💬 NLP

Labels or Input? Rethinking Augmentation in Multimodal Hate Detection

This paper proposes an end-to-end pipeline that enhances compact Vision-Language Models for multimodal hate detection by combining structured prompts, granular labeling, and a novel counterfactual data augmentation framework, thereby achieving robust performance on nuanced content without relying on costly large models.

Original authors: Sahajpreet Singh, Kokil Jaidka, Subhayan Mukerjee

Published 2026-01-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sahajpreet Singh, Kokil Jaidka, Subhayan Mukerjee

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 the internet is a giant, chaotic digital town square. In this square, people share jokes, news, and memes. But sometimes, hidden inside a funny picture with a silly caption, there's a mean-spirited insult or a hateful message. This is like a "wolf in sheep's clothing."

For a long time, computers have tried to spot these "hateful memes" to keep the town square safe. But it's a tough job because the hate isn't always obvious; it's often a mix of a harmless picture and a mean joke.

This paper is like a team of detectives trying to figure out the best way to teach computers to spot these hidden wolves without hiring a million expensive experts to do the work. They tested two main strategies: changing the instructions given to the computer and creating better practice tests.

The Two Detective Strategies

1. The "Instruction Manual" Upgrade (Prompt Optimization)

Imagine you are teaching a robot to spot bad jokes.

The Old Way: You just say, "Is this hateful? Yes or No."
The New Way: The authors tried giving the robot a much more detailed instruction manual. Instead of a simple "Yes/No," they asked the robot to rate the hatefulness on a scale from 0 to 9 (like a pain scale) and gave it specific categories to look for, like "is this sexist?" or "is this racist?"

What they found:

If the task is simple (just Yes/No), a short, simple instruction works best.
But if the task is tricky (rating how bad something is), the detailed, structured instructions made the robot much smarter.
The Surprise: For the biggest, smartest robots, giving them these detailed instructions worked almost as well as spending months training them on new data. It's like realizing that a smart student doesn't need to re-learn the whole subject; they just need a better study guide.

2. The "Counterfeit" Practice Test (Multimodal Augmentation)

The second problem the team faced was that the "textbooks" (datasets) the robots were studying were flawed. Many hateful memes in the training data were just bad pictures with mean captions stuck on top. The robots learned to associate "mean words" with "hate," even if the picture was innocent.

To fix this, the team built a factory that creates "counterfeit" practice tests.

How it works: They took a hateful meme (a nice picture + a mean caption) and used AI to swap the mean caption for a nice, neutral one, while keeping the picture exactly the same.
The Goal: This teaches the robot: "Hey, this picture is actually fine! It's only hateful because of the words. If you change the words, the hate disappears."

What they found:

By adding these "neutralized" memes to the training data, the robots became much better at understanding the context. They stopped just looking for bad words and started looking at the whole picture.
This worked especially well for the smarter, larger robots, helping them catch the subtle, tricky jokes that usually slip through the cracks.

The Results: A Better, Cheaper Solution

The paper concludes that you don't always need the most expensive, massive super-computers to solve this problem.

Better Instructions Matter: If you give a smaller, cheaper computer a really good, structured set of instructions (prompts), it can perform almost as well as a giant, expensive one.
Better Data Matters: If you teach computers with "cleaned up" examples (where the hate is removed but the picture stays), they learn to be fairer and less likely to make mistakes based on stereotypes.

The Bottom Line

The authors built a system that creates these "cleaned up" memes and better instructions to help smaller, more affordable computers do a great job at spotting online hate. They proved that by fixing how we talk to the AI and what we teach it, we can build safer online spaces without needing to spend a fortune on massive technology.

Important Note: The team was very careful to say they did not create new hateful content. Their "factory" only took existing bad memes and turned them into good, neutral ones to use as teaching tools. Their goal was to reduce harm, not create more of it.

Technical Summary: Labels or Input? Rethinking Augmentation in Multimodal Hate Detection

Problem Statement

Online hate speech, particularly in the form of multimodal memes, presents a significant societal challenge due to its reliance on subtle text-image interactions, humor, and cultural context. While recent Vision-Language Models (VLMs) show promise, they face limitations in deployment due to high inference costs and persistent failures on nuanced or implicit content. Furthermore, existing datasets often suffer from structural artifacts, such as templated formats and inconsistent supervision, which can reinforce cultural biases and lead to spurious correlations (e.g., associating specific visual features with hate regardless of the caption).

The core research question addresses whether improving prompting strategies (label augmentation) or refining the underlying data (input augmentation) is more effective than standard baselines for multimodal hateful content classification. The authors aim to bridge the gap between the complexity of online harm and the limited supervision provided by current datasets, specifically exploring how small models can be enhanced without relying on costly large-model inference.

Methodology

The study proposes and evaluates two complementary strategies using the Facebook Hateful Memes dataset (containing ~12.9K image-caption pairs).

1. Prompt Optimization (Label Augmentation)

This approach treats prompting as a multivariate design space to enrich supervision signals.

Experimental Design: A full-factorial design ( $3 \times 2 \times 2$ $3 \times 2 \times 2$ ) varying:
- Learning Modality: Text-only fine-tuning vs. Multimodal fine-tuning/prompting.
- Prompt Structure: Simple (direct instruction) vs. Category-based (defining subtypes like misogyny or xenophobia).
- Label Format: Binary (Hate/No Hate) vs. Scaled (0–9 hatefulness score).
Implementation: A teacher model (GPT-4o-mini) generates scaled labels (0–9) for the training data based on structured prompts. These "soft" labels are used to fine-tune both large multimodal models (InternVL2 4B) and smaller unimodal models (BERT, RoBERTa). Incorrect labels inconsistent with binary ground truth are filtered out.

2. Multimodal Data Augmentation (Input Augmentation)

This approach seeks to improve model robustness by generating counterfactually neutral memes to reduce spurious text-image correlations.

Target: Specifically targets "NH" memes (Non-hateful Image + Hateful Caption), where the visual background is benign but the text induces hate.
Pipeline: A multi-agent system involving:
1. Attribution: A zero-shot classifier (Qwen2.5-14B) identifies memes where the caption is the primary source of hate.
2. Context Extraction: A VLM (InternLM-XComposer-2.5-7B) generates a background description of the image, excluding overlaid text.
3. Neutralization: A generative model (GPT-4o-mini) rewrites the hateful caption into a neutral one, conditioned on the background description to ensure semantic coherence.
4. Regeneration: A rendering model (Gemini 2.0 Flash) overlays the new neutral caption onto the original image layout.
Outcome: The pipeline generated 2,479 counterfactually neutral memes, which were merged with the original dataset to create an "Extended" training set.

Key Contributions

Dataset Extension: The creation of a novel benchmark consisting of 2,479 counterfactually neutral memes designed to evaluate robustness against spurious text-image correlations.
Structured Prompt Framework: A demonstration that prompt design, often overlooked in vision-language tasks, significantly impacts performance. The study shows that structured, category-based prompts are crucial for nuanced, scaled annotations, whereas simple prompts suffice for binary tasks.
Debiasing Pipeline: A proof-of-concept multi-agent pipeline for generating semantically coherent, non-hateful meme variants. This serves as a strategy to mitigate bias and improve generalization without generating new harmful content.

Results and Analysis

Prompt Optimization Findings

Prompt Utility: For binary classification, simple prompts outperformed category-based prompts on smaller unimodal models. However, for scaled (nuanced) labels, category-based prompts significantly improved performance across both unimodal and multimodal models.
Prompting vs. Fine-tuning: On the scaled dataset, prompting the large InternVL2 model with category definitions achieved an F1-score of 64.74, which was competitive with and slightly higher than the best fine-tuning result (63.28). This suggests that for granular hate detection, well-structured zero-shot prompting can rival fine-tuning on augmented data.
Unimodal Robustness: Standard unimodal models (BERT, RoBERTa) effectively utilized categorical context to improve performance when labels were scaled rather than binary.

Multimodal Augmentation Findings

Performance Gains: Training on the extended dataset (original + counterfactuals) yielded modest accuracy gains (<1%) but notable improvements in weighted F1-scores, indicating better handling of class imbalance and edge cases.
- Text Models: RoBERTa F1 improved from 47.18 to 52.54 (+5.36).
- Multimodal Models: Qwen2-VL (7B) achieved the highest overall performance, rising from 69.59 to 71.31 F1.
Zero-Shot Capabilities: Larger VLMs (e.g., Qwen2-VL 7B) demonstrated remarkable zero-shot robustness (F1: 69.13), comparable to their fine-tuned performance. Conversely, smaller models (SmolVLM 2B) relied heavily on fine-tuning, showing a significant performance gap between zero-shot and fine-tuned settings.

Data Quality Evaluation

Label Reliability: Human evaluation of the scaled labels showed a 92.5% agreement rate, confirming the reliability of the machine-generated annotations, though disagreements persisted on indirect or context-dependent hate.
Augmentation Quality: Human evaluation of the 2,479 neutral memes confirmed high quality across formatting, background alignment, and caption coherence. However, a small subset of errors (e.g., missing captions, semantic drift, or visual inconsistencies) highlighted the need for manual validation in automated pipelines.

Significance and Claims

The paper argues that robust multimodal hate detection requires attention to both task framing (prompt design) and data integrity (augmentation).

Practical Pathway: The work offers a practical route to more robust and deployable systems that do not rely solely on expensive large-model inference. It demonstrates that small models can be strengthened through structured prompts and targeted data augmentation.
Ethical Augmentation: By generating counterfactually neutral memes rather than new hateful content, the authors provide an ethically grounded method for refining training distributions and reducing overreliance on spurious correlations.
Nuance over Binary: The study emphasizes that moving beyond binary classification to scaled labels, supported by structured prompting, allows for a more nuanced understanding of harmful content, particularly for culturally grounded memes.

The authors conclude that while their methods show promise, future work must address scaling the augmentation pipeline to culturally diverse settings and resolving remaining challenges in context preservation.