Thesis author: Christian Bick
Legacy system modernization is a complex process involving technical, organizational, business and process uncertainties and risks. Knowledge about the associated risks and mitigation strategies is available across academic and grey literature, but is fragmented and often represented inconsistently. For the relationship focused analysis pursued in this thesis, a particular limitation is that relationships between risks and mitigations are frequently left implicit. This makes the knowledge difficult to systematically compare and reuse, limiting its use for LLM-based analysis. To address the fragmented and weakly connected representation of modernization risk knowledge this thesis works by identifying risks and mitigation strategies of holistic legacy system modernization from academic and grey literature across all dimensions and by transforming the extracted knowledge into a structured, machine-readable and relationship-oriented representation that makes entities and their connections explicitly accessible and traceable for the analysis. This resulting knowledge base is implemented as an OKF representation and integrated into a prototype that provides LLM access and knowledge graph visualization. The evaluation shows that external modernization knowledge improves LLM-based analysis compared with providing no additional context, as in the proof of concept evaluation, the mean Overall Score increased from 46.87 without additional context to 87.58 with Full Structured Context and 92.72 with Unstructured Full Context. Adding explicit relationships to the structured risk and mitigation context increased the mean Overall Score from 75.12 to 87.58. Therefore explicit relationships particularly improve the retrieval and reasoning over connected knowledge, including direct and indirect relationships between risks and mitigations. At the same time, the comparison shows that unstructured textual context can retain valuable information. The structured representation, however, provides stronger relational traceability and substantially lower runtime in the evaluated setting. An external evaluation with six experts further supports the usefulness of the complete structured context, particularly for relationship-based analysis. The thesis therefore demonstrates how heterogeneous modernization risk knowledge can be transformed into a structured, traceable and executable representation for LLM analysis. The resulting knowledge model provides a foundation for more systematic and relationship-aware modernization risk analysis while retaining the contextual and evidential information required for human interpretation.
