Abstract

Conceptual models are built to depict and analyze complex systems. They are made of concepts and relationships among these concepts. In a particular domain, conceptual models are helpful for different stakeholders to reach a clear and unified view of domain problems. However, the process of building conceptual models is time-consuming, tedious, and expertise required. To improve the efficiency of the building process, this paper proposes a configurable semantic-based (semi-) automatic conceptual model transformation methodology (SbACMT) that tries to reuse existing conceptual models to generate new models. SbACMT contains three parts: (i) a configurable semantic relatedness computing method building on the structured linguistic knowledge base “ConceptNet” (SRCM-CNet), (ii) a specific meta-model, which follows the Ecore standard, defines the rules of applying SRCM-CNet to different conceptual models to automatically detect transformation mappings, and (iii) a multistep matching and transformation process that employs SRCM-CNet. A case study is carried out to detail the working mechanism of SbACMT. Furthermore, through a systematically analysis of this case study, we validate the performance of SbACMT. We prove that SbACMT can support the automatic transformation process of conceptual models (e.g., class diagrams). The scalability of SbACMT can be improved by adapting the meta-model and predefined syntax transformation rules.

1. Introduction

Conceptual models, which are mainly built to represent the static characteristics of a system, can be used to refer to models formed after a conceptualization or generalization process [1]. A conceptual model represents concepts (entities) and relationships between them; it helps different stakeholders reach a clear and unified overview of a particular system or domain being modeled. In artificial intelligence field, a typical usage of conceptual models is to build expert systems and knowledge-based systems.

To achieve different kinds of purposes, several kinds of conceptual models (with modeling languages) have been proposed, such as “class diagrams,” “data flow models,” “entity-relationship models,” and “business process models.” Considering diverse research fields and practical cases, conceptual modeling theories are defined and relevant modeling tools are employed.

In practice, conceptual models have been built for various purposes, such as simulating groundwater system [1] and improving environmental science curriculum [2]. In the context of computer science, especially in the software engineering field, a typical kind of conceptual models is “class diagrams.”

Class diagrams are built in the requirement analysis phase and being refined in the system design phase; they are defined with the UML (Unified Modeling Language) notations. Based on the object-oriented theories, a class diagram is a complete mapping of the subjects and their relations of a particular domain being analyzed.

Domain-specific conceptual models can be used to generate domain ontologies and knowledge graph (since ontologies can be regarded as the schema of knowledge graph). The creation of semantic web can also benefit from qualified domain-specific (or domain-cross) conceptual models.

However, conceptual modeling is a time-consuming and iteratively evolve process, which is expertise required. Since conceptual models are useful to help solve complex problems, more and more specific conceptual models (in different domains, with diverse modeling languages and for different purposes) have to be built. Consequently, how to effectively build conceptual models attracts attention from both academics and industrial.

An intuitive idea is to merge/integrate the existing qualified conceptual models to get new ones. To realize this idea, model transformation methods shall be adopted. However, there are three main challenges in adopting model transformation methods to solve this problem:(1)Ch1: conceptual models may have different structures and formats. Furthermore, they are defined on different abstract levels. How to deal with the heterogonous problem?(2)Ch2: the semantic content conveyed in conceptual models is complex and may be ambiguous. How to find the same or similarity concepts automatically?(3)Ch3: how to ensure and validate the correctness of the model transformation process and the new generated model?

Considering the three challenges, we propose a configurable semantic-based (semi-) automatic conceptual model transformation method (SbACMT). The main contributions of this work are as follows:(1)Con1: focusing on generating new class diagrams (a kind of conceptual models), we define a meta-model on a higher abstract level to unify the structure and format (syntax).(2)Con2: considering the semantic conveyed in class diagrams, we propose a semantic relatedness computing method to automatically detect potential transformation mappings.(3)Con3: we propose an integrated transformation process that combines the meta-model-based transformation theory and the semantic relatedness computing method.

SbACMT is a general-purpose model transformation methodology. It aims at serving to the transformations among different kinds of conceptual models, instead of focusing on two specific kinds of conceptual models.

This paper is organized in 7 sections. Section 2 presents the preliminaries of this work. Section 3 introduces the meta-model defined in SbACMT and illustrates the semantic relatedness computing method “SRCM-CNet.” Section 4 details the integrated transformation process. Section 5 shows a case study and evaluates the performance of SbACMT. Section 6 gives the related work, while Section 7 draws a conclusion.

2. Preliminaries

2.1. Model and Model Transformation

Model and model transformation are two pilots of model-driven approaches, such as model-driven engineering (MDE), model-driven architecture (MDA), and model-based testing (MBT). These approaches have been adopted by various domains (e.g., enterprise engineering [35] and software engineering [68]).

Focusing on a specific viewpoint, a model provides abstractions of a system that allow people to have a better understanding of and to reason about it [911].

Models have been divided into different categories. Depending on existing forms, there are mathematical models, graph models, text models, etc. In the context of MDA [12], models are categorized as four abstraction levels: meta-meta-model, meta-model, model, and system (subject). Depending on precision, models are divided into three levels, namely: conceptual models, specification models, and implementation models [13]. Also, in MDA, a similar classification is as follows: computation-independent model (CIM), platform-independent model (PIM), and platform-specific model (PSM).

As another pilot of model-driven approaches, model transformation is a process which contains a sequence of activities operating on models. The aim of model transformation is to generate target models based on source models [1416].

Two systematic classifications of model transformation methods are given in [17, 18]. According to the content conveyed in models, model transformations can be divided into two groups: model-to-model and model-to-text.

Depending on the abstraction level of source and target models, model transformation can be divided as follows: horizontal transformation (source and target models belong to the same abstraction level, e.g., CIM and CIM) and vertical transformation (e.g., transform CIMs to PIMs).

Model transformation methods are defined by leveraging certain model transformation technologies, such as XSLT [19], VIATRA [20], QVT [21], and ATL [22]. Table 1 illustrates briefly the above four model transformation technologies.

As stated in [22], “a vast number of model transformations are being developed and organized in complex patterns.” On the one hand, some model transformation technologies provide a wide range of functions but require certain effort to learn to use them properly. On the other hand, some other transformation technologies are capable to be executed automatically; however, the precondition of applying them is strict and the transformation practices built on them have low reusability.

Besides the above model transformation technologies, some other technologies, such as Kermeta (extension of EMOF) [31], MOFM2T (model-to-text) [32], JTL (EMF-based tool) [33], Tefkat (special for MOF models) [34], and MOMENT [35] (model merging), are also being used in practice.

SbACMT focuses on conceptual model transformation. To serve to different kinds of conceptual models, SbACMT is designed to be configurable. It belongs to the horizontal classification. SbACMT aims at finding automatically the shared and same (similar) or relevant concepts (entities) between different conceptual models.

Considering the model transformation theories, SbACMT is meta-model based and employs semantic relatedness computing method as the detecting technology.

2.2. Semantic Relatedness Computing Methods

Semantic relatedness is a form of measurement that quantitatively identifies the relationship between words or concepts based on the similarity or closeness of their meaning [36].

Semantic relatedness computing methods have been widely adopted. For instance, in natural language processing field, they are used to do semantic information retrieval, keyword extraction, etc.

To develop semantic relatedness computing methods, knowledge resource such as semantic thesaurus (e.g., WordNet and Wikipedia) or corpus (e.g., classics and speech drafts) is needed. Furthermore, depending on different kinds of knowledge resources, various computing methods, such as graph based and information content based, have been proposed.

A systematically classification about the knowledge resources is given in [37]. Figure 1 shows a general idea of this classification.

Comparing with the “web-based” resources, both the “linguistically constructed” and “collaboratively constructed” resources provide more precise and credible content.

Considering the development of semantic relatedness computing methods, the linguistically constructed knowledge resources are commonly used in practice.

As stated in [38], the computing methods can be divided into three groups, namely, “graph based,” “context based,” and “temporal.” Table 2 shows a general illustration of this classification.

Many research works employ WordNet as the semantic thesaurus. However, WordNet is no longer continuously updated and maintained. As an evolved semantic thesaurus, ConceptNet is built partially upon WordNet.

Considering the precision issue, this paper employs ConceptNet as the knowledge resource and adopts a graph-based computing method to do semantic relatedness computing method to do semantic relatedness computing.

3. Foundations

3.1. SbACMT_MM

A meta-model is a model that defines modeling rules. It is built by leveraging the meta-modeling languages. Currently, three mature meta-modeling languages are as follows: MOF [21], KM3 [26], and Ecore [43].

Ecore is a lightweight version of MOF, and it is well tool supported. Comparing with KM3, Ecore is more widely used. Focusing on the transformation of UML class diagrams, we define a meta-model “SbACMT_MM” conforming to the Ecore standard.

Different kinds of conceptual models have different structures and formats. The structures and formats of conceptual models concern on the syntax aspect, which is easy to handle (comparing to the semantic aspect). The purpose of defining a meta-model is to unify structure and format, which may affect the transformation accuracy, and is to apply the semantic checking rules in a general way. The risk of adjusting syntax transformation rules is lower than adapting semantic checking rules.

As shown in Figure 2, SbACMT_MM contains three packages: class diagram, semantic relatedness, and mapping rule. The “class diagram” package shows the composition of a class diagram in model transformation context. The “semantic relatedness” package defines how to do semantic relatedness computing between a pair of concepts. The “mapping rule” package defines the mechanism of selecting concept mapping pairs based on semantic relatedness computing results.

A class diagram can be defined as a six-tuple.

3.1.1. CD=<Domain, Class, Attribute, Relationship, Type, Object>

“Domain” represents the problem domain that a class diagram built for; it defines the context of a class diagram. “Class” corresponds to a group of the same or similar objects within a domain. “Attribute” can be regarded as additional information to clearly explain the “Class.” “Relationship” stands for the relations between classes, which has four instances “generalization,” “association,” “aggregation,” and “dependency.” “Attribute” owns “Type,” which stands for a predefined set of types (e.g., integer and string). “Object” stands for the instances of “Class.” The names of “Domain,” “Class,” “Attribute,” and “Relationship” correspond to “Concept” defined in the “Semantic Relatedness” package.

In the “Semantic Relatedness” package, there are four core modeling elements defined as a four-tuple.

3.1.2. SR=<Concept, SemanticRelation, Type, CalculationRules>

“Concept” relates to the names (attribute) used in “class diagram” package. “Semantic relation” stands for a set of semantic relations (e.g., hyponym and part of) between concepts. Semantic relations are divided into two groups: direct semantic relations and iterative semantic relations. Each of the direct semantic relations has an enumeration type, whereas iterative semantic relations are combinations of the direct semantic relations. A concrete value or a specific computing rule, which is represented as “calculation rules,” is assigned to each of the semantic relations.

In the “mapping rule” package, there are two modeling elements: “MappingPair” and “Threshold.” “MappingPair” stands for all the potential mapping concept pairs, i.e., class-to-class, class-to-attribute, and attribute-to-attribute. For each kind of the mapping pairs, two thresholds (as instances of “Threshold”) are defined to distinguish potential mapping levels, i.e., acceptable, considerable, and impossible.

3.2. Semantic Relatedness Computing in SbACMT

ConceptNet is a semantic network based on information from the OMCS (Open Mind Common Sense) database. It is represented as a directed graph whose nodes are concepts and edges are common sense assertions about these concepts. The concepts represent a closely related set of natural language phrases, e.g., noun phrases, verb phrases, adjective phrases, and clauses.

Currently, the latest version of ConceptNet is 5.7, which contains 304 kinds of languages and maintains 34 kinds of semantic relations among concepts.

Figure 3 is a simple illustration of the structure of ConceptNet. As shown in Figure 3, each arrow indicates a start node and an end node of an edge. In the middle of an edge, the relationship and weight between two nodes are shown. The weight means the credibility of the assertion. Between two nodes (concepts), there may exist several edges (semantic relations).

In the SbACMT context, we select 15 (out of 34) semantic relations from ConceptNet to be used to detect the potential mappings (the same or similar concepts).

As shown in Table 3, a value is assigned to each of the relations.

The selection of these fifteen semantic relations and value assigning are completed through a three-step process. First, we invited seven researchers to vote independently and to choose the relations that they thought are appropriate for judging similar and relevant concepts. The semantic relationships that have no fewer than 4 votes are finally selected. Next, we asked the seven researchers to assign a corresponding value to each of these relations based on their experience. Finally, we carried out tests based on the gold standard “RG-65” [40] to adjust these values.

For each of the comparing concept pairs, we define two algorithms to compute the semantic relatedness. Algorithm 1 detects if a direct semantic relation exists, whereas Algorithm 2 detects if an iterative semantic relation exists. The pseudocode of the two algorithms is shown as follows.

/initialize corresponding values of semantic relations and values/
Semantic Relation_Values={(synonym, 0.95), (similarto, 0.9), (), (), …}
/Algorithm 1 detect direct semantic relations/
Function detecting direct Semantic Relation (String A, String B){
int SR_V=0;
for (int i=0; i<Semantic Relation_Values.size (); i++){
  if (Semantic Relation_Value.get (i).get Relation (Concept Net (A, B))){
   if (Semantic Relation_Value.get (i).get Value ()>SR_V)
    SR_V=Semantic Relation_Value.get(i).get Value();
  }
} return SR_V;
}
/Algorithm 2 detect iterative semantic relations/
Function detecting iterative Semantic Relation (String A, String B){
int temp1, temp2, SR_V=0;
string [] inter Strings=Concept Net (A);
for (int i=0; i<inter Strings.size (); i++){
  for(j=0; j<Semantic Relation_Values.size (); j++){
if (Semantic Relation_Value.get (j).get Relation (inter Strings.get (i), B)){
   temp1=Semantic Relation_Value.get (j).get Value ();
   temp2=Semantic Relation_Value.get (i).get Value();
    if (temp1∗temp2> SR_V)
     SR_V=temp1∗temp2;
   }
  }
} return SR_V;
}

If a comparing concept pair cannot be located, a lexical analysis function will be executed first. This function aims to transform the input concept (word) to its origin form, and it deals with two main situations: words with delimiters and anagram (e.g., plural and tense).

“SR_V” is short for semantic relatedness value, which is computed between a pair of comparing concepts (string A and string B).

4. Selecting Potential Mapping Pairs

“SR_V” ranges from 0 to 1. Thresholds on “SR_V” are defined to distinguish potential mapping pairs. Table 4 shows three sets of these thresholds.

The thresholds are adjusted and assigned through a similar process as the one of the selecting semantic relations from ConceptNet.

For a comparing concept pair, if their SR_V is bigger than “-t1,” we assume the two concepts can be matched directly (acceptable). If the SR_V is between “-t1” and “-t2,” the users or domain experts shall be asked to make the matching judgment (considerable). If the SR_V is less than “-t2,” we assume no mapping shall be built. All the considerable mapping pairs will be listed out as potential mapping. Without manual judgments, the potential mappings will be used as transformation rules.

4.1. Matching Process

In SbACMT, model transformation process is divided into two phases: building potential mappings and executing transformations.

Figure 4 illustrates briefly the transformation process. After taking in two models, SbACMT extracts and classifies two groups of “Class” and “Attribute.” The first phase “building potential mappings” contains three main steps: building matching pairs among classes, building matching pairs between classes and attributes, and building matching pairs among attributes. The three steps are executed in sequence. After getting the potential mappings, a manual verification step can be added optionally. Then, the transformation phase, which concerns on those mappings, is to be executed.

4.2. Class-to-Class Mapping Step

The name of a class is regarded as a specific domain concept. A class may own a set of attributes. Classes are connected with each other by relations, which can be regarded as class’s attributes. While comparing two classes, their names are taken into account.

Equation (1) is defined to compute the semantic relatedness between two classes:

In equation (1), stands for the SR_V between Classi (Ci) and Classj (Cj), whereas Classi comes from the source conceptual model and Classj comes from the target conceptual model. C2C_SRV(i, j) equals to SR_V(Ci, Cj), the SR_V between two class names. All the values (calculated automatically) will be compared with thresholds “c2c-t1” and “c2c-t2” to generate potential mappings.

4.3. Class-to-Attribute Mapping Step

After executing the first matching step, some classes may be left unmatched. This step aims to build potential cross-level (granularity) mappings between classes and attributes. Equation (2) is used:

Take one class (or attribute) from the source conceptual model and one attribute (or class) from the target conceptual model, and then apply semantic relatedness computing algorithm between their names. stands for the cross-level comparing result between a specific pair of Concept Ci (or Cj) and attribute Aj (or Ai).

4.4. Attribute-to-Attribute Mapping Step

After executing the two former matching steps, some classes and attributes may still be left unmatched. For the unmatched classes, we regard them as specific parts that cannot be transformed or merged. For the unmatched attributes, equation (3) is used to build potential mappings among them:

The potential attribute matching pairs are only built between those attributes that belong to the potential matched classes. In equation (3), assuming Ai is an attribute of Classn (Cn) and Aj is an attribute of Classm (Cm), the matching algorithm only needs to be executed if there is a potential mapping between Cn and Cm.

4.5. Verification of Potential Mappings

To generate transformation rules, the potential mappings have to be verified. In SbACMT, we propose two ways “verification based on relations” and “manual verification.”

The relations defined in the source conceptual models (class diagrams) are used to verify the automatically generated potential mappings:(1)For the relation “generalization/specialization (is a or is a kind of)”, if the father class is matched then all its children classes should be matched to the same class (or its children classes) in the target class diagram.(2)For the “aggregation/composition (has a or part of)” relation, the classes owning this relation can be matched to the same class in the target class diagram.(3)For the “dependency” and “association” relations, we regard them as “user-defined” type. They are not used as verification measurements.

For instance, if in the source class diagram, two classes Ci and Cn have a generalization relation, and a potential mapping is built between Ci and Cj (from the target class diagram), we will verify that if a mapping between Cn and Cj is automatically generated. This verification step is executed automatically.

For the “manual verification,” a visual interface is provided to the users (or experts) to show the automatically generated mappings. These mappings are divided as two groups: “acceptable” and “considerable.” If there is no manual judgment, all these mappings will be used directly as transformation rules.

5. Use Case

5.1. Use Case Illustration

We evaluate the performance of SbACMT with the following use case. In the use case, two conceptual models (UML class diagrams) “vehicle management” and “traffic management” conforming to Ecore standard are involved. Figure 5 shows the two class diagrams.

The class diagram “vehicle management” contains three classes “person” (with seven attributes), “driving license” (with six attributes) and “vehicle” (with five attributes). The (association) relations between them are as follows: “person” possesses “vehicle” and “driving license.”

The class diagram of “traffic management” contains three classes “automobile” (with four attributes), “ticket” (with five attributes), and “driver” (with four attributes). The relationships are as follows: “ticket” concerns “automobile” and “driver” pays for the “ticket” and drives “automobile.”

SbACMT takes the two class diagrams as inputs and generates the potential mappings (pairs) between them automatically.

5.2. Transforming Process

The first matching step focuses on classes. A compound word (the class name) will be split into single words and then compared with potential target, respectively. For instance “driving license” is split as “driving” and “license”.

Table 5 shows the computing results. All three classes “person,” “driving license,” and “vehicle” are mapped with classes of the target class diagram. Since no classes left unmatched, the second matching step is skipped.

The third matching step focuses on attributes. In this step, all the attributes belonging to the already matched classes are taken into consideration. Tables 69 show the computing results.

6. Evaluation

The evaluation of SbACMT contains two dimensions: (1) focusing on the semantic relatedness computing method (parameters and constraints) and (2) focusing on the correctness of the automatically generated mappings.

6.1. Adjusting SRCM-CNet with Gold Standard RG-65

We use the “RG-65 gold standard” data set to test SRCM-CNet. The testing results are measured with Spearman’s rank correlation coefficient.

As stated in [36], a semantic relatedness computing method is considered reliable if an inter-rater agreement (computing results) of it is over 75%.

Considering Spearman’s rank correlation coefficient, we compare SRCM-CNet with three other semantic relatedness computing methods. Table 10 shows the testing results. SRCM-CNet gets Spearman’s rank correlation coefficient “0.9,” which is better than the other methods.

6.2. Correctness Evaluation

We evaluate the correctness of SbACMT on basis of the use case presented above.

The quantitative evaluation focuses on class-class and attribute-attribute mappings. The retrieved mappings and expected mappings are shown in Tables 11 and 12, respectively.

As shown in Table 12, according to the computing results, the attributes “id” and “name” (both from class “person”) match with the attribute “id” (from class “driver”). However, we only build the mapping between attributes “id” and “id,” since they have a higher SR_V.

Inspired by the work stated in [45], performance measures are based on three indicators: precision, recall, and F-measure:(1)The precision (P) evaluates the quality of the automatic built transformation mappings.(2)The recall (R) evaluates the sufficiency of the automatic built mappings.(3)The F-measure (F) is the weighted harmonic average of P and R. A higher value of F is preferred.

Equation (4) shows the computing rules of “precision,” “recall,” and “F-measure”:

The evaluation results of SbACMT are shown in Figure 6.

The performance evaluation has been performed on a computer with 2.4 GHz, i5-6200U CPU, and 8 G RAM (with Windows 10 OS and Java 9 JDK).

Generally, the precision rate and the recall rate affect each other. Considering the attribute mapping pairs, the quality of the automatic built mappings is high. The precision of automatically built class mappings is mediocre (the number of classes involved in the use case is small).

This section presents the related works in three aspects: the usage of meta-models in practice, semantic relatedness computing methods adopted in practice, and our previous relevant work [47].

7.1. Meta-Model Usage

Meta-models, which are a special kind of models, define the rules of building models and can be used to verify the correctness of the models conforming to them.

Meta-models have been defined in various domains. In [48], a meta-model is defined to adopt organizational concepts to help analyze and design multiagent systems. In [49], a meta-model, which is made of four smaller meta-models, is defined to formalize software protection from man-at-the-end attacks. In [50], an accident process meta-model is defined to describe the hazard-accident process by fault propagation. In [51], a meta-model is defined to argue about compliance of requirements models and to keep track of compliance decision through the requirement modeling. In [52], a meta-model is defined to help effectively build interenterprise collaborations.

The majority of current existing meta-models are defined without strict semantics or syntax. They are presented as class-based graphs or activity diagrams, and natural language is used to describe the context and the constraints, etc. Researchers and engineers focus on the content conveyed by meta-models more than meta-modeling theories and languages, which leads to the low reusability and poor readability of these meta-models. Table 13 presents the comparing results of the above meta-models.

We define SbACMT_MM conforming to the Ecore standard, which ensures the unified semantics and syntax. In the “semantic-based conceptual model transformation” context, SbACMT_MM sets the theoretical foundation.

7.2. Semantic-Based Matching Methods

Employing semantic-based analysis to determine similarity between concepts (entities) is a common research topic in various fields, such as ontology matching and data schemas matching. A certain number of semantic relatedness computing theories and methods are proposed.

In [53], semantic analysis is used to lift meta-models into ontologies to integrate eventually modeling languages. In [54], towards data integration, an automatic schema matching method built on probabilistic lexical annotation technique is proposed. In [55], a new semantic matching method is adopted to find the corresponding entities via similarity matrices. In [56], a two-step method built on state-of-the-art methods is proposed to improve ontology matching by enriching ontology mappings.

Table 14 presents the above research works and shows a comparison between them.

To support general-purpose usage, a certain number of semantic reasoners are proposed and are used in practice. In Table 15, five semantic reasoners are listed and analyzed.

Agreement maker uses different matching algorithms to aggregate their matching results, which is usually called parallel combination.

STROMA uses linguistic, structural techniques to deal with ontology matching. It adopts five matching strategies and weights the results. The system is generic as it can be used for different domains.

ASMOV utilizes lexicographic, structural, and instance-based techniques and WordNet. It incorporates a semantic validation process, which increases the accuracy of the system.

S-Match represents the ontology entity as a logical formula; the matching problem is reduced to the proposition validity problem. The input of the S-Match system is two graph structure patterns. The output is the logical relation between the nodes in the graph.

Compared with existing computing methods and semantic reasoners, we adapted ConceptNet as semantic knowledge base (with rich semantic relations and assertions). Furthermore, focusing on conceptual model (class diagrams) transformation, we tailor the semantic relations defined in ConceptNet and quantify the adopted relations with the RG-65 gold standard. The rules of applying SRCM-CNet are defined in SbACMT_MM.

7.3. Relevant Previous Works

Compared with our previous work [47], the main novelty of this paper is reflected in three aspects.

First, it is the new meta-model “SbACMT_MM”. Towards conceptual model transformation, we define this meta-model by employing the Ecore standard. SbACMT_MM consists of three parts “class diagram,” “semantic relatedness,” and “mapping pair”; it builds the foundation of combing SRCM-CNet into automatic model transformation process.

Second, it is the adopted semantic knowledge base. In our previous work, WordNet is adapted as the basic semantic resources. In this paper, we adapted ConceptNet as the semantic knowledge base and built a local semantic thesaurus (with Neo4J) based on it. Comparing to WordNet, ConceptNet contains more content (e.g., words and semantic relations), being constantly updated and supports multilanguages.

Third, it is the evaluation part. In previous work, we evaluated the performance of the automatic generated transformation mappings only based on manual judgments. In this paper, we adopted the RG-65gold standard and Spearman’s rank correlation coefficient, which is more persuasive.

8. Conclusion

There are many kinds of conceptual models defined and used in diverse domains. Reusing and merging existing conceptual models to generate new conceptual models are always required in practice.

To accelerate the building process of conceptual models, this paper proposes a general configurable (semi-) automatic conceptual model (class diagrams) transformation method. A semantic relatedness computing method, which is mainly built on the structured knowledge base ConceptNet, is employed as the potential mapping-detecting technology.

SbACMT follows the meta-model based model transform theory. A specific meta-model “SbACMT-MM” defines the rules of applying SRCM-Net to generate mappings between conceptual models. Considering the granularity of conceptual models, we also propose an integrated mapping and transformation process in which three matching steps are divided. User effort can be optionally involved in this process at certain points, such as judging the intermediate potential mapping pairs.

SbACMT is configurable in two aspects. First, by adapting the “class diagram” part of SbACMT_MM, SbACMT can support other kinds (besides class diagrams) of conceptual model transformations. Second, by adapting the thresholds assigned in Table 4, SbACMT can serve for customized transformation. Moreover, the values assigned to semantic relationships can also be adjusted by users.

SbACMT aims to serve to transformations of different kinds of conceptual models. It can accelerate the process of building new conceptual models in general.

SbACMT is well tool supported. As shown with the evaluation of the use case, SbACMT can achieve automatic conceptual model transformation with effective performance. A potential research direction of SbACMT concerns about contextual information. While calculating the semantic relatedness between concept pairs, the contextual information shall be taken into account. Two possible solutions are as follows: making use of the assertions (on the edges) of ConceptNet and combining other semantic resources together with ConceptNet.

Data Availability

The data set RG-65 was used to support the findings of this study.

Conflicts of Interest

The authors declare that they have no conflicts of interest.

Acknowledgments

This work was supported by the “Fundamental Research Funds for the Central Universities (no. 3082018NS2018057).”