A Robust Framework for One-Shot Key Information Extraction via Deep Partial Graph Matching.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Text field labelling is crucial for Key Information Extraction (KIE) from structured documents.
• Existing KIE methods struggle with field drift and outlier problems, limiting robustness and performance.

Purpose of the Study:

• To propose a novel framework, Deep Partial Graph Matching (dPGM), to address limitations in current KIE methods.
• To enhance the robustness and accuracy of one-shot KIE tasks.

Main Methods:

• Framing text field labelling as a partial graph matching problem.
• Developing an end-to-end trainable framework (dPGM) that represents documents as graphs.
• Utilizing a combinatorial solver with a one-to-(at most)-one mapping constraint for exact graph matching.

Main Results:

• The dPGM framework demonstrates robustness against field drift and outlier issues.
• Achieved state-of-the-art performance on both public and the newly introduced DKIE datasets.
• Introduced DKIE, a large-scale dataset for one-shot KIE research.

Conclusions:

• Deep Partial Graph Matching (dPGM) offers a robust and effective solution for one-shot Key Information Extraction.
• The proposed method significantly improves upon existing KIE techniques, particularly in handling data variations.
• The release of the DKIE dataset will foster further advancements in the field.