Enhancing Engineering Document Analysis Through Structured Data Mapping

Abstract This paper presents an innovative approach for mapping engineering documents, particularly unstructured PDF files, to knowledge graphs, offering significant benefits for both engineering processes (i.e. such as ‘review-check-approve’ process) and AI-powered development. The method aims at improving efficiency, enhancing decision-making capabilities, reducing document turn-around duration, and thus decreasing costs while increasing productivity. Specific advantages include: Quick overview for reviewers to prioritize tasks - more weight can be given to identified main topics with less time on the side topics.Improved quality: interrupted relationships and isolated entities may indicate inconsistency that is usually overlooked in the traditional review process.A faster review process achieved through common understanding between reviewers, resulting in a subjective time reduction (average 10% improvement) while enhancing quality. Our method leverages advanced natural language processing (NLP) techniques, sentence transformers and large language models (LLMs) to transform complex document content into structured knowledge visual representations. This approach addresses the challenge of efficiently extracting information from unstructured engineering documents by first identifying relevant keywords, that align with the main topics of the document, through a robust initial process. The extracted keywords define the framework upon which the entire knowledge graph is built, ensuring its relevance and utility. Central to our methodology are novel techniques in LLM-based prompt engineering that enhance entity and relationship extraction in complex engineering contexts. These techniques incorporate chain-of-thought reasoning for improved accuracy. Furthermore, we introduce a robust process of relevant keyword extraction as an initial step, which sets the foundation for all subsequent stages from entity extraction to final graph presentation. The resulting knowledge graphs serve as a method of enhanced visual understanding of documents that are often complex and difficult to comprehend through human analysis alone. By extracting patterns and connections that would otherwise be challenging to identify or understand, these graphs facilitate simplified document comprehension and improved data integration. Our proof-of-concept results demonstrate the effectiveness of this method in supporting decision-making processes in the energy sector. Using this methodology as the ground base we outline directions for future AI development applications such as generating contextually relevant responses of the interrogated documents, summary creation and integration with other AI technologies.