Standard taxonomies that are meant to serve as reference specifications of specific scientific domains are developed through extensive review by scientists and experts. For data integration and interoperability needs within Knowledge Graphs (KGs), these taxonomies must be translated into formal ontologies. In this paper we present an ontology design pattern for modeling a scientific taxonomy as an ontology. The focus of the pattern is to 1) capture temporal dynamics of concepts as taxonomies evolve, 2) model the provenance of concepts to add context and enable governance, 3) assist the translation of taxonomic relations to ontological relations appropriately that will empower their use within KGs, and 4) tag provenance and other metadata information to mappings or alignments of uncertainty between concepts in diferent ontologies.
Language-OWL) for 1) both human and machine interpretability of their precise semantics; 2) modeling taxonomic relationships between concepts, such as subsumption, partonomy, membership, hypernymy; 3) creating mappings between concepts across diferent taxonomies; and 4) a more meaningful way to capture, store, navigate, discover, retrieve, and archive information in Knowledge Graphs (KGs).
The framework of standard taxonomies in their informal documents may vary for diferent domains - e.g., for hazards, one could say the metadata framework should capture cause and efect, spatial and temporal characteristics, co-occurring and cascading events, available proactive and recovery measures. But we found that overall, certain discourse elements overlap. Many of these taxonomies have at least one if not several definitions for every concept, each attributed from some reference sources or organizations, along with related and synonymous names. Other attributes that are common include scientific descriptions, units of measurement, vernacular names. Capturing all the rich information from the text document into the ontology is necessary for many purposes e.g., for ontology alignment tasks, where we need as much lexical information as possible for better matching. However, for many reasons (lack of time, and knowledge about ontology engineering, etc.), taxonomy-to-ontology formalization eforts mostly represent all the descriptive semantics either using Resource Description Framework (RDF) or Simple Knowledge Organization System (SKOS) annotation properties such as rdfs:comment, rdfs:label, skos:definition in a haphazard manner, and the structural semantics are altogether confused or neglected. Here below we highlight four characteristics of scientific taxonomies that have influenced the construction of the design pattern presented in this paper. • Science evolves, and as such scientific concepts change over time. Therefore, the temporal dynamics of corresponding taxonomies (definitions, names, addition/removal of concepts) have to be captured in the ontology to assimilate new scientific information and model revisions. • Scientific concepts have canonical names, but may also have many vernacular names in diferent contexts (e.g., based on language, geographic region, organization). Moreover, the definitions of these concepts are prescribed by subject matter experts within their respective domains (or context). Extending the contextual attributes of names and definitions using provenance, versus simply using annotation “labels" provides a richer description of a concept’s profile. It could also enable governance to ensure reference data quality. • Translating a scientific taxonomy into an ontology requires a well thought out approach that addresses not only how to designate classes and properties, but more importantly, how to translate taxonomic relations to structurally and semantically accurate ontological relations. This will enable the ontology to be eficiently used for access, navigation, and retrieval in a KG. • Scientific taxonomies are essentially used for data representation and integration purposes, and therefore may have to be aligned to other analogous or related reference or dataset taxonomies. In this situation, where there is uncertainty in alignment, additional metadata information regarding the alignment (e.g., mapping relation, score, matching technique, matcher) will need to be modeled.
Standards such as SKOS [
Simple Knowledge Organization System (SKOS) [
The skos:Concept class is meant to denote any abstract entity such as an idea, an object, an event. Each unique concept in this class may have several synonymous or vernacular name(s) that can be denoted via their labels. The three annotation properties for representing concept labels are: skos:prefLabel to assign an authoritative name, skos:altLabel for unauthorized name(s) and synonyms, and skos:hiddenLabel for names to be hidden from text-search and visual interfaces. The class skos:Collection (disjoint with skos:Concept) is used to describe labeled or ordered groups of SKOS concepts. The object property skos:member is used to define concept members of a collection.
The object property skos:semanticRelation is used when there is an inherent meaning between two concepts. The two transitive (skos:broaderTransitive, skos:narrowTransitive) and two non-transitive (skos:broader, skos:narrower) sub-properties of skos:semanticRelation are intended to model hypernym-hyponym relations, i.e., one concept is broader or narrower in
scope with another. The object property skos:mappingRelation is used to state mapping (or alignment) between concepts existing in diferent ontologies. Its five sub-properties include skos:closeMatch (symmetric), skos:exactMatch (symmetric and transitive), skos:broadMatch, skos:narrowmatch, and skos:relatedMatch (symmetric).
The structure of a taxonomy can consist of many levels and sub-levels, each representing a specific category of information. Here we discuss some of the common taxonomy organization structures.
A flat taxonomy, also known as an unlayered taxonomy, is simply a list of concepts, without a top-concept. An example of such a taxonomy is the list of hazards in the National Oceanic and Atmospheric Administration (NOAA) Storm Events Database2.
A hierarchical taxonomy-see Fig. 2(a) is represented as a tree, where individual concepts are arranged in a hierarchy. A hierarchy is often thought of as subsumption, but in meta-modeling they indicate many kinds of semantic refinements. An example is the biological hierarchy ranking3, i.e., Domain > Kingdom > Family > Species.
A poly-hierarchical taxonomy-see Fig. 2(b), represents hierarchies having one-to-many
childparent relationships. While constructing a formalization for such a taxonomy, one should not
explicitly construct OWL subsumption relationships as a poly-hierarchy, but it is fine to infer
concepts into a poly-hierarchy [
A faceted taxonomy allows a concept to be classified in multiple ways (sets of attributes), enabling the classification to be ordered in multiple ways, rather than in a single, predetermined
order (as in a strict hierarchy). Each facet is a unique way of characterizing concepts, and therefore represents a collection of concepts. Concepts in a faceted taxonomy are either 1) classified by every facet, or 2) classified across diferent facets and are then related through specific relations, such as partonomy, membership, hypernymy. A faceted taxonomy allows users to locate and discover concepts based on the facets that are important to them, and therefore it does not necessarily have one top concept.
Finally, taxonomies can also be a complex combination, i.e., hierarchical and faceted. The
SKOS models favors the formal representation of taxonomies that 1) have a hierarchical tree
structure, where the skos:hasTopConcept relation can be used to represent the top-most entry
point of a hierarchy; 2) have concepts related through subsumption, hyponym-hypernym,
member-collection, or relatedness relations. But scienticfi taxonomies may interrelate concepts
using relations beyond the scope of SKOS. For example, the relation between
StrongWindTornado denotes functional-parthood [
The Scientific Taxonomy Pattern (STP) as shown in Fig. 4 and as we will describe below is primarily meant to assist scientific taxonomy-to-ontology translation tasks. It is therefore restricted, but meant to be useful for scientific purposes rather than being all-encompassing by using fuzzy names and vague reification.
Concept. The Concept class is intended to denote abstractions of entities, ideas, events, or processes in scientific domains. They have clear, referenced, intentional, and extensional semantics [10]–in the form of definitions and descriptions–that are obtained through scientific or scholarly work. As such, they are identified as the nodes in any taxonomy. Examples of the Concept class are Hazard, Hurricane, and Disease.
ConceptCollection. The ConceptCollection class is used to denote a “bag of concepts”. When a set of concepts have certain similar characteristics, they are organized into a category in a taxonomy, and such a node represents a ConceptCollection. For example, SpecificHazard is a ConceptCollection that represents the set of all individual hazards (Tornado, TropicalStorm etc.). SpecificHazard is also a Concept with specific hazard properties or metrics such as spatial and temporal characteristics, and human impacts. Thus, an entity that is a ConceptCollection is also simultaneously a Concept (Ax:1) having a prescribed name with characteristic attributes4. 4This interpretation of a collection is slightly diferent from SKOS collections, which are intended to only represent groupings of “closely-related” concepts. skos:Collection and skos:Concept are therefore disjoint classes. Membership of a Concept within a ConceptCollection is made using the isMemberOf and its inverse hasMember relations (Ax:2-5). Finally, Concept is also an EntityWithProvenance and an EntityWithTemporalScope (Ax:6-7).
∃ hasMember.Concept ⊑ ConceptCollection
isMemberOf ≡ hasMember—
OrganizationScheme. The OrganizationScheme class is meant to capture the taxonomic or organization structure of a ConceptCollection - represented through the organizedBy relation (Ax:8-10). The structures by which concepts and concept collections can be organized are subsumption-hierarchy, faceted-hierarchy, or a complex combination. Ideally each ontology developed from a taxonomy must have a “root” ConceptCollection entity that will indicate the source scheme from which the descendant concepts and concept collections are derived. This root entity is also the primary form of access into the ontology within a KG. A ConceptCollection may also be representative of a facet, which is a way of organizing a set of concepts (or instances) based on a perspective or purpose that makes more sense to a human or system. In that sense every Concept or ConceptCollection is a facet. When the root ConceptCollection of an ontology is organized according to a faceted organization scheme, each of its direct descendant or member concept collections will now have an organization scheme of its own. Fig. 5 shows the faceted organization scheme for the UNDRR’s HIP taxonomy. The root ConceptCollection class is organized based on faceted-classification scheme, where the members at the next level are the three facets of ConceptCollection classes namely HazardType, HazardCluster, and SpecificHazard that are also organized following a faceted structure.
∃ organizedBy.OrganizationScheme ⊑ ConceptCollection
ConceptCollection ⊑ =1 organizedBy.OrganizationScheme
semanticRelation. The general semanticRelation is included here as a placeholder for any
organizational structure specific-relations between concepts (Ax:11-12). Such relations could
be explicit subsumption relations from RDF, semantic refinement (i.e., hyponym/hypernym)
relations from SKOS, variations of partonomic or membership relations such as those mentioned
in [
Concept ⊑ ∀ semanticRelation.Concept ∃ semanticRelation.Concept ⊑ Concept MappingRelation. Ontology alignment tasks determine mappings between terms in two ontologies. Automated alignment between two ontologies that are constructed from taxonomies rely mostly on their lexical information (concept names, labels, definitions, synonyms, descriptions, etc). Due to lexical ambiguity, the risk of matching two unrelated concepts increases, which may afect precision. Moreover the mappings generated by most alignment techniques represent is-a (i.e., subsumption) or equivalence relations by default. Our experience has shown that the generated mappings may also indirectly indicate other specific semantic relations that could be derived using machine learning approaches. To reliably represent the alignments between ontologies, we need to capture qualitative mappings of varying degrees of complexity, and corresponding quantitative mapping (or similarity) measures. The MappingRelation class, a reification of the mapsOnto property is introduced for this purpose i.e., to attach provenance, to denote diferent kinds of semantic mappings, and to capture uncertainty measure of alignment (Ax:13-17). We note in Ax:16 that the inverse filler of hasMappingRelation must exist and must be a Concept.
∃ definedAs.Concept ⊑ Definition
∃ mapsOnto.MappingRelation ⊑ Concept
MappingRelation ⊑ ∃ hasMappingRelation—.Concept
Definition. We specify Definition as the description of a scientific concept having some level of consensus amongst scientists or experts in the domain. They are obtained from an authoritative source such as international or government agencies, or academic, or other scientific sources with literature reference(s). As reducing ambiguity is a goal of standard taxonomies, there is a need for ontology patterns that can reconcile established meanings in diferent sub-disciplines. For example, the definition of an EnvironmentalDisaster in disaster management tends to focus on causes, whereas from the perspective of climate change or sustainability the focus is on efects [ 11]. This is why standard taxonomies emphasize or mention alternative definitions to capture a better conceptual understanding.
On the other hand, conceptual landscape of domains change over time resulting in taxonomies
that evolve to allow for more precision, representation, coverage, and better consensus. These
changes may be reflected in the form of the addition/replacement/removal of definitions and
concepts, or updated classification structure. For all these reasons, it is necessary that the
formal representations are capable of capturing temporal trends of the taxonomy, but also
empower their functionality as reference specifications by ascribing provenance to concept
names, definitions, and descriptions.
(13)
(14)
(15)
(16)
(17)
(18)
(19)
(20)
(21)
EntityWithProvenance. This concept is left as a placeholder for assigning provenance.
Detailed modeling of provenance is already done in W3C standard, the PROV Ontology [
EntityWithTemporalScope. This concept is left as a placeholder for capturing temporal
information. Detailed modeling of temporal concepts is already done in the OWL-Time ontology
[
Other Metadata Attributes.
UNDRR’s HIP taxonomy [
The template used in this taxonomy is developed both from a scientific perspective (i.e., reviewed by experts for robustness and scientific consensus), and user-driven perspective (i.e., acknowledging the diferent contexts and useful ways by which hazards can be organized). As a result of the latter approach, the resulting HIP taxonomy is not a single hierarchy of concepts that can be organized into a neat subsumption hierarchy, but rather is multi-dimensional (or multi-faceted), with concepts across three diferent facets semantically and meaningfully interlinked–see facets in Fig. 5. In addition to structural information, each HIP captures the following metadata information: hazard name, hazard reference number, definition(s) with reference, synonyms, scientific description, globally used metrics and numeric limits, references to key relevant UN conventions or multilateral treaties, examples, key references to support facts and statements made for each hazard, coordination agency or organisation that provided technical guidance on the hazard.
Fig. 6 shows a subset of the HIP for Extra-TropicalStorm that populates a portion of the pattern, emphasizing how concepts may be connected by reusing relations from SKOS, PROV, and OWL-Time.
Modeling scientific taxonomies as formal ontologies is crucial in their utilization for data integration, discovery, and exploration purposes within KGs. Scientific taxonomies are inherently (24) (25) (26) (27) c ; e R t R a c D i d N n U i A b x i a T T
L r O f I s , t e c V n O o R c P l , ra S e O n K e S g .
e a t n a
c diferent from non-scientific taxonomies in striving for clarity, consistency, and universality in the use of terminologies, and hence there is a need to appropriate model in their ontology evolving conceptual trends, provenance of the standardized semantics, meaningful structural relations, and other scientific metadata. The most commonly used SKOS standard is too informal to optimally represent all these aspects, but can be expanded and used in combination with other W3 standards for more precise representations of domain taxonomies as is desirable in scientific inquiry. As such, we have developed the Scientific Taxonomy Pattern (STP) to augment taxonomy-to-ontology translation tasks by demonstrating how to 1) model temporal dynamics of concepts, their definitions, names, 2) enable semantic governance of reference concepts and associated information, 3) translate taxonomic hierarchical relations to useful ontology relations that extend the scope of simple subsumption and SKOS, 4) capture provenance and mapping metadata for alignment with uncertainty. Additionally, we have demonstrated using an example, how the pattern can be used to develop an ontology for the UNDRR’s HIP taxonomy by deferring to existing W3C standards such as SKOS, PROV, and OWL-Time.
Future work will focus on expanding the semanticRelation property for more refined modeling of structural relations that will be useful for domain modeling needs.
Acknowledgement. This material is based upon work supported by the National Science Foundation under Grant No. 2033521: “KnowWhereGraph: Enriching and Linking Cross-Domain Knowledge Graphs using Spatially-Explicit AI Technologies”. Any opinions expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.
[9] B. Rehder, Categorization as causal reasoning, Cognitive Science 27 (2003) 709–748. [10] W. G. Stock, Concepts and semantic relations in information science, Journal of the
American Society for Information Science and Technology 61 (2010) 1951–1969. [11] P. R. Mulvihill, The ambiguity of environmental disasters, Journal of environmental studies and sciences 11 (2021) 1–5.