[Paper Review] Exploring and Evaluating Attributes, Values, and Structures for Entity Alignment

Zhiyuan Liu, Yixin Cao+4 more|arXiv (Cornell University)|Oct 7, 2020

Topic Modeling29 references9 citations

TL;DR

This paper proposes AttrGNN, a graph neural network that jointly models relation and attribute triples for entity alignment, using a dynamic attributed value encoder and subgraph partitioning to prioritize discriminative attributes. It achieves a 5.10% average Hits@1 improvement over 12 baselines on DBP15k under both standard and hard evaluation settings, demonstrating robustness to name bias.

ABSTRACT

Entity alignment (EA) aims at building a unified Knowledge Graph (KG) of rich content by linking the equivalent entities from various KGs. GNN-based EA methods present promising performances by modeling the KG structure defined by relation triples. However, attribute triples can also provide crucial alignment signal but have not been well explored yet. In this paper, we propose to utilize an attributed value encoder and partition the KG into subgraphs to model the various types of attribute triples efficiently. Besides, the performances of current EA methods are overestimated because of the name-bias of existing EA datasets. To make an objective evaluation, we propose a hard experimental setting where we select equivalent entity pairs with very different names as the test set. Under both the regular and hard settings, our method achieves significant improvements ($5.10\%$ on average Hits@$1$ in DBP$15$k) over $12$ baselines in cross-lingual and monolingual datasets. Ablation studies on different subgraphs and a case study about attribute types further demonstrate the effectiveness of our method. Source code and data can be found at https://github.com/thunlp/explore-and-evaluate.

Motivation & Objective

• To address the underutilization of attribute triples in existing entity alignment methods, which often rely only on structural relations.
• To overcome the challenge of assigning varying importance to different attributes and values based on their discriminative power for alignment.
• To mitigate performance overestimation in existing EA benchmarks, which are biased toward name-based matching.
• To propose a more realistic evaluation protocol by constructing a hard test set with semantically equivalent entities having very different names.
• To demonstrate that non-name attributes, especially literal and numerical attributes, can significantly improve alignment performance in practical scenarios.

Proposed method

• Proposes a unified GNN framework, AttrGNN, that jointly encodes relation triples and attribute triples using a shared message-passing mechanism.
• Introduces an attributed value encoder that dynamically learns attention weights for attributes and values based on their alignment relevance.
• Partitions the KG into four subgraphs—Name, Literal, Digital, and Structural—enabling type-specific similarity measurement and aggregation.
• Employs a mean aggregator to propagate alignment signals from informative attributes through relation-based neighborhood aggregation.
• Applies a hard evaluation setting by selecting equivalent entity pairs with highly dissimilar names to simulate real-world alignment challenges.
• Uses BERT-based encoders for entity names and fine-tunes the model end-to-end with a contrastive loss to optimize alignment.

Experimental results

Research questions

• RQ1 Can jointly modeling relation and attribute triples improve entity alignment performance beyond relying solely on structural relations?
• RQ2 How can different types of attributes (e.g., name, literal, numeric) be weighted dynamically based on their discriminative power for alignment?
• RQ3 To what extent does the performance of existing EA models degrade when evaluated on a hard benchmark with non-similar names?
• RQ4 Can subgraph partitioning based on attribute type improve the model’s ability to capture type-specific attribute similarities?
• RQ5 Does the use of non-name attributes significantly enhance alignment in practical scenarios where name matching fails?

Key findings

• AttrGNN achieves a 5.10% average improvement in Hits@1 on DBP15k compared to 12 baselines under both regular and hard evaluation settings.
• The model outperforms NameBERT and other name-based models significantly on the hard test set, demonstrating robustness to name variation.
• The ablation study shows that the Literal and Structural channels perform nearly as well as the Name channel under the hard setting, proving the value of non-name features.
• Removing the graph partitioning strategy (MixAttrGNN) leads to a significant performance drop, confirming the benefit of type-specific processing.
• The Digital attribute channel performs poorly due to challenges in learning numerical comparisons, suggesting a need for specialized number-aware representations.
• Case studies confirm that attributes with high attention weights are those with shared values (e.g., postalabbreviation, flag), while mismatched units (e.g., area in sq. mi vs. sq. km) lead to low attention and are filtered out.