Knowledge Graph Reasoning with Fusion of Entity Semantic and Structural Information

The Graph Attention Network (GAT) assigns different weights to neighboring entities of the target entity through the introduction of an attention mechanism for information aggregation. This allows it to focus more on the local neighborhood of entities while neglecting the topological structure between entities and relationships in the graph. Additionally, after multi-head attention, the output embeddings are simply concatenated or averaged, resulting in independence among attention heads, which fails to capture important semantic information across different attention heads. To address the issue of GAT not fully exploiting entity structural information and semantic information when applied to knowledge graph (KG) reasoning tasks, we propose a knowledge graph reasoning model that integrates entity semantics and structural information (SSKGR). First, we use TransE to represent entities and relationships as embedding vectors in the same space. Second, we introduce an interactive attention mechanism that re-integrates the multi-head attention in GAT into multiple mixed attention heads, enhancing interaction between attention heads to extract richer semantic information for the target entity. Simultaneously, we utilize Relation Graph Convolutional Networks (R-GCN) to extract structural information of entities and learn the output feature vectors of GAT and R-GCN through a weight matrix. Finally, we use ConvKB as a decoder for scoring. Experimental results on knowledge graph datasets Kinship, NELL-995, and FB15K-237 indicate that the SSKGR model outperforms most comparative models.