Dynamic Entity Alignment with Attribute Integration
Abstract: Entity alignment is a key technology for integrating knowledge graphs (KG). However, existing methods assume KG is static and overlook the fact that KG evolves over time. With the growth of KG, previous alignment results may require adjustments, and new equivalent entities need to be found for the newly added entities. Additionally, new entities often have fewer neighbors, which adds difficulty to their alignment. In this paper, we propose DEA-AttrAlign to address these challenges. The core idea is to quickly generate representations for entities based on their neighborhoods. In cases where neighborhood information for new entities is lacking, we propose utilizing entity attributes and semantic facts derived from triplets as additional alignment information. Extensive experiments show that our approach is more effective than methods based on retraining or inductive learning.
Paper Type: Long
Research Area: Machine Learning for NLP
Research Area Keywords: graph-based methods; word embeddings; representation learning; graphical models;
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 574
Loading