Research map / Probabilistic and scientific modeling / Molecules and proteins
NeuSymEA: Neuro-symbolic Entity Alignment via Variational Inference
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.
Proposes NeuSymEA, a neuro-symbolic entity alignment framework that models joint truth-score probabilities in a Markov random field with weighted cross-KG rules, optimized by variational EM with a neural E-step and logic-deduction-based symbolic M-step.
Problem
Entity alignment between knowledge graphs suffers from substructure heterogeneity and sparsity for symbolic methods, lack of interpretability and uncertainty handling for neural methods, and an exponentially large rule search space for long rules.
Approach
Model the joint probability of all entity-pair alignment indicators in a Markov random field regulated by weighted cross-KG rules, optimize with variational EM (neural model as the variational distribution in E-step, rule weights updated in M-step), and decompose long rules into unit-length sub-rules via logic deduction for efficient inference and explanation.
Key results
- On full DBP15K ZH-EN, NeuSymEA-D reaches 0.801 Hit@1 and NeuSymEA-L 0.785, versus 0.725 for LightEA (neural best baseline) and 0.652 for PRASE.
- With only 1% seed alignments on FR-EN, NeuSymEA-L achieves 0.737 Hit@1 (NeuSymEA-D 0.642), versus 0.480 for EMEA.
- NeuSymEA is claimed to outperform symbolic, neural and neuro-symbolic baselines on both full and condensed DBP15K, with robustness across dataset scale.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NeuSymEA-D instead of LightEA | DBP15K full JA-EN · DBP15K full JA-EN | Hit@1: 0.736 -> 0.806 | Helped |
| Use NeuSymEA-L instead of LightEA | DBP15K full FR-EN · DBP15K full FR-EN | Hit@1: 0.782 -> 0.834 | Helped |
| Use NeuSymEA-D instead of LightEA | DBP15K full ZH-EN · DBP15K full ZH-EN | Hit@1: 0.725 -> 0.801 | Helped |
| Use NeuSymEA-L instead of RREA | DBP15K condensed JA-EN · DBP15K condensed JA-EN | Hit@1: 0.749 -> 0.811 | Helped |
| Use NeuSymEA-L instead of LightEA | DBP15K condensed FR-EN · DBP15K condensed FR-EN | Hit@1: 0.827 -> 0.858 | Helped |
| Use NeuSymEA-D instead of LightEA | DBP15K condensed ZH-EN · DBP15K condensed ZH-EN | Hit@1: 0.770 -> 0.815 | Helped |
| Use NeuSymEA-L instead of LightEA with 1% training pairs | DBP15K FR-EN · DBP15K FR-EN, 1% training pairs | H@1: 0.430 -> 0.737 | Helped |
| Use NeuSymEA-L instead of EMEA with 1% training pairs | DBP15K JA-EN · DBP15K JA-EN, 1% training pairs | H@1: 0.411 -> 0.632 | Helped |
| Use NeuSymEA-L instead of EMEA with 1% training pairs | DBP15K ZH-EN · DBP15K ZH-EN, 1% training pairs | H@1: 0.517 -> 0.676 | Helped |
Surprising findings
- Neural models degrade substantially on the full DBP15K (e.g., Dual-AMN MRR 0.815 to 0.717 on JA-EN), while symbolic models improve on the full version.
Limitations
- Designed only for pairwise alignment between two KGs; multi-KG alignment would require iterative pairwise alignments.
- Rule confidence decreases multiplicatively with rule length, so long rules receive lower confidence scores.
Experimental setup
Datasets: DBP15K OpenEA DBP1M
Benchmarks: DBP15K OpenEA DBP1M
Metrics: Hit@1 Hit@10 MRR precision recall runtime
Baselines: GCNAlign AlignE BootEA RREA Dual-AMN LightEA PEEA PARIS PRASE EMEA
Builds on: variational EM PRASE EMEA LightEA Dual-AMN
Closest papers by meaning
- Entity Alignment with Noisy Annotations from Large Language Models (NeurIPS 2024 · similarity 0.87)
- HLMEA: Unsupervised Entity Alignment Based on Hybrid Language Models (AAAI 2025 · similarity 0.87)
- Logically Consistent Language Models via Neuro-Symbolic Integration (ICLR 2025 · similarity 0.86)
- Improving Soft Unification with Knowledge Graph Embedding Methods (ICML 2025 · similarity 0.85)
- Lambda: Learning Matchable Prior For Entity Alignment with Unlabeled Dangling Cases (NeurIPS 2024 · similarity 0.85)
- Revisit and Outstrip Entity Alignment: A Perspective of Generative Models (ICLR 2024 · similarity 0.85)
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.