💬 NLP

Beyond Binary Classification: Detecting Fine-Grained Sexism in Social Media Videos

This paper introduces FineMuSe, a new Spanish multimodal dataset with fine-grained annotations and a hierarchical taxonomy, to advance sexism detection beyond binary classification by evaluating multimodal LLMs that show competitive performance with humans in identifying nuanced sexism while highlighting challenges in detecting co-occurring types conveyed through visual cues.

Original authors: Laura De Grazia, Danae Sánchez Villegas, Desmond Elliott, Mireia Farrús, Mariona Taulé

Published 2026-02-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Laura De Grazia, Danae Sánchez Villegas, Desmond Elliott, Mireia Farrús, Mariona Taulé

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to spot rude or unfair behavior in a crowded, noisy video chat room.

For a long time, the robots we built were like very strict bouncers. They could only say two things: "You're in" (this is fine) or "You're out" (this is sexist). If a video was clearly hateful, the bouncer kicked it out. But if the sexism was subtle—like a joke that sounded funny but was actually mean, or a comment that sounded like a compliment but was actually a trap—the bouncer would let it slide. They were too binary; they only saw black and white, missing all the shades of gray.

This paper introduces a new, much smarter system called FineMuSe (think of it as a "Fine-Tuned Microscope for Sexism"). Here is how it works, broken down simply:

1. The New Dataset: A "Museum of Mean Moments"

The researchers collected 828 short videos from places like TikTok, YouTube Shorts, and BitChute. But instead of just labeling them "bad" or "good," they created a massive, detailed catalog.

Imagine a detective's board covered in red string. They didn't just say, "This video is sexist." They asked:

Is it a Stereotype? (e.g., "Women are bad at math.")
Is it Denial? (e.g., "There is no such thing as gender inequality anymore.")
Is it Discrimination? (e.g., Attacking someone for being LGBTQ+.)
Is it Objectification? (e.g., Treating a person like a piece of furniture to be looked at.)
Is it a Joke? (Is the sexism hidden inside a punchline?)

They also labeled videos that fight back against sexism or report bad experiences. This creates a rich map of exactly how sexism shows up, not just that it shows up.

2. The Challenge: The "Visual Trap"

The researchers tested the world's smartest AI brains (Large Language Models) on this new map. They gave the AI two types of clues:

The Script: Just the words spoken in the video.
The Full Picture: The words, the tone of voice, and the actual video images.

The Result: The AI got pretty good at reading the script. If someone said, "Women should stay home," the AI knew that was sexist.

The Problem: The AI struggled when the sexism was hidden in the visuals.
Imagine a video where a man says, "I love my wife," which sounds nice. But the video shows him treating her like a servant, carrying her bags while she sits on a throne. The AI read the nice words but missed the mean picture. It's like a person who only listens to the lyrics of a song but doesn't see the angry face of the singer. The AI needs to learn to "watch" the video, not just "read" the transcript.

3. The "Human vs. Robot" Showdown

The researchers also asked: "Can the AI explain why it flagged a video?"
They compared the AI's explanations to those written by human experts.

The Verdict: The AI is surprisingly good at writing explanations! It can say, "This is sexist because it uses a joke to hide a stereotype," almost as well as a human.
The Catch: Sometimes the AI gets the reason right but misses the visual part of the reason. It's like a detective who correctly identifies the crime but forgot to look at the fingerprint on the window.

4. Why This Matters

Think of online moderation like trying to filter water.

Old Way: You have a big net that catches only the biggest rocks (obvious hate speech). The tiny, invisible sand (subtle sexism) slips right through, making the water still dirty.
New Way (FineMuSe): You have a high-tech filter that catches the rocks, the sand, and even the microscopic algae. It understands that sexism isn't just one thing; it's a chameleon that changes colors depending on the platform (TikTok vs. YouTube) and the language (Spanish dialects from Mexico vs. Spain).

The Bottom Line

This paper is a giant leap forward. It tells us that while AI is getting very smart at spotting obvious sexism, it still needs to learn how to "read the room" visually. It's not enough to just hear the words; the AI needs to understand the body language, the context, and the hidden jokes.

By creating this detailed "museum" of examples, the researchers are giving future AI a better textbook to study, hoping to build tools that can finally catch the subtle, sneaky forms of unfairness that have been hiding in plain sight.

1. Problem Statement

Current automated tools for detecting sexism on social media are predominantly limited to binary classification (sexist vs. non-sexist). This approach fails to capture the nuance, context, and specific manifestations of sexism, such as stereotypes, denial of inequality, discrimination, and objectification. Furthermore, existing research often focuses on:

Text-only or meme-based data, neglecting the complex interplay of text, audio, and visual cues in short-form videos.
Single-platform datasets, which limits generalizability across different user communities and linguistic domains.
Lack of fine-grained annotations for rhetorical devices like irony and humor, which are frequently used to mask or subvert sexist content.

The authors argue that without fine-grained, context-sensitive labels and multimodal analysis, subtle forms of sexism remain undetected, and automated moderation tools lack the transparency required to build trust with users and moderators.

2. Methodology

A. Dataset Construction: FineMuSe

The core contribution is FineMuSe, a new multimodal dataset in Spanish designed for fine-grained sexism detection.

Sources: The dataset comprises 828 videos collected from three distinct platforms to ensure diversity:
- TikTok: 367 videos (moderated platform).
- BitChute: 33 videos (platform with high prevalence of hateful content).
- YouTube Shorts: 428 videos (added to extend MuSeD, focusing on short-form content).
Language: Spanish, covering both Peninsular and Latin American varieties.
Annotation Process:
- Expert Annotators: Six highly trained annotators with backgrounds in linguistics and sexism detection were used to ensure high quality.
- Multi-modal Approach: Annotations were performed across three levels: Text (transcripts via Whisper), Audio, and Video.
- Granularity: The dataset includes both binary labels and fine-grained multi-label annotations.

B. Hierarchical Taxonomy

The authors developed a three-level hierarchical taxonomy (Figure 1) to categorize content:

Level 1 (Binary): Sexist vs. Non-sexist.
Level 2 (Fine-Grained Categories):
- Sexist: Stereotypes, Denial of Inequality/Rejection of Feminism, Discrimination (based on sexual orientation/gender identity), and Objectification.
- Non-sexist: Counter-speech (critiquing sexism) and Reported Sexism (narrating personal experiences).
- Note: These categories are non-exclusive; a single video can contain multiple labels.
Level 3 (Rhetorical Devices): Irony and Humor (applicable to both sexist and non-sexist content).

3. Key Contributions

FineMuSe Dataset: The first Spanish multimodal dataset for sexism detection that combines binary and fine-grained annotations across text, audio, and video modalities from multiple platforms.
Comprehensive Taxonomy: A novel hierarchical framework that captures overlapping forms of sexism, non-sexist responses, and rhetorical strategies (irony/humor).
Extensive LLM Evaluation: A benchmarking study evaluating a wide range of Large Language Models (LLMs), including text-only, vision-language (V+L), and video-processing models.
Human-Model Correlation Analysis: An analysis of how model predictions correlate with human annotations and an evaluation of the quality of model-generated explanations.

4. Experimental Results

A. Model Performance

The study evaluated various models (Llama-3, Qwen, GPT-4o, Claude-3.7, Gemini) using a zero-shot prompting strategy.

Binary Detection: Multimodal models (specifically Claude-3.7-Sonnet (V+L)) achieved the highest accuracy (85.8% on Part 1, 83.6% on Part 2), outperforming text-only models. This confirms that visual context is crucial for detecting sexism.
Fine-Grained Detection:
- Scale Matters: Larger models (e.g., Llama-3-70B, GPT-4o) significantly outperformed smaller counterparts.
- Multimodal Advantage: Multimodal models consistently outperformed text-only models, particularly on YouTube Shorts (Part 2), where visual cues were critical.
- Category Difficulty:
  - Easiest: Stereotypes and Inequality (F1 > 69 for top models).
  - Moderate: Discrimination (F1 ~ 81 for top models).
  - Hardest: Objectification. Even the best models struggled (F1 < 50), likely due to the scarcity of such content in the dataset (due to platform moderation) and the difficulty of interpreting visual objectification without explicit verbal cues.

B. Human-Model Alignment

Correlation: The best-performing model (Claude-3.7-Sonnet V+L) showed strong correlation with human annotators for Stereotypes and Discrimination (Pearson $r \approx 0.70$ ).
Weakness: Correlation was low for Objectification ( $r \approx 0.15$ in Part 1), indicating models struggle to detect objectification when conveyed purely through visual cues without explicit text.
Explanation Quality: In a study comparing model-generated justifications to human-written ones, models produced explanations of comparable quality regarding relevance, informativeness, and groundedness, though humans slightly preferred their own explanations in specific cases.

5. Significance and Conclusion

Beyond Binary: The paper demonstrates that moving beyond binary classification is essential for understanding the nuances of online sexism.
Multimodal Necessity: It establishes that effective sexism detection in social media videos requires integrating text, audio, and visual modalities; text-only models miss critical context.
Platform Diversity: By including TikTok, BitChute, and YouTube Shorts, the study highlights how sexism manifests differently across platforms (e.g., more stereotypes on TikTok/BitChute vs. more argumentative inequality discourse on YouTube).
Future Directions: The authors suggest that future work should focus on improving the detection of Objectification via visual cues and expanding data collection for Irony and Humor, which remain challenging due to their subjective nature.

In summary, FineMuSe provides a robust foundation for developing more transparent, accurate, and nuanced automated moderation tools capable of identifying the complex, multi-layered nature of sexism in modern social media.