The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes

By Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, Davide Testuggine. 2020

939 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Social Psychology, Multimodal Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Social Psychology

In Multimodal Learning

Citations in the corpus

HateCheck: Functional Tests for Hate Speech Detection Models (2020, background): “and multimodal content, such as images combined with text, are common on social media and constitute a current research frontier in hate speech detection (Vidgen et al., 2019; Kiela et al., 2020).”
Anticipating Safety Issues in E2E Conversational AI: Framework and Tooling (2021, background): “Multimodal context has shown to be especially beneﬁcial in cases where the meaning is subtle and/or compositional, such as in the HatefulMeme challenge (Kiela et al., 2020) or detecting inappropriate video content as in the MOCHA challenge (Escalante et al., 2021).”
A Multimodal Framework for the Detection of Hateful Memes (2020, methodology): “Late-fusion methods [14, 15] typically utilize unimodal models to process the two signals independently and then combine their features (usually via simple concatenation) before the final classification layer.”
Flamingo: a Visual Language Model for Few-Shot Learning (2022, methodology): “VisDial [20], a visual dialogue benchmark; HatefulMemes [54] a vision and text classification benchmark; NextQA [129] which specially focuses on causality and temporal relation; STAR [128], a multiple-choice question answering task; and RareAct [73], a benchmark measuring compositionality in action ”
MAGMA - Multimodal Augmentation of Generative Models through Adapter-based Finetuning (2021, methodology): “…of structured, task-based datasets [42,56], we also opt to include the training sets of various downstream tasks – adding the train sets of VQA [2], GQA [23], OKVQA [31], VizWiz [17], Hateful Memes [25], and CoCo Captions [9], resulting in a ﬁnal dataset of around 25 million image-text pairs.”
Language Is Not All You Need: Aligning Perception with Language Models (2023, background): “…entailment Accuracy ✓ ✓ COPA [33] Causal reasoning Accuracy ✓ ✓ Rendered SST-2 [34] OCR-free sentiment classification Accuracy ✓ HatefulMemes [35] OCR-free meme classification ROC AUC ✓ Cross-modal transfer RelativeSize [36] Commonsense reasoning (object size) Accuracy ✓ MemoryColor [37]…”
Handling Bias in Toxic Speech Detection: A Survey (2022, background): “Meanwhile, the usage of memes and videos (short clips and long ones) spreading toxic and harmful content has been gaining momentum [50, 72, 73].”
MMICL: Empowering Vision-language Model with Multi-Modal In-Context Learning (2023, methodology): “Few-shot Image Classification includes the HatefulMemes (Kiela et al., 2020) and Bonard-HOI (Jiang et al., 2022) dataset.”
Exploring Visual Engagement Signals for Representation Learning (2021, background): “Hateful Memes Hateful Memes dataset [40] contains multimodal memes including images and text.”
What matters when building vision-language models? (2024, neutral): “General visual question answering VQAv2 (Goyal et al., 2017) 82,772 443,757 1,595,929 5.72% COCO-QA (Ren et al., 2015) 46,287 78,736 286,982 1.47% Visual7W (Zhu et al., 2016) 14,366 69,817 279,268 1.43% A-OKVQA (Schwenk et al., 2022) 16,539 17,056 236,492 1.21% TallyQA (Acharya et al., 2019) 98,680 ”
Generative Multimodal Models are In-Context Learners (2023, methodology): “We evaluate zero-shot and few-shot abilities of Emu2 on OKVQA [53], VQAv2 [28], VizWiz [29], TextVQA [68], and HatefulMemes [35] tasks.”
Preserving integrity in online social networks (2020, methodology): “To address some of these issues, companies have invested significant effort to create datasets that can be used for research purposes, such as the Deepfake Detection Challenge [31] and the Hateful Memes Dataset [56].”
Vision-Language Models for Vision Tasks: A Survey (2023, background): “• Hateful Memes [27] has been proposed for hateful meme classification ( i.e. , image with text) created by Facebook AI.”