University of Hildesheim at SemEval-2022 task 5: Combining Deep Text and Image Models for Multimedia Misogyny Detection

Milan Kalkenings, Thomas MandlNAACL 2022

TL;DR

This paper tackles the challenging task of detecting misogynistic memes by combining deep learning models for both text and image analysis. The researchers found that text-based features, particularly those enriched with semantic associations, were crucial for accurate detection, outperforming image-only approaches.

What Problem Does This Paper Address?

Hate speech, and specifically misogyny, is a pervasive problem on social media. While identifying hateful text has seen progress, the rise of visual content, especially memes, presents a new frontier for detection. Memes often combine images and text in nuanced ways, making it difficult for traditional text-only or image-only systems to accurately identify harmful content. The SemEval-2022 Task 5 (MAMI) aimed to address this by providing a benchmark for multimedia misogyny identification, pushing researchers to develop systems that can understand the interplay between visual and textual elements in memes.

Methods

The University of Hildesheim team developed a multimodal system by combining six individual models. Their approach involved:

  • Text Analysis:
    • BERT: A standard BERT model was used to create transformer-based representations of the meme's text.
  • Semantic Associations: To capture nuanced meanings, they manually constructed a knowledge graph of associations for words commonly found in memes. For each word in a meme's text, they looked up related terms in this graph, enriching the textual representation with world knowledge and potential prejudices.
  • Sentiment Analysis: VADER and TextBlob were employed to extract sentiment and subjectivity scores, hypothesizing that highly emotional texts might correlate with hateful content.
  • Image Analysis:
    • Faster R-CNN: This object detection model was used to identify key objects within the meme's image. The visual features of these detected objects were then fed into a ResNet system.
  • Tile Approach: The image was divided into 196 tiles, each analyzed by a CNN. These features were then processed to create an image embedding.
  • Fusion System: A transformer-based fusion system was used to integrate the embeddings from the individual text and image models. They employed Sequential Forward Selection (SFS) and Sequential Backward Elimination (SBE) to find the optimal combination of models for the final prediction.

Datasets

The study utilized two primary datasets:

  • SemEval-2022 Task 5 (MAMI): This dataset specifically focuses on multimedia misogyny identification.
  • Facebook Hateful Memes Challenge (HM) dataset: This dataset, while broader in scope (hateful memes in general), was used to test the generalization capabilities of their multimodal system.

Key Insights and Ideas

  • Text is King (for Misogyny): The most striking finding is the dominance of text-based features for misogyny detection. Even with sophisticated image analysis, text alone, especially when augmented with semantic associations, performed remarkably well.
  • Semantic Associations Matter: The manual construction of a knowledge graph to capture word associations proved to be a valuable technique. This suggests that understanding the implicit meanings and potential biases associated with words is crucial for identifying subtle forms of misogyny.
  • Multimodal Fusion Improves Performance: While text was dominant, combining it with other modalities (especially object detection and associations) did lead to improved scores compared to single-modality approaches, indicating the potential of multimodal systems.
  • Generalization is a Challenge: Performance on the broader Facebook Hateful Memes dataset was not as strong as on the MAMI dataset, highlighting the difficulty of creating systems that generalize well across different types of hateful content and datasets.

Key Results

  • Individual Model Performance: The text classification system achieved the highest AUC-ROC score (0.6617) among individual models. The associations system also performed strongly (0.6588).
  • Optimal Fusion: For the MAMI dataset, the optimal fusion system using SFS combined text, object detection, and associations, achieving an AUC-ROC of 0.8509.
  • Cross-Dataset Performance: The best overall performance on the MAMI dataset was 0.8421 (AUC-ROC), and on the HM dataset was 0.7146 (AUC-ROC).
  • Impact of Duplicates: Removing memes with identical text improved performance, while removing memes with identical images decreased it, underscoring the importance of textual content.

Limitations and Future Work

  • Manual Association Construction: The manual creation of the knowledge graph for semantic associations is labor-intensive and may not scale easily. Future work could explore automated methods for building such knowledge bases.
  • Generalization: The performance gap between the MAMI and HM datasets indicates that further research is needed to improve the generalization capabilities of multimodal hate speech detection systems.
  • Impact of Individual Components: The authors plan to further analyze the specific contribution of each individual system to the overall fusion performance.
  • Specialized Datasets: They intend to explore whether training with a misogyny-specific text dataset can further benefit a multimodal system.

Summary updated

BibTeX
@inproceedings{kalkenings2022university,
  author = {Milan Kalkenings and Thomas Mandl},
  title = {{University of Hildesheim at SemEval-2022 task 5: Combining Deep Text and Image Models for Multimedia Misogyny Detection}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
}
Figure