Source-linked AI summary
Detecting Hate Speech in Memes Using Multimodal Deep Learning Approaches: Prize-winning solution to Hateful Memes Challenge
Riza Velioglu, Jewgeni Rose
TL;DR
Hateful memes can emerge from combinations of images and text, making detection a multimodal classification problem. The paper fine-tunes VisualBERT on expanded training data and applies majority voting, placing third in the challenge with reported test-set performance while remaining behind human judgment.
Problem
Hateful meaning can arise from the combination of visual and textual content, while the Hateful Memes Challenge provides more than 10,000 newly created multimodal examples for detection research.
Source: §1 · p. 2Method
The solution fine-tunes pre-trained VisualBERT on an expanded training set and applies majority voting across 27 selected models.
Source: §3 · p. 2, §3.1 · p. 3, §3.4 · p. 4Results
0.811 AUROC and 0.765 accuracy on the challenge test set; the approach placed third among 3,173 participants.
Source: §5 · p. 5Takeaways & Limitations
The reported result is considerable, but the authors conclude that it remains far from human judgment accuracy.
Source: §5 · p. 5
Abstract
from arXiv · showhide
Memes on the Internet are often harmless and sometimes amusing. However, by using certain types of images, text, or combinations of both, the seemingly harmless meme becomes a multimodal type of hate speech -- a hateful meme. The Hateful Memes Challenge is a first-of-its-kind competition which focuses on detecting hate speech in multimodal memes and it proposes a new data set containing 10,000+ new examples of multimodal content. We utilize VisualBERT -- which meant to be the BERT of vision and language -- that was trained multimodally on images and captions and apply Ensemble Learning. Our approach achieves 0.811 AUROC with an accuracy of 0.765 on the challenge test set and placed third out of 3,173 participants in the Hateful Memes Challenge.
1 Introduction
Hate speech in memes can arise from the combination of otherwise neutral or humorous images and text, making holistic multimodal understanding important. The paper presents a challenge solution that outperforms its benchmark models but remains behind human accuracy.
1 Introduction
- Memes can become hateful through the combination of visual and linguistic content, even when either component alone is neutral or funny.Humans understand a meme holistically rather than considering its words and picture independently.Source: §1 · pp. 1–2
- The Hateful Memes Challenge introduced more than 10,000 newly created multimodal memes, following a reported gap between state-of-the-art and human accuracy of 64.73% vs. 84.7% [4].The paper notes that the scale of online hate speech cannot be handled by human inspection of every sample.Source: §1 · pp. 1–2
- Benign confounders pair hateful memes with alternative images or text that flip the label, challenging models to avoid relying on unimodal priors.The challenge is designed to measure multimodal understanding and reasoning.Source: §1 · p. 2
- 0.811 AUROC and 0.765 accuracy on the challenge test set; the approach improved all benchmark models and remained behind human accuracy.The authors placed third among 3,173 participants.Source: §1 · p. 2
2 Problem Statement
The problem is binary classification of labeled memes as hateful or not hateful, evaluated within a dataset built for fine-tuning and testing pre-trained multimodal models.
2 Problem Statement
- The dataset contains 8,500 training, 500 dev, and 1,000 seen-test samples, plus a 2,000-sample unseen test set used to determine winners.The 10K-image dataset is intended for fine-tuning and testing large-scale pre-trained multimodal models, not training from scratch.Source: §2 · p. 2
- Performance is measured by area under the receiver operating characteristic curve (ROC AUC).Source: §2 · p. 2
- Hateful versus not-hateful labels make the task a binary classification problem.Source: §2 · p. 2
3 Methods
The solution fine-tunes pre-trained VisualBERT on an expanded meme dataset, encoding image regions alongside language and combining selected models with majority voting.
3 Methods
- The approach uses VisualBERT, a multimodal vision-and-language BERT, and divides the solution into dataset expansion, image encoding, training, and ensemble learning.Source: §3 · p. 2
3.1 Dataset Expansion
- The training set was expanded by 428 memes, including 100 samples absent from the unseen dev set and 328 cherry-picked, manually relabeled Memotion memes.The authors report that most Memotion samples were wrongly labeled and selected additions similar in style and design to challenge memes.Source: §3.1 · pp. 2–3
3.2 Image Encoding
- Each image is encoded as 100 Mask-RCNN region features projected into the textual token-embedding space before the Transformer.The features are 2048-dimensional; Figure 1 depicts image regions and language combined so self-attention can discover implicit alignments.Source: §3.2 · p. 3, Figure 1 · p. 3, Figure 2 · p. 4
3.3 Training
- The authors use VisualBERT pre-trained on Conceptual Captions, fine-tune it on the aggregated training set, and classify from the final layer’s first output.They report that the Conceptual Captions-pretrained model scored noticeably better in their experiments than the COCO-pretrained version; training uses softmax and binary cross-entropy.Source: §3.3 · p. 3
3.4 Ensemble Learning
- The ensemble selects the top 27 models by dev-unseen AUROC and assigns labels by majority vote.For AUROC calculation, it uses the maximum model probability when the vote is hateful and the minimum when it is not hateful; Table 1 identifies the ensemble models’ performances.Source: §3.4 · p. 4, Table 1 · p. 4
4 Experiments and Results
Majority Voting improved both AUROC and accuracy by 2.5%; the authors describe it as combining multiple weaker models into one strong model.
5 Conclusion
The approach fine-tunes pretrained VisualBERT on expanded training data and applies Majority Voting over the 27 best models. It placed third among 3,173 participants, achieving 0.811 AUROC and 0.765 accuracy on the challenge test set while remaining far behind human judgment.
- 0.811 AUROC and 0.765 accuracy were achieved on the challenge test set.Source: §5 · p. 5
- Third place among 3,173 participants was achieved in the Hateful Memes Challenge.Source: §5 · p. 5
- The approach fine-tunes pretrained VisualBERT on expanded training data, then applies Majority Voting over the 27 best models.Source: §5 · p. 5