Balance Relation-Aware Attention Embedding Model for Knowledge Graph Completion

Predicting missing links has great significance in knowledge graph completion because it helps improve the accuracy and completeness of the information in the graph. The recent models employ an attention mechanism to capture the graph network. Previous attention-based models treat similar relation semantics for associated pairs of entities, but relations might have similar names but represent different semantics. Thus, models fail to capture the semantic dissimilarity of relations involved within the triplet entity pair. This paper proposes a new graph attention model capable of effectively learning embeddings by combining relationship features with triplet entity embeddings to balance relationship contributions while capturing meaningful graph structure. The model incorporates relational compositional embeddings to determine distinct attention for neighboring nodes during computation and update the entity embedding. This novel approach outperforms the previously established models on the benchmark dataset.