“Female Astronaut: Because sandwiches won’t make themselves up there”: Towards Multimodal misogyny detection in memes

Smriti Singh, Amritha Haridasan, Raymond MooneyACL 2023

TL;DR

This paper tackles the challenging problem of automatically detecting misogynistic memes, which combine text and images to convey offensive messages. The authors propose a multimodal model, BERT*+ViT, that leverages a BERT model pre-trained on hateful memes and a Vision Transformer (ViT), achieving state-of-the-art performance. Their work also provides a crucial qualitative analysis, shedding light on why current models struggle with this task.

What Problem Does This Paper Address?

Memes have become a powerful tool for communication and shaping beliefs, but they also serve as a vehicle for harmful content, including misogyny. Detecting misogyny in memes is particularly challenging because it's "multimodal," meaning it relies on the interplay between text and images, often incorporating sarcasm and subtle visual cues. Traditional text-based hate speech detection methods are insufficient. While a dataset called MAMI exists for this task, current models still struggle to accurately identify and classify misogynistic memes, especially those with nuanced or subtle offensive content. This paper aims to not only improve detection performance but also understand the limitations of existing models and the benefits of domain-specific pre-training.

Methods

The researchers experimented with several models, categorizing them into unimodal (text-only or vision-only) and multimodal approaches.

  • Unimodal Models: They fine-tuned standard language models like BERT, DeBERTa, and RoBERTa on the meme text, and vision models like CNN, ViT, and Inception on the meme images.
  • Multimodal Models: They explored combinations of text and vision encoders. This included:
  • CLIP: A model known for its ability to understand relationships between images and text.
  • BERT + Inception: Combining BERT for text with Inception for images.
  • BERT + ViT: Combining BERT for text with Vision Transformer for images.
  • VisualBERT: A model designed to integrate visual and textual features.
  • BERT* + ViT: This is their proposed novel approach. It uses a BERT model that has been specifically pre-trained on a dataset of hateful memes (hence BERT*), combined with a Vision Transformer (ViT). This domain-specific pre-training is a key differentiator.

The models were fine-tuned on the MAMI dataset, and their performance was evaluated using precision, recall, and F1 scores. A significant part of their methodology involved a qualitative analysis of misclassified memes to understand model failures.

Datasets

The primary dataset used in this research is the MAMI dataset.

  • MAMI Dataset: This dataset contains 12,000 memes, annotated for misogyny and four sub-classes: shaming, objectification, violence, and stereotype. It's balanced for misogynous and non-misogynous content, with 10,000 samples for training and 1,000 each for testing and validation. The authors note that while inter-annotator agreement for misogyny labeling was moderate, agreement for sub-classification was fair, indicating the inherent difficulty of the task.

Key Insights and Ideas

  • Domain-Specific Pre-training is Crucial: The most significant finding is that pre-training a language model (BERT) on a related task (hateful memes) before fine-tuning it on the specific misogyny detection task dramatically improves performance. This suggests that understanding general hateful content helps in identifying specific forms like misogyny.
  • Multimodality Matters, But Not Always: While multimodal models generally outperform unimodal ones, some combinations didn't yield significant improvements, and in some cases, even performed worse than strong unimodal models. This highlights that simply combining modalities isn't a silver bullet; the architecture and how modalities are fused are critical.
  • Visual Grounding Isn't Enough: The qualitative analysis revealed that models often fail when the misogyny relies on understanding cultural stereotypes or subtle contextual cues that go beyond simply recognizing objects or text. The model needs to grasp the underlying social commentary or implied meaning.
  • Subtlety is the Enemy: Memes with explicit objectification or violence are easier to detect. The real challenge lies in identifying subtle forms of misogyny, particularly those involving nuanced linguistic innuendos or less overt objectification, which current models frequently miss.

Key Results

  • State-of-the-Art Performance: The proposed BERT*+ViT model achieved the best results, outperforming existing state-of-the-art models by over 10% on Task A (misogynous vs. non-misogynous classification) and by 2% on Task B (sub-classification of misogyny).
  • Task A Performance: BERT*+ViT achieved an F1 score of 0.874 on Task A.
  • Task B Performance: BERT*+ViT achieved an F1 score of 0.721 for the Violence sub-class, and generally strong performance across other sub-classes.
  • Qualitative Findings: The analysis of 200 misclassified memes revealed that approximately 80 were due to insufficient "visual grounding" (lack of understanding stereotypes), and about 15 lacked sufficient context for even humans to easily classify. Subtle objectification was also a significant failure point.

Limitations and Future Work

  • Limitations:
    • The current models struggle with subtle misogyny, stereotypes, and context-dependent humor.
  • Some multimodal models perform worse than unimodal ones, indicating a need for better fusion strategies.
  • The dataset's fair agreement on sub-classification suggests the inherent difficulty of the task.
  • Future Work:
    • Extend the dataset to multiple languages to test generalizability.
  • Evaluate the approach on other multimodal media like reels and TikToks.
  • Reduce the computational complexity of training and deployment.
  • Investigate the interpretability of multimodal cues to understand why certain memes are misogynistic.
  • Further pre-training/fine-tuning on larger datasets could improve performance on complex cases.

Summary updated

BibTeX
@inproceedings{singh2023female,
  author = {Smriti Singh and Amritha Haridasan and Raymond Mooney},
  title = {{“Female Astronaut: Because sandwiches won’t make themselves up there”: Towards Multimodal misogyny detection in memes}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure