[Paper Review] A Self-supervised Method for Entity Alignment

Xiao Liu, Haoyun Hong+5 more|arXiv (Cornell University)|Jun 17, 2021

Advanced Graph Neural Networks48 references4 citations

TL;DR

This paper proposes SelfKG, a self-supervised contrastive learning method for entity alignment in knowledge graphs that improves performance by pushing unlabeled negative entity pairs apart rather than solely pulling labeled positive pairs together. Experiments show SelfKG achieves performance comparable to state-of-the-art supervised methods without requiring any labeled alignment data.

ABSTRACT

Entity alignment, aiming to identify equivalent entities across different knowledge graphs (KGs), is a fundamental problem for constructing large-scale KGs. Over the course of its development, supervision has been considered necessary for accurate alignments. Inspired by the recent progress of self-supervised learning, we explore the extent to which we can get rid of supervision for entity alignment. Existing supervised methods for this task focus on pulling each pair of positive (labeled) entities close to each other. However, our analysis suggests that the learning of entity alignment can actually benefit more from pushing sampled (unlabeled) negatives far away than pulling positive aligned pairs close. We present SelfKG by leveraging this discovery to design a contrastive learning strategy across two KGs. Extensive experiments on benchmark datasets demonstrate that SelfKG without supervision can match or achieve comparable results with state-of-the-art supervised baselines. The performance of SelfKG demonstrates self-supervised learning offers great potential for entity alignment in KGs.

Motivation & Objective

• To investigate whether supervision can be eliminated in entity alignment by leveraging self-supervised contrastive learning.
• To explore whether pushing negative entity pairs apart improves alignment performance more than pulling positive pairs together.
• To design a contrastive learning framework that operates solely on unpaired, unlabeled entities across two knowledge graphs.
• To evaluate the effectiveness of self-supervised learning in achieving competitive alignment accuracy without human-annotated alignments.

Proposed method

• SelfKG employs a contrastive learning strategy that explicitly pushes sampled negative entity pairs far apart in the embedding space.
• The method constructs positive and negative pairs from two knowledge graphs without relying on any labeled alignments.
• It uses a contrastive objective that maximizes the separation between negative pairs while preserving the proximity of positive pairs.
• The model leverages structural and semantic features from both knowledge graphs to generate meaningful embeddings for contrastive learning.
• The framework is trained end-to-end using a contrastive loss function that encourages discriminative representations for aligned entities.

Experimental results

Research questions

• RQ1 Can self-supervised contrastive learning outperform supervised methods in entity alignment without any labeled data?
• RQ2 Does emphasizing negative sampling (pushing negatives apart) lead to better alignment performance than focusing only on positive pair pulling?
• RQ3 How effective is a contrastive learning approach that relies solely on unpaired entities from two knowledge graphs?
• RQ4 Can a self-supervised method achieve performance comparable to state-of-the-art supervised baselines in entity alignment?

Key findings

• SelfKG achieves performance comparable to state-of-the-art supervised methods on benchmark knowledge graph datasets without using any labeled alignment data.
• The method demonstrates that pushing negative entity pairs apart contributes more significantly to alignment performance than simply pulling positive pairs together.
• Extensive experiments confirm that self-supervised contrastive learning is a viable and effective alternative to supervised training in entity alignment.
• The results indicate that self-supervised learning offers strong potential for large-scale knowledge graph construction with minimal human annotation.