Hateful Meme Detection Eval

hateful-meme-detection-eval

Robust Adaptation of Large Multimodal Models for Retrieval Augmented Hateful Meme Detection — Mei et al. (2025) (arXiv:2502.13061, 2025)

What this evaluates

Evaluates multimodal models' ability to detect hateful or offensive memes across multiple domains and under low-resource, out-of-distribution conditions. It probes robustness to distribution shifts, adversarial image perturbations, and the effectiveness of retrieval-augmented inference versus standard fine-tuning or in-context learning.

Datasets

• HatefulMemes — total ?; splits: train (-1), test (-1)
• HarMeme — total ?; splits: train (-1), test (-1)
• MAMI — total ?; splits: train (-1), test (-1)
• Harm-P — total ?; splits: train (-1), test (-1)
• MultiOFF — total ?; splits: train (-1), test (-1)
• PrideMM — total ?; splits: train (-1), test (-1)

Metrics

• AUC (primary) — range: [0, 1]
  • Area under the Receiver Operating Characteristic curve, measuring the model's ability to distinguish between hateful and non-hateful classes across all classification thresholds.
• Accuracy — range: [0, 1]
  • Ratio of correctly classified instances (both hateful and non-hateful) to the total number of instances.
• F1 — range: [0, 1]
  • Harmonic mean of precision and recall, calculated as 2 * (precision * recall) / (precision + recall).

Input / output format

Input: Image and associated text/caption for each meme instance.

Output: Binary classification prediction (hateful vs. non-hateful) or class probability scores for AUC computation.

Scoring recipe

import numpy as np
from sklearn.metrics import roc_auc_score, accuracy_score, f1_score

def compute_metrics(y_true, y_pred_proba, threshold=0.5):
    y_pred = (y_pred_proba >= threshold).astype(int)
    auc = roc_auc_score(y_true, y_pred_proba)
    acc = accuracy_score(y_true, y_pred)
    f1 = f1_score(y_true, y_pred)
    return {'AUC': auc, 'Accuracy': acc, 'F1': f1}

Common pitfalls

• Using test set examples for few-shot or retrieval-based inference instead of the training split, which causes data leakage and inflates performance.
• Assuming few-shot in-context learning improves performance; the paper explicitly notes it is largely ineffective for meme classification and can degrade F1 on some datasets.
• Computing AUC for GPT-4o directly; the paper notes that token likelihoods are not accessible for this model, so AUC is reported as '-'.

Evidence (verbatim from paper)

RA-HMD improves upon RGCL with gains of over 4% in AUC and 3% in accuracy on HatefulMemes.

Citation

@misc{mei2025robust,
  title={Robust Adaptation of Large Multimodal Models for Retrieval Augmented Hateful Meme Detection},
  author={Mei et al. (2025)},
  year={2025},
  note={arXiv:2502.13061}
}

• arXiv: 2502.13061