چكيده به لاتين
Knowledge graphs are growing fields in artificial intelligence. These graphs are used in information processing systems. Currently, knowledge graphs are not complete, and the completion of knowledge graphs is one of the research fields on knowledge structures. The task of link prediction in knowledge graphs is to add information to the current knowledge graph by inferring its facts. Link prediction techniques in recent years were significantly accurate by using knowledge graph embedding. Techniques have been proposed in recent years have used deep learning methods that had high computational complexity. These methods are very fast in small graphs, but they have low speed and ample space from RAM in large graphs.
From 2013 to 2019, transitional models, which were fast with little complexity, were introduced. The accuracy of transitional models is significantly lower than deep learning techniques, and these models have been discarded since 2019. In this research, we increased the accuracy of these transitional models so that they become close to deep learning techniques. The importance of this research is to the sustainability of the transitional methods and their high speeds in their needlessness to strong hardware. The proposed method improved the p@10 and MRR on the TransH model, which was 21.1 and 38.6, respectively, by 10.3 and 9.1 on the Freebase knowledge graph, respectively. Additionally, this method improved the two criteria, p@10 and MR in the RotatE model, which was 54.7 and 4274, respectively, for the English word knowledge graph (Wordnet), by 2.3 and 3591, respectively.
