Automatic knowledge exchange between ontologies and semantic graphs

Abstract

This article presents an innovative knowledge management approach for the validation and transfer of knowledge between semantic networks or ontologies and a target ontology represented in web ontology language (OWL) format. This process has been designed for addressing the quality improvement of ontologies automatically created by learning techniques. The knowledge transfer process is a semi-automatic computer aided method to assist the domain expert to improve the target ontology. To validate our proposal, we have used an automatically generated target ontology. We used knowledge transfer from the well-known Babelnet semantic graph and a manually generated ontology to improve the quality of the target ontology. Finally, to show the suitability of our proposal in the ontology-fixing process, we compare the improved target ontology resulting from the application of the proposed validation and knowledge transfer techniques with its original version. We developed an example of our proposal in our OntologyFixer tool, which is available on a GitHub repository (https://github.com/gabyluna/OntologyFixer/tree/ontofixer_v2).

Keywords

Automatic knowledge exchange ontology ontology fixing ontology quality semantic graph

1. Introduction and motivation

Ontologies enable computers to structure and manage information according to semantic criteria to perform computational inference processes. Therefore, there is great interest in building new ontologies that provide semantic representations of different areas of knowledge and eventually make it possible to solve new problems. To this end, in recent years a large number of ontologies have been developed and made available to the scientific community through different repositories. As examples, we highlight NCBO BioPortal,¹ which allows downloading a large number of biomedical ontologies; LOV4IoT,² which shares more than 800 ontologies that model different aspects of Internet of Things (IoT); the Finnish Ontology Library Service ONKI,³ Swoogle⁴ or the Protégé ontology library.⁵ This fact proves that there is a great interest in the creation of ontologies to accurately model the semantics associated with specific problems.

Gruber defines an ontology as ‘A formal explicit specification of a shared conceptualization for a domain of interest’ [1]. Gruber’s definition has been widely adopted in the web semantics community due to its simplicity and comprehensiveness. The formal and explicit specification of an ontology enables large-scale and systematic knowledge management, as well as computer-based processing and reasoning. In 2004, the World Wide Web Consortium (W3C) proposed the Web Ontology Language (OWL)⁶ as the key to the Semantic Web technologies. The OWL standard is a core component of the Semantic Web Protocol Stack for formal knowledge representation and reasoning. It is supported by a wide variety of open tools and frameworks and therefore, it was adopted as the reference ontology representation in our study. In detail, we used OWL DL dialect and took advantage of the knowledge represented and inferred in the ontology, expressed through individuals, classes, object properties and datatype properties defined in OWL DL. OWL DL is also the representation used by the ontology learning tools that we used in this study.

Ontology construction is a cumbersome process that involves domain experts and knowledge engineers. Hence, ontology learning techniques have been introduced to enable the automatic construction of ontologies from documents written in natural language [2,3]. Combining these techniques with texts obtained from reliable sources such as scientific articles, we can obtain ontologies more easily than building them manually from scratch. However, current ontology learning tools have many limitations and the ontologies created by this method have to be fixed at a later stage. In particular, we highlight the identification of unhelpful terms that introduce junk knowledge in the ontology or the inexistence of terms in source documents (texts) that should be present (represented as nodes) in the target ontology, resulting in the lack of specific knowledge representation in the domain. Our contribution specially addresses the later issue. Several recent works have highlighted this problem and have introduced methodologies and tools that help to partially automate the ontology-fixing process [4,5].

The main drawback of the ontology-fixing methodologies is that the process is mainly driven by quality assessments, consistency errors and design defects. Therefore, there is no validation from a semantic point of view of the correctness and completeness of the knowledge contained in the ontology. However, some third-party ontologies may have some knowledge that could be used to perform some validations of the represented knowledge. We could leverage ontology matching/alignment processes [6,7] to identify relevant knowledge from third-party ontologies and use it to fix an ontology. In addition, using similar schemes, relevant knowledge can be extracted from semantic graphs (ontology dictionaries) enabling them as a very useful tool for ontology quality validation and improvement processes. Our proposal includes a simple ontology matching scheme and shows how to take advantage of relevant knowledge discovered in different sources of information to fix semi-automatically generated ontologies. The fixing mechanism has been tested on an ontology used for experimentation in previous ontology-fixing studies. Our experiments show the usefulness of the proposal and its possible application for semi-automatic construction of ontologies.

The remainder of this study is structured as follows: Section 2 summarises previous efforts and proposals to improve ontology-fixing processes. Section 3 presents our proposal to leverage third-party knowledge sources for ontology fixing. Section 4 performs an evaluation of our proposal and its possible application for fixing the ontologies. Finally, Section 5 shows the main conclusions and outlines directions for future work.

2. Background

To correct, improve and enrich the content of ontologies created automatically or semi-automatically with the help of different tools, researchers have proposed over the years different methods and strategies to fix those errors that may occur in different aspects of ontologies such as its structure, functionality, maintainability, semantics, etc. We briefly describe below some studies that address (1) improving the quality of the ontologies, (3) strategies for identifying possible design defects and (3) tools/methodologies for fixing the ontologies.

Bandeira et al. [8] introduced FOCA, a methodology for evaluating ontologies. It is a three-step method that comprises the identification of ontology type, the application of a Goal/Question/Metric approach and the quality assessment. Thus, FOCA implements a role-based calculation of ontology quality according to its type; it includes a questionnaire to accomplish the evaluation and a statistical model that automatically computes the quality of the ontologies.

The OQuaRE [9] framework⁷ is a method to evaluate the quality of ontologies that adapt the SQuaRE standard (originally designed to assess the quality of software products) to the context of ontologies. OQuaRE uses different metrics to assess the quality of the ontologies with regard to different dimensions, including reliability, operability, maintainability, compatibility, transferability, and functional adequacy.

FOEval [10] introduces an evaluation model to choose the ontology that best fits the user requirements. The model allows users to select indicators and assign weights for each one selected among a wide variety of available quality indexes.

OntoClean [11] is a methodology for the validation of the ontological adequacy and logical consistency of taxonomic relations. It uses various concepts of philosophy, such as rigidity, identity criteria and unity, to provide modelling guidelines.

OntoQA [12] is an approach that analyses ontology schemas and their instances (such as knowledge bases) and describes them using a well-defined set of metrics [13] to evaluate the content of the ontologies. One of the features that highlight OntoQA is its flexible technique for classifying ontologies based on their content and relevance to a set of keywords, as well as user preferences.

Su and Gulla [14] addressed semantic problems of ontologies and introduced a method to find semantic correspondences between ontology components to provide a semi-automatic mapping method and a prototype mapping system to support the ontology mapping process to improve semantic interoperability in heterogeneous systems.

Another study [15] introduced a formal technique to control ontology evolution by domain experts and shows two cases of ontology evolution based on ontology templates that guarantees an ontology without inconsistencies.

OOPS! [16] is a web application that provides mechanisms for the automatic detection of potential errors, called pitfalls, to help developers during the validation process. Each pitfall provides the following information: title, description, elements affected and importance level to give the user enough information to identify problems and take appropriate actions.

Nicolas Troquard et al. [17] introduced an ontology repair methodology that preserves most of the original knowledge without additional costs in terms of computational complexity. On the contrary, researchers quickly realised that bringing the concept of smells/pitfalls to the domain of knowledge engineering could help in the process of improving the quality of ontologies. The work of Póveda et al. [18] introduces and classifies a catalogue of 24 pitfalls that usually appear during the process of building ontologies. The classification of pitfalls is based on several works prepared by different authors who have identified common mistakes during the process of modelling ontologies.

We also introduced a methodology to evaluate the quality of the ontologies (quantitatively and graphically) and to correct ontology inconsistencies by minimising design defects [4]. The proposed methodology is based on the Deming cycle and is supported by quality standards that have proven to be effective in the software engineering domain and that have a high potential to be extended to the quality management of knowledge engineering.

With this in mind, we find that the methodologies and tools developed so far could benefit from ontology mismatch corrections by transferring knowledge from knowledge bases such as third-party ontologies designed by experts and ontological dictionaries strongly reinforced and fed from different sources of knowledge. To achieve this goal, we use an ontology alignment/matching scheme. Ontology matching is the process of discovering correspondences between elements (classes, objects, etc.) included in two or more ontologies. As can be observed in some recent literature reviews [6,7,19], ontology matching has been widely studied. Some recent works have introduced complex ontology alignment systems based on BERT [20,21] and neural networks [22].

Ontology alignment schemes have been successfully used to address different tasks such as ontology merging, query answering, data translation or semantic web navigation. We have also found that ontology alignment schemes could be successfully used to address ontology-fixing tasks. Next section introduces our proposal for addressing knowledge transfer and leveraging knowledge exchanged to improve and fix ontologies.

3. Semantic transfer process

In this section, we describe our strategy to improve and validate the semantics of automatically generated ontologies that may have some conceptual and relational shortcomings. The goal of our proposal is to validate and enhance the knowledge-base concepts (classes and individuals) using third-party ontologies or ontological dictionaries (such as Babelnet⁸ or Wordnet⁹). Figure 1 graphically represents the semantic transferring strategy introduced in this work.

Figure 1.

Semantic transferring process.

As shown in Figure 1, the strategy comprises four stages: (1) search for each class or individual concept presence in third-party ontologies or semantic dictionaries; (2) compare the relations between concepts in the knowledge bases (ontologies or dictionaries) with those of the target ontology; (3) when both relations match, proceed to transfer or adapt this knowledge to the target ontology; and (4) validate that the new ontology has improved the quality of the original ontology after the applied changes.

Stages 1 (concept comparison) and 2 (relations comparison) correspond to a two-step ontology matching strategy that identifies relevant knowledge for ontology fixing. The first stage comprises a string-based element-level matcher that identifies possible alignments between nodes of ontologies. Moreover, the second stage implements a taxonomy-based structure-level matcher that is able to refine the results achieved in the first stage. The ontology matching process was designed to be simple and to allow the use of third-party ontologies and semantic graphs.

During the first stage, an iterative process indicated in Figure 1 is carried out with each of the nodes (classes or individuals) of the ontology using the OWL API.¹⁰ The name of the class or individual is used for running a search in other ontologies or ontological dictionaries to find knowledge that can be used for improving the target ontology. In detail, the set of ontology nodes that are initially usable can be computed as shown in Equation (1)

RN = {tpon ϵ TPO, odn ϵ OD, ton ϵ TO | \begin{matrix} ld (name (tpon), name (ton)) < LMD \\ \lor ld (name (odn), name (ton) < LMD) \end{matrix}}

(1)

where tpon are nodes included in third-party ontologies (TPO), odn are nodes included in ontological dictionaries (OD), ton are nodes belonging to the target ontology (TO), name(x) refers to the name attribute of a node, ld(a,b) function computes the Levenshtein distance between the strings a and b and LMD (Levenshtein maximum distance) is the maximum distance allowed for a positive comparation. Figure 2 shows the concept_comparison method that compares the concepts of the knowledge sources with those of the target ontology.

Figure 2.

Algorithm designed to execute the concept comparison (first stage).

The algorithm designed to execute the concept comparison searches for nodes that belong to the target ontology in the third-party ontologies using the (1) concept_comparison_ontology function and in the ontological dictionaries using the (2) concept_comparison_dict function. The first one compares each node name of the target ontology with the node names of third-party ontologies (using OWL API), while the second one searches semantic graphs (using specific APIs provided by each specific representation) for relations involving these nodes. To avoid the effect of misspellings, we use Levenshtein distance [23] to retrieve lexically close nodes. The algorithm only returns those nodes with a Levenshtein distance less or equal to a certain value (see line 01 in Figure 2, which refers to LMD parameter). For example, if we consider the node named ‘action’ in the target ontology, the node named ‘actions’ in a knowledge source will be returned for the next stage because the Levenshtein distance between both node names is less or equal to the value set for LMD (LMD = 1).

Moreover, we also search for the name of each node of the target ontology in external ontological dictionaries. For this, the search will be limited to the grammatical use of the word (node name) as a noun (POS, part of speech). The result of the search will return a list of synsets. However, due to the polysemy of natural language, some synsets may not match the semantic meaning of the target ontology node. The result of executing the first stage is a collection of records that could be used to improve the ontology which consist of third-party ontology nodes and synsets gathered from ontological dictionaries.

During the second stage (relations comparison/disambiguation), we identify and discard records containing invalid (out of scope) knowledge that were collected in the first stage. We consider a record to have invalid knowledge when it fits lexically with the original ontology node but they are semantically distinct, that is, they have different meanings. Therefore, the identification of invalid knowledge involves the application of techniques inspired by word sense disambiguation (WSD) in texts [24]. However, unlike WSD, the identification of invalid knowledge does not involve the analysis of text fragments. Instead, the algorithm must compare the group of nodes connected to the original concept in the target ontology with the group of nodes connected to the same concept in the third-party ontology (or the ontological dictionary) used as a source of knowledge that is connected to the retrieved record for the concept. If there is no correspondence between these two groups of nodes, then the knowledge included in the retrieved record can be ignored. Therefore, an rn node included in the set of RN retrieved nodes is valid when it fulfils the condition shown in Equation (2)

is_valid (rn) \to | connected (rn) \cap connected (ton) | \geq MT

(2)

where connected(x) is the set of nodes connected with x. Exceptionally, when the target ontology has a very specific subject, the disambiguation process can be relaxed and check only if the nodes connected to each record retrieved in the source of knowledge (semantic graph/ontology) are present in the target ontology (regardless of how these nodes are connected in the target ontology). In such situation, a retrieved node is valid when it fulfils the condition included in Equation (3)

is_valid (rn) \to | connected (rn) \cap nodes (ton) | \geq MT

(3)

where nodes(ton) is the set of nodes included in the target ontology. The algorithm implemented to perform these two comparisons is shown in Figure 3.

Figure 3.

Identification of valid records during relations comparison stage.

The function is_valid (Figure 3) determines whether a record retrieved during the previous stage contains valid knowledge. The function computes the ratio of relations of the original node to neighbour nodes in the target ontology that are also present in the retrieved record for the knowledge source. To do so, we use the get_nodenames_connected_with(Node n, Ontology|SemanticGraph g) function that returns a set of names with the nodes connected to node n in the ontology or semantic graph g and get_ontology_node_set(Ontology o) function that computes the set of nodes included in an ontology o. In case of using misspelt ontologies, the intersection function (line 13) can be replaced by specific functions using string similarity functions (such as Levenshtein distance). The condition to relax the semantic disambiguation process (as explained in the previous paragraph) is modelled using the VERY_SPECIFIC_TARGET_ONTOLOGY variable (see lines 02 and 15). Finally, a matching threshold parameter (MT) has been introduced to determine the minimum semantic match (lines 01 and 22).

The third stage uses the valid knowledge included in the records retrieved in the second stage to improve the structure and the semantic consistency of the target ontology. We exploit the knowledge extracted from the connections between each retrieved record and other nodes in the same knowledge source. In the case of semantic graphs that include a wide range of relations between nodes (e.g. meronyms and holonyms), only subclasses, super classes and instances of this type of knowledge source are transferred. Then, the integration of knowledge allows us to verify that the relations retrieved for each specific node fully match those included in the target ontology. We have to transfer the knowledge about relations that do not exist in the target ontology, as well as transfer connected nodes if they do not exist, and fix the existing relations in the case they do not have the same type/direction. When a knowledge transfer is possible, that is, the relations retrieved do not fully match with the target ontology, we can also detect a symptom of concept mismatch that can be corrected using quick-fix strategies [5]. In addition, we can transfer the names of the nodes in case they do not fully match those of the knowledge source (eventually using a quick-fix strategy). By including this knowledge in the ontology, we are greatly modifying its structure and content, so its quality may be compromised or improved. To determine whether the changes can be applied, the ontology must be validated by an expert user (using quick fix) or according to the ontology quality validation process described in the next (fourth) stage.

In the fourth stage, a validation of the ontology quality is performed using the OntologyFixer web tool introduced recently [5]. This web tool includes several modules for fixing errors and ontology quality assessment using quality metrics based on the OQuaRE framework. Correctness and completeness metrics often require more intensive human intervention. To reduce user effort and intervention, our evaluation is performed by calculating nine content/structure metrics: RROnto (Relationship Richness), INROnto (Relations per class), ANOnto (Annotation Richness), CROnto (Class Richness), NOMOnto (Number of Properties), RFCOnto (Response for a Class), CBOOnto (Coupling Between Objects), LCOMOnto (Lack of Cohesion in Methods) and RCOnto (Class Richness). These metrics can be calculated automatically and allow the evaluation of different aspects of ontologies, such as their structure, functional adequacy, compatibility, transferability, operability and maintainability.

Table 1 provides a brief description of each of the metrics used to evaluate the quality of ontologies. All of the selected metrics are available in the OQuaRE framework [25] except for RCOnto, which is provided by the OntoQA framework [12].

Table 1.

Description of quality metrics used.

Metric	Description	Equation
RROnto	Relationship richness	$\frac{\sum^{\| P_{Ci} \|}}{\sum^{(\| R_{Ci} \| + \| P_{Ci} \|)}}$ where P_Ci and R_Ci represent the number of properties and relations of the ith class, respectively, and \|X\| represents the cardinality of the set X
INROnto	Number of relations per class	$\frac{\sum^{\| R_{Ci} \|}}{\| C \|}$ where C is the set of classes of the ontology
ANOnto	Number of annotation properties per class	$\frac{\sum^{\| A_{Ci} \|}}{\| C \|}$ where A_Ci represents the number of annotations per class
CROnto	Number of instances per class	$\frac{\sum^{\| I_{Ci} \|}}{\| C \|}$ where I_Ci represents the set of individuals of the C_i class
NOMOnto	Number of properties per class	$\frac{\sum^{\| P_{Ci} \|}}{\| C \|}$
RFCOnto	Number of properties that can be accessed directly from the class	$\frac{\sum^{\| P_{Ci} \|} + \sum^{\| Su p_{Ci} \|}}{\| C \|}$ where Sup_Ci represents the number of super classes of C_i
CBOOnto	Number of related classes	$\frac{\sum^{\| Su p_{Ci} \|}}{\| C \| - \| R_{Thing} \|}$ where R_Thing is the number of relations of class Thing
LCOMOnto	Lack of cohesion in methods)	$\frac{\sum path (\| C {(leaf)}_{i} \|)}{m}$ where path(\|C(leaf)_i\|) is the length of the path starting from the leaf class c_i to the Thing class and m is the total number of paths in the ontology
RCOnto	Ratio of classes that have instances	$\frac{\| C^{'} \|}{\| C \|}$ where C′ stands for the number of classes that have instances

Metrics are mapped to the range [1, 5] following the scale of values included in Table 2.

Table 2.

Metrics scaling criteria.

Metric/value	1	2	3	4	5
RROnto	[0–20]%	(20–40]%	(40–60]%	(60–80]%	>80%
INROnto	[0–20]%	(20–40]%	(40–60]%	(60–80]%	>80%
ANOnto	[0–20]%	(20–40]%	(40–60]%	(60–80]%	>80%
CROnto	[0–20]%	(20–40]%	(40–60]%	(60–80]%	>80%
NOMOnto	>8	(6–8]	(4,6]	(2,4]	≤2
RFCOnto	>12	(8–12]	(6–8]	(3–6]	[1–3]
CBOOnto	>8	(6–8]	(4–6]	(2–4]	[1–2]
LCOMOnto	>8	(6–8]	(4–6]	(2–4]	≤2
RCOnto	[0–20]%	(20–40]%	(40–60]%	(60–80]%	>80%

The above metrics are plotted in a radar chart where the vertices of the polygon represent each of the measures and the total area is an indicator of overall quality. The value of the area must be compared with that of the original ontology to evaluate the improvement of the ontology quality and, therefore, to support the recommendation of whether or not to apply the changes. This validation can eventually be replaced or complemented by the assistance of an expert user.

The next section includes several examples of application of the automatic knowledge exchange methodology introduced. Although the experiments presented in the next section show the possibility and benefits resulting from an automatic process of knowledge exchange, our proposal is not intended to be a fully automatic ontology-fixing/quality improvement tool. It aims to support and assist the expert and/or knowledge engineer to review and improve the quality of ontologies, with special benefits on ontologies generated by ontology learning tools.

4. A case study

This section provides a detailed description of the experiments performed to demonstrate the utility of the introduced methodology. First subsection describes the target ontology used to evaluate the designed quality improvement strategy. Second subsection presents the results obtained from applying the proposed semantic transfer strategy to the target ontology. Finally, third subsection shows a list of lessons learned from the process of comparing and transferring semantic knowledge to the ontology.

4.1. Experimental data

In the previous study [4], an ontology learning process was carried out to build an ontology from 62 specific articles in the domain of multi-objective optimisation. To this end, different tools were combined to implement the whole process (i.e. information extraction and automatic ontology generation). This ontology (which was automatically generated and contains a broad range of errors, pitfalls and design defects) was shared in Zenodo platform, to be used to run experimental evaluations of quality metrics and ontology-fixing processes [26]. This ontology was successfully used for experimental purposes in the above-mentioned study.

Although there are several repositories from which it is possible to download a large number of ontologies with an acceptable quality for different purposes, to the best of our knowledge, there are no other examples of ontologies that allow validating strategies for fixing and/or assessing the quality of ontologies. Therefore, considering the nature of this work, we found that the previously mentioned ontology will be an excellent case study for the evaluation of our proposal.

4.2. Experimental results

In this subsection, we present some issues detected using the strategy outlined in the previous section with Babelnet and an ontology created by experts [27,28] as sources of knowledge.

The configuration parameters were experimentally optimised to allow the knowledge transferring from a list of nodes (‘optimization’, ‘aggregation’, ‘connection’, ‘decision making’ of Babelnet and ‘preference model’, ‘desirability function’, ‘researcher’, ‘moo’, ‘metaheuristic’, ‘optimisation’, ‘PreferenceIntegration’, ‘PreferenceInformationFromDm’ of the third-party ontology). To obtain those nodes that are orthographically similar to the target ontology and that can be used to extract knowledge, we set LMD = 1 for the execution of the experiments. To select this value, we compared the ‘preference model’ node of the target ontology with the ‘PreferenceModel’ node of the third-party ontology during the experimental run. Those nodes matched with LMD = 1, so we take this configuration value as valid for the first stage. On the contrary, the MT parameter (second stage) was set to 0.15. This is a suitable value to validate that the records achieved in the first stage match nodes of the target ontology. Finally, since the target ontology only incorporates semantic information about multi-objective evolutionary algorithms (MOEA), the value of VERY_SPECIFIC_TARGET_ONTOLOGY variable is set to true to relax the conditions for the disambiguation process (stage two).

As an example, we have chosen the class ‘optimization’ to analyse the knowledge transfer process from the Babelnet semantic graph. When comparing with the ontological dictionary, all the synsets related to the selected class must be returned. In this case, six synsets of type NOUN were returned by the Babelnet API (bn:10424683n, bn:01564896n, bn:02021921n, bn:00059223n, bn:00059223n, bn:03309733n). If more than one synset is obtained as a result, the strategy proposes to carry out a disambiguation process to select the best synset related to the sense of the target ontology. In Table 3, we can see the results of the selection process.

Table 3.

Results achieved after executing the concept comparison (first) and relation comparison (second) stages.

Id synset	Senses	Frequency	Node with coincidence	Ratio of relations shared with original node
bn:00059223n	Optimisation	0
bn:03309733n	Mathematical optimisation, mathematical programming, optimisation	0
bn:01564896n	Programme optimisation, programme optimisation, software optimisation	2	Programme	0.2
bn:05841544n	NA	0
bn:02021921n	NA	0
bn:10424683n	NA	0

NA: not available.

In the second stage of the introduced methodology, we compare the relations of the retrieved senses (synsets) with the original node named ‘optimization’. In particular, we count how many nodes that are connected with the retrieved records are also connected to this node in the target ontology. Finally, we select only the synset bn:01564896n as valid since it has the highest match, whose value is greater than MT. Once the synset bn:01564896n is marked as valid, we retrieve all of its relations using Babelnet API and validate if they exist in the target ontology. Table 4 shows the list of relations found.

Table 4.

Relations retrieved from Babelnet semantic graph for knowledge exchange.

Relation Babelnet	Value
is-a	Optimisation, mathematical optimisation, process
has-kind	Glowworm swarm optimisation, interprocedural optimisation, entry form optimisation, partition alignment, compiler optimisation
has-instance	loop optimisation, loop unrolling

Several relations listed in Table 4 were not found in the target ontology, so they were added to it and subsequently checked whether the application of these changes improve the quality of the ontology. Figure 4 provides a comparison of the relations of the ‘optimization’ node in the target ontology before and after executing the knowledge transfer process.

Figure 4.

Comparison knowledge integration: (a) original relations and (b) ontological dictionary relations.

As shown in Figure 4(a), only two relations (optimisation subclassOf: ‘Thing’ and optimisation subclassOf: ‘action’) were connected with the ‘optimization’ node. On the contrary, in Figure 4(b), some relations are transferred from the ontological dictionary to the target ontology. We can clearly observe that the process provided a better structure and knowledge to the target ontology. This is because both the extracted relations, properties and individuals provide missing information that has been validated by experts.

The same process performed for the ‘optimization’ class was carried out for all other nodes. As a result, 551 classes and 120 individuals were automatically transferred from Babelnet to the target ontology. Table 5 shows some examples of the knowledge transferred from the Babelnet ontological dictionary.

Table 5.

Examples of items transferred from Babelnet.

Nodes (classes)	Nodes (individuals)
Support system subClassOf thing	Conditional relation individualOf logical relation
Search engine subClassOf software	Reflexiveness individualOf logical relation
Conjunction subClassOf connection	Single-valued function individualOf mathematical relation
Operation research subClassOf research	NATO double-track decision individualOf decision-making
Object-oriented programming subClassOf thing	Semantic search individualOf search engine
Simulation software subClassOf application	Desktop search individualOf search engine
Optimality criterion subClassOf criterion	Residual sum of squares individualOf optimality criterion
Modelling subClassOf representation	Robotics simulator individualOf simulation software
Simulation subClassOf modelling	Technology CAD individualOf simulation software
Cosine function subClassOf transcendental function	High-throughput computing individualOf computing
Confidence interval subClassOf statistical method	Sine individualOf cosine function

In addition, the knowledge transfer process was executed using the third-party ontology as source of knowledge. As a result, 74 classes and 324 individuals were automatically transferred from the third-party ontology to the target ontology. Table 6 includes some examples of the knowledge transferred.

Table 6.

Examples of items transferred from our third-party ontology.

Node (classes)	Node (individuals)
Preference model subclassOf thing	Basic ASF individualOf achievement scalarising function
Achievement scalarising function subclassOf preference model	Composite scalarising functions individualOf achievement scalarising function
Decision rules subClassOf preference model	GUESS-ASF individualOf achievement scalarising function
Fuzzy logic subClassOf preference model	STOM-ASF individualOf achievement scalarising function
Knee point subClassOf preference model	Tchebycheff ASF individualOf achievement scalarising function
Light beam search subClassOf preference model	Weighted Euclidean distance ASF individualOf achievement scalarising function
Objective relative importance subClassOf preference model	Wierzbicki ASF individualOf achievement scalarising function
Outranking relation subClassOf preference model	DRSA individualOf decision rules
Polyhedrical cone-based subClassOf preference model	Fuzzy integral individualOf fuzzy logic
Ranking subClassOf preference model	Fuzzy measure individualOf fuzzy logic
Set preference relation subClassOf preference model
Utility function subClassOf preference model

ASF: achievement scalarising function; STOM: satisficing trade-off method; DRSA: dominance-based rough set approach.

As mentioned in Section 3, the fourth stage of the proposed methodology is dedicated to validation of the quality of the ontology with the help of evaluation metrics. To do so, we have used the Ontology fixer tool [5] that displays a radar chart using several metrics and computes the area of the polygon formed when the dots (quality metric values) are connected between them. The larger the area, the higher the quality of the ontology. Figure 5 shows the radar charts before and after knowledge transfer, both with Babelnet and the third-party ontology.

Figure 5.

Quality metrics results. (a) Initial target ontology, (b) result Babelnet knowledge integration and (c) result third-party and ontology knowledge integration.

As evidenced in Figure 5(b), the quality of the ontology resulting from integrating only the knowledge of the ontological dictionary improved 6% (six points) compared with the initial ontology. On the contrary, Figure 5(c) shows that the quality improves to a score of 31.81 when the knowledge of third-party ontology is also incorporated. CBOOnto was the metric that shows the greatest improvement in knowledge transfer from the ontological dictionary, which implies an enhancement in the coupling of the ontology. That is, a more complete and specialised ontology was obtained using the methodology proposed in the article, which means that the knowledge exchange process successfully filled the existing class-based concept gaps. Furthermore, there was also an increase in the CROnto and RCOnto metrics related to the number of individuals per class when transferring knowledge from the third-party ontology. Unlike when we used Babelnet as the source of knowledge, the amount of concrete knowledge (number of individuals) has increased significantly. Finally, we highlight that a quality assessment is not performed with respect to the ontology competency question metrics (CBOOnto is a maintainability metric, CROnto is a functional adequacy metric and RCOnto is a knowledge-base metric). If we had used metrics capable of evaluating this dimension, the quality of the target ontology would achieve a better-quality score.

Indeed, the possibility of adding other external sources of knowledge, in addition to those used in this case study, could lead to an improvement in the ontology quality metrics. Moreover, the process could be run interactively so that experts could specifically select the most relevant knowledge for transfer. This would result in an even more robust ontology. The following subsection summarises the advantages of our contribution, analysing the benefits and computational requirements of our approach.

4.3. Advantages of our proposal

The usage of ontology learning process is useful to reduce the efforts in creating ontologies to model knowledge domains from textual sources. However, ontologies generated through these algorithms often contain some limitations such as missing knowledge that could not be successfully extracted from textual sources and, therefore, needs to be fixed. The ontology correction processes developed previously only considered information related to structural, functional, adaptability, maintainability or operability quality parameters. However, the correction of semantic aspects was not possible since no semantic validation mechanisms had been introduced. In this work, we introduce a knowledge exchange methodology that takes advantage of available knowledge from different sources such as ontology dictionaries and third-party OWL ontologies for fixing purposes. The mechanism takes advantage of a simple ontology matching scheme, which is able to identify relevant knowledge from third-party sources (ontologies or semantic graphs) to fix automatically generated ontologies.

Since the knowledge sources used in this methodology have a structured nature (e.g. Babelnet API), knowledge collection, validation and exchange can be done efficiently and using low computational resources. As long as the methodology can be applied semi-automatically, it can effectively reduce the workload of experts in the ontology-fixing and quality improvement processes.

5. Conclusion and future work

In this study, we introduce a strategy for the validation and transfer of knowledge from ontologies created by third parties (ontological repositories) and ontological dictionaries (e.g. Babelnet, Wordnet) to semi-automatically or automatically generated ontologies that may contain semantic problems affecting their quality and structure. The algorithm comprises an ontology-matching process, which only retrieves information that could be valuable to enrich the content, improve the quality of the ontology and/or validate that the ontology content is appropriate within the domain.

With the support of the OntologyFixer tool developed in a previous research work, we can assess the quality of the target ontology by comparing radar charts that combine different metrics. The area of the radar chart displays the overall quality and allows the user to decide whether the transferred knowledge has a positive or negative impact on the target ontology. Similarly, by using OntologyFixer, we can further improve the quality of the ontology by applying the different quick fixes to address possible remaining issues.

Experimental results have demonstrated the suitability of the methodology for addressing some issues of an automatically generated ontology. The knowledge identification and selection process automatically discovered a collection of Babelnet synsets and third-party ontology nodes (records) that can be transferred to the target ontology. This shows the great potential of using exchange of semantic resources to improve the quality of a target ontology. So, a large number of third-party ontologies shared in public repositories can be used as sources of knowledge for improving other ontologies.

Finally, the knowledge transfer process is computationally simple. This is because (1) it is based on a computationally simple ontology matching method, (2) the knowledge sources considered in this methodology have a structured nature (graph) and finally, (3) the graph structure of knowledge sources matches (in the case of ontologies) or is very similar (in the case of semantic networks) to the target knowledge bases. The simplicity of this method clearly contrasts with the alternative of having an expert available to manually fix the ontologies.

This work may be extended to include new sources of knowledge that can be used for knowledge transfer. In particular, we believe that knowledge engineering may continue to benefit from methods and tools based on software engineering. In this sense, we believe that there is a possibility to adopt methods and technologies used in software engineering (e.g. SQL schema designs of relational databases, UML/XMI¹¹ domain diagrams, etc.) for knowledge transfer processes and/or improvement of the quality of ontologies. Furthermore, the use of semantic graphs allows additional simple validations such as ensuring that the different words or n-grams that make up a synset do not appear as names of different classes or individuals. Finally, the use of complex ontology matching schemes and the comparison of their performance should also be addressed as future work.

Footnotes

Acknowledgements

The SING group thanks CITI (Centro de Investigación, Transferencia e Innovación) from the University of Vigo for hosting its IT infrastructure.

Author contributions

All authors contributed equally to this work.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Conselleria de Cultura, Educación e Universidade (Xunta de Galicia) under the scope of the strategic funding of Competitive Reference Group (grant number ED431C 2022/03-GRC). V.B-.F. acknowledges FCT – Fundação para a Ciência e a Tecnologia, I.P., for its support in the context of projects UIDB/04466/2020 and UIDP/04466/2020.

Notes

ORCID iDs

Vitor Basto-Fernandes

José Ramón Méndez

References

Gruber

TR.

A translation approach to portable ontology specifications. Knowl Acquis 1993; 5: 199–220.

Asim

Wasim

Khan

MUG

et al. A survey of ontology learning techniques and applications. Database. Epub ahead of print 1 January 2018. DOI: 10.1093/database/bay101.

Lourdusamy

Abraham

. A survey on methods of ontology learning from text. In: Jain

Peng

S-L

Alhadidi

et al. (eds) Intelligent computing paradigm and cutting-edge technologies. Cham: Springer International Publishing, 2020, pp. 113–123.

Roldán-Molina

Ruano-Ordás

Basto-Fernandes

et al. An ontology knowledge inspection methodology for quality assessment and continuous improvement. Data Knowl Eng 2021; 133: 101889.

Roldan-Molina

Mendez

Yevseyeva

et al. Ontology fixing by using software engineering technology. Appl Sci 2020; 10: 6328.

Thiéblin

Haemmerlé

Hernandez

et al. Survey on complex ontology matching. Sem Web 2020; 11: 689–727.

Otero-Cerdeira

Rodríguez-Martínez

Gómez-Rodríguez

Ontology matching: a literature review. Exp Syst Appl 2015; 42: 949–971.

Bandeira

Bittencourt

Espinheira

et al. FOCA: a methodology for ontology evaluation. Arxiv:1612.03353[cs], 2017, http://arxiv.org/abs/1612.03353 (accessed 28 February 2022).

Duque-Ramos

Fernández-Breis

Iniesta

et al. Evaluation of the OQuaRE framework for ontology quality. Exp Syst Appl 2013; 40: 2696–2703.

10.

Bachir Bouiadjra

Benslimane

S-M

. FOEval: full ontology evaluation. In: 2011 7th international conference on natural language processing and knowledge engineering, Tokushima, Japan, 27–29 November 2011, pp. 464–468. New York: IEEE.

11.

Guarino

Welty

. An overview of ontoclean. In: Staab

Studer

(eds) Handbook on ontologies. Berlin; Heidelberg: Springer, pp. 201–220.

12.

Tartir

Arpinar

Moore

et al. OntoQA: metric-based ontology quality analysis. In: Proceedings of IEEE workshop on knowledge acquisition from distributed, autonomous, semantically heterogeneous data and knowledge sources, Houstan, TX, 27 November 2005.

13.

Tartir

Arpinar

IB.

Ontology evaluation and ranking using OntoQA. In: International conference on semantic computing (ICSC 2007), Irvine, CA, 17–19 September 2007, pp. 185–192. New York: IEEE.

14.

Gulla

. Semantic enrichment for ontology mapping. In: Meziane

Métais

(eds) Natural language processing and information systems. Berlin; Heidelberg: Springer, pp. 217–228.

15.

Blasko

Kremen

Kouba

Ontology evolution using ontology templates. Open J Sem Web 2015; 2: 16–29.

16.

Poveda-Villalón

Gómez-Pérez

Suárez-Figueroa

MC.

OOPS! (OntOlogy Pitfall Scanner!): an on-line tool for ontology evaluation. Int J Sem Web Inf Syst 2014; 10: 7–34.

17.

Troquard

Confalonieri

Galliani

et al. Repairing ontologies via axiom weakening. Corr 2017: 1711-03430, http://dblp.uni-trier.de/db/journals/corr/corr1711.html#abs-1711-03430

18.

Póveda

Suárez-Figueroa

Gómez-Pérez

. Common pitfalls in ontology development. In: Meseguer

Mandow

Gasca

(eds) Current topics in artificial intelligence. Berlin; Heidelberg: Springer Berlin Heidelberg, pp. 91–100.

19.

Noy

Musen

MA.

The PROMPT suite: interactive tools for ontology merging and mapping. Int J Hum-Comp Stud 2003; 59: 983–1024.

20.

Chen

Antonyrajah

et al. BERTMap: a BERT-based ontology alignment system, http://arxiv.org/abs/2112.02682 (2022, accessed 18 July 2022).

21.

Chen

Antonyrajah

et al. Biomedical ontology alignment with BERT, 2022, pp. 1–12, http://ceur-ws.org/Vol-3063/om2021_LTpaper1.pdf

22.

Chen

Jiménez-Ruiz

Horrocks

et al. Augmenting ontology alignment by semantic embedding and distant supervision. In: Verborgh

Hose

Paulheim

et al. (eds) The semantic web. Cham: Springer International Publishing, pp. 392–408.

23.

Haldar

Mukhopadhyay

Levenshtein distance technique in dictionary lookup methods: an improved approach. arXiv:11011232 [cs, math], http://arxiv.org/abs/1101.1232 (2011, accessed 28 February 2022).

24.

Bevilacqua

Pasini

Raganato

et al. Recent trends in word sense disambiguation: a survey. In: Proceedings of the thirtieth international joint conference on artificial intelligence, Montreal, QC, Canada, 19–27 August 2021. Montreal, QC, Canada: International Joint Conferences on Artificial Intelligence Organization, pp. 4330–4338.

25.

Duque-Ramos

Fernandez-Breis

Stevens

et al. OQuaRE: a SQuaRE-based approach for evaluating the quality of ontologies. J Res Pract IT 2011; 43: 159–176.

26.

Roldan-Molina

Yevseyeva

Ruano-Ordás

et al. A reference OWL ontology for knowledge inspection and automatic quality improvement methods benchmarking. Zenodo. Epub ahead of print 18 December 2021. DOI: 10.5281/ZENODO.5791362.

27.

Yevseyeva

Basto-Fernandes

et al. Building and using an ontology of preference-based multiobjective evolutionary algorithms. In: Trautmann

Rudolph

Klamroth

et al. (eds) Evolutionary multi-criterion optimization. Cham: Springer International Publishing, pp. 406–421.

28.

Yevseyeva

Basto-Fernandes

, et al. (2017). Ontology of Preference-Based Multiobjective Evolutionary Algorithms. Zenodo. Epub ahead of print 20 March 2017. DOI: 10.5281/ZENODO.3269857.