SSKGE: a time-saving knowledge graph embedding framework based on structure enhancement and semantic guidance

Abstract

In knowledge graph embedding, an attempt is made to embed the objective facts and relationships expressed in the form of triplets into multidimensional vector space, facilitating various applications, such as link prediction and question answering. Structure embedding models focus on the graph structure while the importance of language semantics in inferring similar entities and relations is ignored. Semantic embedding models use pretrained language models to learn entity and relation embeddings based on text information, but they do not fully exploit graph structures that reflect relation patterns and mapping attributes. Structure and semantic information in knowledge graphs represent different hierarchical properties that are indispensable for comprehensive knowledge representation. In this paper, we propose a general knowledge graph embedding framework named SSKGE, which considers both the graph structure and language semantics and learns these two complementary characteristics to integrate entity and relation representations. To compensate for semantic embedding approaches that ignore the graph structure, we first design a structure loss function to explicitly model the graph structure attributes. Second, we leverage a pretrained language model that has been fine-tuned by the structure loss to guide the structure embedding approaches in enhancing the semantic information they lack and obtaining universal knowledge representations. Specifically, guidance is provided by a distance function that makes the spatial distribution of the two types of graph embeddings have a certain similarity. SSKGE significantly reduces the time cost of using a pretrained language model to complete a knowledge graph. Common knowledge graph embedding models such as TransE, DistMult, ComplEx, RotatE, PairRE, and HousE have achieved better results with multiple datasets, including FB15k, FB15k-237, WN18, and WN18RR, using the SSKGE framework. Extensive experiments and analyses have verified the effectiveness and practicality of SSKGE.