Degree-Aware Alignment for Entities in Tail

Weixin Zeng, Xiang Zhao, Wei Wang, Jiuyang Tang, Zhen Tan

Abstract

Entity alignment (EA) is to discover equivalent entities in knowledge graphs (KGs), which bridges heterogeneous sources of information and facilitates the integration of knowledge. Existing EA solutions mainly rely on structural information to align entities, typically through KG embedding. Nonetheless, in real-life KGs, only a few entities are densely connected to others, and the rest majority possess rather sparse neighborhood structure. We refer to the latter as long-tail entities, and observe that such phenomenon arguably limits the use of structural information for EA. To mitigate the issue, we revisit and investigate into the conventional EA pipeline in pursuit of elegant performance. For pre-alignment, we propose to amplify long-tail entities, which are of relatively weak structural information, with entity name information that is generally available (but overlooked) in the form of concatenated power mean word embeddings. For alignment, under a novel complementary framework of consolidating structural and name signals, we identify entity's degree as important guidance to effectively fuse two different sources of information. To this end, a degree-aware co-attention network is conceived, which dynamically adjusts the significance of features in a degree-aware manner. For post-alignment, we propose to complement original KGs with facts from their counterparts by using confident EA results as anchors via iterative training. Comprehensive experimental evaluations validate the superiority of our proposed techniques.

Introduction

Knowledge Graphs (KGs) such as DBpedia and Wikidata are essential for modern AI applications, yet they often remain incomplete and siloed. Entity Alignment (EA) aims to bridge these gaps by identifying equivalent entities across different KGs (e.g., matching "Spain" in an English KG to "España" in a Spanish KG).

A significant challenge in current EA research is the "long-tail" problem. Most state-of-the-art models, particularly those utilizing Graph Convolutional Networks (GCNs), rely heavily on the assumption that equivalent entities share similar neighborhood structures. However, in real-world datasets, a vast majority of entities are sparsely connected. When an entity has only one or two neighbors, structural embeddings lack sufficient context for accurate alignment. This research introduces DAT (Degree-aware entity Alignment in Tail) to address this sparsity by leveraging entity names and the "degree" (number of connections) as primary guiding signals.

The DAT Framework

The proposed DAT framework operates in three distinct phases to maximize the utility of available information, especially for tail entities.

The proposed DAT framework, illustrating the three phases: pre-alignment, alignment, and post-alignment.

1. Pre-alignment: Name Representation

To compensate for weak structural data, DAT utilizes entity names. Rather than simple word embeddings, the model employs concatenated power mean word embeddings. By combining various means (arithmetic, geometric, and harmonic), the system captures a semantically rich summary of an entity's name, providing a robust baseline for alignment even when structural context is absent.

2. Alignment: Degree-Aware Co-Attention

The core innovation of DAT is its degree-aware fusion mechanism. The model dynamically calculates weights for structural versus name-based signals. This weighting is guided by the entity's degree: for low-degree "tail" entities, the model places higher trust in the name signal; for high-degree "popular" entities, it relies more on the structural neighborhood.

3. Post-alignment: Iterative KG Completion

Alignment is treated as a multi-round process. Confident matches found in early rounds are used as "anchors" to transfer relations between KGs. This process effectively "thickens" the graph, increasing the degree of previously sparse entities and improving the signal for subsequent training rounds.

Experimental Results

DAT was evaluated on several benchmarks, including SRPRS (which follows real-life degree distributions) and dense datasets like DBP-WD-100K. The results demonstrate significant improvements over 13 state-of-the-art methods.

  • Accuracy Gains: On the EN-FR dataset, DAT achieved a Hits@1 of 75.8%, more than doubling the performance of structural leaders like RSNs (34.8%).
  • Tail Entity Performance: For entities with a degree of 1, DAT's accuracy was approximately four times higher than structural-only methods (57.4% vs 14.3%).
  • Dynamic Weighting Evidence: Experimental data confirms that the model successfully learns to adjust its reliance on structural information based on the entity's connectivity.
Correlation between entity degree and the weight assigned to structural representation (SR) in the EN-FR dataset.

Conclusion

The DAT framework demonstrates that entity names and graph topology are not just secondary features but are deeply intertwined. By recognizing that the importance of structural information is a function of an entity's degree, DAT provides a powerful solution for the long-tail entities that dominate real-world knowledge graphs. Future work may extend this degree-aware fusion to include other attributes such as entity descriptions or visual data.

Summary updated

Abstract

Entity alignment (EA) is to discover equivalent entities in knowledge graphs (KGs), which bridges heterogeneous sources of information and facilitates the integration of knowledge. Existing EA solutions mainly rely on structural information to align entities, typically through KG embedding. Nonetheless, in real-life KGs, only a few entities are densely connected to others, and the rest majority possess rather sparse neighborhood structure. We refer to the latter as long-tail entities, and observe that such phenomenon arguably limits the use of structural information for EA. To mitigate the issue, we revisit and investigate into the conventional EA pipeline in pursuit of elegant performance. For pre-alignment, we propose to amplify long-tail entities, which are of relatively weak structural information, with entity name information that is generally available (but overlooked) in the form of concatenated power mean word embeddings. For alignment, under a novel complementary framework of consolidating structural and name signals, we identify entity's degree as important guidance to effectively fuse two different sources of information. To this end, a degree-aware co-attention network is conceived, which dynamically adjusts the significance of features in a degree-aware manner. For post-alignment, we propose to complement original KGs with facts from their counterparts by using confident EA results as anchors via iterative training. Comprehensive experimental evaluations validate the superiority of our proposed techniques.
BibTeX
@inproceedings{zeng2020degree,
  author = {Weixin Zeng and Xiang Zhao and Wei Wang and Jiuyang Tang and Zhen Tan},
  title = {{Degree-Aware Alignment for Entities in Tail}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2020},
  url = {https://openreview.net/forum?id=FmkoH8BID8Z},
}
Figure