Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Lambda: Learning Matchable Prior For Entity Alignment with Unlabeled Dangling Cases

Hang Yin, Liyao Xiang, Dong Ding, Yuheng He, Yihan Wu, Xinbing Wang, Chenghu Zhou

TL;DR

Experimental results demonstrate that each component contributes to overall performances that are superior to baselines, even when baselines additionally exploit 30\% of dangling entities labeled for training.

Abstract

We investigate the entity alignment (EA) problem with unlabeled dangling cases, meaning that partial entities have no counterparts in the other knowledge graph (KG), and this type of entity remains unlabeled. To address this challenge, we propose the framework \textit{Lambda} for dangling detection and then entity alignment. Lambda features a GNN-based encoder called KEESA with spectral contrastive learning for EA and a positive-unlabeled learning algorithm for dangling detection called iPULE. iPULE offers theoretical guarantees of unbiasedness, uniform deviation bounds, and convergence. Experimental results demonstrate that each component contributes to overall performances that are superior to baselines, even when baselines additionally exploit 30\% of dangling entities labeled for training.

Lambda: Learning Matchable Prior For Entity Alignment with Unlabeled Dangling Cases

TL;DR

Experimental results demonstrate that each component contributes to overall performances that are superior to baselines, even when baselines additionally exploit 30\% of dangling entities labeled for training.

Abstract

We investigate the entity alignment (EA) problem with unlabeled dangling cases, meaning that partial entities have no counterparts in the other knowledge graph (KG), and this type of entity remains unlabeled. To address this challenge, we propose the framework \textit{Lambda} for dangling detection and then entity alignment. Lambda features a GNN-based encoder called KEESA with spectral contrastive learning for EA and a positive-unlabeled learning algorithm for dangling detection called iPULE. iPULE offers theoretical guarantees of unbiasedness, uniform deviation bounds, and convergence. Experimental results demonstrate that each component contributes to overall performances that are superior to baselines, even when baselines additionally exploit 30\% of dangling entities labeled for training.
Paper Structure (33 sections, 38 equations, 7 figures, 13 tables, 1 algorithm)

This paper contains 33 sections, 38 equations, 7 figures, 13 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries and Related Work
  3. Entity Alignment
  4. Positive-Unlabled Learning
  5. Selective Aggregation with Spectral Contrastive Learning
  6. KG Entity Encoder with Selective Aggregation
  7. Spectral Contrastive Learning
  8. Iterative Positive-Unlabeled Learning for Dangling Detection
  9. Experiments
  10. Experiments of iPULE Convergence and Class Prior Estimation
  11. Experiments Unaware of Dangling Entities
  12. Experiments Aware of Dangling Entities
  13. Ablation Studies and Varying Anchor Nodes
  14. Conclusion
  15. Notation
  16. ...and 18 more sections

Figures (7)

  • Figure 1: Examples of dangling entities.
  • Figure 2: The illustration of our framework.
  • Figure 3: Prior estimation GA-DBP15K and DBP2.0. (loss convergence in appendix \ref{['app_7']}).
  • Figure 4: Visualization of entity representations learned by our method on GA16K dataset.
  • Figure 5: The ablation study of entity alignment performance in the consolidated setting on DBP2.0.
  • ...and 2 more figures

Theorems & Definitions (5)

  • proof
  • proof
  • proof
  • proof
  • proof