one language or cultural context may not transfer cleanly to others. Integration into existing moderation pipelines would require routing flagged content to a frame classifier before policy decision, adding 100-200ms latency per item.
Limitations and When Not to Use This
The paper's frame discovery approach assumes that misogyny is expressed through consistent, discoverable linguistic patterns, which may not hold when bad actors actively vary their language to evade detection or when misogyny is implicit and contextual rather than explicit. The method likely requires large, well-labeled datasets of misogynistic content to extract reliable frames, which introduces ethical challenges around collection and annotation—exposing annotators to harmful content and raising questions about whether platform data can be adequately anonymized. Frames discovered from one platform (e.g., Twitter) or language community may not generalize well to others, and the discovered frames are inherently snapshot-in-time artifacts that become outdated as online discourse evolves and new rhetorical strategies emerge. The paper also likely does not address the problem of distinguishing between misogyny and legitimate feminist discourse that reclaims or critiques misogynistic language, which could lead to false positives that harm women's voices on the platform.
Research Context
This work builds on decades of computational linguistics research in frame semantics (following FrameNet and related projects) and more recent work on detecting abusive language and hate speech online. It extends prior NLP work on misogyny detection (such as shared tasks in SemEval) by moving from binary classification to the finer-grained problem of understanding rhetorical construction and ideology. The paper also connects to social science research on how framing shapes discourse and political psychology research on the role of language in spreading harmful ideologies. This research opens directions for more interpretable and adaptive content moderation systems, and may inspire similar frame-discovery approaches for other forms of harmful content (e.g., Islamophobia, transphobia) where understanding rhetorical patterns is as important as detection.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::