🤖 AI

Multi-Field Hybrid Retrieval-Augmented Generation for Maritime Accident Root Cause Analysis

This paper proposes a multi-field hybrid retrieval-augmented generation framework that leverages a structured dataset of 13,329 maritime accident reports to significantly improve both the retrieval of relevant precedents and the quality of automated root cause analysis drafting compared to baseline methods.

Original authors: Seongjin Kim, Sungil Kim

Published 2026-06-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Seongjin Kim, Sungil Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a maritime safety detective trying to solve a new accident. To do your job well, you need to look at thousands of old case files from the past 50 years to see how similar accidents were solved before. This is called Root Cause Analysis (RCA).

The problem is that these old files are messy. They are huge, written in complex legal and technical language, and the important clues are buried in different sections. A human detective might spend days just reading and cross-referencing these files.

This paper introduces a new "AI Detective Assistant" designed to speed up this process. Here is how it works, explained simply:

1. The Problem: The "Wall of Text"

Imagine trying to find a specific recipe in a library where every book is a giant, unbroken wall of text containing the story of a meal, the list of ingredients, and the judge's final verdict all mixed together. If you search for "fire," the computer might show you a story about a fire, but it might also show you a verdict about a fire that has nothing to do with your current case.

In the past, AI systems treated these accident reports like that giant wall of text. They would get confused because the "story" part, the "cause" part, and the "punishment" part were all blended together.

2. The Solution: The "Index Card" System

The researchers took 13,329 old accident reports and turned them into structured "Incident Cards."

Think of this like taking that giant wall of text and cutting it into three distinct, labeled index cards for every single case:

Card A (Summary): What happened? (The story)
Card B (Causes): Why did it happen? (The technical reason)
Card C (Disposition): What was the official ruling? (The outcome)

By separating these, the AI can look for a specific type of clue. If you are looking for why a ship sank, it knows to look at the "Causes" cards, not the "Story" cards.

3. The Search Engine: The "Two-Legged Race"

To find the right cards, the AI uses a hybrid search strategy, which is like having two different detectives working together:

Detective 1 (The Keyword Hunter): This one is great at finding exact matches, like specific legal terms or technical codes (e.g., "propane gas"). It's like searching a library by exact book title.
Detective 2 (The Meaning Hunter): This one understands the idea behind the words. It knows that "ship hit a rock" is the same as "vessel grounded on reef," even if the words are different. It's like searching by the plot of the story.

The system combines the results from both detectives. If one finds a match and the other finds a similar match, they vote on which one is the best. This ensures the AI doesn't miss anything just because the wording was slightly different.

4. The Result: A Better Report

Once the AI finds the best historical examples (the "precedents"), it uses them to write a new report for the current accident.

Without this system: The AI tries to guess the cause based only on its general training. It might make up a plausible-sounding but wrong reason (a "hallucination"), like guessing a fire started because of alcohol fumes when it was actually gas.
With this system: The AI looks at the real, historical "Cause" cards it found. It sees that in similar past cases, the cause was "leaking propane." It then writes a report that says, "Based on similar past cases, this was likely caused by leaking propane."

What the Numbers Say

The researchers tested this against a standard AI that didn't use these special cards or the two-detective search:

Finding the right files: The new system found relevant past cases three times better than the old way.
Writing the report: The reports written with the help of past cases were rated significantly higher in quality and accuracy. They were less likely to make things up and more likely to stick to the facts found in history.

The Bottom Line

This paper shows that by organizing messy legal documents into neat, separated categories and using a smart search that looks at both exact words and general meanings, we can build an AI assistant that helps maritime investigators solve accidents faster and more accurately. It turns a chaotic library of old cases into a well-organized reference tool that prevents the AI from guessing and forces it to rely on evidence.

Technical Summary: Multi-Field Hybrid Retrieval-Augmented Generation for Maritime Accident Root Cause Analysis

Problem Statement
Maritime accident adjudication reports, such as those from the Korea Maritime Safety Tribunal (KMST), contain critical findings for Root Cause Analysis (RCA). However, leveraging these records for new investigations is labor-intensive. The reports span decades and are characterized by dense, fragmented information mixing legal and technical terminology across distinct sections (narratives, causal reasoning, and dispositions). While Large Language Models (LLMs) offer potential for automating narrative synthesis, they are prone to hallucinations in specialized domains, often generating plausible but factually incorrect regulations or causal mechanisms. Furthermore, standard Retrieval-Augmented Generation (RAG) approaches face two specific challenges in this context: (1) treating adjudication reports as monolithic text blocks dilutes field-specific signals necessary for precise retrieval, and (2) evaluating retrieval quality at scale is difficult due to the high cost of obtaining expert relevance labels.

Methodology
The authors propose a multi-field hybrid RAG framework designed for case-informed structured generation. The methodology consists of four primary components:

Structured Knowledge Base Construction:
The system utilizes a dataset of 13,329 KMST adjudication reports (1971–2025). Raw documents are transformed into structured "incident cards" containing three distinct retrievable fields: Summary (incident description), Causes (tribunal's causal reasoning), and Disposition (administrative outcomes). Each card is also enriched with hierarchical L1/L2 cause tags and metadata (e.g., accident type, jurisdiction, year). This process yields approximately 37,007 retrievable chunks.
Rule-Based Taxonomy Tagging:
To facilitate structured generation without extensive manual annotation, the authors employ a weak supervision approach. Using approximately 2,400 dictionary entries and 800 regex patterns, they map keywords and causes from the reports to a standardized KMST cause taxonomy (L1 groups and 24 L2 factors), achieving 97.6% coverage.
Multi-Field Hybrid Retrieval:
The retrieval strategy employs a field-aware hybrid approach. For each of the three document fields, the system runs two complementary retrievers:
- Sparse Retrieval (BM25): For exact keyword matching of technical identifiers.
- Dense Retrieval (BAAI/bge-m3): For capturing semantic paraphrases.
  The results from these six independent rankings (3 fields × 2 methods) are fused using Reciprocal Rank Fusion (RRF) to produce a final ranked list of evidence chunks.
Structured Generation:
The generator (Qwen3-Next-80B-A3B-Instruct) is conditioned on the top- $k$ retrieved chunks. It is tasked with synthesizing a concise cause narrative ( $n$ ) and mapping the incident to standardized taxonomy tags ( $T$ ). The output is constrained to a strict JSON schema with a temperature of 0.0 to minimize variance.

Evaluation Protocol
Due to the lack of large-scale expert relevance labels, the authors introduce a metadata-derived proxy relevance score ($RelScore$) based on keyword similarity, accident type, jurisdiction, and temporal proximity. Retrieval performance is measured using Norm-Recall@100 and nDCG@100. Generation quality is evaluated via ROUGE-L, SBERT cosine similarity, and an LLM-as-a-judge score (1–5 scale) assessing coherence and grounding.

Key Results

Retrieval Performance: The proposed multi-field hybrid strategy significantly outperforms baselines.
- Moving from whole-document indexing to multi-field indexing improved Norm-Recall@100 from 0.1819 to 0.4981 (Dense).
- The hybrid approach (fusing BM25 and Dense) achieved the highest performance with a Norm-Recall@100 of 0.5463 and nDCG of 0.7697, compared to 0.1819 for the whole-document dense baseline.
- The inclusion of the "Disposition" field provided a significant boost, suggesting administrative outcomes carry discriminative signals for disambiguating causally different cases.
Generation Performance: Grounding the generator in retrieved precedents improved the LLM-as-a-judge score from 3.34 to 3.72 (+11.5%). While lexical overlap (ROUGE-L) and semantic similarity saw marginal gains, the qualitative improvement lies in factual consistency and the ability to cite specific precedents.
Case Studies: In specific examples (e.g., stranding and explosion cases), the LLM-only baseline produced hallucinated causal links (e.g., attributing an explosion to alcohol vapors based on crew intoxication). The proposed framework correctly identified technical mechanisms (e.g., propane gas ignition) by leveraging the specific "Causes" and "Disposition" fields from retrieved precedents.

Significance and Contributions
The paper claims that its primary contribution is demonstrating that field-aware RAG can substantially streamline maritime safety investigation workflows. By isolating descriptive narratives from technical and legal reasoning, the system reconstructs the diagnostic reasoning pathways used by expert investigators. The study establishes that:

Structural Alignment Matters: Treating documents as monolithic blocks dilutes critical signals; separating fields enhances retrieval precision.
Hybrid Retrieval is Essential: Combining sparse and dense signals provides robustness against vocabulary mismatches while preserving the retrieval of precise technical terminology.
Evaluation Proxy Viability: A metadata-driven proxy score can serve as a reproducible, systematic baseline for comparing retrieval strategies in specialized administrative domains where gold-standard labels are unavailable.

Ultimately, the framework enables faster precedent search and more consistent, evidence-based RCA drafting, reducing the risk of hallucination in high-stakes maritime safety analysis.