Semantic Propagation Method Enhanced by Structure for Link Prediction

In link prediction, text is widely used due to its rich semantic information. However, existing researches have the problems of high information redundancy, large complexity of algorithms, and lack of integration with structure knowledge of knowledge graph. This paper proposes a component-focused Siamese textual encoder CFSTE, which reduces the negative impact of redundant text description by adding a component focusing module and introducing a Siamese structure to reuse the same entities and relationships to reduce the complexity of the encoder. In addition, the structure knowledge of knowledge graph is also jointed through the integration strategy of interpolation prediction. Our proposed model is tested on multiple datasets and achieves state-of-the-art performance on the overall indicators. On the WN18rr, Hits@1, Hits@3 and MRR are increased by 0.22%, 1.63% and 5.28%, respectively. For FB15k-237, Hits@1, Hits@3, Hits@10 and MRR are increased by 8.76%, 4.99%, 3.04% and 4.68%, respectively.