Enhancing Knowledge Graph Construction from Multilingual Technical Reports using Large Language Models

Switchgear maintenance is a complex task that requires expert knowledge to address a wide range of potential issues. Although technical reports can assist, they are often disorganized and difficult to navigate. Knowledge graphs offer a structured solution, but constructing them from extensive text sources is challenging. Previous research demonstrated that Large Language Models can effectively convert unstructured data into knowledge graphs; however, their performance has been limited to short texts, struggling with larger documents. This work aimed to address that gap by developing a two-step extraction pipeline consisting of Entity Extraction and Relationship Extraction, applied in a few-shot learning context. Preprocessing techniques were used to improve data quality, and entity alignment methods were applied to reduce graph sparsity. The results show that with proper tuning, high-quality knowledge graphs can be generated from lengthy technical reports. These findings provide a pathway for more efficient use of unstructured knowledge in multiple industrial domains, potentially reducing reliance on manual expertise and improving the organization of large-scale knowledge bases.