TL;DR
This paper tackles the challenge of hate speech detection on social media by moving beyond text-only analysis to incorporate visual information from accompanying photos. They propose and evaluate various deep learning fusion techniques, demonstrating that combining text and image features significantly improves hate speech classification accuracy, with attention-based fusion yielding the best results.
What Problem Does This Paper Address?
Social media platforms are rife with user-generated content, which, while fostering connection, also presents a significant challenge in moderating harmful material like hate speech. Hate speech creates an intimidating and exclusionary environment, and in extreme cases, can incite real-world violence. Traditional approaches to hate speech detection have primarily focused on analyzing the textual content of posts. However, this is often insufficient because the meaning of text can be ambiguous or even benign in isolation, but become hateful when paired with an image. For instance, a seemingly innocuous phrase like "these are disgusting parasites" can be interpreted as hate speech if the accompanying photo depicts a specific group of people. This paper addresses the need for a more robust, multimodal approach to hate speech classification by exploring how to effectively fuse text and image information.
Methods
The authors propose a deep multimodal fusion framework that combines a convolutional neural network (CNN) for text classification with features extracted from images.
Text Processing: They utilize a CNN architecture inspired by Kim (2014) for sentence classification. This involves pre-trained word embeddings, a word-level MLP for domain adaptation, 1D convolutions to capture n-grams, and max-pooling to create a fixed-size text representation.
Image Processing: For image features, they pre-train a deep neural network (similar to ResNet) on a large dataset of social media photos. They then extract features from the second-to-last layer, resulting in a 4096-dimensional vector. To efficiently store and retrieve these features, they apply iterative quantization to convert them into a 256-bit binary hash. This hash preserves semantic similarity between images.
Fusion Techniques: The core of their contribution lies in exploring various fusion strategies:
- Basic Concatenation: The simplest approach, where the text representation vector and the photo hash are concatenated.
- Gated Summation: Two variants are explored:
- Simple-gated fusion: A gate mechanism controls the weighted summation of transformed text and photo features.
- Symmetric-gated fusion: An extension that dynamically controls the combination of visual and textual signals using a more complex gating mechanism.
- Bilinear Transformation: A method to integrate information from two vectors into a single vector using a learned weight matrix.
- Attention Mechanism: This technique allows the model to dynamically focus on relevant parts of the text based on the image query. They implement this using a context vector derived from the text's convolutional outputs and a query vector from the photo. They also experiment with:
- Sparsemax vs. Softmax: Comparing sparsemax, which produces sparser attention weights, against the standard softmax.
- Deep vs. Shallow Cloning: Investigating whether to back-propagate gradients through the attention mechanism to influence word embeddings (deep clone) or to stop back-propagation (shallow clone).
Datasets
The authors use seven months of user-reported data from a social network platform. This data is labeled as hate speech or benign based on platform community standards. The last month of data is used for testing, while the first six months are split into 90% training and 10% development sets for hyperparameter tuning and early stopping.
Key Insights and Ideas
- Visual Cues are Crucial: The paper strongly suggests that text alone is often insufficient for accurate hate speech detection, and incorporating visual information from accompanying photos provides a significant boost in performance.
- Attention is Powerful: Attention mechanisms, particularly when combined with deep cloning and sparsemax, offer a more nuanced way to fuse text and image information, leading to the best results. This suggests that the model can learn to focus on specific textual elements that are made salient by the image.
- Simple Fusion is a Good Start: Even a basic concatenation of text and image features provides a substantial improvement over text-only models, highlighting the immediate value of multimodal data.
- Bilinear Fusion is Costly and Not Always Better: While theoretically interesting, bilinear transformation proved to be computationally expensive and did not yield superior results, suggesting a trade-off between complexity and performance.
- Sparsemax for Focused Attention: The use of sparsemax in attention mechanisms appears to be beneficial, potentially by allowing the model to concentrate on more critical textual segments.
Key Results
- The baseline convolutional text model achieved an ROC-AUC of 82.1.
- Basic concatenation of text and photo features improved the ROC-AUC to 84.0.
- The best-performing model, using attention fusion with symmetric-gated fusion, deep cloning, and sparsemax, achieved an ROC-AUC of 84.8. This represents a significant improvement, especially considering that a 0.5 AUC improvement is considered production-quality significant.
- Bilinear transformation resulted in a lower ROC-AUC of 82.7 and was significantly slower (8-10x) than other approaches.
- Deep cloning in attention mechanisms generally outperformed shallow cloning.
Limitations and Future Work
Limitations: * The authors speculate that bilinear transformation might lead to overfitting, which they did not investigate further due to its computational cost. * The paper notes that the text data can be noisy, containing typos, misspellings, and out-of-vocabulary words, which can be a challenge for text-only models.
Future Work: * Investigating fusion with multiple photos per post. * Exploring fusion with other modalities, such as audio and video. * The authors suggest that character-level models could help alleviate issues with noisy text data.
Summary updated