Layout Inference and Table Detection in Spreadsheet Documents

Spreadsheets have found wide use in many different domains and settings. They provide a broad range of both basic and advanced functionalities. In this way, they can support data collection, transformation, analysis, and reporting. Nevertheless, at the same time spreadsheets maintain a friendly and intuitive interface. Additionally, they entail no to very low cost. Well-known spreadsheet applications, such as OpenOffice, LibreOffice, Google Sheets, and Gnumeric, are free to use. Moreover, Microsoft Excel is widely available, with millions of users worldwide. Thus, spreadsheets are not only powerful tools, but also have a very low entrance barrier. Therefore, they have become very popular with novices and professionals alike. As a result, a large volume of valuable data resides in these documents. From spreadsheets, of particular interest are data coming in tabular form, since they provide concise, factual, and to a large extend structured information. One natural progression is to transfer tabular data from spreadsheets to databases. This would allow spreadsheets to become a direct source of data for existing or new business processes. It would be easier to digest them into data warehouses and to integrate them with other sources. Nevertheless, besides databases, there are other means to work with spreadsheet data. New paradigms, like NoDB, advocate querying directly from raw documents. Going one step further, spreadsheets together with other raw documents can be stored in a sophisticated centralized repository, i.e., a data lake. From then on they can serve (on-demand) various tasks and applications. All in all, by making spreadsheet data easily accessible, we can prevent information silos, i.e., valuable knowledge being isolated and scattered in multiple spreadsheet documents. Yet, there are considerable challenges to the automatic processing and understanding of these documents. After all, spreadsheets are designed primarily for human consumption, and as such, they favor customization and visual comprehension. Data are often intermingled with formatting, formulas, layout artifacts, and textual metadata, which carry domain-specific or even user-specific information (i.e., personal preferences). Multiple tables, with different layout and structure, can be found on the same sheet. Most importantly, the structure of the tables is not known, i.e., not explicitly given by the spreadsheet documents. Altogether, spreadsheets are better described as partially structured, with a significant degree of implicit information. In literature, the automatic understanding of spreadsheet data has only been scarcely investigated, often assuming just the same uniform table layout. However, due to the manifold possibilities to structure tabular data in spreadsheets, the assumption of a uniform layout either excludes a substantial number of tables from the extraction process or leads to inaccurate results. In this thesis, we primarily address two fundamental tasks that can lead to more accurate information extraction from spreadsheet documents. Namely, we propose intuitive and effective approaches for layout analysis and table detection in spreadsheets. Nevertheless, our overall solution is designed as a processing pipeline, where specialized steps build on top of each other to discover the tabular data. One of our main objectives is to eliminate most of the assumptions from related work. Instead, we target highly diverse sheet layouts, with one or multiple tables. On the same time, we foresee the presence of textual metadata and other non-tabular data in the sheet. Furthermore, we make use of sophisticated machine learning and optimization techniques. This brings flexibility to our approach, allowing it to work even with complex or malformed tables. Moreover, this intended flexibility makes our approaches transferable to new spreadsheet datasets. Thus, we are not bounded to specific domains or settings.

Level thesis:degree_level

thesis.doctoral

Grantor dc:publisher

Technische Universität Dresden

Year

2020

Author dc:creator

• Koci, Elvis

Contributors dc:contributor

• Lehner, Wolfgang
• Romero, Oscar

• Spreadsheets
• Analysis
• Recognition
• Automatic
• Tabellenkalkulationen
• Analyse
• Erkennung
• Automatisch