Cross-Language Entity Alignment Based on Dual-Relation Graph and Neighbor Entity Screening

Abstract

Graph convolutional network-based methods have become mainstream for cross-language entity alignment. The graph convolutional network has multi-order characteristics that not only process data more conveniently but also reduce the interference of noise effectively. Although the existing methods have achieved good results for the task of cross-language entity alignment, they have often overlooked the same entity names in the real corpus, resulting in an entity-matching result that was not ideal. Therefore, this study proposed a neighboring-entity-screening rule by combining the entity name and the attribute (NENA) to reduce the influence of these issues. We used the NENA-screening rule to filter and delete redundant equivalent entities and to construct a dual-relation graph as auxiliary evidence for scenarios when the attribute information may be insufficient.This study adopted a graph convolutional network in order to embed knowledge graphs and entity names into a unified vector space, and then a down-sampling method was used to extract the neighboring entities of each entity, thus forming sub-graphs of the two knowledge graphs. We embedded the sub-graphs into the GCN, as the new input, and then we used a cross-graph-matching module to finally achieve alignment. Our results on the DBP15K dataset showed that our approach significantly improved the overall entity alignment.On the sub-dataset ZH-EN of DBP15K, the value of Hits@1 improved by 1.38%, as compared to the best approach mentioned in this paper, and it was useful for the construction and completion of the open knowledge graph.

1. Introduction

A knowledge graph is a graph-based data structure. Specifically, a knowledge graph is a modern theory that combines theories and methods of applied mathematics, graphics, information visualization technology, and other disciplines with metrological citation analysis and other methods, and it uses a visualization graph to display the complete knowledge framework of the disciplines to achieve the purpose of multidisciplinary integration. Knowledge graphs encode unstructured data into structured triples and provide an effective scheme to describe the complex relations between concepts and entities. With the continuous deepening and improvement of the knowledge-graph research, knowledge graphs have been widely applied to various tasks related to artificial intelligence, such as intelligent question-answering, intelligent recommendations, and semantic searches. In recent years, knowledge graphs have been constructed continuously, and the number and scale of the knowledge graphs have increased rapidly. Extensive data sources, uneven quality, and diverse languages are the main factors contributing to the diversity and heterogeneity of the knowledge-graph information. Therefore, to solve these issues, cross-language tasks have been invented. Cross-linguistic research are studies developed from contrastive linguistics. Cross-linguistic tasks are based on the large-scale corpora of two or more languages, and the research objects extend from traditional phonemes, vocabulary, and syntax, to pragmatics, rhetorical structures, and even language using in special scenarios. In recent years, in order to support multi-lingual applications, integrating cross-language knowledge graphs has become a necessary step in the knowledge graph research, and entity alignment is critical for this step.

The goal of the entity alignment task [1,2,3,4] has been to find entities that represent the same thing in the real world in multiple knowledge graphs. Formally, it is necessary to establish the relation between these entities and then connect the multi-source and heterogeneous knowledge graphs in order to form a unified knowledge graph, which provides support for various applications based on knowledge graphs. Cross-language entity alignment task refers to the task of finding matching entity-pairs in different languages in a multi-language scenario. For example, the knowledge graphs of different languages are shown in Figure 1. In shown in Figure 1, ZH-KG1 is a Chinese knowledge graph, and EN-KG2 is an English knowledge graph. ZH-KG1 and EN-KG2 not only have different language forms, but they also have different structures. The entities, relations, and attributes in the two knowledge graphs were represented differently by each language.Even if two entities belong to different knowledge graphs, it does not mean that the two entities refer to different things. The “Husak” entity in the English knowledge graph is different from the “胡萨克” entity in the Chinese knowledge graph. However, since they refer to and represent the same concept in the real world, they were regarded as align-able entity-pairs.

In order to achieve the integration of cross-language knowledge graphs more accurately and efficiently on the basis of the classic model TransE [5,6], researchers have proposed approaches such as TransD [7], TransR [8], MtransE [1] and other representative cross-language entity alignment methods. These approaches have made positive contributions to the research of cross-language entity alignment. However, they have relied on machine translations to eliminate language barriers, and therefore, these methods often result in uneven translation quality and lack context to employ as a reference.

At present, most approaches have used a graph convolutional network (GCN), which can greatly reduce the influence of noise caused by language differences. GCN is a graph domain-information-processing method based on deep learning. Because of its good performance and interpretability, it has become a widely used graph analysis method. It has efficiently combined entity-adjacent node information and learning structures to encode knowledge graphs.It is worth noting that although GCN has been widely used for the task of cross-language entity alignment at present, the existing methods have not considered that entities in the real corpora may have the same name but refer to different things. Insufficient supporting materials has easily led to a significant amount of incorrectly matched results.

After studying and summarizing previous research, this study then applied the multi-order characteristics of graph convolutional networks and then used the entity attribute information as evidence to screen high-quality triples. In order to reduce the influence caused by insufficient attribute information, we constructed a dual-relation graph [9,10]. Meanwhile, the central entity was defined as the head entity that occurred at least once in all triples.Then, the cross-language entity alignment was completed by neighboring-sampling and similarity calculations. Our main contributions were the following:

• We proposed a neighboring-entity-screening rule that used the entity name and its attributes as the main evidence for screening. In the neighboring entity set of the central entity, we identified all entities with the same entity name as the central entity. Then, the attribute information of the entities was used as the main evidence to judge whether two entities with the same entity name referred to the same concept in the reality corpus.
• Dual-relation graph was used to make full use of relations. This paper used the dual-relation graph not only to strengthen the role of entity relations but also to avoid the impact of insufficient attribute information and to reduce the errors in the entity screening process.
• A cross-language entity alignment method based on neighboring-entity screening and a dual-relation graph was proposed. We used a graph convolutional network, combined with neighboring-entity-screening rules and a dual-relation graph to realize cross-language entity alignment and achieve excellent results on public datasets, such as DBP15K, which proved the importance of entity screening.

2. Related Work

The emergence of a large number of heterogeneous, multifarious, and complementary knowledge graphs have emphasized the importance of integrating multi-lingual knowledge graphs to achieve comprehensive and complete knowledge graphs. Entity alignment is a key technology to achieve this goal. The existing entity alignment methods can be divided into two categories, traditional entity alignment methods and representation learning-based entity alignment methods.

2.1. Traditional Entity Alignment Methods

Traditional entity alignment methods have mostly applied used feature-engineering techniques, such as string similarity and rule mining, to complete alignment tasks. For example, Ngomo et al. [11] proposed using triangle inequality to calculate the entity similarity, and NIU [12] optimized artificially defined entity-matching rules by using the maximization expectation algorithm.

Traditional entity alignment methods have manually constructed a large amount of feature information and designed attribute-matching rules. Datasets in different domain and attribute categories have required different attribute-matching rules. Therefore, designing different attribute-matching rules for the attribute categories contained in datasets in different domains have been crucial for realizing traditional entity alignment technology. However, these approaches not only require a significant amount of manpower but are also difficult to migrate between multiple datasets.

2.2. Entity Alignment Based on Representation Learning

With the rapid development of entity alignment, traditional approaches have been unable to meet the requirements of high-efficiency and high-accuracy matching. After significant exploration, some researchers found that approaches based on representation learning were more suitable for the task of entity alignment, and the results have been significantly better than the traditional approaches. Entity alignment based on representation learning has had good results for modeling the semantic information of the knowledge graphs. Many entity alignment algorithms have been implemented based on graph representation learning, which encodes all entities in a knowledge graph in order to extract the graph features of the entities. Entity alignment based on representation learning has been divided into two categories, namely translation model-based methods [2,13,14] and graph convolutional network (GCN)-based methods [2,3,4,15]. It is worth noting that most of the following methods have used Chinese–English, French–English, and Japanese–English datasets.

2.2.1. Entity Alignment Based on Translation Model

Translation-based approaches have originated from cross-language word-embedding. Therefore, they also have a core assumption that the entity-embedding of different knowledge graphs have similar distributions, similar to the word-embedding of different languages [16,17]. Translation-based methods usually include two modules: a translation module (TM) and an alignment module (AM) [16], as shown in Figure 2. TM usually refers to translation models, such as TransE [5]. AM is usually divided into two different alignment modules: mapping and sharing. The mapping approach embeds different knowledge graphs into a unified vector space through a linear transformation matrix. For example, KDCoE [18] minimized the distances between the pre-aligned pairs by optimizing one or two linear transformation matrices. The sharing approach embeds different KGs into a unified vector space by allowing each pre-aligned pair to directly share the same embedding [17]. For example, MTransE [5] proposed to minimize the equation $\parallel e_3-e_5\parallel$ for each pre-aligned entity-pair.

Representation learning approaches based on translation models are able to learn vector representations of entities in low-dimensional dense spaces. Inspired by Word2Vec [19], TransE [5,6] interprets relations as head-to-tail translations. The representation learning approach, represented by TransE, was originally used for the structural embedding of entities in a single knowledge graph. TransE’s extension models TransD [7], TransH [20], etc., have performed entity alignment and entity inference in a unified vector space on the basis of learning entity-embedding representations. The closer the distance between the entity-pair in the embedding space, the higher the semantic similarity of the two entities, and the greater the possibility that the entity-pair could be aligned between knowledge graphs.

Inspired by the above approaches, many researchers [1,9,21,22,23] have explored improvements. MtransE [1] employed the TransE model to encode the structure of the knowledge graph into a specific language vector space. However, the model relied on a large number of previously aligned entities, which were often difficult to obtain.In order to reduce the dependence on a priori aligned entities, IPTransE [22] and BootsTransE [23] adopted a semi-supervised iterative strategy, which predicted new alignment entities with a small number of previously aligned entities, used them to augment the previously aligned entity set, and iterated repeatedly until no new alignment entities appeared. However, the iterative process often resulted in error propagation.

2.2.2. Entity Alignment Based on Graph Convolutional Networks

The entity alignment methods based on translation model used the sum of the head entity vector and the relation vector to predict the tail entity vector, but in actual situations, the links between entities and relations may be more complicated. For example, in the process of using this kind of approach for alignment tasks, a significant amount of noise may be generated, which may then affect the accuracy of alignment. The discovery of this issue introduced new challenges. More effective alignment has become the focus to overcome this challenge.In recent years, many researchers have introduced graph convolutional networks (GCNs). The concept of these approaches was derived from the utilization of the global information modeling of graphs. GCN is a multi-layer neural network that operates directly on a graph, which induces node-embedding vectors based on the attributes of its neighborhoods [24].

Approaches based on GCN, such as [9,10,25,26,27,28,29,30,31,32,33,34,35,36,37,38], have enhanced the embedding of entities and their neighbors’ information, and they only require a small number of aligned seed-pairs to transfer similar information to the whole graph. As compared to the traditional entity-alignment methods, approaches based on GCNs not only require relatively less human involvement in the process of feature construction, but also such approaches could be extended to large knowledge graphs. GCN applications have typically embedded the data to be processed into a unified vector space [39,40].

JAPE [9] introduced two embedding modules, namely structure embedding (SE) and attribute embedding (AE), which jointly embedded the structures of two knowledge graphs into a unified vector space, and then they used the attribute correlation in the knowledge graph for further improvements. As compared to the traditional alignment approach, this has provided significant improvements, but it has not been able to effectively address the one-versus-many issue of structural embedding.In response to this issue, Wang [25] proposed the GCN-align model. It embedded the entities of each language into a unified vector space and then learned the embeddings from the structure and attribute information of the entities. This combination resulted in precise alignment results.Inspired by Wang [25], Fan [30] proposed the assignment of different weight values to different attributes on the basis of the GCN-align. This method effectively utilized attribute information and reduced the computational complexity in the matching process, resulting in a more precise alignment.

The introduction and application of the GCN has promoted the rapid development of entity alignment task. In addition to the aforementioned approaches related to the GCN, the RDGCN [10] was a novel relation-aware dual-graph convolutional network, and it integrated relational information by exploiting the close interaction between the knowledge graph and its dual-relations and then further captured neighboring structures to better understand the entity representations. However, the RDGCN ignored the structural differences between different knowledge graphs. The multi-channel graph neural network model (MuGNN) [32] adopted the judgment and the comparison of the knowledge graphs to coordinate the structural differences between different knowledge graphs and used graph-based models to better utilize the seed-pair information. Although MUGNN reduced the impact of the structural differences between the different knowledge graphs, it was unable to fully utilize the attribute information of the entities. Tam [34] proposed an end-to-end and unsupervised cross-language knowledge-graph entity alignment framework, which captured the relation-based correlations between entities by exploiting the multi-order properties of the GCN and then fused different types of information to take advantage of the richness of the knowledge graph. After Tam [34], Wu et al. [41] proposed a new entity alignment framework (neighborhood matching network, NMN) to address the challenge of structural heterogeneity. The NMN estimated the similarities between entities to capture both the topological structures and the neighborhood differences. It provided two innovative components that enabled a better learning representation for entity alignments. Such strategies allowed the NMN to efficiently construct matching-oriented entity representations while ignoring noisy neighbors that would have a negative impact on the alignment results [41].

These approaches have made useful contributions towards advancing entity alignment, and at the same time, they have opened up new directions for the application of GCNs in entity alignment tasks. However, these works have overlooked the existence of triples of the same entity name in the real corpus. The entities in these triples contain a significant amount of relation and attribute information, and this has lead to inexact matches between similar entities. Therefore, based on the above research and the NMN [41] model, we used the multi-order properties of the GCN and combined the dual-relations with neighboring-entity-screening to achieve cross-language entity alignment.

3. Methods

3.1. Problem Definition

Entity alignment is the process of integrating entities that mean the same thing in the real world, according to different knowledge bases, into a unified knowledge base. Although many entity alignment methods for cross-language knowledge graphs have emerged, there is still a significant amount of advancement required based on the matching scores of similar entities, whether in specific or open domains.Therefore, how to effectively improve the accuracy of similar entities is a hotspot in current alignment-task research. The following questions addressed in this paper concerned cross-language entity alignment: (1) how to exploit the advantages of the GCN in cross-language entity alignment; (2) how to use the entities, relations, and attributes more effectively; and (3) how to handle entities with the same entity name. The problem definition diagram is shown in Figure 3. Without the loss of generality, we implemented the task of entity alignment between the two knowledge graphs KG1 and KG2, based on a set of pre-aligned equivalent entities. Our goal was to identify the equivalent entity-pair between KG1 and KG2.

3.2. Symbol Definition

Knowledge graphs represent knowledge about real-world entities as triples. According to the experimental requirements, two kinds of triples in the knowledge graph were formally defined, which were relation triples and attribute triples. A relation triple represented the relation between two entities and consisted of a head entity, a relation, and a tail entity (which could also be understood as a subject, a predicate, and an object). Its basic definition format was $(head,relation,tail)$. For example, the relation triple <$MountFuji,is,mountain$>, from which we could determine the relation between Mount Fuji and mountain. The attribute triple described the attribute characteristics of the entity, and its basic definition format was $(entity,attribute,value)$. For convenience, we denoted the knowledge graph as $KG=(E,R,T)$, where E, R, and T represent the entity set, the relation set, and the triple set, respectively. For the convenience of representation, in form, we denoted the two heterogeneous knowledge graphs as $K{G}_1=(E_1,R_1,T_1)$ and $K{G}_2=(E_2,R_2,T_2)$, where $E_i$, $R_i$, and $T_i$ represent the entity set, the relation set, and the triple set in $K{G}_i(i=1,2)$, respectively. Taking Figure 3 as an example, we obtained two knowledge graphs $K{G}_1=(E_1,R_1,T_1)$ and $K{G}_2=(E_2,R_2,T_2)$. As shown in Figure 3, we found that $E_1$ and $E_2$ were composed of five entities, namely $E_1=\{1,2,3,4,5\}$ and $E_2=\{11,12,13,14,15\}$. Variables $R_1$ and $R_2$ were composed of relations connecting entities in $E_1$ and $E_2$, respectively, and $T_1$ and $T_2$ were composed of triples formed by entities in KG1 and KG2 that have some relations, respectively.

Definition 1.

We combined the two knowledge graphs to form a large knowledge graph and referred to this large knowledge graph as the primal graph (PG).

Definition 2.

The dual-relation graph (DRG) was an un-directed graph, which assumed that relations sharing the same head or tail entity were a vertex and the shared edge’s weight value as the correlation degree of the two relations.The dual-relation graph was then used to strengthen the role of relation so that it could be fully utilized and serve as auxiliary evidence for entity screening.

Definition 3.

Two entities were said to be equivalent entities if they had the same entity name and an inclusion relation between their attributes.

3.3. Entity Alignment Based on Entity Screening and Dual-Relation Graph

In a practical situation, the heterogeneity and the noise of common neighboring entities in the knowledge graphs greatly reduce the potential of entity alignments and make it difficult to obtain useful information. We used the two knowledge graphs in Figure 1 for this example. In EN-KG2, entity “Husak” and entity “Havel” had the same neighboring entity “president”, and we assumed that there was no relation “successor” between “Husak” and “Havel”. In entity alignment, due to the heterogeneity of neighboring entities, it was difficult for us to determine whether neighboring entities could be aligned due to insufficient evidence, and this process would inevitably produce noise, precluding the extraction of useful information to prove that entity “哈维尔” and entity “Havel” could be aligned.In fact, we knew that “电影” and “film” could also be aligned, but as opposed to “胡萨克” and “Husak”, the former had a significantly lower number of neighboring entities and could provide relatively little useful information, making it difficult or even impossible to align the former. The entity’s nearest neighboring entity played an important role in determining whether the entity should be aligned with other entities. However, not all nearest neighboring entities positively contributed to entity alignment. Therefore, we needed to select informative entities as alignment samples. The whole research procedure is shown in Figure 4.

Using the entities that were part of the two knowledge graphs, in Figure 1, for this example, the steps of our method were as follows:

Step 1: Dual-relation graph (DRG) construction. The dual-relation graph used the entity relation as vertex, and an edge was established between the two relations if they shared the same head or tail entity. In order to ensure the obtained dual-relation graph fully included the source and target knowledge graphs, a graph attention network (GAT) was used to ensure the dual-relation graph interacted with the PG for several rounds. The dotted lines in Figure 4 represent multiple rounds of interaction between the PG and the DRG.

Step 2: Neighboring-entity screening. A neighboring-entity-screening rule based on entity name and attribute was proposed. The attribute information of the entity was used to judge whether the neighboring entity was equivalent to the corresponding central entity. If two entities had the same entity name and there was an inclusion relation for their attributes, we considered the two entities to be equivalent. When the attribute information was insufficient, we used the dual diagram as auxiliary evidence for neighboring-entity screening.

Step 3: Knowledge embedding. We embedded the primal graph structure with the names of the entities obtained in step 2 into a unified vector space, and then the resulting entity feature representations in the primal graph were fed to the GCN layers with highway gates in order to capture the neighboring structural information. In order to strengthen the role of the relation between entities and their closer connections, we embedded the primal graph into the GCN.

Step 4: Entity neighbor-sampling. In order to facilitate the matching of the corresponding entities of the two knowledge graphs, we calculated the probability of each neighboring entity being sampled and then use a down-sampling method to obtain the neighboring entity with higher probability ranking to finally obtain the two knowledge sub-graphs.

Step 5: Entity alignment. We embedded the sub-graph obtained in step 4 into the GCN, used the cross-graph matching module [41] for the neighboring-entity matching, and then aggregated the matched neighboring-entity information. Finally, we further calculated the distance between the two central entities to achieve alignment.

3.3.1. Dual Relation Graph Construction

Formally, this paper defined a primal graph (PG) $K{G}_p=(E_p,R_p,T_p)$, where $E_p,R_p,T_p$ represent the entity set, the relation set, and the triple set of the PG, respectively. For a given arbitrary $K{G}_p$, its DRG was defined as $K{G}_d=(E_d,W_d,T_d)$, where $E_d$ represents the set of entities with relations in the PG as nodes, $W_d$ represents the correlation degree of the two relations, and $T_d$ represents a correlation degree triple of the format $(E_{d1},W_d,E_{d2})$. If two different relations were connected by the same head entity (or tail entity), we built an edge between the two relations, and these edges formed the set $W_d$. For each type of relation r in $K{G}_p$, there was a vertex $e_d(e_d\in R_p)$ in $K{G}_d$ that corresponded to r. In order to ensure the DRG more accurately expressed the relation between different $e_d$ in $K{G}_d$, based on the possibility that the two relations in $K{G}_d$ could share a similar head or tail entity in $K{G}_p$, the Jaccard similarity coefficient was used to set the weight value $w_{ij}^r$ for each edge in $W_d$. Inspired by the literature [24], the calculation of weight $w_{ij}^r$ was:

$$He(r_i,r_j)=\frac{\mid H{e}_i\cap H{e}_j\mid }{\mid H{e}_i\cup H{e}_j\mid },Te(r_i,r_j)=\frac{\mid T{e}_i\cap T{e}_j\mid }{\mid T{e}_i\cup T{e}_j\mid }$$

(1)

$$w_{ij}^r=He(r_i,r_j)+Te(r_i,r_j)$$

(2)

where $H{e}_i$ and $T{e}_i$ represent the set of head and tail entities connected with relation $r_i$ ($r_i\in R_p$), respectively. The expressions $He(r_i,r_j)$ and $Te(r_i,r_j)$, respectively, represent the probability that relations $r_i$ and $r_j$ share the same head entity and tail entity.

In order to facilitate this understanding, this study provided an example of constructing a dual-relation graph, as shown in Figure 5. Since it was necessary to judge whether two relations shared the same head entity (or tail entity) when constructing a dual-relation graph, three possible scenarios are summarized in Figure 5. To illustrate the calculation of weight values, we used relations “career” ($r_1$) and “serve as” ($r_2$) in Figure 5 as an example. The head and tail entities connected by “career” were $\{Husak,Skada,Havel\}$ and $\{politician,director\}$, respectively, and those connected by “serve as” were $\{Husak,Havel\}$ and $\left\{president\right\}$, respectively. Relations “career” and “serve as” had the same head and tail entities; therefore, the weight value of the edge between the two relations was $w_{12}^r=0.6$, according to Equations (1) and (2).

It is worth noting that the dashed and solid lines in Figure 5 represent the same correlation degree. Since all edges between any two relations in the dual-relation graph had equal weight values, it was sufficient to connect an edge directly. In order to gain insight into the construction of the dual-relation graph, all edges are marked in Figure 5, and the case of more than one edge is indicated by a dashed line. The use of a dual-relation graph was not only to strengthen the role of relations between entities in entity alignment but also to serve as auxiliary evidence for entity screening in the case of insufficient entity attribute information.

3.3.2. Neighboring-Entity Screening

In the task of entity alignment, it was difficult to obtain accurate matching results only by the information contained alongside the corresponding entities in two knowledge graphs. Therefore, it has been typical to rely on the entity’s neighboring entities to provide more useful information to support the implementation of entity alignment. However, the existing methods [26,27,28,29] directly overlooked situations where the head and tail entities had the same entity name in the relation triple when obtaining the neighboring entities of each entity, but this situation could exist in the real corpus. For example, with relational triples $(Bob,know,Bob)$, it was obvious that the head and tail entities in this triple had the same name, but it could not be determined that the two entities referred to the same person or thing in the real world. What issues presented when we assumed that two entities were equivalent without comparing their overall information? As shown in Figure 6, we present an example.

In Figure 6, two knowledge graphs of different languages are shown. In the knowledge graph EN-KG, there were two entities whose entity names were both “$Tom$”. According to their attribute information, we found that the two entities were not equivalent. If we assumed that two entities were equivalent simply because their entity names were the same, then we would obtain a new knowledge graph for EN-KG’ after deleting one of the two entities. By the entity alignment of the knowledge graph EN-KG’ with the knowledge graph JA-KG, we were likely to obtain entity “雄猫” and entity “$Tom$” referring to the same thing. However, in the actual situation, the entity “$Tom$” in English and the entity “雄猫” in Japanese are not equivalent entities; therefore, the matching result obtained by the processing method in the figure was incorrect. In order to solve this issues, this study proposed a neighboring-entity-screening rule for joint entity names and attributes.

The neighboring-entity-screening rule for joints entity names and attributes was designed to remove neighboring entities that were equivalent to the central entity. Figure 7 shows the steps of the entity screening rule, for which the implementation steps were as follows (for ease of illustration, we set the IDs of the two “Tom” entities to “1” and “2”):

(1)

We first had to determine whether the name of the neighboring entity was the same as the central entity.

(2)

On the premise that the verification result in the first step was true (that is, the entity names of the two entities were the same), we compared the attribute information of the two entities.

As shown in Figure 7, if entity $e_1$ and entity $e_2$ had the same entity name and their attributes were the same (or there was an inclusion relation), then we considered the two entities to be equivalent and retained entity $e_1$ and deleted entity $e_2$. If entities $e_1$ and $e_2$ were only the same entity name, and there was no inclusion relation between their attributes, we assumed that the two entities were not equivalent. However, in this scenario, there could be a special case where an entity would not have any attribute information, and we could then use the dual-relation graph as auxiliary evidence for screening. In the dual-relation graph, (1) when the sum of the correlation degrees of all corresponding relations of entity $e_1$ was equal to the sum of the correlation degrees of all corresponding relations of entity $e_2$, we assumed that the two entities were equivalent. Conversely, (2) if the sum of the corresponding relations was not equal, then we considered the two entities as not equivalent.

3.3.3. Knowledge Embedding

(1). Knowledge Graph Structure Embedding and Entity Name Embedding. In order to learn knowledge graph structure embedding, we were inspired by the literature [41]. Therefore, the multi-order characteristics of a graph convolutional network were used to aggregate the higher degree neighboring structural information of entities. Using pre-trained entity name-embedding to initialize the input entity features of the GCN led to the model learning contextual information about adjacent structures.The GCN embedding structure of our method is shown in Figure 8. We embedded the knowledge graph structure and entity names into the GCN, so that each layer of the GCN contained entities, relations, and attributes of the knowledge graph, which also retained the original structures and entity names.

The knowledge graph structure embedding embedded the entities and relations in the knowledge graph into a continuous vector space, which was convenient for computing, and preserved the structural information in the knowledge graph. In order to facilitate data processing and retain the structural information of the knowledge graph, we used the primal graph as a large input graph, and each GCN layer used a set of entity features as input to update the entity feature representations, as shown in Equation (3).

$$x_i^{\left(l\right)}=ReLU(\sum _{k\in N_i\cup \left\{i\right\}}C{W}^{\left(l\right)}{x}_k^{(l-1)})$$

(3)

Among them, $\{x_1^{\left(l\right)},x_2^{\left(l\right)},x_3^{\left(l\right)},...,x_n^{\left(l\right)}|x_i^{\left(l\right)}\in {\mathbb{R}}^{d^{\left(l\right)}}(1\le i\le n)\}$ ($d^{\left(l\right)}$ represents the dimension of the l-th GCN layer) represents the output node of the l-th GCN layer, namely, the entity features; $N_i$ is the index set of the neighboring entities of the entity $e_i$; and $C=\frac{1}{{\phi }_i}$ is a constant while ${\phi }_i$ represents the normalization constant; $W^{\left(l\right)}\in {\mathbb{R}}^{d^{\left(l\right)}\times d^{(l+1)}}$ represents the weight matrix of the l-th layer; and i and k both represent the index value of the entity. We used the product of the constant C, the weight matrix, and the entity $e_k$ ($k\in N_i\cup \left\{i\right\}$) as the output node of the $l-1$-th layer, and then all the obtained products were summed to obtain the output entity features of the l-th GCN layer.

(2). Dual-relation graph embedding. We built a dual-relation graph (DRG) to not only enable entity screening but also to enhance the role of relations in entity alignment so that relations were fully utilized. To further verify whether the DRG could improve the performance of the entity alignment method, we embedded the DRG in the same vector space as the knowledge graph. In order to distinguish it from $x^{\left(l\right)}$, we denoted the output entity feature using the DRG as $x^{{}^{\prime }\left(l\right)}$, which was calculated as:

$$x_i^{{}^{\prime }\left(l\right)}=ReLU(C{\tilde{D}}^{-\frac{1}{2}}\tilde{A}{\tilde{D}}^{-\frac{1}{2}}{W}^{\left(l\right)}{x}^{{}^{\prime }(l-1)})$$

(4)

where $\tilde{A}=A+I$ is the adjacency matrix of the primal graph $K{G}_p$ with added self-connections, I is an identity matrix as well the degree matrix ${\tilde{D}}_{ii}={\sum }_j{\tilde{A}}_{ij}$ (both i and j represent the index of the entity). We treated $K{G}_p$ as an un-directed graph when constructing A in order to allow the information to flow in both directions. We iteratively compared the results obtained by Equations (3) and (4) and then selected the maximum value as the output vector, as shown in Equation (5).

$$x_i^{\left(l\right)}=x_i^{\left(l\right)}\mid ·\mid x_i^{{}^{\prime }\left(l\right)}$$

(5)

Among them, $x_i^{\left(l\right)}$ represents the output vector generated by the combination of structure embedding and entity-name embedding, $x^{{}^{\prime }\left(l\right)}$ represents the output vector iteratively generated by using the dual-relation graph, and $\mid ·\mid$ is a defined operation symbol that compares the values of two output vectors and outputs the maximum value.

3.3.4. Neighborhood Entity Sampling

The neighboring entities were key to determining whether an entity should be aligned with other entities. However, not all the neighboring entities contributed exactly to entity alignment. To extract useful neighboring entities, we used a down-sampling method (also known as an extraction method) to reduce the amount of calculation and prevent over-fitting in order to sample informative data from the neighboring entity of the central entity, as the neighboring entities of the central entity had already been processed by the entity-screening rules.

This study employed the pre-trained entity name-embedding to initialize the input node of the graph neural network, so that the entity-embedding learned by each layer of the GCN could contain rich structural and semantic information. The neighboring entities of the target entity were sampled through this information. In this way, neighboring entities that had been closer to the contextual semantic information of the target entity were more likely to be sampled. The contextual information of two equivalent entities in a real corpus is usually similar, so the stronger the correlation between the neighboring entity and the central entity, the more meaningful the information it contained, and the more clues it could provide for alignment. Therefore, we needed to sample neighboring entities with strong correlations to the central entity. In order to ensure the sampled neighboring entities were relevant, we defined a conditional probability to calculate the probability of the neighboring entity being sampled, and the adjacent entities with higher sampling probability ranking were selected as the candidate entity set.Formally, for a given arbitrary entity $e_i$, the probability expression for its neighboring entity $e_{ij}$ to be sampled are shown in Equation (6).

$$p(x_{ij}\mid x_i)=softmax(x_i{W}_s{x}_{ij}^T)=\frac{exp(x_i{W}_s{x}_{ij}^T)}{{\sum }_{k\in N_{e_i}}exp(x_i{W}_s{x}_{ik}^T)}$$

(6)

The softmax function is an activation function that uses an exponential function to convert a neural network output value into a probability value. The variables $x_i$ and $x_{ij}$ represented the learned embedding of entity $e_i$ and entity $e_{ij}$, respectively, and i and j were indexes of entity $e_i$ and $e_j$, respectively. The variable $W_s$ denoted the shared weight matrix, and $N_{e_i}$ was the index of the neighboring entity of entity $e_i$.

For each candidate entity in the knowledge graph, whether the neighboring entities of the entity and the neighboring entities that were closely related to the target entity could be found was critical to deciding whether to align the two entities. For an entity in the knowledge graph, if we wanted to find its equivalent entity in the target knowledge graph, we needed to compare its sampled neighboring entities with each candidate entity in the target knowledge graph in order to select the optimal matching entity.

In order to achieve the matching of neighboring entities, we used a low-cost approximate-matching method. Entities in $E_2$ that were closer to $e_i$ ($e_i\in E_1$) were more likely to align with $e_i$ in the embedding space. For an entity $e_j\in E_2$, its probability of being sampled as a candidate for $e_i$ is shown in Equation (7).

$$p(x_j\mid x_i)=\frac{exp({\parallel x_i-x_j\parallel }_{L_1})}{{\sum }_{k\in E_2}exp({\parallel x_i-x_k\parallel }_{L_1})}$$

(7)

Among them, ${\parallel x_i-x_j\parallel }_{L_1}$ used Manhattan distance and the $L_1$-norm to calculate the distance between the learning embedding layers of entities $e_i$ and $e_j$. The $L_1$-norm was chosen because it streamlined the weights to prevent over-fitting.By selectively sampling the nearest neighboring entities, our approach essentially constructed a neighbor discriminant sub-graph for each entity. Through Equations (6) and (7), the entities with high similarity could be preliminarily obtained, which could effectively improve the matching accuracy and reduce the influence of noise.

3.3.5. Entity Alignment

In Section 3.3.4, we obtained the sub-graphs formed by entities with the higher similarities in the two knowledge graphs. We used the sub-graphs of the two knowledge graphs as input to then identify the entity-pairs that could be aligned. We obtained the sparse vector of neighboring entity-embeddings in the sub-graphs and then utilized a cross-graph model [32] for neighboring entity-matching. The cross-graph model [32] utilized the structural information of the knowledge graph, including the cross-graph structure, and embedded entities into a unified vector space to achieve alignment between the knowledge graphs. Formally, we defined $(e_i,e_{i\_j})$ as the entity-pair to be measured, $e_i\in E_1$ and $e_{i\_j}\in E_2$, respectively. The variables $n_i$ and $n_{i\_j}$ represented the neighboring entities of $e_i$ and $e_{i\_j}$, respectively. The probability of neighboring entities could be obtained from Equations (6) and (7), and the neighboring entity with the largest probability was then extracted. On the basis of the literature [10,27,41], we obtained the calculation formula of the cross-graph matching vector of the neighbor $n_i$, as shown in Equations (8) and (9).

$$W_{n_i{n}_{i\_j}}=\frac{exp(x_{n_i}·x_{n_{i\_j}})}{{\sum }_{n_{i\_j}^{{}^{\prime }}\in N_{i\_j}^s}exp(x_{n_i}·x_{n_{i\_j}^{{}^{\prime }}})}$$

(8)

$$m_{n_i}=\sum _{n_{i\_j}\in N_{i\_j}^s}{W}_{n_i{n}_{i\_j}}(x_{n_i}-x_{n_{i\_j}})$$

(9)

where $W_{n_i{n}_{i\_j}}$ represents the attention weight, $m_{n_i}$ represents the matching vector of $n_i$, and $N_{i\_j}^s$ represents the set of sampled neighboring entities. The variables $x_{n_i}$ and $x_{n_{i\_j}}$ denote the GCN output embeddings of $n_i$ and $n_{i\_j}$, respectively. Then, the GCN output embeddings of neighboring entities $n_i$ were concatenated by a weighted matching vector $m_{n_i}$ as:

$${\stackrel{⌢}{x}}_{n_i}=[x_{n_i}\parallel \beta \ast m_{n_i}]$$

(10)

where $[a\parallel b]$ represents the OR operation of a and b. Since the matching vector captured the difference between the two nearest neighboring entities, for each target neighboring entity in the neighboring entity, the attention mechanism could accurately detect the neighboring entity that was most likely to match the target neighboring entity in the entity of another knowledge graph.

Since the neighborhood sampling was based on the GCN output embedding, it was first necessary to pre-train the GCN-based knowledge-graph-embedding model to generate high-quality entity representations. Then, by measuring the distance between the two entities, we could determine if they should be aligned, which was calculated as:

$$dist(e_i,e_j)={\parallel {\stackrel{⌢}{x}}_{e_i}-{\stackrel{⌢}{x}}_{e_j}\parallel }_{L_1}$$

(11)

Equation (11) uses Manhattan distance to calculate the distance between two entities to judge whether the two entities could be aligned.

During training, the distance between the aligned entity-pairs was expected to be as close as possible, and the distance between the negative entity-pairs was expected to be as far as possible; thus, an edge-based scoring function was used, which was calculated as shown in Equation (12).

$$L=\sum _{(i,j)\in S}\sum _{(i^{{}^{\prime }},j^{{}^{\prime }})\in S^{{}^{\prime }}}max\{0,d(i,j)-d(i^{{}^{\prime }},j^{{}^{\prime }})+\gamma \}$$

(12)

where $\gamma >$ 0 is an edge hyper parameter, S is the alignment seed, $S^{{}^{\prime }}$ is the set of negatively aligned entity-pairs produced by nearest-neighbor sampling, and $max\{0,d(i,j)-d(i^{{}^{\prime }},j^{{}^{\prime }})+\gamma \}$ indicates the value cannot be less than 0.

4. Experimental Setup

4.1. Datasets

This study used DBP15K [10] as the experimental dataset. DBP15K was a subset of DBpedia and has been commonly used as a benchmark dataset for entity alignment. DBpedia was a special example of a semantic web application that extracted structured data from Wikipedia entries in order to enhance Wikipedia’s search capabilities and to link other datasets to Wikipedia. DBP15K contained three sub-datasets, namely Chinese–English (ZH-EN), French–English (FR-EN) and Japanese–English (JA-EN). Each of these datasets was constructed by extracting 15,000 aligned entity links from the multi-lingual version of Dbpedia and contained the knowledge in two languages. The statistics of the three datasets are shown in

Table 1. DBP15K dataset.

DBP15K		Entity	Relation	Attribute	Rel.Triples	Att. Triples
ZH-EN	ZH	66,469	2830	8113	153,929	379,684
	EN	98,125	2317	7173	237,674	567,755
JA-EN	JA	65,744	2043	5882	164,373	354,619
	EN	95,680	2096	6066	233,319	497,230
FR-EN	FR	66,858	1379	4547	192,191	528,665
	EN	105,889	2209	6422	278,590	576,543

.

In this paper, we present only part of the data in DBP15K ZH-EN, and the data are presented in the format of the dataset, as shown in Figure 9. In Figure 9, ent_ids_1 and ent_ids_2 represent the correspondence between IDs and entities in the Chinese and English datasets, respectively. Furthermore, rel_ids_1 and rel_ids_2 represent the correspondence between id and relation in Chinese dataset and English dataset, respectively. In addition, triples_1 and triples_2 represent the triples in the Chinese and English datasets, respectively, which were composed of head entity ID, relation, and tail entity ID.

4.2. Comparison Model

Our approach was compared to several existing graph neural network-based approaches for cross-language entity alignment:

(1)

RDGCN [10]: A deep embedding-based technique that considered the attentive interaction between each KG and its dual-relation by passing them through a two-layer GCN with highway gates. The embeddings were then compared directly to obtain the alignment results [10,42].

(2)

JAPE [9]: A shallow embedding-based technique that generated structural (using a TransE model) and attribute embedding. The two types of embedding were then used simultaneously to compute the similarity score between the entities [9,42].

(3)

MUGNN [32]: A novel multi-channel graph neural-network model (MuGNN) that remembered alignment-oriented knowledge graph embeddings by robustly encoding two knowledge graphs through multiple channels.

(4)

KECG [34]: A semi-supervised entity-alignment method combining a knowledge-embedding model and cross-graph model in order to better utilize seed alignments in order to propagate the entire graphs under KG-based constraints.

(5)

GCN-Align [21]: A novel approach for cross-lingual knowledge-graph alignment based on graph convolutional networks that could learn embeddings from the structural and attribute information of entities and then combine the results to obtain accurate alignment.

(6)

NMN [41]: A novel entity-alignment framework, neighborhood-matching network, that captured the topology structure and neighborhood differences of entities by estimating the similarity between entities.

(7)

Dual-AMN [43]: A new KG encoder, dual-attention-matching network, that not only modeled both intra-graph and cross-graph information smartly, but also greatly reduced the computational complexity [43].

(8)

PSR [44]: A novel entity-alignment approach with three new components, which enabled high performance, high scalability, and high robustness.

It is worth noting that our method was the DRG+ESGCN, which was based on entity alignment with a dual-relation graph (DRG) and neighboring-entity-screening rule, combining entity name and attribute (NENA) data. In addition, we conducted an ablation experiment to evaluate the effectiveness of our proposed approach based on the dual-relation graph with neighboring-entity screening. In the ablation experiment, according to whether the experimental method adopted DRG or NENA, the variants of the DRG+ESGCN were obtained as the ESGCN and the DRGCN, respectively, where the ESGGN represented the entity alignment method using NENA-screening rules, and the DRGCN represented the entity alignment using DRG. In addition to this, we considered the baseline method NMN as a variant of the DRG+ESGCN to illustrate the effectiveness of our method. Since our method was an improvement based on the NMN method, it was more effective to illustrate the improvements generated by our method by using the NMN as a variant model.

4.3. Implementation Details

The configuration we used was as follows: $\beta$ = 0.9, and $\gamma$ = 1.0, and the neighbor-sampling stage sampled 5 entities for each entity. In order to set a reasonable number of sampling neighboring entities, this paper used samples 1, 5, 10, and 15 for comparison. The experimental results showed that it was optimal when the parameter was set to 5. The matching effect did not change significantly when the parameter was adjusted from 5 to 10, or even 15. However, considering that there were some entities whose number of neighbors would be less than 10, the number of sampled neighboring entities was set to 5. The dimensions of the embedding layer of the primal graph and the dual-relation graph were d = 300, $d^{\prime }$ = 600, and $\tilde{d}$ = 300. All dimensions of the hidden representations in GCN layers were 300. The learning rate was set to 0.005, and the number of negative samples of each positive one was set to k = 125 (as shown in Figure 10). In Figure 8, when k (k = 75, 100, 125, 150) were different values, the value of Hits@n (n = 1, 10) changed as k changes, and the peak value was reached when k = 125, which was the most appropriate value for k. This study, as well as the literature [9,10,21,32,34,41], used the same training–testing split: A total of 30% of the data were used for training and 70% of the data were used for testing.

4.4. Metrics

This study used Hits@k (k = 1, 10), a widely used metric in existing entity-alignment methods [9,10,21,32,34,41], to evaluate the performance of the entity alignment. A Hits@k score (higher was better) was computed by measuring the proportion of correctly aligned entities, as ranked at the top of the k list.Furthermore, we chose mean rank (MR) and mean reciprocal rank (MRR) as the evaluation metrics. Higher Hits@ k and MRR scores, as well as lower MR scores, indicated better performance.In comparison, Hits@1 represented the degree of matching with the most similar entity to the target entity, and it was equivalent to the precision of widely used conventional entity-alignment methods.

5. Results and Discussion

On the DBP15K dataset, we compared the performance of our approach with several of the models mentioned above, as shown in

Table 2. Performance on DBP15K dataset.

Model	ZH-EN		JA-EN		FR-EN
	Hits@1	Hits@10	Hits@1	Hits@10	Hits@1	Hits@10
JAPE	0.4078	0.7321	0.3723	0.6819	0.3221	0.6658
KECG	0.4770	0.8350	0.4847	0.8499	0.4929	0.8442
MUGNN	0.4773	0.8421	0.4866	0.8573	0.4890	0.8681
RDGCN	0.7105	0.8529	0.7790	0.9069	0.8883	0.9602
NMN	0.7330	0.8605	0.7861	0.9013	0.9031	0.9662
Dual−AMN	0.7403	0.9019	0.7598	0.9490	0.7293	0.9284
PSR	0.8024	0.9140	0.7310	0.9311	0.8033	0.9380
DRG+ESGCN	0.7570	0.9073	0.8070	0.9330	0.9701	0.9730

and

Table 3. Performance on DBP15K dataset (MR/MRR).

Model	ZH-EN		JA-EN		FR-EN
	MR	MRR	MR	MRR	MR	MRR
JAPE	64	0.490	99	0.476	92	0.430
KECG	71.802	0.598	59.706	0.611	41.925	0.609
RDGCN	68.829	0.763	45.728	0.825	17.664	0.915
Dual−AMN	28.630	0.805	11.797	0.830	20.056	0.801
PSR	11.456	0.810	10.931	0.844	7.532	0.852
ESGCN	1.537	0.789	1.511	0.834	1.333	0.927
DRG+ESGCN	1.782	0.811	1.615	0.853	1.359	0.927

. It is worth noting that since, during the experiment, we divided the dataset into training and testing sets, in order to ensure the authenticity and accuracy of the experimental results, we cross-validated the divided training and testing sets on our model. Then, we averaged the results in order to finally obtain the data shown in Table 2 and Table 3.

As shown in Table 2, the DRG+ESGCN achieved better Hits@1 values on DBP15K ZH-EN, which was 0.024 higher than the NMN and 0.042 higher than the RDGCN. In the above comparison method, the RDGCN used the dual-relation graph, but the neighboring entity was not utilized efficiently. Furthermore, although the NMN used neighborhood sampling to optimize the selection of the neighboring entities, it overlooked the negative effect of entities with the same entity name on entity alignment. When comparing the DRG+ESSGCN with the PSR, we found that the PSR results were significantly better than our method, on the dataset DBP15K ZH-EN. However, on the datasets DBP15K FR-EN and DBP15K JA-EN, our method had a significant improvement, as compared to PSR. This showed that although our method outperformed the NMN, its ability to address the language differences between Chinese and English was still insufficient, and we needed to further improve its performance.

As shown in Table 3, as compared to the PSR, the MR and MRR values improved.Based on the methods of [9,10,19,30,32,39], higher MRR scores, combined with lower MR scores, indicated better performance. On the sub-dataset ZH-EN, the MR’s value with the DRG+ESGCN decreased by 9.674, as compared to that of the PSR, while the MRR value increased by 0.001. Furthermore, as compared to the NMN, the MRR value of the DRG+ESGCN was increased by 0.032. The results fully demonstrated the positive contribution of filtering out triples with a head entity name that was the same as the tail entity name for entity alignment.

The Ablation Experiment.

Table 4. The comparison results of the DRG+ESGCN and its variants in the ablation experiments.

Model	ZH-EN		JA-EN		FR-EN
	Hits@10	MRR	Hits@10	MRR	Hits@10	MRR
NMN	0.8605	0.799	0.9031	0.827	0.9662	0.926
ESGCN	0.8786	0.789	0.9099	0.834	0.9689	0.927
DRGCN	0.8679	0.791	0.9054	0.830	0.9673	0.926
DRG+ESGNN	0.8325	0.685	0.8810	0.761	0.9277	0.839
DRG+ESGCN	0.9073	0.811	0.9330	0.853	0.9730	0.927

shows the results of the ablation experiment. It should be noted that in order to illustrate the advantages of the GCN, we exchanged the GCN for a GNN in our method and obtained a new variant, DRG+GNN. As expected, the performance of the different variants without the NENA-screening rule led to the DRG results decreasing. According to the results of the NMN, the ESGCN, and the DRGCN, we found that both the NENA-screening rules and the DRG have a significant improvement on the entity alignment, which also illustrated the effectiveness of our method. By comparing the results of the DRG+ESGNN and the DRG+ESGCN, we found that the results of the DRG+ESGCN were significantly better than those of the DRG+ESGNN. This also showed that the GCN had obvious advantages over the GNN, even under the same conditions.The GCN not only had multi-order characteristics, but it also effectively avoided the translation noise in the translation model. Our proposed NENA-screening rule used the entities and attributes to remove the negative impact of the equivalent entities, to a certain extent, and we used the DRG to strengthen the role of the relations, creating closer connections between entities, so that the entities, the relations, and the attributes could be fully utilized.

In order to further compare the performance of the DRG+ESGCN and the NMN, the specific gravity of the pre-aligned entity-pairs (denoted by SEED) was changed for comparison, as shown in Figure 11. As expected, the results of both models on the three sub-datasets gradually improved with the increasing proportion of previously aligned information. As shown in Figure 11a–c, the value of Hits@1 increased with the increase in SEED proportion. As compared to the sub-datasets JA-EN and FR-EN, the growth trend of the Hits@1 value of the ESGCN in sub-dataset ZH-EN was more obvious than that of the NMN, which was due to the significant language differences between Chinese and English. By the interaction between the dual-relation graph and the primal graph, the information of the neighboring entity was richer, and the connection with the central entity was closer, which reduced the influence of the language differences to a certain extent. As shown in Figure 11, our method reduced the noise caused by the differences between Chinese and English, to some extent, and significantly improved the Chinese–English entity alignment task.

In order to verify the validity of the results in Figure 11a–c, the MR and MRR values were obtained under different SEED values of specific gravity, as shown in Figure 11d–f. As shown in the figure, as the SEED value increased, the value of MR (MRR) continuously decreased (increased), and under the same conditions, the MR and MRR values of the ESGCN were better than those of the NMN. As shown in Figure 11, it was fully verified that our method made a significant contribution to the cross-language entity alignment task, and it also proved the effectiveness of the dual-relation graph and the entity screening for cross-language entity alignment.

6. Conclusions

This paper presented a neighboring-entity-screening rule based on the entity name and attributes for cross-language entity alignment. In addition, the dual-relation graph was used as auxiliary evidence when attribute information was insufficient. Our approach was designed to reduce the influence of noise caused by heterogeneity and linguistic differences between different knowledge graphs, and it effectively promoted the close associations between entities. Our method not only took full advantage of entities, relations, and attributes, but it also effectively eliminated the equivalent neighboring entities, yielding a significant improvement. We performed extensive experiments on real-world datasets and compared the DRG+ESGCN against seven recent GCN-based methods, and we conducted ablation experiments as well. The experimental results showed that the DRG+ESGCN achieved the best performance by consistently outperforming the competitive methods on the same datasets and evaluation metrics. The GCN not only has multi-order characteristics, but it also effectively avoids the interference of translation noise, giving it a significant advantage over GNNs and other methods in entity alignment tasks.

Although our method achieved superior results in cross-language entity alignment, there were still some limitations. In this study, we used a definition of equivalent entities that conformed to our proposed neighboring-entity-screening rules. Whether the attributes of the two entities had an inclusion relation was the basis for determining the equivalence of the two entities. However, in an actual situation, there could be an inclusion relation between the attributes of two entities, yet they still may not be equivalent; thus, this situation would requires deeper and more detailed verification.Although our method reduced the impact of noise caused by language differences to a certain extent, that was only a small improvement, and it could still not effectively mitigate the negative impact of this issue.

In our future work, we will introduce image information. Images contain rich information about entity features in a visual format.We will use multi-modal data that combines text and imagery to further reduce the negative impact of heterogeneity and language differences between knowledge graphs.