Research · 2022
GAT for Text Line Detection
4M-parameter Graph Attention + GRU model for line detection on receipts - state-of-the-art 97.23% F1 on CORD with 100–200× fewer parameters.
97.23%
F1 on CORD (state of the art)
~4M
Parameters
100–200×
Fewer parameters than competing models
Context
Applied research alongside the receipt-extraction work, 2022. Co-authored with David Montero Martín, David Jiménez and Javier Yebes.
Problem
Matching key–value pairs on receipts depends on knowing which OCR detections sit on the same line. Existing line-detection models were built for structured documents and generalised poorly to receipts.
Approach
We modelled OCR detections as a graph and combined Graph Attention layers with Gated Recurrent Units to predict line membership. An ablation study validated each architectural choice.
The paper was presented at the Document Intelligence Workshop at KDD 2022 in Washington DC.
Results
- 97.23% F1 on the CORD dataset (state of the art)
- ~4M parameters - 100–200× smaller than competing models
- Co-authored with David Montero Martín, David Jiménez and Javier Yebes
Stack
- Graph Attention Networks
- GRU
Have a similar problem?
Tell me what you're working on and I'll tell you honestly whether I can help.