Racist or Sexist Meme? Classifying Memes beyond Hateful

Haris Bin Zia, Ignacio Castro, Gareth TysonACL 2021

TL;DR

This paper tackles the challenge of automatically detecting hateful content in internet memes, which often combine images and text to convey offensive messages. They introduce a multimodal pipeline that leverages pre-trained visual and textual embeddings to identify the protected category being attacked (e.g., race, sex) and the type of attack (e.g., contempt, slurs), achieving impressive AUROC scores of 0.96 and 0.97 respectively.

What Problem Does This Paper Address?

Internet memes, while often humorous, are increasingly being used to spread hate speech. Detecting this "hatefulness" is complex because it relies on the interplay between visual and textual elements, making it a multimodal problem. Manually moderating the sheer volume of memes online is infeasible, necessitating automated solutions. Previous work has struggled with the nuances of multimodal hate speech, and this paper aims to improve detection by developing a pipeline that can identify not just if a meme is hateful, but what specific protected group it targets and how it attacks them.

Methods

The authors propose a straightforward yet effective multimodal pipeline.

  1. Text Extraction: The text content is first extracted from the meme image.
  2. Embedding Generation: Both the meme image and its extracted text are then processed through state-of-the-art pre-trained encoders:
  3. Visual: OpenAI's CLIP image encoder (CIMG).
  4. Textual: CLIP text encoder (CTXT), Facebook's LASER, and Google's LaBSE.
  5. Concatenation: The resulting visual and textual embeddings are concatenated to create a combined representation.
  6. Classification: A simple multi-label Logistic Regression classifier is trained on these concatenated embeddings to predict two tasks:
  7. Task A: The protected category attacked (e.g., race, sex, religion, disability, nationality, or pc_empty for non-hateful memes).
  8. Task B: The type of attack (e.g., contempt, mocking, inferiority, slurs, exclusion, dehumanizing, inciting violence, or attack_empty for non-hateful memes).

The novelty lies in the effective combination of these powerful pre-trained multimodal representations with a simple classifier, demonstrating that sophisticated feature extraction can compensate for a less complex classification model.

Datasets

The primary dataset used is the Hateful Memes Challenge dataset (Kiela et al., 2020). This dataset was augmented with newly created fine-grained labels for protected categories and attack types, making it suitable for the authors' specific tasks. The dataset consists of approximately 9,540 annotated memes. The authors used the train split for training, dev (seen) for hyperparameter tuning, and dev (unseen) for reporting results. They also mention evaluation on a blind test set.

Key Insights and Ideas

  • Multimodality is Crucial: The paper strongly supports the idea that understanding hateful memes requires considering both visual and textual information. Unimodal approaches (vision-only or text-only) perform significantly worse than multimodal ones.
  • Pre-trained Embeddings are Powerful: Leveraging state-of-the-art pre-trained models like CLIP, LASER, and LaBSE provides highly effective representations that capture rich semantic and visual information, enabling good performance even with a simple classifier.
  • Simplicity Can Be Effective: The authors demonstrate that a straightforward logistic regression classifier, when fed with rich multimodal embeddings, can achieve performance comparable to or better than more complex models, highlighting the importance of feature engineering.
  • Fine-grained Labels Enhance Detection: Moving beyond a simple "hateful vs. not-hateful" classification to identifying specific protected categories and attack types provides a more nuanced and actionable understanding of online hate.

Key Results

The multimodal pipeline significantly outperforms unimodal approaches.

  • The best performing model, using a combination of CIMG, CTXT, LASER, and LaBSE embeddings, achieved an AUROC of 0.96 for identifying the protected category (Task A) and an AUROC of 0.97 for detecting the type of attack (Task B) on the dev (unseen) split.
  • Even unimodal classifiers showed strong performance, with the vision-only CIMG classifier achieving an AUROC of 0.93 for Task A and 0.94 for Task B, outperforming text-only models.
  • On a blind test set of 1000 memes, the best model achieved an AUROC of 0.90 for Task A and 0.91 for Task B.

Limitations and Future Work

  • Language Specificity: While the embeddings used (LASER, LaBSE) are language-agnostic, the current evaluation was performed solely on a dataset of English memes.
  • Future Work: The authors plan to extend their work to multilingual settings, evaluating their classifier's performance on memes in various languages.

Summary updated

BibTeX
@inproceedings{zia2021racist,
  author = {Haris Bin Zia and Ignacio Castro and Gareth Tyson},
  title = {{Racist or Sexist Meme? Classifying Memes beyond Hateful}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure