Completing the WIN18RR Dataset Based on the RotatE mode

In this paper, we investigate the problem of predicting missing links in knowledge graphs from entity embeddings and relation embeddings. We use the WN18RR dataset as a standard benchmark for knowledge graph completion, a task that is largely dependent upon the recognition and generalization of intricate and relational data modeling and inference patterns, as well as consolidation of multiple relationship types. This paper proposes a novel model of knowledge graph embeddings, RotatE. In terms of relational data modeling and inference, RotatE has the capacity to address and generalize to a variety of relational data patterns. In RotatE, each relation defines a rotation in a complex vector space to transition from the source entity to the target entity for the knowledge graph embedding. The training process for RotatE embedding model employs self-adversarial negative sampling to improve performance. Additionally, RotatE Cosine Similarity (RoCS) ensures scores are bounded, capturing relational nuances. Experimental results show RotatE's scalability and proficiency in modeling diverse relational patterns. In knowledge graph completion tasks on the WN18RR dataset, the optimized RotatE model significantly outperforms existing models. Compared to RotatE, the optimized model improves Hits@l by 3.5% on WN18RR.