Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 28·5 min read
Training-Time Explainability for Multilingual Hate Speech Detection: Aligning Model Reasoning with Human Rationales
TL;DR
A training-time explainability framework was developed to enhance multilingual hate speech detection by aligning AI model reasoning with human rationales. This approach improves classification performance and interpretability, particularly for detecting anti-Muslim hate in culturally specific contexts.
✦ Why It Matters
Implement training-time explainability in your hate speech detection models to improve accuracy and cultural sensitivity.
Key Takeaways
- Adopt training-time explainability for better model interpretability.
- Utilize LIME and attention mechanisms for enhanced explanation quality.
- Focus on culturally specific cues in hate speech detection.
- Evaluate model performance using diverse multilingual datasets.
How It Works
The framework aligns AI model reasoning with human-annotated rationales during training, allowing the model to learn from human insights. Techniques like LIME and attention mechanisms are used to provide interpretable explanations for model predictions, ensuring that the model captures culturally specific cues relevant to hate speech.