Research map / Image generation / GANs
Revisit and Outstrip Entity Alignment: A Perspective of Generative Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Frames embedding-based entity alignment as a generative problem and proposes GEEA, which uses a mutual VAE to align multi-modal KG entities and synthesize new entities with concrete features.
Problem
Embedding-based entity alignment relies on a discriminative objective with few seed pairs; existing GAN-based EEA methods suffer from mode collapse and cannot generate new entities with concrete features (neighborhoods, attributes).
Approach
Views EEA through a generative lens, proving GAN-based EEA optimizes part of an ELBO, and introduces a mutual VAE (M-VAE) with four encode-decode flows, distribution matching on latent variables, and prior/post reconstruction so the model can convert entities across KGs and generate new entities from noise.
Key results
- On DBP15K ZH-EN, GEEA reaches Hits@1 .761, Hits@10 .946, MRR .827, versus .723/.924/.796 for NeoEA (MCLEA).
- On FB15K-DB15K, GEEA achieves MRR .450 versus .393 for MCLEA; GEEA_SMALL still outperforms baselines.
- GEEA surpasses the second-best method by 36.1% in Hits@1 when only 10% of alignments are used for training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GEEA generative objectives to MCLEA | MCLEA · DBP15K ZH-EN | Hits@1: .723 -> .761 | Helped |
| add GEEA generative objectives to MCLEA | MCLEA · DBP15K JA-EN | Hits@1: .721 -> .755 | Helped |
| add GEEA generative objectives to MCLEA | MCLEA · DBP15K FR-EN | Hits@1: .717 -> .776 | Helped |
| add GEEA generative objectives to MCLEA | MCLEA · FB15K-DB15K | MRR: .393 -> .450 | Helped |
| add GEEA generative objectives to MCLEA | MCLEA · FB15K-YAGO15K | MRR: .332 -> .393 | Helped |
| remove prediction matching (ablation) | GEEA · DBP15K ZH-EN | Hits@1: .761 -> .045 | Hurt |
| add distribution matching (ablation) | GEEA · DBP15K ZH-EN | Hits@1: .702 -> .761 | Helped |
| add prior reconstruction (ablation) | GEEA · DBP15K ZH-EN | Hits@1: .746 -> .761 | Helped |
| add post reconstruction (ablation) | GEEA · DBP15K ZH-EN | Hits@1: .750 -> .761 | Helped |
| remove prior reconstruction (ablation) | GEEA · DBP15K ZH-EN entity synthesis | PRE: 0.438 -> $\inf$ | Hurt |
Surprising findings
- GEEA still performs alignment without the explicit EEA (prediction matching) loss, per the ablation on DBP15K ZH-EN.
- Removing the distribution match loss collapses alignment performance (Hits@1 falls from .761 to .045).
Limitations
- Higher parameter count than existing EEA methods.
- Entity synthesis quality is weaker on FB15K-DB15K and FB15K-YAGO15K than on DBP15K, attributed to larger KG heterogeneity.
- Image features are constant pretrained embeddings, so generated images are nearest-neighbor retrievals rather than generated pixels.
Experimental setup
Datasets: DBP15K FB15K-DB15K FB15K-YAGO15K OpenEA 100K
Benchmarks: DBP15K FB15K-DB15K FB15K-YAGO15K OpenEA 100K
Metrics: Hits@1 Hits@10 MRR PRE RE FID
Baselines: MUGNN AliNet decentRL EVA MSNEA MCLEA NeoEA (MCLEA)
Builds on: MCLEA variational autoencoder negative sampling GAN-based entity alignment
Closest papers by meaning
- Entity Alignment with Noisy Annotations from Large Language Models (NeurIPS 2024 · similarity 0.86)
- HLMEA: Unsupervised Entity Alignment Based on Hybrid Language Models (AAAI 2025 · similarity 0.86)
- NeuSymEA: Neuro-symbolic Entity Alignment via Variational Inference (NeurIPS 2025 · similarity 0.85)
- Tackling Uncertain Correspondences for Multi-Modal Entity Alignment (NeurIPS 2024 · similarity 0.84)
- Lambda: Learning Matchable Prior For Entity Alignment with Unlabeled Dangling Cases (NeurIPS 2024 · similarity 0.83)
- An Autoregressive Text-to-Graph Framework for Joint Entity and Relation Extraction (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.