PaperlayerView exact revision on arXiv

Source-linked AI summary

Disentangling Hate in Online Memes

Rui Cao, Ziqing Fan, Roy Ka-Wei Lee, Wen-Haw Chong, Jing Jiang

arXiv:2108.06207v1cs.IRcs.CLcs.MM

TL;DR

Hateful content detection has largely focused on text, although memes combine text and images to convey hate. DisMultiHate disentangles target-entity representations in memes, and it outperformed state-of-the-art baselines on two datasets while case studies demonstrated target identification.

  • Problem

    Hateful memes combine text and images to communicate hate, but existing multimodal classifiers have limited explainability and cannot identify targets or reason about meme context.

    Source: §1 · p. 1, §2.2 · p. 2

  • Method

    DisMultiHate uses self-supervised training to disentangle target-entity representations in memes and uses them as contextual information for classification.

    Source: §1 · p. 2, §4 · p. 3, §4.2 · p. 5

  • Results

    DisMultiHate outperformed state-of-the-art baselines on both datasets, beating the best baseline by 4% (Acc) on FHM and 5% (F1) on MultiOFF.

    Source: §5.2 · p. 7

  • Takeaways & Limitations

    Case studies demonstrated that DisMultiHate can identify target entities and provide contextual explanations for hateful-meme classification results.

    Source: §1 · p. 2

    FHM error analysis found potential dataset noise and a meme labeled non-hateful despite content the authors interpreted as discriminatory toward Catholics and a priest.

    Source: §5.4 · p. 8

Abstract

from arXiv · show

Hateful and offensive content detection has been extensively explored in a single modality such as text. However, such toxic information could also be communicated via multimodal content such as online memes. Therefore, detecting multimodal hateful content has recently garnered much attention in academic and industry research communities. This paper aims to contribute to this emerging research topic by proposing DisMultiHate, which is a novel framework that performed the classification of multimodal hateful content. Specifically, DisMultiHate is designed to disentangle target entities in multimodal memes to improve hateful content classification and explainability. We conduct extensive experiments on two publicly available hateful and offensive memes datasets. Our experiment results show that DisMultiHate is able to outperform state-of-the-art unimodal and multimodal baselines in the hateful meme classification task. Empirical case studies were also conducted to demonstrate DisMultiHate's ability to disentangle target entities in memes and ultimately showcase DisMultiHate's explainability of the multimodal hateful content classification task.

1 INTRODUCTION

Hateful memes can communicate discriminatory messages through the interaction of image and text, while existing multimodal models have limited ability to explain their reasoning or identify targets. DisMultiHate disentangles target-entity representations to improve classification and explainability, and the paper reports stronger benchmark results and case-study evidence of target identification.

1 INTRODUCTION

  • Existing hateful-content research has predominantly focused on text, despite memes being popular, shareable image-and-text content increasingly used to spread hate.The passage notes that memes can be reposted across conversations and contexts.Source: §1 · p. 1
  • Existing multimodal methods combine text and visual features or fine-tune pretrained models, but have limited explainability and cannot reason about meme context.The paper frames target-entity context as important to understanding hateful messages.Source: §1 · pp. 1–2
  • Hateful memes can appear benign when text and image are examined separately, while their combined message may target entities such as gender and religion.The example is described as targeting a female Muslim.Source: §1 · p. 2
  • DisMultiHate outperformed state-of-the-art baselines on both datasets, and case studies demonstrated its ability to identify target entities and provide contextual explanations.The framework disentangles representations of hate-related targets such as race and gender.Source: §1 · p. 2

2 RELATED WORK

Prior hate-speech detection largely focused on text, while multimodal meme methods improved coverage but remained limited in explaining targets and context. This paper applies representation disentanglement to target entities in hateful memes.

2.1 Hate Speech Detection

  • Text-based hate-speech detection left rich multimedia content relatively neglected, motivating work on multimodal hateful memes.Multimodal research grew alongside the release of hateful-meme datasets, including a 10K-meme challenge dataset constructed so unimodal methods would not perform well.Source: §2.1 · p. 2, §2.2 · p. 2

2.2 Multimodal Hate Speech

  • Multimodal meme methods use early fusion or pretrained models, but often cannot identify hate targets or provide context; DisMultiHate learns target representations with self-supervision.Unlike [58], the paper extracts target entities through disentangled latent representations rather than augmenting general entity information as input.Source: §2.2 · p. 2

2.3 Disentangling Representation Learning

  • Although disentangled representations have been applied to recommendation and visual semantics, this paper applies them to target entities in hateful memes.The authors describe this as the first application of representation disentanglement to hateful-meme classification.Source: §2.3 · p. 3

3 PRELIMINARIES

The paper frames hateful-meme classification as predicting a binary label from an image and its text. It evaluates on FHM and MultiOFF, using preprocessing to add entity and demographic information to the meme inputs.

3.1 Problem Definition

  • Given an image and a sequence of words, the task predicts whether their multimodal meme is hateful or non-hateful.The model may predict a confidence score from zero to one and classify the meme as hateful when it exceeds threshold λ.Source: §3.1 · p. 3

3.2 Datasets

  • Evaluation uses FHM, a 10K-meme binary-label dataset, and MultiOFF, a 1K-meme dataset labeled offensive or non-offensive.Because FHM test labels were unavailable, the authors use dev-seen as test; they also consider MultiOFF offensive memes hateful because most contain abusive and discriminatory messages.Source: §3.2 · p. 3

3.3 Data Pre-processing

  • Preprocessing resizes images, extracts and removes meme text, detects entities in cleaned images, and adds demographic labels for images containing human entities.The demographic labels come from FairFace, mapped to the person’s bounding box with the largest overlap with a detected face.Source: §3.3 · p. 3

4 PROPOSED MODEL

DisMultiHate has preprocessing, text-representation, and visual-representation modules that disentangle hate-target information across modalities. A self-supervised matching loss aligns the text and visual target representations, and the resulting representations support hateful-meme classification.

4 PROPOSED MODEL

  • DisMultiHate projects text and visual representations into a latent space whose units encode probabilities for hate-target categories, then uses a self-supervised matching loss to align the modalities.The learned representations are fed to a regression layer to predict whether a meme is hateful.Source: §4 · pp. 3–4

4.1 Text Representation Learning

  • The text module combines meme text, detected entities, and demographic information, then uses BERT and straight-through Gumbel-Softmax to learn a target-focused representation.The representation is used in matching with the visual representation and is fine-tuned for classification.Source: §4.1 · pp. 4–5

4.2 Visual Representation Learning

  • The visual module attends to image-region features using textual representations as guidance, then matches its disentangled target representation to the text representation.The matching loss supervises visual target disentanglement; the updated attended visual representation is used for classification.Source: §4.2 · p. 5

4.3 Classification

  • The classification module concatenates the sentence and attended-image representations, predicts a hateful score with a regression layer, and combines prediction and matching losses during training.A score above the 0.5 threshold is classified as hateful.Source: §4.3 · pp. 5–6

4.4 Implementation Details

  • Implementation uses 2048-dimensional Faster R-CNN image features, BERT-tokenized text sequences of length 64, minibatches of 64, and AdamW optimization.The matching-loss weight μ is 0.05 on FHM and 0.03 on MultiOFF.Source: §4.4 · p. 6

5 EXPERIMENT

The experiment section evaluates DisMultiHate against baselines, then reports quantitative results, qualitative target-retrieval cases, and classification errors and limitations.

5.1 Evaluation Setting

  • The two public datasets use different metrics: FHM uses AUROC and accuracy, while MultiOFF uses weighted F1, precision, and recall; comparisons include multimodal and BERT text-only baselines.Source: §5.1 · p. 6

5.2 Experimental Results

5.3 Case Study

  • Target-query retrieval compares disentangled text and visual representations; retrieved examples used cues such as “dark” and “pick cotton” to identify “Black Man” in grayscale imagery.Source: §5.3 · p. 7, Table 6 · p. 8

5.4 Error Analysis

  • Errors included a false positive linked to “black” plus target “woman,” FHM annotation and noise cases, and a miss requiring deeper reasoning than text or image features supplied.Source: §5.4 · p. 8

6 CONCLUSION

DisMultiHate disentangles hate-related target representations in memes and outperformed state-of-the-art baselines on two public datasets. Case studies demonstrate target disentanglement; future work will add advanced reasoning and evaluate more datasets.

6 CONCLUSION

  • DisMultiHate disentangles hate-related target representations in memes to improve hateful content classification.The paper reports that experiments on two public datasets showed it outperformed state-of-the-art baselines.Source: §6 · p. 8
  • Case studies empirically demonstrate DisMultiHate’s ability to disentangle target information in memes.Source: §6 · p. 8
  • The authors identify advanced reasoning and evaluation on more hateful meme datasets as future work.Source: §6 · p. 8
  • The authors hope target disentanglement will raise awareness of vulnerable groups targeted in real-world hate speech datasets.Source: §6 · p. 8

Related papers

Loading 2108.06207v1…