A Reference Guide to Ontology Best Practices

A Single-stop Assembly of Ontology Tips and Pointers

As we conclude this recent series on ontology tools and building [1], one item stands clear: the relative lack of guidance on how one actually builds and maintains these beasties. While there is much of a theoretic basis in the literature and on the Web, and much of methodologies and algorithms, there is surprisingly little on how one actually goes about creating an ontology.

An earlier posting pointed to the now classic Ontology Development 101 article as a good starting point [2]. Another really excellent starting point is the Protégé 4 user manual [3]. Though it is obviously geared to the Protégé tool and its interface, it also is an instructive tutorial on general ontology (OWL) topics and constructs. I highly recommend printing it out and reading it in full.

Learning by Example

Another way to learn more about ontology construction is to inspect some existing ontologies. Though one may use a variety of specialty search engines and Google to find ontologies [4], there are actually three curated services that are more useful and which I recommend.

The best, by far, is the repository created by the University of Manchester for the now-completed TONES project [5]. TONES has access to some 200+ vetted ontologies, plus a search and filtering facility that helps much in finding specific OWL constructs. It is a bit difficult to filter by OWL 2-compliant only ontologies (except for OWL 2 EL), but other than that, the access and use of the repository is very helpful. Another useful aspect is that the system is driven by the OWL API, a central feature that we recommended in the prior tools landscape posting. From a learning standpoint this site is helpful because you can filter by vocabulary.

An older, but similar, repository is OntoSelect. It is difficult to gauge how current this site is, but it nonetheless provides useful and filtered access to OWL ontologies as well.

These sources provide access to complete ontologies. Another way to learn about ontology construction is from a bottom-up perspective. In this regard, the Ontology Design Patterns (ODP) wiki is the definitive source [6]. This is certainly a more advanced resource, since its premise begins from the standpoint of modeling issues and patterns to address them, but the site is also backed by an active community and curated by leading academics. Besides ontology building patterns, ODP also has a listing of exemplary ontologies (though without the structural search and selection features of the sources above). ODP is not likely the first place to turn to and does not give “big picture” guidance, but it also should be a bookmarked reference once you begin real ontology development.

It is useful to start with fully constructed ontologies to begin to appreciate the scope involved with them. But, of course, how one gets to a full ontology is the real purpose of this post. For that purpose, let’s now turn our attention to general and then more specific best practices.

Sources of Best Practices

As noted above, there is a relative paucity of guidance or best practices regarding ontologies, their construction and their maintenance. However, that being said, there are some sources whereby guidance can be obtained.

To my knowledge, the most empirical listing of best practices comes from Simperl and Tempich [7]. In that 2006 paper they examined 34 ontology building efforts and commented on cost, effectiveness and methodology needs. It provides an organized listing of observed best practices, though much is also oriented to methodology. I think the items are still relevant, though they are now four to five years old. The paper also contains a good reference list.

Various collective ontology efforts also provide listings of principles or such, which also can be a source for general guidance. The OBO (The Open Biological and Biomedical Ontologies) effort, for example, provides a listing of principles to which its constituent ontologies should adhere [8]. As guidance to what it considers an exemplary ontology, the ODP effort also has a useful organized listing of criteria or guidance.

One common guidance is to re-use existing ontologies and vocabularies as much as possible. This is a major emphasis of the OBO effort [9]. The NeOn methodology also suggests guidelines for building individual ontologies by re-use and re-engineering of other domain ontologies or knowledge resources [10]. Brian Sletten (among a slate of emerging projects) has also pointed to the use of the Simple Knowledge Organization System (SKOS) as a staging vocabulary to represent concept schema like thesauri, taxonomies, controlled vocabularies, and subject headers [11].

The Protégé manual [3] is also a source of good tips, especially with regard to naming conventions and the use of the editor. Lastly, the major source for the best practices below comes from Structured Dynamics‘ own internal documentation, now permanently archived. We are pleased to now consolidate this information in one place and to make it public.

General Best Practices

General best practices refer to how the ontology is scoped, designed and constructed. Note the governing perspective in this series has been on lightweight, domain ontologies.

Scope and Content

• Provide balanced coverage of the subject domain. The breadth and depth of the coverage in the ontology should be roughly equivalent across its scope
• Reuse structure and vocabularies as much as possible. This best practice refers to leveraging non-ontological content such as existing relational database schema, taxonomies, controlled vocabularies, MDM directories, industry specifications, and spreadsheets and informal lists. Practitioners within domains have been looking at the questions of relationships, structure, language and meaning for decades. Effort has already been expended to codify many of these understandings. Good practice therefore leverages these existing structural and vocabulary assets (of any nature), and relies on known design patterns
• Embed the domain coverage into a proper context. A major strength of ontologies is their potential ability to interoperate with other ontologies. Re-using existing and well-accepted vocabularies and including concepts in the subject ontology that aid such connections is good practice. The ontology should also have sufficient reference structure for guiding the assignment of what content “is about”
• Define clear predicates (also known as properties, relationships, attributes, edges or slots), including a precise definition. Then, when relating two things to one another, use care in actually assigning these properties. Initially, assignments should start with a logical taxonomic or categorization structure and expand from there into more nuanced predicates
• Ensure the relationships in the ontology are coherent. The essence of coherence is that it is a state of logical, consistent connections, a logical framework for integrating diverse elements in an intelligent way. So while context supplies a reference structure, coherence means that the structure makes sense. Is the hip bone connected to the thigh bone, or is the skeleton askew? Testing (see below) is a major aspect for meeting this best practice
• Map to external ontologies to increase the likelihood of sharing and interoperability. In Structured Dynamics’ case, we also attempt to map at minimum to the UMBEL subject reference structure for this purpose [13]
• Rely upon a set of core ontologies for external re-use purposes; Structured Dynamics tends to rely on a set of primary and secondary standard ontologies [14]. The corollary to this best practice is don’t link indiscriminantly.

Structure and Design

• Begin with a lightweight, domain ontology [15]. Ontologies built for the pragmatic purposes of setting context and aiding disparate data to interoperate tend to be lightweight with only a few predicates, such as isAbout, narrowerThan or broaderThan. But, if done properly, these lighter weight ontologies with more limited objectives can be surprisingly powerful in discovering connections and relationships. Moreover, they are a logical and doable intermediate step on the path to more demanding semantic analysis. Because we have this perspective, we also tend to rely heavily on the SKOS vocabulary for many of our ontology constructs [16]
• Try to structurally split domain concepts from instance records. Concepts represent the nodes within the structure of the ontology (also known as classes, subject concepts or the TBox). Instances represent the data that populates that structure (also known as named entities, individuals or the ABox) [17]. Trying to keep the ABox and TBox separate enables easier maintenance, better understandability of the ontology, and better scalability and incorporation of new data repositories
• Treat many concepts via “punning” as both classes and instances (that is, as either sets or members, depending on context). The “punning” technique enables “metamodeling,” such as treating something via its IRI as a set of members (such as Mammal being a set of all mammals) or as an instance (such as Endangered Mammal) when it is the object of a different contextual assertion. Use of “metamodeling” is often helpful to describe the overall conceptual structure of a domain. See endnote [18] for more discussion on this topic
• Build ontologies incrementally. Because good ontologies embrace the open world approach [19], working toward these desired end states can also be incremental. Thus, in the face of common budget or deadline constraints, it is possible initially to scope domains as smaller or to provide less coverage in depth or to use a small set of predicates, all the while still achieving productive use of the ontology. Then, over time, the scope can be expanded incrementally. Much value can be realized by starting small, being simple, and emphasizing the pragmatic. It is OK to make those connections that are doable and defensible today, while delaying until later the full scope of semantic complexities associated with complete data alignment
• Build modular ontologies that split your domain and problem space into logical clusters. Good ontology design, especially for larger projects, warrants a degree of modularity. An architecture of multiple ontologies often works together to isolate different work tasks so as to aid better ontology management. Also, try to use a core set of primitives to build up more complex parts. This is a kind of reuse within the same ontology, as opposed to reusing external ontologies and patterns. The corollary to this is: the same concepts are not created independently multiple times in different places in the ontology. Adhering to both of these practices tends to make ontology development akin to object-oriented programming
• Assign domains and ranges to your properties. Domains apply to the subject (the left hand side of a triple); ranges to the object (the right hand side of the triple). Domains and ranges should not be understood as actual constraints, but as axioms to be used by reasoners. In general, domain for a property is the range for its inverse and the range for a property is the domain of its inverse. Use of domains and ranges will assist testing (see below) and help ensure the coherency of your ontology
• Assign property restrictions, but do so sparingly and judiciously [20]. Use of property restrictions will assist testing (see below) and help ensure the coherency of your ontology
• Use disjoint classes to separate classes from one another where the logic makes sense and dictates (if not explicitly stated, they are assumed to overlap)
• Write the ontology in a machine-processable language such as OWL or RDF Schema (among others), and
• Aggressively use annotation properties (see next) to promote the usefulness and human readability of the ontology.

Naming and Vocabulary Best Practices

• Name all concepts as single nouns. Use CamelCase notation for these classes (that is, class names should start with a capital letter and not contain any spaces, such as MyNewConcept)
• Name all properties as verb senses (so that triples may be actually read); e.g., hasProperty. Try to use mixedCase notation for naming these predicates (that is, begin with lower case but still capitalize thereafter and don’t use spaces)
• Try to use common and descriptive prefixes and suffixes for related properties or classes (while they are just labels and their names have no inherent semantic meaning, it is still a useful way for humans to cluster and understand your vocabularies). For examples, properties about languages or tools might contain suffixes such as ‘Language‘ or ‘Tool‘ for all related properties
• Provide inverse properties where it makes sense, and adjust the verb senses in the predicates to accommodate. For example, <Father> <hasChild> <Janie> would be expressed inversely as <Janie> <isChildOf> <Father>
• Give all concepts and properties a definition. The matching and alignment of things is done on the basis of concepts (not simply labels) which means each concept must be defined [21]. Providing clear definitions (along with the coherency of its structure) gives an ontology its semantics. Remember not to confuse the label for a concept with its meaning. (This approach also aids multi-linguality). In its own ontologies, Structured Dynamics uses the property of skos:definition, though others such as rdfs:comment or dc:description are also commonly used
• Provide a preferred label annotation property that is used for human readable purposes and in user interfaces. For this purpose, Structured Dynamics uses the property of skos:prefLabel
• Include a class “SemSet”, which means a series of alternate labels and terms to describe the concept. These alternatives include true synonyms, but may also be more expansive and include jargon, slang, acronyms or alternative terms that usage suggests refers to the same concept. The umbel:SemSet construct enables a listing of individual members to be generated that provides the matching set for tagging and information extraction tasks. (As such, also include the prefLabel in the SemSet for proper lookup and tagging purposes.) The SemSet construct is similar to the “synsets” in Wordnet, but with a broader use understanding. This construct is an integral part of Structured Dynamics’ approach to using ontologies for information extraction and tagging of unstructured text
• Try to assign logical and short names to namespaces used for your vocabularies, such as foaf:XXX, umbel:XXX or skos:XXX, with a maximimum of five letters preferred
• Enable multi-lingual capabilities in all definitions and labels. This is a rather complicated best practice in its own right. For the time being, it means being attentive to the xml:lang=”en” (for English, in this case) property for all annotation properties
• (If you disagree with these naming conventions, use your own, but in any event, be consistent!!).

Documentation Best Practices

• Like good software programs, a properly constructed and commented ontology is the first requirement of best practice documentation
• The entire ontology vocabulary should be documented via a dedicated system that allows finding, selecting and editing of any and all ontology terms and their properties
• The methodologies should be documented for ontology construction and maintenance, including naming, selection, completeness and other criteria. Documents such as this one and others in this series provide examples of important supplementary documentation regarding methodology and practice
• Provide a complete TechWiki-like documentation system for use cases, best practices, evaluation and testing metrics, tools installation and use, and all aspects of the ontology lifecycle should be provided and supported [22]
• Develop a complete graph of the ontology and make it available via graph visualization tools to aid understanding of the ontology in its complete aspect [23], and
• Other ample diagrams and flowcharts should also be prepared and made available for knowledge workers’ use. UML diagrams, for example, might be included here, but general workflows and concept relationships should be explicated in any case through visual means. Such diagrams are much easier to understand and follow than the actual ontology specification.

Organizational and Collaborative Best Practices

• Collaboration is an implementation best practice [24]
• Re-use of already agreed-up structures and vocabularies respects prior investments and needs to be emphasized
• Improved processes for consensus making, including tools support, must be found to enable work groups to identify and decide upon terminology, definitions, alternative labels (SemSets), and relations between concepts. These processes need not be at the formal ontology level, but at the level of the concept graph underlying the ontology [24].

Testing Best Practices

• Test new concepts, aided by proper domain, range and property restrictions; by invoking reasoners such that inconsistencies can be determined [25]
• Test new properties, aided by invoking reasoners, which will identify inconsistencies [25]
• Test via external class assignments, by linking to classes in external ontologies, which acts to ‘explode the domain’ [26]
• Use external knowledge bases and ontologies, such as Cyc or UMBEL [27], to conduct coherency testing for the basic structure and relationships in the ontology
• Evolve the ontology specification to include necessary and sufficient conditions [25] aid more complete reasoner testing for consistency and coherence.

Best Practices for Adaptive Ontologies

In the case of ontology-driven applications using adaptive ontologies [28], there are also additional instructions contained in the system (via administrative ontologies) that tell the system which types of widgets need to be invoked for different data types and attributes. This is different from the standard conceptual schema, but is nonetheless essential to how such applications are designed.

• Use the structWSF middleware layer [29] as the abstract access point to:
• • To create, update, delete or otherwise manage data records
  • To browse or view existing records or record sets, based on simple to possible complex selection or filtering criteria, or
  • To take one of these results sets and progress it through various workflows involving specialized analysis, applications, or visualization.
• Supplement the domain ontology with a semantic component ontology for the purposes of guiding data widget display and visualization [30], and
• Supplement the domain ontology with the irON (instance record Object Notation) for dataset exchange and interoperability [31].

The administrative ontologies supporting these applications are managed differently than the standard domain ontologies that are the focus of most of the best practices above. Nonetheless, some of the domain ontology best practices work in tandem with them, the combination of which are called adaptive ontologies.