A Robust Component-Based Template Matching Approach Using Document Layout Graph for Extracting Information

Son Lam Nguyen Vu, Tien Dong Nguyen, Van Hai Pham

Abstract: This paper introduces a novel approach for Key Information Extraction (KIE) from diverse documents, namely CompIE, leveraging the strengths of deep graph neural networks and subgraph mining techniques. Focusing on overcoming the limitations of traditional KIE methods, CompIE employs a flexible strategy for graph construction and matching. Unlike existing models constrained to one-to-one correspondences, our approach effectively handles complex multiple correspondences within documents. The core of CompIE involves two key processes: Sub-Graph Mining, utilizing the Subdue algorithm to identify frequent subgraph patterns, and Sub-Graph Matching, employing a deep graph neural network for aligning subgraphs with document graphs. Notably, the modification of the loss function from softmax to binary allows for multiple mappings using a single component, addressing the challenge of recurrent components in documents. Experimental results on benchmark datasets, CORD and NAF, demonstrate the superior performance of CompIE, achieving accuracy rates of 94.98% and 85.31%, respectively. These results significantly outperform established models such as Yao et al.’s and LayoutLMv2\(_{BASE}\), underscoring CompIE’s potential as a robust and scalable solution in the field of document understanding and information extraction.