Source-linked AI summary
Beyond Accuracy: A Qualitative Analysis of Vision-Language Models for Hate Speech Detection in Memes
Muhammad Jawad Chowdhury, Adiba Hasan, Ishrak Hossain, Shahriar Ivan, Sabbir Ahmed
TL;DR
Hateful meme detection requires interpreting image and text together, including context, irony, and subtle cues that VLMs may miss. This paper evaluates four VLMs under zero-shot and few-shot prompting through quantitative classification and qualitative justification analysis. The results show persistent reasoning vulnerabilities and limited reliability for complex multimodal hate detection, despite GPT-4o mini’s balanced performance.
Problem
Existing evaluations emphasize narrow quantitative metrics, leaving qualitative reasoning and failure modes across VLM architectures and prompting conditions underexamined.
Method
The study evaluates four VLMs on HMCD using zero-shot and few-shot prompts, then analyzes classifications, generated justifications, and recurring reasoning failures.
Results
All tested models showed significant vulnerabilities, while GPT-4o mini exhibited the most balanced performance across the evaluation.
Takeaways & Limitations
Accuracy alone is not a reliable measure of VLM ability to moderate complex multimodal hate speech.
Takeaways & Limitations
The models consistently failed to infer unstated real-world knowledge and relied largely on literal image and text meanings.
Abstract
from arXiv · showhide
Memes have turned out to be a powerful tool through which individuals share their ideas concerning contemporary social and political problems. Their anonymity, as well as their ability to go viral, make them a powerful medium for spreading hate. It remains very difficult to identify such complex and context-dependent hate speech. Although they display excellent performance on multimodal tasks, vision-language models (VLMs) tend to ignore context, irony, and other subtle cues that play a key role in identifying hateful memes. In this work, we present a qualitative analysis of four state-of-the-art VLMs: LLaVA-7B, Qwen-VL, GPT-4o mini, and Claude 3 Haiku. We evaluate these models under zero-shot and few-shot prompting to examine how contextual framing influences their outputs. Our analysis goes beyond simple classification accuracy and focuses on a qualitative evaluation of the models' generated justifications, providing a more in-depth understanding of their thought processes and constraints when dealing with hateful memes.
Beyond Accuracy: A Qualitative Analysis of Vision-Language Models for Hate Speech Detection in Memes
This study examines how four vision-language models detect hateful memes beyond classification accuracy, emphasizing contextual reasoning and generated justifications.
- The paper qualitatively analyzes LLaVA-7B, Qwen-VL, GPT-4o mini, and Claude 3 Haiku on hateful meme detection.
- Zero-shot and few-shot prompting are compared to examine how contextual framing influences model outputs.
- The evaluation focuses on generated justifications to expose models’ thought processes and constraints beyond simple accuracy.
1.1 INTRODUCTION
The introduction frames hateful meme detection as a multimodal, context-dependent problem and identifies a gap in qualitative comparisons of VLM reasoning under varied prompts.
- Hatefulness may emerge from the combined interpretation of a meme’s image and text rather than either modality alone.
- Although VLMs offer contextual and semantic capabilities, meme hate detection remains challenging because it requires joint visual-textual understanding.
- Prior systems struggled with covert or culturally nuanced hate because representation matching did not provide deep contextual reasoning.
- Known failure modes include contextual misinterpretation, keyword oversensitivity, nuanced-hate failures, and correct classifications supported by flawed reasoning.
- Existing evaluations largely emphasize quantitative metrics, leaving qualitative reasoning differences across architectures and prompting conditions underexamined.
- The study evaluates four VLMs, compares zero-shot with context-rich few-shot prompting, and analyzes established reasoning failures.
1.2 METHODOLOGY
The methodology evaluates four VLMs on a balanced HMCD subset under zero-shot and few-shot prompts, combining binary classification metrics with manual justification analysis.
- 1.2.1 Dataset: The study uses HMCD, a benchmark designed for multimodal reasoning and containing benign confounders requiring joint image-text interpretation.
- 1.2.1 Dataset: The full dataset contains 10,000 image-text pairs labeled hateful or non-hateful across five semantic categories.
- 1.2.1 Dataset: A balanced subset of 500 memes contains 250 hateful and 250 non-hateful samples spanning all five semantic categories.
- 1.2.2 Model and Prompting Configurations: Four VLMs are evaluated, including three open-source models and Claude 3 Haiku as a proprietary API-based benchmark.
- 1.2.2 Model and Prompting Configurations: Zero-shot prompts provide instructions without examples, whereas few-shot prompts provide three labeled meme examples with justifications.
- 1.2.3 Task Formulation and Evaluation: Each model-prompt configuration performs binary classification and generates a textual justification on the same 500 memes.
- 1.2.3 Task Formulation and Evaluation: Manual review of justifications categorizes prompt sensitivity, flawed reasoning, contextual misinterpretation, keyword oversensitivity, and nuanced-hate failures.
1.3 EMPIRICAL ANALYSIS AND FINDINGS
The empirical analysis combines manual inspection of model outputs with standard quantitative metrics to assess classification performance and robustness across configurations.
- The evaluation examines correct and incorrect classifications, recurring justification inconsistencies, and standard metrics across all model-prompt configurations.
1.3.1 Qualitative Error Analysis
Qualitative analysis identified recurring VLM failures that accuracy alone can conceal: prompt-induced shifts, flawed rationales, missing real-world inference, keyword-triggered false positives, and missed implicit hate.
- Overview: Qualitative review of misclassified memes identified recurring error archetypes beyond quantitative performance metrics.The analysis examined model justifications to characterize specific failure modes.
- Prompt Sensitivity and Reasoning Failure: LLaVA-7B changed its classification of a Holocaust-alluding meme from non-hateful to hateful when the prompt changed from zero-shot to few-shot.This shift was termed Example-Induced Bias and was linked to overfitting to superficial patterns in non-representative examples.
- Correct Classification But Flawed Reasoning: GPT-4o-mini and LLaVA correctly classified hateful memes while relying on a neutral template or hallucinated phrase instead of the actual hateful text.These cases show that correct labels can arise from flawed heuristics or hallucinated evidence.
- Contextual Misinterpretation: All models classified a context-dependent hateful meme as non-hateful because they read it literally and failed to infer the cruel mockery requiring real-world knowledge.The models identified the scenario as a joke but did not access the necessary commonsense context.
- Failure to Detect Coded or Nuanced Hate: Qwen2-7B and Claude 3 Haiku missed implicit hateful intent when interpreting an ironic meme literally rather than integrating visual distress with its text.The case highlights difficulty recognizing coded or nuanced hate and moving beyond explicit features.
1.3.2 Quantitative Performance Overview
The four VLMs showed heterogeneous performance profiles under zero-shot and few-shot prompting. GPT-4o mini led hateful-class F1, while Qwen2-7B traded high precision for low recall.
- The evaluation used Precision, Recall, and F1-Score for hateful and non-hateful classes under zero-shot and few-shot prompting.Results were reported on a balanced test set in Table 1.2.
- Performance profiles varied substantially across the evaluated VLMs.The differences may reflect variation in training data, architecture, and alignment procedures.
- GPT-4o mini achieved the highest hateful-class F1-Score in both prompt configurations, with scores of 0.59 and 0.62.Its performance reflected a balance between precision and recall.
- Qwen2-7B achieved the highest precision, with scores of 0.70 and 0.71.This profile was associated with reliable text-based hate detection.
- Qwen2-7B had the lowest recall, with scores of 0.34 and 0.32, missing roughly 70% of hateful content.Its literal handling of ambiguous, non-explicit hate contributed to this low recall.
1.3.3 Discussion
The results reveal a gap between apparent task performance and genuine reasoning. Models often reached correct classifications through flawed logic or by omitting real-world knowledge, making accuracy alone insufficient for judging moderation ability.
- All models frequently made repeated reasoning errors, including correct classifications reached through flawed logic.They also often failed to incorporate real-world knowledge.
- Qwen2-7B’s high precision and very low recall reflected an overly literal strategy that missed most subtle hateful cases.Avoiding errors came at the cost of detecting non-explicit hate.
- Accuracy alone is not a reliable measure of a model’s true ability to moderate content.
1.4 CONCLUSION & FUTURE SCOPES
The study finds that current VLMs remain unreliable for complex multimodal hate speech, despite progress and differences across architectures. Future work should test broader settings and strengthen reasoning and mitigation strategies.
- Conclusion: All tested VLMs show significant vulnerabilities that limit their reliability for real-world hate-speech moderation.The models remain fragile when hateful meaning depends on subtle expressions and contextual variation.
- Conclusion: GPT-4o-mini exhibits the most balanced performance, while performance discrepancies persist across architectures.
- Conclusion: A consistent failure to infer unstated real-world knowledge leaves model reasoning dependent on literal image and text meaning.This limits common-sense and emotional understanding during interpretation.
- Future Scopes: Future evaluations should examine targeted tasks, larger model families, alternative prompting strategies, and multilingual datasets.These comparisons can test whether observed failure modes are language-specific or universal.
- Future Scopes: Research should also assess specialized datasets and develop mitigation strategies for precise hate targets, sub-categories, and observed failure modes.