ADNE: Attentive Description-Neighborhood Embedding for Knowledge Graph Completion

Existing knowledge representation learning models primarily focus on factual triplets in knowledge graphs, failing to leverage entity description texts and neighborhood structural information effectively. This paper proposes ADNE (Attentive Description-Neighborhood Embedding), a joint embedding model based on an attention mechanism. ADNE encodes entity description texts using BERT-whitening to generate isotropic text representations, integrates relation-aware neighborhood expansion and graph memory networks to aggregate topological features, and achieves dynamic fusion of multisource information. Experimental results in the public datasets FB15k-237 and WN18RR demonstrate that ADNE outperforms the baseline CoNE model, reducing the mean rank metric (MR) by 48% and 93%, respectively. Furthermore, ADNE’s attentionbased fusion mechanism effectively balances the contributions of textual and structural features, demonstrating its robustness in handling sparse entity representations and long-tail distribution challenges.