A Retrieval-Augmented Generation Framework for Root Cause Analysis in Maritime Accident Adjudication Reports

Abstract

This thesis proposes a retrieval-augmented generation (RAG) framework for the root cause analysis (RCA) of maritime accidents, defining the task as the inference of detailed causal narratives by retrieving and analyzing historically similar adjudication precedents. To address the challenge of providing grounded evidence in specialized domains, we construct a domain-specific knowledge base by converting 13,329 Korea Maritime Safety Tribunal adjudication summaries (1971–2025) into structured Card units. We assign hierarchical cause labels to serve as structural metadata, optimizing the retrieval of relevant precedents across distinct document fields (summary, causes, disposition). By fusing sparse and dense re- trievers via Reciprocal Rank Fusion (RRF), the framework identifies pertinent precedents and synthesizes them into evidence-based explanations. We further design a quantitative evaluation framework using a Metadata Relevance Score to overcome the absence of a manually labeled gold standard. Experiments demonstrate that the proposed Multi-Field Retrieval strategy substantially outperforms baselines in rank- ing quality (nDCG) and coverage. The final framework generates structured RCA outputs comprising textual cause summaries and hierarchical tags, providing a systematic decision-support tool for maritime safety investigations.