Deep Learning for Information Extraction From Digital Documents: An Innovative Approach to Automatic Parsing and Rich Text Extraction From PDF Files

Introduction

Technology refers to the body of information that is used to produce useful products and to design new products, or all the physical processes that transform inputs into outputs and the social arrangements accompanying this transformation. The role of technology in the field of international competition has become so decisive that today the classification of economic development has gradually turned into technology producing and non-producing countries. The most important subject of today's technological development is data and its processing. The huge increase in data volume and data diversity, the fact that even simple daily activities produce data, require very large volumes of warehouses to store this data. At the same time, the processing of these data and the ability to quickly analyze them for the purpose make the dimensions important in terms of data storage and processing capacity. The analysis of data and the size of the data are also factors affecting competition in areas such as trade, energy, and economy. For this reason, to obtain economic value from big data, it is important to process the correct data with advanced analytical methods, to keep up with the new developing technologies with the machine learning method, and to affect the decision-making processes.

In many countries, the ecosystem, which includes public institutions/organizations, the private sector, and academia, sets targets for artificial intelligence applications and quality data. Big data collection platforms, increasing the capacity of analysis and decision making and implementation accordingly, and creating problem pools for sectors are the main ones (Ferrara et al.,2014; Chang et al.,2006). Access to quality data by extracting valuable information from data is a core mission of many businesses intelligence services that require large-scale document processing. Understanding the documents included in the business processes in terms of page layout allows the establishment of the contextual structure at the stage of semantically interpreting the document terms in it. In this book chapter, the focus is on PDF, which is the most preferred document format open to sharing electronically for security reasons, since the possibility of making changes and editing the document is limited. PDF is the most widely used format for book, magazine, newspaper, and article-like document formats and is developed by Adobe Systems. On the other hand, it is largely kept in unstructured PDF format in the scientific literature (Jimeno Yepes et al., 2021). This format, which is very suitable for preserving basic visual elements, also has great difficulties in terms of automatic processing of these visual contents by machines. Extracting non-natural language content such as figures and tables, which often provide a summary and worthwhile information, as well as text in formatted formats with a focus on visuality, from a PDF file is another challenge. To provide comprehensive information extraction, it is necessary to determine the document layout and to parse the visually enriched texts and non-natural language content in a machine-structured format with background and other special formal features. Based on this need, the authors make use of the information contained in PDF documents with a complex layout, image processing, and deep learning methods in this book chapter and propose a solution that enables the document outline to be determined, parsed into elements, defined as a clip, and the title, subtitle, image, and actual text content of the clip itself to be output in reading order.