Source-linked AI summary
Relation-Aware Entity Alignment for Heterogeneous Knowledge Graphs
Yuting Wu, Xiao Liu, Yansong Feng, Zheng Wang, Rui Yan, Dongyan Zhao
TL;DR
Entity-alignment methods often inadequately capture complex relations in heterogeneous KGs, limiting their representations. RDGCN addresses this with attentive primal–dual graph interactions and gated structural aggregation, and achieves the best reported alignment performance across three real-world datasets while using less training data.
Problem
Existing embedding-based entity-alignment methods often fail to properly capture complex relation information in multi-relational KGs.
Method
RDGCN models attentive interactions between a primal entity graph and dual relation graph, while highway-gated GCN layers integrate neighboring structural information.
Results
RDGCN achieves the best alignment performance across three real-world datasets and uses less training data than state-of-the-art methods.
Takeaways & Limitations
Relation-aware embeddings enable RDGCN to learn better entity representations for alignment, including on entities with triangular structures.
Takeaways & Limitations
RDGCN approximates relation representations from averaged head and tail entity representations because direct relation representations are unavailable with limited training data.
Abstract
from arXiv · showhide
Entity alignment is the task of linking entities with the same real-world identity from different knowledge graphs (KGs), which has been recently dominated by embedding-based methods. Such approaches work by learning KG representations so that entity alignment can be performed by measuring the similarities between entity embeddings. While promising, prior works in the field often fail to properly capture complex relation information that commonly exists in multi-relational KGs, leaving much room for improvement. In this paper, we propose a novel Relation-aware Dual-Graph Convolutional Network (RDGCN) to incorporate relation information via attentive interactions between the knowledge graph and its dual relation counterpart, and further capture neighboring structures to learn better entity representations. Experiments on three real-world cross-lingual datasets show that our approach delivers better and more robust results over the state-of-the-art alignment methods by learning better KG representations.
1 Introduction
Entity alignment integrates complementary heterogeneous knowledge graphs, but existing embedding methods often struggle with complex relation information. RDGCN addresses this by modeling interactions between entity and relation graphs while incorporating neighboring structure, achieving stronger results on three real-world datasets.
- Knowledge graphs organize knowledge as head–relation–tail triples and support applications including question answering, text classification, and recommendation.
- Embedding-based alignment methods reduce manual feature construction and scale to large KGs, but many rely on the constrained assumption head + relation ≈tail.
- Triangular structures expose this limitation because TransE may require mutually contradictory equations for the same entities and relation.Satisfying v1 + ra ≈v2 and v2 + ra ≈v3 implies v1 + 2ra ≈v3, conflicting with v1 + ra ≈v3.
- GCN-based alignment improves representation learning but ignores KG relation labels, while R-GCNs can require excessive parameters for real-world KGs.
- RDGCN uses attentive interactions between primal entity and dual relation graphs, with highway-gated GCN layers to integrate relation and neighboring structural information.
- RDGCN significantly outperforms six recently proposed approaches across three real-world datasets, extending DPGCNN to entity alignment.
2 Related Work
Prior graph-learning and entity-alignment research provides the foundations for RDGCN, while leaving relation-aware multi-graph alignment insufficiently addressed. RDGCN builds on neighborhood-aware edge modeling and directly represents relations through a dual graph.
- Graph neural-network research includes spectral and spatial approaches, with GCNs applied to NLP and R-GCNs designed for relational data.
- RDGCN is presented as the first relation-aware multi-graph learning framework for entity alignment, inspired by DPGCNN’s neighborhood-aware edge features.
- Earlier entity-alignment methods often required expert-designed features or external user-contributed sources, motivating embedding-based alternatives.
- RDGCN directly models relation information by constructing a dual relation graph, improving entity embeddings for more accurate alignment.
3 Problem Formulation
The entity-alignment problem concerns finding equivalent entities across two heterogeneous knowledge graphs using a small set of known aligned pairs as training data.
- A knowledge graph is represented as G = (E, R, T), containing sets of entities, relations, and triples.
- Given two heterogeneous KGs and a small set of equivalent entity pairs, the task is to automatically find additional equivalent entities.
- Known aligned entity pairs serve as alignment seeds and can be used as training data.
4 Our Approach: RDGCN
RDGCN represents entities and relations through iterative attention between a primal entity graph and a weighted dual relation graph, then incorporates neighboring structure with gated GCN layers. Final representations support alignment by distance-based training with challenging negative samples.
- Constructing the Dual Relation Graph: RDGCN constructs a dual relation graph whose vertices represent primal-graph relations and whose edges connect relations sharing heads or tails.Dual-edge weights reflect how likely relations share similar heads or tails; graph construction takes less than two minutes per evaluation dataset.
- Dual-Primal Interaction: Graph attention enables repeated dual-primal interactions, allowing relation representations to influence entity representations and mutual improvement across graphs.The dual attention layer uses primal vertex features, while the primal attention layer uses dual representations corresponding to relations in the primal graph.
- Dual-Primal Interaction: Because direct relation representations are unavailable with limited training data, RDGCN approximates each relation by concatenating averaged head- and tail-entity representations.The approximation is derived from entity representations produced by the previous primal attention layer.
- Primal Attention Layer: Entity-name initialization is mixed with primal-attention outputs to preserve alignment evidence from names while learning relation-aware representations.The final interaction-module representation combines outputs from primal attention layers using weighting parameters.
- Incorporating Structural Information: Two-layer GCNs with highway gates incorporate neighboring structure while controlling noise accumulation and preserving useful relation information.The adjacency matrix adds self-connections, and the primal graph is treated as undirected so information flows in both directions.
- Training: Alignment uses distances between final entity representations, training aligned pairs to be closer than negative pairs under a margin-based objective.Negative instances are selected from K-nearest entities rather than by random sampling.
5 Experimental Setup
The evaluation uses three large-scale cross-lingual DBP15K datasets and the same 30% training, 70% testing split as prior work. RDGCN is compared with four ablation variants and six recent alignment methods.
- Datasets: Three DBP15K datasets pair Chinese, English, Japanese, and French DBpedia versions, with 15K pre-aligned entity pairs per dataset.Each dataset contains two knowledge graphs in different languages.
- Datasets: The experiments use a 30% training and 70% testing split consistent with previous work.The setup evaluates alignment from a relatively small set of training pairs against a larger held-out portion.
- Model Variants: Ablation variants isolate two-layer GCNs, relational GCNs, highway gates, and dual-primal interaction modules.The variants include GCN-s, R-GCN-s, HGCN-s, and RD.
6 Results and Discussion
RDGCN achieves the strongest overall alignment performance while remaining robust with limited prior alignments and handling complex relation structures. Ablations show that both dual-primal interactions and GCN layers contribute complementary information.
- Main Results: RDGCN gives the best performance across all metrics and datasets except Hits@10 on DBP15KZH−EN, where BootEA scores 84.75 versus 84.55.The result is reported across the evaluation datasets and metrics.
- Ablation Studies: GCN-s improves over GCN across datasets, while HGCN-s further boosts Hits@1 by over 30% on DBP15KFR−EN.The reported improvements are associated with noisy machine translations and layer-wise highway gates, respectively.
- Ablation Studies: Removing dual and primal attention layers reduces Hits@1 by 1.1% and Hits@10 by 2.02% on DBP15KZH−EN.The interaction modules integrate approximate relation information with entity information between the dual relation and primal graphs.
- Ablation Studies: Removing GCN layers creates an 8.94% Hits@1 gap between RD and RDGCN on DBP15KZH−EN.The GCN layers capture neighboring structural information, complementing the dual-primal modules’ relation modeling.
- Triangular Structures: RDGCN significantly outperforms BootEA on entities in triangular structures across all three datasets.The comparison uses testing instances containing triangular structures.
- Prior Alignments: 17.79% higher Hits@1 is achieved by RDGCN with 10% prior alignments than BootEA with 40% on DBP15KFR−EN.RDGCN remains consistently strong as the proportion of prior alignments varies from 10% to 40%.
7 Conclusions
The paper presents RDGCN for entity alignment over heterogeneous KGs, combining relation information with neighboring structural information. It achieves the best alignment performance across three real-world datasets while using less training data than state-of-the-art methods.
- RDGCN models attentive interactions between the primal knowledge graph and its dual relation graph.This design incorporates complex relation information into entity alignment.
- Gated GCN layers combine relation information with neighboring structural information to learn better entity representations.
- RDGCN achieves the best alignment performance across three real-world datasets while using less training data than state-of-the-art methods.