Enhancing Knowledge Graph Completion by Extracting Potential Positive Examples

Knowledge graph completion (KGC) aims to infer missing links using known facts. This paper proposes a method based on text descriptions (KG-MPP), which leverages the graph structure information between entities to identify potential positive examples within negative examples, reducing the likelihood of the model learning incorrect information. Furthermore, when evaluating model performance, the influence of structural information between entities on model performance was further explored by increasing the scores of entities that are more likely to be correct answers. Experimental results show that KG-MPP achieves impressive results on the WN18RR dataset, with MRR, Hit@3, and Hit@10 reaching 68.2%, 73.4%, and 83.6%, respectively, setting state-of-the-art levels for Hit@3 and Hit@10. Improvements in multiple performance metrics were also observed on the FB15K-237 dataset.