Graph Hierarchy and Language Model-Based Explainable Entity

Alignment of Knowledge Graphs

Kunyoung Kim

, Donggyu Kim

and Mye Sohn

c,*

Department of Industrial Engineering, Sungkyunkwan University, Suwon, Korea

Keywords: Entity Alignment, Explainable AI, Knowledge Graph, Language Model.

Abstract: This paper proposes a Graph hierarchy and Language model-based Explainable Entity Alignment (GLEE)

framework to perform Entity Alignment (EA) between two or more Knowledge Graphs (KGs) required for

solving complex problems. Unlike existing EA methods that generate embedding for entities using KG

structure information to calculate the similarity between entities, the GLEE framework additionally utilizes

graph hierarchy and datatype properties to find entities to be aligned. In the GLEE framework, the

semantically similar hyper-entities of the entities to be aligned are discovered to reflect graph hierarchy in the

alignment. Also, the semantically similar datatype properties and their values of the entities are also utilized

in EA. At this time, language model is utilized to calculate semantic similarity of the hyper-entities or datatype

properties. As a result, the GLEE framework can trustworthy explain why the two entities are aligned using

the subgraphs that consist of similar hyper-entities, semantically identical properties, and their data values.

To show the superiority of the GLEE framework, the experiment is performed using real world dataset to

prove the EA performance and explainability.

1 INTRODUCTION

As knowledge graphs (KGs) are proven to help

integrate and access disparate data sources using

machine learning and inference capabilities, various

KGs have been developed, ranging from KGs for

research to enterprise KGs (Zou, 2020). However,

due to the complexity of the problems to be solved

using KGs, it is generally difficult for a single KG to

meet the diverse knowledge requirements of an

application (Zeng et al., 2021). To leverage different

KGs simultaneously, a research, Entity Alignment

(EA), has been conducted to identify and connect

entities that belong to different KGs but refer to the

same real-world concept (Fanourakis et al., 2023).

Traditionally, EA has been performed using network

and iteration-based approaches, but recently,

Knowledge Graph Representation Learning (KGRL),

which generates embeddings of entities based on

graph structures and calculates the similarity between

entities using the embeddings, has become a

https://orcid.org/0000-0002-3570-8645

https://orcid.org/0009-0002-5753-5084

https://orcid.org/0000-0002-1951-3493

* Corresponding author

representative player (Zeng et al., 2021). However,

KGRL, which is performed based on the structural

similarity of KGs, may encounter the following

difficulties.

First, to solve complex problems in which

multiple elements are intertwined, EA between KGs

with various domains and scopes must be performed.

At this point, it is difficult to assume that all KGs have

a similar structure (Wang et al., 2022). Therefore, a

method that can perform EA regardless of the

structure or scope of KGs is needed.

Second, as a side-effect, it is not easy to explain

the results of EA on two KGs with different structures

or domains. Since different KGs often have different

naming schemes, it is hard to look up the properties

and neighbors of the aligned entities to check whether

the alignment is correct or not (Trisedya et al., 2023).

To help checking the EA results, explanation is

needed.

To do so, this paper proposes a novel Graph

hierarchy and Language model-based Explainable

124

Kim, K., Kim, D. and Sohn, M.

Graph Hierarchy and Language Model-Based Explainable Entity Alignment of Knowledge Graphs.

DOI: 10.5220/0013012300003886

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Explainable AI for Neural and Symbolic Methods (EXPLAINS 2024), pages 124-131

ISBN: 978-989-758-720-7

Entity alignment (GLEE) framework. Even though

graph hierarchies (Zhang et al., 2020) and datatypes

(Kim et al., 2022; Shen et al., 2021) contain crucial

information about entities, this information is not

directly reflected in KGRL EA methods. In the GLEE

framework, they are used to improve the EA

performance and provide explanation. The GLEE

framework consists of three steps.

In the first step, after selecting two KGs to be

aligned, candidate entities that can be aligned with a

specific entity in one KG (hereafter, target entity) are

found in the other KG. To do so, this paper calculates

graph structural similarity (hereafter, S-similarity)

using the existing embedding-based EA method. S-

similarity is not performed for EA directly, but for

finding candidate entities.

In the second step, EA is performed after finding

the entity to be aligned with the target entity among

candidate entities. To do so, this paper devises graph

hierarchical similarity (H-similarity) and datatype

similarity (D-similarity). H-similarity is calculated to

discover common hyper-concepts shared by the target

entity and the candidate entity. It has a larger value

when the same hyper-entities appear in fewer hops.

At this time, BERT (Devlin et al., 2018)-based word

embedding is used on the entity names to determine

whether the two hyper-entities are the same. If two

KGs are from different languages, multilingual BERT

model (Pires et al., 2019) is used. D-similarity is

calculated by discovering common properties of the

entities and counting the set of common properties

with the same datatype values. To determine whether

the two properties are same, BERT-based word

embedding is used on the property names, too. The

second step results are similarities between the target

entity and candidate entities.

In the third step, the EA is performed using the

three kinds of similarities and explanation is

performed using two subgraphs of the target entity

and its aligned entity extracted from each of the two

KGs. The subgraphs are extracted from the different

KGs but have same graph structure and contents. In

GLEE framework, the subgraphs consist of the paths

with the aligned entities as the starting node and the

hyper concept common to both subgraphs as the

ending node, and the datatype properties and their

data values of the aligned entities. The subgraphs are

used to explain the EA results.

The main contribution of this paper is providing a

detailed steps to utilize three kinds of similarity to

improve the EA performance and provide explanation.

Since S-similarity has relatively low computational

cost than H-similarity and D-similarity to find the

similar entities in the large KGs, S-similarity is

utilized in advance to discover the candidate pairs of

entities to be aligned. On the other hand, H-similarity

and D-similarity are more accurate and interpretable

than S-similarity, but it is impossible to directly apply

them to align entities in the large KGs. Therefore,

they are applied in the candidate pairs obtained by

using S-similarity.

This paper is organized as follow. The related

works of this paper is presented in Section 2. In

Section 3 the illustrative scenario is provided to show

how GLEE framework works. Section 4 depicts

GLEE framework with the details. The superiority of

the framework is presented in Section 5 with

experiments. Finally, Section 6 proposes the

conclusions and further research.

2 RELATED WORKS

Many studies have been performed to automatically

align entities of different knowledge graphs (KGs).

TransE (Bordes et al., 2013) treats the relation in a

relationship triple as the translation from head to the

tail entities achieving promising results in one-to-one

relation. However, it cannot consider the multi-hop

relationships nor process the complex relationships

such as one-to-many, many-to-one, many-to-many.

Therefore, many variants of TransE have

subsequently appeared, such as TransR (Lin et al.,

2015), TransH (Wang et al., 2014), TransD (Ji et al.,

2015), TransG (Xiao et al., 2015a), TransA (Xiao et

al., 2015b), and PTransE (Li et al., 2020). These

studies tackle different limitations of TransE and

enhance the ability to model structures within KGs.

Recently, due to the growth of the deep learning

methods, graph convolutional network (GCN) (Kipf

& Welling, 2016)-based methods are widely adopted

for entity alignment. GCN-Align (Wang et al., 2018)

captures the entities’ neighborhood structures for the

first time by utilizing GCNs and achieving promising

results. RN-GCN (Wu et al., 2019a) incorporates

relation information via attentive interactions

between the KG and its dual relation counterpart.

HGCN (Wu et al., 2019b) jointly learns entity and

relation representations without requiring pre-aligned

relations. EMGCN (Tam et al., 2020) perform

unsupervised EA by using both relation and attribute

information.

Several recent studies use semi-supervised

learning that generates new pre-aligned seeds during

the iteration process. MCEA (Qi et al., 2023) uses a

multiscale graph convolution model to embed the

graph and uses intermediate results to guide the

negative sampling process. PEEA (Tang et al., 2023)

Graph Hierarchy and Language Model-Based Explainable Entity Alignment of Knowledge Graphs

125

increases the connections between far-away entities

and labeled ones by incorporating positional

information into the representation learning.

Many studies have been performed to improve the

EA performance. These studies assume that the KGs

have similar structure and perform experiment using

similar KGs (i.e., DBpedia English and DBpedia

French). That is, the performance is not guaranteed

when the KGs have different structure and scope.

Also, they do not provide explanation of EA results

which undermines the reliability.

Figure 1: An illustrative example of the GLEE framework.

3 AN ILLUSTRATIVE EXAMPLE

An illustrative example is used to show how the

GLEE framework conducts EA and how to explain its

results. As depicted in Figure 1, two entities

representing the same real-world concept, the

‘Ramses II tank,’ a specific model of a main battle

tank, are selected from different KGs. The entity

‘dbr:Ramses_Ⅱ_tank’ is from DBpedia and the entity

‘wd:Q17216121’ is from Wikidata.

First, the entities that represent the upper concepts

of ‘Ramses II tank’ are identified to determine their

similarity in the graph hierarchy. At this time, specific

properties that represent relation to upper concepts

(i.e., ‘dbo:type’ and ‘rdf:type’ in DBpedia) are used.

In the given example, the entities have 'main battle

tank' as a common hyper concept. If the same hyper-

concept does not exist in 1-hop relations for the

entities, entities of 2 or more hops are discovered and

compared. Second, to determine their similarity

depending on their datatypes, the common properties

for the two entities are discovered. In the example, the

two entities have common properties, ‘width’ and

‘height.’ In addition, the datatype values for the

properties are also same. Therefore, we can determine

that two entities in different KGs represent the same

real-world concept because they have a common

hyper-concept and datatype. Furthermore, the

explanation using this result is also trustworthy.

4 GLEE FRAMEWORK

4.1 Overall Framework

The overall systematic procedure of the GLEE

framework is depicted in Figure 2. The GLEE

framework consists of three modules: Graph

Structure-based Candidate Entity Discovery Module,

Entity Similarity Assessment Module, and Subgraph-

based Explanation Module.

Figure 2: Systematic procedure of the GLEE framework.

4.2 Graph Structure-Based Candidate

Entity Discovery

In this module, candidate entities that can be aligned

with the target entity on 𝐾𝐺



are found from 𝐾𝐺



. To

do so, the S-similarity between the target entity on

𝐾𝐺



and all entities on 𝐾𝐺



is calculated using the

existing embedding-based EA methods.

The target entity (𝑒





()

) can be simply represented

as follows.

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

126

(𝑒





()

)∈𝐸





(1)

where 𝐸





is the set of entities of 𝐾𝐺



. At this time,

𝑒





()

is an arbitrary entity on 𝐾𝐺



For ( 𝑒





()

) , the S-similarity with 𝑒





()

(for

∀𝑖,𝑖∈ℕ) is calculated using cosine similarity as

follows.

𝑆𝑠𝑖𝑚𝑒





()

,𝑒





()



𝐺𝐸



𝑒





()

∙𝐺𝐸



𝑒





()



𝐺𝐸





𝑒





()



𝐺𝐸



𝑒





()



(2)

where 𝑒





()

represents the 𝑖



entity in 𝐾𝐺



(𝑒





()

∈𝐸



, where 𝐸



is the set of entities of 𝐾𝐺



𝐺𝐸



𝑒





()

 and 𝐺𝐸



𝑒





()

 represent the

embeddings of 𝑒





()

on 𝐾𝐺



and 𝑒





()

in 𝐾𝐺



which are generated by the existing EA method.

𝐺𝐸



𝑒





()

 and 𝐺𝐸



𝑒





()

 represent the

lengths of the vectors 𝐺𝐸



𝑒





()

 and 𝐺𝐸



𝑒





()

.

Based on S-similarity, this paper designates the

top- 𝑘 entities with high S-similarity values as

candidate entities for 𝑒





()

. The algorithm discovering

the candidate entities for 𝑒





()

is as follows.

Algorithm 1.

Input: 𝐾𝐺



, 𝐾𝐺



, and the target entity (𝑒





()

)

Output: a candidate set of top-k entities on 𝐾𝐺



that have high S-similarity values with 𝑒





()

Step 1 Embedding generation. Generate

embeddings of the entities using a KGRL model

Step 2 S-similarity calculation. Calculate the

cosine similarity between the embeddings of the

𝑒





()

and each entity on 𝐾𝐺



(𝑒





,

)

Step 3: Top-k selection. Identify the candidate set

that is composed of the entities in 𝐾𝐺



that has

high similarity with 𝑒





()

4.3 Entity Similarity Assessment

This module comprises two parts: graph hierarchy-

based and datatype-based entity assessment.

4.3.1 Graph Hierarchy-Based Entity

Similarity Assessment

This submodule assesses whether the candidate

entities belong to the same or similar class with 𝑒





()

To do so, graph hierarchy-based similarity (H-

similarity) is calculated between 𝑒





()

and entities in

the candidate set. At this time, H-similarity is

calculated by gradually increasing the number of hops

of the link connected to 𝑒





()

. First, 1-hop hyper

entities of 𝑒





()

(𝑈𝐸



𝑒





()

) discovers as follows.

𝑈𝐸



𝑒





(



)

=



𝑢𝑒





𝑝𝑟





∈𝑈𝑃





, 𝑚∈ℕ

(3)

where <𝑒





(



)

,𝑝𝑟





,𝑢𝑒





> is 𝑚



subject-property-

object (s-p-o) triple of 𝐾𝐺



. 𝑢𝑒





and 𝑝𝑟





are 𝑚



hyper entity of 1-hop and its property. 𝑈𝑃



is the set

of properties of 𝐾𝐺



that depict hyper-hypo relations.

In DBpedia, a representative KG, 'dbo:type' or

'rdf:type' are examples of these properties.

Like above, 1-hop hyper entities of 𝑒





(



)

(𝑈𝐸



𝑒





(



)

) discovers as follows.

(𝑈𝐸



𝑒





(



)

)=



𝑢𝑒





𝑝𝑟





∈𝑈𝑃





,𝑛∈ℕ

(4)

where <𝑒













,𝑝𝑟





,𝑢𝑒





> is 𝑛



s-p-o of 𝐾𝐺



𝑢𝑒





and 𝑝𝑟





are 𝑛



hyper entity of 1-hop and its

property. 𝑈𝑃



is the set of properties of 𝐾𝐺



that

depict hyper-hypo relations (𝑖



∈ℕ).

In the GLEE framework, it is assumed that the

𝑈𝑃



and 𝑈𝑃



are already discovered. Applying the

same method to the 1-hop hyper entities, hyper

entities with 2 or more hops can also be discovered as

follows: 𝑈𝐸



𝑒





()

 and 𝑈𝐸



𝑒





(



)

 (𝑝=

1,..,𝑃,𝑞=1,..,𝑄).

To compare hyper entities of 𝑒





()

and 𝑒





(



)

entity name-based BERT embeddings are generated.

Since entity names of hyper concepts (e.g. 'ship' and

'tank') are typically abstract compared to entity names

of hypo concepts (e.g. 'USS Inchon' and 'Ramses 2

tank'), it can be assumed that BERT works well on

names of hyper entities.

As a next, H-similarity between 𝑒





()

and

𝑒





(



)

(𝐻𝑠𝑖𝑚𝑒





()

,𝑒





(



)

) is calculated based

on the number of hops and BERT-based word

embeddings as follows.

𝐻𝑠𝑖𝑚



𝑒





(



)

,𝑒

𝐾𝐺

(𝑖

′

)



=max

,,,

𝑠𝑖𝑚𝑢𝑒





,𝑢𝑒







𝑀

(

𝑝,𝑞

)

(5)

where

𝑢𝑒





is the 𝑚



entity of p hops in

𝑈𝐸



𝑒





(



)

and 𝑢𝑒





is the 𝑛



entity of q hops in

Graph Hierarchy and Language Model-Based Explainable Entity Alignment of Knowledge Graphs

127

𝑈𝐸





𝑒















. 𝑠𝑖𝑚

(

𝑢𝑒

𝑚

𝑝

,𝑢𝑒

𝑛

𝑞

)

is the BERT

embedding-based cosine similarity between the

names of the

𝑢𝑒





and 𝑢𝑒





. 𝑀

(

𝑝,𝑞

)

represents the

maximum function as follows.

𝑀

(

𝑝,𝑞

)

=

𝑝(𝑝≥𝑞)

𝑞(𝑝<𝑞)

(6)

At this time, the two hyper entities from 𝐾𝐺



and

𝐾𝐺



that maximize the score (argmax

𝑢𝑒

𝑚

𝑝

,𝑢𝑒

𝑛

𝑞





𝑢𝑒

𝑚

𝑝

,𝑢𝑒

𝑛

𝑞





(

,

)

), in

other words, the hyper entities used to calculate the

H-similarity, are defined as the common hyper

entities in 𝐾𝐺



and 𝐾𝐺



and will be utilized in

explanation.

4.3.2 Datatype-Based Entity Similarity

Assessment

In this sub-module, D-similaity is calculated to

compare the object values of 𝑒





(



)

with those of

𝑒





(



)

(

𝑓𝑜𝑟 ∀ 𝑖



)

. At this time, the objects of 𝑒





(



)

and 𝑒





(



)

are limited to those with a ‘datatype

property.’ In other words, the objects are data values,

such as integers, characters, or strings. For all s-p-o

triples of 𝐾𝐺



and 𝐾𝐺



, a set of properties with data

values as objects is found for each as follows.

𝐷𝑃



𝑒





()

=



𝑝𝑟



𝑑



∈𝐷



(7)

𝐷𝑃



𝑒





(



)

=



𝑝𝑟



𝑑



∈𝐷



(8)

where 𝑑



and 𝑑



are 𝑚



and 𝑛



data values of the

objects of s-p-o triples of 𝐾𝐺



and 𝐾𝐺



. 𝐷 is a set of

data values.

For each 𝑝𝑟



(for ∀ 𝑚), this paper finds a

property that is semantically similar to 𝑝𝑟



among

the elements of 𝐷𝑃



𝑒





(



)

. To do so, BERT-based

similarity between the properties is calculated. As a

result, this paper obtains the semantically similar

property to 𝑝𝑟



(𝑠𝑠𝑝

(

𝑝𝑟



)

) in 𝐷𝑃



𝑒





(



)

:

𝑠𝑠𝑝

(

𝑝𝑟



)

=

𝑁𝑈𝐿𝐿, 𝑖𝑓 argmax





𝑠𝑖𝑚

(

𝑝𝑟



,𝑝𝑟



)

<𝛿

argmax





𝑠𝑖𝑚

(

𝑝𝑟



,𝑝𝑟



)

, 𝑜/𝑤

(9)

where 𝑠𝑖𝑚

(

𝑝𝑟



,𝑝𝑟



)

is the cosine similarity between

BERT embeddings of the names of 𝑝𝑟



and 𝑝𝑟



. At

this time, if the argmax





𝑠𝑖𝑚

(

𝑝𝑟



,𝑝𝑟



)

value is

smaller than predefined threshold 𝛿 , 𝑝𝑟



determined to be not the same as 𝑝𝑟



Next, the data value of 𝑠𝑠𝑝

(

𝑝𝑟



)

and that of 𝑝𝑟



are extracted and compared. If the data value is a

string of natural language texts, BERT embedding is

utilized to determine whether they are same or not. As

a result, this paper identifies data value pairs

consisting of a data value of 𝑝𝑟



and that of

𝑠𝑠𝑝

(

𝑝𝑟



)

for all 𝑚. At this time, the two elements of

every pair must be identical or semantically similar.

After counting the number of elements of 𝐷𝑃



𝑒





()



and the identified data pairs, the D-similarity is

calculated as follows.

𝐷𝑠𝑖𝑚

𝑛𝑜(data pairs with same data values)

𝑛𝑜(elements o

𝐷𝑃





𝑒





(



)



)

(10)

The D-similarity calculation procedure is

summarized as following algorithm. candidate

entities for 𝑒





()

is as follows.

Algorithm 2.

Input: sets of the properties of 𝑒





()

and 𝑒





(



)

that have data values as objects

Output: D-similarity between two object values.

Step 1 Property finding For all s-p-o triples of 𝐾𝐺



and 𝐾𝐺



, find a set of properties with data values

as objects and set them to 𝐷𝑃



𝑒





(



)

 and

𝐷𝑃





𝑒















, respectively

Step 2 Property matching For each 𝑝𝑟



, find

semantically identical property of it among the

properties of 𝐷𝑃





𝑒















and set it to 𝑠𝑠𝑝

(

𝑝𝑟



)

(for ∀ 𝑚)

Step 3 Find data value of matched properties

Extract data value linked to 𝑝𝑟



and 𝑠𝑠𝑝

(

𝑝𝑟



)

and

identify whether they are same or not

Step 4: D-similarity Calculation Calculate the D-

similarity based on the ratio of the matched data

values to the total number of datatype properties of

𝑒





(



)

4.4 Subgraph-Based Explanation

Using the S-similarity, H-similarity, and D-similarity,

one of the candidate entities of 𝐾𝐺



is determined as

the aligned entity of

𝑒





()

. The aligned entity of 𝑒





()

(

𝐴𝐸



𝑒





()



)

is determined as follows.

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

128

𝐴𝐸



𝑒





()



=argmax













𝛼∙𝑆𝑠𝑖𝑚



𝑒





()

,𝑒





(



)



+𝛽∙𝐻𝑠𝑖𝑚



𝑒





()

,𝑒





(



)



+ 𝛾∙𝐷𝑠𝑖𝑚



𝑒





()

,𝑒





(



)



(11)

where 𝛼 , 𝛽, and 𝛾 are the hyperparameters ( 0≤

𝛼,𝛽,𝛾≤ 1, 𝛼+𝛽+𝛾=1).

To explain the EA results, this paper utilizes the

subgraphs of the target entity and the corresponding

aligned entity extracted from 𝐾𝐺



and 𝐾𝐺



respectively. The subgraphs are generated with two

phases as follows. First, this paper generates paths

from for the target and aligned entity to the upper

entities they have in common, respectively. Second,

for the target and aligned entity, the paths are

expanded using the semantically similar datatype

properties and their data values, respectively. An

example of the subgraphs is illustrated in Figure 1

As shown in Figure 1, the two subgraphs are

extracted from different KGs but are composed of

semantically similar properties and entities. The

upper entities, properties, and data values (a.k.a. s-p-

o) that are common to the two subgraphs, are used to

explain why two entities in different KGs are aligned.

For example, in Figure 1, ‘dbr:Ramses_Ⅱ_tank’ in

𝐾𝐺



and ‘wd:Q17216121’ in 𝐾𝐺



are aligned

because they have the hyper entity with the same

entity name (main battle tank) and the common

datatype properties with the same data values (width

3.42 and height 2.40).

5 PERFORMANCE EVALUAITON

To demonstrate the GLEE framework's performance,

this paper selects entities of DBpedia and Wikidata.

DBpedia and Wikidata are KGs that are generated

based on Wikipedia but have different term usage and

structure since they are developed and maintained by

different organizations. The selected entities belong

to 5 categories: ‘vessels,’ ‘tanks,’ ‘aircrafts’

‘generals,’ and ‘military units.’

For the entities, we generate the pairs of entities

that one belongs to DBpedia and the other belongs to

Wikidata. Some pairs may consist of entities

depicting the same real-world concept (‘exactly same’

in the table), while others may not. Depending on the

entities, entity pairs are classified into three types:

exactly same pair, intra-category pair, and inter-

category pair. The first is the pairs consist of two

entities which are semantically same, the second are

pairs consist of entities that are not same and belong

to the same category, and the third are pairs consist of

entities that are not same and belong to different

categories.

For the three kinds of entity pairs, this paper

calculates H-similarity and D-similarity. For each

category, this paper calculates the average values of

H-similarity and D-similarity for the three types of

pairs. At this time, D-similarity is calculated for

DBpedia and Wikidata respectively, since D-

similarity depends on the number of datatype

properties of the target entity. One regards the entity

of DBpedia in the pair as the target entity, and the

other regards the entity of Wikidata in the pair as the

target entity. This paper does not calculate S-

similarity in the experiment since it is just applying

existing KGRL methods, and we do not have

contribution to it. The similarities are summarized in

Table 1.

Table 1: Similarity of the pairwise entities.

Category

𝐻𝑠𝑖𝑚

𝐷𝑠𝑖𝑚

(DBpedia)

𝐷𝑠𝑖𝑚

(Wikidata)

Vessels

Exactly

same

0.795 0.733 0.733

Intra-

cate

0.527 0.000 0.000

Inte

cate

0.505 0.000 0.000

Tanks

Exactly

same

0.972 0.403 0.594

Intra-

cate

0.783 0.000 0.000

Inte

cate

0.647 0.000 0.000

Aircrafts

Exactly

same

0.753 0.345 0.424

Intra-

cate

0.629 0.100 0.143

Inte

cate

0.613 0.000 0.000

Generals

Exactly

same

0.683 0.793 0.442

Intra-

cate

0.661 0.000 0.000

Inte

cate

0.479 0.000 0.000

Units

Exactly

same

0.917 0.572 0.642

Intra-

cate

0.734 0.000 0.000

Inte

cate

0.546 0.000 0.000

For all entities, ‘exactly same’ entity pairs have

the highest H-similarity and D-similarity values. It is

proven that the proposed similarities are effective for

EA. In case of H-similarity, ‘intra-category’ has

higher H-similarity value than that of ‘inter-category.’

It is because intra-category pairs have relatively short

paths to the common hyper entity than inter-category

pairs. In case of D-similarity, values of ‘intra-

Graph Hierarchy and Language Model-Based Explainable Entity Alignment of Knowledge Graphs

129

category’ and ‘inter-category’ are almost zero. That

is, it is hard to have the same datatype property with

same data value of the entities that depict different

real-world concepts.

The core contribution of the GLEE framework is

to provide explanation of EA results using graph

hierarchy and datatypes. To show how GLEE

framework explains the EA result, this paper selects

a pair of entities as an example that was aligned in the

experiment. The entities depict ‘Leopard 2’ which is

a tank and extracted the subgraphs from DBpedia and

Wikidata. The subgraphs are depicted in Figure 3.

Even though the two graphs are extracted from the

different KGs, they both have the hyper entities which

depict ‘main battle tank.’ Also, the datatype

properties in the subgraphs have the same meaning

and the datatype values are same. Using these

subgraphs, it can be acknowledged that the aligned

entities depict the same real-world concept.

Figure 3: An example of generated subgraphs.

6 CONCLUSIONS

This paper proposed a method to align the entities of

different KGs not only using the graph structure, but

also using the graph hierarchy and datatypes. In the

proposed method, graph structure-based EA model is

applied to derive the candidate pairs of entities to be

aligned, and the information on graph hierarchy and

datatypes is utilized to find the exact pairs. By doing

so, alignment performance can be improved, and

explanation of the alignment can also be provided.

However, this paper has the following limitations.

First, it cannot provide an explanation in perspective

of graph structural similarity. That is, the information

on the relations with multi-hop neighbor entities is

not provided as an explanation. Second, the impact of

each property, object or datatype to the alignment is

not provided, whether they are critical for alignment

or somewhat meaningless. Therefore, in our further

research, we will develop a method to discover

neighbor entities which are crucial for alignment and

provide weights that depict importance of properties

and objects of the aligned entities.

ACKNOWLEDGEMENTS

This research was supported by Basic Science

Research Program through the National Research

Foundation of Korea (NRF) funded by the Ministry

of Education (RS-2023-00275408).

REFERENCES

Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., &

Yakhnenko, O. (2013). Translating embeddings for

modeling multi-relational data. Advances in neural

information processing systems, 26.

Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018).

Bert: Pre-training of deep bidirectional transformers for

language understanding. arXiv preprint

arXiv:1810.04805.

Fanourakis, N., Efthymiou, V., Kotzinos, D., &

Christophides, V. (2023). Knowledge graph embedding

methods for entity alignment: experimental review.

Data Mining and Knowledge Discovery, 37(5), 2070-

2137.

Ji, G., He, S., Xu, L., Liu, K., & Zhao, J. (2015, July).

Knowledge graph embedding via dynamic mapping

matrix. In Proceedings of the 53rd annual meeting of

the association for computational linguistics and the

7th international joint conference on natural language

processing (volume 1: Long papers) (pp. 687-696).

Kim, J., Kim, K., Sohn, M., & Park, G. (2022). Deep

Model-Based Security-Aware Entity Alignment

Method for Edge-Specific Knowledge Graphs.

Sustainability, 14(14), 8877.

Kipf, T. N., & Welling, M. (2016). Semi-supervised

classification with graph convolutional networks. arXiv

preprint arXiv:1609.02907.

Li, Y., Qin, D., & Yang, X. (2020, December). Path

modeling based on entity-connectivity for knowledge

base completion. In 2020 7th International Conference

on Information Science and Control Engineering

(ICISCE) (pp. 984-989). IEEE.

Lin, Y., Liu, Z., Luan, H., Sun, M., Rao, S., & Liu, S. (2015).

Modeling relation paths for representation learning of

knowledge bases. arXiv preprint arXiv:1506.00379.

Pires, T., Schlinger, E., & Garrette, D. (2019). How

multilingual is multilingual BERT?. arXiv preprint

arXiv:1906.01502.

Qi, D., Chen, S., Sun, X., Luan, R., & Tong, D. (2023). A

multiscale convolutional gragh network using only

structural information for entity alignment. Applied

Intelligence, 53(7), 7455-7465.

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

130

Shen, Y., Li, Z., Wang, X., Li, J., & Zhang, X. (2021,

October). Datatype-aware knowledge graph

representation learning in hyperbolic space. In

Proceedings of the 30th ACM International Conference

on Information & Knowledge Management (pp. 1630-

1639).

Tam, N. T., Trung, H. T., Yin, H., Van Vinh, T., Sakong,

D., Zheng, B., & Hung, N. Q. V. (2020). Entity

alignment for knowledge graphs with multi-order

convolutional networks. IEEE Transactions on

Knowledge and Data Engineering, 34(9), 4201-4214.

Tang, W., Su, F., Sun, H., Qi, Q., Wang, J., Tao, S., & Yang,

H. (2023, February). Weakly supervised entity

alignment with positional inspiration. In Proceedings of

the Sixteenth ACM International Conference on Web

Search and Data Mining (pp. 814-822).

Trisedya, B. D., Salim, F. D., Chan, J., Spina, D., Scholer,

F., & Sanderson, M. (2023). i-Align: an interpretable

knowledge graph alignment model. Data Mining and

Knowledge Discovery, 37(6), 2494-2516.

Wang, Z., Zhang, J., Feng, J., & Chen, Z. (2014, June).

Knowledge graph embedding by translating on

hyperplanes. In Proceedings of the AAAI conference on

artificial intelligence (Vol. 28, No. 1).

Wang, Z., Lv, Q., Lan, X., & Zhang, Y. (2018). Cross-

lingual knowledge graph alignment via graph

convolutional networks. In Proceedings of the 2018

conference on empirical methods in natural language

processing (pp. 349-357).

Wang, S., Wei, X., Dos Santos, C. N., Wang, Z., Nallapati,

R., Arnold, A., & Philip, S. Y. (2021, December).

Knowledge graph representation via hierarchical

hyperbolic neural graph embedding. In 2021 IEEE

International Conference on Big Data (Big Data) (pp.

540-549). IEEE.

Wang, H., Wang, Y., Li, J., & Luo, T. (2022). Degree aware

based adversarial graph convolutional networks for

entity alignment in heterogeneous knowledge graph.

Neurocomputing, 487, 99-109.

Wu, Y., Liu, X., Feng, Y., Wang, Z., Yan, R., & Zhao, D.

(2019a). Relation-aware entity alignment for

heterogeneous knowledge graphs. arXiv preprint

arXiv:1908.08210.

Wu, Y., Liu, X., Feng, Y., Wang, Z., & Zhao, D. (2019b).

Jointly learning entity and relation representations for

entity alignment. arXiv preprint arXiv:1909.09317.

Xiao, H., Huang, M., Hao, Y., & Zhu, X. (2015a). Transg:

A generative mixture model for knowledge graph

embedding. arXiv preprint arXiv:1509.05488.

Xiao, H., Huang, M., Hao, Y., & Zhu, X. (2015b). TransA:

An adaptive approach for knowledge graph embedding.

arXiv preprint arXiv:1509.05490.

Zeng, K., Li, C., Hou, L., Li, J., & Feng, L. (2021). A

comprehensive survey of entity alignment for

knowledge graphs. AI Open, 2, 1-13.

Zhang, Z., Cai, J., Zhang, Y., & Wang, J. (2020, April).

Learning hierarchy-aware knowledge graph

embeddings for link prediction. In Proceedings of the

AAAI conference on artificial intelligence (Vol. 34, No.

03, pp. 3065-3072).

Zou, X. (2020, March). A survey on application of

knowledge graph. In Journal of Physics: Conference

Series (Vol. 1487, No. 1, p. 012016). IOP Publishing.

Graph Hierarchy and Language Model-Based Explainable Entity Alignment of Knowledge Graphs

131