The internationalisation goal for OWL sought to ofer support for multilingual ontologies. User-displayable labels were suggested as a way to realise this, by means of rdfs:label. However, because each label is a language-tagged string, this hampers accurate representation of strings in languages that require grammatical features such as inflected forms and gender. At least eight linguistic models have been proposed to address this key shortcoming, with OntoLex-Lemon now the de facto standard. The purpose of this survey was to determine if there has been any adoption of linguistic models within OWL ontologies. As OWL ontologies are widely used in the biomedical domain, the survey was limited to those ontologies in NCBO BioPortal, a biomedical repository. The results indicate that OntoLex-Lemon was not used in any production OWL ontology at time of review, nor that of any other linguistic model. In addition, the adoption rate of multilingualism in OWL ontologies in BioPortal was observed to be 5%, with English the primary language, followed by French and German.
There has been wide adoption of ontologies in the biomedical domain [
so it was not able to be included in the review.
Copyright © 2022 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
languages is not without challenge. For instance, Keet and Khumalo took steps
to translate SNOMED CT into isiZulu (a Niger-Congo B (‘Bantu’) language),
with the aim to provide localised electronic health records and patient
discharge note generation [
When developing multilingual ontologies, there are several options [
The standard approach is to set multiple language-tagged annotations using
rdfs:label [
Use of either one is determined by the gender of the noun of the class in the
domain position, or the subject of the sentence. As rdfs:labels are relied upon
in question-answering and similar systems [
OntoLex-Lemon and its precursor, Lemon [
The primary purpose of this study was thus to identify the ratio of multilingual OWL ontologies to monolingual ones, the modelling options used for multilingual ontologies, as well as the OWL profile used. A core aspect of multilingualism for ontologies is the identification of elements using URIs. An identifier is a fragment of the URI that uniquely identifies the entity in
question. A URI fragment can be opaque (semantic-free) or descriptive
(meaningful) [
We also analysed coverage of each natural language compared to the total
entity count within an ontology. As part of this work, we used the LLC lang metric
for multilinguality identified by Ell et al. [
In the next section, the methodology for data collection and preparation is described. In Section 3, the ontologies are analysed for their multilingual aspects and discussed. We close with the conclusion in Section 4. 2
Data collection was conducted using the REST API3 provided by BioPortal. The steps to collect the required data are listed below. 1. Using the /ontologies endpoint within Postman4, a list of ontologies were downloaded and saved as a JSON lfie. 2. Using the /categories endpoint, the process in Step 1 was repeated to get the list of categories, with both files then imported into a MySQL database. 3. A script was written which queried each ontology using the /ontologies/{acronym}/latest submission and /ontologies/{acronym}/categories endpoints. The following data was then saved for each ontology: the list of linked categories, the ontology language used, the namespace acronym, URI, status, and description. 4. Only ‘OWL’ ontologies were selected, with all other formats excluded. 5. Only ‘production’ ontologies were selected, with all other statuses excluded. 6. Ontologies in the ‘Upper Level Ontology’ and ‘Vocabularies’ categories were then excluded. 7. Thereafter, each remaining ontology was downloaded, using the /ontologies/{acronym}/download endpoint.
BioPortal provides categories into which each ontology is classified. Ontologies in the ‘Upper Level Ontology’ were excluded in Step 6 as the focus was on domain ontologies. Ontologies in the ‘Vocabularies’ category were also excluded as these are controlled vocabularies and thesauri, which are typically lightweight ontologies at best. For Step 7, if the file size was large and resulted in timeout errors, the file was manually downloaded. If the Download endpoint returned an error for an ontology, the ontology would be manually inspected online. If the latest submitted version had an ‘Error RDF’ annotation, then no further download attempt was made.
Steps 1–7 were run on the 30/12/2020. Step 1 resulted in a dataset of 912
ontologies5. At the end of Step 7, a dataset of 220 ontologies remained. The
review of ontologies for their multilingualism were conducted on this dataset. The
possible options were multilingual labels, linguistic models, and mapping models.
A diagrammatic representation of the multilingual labels option is shown in
Figure 1 for the class ‘Enzyme’ from the Ontology for Biomedical Investigations
(OBI) [
Each ontology file was loaded in memory, using a script to loop through each line therein. For multilingual labels, @ and xml:lang was searched for, and if found, that line was saved to the database. For linguistic models, any line which contained the namespaces: ontolex (OntoLex-Lemon), linginfo, lexonto, lexinfo, gold and lemon was saved to the database. Likewise for mapping models, any line which contained skos was saved to the database. Thereafter, these saved rows were manually inspected for their multilingual aspects. The natural languages used in each ontology were easily identified; however, identification of those as primary language(s) was more subjective. To determine an ontology’s primary language, an ontology was manually inspected, focussing on the labels used, and the consistency thereof for each of its languages. To qualify as language-independent, there would have to be near-equal support for the natural languages therein, as well as opaque URI identifiers. We applied LLC lang per multilingual ontology to support identification of those as primary language ontologies.
Of the multilingual ontologies identified, the OWL prolfie(s) of each were then
determined using the OWL Classifier [
2 DL were examined and if determined to be related to annotations only, the ontology was reclassified as OWL 2 DL as these violations can be patched easily. 3
The dataset of 220 ontologies was analysed on the diferent types of modelling
options. Only 11 were identified as being multilingual, which are listed in Table
1. As can be observed, they all use the multilingual labels option and mostly
opaque identifiers. The latter is likely a follow-on efect from OBO, which has a
principle that identifiers should not have a label meaningful to humans [
We highlight the multilingual aspects of some of the identified ontologies that prompted classification as ‘language-independent’ or ‘primary language’; Table 2 refers. Both CIDOC-CRM and RADLEX are language-independent ontologies. Although CIDOC-CRM does indeed have its rdfs:comments only in English, comments are ignored in parsing, whereas annotations and labels remain “firstclass citizens” of the ontology, and its rdfs:labels were observed to provide near-complete coverage in English, French, German, Greek, Portuguese, Russian and Chinese, with opaque identifiers as well; hence, the classification as being a language-independent ontology. In RADLEX, English and German are the primary languages, with opaque identifiers. The ontology has its own annotation properties: Preferred_name and Preferred_name_German, with both observed to be used in equal proportion. The ontology also has definitions in English only, but as these are imported from Medical Subject Headings (MeSH), the ontology has been classiefid as a language-independent ontology.
COVIDCRFRAPID, ONTOLURGENCES and UBERON are three primary language ontologies. In COVIDCRFRAPID, in addition to English rdfs:labels, the occasional use of Brazilian Portuguese was also observed. The ontology also made use of skos:prefLabel, but only English was indicated. An OWL fragment showing multilingual annotations for COVIDCRFRAPID is shown in Listing 1. 1 Prefix (:= < http :// purl . org / vodan / whocovid19crfsemdatamodel / >) 2 3 AnnotationAssertion ( rdfs : label : Chest_pain " Chest pain "@ en ) 4 AnnotationAssertion ( rdfs : label : Chest_pain " Dor no peito "@pt - br ) 5 AnnotationAssertion ( skos : prefLabel : Chest_pain " Chest pain "@ en ) Listing 1. Example multilingual annotations from COVIDCRFRAPID
In ONTOLURGENCES, French is the primary language, with descriptive identifiers, also in French. The ontology made use of SKOS’ prefLabel, hiddenLabel and definition properties, with a limited number of English prefLabels observed. In UBERON, English is the primary language, with German, Spanish and French used as well. The ontology has used the OboInOwl metamodel to define SKOS-like relations for classes, such as: oboInOwl:hasExactSynonym and oboInOwl:hasRelatedSynonym. It was observed that some exact/related synonyms are provided in other languages, but not all equally. Very few of the English annotations were language-tagged.
We expand on the metrics for multilingual ontologies identified, shown in Table 2. The total number of classes, properties and named individuals for an ontology, prior to any imports, is shown in the ‘Elements’ column. LLC lang is then applied to each applicable natural language, where its coverage is indicated as a percentage relative to the total number of elements for that ontology. Using UBERON as an example, coverage is so low for German, Spanish, and French, that it hardly qualifies as a multilingual ontology.
Size of the ontology varied significantly per multilingual ontology. For
example, CIDOC-CRM required only 372 annotations, but RADLEX required
45 852. RADLEX required significantly more efort to achieve complete
coverage to that of CIDOC-CRM. For some smaller ontologies, such as LABO
and PDRO, where there was incomplete coverage per natural language, the use
of multilingualism was limited at best. Seven of the eleven multilingual
ontologies were identified to have relations with OboInOwl [
Only ‘production’ ontologies were considered, but if ontologies with alternative statuses were also analysed, it is possible that the newer ontologies put into production in the past months may have used an alternative modelling option. The evidence indicates that there has been no uptake of any linguistic model for OWL ontologies. Likewise for mapping models, there was no use of SKOS for multilingual alignments. Due to insuficient data, we were unable to determine why this is the case from the analysis conducted. A possible reason may be insuficient tooling—there are limited tools to support the development, maintenance, and coordination of multilingual ontologies or multiple monolingual ontologies. It is unclear if the low number of multilingual ontologies is a bias in BioPortal or the resources themselves. It might be the case that there is limited need for multilingualism depending on the subject domain, such as a higher relevance in healthcare and administration rather than for biomedical research. We are currently finalising a more comprehensive review of multilingualism in ontologies, including ontologies from other subject domains. Its outcomes will be used in other ontologies and tools, such as the MoReNL project6, whose foundations assist in the development of ontology-driven multilingual tools. 4
The assessment of use of multilingualism in production-level BioPortal ontologies revealed a low adoption rate. The 11 out of the 220 that were multilingual, predominantly used opaque identifiers with multilingual labels. The ratio of multilingual OWL ontologies to monolingual ones was thus identified to be 5%, however coverage of multilingual labels did not exceed 50% for 3 of these ontologies. There was also no adoption of any linguistic model. The OWL profile was identified to be primarily OWL 2 DL, but some ontologies would require minor patching of the annotations to achieve this. Of interest here is that these annotation profile violations were not identified when developing the ontology, so a better mechanism may be needed to address this.
For the modelling options, both multilingual labels and linguistic models present areas of opportunity for future research. Of the ontologies analysed, the reuse of OBO-related ontologies was observed to be substantive, and this presents the opportunity to represent the shared lexicons for each using a linguistic model. Acknowledgements This work was financially supported by Hasso Plattner Institute for Digital Engineering through the HPI Research School at UCT [FGW] and the National Research Foundation (NRF) of South Africa (Grant Number 120852) [CMK].