Knowledge Graph Fusion Using Degree-Aware Iterative Entity Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current entity alignment methods based on structural features are inadequate for real-world knowledge graphs due to their high sparseness and prevalence of long-tail entities with limited structure information, leading to unsatisfactory knowledge graph fusion results.
Innovation Solution
A knowledge graph fusion method utilizing a degree-aware-based co-attention network that combines structure and entity name information through concatenated power-average word vectors, followed by iterative training to enhance entity alignment and improve knowledge graph accuracy and coverage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If entity alignment methods based on structural features are used, then alignment accuracy is improved for dense knowledge graphs, but alignment effectiveness deteriorates for sparse real-world knowledge graphs where most entities have limited structure information
Solution Approach 1:
The patent combines structural similarity and name similarity into a unified entity alignment framework. By merging these two complementary features, the method overcomes the limitation of relying solely on structural information, which is insufficient for long-tail entities in sparse knowledge graphs.
Solution Approach 2:
The patent introduces a degree-aware mechanism that dynamically adjusts the weighting of structural and name similarities based on entity degree. For low-degree entities (long-tail entities), name similarity is weighted more heavily, while for high-degree entities, structural similarity plays a larger role. This parameter adaptation resolves the contradiction between structural alignment accuracy and effectiveness for sparse entities.
2Manufacturing precision
If knowledge graphs with normal distribution are used, then structural feature-based methods achieve best experimental results, but real-world knowledge graphs exhibit long-tail distribution where most entities are sparsely connected
Solution Approach 1:
The patent applies local quality by treating entities with different degrees differently. Instead of using a uniform alignment approach, the method adapts the alignment strategy based on local entity characteristics (degree distribution). For high-degree entities, structural features are emphasized, while for low-degree entities, name features are prioritized, making the method adaptable to real-world long-tail distributions.
Solution Approach 2:
The patent introduces dynamic weighting of structural and name similarities based on entity degree. The importance of different features is not fixed but dynamically adjusted according to the entity's connectivity pattern, enabling the method to adapt to both dense and sparse knowledge graph scenarios.
3Quantity of substance
If only structural information is used for entity alignment, then alignment is effective for entities with rich adjacency information, but alignment fails for entities with little structure information
Solution Approach 1:
The patent introduces name similarity as an intermediary feature that bridges the gap for entities with insufficient structural information. Name similarity serves as a complementary signal that enables alignment of long-tail entities that lack rich adjacency information, while structural similarity continues to drive alignment for well-connected entities.
Data Source
AI summary
Provided is a knowledge graph fusion method based on iterative completion, which includes: obtaining multiple knowledge graphs, and identifying each of entities of the multiple knowledge graphs; performing structure vector representation learning on each of entities to obtain a structure vector of each of entities, and performing entity name vector representation learning on each of entities to obtain an entity name vector of each of entities; determining a structural similarity between the entities according to the structure vector of each of entities, and determining an entity name similarity between the entities according to the entity name vector of each of entities; constructing a degree-aware-based co-attention network, and calculating an entity similarity between fused entities through the degree-aware-based co-attention network; and obtaining a high-confidence entity pair according to the entity similarity between the fused entities, and performing knowledge graph completion by iterative training to obtain fused knowledge graphs.