Similarity-Based Graph Enhanced Text Representation Learning for Electronic Component Knowledge Graph Completion

Abstract

In the electronic component supply chain system, manually built knowledge graph usually lacks the alternative relations among the electronic components. Prevalent graph embedding approaches exhibit strong capability in representing graph elements. However, it's difficult to generalize to never-seen elements due to the graph incompleteness, and the Laplacian-based convolution of GCN limits the information propagation to immediate neighbors. In contrast, the pre-trained encoder have stronger ability to extract semantic information. In this paper, we propose a hybrid encoding approach SiGeTR: Similarity-based Graph Enhanced Text Representation. Based on the approach of structural encoding, it incorporates the textual encoding which employs the text of triples in the graph and contextualized repre-sentations. Meanwhile, we propose to use node similarity based convolution matrices in the GCN to compute node embeddings. In experiments, our methods obtain state-of-the-art performance on the electronic components knowledge graph benchmark dataset and achieve significant results with low resources.