Transfer Learning on Knowledge Graph Construction: A Case Study of Investigating Gas-Mining Risk Report

Abstract

This study addressed the problem of automated Knowledge Graph (KG) construction from unstructured documents, with the assistance of transfer learning. Despite a large amount of effort made to discover KG, how to explore unknown KGs from existing knowledge remains a challenge. In this paper, we firstly formulate the KG detection process as a transfer-learning problem, which consists of two main steps. At first, we pretrain a backbone model using the source domain. Due to sufficient samples from the source domain, this backbone model can be trained better. Second, we migrate this model (from the known domain) to the target domain by fine-tuning key parameters. The fine-tuning operation only requires less computation, which is very efficient. As such, the backbone model can be successfully transferred into the target domain, even with limit training samples. Experimental evaluations using one real-world dataset of gas-mining reports demonstrate the advantages of utilizing the proposed algorithm to construct KG using transferable information.