Eintrag in der Universitätsbibliographie der TU Chemnitz
Volltext zugänglich unter
URN: urn:nbn:de:bsz:ch1-qucosa2-979790
Hüttemann, Sebastian
Dinter, Barbara (Prof. Dr) ; Müller, Roland M. (Prof. Dr) ; Schryen, Guido (Prof. Dr) (Gutachter)
Domain-Specific Knowledge Extraction and Synthesis for Literature Reviews
Kurzfassung in englisch
The social sciences have a knowledge management problem. Researchers need to identify related work to position their research, discover theoretical and practical knowledge gaps, and acquire methodological knowledge about how to conduct research. These tasks have become increasingly resource and time consuming as millions of articles are published each year. The predominant method of identifying related work requires the use of a handful of academic search engines that dominate the market, namely Google Search, Scopus, and Web of Science. However, these products do a poor job of making the available knowledge sufficiently accessible. As a result, hundreds of millions of articles are described by only a handful of metadata points, such as author, title, publication date, or a set of author-defined keywords. Searching for specific topics, research methods, and theories requires crafting database queries that include potential synonyms to check for matching terms in titles, abstracts, and, where available, the full text of articles. With millions of research articles available, current search strategies result in thousands of search results that must be manually reviewed for inclusion in a literature review. The recent rise of general artificial intelligence with tools such as ChatGPT has led to academic search engines attempting to adopt this new technology. However, it has been shown that these tools do not live up to expectations and in particular suffer from the phenomenon of hallucinating information. This requires researchers to carefully review generated search results of such models. Researchers currently rely on tools and technologies owned by private companies or large publishers, which often come with limitations, such as a lack of domain specificity and restricted search and filtering capabilities. But what if researchers could develop their own domain-specific tools and technologies, reducing dependency on these external players and better supporting the literature review process? This thesis consists of eight design science research studies, each of which contributes a building block to answering the overarching research question of how we can design tools and technologies that incorporate domain knowledge to support literature reviews. The research papers describe the design and development of several conceptual and technical artifacts, such as models, methods, and prototypical instantiations, that guide the design of domain-specific solutions for academic knowledge management, exemplified in the information systems discipline. We explore how knowledge representations, such as domain ontologies, can guide the design and functionality of automated knowledge extraction to identify domain-specific knowledge in research articles. Based on the extracted knowledge, we demonstrate how innovative semantic functionalities can be integrated into search engines and literature review processes to support knowledge synthesis and the conduct of literature reviews. This thesis presents a multi-design science research study aiming for generalizing, integrating, and formalizing the results of the individual research papers into design principles and an overarching design theory. The knowledge contributions of this thesis lead to several implications for research and practice. We have developed methods that can support research disciplines in managing domain knowledge more effectively, potentially supporting the development of domain- specific knowledge infrastructures. We have also shown how the developments in this thesis can support the conduct of literature reviews. By generating additional semantic metadata based on domain ontologies, it becomes possible to not only analyze a sample of articles, but the entire population. This thesis also contributes to the evaluation and improvement of Generative AI-based tools and technologies. Based on the semantic metadata extracted from research articles, it becomes possible to create a ground truth that enables the evaluation of Generative AI-based tools by assessing the degree of information hallucination. Furthermore, the approaches outlined in this thesis could be applied to different types of organizations to support knowledge extraction from unstructured data in documents.
| Universität: | Technische Universität Chemnitz | |
| Institut: | Professur Wirtschaftsinformatik - Geschäftsprozess- und Informationsmanagement | |
| Fakultät: | Fakultät für Wirtschaftswissenschaften | |
| Dokumentart: | Dissertation | |
| Betreuer: | Dinter, Barbara | |
| DOI: | doi:10.60687/2025-0143 | |
| SWD-Schlagwörter: | Wissensextraktion , Literaturrecherche , Wirtschaftsinformatik , Information Extraction | |
| Freie Schlagwörter (Deutsch): | Literature Review , Knowledge Synthesis , Knowledge Extraction , Design Science Research , Large Language Models | |
| DDC-Sachgruppe: | Informatik, Informationswissenschaft, allgemeine Werke, Bibliotheks- und Informationswissenschaften | |
| Sprache: | englisch | |
| Tag der mündlichen Prüfung | 08.07.2025 |
