Systematic Approach for Measuring Semantic Relatedness between Ontologies

Abstract

Measuring ontology matching is a critical issue in knowledge engineering and supports knowledge sharing and knowledge evolution. Recently, linguistic scientists have defined semantic relatedness as being more significant than semantic similarities in measuring ontology matching. Semantic relatedness is measured using synonyms and hypernym–hyponym relationships. In this paper, a systematic approach for measuring ontology semantic relatedness is proposed. The proposed approach is developed with a clear and fully described methodology, with illustrative examples used to demonstrate the proposed approach. The relatedness between ontologies has been measured based on class level by using lexical features, defining semantic similarity of concepts based on hypernym–hyponym relationships. For evaluating our proposed approach against similar works, benchmarks are generated using five properties: related meaning features, lexical features, providing technical descriptions, proving applicability, and accuracy. Technical implementation is carried out in order to demonstrate the applicability of our approach. The results demonstrate an achieved accuracy of 99%. The contributions are further highlighted by benchmarking against recent related works.

1. Introduction

Recently, ontology has been described as one of the most widespread modern knowledge representation techniques, which provides a source of precise knowledge; an ontology can be considered as a formal definition of concepts along with relationships between these concepts. In the knowledge engineering domain, the primary aims of ontologies are knowledge representation and knowledge sharing. In the past two decades, ontologies have been used prominently for representing information in a machine-readable and processable way [1].

More recently, ontologies have been used for exchanging business and scientific data via authenticated and standardized knowledge representation structures, replacing older techniques. In this context, there have been many efforts to develop and implement ontologies from scientific and commercial perspectives. The quality of an ontology is highly dependent on both the knowledge that it represents and the format that is used. Knowledge concerning identical concepts could be represented by different syntaxes. Hence, the existence of multiple ontologies with different syntaxes yet identical same semantics is logical and justifiable.

On the other hand, according to [2], the ontology should have a continuous maintenance process to keep it contemporary as a response to the updates and challenges in the domain it models. In addition, usually, users are aided with a specific ontology to accomplish their tasks, for instance, searching data, and at some point, they will need to expand it to adapt to the dynamic nature of their requirements. In addition, [3,4] highlight that finding correspondences between entities of two or more given ontologies, known as “ontology matching,” is a challenging task.

Measuring the similarity between two ontologies is a potential area for innovation and improvement in the ontology industry. Applying intelligent solutions for automatically calculating similarities between ontologies is a challenging task with applications in many real-world applications that require the integration of semantic information [5], such as sentiment analysis of Twitter posts [6,7].

According to [8], semantic similarity could be measured using synonyms, hyponyms, and hypernyms. A hypernym is a generalization of a more specific term; for instance, color is a hypernym of red. A hyponym is a term that is more specific than a given more general term, i.e., a term whose meaning is included in that of another term; for instance, red is a hyponym of color. A synonym of a term is one that has the same meaning; two concepts are considered to be synonymous if they are having the same meaning; for instance, red is a synonym for ruby.

The linguistic-semantic technique is widely used for measuring similarities between ontologies, i.e., matching lexical features between ontologies [9]. The linguistic-semantic technique is concerned with measuring similarities between lexical units and finding taxonomic relations. In knowledge engineering, taxonomic relations in an ontology are denoted by the hierarchical arrangement of concepts based on hypernym/hyponym relations. The development of taxonomic relations serves as the foundation of ontology construction [10].

In the knowledge engineering domain, measuring semantic relations, i.e., semantic relatedness between ontologies can be achieved based on lexical ontology units. In this context, there are numerous research works that investigate measuring ontology similarities, yet, measuring semantic relatedness demands more attention [11]. Semantic relatedness between words in ontologies is the main conceptualization for measuring ontology similarities [12] and is analyzed and defined by examined hypernym/hyponym relations [13].

In this paper, we propose a new approach for defining semantic relatedness between ontologies based on measuring synonym, hyponym, and hypernym components. Illustrative running examples are used for presenting and explaining our approach. Logic rules have been used effectively for representing semantic components of an ontology [14]; therefore, we use logic rules for representing our assumptions and definitions. For proving its applicability, we implement our approach in Python. The accuracy of our approach is evaluated using precision and recall metrics. Finally, we use standard benchmarking for highlighting our contributions, based on comparisons with previous similar works.

The remainder of the paper is organized as follows: Section 2 presents related works. Section 3 and Section 4 discuss the methodology and the evaluation, respectively. Finally, Section 5 provides a discussion and conclusion.

2. Related Works

In this section, we review related works that are used lexical relations, particularly the hypernym–hyponym relationship.

Ref. [15] designed a method for evaluating the inter-relationship between ontology terms using a graph structure of the seed ontology. Reference [16] proposed a method for calculating a ratio between hypernym and hyponym between ontologies by using the Levenshtein distance algorithm; the aim was to define semantic similarity value. Reference [5] illustrates the importance of discovering hypernym relationships between ontologies for measuring semantic similarities. The study [5] developed an automatic approach for discovering hyponym relationships across ontology elements. However, the approach interferes with discovering synonym relationships and thus leads to curtailment of the accuracy. Reference [17] proposed a probabilistic method for contextual advertising based on ontology. This method assigns weights to the candidate instances based on the probability that it is a hypernym of a concept, reflecting the importance of measuring hypernym relationships. Reference [18] addressed the problem of matching foundational ontologies by defining the hypernym relation carried out in the definitions layout, i.e., ontology meta-data. Reference [19] developed an ontology-based approach for labeling influential topics of scientific articles by using synonym, hypernym, and homonym relationships. The study [20] used hypernym and homonym relationships to define entity matching in ontology-based attribute graphs.

Ref. [21] developed a method for calculating the availability of information content in an ontology based on hypernyms-hyponyms taxonomic knowledge. Reference [1] developed a semantic matching framework for the integration of large knowledge bases using the linguistic-semantic technique, using the WordNet electronic dictionary as a data source, and evaluating using the precision and recall technique. Reference [22] developed ad hoc algorithms using probabilistic association to analyze lexical units; these algorithms generate a directed acyclic graph for finding hypernyms. Reference [10] designed a method for extracting taxonomic relations from a domain text by using a combination of the web approach and word sense disambiguation, using search engines for collecting targeted patterns, and matching each pattern with its corresponding hypernym from Wordnet.

Ref. [23] developed an approach for extending a learning ontology-based model by adding missing concepts. The missing concepts are defined by examining associated ontologies and evaluating similarities based on linguistic patterns such as hyponym-hypernym measurement. Reference [24] used the relationship of synonyms, hypernyms, and homonyms for the explanation of ontology queries. The results demonstrated the importance of discovering the synonym, hypernym, and homonym relationships. Reference [25] proposed a method for enhancing query sentences for ontology-oriented search by defining hypernym and homonym relationships. Reference [26] proposed a method to deal with semantic heterogeneity where data are organized into two different layers by defining hypernym and homonym relationships. Reference [27] proposed a model for ontology matching for large-scale ontologies in big data, which is based on probability, logical, and similarities between lexical units.

Ref. [28] addressed the challenge of the continuous changing of the ISO 27001-based security ontology by defining the hypernyms and hyponyms for security concepts. The hypernyms and hyponyms were extracted from DBpedia. Reference [29] used the WordNet lexical database to define hypernym relation similarities by comparing vocabularies one by one. Reference [30] proposed a model for automatically extracting Arabic text by parsing sentences meaning which can be used for developing ontologies. This method used hypernym and hyponym relationships to develop a query triple-answering method. Reference [12] proposed a semantic evaluation dataset for defining semantic similarity and relatedness. This study confirms the importance of word relatedness in measuring semantic similarity. Reference [31] developed an ontology for IT Operations that is applicable to IT dynamic applications. This ontology has been evaluated by measuring its similarity based on linguistic patterns, i.e., by hyponym-hypernym measurement. Reference [32] used synonym, antonym, hypernym, and hyponym relationships for extracting and analyzing Quran concepts. Reference [33] developed an approach to enhance the measurement of semantic relationships between hypernym-based concepts and homogeneous concepts. The results proved the usefulness of integrating multiple classifications in defining ontologies matching. Reference [34] defined three disciplines for ontology matching which are databases, natural language processing, and machine learning. The research proved high potentiality for developing automation tools in ontology matching based on Natural language processing (NLP).

Table 1. Summary of related works.

#	Reference	Purpose	Methodology
1	Zhao and Zhang [15]	Evaluated the inter-relationship between ontologies	Graph structure of the seed ontology
2	Wang et al. [16]	Calculated sematic relationship between ontologies	Levenshtein distance algorithm
3	Huitzil et al. [5]	Discovered hypernym relationships between ontologies	Ad hoc algorithm
4	Chen et al. [17]	Measured hypernym relationships	Probabilistic method for contextual advertising based on ontology
5	Kamel et al. [18]	Defined the hypernym relation carried out in ontology meta-data	Matching foundational ontologies
6	Kim et al. [19]	Labeled influential topics of scientific articles	Defining synonym, hypernym, and homonym relationships
7	Ma et al. [20]	Defined entity matching in ontology	Attributed graphs
8	Batet and Sánchez [21]	calculated availability of Information content in ontology	Defining hypernyms-hyponyms taxonomic knowledge
9	Rinaldi et al. [1]	Integrated large knowledge bases ontologies	Using the linguistic-semantic technique
10	Dong et al. [22]	Analyzed lexical units	Probabilistic association
11	Doan et al. [10]	Extracted taxonomic relations from a domain-text	Using Wordnet as a reference
12	Jabla et al. [23]	Extended ontology by adding missing concepts	By using examining associated ontologies
13	Afuan et al. [24]	Provided explanation on ontology queries.	Define relationship between synonyms, hypernyms, and homonyms
14	Chiến [25]	Enhanced query search for ontology-oriented domains	Defining hypernym and homonym relationships
15	Giunchiglia et al. [26]	Measured semantic heterogeneity	Defining hypernym and homonym relationships
16	Mountasser et al. [27]	Matched for large-scale ontologies	Defining similarities between lexical units using probability, and logical
17	Sanagavarapu et al. [28]	Addressed the challenge continuous changing of the ISO 27001-based security ontology	Defining the hypernyms and hyponyms for security concepts
18	Ocker et al. [29]	Defined hypernym relation similarities between ontologies	By comparing vocabularies one by one.
19	Saber et al. [30]	Automatically extracting text	Ad hoc algorithm
20	Salaev et al. [12]	Developed semantic evaluation dataset	Ad hoc algorithm
21	Uceda-Sosa et al. [31]	Developed an ontology for IT Operations	measuring the similarity based on linguistic patterns
22	Rahman and Rasid [32]	Extracted and analyzed Quran concepts	Defining synonym, antonym, hypernym, and hyponym relationships
23	Chen et al. [33]	Developed an approach to enhance the measurement of semantic relationships between hypernym-based concepts	Ad hoc algorithm
24	Trojahn et al. [34]	Defined three disciplines for ontology matching	Survey
25	Jin et al. [35]	Used standard algorithms for measuring semantic relatedness	Combined factors of semantic similarity and relatedness

summarizes the considered related works in terms of purposes and methodologies used. The work in [35] used standard algorithms for measuring semantic relatedness between ontologies. The approach in [35] associated semantic similarity with relatedness, while our approach provides a flexible approach that measures synonym and homonym relatedness separately.

The aforementioned discussion of selected related works indicates a research gap, specifically a decisive demand for considering semantic relatedness in measuring ontology similarities. In addition, there is a need for defining relatedness with a specific percentage threshold.

3. Methodology

In this section, our methodology to measure the matching between two ontologies is explained in detail, accompanied by illustrated examples. An ontology is described in terms of classes and associated attributes [36]. Hence, our methodology has been developed based on measuring the matching between classes of targeted ontologies. In other words, the methodology starts the measuring process from the low level of ontology, which is the class. In the following, assumptions and definitions have been described using mathematical logic. The letter “O” denotes an ontology, the letter “C” denotes a class, and the letters “DP” denote a data property.

Assumption 1.

∀ O, Ǝ C: ontology(O) ∧class (C) ∧ contain(O, C) ⟹ True

Assumption 2.

∀ C, Ǝ DP: class (C) ∧ data_property(DP) ∧ contain(C, DP) ⟹ True

Assumption 1 states that any ontology has at least one class; Assumption 2 states that any class is described by at least one data property. Our methodology is based on calculating similarities between ontologies’ classes, therefore, any ontology that does not include at least one class, or includes a class without data properties cannot be considered in our approach.

Definition 1.

∀ C1, C2, O1, O2: class(C1) ∧ class(C2) ∧ ontology (O1) ∧ ontology (O2) ∧ C1 ∈ O1 ∧ C2 ∈ O2 ∧ similarity (C1, C2) ⟹ similarity (O1, O2)

Definition 1 states that where class C1 belongs to ontology O1 and class C2 belongs to ontology O2, the similarity between ontologies O1 and O2 is defined as the similarity between classes C1 and C2.

Definition 2.

∀ O1, O2: ontology (O1) ∧ ontology (O2) ⟹ similarity (O1, O2) ≠ similarity (O2, O1)

Definition 2 states that the similarity between ontology “A” and ontology “B” is not equal to the similarity between ontology “B” and ontology “A”.

To measure the similarity between two classes, or between a class and a subclass, the following steps should be executed in order.

• Prepare the properties of each class by applying the normalization process which is:
  • Remove stopping words; e.g.: “of”, “in”, or “the”;
  • Remove class signs; e.g.: suppose the class name is “car” and the property is “car model” then the data property will be “model”;
  • Apply stemming for each data property.
• Apply cartesian product to compare the properties of the two targeted classes. The comparison becomes true if and only if one of the following applies:
  • The properties are synonyms of each other;
  • If one data property has two words and another data property has one word, then the comparison becomes true if there is at least one word similar. For instance, the class “car” has the data property “plate number” and the class “vehicle” has the data property “number”, these two data properties could be considered similar;
  • If the two data properties have two or more words: (1) remove similar words; (2) if the remaining words are synonyms of each other, then similarity becomes true. For instance, consider the data property “close date” and the data property “date open”, then the word “date” will be eliminated; the word “close” is not a synonym of the word “open”, thus these two data properties are not similar. As another example, consider the data property “open date” and the data property “access date”, the word “date” will be eliminated, and the word “open” is a synonym of the word “access”, thus the two data properties are similar.

This proposed approach follows corpus-based approaches where similarity is counted based on the number of overlapping data properties between the two ontologies [37].

We now give illustrated examples of the proposed methodology.

Example 1.

Consider the following two classes: “Scientific article” and “research” which belong to two different ontologies,

Table 2. Data properties of classes scientific article and research.

Class	Data Properties
Scientific article	Article author; submission date; article title, decision of the article; article identifier; domain
Research	Researcher; date of submission; title; decision; research identification; field, publication type

shows the data properties of these two classes.

Before applying the proposed methodology, the data properties of each class are modeled as a vector. We now implement the steps of the proposed methodology:

3.

Remove the stopping words.

The first step is removing the stopping words.

Table 3. Data properties after removing the stopping words.

Properties of Class “Scientific Article”	Properties of Class “Research”
Article author; submission date; article title; decision article; article identifier; domain	Researcher; date submission; title; decision; research identification; field

shows the data properties after removing the stopping words. For instance, the data property “decision of article” has become “decision article”.

4.

Remove the class signs.

The second step is removing the class signs. For instance, the property “article author” has become “author”.

Table 4. Data properties after removing the class signs.

Properties of Class “Scientific Article”	Properties of Class “Research”
Author; submission date; title; decision, identifier; domain	Researcher; date submission; title; decision; paper identification; field

shows data properties after removing the class signs. Table 4 is developed based on results from Table 3, i.e., after the stopping words have been removed in step 1.

5.

Stemming.

The third step is stemming for all data properties. For instance, the data property “identifier” from class “paper” will be stemmed to be “identify”.

Table 5. Data properties after applying stemming.

Properties of Class “Scientific Article”	Properties of Class “Research”
Author; submit date; title; decide, identify; domain	Research; date submit; title; decide; paper identify, field

shows data properties after applying stemming. Table 5 is developed based on results from Table 4, i.e., after removing the class signs.

6.

Apply cartesian product to properties.

The fourth step is to apply cartesian product to the properties. In this step, each property from the class “Scientific article” has been placed with each property from the class “research”. For instance, the data property “author” from class “Scientific article” has been placed against each property class “research”, the result becomes: ((author, research); (author, date submission); (author, title); (author, decision); (author, paper identify), (author, field)).

Table 6. Cartesian product of properties.

(author, research); (author, date submit); (author, title); (author, decide); (author, paper identify), (author, field)

(submit date, research); (submit date, date submit); (submit date, title); (submit date, decide); (submit date, paper identify); (submit date, field)

(title, research); (title, date submit); (title, title); (title, decide); (title, paper identify), (title, field)

(decide, research); (decide, date submit); (decide, title); (decide, decide); (decide, paper identify), (decide, field)

(identify, research); (identify, date submit); (identify, title); (identify, decide); (identify, paper identify), (identify, field), (identify, cost)

(domain, research); (domain, date submit); (domain, title); (domain, decide); (domain, paper identify), (domain, field)

shows the cartesian product of the properties, developed based on the results from Table 5, i.e., after stemming the data properties.

7.

Comparison

The preceding steps illustrate the normalization process and the fifth step is compares each pair of properties to discover similarities. The similarity is calculated based on the conditions detailed in step 2 of the approach.

Table 7. The positive results of comparison process.

Cardinality Pair	Synonym Similarity	Cause
(submit date, submit)	Yes	One data property has two words, and another data property has one word; there is one word is similar
(title, title)	Yes	The two data properties are equal
(decide, decide)	Yes	The two data properties are equal
(identify, paper identify)	Yes	One data property has two words, and another data property has one word; there is one word that is similar
(domain, field)	Yes	The two data properties are synonym

shows the positive results of the comparison process with the appropriate causal reason. Table 7 is developed based on results from Table 6, i.e., the comparison is computed between each cartesian pair.

Definition 3 defines the calculation for similarity percentage between classes, where similarity could be due to either synonyms or hypernyms.

Definition 3.

∀ C1, C2, No, T1, T2: class(C1) ∧ class(C2) ∧ (No = number of similar data properties between C1 and C2) ∧ (T1 = total number of data properties of C1) ∧ (T2 = total number of data properties of C2) ⟹ (similarity(C1, C2) = (No % T1)) ∧ (similarity (C2, C1) = (No % T2)).

In Definition 3, the similarity between two classes A and B is taken from Definition 2: similarity (A, B) = number of similar synonym data properties between classes A and B, divided by the total number of data properties in class A. As stated in Definition 2, the similarity between two classes A and B is not equal to the similarity between B and A. Therefore, the similarity between B and A is calculated as: similarity (B, A) = number of similar data properties between A and B, divided by the total number of data properties in class B.

Using Definition 3, we can now calculate synonym similarity between the two classes “scientific article” and “research”. The results from Table 6 show a total of five similar synonym data properties between class “Scientific Article” and class “research”. In other words, there are five cartesian pairs that are similar. Class “scientific article” has six data properties, five of which are similar to data properties from class “research”; therefore, synonyms_similarity (scientific article, research) = 83.33%. Class “research” has seven data properties, five of them are similar to data properties from class “scientific article”; therefore, synonyms_similarity (research, scientific article) = 71.43%.

In the modern linguistic-semantic technique, semantic relatedness is described by finding taxonomic relations. A taxonomic relation is defined by synonyms, hypernyms, and hyponyms.

Given that hypernyms and hyponyms could be considered two sides of the same coin, we have calculated hypernym relatedness.

Hence, defining similarity based on similar synonyms is not sufficient. We now demonstrate checking for hypernym relation means of an illustrated example.

Example 2.

Consider two classes, A and B.

Table 8. Data properties of classes A and B from example 2.

Class	Data Properties
A	Red, eagle, animal, oak, tea, fruit, a1, a2
B	Color, bird, tree, drink, apple, b1, b2, b3, b4

shows the data properties of these two classes.

Table 9. Result of hypernym similarity of data properties between classes A and B.

Cardinality Pair	Hypernym Similarity
(animal, bird)	Yes
(fruit, apple)	Yes

shows the result of hypernym similarity of data properties between classes A and B.

Table 10. Result of hypernym similarity of data properties between classes B, and A.

Cardinality Pair	Hypernym Similarity
(color, red)	Yes
(bird, eagle)	Yes
(tree, oak)	Yes
(drink, tea)	Yes

shows the result of hypernym similarity of data properties between classes B and A.

Using Definition 3, we calculate hypernym similarity between classes A and B. Class A has eight data properties, with two of them being hypernyms for data properties from Class B; therefore, hypernym_similarity (A, B) = 25%.

Class B has nine data properties, with four of them being hypernyms for data properties from class A; therefore, hypernym_similarity (B, A) = 44.44%.

In the following section, the applicability of our proposed approach has been presented.

Appendix A shows an example of a relatedness measurement between ontology with one class with ontology with hierarchically organized classes. Appendix B shows an example of a relatedness measurement between two published ontologies.

4. Implementation

In computing, examining the applicability of a proposed solution is critical in proving the success, and acceptability of the approach. In this section, the applicability of our proposed approach is demonstrated by means of implementation and benchmarking.

The WordNet lexical database [38], has been used to obtain a set of synonyms and hyponyms for every given word. Figure 1 shows the output from finding all synonyms of the word “submit”.

To implement the calculation of synonyms similarly (as per Example 1), the following steps were taken:

• Select one word from the cardinality pair;
• Extract all synonyms of the selected word from WordNet;
• Check all extracted synonyms against the other word in the cardinality pair. The comparison flag of the cardinality pair becomes true if one word in the cardinality pair is a synonym of another word in the cardinality pair. The full explanation of this scenario has been presented in Example 1 in Section 3.

Figure 2 shows the external Python libraries that have been used in the implementation. Figure 3 shows the algorithm, written in Python, for calculating synonym similarities between two classes.

The algorithm in Figure 3 has been used for implementing Example 1 using the Python programming language and the NLTK library.

The following steps are followed for implementing hypernym similarity:

• Select one word from the cardinality pair;
• Extract all hypernyms of the selected word from WordNet;
• Check all the extracted hypernyms against the other word in the cardinality pair. The comparison status of the cardinality pair becomes true if one word in the cardinality pair is a hypernym of another word in the cardinality pair. The full explanation of this scenario has been presented in Example 2 in Section 3.

Figure 4 shows all hypernyms of the word “tree” as extracted from WordNet.

Figure 5 shows the algorithm, written in Python, for calculating hypernym similarity between classes “A” and “B” and between classes “B” and “A”.

The algorithm in Figure 5 implements Example 2 using the Python programming language and the NLTK library. The implementation of these two algorithms demonstrates the theoretical and practical applicability of our proposed approach.

5. Evaluation

In the following section, accuracy has been validated by using precision, and recall parameters.

In the knowledge engineering domain, in addition to the applicability, accuracy must also be tested and validated for any proposed solution. We consider two facts for validating the accuracy of our proposed approach which are: (1) In this type of research, the accuracy of results is evaluated by measuring the systems’ outputs against reference sets [39]. (2) In semantic similarities systems, measuring the accuracy using precision and recall parameters has been used as a standard approach [40]. Hence, we have considered the two mentioned previously facts for measuring the accuracy of our proposed approach. Regarding the first fact, WordNet has been chosen as a reference set. Regarding the second fact, precision and recall values are calculated.

Our evaluation procedure, which follows our proposed methodology, is as follows:

• Create twenty classes by combining groups of five words in each class randomly. The words are collected by student volunteers’, and each volunteer creates four classes.
• Randomly divide the twenty classes into ten groups, two classes per group. The classes in each group are measured for similarity using our proposed methodology.
• Generate the cartesian pair for each group (two classes) as has been explained in the methodology.
• Calculate the synonym and hypernym similarities for each cartesian pair, as explained in Examples 1 and 2, using the algorithms in Figure 2 and Figure 4.
• Calculate precision, and recall as follows:

  Precision = TP/(TP + FP)

  Recall = TP/(TP + FN)

  where:

  TP = True positive

  FP = False positive

  TN = True negative

  FN = False negative

In this evaluation, we have ten groups of classes, each group having two classes, each class having five words. The synonym and hypernym similarities have been measured for each group. The size of the cartesian products for each group is twenty-five. For the ten groups, we therefore measured 250 cartesian pairs in total.

Table 11. The results of TP, FP, TN, and FN.

TP	30
FT	2
TN	218
FN	0

shows the results of TP, FP, TN, and FN.

• The last step in this evaluation procedure is calculating the accuracy of our proposed approach. Accuracy represents the number of correctly classified data instances over the total number of data instances. Hence, we use the following equation to define accuracy:

  Accuracy = TP + TN/(TP + TN + FP + FN).

Using this equation, accuracy is 99%.

In Section 4 and Section 5, the applicability and accuracy of our proposed approach have been demonstrated and proved. The following section contains the discussion and conclusion.

6. Discussion and Conclusions

The applicability and accuracy of our proposed approach have been demonstrated and proved. The following section contains the discussion and conclusion. In the knowledge engineering domain, knowing the similarity between ontologies is a significant issue and cornerstone in the processes of knowledge evolution and knowledge sharing. In the literature, various techniques have been used for determining the similarity between ontologies, such as deep learning, support vector machine, and computational linguistic methods. Among these, computational linguistic methods are considered the most practical technique. In language computational methods, measuring synonym similarity is the main technique for ontology matching. Recently, scientists in linguistic semantics are considering semantic relatedness as a more accurate metric for semantic matching.

Semantic relatedness can be measured by calculating both synonym and hypernym similarities. Therefore, measuring semantic relatedness between ontologies is providing deep insight into ontologies’ relevancy. In this paper, a systematic approach for measuring ontology semantic relatedness has been introduced.

In this research, we have combined concepts from linguistics and knowledge engineering sciences for measuring relatedness between ontologies. To the best of our knowledge, this is the first research that provides a flexible relatedness measurement methodology between ontologies at the class level.

The proposed approach provides a flexible measurement methodology. The relatedness between two ontologies is measured on a class-based level, i.e., at a low level of the ontologies. Synonym and hypernym similarities are measured separately using the standard WordNet lexical database. This paper provides a methodology with open technical details, which draws a clear picture of how our approach has been developed. First, a formalization of the tokenization of class data properties is described, followed by the creation of the cartesian pair of the two classes being compared. WordNet is then used as a reference model for calculating synonym and hypernym similarities. The applicability of our proposed approach has been demonstrated by implementation using the Python programming language. The accuracy of our proposed approach is demonstrated by calculating the recall and precision parameters, which show the accuracy of our model to be 99%. To the best of our knowledge, this is the first approach to introduce ontology semantic relatedness with full technical details. The benefits of using computational linguistics for measuring ontology matching have been proven over other methodologies; hence, we have adopted computational linguistic approaches, so that our contributions can be compared with others. The contributions of our proposed approach are now highlighted based on a comparison with previous related works.

Scientific benchmarking is used for evaluating our proposed approach with similar works, which consists of five properties: related meaning features, lexical features, providing technical descriptions, proving applicability, and accuracy. We now give further explanation of each parameter:

• Related meaning features: defining semantic similarities of concepts based on synonyms similarity relations.
• Lexical features: defining semantic similarities of concepts based on hypernym–hyponym relations.
• Providing technical descriptions: where technical explanations of the proposed method have been presented, where the reader can develop a deep understanding of the theory and be able to test and evaluate it.
• Proving applicability: providing a programming implementation for the proposed theory.
• Accuracy: where the accuracy of the proposed theory has been tested and presented.

Table 12. The results of the benchmark comparison.

#	Research	Related Meaning Features	Lexical Features	Providing Technical Descriptions	Proving Applicability	Accuracy
1	Zhao and Zhang (2018) [15]	√	√	X	X	X
2	Wang et al. (2018) [16]	√	√	X	X	X
3	Huitzil et al. (2019) [5]	X	√	X	√	√
4	Chen et al. (2019) [17]	√	√	X	X	√
5	Kamel et al. (2019) [18]	√	√	X	X	X
6	Kim et al. (2019) [19]	√	√	X	√	X
7	Ma et al. (2019) [20]	√	√	X	√	X
8	Batet and Sánchez (2020) [21]	X	√	X	√	X
9	Rinaldi et al. (2020) [1]	√	√	X	√	√
10	Dong et al., (2020) [22]	√	√	X	√	X
11	Doan et al. (2020) [10]	√	√	X	√	X
12	Jabla et al. (2021) [23]	√	√	X	X	X
13	Afuan et al. (2021) [24]	√	√	X	X	X
14	Chiến (2021) [25]	√	√	X	X	X
15	Giunchiglia et al. (2021) [26]	√	√	X	X	√
16	Mountasser et al. (2021) [27]	√	√	X	X	X
17	Sanagavarapu et al. (2022) [28]	√	√	X	√	X
18	Ocker et al. (2022) [29]	√	√	X	X	X
19	Saber et al. (2022) [30]	√	√	X	X	X
20	Salaev et al. (2022) [12]	√	√	X	X	√
21	Uceda-Sosa et al. (2022) [31]	√	√	X	√	X
22	Rahman and Rasid (2022) [32]	√	√	X	X	X
23	Chen et al. (2022) [33]	√	√	X	√	X
24	The proposed approach	√	√	√	√	√

shows the results of the scientific benchmarking of our proposed approach with related works. The “√” symbol denotes that this research has considered the property, and “X” signifies that the property has not been considered.

By considering Table 11, it is clear that our proposed approach has successfully satisfied all the benchmark parameters.

In this paper, we have used WordNet as a reference model. In future work, we are planning to use more than one standard reference, such as DBpedia. This way, the similarities are extracted based on more than one standard reference.