Invoice 2 Vec: Creating AI to Read Documents - Mark Landry - H2O AI World London 2018
- Predictive AI
- H2O Driverless AI
Overview
In this session, H2O.ai's Mark Landry and Brandon present Invoice2Vec, a three-month project built with a client to extract structured information, like invoice dates and totals, from business documents. They explain why off-the-shelf OCR alone falls short, including a real example where OCR misread a Euro amount, then detail their approach combining object detection models like Faster R-CNN and RetinaNet with NLP classification in Driverless AI. The talk covers the team's struggle to generate enough varied, accurately labeled training data and their plan to feed results back into Driverless AI as reusable recipes.
Key takeaways
- Why off-the-shelf OCR isn't enough for document extraction, illustrated by a real OCR error that misread a Euro amount as a different currency
- How the team used object detection models, Faster R-CNN and RetinaNet, to locate bounding boxes for fields like invoice dates and numbers
- Why generating good training data was the hardest part, and how the team used document templates to multiply labeled examples while managing overfitting
- How Driverless AI was used to classify extracted text lines against target fields, reaching high accuracy with minimal manual NLP work
- The tradeoffs between two-stage detectors like Faster R-CNN and one-stage detectors like RetinaNet, including RetinaNet's loss function for focusing on hard examples
- The team's plan to feed learnings back into Driverless AI as reusable recipes for document extraction
Start owning your AI