Weakly supervised update of civil aircraft maintenance knowledge graphs through in-context learning

Yini Zhang, Peixuan Lei, Yuxuan Zhang, Huanchun Peng, Qingqian Zhou, Yuanxiang Li

Abstract: Civil aircraft maintenance records document real-world operational issues and solutions. These records contain valuable maintenance expertise. Existing methods for processing these records rely on isolated analyses. They focus only on individual solutions or categories, failing to provide a comprehensive view of large-scale data. Therefore, we propose integrating the empirical knowledge from maintenance records into knowledge graphs to overcome the efficiency limitations. However, in specialized fields like civil aircraft, there are still two main challenges. First, the high cost of expert annotation limits the utilization of unlabeled data. Second, general language models struggle with domain-specific terminology and context, and fine-tuning them requires a large amount of labeled data. To address these issues, we propose a weakly supervised framework that combines in-context learning and two-stage pseudo-label filtering. Our proposed model leverages limited labeled data through in-context learning to enhance semantic understanding, avoiding reliance on unstable prompt engineering. Simultaneously, pseudo-labels are assigned to unlabeled records and refined via 2-stage filtering to ensure reliability. Experiments demonstrate our method achieves competitive performance even with limited annotated data, offering a scalable solution for managing large volumes of maintenance records in domain-specific, data-scarce scenarios and maintaining up-to-date KGs with expertise.

External IDs:doi:10.1007/s42401-025-00404-7