FSPB: Filtered Semantic Paths and BERT for Knowledge Graph Completion

Currently, the field of Knowledge Graph Completion (KGC) faces the challenge of not being able to effectively handle the new entities in inductive scenarios. With the development of pre-training language models (PLMs), many researchers have attempted to use PLMs for inductive scenarios KGC and have achieved excellent results. However, these efforts often use extra information that is too redundant. In addition, the semantic information of knowledge graphs (KGs) is significantly different from ordinary natural language sentences, which makes the use of PLMs for processing KGC tasks unsatisfactory. In this paper, we propose a new model called FSPB (Filtered Semantic Path BERT), which filters out redundant information by selecting paths with high relevance to triple and design a Knowledge Semantic Masking Prediction Module tailored to the semantics of KGs combined with comparative learning. FSPB improves the model’s ability to handle new entities and relations in an inductive scenario, in addition to improving ranking performance in the link prediction task. We conducted experiments under both transductive and inductive scenarios on several datasets, and the results show that FSPB outperforms existing state-of-the-art KGC methods on MRR and hit@1.