Graph Convolutional Neural Network for extracting tabular data of purchase order documents
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Perustieteiden korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
2023-10-09
Department
Major/Subject
Data Science
Mcode
SCI3115
Degree programme
Master's Programme in ICT Innovation
Language
en
Pages
59 + 3
Series
Abstract
This thesis investigates the efficacy of Graph Convolutional Neural Networks (GCN) in extracting tabular data from purchase order documents, a task traditionally challenging due to the complex relationships between document elements. Unlike standard methods, GCN leverages a graph convolutional architecture to capture these intricate relationships, enabling more accurate tabular data extraction. Utilizing two datasets aggregated from multiple clients under strict data protection protocols, the research investigates how well GCNs can understand and parse complex tabular structures in diverse document formats. The GCNs are trained to classify nodes representing pieces of text within a document, and their performance is critically evaluated against conventional methods for this task, represented by the Azure Form Recognizer (AFR) model, which serves as the baseline for comparison. Upon completing the training task, the GCN model resulted in a 90\% F1-Score on the validation set for one of the datasets, while for the other, it reached 75\% F1-Score. The study reveals that, while AFR provides a straightforward solution for structured data extraction, its performance deteriorates when subjected to documents that deviate from the templates it has been trained on. On the other hand, GCNs demonstrate a better generalization capability, making them more suitable for dealing with various document structures. The findings have significant implications for document analysis, proving that GCNs are suitable and advantageous for extracting structured data from complex documents like purchase orders. Moreover, this thesis covers how this technique can be generalized for other types of document analysis tasks.Description
Supervisor
Laaksonen, JormaThesis advisor
Erkkila, TimoKeywords
tabular text extraction, purchase order documents, data science, GCN