Cold Start of Enterprise Knowledge Graph Construction

Rong Duan, Kangxing Hu

Abstract

Enterprise Knowledge Graphs(EKG) is a powerful tool for Enterprise Knowledge Management(EKM). Most EKG construction suffers cold start problem. In reality, EKG construction is an interactive process,in which domain experts provide a small seed graph, and data driven methods are applied to expand the graph. This paper proposes a framework to solve EKG cold start problem by integrating graph form expert knowledge with non-graph form corpus. The proposed framework employs expert knowledge to guide unsupervised learning, and crosses check the quality of expert knowledge simultaneously. A coarser cluster level and finer entity level vectorization is proposed to predict the link between graph nodes and cluster words. And also, a combined strategy is adopted to measure the importance of the predicted link, and provide to the expert to evaluate incrementally. The proposed framework solves the ”labor intensive” EKG cold start construction problem and utilizes expert knowledge efficiently. Simulation is generated to illustrate the properties of defined measurements, and real-world application is discussed to show the challenges in practices.

Download


Paper Citation