Research map / Label-efficient and robust learning / Semi-supervised learning
Pseudo-Label Calibration Semi-supervised Multi-Modal Entity Alignment
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 PCMEA, a semi-supervised multi-modal entity alignment framework combining mutual-information-enhanced cross-modal alignment with momentum contrastive learning and ensemble-based pseudo-label calibration.
Problem
Multi-modal entity alignment between multi-modal knowledge graphs suffers from modal-specific noise in fusion and underuse of unlabeled data under limited seed alignments.
Approach
Maximizes mutual information between joint and uni-modal embeddings to filter modality-specific noise, and couples pseudo-label calibration (dynamic prediction dictionary with agreement across epochs plus data reordering) with momentum-based contrastive learning to reduce error propagation in semi-supervised training.
Key results
- On FB15K-DB15K with 20% seeds, PCMEA reaches Hits@1 0.6763 and MRR 0.7280, above MSNEA's 0.6527 and 0.7080.
- On FB15K-YAGO15K with 20% seeds, PCMEA reaches Hits@1 0.5896 and MRR 0.6460, versus the best baseline MultiJAF at 0.4630 and 0.5540.
- Removing the contrastive learning loss drops FB15K-DB15K Hits@1 from 0.6763 to 0.3610, the largest ablation effect.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PCMEA instead of the strongest reported baseline, MSNEA | PCMEA · FB15K-DB15K (20% seeds) | Hits@1: 0.6527 -> 0.6763 | Helped |
| Use PCMEA instead of the strongest reported baseline, MultiJAF | PCMEA · FB15K-YAGO15K (20% seeds) | Hits@1: 0.4630 -> 0.5896 | Helped |
| Add image modality (ablation) | PCMEA · FB15K-DB15K (20% seeds) | Hits@1: 0.6510 -> 0.6763 | Helped |
| Add bag-of-words embeddings (ablation) | PCMEA · FB15K-DB15K (20% seeds) | Hits@1: 0.5303 -> 0.6763 | Helped |
| Add PLM embeddings (ablation) | PCMEA · FB15K-DB15K (20% seeds) | Hits@1: 0.6502 -> 0.6763 | Helped |
| Add cross-attention (ablation) | PCMEA · FB15K-DB15K (20% seeds) | Hits@1: 0.6521 -> 0.6763 | Helped |
| Add mutual-information loss (ablation) | PCMEA · FB15K-DB15K (20% seeds) | Hits@1: 0.6718 -> 0.6763 | Helped |
| Add alignment loss (ablation) | PCMEA · FB15K-DB15K (20% seeds) | Hits@1: 0.6674 -> 0.6763 | Helped |
| Add contrastive loss (ablation) | PCMEA · FB15K-DB15K (20% seeds) | Hits@1: 0.3610 -> 0.6763 | Helped |
| Add dynamic prediction dictionary (ablation) | PCMEA · FB15K-DB15K (20% seeds) | Hits@1: 0.6647 -> 0.6763 | Helped |
| Add data reordering to pseudo-label calibration (ablation) | PCMEA · FB15K-DB15K (20% seeds) | Hits@1: 0.6727 -> 0.6763 | Helped |
| Add dynamic prediction dictionary and data reordering (ablation) | PCMEA · FB15K-DB15K (20% seeds) | Hits@1: 0.6588 -> 0.6763 | Helped |
Surprising findings
- Removing the image modality only modestly decreases performance (Hits@1 0.6510 vs 0.6763), and the bag-of-words feature matters more than the PLM for relations and attributes.
Limitations
- Relies on manual seed alignments; performance depends on seed ratio.
- Pseudo-label calibration uses a fixed omega=2 and simple agreement check, which is heuristic.
- Experiments limited to two benchmark datasets.
Experimental setup
Datasets: FB15K-DB15K FB15K-YAGO15K
Benchmarks: FB15K-DB15K FB15K-YAGO15K
Metrics: Hits@1 Hits@5 Hits@10 MRR
Baselines: PoE MMEA EVA ACK-MMEA MSNEA MCLEA MultiJAF MEAformer
Builds on: MCLEA MoCo MINE GAT
Closest papers by meaning
- Tackling Uncertain Correspondences for Multi-Modal Entity Alignment (NeurIPS 2024 · similarity 0.92)
- MESED: A Multi-Modal Entity Set Expansion Dataset with Fine-Grained Semantic Classes and Hard Negative Entities (AAAI 2024 · similarity 0.87)
- HLMEA: Unsupervised Entity Alignment Based on Hybrid Language Models (AAAI 2025 · similarity 0.86)
- Entity Alignment with Noisy Annotations from Large Language Models (NeurIPS 2024 · similarity 0.86)
- Tokenization, Fusion, and Augmentation: Towards Fine-grained Multi-modal Entity Representation (AAAI 2025 · similarity 0.85)
- Multiple Heads are Better than One: Mixture of Modality Knowledge Experts for Entity Representation Learning (ICLR 2025 · 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.