Paper deep dive
Knowing When to Trust Images: Reliability-Aware Multi-modal Entity Alignment
Chenxiao Li, Yunhe Feng, Dongfang Liu, Dong Nie, Yan Huang, Heng Fan
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/27/2026, 2:30:50 AM
Summary
The paper introduces RA-MMEA, a novel framework for Multi-modal Entity Alignment (MMEA) that addresses the issue of unreliable visual data. It employs a Dependency-Aware Visual Reliability Prediction (DA-VRP) module to assess image reliability based on multi-modal dependencies and a Stability-Regularized Visual Embedding Generation (SR-VEG) module to generate alternative visual representations for unreliable or missing images, thereby improving alignment performance.
Entities (7)
Relation Signals (8)
RA-MMEA → solvesproblem → MMEA
confidence 98% · we propose a novel Reliability-Aware framework for MMEA (RA-MMEA)
RA-MMEA → containsmodule → SR-VEG
confidence 95% · and stability-regularized visual embedding generation (SR-VEG)
RA-MMEA → containsmodule → DA-VRP
confidence 95% · The core lies in two modules, including dependency-aware visual reliability prediction (DA-VRP)
RA-MMEA → evaluatedon → FB15K-DB15K
confidence 90% · We validate the effectiveness... on... FB15K-DB15K
RA-MMEA → evaluatedon → DBP15K
confidence 90% · and three bilingual DBP15K datasets
SR-VEG → performstask → visual_embedding_generation
confidence 90% · SR-VEG... focuses on producing alternative visual representation
DA-VRP → performstask → visual_reliability_prediction
confidence 90% · DA-VRP... aims to estimate the reliability of an image
RA-MMEA → usesmodel → CLIP
confidence 85% · we apply a pre-trained CLIP model... to extract the visual embedding
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2609.23267v1
- Canonical: https://arxiv.org/abs/2609.23267v1
Trouble viewing inline? Open PDF directly →
Full Text
63,475 characters extracted from source content.
Expand or collapse full text
Knowing When to Trust Images: Reliability-Aware Multi-modal Entity Alignment Chenxiao Li Yunhe Feng Dongfang Liu Dong Nie Yan Huang Heng Fan Affiliation: University of North Texas Affiliation: Purdue University Affiliation: Meta Incchenxiaoli@my.unt.edu heng.fan@unt.edu Abstract 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. 1 Introduction Figure 1: Comparison of existing methods directly fusing image and other modalities in (a) and RA-MMEA that first assesses the reliability of the image and then performs fusion in (b). Best viewed in pdf for all figures. Multi-modal knowledge graphs (MMKGs) employ heterogeneous information such as image, text, and structure to represent real-world entities. Yet, independently constructed MMKGs often contain duplicated or semantically overlapping entities, resulting in fragmented knowledge across graphs. To address this issue, multi-modal entity alignment (MMEA) (Chen et al., 2020; Liu et al., 2019) has been introduced to identify equivalent entities across different MMKGs by leveraging multi-modal information, thus consolidating fragmented MMKGs into unified and complete knowledge resources. In MMEA, visual modality, i.e., images, offers rich appearance information and is crucial for aligning entities. Existing models typically directly fuse the image and other modalities to construct entity representation (see Fig. 1 (a)). However, images in MMKGs are not always reliable. They may be irrelevant to the entity or exhibit a large semantic gap from the text. Directly incorporating such noisy images into multi-modal fusion may compromise the entity representation and consequently degrade alignment. To alleviate this, a straightforward solution is to filter out unreliable images using manually annotated class conflict dictionaries (CCDs) (Shi et al., 2022). However, constructing CCDs entails significant labeling efforts, which is laborious and expensive. Besides, manually designed class conflict rules in CCDs inevitably introduce biases and struggle to generalize to real-world scenarios. In this paper, we propose to learn visual reliability from the entity itself and adaptively handle unreliable images before multi-modal fusion, and introduce a novel Reliability-Aware framework for MMEA (RA-MMEA), that assesses visual reliability and adaptively improves unreliable visual representations for robust entity alignment (see Fig. 1(b)). The key intuition is that image reliability can be inferred from its consistency with the semantic context of an entity, including graph, relation, and attribute. Based on this intuition, we design a dependency-aware visual reliability predictor (DA-VRP) in RA-MMEA, which estimates image reliability from global multi-modal dependency. Reliable images are expected to form coherent dependency with the remaining entity modalities, whereas unreliable images tend to deviate from such dependency and behave as outliers in the multi-modal representation space. DA-VRP first calculates the dependency score of the image with other modalities through dependency relation modeling, and then uses such dependency score to learn a binary reliability predictor that identifies and suppresses unreliable visual modalities. To enable end-to-end optimization with discrete binary indicators, we employ Gumbel-Softmax to predict the binary reliability in a differentiable manner. After filtering unreliable images, the suppressed visual information needs to be compensated. To this end, we present a simple yet effective stability-regularized visual embedding generation module in RA-MMEA, termed SR-VEG, that aims to produce alternative visual representation conditioned on other modalities for multi-modal fusion. Specifically, it adaptively integrates features from other modalities and employs latent-variable modeling to generate visual representations. To preserve dependency consistency under modality removal, we further present a weight-difference penalty that aligns dependency scores computed with and without the visual modality. This regularization stabilizes the overall dependency estimation within an entity, leading to better reconstructed visual features and thereby improving entity alignment. To the best of our knowledge, RA-MMEA is the first method that explicitly models visual reliability in MMEA. Compared with existing methods, RA-MMEA jointly estimates visual reliability and selectively reconstructs only unreliable visual representations before multi-modal fusion. This reliability-aware paradigm prevents unreliable visual information from degrading entity representations while preserving reliable visual information. Extensive experiments on multiple benchmarks demonstrate that RA-MMEA consistently achieves state-of-the-art performance, highlighting the importance of reliability-aware visual modeling for robust entity alignment. In summary, our contributions are as follows: ♠ We propose RA-MMEA, the first reliability-aware framework for MMEA; ♥ We present DA-VRP that leverages global multi-modal dependency within an entity for visual reliability prediction; ♣ We introduce SR-VEG that generates virtual visual embeddings under stability-regularized consistency constraints; ♠ On multiple benchmarks, our RA-MMEA achieves state-of-the-art results. 2 Related Work Multi-modal Entity Alignment (MMEA). Typical MMEA methods primarily focus on improving multi-modal fusion to optimize embedding representations (Zhang et al., 2024). To mitigate noise arising from modality heterogeneity, existing studies (Shi et al., 2022; Li et al., 2025b) employ annotated ontology matching as a filtering criterion to enhance embedding quality. To handle uncertain missing values in fusion, another line of work (Chen et al., 2023b; Li et al., 2023b; Li et al., 2024) proposes generating pseudo-representation for the missing modality in multi-modal fusion. In contrast to these strategies that regulate input data quality or quantity, methods such as (Lin et al., 2022; Chen et al., 2022; Chen et al., 2023a; Li et al., 2025a) adopt contrastive learning or Transformer to precisely model inter-modal relationships, thereby alleviating the adverse effects of multi-modal information imbalance. Despite promising results, the aforementioned methods overlook the importance of addressing unreliable visual modality to improve MMEA. LLMs in Multi-modal Entity Alignment. Large language models (LLMs) have been recently employed for MMEA due to their strong reasoning and understanding capabilities. Chen et al. (2024) employ LLMs to generate and embed attribute abstracts, facilitating attribute alignment. Ji et al. (2025) project multi-modal embeddings based on LLM-based attribute summaries to obtain the potential associations between them. Lu et al. (2025) propose LLM-guided visual alignment and attribute enhancement to achieve semantic filtering and enhancement in MMEA. Li et al. (2026) resolve dual-level noisy correspondence via a dedicated two-fold principle and employ MLLMs for relational reasoning. Despite excellent performance, these LLM-based models may introduce potential information leakage by relying on external world knowledge and entity names. Differently, instead of using external knowledge for MMEA, this work focuses on assessing and leveraging the reliability of visual modality based on the entity itself, aiming to improve multi-modal fusion under unreliable images. 3 Methodology Figure 2: Illustration of the proposed RA-MMEA, which assesses visual modality reliability and adaptively improves unreliable visual representations through generated feature for MMEA. (a) Inputs with image, graph, relation, and attribute; (b) Multi-modal Embedding; (c) Dependency-Aware Visual Reliability Prediction; (d) Stability-Regularized Visual Embedding Generation; and (e) Multi-modal Fusion. Problem Formulation. In MMEA, a multi-modal knowledge graph (MMKG) is formalized as G=(E,R,A,T,I,P)G=(E,R,A,T,I,P), where E,R,A,TE,R,A,T, and I are the sets of entities, relations, attributes, triples, and images, and P=(e,i)∣e∈E,i∈IP=\(e,i) e∈ E,i∈ I\ denotes the set of entity-image pairs. Each entity is associated to multiple attributes and 0 ∼ 1 image. Given two MMKGs G1=(E1,R1,A1,T1,I1,P1)G_1=(E_1,R_1,A_1,T_1,I_1,P_1) and G2=G_2= (E2,R2,A2,T2,I2,P2)(E_2,R_2,A_2,T_2,I_2,P_2), the set of seed alignments across them is defined as H=(e1,e2)∣e1∈E1,e2∈E2,e1≡e2H=\(e_1,e_2) e_1∈ E_1,e_2∈ E_2,e_1≡ e_2\, where ≡ represents the equivalence of two entities. Given a set of pre-aligned entity pairs as training guidance, MMEA aims to match the counterpart entities e1e_1 and e2e_2 describing the same concepts in the real world from distinct multi-modal knowledge graphs. Overview In this paper, we propose RA-MMEA, a reliability-aware framework for robust multi-modal entity alignment. As shown in Fig. 2, given an entity containing visual image, graph structure, relations, and attributes, we first extract their embedding (Sec. 3.1). Afterwards, DA-VRP (Sec. 3.2) is applied to predict the reliability of the image. If the image is reliable, we perform multi-modal fusion (Sec. 3.4) to obtain the entity representation; otherwise, we discard the unreliable visual embedding and apply SR-VEG (Sec. 3.3) to produce a virtual visual embedding for modality fusion (Sec. 3.4). 3.1 Multi-modal Knowledge Embedding In this section, we describe embedding extraction for the visual image, graph structure, relations, and attributes of the entity eke_k. Visual Embedding. Given a visual image of the k-th entity eke_k, we apply a pre-trained CLIP model Radford et al. (2021), followed by fully connected (FC) layers, to extract the visual embedding hkvh_k^v. Graph Structure Embedding. For the graph structure of eke_k, we follow Li et al. (2025a); Chen et al. (2023a) and use graph attention network Veličković et al. (2018) with two attention heads and two layers to extract the graph embedding hkgh_k^g. Relation Embedding. Following Yang et al. (2019); Chen et al. (2023a), for the relations of eke_k, we utilize bag-of-words followed by FC layers to extract the relation embedding hkrh_k^r. Attribute Embedding. For the attribute modality of eke_k, we first employ bag-of-words and BERT Devlin et al. (2019) to generate two attribute embeddings, respectively. These embeddings are then combined through learnable weights and FC layers to obtain the final attribute embedding hkah_k^a. 3.2 Dependency-Aware Visual Reliability Prediction (DA-VRP) In RA-MMEA, we propose DA-VRP to estimate whether the visual modality of an entity is reliable under the global multi-modal context. The key intuition is that reliable images should exhibit coherent semantic dependency with the remaining modalities of an entity, and thus have relatively high dependency score with other modalities. Dependency Relation Modeling (DRM). Specifically, we first perform dependency relation modeling (DRM) to compute the global relevance of each modality with the remaining modalities. More concretely, drawing inspiration from the Transformer architecture Vaswani et al. (2017), we first apply the scalable dot-product to calculate the relevance score with normalization between modalities m and j, as follows, smj=exp(ϕ(hkm)⊤ϕ(hkj)/d)∑n∈Mexp(ϕ(hkm)⊤ϕ(hkn)/d)s_mj= (φ(h_k^m) φ(h_k^j)/d)Σ _n∈ M (φ(h_k^m) φ(h_k^n)/d) (1) where smjs_mj is the relevance score between modality m∈Mm∈ M and j∈Mj∈ M, and M=v,g,r,aM=\v,g,r,a\ with v, g, r, a representing the visual image, graph structure, relation, and attribute modalities. d is a scaling factor, and ϕ(⋅)φ(·) the linear projection. Afterwards, we aggregate all relevance scores of each modality m with normalization to compute its dependency score within the entity eke_k, as follows, wkm=exp(∑j∈Msmj)∑n∈Mexp(∑j∈Msnj)w^m_k= (Σ _j∈ Ms_mj)Σ _n∈ M (Σ _j∈ Ms_nj) (2) where wkmw^m_k denotes the dependency score for modality m. After this, we employ the dependency score wkvw^v_k of visual image to predict the visual reliability. Visual Reliability Prediction. The dependency score wvk∈[0,1]w^k_v∈[0,1] reflects the semantic consistency between the visual modality and the remaining modalities of entity eke_k, and we use it as a task-level proxy for visual reliability. This is motivated by the fact that cross-modal dependency provides a differentiable approximation of semantic consistency, a principled indicator of whether visual information agrees with the overall entity semantics. A higher dependency score indicates that the visual representation is more coherent with the overall entity semantics. In contrast, a lower score suggests that the image may be noisy, semantically inconsistent, or provide limited visual evidence for representing the corresponding entity. To predict the visual reliability, we first construct a probability vector based on wvkw^k_v, as follows, pkv=[1−wkv,wkv]p^v_k=[1-w^v_k,w^v_k] (3) where pk,0v=1−wvkp^v_k,0=1-w^k_v denotes the probability that the image is unreliable, and pk,1v=wkvp^v_k,1=w^v_k represents the probability that the image is reliable. To obtain the binary reliability prediction, a straightforward way is to apply the Bernoulli function on pk,0vp^v_k,0 and pk,1vp^v_k,1. However, this operation is non-differentiable, preventing the entire pipeline from being trained in an end-to-end manner. To handle this, we adopt Gumbel-Softmax Jang et al. (2017); Maddison et al. (2017) to sample from the categorical distribution, as follows, gkv=Gumbel-Softmax(pkv)g^v_k=Gumbel-Softmax (p^v_k ) (4) where gkvg^v_k denotes a binary one-hot vector representing the predicted visual reliability, with gkv=[0,1]g^v_k=[0,1] indicating that the image is reliable and gkv=[1,0]g^v_k=[1,0] indicating that the image is unreliable. By employing the Gumbel-Softmax trick, the entire framework becomes fully end-to-end trainable, thereby improving entity alignment performance. 3.3 Stability-Regularized Visual Embedding Generation (SR-VEG) To compensate for the loss of visual information caused by filtering the unreliable visual embedding, we propose SR-VEG to generate semantically consistent virtual visual embeddings from the remaining reliable modalities. Besides handling filtered visual modalities, it can also naturally generalize to entities that inherently lack visual information. Specifically, we first construct a hybrid representation from the non-visual modalities, including graph structure, relation, and attribute embeddings, with the learned dependency scores from DA-VRP, as follows, hkhyb=⊕m∈T[wkm⋅hkm],T=g,r,ah_k^hyb= _m∈ T [w_k^m· h_k^m ], T=\g,r,a\ (5) Here, ⊕ denotes concatenation. By incorporating modality-aware weights, SR-VEG preserves cross-modal dependency during visual embedding generation. However, directly removing the visual modality may perturb the original multi-modal dependency structure. Although unreliable visual information should be suppressed, its removal can still affect dependency estimation among the remaining modalities, leading to unstable modality weights and potentially degrading the generated visual representation. To alleviate this issue, we introduce a stability regularization term to constrain the discrepancy between dependency weights estimated before and after visual modality removal. For entities whose visual modalities are filtered, we first recompute dependency scores w^km w_k^m (m∈g,r,am∈\g,r,a\) with the embeddings hkgh_k^g, hkrh_k^r, and hkah_k^a using DRM in DA-VRP, and then compare them with the original dependency weights wkgw_k^g, wkrw_k^r, and wkaw_k^a. We retain only positive deviations using a ReLU operation via ℒp=∑m∈Tmax(w^km−wkm,0)L_p=Σ _m∈ T ( w_k^m-w_k^m,0 ) (6) Here, the penalty ℒpL_p does not directly construct the hybrid representation for generation. Instead, it acts as an auxiliary stability constraint that regularizes dependency estimation under modality perturbation. By discouraging excessive dependency shifts after visual removal, SR-VEG maintains more stable cross-modal relationships and improves the generated visual embeddings. After obtaining hkhybh_k^hyb, we use a VAE-style framework Kingma and Welling (2014); Sohn et al. (2015) to generate the pseudo-visual embeddings. Specifically, the encoder first maps the hybrid representation into the parameters of a latent Gaussian distribution via [μk⊕log(σk)2]=ΦEnc(hkhyb)[ _k ( _k)^2]= _Enc(h_k^hyb) (7) where ΦEnc(⋅) _Enc(·) is the encoder implemented with MLP layers, and μk _k and σk _k denote the mean and standard deviation of the latent distribution, respectively. Then, latent variables are sampled through the reparameterization trick, as follows, zk=μk+ϵ⊙σk,ϵ∼(,)z_k= _k+ε _k, ε (0,I) (8) Afterwards, the sampled latent representation is decoded to generate pseudo-visual embeddings via h¯kv=ΦDec(zk) h_k^v= _Dec(z_k) (9) where ΦDec(⋅) _Dec(·) is the decoder implemented with MLP layers, and h¯kv h_k^v denotes the generated pseudo-visual embeddings. h¯kv h_k^v is expected to preserve semantic consistency with the remaining modalities while compensating for the unreliable or missing visual information. The optimization objective of SR-VEG is detailed in Section 3.5. 3.4 Multi-modal Feature Fusion In RA-MMEA, the final entity representation is obtained by fusing embeddings from different modalities in a reliability-aware manner. For each entity, when its visual modality is predicted to be reliable, we directly incorporate the original visual embedding into the entity representation. Otherwise, we replace the unreliable visual embedding with the generated pseudo-visual embedding from SR-VEG. This process can be formulated as follows, hk=wkghkg⊕wkrhkr⊕wkahka⊕h~kvh_k=w_k^gh_k^g w_k^rh_k^r w_k^ah_k^a h_k^v (10) where hkh_k denotes the final representation of entity eke_k. h~kv=gk,1vwkvhkv+gk,0vh¯kv h_k^v=g_k,1^vw_k^vh_k^v+g_k,0^v h_k^v is the reliability-aware visual representation, where gk,1vg_k,1^v and gk,0vg_k,0^v are the two elements of the visual reliability vector gkvg_k^v, corresponding to the reliable and unreliable visual states, respectively. 3.5 Optimization and Inference Optimization. To enhance interaction between modalities, we introduce multi-modal contrastive learning to achieve bidirectional alignment. The loss function is defined as follows, ℒm=−log(pm(eu,et)+pm(et,eu)/2)L_m=- (p_m(e_u,e_t)+p_m(e_t,e_u)/2 ) (11) where pmp_m denotes the alignment probability between two entities eue_u and ete_t. Furthermore, the intra-modal loss ℒIALL_IAL and inter-modal loss ℒICLL_ICL can be represented as follows, ℒIAL=∑m∈MℒmℒICL=∑m∈Mℒ^mL_IAL= _m∈ ML_m\;\;\;\;\;L_ICL= _m∈ M L_m (12) where ℒ^m L_m is the post-fusion variant of ℒmL_m. To alleviate imbalance and enforce constraints on reconstructed features, we jointly optimize the SR-VEG loss function ℒSRL_SR together with the contrastive loss. Specifically, inspired by Chen et al. (2023b), we employ two reconstruction objectives, ℒRevisL^vis_Re and ℒRehybL^hyb_Re, to minimize the distance between the pseudo and true embeddings, thereby emphasizing the differences between different features during the autoencoding process. In addition, minimizing the Kullback-Leibler (KL) divergence ℒKLL_KL encourages the latent space to approximate a Gaussian distribution. Then, we introduce a similarity distillation loss ℒSimL_Sim to quantify the discrepancy between the similarity matrix of the hybrid embeddings and that of the virtual image features. Finally, the overall loss function can be expressed as follows, ℒ=ℒIAL+ℒICL+ℒSR+γ⋅ℒpL=L_IAL+L_ICL+L_SR+γ·L_p (13) where γ controls the strength of ℒpL_p, and ℒSR=ℒKL+ℒRevis+ℒRehyb+ℒSimL_SR=L_KL+L^vis_Re+L^hyb_Re+L_Sim. Inference. We employ the cosine similarity (Sim) to quantify the alignment probability. The similarity matrix between the source entity set E and the target entity set E′E^ is denoted as Sim⟨E,E′⟩Sim E,E^ . 4 Experiments Datasets. We validate the effectiveness and robustness of RA-MMEA on two cross-KG datasets, FB15K-DB15K/YG15K Liu et al. (2019), with 20%, 50%, and 80% reference entity alignments as seed pairs, and three bilingual DBP15K datasets (ZH/JA/FR-EN) (Sun et al., 2017) under the unsupervised setting. Appendix A depicts the statistics of multi-modal datasets. The details of the evaluation metrics are shown in Appendix B. Setting Models Venue FB15K-DB15K FB15K-YG15K Hits@1↑ Hits@10↑ MRR↑ Hits@1↑ Hits@10↑ MRR↑ Non-iterative MMEA (Chen et al., 2020) KSEM’20 0.265 0.541 0.357 0.234 0.480 0.317 EVA (Liu et al., 2021) AAAI’21 0.199 0.448 0.283 0.153 0.361 0.224 MSNEA (Chen et al., 2022) KDD’22 0.114 0.296 0.175 0.103 0.249 0.153 MCLEA (Lin et al., 2022) COLING’22 0.295 0.582 0.393 0.254 0.484 0.332 MoAlign (Li et al., 2023a) EMNLP’23 0.318 0.564 0.409 0.296 0.525 0.378 MEAformer (Chen et al., 2023a) ACM M’23 0.417 0.715 0.518 0.327 0.595 0.417 PMF (Huang et al., 2024) ACL’24 0.539 - 0.620 0.459 - 0.539 RICEA (Li et al., 2025a) ACL’25 0.471 0.720 0.557 0.411 0.658 0.497 SGMEA (Cheng et al., 2025) COLING’25 0.543 0.777 0.625 0.587 0.826 0.670 RULE (Li et al., 2026) ICLR’26 0.444 - 0.538 0.389 - 0.470 HUMEA (Xing et al., 2026a) AAAI’26 0.512 0.764 0.598 0.439 0.699 0.528 RA-MMEA(Ours) RA -MMEA\;(Ours) – 0.8490.849 0.9480.948 0.8790.879 0.7870.787 0.9280.928 0.8350.835 Iterative EVA (Liu et al., 2021) AAAI’21 0.231 0.488 0.318 0.188 0.403 0.260 MSNEA (Chen et al., 2022) KDD’22 0.149 0.392 0.232 0.138 0.346 0.210 MCLEA (Lin et al., 2022) COLING’22 0.395 0.656 0.487 0.322 0.546 0.400 MEAformer (Chen et al., 2023a) ACM M’23 0.578 0.812 0.661 0.444 0.692 0.529 PMF (Huang et al., 2024) ACL’24 0.624 - 0.702 0.543 - 0.620 IBMEA (Su et al., 2024) ACM M’24 0.631 0.813 0.697 0.521 0.708 0.584 PCMEA (Wang et al., 2024) AAAI’24 0.676 0.887 0.728 0.590 0.835 0.646 RICEA (Li et al., 2025a) ACL’25 0.567 0.804 0.652 0.516 0.733 0.593 SGMEA (Cheng et al., 2025) COLING’25 0.661 0.847 0.729 0.750 0.901 0.805 EIEA (Wang et al., 2025) KDD’25 0.769 0.869 0.804 0.830 0.915 0.860 DMEA (Xing et al., 2026b) W’26 0.804 0.896 0.849 0.898 0.957 0.934 RA-MMEA(Ours) RA -MMEA\;(Ours) – 0.9150.915 0.9760.976 0.9370.937 0.8990.899 0.9640.964 0.9360.936 Table 1: Results on two cross-KG datasets using 20% of the reference entity alignments for training. The best results are shown in bold and the second best results are underlined. Implementation Details. RA-MMEA is implemented in PyTorch on a machine with NVIDIA RTX A6000 GPUs. RA-MMEA is trained for 500 epochs, and optionally further trained by 500 additional epochs with iterative training. We apply AdamW optimizer (β1=0.9 _1=0.9, β2=0.999 _2=0.999) with batch size of 3,500. The scaling factor d is set to 17. The parameter γ that controls ℒpL_p is set to 15. Compared Methods. To rigorously evaluate our framework’s effectiveness, we compare it with 16 methods that focus on multi-modal fusion optimization as in Sec. 4.1. Notably, several recent MMEA methods, such as TMEA (Chen et al., 2024), LGEA (Lu et al., 2025), and CLAEA (Ji et al., 2025), introduce external knowledge for MMEA. Since our goal is to design a reliability-aware framework that detects and mitigates unreliable visual modalities by exploiting inherent multi-modal dependency, rather than enhancing inputs with additional external information, we exclude these information-enhancement methods to ensure a fair comparison. For RULE (Li et al., 2026), we adopt the variant without MLLM during testing. 4.1 State-of-the-art Comparison Tab. 1 shows the non-iterative and iterative results on the cross-KG datasets with 20% seed alignments. RA-MMEA consistently outperforms all baselines on both datasets and under both training settings across all metrics. Specifically, on FB15K-DB15K, RA-MMEA improves Hits@1 by 11.1%–30.6% and MRR by 8.8%–25.4% over the second-best baselines. On FB15K-YG15K, it achieves gains of 0.1%–20.0% in Hits@1 and 0.2%–16.5% in MRR. Notably, the improvements are particularly substantial under the non-iterative setting, demonstrating RA-MMEA’s effectiveness even with limited seed alignments. Additional results with 50% and 80% seed alignments are provided in Appendix C, while results on the bilingual DBP15K datasets are reported in Appendix D. Overall, these results demonstrate that RA-MMEA consistently achieves state-of-the-art performance across datasets, seed ratios, and training settings. 4.2 Robustness Evaluation Visual information in real-world scenarios often contains varying degrees of noise, making it important for MMEA models to maintain stable alignment performance as visual information becomes increasingly unreliable. To evaluate this robustness, we construct artificial Entity-Image mismatches on FB15K-DB15K by randomly selecting 20%, 50%, and 80% of entities and shuffling their associated images while preventing them from being reassigned to their original entities, thereby simulating progressively increasing visual noise. Figure 3: Performance of strong baselines across visual noise levels on FB15K-DB15K. As shown in Fig. 3, the performance of all methods gradually decreases as the noise ratio increases, indicating that incorrect Entity-Image associations make MMEA increasingly challenging. Nevertheless, RA-MMEA consistently achieves the best performance across all noise ratios and exhibits substantially smaller performance degradation than MEAformer and RICEA. These results demonstrate that DA-VRP effectively mitigates interference caused by incorrect Entity-Image associations, enabling RA-MMEA to preserve robust multi-modal representations under increasing visual noise and confirming its robustness in challenging noisy MMEA scenarios. 4.3 Visual Noise Detection The above results demonstrate that RA-MMEA remains robust under progressively increasing visual noise. We further investigate whether this advantage stems from the ability of DA-VRP to identify unreliable visual information. Specifically, we treat the artificially injected Entity-Image mismatches as known unreliable samples and measure the proportion of these samples correctly identified as unreliable by DA-VRP. Noise Ratio Ratio=20% Ratio=50% Ratio=80% Detection Rate (%) 75.24 74.87 75.08 Table 2: Detection rate of DA-VRP under varying visual noise ratios. As shown in Tab. 2, DA-VRP achieves detection rates of 75.24%, 74.87%, and 75.08% under visual noise ratios of 20%, 50%, and 80%, respectively, remaining stable at approximately 75% across different noise levels. These results indicate that DA-VRP can effectively identify unreliable visual information based on cross-modal semantic inconsistency, while its detection capability remains largely insensitive to changes in the proportion of injected noise. Since existing MMEA datasets do not provide ground-truth reliability annotations for the original Entity-Image pairs, we evaluate only the artificially injected visual noise with known labels. Nevertheless, the consistent detection performance provides empirical evidence that cross-modal dependency serves as an effective task-level proxy for visual reliability. Further qualitative evidence is provided in Appendix I, where the predicted reliability is consistent with the semantic correspondence between entities and their associated images. DA-VRP SR-VEG Hits@1↑ Hits@10↑ MRR↑ ❶ ✗ ✗ 0.682 0.860 0.733 ❷ ✓ ✗ 0.798 0.886 0.855 ❸ ✗ ✓ 0.701 0.873 0.752 ❹ ✓ ✓ 0.849 0.948 0.879 Table 3: Ablation studies on DA-VRP and SR-VEG. Hits@1↑ Hits@10↑ MRR↑ ❶ DA-VRP w/ BS 0.503 0.762 0.590 ❷ DA-VRP w/ GS 0.849 0.948 0.879 Table 4: Ablation studies on Gumbel-Softmax (GS) and Bernoulli Sampling (BS) in DA-VRP. Hits@1↑ Hits@10↑ MRR↑ ❶ SR-VEG w/o ℒpL_p 0.833 0.941 0.874 ❷ SR-VEG w/ ℒpL_p 0.849 0.948 0.879 Table 5: Ablation studies on ℒpL_p for SR-VEG. Hits@1↑ Hits@10↑ MRR↑ ❶ generation w/ w^km w_k^m 0.830 0.938 0.867 ❷ generation w/ wkmw_k^m 0.849 0.948 0.879 Table 6: Ablation studies on using w^km w_k^m and using wkmw_k^m for hybrid representation in SR-VEG. Hits@1↑ Hits@10↑ MRR↑ ❶ γ=1γ=1 0.834 0.942 0.874 ❷ γ=5γ=5 0.836 0.943 0.875 ❸ γ=10γ=10 0.845 0.946 0.877 ❹ γ=15γ=15 0.849 0.948 0.879 ❺ γ=20γ=20 0.840 0.945 0.879 Table 7: Ablation studies on the parameter γ for ℒpL_p. 4.4 Ablation Study To further analyze RA-MMEA, we conduct ablation studies on FB15K-DB15K with 20% seeds. Our final configuration is highlighted in cyan. Impacts of DA-VRP and SR-VEG. The core of RA-MMEA lies in two key modules, including DA-VRP and SR-VEG. To analyze their impacts, we conduct ablation studies in Tab. 3. As shown, without DA-VRP and SR-VEG, the Hits@1 score is 0.682 (❶). When using DA-VRP alone, the Hits@1 score is largely improved to 0.798 with 11.6% gains (❷); when applying SR-VEG only, the Hits@1 score is boosted to 0.701 with moderate improvements (❸). Using both modules together yields the best performance of 0.849 Hits@1 score with 16.7% gains (❹), confirming their strong complementarity for improving MMEA. Gumbel-Softmax v.s. Bernoulli Sampling. In our DA-VRP, we adopt Gumbel-Softmax to predict the visual reliability, enabling end-to-end training of RA-MMEA. To analyze its effectiveness, we conduct ablation studies by comparing Gumbel-Softmax and Bernoulli sampling for visual reliability prediction. The results are reported in Tab. 4. As shown, we can observe that, when using Bernoulli sampling for reliability prediction, the Hits@1 score is 0.503 (❶). In contrast, when applying Gumbel-Softmax, the Hits@1 score is improved to 0.849 with 34.6% gains (❷). This confirms that differentiable reliability estimation is essential for learning accurate modality-filtering behavior. Impact of ℒpL_p. In SR-VEG, we design a stability regularization term ℒpL_p, which aims to constrain dependency shifts after removing the visual modality. To analyze the impact of ℒpL_p, we conduct the ablation studies in Tab. 5. As shown, without ℒpL_p, the model suffers from degraded dependency estimation, leading to suboptimal pseudo-visual embeddings and performance. Incorporating ℒpL_p consistently improves the Hits@1 score from 0.833 (❶) to 0.849 (❷) with 1.6% gains, demonstrating that stabilizing cross-modal dependency is critical for reliable virtual visual embedding generation. Analysis of wkmw_k^m for Hybrid Representation. In SR-VEG, we apply the original dependency scores wkmw_k^m (m∈g,r,am∈\g,r,a\) to construct the hybrid representation in Eq. (5). To validate its effectiveness, we compare hybrid representation construction with wkmw_k^m and w^km w_k^m in Tab. 6. As shown, using the original dependency scores wkmw_k^m is more stable (❷ v.s. ❶), as they are computed with full multi-modal information, whereas re-estimating weights after removing the visual modality relies on less information and therefore becomes unstable. Impact of γ. To analyze the impact of γ for ℒpL_p, we conduct ablation studies in Tab. 7 by changing the value of γ. As shown in Tab. 7, we can see that, when γ=15γ=15, we show the best performance (❹). Therefore, we empirically set γ to 15. 4.5 Plug-and-Play Capability We integrate DA-VRP and SR-VEG into two MMEA frameworks with similar processing pipelines, MEAformer and RICEA, to evaluate the generalizability and plug-and-play capability of the proposed modules. As shown in Tab. 8, incorporating DA-VRP and SR-VEG consistently improves both frameworks across all evaluation metrics. Specifically, MEAformer achieves improvements of 8.9% and 7.6% in Hits@1 and MRR, respectively, while RICEA obtains corresponding gains of 3.0% and 3.2%. These consistent improvements demonstrate that DA-VRP and SR-VEG are not restricted to a specific MMEA architecture but can be flexibly integrated into existing frameworks to enhance their ability to effectively handle diverse unreliable visual information. Models Hits@1 ↑ Hits@10 ↑ MRR ↑ MEAformer (Chen et al., 2023a) 0.417 0.715 0.518 MEAformer (+) 0.506 0.758 0.594 Improvement (%) 8.9 4.3 7.6 RICEA Li et al. (2025a) 0.471 0.720 0.557 RICEA (+) 0.501 0.755 0.589 Improvement (%) 3.0 3.5 3.2 Table 8: Plug-and-play evaluation of DA-VRP and SR-VEG across representative MMEA frameworks. "+" denotes their variants. 5 Conclusion In this paper, we propose RA-MMEA, a reliability-aware framework for robust multi-modal entity alignment. RA-MMEA estimates visual reliability through cross-modal dependency and generates alternative representations for unreliable images. Extensive experiments demonstrate its effectiveness and robustness, highlighting the importance of reliability-aware visual modeling in MMEA. Limitations Although RA-MMEA demonstrates strong performance across multiple benchmarks, existing MMEA datasets typically provide only limited and relatively static visual information for each entity, which may not fully capture the diversity of entity-related images in real-world scenarios. In practice, the same entity may be associated with images exhibiting different visual content across different contexts, making it challenging to evaluate MMEA models under more diverse visual conditions. Exploring more diverse visual settings could further broaden the applicability of reliability-aware MMEA. It would also provide additional insights into visual reliability across different entity types and scenarios. Such investigations could further enrich our understanding of reliability-aware visual modeling in MMEA. These directions may also offer further insights into how visual information contributes to multi-modal entity alignment. Future work could develop MMEA benchmarks with richer and more diverse visual information, enabling a more comprehensive evaluation of the robustness and generalizability of MMEA methods in real-world scenarios. References Chen et al. (2020) L. Chen, Z. Li, Y. Wang, T. Xu, Z. Wang, and E. Chen MMEA: entity alignment for multi-modal knowledge graph. In Knowledge Science, Engineering and Management: 13th International Conference, KSEM 2020, Hangzhou, China, August 28–30, 2020, Proceedings, Part I 13, p. 134–147. Cited by: Table 10, Table 11, §1, Table 1. Chen et al. (2022) L. Chen, Z. Li, T. Xu, H. Wu, Z. Wang, N. J. Yuan, and E. Chen Multi-modal siamese network for entity alignment. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, p. 118–126. Cited by: Table 10, Table 10, Table 11, Table 11, Table 12, Table 12, §2, Table 1, Table 1. Chen et al. (2024) L. Chen, Y. Sun, S. Zhang, Y. Ye, W. Wu, and H. Xiong Tackling uncertain correspondences for multi-modal entity alignment. Advances in Neural Information Processing Systems 37, p. 119386–119410. Cited by: §2, §4. Chen et al. (2023a) Z. Chen, J. Chen, W. Zhang, L. Guo, Y. Fang, Y. Huang, Y. Zhang, Y. Geng, J. Z. Pan, W. Song, et al. Meaformer: multi-modal entity alignment transformer for meta modality hybrid. In Proceedings of the 31st ACM International Conference on Multimedia, p. 3317–3327. Cited by: Table 10, Table 10, Table 11, Table 11, Table 12, Table 12, Appendix E, Table 15, §2, §3.1, §3.1, Table 1, Table 1, Table 8. Chen et al. (2023b) Z. Chen, L. Guo, Y. Fang, Y. Zhang, J. Chen, J. Z. Pan, Y. Li, H. Chen, and W. Zhang Rethinking uncertainly missing and ambiguous visual modality in multi-modal entity alignment. In International Semantic Web Conference, p. 121–139. Cited by: §2, §3.5. Cheng et al. (2025) J. Cheng, M. Guo, and F. Zhang SGMEA: structure-guided multimodal entity alignment. In Proceedings of the 31st international conference on computational linguistics, p. 7851–7861. Cited by: Table 10, Table 10, Table 11, Table 11, Table 1, Table 1. Devlin et al. (2019) J. Devlin, M. Chang, K. Lee, and K. Toutanova Bert: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), p. 4171–4186. Cited by: §3.1. He et al. (2016) K. He, X. Zhang, S. Ren, and J. Sun Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 770–778. Cited by: Appendix D. Huang et al. (2024) Y. Huang, X. Zhang, R. Zhang, J. Chen, and J. Kim Progressively modality freezing for multi-modal entity alignment. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 3477–3489. Cited by: Table 1, Table 1. Jang et al. (2017) E. Jang, S. Gu, and B. Poole Categorical reparametrization with gumble-softmax. In International Conference on Learning Representations (ICLR 2017), Cited by: §3.2. Ji et al. (2025) Y. Ji, J. Cheng, F. Zhang, and C. Lu Capturing latent modal association for multimodal entity alignment. In Findings of the Association for Computational Linguistics: EMNLP 2025, p. 22278–22293. Cited by: §2, §4. Kingma and Welling (2014) D. P. Kingma and M. Welling Auto-encoding variational bayes. In International Conference on Learning Representations, Cited by: §3.3. Li et al. (2025a) C. Li, J. Cheng, Q. Tong, F. Zhang, and C. Wang Probing relative interaction and dynamic calibration in multi-modal entity alignment. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 28504–28516. Cited by: Table 10, Table 10, Table 11, Table 11, Appendix E, Table 15, §2, §3.1, Table 1, Table 1, Table 8. Li et al. (2025b) C. Li, J. Cheng, Q. Tong, and F. Zhang Exploring the impacts of feature fusion strategy in multi-modal entity alignment. In Proceedings of the 31st International Conference on Computational Linguistics, p. 7809–7818. Cited by: §2. Li et al. (2026) H. Li, Y. Lin, P. Hu, M. Yang, and X. Peng Learning with dual-level noisy correspondence for multi-modal entity alignment. In The Fourteenth International Conference on Learning Representations, Cited by: §2, Table 1, §4. Li et al. (2023a) Q. Li, C. Ji, S. Guo, Z. Liang, L. Wang, and J. Li Multi-modal knowledge graph transformer framework for multi-modal entity alignment. In Findings of the Association for Computational Linguistics: EMNLP 2023, p. 987–999. Cited by: Table 10, Table 11, Table 1. Li et al. (2024) Q. Li, J. Li, J. Wu, X. Peng, C. Ji, H. Peng, L. Wang, and S. Y. Philip Triplet-aware graph neural networks for factorized multi-modal knowledge graph entity alignment. Neural Networks, p. 106479. Cited by: §2. Li et al. (2023b) Y. Li, J. Chen, Y. Li, Y. Xiang, X. Chen, and H. Zheng Vision, deduction and alignment: an empirical study on multi-modal knowledge graph alignment. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 1–5. Cited by: §2. Lin et al. (2022) Z. Lin, Z. Zhang, M. Wang, Y. Shi, X. Wu, and Y. Zheng Multi-modal contrastive representation learning for entity alignment. In Proceedings of the 29th International Conference on Computational Linguistics, p. 2572–2584. Cited by: Table 10, Table 10, Table 11, Table 11, Table 12, Table 12, Appendix D, §2, Table 1, Table 1. Liu et al. (2021) F. Liu, M. Chen, D. Roth, and N. Collier Visual pivoting for (unsupervised) entity alignment. In Proceedings of the AAAI conference on artificial intelligence, Vol. 35, p. 4257–4266. Cited by: Table 10, Table 10, Table 11, Table 11, Table 12, Table 12, Appendix D, Table 1, Table 1. Liu et al. (2019) Y. Liu, H. Li, A. Garcia-Duran, M. Niepert, D. Onoro-Rubio, and D. S. Rosenblum MMKG: multi-modal knowledge graphs. In The Semantic Web: 16th International Conference, ESWC 2019, Portorož, Slovenia, June 2–6, 2019, Proceedings 16, p. 459–474. Cited by: §1, §4. Lu et al. (2025) C. Lu, C. Li, J. Cheng, Y. Ji, G. Chen, and F. Zhang Breaking the noise barrier: llm-guided semantic filtering and enhancement for multi-modal entity alignment. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, p. 33141–33155. Cited by: §2, §4. Maddison et al. (2017) C. J. Maddison, A. Mnih, and Y. W. Teh The concrete distribution: a continuous relaxation of discrete random variables. In International Conference on Learning Representations, Cited by: §3.2. Radford et al. (2021) A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, p. 8748–8763. Cited by: §3.1. Shi et al. (2022) Y. Shi, M. Wang, Z. Zhang, Z. Lin, and Y. Zheng Probing the impacts of visual context in multimodal entity alignment. In Asia-Pacific Web (APWeb) and Web-Age Information Management (WAIM) Joint International Conference on Web and Big Data, p. 255–270. Cited by: §1, §2. Sohn et al. (2015) K. Sohn, H. Lee, and X. Yan Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28. Cited by: §3.3. Su et al. (2024) T. Su, J. Sheng, S. Wang, X. Zhang, H. Xu, and T. Liu Ibmea: exploring variational information bottleneck for multi-modal entity alignment. In Proceedings of the 32nd ACM International Conference on Multimedia, p. 4436–4445. Cited by: Table 10, Table 11, Table 1. Sun et al. (2017) Z. Sun, W. Hu, and C. Li Cross-lingual entity alignment via joint attribute-preserving embedding. In The Semantic Web–ISWC 2017: 16th International Semantic Web Conference, Vienna, Austria, October 21–25, 2017, Proceedings, Part I 16, p. 628–644. Cited by: Appendix D, §4. Vaswani et al. (2017) A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin Attention is all you need. Advances in neural information processing systems 30. Cited by: §3.2. Veličković et al. (2018) P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Liò, and Y. Bengio Graph attention networks. In International Conference on Learning Representations, Cited by: §3.1. Wang et al. (2024) L. Wang, P. Qi, X. Bao, C. Zhou, and B. Qin Pseudo-label calibration semi-supervised multi-modal entity alignment. In Proceedings of the AAAI conference on artificial intelligence, Vol. 38, p. 9116–9124. Cited by: Table 10, Table 11, Table 1. Wang et al. (2025) L. Wang, C. Zhou, and B. Qin Explicit-implicit entity alignment method in multi-modal knowledge graphs. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, p. 2996–3007. Cited by: Table 1. Xing et al. (2026a) Y. Xing, Q. Xie, Y. Liu, Q. Gu, T. Zheng, B. Chong, and T. He On modality weighting and specificity for multi-modal entity alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, p. 27144–27152. Cited by: Table 10, Table 11, Table 1. Xing et al. (2026b) Y. Xing, Q. Xie, Y. Lv, Z. Zhou, Q. Hou, Q. Gu, B. Chong, and T. He Enhancing multi-modal entity alignment via multi-grained decision fusion. In Proceedings of the ACM Web Conference 2026, p. 4047–4056. Cited by: Table 1. Yang et al. (2019) H. Yang, Y. Zou, P. Shi, W. Lu, J. Lin, and X. Sun Aligning cross-lingual entities with multi-aspect information. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), p. 4431–4441. Cited by: §3.1. Zhang et al. (2024) Q. Zhang, Y. Wei, Z. Han, H. Fu, X. Peng, C. Deng, Q. Hu, C. Xu, J. Wen, D. Hu, et al. Multimodal fusion on low-quality data: a comprehensive survey. arXiv preprint arXiv:2404.18947. Cited by: §2. Appendix A Appendix: Datasets Statistics Tab. 9 shows the statistics of datasets, including the number of entities (Ent.), relations (Rel.), attributes (Attr.), number of relation triples (Rel Tr.) and attribute triples (Attr Tr.), number of images (Image), and number of the reference entity pairs (EA pairs). Each entity is associated to multiple attributes and 0 ∼ 1 image. Dataset KG Ent. Rel. Attr. Rel. Tr. Attr. Tr. Image EA pairs FB15K-DB15K FB15K 14,951 1,345 116 592,213 29,395 14,836 12,846 DB15K 12,842 279 225 89,197 48,080 12,812 FB15K-YG15K FB15K 14,951 1,345 116 592,213 29,395 14,836 11,199 YG15K 15,404 32 7 122,886 23,532 11,172 DBP15KZH-EN ZH 19,388 1,701 8,111 70,414 248,035 15,912 15,000 EN 19,572 1,323 7,173 95,142 343,218 14,125 DBP15KJA-EN JA 19,814 1,299 5,882 77,214 248,991 12,739 15,000 EN 19,780 1,153 6,066 93,484 320,616 13,741 DBP15KFR-EN FR 19,661 903 4,547 105,998 273,825 14,174 15,000 EN 19,993 1,208 6,422 115,722 351,094 13,858 Table 9: Statistics of MMEA datasets, with EA pairs representing the reference entity alignments. Appendix B Appendix: Evaluation Metrics We adopt Hits@n and MRR as evaluation metrics for all methods. Hits@n measures the proportion of correct entities ranked within the top n according to similarity scores, while MRR represents the mean reciprocal rank of the correct entities. Higher values of Hits@n and MRR indicate superior performance of the evaluated method. Appendix C Appendix: More Results on Cross-KG Datasets Tabs. 10 and 11 report the results on the cross-KG datasets using 50% and 80% of the reference entity alignments as seed pairs, respectively. RA-MMEA consistently achieves the best performance across all settings. Specifically, with 50% seed alignments, RA-MMEA improves Hits@1 by 9.2%–20.0% and MRR by 7.2%–16.2% over the second-best baselines. With 80% seed alignments, it further achieves gains of 6.0%–13.6% in Hits@1 and 4.4%–11.1% in MRR. As the proportion of seed alignments increases, the performance of most methods improves due to the stronger alignment supervision. Nevertheless, RA-MMEA maintains a clear advantage over existing methods under both non-iterative and iterative settings. Notably, the improvements remain substantial even with 80% seed alignments, where competing methods already achieve strong alignment performance. This indicates that the benefit of reliability-aware visual modeling is not limited to low-supervision scenarios but remains effective when abundant alignment supervision is available. Together with the results using 20% seed alignments in Tab. 1, these results demonstrate that RA-MMEA consistently benefits from reliable visual information across different levels of alignment supervision, further confirming the effectiveness and generalizability of the proposed framework on cross-KG datasets. Setting Models FB15K-DB15K FB15K-YG15K Hits@1↑ Hits@10↑ MRR↑ Hits@1↑ Hits@10↑ MRR↑ Non-iterative MMEA (Chen et al., 2020) 0.417 0.703 0.512 0.403 0.645 0.486 EVA (Liu et al., 2021) 0.334 0.589 0.422 0.311 0.534 0.388 MSNEA (Chen et al., 2022) 0.288 0.590 0.388 0.320 0.589 0.413 MCLEA (Lin et al., 2022) 0.555 0.784 0.637 0.501 0.705 0.574 MoAlign (Li et al., 2023a) 0.576 0.749 0.634 0.550 0.713 0.617 MEAformer (Chen et al., 2023a) 0.619 0.843 0.698 0.560 0.778 0.639 RICEA (Li et al., 2025a) 0.648 0.852 0.721 0.617 0.811 0.687 SGMEA (Cheng et al., 2025) 0.716 0.882 0.775 0.780 0.924 0.832 HUMEA (Xing et al., 2026a) 0.695 0.880 0.763 0.649 0.847 0.720 RA-MMEA(Ours) RA -MMEA\;(Ours) 0.9160.916 0.9780.978 0.9370.937 0.8880.888 0.9690.969 0.9170.917 Iterative EVA (Liu et al., 2021) 0.364 0.606 0.449 0.325 0.560 0.404 MSNEA (Chen et al., 2022) 0.358 0.656 0.459 0.376 0.646 0.472 MCLEA (Lin et al., 2022) 0.620 0.832 0.696 0.563 0.751 0.631 MEAformer (Chen et al., 2023a) 0.690 0.871 0.755 0.612 0.808 0.682 IBMEA (Su et al., 2024) 0.742 0.880 0.793 0.655 0.821 0.714 PCMEA (Wang et al., 2024) 0.738 0.915 0.781 0.670 0.886 0.721 RICEA (Li et al., 2025a) 0.692 0.869 0.757 0.658 0.827 0.720 SGMEA (Cheng et al., 2025) 0.752 0.894 0.802 0.827 0.938 0.868 RA-MMEA(Ours) RA -MMEA\;(Ours) 0.9320.932 0.9840.984 0.9500.950 0.9190.919 0.9770.977 0.9400.940 Table 10: Results on two cross-KG datasets using 50% of the reference entity alignments for training. The best results are shown in bold and the second best results are underlined. Setting Models FB15K-DB15K FB15K-YG15K Hits@1↑ Hits@10↑ MRR↑ Hits@1↑ Hits@10↑ MRR↑ Non-iterative MMEA (Chen et al., 2020) 0.590 0.869 0.685 0.598 0.839 0.682 EVA (Liu et al., 2021) 0.484 0.696 0.563 0.491 0.692 0.565 MSNEA (Chen et al., 2022) 0.518 0.779 0.613 0.531 0.778 0.620 MCLEA (Lin et al., 2022) 0.735 0.890 0.790 0.667 0.824 0.722 MoAlign (Li et al., 2023a) 0.699 0.882 0.773 0.689 0.884 0.769 MEAformer (Chen et al., 2023a) 0.765 0.916 0.820 0.703 0.873 0.766 RICEA (Li et al., 2025a) 0.776 0.916 0.829 0.734 0.892 0.792 SGMEA (Cheng et al., 2025) 0.815 0.931 0.828 0.857 0.951 0.894 HUMEA (Xing et al., 2026a) 0.809 0.935 0.856 0.774 0.920 0.827 RA-MMEA(Ours) RA -MMEA\;(Ours) 0.9510.951 0.9900.990 0.9670.967 0.9390.939 0.9870.987 0.9570.957 Iterative EVA (Liu et al., 2021) 0.491 0.711 0.573 0.493 0.695 0.572 MSNEA (Chen et al., 2022) 0.565 0.810 0.651 0.593 0.806 0.668 MCLEA (Lin et al., 2022) 0.741 0.900 0.802 0.681 0.837 0.737 MEAformer (Chen et al., 2023a) 0.784 0.921 0.834 0.724 0.880 0.783 IBMEA (Su et al., 2024) 0.821 0.922 0.859 0.751 0.890 0.800 PCMEA (Wang et al., 2024) 0.820 0.964 0.858 0.756 0.942 0.802 RICEA (Li et al., 2025a) 0.787 0.919 0.838 0.752 0.899 0.804 SGMEA (Cheng et al., 2025) 0.828 0.921 0.861 0.882 0.967 0.915 RA-MMEA(Ours) RA -MMEA\;(Ours) 0.9580.958 0.9910.991 0.9700.970 0.9420.942 0.9890.989 0.9590.959 Table 11: Results on two cross-KG datasets using 80% of the reference entity alignments for training. The best results are shown in bold and the second best results are underlined. Appendix D Appendix: Unsupervised Training on Bilingual Datasets To validate the performance of our framework on the bilingual domain, we also conducted unsupervised experiments on the bilingual datasets DBP15K (Sun et al., 2017). Unsupervised training was initially introduced by Liu et al. (2021) to mitigate dependence on gold labels. Following Lin et al. (2022), the vision encoders EncvEnc_v are configured as ResNet-152 (He et al., 2016) with a vision feature dimension of dv=2048d_v=2048. We utilize pre-trained 300-d GloVe vectors along with character bigrams for surface representation after applying machine translations for entity names. Models DBP15KZH−ENDBP15K_ ZH-EN DBP15KJA−ENDBP15K_ JA-EN DBP15KFR−ENDBP15K_ FR-EN Hits@1↑ Hits@10↑ MRR↑ Hits@1↑ Hits@10↑ MRR↑ Hits@1↑ Hits@10↑ MRR↑ Non-iterative EVA (Liu et al., 2021) 0.891 0.961 0.917 0.941 0.986 0.958 0.970 0.996 0.982 MSNEA (Chen et al., 2022) 0.859 0.936 0.887 0.921 0.970 0.939 0.954 0.989 0.968 MCLEA (Lin et al., 2022) 0.860 0.950 0.893 0.914 0.975 0.938 0.953 0.990 0.967 MEAformer (Chen et al., 2023a) 0.909 0.974 0.933 0.950 0.990 0.965 0.972 0.997 0.983 RA-MMEA(Ours) RA -MMEA\;(Ours) 0.9420.942 0.9920.992 0.9610.961 0.9710.971 0.9980.998 0.9820.982 0.9910.991 0.9990.999 0.9950.995 Gain over Prior Best (%) 3.3 1.8 2.8 2.1 0.8 1.7 1.9 0.2 1.2 Iterative EVA (Liu et al., 2021) 0.948 0.992 0.964 0.977 0.997 0.985 0.990 1.000 0.995 MSNEA (Chen et al., 2022) 0.871 0.943 0.898 0.927 0.976 0.945 0.959 0.993 0.979 MCLEA (Lin et al., 2022) 0.942 0.994 0.963 0.974 0.999 0.984 0.986 1.000 0.992 MEAformer (Chen et al., 2023a) 0.964 0.9980.998 0.975 0.985 1.000 0.991 0.992 1.000 0.996 RA-MMEA(Ours) RA -MMEA\;(Ours) 0.9690.969 0.9990.999 0.9780.978 0.9890.989 1.0001.000 0.9930.993 0.9980.998 1.0001.000 0.9990.999 Gain over Prior Best (%) 0.5 0.1 0.3 0.4 0.0 0.2 0.6 0.0 0.3 Table 12: Unsupervised results on bilingual datasets are presented. Tab. 12 shows that our framework consistently outperforms all strong baselines on both Hits@1 and MRR, achieving improvements of 0.4%–3.3% and 0.2%–2.8%, respectively, across two settings on three bilingual datasets. Notably, the Hits@10 score reaches 1.0 on the DBP15K− DBP15K_ JA-EN and DBP15K− DBP15K_ FR-EN. Appendix E Appendix: Low-Resource Comparison To investigate the reliance of our framework on alignment labels during training, we evaluate RA-MMEA on FB15K-DB15K with fewer seed alignments. Following Chen et al. (2023a) and Li et al. (2025a), we set the seed alignment ratios to RsaR_sa = 0.01, 0.03, 0.07, 0.11, 0.14, 0.18. For a more intuitive comparison of performance trends, we adopt the same backbone as their methods. Figure 4: Performance of strong baselines with fewer seed alignments on FB15K-DB15K. Fig. 4 presents the comparison with other methods. RA-MMEA consistently outperforms all baselines across all seed alignment ratios, with increasingly pronounced advantages as more seed alignments become available. Notably, even when Rsa=0.01R_sa=0.01, RA-MMEA achieves the best Hits@1 and MRR among all compared methods. When Rsa=0.18R_sa=0.18, RA-MMEA even surpasses RICEA with Rsa=0.20R_sa=0.20. These results not only demonstrate the robustness of RA-MMEA to unreliable visual data but also highlight its potential in low-resource scenarios. Appendix F Appendix: Applying DA-VRP to Textual Modalities To further evaluate DA-VRP’s ability to filter unreliable information beyond the visual modality, we additionally examine its performance on the relational and attribute modalities on FB15K-DB15K with 20% seeds. In this section, we focus solely on identifying and removing unreliable textual modalities, without involving modality reconstruction. Please note that, we do not filter the structural modality, as it provides the essential backbone for modeling cross-modal dependency. Removing structural information would compromise the stability of the alignment framework. Models Hits@1↑ Hits@10↑ MRR↑ RA-MMEA-R 0.708 0.824 0.745 RA-MMEA-A 0.732 0.845 0.811 RA-MMEA-V 0.798 0.886 0.855 Table 13: Performance of applying our reliability prediction method to textual modalities. Tab. 13 reports the results. RA-MMEA-V, RA-MMEA-R, and RA-MMEA-A respectively represent the variants applying reliability prediction on the visual image as proposed in this work, the relation modality, and the attribute modality. Please note that, for all these three variants, we do not perform modality reconstruction. As shown in Tab. 13, DA-VRP continues to provide clear benefits when applied to textual modalities, outperforming all other baselines in Tab. 1. However, textual modalities often contain complementary relational cues, and aggressive filtering may inadvertently discard useful information. In future work, we will develop finer-grained criteria to better separate complementary from redundant cross-modal relationships, enabling more selective modality filtering. Appendix G Appendix: Visual Replacement Strategy Although unreliable visual representations may still contain useful entity-related information, retaining them may also introduce additional noise. We therefore compare two alternative strategies that preserve the original unreliable visual representations: Concatenation and Weighted-Concatenation. Specifically, Concatenation combines the unreliable visual representation with the generated virtual visual representation, while Weighted-Concatenation reduces its contribution according to the estimated reliability. Hits@1↑ Hits@10↑ MRR↑ ❶ Concatenation 0.831 0.941 0.873 ❷ Weighted-Concatenation 0.835 0.944 0.876 ❸ Ours 0.849 0.948 0.879 Table 14: Comparison of strategies for handling unreliable visual representations. As shown in Tab. 14, Weighted-Concatenation slightly outperforms Concatenation (❷ v.s. ❶), indicating that suppressing unreliable visual information can alleviate its negative impact. Nevertheless, our strategy (❸) achieves the best performance across all metrics, improving Hits@1 by 1.4 and 1.8 percentage points over Weighted-Concatenation and Concatenation, respectively. These results suggest that retaining unreliable visual representations may still introduce misleading visual semantics into multi-modal fusion and consequently affect entity representations during the alignment process, especially when visual information is highly unreliable. In contrast, directly replacing them with virtual visual representations generated from the remaining reliable modalities provides more semantically consistent and informative visual representations for accurate entity alignment, validating the visual replacement strategy in RA-MMEA. Appendix H Appendix: Model Complexity Analysis To further evaluate our framework, we compare the complexity of RA-MMEA with two representative lightweight baselines on FB15K-DB15K. As shown in Tab. 15, the training time, learnable parameters, and GPU memory of our RA-MMEA are 24 minutes, 13M, and 16G, which are comparable to the complexities of MEAformer and RICEA. However, compared to these two methods, RA-MMEA achieves superior alignment performance on different metrics (see Tab. 1, Appendix C and Appendix D), showing an effective balance between accuracy and model complexity. Figure 5: Case study of visual reliability on DBP15K, showing representative images predicted as unreliable (left) and reliable (right) by DA-VRP based on their cross-modal dependency with the corresponding entity semantics. The entity associated with each image is shown above the image. Models Train. Time Learnable Paras. GPU Mem. MEAformer (Chen et al., 2023a)* 22min 10M 13.6G RICEA Li et al. (2025a)* 22min 10M 13.2G RA-MMEA (Ours) 24min 13M 16.0G Table 15: Comparison of complexity on FB15K-DB15K. * indicates our reproduced results. Appendix I Appendix: Case Study To qualitatively examine whether DA-VRP can distinguish unreliable from reliable visual information, we present cases from DBP15K in Fig. 5. The left side shows images predicted as unreliable by DA-VRP. For European Parliament, the image is an abstract seating diagram and provides limited visual semantics for representing the entity. The image associated with Compact disc is nearly blank, while Championnat d’Europe de football is represented by a general map of Europe rather than the football competition itself. Similar cases include 2014 AFC Champions League, 2011 Pokémon (anime), and Languages of Spain, whose images provide ambiguous, incomplete, or weakly related visual cues. These cases show that unreliable images may be irrelevant or capture only limited or indirect entity semantics. In contrast, the reliable cases on the right exhibit entity–image semantic consistency. Édouard Daladier is depicted by a portrait, Beaver by an image of the corresponding animal, and University of Texas at Austin by its recognizable institutional identity. Likewise, the images of Python (programming language), Pope Julius I, and Kyrie Irving contain distinctive visual information that directly corresponds to their entity semantics. Compared with the unreliable cases, these images provide more explicit and discriminative visual evidence for identifying the entities. The comparison shows that unreliable images may provide uninformative or weakly related semantics, potentially introducing misleading evidence, whereas reliable images exhibit stronger entity-image semantic consistency. Notably, these examples cover diverse entity types, suggesting that the predicted reliability is not restricted to entity or visual content types. Despite their different visual characteristics, reliable images exhibit clearer semantic correspondence, while unreliable images provide more ambiguous or incomplete visual evidence. The qualitative results show that DA-VRP captures differences in visual reliability through cross-modal dependency. Together with the quantitative results in Sec. 4.3, these cases further support cross-modal dependency as an effective task-level proxy for visual reliability.