-Probabilistic ontologies incorporate uncertain and incomplete information into domain ontologies, allowing uncertainty in attributes of and relationships among domain entities to be represented in a consistent and coherent manner. The probabilistic ontology language PR-OWL provides OWL constructs for representing multi-entity Bayesian network (MEBN) theories. Although compatibility with OWL was a major design goal of PR-OWL, the initial version fell short in several important respects. These shortcomings are addressed by the latest version, PR-OWL 2. This paper provides an overview of the new features of PR-OWL 2 and presents a case study of a probabilistic ontology in the maritime domain. The case study describes the process of constructing a PR-OWL 2 ontology using an existing OWL ontology as a starting point.
Keywords- Probabilistic ontology, Multi-Entity Bayesian networks, PR-OWL, OWL, Maritime domain ontology, Uncertainty Modeling Process for Semantic Technologies
I.
INTRODUCTION
The emphasis on net-centric operations and the shift to
asymmetric warfare have created new challenges for automated
information integration. To meet these challenges, developers
are recognizing the need to combine explicit representation of
domain semantics with the ability to represent and reason with
uncertainty. Probabilistic ontologies allow the representation of
uncertainty about attributes of and relationships among domain
entities. Probabilistic OWL (PR-OWL) [
This research was partially supported by the Office of Naval Research (ONR), under Contract N00173-09-C-4008. Rommel Carvalho would like to thank the Brazilian Office of the Comptroller General (CGU) for their active support since 2008 and for providing the human resources necessary to conduct this research.
connection between a statement in PR-OWL and a statement in OWL is not formalized, it is unclear how to perform the integration of ontologies that contain statements of both formalisms.
Carvalho [
This paper presents an overview of PR-OWL 2, describes the key features that improve compatibility with OWL, discusses an open-source tool for building PR-OWL 2 probabilistic ontologies, and describes a use case of a PR-OWL 2 ontology for maritime domain awareness.
II.
A. PR-OWL 1: An Upper Ontology for MEBN Theories
PR-OWL provides constructs to define probabilistic
ontologies in the OWL ontology language. The initial version,
PR-OWL 1, is an OWL upper ontology for representing
MEBN theories [
As an example of a PR-OWL ontology, Figure 1 shows a
simple probabilistic ontology developed as part of the
PROGNOS (Probabilistic OntoloGies for Net-centric
Operation Systems) project [
Overall Goal: Identify whether a ship is a ship of interest, i.e. if the ship seems to be suspicious in any way. 1.
that the ship is meeting some other ship for no apparent reason.
Query: Does the ship seem to exhibit evasive behavior?
responsive radar and automatic identification system (AIS).
Each of the nine MFrags of Figure 1 addresses a modular component of the knowledge needed to address the above queries. Specifically, probabilistic knowledge about hypotheses related to the identification of a terrorist crewmember is represented in the HasTerroristCrew, TerroristPerson, and ShipCharacteristics MFrags. Knowledge about unusual routes is represented in the UnusualRoute and Meeting MFrags. Finally, knowledge about hypotheses related to evasive behavior is represented in the EvasiveBehavior, EletronicsStatus, and Radar MFrags.
A detailed explanation of this model can be found in [
C. An Open Source Tool for Probabilistic Ontologies
The MFrags shown in Figure 1 are screenshots from the UnBBayes-MEBN [8], an open source, plug-in-based Java application for building and reasoning with probabilistic ontologies based on the PR-OWL/MEBN framework. 1 It features a graphical user interface (GUI), an application programming interface (API) for saving and loading PR-OWL ontologies, reasoning algorithms for processing queries, and plugin support for extensions.
Queries are processed in UnBBayes-MEBN using an
implementation of the situation-specific Bayesian network
(SSBN) construction algorithm described in [
III.
Ideally, it should be possible to use PR-OWL to reason probabilistically about uncertain aspects of an ontology based on the information already available. That is, we would like to 1UnBBayes is available from http://unbbayes.sourceforge.net/ be able to begin with an OWL ontology containing information about a domain, use PR-OWL to define uncertainty about attributes of and relationships among the entities, and apply a probabilistic reasoner to reason with available evidence. For example, we might begin with an OWL ontology containing classes for ships, routes, persons, and other entities mentioned in the MFrags of Figure1. We would then wish to use PR-OWL to define the probability distributions represented in the MFrags.
The difficulty with this idea is that PR-OWL 1 has no mapping between the random variables used in PR-OWL and the properties used in OWL. For example, suppose we have defined an OWL class Ship with property isShipOfInterest, intended to represent whether a ship is a ship-ofinterest. We might want to use the PR-OWL random variable isShipOfInterest(ship) to define the uncertainty associated with this property. We might use the ShipOfInterest MFrag of Figure 1 to specify its probability distribution. However, despite the syntactic similarity between the property name and the random variable name, PR-OWL 1 has no way to specify formally that the random variable isShipOfInterest(ship) defines the uncertainty of the OWL property isShipOfInterest. Thus, even if we had information about whether a particular ship, say Ship379, is a ship-ofinterest, we would not be able to instantiate the random variable isShipOfInterest(ship) for Ship379.
Poole et al. [
The key to building the bridge that connects the deterministic ontology defined in OWL and its probabilistic extension defined in PR-OWL is to understand how to translate one to the other. On the one hand, given a concept defined in OWL, how should its uncertainty be defined in PR-OWL in a way that maintains its semantics defined in OWL? On the other hand, given a random variable defined in PR-OWL, how should it be represented in OWL in a way that respects its uncertainty already defined in PR-OWL?
PR-OWL 2 formalizes the relationship between OWL
properties and PR-OWL random variables using the relation
definesUncertaintyOf [
Another major difference between PR-OWL 1 and PROWL 2 is that the separate definition of entity in PR-OWL is replaced by OWL’s built-in notion of classes and data types. That is, a PR-OWL entity is now identified with either a class or a data type in OWL. Moreover, since OWL supports multiple inheritance, so does PR-OWL 2. Thus, all the control over the type definition and type hierarchy in PR-OWL is delegated to OWL.
In PR-OWL 2, therefore, the possible values or outcomes of a random variable are instances of classes and data types. When specifying that a random variable will have individuals of a class as its possible outcomes, it is reasonable to assume that all known individuals of that class form a set of collectively exhaustive outcomes. However, the assumptions about individuals in OWL are not enough to guarantee these individuals are mutually exclusive. More specifically, although OWL provides a way to express unique names, it also allows two different names to point to the same object in the real world. To address this issue, PR-OWL 2 follows the MEBN and PR-OWL 1 convention, and assumes that every individual has a unique ID associated to it.
We note that there are certain aspects of the full PR-OWL semantics that are not fully captured in OWL-DL, and therefore cannot be handled by OWL-DL reasoners, but are expected to be respected by PR-OWL reasoners. In particular, to specify the restriction that a random variable defines the uncertainty of a property would require OWL Full. For this reason, the restriction is not explicitly represented in PR-OWL, but it is expected to be enforced by a PR-OWL probabilistic reasoner. This enables consistency checking of the deterministic part of a PR-OWL ontology using a DL reasoner.
PR-OWL 2 CASE STUDY
The following case study demonstrates the application of probability to an existing ontology to represent uncertainty in knowledge about instance attributes. In this case, an existing ontology of Western European warships identifies the major characteristics of each combatant class through the attributes of size, sensors, weapons, missions, and nationality. Figure 5 shows an entity-relationship diagram for the ontology. The decision maker is trying to determine the warship class of a contact about which he has limited information. By adding probability to the existing ontology, we can identify the most likely class of ship he is encountering when provided only partial or uncertain information. The model is designed to answer the following query using the following evidence: Overall Goal: Given uncertain or absent attribute information about a specific ship, what is the most likely European warship class that satisfies these attributes?
The entity-relationship diagram of Figure 5 presents a simplified design of the Military Ship Ontology illustrating the primary attributes used to answer these queries. The decision maker desires to know the class of warship that he faces. A class of ships has a consistent hull design and a standardized suite of weapons and sensors. These weapons and sensors work in concert to provide synergy in executing the primary mission of each type of ship. By combining a ship type with the nation that operates it, a logical prediction of warship class may be obtained.
International law of the sea requires that each merchant ship is registered and sails under a single nation for the purpose of regulation, certification, and pollution control. That process is known as flagging, and an individual ship is flagged by a nation. It is not required that a ship is flagged under the same nation as its owner; a “flag of convenience” allows a ship to be operated under an alternate nation to reduce operating costs and regulations. However, warships are always flagged under the nation of ownership.
The Gross Naval Class is a naval schema that delineates warships from merchant ships, and is mutually exclusive. Through identification of weapon and sensor attributes, as well as overall ship size, a Gross Naval Class estimate may be made for the unknown ship. While it can be assumed that all ships have a radar sensor, only military ships have sensors associated with weapons systems. The presence of a weapon system, or a weapon-associated sensor, provides reasonable evidence that a ship is a warship.
Warships are of different types based on their primary mission. Most ships have multiple mission capabilities, but for this ontology we assume the following primary mission areas by ship type:
− Aircraft Carrier (CV, CVN) − Cruiser (CG) − Guided Missile Destroyer (DDG) − Guided Missile Frigate (FFG) Anti-Surface Warfare (ASuW):
− Destroyer (DD) Anti-Submarine Warfare (ASW):
− Frigate (FF) By observing the combination of weapons and sensors, it is possible to infer the most likely mission area. This, combined with an estimate of ship size, provides an indication of the type of warship.
At this point an MTheory is created to determine hasWarshipClass(ship) in the WarshipClass MFrag for some unknown ship. The eight MFrags associated with this determination are shown in Figure 6. Inputs to hasWarshipClass RV are the RVs from the WarshipType and Nationality MFrags, representing the concepts introduced above with the RVs hasWarshipType(ship) and hasFlag(ship). The WarshipType MFrag may be further decomposed into the ShipSize, GrossNavalClass, and PrimaryMission MFrags. The GrossNavalClass MFrag is influenced by both the ShipSize and ShipSensor MFrags through the hasShipSize(ship) and hasSensor(ship) RVs, while the PrimaryMission MFrag is influenced by the ShipSensor and ShipWeapon MFrags with hasSensor(ship) and hasWeapon(ship) RVs. With the MTheory complete as shown in Figure 6, the Local Probability Distribution (LPD) must be populated.
Prior probabilities for the hasFlag RV were obtained from an estimate of merchant ship registrations available through open source information. Similarly, hasShipSize represents a finite and exhaustible set of ship lengths (LengthLess150m, Length150to100m, LengthGreater200m) into which each ship is categorized. Prior probability estimates were again obtained via open source literature. Priors for hasSensor and hasWeapon were obtained through subject-matter-expert review of open source literature and represent the proportion of warships with each of the types of sensors. LPDs for the GrossNavalClass and PrimaryMission MFrags require conditional statements about relationships from the input nodes shown in Figure 6. A detailed description of these relationships is described in a forthcoming paper.
Queries to the Military Ship Probabilistic Ontology are processed in UnBBayes-MEBN using an implementation of the situation-specific Bayesian network (SSBN) construction algorithm. Instances of unknown ships and representative evidence are entered via the OWL ontology through the UnbBayes GUI to reflect partial or uncertain information about ship attributes. These are checked against known characteristics provided by subject-matter experts.
For example, suppose the following evidence is obtained about a ship of interest: • • • •
UID: Surcouf hasNavalGun(Surcouf): True hasFlag(Surcouf): France hasShipSize(Surcouf): <150m
Executing a query of the isWarshipClass node produces the SSBN found in Figure 7. In this case, there is a 68% chance that Surcouf is a member of the French LaFayette Class of frigates, which is the correct classification.
As discussed in Section III, our goal is to begin with an OWL ontology containing information about a domain, use PR-OWL to define uncertainty about attributes of and relationships among the entities, and apply a probabilistic reasoner to reason with available evidence. Using the formalized construct introduced in PROWL-2, we map each of the RVs in the MFrags of the probabilistic ontology to the existing OWL property in the original ontology. This is accomplished through the probabilistic ontology building sequence executed on the UnbBayes software. For example, the WarshipType class in OWL has an object property of hasPrimaryMission. This object property is mapped to the hasPrimaryMission(ship) RV of the PrimaryMission MFrag. Mappings produced for each RV and its associated property in OWL allow us to use PR-OWL to reason probabilistically about uncertain aspects of an existing ontology based on the information already available.
Combining uncertainty reasoning with semantic technology is necessary for robust, interoperable, net-centric fusion and decision support systems. The probabilistic ontology language PR-OWL provides a way to represent and reason with probabilistic ontologies. PR-OWL 2 improves compatibility with OWL in several important respects. Through a case study, this paper describes the construction of a probabilistic ontology obtained by enhancing an existing OWL ontology with probability information.