The Information Artifact Ontology (IAO) is an ontology developed to tag data about information entities, especially ones relevant to the biomedical domain. IAO includes the class information content entity (ICE), which is intended to capture all information entities. However, the direct subclasses (or 'children') of ICE (the 'parent' class) are distinguished in conflicting ways, which disposes extensions of IAO toward asserted multiple inheritance. This, in turn, breaks the well-known asserted single-inheritance rule of best practice for ontology building. We term the kinds of distinctions used to divide ICE into subclasses 'cross-cutting distinctions'. We offer a proposal for resolving IAO's use of cross-cutting distinctions in a way that retains the classes in IAO while adhering to asserted single inheritance and the principles of ontological realism.
The Information Artifact Ontology (IAO) was developed by the Open Biological and Biomedical
Ontology (OBO) Foundry [
One motivation for an information ontology is to accurately record and interpret patient healthcare
data. This challenge becomes especially salient when those data are shared between healthcare
providers at different organizations. Consider a general practitioner that sends records to a specialist to
whom they have referred a patient, or cases where a patient must be transferred from one hospital to
another. There is no guarantee (nor is it likely) that medical institutions collect, organize, label, or store
patient information in the same way. As a result, it may be unclear to a recipient how to understand
shared information. Detailed examples of this are discussed by Werner Ceusters and Barry Smith [
While one aim of an RTS and of ontologies more generally is to reduce the ambiguity in data,
ambiguity can still be introduced if the ontology allows classes to have more than one asserted parent
(called ‘asserted multiple inheritance’) because cross-cutting distinctions allow for lower-level classes
to reasonably be subsumed under more than one parent class. This breaks the principle of best practice
used in realist ontologies known as ‘asserted single inheritance’ [
IAO itself, it should be emphasized, does not suffer from asserted multiple inheritance. However, the use of cross-cutting distinctions within its upper-level class hierarchy means either that (i) asserted multiple inheritance will be needed to extend IAO, or (ii) IAO can be extended in multiple, contradictory ways. Both can lead to the extensions of IAO being incompatible.
Cross-cutting distinctions occur when entity types are subdivided in such a way that at least some
of the new subtypes can consistently be subclasses of two or more super-classes in the primitive class
hierarchy. We assume here that universals (genuine types in reality) do not crosscut. As a quick
example, consider that no genuine subtype of continuant will also be a genuine subtype of occurrent.
We assume the same is true for any genuine types and subtypes. Cross-cutting distinctions correspond
to violations of Rector’s second principle of ontology building, namely that “the principle of
specialization should be subsumption (as opposed to, for example partonomy) and should be based on
the same, or progressively narrower criteria, throughout” [
Consider a simple example. Suppose we have a food ontology including the class fruit, which we divide into four subclasses using two distinct criteria: citrus fruit and non-citrus fruit, on the one hand, and orange-colored fruit and non-orange-colored fruit, on the other. The difficulty is that apricots are orange non-citrus fruits, while limes are green citrus fruits. Thus, each has two categories under which they could fall. See Figure 1 below for an illustration of this point.
Contrast this with a division of fruit that keeps citrus fruit and non-citrus fruit, but that further divides these subclasses into orange citrus fruit, non-orange citrus fruit, orange non-citrus fruit, and non-orange non-citrus fruit. These divisions might seem even more gerrymandered, but they would follow Rector’s second principle above since they take a single criterion (e.g., citrus vs. noncitrus) and start to make it progressively narrower (e.g., orange citrus vs. non-orange citrus). This would lead to only one classification for apricots and only one for limes, preserving asserted single inheritance.
Now, IAO includes several subclasses of ICE, including document, document part, textual entity,
figure, and directive information entity, each of which is relevant to assert in certain contexts. It
would not be a problem to assert these classes, so long as the correct equivalency axioms and class
restrictions are associated with them, because this allows ontology reasoners to infer multiple
inheritance from the axioms and restrictions given. However, because different criteria are used to
identify the different subtypes of ICE (and so the distinctions crosscut), and because the required
equivalency axioms and restrictions are not always given, subclasses (and sub-subclasses, and so on)
further down the ICE branch might belong to more than one primitive child of ICE. This ultimately
hinders the ability to extend the ontology using asserted single inheritance [
IAO uses at least three different criteria to distinguish subtypes of ICE: (1) the content of an ICE; (2) the structure of an ICE; and (3) the purpose of an ICE. The issue is that a single type of ICE can have a structure indicating it should be classified as one sort of subtype of ICE but content that would indicate classifying it as another sort. A good example is the IAO class report.
IAO:Report =def. A document assembled by an author for the purpose of providing information for the
audience. A report is the output of a documenting process and has the objective to be consumed
by a specific audience. Topic of the report is on something that has completed. A report is not
a single figure. [
According to this definition, a report is a document. From IAO’s definition of ‘document’, it follows
that a report is also a “collection of information content entities” [
From within the framework of ontological realism, this indicates (though does not entail) that
document and narrative object are both universal classes, since defined classes (or ‘collections of
particulars’ to use Smith and Ceusters terminology [
A similar cross-cutting problem arises with directive ICEs. A directive ICE is any ICE that indicates to an agent which course of action to take (such as a recipe). Many directive ICEs are also documents. An email instructing someone to perform a certain task is a document and a directive ICE. Similarly, a to-do list for an employee would simultaneously be both a directive ICE and a document. Each of these would also be narrative objects since they are all composed of propositions (though we do not hold that all directive ICEs must be narrative objects). Many kinds of ICE besides document and narrative object are simultaneously documents and directive ICEs.
The problem is that making both directive ICE and document sibling classes under ICE without
making one a defined class requires either that (i) the ontology include asserted multiple inheritance, or
(ii) subclasses of directive ICE cannot also be subclasses of document or narrative object. The former
breaks the principle of asserted single inheritance, while the latter steers us away from ontological
realism since, as noted, it is clear that some directive ICEs are also documents and narrative objects.
We acknowledge that OWL ontologies are built under the open world assumption [
IAO:PlanSpecification =def. A directive information entity with action specifications and objective
specifications as parts that, when concretized, is realized in a process in which the bearer tries
to achieve the objectives by taking the actions specified. [
Thus, plan specification is a subclass of directive ICE. Supposing we want to extend plan specification with the class cooking recipe, we run into a problem. Instances of cooking recipe (we might assume) consist of any plan specification that explains how to cook some food item. This would include cookbooks, web pages on food preparation, and even old, hand-written family recipes. Each of these is a document, meaning that including the most amount of ontological information about these recipes requires cooking recipe to be a subclass of document. But this cannot be done within the current class hierarchy of IAO without violating asserted single inheritance. 3.3.
The IAO class document part is extended with a variety of subtypes. Many of these, however, also fit the definition for other sibling classes of document part. Once again, this can result in either asserted multiple inheritance or departure from ontological realism if the ontology is extended. First, consider the definition of the following subclass of document part.
IAO:MethodsSection = def. A part of a publication about an investigation that is about the study design
of the investigation. [
This class extends document part, but it could also reasonably extend the class textual entity because a methods section is a collection of markings and symbols intended to be understood as words. Thus, there are at least two potential parent classes of methods section. Again, this results from the lack of a uniform method for asserting subclasses of ICE.
Next, consider the definition of the following additional subclass of document part.
IAO:FiguresSection = def. A part of a document that contains one or more figures. [
An instance of figures section may itself consist of but a single figure. This means that instances of figures section might be identical to one or more figures, or they might be any part of a document that contains a figure, such as a methods section. Accordingly, defining figures section in this way could lead to asserted multiple inheritance in the ontology if figures section or methods section have asserted subclasses. This would not be a problem, if figures section were a defined class with associated logical restrictions, which would allow inferred multiple inheritance.
The prevalence of cross-cutting distinctions in IAO suggests that the difference between defined and
primitive classes is not properly acknowledged. Extending IAO is thereby prone to either a hierarchy
including asserted multiple inheritance or a departure from ontological realism. Abiding by ontological
realism means asserting classes based on the way the world actually is [
For example, it seems obvious at first blush that documents have parts (the document you are reading now is divided into many sections, after all), and so it makes intuitive sense to assert document part as a class in an information ontology. However, document parts do not seem to be genuinely different kinds of entities in virtue of being parts of documents. The same way that a human being does not become a different kind of thing when they acquire a new job with a new set of duties and responsibilities. The different parts of the documents are real, to be sure, but it does not follow that they are themselves genuinely different types of things. (Consider that the left half of your scalp is real but left half of scalp is a fiat part of an object and not a genuinely distinct class of entity.)
Differences in writing styles across fields entail that document parts, like introductions, can vary
widely in their role and content. For instance, a methods and a figures section, or a figures section and
an introduction, may overlap according to IAO depending on the content of that section. Additionally,
a single part of a document might have multiple purposes, which would need to be distinguished in the
ontology. Authors seem to write each document part for a particular purpose, and so we should
distinguish the various functions that document parts (or even whole documents) can have. This would
allow us to account for when a single part of a document subsumes the purposes of multiple extensions
of document part. This suggestion follows Rector’s second criterion of normalization, which uses the
same principle of specialization when building the primitive backbone hierarchy of an ontology [
Accordingly, we propose distinguishing ICEs based on the purpose of those ICEs. This principle of
differentiation facilitates constructing a consistent and effective information ontology since all ICEs are
the products of some intentional creation process. Applying this principle leads to three high-level direct
subclasses of ICE, which mirror those in another widely used information ontology: the Common
Core’s Information Entity Ontology (IEO) [
3. Directive ICE =def. a descriptive ICE that prescribes some action. 4.1.
We make directive ICE a child of descriptive ICE because all directive ICEs are necessarily
descriptive. We follow Barton and colleagues in maintaining that directive ICEs “represent how to
perform some actions, and such actions are then performed on the basis of this description” [
We admit that conceiving of all directive ICEs as containing descriptions sets a low bar for an information entity being descriptive. A note reading “please cut the grass” does not describe very much. It contains no description of grass, of cutting, nor of what it would mean to cut grass specifically as opposed to cutting a steak or paper. There is also no way to tell which grass is being referred to simply by what is written on the note – context is needed to understand what is meant.
However, there is another sense in which the desired action is described in the note. In virtue of
specifying that they want the grass cut, the parent has already implicitly described a process type. Using
a knife to cut steak, or a pair of scissors to cut paper, would not satisfy the description of grass cutting.
That the prescription in the note is not satisfied by cutting a steak implies that one’s steak cutting does
not match the description of the process in the note. Hence, the note is, in part, descriptive. Moreover,
as Barton and colleagues note [
Here, one might wonder why designative ICE was not made a parent of descriptive ICE. After all,
it seems as though all descriptive ICEs would need to designate the various entities they describe. This
is certainly plausible, but even if this is true, it does not indicate that descriptive ICEs are themselves
designative ICEs. It only indicates that descriptive ICEs contain designative ICEs as parts. We lack the
space to spell out exactly how this parthood relation works, but other work has been specifically focused
on the mereology of information entities [
Lastly, considering our principle of differentiation for ICEs can shed light on why designative ICE is not a parent of descriptive ICE. The purpose of descriptive ICEs is to describe a state of affairs rather than to designate or name an entity, even if descriptions often require designating entities. This is dissimilar from the relationship between descriptive and directive ICEs because the purpose of the latter is to describe some action(s) that ought to be carried out. Thus, the purpose of directive ICEs is partly descriptive in nature, which justifies directive ICE being a subclass of descriptive ICE. 4.2.
While every ICE will ultimately be an instance of designative, directive, or descriptive ICE, IAO can also retain many of the useful classes it asserts (such as document, textual entity, and document part) by asserting them as defined classes. In Table 1, we propose samples of axioms (class restrictions) that could be added to existing IAO classes to make them proper defined classes. The classes in this table are defined so that they are not included in the primitive class hierarchy, which removes crosscutting distinctions from it. We are neutral about whether the defined classes here are universals. Still, since they are defined classes, whether they correspond to universals will not infringe upon ontological realism. The list is not exhaustive and further work would be needed to make the necessary changes that would eliminate all the cases of cross-cutting distinctions within the primitive class hierarchy of IAO. Still, we hope that this provides an example for future modifications to IAO.
Formalized Axiom ∀(x)[IAO:Document(x) ↔ (ICE(x) ∧ ∃(y)(has_part(x,y) ∧ ICE(y)))] ∀(x)[TextualEntity(x) ↔ (ICE(x) ∧ ∃(y)(is_concretized_by(x,y) ∧ GlyphQuality(y)))] ∀(x)[Figure(x) ↔ (ICE(x) ∧ ∃(y)(is_concretized_as(x,y) ∧ 2DPattern(y)))]
In this paper, we identified and described the problem of cross-cutting distinctions within IAO. These distinctions make it difficult or impossible to extend IAO consistently, according to the principles of asserted single inheritance and ontological realism. For those adhering to these principles of ontology development, we proposed a modification to the top-level structure of IAO. Our proposal eliminates cross-cutting distinctions in IAO’s universal hierarchy, and it develops logical axioms for inferring poly-hierarchies where needed. While there is still ample work to be done, we hope that our proposal can be used to enhance the effectiveness of IAO.
We would like to thank David Limbaugh and Barry Smith for their feedback on this work, as well as four anonymous reviewers for the immensely helpful feedback they provided.