End-to-end Rule Learning from Knowledge Graphs by Ensembling Expert Logical Rules

ACL ARR 2025 February Submission7824 Authors

Abstract: Increasing attention has been paid to learning logical rules automatically on knowledge graphs to explain why a missing fact is inferred. Previous approaches focus on directly learning logical rules from numerous instances, overlooking expert rules that are commonly present in practice. Therefore, we examine the problem of incremental rule learning, which aims to learn new rules via ensembling expert logical rules on knowledge graphs. The challenge of rule learning upon expert rules lies in how to preserve the reasoning semantics of expert rules. We present a framework to allow existing end-to-end rule learning approaches to integrate expert logical rules without losing their logical entailments. In more details, we introduce the notion of complete onehop-transformed set of rules to integrate rules into neural networks for single-step reasoning. To preserve all logical entailments of expert rules, we develop an algorithm based on reasoning path extraction and optimized by backward reasoning to compute a complete onehop-transformed set of rules. Experimental results on four benchmark datasets demonstrate that the incorporation of expert rules significantly enhances the performance of link prediction on knowledge graphs.

Paper Type: Long

Research Area: Machine Learning for NLP

Research Area Keywords: Machine Learning for NLP

Languages Studied: English

Submission Number: 7824