Foundational ontologies meet ontology matching: A survey

1. Introduction

A primary purpose of work on ontology is to create a common computable semantics for concepts in the world. An ontology describes the concepts and the relation between these concepts together with constraints on how to interpret them. Foundational ontologies aim to create a computable semantics for very general concepts and associated terms. With a large number of ontologies having been created for different domains, as well as different conceptualizations for upper, top or foundational ontologies, the creation of shared semantic descriptions is still an issue.

The role of fondational ontologies in ontology construction, matching and integration is manifold. Their potential for clarity in semantics and a rich formalization are important requirements for ontology development improving ontology quality [41,52] and preventing bad ontology design [34,75]. These ontologies may also act as semantic bridges supporting interoperability between ontologies [35,50,51].

In the semantic web and linked data in general, as stated in [3], distinctions such as whether an entity is inherently a class or an individual, or whether it is a physical object or not, are rarely expressed in the linked data, although they have been largely studied and formalised by foundational ontologies. As stated by the authors, while the first distinction showed its practical importance in modelling and meta-modelling approaches, it is also at the basis of knowledge representation formalisms (such as RDF and OWL) for supporting taxonomic reasoning. The second distinction (physical object or not) is essential to represent the physical world and express their expected physical behaviour (e.g., spatial extension, gravity). Such distinctions are however nearly ubiquitous in ontologies and domain data described by these ontologies.

When a foundational ontology is used in the development and integration of domain ontologies – during or after – top-down and bottom-up approaches may be considered [76]. With a top-down approach, the foundational ontology is used as a reference for deriving domain concepts, taking advantage of the knowledge and experience already encoded in it. In a bottom-up approach, one usually matches an existing domain ontology to a foundational one. The latter is more challenging since inconsistencies may exist between domain and foundational ontologies. For example, a domain ontology might be about hypothetical or fictional entities, and a particular foundational ontology might commit only to real entities. Furthermore, one has to deal with different levels of abstraction and also of formalization in the matching process.

Ontology matching is a research area aimed at finding ways to make different ontologies interoperable. The matching process can be seen as the task of generating a set of correspondences (i.e., an alignment) between the entities of different ontologies [18]. Correspondences express relationships between ontology entities. For instance, the concept of Author in one source ontology may be marked as equivalent (that can be interpreted as logical equivalence or more informally as a high level of similarity) to the concept of Writer in one target ontology, or that Writer in the source is subclass of Person in the target. One’s intuitions about the formalizations of those words may be that they are different, but unless an ontology specifies a formalized definition, they are just labels. A linguistic-based matching algorithm might find them sufficiently similar and evaluators of the match would not have an objective basis for declaring it faulty.

Whereas the area of ontology matching has developed in the last decades, the problem of matching ontologies involving foundational ontologies has seen less development regarding automatic solutions [44,70]. This is not surprising since matching foundational and domain ontologies is a highly complex task, even when done manually. It requires the deep identification of the semantic context of concepts and, at a minimum, the identification of subsumption relations, and in a way such that the subsumption relations must of course be consistent with the formalization of the subsuming concept in the upper ontology. In fact, subsumption and other relations are often neglected by most state-of-the-art matchers.

There have been many manual efforts to make sense of how different foundational ontologies relate to other lexical and semantic data bases, and how they improve the process of matching domain ontologies. In this paper, we survey various approaches to ontology matching using foundational ontologies to create shared semantics.

An additional challenge is that there is little agreement on many of the possible goals and methods of ontology construction or the formal languages in which to encode an ontology. Developers of ontologies have variously advocated very small upper ontologies or large ones, very expressive formal logics, or very minimal ones as a way to support fast logical inference. Evaluations and surveys have typically been conducted by the authors themselves, or their collaborators and supporters (with the possible exception of [50]).

Considering this scenario, this paper reviews the following tasks of ontology matching involving foundational ontologies:

We discuss the main strengths and weaknesses of existing approaches and highlight the challenges to be addressed in the future. We consider that this comprehensive study may set the grounds for advancing domain and foundational ontology matching.

The scope of this paper is in using foundational ontologies for matching and integration of ontologies. While this necessarily touches on the topics of how to create or evaluate ontologies themselves, we will address this topic only in the service of evaluation the use case of ontologies for matching. As such, we do not attempt to review all available foundational ontologies, but just study the use of them for matching other ontologies.

The rest of the paper is organised as follows: Section 2 introduces the different foundational ontologies and ontology matching. Section 3–Section 6 discuss the approaches in the categories (i)–(iv) introduced above. Section 7 discusses the open challenges in the field and Section 8 concludes the survey.

2. Background

2.2. Ontology matching

Ontology matching refers to a process that consists of generating an alignment ( A ′ ) from a set of ontologies Ω, usually a pair of ontologies ( Ω = { o 1 , o 2 } ).

Definition 1 (Matching process).

The matching process can be seen as a function f which, from a pair of ontologies o 1 and o 2 to match, an input alignment A, a set of parameters p, and a set resources r, returns an alignment A ′ between o 1 and o 2 : A ′ = f ( o 1 , o 2 , A , p , r )

Each of the elements featured in this definition can have specific characteristics which influence the difficulty of the alignment task.

An alignment (A or A ′ ) consists of a set of correspondences { c 1 , c 2 , … , c x } :

Definition 2 (Alignment).

An alignment A o 1 → o 2 is a set of correspondences { c 1 , c 2 , … , c n } . A o 1 → o 2 is directional between a source ontology o 1 and a target ontology o 2 .

A correspondence expresses a relation r between ontology entities of o 1 and o 2 . Here, the ontology entities are members of the correspondence:

Definition 3 (Correspondence).

A correspondence c i is a tuple ( e 1 , e 2 , r ) . e 1 and e 2 are the members of the correspondence. They can be simple or complex expressions with entities from respectively o 1 and o 2 :

if the correspondence is simple, both e 1 and e 2 are simple expressions.

if the correspondence is complex, at least one of e 1 or e 2 is a complex expression, involving union, intersection, disjunction, cardinality restrictions, etc.

r is a relation, e.g., equivalence (≡), more general (⊒), more specific (⊑), disjointedness (⊥) holding between e 1 and e 2 .

The correspondence ⟨ e 1 , e 2 , r ⟩ asserts that the relation r holds between the ontology entities e 1 and e 2 . One has to distinguish between more and less rigorous interpretations of relations. The equivalence relation, for example, can be interpreted as logical equivalence or more informally as a high level of similarity. Back to the example in the introduction, consider the fragment of the ontologies ekaw14

¹⁴

http://oaei.ontologymatching.org/2019/conference/data/ekaw.owl

and cmt15

¹⁵

http://oaei.ontologymatching.org/2019/conference/data/cmt.owl

in Figs 1 and 2, respectively. The format used to represent the ontologies is described in [83]. The following correspondences can been established between these two ontologies: (i) (ekaw:Paper,cmt:Paper, ≡) is a simple correspondence; (ii) (ekaw:AcceptedPaper, ∃cmt:hasDecision.cmt:Acceptance, ≡) is a complex correspondence with constructors.

OPEN IN VIEWER

Fig. 1.

Fragment of ekaw ontology (adapted from [87]).

OPEN IN VIEWER

Fig. 2.

Fragment of cmt ontology (adapted from [87]).

While the RDF alignment format provided in the Alignment API16

¹⁶

http://alignapi.gforge.inria.fr/format.html

[14] is the format de facto used in the Ontology Alignment Evaluation Initiative (OAEI) campaigns for representing simple alignments, correspondences can also be represented as OWL 2 subclass, equivalence, and disjointedness axioms (with confidence values represented as axiom annotations). Alternatively, the EDOAL17

¹⁷

http://alignapi.gforge.inria.fr/edoal.html

language (Expressive and Declarative Ontology Alignment Language) extends the alignment format in order to represent complex correspondences.

In the following, we discuss the use of foundational ontologies in different matching tasks: (i) matching of foundational ontologies; (ii) matching of foundational ontologies to lexicons; (iii) matching domain ontologies with the help of foundational ontologies; and (iv) matching foundational ontologies to domain ontologies.

As stated in [42], while the purpose of a foundational ontology is to address interoperability among ontologies, the development of different foundational ontologies re-introduces the interoperability problem. As briefly discussed in the previous section, these ontologies have been developed directed at different classes of applications, as well as relying on different theoretical assumptions.

Early work addressed this problem [27,79,85] from different perspectives on the alignments. While [27] compared specific treatments of fundamental issues (as significant discrepancies related to universals and particulars, qualities, constitution and spatio-temporality) and how similar notions apply differently in BFO and DOLCE, [79] compared the primitive relations (dependence, quality, and constitution) between these ontologies. In [85], the alignment between BFO and DOLCE was established in order to conciliate their respective realistic and cognitive points of view and to integrate medical data. While 100% of BFO categories were aligned to DOLCE, 81% of DOLCE categories were aligned to BFO.

More recently, [90] compares BORO and UFO ontologies according to the their metaphysical choices that define their structure and composition. Instead of comparing terms in both ontologies, the authors compare how the two approaches address issues such as identity and dynamic classification, the treatment of relationships (i.e., instances of relational properties), as well as the relation between existence and time in the two approaches. The radical difference between these two ontologies, hence, reflect deeper differences in ways of conceiving reality.

Other studies addressed other foundational ontologies. In [42], alignments between BFO, DOLCE and GFO have been established with automatic matching tools and manually, with substantially fewer alignments found by the matching tools. The alignments in the context of the whole ontology revealed a considerable number of logical inconsistencies. This work has been extended in [73] in two ways: considering matching systems participating in OAEI 2018, and a new pair of aligned foundational ontologies (SUMO and DOLCE). The alignments in [42] and [59] served as a reference alignment to automatically evaluate the matchers. Examples of reference correspondences include: (bfo:Occurrent,dolce-lite:perdurant, ≡) and (dolce-lite:artifact,sumo:Device, ⊒).

Overall, the results found are in line with what has been reported when evaluating the behaviour of matchers in the task of matching domain and foundational ontologies, which would also require identification of subsumption relations [74]. Current tools fail on correctly capturing the semantics behind the ontological concepts, which requires deeper contextualization of the concepts on the basis of their hierarchy and axioms. Addressing the identification of subsumption relations, the approach in [40] relies on extracting hypernym relations from ontology annotations for establishing such kind of correspondences. Results on exploiting lexico-syntactic patterns and definitions layout on DOLCE and SUMO were evaluated on a manually generated subsumption reference.

From another perspective, the core characterization of mereotopology (a theory of physical parts) of SUMO and DOLCE has been studied in [56], relating their axiomatizations via ontology alignments. This included corrections and additions of axioms to the analyzed theories which eliminate unintended models and characterize missing ones. Finding alignments between DOLCE and SUMO was also addressed in [59], where the SmartDOLCE and SmartSUMO ontologies have been developed on the basis of DOLCE and SUMO. The alignment of the just the taxonomic statements from SUMO to DOLCE involved extracting the upper-level of SUMO and the non-trivial task of aligning the remaining concepts to appropriate DOLCE categories.

Aligning foundational ontologies reveals also the problem of matching their different versions. In [78], a method for tracking, explaining and measuring changes between successive versions of BFO1.0, BFO1.1, and BFO2.0 was applied. The aim was to provide a more comprehensive analysis of the changes with respect to the BFOConvert tool18

Formalizeations [10] within the Common Logic Ontology Repository (COLORE), were used in the specification of alignments between upper ontologies in [28]. These alignments serve for the verification of foundational ontologies. Similarly, [11] shows how to apply techniques for ontology verification to link interpretations among ontologies.

Table 2 summarizes work on matching of foundational ontologies. They are mostly manual efforts, with a few of the resulting alignments made available.

Several efforts in equipping lexical resources with foundational ontologies have been made in order to associate a formal semantics to their lexical layer. As stated in [20,21], while WordNet has been used in numerous work as an ontology, where the hyponym relations between word senses are interpreted as subsumptions relation between concepts, it is only serviceable as an ontology if some of its links are interpreted according to a formal semantics that tell us something about the world and not just about language.

For example, WordNet has the sense of “chair” as a hyponym of “seat” but only an English gloss meant for humans to read for each, and no logical semantics that defines the hyponym link, with the result that hyponyms are often incorrectly treated by users as being logically transitive. An ontology would, at a minimum, define the axiom of transitivity, and state that it holds on the hyponym relation. Most ontologies have little more than class/subclass relations however. One might also want to state that a chair is a artifact made by humans with the intent for use in supporting a seated humans. Some upper ontologies have axioms that define their terms, and some large taxonomies have terms for specific things like chairs, but very few have axiomatized large number of detailed objects, processes and relationships. This poses a problem for matching algorithms if there aren’t both detailed formalizations that can be used to objectively determine a correct match and a large inventory so that matches are not to trivially general terms like object or event.

A number of researchers have investigated different ontological problems in treating WordNet as an ontology (e.g., confusion between concepts and individuals, constraints violations, heterogeneous levels of generality, etc.) [20] and provided the WordNet taxonomy with more rigorous semantics. First the WordNet taxonomy was reorganized to meet the OntoClean [31] methodology requirements, and the resulting upper level nouns were then mapped to DOLCE classes representing their highest level categories. This alignment is concentrated on the noun database, since most particulars in DOLCE describe categories whose members are denoted by nouns. The result is the OntoWordNet resource expressing alignments between WordNet 1.6 version and DOLCE Lite Plus. An extension is presented in [24] in order to extract association relations from WordNet, and to interpret those associations in terms of a set of conceptual relations in DOLCE.

Later, this alignment has been updated [25] with a revision of the manual alignments and different versions of DOLCE and WordNet, WordNet 3.0 and (DOLCE UltraLitePlus), which is a simplified version of DOLCE Lite Plus, intended to make classes and properties names more intuitive and express axiomatizations in a simpler way, among other features.

While these works focused mostly on WordNet noun synsets, [81] extended the previous alignments by aligning verbs according to their links to nouns denoting perdurants, transferring to the verb the DOLCE class assigned to the noun that best represents that verb’s occurrence. They argue that many NLP applications need to deal with events, actions, states, and other temporal entities that are usually represented by verbs.

The alignment of WordNet to other foundational ontologies has been also addressed. In [77], a semi-automatic method for aligning WordNet 3.0 and BFO2.0 is described. It adopts previous alignments between WordNet and the KYOTO ontology, whose top layer is based on DOLCE. The method involves manually creating a set of alignments between the ontologies and implementing a set of matching rules.

In [67], the authors report the matching and integration of several background resources and ontologies of varying complexity to the Cyc knowledge base. These resources and ontologies included large pharmaceutical and medical thesauri and large portions of WordNet. For this task, ontologists have been trained with domain experts and interactive clarification dialog-based tools were developed to enable experts to directly match/integrate their ontologies. In [58], SUMO was originally mapped manually to WordNet 1.6 and then manually updated to 3.0.19

SUMO and WordNet were used in a semi-automated process to match the millions of terms in the YAGO20

Finally, in [45], WordNet has been extended by applying the notion of semantic types in order to establish matching rules between the noun synsets of WordNet and the top-level constructs of the UFO ontology. The proposed rules were validated through an experiment with approximately 5,200 sample correspondences and average accuracy of 93%.

Table 3 summarizes the works presented in this section. There are some available alignments, approaches are mostly semi-automatic, with one case of manual alignment and one case of automatic alignment. WordNet is the lexical resource that is considered in all work listed and two different versions are involved in the alignments (1.6 and 3.0).

Foundational ontologies provide a reference for rigorous comparisons of different ontological approaches, and a framework for analysing, harmonizing, matching and integrating existing domain ontologies [59]. In domain ontology matching, in particular, they act as semantic bridges to help the task. For instance, reducing the matching space to the entities under a same category e.g., avoiding matching dolce:endurants to dolce:perdurants.

Despite the potential gain of exploiting foundational ontologies in domain ontology matching, few works have addressed this alternative, possibility due to the still lack of systematic alignments between domain and foundational ontologies. This gain has been quantitatively measured in [51], where a set of algorithms exploiting such semantic bridges are applied. The circumstances of cases where foundational ontologies improve domain ontology matching, with respect to approaches ignoring them, were then studied. The experiments were conducted with SUMO-OWL (a restricted version of SUMO), OpenCyc and DOLCE and demonstrate that overall the alignment via upper ontologies impacts in F-measure positively. Additionally, in [60] a set of alignment patterns based on OntoUML (a conceptual modeling language based on UFO) are applied to a set of alignments generated by matching systems. An analysis of the impact of patterns to avoid common errors was presented.

The semi-automatic LOM matcher [48] combines WordNet synset matching (checking terms from the ontologies to be matched sharing common synsets) and type matching. This strategy employs SUMO to determine the ontological category of each word constituent for matching using the alignments from WordNet synsets to SUMO. LOM takes the source terms and collects the set of SUMO terms that their synsets map to, and then compares the SUMO term sets to their counterpart for each term in the target ontology.

From a manually established alignment between biomedical ontologies and BFO, in [82], a matching approach relies on filtering out correspondences at domain level that relate two different kinds of ontology entities. The matching approach is based on a set of similarity measures and the use of foundational ontology as a parameter for better understanding the conceptual nature of terms within the similarity calculation step. Besides the reported improvement in the results obtained, the introduction of foundational ontologies in the alignment process increased the influence of semantic factors in this task, further expanding the universe of information to be explored during the alignment.

Table 4 summarizes the use of foundational ontologies as an aid to the of matching domain ontologies. Here the automatic approaches are adopted more frequently. Alignments, however, were not found to be available.

Methodologies for constructing ontologies should not neglect the use of foundational ontologies and may better address it in a top-down approach [2,41]. As one example, existing methodologies including OntoUML, allow for creating a domain or core ontology compliant to UFO [37].

While matching foundational ontologies is mostly manually done, with more automation in matching domain ontologies via foundational ones, in this section, both approaches have been performed.

The following sections discuss a series of issues regarding matching of foundational ontologies: the complexity of the task, the automation of systems with capabilities to include such alignments, the lack of evaluation data sets, the evolution of different versions and the problems that poses, the desired variety and lack of expressiveness in the alignments, and finally, multilingualism.

8. Final remarks

Ontology matching has reached some maturity in terms of matching domain ontologies. There is however room for further developments in the adoption of foundational ontologies in the task. Systematically enriching domain ontologies with foundational ones would also promote their use as semantic bridges in the task of matching domain ontologies. One of the difficulties however is the need of specialised knowledge as injecting foundational ontologies in ontology matching, in general, requires deeply understanding the foundational concepts and its relations. Another issue concerns the lack of formal definitions associated to lexicons (comments and labels) helping to understand the precise semantics of each concept. These are among the main brakes to automatising together with the points discussed above.

This paper has provided an overview of the adoption and exploration of foundational ontologies in the task of ontology matching, on different perspectives: work attempting to compare and match foundational ontologies, natural language definitions vs logical statement issues addressed by linking lexicons and foundational ontologies, the role of foundational ontologies as bridges for linking domain ontologies; and equipping existing domain ontologies with foundational distinctions. We have pointed out the limitations to be addressed in order to bring the clarity of semantics of foundational ontologies in ontologies in general.

In a broader scope, semantic web in general and its materialisation with the linked open data initiative still lack such ontological distinction, as recently stated in [3,88]. This has been further corroborated in [5], where it is stated that in the semantic web, there is an increasingly need for serious engagement with ontology, understood as a general theory of the types of entities and relations making up their respective domains of inquiry. However, there is still little interaction between the communities, despite the fact that they share common ambitions in terms of knowledge understanding. This goes beyond the matching task as discussed in this paper in the sense that is has to take into account the data being described by the ontologies.