多模态信息与门控注意力协同驱动的知识图谱补全方法OA

Knowledge graph completion aims to infer missing triples based on existing structural information,thereby enhancing the completeness and reasoning capabilities of the graph.Traditional structure-based embedding models perform well in modeling entity-relation interactions but often struggle with sparse entities and complex semantic relations due to limited representational capacity.With the recent advancements in pretrained language models,incorporating textual semantics of entities has emerged as a promising direction.However,effectively integrating structural and semantic infor-mation remains a challenge.To address the insufficient structural modeling and inadequate semantic fusion in existing methods,this paper proposes a hybrid model named HST-KG(hybrid struct text KG-model),which integrates structural information and textual semantics.Specifically,TuckER is employed as the structural encoder to capture entity-relation interaction patterns,while BERT(bidirectional encoder representations from transformers)is utilized to extract contextual semantics from entity and relation descriptions.Furthermore,a gated attention fusion mechanism is designed to dynamically weight structural and semantic embeddings,enabling adaptive multimodal collaborative modeling.Experimental results on two benchmark datasets,FB15K-237 and WN18RR,demonstrate that the proposed HST-KG outperforms the latest semantic-enhanced model ISA-KGC.On FB15K-237,HST-KG achieves improvements of 2.2 and 6.5 percentage points in Hits@1 and Hits@3,respectively;on WN18RR,it achieves gains of 0.7 percentage points in MRR and 0.8 percentage points in Hits@10.These results validate the effectiveness of the proposed method in complex relation modeling and sparse entity prediction.