A RAG-based Q&A system for ship regulations applying domain adaptation

Various international organizations and classification societies have established extensive regulations to enhance the safety and reliability of ships. These regulations are continuously revised and updated to keep pace with technological advancements, the increasing size of ships, and the tightening of environmental regulations in the shipbuilding and maritime industry. The key contents of regulations are often conveyed through text, as well as interrelated equations and tables. Due to the complexity of these formats and the inherent difficulty of the regulatory content itself, users require considerable time, effort, and a high level of expertise to retrieve and interpret the necessary information accurately. This study proposed a Q&A (Question-and-Answer) system for ship regulations utilizing RAG (Retrieval-Augmented Generation) to address these challenges. The proposed system comprises a framework that retrieves relevant information from a large-scale database and generates reliable answers tailored to the user’s query intent. To achieve this, we constructed ship regulations into a database optimized for searchability and incorporated hierarchical structures to improve search accuracy. A search module was implemented based on semantic analysis techniques to retrieve the most relevant rules effectively. An answer generation module was also implemented to produce appropriate answers grounded in the retrieved rules. Especially, a domain adaptation technique was applied to a general-purpose model, enabling the system to clearly understand the specialized terminology and context within the domain of ship regulations. The Q&A system supports users in quickly navigating ship regulations, facilitating clear comprehension and practical application of the information. The proposed system was applied to the regulations of the Korean Register, demonstrating its capability to promptly and accurately retrieve rules related to user queries and generate reliable answers. • We propose a RAG-based Q&A system specifically designed for ship regulations. • A ship regulations database is constructed using a hierarchical structure comprising text, equations, and tables. • A search module with enhanced comprehension of ship regulations was implemented by applying domain adaptation to BERT. • We construct data for domain adaptation by proposing a method for generating GPL-based training data. • We implement an answer generation module using designed instructions and memory-efficient LLaMA.