摘要： 将语义数据流处理引擎与知识图谱嵌入表示学习相结合，可以有效提高实时数据流推理查询性能，但是现有的知识表示学习模型更多关注静态知识图谱嵌入，忽略了知识图谱的动态特性，导致难以应用于实时动态语义数据流推理任务。为了使知识表示学习模型适应知识图谱的在线更新并能够应用于语义数据流引擎，建立一种基于改进多嵌入空间的动态知识图谱嵌入模型PUKALE。针对传递闭包等复杂推理场景，提出3种嵌入空间生成算法。为了在进行增量更新时更合理地选择嵌入空间，设计2种嵌入空间选择算法。基于上述算法实现PUKALE模型，并将其嵌入数据流推理引擎CSPARQL-engine中，以实现实时语义数据流推理查询。实验结果表明，与传统的CSPARQL和KALE推理相比，PUKALE模型的推理查询时间分别约降低85%和93%，其在支持动态图谱嵌入的同时能够提升实时语义数据流推理准确率。

关键词: 语义数据流, 数据流引擎, 推理, 知识表示学习, 知识图谱

Abstract: The joint use of semantic data flow processing engine and knowledge graph embedding representation learning can effectively improve the performance of real-time data stream reasoning and query.The existing knowledge representation learning models pay more attention to static knowledge graph embedding, but ignore the dynamic features of knowledge graphs, so they are not well suited for real-time dynamic semantic data flow reasoning.In order to make the knowledge representation models adaptable to online update of knowledge graphs and applicable to semantic data flow engine, this paper proposes a dynamic knowledge graph embedding model named PUKALE based on improved multi-embedding space.For complex reasoning scenarios such as transitive closure, three algorithms are proposed for embedding space generation.Then two algorithms are proposed to optimize the selection of embedding space during incremental update.On the basis of these algorithms, the PUKALE model is realized and embedded into the CSPARQL-engine for real-time semantic data flow reasoning and query.The experimental results show that compared with traditional reasoning engines such as CSPARQL and KALE, the proposed PUKALE engine reduces the reasoning and query time by about 85% and 93%.It supports dynamic graph embedding, and can improve the accuracy of real-time semantic data flow reasoning.

Key words: semantic data flow, data flow engine, reasoning, knowledge representation learning, knowledge graph

中图分类号: 

• TP391