Ontology revision is the process of managing an ontology when a new axiom or fact would render it inconsistent. So far, the AGM approach to belief revision has been adapted to work with ontologies. However, when multiple sources are contributing uncertain knowledge about a static domain, an approach that doesn't give priority to incoming information and allows to recover previously discarded axioms is more suited. We describe an ontology revision framework that links symbolic and numerical techniques to allow the consistent evolution of an ontology from the contributions of multiple potentially unreliable sources.
An explicit specification of a conceptualization for a shared domain of discourse
is called an ontology[
Ontology evolution[
Current work on ontology evolution is based on the idea of bringing the
AGM belief change theory[
However, AGM belief revision is not apt to all kind of ontology changes. One of its principles states that incoming information has a priority: it must belong to the new set of beliefs. This principle works well when the new information represent a certain fact: either a realization of the new contingent state of the world, or a correction of a previous error in conceptualization, or a required property of the formalization. The principle can not be accepted when the new information represents a new evidence about the world, supposed to be a fixed static entity, while its description is only partial and uncertain. In particular, it can not be accepted in a distributed environment, where multiple potentially unreliable information sources are present. Not only an information from an external source can not be unconditionally accepted (can you trust everything you hear?); also, there is not always a relation between the arrival order of information and their acceptability.
There are many different possibility to discard the principle of priority to
incoming information. Hansson[
Any previously held piece of knowledge should belong to the current knowledge space if consistent with it.
To circumscribe the work, we will refer to a specific use case. A team of loosely-coordinated domain experts has the duty to build an ontology for their domain. Each team member contributes to the activity building his conceptualization with an editor. The domain is assumed as a fixed static entity, while the conceptualization is constantly changing during the building and refinement process. The work of each member is shared with the other experts in a peer-to-peer way: each member receives the contribution of the other experts. A supporting software must be able to use an ontology revision mechanism to maintain a consistent local ontology to be visualized and used as the basis for further editing. An example of a work session will be shown in section 5.
In the following we will first summarize in an informal way the syntax and semantic of the OWL ontology language (§2). Then, after an introdution to the problem of ontology revision (§3), we will show the proposed revision procedure (§4), both in its symbolical (§4.1) and numerical (§4.2) steps. Finally in §6 we sketch the future research perspectives. 2
The OWL web ontology language[
In order to introduce the problem of ontology revision and to make the work
self-contained, we will give here an informal definition of an ontology language
syntax and semantics, roughly correspondant to OWL Lite. The full formal
semantics and syntact can be found in [
The basic building blocks of an ontology are classes, individuals and properties. A class is related to a set of individuals, called class extension. Properties can be either data-valued, relating individuals to values, or individual-valued, relating individuals to other individuals.
An ontology is a set of class axioms, property axioms and facts.
There are two kinds of class axioms. A class can be defined as either exactly equivalent to the conjunction of a set of superclasses, or as a subclass of the conjunction of a set of superclasses. A superclass can be either another class, or an anonymous class specified giving constraints on properties.
The allowed restrictions on property values are: – all the values must be instances of a class (or from a datatype, in the case of data-valued properties); – some of the values must be instances of a class (or from a datatype, in the case of data-valued properties); – the cardinality must be at least (or at most, or equal to) either 0 or 1.
Property axioms are used to define properties. A property can be given a super-property, allowing the construction of a property hierarchy. Properties can also be given domains and ranges.
Data-valued properties can be specified as partial functional, i.e. with at most a value. Individual-valued properties can be specified to be functional, inversefunctional, symmetric, transitive, or the inverse of another property.
Finally, a fact states that an individual belongs to a class or that an individual’s property has a certain value. 2.2
An OWL interpretations defines: – a class as a collection of individuals, – a datatype as a set of literal values, – a data-valued property as a relation from individuals to literal values, – an individual-valued property as a relation from individuals to other individuals.
An interpretation I satisfies an ontology O if it obeys to all restrictions given by O’s axioms and facts.
An ontology O is consistent if there is at least an interpretation I which satisfies the ontology.
An ontology O entails an ontology O if each interpretation I which satisfies O also satisfies O . 3 3.1
Belief revision is the process of rearranging a knowledge base to preserve
global consistency while accommodating incoming information. In the AGM
theory[
Minimal change The revision process should alter as little as possible the current belief set.
Priority to incoming information The new information p must belong to the new belief set K∗p.
From these principles eight postulates follow. However, neither the rationality principles nor the postulates univocally define revision. 3.2
Ontology revision is defined as a change of components in ontology[
Our approach to ontology revision will be based on belief bases, a set of
sentences not closed under logical consequence, from which a belief set can be
derived[
The choice to represent changes at the level of single axioms is very
finegrained, but it doesn’t forbid to define more complex, higher-level changes[
To choose an ontology revision procedure we have first to understand why an axiom or fact, potentially incompatible with the current ontology, can arrive. We want to point out two different scenarios, demanding a different approach: – The ontology represents the current state of an evolving world, and the new information reflects a change in the world. The consequent change in the representation of the world is called updating. – We have an incomplete, approximate or erroneous representation of a static world. The new information represent a new evidence regarding this world. The consequent change in the representation is properly called revision.
In our scenario, a loosely-coupled group of peers are incrementally building
an ontology for a fixed domain. Thus, the world is not supposed to change, while
the world’s description is constantly evolving as the participants add, refine or
retract classes and properties definitions. This scenario is that of a revision
process and need to be handled within a framework possessing some specific
requisites. The need for those requisites already appeared in a juridical scenario
(incremental building of a proof in court[
Ability to reject incoming axioms. A belief revision system for a
multisource environment should drop the rationality principle of “priority to the
incoming information”, which is not acceptable since the sources are
asynchronous and there is no strict correlation between the chronological
sequence of information and their credibility or importance[
should be able to recover previously discarded pieces of the the ontology if new axioms redeem them. This should be done not only when the new axioms directly support previously rejected axioms, but also when they indirectly support them by disclaiming the axioms that caused their ostracism.
For these reasons we adapt to ontology revision a belief revision framework
that replace the priority to incoming information with the principle of
recoverability[
Belief revision has been approached both as a qualitative syntactic process and as a numerical mathematical issue. Our distributed ontology revision system links symbolic and numerical techniques. Computationally, the ontology revision consists of two steps acting on the axioms of the ontology, and three steps working with numerical weights.
Each peer stores his knowledge about the domain in at least two
repositories[
Given an incoming contribution p (an axiom or a fact) from a source, the evolution process consists of the following steps: 1. detection of minimally unsatisfiable subsets of KB ∪ {p}, called nogoods; 2. generation of the maximally satisfiable subsets of KB ∪ {p}, called goods; 3. revision of the credibility weights of axioms in KB ∪ {p}; 4. choice of a preferred maximally consistent subset of KB ∪ {p} as the new working ontology B ; 5. recalculation of “a posteriori” reliability of sources. 4.1
Step 1 and 2 are symbolical ATMS-style operations[
Nogood detection can be demanded to a reasoner, such as Racer[
An interesting property that the inference engine does not need to calculate the collection of all nogoods (i.e. minimally inconsistent subsets of KB), but just a collection of inconsistent subsets of KB, which is much easier. 4.2
The numerical approach to ontology revision deal with the ontology as a set of weighted axioms. Weights usually are reals between 0 and 1, representing explicitly the credibility of the axioms.
The numbers represent uncertainty caused by the not complete reliability
of the team members 3. As the reliability of the source is strongly related to
the credibility of the information, it is necessary to deal with couples source,
axiom [
Step 3 of the ontology revision process uses the belief function formalism,
as the one used by Shafer and Srivastava for auditing[
The recalculation of credibility values involves all the collected axioms in KB. The incoming axiom p is confronted not just with the current ontology B, but with all KB, so that the weight of axioms in KB ∪ {p} are reviewed in a broader and less prejudicial basis.
Step 4 is the selection of a new ontology B . The new ontology is the maximally consistent subset of KB ∪ {p} with the greater credibility. Since the incoming information causes a recalculation of all the credibility values, and the selected ontology is maximal, it is possible to rescue axioms from KB.
Even when the new contribution is compatible with the working ontology (meaning that B ∪ {p} is satisfiable), not necessarily B = KB ∪ {p}, since the global revision of numerical weight in step 3 may yield a totally different choice of ontology in step 4. A previously rejected set of axioms r can be rescued if p support r against a previously accepted set q.
In general, even when the new ontology B is syntactically equal to the previous B, meaning that p has been rejected, B may have a different credibility distribution (assignment of weights) from B. The incoming contribution p might be rejected even when a new ontology B , different from B, is selected, but B ∪ {p} is still unsatisfiable.
Step 5 uses Bayesian conditioning to determine the probability that a source give correct contributions, gives the new accepted ontology B . The main point is that a reliable source can not give false informations, while an unreliable source may occasionally give correct contributions.
As an alteration of the credibility of an axiom might result in the perturbation of the credibility of all the axioms from the same source, thus causing a completely different ontology to be selected at the next step. 3 Even the contribution from the agent self can be considered not completely reliable, as this depends of the relative trust a contributor has on his work compared to trust on other experts’ works.
We will show two examples, showing the symbolical and numerical steps respectively. In both, we suppose that a group of domain experts are working on an ontology of birds. 5.1
The initial knowledge base KB of one of those experts is made of the axiom Bird F ly and the fact Bird(T weety), where Bird and F ly are classes (the class of individuals that are birds, and the class of individuals that can fly, respectively), while Tweety is an individual. The knowledge base is consistent, so the initial ontology is B = KB.
The expert receive from a colleague (let’s call him source 1) the axiom ¬F ly(T weety). Now KB = {Bird F ly, Bird(T weety), ¬F ly(T weety)} is unsatisfiable. If we would adopt the AGM principle of Priority to Incoming Information, the new working ontology would be chosen among 1. B1 = {Bird F ly, ¬F ly(T weety)} 2. B2 = {Bird(T weety), ¬F ly(T weety)}
If we adopt the Principle of Recoverability instead, we have a third candidate working copy, 3. B3 = B = {Bird
F ly, Bird(T weety)} Next, another expert (let’s call him source 2) affirms F ly(T weety).
If we use the AGM principles, the new working ontology would be, respectively, 1. B1 = {Bird
source 1, 2. B2 = {Bird(T weety), F ly(T weety)}, if B2 was chosen.
F ly, F ly(T weety)}, if B1 was chosen after the input from
If we allow the rejection of the new contribution, after the arrival of the axiom from source 2, we can: 1. Reject the new axiom. Our working copy remain the same as after step 1. 2. Accept the new axiom.
(a) B1 = {Bird F ly, ¬F ly(T weety)}. We recover Bird(T weety), so
B1 = {Bird F ly, Bird(T weety), F ly(T weety)}. (b) B2 = {Bird(T weety), ¬F ly(T weety)}. We recover Bird F ly, so
B2 = {Bird F ly, Bird(T weety), F ly(T weety)}. (c) B3 = {Bird F ly, Bird(T weety)}. This is a simple expansion, so
B3 = {Bird F ly, Bird(T weety), F ly(T weety)}.
The example show that, if we consider the axiom F ly(T weety) more credible than ¬F ly(T weety), our final working ontology would be the same, independently from the choice made at the first step.
The initial knowledge base KB of one expert is made of the axiom Bird F ly and the fact Bird(T weety).
The expert receives from source 1 the axiom ¬F ly(T weety) and chooses as the new working ontology B2 = {Bird(T weety), ¬F ly(T weety)}.
Now suppose source 1 sends us the axiom ¬Bird(T weety). If we reject this
axiom, probably now our confidence in source 1 will be lower, as the credibility
of the information affects the reliability of its source[
A change on the credibility of an axiom provided by a source yields corresponding changes in the credibility of the other axioms provided by the same source, even if they are not logically related with each other. As a consequence of this perturbation, a completely different working ontology might be chosen, in the previous example B3 instead of B2, thus rejecting the previously accepted axiom from source 1. Since all the collected axioms are retained and their weights can change, the new selection might reconsider some previously discarded axiom, whether the incoming contribution is accepted or not.
Probably, the last come contribution decreases the credibility of the axioms it would render unsatisfiable, even in the case it has been rejected. The same when we receive an axiom which already belongs to the working ontology: it is not the case that nothing happened, as AGM fourth postulate of expansion would suggest[6, p. 49], since we are now, in general, more sure about the correctness of the axiom. 6
When a group of peer tries to capture in an ontology a static domain, but their domain’s knowledge is only partial and potentially unreliable, not all the contribution can be taken as unconditionally useful. It is necessary to use an ontology revision procedure that allows to discard the incoming information, if there is no reason to consider it more reliable than other conflicting contributions, and to rescue previously discarded axioms, if they are now compatible with the current selected ontology.
In general, at each step there will be more than a consistent subset of the ontology with maximal size (i.e., a good). There is the need of a rational criteria to choose a good as the new working ontology. If we keep track of axioms’ sources and give to each peer a a-priori reliability value, we can use the belief-function formalism to estimate the reliability of each good and bayesian conditioning to evaluate a new a-posteriori reliability value for the sources.
This work is just the beginning of an analysis of ontology revision process for a distributed environment. Current research work involves the following subjects. Collaborative Ontology Revision At the end of the work each expert has its own version of the domain ontology. To extract the final result of the collective work of the group of interacting experts, a voting mechanism is needed. The integration of the different conceptualizations must not be performed by an external supervisor, but it can be done by the group itself.
Ontology distribution To allow the distribution of individual fragments of the ontology, it must be possible to partition it and then reconstruct it preserving meaning.
For RDF, this bring to the definition of the minimal self-contained graph[
Given the OWL RDF/XML syntax’s use of blank nodes to build complex definitions, a similar concept can be applied to OWL. This decomposition allows the distribution of the ontology between peers, as in the scenario introduced in section 1.
User interaction A software supporting the collaborative building of an ontology must be able to use an ontology revision mechanism to maintain a consistent working ontology. Where inconsistencies arise and there is no other available ranking, the choice among different maximally consistent subsets can only be done by the user.
However, there are other times during the work when an user intervention would be useful. Why don’t allow the user to explicitly mark a part of the ontology as unreliable, not necessarily causing its deletion from the current working set, but determining a change in the distribution of reliability among the sources?
Explicit reliability judgments by an human agent must be taken into account when the system build a credibility ranking among the available sources. Strong time-Independence Even if the new information has no priority for his novelty, a complete independence of axiom’s weights from contributions’ arrival time is not guaranteed. This, given the asynchronous setting, would be a desirable feature of the system.
Ontology versioning Ontology versioning is defined as the ability to handle an
evolving ontology by creating and managing different variants of it[