Self-Supervised Entity Alignment Using Relative Similarity Metrics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep representation learning-based methods for entity alignment in knowledge graphs rely heavily on human-labeled supervision, which is costly and biased, making it impractical for large-scale web applications.
Innovation Solution
A self-supervised entity alignment method that uses a pre-trained model to obtain initial embeddings of entities from different graphs, and then learns entity alignment by pushing non-aligned entities far away using a relative similarity metric, without the need for labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep representation learning-based methods use human-labeled supervision for entity alignment, then alignment accuracy is improved, but cost and time consumption increase significantly
Solution Approach 1:
The system performs self-supervised learning by automatically generating supervision signals from the knowledge graph structure itself. The contrastive loss function uses positive samples (aligned entities) and negative samples (non-aligned entities) drawn from the graph data, eliminating the need for external human labeling while maintaining alignment accuracy through self-generated training signals.
2Measurement precision
If deep representation learning-based methods use human-labeled supervision for entity alignment, then alignment accuracy is improved, but cost increases
Solution Approach 1:
The method eliminates costly human labeling by using self-supervised contrastive learning. The system automatically constructs training pairs from the knowledge graph, where aligned entity pairs serve as positive samples and non-aligned pairs serve as negative samples, generating supervision signals at zero marginal cost.
Solution Approach 2:
The approach uses embedding vectors (digital representations) of entities as proxies for actual entity objects. By operating on these copied representations in a unified vector space, the system can perform alignment computations efficiently without requiring expensive human inspection of the actual entities.
3Area of stationary object
If deep representation learning-based methods are applied to web-scale knowledge graphs, then coverage is improved, but supervision signal quality deteriorates due to bias
Solution Approach 1:
The system generates supervision signals autonomously from the knowledge graph structure itself rather than relying on external human labels. By using the graph's own aligned and non-aligned entity pairs as training samples, the method scales to web-scale coverage while avoiding the bias and inconsistency inherent in human labeling.
Data Source
AI summary
A computer-implemented method for entity alignment. The computer-implemented method includes: obtaining a first plurality of initial embeddings of entities of a first graph and a second plurality of initial embeddings of entities of a second graph based on a pre-trained model, wherein the first plurality of initial embeddings and the second plurality of initial embeddings are in a unified space; and learning entity alignment between the first graph and the second graph over the first plurality of initial embeddings and the second plurality of initial embeddings by at least one encoder to push non-aligned entities far away using a relative similarity metric, wherein negative sampling for non-aligned entities of an entity of one of the first graph or the second graph is performed on the one of the first graph or the second graph during the learning.