ASGEA: Exploiting Logic Rules from Align-Subgraphs for Entity Alignment

TL;DR

This work tackles interpretability in cross-KG entity alignment by introducing Align-Subgraph Entity Alignment (ASgEa). It combines an Align-Subgraph extraction mechanism with a Path-based Graph Neural Network (ASGNN) and a node-level multi-modal attention module to mine and utilize cross-graph logical rules for EA and MM EA. The approach provides interpretable alignment paths and rule-based scoring, yielding state-of-the-art results across multiple benchmarks and modalities, while also offering visualizations of learned rules. The results demonstrate the practical impact of incorporating logic rules and multi-modal information for more accurate and transparent knowledge graph integration.

Abstract

Entity alignment (EA) aims to identify entities across different knowledge graphs that represent the same real-world objects. Recent embedding-based EA methods have achieved state-of-the-art performance in EA yet faced interpretability challenges as they purely rely on the embedding distance and neglect the logic rules behind a pair of aligned entities. In this paper, we propose the Align-Subgraph Entity Alignment (ASGEA) framework to exploit logic rules from Align-Subgraphs. ASGEA uses anchor links as bridges to construct Align-Subgraphs and spreads along the paths across KGs, which distinguishes it from the embedding-based methods. Furthermore, we design an interpretable Path-based Graph Neural Network, ASGNN, to effectively identify and integrate the logic rules across KGs. We also introduce a node-level multi-modal attention mechanism coupled with multi-modal enriched anchors to augment the Align-Subgraph. Our experimental results demonstrate the superior performance of ASGEA over the existing embedding-based methods in both EA and Multi-Modal EA (MMEA) tasks.

Figures (8)

• Figure 1: (i) The entities "Maurice Sendak" in $\mathcal{G}_1$ and "Bill Evans" in $\mathcal{G}_2$ exhibit structurally similar neighbor relationships but they are not candidates for alignment. (ii) Anchors specifically link "Maurice Sendak" instances across the KGs without connecting "Bill Evans" in $\mathcal{G}_2$ to "Maurice Sendak" in $\mathcal{G}_1$. (iii) Our AsgEa is employed to filter out non-relevant neighbors, retaining only essential alignment information .
• Figure 2: The overall framework of AsgEa. On the left, the ASG Extraction is utilized for constructing entity-specific Align-Subgraphs during training and testing. On the right, the AsGnn is designed for mining alignment rules and scoring, while the Align modal score is employed for matching aligned attributes to subsequently obtain node-level multi-modal scores.
• Figure 3: Illustration of different alignment rules.
• Figure 4: An example for dynamic merge. The yellow portion denotes ASG $g(e_u, e_v)$ while the red portion denotes ASG $g(e_u, e_v')$. The overlapping area signifies the parts that can be reused through dynamic programming.
• Figure 5: Hits@1 and MRR performance on FBDB15K and FBYG15K with different $R_{sa}$. "-MM" indicates using multi-modal information(abbreviated as MM), while "-Stru" indicates reliance on graph structure alone (i.e., triples). "w/o value" indicates without attribute values. For complete results and baseline details, see Appendix.