Ontology alignment has been an active research topic for over a decade. Over that time, many developers have focused on creating alignment systems and methods to nd simple 1-to-1 equivalence matches between two ontologies. However, very few alignment systems focus on nding complex correspondences. There are several reasons for this limitation. First, there are no widely accepted alignment benchmarks that contain such complex relationships. Second, the traditional evaluation metrics like precision, recall, and f-measure are not accurate enough to evaluate the performance of a complex alignment system. And third, the approaches most commonly used to nd simple equivalences do not handle the increased computational complexity of nding complex equivalences well. Therefore, it becomes a big challenge for many developers to create and evaluate the systems. In this paper, in order to advance the development of ontology matching, we seek to address the problem by rst developing potential complex alignment benchmarks from realworld ontologies. In addition, we utilize traditional automated alignment systems to suggest complex correspondences, and nally plan to achieve our ultimate goal of creating and evaluating our own complex alignment system based on logical RDF data compression.
Similar to database integration, ontology alignment is an important process in enabling computers to query and reason across the many linked open datasets on the semantic web. This is a di cult challenge because the ontologies underlying di erent linked datasets can vary in terms of subject area coverage, level of abstraction, ontology modeling philosophy, and even language.
In order to solve this problem, an ontology alignment system aims to
identify the entity relationships between two and more ontologies. Such relationships
have a wide range of complexity, from basic 1-to-1 equivalence to arbitrary
mto-n relationships. However, over a decade, the majority of the studies in the
eld still focuses on the simplest end of this scale { nding 1-to-1 equivalence
relations between ontologies. Very few ontology alignment systems and methods
are developed to uncover complex relations. The reasons for this limitation may
lie in the following. First, there are no widely used and accepted ontology
alignment benchmarks that involve complex relations. Without these benchmarks,
even if there were a complex alignment system, it is not only very hard to
evaluate if this system could correctly detect the complex correspondences, but also
it is a challenge to evaluate whether the system is comprehensive enough to
nd all di erent kinds of complex patterns. Second, the traditional approach of
precision, recall, and f-measure does not seem ne-grained enough to evaluate
complex correspondences. A better version of precision and recall is needed [
This work seeks to progress in the direction of fostering the development of research activities in the eld of complex ontology alignment. We rstly show that real-world ontologies involve many complex relations. And based on these real-world datasets, we develop high quality complex alignment benchmarks, including creating complex alignments and categorizing them into complex patterns. In addition, to decrease the complexity of detecting complex relations, we leverage automated alignment systems to uncover and suggest possible complex relations. Moreover, we plan to apply logical RDF compression with the results that are generated by traditional automated alignment systems to create a new complex alignment system. 2
Ontology alignment seeks to address the conceptual heterogeneities between
ontologies. Over a decade, the community of this eld remains on creating and
improving the algorithms of nding simple alignments. The reason that the
researchers do not really dig into the complex alignment for such a long time, is
that the community is still discovering and analyzing how to reach the goal.
Nowadays, the research related to simple alignment has been well studied. It is
actually a good timing to move on to complex ontology alignment, because more
and more good alignment systems and algorithms have been published, and also
more and more data are populated into ontologies and published as linked open
data, the applications that utilize these LODs are required to involve ontology
matching and data integration processes [
This work focuses on addressing the problem from several di erent perspectives. We prepared benchmarks that involve complex relations from real-world ontologies and will try to distribute them as a new track in the Ontology Alignment Evaluation Initiative (OAEI), which was started in 2005 with the intent to allow researchers in the eld to compare the performance of their approaches on a consistent set of benchmarks over time. Since then, it has been more convenient for researchers from di erent organizations to test their methods. Second, as we've seen, it is a di cult challenge to detect complex relations. This work seeks to narrow down this issue by leveraging traditional alignment systems to suggest possible complex candidates. This would be a valuable starting point for determining the exact relation. Moreover, our ultimate goal in this work is to create a complex alignment system that has better performance than traditional automated alignment systems. 3
Regarding the creation of complex alignment benchmarks, Thieblin [
However, even though there are no widely accepted and used benchmarks
that involve complex relations, some researchers still tried to create alignment
systems and evaluate them using their ow manually developed reference
alignment. BLOOMS [
Our work di ers from the above methods in several aspects. First, we focus on real-world ontologies, which we found that these datasets are not only used by academic researchers, but also the industries and governments to develop applications for the usage of normal human life. There are some interesting relations that have not yet been mentioned in the current benchmark from OAEI. In addition, the instance data of these real-world ontologies are ready to be used as additional information to help improve the performance of alignment process. In contrast to this, signi cant instance data is not readily available for most of the OAEI Conference Track ontologies. Moreover, regarding the creation of a complex alignment, instead of comparing each entity in the source ontology to each entity in the target ontology, we apply logical RDF compression to list a set of available rules, and narrow down them based on the suggestion generated by traditional alignment systems to nally output the complex relation. More details are discussed in Section 5.
The research questions that we plan to address are listed as follows: 1. Do real-world ontology alignments contain complex relations? 2. How well do traditional automated alignment systems work on real-world matching tasks that contain complex relations? 3. Can we create an automated alignment system that performs better than traditional alignment systems on nding complex relations that exist between ontologies? Our hypotheses associated to the above research questions are the following: 1. Most real-world ontologies contain many complex relations. 2. Traditional automated alignment systems may not be able to identify complex relationships directly, but they may be able to suggest the atomic entities involved in such relations. 3. A complex alignment system that leverages logical RDF compression can e ectively identify complex relations between ontologies. 5
Hypothesis 1 Our previous work with the NSF EarthCube Initiative and the
US Geological Survey involved the time consuming task of manually aligning
several real-world ontologies. These alignments have been discussed and
evaluated by domain experts and ontology engineers to guarantee that they are of
high quality. We will inventory these alignments, along with any other real-world
alignments we can acquire, to answer the rst research question: Do real-world
ontology alignments contain complex relations?
Hypothesis 2 To answer the second research question, \How well do
traditional automated alignment systems perform on real-world matching tasks that
contain complex relations?", we plan to rst evaluate several state of the art
alignment systems on the alignment tasks mentioned above. Since traditional
alignment systems only attempt to identify simple relations between ontologies,
their performance will be limited to the percentage of the alignments that
involve these types of relations. However, it is possible that these systems, while
they cannot identify the precise relationship that holds between an entity in the
source ontology and two or more entities in the target ontology, they can at
least identify the entities involved in the relationship. For example, in the
relation below, the class M ischaakusaakihiikin in Cree ontology is equivalent to
the intersection of instances of LakeOrP ond and entities that isContainedBy
a SwampOrM arsh in the SWO ontology. While a traditional alignment
system cannot identify things like intersection or value restrictions, it may be able
to determine that LakeOrP ond, isContainedBy, and SwampOrM arsh are
related in some way to M ischaakusaakihiikin. To check this, for each entity es
in source ontology Os, we will use the automated alignment systems to give a
list of candidates et in the target ontology Ot, ordered by the similarity assigned
to them by the alignment system. We will evaluate the performance against the
benchmark using mean reciprocal rank [
Hypothesis 3 As we mentioned, the ultimate goal of this work is to see if
we can create an automated alignment system that e ectively identi es
complex relationships that exist between two ontologies. Our planned approach is
to create an extensional matcher (i.e. one that relies upon instance data) that
leverages logical RDF compression [
The overall work ow is shown in Figure 1. We rst apply the traditional
alignment systems to suggest the candidates as we described above. And then,
we use RDF compression [
In hydrography dataset, it consists of four ontologies. (a) The Surface
Water Ontology (SWO), which was originally presented in [
Ontology Classes Object Properties Data Properties GBO 40 149 49
GMO 156 124 46 is spoken by some of the native inhabitants of northern Canada. The reason of choosing these four ontologies is that Hydro3, HydrOntology and Cree have a large degree of overlap with SWO. Therefore, we utilize these four ontologies to create a reference alignment manually. Table 2 shows the number of classes, and object properties, and data properties. And we found that the 84 out of 197 relations in them are complex relations. We are currently writing a paper about evaluating the performance of traditional automated alignment systems on this dataset. The traditional automated alignment systems are able to nd the simple relationships. But, they may not be able to identify these complex relationships directly. We hypothesize that they may be able to suggest some possible complex relations partially. Moreover, it is able to greatly narrow down the entities and mitigates the high complexity of computation. In this section, we introduce the evaluation plan for each research question. For research question 1, as we showed in Section 6, it is considered successful that we have found many complex alignments in real-world ontologies. In addition, we are also preparing to incorporate the dataset into OAEI as a new track for other researchers accessing it. For research question 2, after achieving the list of entities involved in a complex relation, we will evaluate the performance against the benchmark using mean reciprocal rank as we discussed in Section 5. For research question 3, it is a challenge to evaluate the performance of a complex alignment system. The traditional precision, recall, and f-measure metrics do not seem ne-grained enough. For example, there is a relation between Hydro3 and SWO, if one alignment system identi ed this:
hydro3:Hydrographic_Feature hydro3:Hydrographic_Structure hydro3:Boundary) swo:HydrographicFeature)) and another identi ed this:
hydro3:Hydrographic_Feature hydro3:Island hydro3:Shore) swo:HydrographicFeature)) Based on the reference alignment, we need a metric to consider the rst system \more correct" than the second. We plan to develop a performance metric that more accurately re ects the performance of a complex alignment system. Another challenge is that, to the best of our knowledge, there are no existing complex alignment systems against which to compare our approach. Therefore, we might consider evaluating the performance based on our manually created reference alignment. 8
It is primarily di cult to identify complex relationships between ontologies because of computational complexity. A naive approach would need to compare every entity in the source ontology to every possible combination of entities in the target ontology, which is not feasible. Instead of doing this, our proposed approach has a good chance of success because it is based on a logical RDF compression method that has already been shown to be applicable to large datasets, and we also can further limit the search space by using the output from traditional alignment systems to narrow the focus. There are some re ections. The performance of using logical RDF compression in our alignment system is primarily based on the Abox information in the ontology. It is still not clear that how to apply our alignment algorithm to a more generalized scenario. However, our approach is feasible, and can be a good starting point to achieve the ultimate goal in the future.
Acknowledgments I am very grateful to my advisor Dr. Pascal Hitzler and Dr. Michelle Cheatham for their support and constructive suggestions of this research proposal.