PSQE: A Theoretical-Practical Approach to Pseudo Seed Quality Enhancement for Unsupervised Multimodal Entity Alignment

Yunpeng Hong; Chenyang Bu; Jie Zhang; Yi He; Di Wu; Xindong Wu

cs.IRFeb 26, 2026

PSQE: A Theoretical-Practical Approach to Pseudo Seed Quality Enhancement for Unsupervised Multimodal Entity Alignment

Authors: Yunpeng Hong, Chenyang Bu, Jie Zhang, Yi He, Di Wu, Xindong Wu

Organizations: Key Laboratory of Knowledge Engineering with Big Data (the Ministry of Education of China), Hefei University of Technology, Hefei, Anhui, CN · Department of Data Science, College of William and Mary, Williamsburg, VA, USA · College of Computer and Information Science, Southwest University, Chongqing, CN

Abstract

Multimodal Entity Alignment (MMEA) aims to identify equivalent entities across different data modalities, enabling structural data integration that in turn improves the performance of various large language model applications. To lift the requirement of labeled seed pairs that are difficult to obtain, recent methods shifted to an unsupervised paradigm using pseudo-alignment seeds. However, unsupervised entity alignment in multimodal settings remains underexplored, mainly because the incorporation of multimodal information often results in imbalanced coverage of pseudo-seeds within the knowledge graph. To overcome this, we propose PSQE (Pseudo-Seed Quality Enhancement) to improve the precision and graph coverage balance of pseudo seeds via multimodal information and clustering-resampling. Theoretical analysis reveals the impact of pseudo seeds on existing contrastive learning-based MMEA models. In particular, pseudo seeds can influence the attraction and the repulsion terms in contrastive learning at once, whereas imbalanced graph coverage causes models to prioritize high-density regions, thereby weakening their learning capability for entities in sparse regions. Experimental results validate our theoretical findings and show that PSQE as a plug-and-play module can improve the performance of baselines by considerable margins.

Figures & tables

Figure 1. Comparative analysis of three types of pseudo seed generation on the FR-EN dataset for Multimodal Entity Alignment (MMEA): (1) Type I: Unimodal, (2) Type II: Multimodal, and (3) Type III: Distribution-Aware Multimodal. An interesting phenomenon is observed: Type II achieves higher precision than Type I but performs worse in the downstream MMEA task. In contrast, Type III, due to its better distribution balance, yields the best performance.

Figure 2. Overall framework of PSQE. PSQE optimizes the precision and the graph coverage balance (distribution) of pseudo-alignment seeds in three stages to enhance their quality.

Figure 3. Example of an attraction term for Remark 1:

x

and

y

are correct entity pairs. Wrongly aligned seeds can push away vector representations from the correct entity, resulting in variability from the vector of the correct entity.

Figure 4. Examples of a repulsion term of Remark 2:

x^{-}

is a neighboring entity, and red arrows indicate new negative samples. Negative samples were not added to

x_{7}^{-}

x_{9}^{-}

in the left figure. Unbalanced graph coverage leads to focusing on

x_{1}^{-}

x_{2}^{-}

, and

x_{3}^{-}

during optimization and under-optimization of entities in the region around

x_{8}^{-}

, which are difficult to distinguish, e.g., the bottom right corner of Fig. (a).

Model	Unsup	DBP15K ZH-EN			DBP15K JA-EN			DBP15K FR-EN
Model	Unsup	H@1	H@10	MRR	H@1	H@10	MRR	H@1	H@10	MRR
RDGCN (IJCAI, 2019)	×	.708	.846	.746	.767	.895	.812	.886	.957	.911
RPR-RHGT (IJCAI, 2022)	×	.693	-	.754	.886	-	.912	.889	-	.919
DESAlign (ICDE, 2024)	x	.826	.972	.885	.811	.963	.869	.810	.957	.865
EVA (AAAI, 2021)	✓	.783	.898	.825	.861	.946	.892	.914	.972	.936
EVA+PSQE	✓	.823	.930	.861	.873	.955	.904	.928	.977	.948

Table 1. Comparison of applying PSQE to four recent representative open-source unsupervised MMEA methods (EVA, MCLEA, MEAformer, PCMEA) on the DBP15K dataset. ‘MEAformer+PSQE’ denotes the application of PSQE to MEAformer. ‘Unsup’ indicates whether the method is unsupervised or not. It should be noted that the experimental results of EVA, MCLEA, MEAformer, and PCMEA deviate from the original paper in terms of specific values due to the differences in the setting of the initialization seeds.

Method	DW-V1			DW-V2
Method	H@1	H@10	MRR	H@1	H@10	MRR
COTSAE (AAAI, 2020)	.709	.904	.77	.922	.983	.940
EVA	.880	.965	.914	.880	.963	.913
EVA+PSQE	.915	.971	.937	.900	.972	.929
MCLEA	.903	.975	.932	.896	.970	.926
MCLEA+PSQE	.912	.975	.937	.921	.977	.943

Table 2. Comparison results on the DWY15K dataset.

Figure 5. Case study of the results of MEAformer’s comparison on the unsupervised setting of PSQE vs. UVP on the JA-EN dataset, where correct entity pairs are marked with ‘✓’ and incorrect ones with ‘×’.

Figure 6. Ablation study of PSQE on the FR-EN dataset.

	H@1	MRR	H@1	MRR	H@1	MRR
	MEAformer		EVA		MCLEA
PSQE	.892	.921	.878	.907	.879	.911
PSQE (w/o rel)	.889	.919	.870	.901	.877	.908
PSQE (w/o attr)	.891	.921	.871	.902	.877	.909
PSQE (w/o visual)	.732	.795	.603	.667	.668	.735

Table 3. Component analysis for PSQE on MEAformer, EVA, and MCLEA models under the JA-EN dataset, where “w/o rel” denotes the removal of relational information from the PSQE, “w/o attr” denotes the removal of the attribute information and “w/o visual” denotes the removal of the visual information.

Figure 7. The impact of different clustering numbers in Stage I on the PSQE performance on the FR-EN dataset.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Model	ZH_EN	JA_EN	FR_EN
PSQE	82.5	81.0	81.0
PSQE - MIC	66.2	62.5	65.1

Appendix

Table 4. Effect of the MIC module on pseudo seed precision.

	ZH-EN	JA-EN	FR-EN
stage I	75.3	64.3	66.2
stage II	502.4	457.5	458.6
stage III	5.2	4.7	5.4

Appendix

Table 5. Efficiency analysis of PSQE. PSQE runtime (seconds) statistics for generating pseudo-aligned seeds on the three DBP15K datasets.

Method	ZH_EN	JA_EN	FR_EN
MEAformer	0.838	0.887	0.929
MEAformer+PSQE	0.842	0.892	0.932

Appendix

Table 6. Comparison of H@1 accuracy with and without PSQE under fixed pseudo seed sizes.

Cross-Modal Alignment

Cross-modal alignment maps representations from different data types into compatible spaces so models can compare, transfer, and combine information.

See all 440 papers

Silence and Noise: Resolving Suppression and Overvaluation in Multimodal Entity Alignment
Oct 12, 2026Xizhe Zhang, Jingfeng Zhang, Meng-Fen ChiangCross-Modal Alignment Multimodal KGs
School of Computer Science, The University of Auckland, New Zealand · College of Intelligent Robotics and Advanced Manufacturing/Institute of Trustworthy Embodied AI, Fudan University, China · College of Electrical and Computer Engineering, National Yang Ming Chiao Tung University, Taiwan
Knowing When to Trust Images: Reliability-Aware Multi-modal Entity Alignment
Sep 20, 2026Chenxiao Li, Yunhe Feng, Dongfang Liu +3Cross-Modal Alignment Uncertainty-Aware Multimodal Fusion
University of North Texas · Purdue University · Meta Inc
Multi-Perspective Evidence Synthesis and Reasoning for Unsupervised Multimodal Entity Linking
Apr 22, 2026Mo Zhou, Jianwei Wang, Kai Wang +3
The University of New South Wales Sydney, Australia · Shanghai Jiao Tong University Shanghai, China · University of Technology Sydney Sydney, Australia

Oct 12, 2026J/K move · Enter open · S save

Silence and Noise: Resolving Suppression and Overvaluation in Multimodal Entity Alignment

Xizhe Zhang, Jingfeng Zhang, Meng-Fen Chiang

School of Computer Science, The University of Auckland, New Zealand · College of Intelligent Robotics and Advanced Manufacturing/Institute of Trustworthy Embodied AI, Fudan University, China · College of Electrical and Computer Engineering, National Yang Ming Chiao Tung University, Taiwan

Cross-Modal Alignment Multimodal KGs Missing-Modality Learning

Open paper AskPDF arXiv

Multimodal entity alignment (MMEA) is fundamental to integrating heterogeneous knowledge graphs (KGs), yet existing paradigms remain critically vulnerable to incomplete modalities and cross-graph structural asymmetries. This paper exposes a fundamental oversight in current MMEA research: treating feature missingness and topological discrepancies as isolated preprocessing or denoising challenges, which overlooks their destructive coupling during multimodal fusion. We systematically formulate and evaluate two newly identified failure modes: 1) Detail Missing (DM)-induced overvaluation, where the absence of local modality details causes models to overemphasize uninformative, blurred representations; and 2) Context Imbalance (CI)-induced suppression, where topological discrepancies across KGs allow dense structural contexts to overwhelm weaker but highly discriminative modal signals, triggering high-level semantic drift. To break this cycle of degradation, we introduce ALMEA, a generative and calibration framework that establishes a bidirectional feedback loop between low-level feature reconstruction and high-level latent space alignment. Specifically, Cross-Modality Imputation (CMI) synthesizes missing modality details to mitigate overvaluation, while Latent Semantic Calibration (LSC) aligns cross-graph joint embeddings to suppress topology-driven semantic drift. Under sparse supervision, we develop Controlled Iterative Selection (CIS) to adaptively acquire representative and diverse pseudo-labeled pairs, optimizing latent calibration under constrained budgets. Extensive evaluations on standard benchmarks demonstrate that ALMEA consistently outperforms state-of-the-art baselines. At a 20% supervision rate, ALMEA yields substantial absolute improvements of up to 5.20% in MRR and 6.40% in Hits@1, maintaining robust gains across higher supervision thresholds.

Abstract

Cross-Modal Alignment

Explore similar work

Silence and Noise: Resolving Suppression and Overvaluation in Multimodal Entity Alignment

Knowing When to Trust Images: Reliability-Aware Multi-modal Entity Alignment

Multi-Perspective Evidence Synthesis and Reasoning for Unsupervised Multimodal Entity Linking