Comparisons in End-to-End Pipeline Designs for Customized Document Information Extraction

Abstract

As businesses continue to adapt to the shift toward the digitalization of corporate tasks, one particular remaining financial and temporal bottleneck is the need for manual labor in interpreting digital documents and recording relevant information. Much work and research has been done, utilizing both machine learning techniques and traditional algorithmic approaches, to alleviate the resources required for this task by developing automated solutions for extracting information from such documents. However, current commercially available solutions typically struggle with either generalization to unique document structures or with handling the range of potential details present within a document type. The thesis introduces and compares two distinct end-to-end pipeline architectures combining neural networks with algorithmic techniques to effectively extract custom key-value information, with one focusing on commercial invoices with consistent keys and the other on technical specification sheets with variable keys. With accuracy, generalizability, and modularity as priorities, their use cases, benefits, and limitations are explored alongside comparisons with existing commercial solutions.