Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Sep 9·5 min read
When Superpixels Fail on Documents: A Study of Segmentation for LIME Explanations
TL;DR
Document-aware segmentation methods significantly improve the reliability of LIME explanations for document image classification, outperforming traditional superpixel approaches. This study highlights the importance of aligning segmentation with document structure for better interpretability.
✦ Why It Matters
Adopt document-aware segmentation techniques in LIME to enhance the interpretability of your document image classifiers today.
Key Takeaways
- Use document-aware segmentations for better LIME explanations.
- Evaluate the impact of segmentation on explanation reliability.
- Consider OCR bounding boxes for document image classification.
- Be aware of biases in model explanations related to document features.
How It Works
The study introduces document-aware segmentations that utilize OCR bounding boxes to better align with document structures. This approach contrasts with traditional superpixel methods, which often misrepresent text and layout, leading to less reliable explanations.