SPR: A Similar Projection Revisor for Complex Logical Reasoning over Knowledge Graphs

Tingting Wang, Yuxuan Tang, Ruolin Li, Duo Yu, Bowen Feng, Feng Ding, Shuo Yu, Yanming Shen

Abstract: Complex logical reasoning over knowledge graphs (KGs) aims to infer target answers by first-order logic queries. Compared to link prediction or multi-hops reasoning tasks, complex logical reasoning is more related to users need of information retrieval. However, existing complex logical reasoning methods only pay attention to modeling entities, relations or operators, and ignore the assistance of auxiliary information in KGs. To address this issue, we propose Similar Projection Revisor (SPR), a component based on similar triples projection for complex logical reasoning tasks. Firstly, SPR utilizes auxiliary information in KGs, in terms of relations in triples which are as same as the reasoning step. Secondly, SPR revises the bias projection in the current query via projection revising module. As a pluggable component, SPR can be embedded in complex logical reasoning baselines to improve their performance without changing baselines structure. Experimental results on two benchmark datasets demonstrate that SPR can increase baselines performance by 1.2% and 1.1% respectively.

External IDs:dblp:conf/IEEEwisa/WangTLYFDYS24