DIOWG Subgroup on BFO 2020

The purpose of this page is to provide a record of the thinking and consensus of the BFO 2020 Subgroup of the DIOWG and serve as a reference for those looking to better understand issues surrounding relations in BFO.

The BFO 2020 Subgroup is tasked with reviewing new features in the upcoming version of BFO, BFO 2020 and recommending a policy on relation reuse to the DIOWG. To that end, we have been developing an understanding of the strengths and drawbacks of the temporalized relations strategy compared to the RO relations widely in use in the community.

Relations

This section describes the BFO 2020 Subgroup's consensus on the drawbacks of Temporally Unspecified Relations and Temporalized Relations. In Background, we'll provide some background on the general problem both strategies address and a quick overview of each type of relations.

Background

Many of the relationships we seek to model hold at some times, but not at others. For instance, someone may be a member of an organization for one year and not be a member of that organization before or after that year.

We've reviewed two strategies to represent these three-place reference relations using binary OWL relations: temporally unspecified relations and temporalized relations.

Temporally unspecified relations do not specify when the relation in question holds. The OBO Relations Ontology (RO) notably uses temporally unspecified relations:

A part_of B

A member_of B

We'll refer to this style of relation with the more general 'temporally unspecified' relation because many ontologies within the DoD/IC, such as the Common Core Ontologies (CCO) Extended Relations Ontology and several modules from the Defense Intelligence Core Ontology, use non-RO temporally unspecified relations as well.

The other strategy we've reviewed is temporalized relations (TRs). Temporalized relations specify whether the relation holds at all times or only at some time in the semantics of the relations:

A 'has part at some time' B

A 'has part at some time' B

Note that the at-all-times temporalized relations hold for all times at which A exists, not all times.

There is a design pattern that allows for the creation of more specific TRs (eg., 'part of in March 20203' or 'part of while machine is operating'), but for current purposes we'll restrict our attention to the generic TRs.

Temporally Unspecified Relations (RO-Style)

Temporal Ambiguity

The primary drawback of temporally unspecified relations is that they are massively ambiguous with respect to when they hold. Consider the following statement:

Liver01 part_of Lynn

If this appears in an ontology or dataset, it's impossible know whether Liver01 has been a part of Lynn as long as it has existed, whether it became of Lynn at some later point (say, through a transplant), or whether it is still part of Lynn.

This ambiguity poses a number of interoperability and reasoning problems, particularly when aggregating data from a number of sources:

• Failures of transitivity, which can introduce falsehoods during reasoning.

• Property chains will not work as expected, and can also introduce falsehoods.

The severity of these problems will depend on the use case. When aggregating data from multiple sources and using automated reasoning, especially data gathered at different times, these issues can be quite serious. On the other hand, there are common use cases for which these issues are trivial (e.g., where one is using ontology terms to specify a model or to tag data or when datasets are strictly segregated by temporal interval).

Stipulative Semantics for Subclass Axioms

RO Documentation says that class axioms using RO relations are to be interpreted as holding for "all snapshots in time". This is an attempt to capture permanent generic relatedness, but it's important to note that this stipulation is logically inert. Since the temporal aspect is not represented in the logic, it will have no effect on reasoning.

The stipulative temporal semantics also compounds the ambiguity of temporally unspecified relations. RO includes an annotation property used to indicate that a given axiom holds "at all snapshots". It's unclear then, whether only annotated axioms should be interpreted in this way or whether any RO subclass axiom should be interpreted this way. Similarly, temporally unspecified relations are not only used within RO. Its not clear whether these relations, especially those that descend from RO, should be interpreted to hold at "all snapshots", despite developers not having indicated this, or whether we should restrict this interpretation to RO.

Temporalized Relations

Current Defined Classes Cannot be Expressed

Temporalized relations pose problems for many defined classes currently in use. Consider the equivalency axiom using the temporally unspecified member_of:

Lawyer EquivalentClass ('agent' and member_of some 'bar association')

Neither the all-times nor the some-times temporalized relation will be appropriate for this axiom. The at some times version is too broad:

Lawyer EquivalentClass ('agent' and 'member-of-at-some-time' some 'bar association')

This axiom will include retired and disbarred lawyers (and even future lawyers) in the Lawyer class. The at-all-times axiom, on the other hand, will exclude lawyers:

Lawyer EquivalentClass ('agent' and 'member-of-at-all-times' some 'bar association')

Since no lawyer has always been a member of a bar association, all current lawyers will be excluded from the Lawyer class.

Note: This is also an issue for non-rigid primitive classes, but as these are typically avoided in favor of defined classes within the IC/DoD, we focus here on defined classes.

No General Use Transitive Parthood

The three-place parthood relation, between a whole, a part, and a time, is transitive. The part-of-at-all-times relation is transitive, but the part-of-at-some-time relation is not. In many cases, part-of-at-all-times will be too strong, since objects gain and lose parts. There is no general-use temporalized relation that is both transitive and appropriate to describe parthood relations that hold for a limited period of time.

It is possible to make a custom transitive temporalized relation that holds only in the time interval of interest ('part of in March 2023'), but these will be limited in their applicability and reuse.

Additional Complexity

Temporalized relations introduce complexity to ontologies and ontology workflows. First, when writing axioms and property chains, developers must choose the appropriate TR. This complexity is compounded when there are more than one temporalized relation in an axiom or property chain. Developers must then choose a TR for each relation in the axiom or property chain, and ensure that, taken together, the axiom has the intended interpretation. This can often be quite tricky. This complexity is an obstacle for ontology workflows that involve having non-ontologists input axioms and property chains, and will require a substantial amount of additional training and oversight.