Organizations: 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
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.
Figures & tables
Figure 1: Overview of DM/CI challenges in MMEA benchmarks and the proposed ALMEA framework. Top: Two illustrative aligned entity pairs with Detail Missing (DM) and Context Imbalance (CI). “x” denotes a missing modality, and CI is measured using the modality-wise contextual discrepancies Δ(g,r,a) between ground-truth pairs across the source and target KGs. Middle: Dataset statistics on FB15K-DB15K and FB15K-YAGO15K: overall DM rates for visual/attribute modalities (left) and CI distributions over all ground-truth pairs (middle/right), grouped into Low/Mid/High regions by the averaged discrepancy score. Bottom: Our solutions to DM and CI. Entities are fused into joint embeddings (hi,hj) with modality weights πm for nearest-neighbor retrieval. Cross-Modality Imputation (CMI) restores missing low-level details via a conditional VAE to mitigate DM-induced overvaluation, whereas Latent Semantic Calibration (LSC) aligns latent semantics across KGs to alleviate CI-induced suppression.
Figure 2: Representative failure cases of multimodal entity alignment under DM and CI. Each row shows modality weights πm ( m∈{v,g,r,a} ) for the source entity ( S ), the top-1 incorrect target match ( T , wrong), and the ground-truth target at its retrieved rank ( T , GT@r). Red “x” denotes a missing modality. Cases 1 and 3 are CI-dominated failures, whereas Case 2 is dominated by DM. Abbreviations: MP = "Member of Parliament", LL = "Lucy Liu", ZQ = "Zachary Quinto", CCA = "Chelsea College of Arts", and ITFC = "Ipswich Town F.C".
FB15K-DB15K (20%)
FB15K-DB15K (50%)
FB15K-DB15K (80%)
Methods
MRR(%)
Hits@1(%)
Hits@10(%)
MRR(%)
Hits@1(%)
Hits@10(%)
MRR(%)
Hits@1(%)
Hits@10(%)
TransE
13.40
0 7.80
24.00
30.60
23.00
44.60
50.70
42.60
65.90
GCN-align
0 8.70
0 5.30
17.40
29.30
22.60
43.50
47.20
41.40
63.50
SEA
25.50
17.00
42.50
47.00
37.30
65.70
50.50
51.20
78.40
MMEA
35.70
26.50
54.10
51.20
41.70
70.30
68.50
59.00
86.90
EVA
18.48
11.90
31.60
26.14
18.06
42.32
35.34
25.98
53.48
Table 1: Cross-MMKG alignment accuracy under 20%, 50%, and 80% supervision. Bold represents two variants and the improvement, while underlining indicates the strongest baseline; ∗ denotes results reproduced from released code, and the shade of each block represents the result of a paired t -test with ALMEA, based on five matched runs: significant , highly significant , extremely significant , and not significant . Improv.: A Comparison of ALMEA and SimDiff
Setting
Methods
DBP15KZH−EN
DBP15KJA−EN
DBP15KFR−EN
MRR(%)
Hits@1(%)
Hits@10(%)
MRR(%)
Hits@1(%)
Hits@10(%)
MRR(%)
Hits@1(%)
Hits@10(%)
w/o SF
BootEA
70.3
62.9
84.7
70.1
62.2
85.4
73.1
65.3
87.4
EVA
81.4
76.1
90.7
81.7
76.2
91.3
84.7
79.3
94.2
MCLEA
86.5
81.6
94.8
86.5
81.2
95.2
88.5
83.4
97.5
SimDiff
87.7
82.9
96.3
88.3
83.3
96.6
90.5
86.1
98.0
ALMEA w/o CIS
81.8
74.7
94.4
81.3
74.5
93.3
82.7
75.9
94.7
Table 2: Multilingual MMKGs Alignment Accuracy Comparison. The results comparison on three bilingual datasets in different settings: without using Surface Forms (w/o SF), with Surface Forms (w/ SF). Under the unsupervised (unsup) setting, ALMEA-V and ALMEA-N use visual and name similarities, respectively, to initialize the alignment space. Bold values highlight the ALMEA variants rather than the column-wise best results.
MCLEA
GEEA
MEAformer
OTMEA
SimDiff
A -
A +
Params (M)
9.06
47.46
10.58
11.02
10.31
14.34
14.34
Iter/s ↑
3.94
0.73
0.74
0.69
0.36
1.28
1.12
Table 3: Training efficiency. A - /A + denote ALMEA without/with CIS; higher Iter/s indicates faster training.
Figure 3: Effectiveness of diversity in iterative learning. t-SNE plots show the sub-dataset (blue) and the CIS-selected subset (red), comparing selections without and with diversity score D .
Variants
FB15K-DB15K
FB15K-YAGO15K
MRR
Hits@1
Hits@10
MRR
Hits@1
Hits@10
ALMEA w/o CIS
84.83
80.45
93.28
80.10
74.48
90.16
Mod.
w/o DMDS
82.80
77.13
90.94
78.43
72.45
88.83
w/o LSC
81.73
76.41
89.89
77.06
71.21
87.48
Loss
w/o LDML
82.96
77.80
91.36
78.33
72.30
88.86
w/o LURL
83.46
78.50
91.51
78.56
72.67
88.99
Table 4: Ablation study. All metrics are reported in %. Results are grouped into ALMEA w/o CIS and ALMEA w/ CIS . ALMEA w/o CIS is tested under 80% supervision, where “Mod.”, “Loss”, and “Mdl.” denote module-, loss-, and modality-level ablations, respectively. ALMEA w/ CIS is tested under 20% supervision. DMDS denotes DM Disturbance Simulation.
Source
Modal information
Neighborhood
Ground Truth
Qantas, Jakarta, Cairns, Darwin_Northern_Territory, Canberra, Brisbane
number_of_runways, longitude, mean_elevation, latitude
Table 5: Qualitative comparison of imputed modalities for Perth Airport on FB15K. For each imputed modality, the nearest neighbors are retrieved in the embedding space. Blue text denotes retrieved items that exactly match the ground truth, while underlined text denotes additional non-ground-truth neighbors returned by KNN.
Figure 4: Sensitivity analysis of the sparsification factor α .
Figure 5: Candidate Pairs Quality Analysis of CIS. True positive, false positive, and duplication rates are shown over τ∈[0,0.2] for α∈{0.1,0.35,1} , without D ((a)–(c)) and with D ((d)–(f)).
Figure 6: Downstream Gains of CIS in Low-Resource Settings.
Data
Variants
20%
50%
80%
MRR
H@1
H@10
MRR
H@1
H@10
MRR
H@1
H@10
DB
M
55.4
46.5
72.2
73.3
66.4
85.4
82.4
77.5
90.5
E
59.1
50.3
75.5
73.5
66.7
85.4
83.0
78.0
90.7
C
59.4
50.8
75.6
74.2
67.4
86.2
83.8
79.1
92.4
YG
M
44.3
35.9
60.4
66.0
58.6
79.4
77.0
71.0
87.3
E
50.2
41.6
66.6
67.1
59.7
80.9
78.0
71.7
88.5
Table 6: CIS selection strategies at 20%, 50%, and 80% supervision. DB and YG denote FB15K-DB15K and FB15K-YAGO15K; M, E, and C denote the margin, entropy, and CIS variants of ALMEA, respectively.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
G
∣E∣
∣R∣
#Triples( R )
A
#Triples( A )
∣V∣
∣GT∣ (FB15K-DB15K)
∣GT∣ (FB15K-YAGO15K)
FB15K
14,951
1,345
592,213
116
29,385
13,444
12,846
11,199
DB15K
12,842
279
89,197
225
48,080
12,837
12,846
–
YAGO15K
15,404
32
122,886
7
23,532
11,194
–
11,199
Appendix
Table 7: Dataset Summary. ∣GT∣ represents the ground-truth entity alignments, indicating the number of aligned entity pairs between the source and target datasets. Specifically, ∣GT∣ (FB15K-DB15K) denotes the number of aligned entity pairs between FB15K and DB15K, while ∣GT∣ (FB15K-YAGO15K) denotes the aligned pairs between FB15K and YAGO15K. Some source entities may lack image modality or a corresponding target entity.
Datasets
Ttotal
Twarm
Tinter
Tr
Titer
batch-size
masked dropping rate
learning rate
sparsification factor α (20%, 50%, 80%)
FB15K-DB15K
750
500
50
5
1000
3,500
0.45
0.001
[0.35, 0.45, 0.95]
FB15K-YAGO15K
550
300
50
5
1000
3,500
0.50
0.001
[0.25, 0.35, 0.45]
Appendix
Table 8: Hyperparameter settings in the main experiments. Training Configuration Settings
FB15K-DB15K
FB15K-YAGO15K
Variant
MRR
Hits@1
Hits@10
MRR
Hits@1
Hits@10
ALMEA w/o CIS
84.83
80.45
93.28
80.10
74.48
90.16
w/o LDML
82.96
77.80
91.36
78.33
72.30
88.86
w/o LURL
83.46
78.50
91.51
78.56
72.67
88.99
w/o LCMAL
83.43
78.25
91.38
78.33
72.45
88.72
Appendix
Table 9: Auxiliary-loss ablation under 80% supervision. All metrics are reported in %.
Dataset & Ratio
MCLEA
MSNEA
GEEA
MEAformer
OTMEA
SimDiff (SoTA)
FB15K-DB15K (20%)
**
**
**
FB15K-DB15K (50%)
**
**
FB15K-DB15K (80%)
—
FB15K-YAGO15K (20%)
**
FB15K-YAGO15K (50%)
**
—
FB15K-YAGO15K (80%)
—
Appendix
Table 10: Summary of statistical significance (paired t-test over 5 runs). Significance levels: * = p<0.05 , ** = p<0.01 , *** = p<0.001 , — = not significant.
# Attribute : iso_numeric , longitude , latitude , area , date_founded , population_number , calling_code , populationDensity
Appendix
Table 11: Qualitative comparison of imputed modalities for Bulgaria on FB15K. For each imputed modality, the nearest neighbors are retrieved in the embedding space. Blue text denotes retrieved items that exactly match the ground truth, while underlined text denotes additional non-ground-truth neighbors returned by KNN.
Table 12: Qualitative comparison of imputed modalities for Mayor on FB15K. For each imputed modality, the nearest neighbors are retrieved in the embedding space. Blue text denotes retrieved items that exactly match the ground truth, while underlined text denotes additional non-ground-truth neighbors returned by KNN.
Table 13: Qualitative comparison of imputed modalities for River Thames on FB15K. For each imputed modality, the nearest neighbors are retrieved in the embedding space. Blue text denotes retrieved items that exactly match the ground truth, while underlined text denotes additional non-ground-truth neighbors returned by KNN.
Figure 7: Random recovery case under Limited Supervision. With only 5% of the supervision and just 2% of the budget, CIS elevated GT’s ranking to top 1.
FB15K-DB15K (20%)
FB15K-DB15K (50%)
FB15K-DB15K (80%)
Methods
MRR(%)
Hits@1(%)
Hits@10(%)
MRR(%)
Hits@1(%)
Hits@10(%)
MRR(%)
Hits@1(%)
Hits@10(%)
ALMEA w/ margin
55.42
46.52
72.24
73.30
66.40
85.39
82.40
77.53
90.54
ALMEA w/ entropy
59.09
50.27
75.45
73.52
66.74
85.40
82.95
77.98
90.74
ALMEA w/ CIS
59.43
50.79
75.64
74.16
67.40
86.16
83.80
79.08
92.41
Appendix
Table 14: CIS Selection Strategy Analysis under Limited Supervision. Results are under 20%, 50%, and 80% valid alignment pairs settings.
FB15K-DB15K (20%)
FB15K-DB15K (50%)
FB15K-DB15K (80%)
Methods
MRR(%)
Hits@1(%)
Hits@10(%)
MRR(%)
Hits@1(%)
Hits@10(%)
MRR(%)
Hits@1(%)
Hits@10(%)
EVA
18.70
12.43
31.83
25.3
17.7
40.5
35.3
25.8
53.3
MCLEA
54.0
44.89
71.54
64.9
56.11
78.87
77.9
70.97
88.17
MSNEA
60.76
52.21
70.39
64.23
56.04
79.89
63.71
55.79
78.62
MEAformer
66.3
57.72
81.27
74.8
68.21
86.49
82.8
77.55
88.92
SimDiff
67.8
60.16
81.89
77.1
71.55
87.21
84.41
80.79
91.01
Appendix
Table 15: CIS under unlimited budget, also be described as iterative learning (IL). Here, ALMEA w/ CIS stands for ALMEA with Controlled Iterative Selection (using a limited budget). ALMEA w/ IL denotes ALMEA with Iterative Learning (use an unlimited budget).
FB15K-DB15K (20%)
Methods
MRR(%)
Hits@1(%)
Hits@10(%)
ALMEA (Imbalance-0.05 × )
55.06
45.73
72.80
ALMEA (Imbalance-0.1 × )
55.51
46.21
73.36
ALMEA (Imbalance-0.5 × )
59.09
50.27
75.76
ALMEA
59.43
50.79
75.64
Appendix
Table 16: CIS robustness under imbalanced candidate spaces. We gradually reduce the number of target entities to simulate the scale difference between the source and target KGs, and report the resulting alignment performance.
The visual modality, i.e., images, plays a key role in multi-modal entity alignment (MMEA). Existing approaches often directly fuse the image with other modalities to align different entities. Although simple, such strategies overlook the potential noise in the images and their semantic misalignment with corresponding entities, resulting in suboptimal fusion and degraded performance. Addressing this, we propose a novel Reliability-Aware framework for MMEA (RA-MMEA), which assesses visual reliability and adaptively improves unreliable visual representations for robust entity alignment. The core lies in two modules, including dependency-aware visual reliability prediction (DA-VRP) and stability-regularized visual embedding generation (SR-VEG). The former aims to estimate the reliability of an image by leveraging multi-modal dependency within the entity, while the latter focuses on producing alternative visual representation conditioned on semantics encoded in textual modalities for multi-modal fusion. Compared to current methods, RA-MMEA enables more reliable visual representations for modality fusion, thereby improving performance. In extensive experiments, RA-MMEA achieves state-of-the-art results, verifying the importance of reliable visual modality for entity alignment and the effectiveness of RA-MMEA. The code and results will be released.
Chenxiao Li, Yunhe Feng, Dongfang Liu +3
University of North Texas · Purdue University · Meta Inc
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.
Yunpeng Hong, Chenyang Bu, Jie Zhang +3
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
Learning effective multimodal entity representations is fundamental for reasoning tasks such as multimodal knowledge graph completion (MMKGC). However, existing methods often suffer from semantic over-smoothing within modalities and ineffective noise filtration across modalities, particularly under sparse or ambiguous conditions. To overcome these limitations, we propose PrismF, a unified framework that synergizes multi-perspective enhancement with progressive fusion to extract stronger signals from diverse inputs. PrismF enhances fine-grained intra-modal semantics through a multi-perspective mechanism that decomposes each modality into complementary views and constrains them with a decoupling loss to reduce representation collapse. Furthermore, it improves cross-modal integration through a progressive fusion strategy that dynamically calibrates inter-modal interactions, enabling the model to emphasize informative signals while suppressing noisy or unreliable ones. Extensive experiments on three public benchmarks show that PrismF achieves the strongest overall performance, including relative improvements of 4.04% in MRR and 11.17% in Hits@1 on KVC16K. Our code can be found at https://github.com/HubuKG/PrismF.
Chenyi Xiong, Yan Zhang, Jing Hu +5
School of Computer Science, Hubei University Wuhan, China · School of Cyber Science and Technology, Hubei University Wuhan, China · Institute of Vocational Education, Guangdong Industry Polytechnic University Guangzhou, China +1