关键词: 知识图谱, 邻居编码, 补全研究, 知识表示学习

Abstract: Most knowledge graph completion models transform information into static triples that are easy to process, ignoring the dynamic attributes and information of entities and relations under different semantic conditions, which leads to the deficiency of the ability of the model to analyze and discover contextual information. To solve these problems, this paper proposes a knowledge graph completion model for dynamically assigning attention scores (DASKGC), which adaptively adjusts the matching score for each entity and relationship. The model uses a neighbor information exchange encoder to obtain the role information of entities under different semantics. Then the path-matching method is used to obtain the exact relationship between entities. Finally, the loss function is used to update the correlation scores of triples under different semantics. In the experimental results, the MMR of the proposed DASKGC on the dataset Nell-995 reached 89.5%, and the Hits@1 on the dataset DDB14 and FB15K-237 reached 93.9% and 92.4%, respectively. Other Hits indexes also had a good performance.