| Abstract [eng] |
Incorrect or ambiguous software requirements are a well-documented source of wasted effort and delivery failures in software projects, yet automated validation tools that combine methodological rigor with practical applicability remain scarce. This Master's thesis proposes a requirements compliance validation method based on Graph Retrieval-Augmented Generation (Graph RAG), in which ISO/IEC/IEEE 29148:2018-aligned compliance criteria – content, semantic, and data compliance – are evaluated by a Large Language Model (LLM) against a domain knowledge graph constructed from project documentation. A systematic literature review confirmed that Graph RAG has not previously been applied to requirement compliance validation as a standalone method, establishing a clear research gap addressed by this paper. The proposed method was implemented as a configurable validation tool and evaluated against real-world requirements from a cyber-security project. Experimental results demonstrate that the method is capable of achieving F1 scores of 0.852 (content), 0.974 (semantic), and 0.903 (data). The method exhibits a consistent conservative bias, positioning it as an effective screening tool that surfaces requirements warranting human review rather than a fully autonomous arbiter of requirement quality. |
