Representation learning approach for understanding structured documents.

Area of Science:

• Computer Science
• Artificial Intelligence
• Information Retrieval

Background:

• Current document understanding methods struggle with complex layouts and logical connections between elements.
• Existing approaches fail to capture deep semantic relationships across diverse document structures.

Purpose of the Study:

• Introduce a novel representation learning framework, Document Relationship Entity Embedding Learner (D-REEL), to model intricate semantic relationships within documents.
• Enhance document understanding by accurately assessing semantic correlations between document fields, irrespective of their spatial arrangement.

Main Methods:

• D-REEL generates extraction candidates and learns dense vector representations (embeddings) for each.
• Compares embeddings to assess semantic correlations and determine article relationships.
• Combines spatial information with domain-specific schemas for precise extraction and correlation scoring.
• Proposes Semantic Structural Congruence (SSC) metric using location-agnostic localization.

Main Results:

• Significant improvements in correlation accuracy and extraction performance demonstrated on public datasets.
• Achieved an average mAP increment of 2-3% for correlation accuracy.
• Obtained an SSC improvement of almost 10% for the PRIMA dataset, indicating effective relationship measurement.

Conclusions:

• D-REEL offers a robust solution for understanding complex document layouts and semantic relationships.
• The framework enables precise extraction and reliable correlation scoring, even with irregular document structures.
• The proposed SSC metric effectively quantifies semantic structural congruence in documents.