In today's semantic web, ontology fragmentation and modularization are considered as important tasks due to the size and complexity of prime ontologies. At the same time, the ontology alignment community thrives with solutions for discovering and producing alignments between semantically related concepts. However, these are seldom used in bulk and their subsequent re-use is somewhat problematic. In this paper we set to explore these issues from a practical viewpoint: use ontology alignments to inform an ontology fragmentation strategy for the bene¯t of exposing and distributing rich ontology aligned fragments.
Semantic integration and interoperability solutions are becoming increasingly
important in dynamic and distributed environments where information
originates from various stakeholders. Ontology mapping is seen as one of the core
technologies in this space [
One of the reasons for this phenomenon is the relatively high computational cost for processing ontology alignments. W3C's omnipresent owl:sameAs statement is only a starting point for what it turns out to be a high computational cost and time consuming reasoning task to be undertaken by a DL reasoner. That is because the owl:sameAs statement merely re°ects the logical equivalence of the two referenced terms, and thus points to a possible semantic similarity, but does not provide any further information about their provenance and semantics. To take full advantage of the semantic similarity we need to consider more than just reporting the factual equivalence between two terms in an application.
Some proposals to overcome the high computational cost and high latency in performing an inference cycle, is to fragment or modularize the original ontologies. Others are considering enhancing the existing OWL vocabulary with richer 3 An up-to-date overview of the ¯eld is re°ected in http://www.ontologymatching.org/ 4 See, for example, the SKOS vocabulary: http://www.w3.org/TR/skos-reference/ constructs that enable more provenance information and semantics to be exposed when semantic similarity is reported. However, these are heavy weight approaches that pose certain assumptions on the domain and applications.
We advocate a practical and ready-to-use with current technology solution to this problem: use existing alignments and OWL taxonomic reasoning to identify fragments from the original ontologies that capture the immediate provenance and semantics of the aligned terms. We then propose to extract those fragments using standard W3C rules and query technology that is easy to reuse and replicate in di®erent scenarios. The extracted fragments are bundled together in self-contained and well de¯ned portable OWL fragments. These can then be accessed and re-used at a lower cost than that of accessing and re-using the original ontologies.
In the next section we elaborate on the requirements for operational ontology fragments drawn from our experience in an international research collaboration programme (section 2). Our proposed Portable Ontology Aligned Fragments (POAF) work is described in section 3 along with a concise example case (section 3.1). We report on related work in section 4 before concluding the paper in section 5. 2
The work described in this paper forms part of the International Technology Alliance project Semantic Integration and Collaborative Planning5. The overall aim of this project is to enable the integration of heterogeneous and physically distributed information content in semantically-coherent ways. This need arises from the information demands of coalition operations. A key concern in these operations relates to the physical distribution and semantic heterogeneity of relevant information content: physical distribution makes information di±cult to search, retrieve and manage, while semantic heterogeneity makes information di±cult to integrate and understand.
Semantic integration and interoperability solutions have been studied and
applied to a variety of domains [
One way of achieving this, is to fragment the original knowledge assets involved in a semantic integration solution. Our experiences from ontology mapping shows that only a small fraction of the referenced ontologies is used in the produced alignments. That opens the road for an ontology alignment informed 5 More information available from: http://usukita.org/ fragmentation task that meets the strict operational constraints in a coalition context. In this scenario, the ontology alignments become the focal point, both as enablers of semantic integration solutions as well as triggers for fragmenting the original knowledge assets into meaningful and semantically coherent chunks of knowledge. These semantically coherent fragments will mirror the original ontologies but they are smaller in size and complexity and thus, easier to process in a operationally useful time frame.
In the next section we describe a novel mechanism for extracting those fragments from ontologies using ontology alignments as the trigger. 3
POAF aims to increase the usability, tracking of provenance, and portability of ontology alignment products (typically a number of owl:sameAs statements) among interested parties. It is a post ontology alignment process. We assume that an ontology alignment or mapping tool has been executed and delivered a set of ontology alignments using the W3C's standard notation: owl:sameAs, owl:equivalentClass and owl:equivalentProperty. These notations enable us to indicate semantic similarity of two OWL constructs (ranging from individuals to classes and properties, depending on the type of OWL language used). When they encounter owl:sameAs or similar statements, automated reasoners make use of that information and access the aligned terms in order to complete their reasoning process. However, there are issues with this approach which call for a lighter and more e±cient way of sharing aligned terms: { Availability and access of the aligned OWL ontologies: if the ontologies where the owl:sameAs referenced terms belong are not accessible (i.e.: due to network outage, bandwidth restrictions or interference concerns in a battlespace operational context) or not available at the time the reasoner tries to conduct its inference cycle, this could cause a break in the reasoning and consequently bring the inference process to a halt, causing the reasoner to abandon this task (some advanced DL reasoners could resume at another point but the particular inference will not be concluded); { Unnecessary reasoning steps: when a reasoner visits the ontologies where the aligned terms originate from - depending on the type of reasoning task - it is likely that unnecessary crawling of the OWL ontologies will occur. Even if the task is to simply resolve the name of the aligned concept, visiting the originating ontologies will add unnecessary time to the processing task; { Fragmented and distributed factual knowledge base: When a reasoner tries to perform a task where multiple owl:sameAs statements are involved and point to a number of di®erent ontologies, the reasoner will have to collate information from di®erent URI addresses. Although today's DL reasoners can cope with this task, it is an unnecessary resource load for the system and could a®ect its performance; { Di±cult to inspect and track provenance information: When a number of owl:sameAs statements are used to convey semantic similarity information for the aligned concepts, it is di±cult to track their provenance. This is becoming increasingly di±cult as more and more ontologies are involved. At the very least it is often not feasible to inspect the origin of the aligned concepts using eyeball checking in a casual fashion. It is likely that an engineer will need to employ a reasoner to do this task, which brings us back to the problems mentioned before: that of unnecessary reasoning steps, overload of time and bandwidth resources and knowledge base fragmentation.
Based on these observations, and our experiences with designing, developing,
deploying and using ontology mapping systems [
We implemented these steps in a process to operationalise the idea of POAF. We depict this process in Figure 2.
The ¯rst step of the POAF generation process involves the capture fragment task: we use OWL semantics and the ontology subsumption relations to identify related fragments. These are fed into the second step, Use W3C technology to extract fragment, where the actual extraction occurs. We opted for W3C technology so that we increase our interoperability potential with other tools and technologies. We use SPARQL queries, SWRL rules and HP's Jena Framework. The next step is the Bundle Fragment step. The aim of this task is to construct well formed OWL fragments so that DL reasoners and external tools can use them as a substitute for the original ontologies, depending on the task and scope of application. Finally, the last step is to provide as much provenance information as possible in the Provenance aggregator task. In this task we collate the original namespaces with POAF speci¯c ones to distinguish between declared and inferred statements in the POAF ¯les. 3.1
An example case POAF, albeit an early stage of development, has been deployed in experimental scenarios in the context of the "Semantic Integration and Collaborative Planning" project of the International Technology Alliance programme. In one of these scenarios, we aligned two OWL ontologies: terrorism.owl6 and tkb.owl7 for the sake of improving interoperability between agents using these ontologies in a common task context. We used an ontology mapping tool to discover potential alignments between the two ontologies. Although this step is not part of POAF, for the sake of completeness we provide brief information on the process: we chose to work with CMS (CROSI Mapping System), an open source comprehensive ontology mapping system8 . We include in the screenshot below, small fragments of the original ontologies and one of the proposed alignments, as delivered by CMS in owl:sameAs format:
We set the CMS system to the following matching criteria: use 1 matcher (StructurePlus, an algorithm that uses ontology structure information along with linguistic features), no weight tuning up (not applicable in our case as we used only 1 matcher), output the top 20% of results only and use OWL format for the output9 .
Going back to the steps outlined before, we now need to "Capture a fragment of the ontology that is directly relevant to the aligned term ". We chose the following OWL constructs to be directly relevant to any aligned term: subclasses, data and object properties. We identify subclasses by default taxonomic reasoning whereas for object and data type properties we take into account the domain and range type of the property. If the value of the rdfs:domain and rdfs:range matches the aligned class, then we earmark this property for inclusion in the POAF structure. This basic information is deemed to be enough to 6 This is an OWL ontology that describes concepts in the terrorism domain, available from http://counterterror.mindswap.org/2005/terrorism.owl 7 this is an OWL ontology in the terrorism domain that describes main concepts and facts on terrorism attacks. Available on request. 8 available from http://sourceforge.net/projects/ontologymapping/ 9 more on the CMS system and the alignment algorithms it uses can be found here: http://www.aktors.org/crosi/ capture the direct provenance of the aligned term and eventually will constitute a POAF for that term. Note that we do not aim to capture a comprehensive fragment that can act as a substitute for an entire ontology: POAF structures should only be used as elaborate summaries of aligned fragments that include their directly related constructs (immediate subclass and their properties). We chose not to include immediate superclasses, if they exist, due to the intuitive denotational semantic distance: superclasses are, in general, more abstract and this could introduce semantic discrepancies in the POAF structure. On the other hand, a direct subclass is a safer option as subclasses are, in general, more speci¯c and, intuitively speaking, we expect them to be semantically subsumed by the aligned class.
After we decided what to include in a POAF structure, the next step is
to: "Use the aligned term, and OWL semantics, to identify and capture those
fragments ". At this stage we had several options for extracting fragments from an
ontology. As our work is not directly targeting ontology modularization, per se,
but rather an informed selection of ontology constructs, we decided to stray away
from graph traversal techniques used by ontology modularization practitioners
[
We see rules technology as the best candidate here, as rules can be edited, re-de¯ned and re-used in di®erent settings. Also, W3C already backs certain semantic web rules technology: SWRL and RIF, most notably. Another option here, is to use query technology, especially SPARQL, as it meets the criteria set for rules (editable, re-usable) and enjoys strong backing from the W3C standards community.
To satisfy these requirements, we opted for a solution where we use a combination of SWRL rules and SPARQL queries to identify and capture the fragments. We decided to provide both as SWRL captures the logic for identifying the fragment, but its executability is cumbersome with only a handful of theorem provers10 and closed environments o®ering SWRL reasoning11; SPARQL on the other hand, provides the functionality we are after in a semantic query format using the CONSTRUCT clause. In addition to this, SPARQL queries are easily re-usable in di®erent settings.
To illustrate the result of our "capturing the fragments" task, in Figure 3 we include a screenshot of the eligible constructs from the original ontologies.
As we can see, both terrorism.owl:City and tbl.owl:City have object properties declared: locatedInCity for terrorism.owl:City and incidentOccuredHere for tbl.owl:City. Also, terrorism.owl:City has two subclasses: Town and Municipality whereas tkb.owl:City has CapitalCity as a subclass. Note that these are only a fraction of the original object property and subclass declarations 10 like the Hoolet theorem prover: http://owl.man.ac.uk/hoolet/ 11 like the Protege SWRLTab: http://protege.cim3.net/cgi-bin/wiki.pl?SWRLTab due to space limitations. This is the basic information (subclasses, properties) we aim to capture in order to populate the POAF structure for this alignment.
We do that by using SWRL rules and SPARQL CONSTRUCT queries. For example, to capture object property information we use the SWRL rule depicted in Figure 4, which can also be expressed as a SPARQL CONSTRUCT query, depicted in Figure 5.
Note that the query above captures only the rdfs:domain information of an object property statement; a similar query is used for rdfs:range statements.
The next step in our process is to: "Bundle these fragments in OWL compatible snippets of code". To do that we simply execute the SWRL rule over the original ontologies or ¯re the SPARQL query via a custom built endpoint. This will capture the information we want to include in the POAF structure (subclass, properties). The results are depicted in Figure 6.
As we can see, class Municipality from terrorism.owl has been associated with class City from tkb.owl. Similarly, class Town from terrorism.owl has been associated with City from tkb.owl. On the other hand, class Capital City from tkb.owl has been associated with class City from terrorism.owl. Similar correspondences between object properties have been produced: the IncidentOccuredHere object property from tkb.owl is now associated with class City from terrorism.owl and locatedinCity from terrorism.owl is associated with class City from tkb.owl. Note that all of the original declared associations of subclass and object properties are retained for semantic consistency. These associations are driven by the ontology mapping results of terrorism.owl:City being declared as owl:sameAs to tkb.owl:City, and it is legitimate to assume that they share the same properties and subclasses. The POAF structure depicted in Figure 6 merely re°ects this and we built it automatically.
The ¯nal step in the POAF process is to: "include as much provenance information as possible in these fragments, so that tracking and tracing of the original ontology is feasible.". This is essentially a safeguarding mechanism against unwarranted inferences from reasoners attempting to parse the POAF structure. Currently, we use the standard rdfs:subclassOf construct to denote subclass relationship. But we also store the POAF inferred subclass relationships with a designated poaf:inferred construct to distinguish between declared and inferred information. This will ease tracking and evaluation of POAF structures. We are also planning to collect and package all the original namespaces in a designated construct so that we facilitate the tracing of provenance information for bulk processing of POAF structures. The current setting in the POAF work can be enhanced with the provision of a customized front end for selecting the ontology constructs to be extracted and included in a POAF structure. For example, our default option is to extract the direct subclass of the aligned class. However, depending on the application scenario and inference needed, one might need to extract all the subclasses of the aligned class or even some or all of the super-classes. This sort of custom tuning of the taxonomy hierarchy is feasible with advanced Semantic Web APIs, like HP's Jena Framework12. Using a customized extraction mechanism we should be able to extract any kind of ontology construct that is somehow related with the aligned class and goes beyond the direct subclass and object/data type property as is the default setting.
We are planning to work on a meta-OWL extension for expressing subclass information that is inferred by the POAF structure. Currently, we simply store such information with a designated construct (poaf:inferred). But eventually, we would like to replace the standardized rdfs:subclassOf with an equivalent construct to distinguish between declared and inferred information. Any extension of that sort will have to be OWL compatible and parseable by standard DL reasoners, however.
Finally, we also plan to work on the SPARQL front by distributing the generated SPARQL CONSTRUCT queries which encapsulate both the fragment extraction logic but also the original semantic similarity information. Currently, we use those in a sandbox application where the POAF program initiates the process. In a distributed scenario though, we want to achieve the same functionality with a series of SPARQL CONSTRUCT clauses that capture the desired information, and build the POAF structure. 4
POAF is in principle an ontology fragmentation method. However it is distinctly
di®erent from typical fragmentation techniques in that it makes use of a strong
pre-requisite condition for enacting fragmentation: the availability of ontology
alignments. Another di®erentiating factor is that POAF aims to increase the
potential for sharing, using and distributing ontology alignments across
multiple systems in a dynamic networked environment. The ontology fragmentation
aspect is a side e®ect of this functionality and we do not intend to use it as its
primary function. Taking into account these distinct characteristics of POAF we
identify the following related research: Euzenat and colleagues use the notion of
ontology alignments to inform the syntax for expressing ontology modules [
Similarly, taxonomic reasoning is deployed in [
Others have argued that relaxing the subsumption dependencies between
ontology constructs can improve the expressivity of modular languages [
In [
In this paper we proposed a light weight mechanism for sharing and distributing enhanced ontology alignment information. We make use of the generated alignments between two ontologies and we expose their immediately related terms in a well de¯ned and concise OWL structure, which we dubbed Portable Ontology Alignment Fragments (POAF). Our work taps into the area of ontology fragmentation, but is not directly related to a fragmentation strategy. Rather, we aim to highlight the importance of distributing more information when reporting semantic similarity between two aligned terms. This information could enhance uptake and re-use of ontology alignments, We also contribute to the adoption of semantic web technologies in scenarios where operational time constraints and resource overheads are important considerations. A POAF based solutions aims to reduce the time needed to process ontologically aligned structures as it encapsulates the aligned concepts and their immediately related concepts in small and easily manageable fragments.
POAF can also be used as a quality assessment tool for ontology alignment. Since the generated POAF structure exposes the related terms of the aligned terms (subclass and properties), this information can be used to semantically sanitize and proof-check the proposed alignment. This process could be a manual or a semi-automated one depending on the assessment task.
This research was sponsored by the US Army Research laboratory and the UK Ministry of Defence and was accomplished under Agreement Number W911NF-06-3-0001. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the o±cial policies, either expressed or implied, of the US Army Research Laboratory, the U.S. Government, the UK Ministry of Defence, or the UK Government. The US and UK Governments are authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation hereon.