TL;DR
This paper benchmarks various approaches for detecting hateful memes, a challenging multimodal task. They found that while text-based BERT models perform well, an ensemble of text and image BERT models achieves the best results, highlighting the importance of multimodal understanding.
What Problem Does This Paper Address?
Hateful memes, which combine images and text to spread hate speech, are a growing concern online. Detecting them is difficult because the hateful message often arises from the interplay between the visual and textual elements, making it hard for models that only analyze one modality to succeed. The Hateful Memes Challenge aims to push the boundaries of multimodal hate detection by providing a dataset and tasks that require understanding both image and text. This paper addresses the need to benchmark different approaches for this task, comparing text-only, image-text fusion, and retrieval-based methods.
Methods
The authors experimented with several methods:
- Retrieval-based methods (MNN): These methods use a knowledge base (the training dataset) to find similar memes and transfer their labels. They explored two variants: one using FastText embeddings for text and DenseNet for images (MNN:base), and another using BERT embeddings for text (MNN:BERT).
- BERT-based methods:
- TxtBERT: A standard BERT model trained on text only.
- CaptionBERT: Uses text and automatically generated image captions.
- ImgBERT: Combines TxtBERT with image embeddings from a CNN.
- Retrieval Augmented Classification (RAC): This approach augments BERT models with retrieved text. TxtRAC combines TxtBERT with retrieved text, while Txt+Img RAC combines ImgBERT with retrieved text.
- Ensemble: An ensemble of TxtBERT and ImgBERT was created by averaging their prediction scores.
Datasets
The primary dataset used is the Hateful Memes Challenge dataset, originally introduced by Facebook AI. This dataset consists of image-text pairs annotated for hate speech, protected categories (race, disability, religion, nationality, sex), and attack types (contempt, mocking, inferiority, slurs, exclusion, dehumanizing, inciting violence). The authors created their own training, validation, and development sets by splitting the provided data, and used a separate unseen test set for final evaluation.
Key Insights and Ideas
- Text is powerful, but not enough: Text-based BERT models (TxtBERT) achieve surprisingly strong performance, indicating the significant role of textual content in identifying hateful memes.
- Image-text synergy is crucial: While text-only models are good, combining text with image information (ImgBERT) or using automatically generated captions (CaptionBERT) can further improve performance, especially when the hateful message is derived from the image-text interaction.
- Retrieval is tricky: Simple retrieval methods (MNN) and even retrieval-augmented BERT models (TxtRAC, Txt+Img RAC) did not consistently outperform simpler fusion methods. This suggests that the complexity and diversity of the dataset might make it difficult for retrieval to reliably find relevant information.
- Ensembling wins: An ensemble of TxtBERT and ImgBERT achieved the best overall performance in terms of ROC AUC on two out of three tasks, demonstrating the benefit of combining diverse model predictions.
Key Results
- The ensemble of TxtBERT and ImgBERT achieved the best ROC AUC scores for the hate detection task and the attack type task on their development set.
- TxtBERT was the best performing model for the protected category task on the development set.
- On the hidden test set, TxtBERT achieved the best score for the protected category task, while the Ensemble was best for the attack type task.
- ImgBERT outperformed TxtBERT in only 8% of hateful memes on the development set, suggesting that the image alone often doesn't add significant value beyond what text provides, unless the text is benign and the image carries the hateful message.
Limitations and Future Work
- Imbalanced dataset: The dataset is imbalanced, with a majority of non-hateful memes. The models struggled to achieve high performance on the "empty" (non-hateful) labels compared to other labels.
- Retrieval challenges: The authors note that retrieval methods did not perform as well as expected, possibly due to the complex context and diversity of the meme texts.
- Ensemble limitations: The authors acknowledge that their ensemble method might not be optimal and suggest exploring checkpoint-based ensembles for a fairer comparison.
- Future work: The authors plan to experiment with larger pre-trained vision and language transformer models, explore different retrieval sources, and investigate explainability approaches for multimodal methods.
Summary updated