Understanding large ontologies, with diverse semantics and modelling practices, is still an issue, and has an impact on many ontology engineering tasks. While existing methods summarise ontologies by extracting the most important nodes or subgraphs, a complete overview of an ontology, and a comparison between multiple ontologies, are not supported. Based on the hypothesis that ontologies are designed as compositions of patterns, this paper presents a research proposal for developing a method able to extract conceptual components from multiple ontologies and the observed ontology design patterns implementing them.
different modelling styles): e.g. the location component (being located at a place)
can be implemented in one or more ontologies as a binary (e.g. hasLocation) and
a n-ary relation (e.g. Location) between 3 arguments (Time, Object, Place), or
can be specialised as the location of a specific object (e.g. a building).
ODPs, being reusable template solutions to recurrent modelling problems [
Based on these premises, this research project aims at providing methods and tools for supporting ontology understanding, and indirectly other ontology engineering tasks, by (i) identifying modelling problems (CCs) common to multiple ontologies and implemented with specific modelling solutions (ODPs), (ii) analysing these solutions and comparing them with documented ODPs, (iii) automatically annotating them to make them recognisable. 2
Automatically extracting CCs and their corresponding ODPs from KGs can
provide a basis for novel approaches to support ontology engineering tasks.
Ontology selection. An ontology designer may need to reuse existing
ontologies for modelling her data. The patterns implemented in an ontology reflect the
modelling problems that it addresses, rather than the collections of concepts it
contains. Therefore, a modularised, pattern-based and topic-centered description
of ontologies should support a better understanding, hence selection of them.
Ontology visualisation. Ontology visualisation is an important tool for
working with ontologies. Most of the existing frameworks show node-link views with
a focus on class hierarchies [
ODP-based ontology engineering. Currently, ontology patterns are
developed manually and collected in repositories such as ODP Portal. Mostly based
on a top-down approach, these repositories unavoidably lack completeness [
This research requires tackling different topics that have been investigated in the literature.
ODPs. Pattern-based ontology engineering methodologies, such as [
ODPs detection. These methods focus on identifying parts of ontologies reusing
s-o-t-a ODPs. [
Ontology selection and understanding. Users can browse terms from
different ontologies on catalogues of ontologies (e.g. BioPortal) and semantic search
engines (e.g. Schemapedia). Comparison of ontologies and ODP-based filtering
of the results are not supported. Most ontology summarisation approaches [
Ontology partitioning. Modularisation approaches (e.g. [
Complex ontology matching. Complex ontology alignments [
My research project aims at addressing the following research questions, driven by the respective hypotheses.
RQ1: What are the Conceptual Components, and the ODPs implementing them, used for modelling knowledge graphs on the web? RQ1.1: Is the boundary of an ODP identifiable? RQ1.2: How to assess that fragments from different ontologies address the same modelling problem? RQ2: Considering the ODPs observed on the web and the ODPs defined in online catalogues of patterns: do they match? RQ3: How to automatically annotate ODPs used in a KG? (RQ1) H1: A combination of designated heuristics can support the identification of the boundary of implemented ODPs (RQ1) H2: Densely connected subgraphs detected in ontologies, grouped based on the terminology used in the literal data describing their ontology entities, may indicate modelling problems addressed by different OWL implementations. (RQ2) H1: Matching observed patterns with catalogues’ ones gives a measure of how much, and how, predefined patterns are reused in practice. (RQ3) H1: A designated language can annotate ODPs observed in KGs: their components, attributes, relations. 5
The intuition behind this research is that an ontology is (un)intentionally developed as a composition of ontology design patterns, intended as modelling solutions observed in existing ontologies, regardless their correctness or quality. I hypothesise that these patterns can be empirically observed based on their density and their vocabulary. I expect (i) the density of their internal connections (i.e. connections between entities of the same ODP) to be higher than the density of the connections between entities from different ODPs; (ii) the combination of the words describing an ODP to evoke the relation it represents (e.g an ODP implementing the location component may possibly include terms such as place and is located at ). Community detection can recognise phenomena such as (i), while text clustering allows me to exploit the vocabulary of ODPs in order to group them based on their similarity, which may indicate they are addressing the same modelling problem.
CH ontologies corpus. As an empirical basis, I built a corpus of schemas modelling Cultural Heritage (CH). The choice of selecting ontologies from this domain is motivated by (i) my previous experience in developing CH ontologies, and (ii) the general complexity of concepts and diversity of terminology of this domain. Indeed, there are many different domains that are related to the already vast domain of CH (e.g. geology for archaeological properties), heterogeneous types of cultural heritage (musical, architectural, etc.), and different cultural institutions have often their own classifications and terminologies. 43 ontologies have been selected from the literature, using a catalogue of general purpose ontologies, i.e. LOV, and publishing an online survey.
Method. For extracting CCs and ODPs from multiple ontologies, I adopt an
empirical approach that combines community detection, word sense
disambiguation, frame detection, clustering techniques (see Figure 1). The same method can
be exploited in order to group ODPs from actual ontologies and catalogues’ ones.
Intensional ontology graphs. In order to transform an ontology into a graph that
can be processed by community detection algorithms (undirected and unlabelled
graphs) while preserving the formalisation of the ontology conceptualisation, I
formally represent it as an intensional graph, i.e. a graph aiming at encoding
the intensional level of an ontology. An edge :p is generated between every two
nodes that are domain and range of a property :p. As for property restrictions on
classes, an edge :p links the class local to the restriction and the class in the
restriction expression. Finally, for each edge :p between two nodes :n1 and :n2, two
unlabelled edges are generated: the first between :n1 and a new node :n1-p-n2,
the second between :n1-p-n2 and :n2, in order to preserve the context of use of :p.
Community detection. Community detection splits networks into groups of nodes,
such that there is a higher density of edges within groups than between them:
since I expect the links between the entities involved in one ODP to be denser
than the links between entities from different ODPs, communities could
identify possible patterns. A modified version of a s-o-t-a algorithm [
Based on this method, it is possible to build a catalogue of conceptual components and observed ODPs, where each conceptual component links to its associated ODPs within the ontologies: e.g. see the catalogue from the CH corpus. Experiments. I run the method described in Method on both the CH corpus and another corpus (Conf ) from the Conference evaluation track of the Ontology Alignment Evaluation Initiative (OAEI). Through a manual inspection of the communities, I found that there are recognisable patterns common to many communities, which correlate with the modelling practice adopted for a specific ontology (fragment). Some communities ( 5% of CH and 1% of Conf) can be identified as missing a conceptual unity, because of poor axiomatisation in the source ontology: e.g. communities grouping semantically heterogeneous properties that are not involved in any restrictions. Instead, the majority of the communities have a good level of semantic coherence e.g. representing the acquisition of a cultural property from a previous owner, or the membership in a conference. The clusters identified from both corpora represent different conceptual components, at different levels of abstraction: components as event, categorization, intentionally act are present in both corpora, while other components specific to the domain emerge (e.g. performing arts in CH, award in Conf). In both corpora, some clusters could be either split or merged. Even if my goal is not to produce ontology alignments, I evaluated these initial results against the ontology matching (OM) task, based on the hypothesis that, given a pair of similar entities to be aligned, they should belong to either the same cluster or two related clusters. The results of my method are compared with alignments asserted within the ontologies of the corpus; alignments generated by an OM tool with good performances; gold standard alignments on the Conf corpus in the OAEI 2020 conference track. This experiment shows that clusters and their relations may be used to improve the performance of alignment algorithms, significantly reducing the dimension of the task. 1Abstract conceptualisations of objects or events that may be evoked by words. Annotating ODPs. I am working on an extended version of OPLa annotation language, able to annotate an ontology with its implemented ODPs and CCs. 6
The method will be evaluated on: (i) the communities as possible ontology design patterns, (ii) the clusters of communities as possible conceptual components (RQ1). The evaluation will be both manual and automatic. A first evaluation of the initial experiments, consisting in manually inspecting communities and clusters detected on the CH and Conf corpora and in analysing the results in the context of the ontology matching task, has been already performed. A corpus of ontologies annotated with the ODPs they implement will be built in order to have a gold standard for comparing our automatically detected ODPs. This activity will test if the extended annotation language is fit for the purpose (RQ3). An experiment will be designed that clusters community from ontology corpora and catalogues’ ODPs: based on the terminology, observed and s-o-t-a ODPs addressing the same CC should end up in the same cluster (RQ2). The results produced will go also through an evaluation phase by humans, in order to empirically assess the degree of agreement between automatic and manual detection of ontology patterns and conceptual components, possibly designing crowdsourcing tasks. Even if a user-based evaluation of the catalogue would be valuable, it is not an easy user study to be designed, as it would be evaluating key concepts detection, since it requires involving experts in ontology design based on patterns. Therefore, an indirect user-based evaluation will be considered, by e.g.: (i) evaluating a pattern-based visualisation tool that will be based on my method; (ii) using my method as a basis for ontology selection tasks. These activities will be integrated within the EU H2020 project Polifonia, which will realise an ecosystem of computational methods and tools for the European musical heritage (MH) knowledge on the web: this will provide me with an additional large-scale experimental basis. 7
Preliminary results of this research proposal are already satisfying and show the potential of the method, but they still need to be improved.
Future work will include different activities. The intensional graphs need to include also all class expressions, in order to maximise the detection of relevant communities and limit loss of information. A refinement of the conceptual components (by merge or split of detected clusters) based on heuristics to be defined will be a next step. I will study a method to match observed patterns, detected in actual ontologies, to online catalogues’s patterns, which may have a great impact for supporting interoperability. Moreover, the extended version of OPLa will be necessary for automatically annotating ODPs in ontologies. Acknowledgements. This research is conducted under the supervision of Dr. Valentina Presutti and is supported by the CNR (E-RIHS infrastructure).