重庆理工大学学报2025，Vol.39Issue(17)：142-149,8.DOI:10.3969/j.issn.1674-8425(z).2025.09.018
MCRF:一种知识图谱问答的多尺度级联推理框架
MCRF:A multi-scale cascading reasoning framework for knowledge graph question answering
摘要
Abstract
Each hop in an automated question-answering multi-hop inference task focuses on different parts of a problem,but current research fails to address the fine-grained dynamic representation problem and capture the relational path.Therefore,this paper designs a multi-scale cascading inference framework MCRF.It is a path-based method,calculating the relationship and entity score of the next hop according to the eigenvector in each hop,and obtaining the path corresponding to the final entity as the inference path after several jumps.The upper and lower levels are connected and information is transmitted through the cyclic self-attention of the problem weight matrix.The upper level of the framework gets the dynamic representation of the problem,and the lower level merges and updates the relational vector on a multi-scale.Experiments conducted on WebQSP,MetaQA and CompWebQ datasets show MCRF increases Hit@1 by 1.3％,0.2％and 0.3％compared with the latest baseline model.关键词
知识图谱问答/多跳推理/关系增强/细粒度捕捉Key words
knowledge graph question answering/multi-hop inference/relationship enhancement/fine-grained capture分类
信息技术与安全科学引用本文复制引用
胡洋,卢玲,郑英豪,方纪祥,王爱娟..MCRF:一种知识图谱问答的多尺度级联推理框架[J].重庆理工大学学报,2025,39(17):142-149,8.基金项目
国家自然科学基金项目(62103070) （62103070）
重庆理工大学研究生创新基金项目(gzlcx20243184) （gzlcx20243184）
