Open-World Knowledge Graph Based on Text Knowledge Transformation and Alignment

With the development of knowledge graph completion technology, many models have emerged that can predict missing relations in knowledge graphs with high accuracy. However, these relation prediction models only perform well in predicting relations that have already appeared in the knowledge graph, and perform poorly for new relations that have not appeared in the knowledge graph. In this paper, we propose an open-domain knowledge graph relation prediction model for predicting new relations. The proposed model genarates a relation embedding that can be aligned into the knowledge graph embedding by learning the semantics of a new relation from a textual description of that relation. With this relation embedding, relation prediction can be achieved based on knowledge graph embedding. Since this model considers both neighbor entity features and neighbor topology features of relations when extracting features from text, it performs well on open-domain knowledge graph completion tasks. Experiments show that this model exhibits an excellent ability to predict new relations without a large-scale training dataset.