TL;DR
This paper tackles the challenge of hate speech (HS) classification models failing to generalize across different datasets due to varying annotation schemes and data distributions. They propose an "extremely weak supervision" approach using only category names, which significantly improves cross-dataset transferability and reduces reliance on manual annotation.
What Problem Does This Paper Address?
Hate speech detection is a critical but challenging NLP task. A major hurdle is that supervised HS classification models trained on one dataset often perform poorly on others. This lack of generalizability stems from several issues:
- Definitional Ambiguity: What constitutes hate speech can vary based on cultural norms and individual interpretations.
- Annotation Challenges: Subjectivity, annotator bias, and differing annotation schemes lead to inconsistent labels, even for similar content.
- Data Distribution Mismatch: Datasets are often created using specific search queries or focusing on explicit hate speech, leading to a gap between training data and real-world, often more subtle, hate speech.
These problems make it difficult to compare models trained on different datasets and hinder their deployment in diverse real-world scenarios. The authors aim to create HS classification models that are more robust and adaptable to these variations.
Methods
The core of the paper's contribution lies in its adoption and adaptation of extremely weak supervision. Instead of relying on labeled examples, this approach uses only the category names (e.g., "Sexist," "Racist") as supervision signals.
The authors leverage and modify the X-Class model. X-Class works by:
- Estimating Category Representations: It iteratively expands category representations by finding words in an unlabeled dataset that are semantically similar to the category name.
- Document Representation: Each document is represented by weighting its words based on their similarity to the category representations.
- Pseudo-Labeling: A Gaussian Mixture Model (GMM) is used to cluster documents, and the most confident pseudo-labeled documents are used to train a final classifier.
The authors introduce a modification called X-Class Agree. This refinement addresses issues with GMM's global density estimation, particularly for low-frequency categories. X-Class Agree combines the GMM's assignment with a direct representation-based prediction, ensuring that a document is only pseudo-labeled if both methods agree, leading to more confident and accurate pseudo-labels.
Datasets
The experiments were conducted on two popular HS datasets:
- Waseem dataset: Contains tweets with sexist and racist content.
- SBIC dataset: A larger dataset with posts from various sources (Reddit, Twitter, hate sites) annotated with free-text target groups.
The authors carefully handled the relationship between the datasets' taxonomies, defining three cases: 1-to-1, N-to-1, and N-to-N mappings. For their experiments, they focused on scenarios where category mappings were either available or unavailable, particularly highlighting the N-to-N case where direct supervised learning is challenging.
Key Insights and Ideas
- Weak Supervision is Key for Generalization: Extremely weak supervision, using only category names, proves surprisingly effective for cross-dataset HS classification, outperforming traditional supervised methods in challenging scenarios.
- Category Representation as Class Definition: The paper frames the learned category representation (and its associated keywords) as the model's interpretation of a class definition. This allows for "re-annotating" a dataset by simply deriving category representations from a different source, enabling cross-dataset comparisons without manual relabeling.
- Data Distribution vs. Class Definition: The study effectively disentangles the impact of data distribution and class definition on cross-dataset performance. They found that providing a category representation derived from the target dataset significantly improves performance, even when using source-dataset documents.
- Limitations of Keyword-Based Methods: While X-Class uses keywords, the authors acknowledge that relying solely on them can be noisy, especially in cross-dataset settings where the target dataset might lack relevant keywords or have noisy ones.
Key Results
- In-Dataset Performance: X-Class (and its variant X-Class Agree) achieved competitive performance, often matching or exceeding supervised baselines like BERT and SVM, especially in the SBIC dataset.
- Cross-Dataset Performance: In the "Waseem → SBIC" setting (where the target taxonomy is finer-grained and a direct mapping is impossible), X-Class Agree achieved a notable accuracy of 69.8% by retraining on unlabeled SBIC data with target category representations, significantly outperforming other methods in this challenging scenario.
- Impact of Category Representation: In cross-dataset settings, using category representations derived from the target dataset (rather than the source) led to substantial improvements. For instance, in the "Waseem → SBIC" experiment, accuracy increased from 60.7% to 75.1% by using target category representations.
- Keyword Voting: Using X-Class's expanded keywords improved simple keyword voting baselines, though precision sometimes dropped due to noisy keywords.
Limitations and Future Work
- Monolingual Model: The study used a monolingual BERT model, limiting its applicability to English. Future work could explore multilingual models.
- Social Media Language: The use of social media language, including code-switching, might degrade performance.
- HS Detection vs. Classification: The current work focuses on classifying HS categories. Extending it to HS detection (identifying if a post contains HS at all) could be done by incorporating an "Others" category or a pipeline approach.
- Richer Representations: The authors suggest exploring richer category representations beyond simple keyword bags, potentially leveraging open-domain concept extraction.
- Zero-Shot and Open-World Classification: Future work could investigate generalizing to entirely new, unseen categories.
- Large Language Models (LLMs): The authors note that larger LLMs might exhibit significantly better zero-shot performance, a direction for future exploration.
- Ethical Considerations: The paper emphasizes the need for thorough evaluation before deploying weakly-supervised models, especially given their potential to misinterpret data for under-represented groups.
Summary updated