Simple knowledge graph completion model based on PU learning and prompt learning

Abstract

Knowledge graphs (KGs) are important resources for many artificial intelligence tasks but usually suffer from incompleteness, which has prompted scholars to put forward the task of knowledge graph completion (KGC). Embedding-based methods, which use the structural information of the KG for inference completion, are mainstream for this task. But these methods cannot complete the inference for the entities that do not appear in the KG and are also constrained by the structural information. To address these issues, scholars have proposed text-based methods. This type of method improves the reasoning ability of the model by utilizing pre-trained language (PLMs) models to learn textual information from the knowledge graph data. However, the performance of text-based methods lags behind that of embedding-based methods. We identify that the key reason lies in the expensive negative sampling. Positive unlabeled (PU) learning is introduced to help collect negative samples with high confidence from a small number of samples, and prompt learning is introduced to produce good training results. The proposed PLM-based KGC model outperforms earlier text-based methods and rivals earlier embedding-based approaches on several benchmark datasets. By exploiting the structural information of KGs, the proposed model also has a satisfactory performance in inference speed.