[Paper Review] The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes

Douwe Kiela, Hamed Firooz+5 more|arXiv (Cornell University)|May 10, 2020

Hate Speech and Cyberbullying Detection94 references146 citations

TL;DR

This paper introduces a 10k-meme multimodal dataset for hate speech detection in memes, designed to require true multimodal reasoning and evaluating various unimodal and multimodal models, with humans outperforming current models.

ABSTRACT

This work proposes a new challenge set for multimodal classification, focusing on detecting hate speech in multimodal memes. It is constructed such that unimodal models struggle and only multimodal models can succeed: difficult examples ("benign confounders") are added to the dataset to make it hard to rely on unimodal signals. The task requires subtle reasoning, yet is straightforward to evaluate as a binary classification problem. We provide baseline performance numbers for unimodal models, as well as for multimodal models with various degrees of sophistication. We find that state-of-the-art methods perform poorly compared to humans (64.73% vs. 84.7% accuracy), illustrating the difficulty of the task and highlighting the challenge that this important problem poses to the community.

Motivation & Objective

• Define a hate speech detection task that requires multimodal reasoning beyond unimodal cues.
• Construct a licensed, reconstructive meme dataset with benign confounders to challenge unimodal priors.
• Provide clear evaluation metrics and baselines to benchmark multimodal models vs. humans.
• Analyze annotator agreement and dataset properties to understand hate categories and attack types.
• Offer a public benchmark to drive progress in multimodal understanding and real-world hate speech mitigation.

Proposed method

• Construct a challenge set by reconstructing memes with licensed Getty images to preserve meaning.
• Annotate memes for hatefulness using a 3-point scale (definitely hateful, not sure, definitely not hateful) and derive a binary label.
• Create benign confounders (both image and text) to flip labels and enforce multimodal reliance.
• Split data into dev/test/fine-tuning sets with balanced dev/test distributions and report ROC AUC and accuracy.
• Evaluate a range of unimodal and multimodal models (image-only, text-only, early/mid/late fusion, and multimodal pretraining variants).
• Provide baseline performance figures across models to illustrate the gap to human performance.

Experimental results

Research questions

• RQ1 Can multimodal models surpass unimodal baselines on hate speech detection in memes when benign confounders are present?
• RQ2 How does multimodal pretraining vs. unimodal pretraining affect performance on this task?
• RQ3 What is the gap between current models and human performance on this benchmark?
• RQ4 What roles do image and text cues respectively play in detecting hateful memes?
• RQ5 How reliable are annotations and what is inter-annotator agreement on hatefulness judgments?

Key findings

• Unimodal models struggle relative to multimodal models, with text-based signals providing a modest edge over image-only signals on this task.
• Multimodal models generally outperform unimodal baselines, with early fusion architectures yielding better results than late fusion.
• Multimodal pretraining shows only modest gains over unimodally pretrained baselines, indicating room for improvement in multimodal learning.
• Human performance on the dataset is substantially higher than state-of-the-art models, highlighting the task’s difficulty.
• Annotator agreement is moderate (Cohen’s kappa 68.4), reflecting the challenge in classifying hate speech under the paper’s definition.
• The dataset includes diverse meme types (multimodal hate, unimodal hate, benign confounders, random non-hateful) to stress true multimodal reasoning.