Training-Time Explainability for Multilingual Hate Speech Detection: Aligning Model Reasoning with Human Rationales
Proposes a training-time explainability framework that aligns model reasoning with human-annotated rationales for multilingual hate speech detection Evaluates on two datasets: HateXplain (English) and BullySent (Hinglish), targeting culturally coded anti-Muslim hate speech Uses multiple interpretability methods including LIME, Integrated Gradients, Grad X Input, and attention mechanisms to assess explanation quality Gradient- and attention-based regularization improve F-scores while enhancing bo
Analysis
TL;DR
- Proposes a training-time explainability framework that aligns model reasoning with human-annotated rationales for multilingual hate speech detection
- Evaluates on two datasets: HateXplain (English) and BullySent (Hinglish), targeting culturally coded anti-Muslim hate speech
- Uses multiple interpretability methods including LIME, Integrated Gradients, Grad X Input, and attention mechanisms to assess explanation quality
- Gradient- and attention-based regularization improve F-scores while enhancing both plausibility and faithfulness of model explanations
- Demonstrates that integrating explainability during training—not just post-hoc—enables detection of implicit, culturally specific hate cues
Why It Matters
This research addresses a critical gap in AI content moderation: the tension between model accuracy and interpretability when dealing with culturally nuanced, multilingual hate speech. For AI practitioners building moderation systems, it demonstrates that explainability can be baked into training to simultaneously boost performance and trustworthiness, rather than treating it as an afterthought.
Technical Details
- Framework: Training-time explainability that regularizes model objectives using human-annotated rationales, aligning internal representations with human reasoning patterns
- Datasets: HateXplain (English hate speech with explanations) and BullySent (Hinglish code-mixed dataset), both reflecting anti-Muslim hate speech patterns
- Interpretability Methods: LIME (Local Interpretable Model-agnostic Explanations), Integrated Gradients, Grad X Input, and attention-based analysis for cross-method agreement evaluation
- Evaluation Metrics: Classification F-scores, explanation plausibility, faithfulness, and cross-method consistency to measure both predictive and interpretive quality
- Key Innovation: Gradient- and attention-based regularization terms added to the training loss to encourage the model to attend to culturally relevant linguistic cues
Industry Insight
- Content moderation systems should prioritize training-time explainability over post-hoc methods to achieve both higher accuracy and genuine interpretability, especially for culturally sensitive domains
- Multilingual and code-mixed datasets like BullySent represent an underexplored but critical frontier for hate speech detection, as a significant portion of online hate operates across language boundaries
- The finding that explainability regularization improves F-scores challenges the common assumption that interpretability comes at the cost of performance—organizations can pursue both simultaneously
Disclaimer: The above content is generated by AI and is for reference only.