Tables are omnipresent. They appear in scientific articles, business reports, web pages, and many other document types. In a world where data is increasingly valuable, the need for machines to access and understand tabular information is growing rapidly. However, their variety, irregular layouts, implicit semantics, and human-centric design make automated interpretation difficult. Reliable extraction and understanding are therefore essential for search, analytics, and knowledge-driven applications. Making tables machine-readable differs from traditional text processing, requiring specialized techniques that can handle the structure and semantics of tabular data. Nevertheless, automating table understanding enables new possibilities for knowledge extraction, data integration, and intelligent information retrieval. This thesis investigates how tables can be systematically extracted, interpreted, and integrated into downstream knowledge-intensive applications. It formalizes the previous notions of table structures and embeds them into the broader document context with the help of the new S2Doc format to unify heterogeneous representations. It further presents KIETA, a modular hybrid pipeline framework for table processing built on S2Doc. This framework defines a structure that allows combining any implemented components to build table extraction and document understanding workflows. KIETA includes several rule-based and machine-learning-based modules for table detection, structure recognition, functional analysis, and semantic annotation. In addition to that, a web application enabling interactive editing, inspection, and workflow orchestration has been developed to facilitate rapid prototyping and evaluation of table processing pipelines. Of note is the table editor, allowing users to visually inspect and correct table structures, while any change is automatically reflected in the underlying S2Doc representation. Several experiments compare different table processing strategies on diverse real-world documents. This includes an extensive evaluation of different table extraction pipelines, including end-to-end machine-learning-based approaches as well as hybrid methods combining machine learning and rule-based components. The results highlight the strengths and weaknesses of each approach and demonstrate the advantages of a universal document representation and the modular pipeline design. Furthermore, challenges like error propagation that are inherent to such a modular setup are analyzed. In two additional experiments, the capabilities of LLMs regarding table understanding are investigated, and it is explored whether specialized table processing pipelines can work in synergy with LLMs to improve overall performance. The first experiment evaluates the impact of preprocessing tables with a KIETA pipeline on the accuracy of LLM-based table question answering. The second experiment directly compares the table understanding capabilities of LLMs with specialized table processing pipelines using the task of knowledge graph construction from tables. This knowledge graph is evaluated by constructing a leaderboard from the results. Finally, the thesis explores the applicability of the proposed methods in a real-world use case. It focuses on constructing a knowledge graph from tables of publications mentioned in the related work section of this thesis, which is then used to create an overview table. The work concludes by outlining remaining challenges in universal table understanding and opportunities for integrating structured extraction with large language models.
Sebastian Kempf (Thu,) studied this question.