TL;DR
This paper tackles the challenging task of automatically identifying misogynistic content in online memes, which combine images and text. The authors enhance the powerful UNITER multimodal model by incorporating image sentiment analysis and graph convolutional networks (GCNs) to capture richer contextual information, achieving competitive results in the SemEval-2022 MAMI task.
What Problem Does This Paper Address?
The proliferation of social media has led to an explosion of online content, including hate speech and offensive language. Misogyny, a specific form of hate speech targeting women, is a significant concern, yet it has received less attention than broader hate speech detection. Memes, a popular form of online communication, present a unique challenge for detection because their meaning is often derived from the interplay between an image and its accompanying text, making them difficult for unimodal (text-only or image-only) systems to understand. This paper addresses the need for robust, multimodal systems capable of identifying misogynistic memes by leveraging both visual and textual cues.
Methods
The core of the proposed approach is the pre-trained multimodal UNITER model. The authors explore two primary enhancements:
- Visual Sentiment-enhanced UNITER: This method augments UNITER by adding an image sentiment classifier. The intuition is that misogynistic content often correlates with negative sentiments, and by explicitly incorporating image sentiment, the model can better grasp the overall tone of the meme. A pre-trained VGG-19 model, fine-tuned on a sentiment dataset, is used for this purpose.
- VGCN-enhanced UNITER: This approach leverages graph convolutional networks (GCNs) to incorporate external knowledge. A heterogeneous graph is constructed where nodes represent text tokens and objects detected in the image. This graph is built from a large dataset of meme templates and their captions, capturing co-occurrence relationships. A GCN is then applied to this graph to generate embeddings that are fused with UNITER's existing representations. This aims to provide a richer understanding of meme content beyond what's directly present in the image-text pair.
An ensemble model is also developed, combining multiple versions of the two proposed architectures trained on different data splits using a soft voting scheme. This ensemble aims to smooth out performance variations and improve overall robustness.
Datasets
The primary dataset used for training and evaluation is from the SemEval-2022 Task 5: MAMI (Multimedia Automatic Misogyny Identification) shared task. This dataset consists of memes with associated text captions, labeled for misogynous content.
- Sub-task A: Binary classification of misogynous vs. non-misogynous memes.
- Sub-task B: Multi-label classification distinguishing between types of misogyny: stereotype, shaming, objectification, and violence.
An external dataset of 3,000 meme templates and their captions was used to build the graph for the VGCN-enhanced model.
Key Insights and Ideas
- Multimodality is Crucial: The paper strongly emphasizes that understanding memes requires processing both visual and textual information. Seemingly benign text can become offensive when paired with a specific image, and vice-versa.
- External Knowledge Enhances Detection: Incorporating information from a graph built on external meme data (VGCN) proved beneficial, suggesting that understanding meme culture and common associations can improve misogyny detection.
- Sentiment as a Signal: Explicitly adding image sentiment analysis to the model helped capture an important affective dimension of the memes, which is often correlated with offensive content.
- Ensembling for Robustness: Combining multiple models, even those with similar architectures but trained on different splits, effectively smoothed out performance variations and led to the best overall results.
- Data Cleaning Trade-offs: While removing website watermarks and timestamps from text might seem like good practice for generalization, it surprisingly decreased performance, indicating these elements can sometimes serve as useful clues for misogyny detection in this specific dataset.
Key Results
The UPB team achieved competitive results in the SemEval-2022 Task 5: MAMI. Their best model, the ensemble, achieved:
- Sub-task A (Binary Identification): An F1-score of 71.4%.
- Sub-task B (Type Identification): An F1-score of 67.3%.
These scores placed their team in the upper third of the leaderboard, demonstrating the effectiveness of their multimodal approach. The paper also presents detailed results for individual models and ensemble components, showing variations in performance across different configurations.
Limitations and Future Work
- Over-prediction Tendency: The ensemble model showed a tendency to over-predict misogyny, particularly in cases of aggressive language or when visual cues were ambiguous.
- Bias in Training Data: The authors note that imbalances in the training data, such as the disproportionate appearance of certain visual objects ("woman," "eyebrows") in misogynous memes, can lead to misclassifications.
- Contextual Nuance: Memes requiring deep background knowledge or subtle cultural understanding remain challenging to detect.
- Future Directions: The authors plan to continue research on detecting abusive and hateful content, explore other pre-trained multimodal models, and investigate task-adaptive pre-training. They also suggest extending the research to other modalities like video and audio to gain a broader understanding of toxic content spread.
Summary updated