The ROC ontology for country responses to COVID-19 provides a model for collecting, linking and sharing data on the COVID-19 pandemic. It follows semantic standardization (W3C standards RDF, OWL, SPARQL) for the representation of concepts and creation of vocabularies. ROC focuses on country measures and enables the integration of data from heterogeneous data sources. The proposed ontology is intended to facilitate statistical analysis to study and evaluate the e ectiveness and side e ects of government responses to COVID-19 in di erent countries. The ontology contains data collected by OxCGRT from publicly available information. This data has been compiled from information provided by ECDC for most countries, as well as from various repositories used to collect data on COVID-19.
Almost all countries are a ected by the (second wave) COVID-19 pandemic. Many organisations have setup systems and projects to collect and publish global data on the impact of the virus allowing to monitor the pandemic. Most countries have implemented responsive measures as reaction to COVID-19. The impact of the pandemic and policies implemented di er widely among countries. Meanwhile data on the infection spread and government responses during the rst phase of the pandemic (March to July 2020) is available to enable research on the e ects of individual policies.
Policy makers face tough decisions on how to deal with the pandemic, since responses considered most e ective have also signi cant negative impacts on economy and social life. This is strongly re ected in the measures implemented by the countries. The policies implemented and the reactions of society vary greatly among countries and cities. While many people support their government strategy, others reject measures included. Empirical studies need to be conducted to support decision-makers and information campaigns. Which responses have proven to be e ective in containing COVID-19 and what are their e ects on the economy and society? are questions that apply to individual countries as well as to cultural, political and geographical regions.
To enable empirical research, global data on infections, recoveries, death rates and policies implemented in various countries and regions is published via various sources. The challenge for data scientists and epidemiologists is that these data sources do not share common standards or methodologies for reporting their data. The reported data is in uenced by time zones/holidays, di erent political and/or economic incentives, variation in counting methods1, discovered / undiscovered numbers, etc. In addition, in many cases country's reports differ depending on the reporting institute. For example, the numbers of infected people in Germany di ers between Johns Hopkins Coronavirus Resource Center2 (USA) and Robert Koch institute3 (GER). Unfortunately, this di erence is not comprehensible. To study this valuable information and perform statistical analysis on it, a common standard for harmonizing and reporting the data is required.
Many researchers are taking up this challenge, and the rst solutions to this
problem have already been developed. The COviD-19 ontology (CODO) is an
ontology for organizing cases data and patient information and aims to use the
technology of knowledge graphs to analyse the pandemic [
For the ontology development process, we followed a hybrid approach, which is driven by available data on the one hand and questions of interest that should be answerable by the ontology on the other hand. We mainly collected data from 1 cause and place of death -at home or in hospital2 https://coronavirus.jhu.edu/map.html 3 https://rki.de/DE/Content/InfAZ/N/Neuartiges Coronavirus/Fallzahlen.html OxCGRT4, ILO5 and ECDC6. ECDC provides data on infection rates, OxCGRT systematically monitors government actions related to the pandemic and ILO provides economic indicators focused on the labour market. The modeling of the data consisted of the following steps: manual review and merging of the data. Concepts were extracted and linked using logical relationships. To design and create the ontology, we used Protege and then ingested the data based on the ROC ontology.
The remainder of the paper is organized as follows: Section 2 contains a review on related work. Section 3 describes the development methodology. Section 4 introduces the ROC ontology and its concepts. Section 5 describes the data ingestion process and querying capabilities. Section 6 concludes the paper with a short summary and future work. 2
This section provides an overview on related work investigating ontologies related to diseases, especially relevant to COVID-19.
The Human Disease Ontology (DO)7 classi es thousands of human diseases
moving to a multi-editor model in web ontology language enabling collaboration
of several working groups, and has recently been extended by DOID8
including COVID-19 concepts [
One of the most relevant work is the Ontology for cases and patient
information (CODO)9, which initiated the development of CIDO. It provides a
standards-based and comprehensive open source model for data collection on
the COVID 19 pandemic. The ontology is very well suited for the integration
of data from heterogeneous data sources and thus represents one of the major
inspirations for our work. Furthermore, our methodology follows the procedure
for term de nition described in the work. This is based on the reuse of concepts
from other leading vocabularies and the use of the W3C standards RDF, OWL,
SWRL and SPARQL. The evaluation of CODO was conducted on the basis of
data received from the Indian government [
Of course, the ontologies (COVIDCRFRAPID)10 of the World Health
Organization (WHO) as data model for the case COVID-19 RAPID, the ontologies
Kg-COVID-1911 for the creation of knowledge graphs including SARS-COV-2
4 https://www.bsg.ox.ac.uk/
5 https://www.ilo.org/global/lang{en/index.htm
6 https://www.ecdc.europa.eu/
7 http://www.disease-ontology.org
8 https://disease-ontology.org/term/DOID:11725/
9 https://github.com/biswanathdutta/CODO
10 https://bioportal.bioontology.org/ontologies/COVIDCRF RAPID
11 https://github.com/Knowledge-Graph-Hub/kg-covid-19
and the ontology Linked-Data COVID-1912 are also relevant. These ontologies
contain conceptualizations (and partially instance data) related to COVID-19
cases and are primarily aimed at software applications such as question &
answering or monitoring dashboards [
All these works support the containment of the pandemic, but more importantly, the works build on and complement each other. The focus so far has been mainly on gathering cases, symptoms and information from infected people, for example to identify and isolate hot spots. However, one eld has not been covered so far, namely the eld of countries government responses. It is precisely this knowledge gap that our work with the ROC ontology aims to ll.
In the next sections we describe the methodology for the development of the ROC ontology based on the state-of-the-art principles, its structure and evaluation. 3
The artifact created and investigated in this work is an ontology authored using the Web Ontology Language 2 (OWL 2)13, the semantics of which are based on description logics.
The main purpose of the ontology, as pointed out in Section 1, is the integration of public data on national responses to the COVID-19 pandemic and to provide a layer of interoperability between di erent and diverse resources as well as to answer interesting questions from the data. The development of the ontology was therefore bottom-up data-driven as well as top-down application driven. A further requirement was the integration of existing ontologies in the domain of COVID-19 in order to avoid redundancy as well as to leverage knowledge gained from the integration of existing knowledge (such as the combination of data about national responses with case data).
A multitude of ontology development methods exist, comprising, but not
limited to METHONTOLOGY [
The selection of the ontology development method was driven by the nonfunctional requirements to (a) integrate existing data sources and (b) integrate existing ontological sources as well as (c) by content-related requirements, which are formulated in the form of competency questions.
We decided to use the NeOn methodology [
The ontology development process therefore consists of three main activities, which start in parallel and which are divided into subtasks, some of which interact (see Figure 1):
Requirements Analysis Data-driven development
CODO integration
T1.1-1.3 Purpose, scope, language, users
T1.4-1.7 Requirements /
CQs
T1.8 Terminology extraction
T2.1 Data gathering
T2.2 Conceptual abstraction
T2.3 Information exploration
T2.7-2.8 Formalization, implementation
ROC
T2.4-2.6 Transformation and
refinement CODO Alignment
T3.1-3.3 External statement selection
T3.4 External statement mapping
T3.5
Consistency check
Activities 2 and 3 resulted in the ROC ontology, while activity 3 resulted in a set of import and mapping axioms (which are also part of the ROC ontology). 4
The ontology14 was designed around a set of competency questions, for instance, one can ask the following: CQ1 Which countries do establish a certain response? CQ2 At which incidence level do individual countries establish certain responses / response levels? CQ3 How long do individual countries keep their response measures active? CQ4 Were countries which established responses at low incidence levels able to avoid high incidence rates? CQ5 Is there an e ect of certain responses on infection rates? Can we measure e ect or delay of e ect? 14 The current version 1.0 of the ROC ontology is publicly available at http://qurator-csi.de/ontologies/roc
The ontology consists of 27 OWL classes, 10 object properties, 42 data properties and 3 annotation properties. Its expressivity is ALE HI(D), i.e., AL with full existential quali cation, role hierarchy, inverse roles and data types, making it a member of the OWL 2 DL pro le. The central domain concepts of the ROC ontology are the indicators as de ned by the Oxford Covid-19 Government Response Tracker (OxCGRT) mentioned in Section 1.
The values for each of these indicators is acquired and stored in a data record. A record is represented in the ROC ontology as an instance of the class ResponseStatistics, which is a subclass of the CODO class CountryWiseStatistics (see Figure 2a).
In alignment with the structure of the data sources, which assign numerical values to each of these indicators, we modeled these indicators as data properties. We established a data property hierarchy re ecting the taxonomy of OxCGRT coding categories (containment and closure (C), economic response (E), health systems (H), and miscellaneous (M)) (see Figure 2b). Creating common super properties for each category of response values allows for a Description Logic reasoner to infer that if any of the response values in a certain category has a value, then the super data property (representing the whole category) has a value. (b) The hierarchy of data properties representing the OxCGRT response (a) A response statistics instance with acquired response values values 5
ROC in use This section describes how the ontology is used to ingest a set of data collected from multiple sources. It outlines the data transformation process and how to query the resulting RDF knowledge base (KB) to answer competency questions.
We manually reviewed and merged data collected from OxCGRT, ILO and
ECDC. We then transformed data coming from those di erent sources into RDF
based on the ROC ontology. The resulting KB serves as an integrated view to
answer queries spreading over all required sources. For data transformation, we
made use of the Karma integration tool [
We devised an initial set of queries to answer the competency questions over the RDF data. As an example, for the question "List the countries and their respective health responses?" the corresponding query would be the one depicted in Listing 1.
PREFIX roc : <http :// qurator - csi . de / ontologies / covid / responses #> PREFIX codo : <http :// www . isibang . ac . in / ns / codo #> SELECT ? country AVG (? testing_policy ) AVG (? contact_tracing ) SUM (? investment_healthcare ) SUM (? investment_in_vaccines ) AVG (? facial_coverings ) WHERE { ? country codo : countryWiseStatistics ? stats .
? stats roc : h2_testing_policy ? testing_policy ; 15 https://www.w3.org/TR/r2rml/ 16 https://virtuoso.openlinksw.com/ roc : h3_contact_tracing ? contact_tracing ; roc : h4_emergency_investment_in_healthcare ? investment_healthcare ; roc : h5_investment_in_vaccines ? investment_in_vaccines ; roc : h6_facial_coverings ? facial_coverings . } GROUP BY ? country Listing 1: SPARQL query for question:"List the countries and their respective health responses?"
The query results are depicted in Figure 4. We can see for instance that Germany has a higher testing policy index and the highest emergency investments in healthcare and vaccines. We can also note that Sweden did not implement a facial covering policy. Based on the correlation between countries stringency and actual statistics, the e ectiveness or failure of implemented measures can be derived, informing government future decisions. The present work focuses on country responses against COVID-19 and proposes a novel ontology ROC to enable the integration of data from heterogeneous data sources and answer interesting questions. This facilitates statistical analysis to investigate and evaluate the e ectiveness and side e ects of such responses. The ontology consists of 27 OWL classes, 10 object properties, 42 data properties and 3 annotation properties. The data collected by OxCGRT, ILO and ECDC were manually reviewed and merged. Then we converted data from these di erent sources into RDF based on the ROC ontology. The resulting RDF serves as an integrated view to answer queries that span all required sources. The resulting KB contains 1850 instances of data for Germany, Jordan and Sweden collected between January and early November 2020. We uploaded the RDF data to the Virtuoso Triple Store and made it accessible and queryable through a SPARQL endpoint.
Given the fact that most experts are not familiar with SPARQL nor with
Semantic Web technologies, we plan to connect our Controlled Natural Language
querying system [
Other factors could in uence the e ectiveness of country responses. These are either di cult to capture17 or relatively easy to determine18. Consideration of such factors would lead to the extension of the properties of the concepts/terms in order to increase the semantic expressiveness of the ontology.
Lastly, a technical and a goal-based evaluation of the e ectiveness of the approach is a challenge to be addressed by future work.
The research presented in this article is partially funded by the German Federal Ministry of Education and Research (BMBF) through the project QURATOR (Unternehmen Region, Wachstumskern, grant no. 03WKDA1F). http://qurator.ai