Основні поняття
The author proposes the concept of tree-like rules to enhance rule-based methods' reasoning abilities by refining chain-like rules. The approach involves transforming chain-like rules into tree-like rules through an effective framework.
Анотація
The content introduces the concept of refining chain-like rules into tree-like rules on knowledge graphs to improve reasoning abilities. Existing studies have primarily focused on learning chain-like rules, limiting their semantic expressions and prediction accuracy. The proposed framework aims to expand the application scope and enhance reasoning abilities by introducing tree-like rules. Experimental comparisons demonstrate that refined tree-like rules consistently outperform chain-like rules in link prediction tasks across various datasets.
Статистика
"Experimental results show that tree-like rules continuously outperform chain-like rules on link prediction tasks for different sources of chain-like rules on different KGs."
"On the UMLS dataset, tree-like rules demonstrate a significant outperformance compared to Anyburl chain-like rules, with an impressive 7.79% improvement in MRR."
"For each variable, branch atoms with top k = 5 scores are selected to refine the rule."
Цитати
"Our refined tree-like rules consistently outperform original chain-like rules on KG reasoning tasks."
"The proposed framework effectively refines chain-like rules from different methods into higher-quality tree-like rules on different knowledge graphs."