[Paper Review] EchoEA: Echo Information between Entities and Relations for Entity Alignment

Xueyuan Lin, E Haihong+2 more|arXiv (Cornell University)|Jul 7, 2021

Advanced Graph Neural Networks28 references4 citations

TL;DR

EchoEA proposes a deeper, four-level self-attention GNN encoder named Echo to enhance entity representation by echoing information between entities and relations, while introducing the Attribute-combined Bi-directional Global-filtered Strategy (ABGS) to generate high-quality semi-supervised training data. The method achieves an average Hits@1 of 96% across three cross-lingual datasets, significantly outperforming state-of-the-art GNN-based methods.

ABSTRACT

Entity alignment (EA) plays an important role in automatically integrating knowledge graphs (KGs) from multiple sources. Recent approaches based on Graph Neural Network (GNN) obtain entity representation from relation information and have achieved promising results. Besides, more and more methods introduce semi-supervision to ask for more labeled training data. However, two challenges still exist in GNN-based EA methods: (1) Deeper GNN Encoder: The GNN encoder of current methods has limited depth (usually 2-layers). (2) Low-quality Bootstrapping: The generated semi-supervised data is of low quality. In this paper, we propose a novel framework, Echo Entity Alignment (EchoEA), which leverages 4-levels self-attention mechanism to spread entity information to relations and echo back to entities. Furthermore, we propose attribute-combined bi-directional global-filtered strategy (ABGS) to improve bootstrapping, reduce false samples and generate high-quality training data. The experimental results on three real-world cross-lingual datasets are stable at around 96\% at hits@1 on average, showing that our approach not only significantly outperforms the state-of-the-art GNN-based methods, but also is universal and transferable for existing EA methods.

Motivation & Objective

• To address the limitations of shallow GNN encoders (typically 2-layer) in entity alignment, which suffer from oversmoothing and limited receptive fields.
• To overcome low-quality bootstrapping in semi-supervised entity alignment, where noisy false positives and negatives degrade model performance.
• To enhance the interaction between entities and relations by enabling bidirectional information flow through a deeper architecture.
• To develop a robust, universal, and transferable framework that improves both representation learning and data efficiency in cross-lingual entity alignment.

Proposed method

• Design a four-level self-attention mechanism in the Echo encoder to propagate entity information to relations and echo it back, achieving a 5-layer GNN depth beyond typical 2-layer models.
• Introduce the Attribute-combined Bi-directional Global-filtered Strategy (ABGS), which uses both local and global alignment to filter and refine both positive and negative bootstrapped samples.
• Leverage attribute similarity and value similarity to enhance the quality of initial seed alignments and guide iterative bootstrapping.
• Apply global filtering to remove low-confidence predictions, reducing false positives and false negatives during iterative training.
• Use a symmetric message-passing mechanism that models the mutual contribution of entities and relations, improving representation learning.
• Combine the Echo encoder with an iterative semi-supervision loop that dynamically expands the training set with high-confidence predictions.

Experimental results

Research questions

• RQ1 Can a deeper GNN architecture with multi-level attention improve entity alignment performance beyond the typical 2-layer GNNs?
• RQ2 Does incorporating both positive and negative samples in bootstrapping lead to better generalization and reduced noise compared to methods that only use positive samples?
• RQ3 Can global and local alignment signals be effectively combined to filter and improve the quality of generated training data in iterative entity alignment?
• RQ4 How does the Echo model’s bidirectional information flow between entities and relations affect representation learning and alignment accuracy?
• RQ5 To what extent does the ABGS strategy reduce false positive and false negative rates in bootstrapped training data compared to state-of-the-art methods?

Key findings

• EchoEA achieves an average Hits@1 of approximately 96% across three real-world cross-lingual knowledge graph datasets, significantly outperforming existing GNN-based methods.
• The model reaches a peak Hits@1 of 97.91% on the FR_EN dataset, demonstrating strong performance on challenging cross-lingual alignment tasks.
• The ABGS strategy reduces both the false positive rate and false negative rate by about 50% compared to the state-of-the-art method MRAEA, indicating higher-quality bootstrapped data.
• The Echo encoder with four-level attention outperforms RAGA, even when the Echo component alone (w/o ABGS) surpasses RAGA’s performance, proving the effectiveness of the architecture.
• Ablation studies confirm that the PAN, CAN, and EN components of Echo contribute significantly to performance, with PAN being the most important, followed by CAN and EN.
• The entity utilization rate in EchoEA reaches around 95% over 100 training epochs, indicating efficient use of generated bootstrapped samples, and consistently outperforms MRAEA in both data quality and coverage.